This action might not be possible to undo. Are you sure you want to continue?
Student’s Manual
Further Mathematics for
Economic Analysis
2
nd
edition
Arne Strøm
Knut Sydsæter
Atle Seierstad
Peter Hammond
For further supporting resources please visit:
www.pearsoned.co.uk/sydsaeter
Preface
This student’s solutions manual accompanies Further Mathematics for Economic Analysis (2nd edition, FT
Prentice Hall, 2008). Its main purpose is to provide more detailed solutions to the problems marked with
⊂SM ⊃
in the text. The Manual should be used in conjunction with the answers in the book. In some few cases only
a part of the problem is done in detail, because the rest follows the same pattern.
At the end of this manual there is a list of misprints and other errors in the book, and even one or two in
the errata list in the preliminary and incomplete version of this manual released in September this year. We
would greatly appreciate suggestions for improvements from our readers as well as help in weeding out the
inaccuracies and errors that probably remain.
Oslo and Coventry, October 2008
Arne Strøm (arne.strom@econ.uio.no)
Knut Sydsæter (knut.sydsater@econ.uio.no)
Atle Seierstad (atle.seierstad@econ.uio.no)
Peter Hammond (hammond@stanford.edu)
Contents
1 Topics in Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Multivariable Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3 Static Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4 Topics in Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5 Differential Equations I: FirstOrder Equations in One Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6 Differential Equations II: SecondOrder Equations and Systems in the Plane . . . . . . . . . . . . . . . . . . 49
7 Differential Equations III: HigherOrder Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
8 Calculus of Variations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
9 Control Theory: Basic Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
10 Control Theory with Many Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
11 Difference Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
12 Discrete Time Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
13 Topology and Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
14 Correspondences and Fixed Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
A Sets, Completeness, and Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
B Trigonometric Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Corrections to the book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Version 1.0 28102008 879
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 1 TOPI CS I N L I NEAR AL GEBRA 1
Chapter 1 Topics in Linear Algebra
1.2
1.2.3 Let A =
⎛
⎝
1 2 0
0 1 1
1 0 1
⎞
⎠
be a matrix with the three given vectors as columns. Cofactor expansion of
[A[ along the ﬁrst row yields
1 2 0
0 1 1
1 0 1
= 1
1 1
0 1
−2
0 1
1 1
÷0 = 1 −2(−1) = 3 ,= 0
By Theorem 1.2.1 this shows that the given vectors are linearly independent.
1.2.6 Part (a) is just a special case of part (b), so we will only prove (b). To show that a
1
, a
2
, . . . , a
n
are
linearly independent it sufﬁces to show that if c
1
, c
2
, . . . , c
n
are real numbers such that
c
1
a
1
÷c
2
a
2
÷· · · ÷c
n
a
n
= 0
then all the c
i
have to be zero. So suppose that we have such a set of real numbers. Then for each i = 1,
2, . . . , n, we have
a
i
· (c
1
a
1
÷c
2
a
2
÷· · · ÷c
n
a
n
) = a
i
· 0 = 0 (1)
Since a
i
·a
j
= 0 when i ,= j, the lefthand side of (1) reduces to a
i
·(c
i
a
i
) = c
i
a
i

2
. Hence, c
i
a
i

2
= 0.
Because a
i
,= 0 we have a
i
 ,= 0, and it follows that c
i
= 0.
1.3
1.3.1 (a) The rank is 1. See the answer in the book.
(b) The minor formed from the ﬁrst two columns in the matrix is
1 3
2 0
= −6 ,= 0. Since this minor
is of order 2, the rank of the matrix must be at least 2, and since the matrix has only two rows, the rank
cannot be greater than 2, so the rank equals 2.
(c) The ﬁrst two rows and last two columns of the matrix yield the minor
−1 3
−4 7
= 5 ,= 0, so the rank
of the matrix is at least 2. On the other hand, all the four minors of order 3 are zero, so the rank is less
than 3. Hence the rank is 2. (It can be shown that r
2
= 3r
1
÷r
3
, where r
1
, r
2
, and r
3
are the rows of the
matrix.)
An alternative argument runs as follows: The rank of a matrix does not change if we add a multiple
of one row to another row, so
⎛
⎝
1 2 −1 3
2 4 −4 7
−1 −2 −1 −2
⎞
⎠
−2 1
←
←
∼
⎛
⎝
1 2 −1 3
0 0 −2 1
0 0 −2 1
⎞
⎠
.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
2 CHAPTER 1 TOPI CS I N L I NEAR AL GEBRA
Here ∼ means that the last matrix is obtained from the ﬁrst one by elementary row operations. The last
matrix obviously has rank 2, and therefore the original matrix also has rank 2.
(d) The ﬁrst three columns of the matrix yield the minor
1 3 0
2 4 0
1 −1 2
= −4 ,= 0, so the rank is 3.
(e)
2 1
−1 4
= 9 ,= 0, so the rank is at least 2. All the four minors of order 3 are zero, so the rank must
be less than 3. Hence the rank is 2. (The three rows, r
1
, r
2
, and r
3
, of the matrix are linearly dependent,
because r
2
= −14r
1
÷9r
3
.)
(f) The determinant of the whole matrix is zero, so the rank must be less than 4. On the other hand, the
ﬁrst three rows and the ﬁrst three columns yield the minor
1 −2 −1
2 1 1
−1 1 −1
= −7 ,= 0
so the rank is at least 3. It follows that the matrix has rank 3.
1.3.2 (a) The determinant is (x ÷1)(x −2). The rank is 3 if x ,= −1 and x ,= 2. The rank is 2 if x = −1
or x = 2.
(a) By cofactor expansion along the ﬁrst row, the determinant of the matrix A =
⎛
⎝
x 0 x
2
−2
0 1 1
−1 x x −1
⎞
⎠
is
[A[ = x · (−1) −0 · 1 ÷(x
2
−2) · 1 = x
2
−x −2 = (x ÷1)(x −2)
If x ,= −1 and x ,= 2, then [A[ ,= 0, so the rank of A equals 3. If x = −1 or x = 2, then [A[ = 0 and
r(A) ≤ 2. On the other hand, the minor we get if we strike out the ﬁrst row and the third column in A is
0 1
−1 x
= 1 ,= 0 for all x, so r(A) can never be less than 2.
Conclusion: r(A) =
¸
2 if x = −1 or x = 2
3 otherwise
(b) A little calculation shows that the determinant of the matrix is t
3
÷4t
2
−4t −16, and if we note that
this expression has t ÷4 as a factor, it follows that the determinant is
t
3
÷4t
2
−4t −16 = t
2
(t ÷4) −4(t ÷4) = (t
2
−4)(t ÷4) = (t ÷2)(t −2)(t ÷4)
Thus, if t does not equal any of the numbers −2, 2, and −4, the rank of the matrix is 3.
If we strike out the second rowand the ﬁrst column of the matrix, we get the minor
5 6
1 t ÷4
= 5t ÷14,
which is different from 0 for all the three special values of t that we found above, and thus the rank of
the matrix is
¸
2 if t = −4, −2, or 2
3 otherwise
(c) The ﬁrst and third rows are identical, as are the second and fourth. But the ﬁrst two rows are always
linearly independent. So the rank is 2 for all values of x, y, z, and w.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 1 TOPI CS I N L I NEAR AL GEBRA 3
1.4
1.4.2 (a) It is clear that x
1
= x
2
= x
3
= 0 is a solution. The determinant of the coefﬁcient matrix is
D =
1 −1 1
1 2 −1
2 1 3
= 9 ,= 0
Hence, by Cramer’s rule the solution is unique and the system has 0 degrees of freedom. (This agrees
with Theorem 1.4.2: since the rank of the coefﬁcient matrix is 3 and there are 3 unknowns, the system
has 3 −3 = 0 degrees of freedom.)
(b) By Gaussian elimination (or other means) we ﬁnd that the solution is x
1
= a, x
2
= −a, x
3
= −a,
and x
4
= a, with a arbitrary. Thus, there is one degree of freedom. (We could get the number of degrees
of freedom from Theorem 1.4.2 in this case, too. The minor formed from the ﬁrst three columns of the
coefﬁcient matrix has determinant −3 ,= 0, so the rank of the coefﬁcient matrix is 3. Since there are 4
unknowns, the system has 4 −3 = 1 degree of freedom.)
1.4.3 The determinant of the coefﬁcient matrix is a
2
−7a = a(a −7). Thus, if a ,= 0 and a ,= 7, the system
has a unique solution. If a = 0 or a = 7, the rank of the coefﬁcient matrix is 2 (why?), and the system
either has solutions with 1 degree of freedom or has no solutions at all, depending on the value of b. One
way of ﬁnding out is by Gaussian elimination, i.e. by using elementary row operations on the augmented
coefﬁcient matrix:
⎛
⎝
1 2 3 1
−1 a −21 2
3 7 a b
⎞
⎠
1 −3
←
←
∼
⎛
⎝
1 2 3 1
0 a ÷2 −18 3
0 1 a −9 b −3
⎞
⎠
←
−(a ÷2)
∼
⎛
⎝
1 2 3 1
0 0 −a
2
÷7a −ab −2b ÷3a ÷9
0 1 a −9 b −3
⎞
⎠
←
←
∼
⎛
⎝
1 2 3 1
0 1 a −9 b −3
0 0 a(7 −a) −ab −2b ÷3a ÷9
⎞
⎠
This conﬁrms that as long as a ,= 0 and a ,= 7, the system has a unique solution for any value of b. But if
a = 0 or a = 7, the system has solutions if and only if −ab −2b ÷3a ÷9 = 0, and then it has solutions
with 1 degree of freedom. (The rank of the coefﬁcient matrix is 2 and there are 3 unknowns.)
If a = 0, then the system has solutions if and only if b = 9/2.
If a = 7, then the system has solutions if and only if −9b ÷30 = 0, i.e. b = 10/3.
1.4.6 (a) [A
t
[ = (t −2)(t ÷3), so r(A
t
) = 3 if t ,= 2 and t ,= −3. Because the upper left 2 2 minor of
A
t
is −1 ,= 0, the rank of A
t
can never be less than 2, so r(A
2
) = 2, r(A
−3
) = 2.
(b) Let x =
⎛
⎝
x
1
x
2
x
3
⎞
⎠
. The vector equation A
−3
x =
⎛
⎝
11
3
6
⎞
⎠
is equivalent to the equation system
x
1
÷ 3x
2
÷2x
3
= 11 (1)
2x
1
÷ 5x
2
−3x
3
= 3 (2)
4x
1
÷10x
2
−6x
3
= 6 (3)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
4 CHAPTER 1 TOPI CS I N L I NEAR AL GEBRA
Equation (3) is obviously equivalent to (2), so we can remove it, and then we are left with the two equations
x
1
÷3x
2
÷2x
3
= 11 (1)
2x
1
÷5x
2
−3x
3
= 3 (2)
We can consider these equations as an equation system with x
1
and x
2
as the unknowns:
x
1
÷3x
2
= 11 −2x
3
(1
/
)
2x
1
÷5x
2
= 3 ÷3x
3
(2
/
)
For each value of x
3
this system has the unique solution x
1
= 19x
3
− 46, x
2
= −7x
3
÷ 19. Thus the
vector equation A
3
x = (11, 3, 6)
/
has the solution
x = (19s −46, −7s ÷19, s)
/
where s runs through all real numbers.
1.5
1.5.1 For convenience, let A, B, . . . , F denote the matrices given in (a), (b), . . . , (f), respectively.
(a) The characteristic polynomial of the matrix A =
2 −7
3 −8
is
[A −λI[ =
2 −λ −7
3 −8 −λ
= λ
2
÷6λ ÷5 = (λ ÷1)(λ ÷5)
so the eigenvalues of A are λ
1
= −1 and λ
2
= −5. The eigenvectors corresponding to an eigenvalue λ
are the vectors x =
x
y
,= 0 that satisfy Ax = λx, i.e.
2x −7y = −x
3x −8y = −y
¸
⇐⇒ 3x = 7y for λ = −1
and
2x −7y = −5x
3x −8y = −5y
¸
⇐⇒ 7x = 7y for λ = −5
This gives us the eigenvectors v
1
= s
7
3
and v
2
= t
1
1
, where s and t are arbitrary real numbers
(different from 0).
(b) The characteristic equation of B is [B − λI[ =
2 −λ 4
−2 6 −λ
= λ
2
− 8λ ÷ 20 = 0. B has two
complex eigenvalues, 4 ±2i, and no real eigenvalues.
(c) The characteristic polynomial of Cis [C−λI[ = λ
2
−25, and we see immediately that the eigenvalues
are λ
1
= 5 and λ
2
= −5. The eigenvectors are determined by the equation systems
x ÷4y = 5x
6x − y = 5y
¸
⇐⇒ x = y and
x ÷4y = −5x
6x − y = −5y
¸
⇐⇒ y = −
3
2
x
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 1 TOPI CS I N L I NEAR AL GEBRA 5
respectively, so the eigenvectors are
v
1
= s
1
1
and v
2
= t
−2
3
where s and t are arbitrary nonzero numbers.
(d) The characteristic polynomial of D is
[D −λI[ =
2 −λ 0 0
0 3 −λ 0
0 0 4 −λ
= (2 −λ)(3 −λ)(4 −λ)
The eigenvalues are obviously λ
1
= 2, λ
2
= 3, λ
3
= 4, and the corresponding eigenvectors are
v
1
= s
⎛
⎝
1
0
0
⎞
⎠
, v
2
= t
⎛
⎝
0
1
0
⎞
⎠
, v
3
= u
⎛
⎝
0
0
1
⎞
⎠
where s, t , u are arbitrary nonzero numbers. (Note: The eigenvalues of a diagonal matrix are always
precisely the diagonal elements, and (multiples of) the standard unit vectors will be eigenvectors. But if
two or more of the diagonal elements are equal, there will be other eigenvectors as well. An extreme case
is the identity matrix I
n
: all (nonzero) nvectors are eigenvectors for I
n
.)
(e) The characteristic polynomial of E is
2 −λ 1 −1
0 1 −λ 1
2 0 −2 −λ
= −λ
3
÷λ
2
÷2λ = −λ(λ
2
−λ −2)
The eigenvalues are the roots of the equation −λ(λ
2
−λ−2) = 0, namely λ
1
= −1, λ
2
= 0 and λ
3
= 2.
The eigenvectors corresponding to λ
1
= −1 are solutions of
Ex = −x ⇐⇒
⎧
⎪
⎨
⎪
⎩
2x
1
÷x
2
− x
3
= −x
1
x
2
÷ x
3
= −x
2
2x
1
−2x
3
= −x
2
⇐⇒
¸
x
1
=
1
2
x
3
x
2
= −
1
2
x
3
⇐⇒
¸
x
2
= −x
1
x
3
= 2x
1
so they are of the form v
1
= s
⎛
⎝
1
−1
2
⎞
⎠
. Similarly, v
2
= t
⎛
⎝
1
−1
1
⎞
⎠
and v
3
= u
⎛
⎝
2
1
1
⎞
⎠
are the eigenvectors
corresponding to λ
2
= 0 and λ
3
= 2.
(f) The characteristic polynomial of F is
1 −λ −1 0
−1 2 −λ −1
0 −1 1 −λ
= −λ
3
÷4λ
2
−3λ = −λ(λ
2
−4λ ÷3) = −λ(λ −1)(λ −3)
The eigenvalues are therefore λ
1
= 0, λ
2
= 1, and λ
3
= 3. By the same method as above we ﬁnd that
the corresponding eigenvectors are v
1
= s
⎛
⎝
1
1
1
⎞
⎠
, v
2
= t
⎛
⎝
−1
0
1
⎞
⎠
, and v
3
= u
⎛
⎝
1
−2
1
⎞
⎠
.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
6 CHAPTER 1 TOPI CS I N L I NEAR AL GEBRA
1.5.2 (a) X
/
AX = X
/
(AX) = (x, y, z)
⎛
⎝
ax ÷ay
ax ÷ay
bz
⎞
⎠
= (ax
2
÷ay
2
÷2axy ÷bz
2
) (a 1 1 matrix),
A
2
=
⎛
⎝
2a
2
2a
2
0
2a
2
2a
2
0
0 0 b
2
⎞
⎠
, A
3
=
⎛
⎝
4a
3
4a
3
0
4a
3
4a
3
0
0 0 b
3
⎞
⎠
(b) The characteristic polynomial of A is p(λ) =
a −λ a 0
a a −λ 0
0 0 b −λ
= (λ
2
−2aλ)(b −λ), so the
eigenvalues of A are λ
1
= 0, λ
2
= 2a, λ
3
= b.
(c) From (b) we get p(λ) = −λ
3
÷ (2a ÷ b)λ
2
− 2abλ. Using the expressions for A
2
and A
3
that we
found in part (a), it is easy to show that p(A) = −A
3
÷(2a ÷b)A
2
−2abA = 0.
1.5.4 (a) The formula in Problem 1.9.7(b) yields
[A −λI[ =
4 −λ 1 1 1
1 4 −λ 1 1
1 1 4 −λ 1
1 1 1 4 −λ
= (3 −λ)
4
1 ÷
4
3 −λ
= (3 −λ)
3
(7 −λ)
Hence, the eigenvalues of A are λ
1
= λ
2
= λ
3
= 3, λ
4
= 7.
(b) An eigenvector x = (x
1
, x
2
, x
3
, x
4
)
/
of A corresponding to the eigenvalue λ = 3 must satisfy the
equation system (A −3I)x = 0. The 4 equations in this system are all the same, namely
x
1
÷x
2
÷x
3
÷x
4
= 0
The system has solutions with 4 −1 = 3 degrees of freedom. One simple set of solutions is
x
1
=
⎛
⎜
⎜
⎝
1
−1
0
0
⎞
⎟
⎟
⎠
, x
2
=
⎛
⎜
⎜
⎝
1
0
−1
0
⎞
⎟
⎟
⎠
. x
3
=
⎛
⎜
⎜
⎝
1
0
0
−1
⎞
⎟
⎟
⎠
These three vectors are obviously linearly independent because if
c
1
x
1
÷c
2
x
2
÷c
3
x
3
=
⎛
⎜
⎜
⎝
c
1
÷c
2
÷c
3
−c
1
−c
2
−c
3
⎞
⎟
⎟
⎠
is the zero vector, then c
1
= c
2
= c
3
= 0.
1.5.5 (c) If λ is an eigenvalue for C with an associated eigenvector x, then C
n
x = λ
n
x for every natural
number n. If C
3
= C
2
÷C, then λ
3
x = λ
2
x÷λx, so (λ
3
−λ
2
−λ)x = 0. Then λ
3
−λ
2
−λ = 0, because
x ,= 0. If C÷I
n
did not have an inverse, [C÷I
n
[ = 0. Then λ = −1 would be an eigenvalue for C, and
so we would have λ
3
−λ
2
−λ = 0, which is not true for λ = −1. Hence −1 is not an eigenvalue for C,
and consequently C ÷I
n
has an inverse.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 1 TOPI CS I N L I NEAR AL GEBRA 7
1.6
1.6.1 (a) Let A =
2 1
1 2
. The characteristic polynomial of A is
p(λ) =
2 −λ 1
1 2 −λ
= (2 −λ)
2
−1 = λ
2
−4λ ÷3 = (λ −1)(λ −3)
Thus, the eigenvalues are λ
1
= 1 and λ
2
= 3. The associated eigenvectors with length 1 are uniquely
determined up to sign as
x
1
=
1
2
√
2
−
1
2
√
2
and x
2
=
1
2
√
2
1
2
√
2
This yields the orthogonal matrix
P =
1
2
√
2
1
2
√
2
−
1
2
√
2
1
2
√
2
(It is easy to verify that P
/
P = I, i.e. P
−1
= P
/
.) We therefore have the diagonalization
P
−1
AP =
1
2
√
2 −
1
2
√
2
1
2
√
2
1
2
√
2
2 1
1 2
1
2
√
2
1
2
√
2
−
1
2
√
2
1
2
√
2
=
1 0
0 3
(b) Let B =
⎛
⎝
1 1 0
1 1 0
0 0 2
⎞
⎠
. The characteristic polynomial of B is
1 −λ 1 0
1 1 −λ 0
0 0 2 −λ
= (2 −λ)
(1 −λ)
2
−1
= (2 −λ)(λ
2
−2λ) = −λ(λ −2)
2
(use cofactor expansion of the ﬁrst determinant along the last row or the last column). The eigenvalues
are λ
1
= 0 and λ
2
= λ
3
= 2. It is easily seen that one eigenvector associated with the eigenvalue
λ
1
= 0 is x
1
= (1, −1, 0)
/
. Eigenvectors x = (x
1
, x
2
, x
3
)
/
associated with the eigenvalue 2 are given by
(B −2I)x = 0, i.e.
−x
1
÷x
2
= 0
x
1
−x
2
= 0
0 = 0
This gives x
1
= x
2
, x
3
arbitrary. One set of linearly independent eigenvectors with length 1 is then
1
√
2
⎛
⎝
1
−1
0
⎞
⎠
,
1
√
2
⎛
⎝
1
1
0
⎞
⎠
,
⎛
⎝
0
0
1
⎞
⎠
Fortunately, these three vectors are mutually orthogonal (this is not automatically true for two eigenvectors
associated with the same eigenvalue), and so we have a suitable orthogonal matrix
P =
⎛
⎜
⎝
1
2
√
2
1
2
√
2 0
−
1
2
√
2
1
2
√
2 0
0 0 1
⎞
⎟
⎠
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
8 CHAPTER 1 TOPI CS I N L I NEAR AL GEBRA
It is now easy to verify that P
−1
= P
/
and
P
−1
BP =
⎛
⎝
0 0 0
0 2 0
0 0 2
⎞
⎠
(c) The characteristic polynomial of C =
⎛
⎝
1 3 4
3 1 0
4 0 1
⎞
⎠
is
1 −λ 3 4
3 1 −λ 0
4 0 1 −λ
= 4
3 1 −λ
4 0
÷(1 −λ)
1 −λ 3
3 1 −λ
= (1 −λ)
(1 −λ)
2
−25
(cofactor expansion along the last column), and so the eigenvalues are λ
1
= 1, λ
2
= 6, and λ
3
= −4. An
eigenvector x = (x, y, z)
/
corresponding to the eigenvalue λ must satisfy
Cx = λx ⇐⇒
x ÷3y ÷4z = λx
3x ÷ y = λy
4x ÷ z = λz
One set of unnormalized eigenvectors is
u =
⎛
⎝
0
−4
3
⎞
⎠
, v =
⎛
⎝
5
3
4
⎞
⎠
, w =
⎛
⎝
−5
3
4
⎞
⎠
with lengths u = 5, v = w = 5
√
2, and a corresponding orthogonal matrix is
P =
⎛
⎜
⎝
0
5
10
√
2 −
5
10
√
2
−
4
5
3
10
√
2
3
10
√
2
3
5
4
10
√
2
4
10
√
2
⎞
⎟
⎠
A straightforward calculation conﬁrms that P
/
CP = diag(1, 6, −4) = diag(λ
1
, λ
2
, λ
3
).
1.6.5 For the given A, we have A
2
= 5A−5I. Therefore A
3
= A
2
A = (5A−5I)A = 5A
2
−5A = 20A−25I
and A
4
= 20A
2
−25A = 75A −100I =
50 75
75 125
.
1.7
1.7.5 (a) It is clear that Q(x
1
, x
2
) ≥ 0 for all x
1
and x
2
and that Q(x
1
, x
2
) = 0 only if x
1
= x
2
= 0, so Q
is positive deﬁnite.
(b) The symmetric coefﬁcient matrix of Q is
⎛
⎝
5 0 1
0 2 1
1 1 4
⎞
⎠
. The leading principal minors are
D
1
= 5, D
2
=
5 0
0 2
= 10, D
3
=
5 0 1
0 2 1
1 1 4
= 33
Since all the leading principal minors are positive, it follows from Theorem 1.7.1 that Q is positive
deﬁnite. (An alternative way to see this is to write Q as a sum of squares: Q(x
1
, x
2
, x
3
) = (x
1
÷x
3
)
2
÷
(x
2
÷x
3
)
2
÷4x
2
1
÷x
2
2
÷2x
2
3
is obviously nonnegative and it is zero only if all the square terms are zero.
But it is not always easy to see how to rewrite a quadratic form in this fashion.)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 1 TOPI CS I N L I NEAR AL GEBRA 9
(c) Since Q(x
1
, x
2
) = −(x
1
− x
2
)
2
≤ 0 for all x
1
and x
2
, Q is negative semideﬁnite. But Q is not
deﬁnite, since Q(x
1
, x
2
) = 0 whenever x
1
= x
2
.
(d) The symmetric coefﬁcient matrix of Q is
⎛
⎝
−3 1 0
1 −1 2
0 2 −8
⎞
⎠
, and the leading principal minors are
D
1
= −3 < 0, D
2
= 2 > 0, and D
3
= −4 < 0. By Theorem 1.7.1, Q is negative deﬁnite.
1.7.7 (a) The symmetric coefﬁcient matrix of Q is
A =
a
11
a
12
a
21
a
22
=
3 −(5 ÷c)/2
−(5 ÷c)/2 2c
Since a
11
> 0, the form can never be negative semideﬁnite. It is
(i) positive deﬁnite if a
11
> 0 and [A[ > 0,
(ii) positive semideﬁnite if a
11
≥ 0, a
22
≥ 0. and [A[ ≥ 0,
(iii) indeﬁnite if [A[ < 0.
The determinant of A is
[A[ = 6c −
1
4
(5 ÷c)
2
= −
1
4
(c
2
−14c ÷25) = −
1
4
(c −c
1
)(c −c
2
)
where c
1
= 7 − 2
√
6 ≈ 2.101 and c
2
= 7 ÷ 2
√
6 ≈ 11.899 are the roots of the quadratic equation
c
2
−14c ÷25 = 0. It follows that
(1) Q is positive deﬁnite if c
1
< c < c
2
, i.e. if c lies in the open interval (c
1
, c
2
);
(2) Q is positive semideﬁnite if c
1
≤ c ≤ c
2
, i.e. if c lies in the closed interval [c
1
, c
2
];
(3) Q is indeﬁnite if c < c
1
or c > c
2
, i.e. if c lies outside the closed interval [c
1
, c
2
].
1.7.10 (b) If x satisﬁes the Lagrange conditions, then Q(x) = x
/
Ax = x
/
(λx) = λx
/
x = λx
2
= λ,
because the constraint is simply x
2
= 1. Hence, the maximum and minimum values of Q(x) are
simply the largest and smallest eigenvalue of Q, which are λ
1
= 9 and λ
2
= −5. The corresponding
maximum and minimum points (eigenvectors) are ±
1
2
√
2 (1, 1) and ±
1
2
√
2 (1, −1), respectively.
1.8
1.8.3 Negative deﬁnite subject to the constraint, by Theorem 1.8.1:
0 0 1 1 1
0 0 4 −2 1
1 4 −5 1 2
1 −2 1 −1 0
1 1 2 0 −2
= −180 <
0.
1.8.4 Positive deﬁnite subject to the constraint, by Theorem 1.8.1:
0 0 1 2 1
0 0 2 −1 −3
1 2 1 1 0
2 −1 1 1 0
1 −3 0 0 1
= 25 > 0.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
10 CHAPTER 1 TOPI CS I N L I NEAR AL GEBRA
1.9
1.9.3 The obvious partitioning to use in (a) is with A
11
as 22 matrix in the upper left corner of A, and in (b)
it is natural to let A
11
be the upper left 11 matrix. In both cases the partitioned matrix will be of the form
A
11
0
0 A
22
, i.e. A
12
and A
21
will both be zero matrices. If we use formula (1.9.4) we get = A
22
,
whereas formula (1.9.5) gives
˜
= A
11
. In either case we ﬁnd that
A
11
0
0 A
22
−1
=
A
−1
11
0
0 A
−1
22
,
and the answers to (a) and (b) are as given in the book.
The matrix in (c) can be partitioned as A =
A
11
A
12
A
21
A
22
=
I
4
v
v
/
I
1
, where I
4
and I
1
are the
identity matrices of orders 4 and 1, respectively, and v
/
= (1, 1, 1, 0) is the transpose of the 4 1
matrix (column vector) v. If we use formula (1.9.4) in the book, we get = I
1
−v
/
v = (−2) = −2I
1
.
Then A
−1
11
÷A
−1
11
A
12
−1
A
21
A
−1
11
= I
4
−
1
2
vv
/
= I
4
−
1
2
⎛
⎜
⎜
⎝
1 1 1 0
1 1 1 0
1 1 1 0
0 0 0 0
⎞
⎟
⎟
⎠
=
1
2
⎛
⎜
⎜
⎝
1 −1 −1 0
−1 1 −1 0
−1 −1 1 0
0 0 0 2
⎞
⎟
⎟
⎠
,
and −
−1
A
21
A
−1
11
=
1
2
I
1
v
/
I
4
=
1
2
v
/
=
1
2
(1, 1, 1, 0). Further, −A
−1
11
A
12
−1
=
1
2
v, so we ﬁnally get
A
−1
=
I
4
−
1
2
vv
/ 1
2
v
1
2
v
/
−
1
2
I
1
=
1
2
⎛
⎜
⎜
⎜
⎜
⎝
1 −1 −1 0 1
−1 1 −1 0 1
−1 −1 1 0 1
0 0 0 2 0
1 1 1 0 −1
⎞
⎟
⎟
⎟
⎟
⎠
.
An alternative partitioning is A =
I
3
W
W
/
I
2
, where W
/
=
0 0 0
1 1 1
. Then = I
2
−WW
/
=
I
2
−
0 0
0 3
=
1 0
0 −2
, so
−1
=
1
2
2 0
0 −1
. A bit of calculation shows that −
−1
W
/
=
1
2
W
and WW
/
=
⎛
⎝
1 1 1
1 1 1
1 1 1
⎞
⎠
, so A
−1
=
I
3
−
1
2
WW
/ 1
2
W
1
2
W
/
−1
=
1
2
2I
3
−WW
/
W
W
/
2
−1
, as before.
(Note that because A is symmetric, A
−1
must also be symmetric. Thus the upper right submatrix of A
−1
must be the transpose of the lower left submatrix, which helps us save a little work.)
1.9.4 The matrix B =
⎛
⎜
⎜
⎝
1 −x
1
. . . −x
n
x
1
a
11
. . . a
1n
.
.
.
.
.
.
.
.
.
.
.
.
x
n
a
n1
. . . a
nn
⎞
⎟
⎟
⎠
can be partitioned as B =
I
1
−X
/
X A
, where A and X
are as given in the problem. We evaluate the determinant of B by each of the formulas (6) and (7): By
formula (6),
[B[ = [I
1
[ · [A −XI
−1
1
(−X
/
)[ = [A ÷XX
/
[
and by formula (7),
[B[ = [A[ · [I
1
−(−X
/
)A
−1
X[ = [A[ · [I
1
÷X
/
A
−1
X[
where the last factor is the determinant of a 1 1 matrix and therefore equal to the single element
1 ÷X
/
A
−1
X of that matrix.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 1 TOPI CS I N L I NEAR AL GEBRA 11
1.9.6 (a) Let A =
A
11
A
12
0 A
22
, where A
11
is a k k matrix and A
22
is (n−k) (n−k). We will consider
two ways to demonstrate that [A[ = [A
11
[ [A
22
[.
(I) By deﬁnition, the determinant of the n n matrix A is a sum
A
of n! terms. Each of these terms
is the product of a sign factor and n elements from A, chosen so that no two elements are taken from
the same row of A or from the same column. The sign factor ±1 is determined by the positions of the
elements selected. (See EMEA or almost any book on linear algebra for details.)
A term in this sum will automatically be zero unless the factors from the ﬁrst k columns are taken
from A
11
and the last n −k factors from A
22
. If the factors are selected in this way, the term in question
will be a product of one term in the sum
1
making up [A
11
[ and one from the sum
2
that makes up
[A
22
[. (The sign factors will match.) All such pairs of terms will occur exactly once and so
A
=
1
2
,
that is, [A[ = [A
11
[ [A
22
[.
(II) Suppose that A is upper triangular, i.e. all elements below the main diagonal are 0. Then A
11
and
A
22
are also upper triangular. We know that the determinant of a triangular matrix equals the product
of the elements on the main diagonal. (Just think of cofactor expansion along the ﬁrst column, then the
second column, etc.) In this case it is clear that [A[ = [A
11
[ [A
22
[.
Of course, in the general case A need not be upper triangular at all, but we can make it so by means
of elementary row operations, more speciﬁcally the operations of (i) adding a multiple of one row to
another row and (ii) interchanging two rows. Operation (i) does not affect the value of the determinant,
while operation (ii) multiplies the value by −1. We perform such operations on the ﬁrst k rows of A in
such a way that A
11
becomes upper triangular, and then operate on the last n − k rows of A such that
A
22
becomes upper triangular. The number σ of sign changes that [A[ undergoes is then the sum of the
numbers σ
1
and σ
2
of sign changes inﬂicted on [A
11
[ and [A
22
[, respectively. By the formula we showed
for the upper triangular case, (−1)
σ
[A[ = (−1)
σ
1
[A
11
[ · (−1)
σ
2
[A
22
[, and since σ
1
÷ σ
2
= σ, we get
[A[ = [A
11
[ [A
22
[ in the general case too.
To show the formula
A
11
0
A
21
A
22
= [A
11
[ [A
22
[, simply look at the determinant
A
/
11
A
/
21
0 A
/
22
of the
transposed matrix.
(b) The equality follows by direct multiplication. By the result in (a), the ﬁrst factor on the left has
determinant [I
k
[ [I
n−k
[ = 1, and so, by (a) again,
A
11
A
12
A
21
A
22
=
A
11
−A
12
A
−1
22
A
21
0
A
21
A
22
= [A
11
−A
12
A
−1
22
A
21
[ [A
22
[
1.9.7 (a) With D =
I
n
A
0 I
m
and E =
I
n
−A
B I
m
we get
DE =
I
n
÷AB 0
B I
m
and ED =
I
n
0
B I
m
÷BA
Cofactor expansion of the determinant of DE along each of the last m columns shows that [DE[ =
[I
n
÷AB[. Similarly, cofactor expansion along each of the ﬁrst n rows shows that [ED[ = [I
m
÷BA[.
Alternatively, we could use formula (7) with A
22
= I
m
to evaluate [DE[ and formula (6) with A
11
= I
n
to evaluate [ED[.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
12 CHAPTER 2 MUL TI VARI ABL E CAL CUL US
(b) With A and B as in the hint, AB is an n n matrix with every column equal to A. Therefore
F = I
n
÷AB =
⎛
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎝
a
1
a
1
−1
1
a
1
−1
. . .
1
a
1
−1
1
a
2
−1
a
2
a
2
−1
. . .
1
a
2
−1
.
.
.
.
.
.
.
.
.
.
.
.
1
a
n
−1
1
a
n
−1
. . .
a
n
a
n
−1
⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
and
G =
⎛
⎜
⎜
⎝
a
1
1 . . . 1
1 a
2
. . . 1
.
.
.
.
.
.
.
.
.
.
.
.
1 1 . . . a
n
⎞
⎟
⎟
⎠
=
⎛
⎜
⎜
⎝
a
1
−1 0 . . . 0
0 a
2
−1 . . . 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 . . . a
n
−1
⎞
⎟
⎟
⎠
(I
n
÷AB)
From the result in (a) it follows that
[G[ = (a
1
−1)(a
2
−1) · · · (a
n
−1) [I
n
÷AB[ = (a
1
−1)(a
2
−1) · · · (a
n
−1) [I
1
÷BA[
= (a
1
−1)(a
2
−1) · · · (a
n
−1)
1 ÷
n
¸
i=1
1
a
i
−1
Chapter 2 Multivariable Calculus
2.1
2.1.3 (a) The unit vector in the direction given by v = (1, 1) is a =
1
v
v =
1
√
2
(1, 1). By formula (2.1.8)
the directional derivative of f in this direction at (2, 1) is
f
/
a
(2, 1) = ∇f (2, 1) · a =
1
√
2
∇f (2, 1) · (1, 1) =
1
√
2
(2, 1) · (1, 1) =
3
√
2
=
3
√
2
2
Note: It is pure coincidence that the gradient of f at (2,1) equals (2,1).
(b) The gradient of g is ∇g(x, y, z) = ((1 ÷ xy)e
xy
− y, x
2
e
xy
− x, −2z), and the unit vector in the
direction given by (1, 1, 1) is b =
1
√
3
(1, 1, 1). Formula (2.8.1) gives
g
/
b
(0, 1, 1) = ∇g(0, 1, 1) · b =
1
√
3
(0, 0, −2) · (1, 1, 1) = −
2
√
3
= −
2
√
3
3
2.1.5 (a) The vector from (3, 2, 1) to (−1, 1, 2) is (−1, 1, 2) −(3, 2, 1) = (−4, −1, 1), and the unit vector
in this direction is a =
1
√
18
(−4, −1, 1). The gradient of f is
∇f (x, y, z) =
y ln(x
2
÷y
2
÷z
2
)÷
2x
2
y
x
2
÷y
2
÷z
2
, x ln(x
2
÷y
2
÷z
2
)÷
2xy
2
x
2
÷y
2
÷z
2
,
2xyz
x
2
÷y
2
÷z
2
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 2 MUL TI VARI ABL E CAL CUL US 13
By formula (2.1.8) the directional derivative of f at (1, 1, 1) in the direction a is
f
/
a
(1, 1, 1) = (1/
√
18 ) ∇f (1, 1, 1) · (−4, −1, 1)
= (1/
√
18 ) (ln 3 ÷2/3, ln 3 ÷2/3, 2/3) · (−4, −1, 1) = −(5 ln 3 ÷8/3)/
√
18
(b) At (1, 1, 1) the direction of fastest growth for f is given by ∇f (1, 1, 1) = (ln 3÷2/3, ln 3÷2/3, 2/3).
2.1.6 The unit vector in the direction of maximal increase of f at (0, 0) is a =
1
√
10
(1, 3) and therefore
∇f (0, 0) = t a for a number t ≥ 0. We also know that f
/
a
(0, 0) = 4. On the other hand, by (2.1.8),
f
/
a
(0, 0) = ∇f (0, 0) · a = t a · a = t a
2
= t . Hence, t = 4 and
∇f (0, 0) = 4a =
4
√
10
(1, 3) =
4
√
10
10
(1, 3) =
2
√
10
5
(1, 3)
2.1.9 (a) We know that y
/
= −F
/
1
(x, y)/F
/
2
(x, y), and by the chain rule for functions of several variables,
y
//
=
d
dx
(y
/
) = −
∂
∂x
F
/
1
(x, y)
F
/
2
(x, y)
dx
dx
−
∂
∂y
F
/
1
(x, y)
F
/
2
(x, y)
dy
dx
= −
F
//
11
F
/
2
−F
/
1
F
//
21
(F
/
2
)
2
· 1 −
F
//
12
F
/
2
−F
/
1
F
//
22
(F
/
2
)
2
−
F
/
1
F
/
2
=
−(F
//
11
F
/
2
−F
/
1
F
//
21
)F
/
2
÷(F
//
12
F
/
2
−F
/
1
F
//
22
)F
/
1
(F
/
2
)
3
=
−F
//
11
(F
/
2
)
2
÷2F
//
12
F
/
1
F
/
2
−F
//
22
(F
/
1
)
2
(F
/
2
)
3
(Remember that F
//
21
= F
//
12
.) Expanding the determinant in the problem yields precisely the numerator
in the last fraction.
2.2
2.2.6 (a) If x and y are points in S and λ ∈ [0, 1], then x ≤ r, y ≤ r, and by the triangle inequality,
λx ÷ (1 − λ)y ≤ λx ÷ (1 − λ)y ≤ λr ÷ (1 − λ)r = r. Hence, λx ÷ (1 − λ)y belongs to S. It
follows that S is convex.
(b) S
1
is the interior of the ball S, i.e. what we get when we remove the spherical “shell” {x : x = r}
from S. The triangle inequality shows that S
1
is convex. The set S
2
is the spherical shell we mentioned,
while S
3
consists of the shell and the part of ޒ
n
that lies outside S. Neither S
2
nor S
3
is convex.
2.2.7 (a) Let us call a set S of numbers midpoint convex if
1
2
(x
1
÷ x
2
) whenever x
1
and x
2
belong to S.
The set S = ޑof rational numbers is midpoint convex, but it is not convex, for between any two rational
numbers r
1
and r
2
there are always irrational numbers. For example, let t = r
1
÷(r
2
−r
1
)/
√
2. Then t
lies between r
1
and r
2
. If t were rational, then
√
2 = (r
2
− r
1
)/(t − r
1
) would also be rational, but we
know that
√
2 is irrational.
(b) Suppose S is midpoint convex and closed. Let x
1
and x
2
be points in S, with x
1
< x
2
, and let λin (0, 1).
We shall prove that the point z = λx
1
÷(1−λ)x
2
belongs to S. Since S is midpoint convex, it must contain
y
1
=
1
2
(x
1
÷x
2
), and then it contains y
21
=
1
2
(x
1
÷y
1
) =
3
4
x
1
÷
1
4
x
2
and y
22
=
1
2
(y
1
÷x
2
) =
1
4
x
1
÷
3
4
x
2
.
We can continue in this fashion, constructing new midpoints between the points that have already been
constructed, and we ﬁnd that S must contain all points of the form
k
2
n
x
1
÷
2
n
−k
2
n
x
2
=
k
2
n
x
1
÷
1 −
k
2
n
x
2
, n = 1, 2, 3, . . . , k = 0, 1, . . . , 2
n
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
14 CHAPTER 2 MUL TI VARI ABL E CAL CUL US
Thus, for each n we ﬁnd 2
n
÷ 1 evenly spaced points in the interval [x
1
, x
2
] that all belong to S. The
distance between two neighbouring points in this collection is (x
2
− x
1
)/2
n
. Now let r be any positive
number, and consider the open rball (open interval) B = B(z: r) = (z −r, z ÷r) around z. Let n be so
large that (x
2
− x
1
)/2
n
< 2r. Then B must contain at least one of the points (k/2
n
)x
1
÷ (1 − k/2
n
)x
2
constructed above, so B contains at least one point from S. It follows that z does indeed belong to
cl(S) = S.
2.2.8 Let S be a convex subset of ޒ containing more than one point, and let a = inf S, b = sup S (where
a and b may be ﬁnite or inﬁnite). Then a < b. In order to prove that S is an interval, it sufﬁces to show
that S contains the open interval (a, b). Let x be a point in (a, b). Since x < b = sup S, there exists a β
in S with x < β. Similarly, there is an α in S with α < x. Then x is a convex combination of α and β,
and since S is convex, x belongs to S.
2.3
2.3.5 (a) z
//
11
(x, y) = −e
x
− e
x÷y
, z
//
12
(x, y) = −e
x÷y
, and z
//
22
(x, y) = −e
x÷y
, so z
//
11
(x, y) < 0 and
z
//
11
z
//
22
−(z
//
12
)
2
= e
2x÷y
> 0. By Theorem 2.3.1, z is a strictly concave function of x and y.
(b) z is strictly convex, because z
//
11
= e
x÷y
÷e
x−y
> 0 and z
//
11
z
//
22
−(z
//
12
)
2
= 4e
2x
> 0.
(c) w = u
2
, where u = x ÷ 2y ÷ 3z. So w is a convex function of an afﬁne function, hence convex
according to (2.3.8). It is not strictly convex, however, because it is constant on every plane of the form
x ÷2y ÷3z = c, and therefore constant along each line joining two points in such a plane.
2.3.6 (b) Let λ
2
> λ
1
> 0 and deﬁne μ = λ
1
/λ
2
. Then μ ∈ (0, 1) and by the concavity of f we have
μf (λ
2
x) ÷ (1 − μ)f (0) ≤ f (μλ
2
x ÷ (1 − μ)0), i.e. (λ
1
/λ
2
)f (λ
2
x) ≤ f (λ
1
x), so f (λ
2
x)/λ
2
≤
f (λ
1
x)/λ
1
. It also follows that if f is strictly concave, then f (λx)/λ is strictly decreasing as a function
of λ.
(c) Take any x ,= 0 in the domain of f . Then x ,= 2x and f (
1
2
x ÷
1
2
2x) = f (
3
2
x) =
3
2
f (x) =
1
2
f (x) ÷f (x) =
1
2
f (x) ÷
1
2
f (2x). Therefore f cannot be strictly concave.
2.3.8 The challenge here is mainly in getting the derivatives right, but with care and patience you will ﬁnd
that
f
//
11
(v
1
, v
2
) = −Pv
2
2
, f
//
12
(v
1
, v
2
) = Pv
1
v
2
, f
//
22
(v
1
, v
2
) = −Pv
2
1
where
P = (ρ ÷1)δ
1
δ
2
A(v
1
v
2
)
−ρ−2
δ
1
v
−ρ
1
÷δ
2
v
−ρ
2
−(1/ρ)−2
These formulas show that for all v
1
and v
2
, the “direct” second derivatives f
//
11
and f
//
22
have the same sign
as −(ρ ÷1). Also, f
//
11
f
//
22
−(f
//
12
)
2
= 0 everywhere. It follows from Theorem 2.3.1 that f is convex if
ρ ≤ −1, and concave if ρ ≥ −1. If ρ = −1 then f (v
1
, v
2
) = A(δ
1
v
1
÷δ
2
v
2
) is a linear function, which
indeed is both convex and concave.
The equation f
//
11
f
//
22
−(f
//
12
)
2
= 0 is a consequence of the fact that f is homogeneous of degree 1,
see e.g. Section 12.6 on homogeneous functions in EMEA. Since f is homogeneous of degree 1, it is
linear along each ray from the origin and therefore it cannot be strictly convex or strictly concave for any
value of ρ.
2.3.9 (a) This is mainly an exercise in manipulating determinants. If you feel that the calculations below
look frightening, try to write them out in full for the case k = 3 (or k = 2). Note that z
//
ij
= a
i
a
j
z/x
i
x
j
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 2 MUL TI VARI ABL E CAL CUL US 15
for i ,= j, and z
//
ii
= a
i
(a
i
−1)z/x
2
i
. Rule (1.1.20) tells us that a common factor in any column (or row)
in a determinant can be “moved outside”. Therefore,
D
k
=
z
//
11
z
//
12
. . . z
//
1k
z
//
21
z
//
22
. . . z
//
2k
.
.
.
.
.
.
.
.
.
.
.
.
z
//
k1
z
//
k2
. . . z
//
kk
=
a
1
(a
1
−1)
x
2
1
z
a
1
a
2
x
1
x
2
z · · ·
a
1
a
k
x
1
x
k
z
a
2
a
1
x
2
x
1
z
a
2
(a
2
−1)
x
2
2
z · · ·
a
2
a
k
x
2
x
k
z
.
.
.
.
.
.
.
.
.
.
.
.
a
k
a
1
x
k
x
1
z
a
k
a
2
x
k
x
2
z · · ·
a
k
(a
k
−1)
x
2
k
z
(1)
=
a
1
a
2
. . . a
k
x
1
x
2
. . . x
k
z
k
a
1
−1
x
1
a
1
x
1
· · ·
a
1
x
1
a
2
x
2
a
2
−1
x
2
· · ·
a
2
x
2
.
.
.
.
.
.
.
.
.
.
.
