i=1
(bt
2/5
i
r
i
)
2
.
Take the derivative with respect to b and set it equal to zero to get
b =
r
i
t
0.4
i
t
0.8
i
= 569.5695.
Then the energy in kilotons is
E =
1
4.186 10
12
b
5
= 17.89.
1
2 CHAPTER 1. SCALING, DIMENSIONAL ANALYSIS (SECS. 1 & 2)
Another method is to average. We have
E =
r
5
t
2
.
Substitute each data point to get
E
i
=
1.25r
5
i
t
2
i
, i = 1, 2, . . . , 8.
Now average the E
i
and divide by 4.186 10
12
to get E = 18.368.
4. The variables are t, r, , e, P. We already know one dimensionless quantity
1
= r
5
/et
2
. Try to nd another that uses P, which is a pressure, or force
per unit area, that is, mass per length per timesquared. By inspection,
2
=
P
gr
is another dimensionless quantity. Thus we have
f(r
5
/et
2
,
P
gr
) = 0.
Now we cannot isolate the r and t variables in one dimensionless expres
sion. If we solve for the rst dimensionless quantity we get
r
5
/et
2
= F(
P
gr
).
Then
r =
_
et
2
_
1/5
F(
P
gr
).
Because the second dimensionless variable contains r in some unknown
manner, we cannot conclude that r varies like t
2/5
. However, if one can
argue that the ambient pressure is small and can be neglected, then we
can set P = 0 and obtain the result
r =
_
et
2
_
1/5
F(0),
which does imply that r varies like t
2/5
.
5. If x =
1
2
gt
2
, then = x/gt
2
is dimensionless and the physical law is
=
1
2
. If we include mass, then m must be some function of t, x, g, which
is impossible.
6. If x =
1
2
gt
2
+ vt, then, by inspection, y = x/gt
2
and s = v/gt are
dimensionless. Dividing the equation by gt
2
gives the dimensionless form
y = 1/2 +s.
1.1. DIMENSIONAL ANALYSIS 3
Exercises, page 17
1. Assume that f(v, , g) = 0. If is dimensionless
[] = [v
1
2
g
3
],
= (LT
1
)
1
L
2
(LT
2
)
3
.
Thus we have the homogeneous system
1
+
2
+
3
= 0,
1
2
3
= 0.
The rank of the coecient matrix is one, so there is one dimensionless
variable. Notice that (2, 1, 1) is a solution to the system, and thus
= g/v
2
.
By the Pi theorem, F() = 0 or g/v
2
= Const.
2. Two dimensionless variables are
V
m
,
S
V
2/3
.
Therefore
V
m
= f
_
S
V
2/3
_
.
3. Pick length, time and mass as fundamental and write
x =
1
x, t =
2
t, m =
3
m.
Then write v =
1
1
2
v, and so on for the other variables. Show that
v
2
9
r
2
g
1
(1
l
/) =
1
1
2
(v
2
9
r
2
g
1
(1
l
/))
So, by denition, the law is unit free.
4. Select M, L, and T (mass, length, and time) as fundamental dimensions.
5. There is only one dimensionless variable among E, P, and A, namely
PA
3/2
/E. Thus,
PA
3/2
/E = const.
6. The two dimensionless variables are
at
L
,
bt
.
4 CHAPTER 1. SCALING, DIMENSIONAL ANALYSIS (SECS. 1 & 2)
7. Dimensionless quantities are
e
,
E
v
2
.
Thus, by the pi theorem,
v =
Ef
_
e
_
.
8. Select M, L, and T (mass, length, and time) as fundamental dimensions.
9. Assume there is a physical law f(T, V, C, Y, r) = 0. We have
= T
1
V
2
C
3
Y
4
r
5
and so
1 = T
1
(L
3
)
2
(ML
3
)
3
(MT
1
)
4
(MT
1
V
3
)
5
.
Setting the powers of T, L, and M equal to zero and solving gives
1
=
4
+
5
,
2
=
4
,
3
=
4
5
.
This leads to two dimensionless quantities
TY
V C
,
Tr
C
.
10. Select M, L, and T (mass, length, and time) as fundamental dimensions.
11. Select M, L, and T (mass, length, and time) as fundamental dimensions.
12. Pick length L, time T, and mass M as fundamental dimensions. Then
the dimension matrix has rank three and there are 5 3 = 2 dimension
less variables; they are given by
1
= and
2
= R
l
/
P. Thus
f(
1
,
2
) = 0 implies
= R
1
_
P/
l
G()
for some function G.
13. The dimensions are
[E] =
energy
mass
, [T] = temp, [k] =
energy
mass temp
.
It is clear there is only one dimensionless variable, = E/kT. Thus
E/kT = const.
1.1. DIMENSIONAL ANALYSIS 5
14. We have dimensions
[F] = MLT
1
, [V ] = LT
1
, [C] = L
3
T
1
, [K] = ML
1
T
2
.
Assume a physical law f(F, V, C, K) = 0. If is dimensionless, then
= F
1
V
2
C
3
K
4
.
This gives
1 = (MLT
1
)
1
(LT
1
)
2
(L
3
T
1
)
3
(ML
1
T
2
)
4
.
This leads to the system of equations
1
+
2
+ 3
3
4
= 0,
2
1
3
2
4
= 0,
1
+
4
= 0.
This system has rank 3 and so there is one solution, (1, 1, 1, 1), which
gives the dimensionless variable CK/FV = const.
15. We have dimensions
[w] = L, [C
0
] = [C 1] = ML
3
, [d] = L
2
T
1
, [] = ML
2
T
1
.
Assume a physical law f(w, C
0
, C
1
, d, ) = 0. If is dimensionless, then
= w
1
C
2
0
C
3
1
d
5
.
This gives
1 = L
1
(ML
3
)
2
(ML
3
)
3
(L
2
T
1
)
4
(ML
2
T
1
)
5
.
This leads to the system of equations
1
3
2
3
3
+ 2
4
2
5
= 0,
2
+
3
+
5
= 0,
4
+
5
= 0.
The system has rank 3 and so there are two independent solutions (0, 1, 1, 0, 0)
and (1, 1, 0, 1, 1). This gives dimensionless variables
C
0
C
1
,
w
dC
0
.
Therefore
w
dC
0
= G
_
C
0
C
1
_
,
which gives the form of the ux ,
=
dC
0
w
G
_
C
0
C
1
_
.
6 CHAPTER 1. SCALING, DIMENSIONAL ANALYSIS (SECS. 1 & 2)
1.2 Scaling
Exercises, page 30
1. In (a) we have u = Asint and so u
 = A. Then we have t
c
= 1/. In (b) we have u = Ae
t
and u
= Ae
t
. Then t
c
= max u/ max u
= (1 t)Ae
t
. The maximum of u occurs at
t = 1/ and is M = A/e. To nd the maximum of u
we calculate the
second derivative to get u
= A(t 2)e
t
. So the maximum derivative
occurs at t = 2/ or at an endpoint. It is easily checked that the maximum
derivative occurs at t = 0 and has value max u
= exp(t/)/. Then t
c
= max u/ max u
 =
2/
1
= 2. The time scale is very small, indicating rapid change in a small
interval. But a graph shows that that this rapid decrease occurs only in a
small interval near t = 0; in most of the interval the changes occur slowly.
Thus two time scales are suggested, one near the origin and one out in the
interval where t is order one.
3. We have
m
= ax
2
bx
3
, m = x
3
.
Thus
(x
3
)
= 3x
2
x
= ax
2
bx
3
,
giving
x
=
a
3
b
3
x.
We have a given in mass per time per lengthsquared and b in mass per
time per volume. Scaling time by /b and length by a/b leads to the
dimensionless model
y
=
1
3
1
3
y.
If x(0) = 0 then y(0) = 0 and the solution to the dimensionless model is
y() = 1 e
/3
.
Yes, this is a reasonable model. The organism grows exponentially toward
a limiting value. This is, in fact, observed with most organisms.
4. The constants in the problem, V , k, and a have dimensions
[V ] =
L
T
, [k] =
M
T
2
, [a] =
M
TL
.
One time scale is
_
m/k which is based on damping. Another is
_
m/aV ,
which is based on the restoring force. To rescale, let T and L be scales to
1.2. SCALING 7
be chosen and let y = x/L and = t/T. Then the model becomes
y
=
aTL
m
yy

T
2
k
m
y, y(0) = 0, y
(0) =
V T
L
.
We want the restoring force to be small and have the small coecient.
Therefore take T =
_
m/aV . Then we get
y
=
aL
_
m/aV
m
yy

k
aV
y, y(0) = 0, y
(0) =
V
_
m/aV
L
Now choose the length scale L so that the coecient of yy
 is one. The
dierential equation then becomes
y
= yy
 y, y(0) = 0, y
(0) = 1,
k
aV
.
So, the small coecient is in front of the small damping term.
5. The dimensions of the constants are
[I] =
ML
T
, [a] =
M
T
, [k] =
M
T
2
L
.
Letting u = x/(I/a) and = t/T, where T is yet to be determined, we
get
m
aT
u
= u
kTI
2
a
3
u, u(0) = 0,
m
aT
u
(0) = 1.