.
a
k
x
k
a
k
x
k
· · ·
a
k
−1
x
k
(2)
=
a
1
a
2
. . . a
k
(x
1
x
2
. . . x
k
)
2
z
k
a
1
−1 a
1
· · · a
1
a
2
a
2
−1 · · · a
2
.
.
.
.
.
.
.
.
.
.
.
.
a
k
a
k
· · · a
k
−1
where equality (1) holds because a
j
z/x
j
is a common factor in column j for each j and equality (2) holds
because 1/x
i
is a common factor in row i for each i.
(b) More determinant calculations. Let s
k
=
¸
k
i=1
a
i
= a
1
÷· · · ÷a
k
. We use the expression for D
k
that
we found in part (a), and add rows 2, 3, . . . , k to the ﬁrst row. Then each entry in the ﬁrst row becomes
equal to s
k
−1. Afterwards we take the common factor s
k
−1 in row 1 and move it outside.
D
k
=
a
1
a
2
. . . a
k
(x
1
x
2
. . . x
k
)
2
z
k
s
k
−1 s
k
−1 · · · s
k
−1
a
2
a
2
−1 · · · a
2
.
.
.
.
.
.
.
.
.
.
.
.
a
k
a
k
· · · a
k
−1
= (s
k
−1)
a
1
a
2
. . . a
k
(x
1
x
2
. . . x
k
)
2
z
k
1 1 · · · 1
a
2
a
2
−1 · · · a
2
.
.
.
.
.
.
.
.
.
.
.
.
a
k
a
k
· · · a
k
−1
Now subtract column 1 from all the other columns. Rule (1.1.22) says that this does not change the value
of the determinant, so
D
k
= (s
k
−1)
a
1
a
2
. . . a
k
(x
1
x
2
. . . x
k
)
2
z
k
1 0 · · · 0
a
2
−1 · · · 0
.
.
.
.
.
.
.
.
.
.
.
.
a
k
0 · · · −1
= (−1)
k−1
(s
k
−1)
a
1
a
2
. . . a
k
(x
1
x
2
. . . x
k
)
2
z
k
(c) By assumption, a
i
> 0 for all i, so if
¸
n
i=1
a
i
< 1, then s
k
=
¸
k
i=1
a
i
< 1 for all k. Therefore D
k
has the same sign as (−1)
k
. It follows from Theorem 2.3.2(b) that f is strictly concave.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
16 CHAPTER 2 MUL TI VARI ABL E CAL CUL US
2.4
2.4.3 We shall show that Jensen’s inequality (2.4.2) holds for all natural numbers m, not just for m = 3. Let
f be a function deﬁned on a convex set S in ޒ
n
and for each natural number m let A(m) be the following
statement: “The inequality
f (λ
1
x
1
÷· · · ÷λ
m
x
m
) ≥ λ
1
f (x
1
) ÷· · · ÷λ
m
f (x
m
)
holds for all x
1
, . . . , x
m
in S and all λ
1
≥ 0, . . . , λ
m
≥ 0 with λ
1
÷· · · ÷λ
m
= 1.”
We shall prove that A(m) is true for every natural number m. It is obvious that A(1) is true, since
it just says that f (x) = f (x), and A(2) is also true, since f is concave. Now suppose that A(k) is true,
where k is some natural number greater than 1. We shall prove that A(k ÷1) is also true.
Let x
1
, . . . , x
k÷1
be points in S and let λ
1
, . . . , λ
k÷1
be nonnegative numbers with sum 1. We can
assume that λ
k÷1
> 0, for otherwise we are really in the case m = k. Then μ = λ
k
÷λ
k÷1
> 0 and we
can deﬁne y = (1/μ)(λ
k
x
k
÷ λ
k÷1
x
k÷1
). The point y is a convex combination of x
k
and x
k÷1
, and so
y ∈ S. By A(k) we have
f (λ
1
x
1
÷· · · ÷λ
k÷1
x
k÷1
) = f (λ
1
x
1
÷· · · ÷λ
k−1
x
k−1
÷μy)
≥ λ
1
f (x
1
) ÷· · · ÷λ
k−1
f (x
k−1
) ÷μf (y)
(∗)
Moreover, since f is concave, and (λ
k
/μ) ÷(λ
k÷1
/μ) = 1,
μf (y) = μf
λ
k
μ
x
k
÷
λ
k÷1
μ
x
k÷1
≥ μ
λ
k
μ
f (x
k
) ÷
λ
k÷1
μ
f (x
k÷1
)
= λ
k
f (x
k
) ÷λ
k÷1
f (x
k÷1
)
and this inequality together with (∗) yields
f (λ
1
x
1
÷· · · ÷λ
k÷1
x
k÷1
) ≥ λ
1
f (x
1
) ÷· · · ÷λ
k−1
f (x
k−1
) ÷λ
k
f (x
k
) ÷λ
k÷1
f (x
k÷1
)
which shows that A(k ÷1) is true. To sum up, we have shown that: (i) A(2) is true, (ii) A(k) ⇒A(k ÷1)
for every natural number k ≥ 2. It then follows by induction that A(m) is true for all m ≥ 2. Note that
both (i) and (ii) are necessary for this conclusion.
2.4.5 Let x, y belong to S and let λ ∈ [0, 1]. Then z = λx ÷ (1 − λ)y belongs to S. To show that f is
concave, it is sufﬁcient to show that λf (x) ÷(1 −λ)f (y) ≤ f (z). By assumption, f has a supergradient
p at z, and therefore
f (x) −f (z) ≤ p · (x −z) and f (y) −f (z) ≤ p · (y −z)
Since both λ and 1 −λ are nonnegative, it follows that
λf (x) ÷(1 −λ)f (y) −f (z) = λ[f (x) −f (z)] ÷(1 −λ)[f (y) −f (z)]
≤ p · [λ(x −z) ÷(1 −λ)(y −z)] = p · 0 = 0
by the deﬁnition of z.
2.4.6 Theorem 2.4.1(c) tells us that f (x, y) = x
4
÷y
4
is a strictly convex function of (x, y) if and only if
f (x, y) −f (x
0
, y
0
) > ∇f (x
0
, y
0
) · (x −x
0
, y −y
0
) whenever (x, y) ,= (x
0
, y
0
). Since ∇f (x
0
, y
0
) =
(4x
3
0
, 4y
3
0
), we have
f (x, y) −f (x
0
, y
0
) −∇f (x
0
, y
0
) · (x −x
0
, y −y
0
) = p(x, x
0
) ÷p(y, y
0
)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 2 MUL TI VARI ABL E CAL CUL US 17
where p(t, t
0
) = t
4
− t
4
0
− 4t
3
0
(t − t
0
) for all t and t
0
. Now, p(t, t
0
) = (t − t
0
)(t
3
÷ t
2
t
0
÷ t t
2
0
− 3t
3
0
).
The second factor here is 0 if t = t
0
, so it is divisible by t −t
0
, and polynomial division yields
p(t, t
0
) = (t −t
0
)(t −t
0
)(t
2
÷2t t
0
÷3t
2
0
) = (t −t
0
)
2
[(t ÷t
0
)
2
÷2t
2
0
]
The expression in square brackets is strictly positive unless t = t
0
= 0, so it follows that p(t, t
0
) > 0
whenever t ,= t
0
. Hence, p(x, x
0
) ÷p(y, y
0
) > 0 unless both x = x
0
and y = y
0
.
2.5
2.5.2 (a) f is linear, so it is concave and therefore also quasiconcave.
(c) The set of points for which f (x, y) ≥ −1 is P
−1
= {(x, y) : y ≤ x
−2/3
}, which is not a convex
set (see Fig. M2.5.2(c), where P
−1
is the unshaded part of the plane), so f is not quasiconcave. (It is
quasiconvex in the ﬁrst quadrant, though.)
(d) The polynomial x
3
÷x
2
÷1 is increasing in the interval (−∞, −2/3], and decreasing in [−2/3, 0].
So f is increasing in (−∞, −2/3] and decreasing in [−2/3, ∞). (See Fig. M2.5.2(d).) Then the upper
level sets must be intervals (or empty), and it follows that f is quasiconcave. Alternatively, we could use
the result in the note below.
y
2
1
1
2
3
4
x
4 3 2 1 1 2 3 4
f (x) ≤ −1 f (x) ≤ −1
y
1
1
2
3
x
3 2 1 1 2 3
y = x
3
÷x
2
÷1
y = f (x)
Figure M2.5.2(c) Neither P
a
nor P
a
is convex for a = −1.
Figure M2.5.2(d) The graph of f .
A note on quasiconcave functions of one variable
It is shown in Example 2.5.2 in the book that a function of one variable that is increasing or decreasing
on a whole interval is both quasiconcave and quasiconvex on that interval. Now suppose f is a function
deﬁned on an interval (a, b) and that there is a point c in (a, b such that f is increasing on (a, c] and
decreasing on [c, b). Then f is quasiconcave on the interval (a, b). This follows because the upper level
sets must be intervals. Alternatively we can use Theorem 2.5.1(a) and note that if x and y are points in
(a, b), then f (z) ≥ min{f (x), f (y)} for all z between x and y. We just have to look at each of the three
possibilities x < y ≤ c, x ≤ c < y, and c < x < y, and consider the behaviour of f over the interval
[x, y] in each case. This argument also holds if a = −∞ or b = ∞, and also in the case of a closed
or halfopen interval. Similarly, if f is decreasing to the left of c and increasing to the right, then f is
quasiconvex.
We could use this argument in Problem 2.5.2(d), for instance.
2.5.6 Since f is decreasing and g is increasing, it follows from Example 2.4.2 that both of these functions
are quasiconcave as well as quasiconvex. Their sum, f (x) ÷ g(x) = x
3
− x, is not quasiconcave,
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
18 CHAPTER 2 MUL TI VARI ABL E CAL CUL US
however. For instance, it is clear from Fig. A2.5.6 in the answer section of the book that the upper level
set P
0
= {x : f (x) ÷g(x) ≥ 0} = [−1, 0] ∪[1, ∞), which is not a convex set of ޒ. Similarly, the lower
level set P
0
= (−∞, −1] ∪ [0, 1] is not convex, and so f ÷g is not quasiconvex either.
2.5.7 (a) Let f be singlepeaked with a peak at x
∗
. We want to prove that f is strictly quasiconcave. That
is, we want to prove that if x < y and z is a point strictly between x and y, then f (z) > min{f (x), f (y)}.
There are two cases to consider:
(A) Suppose z ≤ x
∗
. Since x < z ≤ x
∗
and f is strictly increasing in the interval [x, x
∗
], we have
f (z) > f (x).
(B) Suppose x
∗
< z. Then x
∗
< z < y, and since f is strictly decreasing in [x
∗
, y], we get f (z) > f (y).
In both cases we get f (z) > min{f (x), f (y)}, and so f is strictly quasiconcave
(b) No. Even if f is concave, it may, for example, be linear in each of the intervals (−∞, x
∗
] and
[x
∗
, ∞), or in parts of one or both of these intervals, and in such cases f cannot be strictly concave.
2.5.11 With apologies to the reader, we would like to change the name of the function from f to h. So
h(x) is strictly quasiconcave and homogeneous of degree q ∈ (0, 1), with h(x) > 0 for all x ,= 0 in
K and h(0) = 0, and we want to prove that h is strictly concave in the convex cone K. Deﬁne a new
function f by f (x) = h(x)
1/q
. Then f (x) is also strictly quasiconcave (use the deﬁnition) and satisﬁes
the conditions of Theorem 2.5.3. Let x ,= y and let λ ∈ (0, 1).
Assume ﬁrst that x and y do not lie on the same ray from the origin in ޒ
n
. Then both f (x) and f (y)
are strictly positive and we can deﬁne α, β, μ, x
/
, and y
/
as in the proof of Theorem 2.5.3. We get x
/
,= y
/
and f (x
/
) = f (y
/
), and (see the original proof)
f (λx ÷(1 −λ)y) = f (μx
/
÷(1 −μ)y
/
) > f (x
/
) = f (y
/
) = μf (x
/
) ÷(1 −μ)f (y
/
)
= (μβ/α)f (x) ÷(β −βμ)f (y) = λf (x) ÷(1 −λ)f (y)
with strict inequality because f is strictly quasiconcave. Since 0 < q < 1, the qth power function t .→t
q
is strictly increasing and strictly concave, and therefore
h(λx ÷(1 −λ)y) =
f (λx ÷(1 −λ)y)
q
>
λf (x) ÷(1 −λ)f (y)
q
> λf (x)
q
÷(1 −λ)f (y)
q
= λh(x) ÷(1 −λ)h(y)
It remains to show that h(λx ÷ (1 − λ)y) > λh(x) ÷ (1 − λ)h(y) in the case where x and y lie on the
same ray from the origin. We can assume that x ,= 0 and y = t x for some nonnegative number t ,= 1.
Since the qth power function is strictly concave, (λ ÷(1 −λ)t )
q
> λ1
q
÷(1 −λ)t
q
= λ ÷(1 −λ)t
q
. It
follows that
h(λx ÷(1 −λ)y) = h
(λ ÷(1 −λ)t )x
=
λ ÷(1 −λ)t
q
h(x)
>
λ ÷(1 −λ)t
q
h(x) = λh(x) ÷(1 −λ)h(t x) = λh(x) ÷(1 −λ)h(y)
2.6
2.6.1 (a) f
/
1
(x, y) = ye
xy
, f
/
2
(x, y) = xe
xy
, f
//
11
(x, y) = y
2
e
xy
. f
//
12
(x, y) = e
xy
÷ xye
xy
, f
//
22
(x, y) =
x
2
e
x,y
. With the exception of f
//
12
, these derivatives all vanish at the origin, so we are left with the quadratic
approximation f (x, y) ≈ f (0, 0) ÷f
//
12
(0, 0)xy = 1 ÷xy.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 2 MUL TI VARI ABL E CAL CUL US 19
(b) f
/
1
(x, y) = 2xe
x
−
y
2
, f
/
2
(x, y) = −2ye
x
2
−y
2
, f
//
11
(x, y) = (2÷4x
2
)e
x
2
−y
2
, f
//
12
(x, y) = −4xye
x
2
−y
2
,
f
//
22
(x, y) = (−2 ÷ 4y
2
)e
x
2
−y
2
. Hence, f
/
1
(0, 0) = f
/
2
(0, 0) = 0, f
//
11
(0, 0) = 2, f
//
12
(0, 0) = 0,
f
//
22
(0, 0) = −2.
Quadratic approximation: f (x, y) ≈ f (0, 0) ÷
1
2
f
//
11
(0, 0)x
2
÷
1
2
f
//
22
(0, 0)y
2
= 1 ÷x
2
−y
2
.
(c) The ﬁrst and secondorder derivatives are: f
/
1
(x, y) =
1
1 ÷x ÷2y
, f
/
2
(x, y) =
2
1 ÷x ÷2y
,
f
//
11
(x, y) = −
1
(1 ÷x ÷2y)
2
, f
//
12
(x, y) = −
2
(1 ÷x ÷2y)
2
, f
//
22
(x, y) = −
4
(1 ÷x ÷2y)
2
.
At (0, 0) we get f (0, 0) = 1, f
/
1
(0, 0) = 1, f
/
2
(0, 0) = 2, f
//
11
(0, 0) = −1, f
//
2
(0, 0) = −2, f
//
22
(0, 0) =
−4, and the quadratic approximation to f (x, y) around (0, 0) is
f (x, y) ≈ f (0, 0) ÷f
/
1
(0, 0)x ÷f
/
2
(0, 0)y ÷
1
2
f
//
11
(0, 0)x
2
÷f
//
12
(0, 0)xy ÷
1
2
f
//
22
(0, 0)y
2
= 1 ÷x ÷2y −
1
2
x
2
−2xy −2y
2
(There is a misprint in the answer in the ﬁrst printing of the book.)
2.6.4 z is deﬁned implicitly as a function of x and y by the equation F(x, y, z) = 0, where F(x, y, z) =
ln z −x
3
y ÷xz −y. With x = y = 0 we get ln z = 0, so z = 1. Since F
/
3
(x, y, z) = 1/z ÷x = 1 ,= 0 at
(x, y, z) = (0, 0, 1), z is a C
1
function of x and y around this point. In order to ﬁnd the Taylor polynomial
we need the ﬁrst and secondorder derivatives of z with respect to x and y. One possibility is to use the
formula z
/
x
= −F
/
1
(x, y, z)/F
/
3
(x, y, z) and the corresponding formula for z
/
y
, and then take derivatives of
these expressions to ﬁnd the secondorder derivatives. That procedure leads to some rather nasty fractions
with a good chance of going wrong, so instead we shall differentiate the equation ln z = x
3
y −xz ÷y,
keeping in mind that z is a function of x and y. We get
z
/
x
/z = 3x
2
y −z −xz
/
x
=⇒ (1 ÷xz)z
/
x
= 3x
2
yz −z
2
(1)
z
/
y
/z = x
3
−xz
/
y
÷1 =⇒ (1 ÷xz)z
/
y
= x
3
z ÷z (2)
Taking derivatives with respect to x and y in (1), we get
(z ÷xz
/
x
)z
/
x
÷(1 ÷xz)z
//
xx
= 6xyz ÷3x
2
yz
/
x
−2zz
/
x
(3)
xz
/
y
z
/
x
÷(1 ÷xz)z
//
xy
= 3x
2
z ÷3x
2
yz
/
y
−2zz
/
y
(4)
and differentiating (2) with respect to y, we get
xz
/
y
z
/
y
÷(1 ÷xz)z
//
yy
= x
3
z
/
y
÷z
/
y
(5)
Now that we have ﬁnished taking derivatives, we can let x = y = 0 and z = 1 in the equations we have
found. Equations (1) and (2) give z
/
x
= −1 and z
/
y
= 1 (at the particular point we are interested in), and
then (3)–(5) give z
//
xx
= 3, z
//
xy
= −2, and z
//
yy
= 1. The quadratic approximation to z around (0, 0) is
therefore
z ≈ 1 −x ÷y ÷
3
2
x
2
−2xy ÷
1
2
y
2
(The ﬁrst printing of the book has a misprint in the answer to this problem, too—the coefﬁcient of x
2
is
wrong.)
2.7
2.7.2 (a) F(x, y, z) = x
3
÷ y
3
÷ z
3
− xyz − 1 is obviously C
1
everywhere, and F
/
3
(0, 0, 1) = 3 ,= 0,
so by the implicit function theorem the equation deﬁnes z as a C
1
function g(x, y) in a neighbourhood
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
20 CHAPTER 2 MUL TI VARI ABL E CAL CUL US
of (x
0
, y
0
, z
0
) = (0, 0, 1). To ﬁnd the partial derivatives g
/
1
and g
/
2
there is no need to use the matrix
formulation in Theorem 2.7.2. After all, g is just a single realvalued function, and the partial derivative
g
/
1
(x, y) is just what we get if we treat y as a constant and take the derivative of g with respect to x.
Thus, g
/
1
(x, y) = −F
/
1
(x, y, z)/F
/
3
(x, y, z) = −(3x
2
− yz)/(3z
2
− xy) and, in particular, g
/
1
(0, 0) =
−F
/
1
(0, 0, 1)/F
/
3
(0, 0, 1) = 0. Likewise, g
/
2
(0, 0) = 0.
(b) As in part (a), F
/
3
is C
1
everywhere and F
/
3
(x, y, z) = e
z
− 2z ,= 0 for z = 0, so the equa
tion F = 0 deﬁnes z as a C
1
function g(x, y) around (x
0
, y
0
, z
0
) = (1, 0, 0). We get g
/
1
(x, y) =
−F
/
1
(x, y, z)/F
/
3
(x, y, z) = 2x/(e
z
− 2z) and g
/
2
(x, y, z) = −F
/
2
/F
/
3
= 2y/(e
z
− 2z), so g
/
1
(1, 0) = 2
and g
/
2
(1, 0) = 0.
2.7.3 The given equation system can be written as f(x, y, z, u, v, w) = 0, where f = (f
1
, f
2
, f
3
) is the
function ޒ
3
ޒ
3
→ޒ
3
given by f
1
(x, y, z, u, v, w) = y
2
−z ÷u −v −w
3
÷1, f
2
(x, y, z, u, v, w) =
−2x ÷ y − z
2
÷ u ÷ v
3
− w ÷ 3, and f
3
(x, y, z, u, v, w) = x
2
÷ z − u − v ÷ w
3
− 3. The Jacobian
determinant of f with respect to u, v, w is
∂(f
1
, f
2
, f
3
)
∂(u, v, w)
=
1 −1 −3w
2
1 3v
2
−1
−1 −1 3w
2
= 6w
2
−2
This determinant is different from 0 at P, so according to Theorem 2.7.2 the equation system does deﬁne
u, v, and w as C
1
functions of x, y, and z. The easiest way to ﬁnd the partial derivatives of these functions
with respect to x is probably to take the derivatives with respect to x in each of the three given equations,
remembering that u, v, w are functions of x, y, z. We get
u
/
x
−v
/
x
−3w
2
w
/
x
= 0
−2 ÷u
/
x
÷3v
2
v
/
x
−w
/
x
= 0
2x −u
/
x
−v
/
x
÷3w
2
w
/
x
= 0
⎫
⎪
⎬
⎪
⎭
, so at P we get
⎧
⎪
⎨
⎪
⎩
u
/
x
−v
/
x
−3w
/
x
= 0
u
/
x
− w
/
x
= 2
−u
/
x
−v
/
x
÷3w
/
x
= −2
The unique solution of this system is u
/
x
= 5/2, v
/
x
= 1, w
/
x
= 1/2. (In the ﬁrst printing of the book w
/
x
was incorrectly given as 5/2.)
2.7.4 The Jacobian determinant is
f
/
u
f
/
v
g
/
u
g
/
v
=
e
u
cos v −e
u
sin v
e
u
sin v e
u
cos v
= e
2u
(cos
2
v ÷sin
2
v) = e
2u
, which
is nonzero for all u (and v).
(a) e
u
,= 0, so the equations imply cos v = sin v = 0, but that is impossible because cos
2
v ÷sin
2
v = 1.
(b) We must have cos v = sin v = e
−u
, and since cos
2
v ÷ sin
2
v = 1 we get 2 cos
2
v = 1 and
therefore sin v = cos v =
√
1/2 =
1
2
√
2. (Remember that cos v = e
−u
cannot be negative.) The only
values of v that satisfy these equations are v = (
1
4
÷ 2k)π, where k runs through all integers. Further,
e
u
= 1/ cos v =
√
2 gives u =
1
2
√
2.
2.7.6 The Jacobian is x
1
. We ﬁnd that x
1
= y
1
÷ y
2
, x
2
= y
2
/(y
1
÷ y
2
) (provided y
1
÷ y
2
,= 0). The
transformation maps the given rectangle onto the set S in the y
1
y
2
plane given by the inequalities
(i) 1 ≤ y
1
÷y
2
≤ 2, (ii)
1
2
≤
y
2
y
1
÷y
2
≤
2
3
The inequalities (i) show that y
1
÷y
2
> 0, and if we multiply by 6(y
1
÷y
2
) in (ii) we get the equivalent
inequalities
3(y
1
÷y
2
) ≤ 6y
2
≤ 4(y
1
÷y
2
) ⇐⇒ y
1
≤ y
2
≤ 2y
1
(iii)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 2 MUL TI VARI ABL E CAL CUL US 21
It follows that S is a quadrilateral with corners at (1/2, 1/2), (1, 1), (2/3, 4/3), and (1/3, 2/3). See ﬁgure
M2.7.6.
y
2
y
1
y
2
= 2y
1
y
2
= y
1
y
1
÷y
2
= 2
y
1
÷y
2
= 1
2 1
1
2
S
Figure M2.7.6
2.7.8 (a) J =
∂
∂r
(r cos θ)
∂
∂θ
(r cos θ)
∂
∂r
(r sin θ)
∂
∂θ
(r sin θ)
=
cos θ −r sin θ
sin θ r cos θ
= r
(b) J ,= 0 everywhere except at the origin in the rθplane, so T is locally onetoone in A. But it is not
globally onetoone in A, since we have, for example, T (r, 0) = T (r, 2π).
2.7.10 (a) Taking differentials in the equation system
1 ÷(x ÷y)u −(2 ÷u)
1÷v
= 0
2u −(1 ÷xy)e
u(x−1)
= 0
(1)
we get
u(dx ÷dy) ÷(x ÷y) du −e
(1÷v) ln(2÷u)
¸
ln(2 ÷u) dv ÷
1 ÷v
2 ÷u
du
¸
= 0
2 du −e
u(x−1)
(y dx ÷x dy) −(1 ÷xy)e
u(x−1)
(x −1) du ÷u dx
= 0
If we now let x = y = u = 1 and v = 0, we get
2 du ÷dx ÷dy −3 ln 3 dv −du = 0
and
2 du −dx −dy −2 dx = 0
Rearranging this system gives
du −3 ln 3 dv = −dx −dy
2 du = 3 dx ÷dy
with the solutions
du =
3
2
dx ÷
1
2
dy and dv =
1
3 ln 3
5
2
dx ÷
3
2
dy
Hence,
u
/
x
(1, 1) =
∂u
∂x
(1, 1) =
3
2
and v
/
x
(1, 1) =
∂v
∂x
(1, 1) =
5
6 ln 3
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
22 CHAPTER 2 MUL TI VARI ABL E CAL CUL US
(Alternatively we could have used the formula for the derivatives of an implicit function, but it is still a
good idea to substitute the values of x, y, u, and v after ﬁnding the derivatives.)
(b) Deﬁne a function f by f (u) = u − ae
u(b−1)
. Then f (0) = −a and f (1) = 1 − ae
b−1
. Since
b ≤ 1, we have e
b−1
≤ e
0
= 1. It follows that ae
b−1
≤ a ≤ 1, so f (1) ≥ 0. On the other hand,
f (0) ≤ 0, so the intermediate value theorem ensures that f has at least one zero in [0, 1]. Further,
f
/
(u) = 1 −a(b −1)e
u(b−1)
≥ 1, because a(b −1) ≤ 0. Therefore f is strictly increasing and cannot
have more than one zero, so the solution of f (u) = 0 is unique.
(c) For given values of x and y, let a = (1÷xy)/2 and b = x. The equation in part (b) is then equivalent
to the second equation in system (1). Thus we get a uniquely determined u, and this u belongs to [0, 1].
(Note that, when x and y lie in [0, 1], then the values of a and b that we have chosen also lie in [0, 1].)
The ﬁrst equation in (1) now determines v uniquely, as
v = −1 ÷
ln(1 ÷(x ÷y)u)
ln(2 ÷u)
2.8
2.8.1 (a) The system has 7 variables, Y, C, I, G, T , r, and M, and 4 equations, so the counting rule says
that the system has 7 −4 = 3 degrees of freedom.
(b) We can write the system as
f
1
(M, T, G, Y, C, I, r) = Y −C −I −G = 0
f
2
(M, T, G, Y, C, I, r) = C −f (Y −T ) = 0
f
3
(M, T, G, Y, C, I, r) = I −h(r) = 0
f
4
(M, T, G, Y, C, I, r) = r −m(M) = 0
(∗)
Suppose that f , h, and m are C
1
functions and that the system has an equilibrium point (i.e. a solution
of the equations), (M
0
, T
0
, G
0
, Y
0
, C
0
, I
0
, r
0
). The Jacobian determinant of f = (f
1
, f
2
, f
3
, f
4
) with
respect to (Y, C, I, r) is
∂(f
1
, f
2
, f
3
, f
4
)
∂(Y, C, I, r)
=
1 −1 −1 0
−f
/
(Y −T ) 1 0 0
0 0 1 −h
/
(r)
0 0 0 1
= 1 −f
/
(Y −T )
so by the implicit function theorem the system (∗) deﬁnes Y, C, I, and r as C
1
functions of M, T , and
G around the equilibrium point if f
/
(Y −T ) ,= 1.
Note that the functions hand mhave no inﬂuence on the Jacobian determinant. The reason is that once
M, T , and G are given, the last equation in (∗) immediately determines r, and the next to last equation
then determines I. The problem therefore reduces to the question whether the ﬁrst two equations can
determine Y and T when the values of the other variables are given. The implicit function theorem tells
us that the answer will certainly be yes if the Jacobian determinant
∂(f
1
, f
2
)
∂(Y, C)
=
1 −1
−f
/
(Y −T ) 1
= 1 −f
/
(Y −T )
is nonzero, i.e. if f
/
(Y −T ) ,= 1.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 3 STATI C OPTI MI ZATI ON 23
2.8.2 (a) The Jacobian determinants are
∂(u, v)
∂(x, y)
=
u
/
x
u
/
y
v
/
x
v
/
y
=
e
x÷y
e
x÷y
4x ÷4y −1 4x ÷4y −1
= 0 (i)
∂(u, v)
∂(x, y)
=
u
/
x
u
/
y
v
/
x
v
/
y
=
1
y
−x
y
2
−2y
(y ÷x)
2
2x
(x ÷y)
2
=
2x
y(y ÷x)
2
−
2x
y(y ÷x)
2
= 0 (ii)
(b) (i) It is not hard to see that v = 2(x ÷y)
2
−(x ÷y) and x ÷y = ln u, so v = 2(ln u)
2
−ln u.
(ii) We have x = uy, so
v =
y −uy
y ÷uy
=
1 −u
1 ÷u
2.8.3 We need to assume that the functions f and g are C
1
in an open ball around (x
0
, y
0
). For all (x, y) in
A the Jacobian determinant is
∂(u, v)
∂(x, y)
=
∂u
∂x
∂u
∂y
∂v
∂x
∂v
∂y
=
f
/
1
(x, y) f
/
2
(x, y)
g
/
1
(x, y) g
/
2
(x, y)
= 0
so f
/
2
(x, y)g
/
1
(x, y) = f
/
1
(x, y)g
/
2
(x, y). Since f
/
1
(x
0
, y
0
) ,= 0, the equation G(x, y, u) = u−f (x, y) =
0 deﬁnes x as a function x = ϕ(y, u) in an open ball around (y
0
, u
0
), where u
0
= f (x
0
, y
0
). Within
this open ball we have ϕ
/
1
(y, u) = ∂x/∂y = −(∂G/∂y)/(∂G/∂x) = −f
/
2
(x, y)/f
/
1
(x, y). Moreover,
v = g(x, y) = g(ϕ(y, u), y). If we let ψ(y, u) = g(ϕ(y, u), y), then ψ
/
1
(y, u) = g
/
1
ϕ
/
1
÷ g
/
2
=
−f
/
2
g
/
1
/f
/
1
÷g
/
2
= −f
/
1
g
/
2
/f
/
1
÷g
/
2
= 0. Thus, ψ(y, u) only depends on u, and v = ψ(y, u) is a function
of u alone. Hence, v is functionally dependent on u.
(If we let F(u, v) = v −ψ(y
0
, u), then F satisﬁes the requirements of deﬁnition (2.8.5).)
Chapter 3 Static Optimization
3.1
3.1.3 (a) The ﬁrstorder conditions for maximum are
1
3
pv
−2/3
1
v
1/2
2
−q
1
= 0,
1
2
pv
1/3
1
v
−1/2
2
−q
2
= 0
with the unique solution
v
∗
1
=
1
216
p
6
q
−3
1
q
−3
2
, v
∗
2
=
1
144
p
6
q
−2
1
q
−4
2
The objective function pv
1/3
1
v
1/2
2
− q
1
v
1
− q
2
v
2
is concave, cf. the display (2.5.6) on Cobb–Douglas
functions, and therefore (v
∗
1
, v
∗
2
) is a maximum point.
(b) The value function is
π
∗
(p, q
1
, q
2
) = p(v
∗
1
)
1/3
(v
∗
2
)
1/2
−q
1
v
∗
1
−q
2
v
∗
2
=
1
432
p
6
q
−2
1
q
−3
2
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
24 CHAPTER 3 STATI C OPTI MI ZATI ON
and it follows that
∂π
∗
(p, q
1
, q
2
)
∂p
=
1
72
p
5
q
−2
1
q
−3
2
= (v
∗
1
)
1/3
(v
∗
2
)
1/2
∂π
∗
(p, q
1
, q
2
)
∂q
1
= −
1
216
p
6
q
−3
1
q
−3
2
= −v
∗
1
,
∂π
∗
(p, q
1
, q
2
)
∂q
2
= −
1
144
p
6
q
−2
1
q
−4
2
= −v
∗
2
3.1.5 The ﬁrstorder conditions for maximum are
f
/
1
(x, y, r, s) = r
2
−2x = 0, f
/
2
(x, y, r, s) = 3s
2
−16y = 0
with the solutions x
∗
(r, s) =
1
2
r
2
and y
∗
(r, s) =
3
16
s
2
. Since f (x, y, r, s) is concave with respect to
(x, y), the point (x
∗
, y
∗
) is a maximum point. Moreover,
f
∗
(r, s) = f (x
∗
, y
∗
, r, s) =
1
4
r
4
÷
9
32
s
4
so
∂f
∗
(r, s)
∂r
= r
3
,
∂f
∗
(r, s)
∂s
=
9
8
s
3
On the other hand,
∂f (x, y, r, s)
∂r
= 2rx,
∂f (x, y, r, s)
∂s
= 6sy
so
¸
∂f (x, y, r, s)
∂r
¸
(x,y)=(x
∗
,y
∗
)
= 2rx
∗
= r
3
,
¸
∂f (x, y, r, s)
∂s
¸
(x,y)=(x
∗
,y
∗
)
= 6sy
∗
=
9
8
s
3
in accordance with the envelope result (3.1.3).
3.1.6 We want to maximize
π(v, p, q, a) = π(v
1
, . . . , v
n
: p, q
1
, . . . , q
n
, a
1
, . . . , a
n
) =
n
¸
i=1
(pa
i
ln(v
i
÷1) −q
i
v
i
)
with respect to v
1
, . . . , v
n
for given values of p, q = (q
1
, . . . , q
n
), and a = (a
1
, . . . , a
n
). Since ∂π/∂v
i
=
pa
i
/(v
i
÷1) −q
i
, the only stationary point is v
∗
= v
∗
(p, q, a) = (v
∗
1
, . . . , v
∗
n
), where v
∗
i
= pa
i
/q
i
−1.
Since π is concave with respect to v
1
, . . . , v
n
, this is a maximum point. The corresponding maximum
value is
π
∗
(p, q
1
, . . . , q
n
, a
1
, . . . , a
n
) =
n
¸
i=1
(pa
i
ln(v
∗
i
÷1) −q
i
v
∗
i
) =
n
¸
i=1
pa
i
ln
pa
i
q
i
−q
i
pa
i
q
i
−1
=
n
¸
i=1
(pa
i
ln p ÷pa
i
ln a
i
−pa
i
ln q
i
−pa
i
÷q
i
)
Easy calculations now yield
∂π
∗
(p, q, a)
∂p
=
n
¸
i=1
a
i
ln
pa
i
q
i
=
n
¸
i=1
a
i
ln(v
∗
i
÷1) =
¸
∂π(v, p, q, a)
∂p
¸
v=v
∗
(p,q,a)
∂π
∗
(p, q, a)
∂q
i
= −
pa
i
q
i
÷1 = −v
∗
i
=
¸
∂π(v, p, q, a)
∂q
i
¸
v=v
∗
(p,q,a)
∂π
∗
(p, q, a)
∂a
i
= p ln
pa
i
q
i
= p ln(v
∗
i
÷1) =
¸
∂π(v, p, q, a)
∂a
i
¸
v=v
∗
(p,q,a)
in accordance with formula (3.1.3).
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 3 STATI C OPTI MI ZATI ON 25
3.2
3.2.1 The ﬁrstorder conditions are
f
/
1
(x
1
, x
2
, x
3
) = 2x
1
− x
2
÷2x
3
= 0
f
/
2
(x
1
, x
2
, x
3
) = −x
1
÷2x
2
÷ x
3
= 0
f
/
3
(x
1
, x
2
, x
3
) = 2x
1
÷ x
2
÷6x
3
= 0
The determinant of this linear equation system is 4, so by Cramer’s rule it has a unique solution. This
solution is of course (x
1
, x
2
, x
3
) = (0, 0, 0). The Hessian matrix is (at every point)
H = f
//
(x
1
, x
2
, x
3
) =
⎛
⎝
2 −1 2
−1 2 1
2 1 6
⎞
⎠
with leading principal minors D
1
= 2, D
2
=
2 −1
−1 2
= 3, and D
3
= [H[ = 4, so (0, 0, 0) is a local
minimum point by Theorem 3.2.1(a).
3.2.3 (a) The ﬁrstorder conditions are
f
/
1
(x, y, z) = 2x ÷2xy = 0
f
/
2
(x, y, z) = x
2
÷2yz = 0
f
/
3
(x, y, z) = y
2
÷2z −4 = 0
This system gives ﬁve stationary points: (0, 0, 2), (0, ±2, 0), (±
√
3, −1, 3/2). The Hessian matrix is
f
//
(x, y, z) =
⎛
⎝
2 ÷2y 2x 0
2x 2z 2y
0 2y 2
⎞
⎠
with leading principal minors D
1
= 2 ÷2y, D
2
= 4(1 ÷y)z −4x
2
, and D
3
= 8(1 ÷y)(z −y
2
) −8x
2
.
The values of the leading principal minors at the stationary points are given in the following table:
D
1
D
2
D
3
(0, 0, 2) 2 8 16
(0, 2, 0) 6 0 −96
(0, −2, 0) −2 0 32
(
√
3, −1, 3/2) 0 −12 −24
(−
√
3, −1, 3/2) 0 −12 −24
It follows from Theorem 3.2.1 that (0, 0, 2) is a local minimum point, and all the other stationary points
are saddle points.
(b) The stationary points are the solutions of the equation system
f
/
1
(x
1
, x
2
, x
3
, x
4
) = 8x
2
−8x
1
= 0
f
/
2
(x
1
, x
2
, x
3
, x
4
) = 20 ÷8x
1
−12x
2
2
= 0
f
/
3
(x
1
, x
2
, x
3
, x
4
) = 48 −24x
3
= 0
f
/
4
(x
1
, x
2
, x
3
, x
4
) = 6 −2x
4
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
26 CHAPTER 3 STATI C OPTI MI ZATI ON
The ﬁrst equation gives x
2
= x
1
, and then the second equation gives 12x
2
1
−8x
1
−20 = 0 with the two
solutions x
1
= 5/3 and x
1
= −1. The last two equations determine x
3
and x
4
. There are two stationary
points, (5/3, 5/3, 2, 3) and (−1, −1, 2, 3). The Hessian matrix is
f
//
(x
1
, x
2
, x
3
, x
4
) =
⎛
⎜
⎜
⎝
−8 8 0 0
8 −24x
2
0 0
0 0 −24 0
0 0 0 −2
⎞
⎟
⎟
⎠
and the leading principal minors of the Hessian are
D
1
= −8, D
2
= 192x
2
−64 = 64(3x
2
−1), D
3
= −24D
2
, D
4
= 48D
2
At (−1, −1, 2, 3) we get D
2
< 0, sothis point is a saddle point. The other stationarypoint, (5/3, 5/3, 2, 3),
we get D
1
< 0, D
2
> 0, D
3
< 0, and D
4
> 0, so this point is a local maximum point.
3.3
3.3.2 (a) The admissible set is the intersection of the sphere x
2
÷y
2
÷z
2
= 216 and the plane x÷2y÷3z = 0,
which passes through the center of the sphere. This set (a circle) is closed and bounded (and nonempty!),
and by the extreme value theorem the objective function does attain a maximum over the admissible set.
The Lagrangian is L(x, y, z) = x ÷ 4y ÷ z − λ
1
(x
2
÷ y
2
÷ z
2
− 216) − λ
2
(x ÷ 2y ÷ 3z), and the
ﬁrstorder conditions are:
(i) 1 −2λ
1
x −λ
2
= 0, (ii) 4 −2λ
1
y −2λ
2
= 0, (iii) 1 −2λ
1
z −3λ
2
= 0
From (i) and (ii) we get λ
2
= 1 − 2λ
1
x = 2 − λ
1
y, which implies λ
1
(y − 2x) = 1. Conditions (i)
and (iii) yield 1 − 2λ
1
x =
1
3
−
2
3
λ
1
z = 0, which implies λ
1
(
2
3
z − 2x) = −
2
3
. Multiply by −
3
2
to get
λ
1
(3x −z) = 1.
It follows that y − 2x = 3x − z, so z = 5x − y. Inserting this expression for z in the constraint
x ÷ 2y ÷ 3z = 0 yields 16x − y = 0, so y = 16x and z = −11x. The constraint x
2
÷ y
2
÷ z
2
= 216
then yields (1 ÷ 256 ÷ 121)x
2
= 216, so x
2
= 216/378 = 4/7, and x = ±
2
7
√
7. Hence there are two
points that satisfy the ﬁrstorder conditions:
x
1
=
2
7
√
7,
32
7
√
7, −
22
7
√
7
, x
2
=
−
2
7
√
7, −
32
7
√
7,
22
7
√
7
The multipliers are then λ
1
= 1/(y −2x) = 1/(14x) = ±
1
28
√
7 and λ
2
= 1 −2λ
1
x =
6
7
.
The objective function, f (x, y, z) = x ÷ 4y ÷ z, attains it maximum value f
max
=
108
7
√
7 at x
1
,
with λ
1
=
1
28
√
7, λ
2
=
6
7
. (The point x
2
is the minimum point and f
min
= −
108
7
√
7.)
Comment: It is clear that the Lagrangian is concave if λ
1
> 0 and convex if λ
1
< 0. Therefore
part (b) of Theorem 3.3.1 is sufﬁcient to show that x
1
is a maximum point and x
2
is a minimum point in
this problem, so we did not need to use the extreme value theorem.
(b) Equation (3.3.10) shows that f
∗
≈ λ
1
· (−1) ÷λ
2
· 0.1 = −
1
28
√
7 ÷
0.6
7
≈ −0.009.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 3 STATI C OPTI MI ZATI ON 27
3.3.4 (a) With the Lagrangian L(x, y) =
1
2
ln(1 ÷x
1
) ÷
1
4
ln(1 ÷x
2
) −λ(2x
1
÷3x
2
−m), the ﬁrstorder
conditions are
(i)
1
2(1 ÷x
1
)
−2λ = 0, (ii)
1
4(1 ÷x
2
)
−3λ = 0
Equations (i) and (ii) yield
λ =
1
4(1 ÷x
1
)
=
1
12(1 ÷x
2
)
=⇒ 1 ÷x
1
= 3(1 ÷x
2
) ⇐⇒ x
1
−3x
2
= 2
Together with the constraint 2x
1
÷3x
2
= mthis gives x
1
= x
∗
1
(m) =
1
3
(m÷2), x
2
= x
∗
2
(m) =
1
9
(m−4),
and then λ =
3
4
(m÷5)
−1
.
(b) U
∗
(m) =
1
2
ln(1÷x
∗
1
(m))÷
1
4
ln(1÷x
∗
2
(m)) =
1
2
ln(
1
3
(m÷5))÷
1
4
ln(
1
9
(m÷5)) =
3
4
ln(m÷5)−ln 3,
so dU
∗
/dm =
3
4
(m÷5)
−1
= λ.
3.3.6 (a) The two constraints determine an ellipsoid centred at the origin and a plane through the origin,
respectively. The admissible set is the curve of intersection of these two surfaces, namely an ellipse. This
curve is closed and bounded (and nonempty), so the extreme value theorem guarantees the existence of
both maximum and minimum points. It is not very hard to show that the matrix g
/
(x) in Theorem 3.3.1
has rank 2 at all admissible points, so the usual ﬁrstorder conditions are necessary.
Lagrangian: L(x, y, z) = x
2
÷ y
2
÷ z
2
− λ
1
(x
2
÷ y
2
÷ 4z
2
− 1) − λ
2
(x ÷ 3y ÷ 2z). Firstorder
conditions:
(i) 2x −2λ
1
x −λ
2
= 0, (ii) 2y −2λ
1
y −3λ
2
= 0, (iii) 2z −8λ
1
z −2λ
2
= 0
From (i) and (ii) we get (iv) λ
2
= 2(1 −λ
1
)x =
2
3
(1 −λ
1
)y.
(A) If λ
1
= 1, then λ
2
= 0 and (iii) implies z = 0. The constraints reduce to x
2
÷y
2
= 1 and x÷3y = 0,
and we get the two candidates
(x, y, z) =
±
3
10
√
10 , ∓
1
10
√
10 , 0
(∗)
(B) If λ
1
,= 1, then (iv) implies y = 3x, and the second constraint gives z = −5x. The ﬁrst constraint
then yields x
2
÷9x
2
÷100x
2
= 1 which leads to the two candidates
(x, y, z) =
±
1
110
√
110, ±
3
110
√
110 , ∓
5
110
√
110
(∗∗)
In this case the multipliers λ
1
and λ
2
can be determined from equations (i) and (iii), and we get λ
1
=
7
22
and λ
2
=
15
11
x.
The objective function, x
2
÷y
2
÷z
2
, attains its maximum value 1 at the points (∗), while the points
(∗∗) give the minimum value 7/22. It is worth noting that with λ
1
= 1 the Lagrangian is concave (linear,
in fact), so Theorem3.3.1(b) shows that the points in (∗) are maximumpoints, even if we do not check the
rank condition in Theorem 3.3.1(a) (but without that rank condition we cannot be sure that the ﬁrstorder
conditions are necessary, so there might be other maximum points beside the two that we have found).
(There is a much simpler way to ﬁnd the maximum points in this problem: Because of the ﬁrst
constraint, the objective function x
2
÷y
2
÷z
2
equals 1−3z
2
, which obviously has a maximumfor z = 0.
We then just have to solve the equations x
2
÷y
2
= 1 and x ÷3y = 0 for x and y.)
(b) f
∗
≈ 1 · 0.05 ÷0 · 0.05 = 0.05.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
28 CHAPTER 3 STATI C OPTI MI ZATI ON
3.3.10 There was a misprint in this problem in the ﬁrst printing of the book: The constraints should all be
equality constraints, so please correct g
j
(x, r) ≤ 0 to g
j
(x, r) = 0 for j = 1, . . . , m.
Now consider two problems, namely the problem in (3.3.13) and the problem
max f (x, r) subject to
¸
g
j
(x, r) = 0,
r
i
= b
m÷i
,
j = 1, . . . , m
i = 1, . . . , k
(∗)
The Lagrangian for problem (3.3.13) is L(x, r) = f (x, r) −
¸
m
j=1
λ
j
g
j
(x, r) and the Lagrangian for
problem(∗) is
¯
L(x, r) = f (x, r) −
¸
m
j=1
λ
j
g
j
(x, r) −
¸
k
i=1
λ
m÷i
(r
i
−b
m÷i
). The ﬁrstorder conditions
for maximum in problem (∗) imply that
∂
¯
L(x, r)
∂r
i
=
∂f (x, r)
∂r
i
−
m
¸
j=1
∂g
j
(x, r)
∂r
i
−λ
m÷i
= 0, i = 1, . . . , k
Equation (3.3.9) implies that ∂f
∗
(r)/∂r
i
= λ
m÷i
, and so
∂f
∗
(r)
∂r
i
= λ
m÷i
=
∂f (x, r)
∂r
i
−
m
¸
j=1
∂g
j
(x, r)
∂r
i
=
∂L(x, r)
∂r
i
3.4
3.4.3 Lagrangian: L(x, y, z) = x ÷y ÷z−λ
1
(x
2
÷y
2
÷z
2
−1)−λ
2
(x −y −z−1). Firstorder conditions:
(i) 1 −2λ
1
x −λ
2
= 0, (ii) 1 −2λ
1
y ÷λ
2
= 0, (iii) 1 −2λ
1
z ÷λ
2
= 0
Equations (ii) and (iii) give 2λ
1
y = 2λ
1
z. If λ
1
= 0, then (i) and (ii) yield 1 −λ
2
= 0 and 1 ÷λ
2
= 0.
This is clearly impossible, so we must have λ
1
,= 0, and therefore y = z. We solve this equation together
with the two constraints, and get the two solutions
(x
1
, y
1
, z
1
) = (1, 0, 0), λ
1
= 1, λ
2
= −1
(x
2
, y
2
, z
2
) = (−
1
3
, −
2
3
, −
2
3
), λ
1
= −1, λ
2
=
1
3
In the secondderivative test (Theorem 3.4.1) we now have n = 3 and m = 2, so all we need check is
B
3
(x, y, z) =
0 0 2x 2y 2z
0 0 1 −1 −1
2x 1 −2λ
1
0 0
2y −1 0 −2λ
1
0
2z −1 0 0 −2λ
1
A little tedious computation yields B
3
(x
1
, y
1
, z
1
) = −16 and B
3
(x
2
, y
2
, z
2
) = 16. Hence (x
1
, y
1
, z
1
) is
a local maximum point and (x
2
, y
2
, z
2
) is a local minimum point.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 3 STATI C OPTI MI ZATI ON 29
3.5
3.5.2 (a) The Lagrangian is L(x, y) = ln(x ÷1) ÷ln(y ÷1) −λ
1
(x ÷2y −c) −λ
2
(x ÷y −2), and the
necessary Kuhn–Tucker conditions for (x, y) to solve the problem are:
L
/
1
(x, y) =
1
x ÷1
− λ
1
−λ
2
= 0 (1)
L
/
2
(x, y) =
1
y ÷1
−2λ
1
−λ
2
= 0 (2)
λ
1
≥ 0, and λ
1
= 0 if x ÷2y < c (3)
λ
2
≥ 0, and λ
2
= 0 if x ÷ y < 2 (4)
Of course, (x, y) must also satisfy the constraints
x ÷2y ≤ c (5)
x ÷ y ≤ 2 (6)
(b) Let c = 5/2. We consider the four possible combinations of λ
1
= 0, λ
1
> 0, λ
2
= 0, and λ
2
> 0.
(I) λ
1
= λ
2
= 0. This contradicts (1), so no candidates.
(II) λ
1
> 0, λ
2
= 0. From(3) and (5), x÷2y = 5/2. Moreover, by eliminatingλ
1
from(1) and (2) we get
x ÷1 = 2y ÷2. The last two equations have the solution x = 7/4, y = 3/8. But then x ÷y = 17/8 > 2,
contradicting (6). No candidates.
(III) λ
1
= 0, λ
2
> 0. From (4) and (6), x ÷ y = 2. Moreover, eliminating λ
2
from (1) and (2) we get
x = y, and so x = y = 1. But then x ÷2y = 3 > 5/2, contradicting (5). No candidates.
(IV) λ
1
> 0, λ
2
> 0. Then x ÷y = 2 and x ÷2y = 5/2, so x = 3/2, y = 1/2. We ﬁnd that λ
1
= 4/15
and λ
2
= 2/15, so this is a candidate, and the only one.
The Lagrangian is obviously concave in x and y, so (x, y) = (3/2, 1/2) solves the problem.
(c) If we assume that V(c) is a differentiable function of c, then formula (3.5.6) yields V
/
(5/2) = λ
1
=
4/15.
A direct argument can run as follows: For all values of c such that λ
1
and λ
2
are positive, x and y
must satisfy the constraints with equality, and so x = 4 − c and y = c − 2. Then equations (1) and (2)
yield
λ
1
=
1
y ÷1
−
1
x ÷1
=
1
c −1
−
1
5 −c
and λ
2
=
2
x ÷1
−
1
y ÷1
=
2
5 −c
−
1
c −1
It is clear that these expressions remain positive for c in an open interval around c = 5/2. (More precisely,
they are both positive if and only if 7/3 < c < 3.) For such c, the derivative of the value function is
V
/
(c) =
∂V
∂x
dx
dc
÷
∂V
∂y
dy
dc
= −
1
1 ÷x
÷
1
1 ÷y
= λ
1
3.5.3 We reformulate the problem as a standard maximization problem:
maximize −4 ln(x
2
÷2) −y
2
subject to
¸
−x
2
−y ≤ −2
−x ≤ −1
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
30 CHAPTER 3 STATI C OPTI MI ZATI ON
The Lagrangian is L(x, y) = −4 ln(x
2
÷ 2) − y
2
− λ
1
(−x
2
− y ÷ 2) − λ
2
(−x ÷ 1), so the necessary
Kuhn–Tucker conditions together with the constraints are:
L
/
1
= −
8x
x
2
÷2
÷2λ
1
x ÷λ
2
= 0 (i)
L
/
2
= −2y ÷λ
1
= 0 (ii)
λ
1
≥ 0, and λ
1
= 0 if x
2
÷y > 2 (iii)
λ
2
≥ 0, and λ
2
= 0 if x > 1 (iv)
x
2
÷y ≥ 2 (v)
x ≥ 1 (vi)
We try the four possible combinations of zero or positive multipliers:
(A) λ
1
= 0, λ
2
= 0. From (i) we see that x = 0, which contradicts x ≥ 1.
(B) λ
1
= 0, λ
2
> 0. From (iv) and (vi), x = 1, and (ii) gives y = 0. This contradicts (v).
(C) λ
1
> 0, λ
2
= 0. From (iii) and (v) we get x
2
÷y = 2. Equation (i) gives λ
1
= 4/(x
2
÷2), and then
(ii) gives y = λ
1
/2 = 2/(x
2
÷2). Inserted into x
2
÷y = 2, this gives x
4
= 2, or x =
4
√
2. It follows
that y = 2/(
√
2 ÷2) = 2 −
√
2, and λ
1
= 4 −2
√
2. So (x, y, λ
1
, λ
2
) = (
4
√
2, 2 −
√
2, 4 −2
√
2, 0)
is a candidate.
(D) λ
1
> 0, λ
2
> 0. Then (iii)–(vi) imply x
2
÷ y = 2 and x = 1. So x = y = 1. Then from (i),
λ
2
÷2λ
1
= 8/3 and (ii) gives λ
1
= 2. But then λ
2
= 8/3 −4 < 0. Contradiction.
Thus the only possible solution is the one given in (C), and the minimumvalue of f (x, y) = 4 ln(x
2
÷2)÷
y
2
under the given constraints is f (
4
√
2, 2−
√
2 ) = 4 ln(
√
2÷2)÷(2−
√
2 )
2
= 4 ln(
√
2÷2)÷6−4
√
2.
3.5.4 With the Lagrangian L(x, y) = −(x −a)
2
−(y −b)
2
−λ
1
(x −1)−λ
2
(y −2) we get the Kuhn–Tucker
necessary conditions for (x, y) to be a maximum point:
L
/
1
(x, y) = −2(x −a) −λ
1
= 0 (i)
L
/
2
(x, y) = −2(y −b) −λ
2
= 0 (ii)
λ
1
≥ 0, and λ
1
= 0 if x < 1 (iii)
λ
2
≥ 0, and λ
2
= 0 if y < 2 (iv)
Of course, a maximum point (x, y) must also satisfy the constraints (v) x ≤ 1 and (vi) y ≤ 2.
We try the four possible combinations of λ
1
= 0, λ
1
> 0, λ
2
= 0, and λ
2
> 0.
(A) λ
1
= 0, λ
2
= 0. Equations (i) and (ii) give x = a and y = b. Because of the constraints this is
possible only if a ≤ 1 and b ≤ 2.
(B) λ
1
= 0, λ
2
> 0. We get x = a and y = 2. Constraint (v) implies a ≤ 1, and equation (ii) yields
b = y ÷
1
2
λ
2
> y = 2.
(C) λ
1
> 0, λ
2
= 0. This gives x = 1, y = b, a = 1 ÷
1
2
λ
1
> 1, and b ≤ 2.
(D) λ
1
> 0, λ
2
> 0. With both multipliers positive, both constraints must be satisﬁed with equality:
x = 1, y = 2. We also get a = 1 ÷
1
2
λ
1
> 1 and b = 2 ÷
1
2
λ
2
> 2.
We observe that in each of the four cases (A)–(D) there is exactly one point that satisﬁes the Kuhn–
Tucker conditions, and since the objective function is concave, these points are maximum points. Which
case applies depends on the values of a and b. The solution can be summarized as x
∗
= min{a, 1},
y
∗
= min{b, 2}.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 3 STATI C OPTI MI ZATI ON 31
The admissible set in this problem is the same as in Example 3.5.1, and the maximization problem
is equivalent to ﬁnding an admissible point as close to (a, b) as possible. It is readily seen that the point
(x
∗
, y
∗
) given above is the optimal point. In particular, in case (A) the point (a, b) itself is admissible
and is also the optimal point.
3.5.6 (a) With the Lagrangian L(x, y) = x
5
− y
3
− λ
1
(x − 1) − λ
2
(x − y) the Kuhn–Tucker conditions
and the constraints are
L
/
1
(x, y) = 5x
4
−λ
1
−λ
2
= 0 (i)
L
/
2
(x, y) = −3y
2
÷λ
2
= 0 (ii)
λ
1
≥ 0, and λ
1
= 0 if x < 1 (iii)
λ
2
≥ 0, and λ
2
= 0 if x < y (iv)
x ≤ 1 (v)
x ≤ y (vi)
Consider the four possible combinations of zero or positive multipliers:
(A) λ
1
= 0, λ
2
= 0. Equations (i) and (ii) give x = 0 and y = 0. Thus (x
1
, y
1
) = (0, 0) is a candidate
for optimum.
(B) λ
1
= 0, λ
2
> 0. Since λ
2
> 0, the complementary slackness condition (iv) tells us that we cannot
have x < y, while constraint (vi) says x ≤ y. Therefore x = y. Equations (i) and (ii) then give
5x
4
= 0 ÷λ
2
= 3y
2
, and thus 5y
4
= 3y
2
. Since λ
2
,= 0 we have y ,= 0, and therefore 5y
2
= 3, so
x = y = ±
√
3/5 = ±
√
15/25 = ±
1
5
√
15. We get two new candidates, (x
2
, y
2
) = (
1
5
√
15,
1
5
√
15 )
and (x
3
, y
3
) = (−
1
5
√
15, −
1
5
√
15 ).
(C) λ
1
> 0, λ
2
= 0. Now (ii) gives y = 0, while (iii) and (v) give x = 1. But this violates constraint
(vi), so we get no new candidates for optimum here.
(D) λ
1
> 0, λ
2
> 0. The complementary slackness conditions show that in this case both constraints
must be satisﬁed with equality, so we get one new candidate point, (x
4
, y
4
) = (1, 1)
Evaluating the objective function h(x, y) = x
5
− y
3
at each of the four maximum candidates we have
found shows that
h(x
1
, y
1
) = h(x
4
, y
4
) = 0
h(x
2
, y
2
) = x
5
2
−y
3
2
= x
5
2
−x
3
2
= (x
2
2
−1)x
3
2
= −
2
5
x
3
2
= −
6
125
√
15
h(x
3
, y
3
) = h(−x
2
, −y
2
) = −h(x
2
, y
2
) =
6
125
√
15
(For the last evaluation we used the fact that h is an odd function, i.e. h(−x, −y) = −h(x, y).) Hence, if
there is a maximumpoint for h in the feasible set, then (x
3
, y
3
) is that maximumpoint. Note that although
the feasible set is closed, it is not bounded, so it is not obvious that there is a maximum. But part (b) of
this problem will show that a maximum point does exist.
(b) Let h(x, y) = x
5
−y
3
and deﬁne f (x) = max
y≥x
h(x, y). Since h(x, y) is strictly decreasing with
respect to y, it is clear that we get the maximum when y = x, so f (x) = h(x, x) = x
5
−x
3
. To ﬁnd the
maximum of f (x) for x ≤ 1 we take a look at
f
/
(x) = 5x
2
−3x
2
= 5x
2
(x
2
−
3
5
) = 5x
2
(x ÷x
2
)(x −x
2
)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
32 CHAPTER 3 STATI C OPTI MI ZATI ON
where x
2
=
√
3/5 =
1
5
√
15 (as in part (a)). It is clear that
f
/
(x) > 0 if x ∈ (−∞, −x
2
) or x ∈ (x
2
, ∞)
f
/
(x) < 0 if x ∈ (−x
2
, 0) or x ∈ (0, x
2
)
Therefore f is strictly increasing in the interval (−∞, −x
2
], strictly decreasing in [−x
2
, x
2
], and strictly
increasing again in [x
2
, ∞). It follows that f (x) will reach its highest value when x = −x
2
or when
x = 1. (Draw a graph!) Since f (−x
2
) =
6
125
and f (1) = 0, the maximum point is −x
2
(= x
3
).
So why does this show that the point (x
3
, y
3
) that we found in part (a) really is a maximum point in
that problem? The reason is that for every point (x, y) with x ≤ 1 and x ≤ y we have
h(x, y) ≤ f (x) ≤ f (x
3
) = h(x
3
, x
3
) = h(x
3
, y
3
)
3.6
3.6.2 Lagrangian: L(x, y) = xy ÷x ÷y −λ
1
(x
2
÷y
2
−2) −λ
2
(x ÷y −1). Firstorder conditions:
L
/
1
(x, y) = y ÷1 −2λ
1
x −λ
2
= 0 (i)
L
/
2
(x, y) = x ÷1 −2λ
1
y −λ
2
= 0 (ii)
λ
1
≥ 0, and λ
1
= 0 if x
2
÷y
2
< 2 (iii)
λ
2
≥ 0, and λ
2
= 0 if x ÷y < 1 (iv)
It is usually a good idea to exploit similarities and symmetries in the ﬁrstorder conditions. In the present
case, we can eliminate λ
2
from (i) and (ii) to get
y −2λ
1
x = x −2λ
1
y ⇐⇒ (1 ÷2λ
1
)(y −x) = 0
Since 1÷2λ
1
≥ 1, this implies y = x for any point that satisﬁes the ﬁrstorder conditions. Now consider
the various combinations of zero or positive values of λ
1
and λ
2
:
(A) λ
1
= 0, λ
2
= 0. Equations (i) and (ii) give (x, y) = (−1, −1).
(B) λ
1
= 0, λ
2
> 0. Since λ
2
> 0, we must have x÷y = 1, and since x = y, we get (x, y) = (1/2, 1/2).
(C) λ
1
> 0, λ
2
= 0. Now x
2
÷ y
2
= 2 because λ
1
> 0, and since x = y we get x = y = ±1. The
point (1, 1) violates the constraint x ÷ y ≤ 1. If x = y = −1, then equation (i) yields 2λ
1
= λ
2
= 0,
contradicting the assumption λ
1
> 0. Thus there are no candidate points in this case. (We did get the
point (−1, −1) in case (A) above.)
(D) λ
1
> 0, λ
2
> 0. In this case both constraints must be active, i.e. x
2
÷y
2
= 1 and x ÷y = 1. Since
x = y, the ﬁrst constraint yields x = y = 1/2, but then x
2
÷y
2
,= 2. So no candidates in this case either.
Comparing the values of the objective function xy ÷ x ÷ y at the two points (−1, −1) and (1/2, 1/2)
shows that (1/2, 1/2) must be the maximum point. (The extreme value theorem guarantees that there
really is a maximum point.)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 3 STATI C OPTI MI ZATI ON 33
3.7
3.7.1 (a) The Lagrangian is L(x, y) = 100 −e
−x
−e
−y
−e
−z
−λ
1
(x ÷y ÷z −a) −λ
2
(x −b) and the
Kuhn–Tucker conditions are
e
−x
−λ
1
−λ
2
= 0 (i)
e
−y
−λ
1
= 0 (ii)
e
−z
−λ
1
= 0 (iii)
λ
1
≥ 0, and λ
1
= 0 if x ÷y ÷z < a (iv)
λ
2
≥ 0, and λ
2
= 0 if x < b (v)
Equations (ii) and (iii) imply that λ
1
> 0 and y = z. From (iv) we get x ÷y ÷z = a, so x ÷2y = a.
(A) Suppose λ
2
= 0. Then (i) and (ii) imply λ
1
= e
−x
and x = y. Hence x ÷ y ÷ z = 3x = a, so
x = a/3, and therefore a/3 ≤ b, i.e. a ≤ 3b.
(B) Suppose λ
2
> 0. Condition (v) now implies x = b, and so y = (a − x)/2 = (a − b)/2. Then
λ
1
= e
−y
= e
−(a−b)/2
, and(i) yields λ
2
= e
−b
−e
−(a−b)/2
. Since λ
2
> 0, we must have −b > −(a−b)/2,
i.e. a > 3b.
Thus, the Kuhn–Tucker conditions have a unique solution in each of the two cases a ≤ 3b and a > 3b.
The Lagrangian is concave, so we know that the points we have found really are optimal.
(b) See the answer in the book for the evaluation of ∂f
∗
(a, b)/∂a and ∂f
∗
(a, b)/∂b. Strictly speaking,
if a = 3b, you must consider the onesided derivatives and show that the left and right derivatives are the
same, but things work out all right in that case too.
3.7.3 (a) Consider the maximization problem and write it in standard form as
maximize f (x, y) = x
2
÷y
2
subject to
¸
2x
2
÷4y
2
≤ s
2
−2x
2
−4y
2
≤ −r
2
(∗)
(∗∗)
We use the Lagrangian L(x, y, r, s) = x
2
÷y
2
−λ
1
(2x
2
÷4y
2
−s
2
) ÷λ
2
(2x
2
÷4y
2
−r
2
), and get the
Kuhn–Tucker conditions
L
/
1
= 2x −4λ
1
x ÷4λ
2
x = 0 (i)
L
/
2
= 2y −8λ
1
y ÷8λ
2
y = 0 (ii)
λ
1
≥ 0, and λ
1
= 0 if 2x
2
÷4y
2
< s
2
(iii)
λ
2
≥ 0, and λ
2
= 0 if 2x
2
÷4y
2
> r
2
(iv)
If λ
1
= 0, then (i) and (ii) would yield
(2 ÷4λ
2
)x = 0 =⇒ x = 0, (2 ÷8λ
2
)y = 0 =⇒ y = 0
which contradicts the constraint 2x
2
÷4y
2
≥ r
2
. Therefore we must have λ
1
> 0 and 2x
2
÷4y
2
= s
2
.
Moreover, we cannot have λ
2
> 0, for that would imply 2x
2
÷ 4y
2
= r
2
< s
2
, which contradicts
2x
2
÷4y
2
= s
2
. Hence, λ
1
> 0, λ
2
= 0. Equations (i) and (ii) reduce to
(i
/
) (2 −4λ
1
)x = 0 and (ii
/
) (2 −8λ
1
)y = 0
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
34 CHAPTER 3 STATI C OPTI MI ZATI ON
If x = 0, then y ,= 0 (because 2x
2
÷ 4y
2
= s
2
> 0), and (ii
/
) implies λ
1
= 1/4. If x ,= 0, then (i
/
)
implies λ
1
= 1/2. We are left with the two possibilities
(A) λ
1
= 1/2, y = 0, x = ±
1
2
√
2 s (B) λ
1
= 1/4, x = 0, y = ±
1
2
s
Case (A) gives the optimum points, (x
∗
, y
∗
) = (±
1
2
√
2 s, 0), and f
∗
(r, s) = f (x
∗
, y
∗
) =
1
2
s
2
. To verify
the envelope result (3.7.5), note that
∂L(x, y, r, s)/∂r = −2λ
2
r, ∂L(x, y, r, s)/∂s = 2λ
1
s
If we insert the optimal values of x
∗
and y
∗
and the corresponding values λ
1
= 1/2 and λ
2
= 0 of the
multipliers, we get
∂L(x
∗
, y
∗
, r, s)/∂r = 0 = ∂f
∗
(r, s)/∂r, ∂L(x
∗
, y
∗
, r, s)/∂s = s = ∂f
∗
(r, s)/∂s
in accordance with (3.7.5).
(b) The minimization problemis equivalent to maximizing g(x, y) = −f (x, y) subject to the constraints
(∗) and (∗∗) in part (a). We get a new Lagrangian L(x, y, r, s) = −x
2
− y
2
− λ
1
(2x
2
÷ 4y
2
− s
2
) ÷
λ
2
(2x
2
÷4y
2
−r
2
), and the ﬁrstorder conditions are as in (a), except that (i) and (ii) are replaced by
L
/
1
= −2x −4λ
1
x ÷4λ
2
x = 0 (i
/
)
L
/
2
= −2y −8λ
1
y ÷8λ
2
y = 0 (ii
/
)
The solution proceeds along the same lines as in (a), but this time λ
2
= 0 is impossible, so we get λ
2
> 0
and λ
1
= 0. The optimum points are (x
∗
, y
∗
) = (0, ±
1
2
r), with λ
2
= 1/4, and g
∗
(r, s) = g(x
∗
, y
∗
) =
−(x
∗
)
2
−(y
∗
)
2
= −
1
4
r
2
, and the minimum value of f is f
∗
(r, s) = −g
∗
(r, s) =
1
4
r
2
. The equations in
(3.7.5) now become
L
/
r
(x
∗
, y
∗
, r, s) = −2λ
2
r = −
1
2
r =
∂g
∗
∂r
= −
∂f
∗
∂r
, L
/
s
(x
∗
, y
∗
, r, s) = 2λ
1
s = 0 =
∂g
∗
∂s
= −
∂f
∗
∂s
(c) The admissible set is the area between two ellipses, and the problems in (a) and (b) are equivalent to
ﬁnding the largest and the smallest distance from the origin to a point in this admissible set.
3.7.4 Let r and s be points in the domain of f
∗
, let λ ∈ [0, 1], and put t = λr÷(1−λ)s. We want to prove that
f
∗
(t) ≥ λf
∗
(r)÷(1−λ)f
∗
(s). There are points x and y such that f
∗
(r) = f (x, r) and f
∗
(s) = f (y, s).
Let w = λx ÷(1 −λ)y. Since g is convex, we have g(w, t) λg(x, r) ÷(1 −λ)g(y, s) 0, so (w, t)
is admissible. And since f is concave, we have
f
∗
(t) ≥ f (w, t) = f (λ(x, r) ÷(1 −λ)(y, s))
≥ λf (x, r) ÷(1 −λ)f (y, s) = λf
∗
(r) ÷(1 −λ)f
∗
(s)
3.8
3.8.2 (a) Lagrangian: L(x, y) = xy −λ(x ÷2y −2). Kuhn–Tucker conditions:
∂L/∂x = y − λ ≤ 0 (= 0 if x > 0) (i)
∂L/∂y = x −2λ ≤ 0 (= 0 if y > 0) (ii)
λ ≥ 0 (= 0 if x ÷2y < 2) (iii)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 3 STATI C OPTI MI ZATI ON 35
There are admissible points where xy > 0, so we cannot have x = 0 or y = 0 at the optimum point
or points. It follows that (i) and (ii) must be satisﬁed with equality, and λ must be positive. Hence,
x = 2λ = 2y and x ÷ 2y = 2, so x = 1, y = 1/2, λ = 1/2. (The extreme value theorem guarantees
that there is a maximum, since the admissible set is closed and bounded. It is the closed line segment
between (2, 0) and (0, 1).)
(b) As in part (a), there are admissible points where x
α
y
β
> 0, so we may just as well accept (2, 0) and
(0, 1) as admissible points and replace the constraints x > 0 and y > 0 by x ≥ 0 and y ≥ 0. Then the
extreme value theorem guarantees that there is a maximum point in this case too, and it is clear that both
x and y must be positive at the optimum. With the Lagrangian L(x, y) = x
α
y
β
−λ(x ÷2y −2) we get
the Kuhn–Tucker conditions
∂L/∂x = αx
α−1
y
β
− λ ≤ 0 (= 0 if x > 0) (i)
∂L/∂y = βx
α
y
β−1
−2λ ≤ 0 (= 0 if y > 0) (ii)
λ ≥ 0 (= 0 if x ÷2y < 2) (iii)
It is clear that (i), (ii), and (iii) must all be satisﬁed with equality, and that λ > 0. From (i) and (ii) we
get βx
α
y
β−1
= 2λ = 2αx
α−1
y
β
, so βx = 2αy. This equation, combined with (iii) yields the solution:
x = 2α/(α ÷β), y = β/(α ÷β).
(If we had not extended the admissible set to include the end points, then we could not have used the
extreme value theorem to guarantee a maximum, but with the conditions on α and β the Lagrangian is
concave, so we could still be certain that the point we have found is a maximum point. But the argument
above, with a closed and bounded admissible set, works for all positive values of α and β, even if L is
not concave.)
3.8.3 With the Lagrangian L(x, y, c) = cx ÷y −λ(x
2
÷3y
2
−2) we get the Kuhn–Tucker conditions:
L
/
1
= c −2λx ≤ 0 (= 0 if x > 0) (i)
L
/
2
= 1 −6λy ≤ 0 (= 0 if y > 0) (ii)
λ ≥ 0 (= 0 if x
2
÷3y
2
< 2) (iii)
If λ = 0, then (ii) implies 1 ≤ 0, but that is impossible. Hence, λ > 0 and x
2
÷ 3y
2
= 2. Further, (ii)
implies 6λy ≥ 1, so y > 0. Therefore (ii) is an equality and y = 1/6λ.
(A) If x = 0, then (i) implies c ≤ 2λx = 0. Further, 3y
2
= 2, so y =
√
2/3 =
√
6/3, and λ = 1/6y =
√
6/12.
(B) If x > 0, then (i) is satisﬁed with equality and c = 2λx > 0, and x = c/2λ. The equation
x
2
÷3y
2
= 2 then leads to λ =
6(3c
2
÷1)/12, x = 6c/
6(3c
2
÷1), and y = 2/
6(3c
2
÷1).
Since the admissible set is closed and bounded and the objective function f (x, y) = cx ÷y is continuous,
the extreme value theorem guarantees that there is a maximum point for every value of c. The cases (A)
and (B) studied above show that the Kuhn–Tucker conditions have exactly one solution in each case, so
the solutions we found above are the optimal ones.
If c ≤ 0, then we are in case (A) and f
∗
(c) = cx
∗
÷y
∗
=
√
6/3.
If c > 0, then we are in case (B) and f
∗
(c) = cx
∗
÷y
∗
=
6(3c
2
÷1)/3.
The value function f
∗
(c) is obviously continuous for c ,= 0, and because lim
c→0
÷ f
∗
(c) =
√
6/3 =
f
∗
(0) = lim
c→0
− f
∗
(c), it is continuous at c = 0 too. The value function is differentiable at all c ,= 0,
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
36 CHAPTER 3 STATI C OPTI MI ZATI ON
and it is not hard to show that both onesided derivatives of f
∗
(c) at c = 0 are 0, so f
∗
is differentiable
there too.
For c ≤ 0 we get (f
∗
)
/
(c) = 0 = x
∗
, and a little calculation shows that (f
∗
)
/
(c) = x
∗
for all c > 0 as
well. Thus (f
∗
)
/
(c) = L
/
3
(x
∗
, y
∗
, c) for all c in accordance with equation (3.7.5).
3.8.5 See Fig. A3.8.5 in the answer section of the book. Since the Lagrangian
L(x, y) = x ÷y −
1
2
(x ÷y)
2
−
1
4
x −
1
3
y −λ
1
(x −5) −λ
2
(y −3) −λ
3
(−x ÷2y −2)
is concave, a point that satisﬁes Kuhn–Tucker conditions must be a maximum point. The objective
function has no stationary points, so any maximum points must lie on the boundary of S. The Kuhn–
Tucker conditions are:
L
/
1
(x, y) = 1 −(x ÷y) −
1
4
−λ
1
÷ λ
3
≤ 0 (= 0 if x > 0) (i)
L
/
2
(x, y) = 1 −(x ÷y) −
1
3
−λ
2
−2λ
3
≤ 0 (= 0 if y > 0) (ii)
λ
1
≥ 0 (= 0 if x < 5) (iii)
λ
2
≥ 0 (= 0 if y < 3) (iv)
λ
3
≥ 0 (= 0 if −x ÷2y < 2) (v)
The solution is x = 3/4, y = 0, with λ
1
= λ
2
= λ
3
= 0.
Once we have foundor beentoldabout this point it is easytocheckthat it satisﬁes (i)–(v), but otherwise
it can be a very tedious job to go through all possible combinations of zero or positive multipliers as well
as x = 0 or x > 0 and y = 0 or y > 0. In this problem there are 32 different combinations to check if
we do not see any shortcuts. In fact it would probably be more efﬁcient to check each of the ﬁve straight
line segments that form the boundary of S. But it would hardly be practical to do that in a problem with
more than two variables, because it quickly becomes difﬁcult to visualize the geometry of the admissible
set.
3.8.6 (a) The last inequality in (∗) gives
m
¸
j=1
λ
∗
j
(g
j
(x
∗
) −b
j
) ≥
m
¸
j=1
λ
j
(g
j
(x
∗
) −b
j
) for all λ 0 (∗∗)
If g
k
(x
∗
) > b
k
for some k, then
¸
m
j=1
λ
j
(g
j
(x
∗
) −b
j
) can be made arbitrary large by choosing λ
k
large
and λ
j
= 0 for all j ,= k. Hence, g
j
(x
∗
) ≤ b
j
, j = 1, . . . , m. By choosing all λ
j
equal to 0 in (∗∗),
we get
¸
m
j=1
λ
∗
j
(g
j
(x
∗
) − b
j
) ≥ 0. Now, λ
∗
j
≥ 0 and g
j
(x
∗
) ≤ b
j
for every j, so each λ
∗
j
(g
j
(x
∗
) − b
j
)
must be zero, and then
¸
m
j=1
λ
∗
j
(g
j
(x
∗
) − b
j
) = 0. Finally, whenever x is admissible, the inequality
¨
L(x, λ
∗
) ≤
¨
L(x
∗
, λ
∗
) implies that f (x)−f (x
∗
) ≤
¸
m
j=1
λ
∗
j
[g
j
(x)−g
j
(x
∗
)] =
¸
m
j=1
λ
∗
j
[g
j
(x)−b
j
] ≤ 0.
Therefore f (x) ≤ f (x
∗
), so x
∗
solves problem (1).
(b) Proof of the second inequality in (∗): Under the given assumptions,
¨
L(x
∗
, λ
∗
) −
¨
L(x
∗
, λ) =
¸
m
j=1
λ
j
[g
j
(x
∗
) −b
j
] −
¸
m
j=1
λ
∗
j
[g
j
(x
∗
) −b
j
] =
¸
m
j=1
λ
j
[g
j
(x
∗
) −b
j
] ≤ 0 when λ ≥ 0. Since the ﬁrst
inequality in (∗) is assumed, we have shown that (x
∗
, λ
∗
) is a saddle point for
¨
L.
3.9
3.9.1 (A) implies that π(x
∗
) ≥ π(ˆ x), i.e. f (x
∗
) −
¸
m
j=1
λ
j
g
j
(x
∗
) ≥ f (ˆ x) −
¸
m
j=1
λ
j
g
j
(ˆ x). But, because ˆ x
also solves (3.9.1), f (ˆ x) = f (x
∗
) and then
¸
m
j=1
λ
j
g
j
(ˆ x) ≥
¸
m
j=1
λ
j
g
j
(x
∗
). Thus, because λ
j
≥ 0 and
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 3 STATI C OPTI MI ZATI ON 37
g
j
(ˆ x) ≤ b
j
, j = 1, . . . , m, and also because of (3.9.5), we have
m
¸
j=1
λ
j
b
j
≥
m
¸
j=1
λ
j
g
j
(ˆ x) ≥
m
¸
j=1
λ
j
g
j
(x
∗
) =
m
¸
j=1
λ
j
b
j
(∗)
Here the two middle terms, being squeezed between two equal numbers, must themselves be equal.
Therefore f (ˆ x) −
¸
m
j=1
λ
j
g
j
(ˆ x) = f (x
∗
) −
¸
m
j=1
λ
j
g
j
(x
∗
) ≥ f (x) −
¸
m
j=1
λ
j
g
j
(x) for all x 0,
proving (A). Also, if g
k
(ˆ x) < b
k
and λ
k
> 0 for any k, then
¸
m
j=1
λ
j
(g
j
(ˆ x) −b
j
) < 0, which contradicts
(∗). Thus ˆ x satisﬁes (A)–(C).
3.10
3.10.1 (a) Since the admissible set is closed and bounded, there is at least one maximum point in this
problem. We use Theorem 3.8.3 (necessary ﬁrstorder conditions for problems with mixed constraints).
With L(x, y, z, a) = x
2
÷y
2
÷z
2
−λ(2x
2
÷y
2
÷z
2
−a
2
) −μ(x ÷y ÷z), the necessary conditions
are:
∂L/∂x = 2x −4λx −μ = 0 ⇐⇒ 2(1 −2λ)x = μ (i)
∂L/∂y = 2y −2λy −μ = 0 ⇐⇒ 2(1 −λ)y = μ (ii)
∂L/∂z = 2z −2λz −μ = 0 ⇐⇒ 2(1 −λ)z = μ (iii)
λ ≥ 0, and λ = 0 if 2x
2
÷y
2
÷z
2
< a
2
(iv)
There is no sign restriction on μ, since the corresponding constraint is an equality, not an inequality.
If λ = 0, we get x = y = z = μ/2, which implies 3x = 0, so x = y = z = μ = 0. But the point
(x, y, z) = (0, 0, 0) is obviously not a maximum point but a global minimum point in this problem.
If λ > 0, then 2x
2
÷y
2
÷z
2
= a
2
because of complementary slackness. There are two possibilities:
λ = 1 and λ ,= 1.
(A) If λ ,= 1 then (ii) and (iii) imply y = z, so the constraints yield x ÷2y = 0 and 2x
2
÷2y
2
= a
2
,
with the solutions (x, y, z) = (±
1
5
√
10 a, ∓
1
10
√
10 a, ∓
1
10
√
10 a). Since x = −2y, equations (i) and (ii)
yield −4(1 −2λ)y = μ = 2(1 −λ)y, so λ = 3/5.
(B) If λ = 1, then μ = 0 and x = 0, so y ÷z = 0 and y
2
÷z
2
= a
2
, with y = −z =
a
2
/2. This
gives the two points (x, y, z) = (0, ±
1
2
√
2 a, ∓
1
2
√
2 a).
If we evaluate the objective function x
2
÷y
2
÷z
2
at each of the points we have found, we ﬁnd that
the two points (x
∗
, y
∗
, z
∗
) = (0, ±
1
2
√
2 a, ∓
1
2
√
2 a), with λ = 1, μ = 0 both solve the problem. We
have found several other points that satisfy the necessary conditions for a maximumbut are not maximum
points. Such is often the case when the Lagrangian is not concave.
(b) f
∗
(a) = a
2
, so df
∗
(a)/da = 2a, and ∂L(x
∗
, y
∗
, z
∗
, a)/∂a = 2λa = 2a.
3.11
3.11.1 Assume ﬁrst that f is a function of just one variable, and let x
0
be any point in A ⊆ ޒ. There are
three cases to consider:
(I) If f (x
0
) > 0, then f (x) > 0 for all x in an open interval U around x
0
(because f is continuous).
Then, for all x in U we have f
÷
(x)
2
= f (x)
2
, and so (d/dx)(f
÷
(x)
2
) = 2f (x)f
/
(x) = 2f
÷
(x)f
/
(x).
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
38 CHAPTER 4 TOPI CS I N I NTEGRATI ON
(II) If f (x
0
) < 0, then there is an open interval V around x
0
such that for all x in V we have f (x) < 0
and f
÷
(x) = 0, and then (d/dx)(f
÷
(x)
2
) = 0 = 2f
÷
(x)f
/
(x).
(III) (The most interesting case.) If f (x
0
) = 0, let K = [f
/
(x
0
)[. There is then an open interval W
around x
0
such that for all x ,= x
0
in W,
f (x) −f (x
0
)
x −x
0
−f
/
(x
0
)
< 1, which implies
f (x) −f (x
0
)
x −x
0
< [f
/
(x
0
)[ ÷1 = K ÷1
and therefore
[f (x)[ = [f (x) −f (x
0
)[ < (K ÷1)[x −x
0
[
Then for all x ,= x
0
in W, we get
f
÷
(x)
2
−f
÷
(x
0
)
2
x −x
0
=
f
÷
(x)
2
x −x
0
≤
f (x)
2
x −x
0
< (K ÷1)
2
[x −x
0
[ →0 as x →x
0
and so ((f
÷
)
2
)
/
(x
0
) = 0 = 2f
÷
(x
0
)f
/
(x
0
).
Thus, in all three cases we have shown that (d/dx)(f
÷
(x)
2
) = 2f
÷
(x)f
/
(x). This result immediately
carries over to the partial derivatives of f
÷
(x)
2
if f is a function ޒ
n
→ޒ:
∂
∂x
i
(f
÷
(x))
2
= 2f
÷
(x)
∂
∂x
i
f (x)
and the gradient of f
÷
(x)
2
is
∇(f
÷
(x)
2
) =
∂
∂x
1
(f
÷
(x))
2
, . . . ,
∂
∂x
n
(f
÷
(x))
2
= 2f
÷
(x)
∂
∂x
1
f (x), . . . ,
∂
∂x
n
f (x)
= 2f
÷
(x)∇f (x)
Note that f need not really be C
1
. All that is needed is that all the ﬁrstorder partial derivatives of f exist.
Chapter 4 Topics in Integration
4.1
4.1.5 (a) Expand the integrand. We get
9
4
(
√
x −1)
2
x
dx =
9
4
x −2
√
x ÷1
x
dx =
9
4
1 −
2
√
x
÷
1
x
dx =
9
4
(x −4
√
x ÷ln x) = 1 ÷ln
9
4
(b) With u = 1 ÷
√
x we get x = (u −1)
2
, dx = 2(u −1) du, and
1
0
ln(1 ÷
√
x ) dx =
2
1
2(u −1) ln u du =
2
1
(u −1)
2
ln u −
2
1
(u −1)
2
u
du
= ln 2 −
2
1
1
2
u
2
−2u ÷ln u
=
1
2
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 4 TOPI CS I N I NTEGRATI ON 39
(c) Let u = 1 ÷x
1/3
. Then x = (u −1)
3
, dx = 3(u −1)
2
du, and
27
0
x
1/3
1 ÷x
1/3
dx =
4
1
u −1
u
3(u −1)
2
du =
4
1
3u
3
−9u
2
÷9u −3
u
du = · · · =
45
2
−3 ln 4
4.2
4.2.2 We use formula (4.2.1) and then introduce u = αx
2
as a new variable, assuming α ,= 0. This yields
F
/
(α) =
1
0
∂
∂α
(xe
αx
2
) dx =
1
0
x
3
e
αx
2
dx =
1
2α
2
α
0
ue
u
du =
1
2α
2
α
0
(ue
u
−e
u
) =
αe
α
−e
α
÷1
2α
2
Direct calculation of F yields F(α) =
1
2α
1
0
e
αx
2
=
e
α
−1
2α
, which gives F
/
(α) =
αe
α
−e
α
÷1
2α
2
,
conﬁrming the result above.
We assumed above that α ,= 0. If α = 0, then formula (4.2.1) yields F
/
(0) =
1
0
x
3
dx = 1/4. To
get the answer by differentiating F directly, we need to know that F(0) =
1
0
x dx = 1/2. Then
F
/
(0) = lim
α→0
F(α) −F(0)
α
= lim
α→0
e
α
−1 −α
2α
2
=
“0
0
”
= · · · =
1
4
as it should be.
4.2.6 By Leibniz’s formula (Theorem 4.2.1),
˙ x(t ) = e
−δ(t −τ)
y(t ) ÷
t
−∞
−δe
−δ(t −τ)
y(τ) dτ = y(t ) −δx(t )
4.2.8 See the answer in the book. In order to use Leibniz’s formula (Theorem 4.2.2 in this case) we need
to know that there exist functions p(τ) and q(τ) such that
t
−∞
p(τ) dτ and
t
−∞
q(τ) dτ converge, and
such that [f
/
(t −τ)k(τ)[ ≤ p(τ) and [G
/
t
(τ, t )[ ≤ q(τ) for all τ ≤ t . Since G
/
(τ, t ) = −k(τ)f (t −τ),
the inequality for G
/
t
boils down to [k(τ)f (t −τ)[ ≤ q(τ) for all τ ≤ t .
4.2.10 (a) g
/
(Q) = c ÷h
Q
0
f (D) dD −p
a
Q
f (D) dD, and g
//
(Q) = (h ÷p)f (Q) ≥ 0.
(b) Since
a
0
f (D) dD = 1, we have
a
Q
f (D) dD =
a
0
f (D) dD −
Q
0
f (D) dD = 1 −
Q
0
f (D) dD
and therefore
g
/
(Q) = c −p ÷(h ÷p)
Q
0
f (D) dD
Since Q
∗
is the minimum point of g(Q), it must be a stationary point—that is, g
/
(Q
∗
) = 0. Therefore
c −p ÷(h ÷p)F(Q
∗
) = 0, which implies F(Q
∗
) = (p −c)/(p ÷h).
4.3
4.3.2 From the functional equation (4.3.2), (
3
2
) = (
1
2
÷1) =
1
2
(
1
2
) =
1
2
√
π. The given formula is thus
correct for n = 1. Suppose it is correct for n = k. Then, using (4.3.2),
(k ÷1 ÷
1
2
) = ((k ÷
1
2
) ÷1) = (k ÷
1
2
)(k ÷
1
2
) = (k ÷
1
2
)
(2k −1)!
2
2k−1
(k −1)!
√
π
=
2k ÷1
2
(2k −1)!
2
2k−1
(k −1)!
√
π =
(2k −1)! 2k(2k ÷1)
2 · 2k · 2
2k−1
(k −1)!
√
π =
(2k ÷1)!
2
2k÷1
k!
√
π
Thus the proposed formula is valid also for n = k ÷1. By mathematical induction it is true for all natural
numbers n.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
40 CHAPTER 4 TOPI CS I N I NTEGRATI ON
4.3.5 (a) Introduce u = λx as a new variable. Then
∞
−∞
f (x) dx =
λ
α
(α)
∞
0
x
α−1
e
−λx
dx =
λ
α
(α)
∞
0
(u/λ)
α−1
e
−u
1
λ
du =
1
(α)
∞
0
u
α−1
e
−u
du = 1
(b) M(t ) =
∞
−∞
e
t x
f (x) dx =
λ
α
(α)
∞
0
x
α−1
e
−(λ−t )x
dx. With u = (λ − t )x as a new variable we
get
M(t ) =
λ
α
(α)
∞
0
u
α−1
e
−u
(λ −t )
α
du =
λ
α
(α)
(α)
(λ −t )
α
=
λ
λ −t
α
= λ
α
(λ −t )
−α
Differentiation gives M
/
(t ) = αλ
α
(λ −t )
−α−1
and in general
M
(n)
(t ) = α(α ÷1) · · · (α ÷n −1)λ
α
(λ −t )
−α−n
=
(α ÷n)
(α)
λ
α
(λ −t )
α÷n
Hence,
M
/
(0) =
α
λ
and M
(n)
(0) =
α(α ÷1) · · · (α ÷n −1)
λ
n
=
(α ÷n)
(α)λ
n
(Alternatively we could have used the formula M
(n)
(0) =
∞
−∞
x
n
f (x) dx from Problem 4.2.5.)
4.4
4.4.2 The inner integral is
1
x
2
b
0
1
x
e
y/x
dy =
1
x
2
y=b
y=0
e
y/x
=
1
x
2
e
b/x
−
1
x
2
, so I =
a
1
1
x
2
e
b/x
−
1
x
2
dx.
With w =
1
x
, we get x =
1
w
and dx = −
1
w
2
dw. Therefore I =
1/a
1
(w
2
e
bw
− w
2
)
−
1
w
2
dw =
1/a
1
(−e
bw
÷1) dw =
1
b
(e
b
−e
b/a
) ÷
1
a
−1.
4.4.3 The integral I =
R
f (x, y) dx dy is
I =
1
0
a
0
2k
(x ÷y ÷1)
3
dx
dy =
1
0
x=a
x=0
−
k
(x ÷y ÷1)
2
dy
=
1
0
k
(y ÷1)
2
−
k
(y ÷a ÷1)
2
dy =
1
0
−
k
y ÷1
k
y ÷a ÷1
= −
k
2
÷
k
a ÷2
÷k −
k
a ÷1
=
k(a
2
÷3a)
2(a
2
÷3a ÷2)
The integral equals 1 if k = k
a
=
2(a
2
÷3a ÷2)
a
2
÷3a
= 2 ÷
4
a
2
÷3a
. Obviously, k
a
> 2 if a > 0.
4.4.5 The innermost integral is
1
0
(x
2
1
÷x
2
2
÷· · · ÷x
2
n
) dx
1
=
1
0
[
1
3
x
3
1
÷x
1
(x
2
2
÷x
2
3
÷· · · ÷x
2
n
)] =
1
3
÷x
2
2
÷x
2
3
÷· · · ÷x
2
n
Next,
1
0
(
1
3
÷x
2
2
÷x
2
3
÷· · · ÷x
2
n
) dx
2
=
1
0
[
1
3
x
2
÷
1
3
x
3
2
÷x
2
(x
2
3
÷x
2
4
÷· · · ÷x
2
n
)] =
2
3
÷x
2
3
÷x
2
4
÷· · · ÷x
2
n
etc. By induction, I = n/3.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 4 TOPI CS I N I NTEGRATI ON 41
4.5
4.5.2 If y ∈ [0, 1], then x runs from 0 to y, and if y ∈ [1, 2], then x must run from
√
y −1 to 1. Thus,
V =
1
0
y
0
xy
2
dx
dy ÷
2
1
1
√
y−1
xy
2
dx
dy =
1
0
y
4
2
dy ÷
2
1
y
2
−
y
3
2
dy =
67
120
4.5.6 [x −y[ = x −y if x ≥ y and y −x if x < y. Hence (see Fig. A4.5.6(a) in the book),
1
0
1
0
[x −y[ dx dy =
A
[x −y[ dx dy ÷
B
[x −y[ dx dy
=
1
0
¸
y
0
(y −x) dx ÷
1
y
(x −y) dx
¸
dy =
1
0
¸ x=y
x=0
yx −
1
2
x
2
÷
x=1
x=y
1
2
x
2
−yx
¸
dy
=
1
0
¸
y
2
−
1
2
y
2
÷
1
2
−y −
1
2
y
2
÷y
2
¸
dy =
1
0
y
2
−y ÷
1
2
dy =
1
0
1
3
y
3
−
1
2
y
2
÷
1
2
y
=
1
3
4.6
4.6.1 (a) Use the same subdivision as in Example 4.6.1, except that j = 0, . . . , 2n −1. Then
(2x
∗
i
−y
∗
j
÷1) x
i
y
j
=
2
i
n
−
j
n
÷1
1
n
1
n
= 2
i
n
3
−
j
n
3
÷
1
n
2
and
2n−1
¸
j=0
n−1
¸
i=0
2
i
n
3
−
j
n
3
÷
1
n
2
= 2
1
n
3
2n−1
¸
j=0
n−1
¸
i=0
i
−
1
n
3
n−1
¸
i=0
2n−1
¸
j=0
j
÷
1
n
2
2n−1
¸
j=0
n−1
¸
i=0
1
= 2
1
n
3
2n−1
¸
j=0
1
2
n(n −1) −
1
n
3
n−1
¸
i=0
1
2
(2n −1)2n ÷
1
n
2
2n−1
¸
j=0
n
= 2
1
n
3
1
2
n(n −1)2n −
1
n
3
1
2
(2n −1)2nn ÷
1
n
2
n2n = 2 −
1
n
→2
as n →∞.