If the mass is small, we want to choose T so that the coecient of the
u
term is small. So, select T that makes the restoring force term have
coecient 1. Thus, take
T =
a
3
kI
2
.
The model then becomes
u
= u
u, =
mkI
2
a
4
.
6. (a) The constant a must be budwormssquared because it is added to such
a term in the denominator. The entire predation term must be budworms
per time, and so b must have dimensions budworms per time. (b) The
parameter a denes the place where the predation term makes a signicant
rise. Thus it indicates the threshold where the number of budworms is
plentiful so that predation kicks in; there are enough budworms to make
the birds interested. (c) The dimensionless equation is
dN
d
= sN(1 N/q)
N
2
1 +N
2
.
8 CHAPTER 1. SCALING, DIMENSIONAL ANALYSIS (SECS. 1 & 2)
(d) To nd equilibrium solutions we set
sN(1 N/q)
N
2
1 +N
2
= 0.
At this point we can use a calculator or a computer algebra program
like Maple or Mathematica to solve the equation for N. Observe that we
obtain a fourth degree polynomial equation when we simplify this algebraic
equation:
sN(1 N/q)(1 +N
2
) N
2
= 0.
When s = 12 and q = 0.25 the equilibrium populations are N = 0, /, 0.261.
When s = 0.4 and q = 35 the equilibrium populations are N = 0, 0.489, 2.218, 32.29.
7. Introduce the following dimensionless variables:
m = m/M, x = x/R, t = t/T, v = v/V,
where T and V are to be determined. In dimensionless variables the
equations now take the form
m
=
T
M
,
x
=
V T
R
v,
v
=
T
MV
1
m
Tg
V (1 x)
2
.
To ensure that the terms in the velocity and acceleration equations are
the same order, with the gravitational term small, pick
V T
R
=
T
MV
,
which gives
V =
_
R/M
as the velocity scale.
8. The dierential equation is
c
=
q
V
(c
i
c) kc
2
, c(0) = c
0
.
Here k is a volume per mass per time. Choosing dimensionless quantities
via
C = c/c
i
, = t/(V/q),
the model equation becomes
dC
d
= (1 C) bC
2
, C(0) = ,
where = c
0
/c
i
and b = kV c
i
/q. Solve this initial value problem using
separation of variables.
1.2. SCALING 9
9. The quantity q is degrees per time, k is time
1
, and is degrees (one can
only exponentiate a pure number). Introducing dimensionless variables
T = T/T
f
, = t/(T
f
/q),
we obtain the dimensionless model
dT
d
= e
E/T
(1 T), T(0) = ,
where = T
0
/T
f
, E = /T
f
, and = kT
f
/q.
10. Let h be the height measured above the ground. Then Newtons second
law gives
mh
= mg a(h
)
2
, h(0) = 0, h
(0) = V.
Choose new dimensionless time and distance variables according to
= t/(V/g), y = h/(V
2
/g).
Then the dimensionless model is
y
= 1 (y
)
2
, y(0) = 0, y
(0) = 1,
where prime is a derivative and = aV
2
/mg.
11. The model is
mx
=
k
x
2
e
t/a
, x(0) = L, x
(0) = 0.
We have [a] = T and [k] =
ML
3
T
2
. Two time scales are
a,
_
mL
3
k
.
12. The model is
mx
= kxe
t/a
, x(0) = L, x
(0) = V.
Let = t/a and y = x/L be dimensionless variables. Then
y
= ye
, y(0) = 1, y
(0) = ,
where = ka
2
/m and = V a/L.
13. The dynamics is given by
x
= y(1 y) if <
f
,
and
y
= y(1 y) y if >
f
.
Here, y = x/K and = rt. Also, = qB/r where b(t) = B.
Setting y
= 0 for >
f
gives the constant population y = 1 . If we
solve for y = y() for <
f
, then we can set y(
f
) = 1 to obtain
f
as a function of .
14. The mass times acceleration is the force, or
ms
= mg sin,
where the force is the tangential component of the force mg along the arc
of the path. But s = L, and so
mL
= mg sin.
Then
+
g
L
sin = 0, (0) =
0
,
(0) = 0.
We can scale theta by
0
and time by
_
L/g. Then the model becomes
+
1
0
sin(
0
) = 0, (0) = 1,
(0) = 0.
15. Three time scales are
_
L/g,
1
0
,
1
k
.
These time scales involve the eect of gravity (undamped oscillations),
the angular frequency caused by the initial angular velocity, and the time
scale of the damping.
Chapter 2
Perturbation Methods
2.1 Regular Perturbation
Exercises, page 100
1. Since the mass times the acceleration equals the force, we have my
=
ky a(y
)
2
. The initial conditions are y(0) = A, y
(0) = 0. Here, a
is assumed to be small. The scale for y is clearly the amplitude A. For
the time scale choose
_
m/k, which is the time scale when no damping is
present. Letting y = y/A, = t/
_
m/k be new dimensionless variables,
the model becomes
y
+(y
)
2
+y = 0,
where aA/m. The initial data is y(0) = 1, y
u = tu, u(0) = 1, u
(0) = 1.
A twoterm perturbation expansion is given by
y(t) = e
t
+
1
8
(e
t
e
t
(1 + 2x + 2x
2
)).
A sixterm Taylor expansion is
y(t) = 1 t +
1
2
t
2
+
1
6
x
3
+
1 2
24
x
4
1 4
120
x
5
.
Plots show the superior performance of the twoterm perturbation approx
imation.
3. We have e
t
= o(
1
t
2
) as t because (using LHopitals rule)
lim
t
e
t
1
t
2
= lim
t
t
2
e
t
= 0.
11
12 CHAPTER 2. PERTURBATION METHODS
4. Using the binomial theorem,
(1 +y)
3/2
= 1
3
2
y +
(3/2)(5/2)
2!
2
y
2
+ .
Now substitute y = y
0
+y
1
+
2
y
2
+ and expand.
5. (a) We have
t
2
tanht
t
2
= tanht < 1
for large t. Thus t
2
tanht = 0(t
2
) as t .
(b) We have
lim
t
e
t
1
= 0,
which proves the order relation.
(c) The order relation follows from the inequality
_
(1 )
1 1
for small, positive .
(d) The idea is to expand cos in a power series to get
1 (1
2
/2 +
4
/4! )
=
2
/2
4
/4! + )
=
1
2
3/2
1
1
2
/12 +
.
But, using the geometric series,
1
1
2
/12 +
= 1 + 0(
2
),
which gives the result.
(e) We have
t
t
2
=
1
t
1
for large t. So the ratio is bounded, proving the assertion.
(f) By Taylors expansion,
e
1 = 1 + + 0(
2
) 1 = 0().
2.1. REGULAR PERTURBATION 13
(g) Expand the integrand in a Taylor series and integrate termbyterm
to get
_
0
e
x
2
dx =
_
0
(1 x
2
+
x
2
2!
+ )dx
=
1
3
3
+
= 0.()
An alternate method is to notice that e
x
2
1, which gives
_
0
e
x
2
dx .
Thus
_
0
e
x
2
dx
1
(h) Observe that
lim
0
e
tan
= 1.
Because the limit exists, the function must be bounded in a neighborhood
of = 0, which implies the result.
(i) Notice that
e
p
= e
p
0
as (exponentials decay faster than power functions grow). One
can use LHospitals rule to show this.
(j) Notice that
ln
p
= p ln 0
as 0 since power functions go to zero faster that logarithms grow near
zero. Again, use LHospitals rule to verify this fact.
6. Substitute x = x
0
+x
1
+ into the nonlinear equation to get
h() (x
0
+x
1
+ , ) = 0
By Taylors expansion
h() = h(0) +h
(0) +
1
2
h
(0)
2
+
Using the chain rule we can compute these derivatives of h at = 0 and
thus expand the equation in powers of . We get
(x
0
) = 0, x
1
=
(x
0
, 0)
x
(x
0
, 0)
,
and so on. To obtain x
1
we clearly require
x
(x
0
, 0) = 0.
14 CHAPTER 2. PERTURBATION METHODS
7. (x + 1)
3
= x. Set x = x
0
+x
1
+ and expand.
8. (a) Let = t, with = 1 +
1
+ . Then the problem becomes
2
y
+y = y
2
(y
)
2
, y(0) = 1, y
(0) = 0.
Here, prime denotes a derivative. Now assume a regular perturba
tion expansion. The leading order problem is
y
0
+y
0
= 0, y
0
(0) = 1, y
0
(0) = 0,
which has solution
y
0
() = cos .
The next order problem is
y
1
+y
1
= 2
1
y
0
+y
0
(y
0
)
2
, y
1
(0) = 0, y
1
(0) = 0.
The equation simplies to
y
1
+y
1
= (
1
4
+ 2
1
) cos
1
4
cos 3.
To eliminate the secular term take
1
= 1/8. Then, to leading
order,
y
0
(t) = cos
__
1
1
8
_
t
_
.
(b) Let = t, with = 3 +
1
+ . Then the problem becomes
2
y
+ 9y = 3y
3
, y(0) = 0, y
(0) = 1.
Here, prime denotes a derivative. Now assume a regular perturba
tion expansion. The leading order problem is
y
0
+y
0
= 0, y
0
(0) = 0, y
0
(0) =
1
3
,
which has solution
y
0
() =
1
3
sin.