4.7
y
x
P
1
P
2
P
3
Q
1
Q
2
Q
3
Figure M4.7.2
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
42 CHAPTER 4 TOPI CS I N I NTEGRATI ON
4.7.2 Assume for convenience that the points P
i
= (x
i
, y
i
), i = 1, 2, 3, all lie in the ﬁrst quadrant, and that
x
1
≤ x
2
≤ x
3
. Figure M4.7.2 shows the triangle P
1
P
2
P
3
together with the normals from the P
i
to the
corresponding points Q
i
on the xaxis. For each pair (i, j) with i < j, the points P
i
, Q
i
, Q
j
, and P
j
form
a quadrilateral with two parallel sides (called a trapezium in Britain, a trapezoid in the US), whose area is
T
ij
=
1
2
(x
j
−x
i
)(y
i
÷y
j
). If P
2
lies below the line P
1
P
3
, then the area of the triangle is T
13
−T
12
−T
23
,
and an easy computation shows that this equals A =
1
2
1 x
1
y
1
1 x
2
y
2
1 x
3
y
3
. If P
2
lies above P
1
P
3
, then the area
of the triangle is T
12
÷T
23
−T
13
= −A. In either case the area equals [A[.
If the x
i
are in a different order from what we assumed above, we can renumber them. That may
change the sign of the determinant but not its absolute value. Finally, if the triangle does not lie in the
ﬁrst quadrant, we can move it there by a parallel translation. Such a translation will not change the area
of the triangle, nor will it change the value of the determinant, since we are just adding multiples of the
ﬁrst column to the other two columns.
4.7.3 (b) In this problem it is convenient to use polar coordinates centred at (0, 1), so let x = r cos θ,
y = 1 ÷r sin θ. The Jacobian is ∂(x, y)/∂(r, θ) = r in this situation too, and
A
x
2
dx dy =
2π
0
1/2
0
r
3
cos
2
θ dr
dθ =
2π
0
r=1/2
r=0
r
4
4
cos
2
θ
dθ =
1
64
2π
0
cos
2
θ dθ =
π
64
4.7.5 (b) Introduce new variables u = y − 2x, v = 3x ÷ y. Then x = −
1
5
u ÷
1
5
v, y =
3
5
u ÷
2
5
v and
J =
∂(x, y)
∂(u, v)
= −
1
5
. It follows that
A
2
(x ÷y) dx dy =
8
4
1
−1
2
5
u ÷
3
5
v
[J[ du
dv =
8
4
1
−1
2
5
u ÷
3
5
v
1
5
du
dv = 144/25
4.8
4.8.1 (a) Use polar coordinates and let A
n
= {(x, y) : 1 ≤ x
2
÷y
2
≤ n
2
}. Then
A
n
(x
2
÷y
2
)
−3
dx dy =
2π
0
n
1
r
−6
r dr
dθ
∗
= 2π
n
1
r
−5
dr =
1
2
π(1−n
−4
) →
1
2
π as n →∞
About the equality
∗
=: The integral J
n
=
n
1
r
−6
r dr is independent of θ, therefore
2π
0
J
n
dθ = 2πJ
n
.
(b) With polar coordinates and with A
n
as in part (a), we get
I
n
=
A
n
(x
2
÷y
2
)
−p
dx dy = 2π
n
1
r
1−2p
dr =
⎧
⎨
⎩
2π
n
2−2p
−1
2 −2p
if p ,= 1
2π ln n if p = 1
If p > 1, then I
n
→π/(p −1) as n →∞, but if p ≤ 1, then I
n
→∞.
4.8.4 Note that the integrand takes both positive and negative values in the domain of integration. In the
calculations of the two iterated integrals the positive and the negative parts will more or less balance
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 5 DI F F ERENTI AL EQUATI ONS I : F I RST ORDER EQUATI ONS I N ONE VARI ABL E 43
each other, but not in exactly the same way. The two iterated integrals
d
1
b
1
y(y ÷ x)
−3
dx dy and
b
1
d
1
y(y ÷x)
−3
dy dx both tend to ∞as b and d tend to ∞, and so do the integrals of x(y ÷x)
−3
. The
trouble arises when we try to take the difference. That leads us into an ∞− ∞situation that does not
lead to any deﬁnite value.
4.8.6 (b) Introduce new variables u = x ÷ y, v = x − y. Then x =
1
2
(u ÷ v), y =
1
2
(u − v), and the
Jacobian determinant is ∂(x, y)/∂(u, v) = −1/2. The square B
/
n
= [−n, n] [−n, n] in the uvplane
corresponds to the square B
n
with corners (2n, 0), (0, 2n), (−2n, 0), and (0, −2n) in the xyplane,
and I
n
=
B
n
e
−(x−y)
2
1 ÷(x ÷y)
2
dx dy =
n
−n
n
−n
e
−v
2
1 ÷u
2
1
2
du
dv =
n
−n
e
−v
2
dv
n
−n
1
2
1
1 ÷u
2
du =
n
−n
e
−v
2
dv · arctan n. Let n → ∞. Then I
n
→
∞
−∞
e
−v
2
dv ·
π
2
=
√
π ·
π
2
=
π
3/2
2
. (Here we used
Poisson’s integral formula (4.3.3).)
4.8.7 (b) With polar coordinates and with the sets A
n
as in Example 3,
A
−ln(x
2
÷y
2
)
x
2
÷y
2
dx dy = lim
n→∞
π/2
0
1
1/n
−
ln r
2
r
r dr
dθ = −π lim
n→∞
1
1/n
ln r dr = · · · = π
Chapter 5 Differential Equations I: Firstorder Equations
in One Variable
5.1
5.1.6 The statement of the problemis slightly inaccurate. Instead of “for all t ” it should have said “for all t in
some open interval I around 0”. With that modiﬁcation the answer in the book is quite correct. (Actually,
the given differential equation has no solution deﬁned on the entire real line. One can show that, with the
initial condition x(0) = 0, the equation has the solution x = tan(
1
2
t
2
) over (−
√
π,
√
π), but this solution
cannot be extended to any larger interval because x(t ) runs off to inﬁnity as t approaches either endpoint.
5.2
x
1
t
1
x
2
÷y
2
= 4, x > 0
Figure M5.2.2
5.2.2 The solution curves are semicircles (not full circles) of the form t
2
÷ x
2
= C, x ,= 0, with C an
arbitrary positive constant. (This can be shown by direct differentiation, or by solving the separable
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
44 CHAPTER 5 DI F F ERENTI AL EQUATI ONS I : F I RST ORDER EQUATI ONS I N ONE VARI ABL E
equation ˙ x = −t /x using the method set out in the next section.) The integral curve through (0, 2) is
given by t
2
÷x
2
= 4, x > 0, in other words it is the graph of the function x(t ) =
√
4 −t
2
over the interval
(−2, 2). The lower semicircle shown in Fig. M5.2.2 is the graph of another function, x(t ) = −
√
4 −t
2
,
which is also a solution of the differential equation, but not the one you were asked to ﬁnd. (The ﬁgure
in answer in the book is misleading since it shows a full circle.)
5.3
5.3.3 (a) One constant solution, x ≡ 0. Otherwise, separating the variables,
dx
x
=
1 −t
t
dt =
1
t
−1
dt =⇒ ln [x[ = ln [t [ −t ÷C
1
Hence, [x[ = e
ln [t [−t ÷C
1
= e
ln [t [
e
−t
e
C
1
= C
2
[t [e
−t
, so x = Ct e
−t
, where C = ±C
2
= ±e
C
1
. The
integral curve through (t
0
, x
0
) = (1, 1/e) is x = t e
−t
.
(b) One constant solution, x ≡ 0. Otherwise,
˙ x
x
=
t
2
1 ÷t
3
=⇒ ln [x[ =
1
3
ln [1 ÷t
3
[ ÷C
1
=⇒ x = C
3
1 ÷t
3
Integral curve through (t
0
, x
0
) = (0, 2) for C = 2.
(c) No constant solutions.
x dx =
t dt =⇒
1
2
x
2
=
1
2
t
2
÷C
1
=⇒x
2
= t
2
÷C, where C = 2C
1
.
An integral curve through (t
0
, x
0
) = (
√
2, 1) must have C = −1 and x > 0, so x =
√
t
2
−1.
(d) The equation is equivalent to ˙ x = e
−2t
(x ÷ 1)
2
. One constant solution, x ≡ −1. The nonconstant
solutions are found by separating the variables:
dx
(x ÷1)
2
=
e
−2t
dt =⇒ −
1
x ÷1
= −
1
2
e
−2t
÷C
1
=⇒ x =
1 −e
−2t
÷C
1 ÷e
−2t
−C
where C = 1 ÷2C
1
. For the solution to pass through (t
0
, x
0
) we must have 1 −e
0
÷C = 0, i.e. C = 0,
so x =
1 −e
−2t
1 ÷e
−2t
.
5.3.6 For convenience we shall suppose that x > 0 and t > 0.
(a) The equation
t ˙ x
x
= a is separable, and we get
dx
x
=
a
t
dt =⇒ ln x = a ln t ÷C
1
= ln t
a
÷C
1
=⇒ x = e
ln x
= t
a
· e
C
1
= Ct
a
where C = e
C
1
.
(b) Here,
t ˙ x
x
= at ÷b. That gives
dx
x
=
at ÷b
t
dt =
a ÷
b
t
dt =⇒ ln x = at ÷b ln t ÷C
1
=⇒ x = Ce
at
t
b
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 5 DI F F ERENTI AL EQUATI ONS I : F I RST ORDER EQUATI ONS I N ONE VARI ABL E 45
(c) The equation
t ˙ x
x
= ax ÷b gives
˙ x
x(ax ÷b)
=
1
t
. Since
1
x(ax ÷b)
=
1
b
1
x
−
a
ax ÷b
, we get
1
x
−
a
ax ÷b
dx =
b
t
dt =⇒ ln x −ln [ax ÷b[ = b ln t ÷C
1
=⇒
x
ax ÷b
= Ct
b
where C = ±e
C
1
, and ﬁnally x =
Cbt
b
1 −Cat
b
.
5.3.8 (a) Let P = An
α
0
a
b
and Q = αv ÷ ε. Separating the variables we get
K
c−b
dK = P
e
Qt
dt .
Integration yields
K
1−b÷c
1 −b ÷c
=
P
Q
e
Qt
÷ C
1
. Hence, K =
¸
P
Q
(1 − b ÷ c)e
Qt
÷ C
¸
1/(1−b÷c)
, where
C = C
1
(1 −b ÷c).
(b) We separate the variables and get
x dx
(β −αx)(x −a)
=
dt . The hint in the problem gives
β dx
β −αx
÷
a dx
x −a
= (β −αa)
dt ⇐⇒ −
β
α
ln [β −αx[ ÷a ln [x −a[ = (β −αa)t ÷C
1
⇐⇒
β
α
ln [β −αx[ −a ln [x −a[ = ln [β −αx[
β/α
÷ln [x −a[
−a
= −(β −αa)t −C
1
⇐⇒ [β −αx[
β/α
[x −a[
−a
= e
−(β−αa)t −C
1
= e
−(β−αa)t
e
−C
1
= Ce
(αa−β)t
where C = e
−C
1
. We have the same answer as in the book (since [β −αx[ = [αx −β[).
5.3.9 (a) Note ﬁrst that L = L
0
e
λt
= [L
α
0
e
αλt
]
1/α
, so as t →∞,
K
L
=
¸
K
α
0
÷(sA/λ)L
α
0
(e
αλt
−1)
¸
1/α
[L
α
0
e
αλt
]
1/α
=
¸
K
α
0
L
α
0
e
αλt
÷
sA
λ
(1 −e
−αλt
)
¸
1/α
→
sA
λ
1/α
and
X
L
=
AK
1−α
L
α
L
= A
K
L
1−α
→A
sA
λ
(1−α)/α
= A
1/α
s
λ
(1−α)/α
(b) The equation is separable, dK/dt = sAK
1−α
L
α
= sAK
1−α
b
α
(t ÷a)
α
. We get
K
α−1
dK = sAb
α
(t ÷a)
pα
dt =⇒
1
α
K
α
=
sAb
α
pα ÷1
(t ÷a)
pα÷1
÷C
The initial condition gives C =
1
α
K
α
0
−
sAb
α
pα ÷1
a
pα÷1
, and so
K =
¸
K
α
0
÷sαAb
α
(t ÷a)
pα÷1
−a
pα÷1
/(pα ÷1)
¸
1/α
It follows that
K
L
=
¸
K
α
0
÷
sαAb
α
pα ÷1
(t ÷a)
pα÷1
−a
pα÷1
¸
1/α
¸
b
α
(t ÷α)
pα
¸
1/α
=
¸
K
α
0
b
α
(t ÷α)
pα
÷
sαA
pα ÷1
t ÷a −
a
pα÷1
(t ÷α)
pα
¸
1/α
→∞ as t →∞
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
46 CHAPTER 5 DI F F ERENTI AL EQUATI ONS I : F I RST ORDER EQUATI ONS I N ONE VARI ABL E
5.4
5.4.6 Use formula (5.4.6) to solve these equations.
(b) Here
a(t ) dt = −
(1/t ) dt = −ln t , and (5.4.6) yields the solution x = Ct ÷t
2
.
(c) In this case,
a(t ) dt = −
1
2
ln(t
2
−1), and (5.4.6) yields the solution x = C
√
t
2
−1 ÷t
2
−1.
(d) Here a(t ) = −2/t , b(t ) = −2a
2
/t
2
,
a(t ) dt = −2 ln t , which leads to x = Ct
2
÷2a
2
/3t .
5.4.9 From x = X/N, by logarithmic differentiation, ˙ x/x =
˙
X/X −
˙
N/N. Moreover, (ii) implies that
˙
X/X = a
˙
N/N, so ˙ x/x = (a−1)
˙
N/N = (a−1)[α−β(1/x)]. It follows that ˙ x = (a−1)αx −(a−1)β.
The solution is x(t ) = [x(0) − β/α]e
α(a−1)t
÷ β/α. Then (ii) and x = X/N together imply that
N(t ) = [x(t )/A]
1/(a−1)
, X(t ) = A[N(t )]
a
. If 0 < a < 1, then x(t ) → β/α, N(t ) → (β/αA)
1/(a−1)
,
and X(t ) →A(β/αA)
a/(a−1)
as t →∞.
5.4.10 (b) It sufﬁces to note that (1 −e
−ξt
)/ξ > 0 whenever ξ ,= 0 (look at ξ > 0 and ξ < 0 separately).
Then apply this with ξ = ασ − μ. Faster growth per capita is to be expected because foreign aid
contributes positively.
(c) Using equation (∗∗), we get x(t ) =
¸
x(0) ÷
σ
ασ −μ
H
0
N
0
¸
e
−(ρ−ασ)t
÷
σ
μ −ασ
H
0
N
0
e
(μ−ρ)t
.
Note that, even if ασ < ρ, x(t ) is positive and increasing for large t as long as μ > ρ. So foreign aid
must grow faster than the population.
5.5
5.5.2 With f (t, x) = 1 and g(t, x) = t /x ÷2, we have (g
/
t
−f
/
x
)/f = 1/x, so we are in Case (II). It follows
from (5.5.11) that we can choose β(x) = exp(
(1/x) dx) = exp(ln x) = x as an integrating factor.
Hence, x ÷(t ÷2x) ˙ x = 0 is exact, and (8) easily yields h(t, x) = t x ÷x
2
−t
0
x
0
−x
2
0
. The solution of
the differential equation is obtained from t x ÷ x
2
= C, where C is a constant. We assumed that t > 0
and x > 0, so C will be positive, and the solution of t x ÷x
2
= C is x = −
1
2
t ÷
1
4
t
2
÷C.
5.6
5.6.1 (a) With t > 0, the given equation is equivalent to the Bernoulli equation ˙ x ÷(2/t )x = x
r
with r = 2.
Let z = x
1−r
= x
−1
= 1/x, so that x = 1/z. Then ˙ x = −z
−2
˙ z and the differential equation becomes
−z
−2
˙ z ÷(2/t )z
−1
= z
−2
⇐⇒ ˙ z −(2/t )z = −1
whose solution is z = Ct
2
÷t . Thus x = (Ct
2
÷t )
−1
.
(b) The equation is a Bernoulli equation as in (5.6.1), with r = 1/2. Thus we substitute z = x
1−1/2
=
x
1/2
, i.e. x = z
2
. Then ˙ x = 2z˙ z and the given equation becomes
2z˙ z = 4z
2
÷2e
t
z ⇐⇒ ˙ z −2z = e
t
(Recall that x is assumed to be positive, and therefore z > 0.) Formula (5.4.4) yields
z = e
2t
(C ÷
e
−2t
e
t
dt ) = e
2t
(C ÷
e
−t
dt ) = e
2t
(C −e
−t
) = Ce
2t
−e
t
.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 5 DI F F ERENTI AL EQUATI ONS I : F I RST ORDER EQUATI ONS I N ONE VARI ABL E 47
(Alternatively we could go back to the method that was used to deduce (5.4.4) and calculate like this:
˙ z −2z = e
t
⇐⇒ (˙ z −2z)e
−2t
= e
−t
⇐⇒
d
dt
(ze
−2t
) = e
−t
⇐⇒ ze
−2t
= −e
−t
÷C, etc.)
The solution of the given equation is therefore
x = z
2
= (Ce
2t
−e
t
)
2
(c) As in part (a), we substitute x = 1/z, with ˙ x = −z
−2
dz. This leads to the differential equation
˙ z −(1/t )z = −(ln t )/t
Formula (5.4.6) yields the solution
z = e
ln t
C −
e
−ln t
ln t
t
dt
= t
C −
ln t
t
2
dt
= Ct ÷ln t ÷1 =⇒ x = (Ct ÷ln t ÷1)
−1
(The answer in the book also lists x ≡ 0 as a solution, and it certainly satisﬁes the equation, but the
problem explicitly calls for solutions with x > 0.)
5.6.3 Introducing z = K
1−b
as a new variable, we ﬁnd that (see (5.6.3)) ˙ z ÷ Pz = Qe
(av÷ε)t
, where
P = αδ(1 − b) and Q = αAn
a
0
(1 − b). According to (5.4.4), the solution of this linear differential
equation is
z = Ce
−Pt
÷Qe
−Pt
e
Pt
e
(av÷ε)t
dt = Ce
−Pt
÷Qe
−Pt
e
(av÷ε÷P)t
dt
= Ce
−Pt
÷Qe
−Pt
1
av ÷ε ÷P
e
(av÷ε÷P)t
= Ce
−Pt
÷
Qe
(av÷ε)t
av ÷ε ÷P
Insert the values of P and Q. Then K = z
1/(1−b)
gives the answer in the book.
5.6.4 Introduce z = K
1−α
as a new variable. By (5.6.3), we get the equation ˙ z −γ
2
(1 −α)z = γ
1
b(1 −α).
According to (5.4.3), the solution is z = Ce
γ
2
(1−α)t
− γ
1
b/γ
2
. Then K = z
1/(1−α)
gives the answer in
the book.
5.6.7 The equation is of the form ˙ x = g(x/t ) with g(z) = 1 ÷ z − z
2
. According to Problem 5.6.6,
the substitution z = x/t leads to the separable equation t ˙ z = g(z) − z = 1 − z
2
. This has the two
constant solutions z = −1 and z = 1. To ﬁnd the other solutions, we separate the variables and get
dz
1 −z
2
=
dt
t
. By a wellknown identity,
1
2
1
1 ÷z
÷
1
1 −z
dz =
dt
t
÷ C. Integration
yields
1
2
ln [1 ÷z[ −
1
2
ln [1 −z[ = ln [t [ ÷C, so
ln
1 ÷z
1 −z
= 2 ln [t [ ÷2C = ln t
2
÷2C =⇒
1 ÷z
1 −z
= At
2
where A = ±e
2C
. Solving for z gives z =
At
2
−1
At
2
÷1
, and ﬁnally x = t z =
At
3
−t
At
2
÷1
. In addition we have
the two solutions x = −t and x = t , corresponding to z = ±1.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
48 CHAPTER 5 DI F F ERENTI AL EQUATI ONS I : F I RST ORDER EQUATI ONS I N ONE VARI ABL E
5.7
5.7.3 (a) This is a separable equation with solution x(t ) = (1÷Ae
t
)/(1−Ae
t
) where A = (x
0
−1)/(x
0
÷1)
for x
0
,= −1. For x
0
= −1 we get the constant solution x(t ) ≡ −1. For x
0
,= 1, x(t ) →−1 as t →∞.
If x
0
> 1, which occurs when 0 < A < 1, then x(t ) → ∞ as t → (−ln A)
−
, and x(t ) → −∞ as
t →(−ln A)
÷
. See Fig. A5.7.3(a) in the answer section of the book for some integral curves.
5.7.4 (a) ∂k
∗
/∂s = f (k
∗
)/[λ − sf
/
(k
∗
)] > 0 and ∂k
∗
/∂λ = −k
∗
/[λ − sf
/
(k
∗
)] < 0 when λ > sf
/
(k
∗
).
In equilibrium, capital per worker increases as the savings rate increases, and decreases as the growth
rate of the work force increases. From F(K, L) = Lf (k) with k = K/L, we obtain F
/
K
(K, L) =
Lf
/
(k)(1/L) = f
/
(k).
(b) From equations (i) to (iv), c = (X −
˙
K)/L = (1 − s)X/L = (1 − s)f (k). But sf (k
∗
) = λk
∗
, so
when k = k
∗
we have c = f (k
∗
)−λk
∗
. The necessary ﬁrstorder condition for this to be maximized w.r.t.
k
∗
is that f
/
(k
∗
) = λ. But F(K, L) = Lf (k) and so F
/
K
= Lf
/
(k)dk/dK = f
/
(k) because k = K/L
with L ﬁxed. Thus ∂F/∂K = λ.
(c) See the answer in the book.
5.8
5.8.4 The equation is separable, and the solution through (t
0
, x
0
) = (0,
1
2
) is x = 1/(1 ÷e
−t
). If condition
(3) in Theorem 5.8.3 were satisﬁed, then for every t , there would exist numbers a(t ) and b(t ) such that
[x(1 − x)[ ≤ a(t )[x[ ÷ b(t ) for all x. But then [1 − x[ ≤ a(t ) ÷ b(t )/[x[, which clearly is impossible
when x is sufﬁciently large. Similarly, (4) implies x
2
(1−x) ≤ a(t )x
2
÷b(t ), so 1−x ≤ a(t ) ÷b(t )/x
2
,
which becomes impossible as x →−∞.
5.8.5 See Fig. M5.8.5. For t < a, ˙ ϕ(t ) = −2(t −a) = 2(a −t ) = 2
(a −t )
2
= 2
√
[ϕ(t )[. The argument
for t > b is similar. For t in (a, b) we have ˙ ϕ(t ) = 0 = 2
√
[ϕ(t )[. For t < a, (ϕ(t ) −ϕ(a))/(t −a) =
−(t − a)
2
/(t − a) = −(t − a) = a − t , and for t slightly larger than a, (ϕ(t ) − ϕ(a))/(t − a) = 0.
It follows that when t is near a,
(ϕ(t ) − ϕ(a))/(t − a)
≤ [t − a[, so ϕ is differentiable at a, and
˙ ϕ(a) = lim
t →a
(ϕ(t ) − ϕ(a))/(t − a) = 0 = 2
√
[ϕ(a)[. In the same way we show that the differential
equation is satisﬁed at t = b.
x
−3
−2
−1
1
2
3
t
−4 −3 −2 −1 1 2 3 4
x = ϕ(t )
Figure M5.8.5
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 6 / DI F F ERENTI AL EQUATI ONS I I : SECOND ORDER EQUATI ONS 49
6 Differential Equations II: SecondOrder Equations and
Systems in the Plane
6.1
6.1.4 In each of these three equations, let u = ˙ x. That will give simple ﬁrstorder equations for u which
you can solve, and afterwards ﬁnd x as x =
u dt .
(a) Putting u = ˙ x, we get ˙ u ÷ 2u = 8, which has the solution (see (5.4.3)) u = Ce
−2t
÷ 4. But then
x =
(Ce
−2t
÷4) dt = −
1
2
Ce
−2t
÷4t ÷B = Ae
−2t
÷4t ÷B, with A = −
1
2
C.
(b) With u = ˙ x, we get ˙ u −2u = 2e
2t
, with the solution (see (5.4.4))
u = Ce
2t
÷e
2t
e
−2t
2e
2t
dt = Ce
2t
÷e
2t
2 dt = Ce
2t
÷e
2t
2t = (C ÷2t )e
2t
Integration by parts then yields
x =
(C ÷2t )e
2t
dt =
1
2
(C ÷2t )e
2t
−
1
2
2e
2t
dt
=
1
2
(C ÷2t )e
2t
−
1
2
e
2t
÷B =
1
2
(C −1)e
2t
÷t e
2t
÷B = Ae
2t
÷t e
2t
÷B
with A =
1
2
(C −1).
(c) Let u = ˙ x. Then ˙ u − u = t
2
, which has the solution (see (5.4.4)) u = Ae
t
÷ e
t
e
−t
t
2
dt . Using
integration by parts twice gives
e
−t
t
2
dt = −t
2
e
−t
− 2t e
−t
− 2e
−t
, and so u = Ae
t
− t
2
− 2t − 2.
Then x =
(Ae
t
−t
2
−2t −2) dt = Ae
t
−
1
3
t
3
−t
2
−2t ÷B.
6.1.6 (a) Suppose x = ϕ(t ) is a solution of ´ x = F(x, ˙ s). We know that if ˙ x = ˙ ϕ(t ) ,= 0, then the equation
x −ϕ(t ) = 0 deﬁnes t as a function of x, with
dt
dx
= −
(∂/∂x)(x −ϕ(t ))
(∂/∂t )(x −ϕ(t ))
= −
∂
∂x
(x −ϕ(t ))
∂
∂t
(x −ϕ(t ))
=
1
˙ ϕ(t )
=
1
˙ x
Therefore ˙ x = 1/t
/
, where t
/
= dt /dx. (The prime
/
denotes differentiation with respect to x.) Now the
chain rule gives
´ x =
d
dt
( ˙ x) =
d
dt
1
t
/
=
d
dx
1
t
/
dx
dt
= −
t
//
(t
/
)
2
1
t
/
= −
t
//
(t
/
)
3
and the differential equation ´ x = F(x, ˙ x) becomes
−
t
//
(t
/
)
3
= F(x, 1/t
/
) ⇐⇒ t
//
= −(t
/
)
3
F(x, 1/t
/
)
(b) (i) Obviously, x(t ) ≡ C is a solution for any constant C. The equation for t
//
in (a) becomes
t
//
= −(t
/
)
3
(−x)(1/t
/3
) = x, or d
2
t /dx
2
= x. Integration yields dt /dx =
1
2
x
2
÷ A, and further
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
50 CHAPTER 6 / DI F F ERENTI AL EQUATI ONS I I : SECOND ORDER EQUATI ONS
integration results in t =
1
6
x
3
÷Ax ÷B, where A and B are arbitrary constants. A nonconstant solution
of the equation ´ x = −x ˙ x
3
is therefore given implicitly by the equation x
3
÷A
1
x ÷B
1
= 6t .
(ii) In this case, x(t ) ≡ C is a solution for every constant C ,= 0. For solutions with ˙ x ,= 0 we use the
transformation in (a) and get
t
//
= −(t
/
)
3
1/(t
/
)
2
x
= −
t
/
x
⇐⇒
t
//
t
/
= −
1
x
⇐⇒ ln [t
/
[ = −ln [x[ ÷A
1
This yields t
/
= A
2
/x, where A
2
= ±e
A
1
, and then t =
t
/
dx = A
2
ln [x[ ÷ B
1
. This yields ln [x[ =
At ÷B
2
, where A = 1/A
2
and B
2
= −B
1
/A
2
. Finally, x = ±e
B
2
e
At
= Be
At
.
(Note that the constants A and B here are different from 0. If we let A = 0 in Be
at
, we recover the
constant solutions mentioned at the beginning.)
6.2
6.2.3 (a) Direct calculations show that ´ u
1
÷ ˙ u
1
−6u
1
= 0 and ´ u
2
÷ ˙ u
2
−6u
2
= 0. Since u
1
and u
2
are not
proportional, Theorem 6.2.1(a) says that the general solution of ´ x ÷ ˙ x − 6x = 0 is x = Au
1
÷ Bu
2
=
Ae
2t
÷Be
−3t
.
(b) Theorem 6.2.1(b) now tells us that the general solution is Au
1
÷Bu
2
÷u
∗
, where u
∗
is any particular
solution of the equation. Since the righthand side of the equation is a polynomial of degree 1, we try to ﬁnd
a particular solution of the formu
∗
= Ct ÷D. We get ´ u
∗
÷˙ u
∗
−6u
∗
= C−6(Ct ÷D) = −6Ct ÷(C−6D),
so u
∗
is a solution if and only if −6C = 6 and C = 6D, in other words if and only if C = −1 and
D = −1/6. Thus the general solution of the given equation is x = Ae
2t
÷Be
−3t
−t −1/6.
6.2.5 Let x = (t ÷k)
−1
. Then ˙ x = −(t ÷k)
−2
, ´ x = 2(t ÷k)
−3
, and
(t ÷a)(t ÷b) ´ x ÷2(2t ÷a ÷b) ˙ x ÷2x
= (t ÷k)
−3
¸
2(t ÷2)(t ÷b) −2(2t ÷a ÷b)(t ÷k) ÷2(t ÷k)
2
¸
= (t ÷k)
−3
2
¸
k
2
−(a ÷b)k ÷ab
¸
= (t ÷k)
−3
2(k −a)(k −b)
Thus x = (t ÷k)
−1
solves the given differential equation if and only if k = a or k = b. Since a ,= b, the
functions u
1
= (t ÷a)
−1
and u
2
= (t ÷b)
−1
are not proportional, and the general solution of the given
equation is x = Au
1
÷Bu
2
= a(t ÷a)
−1
÷B(t ÷b)
−1
.
6.3
6.3.2 (a) The characteristic equation is r
2
− 1 = 0, with roots r = 1 and r = −1. The general solu
tion of the corresponding homogeneous differential equation is therefore (Case (I) in Theorem 6.3.1),
x = Ae
t
÷ Be
−t
. To ﬁnd a particular solution of the given equation we try u
∗
(t ) = p sin t ÷ q cos t .
Then ˙ u
∗
= p cos t − q sin t , and ´ u
∗
= −p sin t ÷ q cos t . Inserting this into the given equation yields
−2p sin t − 2q cos t = sin t . Thus p = −1/2 and q = 0. The general solution of ´ x − x = sin t is
therefore x = Ae
t
÷Be
−t
−
1
2
sin t .
(b) The general solution of the homogeneous equation is again x = Ae
t
÷Be
−t
. Since e
−t
is a solution
of the homogeneous equation, we try u
∗
(t ) = pt e
−t
. Then ˙ u
∗
(t ) = pe
−t
−pt e
−t
and ´ u
∗
(t ) = −pe
−t
−
pe
−t
÷ pt e
−t
, which inserted into the given equation yields −2pe
−t
= e
−t
, so p = −1/2, and the
general solution is x = Ae
t
÷Be
−t
−
1
2
t e
−t
.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 6 / DI F F ERENTI AL EQUATI ONS I I : SECOND ORDER EQUATI ONS 51
(c) The characteristic equation is r
2
−10r ÷25 = (r −5)
2
= 0, so r = 5 is a double root. The general
solution of the homogeneous equation is therefore (Case (II) in Theorem 6.3.1), x = Ae
5t
÷ Bt e
5t
. To
ﬁnd a particular solution we try u
∗
(t ) = pt ÷ q, and ﬁnd p = 2/75, q = 3/125. Hence, the general
solution of the given equation is x = Ae
5t
÷Bt e
5t
÷
2
75
t ÷
3
125
.
6.3.3 (a) The characteristic equation is r
2
÷2r ÷1 = (r ÷1)
2
= 0, so r = −1 is a double root. The general
solution of the homogeneous equation is therefore (Case (II) in Theorem 6.3.1), x = Ae
−t
÷Bt e
−t
. To
ﬁnd a particular solution we try u
∗
(t ) = Ct
2
÷Dt ÷E. Then ˙ u
∗
= 2Ct ÷D, and ´ u
∗
= 2C. Inserted into
the given equation this yields 2C÷4Ct ÷2D÷Ct
2
÷Dt ÷E = t
2
, or Ct
2
÷(4C÷D)t ÷2C÷2D÷E = t
2
.
Equating like powers of t yields C = 1, 4C ÷D = 0, and 2C ÷2D ÷E = 0. It follows that D = −4,
and E = 6. So the general solution is x = Ae
−t
÷ Bt e
−t
÷ t
2
− 4t ÷ 6. The constants A and B are
determined by the initial conditions. The condition x(0) = 0 yields A ÷ 6 = 0, so A = −6. Since
˙ x(t ) = −Ae
−t
÷ Be
−t
− Bt e
−t
÷ 2t − 4, the condition ˙ x(0) = 1 implies −A ÷ B − 4 = 1, and so
B = A ÷5 = −1. The required solution is x = −6e
t
−t e
t
÷t
2
−4t ÷6.
(b) The characteristic equation is r
2
÷4 = 0, so r = ±2i are the complex roots. The general solution of
the homogeneous equation is therefore (Case (III) in Theorem 6.3.1), x = Asin 2t ÷ B cos 2t . To ﬁnd
a particular solution we try u
∗
(t ) = Ct ÷ D, and ﬁnd C = 1 and D = 1/4. It follows that the general
solution is x = Asin 2t ÷B cos 2t ÷t ÷1/4. To ﬁnd the solution with the given initial conditions we must
determine A and B. The initial condition x(π/2) = 0 gives Asin π ÷B cos π ÷π/2 ÷1/4 = 0. Since
sin π = 0 and cos π = −1, we ﬁnd B = π/2 ÷1/4. Since ˙ x = 2Acos 2t −2B sin 2t ÷1, the equation
˙ x(π/2) = 0 gives −2A ÷1 = 0. Therefore A = 1/2, and so x =
1
2
sin 2t ÷(
1
4
÷
1
2
π) cos t ÷t ÷
1
4
.
6.3.4 Since the righthand side is a linear function of t , we try to ﬁnd a particular solution of the form
u
∗
= Pt ÷Q. We get ˙ u
∗
= P and ´ u
∗
= 0, so we must have
γ [β ÷α(1 −β)]P −γ δ
∗
Pt −γ δ
∗
Q = −γ δ
∗
kt −γ δ
∗
L
0
It follows that P = k and Q = L
0
÷[β ÷α(1 −β)]k/δ
∗
.
The characteristic equation is r
2
÷γ [β ÷α(1 −β)]r −γ δ
∗
= 0, with roots
r = −
1
2
γ [β ÷α(1 −β)] ±
1
2
γ
2
[β ÷α(1 −β)]
2
÷4γ δ
∗
Oscillations occur if the characteristic roots are complex, i.e. if and only if
1
4
γ
2
[β÷α(1−β)]
2
÷γ δ
∗
< 0.
6.3.8 (a) This is an Euler differential equation. With x = t
r
, we get ˙ x = rt
r−1
, and ´ x = r(r − 1)t
r−2
.
Inserting this into the given equation and cancelling t
r
gives the equation r
2
÷ 4r ÷ 3 = 0, with roots
r = −1 and r = −3. The general solution is therefore x = At
−1
÷Bt
−3
.
(b) Substituting x = t
r
into the homogeneous equation yields the equation r
2
−4r ÷3 = (r −1)(r −3).
The general solution of the homogeneous equation is therefore x = At ÷ Bt
3
. To ﬁnd a particular
solution of the nonhomogeneous equation, t
2
´ x − 3t ˙ x ÷ 3x = t
2
, we try u
∗
(t ) = Pt
2
÷ Qt ÷ R. Then
˙ u
∗
= 2Pt ÷Q, and ´ u
∗
= 2P. Inserted into the given equation this yields
2Pt
2
−3t (2Pt ÷Q) ÷3Pt
2
÷3Qt ÷3R = t
2
⇐⇒ −Pt
2
÷3R = t
2
This holds for all t if and only if P = −1 and R = 0. (Note that Qdid not appear in the last equation. That
is because Qt is a solution of the homogeneous equation.) One particular solution of the nonhomogeneous
equation is therefore u
∗
= −t
2
, and so the general solution is x = At ÷Bt
3
−t
2
.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
52 CHAPTER 6 / DI F F ERENTI AL EQUATI ONS I I : SECOND ORDER EQUATI ONS
6.3.10 By Leibniz’s rule, ´ p = a[D(p(t )) −S(p(t ))] = a(d
1
−s
1
)p ÷a(d
0
−s
0
). In other words, p must
satisfy ´ p ÷a(s
1
−d
1
)p = a(d
0
−s
0
). Note that a(s
1
−d
1
) > 0, and see the answer in the book.
6.4
6.4.3 Write the equation as ´ p(t ) − λp(t ) = k, where λ = γ (a − α). If λ > 0, then the solution is
p(t ) = Ae
rt
÷ Be
−rt
− k/r
2
, where r =
√
λ; if λ = 0, then the solution is p(t ) = At ÷ B ÷
1
2
kt
2
; if
λ < 0, then the solution is p(t ) = Acos
√
−λ t ÷B sin
√
−λ t −k/λ.
The equation is not stable for any values of the constants. This is obvious from the form of the
solutions—if λ ≥ 0, the corresponding homogeneous equation has solutions that run off to inﬁnity, and
if λ < 0 the solutions oscillate with a ﬁxed amplitude.
We can also see the instability fromthe criterion in (6.4.2): The characteristic equation is r
2
−λ = 0.
If λ > 0, this equation has two real solutions, one positive and one negative. If λ = 0, the characteristic
roots are r
1
= r
2
= 0. If λ < 0, the equation has complex roots with real part equal to 0.
6.5
6.5.2 (a) If we add the two equations we get ˙ x ÷ ˙ y = (a ÷b)(x ÷y). Hence x ÷y = Ce
(a÷b)t
. Because of
the initial conditions x(0) =
1
2
, y(0) =
1
2
, we must have C = 1. This implies ˙ x = a(x ÷y) = ae
(a÷b)t
and ˙ y = be
(a÷b)t
.
If a ÷b ,= 0, then
x =
ae
(a÷b)t
a ÷b
÷A, y =
be
(a÷b)t
a ÷b
÷B
for suitable constants A and B. The initial conditions yield
A = x(0) −
a
a ÷b
=
1
2
−
a
a ÷b
=
b −a
2(a ÷b)
, B = y(0) −
b
a ÷b
=
1
2
−
b
a ÷b
=
a −b
2(a ÷b)
and so
x(t ) =
2ae
(a÷b)t
÷b −a
2(a ÷b)
, y(t ) =
2be
(a÷b)t
÷a −b
2(a ÷b)
If a ÷b = 0, then ˙ x = a and ˙ y = b = −a, and therefore
x =
1
2
÷at, y =
1
2
−at
(b) The ﬁrst equation gives y = ˙ x −2t x, and if we use this in the second equation we get
´ x −2x −2t ˙ x = −2t −2x ⇐⇒ ´ x −2t ˙ x = −2t
This is a ﬁrstorder linear equation for ˙ x with general solution ˙ x = Ce
−t
2
÷ 1. From the ﬁrst equation
and the initial conditions we get ˙ x(0) = 0 ÷y(0) = 1, so C = 0. Therefore ˙ x = 1. Because x(0) = 1,
we get x = t ÷1, and y = ˙ x −2t x = 1 −2t (t ÷1) = −2t
2
−2t ÷1.
(c) The ﬁrst equation yields y = −
1
2
˙ x ÷
1
2
sin t . From the second equation we then get
−
1
2
´ x ÷
1
2
cos t = 2x ÷1 −cos t ⇐⇒ ´ x ÷4x = −2 ÷3 cos t (∗)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 6 / DI F F ERENTI AL EQUATI ONS I I : SECOND ORDER EQUATI ONS 53
The corresponding homogeneous equation, ´ x ÷4x = 0, has the general solution x = Acos 2t ÷B sin 2t .
To ﬁnd a particular solution of (∗) we try u
∗
= C ÷ Dcos t ÷ E sin t . Then ´ u
∗
= −Dcos t − E sin t
and ´ u
∗
÷ 4u
∗
= 4C ÷ 3Dcos t ÷ 3E sin t . It follows that C = −1/2, D = 1, and E = 0. Thus
the general solution of (∗) is x = Acos 2t ÷ B sin 2t −
1
2
÷ cos t , and we get y = −
1
2
˙ x ÷
1
2
sin t =
−Acos 2t ÷B sin 2t ÷sin t . The initial conditions x(0) = y(0) = 0 yield B −
1
2
÷1 = 0 and −A = 0,
so the solutions we are looking for are
x = −
1
2
cos 2t ÷cos t −
1
2
, y = −
1
2
sin 2t ÷sin t
6.5.3 The ﬁrst equation gives p = e
−2t
( ˙ x−x). Then ˙ p = −2e
−2t
( ˙ x−x)÷e
−2t
( ´ x−˙ x) = e
−2t
( ´ x−3˙ x÷2x).
If we insert these expressions for p and ˙ p into the second equation, we get
e
−2t
( ´ x −3˙ x ÷2x) = 2e
−2t
x −e
−2t
( ˙ x −x) = e
−2t
(3x − ˙ x) =⇒ ´ x −2˙ x −x = 0
The general solution of the last equation is x = Ae
(1÷
√
2 )t
÷Be
(1−
√
2 )t
, where 1±
√
2 are the roots of the
characteristic equation, r
2
−2r −1 = 0. A straightforward calculation gives ˙ x −x = A
√
2e
(1÷
√
2 )t
−
B
√
2e
(1−
√
2 )t
and then p = e
−2t
( ˙ x −x) = A
√
2e
(
√
2−1)t
−B
√
2e
(−
√
2−1)t
.
6.5.4 From the ﬁrst equation, σ = απ − ˙ π. Inserting this into the second equation, we get
α ˙ π − ´ π = π −
α
β
π ÷
1
β
˙ π ⇐⇒ ´ π ÷
1
β
−α
˙ π ÷
1 −
α
β
π = 0 (∗)
which is a secondorder differential equation for π with constant coefﬁcients. The characteristic equation
is r
2
÷(1/β −α)r ÷(1−α/β) = 0, with roots r
1,2
=
1
2
(α−1/β) ±
1
2
(α ÷1/β)
2
−4. If α÷1/β > 2,
then r
1
and r
2
are real and different, and the general solution of (∗) is π = Ae
r
1
t
÷Be
r
2
t
. It follows that
σ = απ − ˙ π = (α −r
1
)Ae
r
1
t
÷(α −r
2
)Be
r
2
t
.
6.6
6.6.3 (ii) The equilibrium point of the linear system ˙ x = x ÷ 2y, ˙ y = −y, is (x
∗
, y
∗
) = (−5, 2). Let
z = x ÷5 and w = y −2 (cf. example 6.5.4). The given equation system is equivalent to the system
˙ w = w ÷2z
˙ z = −z
⇐⇒
˙ w
˙ z
= A
w
z
(∗)
with coefﬁcient matrix A =
1 2
0 −1
. Since the trace of Ais 0, the systemis not globally asymptotically
stable. The eigenvalues of Aare λ
1
= 1 and λ
2
= −1, with corresponding eigenvectors
1
0
and
1
−1
.
According to (6.5.9), the solution of (∗) is
w
z
= Ae
t
1
0
÷Be
−t
1
−1
=
Ae
t
÷Be
−t
−Be
−t
. The
solution of the given system is therefore x = z − 5 = Ae
t
÷ Be
−t
− 5 and y = w ÷ 2 = −Be
−t
÷ 2,
the same solution as in (i).
It is clear that if A ,= 0, then x does not converge to the equilibriumvalue as t →∞, which conﬁrms
that the system is not asymptotically stable.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
54 CHAPTER 6 / DI F F ERENTI AL EQUATI ONS I I : SECOND ORDER EQUATI ONS
6.6.4 (a) From the ﬁrst equation, y =
1
2
( ˙ x −ax −α). The second equation then yields
1
2
( ´ x −a ˙ x) = 2x ÷
1
2
a( ˙ x −ax −α) ÷β ⇐⇒ ´ x −2a ˙ x ÷(a
2
−4)x = 2β −αa (♦)
The characteristic equation is r
2
−2ar ÷a
2
−4 = 0, which has the roots r
1,2
= a ±2. It follows that the
homogeneous equation associated with (♦) has the general solution x
H
= Ae
(a−2)t
÷ Be
(a÷2)t
. Hence,
the general solution of equation (♦) itself is x
H
÷u
∗
, where u
∗
is any particular solution of (♦). Since the
righthand side of the equation is a constant, we look for a suitable constant u
∗
, which turns out to be u
∗
=
(2β−αa)/(a
2
−4). Thus, the general solution of (♦) is x = Ae
(a−2)t
÷Be
(a÷2)t
÷(2β −αa)/(a
2
−4).
Then the equation y =
1
2
( ˙ x −ax −α) yields y = −Ae
(a−2)t
÷Be
(a÷2)t
÷(2α −aβ)/(a
2
−4).
(b) The equilibrium point (x
∗
, y
∗
) is given by the equation system
ax
∗
÷2y
∗
÷α = 0, 2x
∗
÷ay
∗
÷β = 0
Easy calculations show that x
∗
= (2β − αa)/(a
2
− 4) and y
∗
= (2α − aβ)/(a
2
− 4). Of course, this
is just the stationary solution of the system in part (a), given by A = B = 0. The equilibrium point is
globally asymptotically stable if and only if e
(a−2)t
and e
(a÷2)t
both tend to 0 as t →∞, and this happens
if and only if both a −2 and a ÷2 are negative, i.e. if and only if a < −2.
An alternative way to check stability is to consider the trace and determinant of the coefﬁcient matrix
A =
a 2
2 a
. We get tr(A) = 2a and [A[ = a
2
−4, and Theorem 6.6.1 says that the equilibrium point
is globally asymptotically stable if and only if 2a < 0 and a
2
−4 > 0, which is equivalent to a < −2.
(c) With a = −1, α = −4, and β = −1, the general solution of the system is x = Ae
−3t
÷ Be
t
÷ 2,
y = −Ae
−3t
÷ Be
t
÷ 3. It is clear that this will converge to the equilibrium point (2, 3) if and only if
B = 0. One such solution is then x = e
−3t
÷ 2, y = −e
−3t
÷ 3. This solution moves towards (2, 3)
along the line x ÷y = 5, and so do all the convergent solutions, i.e. the solutions with B = 0.
6.7
6.7.4 (a) See Figs. A6.7.4(a) and (b) in the answer section of the book. The system has a single equilibrium
point, namely (0, 0). It seems clear from the phase diagram that this is not a stable equilibrium. Indeed,
the equations show that any solution through a point on the yaxis below the origin will move straight
downwards, away from (0, 0).