The next order problem is
9y
1
+ 9y
1
= 3y
2
0
6
1
y
0
, y
1
(0) = 0, y
1
(0) =
1
9
.
The equation simplies to
y
1
+y
1
=
1
9
_
1
12
+ 2
1
_
sin
1
9 36
sin3.
To eliminate the secular term take
1
= 1/24. Then, to leading
order,
y
0
(t) =
1
3
sin
__
3
1
24
_
t
_
.
2.1. REGULAR PERTURBATION 15
9. The equation with can be handled with a perturbation series in . After
determining the coecients, one can substitute = 0.001.
10. If we ignore 0.01x in the rst equation then y = 0.1. Then, from the second
equation x = 0.9. Checking these values by substituting back into the rst
equation gives 0.01(0.9) + 0.1 = 0.109, so the approximation appears to
be good. But the exact solution is x = 190, y = 1. So the approximation
is in fact terrible. What went wrong? Since x = 90 the rst term in the
rst equation is 0.01x = 0.9, which is not small compared to the two
other terms in the rst equation. Thus the rst term cannot be neglected.
11. Letting h = h
0
+h
1
+ gives
h
0
+h
1
+
2
h
2
+ = 1 + 2h
0
(3h
0
2h
1
)
2
+ .
Here we used the binomial theorem to expand the right hand side. We
also have the initial conditions
h
0
(0) = 0, h
0
(0) = 1; h
1
(0) = h
1
(0) = 0, . . . .
The equations are
h
0
= 1, h
1
= 2h
0
, h
2
= (3h
0
2h
1
), .
Now we can solve consecutively and get h
0
, h
1
, h
2
, . . .. Once the expansion
is obtained, solve h
(t) = 0 to get t
m
.
12. The initial value problem is
my
= ay
kye
rt
, y(0) = y
0
, y
(0) = 0.
The quantity m is mass, y
0
is length, r is time
1
, a is mass per time, and
k is mass per timesquared. To nondimensionalize, take u = y/y
0
and
= t/(m/a). Then
u
= u
ue
, u(0) = 1, u
(0) = 0,
where
=
km
a
2
, =
mr
a
.
To leading order we have u
0
= u
0
which gives u
0
(t) = A+Be
. From
the initial conditions A = 1 and B = 0, giving
u
0
(t) = 1.
At the next order we have u
1
= u
1
e
+
1
2
e
.
Use the initial conditions to determine A and B. (Here we are assuming
= 1.) Continue in this manner.
16 CHAPTER 2. PERTURBATION METHODS
13. Consider the equation
y
= y cos
x
2
.
Assuming a perturbation expansion, we get
y
0
= 0, y
1
= y
0
cos
x
2
, . . . .
Easily, y
0
(t) = x, and so on.
14. We have
y
0
+y
1
+ = e
eps/(y
0
+y
1
+ )
To leading order we have y
0
= 1 which gives, using the initial condition,
y
0
(t) = x + 1. To get higher order terms, use the expansion
exp
_
y
0
+y
1
+
_
= 1
y
0
+
y
1
y
2
0
2
+ .
15. Let =
0
+
1
+ . Then, substituting, gives
0
+
1
+ +
1
(
0
2
+ (
2
1
3
0
)
3
+ ) = 0.
Then
0
+
0
= 0,
1
+
1
= 0,
2
+
2
=
1
3
0
, and so on. The initial
conditions give
0
= cos and
1
= 0. Then
2
+
2
=
1
3
cos .
For part (b), multiply the equation by
to get
+
1
sin()
= 0.
This is the same as
1
2
(
2
)
2
(cos())
= 0.
Therefore
1
2
2
cos() = C.
From the initial conditions we get C =
1
2
cos , and therefore the last
equation can be written
2
d
_
cos() cos
= d.
Now integrate over onefourth of a period P to get
2
_
1
0
ds
_
cos(s) cos
=
P
4
.
To get the expansion P = 2 +
2
2
/8 + , expand the integrand in
powers of using the binomial theorem.
2.1. REGULAR PERTURBATION 17
16. The threeterm approximation is given by
y
a
(t) = t +
t
4
12
+
t
7
504
2
.
The error is
E(t, ) = y
a
ty
a
=
3
t
8
504
.
Clearly the approximation is not uniform on t 0.
17. The leading order solution is
y
0
(t) =
t(1 lnt).
Substituting into the ODE gives
Ly
0
=
2
(1 + ln t)t
3/2
.
We nd
Ly
0
 0.0448.
Thus one would expect y
0
to be a good approximation on 0 t e.
18. Substitute the series u = u
0
+u
1
+ into the dierential equation and
initial conditions:
u
0
+u
1
+ +u
0
+u
1
+ =
1
1 +u
0
+
= 1 u
0
+ 0(
2
).
To get the last step we used the geometric series expansion. Now collecting
the coecient of
0
we get the leading order problem
u
0
+u
0
= 1, u
0
(0) = 0.
Collecting the coecients of we get
u
1
+u
1
= u
0
, u
1
(0) = 0.
The solution to the leading order problem (a linear equation) is
u
0
= 1 e
t
.
The the next order problem becomes
u
1
+u
1
= e
t
1, u
1
(0) = 0.
This linear equation has solution
u
1
= (t + 1)e
t
1.
Therefore, a twoterm approximation is
u(t) = 1 e
t
+((t + 1)e
t
1) + .
18 CHAPTER 2. PERTURBATION METHODS
19. Let s be the speed of the wildebeest and =
_
_
dx
dt
_
2
+
_
dy
dt
_
22
=
dx
dt
_
1 +y
2
be the speed of the lion. The velocity of the lion is
_
dy
dx
,
dy
dt
_
=
_
(a x)
2
+ (b +st y)
2
(a x, b +st y) .
Now,
y
=
b +st y
a x
, y
=
dy
dx
.
Therefore,
d
dt
y
= y
dx
dt
=
(a x)(s
dy
dt
) +
dx
dt
(b +st y)
(a x)
2
.
This simplies to
(a x)y
=
s
dy
dt
dx
dt
+
dy
dx
=
s
dx
dt
=
s
_
1 +y
2
.
So we have
(a x)y
=
_
1 +y
2
, y(0) = 0, y
(0) =
b
a
.
Assuming
y = y
0
(x) +y
1
(x) + ,
the leading order problem is
(a x)y
0
= 0, y
0
(0) = 0, y
0
(0) =
b
a
,
which has solution
y
0
(x) =
b
a
x.
At next order
(a x)y
1
= 1, y
1
(0) = 0, y
1
(0) = 0.
Now,
y
1
= ln(a x) +C.
Using the second initial condition, C = lna. Now integrate again to get
y
1
(x) =
_
x
0
ln(a )d +xlna.
Then
y(x) =
b
a
x +
_
_
x
0
ln(a )d +xlna
_
+ .
2.2. SINGULAR PERTURBATION 19
2.2 Singular Perturbation
Exercises, page 111
1. (a) Consider the equation
x
4
+x
3
x
2
+ 2x 1 = 0
If x = O(1) then x
2
2x+1 0 which means x 1, 1, a double root
near one. To nd the remaining roots assume a dominant balance
x
4
x
2
with the remaining terms small. Then x = O(1/
). This
is a consistent balance because x
4
, x
2
= O(1/) and the remain
ing three terms are are order one and small in comparison. So the
dominant balance is
x
4
x
2
0,
which gives x 1/
.
(b) The equation
x
3
+x 2 = 0
has an order one root x 2. The consistent dominant balance is
x
3
x which gives x = O(1/
)
So, we are rescaling. Then the equation becomes
y
3
+y 2
= 0
and the small term appears where it should in the equation. Now
assume y = y
0
+y
1
+ .
In terms of x,
x =
i
1 + .
(c) The equation
2
x
6
x
4
x
3
+ 8 = 0
has three order one roots as solutions of x
3
+ 8 = 0, or
x 2, 1
3i.
To nd the other roots, the dominant balance is
2
x
6
x
3
0, which
gives
x
1
2/3
e
2i/3
,
1
2/3
e
2i/3
.
(d) The equation
x
5
+x
3
1 = 0
has three order one roots (the cube roots of one) given by
x 1,
1
2
3i.
The dominant balance for the remaining roots is between the rst
two terms which gives
x
i
.
2. Follow the given hint.
3. Observe that the equation can be written as a quadratic in :
2
2
+x +x
3
= 0.
Thus
=
1
4
(x
_
x
2
8x
3
).
These two branches can be graphed on a calculator, and the graph shows
that there is just one negative value for x, near x = 0, in the case that
is small and positive. Thus assume
x = x
1
+x
2
2
+ .
Substituting into the given equation gives
x = 2 + 8
2
+ .
2.3. BOUNDARY LAYER ANALYSIS 21
4. The problem is
y
+y
+y = 0, y(0) = 0, y(1) = 0.
When = 0 we get y
1 4) = 1 +O(),
m
2
=
1
2
(1
1 4) =
1
+O(1).
Here we have used the binomial expansion
1 +x = 1 +x/2 +O(x
2
) for
small x. Note that one of the roots is order one, and one of the roots is
large. So the general solution is
y(t) = c
1
e
m1t
+c
2
e
m2t
.
Applying the boundary conditions gives the exact solution
y(t) =
e
m1t
e
m2t
e
m
1
e
m
2
.
Sketches of this solution show a boundary layer near t = 0 where there is
a rapid increase in y(t). Observe that e
m
1
>> e
m
2
. If t = O(1) then
e
m
1
t
e
t
, e
m
2
t
0,
and thus
y(t)
e
m
1
t
e
m1
e
1t
,
which is an outer approximation. If t = O(), then
y(t)
e
(1+O())t
e
(1/+O(1))t
e
m1
e
O()
e
O()
e
t/
e
m1
1 e
t/
e
1
.
This is an inner approximation near t = 0.
2.3 Boundary Layer Analysis
Exercises, page 121
22 CHAPTER 2. PERTURBATION METHODS
1. (a) Consider
y
+ 2y
+y = 0, y(0) = 0, y(1) = 1.
By Theorem 2.9 there is a boundary layer at x = 0. Setting = 0
gives the outer approximation y(x) = Ce
x/2
. Use the right bound
ary condition to nd C =
ee
x/2
= e
(1x)/2
.
The thickness of the layer is () = and the inner equation is
Y
+ 2Y
+Y = 0.
The leading order inner approximation is
Y
i
() = a +be
2
, =
x
.
Now, Y
i
(0) = a +b = 0, and matching gives a =
e.
(b) We have
y
+y
+y
2
= 0, y(0) = 1/4, y(1) = 1/2.
There is an expected layer at x = 0. Setting = 0 and solving gives
the outer solution
y
0
(x) =
1
x + 2
.
We applied the right boundary condition. In the boundary layer set
= x/ and Y () = y(x). Then the inner equation is
Y
+Y
+Y
2
= 0.
To leading order we have Y
i
+Y
i
= 0 with Y
i
(0) = 1/4. So the inner
approximation is
Y
i
() = A(1 e
).
Matching gives A = 1/2 and so the uniform approximation is
y(x) =
1
x + 2
1
4
e
x/
.
(c) We have
y
+ (1 +x)y
2
+ (1 +)
1
= 1.
2.3. BOUNDARY LAYER ANALYSIS 23
Balancing gives = and the inner rstorder approximate equation
is
Y
i
+Y
i
= 0, Y
i
(0) = 0.
So the inner solution is
Y
i
() = A(1 e
).
Matching gives A = 1 and so the uniform approximation is
y(x) = ln(x + 1) e
x/
.
(d) The problem is
y
+ (1 +t)y
+y = 0, y(0) = 0, y(1) = 1.
By Theorem 3.1 in the text, there is a boundary layer at zero. The
outer solution is y
0
= 2/(t + 1). In the boundary layer set = t/
and Y () = y(t). Then the inner equation is
Y
+ +Y
+Y = 0.
To leading order we have Y
i
+ Y
i
= 0 with Y
i
(0) = 0. So the inner
approximation is
Y
i
() = A(1 e
).
Matching gives A = 2 and so the uniform approximation is
y(t) =
2
t + 1
+ 2(1 e
t/
) 2.
(e) The problem is
y
+t
1/3
y
+y = 0, y(0) = 0, y(1) = e
3/2
.
There is a boundary layer near t = 0. The outer solution is
y
0
(t) = exp(1.5t
2/3
).
In the inner region set = t/(). Then the dominant balance is
between the rst and second terms and =
3/4
. So the inner ap
proximation to leading order is
Y
i
+
1/3
Y
i
= 0.
Solving gives
Y
i
() = c
_
0
exp(0.75s
4/3
)ds.
Pick an intermediate variable to be = t/
+xy
xy = 0, y
(0) = 1, y(1) = e.
By Theorem 2.9 there is a layer at x = 0. The outer approximation
is
y
0
(x) = e
1x
.
Letting = x/() in the layer, we nd from dominant balance that
() =
i
+Y
i
= 0.
Then
Y
i
() = a
_
0
e
s
2
/2
ds +b.
Then Y
i
(0) = b = 0. Matching gives
a = e
__
0
eds
_
1
.
(g) The problem is
y
+ 2y
+e
y
= 0, y(0) = y(1) = 0
There is a layer at zero. The outer solution is
y
0
(t) = ln
t + 1
2
In the boundary layer the rst two terms dominate and () = .
The inner solution is
Y
i
() = A(1 e
2
)
Matching gives A = ln2. The uniform approximation is
y(t) = ln2(1 e
2t/
) ln
t + 1
2
ln2.
(h) The problem is
y
(2 t
2
)y = 1, y(1) = y(1) = 1
Now there are two layers near t = 1 and t = 1. The outer solution,
which is valid in the interval (1, 1), away from the layers is
y
0
(t) =
1
2 t
2
2.3. BOUNDARY LAYER ANALYSIS 25
In the layer near t = 1 set = (1 t)/(). We nd =
with
inner equation, to leading order, Y
i
Y
i
= 1. The inner solution is
Y
i
() = 1 +ae
(1 +a)e
i
)
i
= 1. The inner
solution is
Y
i
() = 1 +be
(1 +b)e
e
(t+1)/
b(x)y
= 0, y(0) = , y(1) = .
There is an expected layer at x=1 because the coecient of y
is
negative. Therefore the outer solution is
y
0
(x) = const. = .
Now make the change of variables
=
1 x
()
.
Then x = 1 and the dierential equation becomes
2
Y
+ [b(1) b
(1) + )
1
= 0.
Balancing terms gives = and the leading order inner equation is
Y
i
+b(1)Y
i
= 0.
We have Y
i
(0) = . The solution is
Y
i
() = A+ ( A)e
b(1)
.
Matching gives A = . Then, a uniform approximation is
y(x) = + ( )e
b(1)x/
.
(j) Consider
y
4( x
2
)y = cos x, y(0) = 0, y(/2) = 1.
26 CHAPTER 2. PERTURBATION METHODS
2. The solution is
u(x) = a sin(x/
) +b cos(x/
)
where a and b are determined uniquely by the boundary conditions. This
a very rapidly oscillating function over the entire interval. To apply per
turbation methods we set = 0 to get the outer solution u
(
x) = 0. This
constant solution cannot be matched to rapid oscillations.
3. See problem 1(g).
4. Consider
u
(2x + 1)u
+ 2u = 0, u(0) = 1, u(1) = 0.
5. The problem is
y
+
1
x
y
+y = 0, y(0) = 1, y
(0) = 0
which is an initial value problem and appears to be singular. But, the
outer solution is
y
0
(x) = Ce
x
2
/2
and it is observed that it satises both initial conditions when C = 1. It
also satises the ODE uniformly, i.e.,
y
0
+
1
x
y
0
+y
0
= (x
2
1)e
x
2
/2
= O()
Thus it provides a uniform approximation and the problem does not have
a layer. It is instructive to try to put a layer at x = 0; one nds that no
scaling is possible.
6. Consider
y
+
_
x
1
2
_
y = 0, y(0) = 1, y(1) = 2.
There is a layer at both x = 0 and x = 1. The outer solution is y
0
(x) = 0.
By dominant balancing, the width of the layer in both cases is () =
_
().
The inner equation at x = 0 is
Y
l
1
2
Y
l
= 0,
where
=
x
.
The inner equation at x = 1 is
Y
r
1
2
Y
r
= 0,
2.3. BOUNDARY LAYER ANALYSIS 27
where
=
1 x
.
. Using the appropriate boundary conditions gives the left and right inner
solutions
Y
l
() = e
/
2
, Y
r
() = e
/
2
.
The uniform approximation is
y(x) = e
x/
2
+e
(1x)/
2
.
7. There is a layer at x = 0. The outer solution is y
0
(x) = exp(x). The
layer has width () = and the leading order inner problem is
Y
i
+Y
i
= 0.
It has solution Y
i
() = A + Bexp(). The boundary condition gives
A +B = 1. Thus Y
i
() = A + (1 A) exp(). Matching give A=1 and
the uniform approximation is
y(x) = 2e
x/
e
x
.
8. Assuming a layer at x = 0 we obtain outer solution
y
0
(x) = f
(x) f
(1).
Now assume = x/. In transforming the equation, we need
1
f
()
=
1
f
(0) +f
(0) +
=
1
f
(0)
f
(0) + .
The dominant balance is = and the leading order approximation is
Y
1
f
(0)
Y = 0,
giving
Y () = A+Be
/f
(0)
.
Here we need f
(b) = 0 automatically. We
try a layer at x = 0 by dening the scaling
= x/().
28 CHAPTER 2. PERTURBATION METHODS
Then the equation becomes
+ cos(b )U = 1.
We have to expand everything in terms of . We have
cos(b x) = cos b cos() + sinb sin()
= cos b(1
1
2
()
2
+ ) + sin b(
1
6
()
3
+ ).
Substituting into the dierential equation leads to a dominant balance
giving
=
.
The leading order inner equation is
U
+ (cos b)U = 1.
The general solution is
U() = Ae
(cos b)
+Be
(cos b)
+
1
cos b
.
Now, U(0) = A+B +
1
cos b
= 0. Thus B =
_
A+
1
cos b
_
. For matching to
work we must have A =
1
cos b
. Then we have the uniform approximation
u(x) =
1
cos b
e
(cos b)x/
+
1
cos b
+
1
cos(b x)
1
cos b
=
1
cos b
e
(cos b)x/
+
1
cos(b x)
.