(b) The ﬁrst equation has the general solution x = Ae
−t
, and the initial condition x(0) = −1 implies
A = −1, so x = −e
−t
. The second equation is the system the becomes ˙ y = e
−t
y − y
2
, which is a
Bernoulli equation.
With z = y
1−2
= y
−1
= 1/y we get y = 1/z, ˙ y = −˙ z/z
2
, and
−˙ z/z
2
= e
−t
(1/z) −1/z
2
=⇒ ˙ z ÷e
−t
z = 1
If we use formula (5.4.7) with a(t ) = e
−t
, b(t ) = 1, t
0
= 1, and z
0
= 1, we get −
t
s
a(ξ) dξ = e
−t
−e
−s
and
z = e
e
−t
−1
÷
t
0
e
e
−t
−e
−s
ds = e
e
−t
e
−1
÷
t
0
e
−e
−s
ds
y =
e
−e
−t
e
−1
÷
t
0
e
−e
−s
ds
For all s ≥ 0 we have e
−s
≤ 1, so −e
−s
≥ −1 and e
−e
−s
≥ e
−1
. It follows that for t > 0 we have
t
0
e
−e
−s
ds ≥ e
−1
t , e
−e
−t
≤ e
0
= 1 and y(t ) ≤ 1/(e
−1
÷t e
−1
) →0 as t ∞.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 6 / DI F F ERENTI AL EQUATI ONS I I : SECOND ORDER EQUATI ONS 55
6.7.6 The equation ˙ x = −x has the general solution x = Ae
−t
. The initial condition x(0) = 1 then implies
x(t ) = e
−t
. The second equation becomes ˙ y = −e
−2t
y. This is a separable equation and we get
dy
y
= −
e
−2t
dt =⇒ ln [y[ =
1
2
e
−2t
÷C
The initial condition y(0) = 1 yields C = −
1
2
, and we get y(t ) = e
(e
−2t
−1)/2
. As t → ∞, the point
(x(t ), y(t )) tends to (0, e
−1/2
).
6.8
6.8.5 Let f (Y, K) = α
I (Y, K) − S(Y, K)
and g(Y, K) = I (Y, K). The matrix A in Theorem 6.8.2 is
A(Y, K) =
f
/
Y
f
/
K
g
/
Y
g
/
K
=
α(I
/
Y
−S
/
Y
) α(I
/
K
−S
/
K
)
I
/
Y
I
/
K
. Now, tr
A(Y, K)
= α(I
/
Y
−S
/
Y
) ÷I
/
K
< 0
and [A(Y, K)[ = α(I
/
Y
− S
/
Y
)I
/
K
− α(I
/
K
− S
/
K
)I
/
Y
= α(I
/
Y
S
/
K
− I
/
K
S
/
Y
) > 0 by the assumptions in the
problem. Finally, f
/
Y
g
/
K
= α(I
/
K
−S
/
K
)I
/
Y
< 0 everywhere, so an equilibrium point for the system must
be globally asymptotically stable according to Olech’s theorem.
6.8.6 See the answer in the book for the stability question.
K must satisfy the Bernoulli equation
˙
K = sK
α
−δK. This corresponds to the equation in Problem
5.6.4, with γ
1
b = s and γ
2
= −δ. The general solution is therefore K(t ) =
¸
Ce
−δ(1−α)t
÷ s/δ
¸
1/(1−α)
.
The initial condition K(0) = K
0
yields C = K
1−α
0
− s/δ. Since δ(1 − α) > 0, the solution K(t ) =
¸
(K
1−α
0
−s/δ)e
−δ(1−α)t
÷s/δ
¸
1/(1−α)
tends to (s/δ)
1/(1−α)
= K
∗
as t →∞.
6.9
6.9.2 Write the system as
˙
k = F(k, c) = f (k) −δk −c, ˙ c = G(k, c) = −c(r ÷δ −f
/
(k))
The Jacobian matrix at (k, c) is
A(k, c) =
F
/
k
F
/
c
G
/
k
G
/
c
=
f
/
(k) −δ −1
cf
//
(k) f
/
(k) −r −δ
The origin (k, c) = (0, 0) is an equilibrium point, but not the one the authors had in mind. At an
equilibrium point (k
∗
, c
∗
) with c
∗
> 0 we have c
∗
= f (k
∗
) −δk
∗
and f
/
(k
∗
) = r ÷δ. Then
A(k
∗
, c
∗
) =
f
/
(k
∗
) −δ −1
cf
//
(k
∗
) 0
=
r −1
cf
//
(k
∗
) 0
has determinant [A(k
∗
, c
∗
)[ = cf
//
(k
∗
) < 0, so (k
∗
, c
∗
) is a saddle point.
6.9.3 (a) (x
0
, y
0
) = (4/3, 8/3). It is a saddle point because the Jacobian at (4/3, 8/3) is
A =
y −x −2 x
y
2
/2x
2
1 −y/x
=
−2/3 4/3
2 −1
, with determinant [A[ = −2.
(b) See Fig. A6.9.3 in the answer section of the book.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
56 7 DI F F ERENTI AL EQUATI ONS I I I : HI GHER ORDER EQUATI ONS
6.9.4 (a) The equilibrium points are the solutions of the equation system
(i) y
2
−x = 0, (ii) 25/4 −y
2
−(x −1/4)
2
= 0
Substituting x for y
2
in (ii) leads to the quadratic equation x
2
÷ x/2 − 99/16 = 0, which has the
solutions x = (−1 ± 10)/4. Since x = y
2
must be nonnegative, we get x = 9/4 and y = ±3/2.
The Jacobian at (x, y) is A(x, y) =
−1 2y
−2x ÷1/2, −2y
, so A(9/4, 3/2) = A
1
=
−1 3
−4 −3
and
A(9/4, −3/2) = A
2
=
−1 −3
−4 3
. The determinants and traces of these matrices are [A
1
[ = 15,
tr(A
1
) = −4, [A
2
[ = −15, and tr(A
2
) = 2. It follows that (9/4, 3/2) is locally asymptotically stable,
whereas (9/4, −3/2) is a saddle point. These conclusions are conﬁrmed by the solution to part (b). See
Fig. A6.9.4 in the answer section of the book.
7 Differential Equations III: HigherOrder Equations
7.1
7.1.3 The homogeneous equation ´ x ÷ x = 0 has the two linearly independent solutions u
1
= sin t and
u
2
= cos t . To ﬁnd the solution of the equation ´ x ÷x = 1/t we use the method of variation of parameters.
Let x = C
1
(t ) sin t ÷C
2
(t ) cos t . The two equations to determine
˙
C
1
(t ) and
˙
C
2
(t ) are
˙
C
1
(t ) sin t ÷
˙
C
2
(t ) cos t = 0
˙
C
1
(t ) cos t −
˙
C
2
(t ) sin t = 1/t
This is a linear equation system in the unknowns
˙
C
1
(t ) and
˙
C
2
(t ), and the determinant of the system is
sin t cos t
cos t −sin t
= −sin
2
t −cos
2
t = −1. (See Section B.1.) Cramer’s rule gives
˙
C
1
(t ) =
cos t
t
,
˙
C
2
(t ) = −
sin t
t
It follows that u
∗
(t ) = sin t
cos t
t
dt − cos t
cos t
t
dt is the general solution of the given equation
(provided we include an arbitrary constant of integration in each of the two integrals).
7.2
7.2.2 Integer roots of the characteristic equation r
3
−r
2
−r ÷1 = 0 must divide the constant term1 (EMEA,
Note 4.7.2), so the only possibilities are ±1, and we see that both r = −1 and r = 1 are roots. Moreover,
(r
3
−r
2
−r ÷1)(r −1) = (r
2
−1), so r
3
−r
2
−r ÷1 = (r −1)(r
2
−1) = (r −1)
2
(r ÷1). According to the
general method for ﬁnding linearly independent solutions to linear homogeneous equations with constant
coefﬁcients, the general solution of the associated homogeneous equation is x
H
= (A ÷ Bt )e
t
÷ Ce
−t
.
Looking at the righthand side of the given nonhomogeneous equation, it might seem natural to try
u
∗
= (D ÷ Et )e
−t
as a particular solution. But that does not work because r = −1 is a root in
the characteristic equation. We must therefore increase the degree of the polynomial factor and try
with a quadratic polynomial instead. Let u
∗
= (Et ÷ Ft
2
)e
−t
. A bit of tedious algebra gives ˙ u
∗
=
(E÷(2F−E)t −Ft
2
)e
−t
, ´ u
∗
= (2F−2E÷(E−4F)t ÷Ft
2
)e
−t
,
...
u
∗
= (3E−6F÷(6F−E)t −Ft
2
)e
−t
,
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
7 DI F F ERENTI AL EQUATI ONS I I I : HI GHER ORDER EQUATI ONS 57
and ﬁnally
...
u
∗
− ´ u
∗
− ˙ u
∗
÷u
∗
= (4E −8F ÷8Ft )e
−t
. This equals 8t e
−t
if F = 1 and E = 2, so the
general solution of the given equation is x = x
H
÷u
∗
= (A÷Bt )e
t
÷(C ÷2t ÷t
2
)e
−t
. Requiring this
solution to satisfy the initial conditions gives A = −1, B = 1, C = 1, and so x = (t −1)e
t
÷(t ÷1)
2
e
−t
.
Why did we not include a termof the formDe
−t
in the tentative solution u
∗
? The answer is that De
−t
is a solution of the homogeneous equation for every value of D, so Dwould not appear in
...
u
∗
−´ u
∗
−˙ u
∗
÷u
∗
.
7.2.3 Differentiating the equation w.r.t. t gives (using the product rule and (4.1.5))
...
K = (γ
1
κ ÷γ
2
)
´
K ÷(γ
1
σ ÷γ
3
)μ
0
μe
μt
t
0
e
−μτ
˙
K(τ)dτ ÷(γ
1
σ ÷γ
3
)μ
0
e
μt
e
−μt
˙
K(t ) (∗)
From the given equation, (γ
1
σ ÷γ
3
)μ
0
e
μt
t
0
e
−μτ
˙
K(τ)dτ =
´
K −(γ
1
κ ÷γ
2
)
˙
K. Inserting this into (∗)
yields the equation
...
K − p
´
K ÷ q
˙
K = 0 given in the answer in the book. One root of the characteristic
equation r
3
− pr
2
÷ qr = 0 is r
3
= 0. The other two are the roots r
1
and r
2
of r
2
− pr ÷ q = 0,
and they are real, nonzero, and different if p
2
− 4q > 0 and q ,= 0. If these conditions are satisﬁed, it
follows from the theory in this section that the general solution of the differential equation is of the form
K(t ) = C
1
e
r
1
t
÷C
2
e
r
2
t
÷C
3
e
r
3
t
.
7.3
7.3.2 The roots of the characteristic equation r
3
÷4r
2
÷5r ÷2 = 0 are r
1
= r
2
= −1 and r
3
= −2, which
are all negative, so global asymptotic stability also follows from Theorem 7.3.1.
7.4
7.4.1 (i) The systemis: (a) ˙ x
1
= −x
1
÷x
2
÷x
3
, (b) ˙ x
2
= x
1
−x
2
÷x
3
, (c) ˙ x
3
= x
1
÷x
2
÷x
3
. Differentiating
(a) w.r.t. t and inserting from (b) and (c) gives (d) ´ x
1
÷ ˙ x
1
−2x
1
= 2x
3
. Differentiating once more w.r.t.
t and inserting from (c) gives
...
x
1
÷ ´ x
1
− 2˙ x
1
= 2˙ x
3
= 2x
1
÷ 2(x
2
÷ x
3
) = 2x
1
÷ 2( ˙ x
1
÷ x
1
), using
(a) again. Thus the differential equation for x
1
is (e)
...
x
1
÷ ´ x
1
−4˙ x
1
−4x
1
= 0. Since the characteristic
polynomial is (r ÷ 1)(r ÷ 2)(r − 2), the general solution is x
1
= C
1
e
−t
÷ C
2
e
−2t
÷ C
3
e
2t
. From (d)
we ﬁnd x
3
=
1
2
( ´ x
1
÷ ˙ x
1
− 2x
1
) = −C
1
e
−t
÷ 2C
3
e
2t
. We then ﬁnd x
2
from (a): x
2
= ˙ x
1
÷ x
1
− x
3
=
C
1
e
−t
−C
2
e
−2t
÷C
3
e
2t
.
(ii) We write the system as ˙ x = Ax, where A =
⎛
⎝
−1 1 1
1 −1 1
1 1 1
⎞
⎠
and x =
⎛
⎝
x
1
x
2
x
3
⎞
⎠
. The eigenvalues of
A are the solutions of the equation
−1 −λ 1 1
1 −1 −λ 1
1 1 1 −λ
= 0, and we ﬁnd them to be λ
1
= −1,
λ
2
= −2, and λ
3
= 2. (These are the same as the solutions of the characteristic equation of the differential
equation (e). This is no coincidence. See the remark above Theorem 6.6.1 concerning secondorder
systems.) The eigenvectors associated with the eigenvalues are determined by the three systems
x
2
÷ x
3
= 0
x
1
÷ x
3
= 0
x
1
÷x
2
÷2x
3
= 0
,
x
1
÷x
2
÷ x
3
= 0
x
1
÷x
2
÷ x
3
= 0
x
1
÷x
2
÷3x
3
= 0
,
−3x
1
÷ x
2
÷x
3
= 0
x
1
−3x
2
÷x
3
= 0
x
1
÷ x
2
−x
3
= 0
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
58 7 DI F F ERENTI AL EQUATI ONS I I I : HI GHER ORDER EQUATI ONS
The following vectors are solutions of these systems:
v
1
=
⎛
⎝
1
1
−1
⎞
⎠
, v
2
=
⎛
⎝
1
−1
0
⎞
⎠
, v
3
=
⎛
⎝
1
1
2
⎞
⎠
The solution of the given system of differential equations is then
⎛
⎝
x
1
x
2
x
3
⎞
⎠
= C
1
e
−t
⎛
⎝
1
1
−1
⎞
⎠
÷C
2
e
−2t
⎛
⎝
1
−1
0
⎞
⎠
÷C
3
e
2t
⎛
⎝
1
1
2
⎞
⎠
We see that we have arrived at the same solutions as above.
(iii) The resolvent, with t
0
= 0, is
P(t, 0) =
⎛
⎜
⎝
1
3
e
−t
÷
1
2
e
−2t
÷
1
6
e
2t 1
3
e
−t
−
1
2
e
−2t
÷
1
6
e
2t
−
1
3
e
−t
÷
1
3
e
2t
1
3
e
−t
−
1
2
e
−2t
÷
1
6
e
2t 1
3
e
−t
÷
1
2
e
−2t
÷
1
6
e
2t
−
1
3
e
−t
÷
1
3
e
2t
−
1
3
e
−t
÷
1
3
e
2t
−
1
3
e
−t
÷
1
3
e
2t 1
3
e
−t
÷
2
3
e
2t
⎞
⎟
⎠
The ith column of P(t, 0) is a vector (x
1
(t ), x
2
(t ), x
3
(t ))
/
of particular solutions of the system such that
(x
1
(0), x
2
(0), x
3
(0))
/
is the ith standard unit vector e
i
. In particular, P(0, 0) = I
3
. It is not hard to verify
that AP(t, 0) = (d/dt )P(t, 0).
7.5
7.5.4 (a) Equation (∗) is separable, so
g
1
(x)
f
1
(x)
dx =
f
2
(y)
g
2
(y)
dy ÷ C for some constant C. Therefore,
the function H(x, y) is constant along each solution curve for the system.
(b) H(x, y) =
bx −h
x
dx −
k −ay
y
dy =
b −
h
x
dx −
k
y
−a
dy =
bx −hln x −(k ln y −ay) ÷C = b(x −x
0
ln x) ÷a(y −y
0
ln y) ÷C,
where x
0
= h/b and y
0
= k/a.
7.5.5 We can write the system as
˙ x = x(k −ay −εx), ˙ y = y(−h ÷bx −δy)
It is clear that a point (x
0
, y
0
) with x
0
,= 0 and y
0
,= 0 is an equilibrium point if and only if
bx
0
−δy
0
= h
εx
0
÷ay
0
= k
(∗)
The determinant of this system is ab ÷ δε ,= 0, so it has a unique solution, and Cramer’s rule gives the
solution as x
0
= (ah ÷kδ)/(ab ÷δε), y
0
= (bk −hε)/(ab ÷δε), as given in the problem.
We claim that the function
L(x, y) = b(x −x
0
ln x) ÷a(y −y
0
ln y) −b(x
0
−x
0
ln x
0
) ÷a(y
0
−y
0
ln y
0
)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
7 DI F F ERENTI AL EQUATI ONS I I I : HI GHER ORDER EQUATI ONS 59
is a strongLiapunovfunctionfor the system, with(x
0
, y
0
) as its minimumpoint. First note that L(x
0
, y
0
) =
0. Moreover, L
/
x
(x, y) = b(1 − x
0
/x) and L
/
y
(x, y) = a(1 − y
0
/y) are both 0 at (x
0
, y
0
) and L
//
xx
=
bx
0
/x
2
> 0, L
//
yy
= ay
0
/y
2
> 0, and L
//
xy
= 0, so L(x, y) is strictly convex and has a unique minimum
at (x
0
, y
0
). Finally,
˙
L = b
1 −x
0
/x
˙ x ÷a
1 −y
0
/y
˙ y = b(k −ay −εx)(x −x
0
) ÷a(−h ÷bx −δy)(y −y
0
)
= b(εx
0
÷ay
0
−ay −εx)(x −x
0
) ÷a(δy
0
−bx
0
÷bx −δy)(y −y
0
)
= −εb(x −x
0
)
2
−aδ(y −y
0
)
2
which is negative for (x, y) ,= (x
0
, y
0
). We used the equations k = εx
0
÷ ay
0
and h = −δy
0
÷ bx
0
from (∗). We have proved that L is a strong Liapunov function for the system, so (x
0
, y
0
) is locally
asymptotically stable.
7.7
7.7.1 (a) By integration, z =
1
4
x
4
÷
1
2
x
2
y
2
−e
x
y ÷ϕ(y), where ϕ(y) is an arbitrary function and plays the
role of a constant of integration when we integrate with respect to x.
(b) The recipe for solving equations of the form (7.7.2) leads to the solution z = 3x ÷ϕ(y −2x), where
ϕ is an arbitrary differentiable function. It looks as if x has a special role here, but that is an illusion.
If we use the recipe with y instead of x as the independent variable in equations (7.7.3), we are led to
z = 3y/2 ÷ψ(x −y/2), where y seems to be singled out. Actually, these solutions are just two ways of
writing the same thing. The functions ϕ and ψ are related by the equation ψ(u) = ϕ(−2u) ÷3u.
(c) The equations in (7.7.3) are both separable, dy/dx = y
2
/x
2
and dz/dx = z
2
/x
2
. The solutions are
−1/y = −1/x ÷C
1
, −1/z = −1/x ÷C
2
. The general solution is therefore (1/x −1/y, 1/x −1/z)
= 0, or 1/z = 1/x − ϕ(1/x − 1/y), and hence z =
x
1 −xϕ(1/x −1/y)
, where ϕ is an arbitrary
differentiable function.
7.7.3 (a) The equations in (7.7.3) are dy/dx = −y/x and dz/dx = 1. The latter equation gives z = x÷C
1
,
and so z −x = C
1
. The ﬁrst equation is separable with solution xy = C
2
. The general solution of (∗) is
given by (z −x, xy) = 0. Solving this equation for the ﬁrst variable yields z = x ÷ϕ(xy), where ϕ is
an arbitrary differentiable function.
(b) The condition f (x, 1) = x
2
implies that x ÷ϕ(x) = x
2
. Thus ϕ(x) = −x ÷x
2
for all x, and hence
f (x, y) = x ÷ϕ(xy) = x −xy ÷x
2
y
2
.
7.7.7 The equations in (7.7.3) are dv
2
/dv
1
= v
2
/v
1
and dx/dv
1
= xε(x)/v
1
, with the solutions v
2
/v
1
= C
1
,
f (x) −ln v
1
= C
2
, where f (x) =
(1/xε(x)) dx. Since f
/
(x) = 1/xε(x) > 0, f is strictly increasing,
and has an inverse f
−1
that is also strictly increasing. The general solution is (v
2
/v
1
, f (x)−ln v
1
) = 0,
or f (x) = ln v
1
÷ ϕ(v
2
/v
1
). Hence, x = f
−1
(ln v
1
÷ ϕ(v
2
/v
1
)). Deﬁne g(v
1
, v
2
) = e
ln v
1
÷ϕ(v
2
/v
1
)
=
v
1
e
ϕ(v
2
/v
1
)
. Then g is homogeneous of degree 1, and we see that x = f
−1
(ln(g(v
1
, v
2
)). The composition
F of the two increasing functions f
−1
and ln is increasing. It follows that x = F(g(v
1
, v
2
)) is homothetic.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
60 8 CAL CUL US OF VARI ATI ONS
8 Calculus of Variations
8.2
8.2.3 (a) With F(t, x, ˙ x) = x
2
÷ ˙ x
2
÷ 2xe
t
we get F
/
x
= 2x ÷ 2e
t
and F
/
˙ x
= 2˙ x, and the Euler equation
becomes
2x ÷2e
t
−
d
dt
(2˙ x) = 0 ⇐⇒ 2x ÷2e
t
−2´ x = 0 ⇐⇒ ´ x −x = e
t
(b) With F(t, x, ˙ x) = −e
˙ x−ax
we get F
/
x
= ae
˙ x−ax
and F
/
˙ x
= −e
˙ x−ax
. The Euler equation is
ae
˙ x−ax
÷
d
dt
e
˙ x−ax
= 0 ⇐⇒ ae
˙ x−ax
÷e
˙ x−ax
( ´ x −a ˙ x) = 0 ⇐⇒ ´ x −a ˙ x ÷a = 0
(c) Here F(t, x, ˙ x) = [(x − ˙ x)
2
÷x
2
]e
−at
, so F
/
x
= [2(x − ˙ x) ÷2x]e
−at
and F
/
˙ x
= −2(x − ˙ x)e
−at
. The
Euler equation becomes
[2(x − ˙ x) ÷2x]e
−at
÷
d
dt
[2(x − ˙ x)e
−at
] = 0
⇐⇒ [2(x − ˙ x) ÷2x]e
−at
÷2( ˙ x − ´ x)e
−at
−2a(x − ˙ x)e
−at
= 0
⇐⇒ 2(x − ˙ x) ÷2x ÷2( ˙ x − ´ x) −2a(x − ˙ x) = 0
⇐⇒ ´ x −a ˙ x ÷(a −2)x = 0
(d) With F(t, x, ˙ x) = 2t x ÷3x ˙ x ÷t ˙ x
2
we get F
/
x
= 2t ÷3˙ x and F
/
˙ x
= 3x ÷2t ˙ x, and the Euler equation
becomes
2t ÷3˙ x −
d
dt
(3x ÷2t ˙ x) = 0 ⇐⇒ 2t ÷3˙ x −3˙ x −2˙ x −2t ´ x = 0 ⇐⇒ t ´ x ÷ ˙ x = t
It is worth noting that this is an exact equation for ˙ x because it says that (d/dt )(t ˙ x) = t . Hence,
t ˙ x =
1
2
t
2
÷C, which implies ˙ x =
1
2
t ÷C/t , and so x =
1
4
t
2
÷C ln [t [ ÷A, where C and Aare constants.
8.2.4 With F(t, x, ˙ x) = x
2
÷2t x ˙ x ÷ ˙ x
2
we get F
/
x
= 2x ÷2t ˙ x and F
/
˙ x
= 2t x ÷2˙ x, so the Euler equation
(8.2.2) is
2x ÷2t ˙ x −
d
dt
(2t x ÷2˙ x) = 0 ⇐⇒ 2x ÷2t ˙ x −(2x ÷2t ˙ x ÷2´ x) = 0 ⇐⇒ ´ x = 0
The general solution of ´ x = 0 is x = At ÷B, and the boundary conditions x(0) = 1 and x(1) = 2 yield
A = B = 1. Thus the only admissible function that satisﬁes the Euler equation is x = t ÷1.
We have F
//
xx
= 2 > 0, F
//
x ˙ x
= 2t , and F
//
˙ x ˙ x
= 2 > 0, and since F
//
xx
F
//
˙ x ˙ x
−(F
//
x ˙ x
)
2
= 4 −4t
2
≥ 0 for
all t in [0, 1], it follows that F(t, x, ˙ x) is convex with respect to (x, ˙ x) as long as t ∈ [0, 1]. Hence, by
Theorem 8.3.1, x = t ÷1 is the optimal solution of the problem.
8.2.6 Let F(t, x, ˙ x) =
√
1 ÷ ˙ x
2
. Then F
/
x
= 0 and F
/
˙ x
=
˙ x
√
1 ÷ ˙ x
2
. The Euler equation, F
/
x
−
d
dt
F
/
˙ x
= 0,
therefore reduces to
d
dt
˙ x
√
1 ÷ ˙ x
2
= 0, i.e.
˙ x
√
1 ÷ ˙ x
2
= C (constant)
This implies
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
8 CAL CUL US OF VARI ATI ONS 61
˙ x
2
1 ÷ ˙ x
2
= C
2
, so ˙ x = C
1
(=
C
√
1 −C
2
)
Since ˙ x is a constant, x is a linear function of t :
x = C
1
t ÷C
2
.
Thus the graph of x is a straight line, namely the straight line through the points (t
0
, x
0
) and (t
1
, x
1
). Of
course, this is precisely what we would expect: the shortest curve between two points is the straight line
segment between them. The function x is given by the equation
x(t ) = x
0
÷
x
1
−x
0
t
1
−t
0
(t −t
0
)
Note also that F
//
˙ x ˙ x
= 1/(1 ÷ ˙ x
2
)
3/2
> 0, so F is convex with respect to (x, ˙ x), and Theorem 8.3.1 in
Section 8.3 shows that the function we have found really does give the minimum, at least among the
admissible C
2
functions.
8.3
8.3.2 (a) The objective function is F(t, x, ˙ x) = U( ¨ c − ˙ xe
rt
), and
∂F
∂x
= 0,
∂F
∂ ˙ x
= −U
/
( ¨ c − ˙ xe
rt
)e
rt
.
As ∂F/∂x = 0, the Euler equation reduces to
d
dt
−U
/
( ¨ c − ˙ xe
rt
)e
rt
= 0, so U
/
( ¨ c − ˙ xe
rt
)e
rt
= K (a constant)
(Evaluating the derivative
d
dt
−U
/
( ¨ c − ˙ xe
rt
)e
rt
above leads to the equation
−U
//
( ¨ c − ˙ xe
rt
)(−´ xe
rt
−r ˙ xe
rt
)e
rt
−rU
/
( ¨ c − ˙ xe
rt
)e
rt
= 0
This can be simpliﬁed to
´ x ÷r ˙ x =
rU
/
( ¨ c − ˙ xe
rt
)
U
//
( ¨ c − ˙ xe
rt
)
e
−rt
which will most likely be harder to solve.)
(b) It follows from part (a) that
U
/
( ¨ c − ˙ xe
rt
) = Ke
−rt
If U(c) = −e
−vc
/v, then U
/
(c) = e
−vc
, and we get
exp(−v ¨ c ÷v ˙ xe
rt
) = Ke
−rt
−v ¨ c ÷v ˙ xe
rt
= ln K −rt
˙ xe
rt
= C −rt /v (C = ¨ c ÷(ln K)/v)
˙ x = (C −rt /v)e
−rt
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
62 8 CAL CUL US OF VARI ATI ONS
Integration by parts yields
x = x(t ) =
(C −rt /v)e
−rt
dt = A ÷(B ÷t /v)e
−rt
where B = (1 −vC)/rv. Finally, the constants A and B are determined by the equations
A ÷B = x(0) = x
0
, A ÷(B ÷T/v)e
−rT
= x(T ) = 0
U is concave because U
//
(c) = −ve
−vc
< 0. Hence, Theorem 8.3.1 shows that the solution we have
found is optimal.
8.3.4 (a) WithF(t, y, ˙ y) = ln
¸
y−σ ˙ y−¨ zl(t )
¸
we get F
/
y
= 1/[y−σ ˙ y−¨ zl(t )] andF
/
˙ y
= −σ/[y−σ ˙ y−¨ zl(t )].
The Euler equation is then
1
y −σ ˙ y − ¨ zl(t )
−
d
dt
¸
−σ
y −σ ˙ y − ¨ zl(t )
¸
= 0 ⇐⇒
1
y −σ ˙ y − ¨ zl(t )
−
σ( ˙ y −σ ´ y − ¨ z
˙
l(t ))
[y −σ ˙ y − ¨ zl(t )]
2
= 0
⇐⇒ y −σ ˙ y − ¨ zl(t ) −σ ˙ y ÷σ
2
´ y ÷σ ¨ z
˙
l(t ) = 0 ⇐⇒ ´ y −
2
σ
˙ y ÷
1
σ
2
y =
¨ z
σ
2
¸
l(t ) −σ
˙
l(t )
¸
(b) With l(t ) = l
0
e
αt
the Euler equation becomes
´ y −
2
σ
˙ y ÷
1
σ
2
y =
¨ z(1 −ασ)l
0
σ
2
e
αt
(∗)
The characteristic equation r
2
− (2/σ)r ÷ 1/σ
2
= 0 has the double root r
1
= r
2
= 1/σ, so the
corresponding homogeneous differential equation has the general solution y
H
= Ae
t /σ
÷Bt e
t /σ
. To ﬁnd
a particular integral of (∗) we try with a function u
∗
= Ce
αt
. We get
´ u
∗
−
2
σ
˙ u
∗
÷
1
σ
2
u
∗
= α
2
Ce
αt
−
2α
σ
Ce
αt
÷
1
σ
2
Ce
αt
= C
(1 −ασ)
2
σ
2
e
αt
It follows that u
∗
is a solution of (∗) if and only if C = ¨ zl
0
/(1 −ασ), and the general solution of (∗) is
y = y
H
÷u
∗
= Ae
t /σ
÷Bt e
t /σ
÷
¨ zl
0
1 −ασ
e
αt
(If ασ = 1, then (∗) is homogeneous and the general solution is y
H
.)
8.4
8.4.1 With F(t, K,
˙
K) = e
−t /4
ln(2K −
˙
K) we get F
/
K
= 2e
−t /4
/(2K −
˙
K) and F
/
˙
K
= −e
−t /4
/(2K −
˙
K).
The Euler equation is
e
−t /4
2
2K −
˙
K
−
d
dt
−e
−t /4
1
2K −
˙
K
= 0 ⇐⇒
2e
−t /4
2K −
˙
K
−
e
−t /4
4(2K −
˙
K)
−
e
−t /4
(2
˙
K −
´
K)
(2K −
˙
K)
2
= 0
⇐⇒
e
−t /4
4(2K −
˙
K)
2
¸
8(2K −
˙
K) −(2K −
˙
K) −4(2
˙
K −
´
K)
¸
= 0
⇐⇒ 4
´
K −15
˙
K ÷14K = 0
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
8 CAL CUL US OF VARI ATI ONS 63
The characteristic equation of the Euler equation is 4r
2
− 15r ÷ 14 = 0, which has the roots r
1
= 2
and r
2
= 7/4. Hence the general solution of the Euler equation is K = Ae
2t
÷ Be
7t /4
. The boundary
conditions yield the equations
K(0) = A ÷B = K
0
, K(T ) = Ae
2T
÷Be
7T/4
= K
T
By means of Cramer’s rule or by other methods you will ﬁnd that
A =
K
T
−e
7T/4
K
0
e
2T
−e
7T/4
, B =
e
2T
K
0
−K
T
e
2T
−e
7T/4
8.4.2 (a) Let F(t, x, ˙ x) =
1
100
t x − ˙ x
2
e
−t /10
. Then F
/
x
= t e
−t /10
/100 and F
/
˙ x
= −2˙ xe
−t /10
, and the
Euler equation becomes
t
100
e
−t /10
−
d
dt
−2˙ xe
−t /10
= 0 ⇐⇒
t
100
e
−t /10
÷2´ xe
−t /10
−
2
10
˙ xe
−t /10
= 0
⇐⇒ ´ x −
1
10
˙ x = −
1
200
t (∗)
The general solution of the corresponding homogeneous equation is
x
H
= A ÷Be
t /10
To ﬁnd a particular solution u
∗
of (∗) we try with u = Pt
2
÷Qt . Then ˙ u = 2Pt ÷Q and ´ u = 2P, and
if we insert this into (∗) we get
2P −
P
5
t −
Q
10
= −
1
200
t
This yields P = 5/200 = 1/40 and Q = 20P = 1/2, so the general solution of the Euler equation is
x = A ÷Be
t /10
÷
1
40
t
2
÷
1
2
t
The boundary conditions x(0) = 0 and x(T ) = S yield the equations
A ÷B = 0, A ÷Be
T/10
÷T
2
/40 ÷T/2 = S
with the solution
A = −B =
T
2
/40 ÷T/2 −S
e
T/10
−1
Since F(t, x, ˙ x) is concave with respect to (x, ˙ x), this is indeed an optimal solution.
(b) With T = 10 and S = 20 we get B = −5/2 ÷5 −20/(e −1) = 25/2(e −1), and the optimal
solution is
x =
25(e
t /10
−1)
2(e −1)
÷
1
40
t
2
÷
1
2
t
8.4.3 With F(t, K,
˙
K) = U(C, t ) we get F
/
K
= U
/
C
C
/
K
= U
/
C
(f
/
k
−δ) and F
/
˙
K
= U
/
C
· (−1) = −U
/
C
. The
Euler equation is
U
/
C
(f
/
K
−δ) ÷
d
dt
U
/
C
= 0 ⇐⇒ U
/
C
(f
/
K
−δ) ÷U
//
CC
˙
C ÷U
//
Ct
= 0
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
64 8 CAL CUL US OF VARI ATI ONS
The Euler equation implies
˙
C = −[U
//
Ct
÷U
/
C
(f
/
K
−δ)]/U
//
CC
, and therefore
˙
C
C
= −
U
//
Ct
÷U
/
C
(f
/
K
−δ)
CU
//
CC
= −
1
ˇ ω
U
//
Ct
U
/
C
÷f
/
K
−δ
where ˇ ω = CU
//
CC
/U
/
C
is the elasticity of marginal utility with respect to consumption.
8.4.4 (a) F(t, p, ˙ p) = pD(p, ˙ p) −b(D(p, ˙ p)) yields F
/
p
= D÷pD
/
p
−b
/
(D)D
/
p
= D÷[p −b
/
(D)]D
/
p
and F
/
˙ p
= [p −b
/
(D)]D
/
˙ p
, so the Euler equation is
F
/
p
−
d
dt
(F
/
˙ p
) = 0 ⇐⇒ D ÷
¸
p −b
/
(D)
¸
D
/
p
−
d
dt
¸
[p −b
/
(D)]D
/
˙ p
¸
= 0
(b) With b(x) = αx
2
÷βx÷γ and x = D(p, ˙ p) = Ap÷B ˙ p÷C, we get b
/
(x) = 2αx÷β, ∂D/∂p = A,
and ∂D/∂ ˙ p = B. Insertion into the Euler equation gives
Ap ÷B ˙ p ÷C ÷[p −2α(Ap ÷B ˙ p ÷C) −β]A −
d
dt
¸
p −2α(Ap ÷B ˙ p ÷C) −β
B
¸
= 0
which yields Ap ÷B ˙ p ÷C ÷Ap −2αA
2
p −2αAB ˙ p −2αAC −βA−B ˙ p ÷2αAB ˙ p ÷2αB
2
´ p = 0.
Rearranging the terms we can write the Euler equation as
´ p −
A
2
−A/α
B
2
p =
βA ÷2αAC −C
2αB
2
and it is easy to see that we get the answer given in the book.
8.5
8.5.2 (a) The Euler equation is
(−2˙ x −10x)e
−t
−
d
dt
[(−2˙ x −2x)e
−t
] = 0 ⇐⇒ e
−t
(−2˙ x −10x ÷2´ x ÷2˙ x −2˙ x −2x) = 0
⇐⇒ ´ x ÷ ˙ x −6x = 0
The general solution of this equation is x = Ae
3t
÷Be
−2t
, and the boundary conditions yield the equations
x(0) = A ÷B = 0, x(1) = Ae
3
÷Be
−2
= 1
which have the solution A = −B = 1/(e
3
−e
−2
). The integrand F(t, x, ˙ x) = (10− ˙ x
2
−2x ˙ x −5x
2
)e
−t
is concave with respect to (x, ˙ x) (look at the Hessian or note that −˙ x
2
−2x ˙ x −5x
2
= −( ˙ x −x)
2
−4x
2
),
so the solution we have found is optimal (Theorem 8.5.1).
(b) (i) With x(1) free, we get the transversality condition
¸
F
/
˙ x
¸
t =1
= 0, which implies
˙ x(1) ÷x(1) = 0 ⇐⇒ 3Ae
3
−2Be
−2
÷Ae
3
÷Be
−2
= 0 ⇐⇒ 4Ae
3
−Be
−2
= 0
Together with the equation A÷B = x(0) = 0, this yields A = B = 0, so the optimal solution is x ≡ 0.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
9 CONTROL THEORY: BASI C TECHNI QUES 65
(ii) With the terminal condition x(1) ≥ 2 the transversality condition is
¸
F
/
˙ x
¸
t =1
≤ 0, with
¸
F
/
˙ x
¸
t =1
= 0
if x(1) > 2. In our case this reduces to
−˙ x(1) −x(1) ≤ 0, and −˙ x(1) −x(1) = 0 if x(1) > 2
Since x(t ) = Ae
3t
−Ae
−2t
, we get
−A(4e
3
÷e
−2
) ≤ 0, and −A(4e
3
÷e
−2
) = 0 if x(1) = A(e
3
−e
−2
) > 2
From the last condition it is clear that we cannot have A(e
3
− e
−2
) > 2, for then we must have A = 0,
and that is impossible. Therefore A(e
3
−e
−2
) = 2 and x =
2(e
3t
−e
−2
)
e
3
−e
−2
.
8.5.4 With U(C) = a −e
−bC
, we get U
/
(C) = be
−bC
and U
//
(C) = −b
2
e
−bC
, so equation (∗) in Example
8.5.3 reduces to
´
A − r
˙
A = (ρ − r)/b. Putting z =
˙
A, we get ˙ z − rz = (ρ − r)/b. This is a linear
ﬁrstorder differential equation with constant coefﬁcients, and the solution is z = Pe
rt
− (ρ − r)/br.
Then A =
z dt = Ke
rt
−(ρ −r)t /br ÷L, where K = P/r and L are constants. These constants are
determined by the boundary conditions A(0) = A
0
and A(T ) = A
t
.
8.5.5 (a) The conditions are −(d/dt )[C
/
˙ x
(t, ˙ x)e
−rt
] = 0 and C
/
˙ x
(5, ˙ x(5)) ≥ 0 (= 0 if x(5) > 1500). It
follows that C
/
˙ x
(t, ˙ x) = Ke
−rt
, where K is a constant.
(b) It follows from part (a) that if r = 0 then C
/
˙ x
(t, ˙ x) = K must be constant. With C(t, u) = g(u) it
means that g
/
( ˙ x) must be constant. Since g
/
is strictly increasing, ˙ x must also be a constant, so x = At .
We must have 5A ≥ 1500. Since C
/
˙ x
(t, ˙ x) = g
/
( ˙ x) = g
/
(A) > 0, the transversality condition shows
that x(5) = 1500, so A = 300 and x = 300t . Thus planting will take place at the constant rate of 300
hectares per year.
9 Control Theory: Basic Techniques
9.2
9.2.3 We transform the problem to a maximization problem by maximizing
1
0
[−x(t ) − u(t )
2
] dt . The
Hamiltonian H(t, x, u, p) = −x − u
2
− pu is concave in (x, u), so according to Theorem 9.2.1, the
following conditions are sufﬁcient for optimality (using (9.2.7)):
(i) H
/
u
(t, x
∗
(t ), u
∗
(t ), p(t )) = −2u
∗
(t ) −p(t ) = 0;
(ii) ˙ p(t ) = −H
/
x
(t, x
∗
(t ), u
∗
(t ), p(t )) = 1, p(1) = 0;
(iii) ˙ x
∗
(t ) = −u
∗
(t ), x
∗
(0) = 0.
From (ii), p(t ) = t ÷A for some constant A, and p(1) = 0 gives A = −1 and so p(t ) = t −1. From (i)
we have u
∗
(t ) = −
1
2
p(t ) =
1
2
(1−t ). It remains to determine x
∗
(t ). From(iii), ˙ x
∗
(t ) = −u
∗
(t ) =
1
2
t −
1
2
,
and so x
∗
(t ) =
1
4
t
2
−
1
2
t ÷B. With x
∗
(0) = 0, we get B = 0. So the optimal solution is u
∗
(t ) =
1
2
(1−t ),
x
∗
(t ) =
1
4
t
2
−
1
2
t , with p(t ) = t −1.
9.2.4 The Hamiltonian H(t, x, u, p) = 1−4x−2u
2
÷puis concave in (x, u), so according toTheorem9.2.1
(using (9.2.7), the following conditions are sufﬁcient for optimality:
(i) H
/
u
(t, x
∗
(t ), u
∗
(t ), p(t )) = −4u
∗
(t ) ÷p(t ) = 0;
(ii) ˙ p(t ) = −H
/
x
(t, x
∗
(t ), u
∗
(t ), p(t )) = 4, p(10) = 0;
(iii) ˙ x
∗
(t ) = u
∗
(t ), x
∗
(0) = 0.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
66 9 CONTROL THEORY: BASI C TECHNI QUES
From(ii), p(t ) = 4t ÷Afor some constant Aand p(10) = 40÷A = 0, so A = −40 and p(t ) = 4t −40.
From (i) we have u
∗
(t ) =
1
4
(4t −40) = t −10. From (iii), ˙ x
∗
(t ) = u
∗
(t ) = t −10, and with x
∗
(0) = 0,
we get x
∗
(t ) =
1
2
t
2
−10t .
9.2.5 The Hamiltonian H(t, x, u, p) = x−u
2
÷p(x÷u) is concave in (x, u), so according toTheorem9.2.1
(using (9.2.7), the following conditions are sufﬁcient for optimality:
(i) H
/
u
(t, x
∗
(t ), u
∗
(t ), p(t )) = −2u
∗
(t ) ÷p(t ) = 0;
(ii) ˙ p(t ) = −H
/
x
(t, x
∗
(t ), u
∗
(t ), p(t )) = −1 −p(t ), p(T ) = 0;
(iii) ˙ x
∗
(t ) = x
∗
(t ) ÷u
∗
(t ), x
∗
(0) = 0.
From(ii) we get the linear differential equation ˙ p(t )÷p(t ) = −1. The general solutionis p(t ) = Ae
−t
−1,
and p(T ) = 0 gives A = e
T
, so p(t ) = e
T −t
− 1. From (i) we have u
∗
(t ) =
1
2
(e
T −t
− 1). Finally,
from (iii), ˙ x
∗
(t ) = x
∗
(t ) ÷
1
2
(e
T −t
−1), with general solution x
∗
(t ) = Be
t
÷
1
2
e
t
e
−t
(e
T −t
−1) dt =
Be
t
−
1
4
e
T −t
÷
1
2
. The condition x
∗
(0) = 0 gives B =
1
4
e
T
−
1
2
. Thus, x
∗
(t ) =
1
4
e
T ÷t
−
1
4
e
T −t
−
1
2
e
t
÷
1
2
.
9.2.6 (b) Conditions (7) and (5) reduce to
(i) I
∗
(t ) =
1
2
p(t );
(ii) ˙ p(t ) −0.1p(t ) = −1 ÷0.06K
∗
(t ).
Moreover, K
∗
(t ) = I
∗
(t ) −0.1K
∗
(t ) =
1
2
p(t ) −0.1K
∗
(t ). Thus (K
∗
(t ), p(t )) must satisfy
(iii)
˙
K =
1
2
p −0.1K, and
(iv) ˙ p −0.1p = −1 ÷0.06K.
From (iii) we get p = 2
˙
K ÷ 0.2K, and then ˙ p = 2
´
K ÷ 0.2
˙
K. Inserting these results into (iv) and
simplifying yields
´
K −0.04K = −0.5.
9.4
9.4.2 The Hamiltonian H(t, x, p, u) = 1 − x
2
− u
2
÷ pu is concave in (x, u), so according to Theorem
9.4.2 (using (9.2.7), the following conditions are sufﬁcient for optimality:
(i) H
/
u
(t, x
∗
(t ), u
∗
(t ), p(t )) = −2u
∗
(t ) ÷p(t ) = 0;
(ii) ˙ p(t ) = −H
/
x
(t, x
∗
(t ), u
∗
(t ), p(t )) = 2x
∗
(t );
(iii) p(1) ≥ 0, and p(1) = 0 if x
∗
(1) > 1;
(iv) ˙ x
∗
(t ) = u
∗
(t ), x
∗
(0) = 0, x
∗
(1) ≥ 1.
From ˙ p(t ) = 2x
∗
(t ) we get ´ p(t ) = 2˙ x
∗
(t ) = 2u
∗
(t ) = p(t ). It follows that p(t ) = Ae
t
÷ Be
−t
for
suitable constants A and B. Furthermore, x
∗
(t ) =
1
2
˙ p =
1
2
(Ae
t
− Be
−t
), and since x
∗
(0) = 0 gives
B = A, we have x
∗
(t ) =
1
2
A(e
t
− e
−t
), u
∗
(t ) = ˙ x
∗
(t ) =
1
2
A(e
t
÷ e
−t
), and p(t ) = A(e
t
÷ e
−t
). If
x
∗
(1) > 1, we must have p(1) = 0, and therefore A = 0, but that would give x
∗
(t ) = 0 for all t , which
contradicts x
∗
(1) ≥ 1. Therefore we must have x
∗
(1) = 1, which gives A = 2/(e −e
−1
) = 2e/(e
2
−1).
(The value of A here is twice the value of A given in the book.)
9.4.3 (a) As in Example 9.2.1, we have p(t ) = −
1
2
(T
2
− t
2
), but this time u
∗
(t ) is the value of u that
maximizes H(t, x
∗
(t ), u, p(t )) = 1−t x
∗
(t ) −u
2
÷p(t )u for u ∈ [0, 1]. Note that H
/
u
= −2u÷p(t ) =
−2u−
1
2
(T
2
−t
2
) < 0 for all t < T . Thus the optimal choice of u must be u
∗
(t ) ≡ 0, and then x
∗
(t ) ≡ x
0
.
(b) Also in this case H
/
u
= −2u ÷ p(t ) = −2u −
1
2
(T
2
− t
2
) and H
//
uu
= −2. So for each t in [0, T ],
the optimal u
∗
(t ) must maximize the concave function H(t, x
∗
(t ), u, p(t )) for u ∈ [−1, 1]. Note that
H
/
u
= 0 when u = −
1
4
(T
2
− t
2
), and this nonpositive number is an interior point of [−1, 1] provided
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
9 CONTROL THEORY: BASI C TECHNI QUES 67
−
1
4
(T
2
−t
2
) > −1, i.e. t >
√
T
2
−4. If t ≤
√
T
2
−4, then u = −1 is optimal. (Choosing u
∗
(t ) = 1
cannot be optimal because H
/
u
is negative for u = 1.) For the rest see the answer in the book.