10. The outer solution is clearly
u
0
(x) = 0,
with a layer at x = 1. (Theorem 2.9 applies.) In the layer,
=
1 x
.
Then the dierential equation transforms to
1
a(1 )
1
= f(x).
Expanding,
U
+ (a(1) a
(1) + )U
= f(x).
To leading order,
U
i
+ (a(1) a
(1) + )U
i
= 0,
which gives
U
i
() = A+Be
a(1)
.
The boundary condition is U
i
(0) = A+Be
a(1)
= f(1)/a(1).
2.3. BOUNDARY LAYER ANALYSIS 29
11. See problem 12e.
12. (a) By Theorem 3.1 the problem
y
+ (1 +x
2
)y
x
3
y = 0, y(0) = y(1) = 1
has a layer of order = at t = 0. The outer solution is
y
0
(x) =
_
2
1 +x
2
exp(
x
2
1
2
)
The inner solution is
Y
i
() = A+ (1 A)e
Matching gives A =
_
2/e.
(b) By Theorem 3.1 the problem
y
+
2
t
y
y = 0, y(0) = 0, y
(1) = 1
If we assume a layer at t = 1 the outer solution is y
0
(t) = 0. In the
inner region near t = 1 the dominant balance is between the rst and
last terms and the width of the layer is =
. The inner variable
is = (1 t)/. The inner solution is
Y
i
() = (
+b)e
+be
e
(t1)/
+
1
x + 1
y
+y = 0, y(0) = 0, y(1) = 1.
The outer approximation is y
0
(t) =const., so there are several pos
sibilities. But, because the y
()
2
Y
+
1
(1 +())
1
()
Y
+Y = 0.
Expanding the second term in a geometric series gives
()
2
Y
+ (1 () + )
1
()
Y
+Y = 0.
Balancing gives () = and the leading order equation is
Y
i
+Y
i
= 0,
which gives
Y
i
() = a(1 e
).
Matching gives a = 1. Therefore a uniform approximation is
y
u
(x) = 1 + 1 e
x/
1 = 1 e
x/
.
This approximation satises y
u
(0) = 0, y
u
(1) = 1exp(1/), which
is one minus an exponentially small term. Substituting into the dif
ferential equation gives
y
u
+
1
x + 1
y
u
+y
u
= + exponentially small term.
Boundary layer at the right boundary. Just as illustration we
show how to proceed if the dierential equation is
y
1
x + 1
y
+y = 0
2.4. INITIAL LAYERS 31
with the same boundary conditions. Now the y
coecient is negative
and we expect a layer at x = 1. Therefore the outer solution is
y
0
(x) = 0 (from applying the left boundary condition). Now let
=
1 x
()
, (note the change)
Then the inner problem is
()
2
Y
(1 () + )
1
()
Y
+Y = 0.
Note the minus sign that appears when transforming y
. Then, dom
inant balance forces () = and the leading order equation is
Y
i
+Y
i
= 0, Y
i
(0) = 1.
The inner solution is
Y
i
() = a +be
.
Applying the boundary condition gives a +b = 1, so
Y
i
() = a + (1 a)e
).
Matching gives a = 0 because y
0
(x) 0 as x 1. Therefore a
uniform approximation is outer + inner  common limit, or
y
u
(x) = e
(1x)/
.
2.4 Initial Layers
Exercises, page 133
1. The equation is
y
+y = e
t
, y(0) = 2.
Set = 0 to obtain the outer solution y
0
(t) = e
t
, away from t = 0.
Rescale near zero via = t/(). Then, in the usual way, we nd () =
and the leading order inner problem is
Y
i
+Y
i
= 1.
This has solution Y
i
() = 1 + Ce
+b(t)y
+y = 0, y(0) = 1, y
(0) =
+.
The outer solution is
y
0
(t) = C exp
_
_
t
0
b(z)
1
dz
_
.
Near t = 0 set = (). Then
()
2
Y
+
1
()
(b(0) +b
(0)() + )Y
+Y = 0.
The dominant balance gives () = and the inner problem, to leading
order is
Y
i
+b(0)Y
i
= 0.
This has general solution Y
i
= A + Be
b(0)
. From Y
i
(0) = 1 we get
A+B = 1. The other initial condition leads to Y
i
(0) = . Therefore the
inner approximation in the initial layer is
Y
i
() =
_
1 +
b(0)
_
b(0)
e
b(0)
.
The matching condition gives C = 1 +
b(0)
. So, a uniform approximation
is
y(t) =
_
1 +
b(0)
_
exp
_
_
t
0
b(z)
1
dz
_
b(0)
e
b(0)t/
.
3. The problem is
y
+ (t + 1)
2
y
= 1, y(0) = 1, y
(0) = 1.
The outer approximation is
y
0
(t) =
1
t + 1
+C.
Rescaling in the initial layer gives
()
2
Y
+
1
()
(1 +())
2
Y
= 1.
The dominant balance is = and to leading order we have
Y
+Y
= 0,
which gives
Y
i
() = A+Be
.
2.4. INITIAL LAYERS 33
Now, Y (0) = 1 gives A + B = 1. And, the other initial condition gives
Y
.
Matching gives C=3. Then the uniform approximation is
y(t) = e
t/
+ 3
1
t + 1
.
4. The damped oscillator is governed by
my
+ay
+ky
3
= 0, y(0) = 0, my
(0) = I.
Let = t/(a/k) and u = y/(I/a) to get
u
+Y
+u
+u = 0, u(0) = 0, u
(0) = 1.
The standard singular perturbation method with an initial layer at t = 0
leads to the approximation
u() = e
e
/
.
5. (a) In this case the system is
x
= y sinx, y
= x
2
y +y
3
with initial conditions x(0) = k, y(0) = 0. Setting = 0 we get the
outer equations
x
0
= y
0
, 0 = x
2
0
y
0
Here we can choose y
0
= 0 and x
0
= k and the initial conditions are
met. So this problem does not have an initial layer. It is a regular
perturbation problem with leading order solution
x
0
(t) = k, y
0
(t) = 0
It is instructive for the student to assume a layer near t = 0 and carry
out the analysis to nd that the inner approximation agrees with the
outer approximation.
(b) The problem is
u
= v, v
= u
2
v, u(0) = 1, v(0) = 0
Assume a boundary layer near t = 0. Then the outer problem is
u
0
= v
0
, v
0
= u
2
0
Then
u
0
= u
2
0
34 CHAPTER 2. PERTURBATION METHODS
which has solution (separate variables)
u
0
(t) =
1
t +c
, v
0
(t) =
1
(t +c)
2
In the boundary layer take = t/. Then the inner problem is
U
= V, V
= U
2
V, U(0) = 1, V (0) = 0
Thus, setting = 0 and solving yields the inner approximation
U() = const. = 1, V () = e
1
Matching gives
lim
t0
u
0
(t) = lim
U()
or 1/c = 1. Hence, c = 1. Then the uniform approximation is
u =
1
t + 1
+ 1 1 =
1
t + 1
and
v =
1
(t + 1)
2
+e
t/
1 (1) =
1
(t + 1)
2
+e
t/
6. The governing equations are
a
= k
f
a +k
b
b, b
= k
f
a k
b
b.
Therefore a + b is constant, and so a + b = a
0
, giving b = a
0
a. Then
the aequation becomes
a
= (kf +k
b
)a +k
b
a
0
,
which has general solution
a(t) = Ce
(k
b
+k
f
)t
+
a
0
k
b
k
b
+k
f
.
The constant C can be determined from the initial condition.
7. The governing equations for the reaction X +Y Z are
x
= kxy, y
= kxy.
Therefore x y = C, where C is constant. Hence,
x
= rk(T)x, T = T
0
+h(x x
0
).
Making all the suggested changes of variables gives
= e
A
e
A/
(1 + ), (0) = 1,
where A = E/RT
0
and = hx
0
/T
0
. For small A take
=
0
+
1
A+
2
A
2
+ .
To leading order
0
= 1 +
0
,
(
0) = 1,
which has solution
0
() = (1 1/)e
t
+
1
.
For large A take
=
0
+
1
1
A
+
2
1
A
2
+ .
9.
2.5 WKB Approximation
Exercises, page 141
1. Letting = 1/
we have
2
y
(1 +x
2
)
2
y = 0, y(0) = 0, y
(0) = 1.
This is the nonoscillatory case. From equation (2.96) of the text the
WKB approximation is, after applying the condition y(0) = 0,
y
WKB
=
c
1
1 +x
2
_
exp
_
_
x
0
(1 +
2
)
2
d
_
exp
_
_
x
0
(1 +
2
)
2
d
__
=
2c
1
1 +x
2
sinh
_
_
x
0
(1 +
2
)
2
d
_
.
Applying the condition y
(0) = 1 gives c
1
= 1/2.
2. Letting = 1/
we have
y
+(x +)
4
y = 0, y(0) = y() = 0.
This is the oscillatory case and the method in Example 2.15 applies. The
large eigenvalues are
n
=
_
n
_
0
(x +)
2
dx
_
=
9n
2
49
4
.
36 CHAPTER 2. PERTURBATION METHODS
The eigenfunctions are
C
x +
sin
_
3n
7
2
_
x
0
( +)
2
d
_
=
C
x +
sin
_
3n
7
2
(
x
3
x
+x
2
+
2
x)
_
.
3. Let = 1/
2
and rewrite the problem as
2
y
2
d
2
y
dt
2
+q()
2
y = 0.
We can think of the equation y
+ q(t)
2
y = 0 as a harmonic oscillator
where the frequency is q(t), which is timedependent. If is small, it
will take a large time t before there is signicant changes in the frequency.
Thinking of it dierently, if q(t) is a given frequency, then the graph of
q(t) will be stretched out; so the frequency will vary slowly.
7. Here we apply the ideas in Example 2.15 with k(x) = e
2x
and = 1/
.
Then the WKB approximation is
y
WKB
=
c
1
e
x
sin
_
2
(e
2x
1)
_
+
c
2
e
x
cos
_
2
(e
2x
1)
_
.
Applying the condition y(0) = 0 gives c
2
= 0. Then
y
WKB
=
c
1
e
x
sin
_
2
(e
2x
1)
_
Then y(1) = 0 gives
sin
_
2
(e 1)
_
= 0,
and this forces
=
4
2
(e 1)
2
for large n. [Note the typographical errorthe boundary conditions should
be homogeneous.]
2.6. ASYMPTOTIC EXPANSION OF INTEGRALS 37
8. Let = 1/ to obtain
2
y
+ (x
2
+
2
x)y = 0.
Now let y = exp(iu/) and proceed as in the derivation of the WKB
approximation.
2.6 Asymptotic Expansion of Integrals
Exercises, page148
1.
2. Making the substitution t = tan
2
gives
I() =
_
/2
0
e
tan
2
d =
1
2
_
0
e
t
dt
(1 +t)
t
Now Watsons lemma (Theorem 6.1) applies. But we proceed directly by
expanding 1/(1 +t) in its Taylor series
1
1 +t
= 1 t +t
2
which gives
I() =
1
2
_
0
e
t
t
(1 t +t
2
)dt
Now let u = t. This gives
I() =
1
2
_
0
e
u
_
1
+
u
3/2
2
+
_
Then, using the dention of the gamma function,
I() =
1
2
((
1
2
)
1
(
3
2
) +
1
2
(
5
2
) + )
3. Assume that g has a maximum at b with g
(b)(t b) +
The integral becomes
I() =
_
b
a
f(t)e
g(t)
dt
=
_
b
a
f(t)e
(g(b)+g
(b)(tb)+ )
dt
f(b)e
g(b)
_
b
a
e
g
(b)(tb)
dt
38 CHAPTER 2. PERTURBATION METHODS
Now make the substitution v = g
(b)(t b) to obtain
I() f(b)e
g(b)
1
g
(b)
_
0
g
(b)(ab)
e
v
dv
or
I()
f(b)e
g(b)
g
(b)
_
0
e
v
dv
or
I()
f(b)e
g(b)
g
(b)
for large .
If the maximum of g occurs at t = a with g
t
4
2
+
t
6
3
)dt
Now let u = t and we get
I() =
1
_
0
e
u
(
u
2
2
u
4
2
4
+
u
6
3
6
+ )dt
Using the denition of the gamma function, we obtain
I() =
1
(
2!
2
4!
2
4
+
6!
3
6
+ )
(b) Let g(t) = 2t t
2
. This function has its maximum at t = 1 where
g
(1) = 0 and g
1 +t. Then
I() =
_
1
0
1 +te
(2tt
2
)
dt
1
2
f(1)e
g(1)
2
g
(1)
=
_
2
e
(c) Let g(t) = 1/(1+t). This function has its maximum, with a negative
derivative, at t = 1. Thus Exercise 6.3 holds. With f(t) =
3 +t
we have
I()
f(1)e
g(1)
g
(1)
=
8
e
/2
2.6. ASYMPTOTIC EXPANSION OF INTEGRALS 39
5. We have
(x + 1) =
_
0
u
x
e
u
du
Integrate by parts by letting r = u
x
and ds = e
u
du. Then the integral
becomes
_
0
u
x
e
u
du = x
_
0
e
u
u
x1
du = x(x)
6.
7.
8.
9.
10.
11.
12. (b) Using the fact that e
t
< 1 for t > 0, we have
r
n
() = n!
_
e
t
t
n+1
dt
n!
_
1
t
n+1
dt
= n!
1
nt
n

= (n 1)!
1
n
0
as .
(c) Observe that
_
e
t
t
n+1
dt
e
n+1
Then the ratio of r
n
to the last term of the expansion is
r
n
()
(n 1)!e
/
n
n
2
2
_
0
1
t +)
3
e
t
dt
Now integrate by parts again via
u =
1
t +)
3
, dv = e
t
Then
I() =
1
2
2
3
+ 6
_
0
1
t +)
4
e
t
dt
Continuing in the same manner gives
I() =
1
2
2!
3
+
3!
2
+ +
n!
n1
(1)
n+1
+ (n + 1)!
_
0
1
t +)
n+2
e
t
dt
Chapter 3
Calculus of Variations
3.1 Variational Problems
Exercises page 158
1. The functional is
J(y) =
_
1
0
(y
siny (y +t)
2
)dt.
First note that if y(t) = t then J(y) = 2/. Now we have to show that
J(y) < 2/ for any other y. To this end,
J(y) =
_
1
0
(y
siny (y +t)
2
)dt
_
1
0
y
sinydt
=
1
_
1
0
(cos y)
dt
=
1
.
2. Hint: substitute
y(x) = x +c
1
x(1 x) +c
2
x
2
(1 x)
into the functional J(y) to obtain a function F = F(c
1
, c
2
) of the two
variables c
1
and c
2
. Then apply ordinary calculus techniques to F to nd
the values that minimize F, and hence J. That is, set F = 0 and solve
for c
1
and c
2
.
41
42 CHAPTER 3. CALCULUS OF VARIATIONS
3.2 Necessary Conditions for Extrema
Exercises, page 166
1. (a) The set of polynomials of degree 2 is a linear space. (b) The set
of continuous functions on [0, 1] satisfying f(0) = 0 is a linear space. (c)
The set of continuous functions on [0, 1] satisfying f(0) = 1 is not a linear
space because, for example, the sum of two such functions is not in the
set.
2. We prove that
y
1
=
_
b
a
y(x)dx
is a norm on the set of continuous functions on the interval [a, b]. First
y
1
=
_
b
a
y(x)dx = 
_
b
a
y(x)dx =  y
1
.
Next, if y
1
= 0 i
_
b
a
y(x)dx = 0 i y(x) = 0. Finally, the triangle
inequality if proved by
y +v
1
=
_
b
a
y(x) +v(x)dx
_
b
a
(y(x) +v(x))dx
=
_
b
a
y(x)dx +
_
b
a
v(x)dx = y
1
+v
1
.
The proof that the maximum norm is, in fact, a norm, follows in the same
manner. To prove the triangle inequality use the fact that the maximum
of a sum is less than or equal to the sum of the maxima.
3. Let y
1
= 0 and y
2
= 0.01 sin1000x. Then
y
1
y
2

s
= max 0.01 sin1000x = 0.01
and
y
1
y
2

w
= max 0.01 sin1000x + max (0.01)(1000) cos 1000x = 10.01.
4. We have
J(y
0
, h) = lim
0
J(y
0
+h) J(w)
= lim
0
J(y
0
+h) J(w)
= lim
0
J(y
0
+h) J(w)
= J(y
0
, h).
3.2. NECESSARY CONDITIONS FOR EXTREMA 43
5. (a) not linear; (b) not linear; (c) not linear; (d) not linear; (e) linear; (f)
not linear.
6. An alternate characterization of continuity of a functional that is often
easier to work with is: a functional J on a normed linear space with norm
 is continuous at y if for any sequence of functions y
n
with y
n
y 0
we have J(y
n
) J(y) as n 0.
(a) Now let y
n
y
w
0. Then y
n
y
s
0 (because v
s
v
w
).
By assumption J is continuous at y in the strong norm, and therefore
J(y
n
) J(y). So J is continuous in the weak norm.
(b) Consider the arclength functional J(y) =
_
b
a
_
1 + (y
)
2
dx. If two
functions are close in the strong norm, i.e., if the maximum of their
dierence is small, then it is not necessarily true that their arclengths
are close. For example, one may oscillate rapidly while the other does
not. Take y = 0 and y = n
1
sinnx for large n. These two functions
are close in the strong norm, but not the weak norm.
7. (a) J(y, h) =
_
b
a
(hy
+yh
)dx.
(b) J(y, h) =
_
b
a
(2h
+ 2h)dx.
(c) J(y, h) = e
y(a)
h(a).
(d) See Exercise 3.6.
(e) J(y, h) =
_
b
a
h(x) sinx dx.
(f) J(y, h) =
_
b
a
2y
dx +G
(y(b))h(b).
8. Let y
n
y. Then J(y
n
) = J(y
n
y)+J(y) J(y) because J(y
n
y) 0
(since y
n
y 0 and J is continuous at zero by assumption).