9.4.4 (a) Since H(t, x, p, u) = x ÷ pu is concave in (x, u), the following conditions are sufﬁcient for
optimality:
(i) u = u
∗
(t ) maximizes p(t )u for u ∈ [0, 1];
(ii) ˙ p = −1;
(iii) ˙ x
∗
(t ) = u
∗
(t ), x
∗
(0) = 0, x
∗
(10) = 2.
From (ii) we get p(t ) = A − t for some constant A, and (i) implies that u
∗
(t ) = 1 if p(t ) > 0 and
u
∗
(t ) = 0 if p(t ) < 0. Now u
∗
(t ) ≡ 0 and u
∗
(t ) ≡ 1 are both impossible because ˙ x
∗
(t ) = u
∗
(t ) and
x
∗
(0) = 0 contradict x
∗
(10) = 2. Since p(t ) is strictly decreasing, we conclude that for some t
∗
in
(0, 10) we have p(t ) > 0, and thus u
∗
(t ) = 1, for t in [0, t
∗
], and p(t ) < 0 with u
∗
(t ) = 0 for t in
(t
∗
, 10]. At t
∗
we have p(t
∗
) = 0, so A = t
∗
and p(t ) = t
∗
−t . We see that x
∗
(t ) = t for t in [0, t
∗
] and
x
∗
(t ) = t
∗
for t in (t
∗
, 10]. Since x
∗
(10) = 2, we conclude that t
∗
= 2, and the solution is as given in
the book.
(b) As in (a) we ﬁnd that for some t
∗
in (0, T ), u
∗
(t ) = 1 in [0, t
∗
] and u
∗
(t ) = 0 for t in (t
∗
, T ]. Then
x
∗
(t ) = t ÷x
0
for t in [0, t
∗
] and x
∗
(t ) = t
∗
÷x
0
for t in (t
∗
, T ]. Since x
∗
(T ) = x
1
, we conclude that
t
∗
= x
1
−x
0
, and the solution is as given in the book.
9.4.6 The Hamiltonian H(t, x, u, p) = [10u − u
2
− 2]e
−0.1t
− pu is concave in (x, u), so according to
Theorem 9.4.2, the following conditions are sufﬁcient for optimality:
(i) u
∗
(t ) maximizes H(t, x
∗
(t ), u, p(t )) = [10u −u
2
−2]e
−0.1t
−p(t )u for u ≥ 0;
(ii) ˙ p(t ) = −H
/
x
(t, x
∗
(t ), u
∗
(t ), p(t )) = 0;
(iii) p(5) ≥ 0, and p(5) = 0 if x
∗
(5) > 0;
(iv) ˙ x
∗
(t ) = −u
∗
(t ), x
∗
(0) = 10, x
∗
(5) ≥ 0.
From (ii) we have p(t ) ≡ ¨ p for some constant ¨ p. Suppose u
∗
(t ) > 0 for all t in [0, 5]. Then (i) is
equivalent to H
/
u
(t, x
∗
(t ), u
∗
(t ), p(t )) = 0, that is,
(v) [10 −2u
∗
(t )]e
−0.1t
= ¨ p.
If ¨ p = 0, then (v) implies that u
∗
(t ) ≡ 5, and then ˙ x
∗
(t ) = −5. With x
∗
(0) = 10, we get x
∗
(t ) = 10−5t ,
which gives x
∗
(5) = −15, a contradiction. Hence ¨ p > 0 and from (iii) (and x
∗
(5) ≥ 0) we conclude
that x
∗
(5) = 0. From (v) we have u
∗
(t ) = 5 −
1
2
¨ pe
0.1t
, and then ˙ x
∗
(t ) = −u
∗
(t ) with x
∗
(0) = 10
gives x
∗
(t ) = 5 ¨ p(e
0.1t
− 1) − 5t ÷ 10. The condition x
∗
(5) = 0 gives ¨ p = 3/(e
0.5
− 1). Since all the
conditions (i)–(iv) are satisﬁed by the pair (x
∗
(t ), u
∗
(t )), we have found an optimal solution.
9.4.7 (b) The Hamiltonian H(t, x, u, p) = −(ax ÷bu
2
) ÷pu is concave in (x, u). To check the maximum
condition (9.4.5) inTheorem9.4.1, it sufﬁces to check that, if u
∗
(t ) > 0, then (H
/
u
)
∗
= −2bu
∗
(t )÷p(t ) =
0, and if u
∗
(t ) = 0, then (H
/
u
)
∗
= −2bu
∗
(t ) ÷p(t ) ≤ 0. Also p(t ) = at ÷A for some constant A.
Suppose u
∗
(t ) > 0 for all t in [0, T ]. Then u
∗
(t ) = p(t )/2b = (1/2b)(at ÷ A). Moreover,
˙ x
∗
(t ) = u
∗
(t ) = (1/2b)(at ÷ A), so x
∗
(t ) = (1/2b)(
1
2
at
2
÷ At ) ÷ C. Here x
∗
(0) = 0 yields C = 0,
and x
∗
(T ) = B gives A = 2bB/T −aT/2. Thus
u
∗
(t ) = a(2t −T )/4b ÷B/T and x
∗
(t ) = at (t −T )/4b ÷Bt /T
Note that u
∗
(t ) is increasing in t , and u
∗
(0) = B/T −aT/4b ≥ 0 if and only if B ≥ aT
2
/4b. So in this
case we have found an optimal solution, because it is easy to check that all the conditions in Theorem
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
68 9 CONTROL THEORY: BASI C TECHNI QUES
9.4.2 are satisﬁed. (This solution is valid if the required total production B is large relative to the time
period T available, and the storage cost a is sufﬁciently small relative to the unit production cost b. See
Kamien and Schwartz (1991) for further economic interpretations.)
Suppose B < aT
2
/4b. Then we guess that production is postponed in an initial time period. We
cannot have u
∗
(t ) ≡ 0 because then x
∗
(t ) ≡ 0, which contradicts x
∗
(T ) = B > 0 (assumed implicitly).
If u
∗
(t
∗÷
) > 0, thenbycontinuityof p(t ), the unique maximizer of H wouldbe > 0 for some t < t
∗
, witht
close to t
∗
. So u
∗
(t
∗÷
) = 0, and u
∗
(t ) = x
∗
(t ) ≡ 0 in [0, t
∗
]. In particular, x
∗
(t
∗
) = 0. In (t
∗
, T ] we have
as before u
∗
(t ) = (1/2b)(at ÷A), and since u
∗
(t
∗÷
) = 0, we have A = −at
∗
, and u
∗
(t ) = (a/2b)(t −t
∗
).
Then ˙ x
∗
(t ) = u
∗
(t ) = (a/2b)(t −t
∗
), andwithx
∗
(t
∗
) = 0, x
∗
(t ) = (a/4b)(t −t
∗
)
2
. Nowt
∗
is determined
by the condition x
∗
(T ) = B, which gives (a/4b)(T − t
∗
)
2
= B, or T − t
∗
= ±
√
4bB/a. The minus
sign would make t
∗
> T , so we must have t
∗
= T − 2
√
bB/a. Note that t
∗
> 0 ⇐⇒ B < aT
2
/4b.
We end up with the following suggestion for an optimal solution:
u
∗
(t ) =
⎧
⎨
⎩
0 if t ∈ [0, t
∗
]
a(t −t
∗
)
2b
if t ∈ (t
∗
, T ]
, x
∗
(t ) =
⎧
⎨
⎩
0 if t ∈ [0, t
∗
]
a(t −t
∗
)
2
4b
if t ∈ (t
∗
, T ]
, p(t ) = a(t −t
∗
)
with t
∗
= T −2
√
bB/a.
It remains to check that the proposed solution satisﬁes all the conditions in Theorem 9.4.2. The
pair (x
∗
, u
∗
) is admissible and satisﬁes the terminal condition x
∗
(T ) = B, and ˙ p = a = −(H
/
x
)
∗
. It
remains only to check the maximum condition. For t in (t
∗
, T ] we have u
∗
(t ) > 0 and we see that
(H
/
u
)
∗
= −2bu
∗
(t ) ÷ p(t ) = 0. For t in [0, t
∗
], we have u
∗
(t ) = 0 and (H
/
u
)
∗
= −2bu
∗
(t ) ÷ p(t ) =
a(t −t
∗
) ≤ 0, as it should be. So we have found an optimal solution.
Note: When we use sufﬁcient conditions, we can use “wild” guesses about the optimal control and
the optimal path, as long as we really check that all the conditions in the sufﬁciency theorem are satisﬁed.
9.4.8 The Hamiltonian is H(t, x, p, u) = x
2
− 2u ÷ pu = x
2
÷ (p − 2)u. (Since x(2) is free, we put
p
0
= 1.) The maximum principle (Theorem 9.4.1) gives the following necessary conditions:
(i) u = u
∗
(t ) maximizes (p(t ) −2)u for u ∈ [0, 1];
(ii) ˙ p = −H
/
x
(t, x
∗
(t ), u
∗
(t ), p(t )) = −2x
∗
(t ), p(2) = 0;
(iii) ˙ x
∗
(t ) = u
∗
(t ), x
∗
(0) = 1.
From (i) we see that p(t ) > 2 ⇒ u
∗
(t ) = 1 and p(t ) < 2 ⇒ u
∗
(t ) = 0. The function x
∗
must
be increasing, because ˙ x
∗
(t ) ≥ 0. Hence ˙ p(t ) = −2x
∗
(t ) ≤ −2x
∗
(0) = −2 < 0. Consequently p is
strictly decreasing. Since p(2) = 0, we have p(t ) > 0 for t in [0, 2). Because ˙ p ≤ −2, we see that
p(2) −p(0) =
2
0
˙ p(t ) dt ≤
2
0
(−2) dt = −4, so p(0) ≥ p(2) ÷4 = 4. There is therefore a unique t
∗
in (0, 2) with p(t
∗
) = 2, and
p(t )
¸
> 2 for t ∈ [0, t
∗
)
< 2 for t ∈ (t
∗
, 2]
=⇒ u
∗
(t ) =
¸
1 for t ∈ [0, t
∗
]
0 for t ∈ (t
∗
, 2]
For t ≤ t
∗
, we have ˙ x
∗
(t ) = u
∗
(t ) = 1, so x
∗
(t ) = x
∗
(0) ÷ t = 1 ÷ t . Moreover, ˙ p(t ) =
−2x(t ) = −2−2t , which gives p(t ) = −2t −t
2
÷C
0
for a suitable constant C
0
. Since x
∗
is continuous,
x
∗
(t
∗
) = 1 ÷t
∗
(the limit of x
∗
(t ) as t approaches t
∗
from the left).
For t ≥ t
∗
, we have ˙ x
∗
(t ) = u
∗
(t ) = 0, so x
∗
(t ) remains constant, x
∗
(t ) = x
∗
(t
∗
) = 1 ÷ t
∗
, and
˙ p(t ) = −2x
∗
(t ) = −2 −2t
∗
, so p(t ) = −2(1 ÷t
∗
)t ÷C
1
, where C
1
= p(2) ÷2(1 ÷t
∗
)2 = 4(1 ÷t
∗
).
Now, p(t
∗
) must equal 2, so −2(1÷t
∗
)t
∗
÷4(1÷t
∗
) = 2, or (t
∗
)
2
−t
∗
−1 = 0. This quadratic equation
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
9 CONTROL THEORY: BASI C TECHNI QUES 69
gives t
∗
= (1±
√
5 )/2, and since t
∗
must be positive, t
∗
= (1÷
√
5 )/2. We knowthat for t ≥ t
∗
we have
p(t ) = −2(1 ÷t
∗
)t ÷C
1
= −2(1 ÷t
∗
)t ÷4(1 ÷t
∗
) = 2(1 ÷t
∗
)(2 −t ) = (3 ÷2
√
5 )(2 −t ). We also
know that there is a constant C
0
such that p(t ) = −2t −t
2
÷C
0
for t ≤ t
∗
. To determine the constant C
0
,
we use the equation p(t
∗
) = 2, which gives C
0
−2t
∗
−(t
∗
)
2
= 2, or C
0
= 2 ÷2t
∗
÷(t
∗
)
2
=
9
2
÷
3
2
√
5.
The only possible solution is spelled out in the answer in the book.
9.5
9.5.1 With F(t, x, ˙ x) = 2xe
−t
−2x ˙ x − ˙ x
2
we get F
/
x
= 2e
−t
−2˙ x and F
/
˙ x
= −2x −2˙ x. The Euler equation
is then 2e
−t
− 2˙ x −
d
dt
(−2x − 2˙ x) = 0, or 2e
−t
− 2˙ x ÷ 2˙ x ÷ 2´ x = 0, which reduces to ´ x = −e
−t
. It
follows that ˙ x = e
−t
÷A and x = −e
−t
÷At ÷B for suitable constants A and B. From the boundary
conditions x(0) = 0 and x(1) = 1 we get A = e
−1
and B = 1, so the only admissible solution of the
Euler equation is x
∗
(t ) = −e
−t
÷e
−1
t ÷1.
The associated control problem is
max
1
0
(2xe
−t
−2xu −u
2
) dt, ˙ x = u, x(0) = 0, x(1) = 1, u ∈ (−∞, ∞)
The Hamiltonian (with p
0
= 1) is then H(t, x, u, p) = 2xe
−t
−2xu−u
2
÷pu. (Incidentally, if p
0
= 0,
then the Hamiltonian would simply be pu. If p(t )u has a maximum for some u in (−∞, ∞), then
p(t ) = 0, but p(t ) = 0 is impossible when p
0
= 0.) We have H
/
u
= −2x −2u÷p and H
/
x
= 2e
−t
−2u.
Since the control region is open, H
/
u
(t, x
∗
(t ), u
∗
(t ), p(t )) = 0, so (i) u
∗
(t ) =
1
2
p(t ) −x
∗
(t ). Moreover
(ii) ˙ p(t ) = −H
/
x
(t, x
∗
(t ), u
∗
(t ), p(t )) = −2e
−t
÷ 2u
∗
(t ) = −2e
−t
÷ 2˙ x
∗
(t ). From ˙ x
∗
(t ) = u
∗
(t ) we
get ´ x
∗
(t ) = ˙ u
∗
(t ) =
1
2
˙ p − ˙ x
∗
(t ) = −e
−t
, using (i) and (ii). Hence x
∗
(t ) must be a solution of the
Euler equation that we found in (a). As we found in (a), there is precisely one solution that also satisﬁes
the boundary conditions on x
∗
(t ). Now that we know x
∗
(t ), the control function u
∗
(t ) and the adjoint
function p(t ) are given by the equations above.
9.5.2 With F(t, x, ˙ x) = 3−x
2
−2˙ x
2
, F
/
x
= −2x and F
/
˙ x
= −4˙ x, so the Euler equation F
/
x
−(d/dt )F
/
˙ x
= 0
reduces to ´ x −
1
2
x = 0. The characteristic equation r
2
=
1
2
has solutions r
1
=
1
2
√
2 and r
2
= −
1
2
√
2.
So if x
∗
(t ) solves the problem, then we must have x
∗
(t ) = Ae
1
2
√
2 t
÷Be
−
1
2
√
2 t
.
Since x
∗
(0) = 1, we have A÷B = 1, or B = 1−A, and, moreover, x
∗
(2) = Ae
√
2
÷(1−A)e
−
√
2
≥ 4
requires
A ≥ (4 −e
−
√
2
)/(e
√
2
−e
−
√
2
) = (4e
√
2
−1)/(e
2
√
2
−1) ≈ 0.97 (∗)
We now invoke the transversality condition (8.5.3):
(F
/
˙ x
)
∗
t =2
= −4˙ x
∗
(2) ≤ 0, with −4˙ x
∗
(2) = 0 if x
∗
(2) > 4
Equivalently, ˙ x
∗
(2) ≥ 0, with ˙ x
∗
(2) = 0 if x
∗
(2) > 4. Since ˙ x
∗
(t ) =
1
2
√
2[Ae
1
2
√
2 t
−(1 −A)e
−
1
2
√
2 t
],
we have ˙ x
∗
(2) =
1
2
√
2[Ae
√
2
− (1 − A)e
−
√
2
] = 0 provided A = e
−
√
2
/(e
√
2
÷ e
−
√
2
) ≈ 0.06, contra
dicting (∗). We conclude that x
∗
(2) = 4 and thus A = (4e
√
2
−1)/(e
2
√
2
−1). The function F is (for t
ﬁxed) concave in (x, ˙ x), so we have found the solution.
The control problem is
max
2
0
(3 −x
2
−2u
2
) dt, ˙ x = u, x(0) = 1, x(2) ≥ 4
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
70 9 CONTROL THEORY: BASI C TECHNI QUES
The Hamiltonian is H = 3 − x
2
− 2u
2
÷ pu, so H
/
x
= −2x and H
/
u
= −4u ÷ p. If (x
∗
(t ), u
∗
(t )) is
optimal then (i) u
∗
(t ) =
1
4
p(t ); (ii) ˙ p(t ) = 2x
∗
(t ); (iii) ˙ x
∗
(t ) = u
∗
(t ). Differentiating (iii) yields
´ x
∗
(t ) = ˙ u
∗
(t ) =
1
4
˙ p(t ) =
1
4
2x
∗
(t ) =
1
2
x
∗
(t ), which is the same differential equation as before. The rest
is easy.
9.5.3 With F(t, x, ˙ x) = (−2˙ x − ˙ x
2
)e
−t /10
, F
/
x
= 0 and F
/
˙ x
= (−2 − ˙ x)e
−t /10
, so the Euler equation is
−
d
dt
(−2 − ˙ x)e
−t /10
= 0. See the answer in the book. The function F is (for t ﬁxed) concave in (x, ˙ x),
so we have found the solution.
The Hamiltonian for the control problemis H = (−2u−u
2
)e
−t /10
÷pu. Here ˙ p(t ) = −(H
/
x
)
∗
= 0,
so p(t ) = ¨ p, a constant. Moreover, (H
/
u
)
∗
= (−2 − 2u
∗
(t ))e
−t /10
÷ p(t ) = 0, and thus u
∗
(t ) =
1
2
¨ pe
t /10
−1. Since ˙ x
∗
(t ) = u
∗
(t ), integration gives x
∗
(t ) = 5 ¨ pe
t /10
−t ÷A. The boundary conditions
yield ¨ p = 0 and A = 1, so we get the same solution.
9.6
9.6.1 (a) The Hamiltonian H(t, x, u, p) = x −
1
2
u
2
÷ pu is concave in (x, u), so according to Theorem
9.2.2 (see (9.2.7)), the following conditions are sufﬁcient for optimality:
(i) H
/
u
(t, x
∗
(t ), u
∗
(t ), p(t )) = −u
∗
(t ) ÷p(t ) = 0;
(ii) ˙ p(t ) = −H
/
x
(t, x
∗
(t ), u
∗
(t ), p(t )) = −1 with p(T ) = 0;
(iii) ˙ x
∗
(t ) = u
∗
(t ), x
∗
(0) = x
0
.
From (i) and (ii) we have u
∗
(t ) = p(t ) = T −t . Since ˙ x
∗
(t ) = u
∗
(t ) = T −t and x
∗
(0) = x
0
, we get
x
∗
(t ) = T t −
1
2
t
2
÷x
0
.
(b) The optimal value function is
V(x
0
, T ) =
T
0
x
∗
(t ) −
1
2
u
∗
(t )
2
dt =
T
0
T t −
1
2
t
2
÷x
0
−
1
2
(T −t )
2
dt (∗)
Integrating we ﬁnd V(x
0
, T ) =
t =T
t =0
(
1
2
T t
2
−
1
6
t
3
÷x
0
t ÷
1
6
(T −t )
3
) =
1
6
T
3
÷x
0
T , so ∂V/∂x
0
= T =
p(0). (If we were asked only to ﬁnd ∂V/∂x
0
, it would be easier to differentiate (∗) with respect to x
0
under the integral sign: ∂V/∂x
0
=
T
0
dt = T .)
The value of the Hamiltonian “along the optimal path” is H
∗
(t ) = H(t, x
∗
(t ), u
∗
(t ), p(t )) =
x
∗
(t ) −
1
2
u
∗
(t )
2
÷p(t )u
∗
(t ) = T t −
1
2
t
2
÷x
0
−
1
2
(T −t )
2
÷(T −t )
2
, and so H
∗
(T ) =
1
2
T
2
÷x
0
. We
see that ∂V/∂T =
1
2
T
2
÷x
0
= H
∗
(T ).
9.6.4 (a) The Hamiltonian is H(t, x, u, p) = 2x
2
e
−2t
− ue
t
÷ pue
t
= 2x
2
e
−2t
÷ (p − 1)ue
t
. (Recall
that x(T ) is free.) Suppose that (x
∗
(t ), u
∗
(t )) is an admissible pair with adjoint function p(t ). Since
H
/
x
= 4xe
−2t
, the differential equation for p(t ) is ˙ p(t ) = −4e
−2t
x
∗
(t ). For all t in [0, T ], we have
˙ x
∗
(t ) = u
∗
(t )e
t
≥ 0, and hence x
∗
(t ) ≥ x
∗
(0) = 1 for all t in [0, T ]. Therefore, ˙ p(t ) < 0, so p(t ) is
strictly decreasing. Fromthe maximumprinciple, for every t in [0, T ], u
∗
(t ) is the value of u in [0, 1] that
maximizes H(t, x
∗
(t ), u, p(t )) = 2(x
∗
(t ))
2
e
−2t
÷(p(t ) −1)ue
t
. Since 2(x
∗
(t ))
2
e
−2t
does not depend
on u, u
∗
(t ) is the value of u in [0, 1] that maximizes (p(t ) −1)u. Thus, u
∗
(t ) = 1 if p(t ) > 1, u
∗
(t ) = 0
if p(t ) < 1. Since p(T ) = 0, we have p(t ) < 1 for t close to T . There are two possibilities:
Case A: Suppose p(0) ≤ 1. Since p is strictly decreasing, we get p(t ) < 1, and hence u
∗
(t ) = 0 for all
t > 0. It follows that ˙ x
∗
(t ) = u
∗
(t ) = 0 and x
∗
(t ) ≡ x
∗
(0) = 1 for all t in [0, T ].
It remains to determine p(t ). We have ˙ p(t ) = −4e
−2t
x
∗
(t ) = −4e
−2t
, and so p(t ) = 2e
−2t
÷ C,
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
9 CONTROL THEORY: BASI C TECHNI QUES 71
where the constant C is determined by p(T ) = 0, and hence C = −2e
−2T
. It follows that p(t ) =
2(e
−2t
−e
−2T
) = 2e
−2t
(1−e
−2(T −t )
). Since we have assumed p(0) ≤ 1, we must have 2(1−e
−2T
) ≤ 1.
This gives e
−2T
≥ 1/2, and so −2T ≥ ln(1/2) = −ln 2, or T ≤
1
2
ln 2.
Case B: Suppose p(0) > 1. Then there exists a point t
∗
∈ (0, T ) with p(t
∗
) = 1. In this case p(t ) > 1
for t < t
∗
and p(t ) < 1 for t > t
∗
. This gives
u
∗
(t ) =
¸
1 if 0 ≤ t ≤ t
∗
0 if t
∗
< t ≤ 1
Since ˙ x
∗
(t ) = u
∗
(t ) and x
∗
(0) = 1, we get x
∗
(t ) = e
t
if 0 ≤ t ≤ t
∗
and x
∗
(t ) = e
t
∗
if t
∗
≤ t ≤ 1. Thus,
˙ p(t ) = −4e
−2t
x
∗
(t ) = −4e
−t
if 0 ≤ t < t
∗
, and ˙ p(t ) = −4e
t
∗
−2t
if t
∗
< t ≤ 1. Integration gives
p(t ) =
¸
4e
−t
÷C
1
if 0 ≤ t < t
∗
2e
t
∗
−2t
÷C
2
if t
∗
< t ≤ 1
Since p is continuous, both expressions for p(t ) are valid for t = t
∗
. Moreover, p(T ) = 0, so we get the
equations
(i) p(t
∗
) = 1 = 4e
−t
∗
÷C
1
(ii) p(t
∗
) = 1 = 2e
−t
∗
÷C
2
(iii) p(T ) = 0 = 2e
t
∗
−2T
÷C
2
From (i) and (iii) we get C
1
= 1 −4e
−t
∗
and C
2
= −2e
t
∗
−2T
. Hence,
p(t ) =
¸
4(e
−t
−4e
−t
∗
) ÷1 if 0 ≤ t ≤ t
∗
2e
t
∗
(e
−2t
−e
−2T
) if t
∗
≤ t ≤ 1
It remains to determine t
∗
. From (ii) and (iii) we get 1 = 2e
−t
∗
−2e
t
∗
−2T
. Multiplying this equation by
e
t
∗
yields
e
t
∗
= 2 −2e
2t
∗
−2T
(iv)
Multiplying with
1
2
e
2T
and rearranging gives (e
t
∗
)
2
÷
1
2
e
t
∗
e
2T
−e
2T
= 0. This is a second degree equation
for determining e
t
∗
, and we ﬁnd
e
t
∗
= −
1
4
e
2T
÷
1
16
e
4T
÷e
2T
and so t
∗
= ln
1
16
e
4T
÷e
2T
−
1
4
e
2T
(Since e
t
∗
is positive, we must take the positive square root.) This solution makes sense if and only if
t
∗
> 0, i.e. if and only if
1
16
e
4T
÷e
2T
−
1
4
e
2T
> 1. This inequality is equivalent to
1
16
e
4T
÷ e
2T
>
(
1
4
e
2T
÷1)
2
=
1
16
e
4T
÷
1
2
e
2T
÷1, and so e
2T
> 2 ⇐⇒ T >
1
2
ln 2. For the summing up see the book.
(b) We have to consider two cases A and B. Note that in both cases u
∗
(T ) = 0 (and p(T ) = 0), and so
H
∗
(T ) = H(T, x
∗
(T ), u
∗
(T ), p(T )) = 2x
∗
(T )
2
e
−2T
.
Case A: If T ≤
1
2
ln 2, then x
∗
(T ) = 1, H
∗
(T ) = 2e
−2t
, and
V(T ) =
T
0
2(x
∗
(t ))
2
e
−2t
−u
∗
(t )e
t
dt =
T
0
2e
−2t
dt =
T
0
−e
−2t
= 1 −e
−2T
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
72 9 CONTROL THEORY: BASI C TECHNI QUES
This implies V
/
(T ) = 2e
−2T
= H
∗
(T ).
Case B: For T >
1
2
ln 2, we have x
∗
(T ) = x
∗
(t
∗
) = e
t
∗
. This gives H
∗
(T ) = 2e
2t
∗
−2T
. Furthermore,
V(T ) =
T
0
2(x
∗
(t ))
2
e
−2t
−u
∗
(t )e
t
dt
=
t
∗
0
(2 −e
t
) dt ÷
T
t
∗
2e
2t
∗
−2t
dt = 2t
∗
−e
t
∗
÷1 −e
2t
∗
−2T
÷1
Note that t
∗
depends of T . Hence we get
V
/
(T ) = 2
dt
∗
dT
−e
t
∗ dt
∗
dT
−e
2t
∗
−2T
(2
dt
∗
dT
−2) = (2 −e
t
∗
−2e
2t
∗
−2T
)
dt
∗
dT
÷2e
2t
∗
−2T
Equation (iv) shows that 2 −e
t
∗
−2e
2t
∗
−2T
= 0, and therefore
V
/
(T ) = 0 ÷2e
2t
∗
−2T
= H
∗
(T )
as expected. (Alternatively we could have inserted the expressions found for t
∗
and e
t
∗
into the expression
for V(T ), and found V(T ) in terms of T . The algebra required is heavy.)
9.7
9.7.1 (a) The Hamiltonian H = 100 − x −
1
2
u
2
÷ pu is concave in (x, u) and u ∈ (−∞, ∞), so the
following conditions are sufﬁcient for optimality:
(i) (H
/
u
)
∗
= −u
∗
(t ) ÷p(t ) = 0;
(ii) ˙ p(t ) = 1;
(iii) ˙ x
∗
(t ) = u
∗
(t ), x
∗
(0) = x
0
, and x
∗
(1) = x
1
.
From (ii) we get p(t ) = t ÷ A, and (i) and (iii) give ˙ x
∗
(t ) = u
∗
(t ) = p(t ) = t ÷ A. Integrating and
using the two boundary conditions gives B = x
0
, A = x
1
−x
0
−
1
2
, so the solution is as given in the book.
(b) V = V(x
0
, x
1
) =
1
0
(100 −x
∗
(t ) −
1
2
u
∗
(t )
2
) dt
=
1
0
¸
100 −
1
2
t
2
−(x
1
−x
0
−
1
2
)t −x
0
−
1
2
t ÷x
1
−x
0
−
1
2
2
¸
dt .
Differentiating w.r.t. x
0
under the integral sign yields
∂V/∂x
0
=
1
0
(t −1 ÷t ÷x
1
−x
0
−
1
2
) dt =
1
0
(2t −
3
2
÷x
1
−x
0
) dt =
1
0
(t
2
−
3
2
t ÷(x
1
−x
0
)t ) =
x
1
− x
0
−
1
2
= p(0), as it should be. In the same way, ∂V/∂x
1
=
1
0
[−t −
t ÷ x
1
− x
0
−
1
2
)] dt =
1
0
(−2t −x
1
÷x
0
÷
1
2
) dt =
1
0
(−t
2
−x
1
t ÷x
0
t ÷
1
2
t ) = −x
1
÷x
0
−
1
2
= −p(1), as it should be.
9.7.2 (a) The Hamiltonian is H = (1 − s)
√
k ÷ ps
√
k =
√
k ÷
√
k(p − 1)s. (In (b) we use the Arrow
theorem, so we assume p
0
= 1.) The maximum principle (Theorem 9.4.1) gives the conditions:
(i) s = s
∗
(t ) maximizes
√
k
∗
(t ) ÷
√
k
∗
(t )(p(t ) −1)s for s ∈ [0, 1];
(ii) ˙ p(t ) = −H
/
k
(t, k
∗
(t ), s
∗
(t ), p(t )) = −
1
2
√
k
∗
(t )
[1 ÷s
∗
(t )(p(t ) −1)], p(10) = 0;
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
9 CONTROL THEORY: BASI C TECHNI QUES 73
(iii)
˙
k
∗
(t ) = s
∗
(t )
√
k
∗
(t ), k
∗
(0) = 1.
Since s
∗
(t ) ≥ 0, it follows from (iii) that k
∗
(t ) ≥ 1 for all t . Then we see from (i) that
s
∗
(t ) =
¸
1 if p(t ) > 1
0 if p(t ) < 1
(iv)
By studying (ii) and (iv) we see that ˙ p(t ) < 0 for all t . (If p(t ) > 1, then s
∗
(t ) = 1 and from (ii),
˙ p(t ) = −p(t )/2
√
k
∗
(t ) < 0. (What if p(t ) < 1?) Thus p(t ) is strictly decreasing.
Suppose p(t ) < 1for all t in[0, 10]. Thens
∗
(t ) ≡ 0 andk
∗
(t ) ≡ 1. From(ii) this gives ˙ p(t ) = −1/2,
and with p(10) = 0 we get p(t ) = 5−t /2. But then p(t ) > 1 for t < 8, a contradiction. We conclude that
there must exist a t
∗
in (0, 10) such that p(t
∗
) = 1. Then s
∗
(t ) = 1 on [0, t
∗
] and s
∗
(t ) = 0 on (t
∗
, 10].
But then
˙
k
∗
(t ) =
√
k
∗
(t ) on [0, t
∗
] and
˙
k
∗
(t ) = 0 on [0, t
∗
]. By integrating the differential equation for
k
∗
(t ) on [0, t
∗
], we ﬁnd 2
√
k
∗
(t ) = t ÷ C. With k
∗
(0) = 1 we get C = 2, so k
∗
(t ) = (
1
2
t ÷ 1)
2
. On
(t
∗
, 10] we have k
∗
(t ) = (
1
2
t
∗
÷1)
2
since k
∗
(t ) is continuous. Thus,
s
∗
(t ) =
¸
1 if [0, t
∗
]
0 if (t
∗
, 10]
, k
∗
(t ) =
¸
(
1
2
t ÷1)
2
if [0, t
∗
]
(
1
2
t
∗
÷1)
2
if (t
∗
, 10]
(v)
On(t
∗
, 10] we get from(ii) that ˙ p(t ) = −1/2
√
k
∗
(t ) = −1/(t
∗
÷2), andsop(t ) = −t /(t
∗
÷2)÷D. Since
p(10) = 0, this implies that p(t ) = (10−t )/(t
∗
÷2) on (t
∗
, 10]. But p(t
∗
) = 1, so (10−t )/(t
∗
÷2) = 1,
from which it follows that t
∗
= 4. It remains only to ﬁnd p(t ) on [0, t
∗
]. On this interval ˙ p(t ) =
−p(t )/2
√
k
∗
(t ) = −p(t )/(t ÷2). The solution of this separable equation is p(t ) = E/(t ÷2), for some
constant E. But p(4) = 1, so E = 6. We have found the only possible solution. See the answer in the
book.
(b) See the answer in the book.
9.7.3 (a) The Hamiltonian H(t, x, u, p) = e
−βt
√
u ÷ p(αx(t ) − u) is concave in (x, u), so according to
Theorem 9.7.1, the following conditions are sufﬁcient for optimality:
(i) u = u
∗
(t ) maximizes H(t, x
∗
(t ), u, p(t )) = e
−βt
√
u ÷p(t )(αx
∗
(t ) −u) for u ≥ 0;
(ii) ˙ p(t ) = −H
/
x
(t, x
∗
(t ), u
∗
(t ), p(t )) = −αp(t );
(iii) ˙ x
∗
(t ) = αx
∗
(t ) −u
∗
(t ), x
∗
(0) = 1, x
∗
(T ) = 0.
From (ii) we get p(t ) = Ae
−αt
. Deﬁne the function g by g(u) = e
−βt
√
u − Ae
−αt
u. Then g
/
(u) =
e
−βt
(1/2
√
u) − Ae
−αt
, and we see that g
/
(u) > 0 when u is slightly larger than 0. This means that
u
∗
(t ) = 0 cannot maximize g, or the Hamiltonian, for any t . Hence u
∗
(t ) > 0 and g
/
(u
∗
(t )) = 0, so
e
−βt
2
√
u
∗
(t )
= Ae
−αt
or u
∗
(t ) =
1
4A
2
e
2(α−β)t
From (iii) we get the following linear differential equation for x
∗
(t ):
˙ x
∗
(t ) = αx
∗
(t ) −
1
4A
2
e
2(α−β)t
with solution x
∗
(t ) = Ce
αt
−
1
4A
2
(α −2β)
e
2(α−β)t
The two constants A and C are determined by the boundary conditions in (iii). The explicit expressions
for x
∗
(t ), u
∗
(t ), and p(t ) can be found in the answer in the book. (There is a misprint in the formula for
x
∗
(t ): the ﬁrst fraction must be multiplied by e
αt
.)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
74 9 CONTROL THEORY: BASI C TECHNI QUES
(b) If x(T ) = 0 is replaced by x(T ) ≥ 0, the only change in the conditions (i)–(iii) is that (ii) is
augmented by p(T ) ≥ 0, with p(T ) = 0 if x
∗
(T ) > 0. Again p(t ) = Ae
−αt
, so p(T ) = 0 would imply
Ae
−αT
= 0, so A = 0 and p(t ) ≡ 0. The maximum condition (i) would then imply that u
∗
(t ) for u ≥ 0
would maximize e
−βt
√
u, which has no maximum. Hence x
∗
(T ) = 0, and the solution is as in (a).
9.8
9.8.1 (b) In this problem we use Theorem 9.8.1 and we try to ﬁnd the only possible solution to the problem.
The Hamiltonian H = p
0
(−9 −
1
4
u
2
) ÷ p(t )u is concave in (x, u), but this does not guarantee that a
solution to the necessary conditions is optimal. (See Note 9.8.1.)
Suppose (x
∗
(t ), u
∗
(t )) is an optimal pair deﬁned on [0, t
∗
]. Then there exists a continuous function
p(t ) and a number p
0
, which is either 0 or 1, such that for all t in [0, t
∗
] we have (p
0
, p(t )) ,= (0, 0) and
(i) u = u
∗
(t ) maximizes H(t, x
∗
(t ), u, p(t )) = p
0
(−9 −
1
4
u
2
) ÷p(t )u for u ∈ ޒ;
(ii) ˙ p(t ) = −H
/
x
(t, x
∗
(t ), u
∗
(t ), p(t )) = 0;
(iii) ˙ x
∗
(t ) = u
∗
(t ), x
∗
(0) = 0, x
∗
(T ) = 16;
(iv) H(t
∗
, x
∗
(t
∗
), u
∗
(t
∗
), p(t
∗
)) = p
0
(−9 −
1
4
u
∗
(t
∗
)
2
) ÷p(t )u
∗
(t
∗
) = 0.
Since H is concave in (x, u), condition (i) is equivalent to (H
/
u
)
∗
= −p
0
1
2
u
∗
(t ) ÷p(t ) = 0. Thenp
0
= 0
implies p(t ) = 0, which contradicts (p
0
, p(t )) ,= (0, 0). Hence p
0
= 1, and u
∗
(t ) = 2p(t ). From (ii)
we have p(t ) = ¨ p for some constant ¨ p, and so u
∗
(t ) = 2 ¨ p. Moreover, ˙ x
∗
(t ) = u
∗
(t ) = 2 ¨ p. Integrating
and using x
∗
(0) = 0 this gives x
∗
(t ) = 2 ¨ pt , and x
∗
(t
∗
) = 16 yields ¨ p = 8/t
∗
. Finally, (iv) implies
−9 −
1
4
(2 ¨ p)
2
÷ ¨ p2 ¨ p = 0, or ¨ p
2
= 9. Here ¨ p = −3 gives t
∗
= 8/ ¨ p < 0. So the only possible solution
is ¨ p = 3, Then t
∗
= 8/3, u
∗
(t ) = 6, and x
∗
(t ) = 6t .
9.8.2 Consider the case B ≥ aT
2
/4b. Then fromProblem9.4.7, u
∗
(t ) = a(2t −T )/4b÷B/T and x
∗
(t ) =
at (t −T )/4b ÷Bt /T and p(t ) = 2bu
∗
(t ). To determine the optimal T
∗
, the condition is H(T
∗
) = 0,
or ax
∗
(T
∗
) ÷ b(u
∗
(T
∗
))
2
= p(T
∗
)u
∗
(T
∗
) = 2b(u
∗
(T
∗
))
2
. This reduces to aB/b = (u
∗
(T
∗
))
2
, or
u
∗
(T
∗
) =
√
aB/b, that is aT
∗
/4b ÷ B/T
∗
=
√
aB/b. Solving this equation for T
∗
gives the unique
solution T
∗
= 2
√
bB/a. (Note that this is the positive solution T of B = aT
2
/4b.)
If B < aT
2
/4b, we ﬁnd that the equation H
∗
(T
∗
) = 0 does not determine T
∗
. (To save on storage
costs, the ﬁrm waits until t = T −2
√
Bb/a to start production, and then produces at an optimal rate until
delivery time T . Note that no discounting is assumed, so waiting costs nothing.)
9.9
9.9.2 The current value Hamiltonian H
c
= 10u − u
2
− 2 − λu is concave in (x, u), so the following
conditions are sufﬁcient for optimality:
(i) u = u
∗
(t ) maximizes H
c
(t, x
∗
(t ), u, λ(t )) = 10u −u
2
−2 −λ(t )u for u ≥ 0;
(ii)
˙
λ(t ) −0.1λ = −(H
c
)
/
x
(t, x
∗
(t ), u
∗
(t ), λ(t )) = 0;
(iii) λ(5) ≥ 0, with λ(5) = 0 if x
∗
(5) > 0;
(iv) ˙ x
∗
(t ) = −u
∗
(t ), x
∗
(0) = 10, x
∗
(5) ≥ 0.
From (ii) we have λ(t ) = Ae
0.1t
for some constant A. Suppose u
∗
(t ) > 0 for all t in [0, 5]. Then (i) is
equivalent to (H
c
)
/
u
(t, x
∗
(t ), u
∗
(t ), λ(t )) = 10 −2u
∗
(t ) −Ae
0.1t
= 0, that is,
(v) u
∗
(t ) = 5 −
1
2
Ae
0.1t
.
Then ˙ x
∗
(t ) = −5 ÷
1
2
Ae
0.1t
, and integration gives x
∗
(t ) = −5t ÷ 5Ae
0.1t
÷ B, where B is a new
constant. The initial condition x
∗
(0) = 10 yields B = 10 − 5A, so x
∗
(t ) = 5A(e
0.1t
− 1) − 5t ÷ 10.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
9 CONTROL THEORY: BASI C TECHNI QUES 75
Next look at condition (iii) which reduces to A ≥ 0, with A = 0 if x
∗
(5) > 0. But A = 0 would imply
x
∗
(t ) = −5t ÷10, which for t = 5 yields the contradiction x
∗
(5) < 0. Thus A > 0 and x
∗
(5) = 0. But
then 0 = 5A(e
0.5
−1)−15, so A = 3/(e
0.5
−1), and the same solutions for x
∗
(t ) and u
∗
(t ) as in Problem
9.4.6 are found. Since all the conditions (i)–(iv) are satisﬁed by the admissible pair (x
∗
(t ), u
∗
(t )), with
the given λ(t ), we have found the solution.
9.9.3 The current value Hamiltonian H
c
= −2u−u
2
÷λu is concave in (x, u), and u ∈ ޒ, so the following
conditions are sufﬁcient for optimality:
(i) (H
c
)
/
u
= −2 −2u
∗
(t ) ÷λ = 0;
(ii)
˙
λ(t ) −0.1λ = −(H
c
)
/
x
(t, x
∗
(t ), u
∗
(t ), λ(t )) = 0;
(iii) ˙ x
∗
(t ) = u
∗
(t ), x
∗
(0) = 1, x
∗
(1) = 0.
From(ii) we have λ(t ) = Ae
0.1t
for some constant A. From(ii) we have u
∗
(t ) =
1
2
λ(t )−1 =
1
2
Ae
0.1t
−1.
Then ˙ x
∗
(t ) =
1
2
Ae
0.1t
−1, with solution x
∗
(t ) = 5Ae
0.1t
−t ÷B. The constants Aand B are determined
from the boundary conditions, and we get the same solution as in Problem 9.5.3.
9.10
9.10.2 With Aas the state variable, the Hamiltonian is H(t, A, u, p) = U(rA(t ) ÷w−u(t ))e
−ρt
÷pu, and
the scrap value function is S(T, A) = e
−ρT
ϕ(A). (We assume that λ
0
= 1.) Moreover, w is a constant.
With the assumptions in Example 8.5.3, the utility function U has U
/
> 0 and U
//
< 0. This implies
that H(t, A, u, p) is concave in (A, u). Since ϕ(A) is also concave, a set of sufﬁcient conditions for
optimality is then (assuming interior solution in (i)):
(i) H
/
3
(t, A
∗
(t ), u
∗
(t ), p(t )) = 0, or p(t ) = U
/
(rA
∗
(t ) ÷w −u
∗
(t ))e
−ρt
;
(ii) ˙ p = −H
/
2
(t, A
∗
(t ), u
∗
(t ), p(t )) = −rU
/
(rA
∗
(t ) ÷w −u
∗
(t ))e
−ρt
;
(iii) p(T ) = S
/
A
(T, A
∗
(T )) = e
−ρT
ϕ
/
(A
∗
(T ));
(iv)
˙
A
∗
(t ) = u
∗
(t ), A
∗
(0) = A
0
.
In the answer in the book we go a little further. We differentiate the expression for p(t ) in (i) w.r.t. t and
equate it to the expression for ˙ p(t ) in (ii). Using u
∗
(t ) =
˙
A
∗
(t ), this gives
U
//
(rA
∗
÷w −
˙
A
∗
)(r
˙
A
∗
−
´
A
∗
)e
−ρt
−ρU
/
(rA
∗
÷w −
˙
A
∗
)e
−ρt
= −rU
/
(rA
∗
÷w −
˙
A
∗
)e
−ρt
Multiplying by −e
ρt
and rearranging gives
´
A
∗
−r
˙
A
∗
÷(ρ −r)U
/
/U
//
= 0. Combining (i) and (iii), we
also see that ϕ
/
(A
∗
(T )) = U
/
(rA
∗
(T ) ÷w −u
∗
(T )).
9.10.3 Compared with Problem 9.7.2 the only difference is that instead of the condition “x(10) free”, we
have nowincluded a scrap value in the objective function. The scrap value function is S(k) = 10
√
k, with
S
/
(k) = 5/
√
k. Conditions (i)–(iv) in the answer to Problem 9.7.2 are still valid except that p(10) = 0
in (ii) is now replaced by
p(10) =
5
√
k
∗
(10)
(ii)
/
Again p(t ) is strictly decreasing, and p(t ) < 1 for all t in [0, 10] is again easily seen to be impossible.
Suppose p(t ) > 1 for all t in [0, 10]. (This was not an option when we required p(10) = 0.) Then
s
∗
(t ) ≡ 1 and
˙
k
∗
(t ) =
√
k
∗
(t ), with k
∗
(0) = 1. It follows that 2
√
k
∗
(t ) = t ÷2, or k
∗
(t ) = (
1
2
t ÷1)
2
. In
particular, k
∗
(10) = 36. Then (ii)
/
gives p(10) = 5/6 < 1 a contradiction. We conclude that there must
exist a t
∗
in (0, 10) such that p(t
∗
) = 1, and (v) in the answer to Problem 9.7.2 is still valid (although
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
76 9 CONTROL THEORY: BASI C TECHNI QUES
with a different t
∗
).
On (t
∗
, 10] we again get p(t ) = −t /(t
∗
÷ 2) ÷ D and 1 = p(t
∗
) = −t
∗
/(t
∗
÷ 2), which implies
that D = (2t
∗
÷2)/(t
∗
÷2), and thus p(t ) = −t /(t
∗
÷2) ÷(2t
∗
÷2)/(t
∗
÷2). In particular, p(10) =
(2t
∗
−8)/(t
∗
÷2). We now use (ii)
/
to determine t
∗
:
2t
∗
−8
t
∗
÷2
=
5
1
2
t
∗
÷1
=
10
t
∗
÷2
from which it follows that t
∗
= 9. The rest is routine, see the answer in the book. (Again the Arrow
condition is satisﬁed, and it is valid also in this case.)
9.10.4 (a) This is a problemwithscrapvalue functionS(x) =
1
2
x. We useTheorem9.10.1. Let (x
∗
(t ), u
∗
(t ))
be an admissible pair. With p
0
= 1, the Hamiltonian is H(t, x, u, p) = x−u÷pu = x÷(p−1)u. (Note
that the scrap value function is not to be included in the Hamiltonian.) If (x
∗
(t ), u
∗
(t )) is an optimal pair
in the problem, and p(t ) is the adjoint function, then according to (B) in Theorem 9.10.1,
˙ p(t ) = −
∂H
∗
∂x
= −1 and p(1) = S
/
(x
∗
(1)) =
1
2
(i)
Moreover, u
∗
(t ) = 1 if p(t ) > 1 and u
∗
(t ) = 0 if p(t ) < 1. From (i) we get
p(t ) =
3
2
−t
We see that the strictly decreasing function p(t ) is equal to 1 at t = 1/2. Hence
u
∗
(t ) =
¸
1 if t ∈ [0, 1/2]
0 if t ∈ (1/2, 1]
Since ˙ x
∗
(t ) = u
∗
(t ) and x(0) = 1/2, we get
x
∗
(t ) =
¸
t ÷1/2 if t ∈ [0, 1/2]
1 if t ∈ (1/2, 1]
We have found the optimal solution because H(t, x, u, p(t )) and S(x) = x/2 are concave (in fact linear)
functions of (x, u).