9. We have
J(y +h) =
_
b
a
_
x(y
+h
)
2
+ (y +h) sin(y
+h
)
2
_
dx.
Now take the second derivative of this function of with respect to and
then set = 0. We obtain
2
J =
_
b
a
(2x(h
)
2
y(h
)
2
siny
+ 2hh
cos y
)dx.
10. Here the functional is
J(y) =
_
1
0
(x
2
y
2
+ (y
)
2
)dx.
Then
J(y, h) =
_
1
0
(2yh + 2y
)dx.
44 CHAPTER 3. CALCULUS OF VARIATIONS
Thus
J(x, x
2
) =
3
2
,
and
J = J(y +h) J(y) = J(x +x
2
) J(x) = etc.
11. J(y) =
_
1
0
(1 +x)y
2
dx. We nd
J(y, h) = 2
_
1
0
(1 +x)y
dx.
Now substitute the given y to show J(y, h) = 0 for appropriate h.
12. In this case
J(y) =
_
2
0
(y
)
2
dx.
Then
J(y +h) =
_
2
0
(1 + cos x)
2
dx = 2 +
2
_
2
0
cos
2
xdx.
Then
d
d
J(y +h) = 2
_
2
0
cos
2
xdx,
and so
d
d
J(y +h) 
=0
= 0.
Thus, by denition, J is stationary at y = x in the direction h = sinx.
The family of curves y +h is shown in the gure.
13. Here
J(y) =
_
1
0
(3y
2
+x)dx +y(0)
2
.
Then
J(y, h) =
_
1
0
6yh dx + 2y(0)h(0).
Substituting y = x and h = x + 1 gives J = 5.
3.3 The Simplest Problem
Exercises, page 175
1. (a) The Euler equation reduces to an identity (0 = 0), and hence every
C
2
function is an extremal. (b) The Euler equation reduces to an identity
(0 = 0), and hence every C
2
function is an extremal. (c) The Euler
equation reduces to y = 0, which is the only extremal. Remember, by
denition, solutions to the Euler equation are extremals, regardless of the
boundary conditions.
3.3. THE SIMPLEST PROBLEM 45
2. (a) The Euler equation is
L
y
d
dx
L
y
=
d
dx
(2y
/x
3
) = 0
Thus
y(x) = Ax
4
+B
(b) The Euler equation is
y
y = e
x
The general solution is
y(x) = Ae
x
+Be
x
+
x
2
e
x
3. The Euler equation is
L
y
d
dx
L
y
= f
y
_
1 + (y
)
2
d
dx
fy
_
1 + (y
)
2
= 0.
Taking the total derivative and then multiplying by
_
1 + (y
)
2
gives
f
y
y
f
x
fy
+
f(y
)
2
y
1 + (y
)
2
= 0
This reduces to
f
y
y
f
x
fy
1 + (y
)
2
= 0
4. The Euler equation is
d
dx
L
y
= 0
or
y
x
_
1 + (y
)
2
= c
Solving for y
+ 2qY h +ph
2
+qh
2
]dx
>
_
b
a
[2pY
+ 2qY h]dx
= 2pY
h
b
a
_
b
a
[(2pY
h 2qY h]dx
= 0.
The last two lines follow from integration by parts and the fact that Y
satises the Euler equation
(pY
qY = 0.
6. Let h C
2
. Then
J(y, h) =
_
b
a
_
b
a
K(s, t)[y(s)h(t) +h(s)y(t)]ds dt
+2
_
b
a
y(t)h(t)dt 2
_
b
a
h(t)f(t)dt
Now, using the symmetry of K and interchanging the order of integration
allows us to rewrite the rst integral as
2
_
b
a
_
b
a
K(s, t)y(s)h(t)dsdt
Then
J(y, h) = 2
_
b
a
_
_
b
a
K(s, t)y(s)ds +y(t) f(t)
_
h(t)dt
Thus
_
b
a
K(s, t)y(s)ds +y(t) f(t) = 0
This is a Fredholm integral equation (see Chapter 4) for y.
7. The Euler equation is
((1 +x)y
= 0
or
y
= c
1
/(1 +x)
Integrating again
y(x) = c
1
ln(1 +x) +c
2
3.3. THE SIMPLEST PROBLEM 47
If y(0) = 0, y(1) = 1 then we get
y(x) =
ln(1 +x)
ln2
If the boundary condition at x = 1 is changed to y
d
dt
(L y
L
y
) = L
t
dL
dt
+y
dL
y
dt
+y
L
y
= L
t
L
t
L
y
y
L
y
y
+y
dL
y
dt
+y
L
y
= y
(L
y
d
dt
L
y
)
10. The minimal surface of revolution is found by minimizing
J(y) =
_
b
a
2y
_
1 + (y
)
2
dx
Because the integrand does not depend explicitly on x, a rst integral is
L y
L
y
= c
Upon expanding and simplifying, this equation leads to
dy
dx
=
_
k
2
y
2
1
48 CHAPTER 3. CALCULUS OF VARIATIONS
Now separate variables and integrate while using the fact that
d
du
cosh
1
u =
1
u
2
1
.
11. The Euler equation becomes
x
2
y
+ 2xy
y = 0
which is a CauchyEuler equation. Its solution is
y(x) = Ax
(1+
5)/2
+Bx
(1
5)/2
(recall that a CauchyEuler equation can be solved by trying power func
tions, y = x
m
for some m).
12. To nd Eulers equation, use the fact that L is independent of x and
therefore L y
L
y
= C.
13. Using the Euler equation it is straight forward to see that the extremal
is Y = 0, giving the value J(Y ) = 4. Take, for example, y = sin x.
Then J(y) > 4. Hence, Y = 0 does not give a local maximum. Because
the extremals are only necessary conditions, there is no guarantee that
Y = 0 provides a minimum either. Note also that, for any y, we have
J(y) =
_
1
0
[y
2
+ (y
2)
2
] dx 0.
14. Substitution of r into the integral gives the variational problem
E(y) = 2
_
T
0
e
t
_
y y
dt max.
The Euler equation is
1
y y
d
dt
1
y y
= 0.
This simplies to
y
+ ( + 2 2)y
+ 2(1 )y = 0,
which is a linear equation with constant coecients.
3.4 Generalizations
Exercises, page 184
1. (a) The Euler equations are
8y
1
y
2
= 0, 2y
2
y
1
= 0.
3.4. GENERALIZATIONS 49
Eliminating y
2
gives
y
(4)
1
16y
1
= 0.
The characteristic equation is m
4
16 = 0, which has roots m =
2, 2i. Therefore the solution is
y
1
(x) = ae
2x
+be
2x
+c cos 2x +d sin2x,
where a, b, c, d are arbitrary constants. Then y
2
= 0.5y
1
, and the four
constants a, b, c, d can be computed from the boundary conditions.
(b) The Euler equation is
y
(4)
= 0,
and thus the extremals are
y(x) = a +bx +cx
2
+dx
3
.
The four constants a, b, c, d can be computed from the boundary con
ditions.
(c) The Euler equation is y
(4)
2y
+y = 0.
(d) The Euler equation is y
(4)
= 0.
(e) The Euler equation is y
(4)
y
(3)
y
y = 0.
2. The Euler equation for J(y) =
_
L(x, y, y
, y
)dx is
L
y
(L
y
)
+ (L
y
)
= 0
If L
y
= 0 then clearly
L
y
(L
y
)
= const.
If L
x
= 0, then expand all the derivatives and use the Euler equation to
show
d
dx
(L y
(L
y
(L
y
) y
L
y
)) = 0
3. The Euler equation is
L
M
d
dx
L
M
= 2a(aM M
b) +
d
dx
2(aM M
) = 0,
or
M
a
2
M = ab.
The general solution is
M(t) = Ae
at
+Be
at
+
b
a
.
The left boundary condition is M(0) = M
0
; the right boundary condition
is the natural boundary condition L
M
= 0 at t = T, or
aM(T) M
(T) = b.
50 CHAPTER 3. CALCULUS OF VARIATIONS
4. The two Euler equations (expanded out) are
L
y
y
y
+L
y
z
z
= 0
L
z
y
y
+L
y
z
z
= 0
It is given that the determinant of the coecient matrix of this system is
nonzero. Therefore the only solution is y
= z
+ 2y
+y = 0,
giving extremals
y(x) = ae
x
+bxe
x
.
The boundary conditions are y(0) = 1 and y
(3) = 0.
(c) The Euler equation is
x
2
y
+ 2xy
+
1
4
y = 0,
which is a CauchyEuler equation. Assuming solutions of the form
y = t
m
gives the characteristic equation
m(m1) + 2m+
1
4
= 0,
which has a real double root m =
1
2
. Thus the general solution is
y(x) = a
1
x
+b
1
x
lnx.
The given boundary condition is y(1) = 1; the natural boundary
condition at x = e is y
2y
= 1.
(e) The extremals are
y(x) = ax +b.
The boundary conditions are y(0) = 1 and y
(1) +y(1) = 0.
3.5. THE CANONICAL FORMALISM 51
6. The natural boundary condition is
L
y
(b, y(b), y
(b)) +G
(y(b)) = 0.
7. The Lagrangian is L =
_
1 k
2
+y
2
ky
_
1 k
2
y
2
+k = C.