(b) The scrap value function is now
¨
S(x) = −
1
4
(x −2)
2
, but the Hamiltonian is as in (a). Condition (i)
is replaced by
˙ p(t ) = −1 and p(1) =
¨
S
/
(x
∗
(1)) = −
1
2
(x
∗
(1) −2) = 1 −
1
2
x
∗
(1) (ii)
From (ii) we see that p(t ) = −t ÷C and since p(1) = 1 −
1
2
x
∗
(1), we have
p(t ) = −t ÷2 −
1
2
x
∗
(1) (iii)
Since x
∗
(0) = 1/2 and 0 ≤ ˙ x
∗
≤ 1, it is clear that 1/2 ≤ x
∗
(1) ≤ 3/2. Condition (iii) therefore gives
1/4 ≤ p(1) ≤ 3/4
Since ˙ p(t ) = −1 for all t , we have p(t ) = p(0) −t . Hence, p(0) = p(1) ÷1 and
5/4 ≤ p(0) ≤ 7/4
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
9 CONTROL THEORY: BASI C TECHNI QUES 77
Condition (ii) is still valid, so there exists a unique number t
∗
between 0 and 1 such that p(t
∗
) = 1, and
u
∗
(t ) =
¸
1 if t ∈ [0, t
∗
],
0 if t ∈ (t
∗
, 1],
x
∗
(t ) =
¸
1/2 ÷t if t ∈ [0, t
∗
]
1/2 ÷t
∗
if t ∈ (t
∗
, 1]
To determine t
∗
we use the transversality condition. We know that p(t ) = p(0) − t = p(1) ÷ 1 − t ,
so p(1) = p(t ) ÷ t − 1 for all t . Since p(t
∗
) = 1, we get p(1) = t
∗
. The transversality condition then
gives t
∗
= p(1) = 1 −
1
2
x
∗
(1) = 1 −
1
2
1
2
÷t
∗
=
3
4
−
1
2
t
∗
, which means that t
∗
=
1
2
. The optimal pair
(x
∗
(t ), u
∗
(t )) is therefore exactly as in (a), but this is accidental. A different scrap value function would
usually have given another solution.
9.10.5 (a) The Hamiltonian H(t, x, u, p) = −u
2
− px ÷ pu is concave in (x, u) and the scrap value
function S(x) = −x
2
is concave in x. The following conditions are therefore sufﬁcient for optimality:
(i) H
/
3
(t, x
∗
(t ), u
∗
(t ), p) = −2u
∗
(t ) ÷p(t ) = 0, so u
∗
(t ) =
1
2
p(t );
(ii) ˙ p(t ) = −∂H
∗
/∂x = p(t ) and p(T ) = −S
/
(x
∗
(T )) = −2x
∗
(T );
(iii) ˙ x
∗
(t ) = −x
∗
(t ) ÷u
∗
(t ), x
∗
(0) = x
0
.
From (ii) we get p(t ) = Ce
t
for an appropriate constant C, and since p(T ) = −2x
∗
(T ), we get
p(t ) = −2x
∗
(T )e
t −T
and also u
∗
(t ) =
1
2
p(t ) = −x
∗
(T )e
t −T
. The differential equation
˙ x
∗
(t ) = −x
∗
(t ) ÷u
∗
(t ) = −x
∗
(t ) −x
∗
(T )e
t −T
has the general solution
x
∗
(t ) = De
−t
−
1
2
x
∗
(T )e
t −T
From x
∗
(T ) = De
−T
−
1
2
x
∗
(T ) we get D =
3
2
x
∗
(T )e
T
. The initial condition x
∗
(0) = x
0
gives
D −
1
2
x
∗
(T )e
−T
=
3
2
x
∗
(T )e
T
−
1
2
x
∗
(T )e
−T
=
1
2
x
∗
(T )(3e
T
−e
−T
) = x
0
so
x
∗
(T ) =
2x
0
3e
T
−e
−T
=
2x
0
e
T
3e
2T
−1
and D =
3
2
x
∗
(T )e
T
=
3x
0
e
2T
3e
2T
−1
Thus the optimal solution is
u
∗
(t ) = −
2x
0
e
t
3e
2T
−1
, x
∗
(t ) =
x
0
(3e
2T −t
−e
t
)
3e
2T
−1
, p(t ) = −
4x
0
e
t
3e
2T
−1
(b) We have
V(x
0
, T ) = −
T
0
(u
∗
(t ))
2
dt −(x
∗
(T ))
2
= −
4x
2
0
(3e
2T
−1)
2
¸
T
0
e
2t
dt ÷e
2T
¸
= −
4x
2
0
(3e
2T
−1)
2
¸
T
0
1
2
e
2t
÷e
2T
¸
= −
4x
2
0
(3e
2T
−1)
2
¸
3
2
e
2T
−
1
2
¸
= −
2x
2
0
3e
2T
−1
It follows that
∂V
∂x
0
= −
4x
0
3e
2T
−1
= p(0) and
∂V
∂T
=
12x
2
0
e
2T
(3e
2T
−1)
2
We see from the solutions in (a) that p(T ) = −2x
∗
(T ) and u
∗
(t ) = −x
∗
(T ), so H
∗
(T ) = −u
∗
(T )
2
÷
p(T )(−x
∗
(T ) ÷u
∗
(T )) = 3x
∗
(T )
2
= ∂V/∂T .
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
78 9 CONTROL THEORY: BASI C TECHNI QUES
9.10.6 The current value Hamiltonian H
c
(t, x, u, λ) = −(x −u)
2
÷λ(u−x ÷a) is concave in (x, u). The
scrap value function S(x) = −x
2
is concave in x. The following conditions are therefore sufﬁcient for
an admissible pair (x
∗
, u
∗
) to solve the problem:
(i) ∂(H
c
)
∗
/∂u = 0 , i.e. λ(t ) = −2(x
∗
(t ) −u
∗
(t ));
(ii)
˙
λ(t ) −rλ = −∂(H
c
)
∗
/∂x = 2(x
∗
(t ) −u
∗
(t )) ÷λ(t ), and λ(T ) = −2x
∗
(T );
(iii) ˙ x
∗
(t ) = u
∗
(t ) −x
∗
(t ) ÷a, x
∗
(0) = 0.
From (i) and (ii) it follows that
˙
λ(t ) = rλ, so λ(t ) = Ae
rt
, where A is a constant. Then (i) yields
x
∗
(t )−u
∗
(t ) = −
1
2
Ae
rt
. But then ˙ x
∗
(t ) = u
∗
(t )−x
∗
(t )÷a =
1
2
Ae
rt
÷a, so x
∗
(t ) = (A/2r)e
rt
÷at ÷B.
The initial condition x
∗
(0) = 0 gives B = −A/2r, so
x
∗
(t ) =
A
2r
(e
rt
−1) ÷at (iv)
From(ii), x
∗
(T ) = −
1
2
λ(T ) = −
1
2
Ae
rT
, and so (iv) with t = T yields −
1
2
Ae
rT
= (A/2r)(e
rT
−1)÷aT .
Solving for A yields
A = −
2arT
e
rT
(1 ÷r) −1
The expressions for λ(t ), x
∗
(t ), and u
∗
(t ) follow easily. See the answer in the book.
9.11
9.11.1 The current value Hamiltonian H
c
= ln u − λ(0.1x − u) is concave in (x, u), and u > 0, so the
following conditions are sufﬁcient for optimality:
(i) (H
c
)
/
u
(t, x
∗
, u
∗
, λ(t )) = 1/u
∗
(t ) ÷λ = 0;
(ii)
˙
λ(t ) −0.2λ = −(H
c
)
/
x
(t, x
∗
, u
∗
, λ(t )) = −0.1λ;
(iii) (a) lim
t →∞
λ(t )e
−0.2t
(−x
∗
(t )) ≥ 0;
(b) There exists a number M such that [λ(t )e
−0.2t
[ ≤ M for all t ≥ 0;
(c) There exists a number t
/
such that λ(t ) ≥ 0 for all t ≥ t
/
;
(iv) ˙ x
∗
(t ) = 0.1x
∗
(t ) −u
∗
(t ), x
∗
(0) = 10, lim
t →∞
x
∗
(t ) ≥ 0.
From (ii) we get
˙
λ(t ) = 0.1λ(t ), and thus λ(t ) = Ae
0.1t
, for some constant A. Condition (i) yields
u
∗
(t ) = 1/λ(t ) = e
−0.1t
/A. Then from (iv), ˙ x
∗
(t ) = 0.1x
∗
(t ) − e
−0.1t
/A. The solution of this linear
differential equation, with x
∗
(0) = 10, is easily seen to be x
∗
(t ) = (10 − 5/A)e
0.1t
÷ 5e
−0.1t
/A.
Condition (iv) requires lim
t →∞
x
∗
(t ) = lim
t →∞
[(10 − 5/A)e
0.1t
÷ 5e
−0.1t
/A] ≥ 0. This obviously
requires 10 − 5/A ≥ 0. From (iii)(c) we see that we must have A ≥ 0, and so A ≥ 1/2. However,
referring to (iii)(a),
λ(t )e
−0.2t
(−x
∗
(t )) = Ae
0.1t
e
−0.2t
(−(10 −5/A)e
0.1t
−5e
−0.1t
/A) = −(10A −5) −5e
−0.2t
The limit of this expression can be ≥ 0 only if 10A−5 ≤ 0, i.e. A ≤ 1/2. Thus we must have A = 1/2.
Then u
∗
(t ) = 2e
−0.1t
, x
∗
(t ) = 10e
−0.1t
, with λ(t ) =
1
2
e
0.1t
.
It remains to check that the conditions in (iii) are satisﬁed. First, (iii)(a) is satisﬁed because we
have lim
t →∞
1
2
e
0.1t
e
−0.2t
(−10e
−0.1t
) = −5 lim
t →∞
e
−0.2t
= 0. Since [λ(t )e
−0.2t
[ = [
1
2
e
0.1t
e
−0.2t
[ =
1
2
e
−0.1t
≤
1
2
for all t ≥ 0, (iii)(b) is also satisﬁed. Finally, λ(t ) =
1
2
e
0.1t
≥ 0 for all t ≥ 0. We have
therefore found the solution.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
9 CONTROL THEORY: BASI C TECHNI QUES 79
9.11.2 The current value Hamiltonian (with λ
0
= 1) is H
c
= x(2 − u) ÷ λuxe
−t
, and the following
conditions must be satisﬁed for (x
∗
(t ), u
∗
(t )) to be optimal:
(i) u = u
∗
(t ) maximizes H
c
(t, x
∗
(t ), u, λ(t )) = 2x
∗
(t ) ÷x
∗
(t )(e
−t
λ(t ) −1)u for u ∈ [0, 1];
(ii)
˙
λ(t ) −λ(t ) = −∂(H
c
)
∗
/∂x = −(2 −u
∗
(t )) −λ(t )u
∗
(t )e
−t
;
(iii) ˙ x
∗
(t ) = u
∗
(t )x
∗
(t )e
−t
, x
∗
(0) = 1.
Since ˙ x
∗
(t ) ≥ 0 and x
∗
(0) = 1, x
∗
(t ) ≥ 1 for all t . Looking at (i) we see that
u
∗
(t ) =
¸
1 if λ(t )e
−t
> 1
0 if λ(t )e
−t
< 1
We guess that p(t ) = λ(t )e
−t
→ 0 as t → ∞. Then λ(t )e
−t
< 1 on some maximal interval (t
∗
, ∞).
On this interval u
∗
(t ) = 0 and
˙
λ(t ) −λ(t ) = −2, and so λ(t ) = Ce
t
÷2, or λ(t )e
−t
= C ÷2e
−t
, which
tends to 0 as t → ∞only if C = 0. Then λ(t ) = 2. Note that λ(t )e
−t
∗
= 2e
−t
∗
= 1 when t
∗
= ln 2,
and we propose that u
∗
(t ) = 1 on [0, ln 2], u
∗
(t ) = 0 on (ln 2, ∞). Then ˙ x
∗
(t ) = x
∗
(t )e
−t
on [0, ln 2],
˙ x
∗
(t ) = 0 on (ln 2, ∞). On [0, ln 2],
dx
∗
(t )/x
∗
(t ) =
e
−t
dt , so ln x
∗
(t ) = −e
−t
÷ A, and with
x
∗
(0) = 1 this gives x
∗
(t ) = e
1−e
−t
. On (t
∗
, ∞) we have x
∗
(t ) = e
1−e
−ln 2
= e
1−1/2
= e
1/2
.
On [0, ln 2] we have
˙
λ(t ) ÷ (e
−t
− 1)λ(t ) = −1. We use formula (5.4.6) with a(t ) = e
−t
− 1 and
b(t ) = −1. Then
a(t ) dt =
(e
−t
−1) dt = −e
−t
−t and
λ(t ) = Ce
e
−t
÷t
÷e
e
−t
÷t
e
−e
−t
−t
(−1) dt = Ce
e
−t
÷t
−e
e
−t
÷t
e
−e
−t
e
−t
dt = Ce
e
−t
÷t
−e
t
because the last integral is obviously equal to e
−e
−t
. Since λ(ln 2) = 2 we ﬁnd that C = 2e
−1/2
. We have
obtained the same answers as in the book, keeping in mind that p(t ) = λ(t )e
−t
.
It does not change the problem if x(∞) ≥ 0 is added as a restriction in the problem. Then (B) and
(C) in Note 9.11.3 are trivially satisﬁed, and the expression in (A) reduces to lim
t →∞
2e
−t
(0 − e
1/2
),
which is clearly 0. The Arrow concavity condition holds in this problem and this yields optimality also
in the inﬁnite horizon case.
9.11.4 The current value Hamiltonian H
c
= (x − u) ÷ λue
−t
is concave in (x, u). The problem is not
affected by introducing the requirement that lim
t →∞
x
∗
(t ) ≥ 0. So the following conditions are sufﬁcient
for optimality:
(i) u = u
∗
(t ) maximizes x
∗
(t ) ÷e
−t
(λ(t ) −e
t
)u for u ∈ [0, 1];
(ii)
˙
λ(t ) −λ(t ) = −(H
c
)
/
x
(t, x
∗
, u
∗
, λ(t )) = −1;
(iii) (a) lim
t →∞
λ(t )e
−t
(−x
∗
(t )) ≥ 0;
(b) There exists a number M s.t. [λ(t )e
−t
[ ≤ M for all t ≥ 0;
(c) There exists a number t
/
s.t. λ(t ) ≥ 0 for all t ≥ t
/
;
(iv) ˙ x
∗
(t ) = u
∗
(t )e
−t
, x
∗
(−1) = 0.
From (ii) it follows that λ(t ) = Ae
t
÷1. We guess that p(t ) = e
−t
λ(t ) = e
−t
(Ae
t
÷1) = A÷e
−t
→0
as t → ∞, so A = 0. From (i) we see that u
∗
(t ) = 1 if e
t
< 1, and u
∗
(t ) = 0 if e
t
> 1. It follows that
u
∗
(t ) = 1 in [−1, 0] and u
∗
(t ) = 0 in (0, ∞). Then from (iv), we get x
∗
(t ) = e − e
−t
in [−1, 0] and
x
∗
(t ) = e − 1 in (0, ∞). The conditions in (iii) are obviously satisﬁed, so we have found the optimal
solution. (The answer in the book is wrong.)
9.12
9.12.1 (a) The current value Hamiltonian H
c
= ax −
1
2
u
2
÷ λ(−bx ÷ u) is concave in (x, u), so the
following conditions are sufﬁcient for optimality:
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
80 9 CONTROL THEORY: BASI C TECHNI QUES
(i) (H
c
)
/
u
(t, x
∗
(t ), u
∗
(t ), λ(t )) = −u
∗
(t ) ÷λ(t ) = 0;
(ii)
˙
λ(t ) −rλ(t ) = −(H
c
)
/
x
(t, x
∗
(t ), u
∗
(t ), λ(t )) = −a ÷bλ(t );
(iii) lim
t →∞
λ(t )e
−rt
(x(t ) −x
∗
(t )) ≥ 0 for all admissible x(t );
(iv) ˙ x
∗
(t ) = −bx
∗
(t ) ÷u
∗
(t ), x
∗
(0) = 10.
From (i) we get u
∗
(t ) = λ(t ), and thus ˙ x
∗
(t ) = −bx
∗
(t ) ÷λ(t ), and x = x
∗
and λ must satisfy
˙ x = −bx ÷λ = F(x, λ),
˙
λ = (b ÷r)λ −a = G(x, λ) (∗)
The equilibrium point is ( ¨ x,
¨
λ) = (a/b(b ÷r), a/(b ÷r).
(b) The Jacobian matrix of (∗) is
F
/
x
F
/
λ
G
/
x
G
/
λ
=
−b 1
0 b ÷r
, with eigenvalues −b and b ÷r. Since
they are real and of opposite signs, the equilibrium point is a saddle point. To prove sufﬁciency, we can
restrict attention to admissible x(t ) satisfying lim
t →∞
x(t )e
−rt
≥ 0, because if the limit is < 0, then the
value of the criterion is −∞. For such x(t ), condition (iii) is evidently satisﬁed.
(c) V =
∞
0
[ax
∗
(t ) −
1
2
(u
∗
(t ))
2
]e
−rt
dt =
∞
0
ax
0
e
−(b÷r)t
dt ÷terms that do not depend on x
0
. But
then ∂V/∂x
0
=
∞
0
ae
−(b÷r)t
dt = a/(b ÷r) = λ(0).
9.12.2 From the answer to Problem 9.9.1 with x(0) = 1, we ﬁnd x
∗
= Ae
(1÷
√
2 )t
÷(1 −A)e
(1−
√
2 )t
. and
λ = A
√
2 e
(
√
2÷1)t
− (1 − A)
√
2 e
(1−
√
2 )t
. The adjoint variable p(t ) = λ(t )e
−2t
tends to 0 as t → ∞
if and only if A = 0. Then x
∗
= e
(1−
√
2 )t
and λ(t ) = −
√
2 e
(1−
√
2 )t
(with p(t ) = −
√
2 e
(−1−
√
2 )t
). To
prove sufﬁciency, note that if lim
t →∞
x
2
(t )e
−2t
= lim
t →∞
(x(t )e
−t
)
2
,= 0, then the objective function
is −∞. We can therefore restrict attention to admissible solutions for which lim
t →∞
x(t )e
−t
= 0. Then
condition (iii) is satisﬁed and we have found the optimal solution.
9.12.3 (a) With H
c
= ln C ÷λ(AK
α
−C), ∂(H
c
)
∗
/∂C = 0 implies 1/C
∗
−λ = 0, or C
∗
λ = 1. Taking
ln of each side and differentiating w.r.t. t yields
˙
C
∗
/C
∗
÷
˙
λ/λ = 0. Also,
˙
λ − rλ = −∂(H
c
)
∗
/∂K =
−λαA(K
∗
)
α−1
or, equivalently,
˙
λ/λ = r − αA(K
∗
)
α−1
. It follows that if K = K
∗
(t ) > 0 and C =
C
∗
(t ) > 0 solve the problem, then the second equation in (∗) holds. (The ﬁrst equation is part of the
problem.)
(b) The Jacobian matrix evaluated at (400, 40) is J =
1/20 −1
−1/400 0
and [J[ = −1/400 < 0, so the
equilibrium point is a saddle point.
(c) If K
0
= 100 and T = ∞the solution curve converges towards the equilibrium point. For sufﬁcient
conditions, see Note 9.11.3.
9.12.4 (a) The current value Hamiltonian H
c
= −(x−1)
2
−
1
2
u
2
÷λ(x−u) is concave in (x, u). Since u ∈ ޒ,
the maximumconditionreduces to (i) (H
c
)
/
u
(t, x
∗
, u
∗
, λ(t )) = −u
∗
(t )−λ = 0. The differential equation
for λ is (ii)
˙
λ(t ) −λ(t ) = −(H
c
)
/
x
(t, x
∗
, u
∗
, λ(t )) = 2x
∗
(t ) −2 −λ. Of course, ˙ x
∗
(t ) = x
∗
(t ) −u
∗
(t ).
It follows that the optimal pair x
∗
(t ), u
∗
(t )) must satisfy
˙ x = F(x, λ) = x −u = x ÷λ
˙
λ = G(x, λ) = 2x −2
(∗)
The Jacobian matrix of (∗) is
F
/
x
F
/
λ
G
/
x
G
/
λ
=
1 1
2 0
. The determinant is −2, so the equilibrium point
(1, −1) is a saddle point. The eigenvalues are −1 and 2.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
10 CONTROL THEORY WI TH MANY VARI ABL ES 81
(b) Solving the ﬁrst equation in (∗) for λ yields λ = ˙ x −x, and then
˙
λ = ´ x − ˙ x. Inserted into the second
equation in (∗) yields ´ x − ˙ x −2x = −2. The general solution of this equation is x(t ) = Ae
−t
÷Be
2t
÷1,
and x(0) = 1/2 yields B = −A −1/2, so
x(t ) = Ae
−t
−(A ÷1/2)e
2t
÷1
Using λ = ˙ x −x, we ﬁnd the corresponding solution for λ,
λ(t ) = −2Ae
−t
−(A ÷1/2)e
2t
−1
If A ,= −1/2 we see that x(t ) and λ(t ) both diverge as t → ∞. For A = −1/2, it follows that
x(t ) = −(1/2)e
−t
÷1 →1 and λ(t ) = e
−t
−1 →−1 as t →∞. Note that fromx(t ) = −(1/2)e
−t
÷1
and λ(t ) = e
−t
−1 we ﬁnd that (x, λ) = (x(t ), λ(t )) satisﬁes λ = −2x ÷1.
It remains to prove that condition (d) in Theorem 9.11.1 is satisﬁed. Note that the integrand in the
objective function can be written (−x
2
÷ 2x − 1 −
1
2
u
2
)e
−t
, so we need only consider admissible x(t )
for which lim
t →∞
x(t )e
−t
≥ 0, because if lim
t →∞
x(t )e
−t
< 0 the objective function is −∞.
10 Control Theory with Many Variables
10.1
10.1.3 To prove that g is concave, let x
1
, x
2
, λ ∈ (0, 1) and choose u
1
, u
2
in U
x
such that g(x
1
) = F(x
1
, u
1
),
g(x
2
) = F(x
2
, u
2
). Then g(λx
1
÷(1 −λ)x
2
) ≥ F(λx
1
÷(1 −λ)x
2
, λu
1
÷(1 −λ)u
2
) ≥ λF(x
1
, u
1
) ÷
(1 −λ)F(x
2
, u
2
) = λg(x
1
) ÷(1 −λ)g(x
2
). Note that λu
1
÷(1 −λ)u
2
∈ U
λx
1
÷(1−λ)x
2
.
10.2
10.2.2 Suppose T < 2/a. If t ∈ [0, T − 2/a], then u
∗
(t ) = 1 and then ˙ x
∗
1
(t ) = ax
∗
1
(t ) with x
∗
1
(0) = x
0
1
.
If follows that x
∗
1
(t ) = x
0
1
e
at
. In particular, x
∗
1
(T − 2/a) = x
0
1
e
aT −2
. Moreover, we have ˙ x
∗
2
(t ) =
a(1 −u
∗
(t ))x
∗
1
(t ) = 0, and so x
∗
2
(t ) = x
0
2
.
If t ∈ (T −2/a, T ], then u
∗
(t ) = 0 and ˙ x
∗
1
(t ) = 0 with x
∗
1
(T −2/a) = x
0
1
e
aT −2
, so x
∗
1
(t ) = x
0
1
e
aT −2
.
Moreover, ˙ x
∗
2
(t ) = ax
0
1
e
aT −2
, so integration gives x
∗
2
(t ) = ax
0
1
e
aT −2
t ÷B, with the boundary condition
x
∗
2
(T −2/a) = x
0
2
determining B. In fact, we get x
∗
2
(t ) = x
0
2
÷ax
0
1
e
aT −2
(t −(T −2/a)).
10.2.3 (a) There are two state variables x
1
and x
2
, so we introduce two adjoint variables p
1
and p
2
. There are
also two control variables u
1
and u
2
. The Hamiltonian H =
1
2
x
1
÷
1
5
x
2
−u
1
−u
2
÷p
1
u
1
÷p
2
u
2
is linear
and hence concave. The following conditions are sufﬁcient for the admissible quadruple (x
∗
1
, x
∗
2
, u
∗
1
, u
∗
2
)
to be optimal:
(i) (u
1
, u
2
) = (u
∗
1
(t ), u
∗
2
(t )) maximizes (p
1
(t ) −1)u
1
÷(p
2
(t ) −1)u
2
for u
1
∈ [0, 1], u
2
∈ [0, 1];
(ii) ˙ p
1
(t ) = −(H
/
x
1
)
∗
= −1/2, p
1
(T ) = 0, ˙ p
2
(t ) = −(H
/
x
2
)
∗
= −1/5, p
2
(T ) = 0;
We see immediately from (ii) that p
1
(t ) =
1
2
(T −t ) and p
2
(t ) =
1
5
(T −t ). Note that p
1
(0) =
1
2
T > 1
since T > 5. Thus p
1
(t ) strictly decreases from a level higher than 1 at t = 0 to 0 at t = T . Looking
at (i) we see that u
∗
1
(t ) = 1 for p
1
(t ) > 1 and 0 for p
1
(t ) < 1. Since p
1
(t
∗
) = 1 when
1
2
(T − t
∗
) = 1,
we get t
∗
= T −2. In the same way, p
2
(0) =
1
5
T > 1 since T > 5. Thus p
2
(t ) strictly decreases from
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
82 10 CONTROL THEORY WI TH MANY VARI ABL ES
a level higher than 1 at t = 0 to 0 at t = T . Looking at (i) we see that u
∗
2
(t ) = 1 for p(t ) > 1 and 0 for
p(t ) < 1. Since p
2
(t
∗∗
) = 1 when
1
5
(T −t
∗∗
) = 1, we get t
∗∗
= T −5. We conclude that
u
∗
1
(t ) =
¸
1 if t ∈ [0, T −2]
0 if t ∈ (T −2, T ]
u
∗
2
(t ) =
¸
1 if t ∈ [0, T −5]
0 if t ∈ (T −5, T ]
The corresponding values for (x
∗
1
(t ), x
∗
2
(t )) are easily worked out. See the answer in the book.
(b) The scrap value function is S(x
1
, x
2
) = 3x
1
÷ 2x
2
. The Hamiltonian and the differential equations
for p
1
and p
2
are the same as in (a). The transversality conditions are changed. In fact, according to
Theorem 10.1.5 (C)(c’), we have p
1
(T ) = S
/
1
(x
∗
1
(t ), x
2
(t )) = 3 and p
2
(T ) = S
/
2
(x
∗
1
(t ), x
2
(t )) = 2.
Then p
1
(t ) = 3 ÷
1
2
(T − t ) and p
2
(t ) = 2 ÷
1
5
(T − t ). For t ∈ [0, T ] we see that p
1
(t ) and p
2
(t ) are
both greater than 1, and condition (i) in (a) implies that u
∗
1
(t ) = u
∗
2
(t ) = 1 for all t ∈ [0, T ], and thus
x
∗
1
(t ) = x
∗
2
(t ) = t .
10.2.4 The Hamiltonian H = x
2
÷c(1 −u
1
−u
2
) ÷p
1
au
1
÷p
2
(au
2
÷bx
1
) is linear and hence concave.
The control region is U = {(u
1
, u
2
) : 0 ≤ u
1
, 0 ≤ u
2
, u
1
÷ u
2
≤ 1}. The following conditions are
sufﬁcient for a quadruple (x
∗
1
, x
∗
2
, u
∗
1
, u
∗
2
) to be admissible and optimal:
(i) (u
1
, u
2
) = (u
∗
1
(t ), u
∗
2
(t )) maximizes (ap
1
(t ) −c)u
1
÷(ap
2
(t ) −c) for (u
1
, u
2
) ∈ U;
(ii) ˙ p
1
(t ) = −(H
/
x
1
)
∗
= −bp
2
(t ), p
1
(T ) = 0, ˙ p
2
(t ) = −(H
/
x
2
)
∗
= −1, p
2
(T ) = 0;
(iii) ˙ x
∗
1
(t ) = au
∗
1
(t ), x
∗
1
(0) = x
0
1
, ˙ x
∗
2
(t ) = au
∗
2
(t ) ÷bx
∗
1
(t ), x
∗
2
(0) = x
0
2
.
From (ii) we see that p
2
(t ) = T − t . Therefore ˙ p
1
(t ) = −b(T − t ). Integrating and using p
1
(T ) = 0,
we get p
1
(t ) =
1
2
b(T −t )
2
. To ﬁnd the optimal controls we must for each t in [0, T ] solve the problem
max ϕ(u
1
, u
2
) = max
¸
¸
1
2
ab(T −t )
2
−c
¸
u
1
÷
¸
a(T −t ) −c
¸
u
2
s.t. (u
1
, u
2
) ∈ U
¸
(∗)
The control region U is the closed triangular region with corners at (0, 0), (1, 0), and (0, 1). Since ϕ is
a linear function it will attain its maximum over U at one of those corners. (In some cases there may be
more than one maximum point, but even then at least one corner will be a maximum point.) Note that
ϕ(0, 0) = 0, ϕ(1, 0) =
1
2
ab(T − t )
2
− c, and ϕ(0, 1) = a(T − t ) − c. In the chains of equivalences
below it is understood that either the top inequality holds all the way or the bottom inequality holds all
the way. We see that (with t < T )
ϕ(1, 0) ≷ ϕ(0, 1) ⇐⇒
1
2
ab(T −t )
2
−c ≷ a(T −t ) −c ⇐⇒ t ≶ T −2/b
ϕ(0, 1) ≷ ϕ(0, 0) ⇐⇒ a(T −t ) −c ≷ 0 ⇐⇒ t ≶ T −c/a
ϕ(1, 0) ≷ ϕ(0, 0) ⇐⇒ T −t ≷
2c/ab ⇐⇒ t ≶ T −
2c/ab
Putting all this together (note that the assumption T −c/a > T −2/b in the problem implies c/a < 2/b,
and then c/a <
√
2c/ab < 2/b), we ﬁnd that a set of optimal controls are
(u
∗
1
(t ), u
∗
2
(t )) =
⎧
⎪
⎨
⎪
⎩
(1, 0) if t ∈ [0, T −2/b]
(0, 1) if t ∈ (T −2/b, T −c/a]
(0, 0) if t ∈ (T −c/a, T ]
(∗∗)
(For values of t in the corresponding open intervals, the optimal values of u
1
and u
2
are uniquely
determined; for t = T −2/b, t = T −c/a, and t = T we choose the values so that the control functions
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
10 CONTROL THEORY WI TH MANY VARI ABL ES 83
become leftcontinuous.) The corresponding values of x
∗
1
(t ) and x
∗
2
(t ) follow from (iii) and (∗), but the
expressions on the interval (T −c/a, T −2/b) get very messy.
10.2.5 There are two state variables x
1
and x
2
, so we introduce two adjoint variables p
1
and p
2
. The
Hamiltonian is H(t, x
1
, x
2
, u, p
1
, p
2
) = x
1
− cx
2
÷ u
0
− u ÷ p
1
u ÷ p
2
bx
1
. (We have p
0
= 1, since
p
1
(T ) = p
2
(T ) = 0.) The Hamiltonian is concave in (x
1
, x
2
, u), so according to Theorem 10.1.2 the
admissible triple (x
∗
1
, x
∗
2
, u
∗
) is optimal provided for each t in [0, T ],
(i) u = u
∗
(t ) maximizes H(t, x
∗
1
(t ), x
∗
2
(t ), u, p
1
(t ), p
2
(t )) = x
∗
1
(t ) −cx
∗
2
(t ) ÷u
0
−u÷p
1
(t )u÷
p
2
(t )bx
∗
1
(t ) = ÷p
2
(t )bx
∗
1
(t ) ÷(p
1
(t ) −1)u for u in [0, u
0
] (where does not depend on u).
(ii) ˙ p
1
(t ) = −∂H
∗
/∂x
1
= −1 −bp
2
(t ), p
1
(T ) = 0 and ˙ p
2
(t ) = −∂H
∗
/∂x
2
= c, p
2
(T ) = 0
From (i) we see that p
1
(t ) > 1 ⇒ u
∗
(t ) = u
0
, p
1
(t ) < 1 ⇒ u
∗
(t ) = 0. Moreover, from (ii),
p
2
(t ) = c(t −T ), and ˙ p
1
(t ) = −1−bp
2
(t ) = −1−bc(t −T ), which gives p
1
(t ) = T −t −
1
2
bc(t −T )
2
=
−
1
2
bct
2
÷ (bcT − 1)t ÷ T −
1
2
bcT
2
, since p
1
(T ) = 0. Note that p
1
(t ) is a quadratic polynomial in t ,
with p
1
(0) = T (1 −
1
2
bcT ) and maximum at t
1
= T −1/bc (since ˙ p
1
(t
1
) = 0). The maximum value of
p
1
is p
1
(t
1
) = 1/2bc.
We restrict our attention to the main case bcT > 2 and 2bc < 1. (The other cases are much simpler.)
We then get p
1
(0) = T (1 −
1
2
bcT ) < 0, p
1
(t
1
) = 1/2bc > 1, p
1
(T ) = 0. The graph of p is shown in
Figure 10.2.5.
t
p
1
t
∗
t
∗∗
T
Figure 10.2.5
There will be two points t
∗
and t
∗∗
, such that 0 < t
∗
< t
∗∗
< T and p
1
(t
∗
) = p
1
(t
∗∗
) = 1. Now,
p
1
(t ) =
1
2
bct
2
−(bcT −1)t −T ÷
1
2
bcT
2
÷1 = 0 ⇐⇒ t = T −
1
bc
±
1
bc
√
1 −2bc
This gives
t
∗
= T −
1
bc
−
1
bc
√
1 −2bc , t
∗∗
= T −
1
bc
÷
1
bc
√
1 −2bc
The optimal control u
∗
and the corresponding x
∗
1
(t ) are given by
u
∗
(t ) =
⎧
⎨
⎩
0 for t in [0, t
∗
],
1 for t in (t
∗
, t
∗∗
]
0 for t in (t
∗∗
, T ]
, x
∗
1
(t ) =
⎧
⎨
⎩
x
0
1
for t in [0, t
∗
],
t −t
∗
÷x
0
1
for t in (t
∗
, t
∗∗
]
t
∗∗
−t
∗
÷x
0
1
for t in (t
∗∗
, T ]
The expression for x
∗
2
(t ) is somewhat messy:
x
∗
2
(t ) =
⎧
⎨
⎩
bx
0
1
t ÷x
0
2
for t in [0, t
∗
],
1
2
bt
2
÷b(x
0
1
−t
∗
)t ÷x
0
2
÷
1
2
b(t
∗
)
2
for t in (t
∗
, t
∗∗
]
b(t
∗∗
−t
∗
)t ÷bx
0
1
t ÷
1
2
b[(t
∗
)
2
−(t
∗∗
)
2
] ÷x
0
2
for t in (t
∗∗
, T ]
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
84 10 CONTROL THEORY WI TH MANY VARI ABL ES
10.2.7 There are two state variables and one control variable. We shall use Theorem10.1.2. The Hamiltonian
is H(t, x, y, u, p
1
, p
2
) = x ÷(p
1
÷p
2
−
1
2
)u. Suppose (x
∗
, y
∗
, u
∗
) is an admissible triple which solves
the problem. Then:
(i) For each t in [0, 2], u = u
∗
(t ) maximizes
H(t, x
∗
(t ), y
∗
(t ), u, p
1
(t ), p
2
(t )) = x
∗
(t ) ÷(p
1
(t ) ÷p
2
(t ) −
1
2
)u for u in [0, 1];
(ii) ˙ p
1
(t ) = −∂H
∗
/∂x = −1, ˙ p
2
(t ) = −∂H
∗
/∂y = 0;
(iii) (a) p
1
(2) = 0,
(b) p
2
(2) ≤ 0, and p
2
(2) = 0 if y
∗
(2) < 1;
(iv) ˙ x
∗
(t ) = u
∗
(t ), x
∗
(0) = 1, ˙ y
∗
(t ) = u
∗
(t ), y
∗
(0) = 0.
From (ii), ˙ p
1
= −1, and since p
1
(2) = 0, we have p
1
(t ) = 2 −t . Because ˙ p
2
= 0, p
2
(t ) = ¨ p
2
, where
¨ p
2
is a constant. Hence, p
1
(t ) ÷p
2
(t ) −
1
2
= 2−t ÷ ¨ p
2
−
1
2
= t
∗
−t, where t
∗
= 3/2÷ ¨ p
2
is a constant.
Condition (i) gives
u
∗
(t ) =
¸
1 if t < t
∗
0 if t > t
∗
If t
∗
= 2, then u
∗
(t ) ≡ 1 and y
∗
(t ) ≡ t , and then y
∗
(2) > 1, which is impossible.
If t
∗
= 0, then u
∗
(t ) ≡ 0 and y
∗
(t ) ≡ 0, which gives y
∗
(2) = 0 < 1. In this case ¨ p
2
= p
2
(2) = 0
because of (iii). Hence, t
∗
= 3/2, contradicting t
∗
= 0.
Thus we must have 0 < t
∗
< 2, and then x
∗
(t ) −1 = y
∗
(t ) = t for t ≤ t
∗
and x
∗
(t ) −1 = y
∗
(t ) = t
∗
for t > t
∗
. In particular, y
∗
(2) = t
∗
= 3/2 ÷ ¨ p
2
. It remains to determine ¨ p
2
and t
∗
. Since y
∗
(2) ≤ 1,
we must have 3/2 ÷ ¨ p
2
≤ 1, i.e. ¨ p
2
≤ −1/2 < 0. From (iii), y
∗
(2) = 1, and so ¨ p
2
= −1/2 and t
∗
= 1.
The Hamiltonian is concave in (x, y, u) (in fact linear), so we have found a solution:
u
∗
(t ) =
¸
1 if t ≤ 1
0 if t > 1
, x
∗
(t ) −1 = y
∗
(t ) =
¸
t if t ≤ 1
1 if t > 1
, p
1
(t ) = 2 −t, p
2
(t ) = −
1
2
10.3
10.3.1 The Hamiltonian H = (x −u)e
−rt
÷p(t )ue
−t
is concave in (x, u). If (x
∗
(t ), u
∗
(t )) is optimal, then
(i) u = u
∗
(t ) maximizes (e
−t
p(t ) −e
−rt
)u for u ∈ [0, 1];
(ii) ˙ p(t ) = −H
/
x
(t, x
∗
(t ), u
∗
(t ), p(t )) = −e
−rt
;
(iii) ˙ x
∗
(t ) = u
∗
(t )e
−t
, x
∗
(0) = x
0
≥ 0.
From (ii) we get p(t ) = (1/r)e
−rt
÷ C. There is no restriction on x(t ) as t → ∞, so we guess that
p(t ) →0 as t →∞. Then we must have C = 0, so p(t ) = (1/r)e
−rt
. With this choice of p(t ), we see
that u = u
∗
(t ) maximizes (1/r)e
−rt
(e
−t
−r)u for u ∈ [0, 1]. Thus, u
∗
(t ) = 1 if e
−t
> r and u
∗
(t ) = 0
if e
−t
< r. We have e
−t
∗
= r when t
∗
= −ln r, and then we see that (draw a picture!)
u
∗
(t ) =
¸
1 if t ∈ [0, −ln r]
0 if t ∈ (−ln r, ∞)
From (iii) we ﬁnd that x
∗
(t ) = x
0
÷1 −e
−t
in [0, −ln r], while x
∗
(t ) = x
0
÷1 −r in (−ln r, ∞). Note
that since ˙ x ≥ 0 for all t ≥ 0 and x(0) ≥ 0, we have x(t ) ≥ x
0
for all t ≥ 0.
We have p(t ) = e
−rt
/r ≥ 0, so according to Note 10.3.2 it remains only to verify (10.3.10a). In
fact, we see that lim
t →∞
p(t )(x
0
−x
∗
(t )) = lim
t →∞
(e
−rt
/r)(x
0
−x
0
−1 ÷r) = 0.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
10 CONTROL THEORY WI TH MANY VARI ABL ES 85
10.3.2 (a) The model is closely related to the model in Example 10.2.2. The Hamiltonian is H = x
2
e
−rt
÷
p
1
aux
1
÷p
2
a(1 −u)x
1
. If (x
∗
1
(t ), x
∗
2
(t ), u
∗
(t )) is optimal, then
(i) u = u
∗
(t ) maximizes ax
∗
1
(t )(p
1
(t ) −p
2
(t ))u for u ∈ [0, 1];
(ii) ˙ p
1
(t ) = −ap
1
(t )u
∗
(t ) −ap
2
(t )(1 −u
∗
(t )), ˙ p
2
(t ) = −e
−rt
;
(iii) ˙ x
∗
1
(t ) = au
∗
(t )x
∗
1
(t ), x
∗
1
(0) = x
0
1
> 0, ˙ x
∗
2
(t ) = a(1 −u
∗
(t ))x
∗
1
(t ), x
∗
2
(0) = 0.
From (ii) we get p
2
(t ) = e
−rt
/r ÷ C. There is no restriction on x
2
(t ) as t → ∞, so we guess that
p
2
(t ) →0 as t →∞. Then we must have C = 0, so p
2
(t ) = e
−rt
/r.
From (i) we see that
u
∗
(t ) =
¸
1 if p
1
(t ) > e
−rt
/r
0 if p
1
(t ) < e
−rt
/r
Suppose p
1
(0) > p
2
(0) = 1/r. Then u
∗
(t ) = 1 to the immediate right of t = 0 and ˙ p
1
(t ) = −ap
1
(t ), so
p
1
(t ) = p
1
(0)e
−at
> (1/r)e
−at
> (1/r)e
−rt
, since r > a, and we see that we must have p
1
(t ) > p
2
(t )
for all t ≥ 0. Then u
∗
(t ) ≡ 1 and from (iii) we have x
∗
2
(t ) ≡ 0, and the objective function is 0. This is
not the optimal solution. (In the terminology of Example 10.2.2 the total discounted consumption is 0.)
If p
1
(0) = p
2
(0), then ˙ p
1
(0) = −ap
2
(0) = −a/r > −1 = ˙ p
2
(0), and again we see that p
1
(t ) > p
2
(t )
for all t ≥ 0. Suppose p
1
(0) < 1/r. Then u
∗
(t ) = 0 for t close to 0. Let us see if we can have u
∗
(t ) = 0
for all t . Then ˙ p
1
(t ) = −ap
2
(t ) = −(a/r)e
−rt
and so p
1
(t ) = (a/r
2
)e
−rt
÷ D. Again we must
have D = 0 and then p
1
(t ) = (a/r
2
)e
−rt
< (1/r)e
−rt
= p
2
(t ). Finally, using (10.3.10(a)), we have
p
1
(t )(x
0
1
−x
∗
1
(t )) = 0 and p
2
(t )(0 −x
∗
2
(t )) = (1/r)e
−rt
(−ax
0
1
) →0 as t →∞. As in Example 10.2.2
the Arrow condition is satisﬁed, so we have found the optimal solution. Note that using the interpretation
in Example 10.2.2, in this case the discount factor r is so high that it is optimal with no further investment
in the investment sector, which leads to consumption increasing at a constant rate.
(b) Choose the control u(t ) = b/a, with 0 < r < b < a. Then u(t ) ∈ [0, 1] and we seek corresponding
admissible state variables. From ˙ x
1
(t ) = a(b/a)x
1
(t ) = bx
1
(t ), with x
1
(0) = x
0
1
, we ﬁnd x
1
(t ) = x
0
1
e
bt
.
Then ˙ x
2
(t ) = (a − b)x
0
1
e
bt
, so x
2
(t ) = (a − b)x
0
1
(1/b)e
bt
÷ C. With x
2
(0) = 0, we ﬁnd that C =
−(a −b)(1/b)x
0
1
, so x
2
(t ) = (a −b)(1/b)x
0
1
(e
bt
−1). Then the objective function is
∞
0
x
2
(t )e
−rt
dt =
∞
0
[(a − b)(1/b)x
0
1
e
(b−r)t
− e
−rt
] dt =
∞
0
[(a − b)(1/b)x
0
1
[(1/(b − r)]e
(b−r)t
÷ e
−rt
/r]. By using
0 < r < b < a we see that the integral diverges.
10.4
10.4.2 First we apply Theorem 10.4.1. The control region U = [0, 1] is convex and the condition in Note
10.4.2 is satisﬁed because [ux[ ≤ [x[ since [u[ ≤ 1. The set N(t, x) = {((1−u)x
2
÷γ, ux) : γ ≤ 0, u ∈
[0.1]} is convex (a rectangle) by the same argument as in Example 10.4.1. Thus there exists an optimal
control.
The Hamiltonian is H = (1 −u)x
2
÷pux = x
2
÷ux(p −x), and if (x
∗
(t ), u
∗
(t )) is optimal then:
(i) u = u
∗
(t ) maximizes x
∗
(t )(p(t ) −x
∗
(t ))u for u ∈ [0, 1];
(ii) ˙ p(t ) = −(H
/
x
)
∗
= −2(1 −u
∗
(t ))x
∗
(t ) −p(t )u
∗
(t ), p(1) = 0;
(iii) ˙ x
∗
(t ) = u
∗
(t )x
∗
(t ), x
∗
(0) = x
0
> 0.
From (iii) we see that ˙ x
∗
(t ) ≥ 0 so x
∗
(t ) ≥ x
0
> 0. Thus (i) implies that
u
∗
(t ) =
¸
1 if p(t ) > x
∗
(t )
0 if p(t ) < x
∗
(t )
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
86 10 CONTROL THEORY WI TH MANY VARI ABL ES
We claim that p(t ) is strictly decreasing. In fact, for those t where p(t ) > x
∗
(t ) we have u
∗
(t ) = 1 and
˙ p(t ) = −p(t ) < 0. For those t where p(t ) < x
∗
(t ) we have u
∗
(t ) = 0 and ˙ p(t ) = −2x
∗
(t ) < 0. Since
x
∗
(t ) is increasing from the level x
0
> 0 and p(t ) is strictly decreasing and is 0 at t = 1, we have to have
p(t ) < x
∗
(t ) in some interval (t
∗
, 1]. Then u
∗
(t ) = 0 in this interval.
If t
∗
= 0, then u
∗
(t ) ≡ 0, x
∗
(t ) ≡ x
0
, and ˙ p(t ) ≡ −2x
0
. Since p(1) = 0 we get p(t ) = 2x
0
(1 −t ).