Solving gives y(x) = Ax +B. Now y(0) = 0 forces B = 0 and y(x) = Ax.
Now apply the natural boundary condition L
y
= 0 at x = b. We get
y
_
1 k
2
y
2
+k =
A
1 k
2
A
+k = 0.
This gives
A =
_
k
2
(1 k
2
)
1 +k
2
.
8. The natural boundary condition is (see Problem 6)
9y
(2) + 9y(2) = 2.
3.5 The Canonical Formalism
Exercises, page 196
1. The Hamiltonian is
H(t, y, p) =
p
2
4r(t)
q(t)y
2
Hamiltons equations are
y
=
p
2r(t)
, p
= 2q(t)y
2. Here we have
J(y) =
_
_
(t
2
+y
2
)(1 + (y
)
2
)
We nd
p = L
y
=
_
t
2
+y
2
y
_
1 + (y
)
2
which yields
(y
)
2
=
p
2
t
2
+y
2
p
2
52 CHAPTER 3. CALCULUS OF VARIATIONS
The Hamiltonian simplies, after some algebra, to
H(t, y, p) =
_
t
2
+y
2
p
2
Then Hamiltons equations are
dy
dt
=
p
_
t
2
+y
2
p
2
,
dp
dt
=
y
_
t
2
+y
2
p
2
Dividing the two equations gives
dy
dp
=
p
y
Integrating yields
y
2
p
2
= const
These are hyperbola in the yp phase plane.
3. Hamiltons equations for the pendulum are
=
p
ml
2
, p
= mgl sin
4. (a) The potential energy is the negative integral of the force, or
V (y) =
_
(
2
y +ay
2
)dy =
1
2
2
y
2
1
3
ay
3
The Lagrangian is L =
1
2
m(y
)
2
V (y). The Euler equation coincides
with Newtons second law:
my
=
2
y +ay
2
(b) The momentum is p = L
y
= my
, which gives y
= p/m. The
Hamiltonian is
H(y, p) =
1
2
(p/m)
2
+V (y)
which is the kinetic plus the potential energy, or the total energy of
the system. Yes, energy is conserved (L is independent of time).
(c) We have H = E for all time t, so at t = 0 we have
1
2
(p(0)/m)
2
+V (y(0)) =
2
10
If y(0) = 0 we can solve for p(0) to get the momentum at time zero;
but this gives
y
(0) =
_
2
/5m
3.5. THE CANONICAL FORMALISM 53
(d) The potential energy has a local maximum at y =
2
/a and is equal
to
V
max
= V (
2
/a) =
6
6a
2
Note that V = 0 at y = 0, 3
2
/2a. If E < V
max
then we obtain
oscillatory motion; if E > V
max
, then the motion is not oscillatory.
Observe that the phase diagram (the solution curves in ypspace, or
phase space), can be found by graphing
p =
_
2m(E V (y)
for various constants E.
5. The Euler equations are
F
y
d
dt
F
y
= 0, F
p
d
dt
F
p
= 0,
where F = py
d
dt
p = 0, y
H
p
= 0.
6. Here the force is F(t, y) = ke
t
/y
2
. We can dene a potential by V (t, y) =
_
F(t, y)dy = ke
t
/y. The Lagrangian is
L(t, y, y
) =
m
2
(y
)
2
V (t, y)
The Euler equation is
my
ke
t
/y
2
= 0
which is Newtons second law of motion. The Hamiltonian is
H(t, y, p) =
p
2
2m
+
k
y
e
t
which is the total energy. Is energy conserved? We can compute dH/dt
to nd
dH
dt
= ke
t
/y = 0
So the energy is not constant.
7. The kinetic energy is T = m(y
)
2
/2. The force is mg (with a plus sign since
positive distance is measured downward). Thus V (y) = mgy. Then the
Lagrangian is L = m(y
)
2
/2 +mgy.
8. We have, for example,
L
xi
d
dt
L
x
i
=
V
x
i
d
dt
(mx
i
) = 0,
or
mx
i
= F
i
.
54 CHAPTER 3. CALCULUS OF VARIATIONS
9. We have r
)
2
+r
2
(
)
2
) =
m
2
(
2
+ (l t)
2
(
)
2
),
and the potential energy is
V = mgh = mg(l r cos ) = mg(l (l t) cos .)
The Lagrangian is L = T V . The Euler equation, or the equation of
motion, is
g sin + (l t)
2 = 0.
One can check that the Hamiltonian is not the same as the total energy;
energy is not conserved in this system. If fact, one can verify that
d
dt
(T +V ) = mgcos = 0.
10. Follow the instructions.
11. The EmdenFowler equation is
y
+
2
t
y
+y
5
= 0.
Multiply by the integrating factor t
2
to write the equation in the form
(t
2
y
+t
2
y
5
= 0.
Now we can identify this with the Euler equation:
L
y
= t
2
y
5
, L
y
= t
2
y
.
Integrate these two equations to nd
L =
1
2
t
2
(y
)
2
1
6
t
2
y
6
+(t)
12. Multiply y
+ay
+be
at
= 0
Now identify the terms in this equation with the terms in the Euler equa
tion as in Exercise 5.14. Finally we arrive at a Lagrangian
L = e
at
_
m
2
(y
)
2
by
_
There are many Lagrangians, and we have chosen just one by selecting
the arbitrary functions.
3.6. ISOPERIMETRIC PROBLEMS 55
13. Follow Example 3.27 in the book with m, a, k and y replaced by L, R, C
1
and I, respectively.
14. Multiplying the given equation by expP(t) makes the sum of the rst two
terms a total derivative and we get
d
dt
(e
P(t)
y
) +e
P(t)
f(y) = 0.
Comparing to the Euler equation we take
L
y
= e
P(t)
f(y), L
y
= e
P(t)
y
.
Integrating the rst gives
L = e
P(t)
F(y) +(t, y
),
where is arbitrary. Plugging this into the second equation we get
y
= e
P(t)
y
.
Integrating,
=
1
2
y
2
e
P(t)
+(t),
where is arbitrary. Thus,
L = e
P(t)
_
F(y)
1
2
y
2
_
+(t).
3.6 Isoperimetric Problems
Exercises, page 203
1. Form the Lagrangian
L
= (y
)
2
+y
2
.
The Euler equation reduces to
y
y = 0, y(0) = y() = 0
The extremals are given by
y
n
(x) =
_
2/ sinnx, n = 1, 2, 3, . . . .
2. Let L
= x
2
+ (y
)
2
+y
2
. Then the Euler equation for L
is
y
+y = 0, y(0) = y(1) = 0.
If 0 then this boundary value problem has only trivial solutions. If
< 0, say =
2
, then the problem has nontrivial solutions
y
n
(x) = B
n
sinnx, n = 1, 2, . . . ,
56 CHAPTER 3. CALCULUS OF VARIATIONS
where the B
n
are constants. Applying the constraint gives
_
1
0
B
2
n
sin
2
nx dx = 2.
Thus B
n
= 2 for all n. So the extremals are
y
n
(x) = 2 sin nx, n = 1, 2, . . . .
3. The Euler equations become
L
y
1
d
dx
L
y
1
= 0, L
y
2
d
dx
L
y
2
= 0,
where L
= L +G.
4. Form the Lagrangian
L
= xy
yx
+
_
(x
)
2
+ (y
)
2
.
Because the Lagrangian does not depend explicitly on t, a rst integral is
given by
L
x
y
L
y
= C.
5. The problem is to minimize
J(y) =
_
1
0
_
1 + (y
)
2
dx, y(0) = y(1) = 0
subject to the constraint
_
1
0
y(x)dx = A.
Form the Lagrangian
L
=
_
1 + (y
)
2
+y
Because the Lagrangian does not depend explicitly on x, a rst integral is
given by
L
L
y
= c.
Expanding out this equation leads to
y
1 (y c)
2
(y c)
2
.
Separating variables and integrating, and then using the substitution u =
1 (y c)
2
, gives
1
2
_
du
u
= x +c
1
.
3.6. ISOPERIMETRIC PROBLEMS 57
Thus
(x +c
1
)
2
+ (y c/)
2
=
1
2
,
which is a circle. Now the two boundary conditions and the constraint
give three equations for the constants c, c
1
, .
6. Form the Lagrangian
L
= p(y
)
2
+qy
2
+ry
2
.
Then the Euler equation corresponding to L
is
L
y
(L
y
)
= 0,
or
(py
)
_
b
a
L(t, y, g(t, y), y
, g
t
+g
y
y
)dt.
Now form the Euler equation for F. We get
F
y
d
dt
F
y
= 0,
or, in terms of L,
L
y
+L
z
g
y
+L
z
(g
ty
+g
yy
y
)
d
dt
(L
y
+L
z
g
y
) = 0.
This simplies to
L
y
d
dt
L
y
+g
y
(L
z
d
dt
L
z
= 0.
Also G(t, y, g(t, y)) = 0, and taking the partial with respect to y gives
G
y
+G
z
g
y
= 0.
Thus
L
y
d
dt
L
y
G
y
=
L
z
d
dt
L
z
G
y
.
Now these two expressions must be equal to the same function of t, that
is
L
y
d
dt
L
y
= (t)G
y
, L
z
d
dt
L
z
= (t)G
z
.