But then p(0) = 2x
0
, contradicting p(0) ≤ x(0) = x
0
. With t
∗
> 0,
u
∗
(t ) =
¸
1 if t < t
∗
0 if t > t
∗
and x
∗
(t ) =
¸
x
0
e
t
if t < t
∗
x
0
e
t
∗
if t > t
∗
In (t
∗
, 1] we have ˙ p(t ) = −2x
∗
(t ) = −2x
0
e
t
∗
and p(1) = 0, so p(t ) = 2x
0
e
t
∗
(1 − t ). But at t
∗
we
have p(t
∗
) = x
∗
(t
∗
), so 2x
0
e
t
∗
(1 − t
∗
) = x
0
e
t
∗
, from which it follows that t
∗
= 1/2. We ﬁnd that
p(t ) = x
0
e
1−t
in [0, 1/2]. We have found the optimal solution.
10.4.3 The Hamiltonian is H = p
0
x
2
÷p(1 −u
2
). According to the maximum principle, if (x
∗
(t ), u
∗
(t ))
is an optimal pair, there exist a continuous and piecewise differentiable function p(t ) and a constant p
0
,
either 0 or 1, such that (p
0
, p(t )) ,= (0, 0) and
(i) u = u
∗
(t ) maximizes p
0
(x
∗
)
2
÷p(t )(1 −u
2
) for u ∈ [−1, 2];
(ii) ˙ p(t ) = −H
/
x
(t, x
∗
(t ), u
∗
(t ), p(t )) = −2p
0
x
∗
(t );
(iii) ˙ x
∗
(t ) = 1 −(u
∗
(t ))
2
, x
∗
(0) = x
∗
(1) = 4.
Suppose p
0
= 0. Then from (ii), p(t ) = ¨ p ,= 0, and u
∗
(t ) maximizes − ¨ pu
2
for u ∈ [−1, 2]. If ¨ p > 0,
then u
∗
(t ) ≡ 0 and ˙ x
∗
(t ) ≡ 1, so with x
∗
(0) = 4, x
∗
(t ) = t ÷4, and then x
∗
(1) ,= 4. On the other hand,
if ¨ p < 0, then u
∗
(t ) ≡ 2 and ˙ x
∗
(t ) ≡ −3, so with x
∗
(0) = 4, x
∗
(t ) = −3t ÷ 4, and then x
∗
(1) ,= 4.
Thus, assuming p
0
= 0 leads to contradictions, so p
0
= 1.
An optimal control u
∗
(t ) must maximize p(t )(1−u
2
) for u ∈ [−1, 2]. Then we see that if p(t ) > 0,
one should use u
∗
(t ) = 0, and if p(t ) < 0, then one should use u
∗
(t ) = 2. (In neither case should one
use u
∗
(t ) = −1!) From (iii) and u ∈ [−1, 2] we see that ˙ x
∗
(t ) is never less than −3. Since x
∗
(0) = 4,
it means that x
∗
(t ) is always ≥ 1. (Formally, x
∗
(t ) − x
∗
(0) =
t
0
˙ x
∗
(τ) dτ ≥
t
0
(−3) dτ = −3t , so
x
∗
(t ) ≥ x
∗
(0) −3t = 4 −3t , which is ≥ 1 in [0, 1].) From (ii) it follows that ˙ p(t ) = −2x
∗
(t ) < 0, so
p(t ) is strictly decreasing.
Suppose p(t ) < 0 for all t ∈ [0, 1]. Then u
∗
(t ) ≡ 2 and we get a contradiction to x
∗
(1) = 4. If
p(t ) > 0 for all t ∈ [0, 1], then u
∗
(t ) ≡ 0 and again we get a contradiction to x
∗
(1) = 4. We conclude
that the strictly decreasing function p(t ) is > 0 in some interval [0, t
∗
] and < 0 in (t
∗
, 1]. Then
u
∗
(t ) =
¸
0 if t ∈ [0, t
∗
]
2 if t ∈ (t
∗
, 1]
In [0, t
∗
] we have ˙ x
∗
(t ) = 1 and since x
∗
(0) = 4, x
∗
(t ) = t ÷ 4. In (t
∗
, 1] we have ˙ x
∗
(t ) = −3, and
since x
∗
(1) = 4, we have x
∗
(t ) = −3t ÷ 7. Now p(t ) is continuous at t
∗
, so t
∗
÷ 4 = −3t
∗
÷ 7, so
t
∗
= 3/4.
It remains to ﬁnd p(t ). On [0, 3/4] we have ˙ p(t ) = −2x
∗
(t ) = −2t −8, and so p(t ) = −t
2
−8t ÷C.
Since p(3/4) = 0, C = 105/16. In (3/4, 1] we have ˙ p(t ) = −2x
∗
(t ) = 6t − 14, and so p(t ) =
3t
2
− 14t ÷ D. Since p(3/4) = 0, D = 141/16. The answer is summed up in the answer in the book.
(Note the misprint in line 3 of the answer to Problem 10.4.3 in the book: . . . When u = 2, ˙ x = −3 . . . )
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
10 CONTROL THEORY WI TH MANY VARI ABL ES 87
10.6
10.6.1 (a) H = −
1
2
u
2
−x −pu and L = H ÷q(x −u). H is concave in (x, u) and h(t, x, u) = x −u is
linear and therefore quasiconcave. The conditions in Theorem 10.6.1 are
(i) ∂L
∗
/∂u = −u
∗
(t ) −p(t ) −q(t ) = 0;
(ii) q(t ) ≥ 0 (= 0 if x
∗
(t ) > u
∗
(t ));
(iii) ˙ p(t ) = 1 −q(t ), with p(2) = 0.
(b) In [0, t
∗
] we have x
∗
(t ) = u
∗
(t ), so ˙ x
∗
(t ) = −u
∗
(t ) = −x
∗
(t ), and with x
∗
(0) = 1 we get x
∗
(t ) =
u
∗
(t ) = e
−t
. From (i) we have q(t ) = −u
∗
(t ) −p(t ) = −e
−t
−p(t ), so (iii) gives ˙ p(t ) = 1 −q(t ) =
1÷e
−t
÷p(t ). This linear differential equation has the solution p(t ) = Ae
t
−1−
1
2
e
−t
. Fromthe argument
in the problem, q(t
∗ −
) = 0. Thus q(t
∗ −
) = −e
−t
∗
−p(t
∗
) = 0, so p(t
∗
) = −e
−t
∗
= Ae
t
∗
−1 −
1
2
e
−t
∗
,
or Ae
t
∗
= 1 −
1
2
e
−t
∗
.
In (t
∗
, 1] we have q(t ) = 0 and ˙ p(t ) = 1, and since p(2) = 0, we get p(t ) = t −2. Then from (i),
u
∗
(t ) = −p(t ) = 2 −t , and so ˙ x
∗
(t ) = t −2 and then x
∗
(t ) =
1
2
t
2
−2t ÷B. In particular, since x
∗
(t )
is continuous at t
∗
, x
∗
(t
∗
) =
1
2
(t
∗
)
2
− 2t
∗
÷ B = e
−t
∗
. This gives B = e
−t
∗
−
1
2
(t
∗
)
2
÷ 2t
∗
. Finally,
since p(t ) is continuous at t
∗
, we have t
∗
−2 = Ae
t
∗
−1 −
1
2
e
−t
∗
= 1 −
1
2
e
−t
∗
−1 −
1
2
e
−t
∗
= −e
−t
∗
,
so e
−t
∗
= 2 − t
∗
. By using this relationship you will see that the values of A and B are the same as in
the answer in the book.
10.6.2 This is a problemof the form(10.6.1)–(10.6.4). The Lagrangian (10.6.5) is here L = H÷q(x−u) =
x −
1
2
u
2
÷ pu ÷ q(x − u). Note that the Hamiltonian H = x −
1
2
x
2
÷ pu is concave (in fact linear)
in (x, u) and that h = x − u is quasiconcave (in fact linear) in (x, u). According to Theorem 10.6.1,
the following conditions are sufﬁcient for (x
∗
(t ), u
∗
(t )) to be optimal: There exist a continuous function
p(t ) and a piecewise continuous function q(t ) such that
(i) ∂L
∗
/∂u = −u
∗
(t ) ÷p(t ) −q(t ) = 0;
(ii) q(t ) ≥ 0, with q(t ) = 0 if x
∗
(t ) > u
∗
(t );
(iii) ˙ p(t ) = −∂L
∗
/∂x = −1 −q(t ), p(2) = 0;
(iv) ˙ x
∗
(t ) = u
∗
(t ), x
∗
(0) = 1.
We guess that u
∗
(t ) = x
∗
(t ) on some interval [0, t
∗
], and then (iv) gives ˙ x
∗
(t ) = x
∗
(t ), with x
∗
(0) = 1,
so x
∗
(t ) = e
t
= u
∗
(t ). Then from (i) we have q(t ) = p(t ) − e
t
, and (iii) gives ˙ p(t ) = −1 − q(t ) =
−1 − p(t ) ÷ e
t
, or ˙ p(t ) ÷ p(t ) = −1 ÷ e
t
. The solution of this linear differential equation is p(t ) =
Be
−t
÷
1
2
e
t
−1, and then q(t ) = Be
−t
÷
1
2
e
t
−1 −e
t
= Be
−t
−
1
2
e
t
−1.
On (t
∗
, 2] we guess that x
∗
(t ) > u
∗
(t ). Then from (ii) we have q(t ) = 0, and so (iii) gives p(t ) =
2−t . Then from(i) u
∗
(t ) = p(t ) −q(t ) = 2−t , and (iv) gives ˙ x
∗
(t ) = 2−t , so x
∗
(t ) = −
1
2
t
2
÷2t ÷A.
Since x
∗
(t ) is continuous at t
∗
, we have −
1
2
(t
∗
)
2
÷ 2t
∗
÷ A = e
t
∗
, so A = e
t
∗
÷
1
2
(t
∗
)
2
− 2t
∗
. Since
p(t ) is also continuous at t
∗
, we get Be
−t
∗
÷
1
2
e
t
∗
− 1 = 2 − t
∗
, so Be
−t
∗
= −
1
2
e
t
∗
÷ 3 − t
∗
. Finally,
since q(t
∗−
) = 0 (see the argument in the previous problem), we get Be
−t
∗
=
1
2
e
t
∗
÷ 1. From the last
two equalities we get e
t
∗
= 2 −t
∗
, i.e. t
∗
≈ 0.44. To conﬁrm that all the conditions (i)–(iv) are satisﬁed,
we should verify that q(t ) = (
1
2
e
2t
∗
÷e
t
∗
)e
−t
−
1
2
e
t
−1 is nonnegative in [0, t
∗
]. This is the case because
we ﬁnd that ˙ q(t ) < 0 and q(t
∗
) = 0. The solution is summed up in the answer in the book.
10.6.3 Note that there are two constraints, h
1
= u − c ≥ 0 and h
2
= ax − u ≥ 0. The Lagrangian is
L = H ÷ q
1
(u − c) ÷ q
2
(ax − u) = u ÷ p(ax − u) ÷ q
1
(u − c) ÷ q
2
(ax − u). The Hamiltonian is
concave (in fact linear) in (x, u) and h
1
and h
2
are quasiconcave (in fact linear) in (x, u). According to
Theorem10.6.1, the following conditions are sufﬁcient for an admissible pair (x
∗
(t ), u
∗
(t )) to be optimal:
There exist a continuous function p(t ) and piecewise continuous functions q
1
(t ) and q
2
(t ) such that
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
88 10 CONTROL THEORY WI TH MANY VARI ABL ES
(i) ∂L
∗
/∂u = 1 −p(t ) ÷q
1
(t ) −q
2
(t ) = 0;
(ii) q
1
(t ) ≥ 0, with q
1
(t ) = 0 if u
∗
(t ) > c;
(iii) q
2
(t ) ≥ 0, with q
2
(t ) = 0 if ax
∗
(t ) > u
∗
(t );
(iv) ˙ p(t ) = −∂L
∗
/∂x = −ap(t ) −aq
2
(t );
(v) p(T ) ≥ 0 with p(T ) = 0 if x
∗
(T ) > x
T
;
(vi) ˙ x
∗
(t ) = ax
∗
(t ) −u
∗
(t ), x
∗
(0) = x
0
, x
∗
(T ) ≥ x
T
;
(vii) c ≤ u
∗
(t ) ≤ ax
∗
(t )
In the similar model in Example 10.6.2 it was optimal to start out with u
∗
(t ) = c in an initial interval,
and then keep x
∗
(t ) constant, so let us try the same here:
u
∗
(t ) =
¸
c if t ∈ [0, t
/
]
ax
∗
(t ) if t ∈ (t
/
, T ]
, then x
∗
(t ) =
¸
(x
0
−c/a)e
at
÷c/a if t ∈ [0, t
/
]
(x
0
−c/a)e
at
/
÷c/a if t ∈ (t
/
, T ]
(viii)
In the interval [0, t
/
] we have u
∗
(t ) < ax
∗
(t ) because c < (ax
0
−c)e
at
÷c, and then (iii) gives q
2
(t ) = 0,
so we have ˙ p(t ) = −ap(t ). In the interval (t
/
, T ] we have u
∗
(t ) = ax
∗
(t ) = (ax
0
−c)e
at
/
÷c > c, so
according to (ii), q
1
(t ) = 0. Then (i) gives q
2
(t ) = 1 −p(t ), which inserted into (iv) gives ˙ p(t ) = −a.
We claim moreover that p(t
/
) must be 1. In fact from (i), since q
2
(t
/−
) = 0, we have 1 − p(t
/−
) =
−q
1
(t
/−
) ≤ 0 and since q
1
(t
/÷
) = 0, we have 1 − p(t
/÷
) = q
2
(t
/÷
) ≥ 0. Because p(t ) is continuous,
p(t
/
) = 1. Then we get
˙ p(t ) =
¸
−ap if t ∈ [0, t
/
]
−a if t ∈ (t
/
, T ]
and p(t
/
) = 1 ⇒ p(t ) =
¸
e
−a(t −t
/
)
if t ∈ [0, t
/
]
a(t
/
−t ) ÷1 if t ∈ (t
/
, T ]
(ix)
The corresponding values of q
1
(t ) and q
2
(t ) are
q
1
(t ) =
¸
e
−a(t −t
/
)
−1 if t ∈ [0, t
/
]
0 if t ∈ (t
/
, T ]
, q
2
(t ) =
¸
0 if t ∈ [0, t
/
]
a(t −t
/
) if t ∈ (t
/
, T ]
(x)
Case A: x
∗
(T ) > x
T
. Then p(T ) = 0 and thus a(t
/
− T ) ÷ 1 = 0, so t
/
= t
A
, where t
A
= T − 1/a.
The optimal solution is given by (viii), p(t ) by (ix), and q
1
(t ) and q
2
(t ) by (x), all with t
/
replaced by
T − 1/a. It is a useful exercise to check that all the conditions in (i)–(vii) are satisﬁed. In particular,
check that q
1
(t ) and q
2
(t ) are both ≥ 0. We must also check that x
∗
(T ) > x
T
. We see from (viii) that
this is equivalent to (x
0
−c/a)e
a(T −1/a)
÷c/a > x
T
, or t
A
> t
B
, with t
B
deﬁned in (xi) below.
Case B: x
∗
(T ) = x
T
. Then from (viii) we get (x
0
− c/a)e
at
/
÷ c/a = x
T
, which solved for t
/
gives
t
/
= t
B
, where
t
B
=
1
a
ln
x
T
−c/a
x
0
−c/a
(xi)
Note that from (v) we need to have p(T ) ≥ 0. The formula for p(t ) in (ix) gives a(t
b
−T ) ÷1 ≥ 0, or
t
B
≥ t
A
. The solution with t
/
= t
B
is valid if the last inequality is satisﬁed.
The ﬁnal conclusion is given in the answer in the book. (In line 2 of the answer to Problem 10.6.3,
replace x
∗
by x
∗
and t
∗
by t
/
.)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
10 CONTROL THEORY WI TH MANY VARI ABL ES 89
10.6.4 The Lagrangian is L = H ÷ q
1
(1 − u) ÷ q
2
(1 ÷ u) ÷ q
3
(2 − x − u) = x ÷ p(x ÷ u) ÷ q
1
(1 −
u) ÷q
2
(1 ÷u) ÷q
3
(2 −x −u). The Hamiltonian H = x ÷p(x ÷u) is concave (in fact linear) in (x, u)
and h
1
= 1 −u, h
2
= 1 ÷u, and h
3
= 2 −x −u are quasiconcave (in fact linear) in (x, u). According
to Theorem 10.6.1, the following conditions are sufﬁcient for (x
∗
(t ), u
∗
(t )) to be an optimal pair: there
exist a continuous function p(t ) and piecewise continuous functions q
1
(t ), q
2
(t ), and q
3
(t ), such that
(i) ∂L
∗
/∂u = p(t ) −q
1
(t ) ÷q
2
(t ) −q
3
(t ) = 0;
(ii) q
1
(t ) ≥ 0, with q
1
(t ) = 0 if u
∗
(t ) < 1;
(iii) q
2
(t ) ≥ 0, with q
2
(t ) = 0 if u
∗
(t ) > −1;
(iv) q
3
(t ) ≥ 0, with q
3
(t ) = 0 if u
∗
(t ) ÷x
∗
(t ) < 2;
(v) ˙ p(t ) = −∂L
∗
/∂x = −1 −p(t ) ÷q
3
(t ), p(1) = 0;
(vi) ˙ x
∗
(t ) = x
∗
(t ) ÷u
∗
(t ), x
∗
(0) = 0, i.e. (x
∗
, u
∗
) is admissible.
If we disregard the constraint x ÷u ≤ 2, this is the problem solved in Example 9.4.1, whose solution was
(x(t ), u(t )) = (e
t
−1, 1). Note that in this case x(t ) ÷u(t ) = e
t
≤ 2 as long as t ≤ ln 2. We guess that
this is the optimal solution on the interval [0, ln 2] in the present problem too. In fact, we will try to guess
the optimal solution and then verify its optimality by checking that all the conditions (i)–(vi) are satisﬁed.
So we start out with (x
∗
(t ), u
∗
(t )) = (e
t
−1, 1) on [0, ln 2]. At t = ln 2 we have x
∗
(ln 2) = e
ln 2
−1 = 1,
and, looking at the objective function, for t > ln 2 it seems optimal to increase x(t ) as fast as the constraint
x ÷ u ≤ 2 allows, i.e. putting ˙ x
∗
(t ) = u
∗
(t ) ÷ x
∗
(t ) = 2, as long as the h
1
and h
2
constraints are not
violated. Now, with ˙ x
∗
(t ) = 2 on [ln 2, 1], and x
∗
(ln 2) = 1, we get x
∗
(t ) = 2t ÷ 1 − 2 ln 2. Then
u
∗
(t ) = 2 − x
∗
(t ) = 1 ÷ 2 ln 2 − 2t , and it is easy to verify that u
∗
(t ) = 1 ÷ 2 ln 2 − 2t takes values
in (−1, 1) when t ∈ (ln 2, 1]. The suggestion we have for an optimal solution is therefore: In [0, ln 2],
(x
∗
(t ), u
∗
(t )) = (e
t
−1, 1), in (ln 2, 1], (x
∗
(t ), u
∗
(t )) = (2t ÷1 −2 ln 2, 1 ÷2 ln 2 −2t ).
We knowthat the suggestedsolutionis admissible. It remains toﬁndappropriate multipliers satisfying
(i)–(v).
In the interval (ln 2, 1], u
∗
(t ) ∈ (−1, 1), so from (ii) and (iii), q
1
(t ) = q
2
(t ) = 0. Then (i) gives
p(t ) = q
3
(t ) and from (v), ˙ p(t ) = −1 with p(1) = 0, so p(t ) = 1 −t . In particular, p(ln 2) = 1 −ln 2.
In the interval [0, ln 2), u
∗
(t ) = 1 > −1 and x
∗
(t ) ÷ u
∗
(t ) = e
t
< 2. Then from (iii) and (iv),
q
2
(t ) = q
3
(t ) = 0. Then (v) gives ˙ p(t ) = −1 −p(t ). Solving the linear differential equation on [0, ln 2]
with p(ln 2) = 1 − ln 2 gives p(t ) = (4 − 2 ln 2)e
−t
− 1. Then from (i), q
1
(t ) = p(t ). The complete
suggestion for an optimal solution is therefore:
u
∗
(t ) x
∗
(t ) p(t ) q
1
(t ) q
2
(t ) q
3
(t )
t ∈ [0, ln 2] e
t
−1 1 (4 −2 ln 2)e
−t
−1 (4 −2 ln 2)e
−t
−1 0 0
t ∈ (ln 2, 1] 2t ÷1 −2 ln 2 1 ÷2 ln 2 −2t 1 −t 0 0 1 −t
Having checked that (x
∗
(t ), u
∗
(t )) satisﬁes all the conditions (i)–(vi), we conclude that (x
∗
(t ), u
∗
(t )) is
optimal. Note that q
1
(t ) and q
3
(t ) have jump discontinuities at t = ln 2.
10.7
10.7.1 We maximize
5
0
(−u −x) dt , so the Lagrangian is L = H ÷qx = −u −x ÷p(u −t ) ÷qx. Here
H is concave in (x, u) and h(t, x) = x is quasiconcave, so by Theorem 10.7.1, the conditions (i)–(vi)
below are sufﬁcient for (x
∗
(t ), u
∗
(t )) to be optimal:
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
90 10 CONTROL THEORY WI TH MANY VARI ABL ES
(i) u = u
∗
(t ) maximizes −u −x
∗
(t ) ÷p(t )(u −t ) = −x
∗
(t ) −tp(t ) ÷u(p(t ) −1) for u ≥ 0;
(ii) q(t ) ≥ 0, with q(t ) = 0 if x
∗
(t ) > 0;
(iii) ˙ p(t ) = −∂L
∗
/∂x = 1 −q(t ), p(5) = 0;
(iv) p(5
−
) −p(5) = β;
(v) β ≥ 0, with β = 0 if x
∗
(5) > 0;
(vi) ˙ x
∗
(t ) = u
∗
(t ) −t , x
∗
(0) = 1.
Since we want to keep x
∗
(t ) down, we put u
∗
(t ) = 0 in some interval [0, t
∗
]. Then ˙ x
∗
(t ) = −t with
x
∗
(0) = 1, so x
∗
(t ) = −
1
2
t
2
÷ 1. We see that x
∗
(t ) is decreasing and is 0 at t
∗
=
√
2. In [0,
√
2) we
have by (ii), q(t ) = 0. Then (iii) gives ˙ p(t ) = 1, and thus p(t ) = t ÷A, for some constant A.
In order still to keep x
∗
(t ) down, we try u
∗
(t ) = t in (
√
2, 5]. Then ˙ x
∗
(t ) = 0 and thus x
∗
(t ) =
x
∗
(
√
2) = 0. For u
∗
(t ) = t to be the maximizer in (i) one has to have p(t ) = 1, in particular p(5
−
) = 1.
Since p(t ) is continuous at t =
√
2, we have
√
2 ÷A = 1, so A = 1 −
√
2. From (iii) we get q(t ) = 1.
Finally, since p(5) = 0, (iv) gives β = 1. Now all the conditions (i)–(vi) are satisﬁed, so (x
∗
(t ), u
∗
(t ))
is optimal.
10.7.2 The Lagrangian is L = H ÷qx = 1 −x ÷pu÷qx. Here H is concave in (x, u) and h(t, x) = x is
quasiconcave, and the conditions (i)–(vi) below are therefore sufﬁcient for (x
∗
(t ), u
∗
(t )) to be optimal:
(i) u = u
∗
(t ) maximizes 1 −x
∗
(t ) ÷p(t )u for u ∈ [−1, 1];
(ii) q(t ) ≥ 0, with q(t ) = 0 if x
∗
(t ) > 0;
(iii) ˙ p(t ) = −∂L
∗
/∂x = 1 −q(t ), p(2) = 0;
(iv) p(2
−
) −p(2) = β;
(v) β ≥ 0, with β = 0 if x
∗
(2) > 0;
(vi) ˙ x
∗
(t ) = u
∗
(t ), x
∗
(0) = 1.
We start by putting u
∗
(t ) = −1. Then x
∗
(t ) = 1 −t is decreasing and is 0 at t
∗
= 1. In [0, 1) we have
by (ii), q(t ) = 0. Then (iii) gives ˙ p(t ) = 1, and thus p(t ) = t ÷A, for some constant A.
In order still to keep x
∗
(t ) down put u
∗
(t ) = 0 in (1, 2]. Then with x
∗
(1) = 0 we get x
∗
(t ) = 0. For
u
∗
(t ) = 0 to be the maximizer in (i) for t ∈ (1, 2), one has to have p(t ) = 0, in particular p(2
−
) = 0.
Then since p(2) = 0, we get from (iv) that β = 0. Since p(t ) is continuous at t = 1, p(1) = 1 ÷A = 0,
so A = −1. Finally, from (iii) we get q(t ) = 1. Now all the conditions (i)–(vi) are satisﬁed, so the
optimal solution is:
(x
∗
(t ), u
∗
(t ), p(t )) =
¸
(1 −t, −1, t −1) if t ∈ [0, 1]
(0, 0, 0) if t ∈ (1, 2]
, q(t ) =
¸
0 if t ∈ [0, 1]
1 if t ∈ (1, 2]
with β = 0.
10.7.3 The Lagrangian is L = H ÷qx = −u
2
−x ÷pu÷qx. Here H is concave in (x, u) and h(t, x) = x
is quasiconcave, so the conditions (i)–(vi) below are therefore sufﬁcient for (x
∗
(t ), u
∗
(t )) to be optimal:
(i) u = u
∗
(t ) maximizes −x
∗
(t ) ÷p(t )u −u
2
for u ∈ ޒ;
(ii) q(t ) ≥ 0, with q(t ) = 0 if x
∗
(t ) > 0;
(iii) ˙ p(t ) = −∂L
∗
/∂x = 1 −q(t ), p(10) = 0;
(iv) p(10
−
) −p(10) = β;
(v) β ≥ 0, with β = 0 if x
∗
(10) > 0;
(vi) ˙ x
∗
(t ) = u
∗
(t ), x
∗
(0) = 1.
Since H is concave in u and u ∈ ޒ, (i) is equivalent to (H
/
u
)
∗
= p(t ) − 2u
∗
(t ) = 0, so u
∗
(t ) =
1
2
p(t ).
We have x
∗
(0) = 1. Let [0, t
∗
] be the maximal interval where x
∗
(t ) > 0. Then from (ii), q(t ) = 0,
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
10 CONTROL THEORY WI TH MANY VARI ABL ES 91
and (iii) gives ˙ p(t ) = 1, and thus p(t ) = t ÷ A. Then u
∗
(t ) =
1
2
(t ÷ A) and ˙ x
∗
(t ) =
1
2
(t ÷ A), so
x
∗
(t ) =
1
4
(t ÷ A)
2
÷ B. Since x
∗
(0) = 1, B = −
1
4
A
2
÷ 1. If t
∗
= 10, then p(t ) = t − 10 since
p(10) = 0, and u
∗
(t ) =
1
2
(t − 10) < 0 for t < 10, contradicting u
∗
(t ) ≥ 0. Thus t
∗
< 10 and
x
∗
(t
∗
) = 0, and so x
∗
(t
∗
) =
1
4
(t
∗
÷ A)
2
−
1
4
A
2
÷ 1 = 0. The function x
∗
(t ) must have a minimum at
t
∗
so ˙ x
∗
(t
∗
) =
1
2
(t
∗
÷A) = 0, and so A = −t
∗
< 0. Hence x
∗
(t
∗
) = 1 −
1
4
A
2
= 0, so A = −2. With
p(t ) = t −2 we have u
∗
(t ) =
1
2
(t −2), and x
∗
(t ) =
1
4
(t −2)
2
in [0, 2].
Looking at the objective function, when x
∗
(t ) has become 0, it is obvious that we need to keep
u
∗
(t ) = 0 on (2, 10]. So u
∗
(t ) = x
∗
(t ) = 0, and then p(t ) = 0. Then from (iii), q(t ) = 1 and (iv) gives
β = p(10
−
) = 0 since p(10) = 0. Now all the conditions (i)–(vi) are satisﬁed, so the optimal solution is
(x
∗
(t ), u
∗
(t ), p(t )) =
¸
(
1
4
(t −2)
2
,
1
2
(t −2), t −2) if t ∈ [0, 2]
(0, 0, 0) if t ∈ (2, 10]
, q(t ) =
¸
0 if t ∈ [0, 2]
1 if t ∈ (2, 10]
with β = 0.
10.7.4 (a) The Hamiltonian H = (4 − t )u ÷ pu is concave in (x, u), so the following conditions are
sufﬁcient for (x
∗
(t ), u
∗
(t )) to be optimal:
(i) u = u
∗
(t ) maximizes
p(t ) −(t −4)
u for u ∈ [0, 2];
(ii) ˙ p(t ) = −∂H
∗
/∂x = 0;
(iii) ˙ x
∗
(t ) = u
∗
(t ), x
∗
(0) = 1, x
∗
(3) = 3.
From (ii) we get p(t ) = ¨ p for some constant ¨ p. Condition (i) implies that we must have u
∗
(t ) = 2 if
¨ p > t − 4 and u
∗
(t ) = 0 if ¨ p < t − 4. If ¨ p < t − 4 for all t in [0, 3], then u
∗
(t ) ≡ 0, and from (iii)
we have x
∗
(t ) ≡ 1, contradicting x
∗
(3) = 3. In the same way we see that ¨ p > t − 4 for all t in [0, 3]
is impossible. Hence we have to choose u
∗
(t ) = 2 in some interval [0, t
∗
] and u
∗
(t ) = 0 in (t
∗
, 3], with
t
∗
− 4 = ¨ p. Now from (iii) we have x
∗
(t ) = 2t ÷ 1 in [0, t
∗
], and x
∗
(t ) = 2t
∗
÷ 1 in (t
∗
, 3]. Since
x
∗
(3) = 2t
∗
÷ 1 = 3, we see that t
∗
= 1, and then p(t ) = ¨ p = t
∗
− 4 = −3. It is clear that we have
found the optimal solution, since all the conditions (i)–(iii) are satisﬁed.
(b) The Lagrangian is L = H ÷q(t ÷1 −x) = (4 −t )u ÷pu ÷q(t ÷1 −x). Here H is concave in
(x, u) and h(t, x) = t ÷1 −x is quasiconcave, so the conditions (i) and (iii) in (a) in addition to
(ii)
/
q(t ) ≥ 0, with q(t ) = 0 if t ÷1 > x
∗
(t ),
(iii)
/
˙ p(t ) = −∂L
∗
/∂x = q(t ),
(iv) p(3
−
) −p(3) = −β,
(v) β ≥ 0, with β = 0 if 4 −x
∗
(3) > 0,
are sufﬁcient for (x
∗
(t ), u
∗
(t )) to be optimal.
The objective function indicates that we should keep u
∗
(t ) as large as possible, especially at the
beginning. But having u
∗
(t ) larger than 1 will cause x
∗
(t ) to violate the constraint x ≤ t ÷ 1, so we
suggest u
∗
(t ) = 1, and then x
∗
(t ) = t ÷1 in some interval [0, t
∗
]. Note that according to (i), u
∗
(t ) = 1
can only maximize the Hamiltonian in [0, t
∗
] provided p(t ) = t −4. From (iii)
/
we further get q(t ) = 1
in [0, t
∗
]. With x
∗
(t ) = t ÷1 we get x
∗
(2) = 3, and since x
∗
(3) = 3 and ˙ x
∗
(t ) ≥ 0, we must have t
∗
≤ 2.
In fact, we suggest t
∗
= 2 and then u
∗
(t ) = 0 in (2, 3]. From ˙ x
∗
(t ) ≥ 0, we get x
∗
(t ) ≤ x
∗
(3) = 3 and
then h(t, x
∗
(t )) = t ÷1−x
∗
(t ) ≥ t −2. It follows that h(t, x
∗
(t )) > 0 for t in (2, 3]. But by (ii)
/
we have
q(t ) = 0 in (2, 3]. Then (iii)
/
yields p(t ) = ¨ p for some constant ¨ p. Since p(t ) is continuous at t = 2,
p(2
−
) = 2 −4 = −2 = ¨ p. It remains to determine β. We see that h(3, x
∗
(3)) = 3 ÷1 −3 = 1 > 0, so
from (v) we get β = 0, and (iv) gives p(3
−
) = p(3) = −2.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
92 11 DI F F ERENCE EQUATI ONS
The optimal solution is spelled out the answer in the book. It is a useful exercise to check carefully
that all the conditions (i), (ii)
/
, (iii)
/
, (iv), and (v) are now satisﬁed. Note in particular that for t in (2, 3]
the expression (p(t ) −(t −4))u = (−2 −(t −4))u = (2 −t )u is maximized by u = u
∗
(t ) = 0, since
2 −t is negative in (2, 3].
11 Difference Equations
11.1
11.1.2 In parts (a)–(f) the solution is given by x
t
= a
t
(x
0
− x
∗
) ÷ x
∗
, cf. formula (11.1.5). In (g)–(i) the
solution is x
t
= x
0
÷t b.
(a) Since 0 < a < 1, the power a
t
decreases monotonically towards 0 as a limit as t → ∞. Because
x
0
−x
∗
< 0 it follows that x
t
will increase monotonically towards the limit x
∗
.
(b) a
t
alternates between negative and positive values and tends to 0. We get damped oscillations around
the limit x
∗
.
(c) x
t
increases monotonically and faster and faster towards ∞.
(d) Since a < −1, the powers a
t
alternate between negative and positive values, while [a
t
[ tends to ∞
as t →∞. Thus we get explosive oscillations about x
∗
.
(e) The solution is constant, x
t
= x
∗
for all t .
(f) Oscillations around x
∗
with constant amplitude.
(g) x
t
= x
0
÷t b increases (linearly) towards ∞.
(h) The solution is monotonically (linearly) decreasing towards −∞.
(i) x
t
= x
0
for all t .
11.2
11.2.3 (a) Let the remaining debt on 1 January in year n be L
n
. Then L
0
= L. Since the payment on the
principal in year n is L
n−1
−L
n
and the interest is rL
n−1
, we have L
n−1
−L
n
=
1
2
rL
n−1
, n = 1, 2, . . . .
The solution is L
n
= (1 −
1
2
r)
n
L.
(b) (1 −
1
2
r)
10
L =
1
2
L implies that r = 2 −2 · 2
−1/10
≈ 0.133934
(c) The payment in year n will be L
n−1
− L
n
÷ rL
n−1
=
3
2
rL
n−1
=
3
2
r(1 −
1
2
r)
n−1
L. The loan will
never be completely paid since L
n
> 0 for all n (but it does tend to 0 in the limit as n →∞).
11.2.4 Let r
t
be the interest rate in period t , a
t
the repayment, and b
t
the outstanding balance. Then b
t ÷1
=
(1 ÷r
t
)b
t
−a
t ÷1
, where b
0
= K. We get
b
1
= (1 ÷r
0
)b
0
−a
1
= (1 ÷r
0
)K −a
1
b
2
= (1 ÷r
1
)b
1
−a
2
= (1 ÷r
1
)(1 ÷r
0
)K −(1 ÷r
1
)a
1
−a
2
b
3
= (1 ÷r
2
)b
2
−a
3
= (1 ÷r
2
)(1 ÷r
1
)(1 ÷r
0
)K −(1 ÷r
2
)(1 ÷r
1
)a
1
−(1 ÷r
2
)a
2
−a
3
· · · · · · · · ·
The pattern is clear, and it can be shown by induction that
b
t
=
t −1
¸
s=0
(1 ÷r
s
)K −
t −1
¸
s=1
¸
t −1
¸
k=s
(1 ÷r
k
)a
s
¸
−a
t
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
11 DI F F ERENCE EQUATI ONS 93
It is reasonable to suppose that if the interest rate changes, then the repayments also change to ensure
that b
t ÷1
< b
t
, i.e. a
t ÷1
> r
t
b
t
. If the repayment in period t ÷1 is less than the interest accrued during
period t , i.e. if a
t ÷1
< r
t
b
t
, then the outstanding debt will increase, and if this situation continues, the
loan will never be paid off.
11.3
11.3.1 (a) x
t ÷1
= A÷B 2
t ÷1
= A÷2B 2
t
and x
t ÷2
= A÷B 2
t ÷2
= A÷4B 2
t
, so x
t ÷2
−3x
t ÷1
÷2x
t
=
A ÷4B 2
t
−3A −6B 2
t
÷2A ÷2B 2
t
= 0 for all t .
(Section 11.4 shows how to ﬁnd this solution, in the form x
t
= A1
t
÷B 2
t
.)
(b) With x
t
= A3
t
÷B4
t
we get x
t ÷1
= 3A3
t
÷4B4
t
, x
t ÷2
= 9A3
t
÷16B4
t
, and x
t ÷2
−7x
t ÷1
÷12x
t
=
9A3
t
÷16B4
t
−21A3
t
−28B4
t
÷12A3
t
÷12B4
t
= 0.
11.3.5 We shall prove that
u
(1)
t
and u
(2)
t
are linearly dependent ⇐⇒
u
(1)
0
u
(2)
0
u
(1)
1
u
(2)
1
= 0
Proof of ⇒: If the solutions u
(1)
t
and u
(2)
t
are linearly dependent, then there exist constants c
1
and c
2
, not
both equal to 0, such that c
1
u
(1)
t
÷c
2
u
(2)
t
= 0 for all t . This holds, in particular, for t = 0 and t = 1, and
so the columns of the determinant above are linearly dependent, and the determinant must be 0.
Proof of ⇐: If the determinant is zero, the columns are linearly dependent, so there exist constants c
1
and c
2
, not both 0, such that
c
1
u
(1)
t
÷c
2
u
(1)
t
= 0 (∗)
for t = 0 and for t = 1. Now suppose that (∗) holds for t = 0, 1, . . . , T − 1, where T is some integer
greater than 1. Then
c
1
u
(1)
T
÷c
2
u
(2)
T
= −a
t
¸
c
1
u
(1)
T −1
÷c
2
u
(2)
T −1
¸
−b
t
¸
c
1
u
(1)
T −2
÷c
2
u
(2)
T −2
¸
= 0
so (∗) holds for t = T also. It follows by induction that (∗) holds for all t ≥ 0. Hence, u
(1)
t
and u
(2)
t
are
linearly dependent.
11.3.6 (a) From Problem 2 we can ﬁnd the linearly independent solution u
(1)
t
= 1 and u
(2)
t
= t of the
homogeneous equation x
t ÷2
− 2x
t ÷2
÷ x
t
= 0. Then D
t
= u
(1)
t
u
(2)
t ÷1
− u
(1)
t ÷1
u
(2)
t
= (t ÷ 1) − t = 1 for
all t , and we get
u
∗
t
= −u
(1)
t
t
¸
k=1
c
k−1
u
(2)
k
÷u
(2)
t
t
¸
k=1
c
k−1
u
(1)
k
= −
t
¸
k=1
kc
k−1
÷t
t
¸
k=1
c
k−1
as a particular solution of x
t ÷2
−2x
t ÷1
÷x
t
= c
t
. The general solution is then x
t
= A ÷Bt ÷u
∗
t
.
(b) With c
t
= t , the particular solution u
∗
t
in part (a) becomes
u
∗
t
= −
t
¸
k=1
k(k −1) ÷t
t
¸
k=1
(k −1) = −
1
3
(t −1)t (t ÷1) ÷
1
2
t
2
(t −1) =
1
6
t (t −1)(t −2)
The necessary summation formulas
¸
t
k=1
k(k −1) =
1
3
(t −1)t (t ÷1) and
¸
t
k=1
(k −1) =
1
2
t (t −1) are
easily proved by induction. It is also easy to check that u
∗
t
really is a solution of the difference equation.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
94 11 DI F F ERENCE EQUATI ONS
11.4
11.4.2 (a) The characteristic equation m
2
÷2m÷1 = (m÷1)
2
= 0 has the double root m = −1, so the
general solution of the associated homogeneous equation is x
H
t
= (A ÷ Bt )(−1)
t
. We ﬁnd a particular
solution of the nonhomogeneous equation by inserting u
∗
t
= P2
t
. This yields P = 1, and so the general
solution of the given equation is x
t
= (A ÷Bt )(−1)
t
÷2
t
.
(b) By using the method of undetermined coefﬁcients to determine the constants P, Q, and R in the
particular solution u
∗
t
= P5
t
÷ Qcos
π
2
t ÷ Rsin
π
2
t , we obtain P =
1
4
, Q =
3
10
, and R =
1
10
. So the
general solution to the given equation is x
t
= A ÷B2
t
÷
1
4
5
t
÷
3
10
cos
π
2
t ÷
1
10
sin
π
2
t .
11.4.4 Since 1÷a ÷b = 0, we have b = −1−a, so we are looking for a particular solution of the equation
x
t ÷2
÷ax
t ÷1
−(a ÷1)x
t
= c. A constant function will be a solution of the corresponding homogeneous
function, so that will not work (unless c = 0). Let us try a function of the form u
∗
t
= Dt . We get
u
∗
t ÷2
÷au
∗
t ÷1
−(a ÷1)u
∗
t
= D(t ÷2) ÷aD(t ÷1) −(a ÷1)Dt = D(a ÷2)
Thus, D = c/(a÷2) can be used unless a = −2. If a = −2, the difference equation is x
t ÷2
−2x
t ÷1
÷x
t
=
c, and we look for a particular solution of the form u
∗
t
= Dt
2
. In this case we get
u
∗
t ÷2
−2u
∗
t ÷2
÷u
∗
t
= D(t ÷2)
2
−2D(t ÷1)
2
÷Dt
2
= 2D
and the desired value of D is D = c/2.
11.4.6 If b =
1
4
a
2
and x
t
= u
t
(−a/2)
t
, then the lefthand side of equation (11.4.1) becomes
x
t ÷2
÷ax
t ÷1
÷
1
4
a
2
x
t
= u
t ÷2
(−a/2)
t ÷2
÷au
t ÷1
(−a/2)
t ÷1
÷
1
4
a
2
u
t
(−a/2)
t
=
1
4
a
2
(−a/2)
t
(u
t ÷2
−2u
t ÷1
÷u
t
)
which is 0 if u
t ÷2
− 2u
t ÷1
÷ u
t
= 0. The general solution of this equation is u
t
= A ÷ B t , so
x
t
= u
t
(−a/2)
t
= (A ÷B t )(−a/2)
t
, which is the result claimed for case II in Theorem 11.4.1.
11.4.9 (a) It seems natural to try a function of the form Y
∗
t
= C(1 ÷g)
t
. We get
Y
∗
t ÷2
−(b÷k)Y
t ÷1
÷kY
∗
t
= C(1÷g)
t
[(1÷g)
2
−(b÷k)(1÷g)÷k] = C(1÷g)
t
[(1÷g)
2
−b(1÷g)−kg]
This shows that Y
∗
t
=
a(1 ÷g)
t
(1 ÷g)
2
−(b ÷k)(1 ÷g) ÷k
(if the denominator of this fraction is nonzero).
(b) The equation m
2
−(b ÷k)m÷k = 0 has two complex roots if and only if (b ÷k)
2
−4k < 0.
(c) Part (III) of Theorem11.4.1 shows that the growth factor of the oscillations is r =
√
k. The oscillations
are damped if and only if r < 1, i.e. if and only if k < 1.
11.4.11 Claim: If
1
4
a
2
≥ b, then both roots of the equation f (m) = m
2
÷ am ÷ b = 0 lie in the interval
(−1, 1) if and only if [a[ < 1 ÷b and b < 1.
Proof: Both roots belong to (−1, 1) if and only if f (−1) > 0, f (1) > 0, f
/
(−1) < 0, and f
/
(1) > 0.
(Draw a picture!) These four inequalities are equivalent to 1 −a ÷b > 0, 1 ÷a ÷b > 0, −2 ÷a < 0,
and a ÷2 > 0, which in turn are equivalent to [a[ < 1 ÷b and [a[ < 2.
If [a[ < 2, then b ≤
1
4
a
2
< 1. On the other hand, if [a[ < 1 ÷b and b < 1, then [a[ < 2.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
11 DI F F ERENCE EQUATI ONS 95
11.5
11.5.3 The points (a
1
, a
2
) that satisfy the three inequalities in (∗∗) are the points that lie above each of the
lines given by a
2
= −1 − a
1
and a
2
= −1 ÷ a
1
and below the line a
2
= 1. These points are precisely
the points lying in the interior of the triangle formed by those three lines, i.e. the triangle with corners at
(−2, 1), (0, −1), and (2, 1).
11.5.5 The characteristic equation is m
2
÷ (σβ/α − 2)m ÷ (1 − σβ) = 0. A necessary and sufﬁcient
condition for the roots of this equation to be complex (more precisely: not real) is
σβ
α
−2
2
< 4(1 −σβ) ⇐⇒
σ
2
β
2
α
2
−
4σβ
α
÷4 < 4 −4σβ ⇐⇒ σ
2
β
2
< 4ασβ −4α
2
σβ
⇐⇒ σβ < 4α −4α
2
= 4α(1 −α)
The difference equation is globally asymptotically stable if and only if the inequalities (∗∗) in Section
11.5 are satisﬁed when a
1
= σβ/α −2 and a
2
= 1 −σβ. This gives the conditions
1 ÷
σβ
α
−2
÷1 −σβ > 0 and 1 −
σβ
α
−2
÷1 −σβ > 0 and σβ > 0
⇐⇒
σβ
α
−σβ > 0 and 4 >
σβ
α
÷σβ
⇐⇒ α < 1 and (1 ÷α)σβ < 4α
(Remember that α, β, and σ are all positive.)
11.6
11.6.1 (a) From the given equations we get x
t ÷2
= 2y
t ÷1
= x
t
, and the initial conditions give x
0
= 1,
x
1
= 2y
0
= 2. The characteristic equation of x
t ÷2
−x
t
= 0 is m
2
−1 = 0, which has the roots m
1
= 1,
m
2
= −1.
The general solution of x
t ÷2
−x
t
= 0 is x
t
= A÷B(−1)
t
, and the initial conditions imply A÷B = 1,
A−B = 2, so A =
3
2
and B = −
1
2
. Thus, x
t
=
3
2
−
1
2
(−1)
t
. This, in turn, gives y
t
=
1
2
x
t ÷1
=
3
4
÷
1
4
(−1)
t
.
(b) We ﬁrst eliminate z. The ﬁrst equation yields z
t
= −x
t ÷1
−y
t
÷1. Using this in the second and third
equations, we get
(i) y
t ÷1
= −x
t
÷x
t ÷1
÷y
t
−1 ÷t and (ii) −x
t ÷2
−y
t ÷1
÷1 = −x
t
−y
t
÷2t
Equation (ii) implies (iii) y
t ÷1
= −x
t ÷2
÷1 ÷x
t
÷y
t
−2t , and then (i) and (iii) imply
−x
t
÷x
t ÷1
÷y
t
−1 ÷t = −x
t ÷2
÷1 ÷x
t
÷y
t
−2t, and so (iv) x
t ÷2
÷x
t ÷1
−2x
t
= 2 −3t
By a stroke of good luck y does not appear in (iv). The characteristic equation of (iv) is m
2
÷ m − 2
with the roots m
1
= 1 and m
2
= −2, and so the homogeneous equation corresponding to (iv) has the
general solution x
H
t
= A ÷ B(−2)
t
. For a particular solution u
∗
t
of (iv) itself we try with a quadratic
polynomial u
∗
= Dt ÷ Et
2
. We get u
∗
t
=
3
2
t −
1
2
t
2
, so the general solution of (iv) is x
t
= x
H
t
÷ u
∗
t
=
A ÷B(−2)
t
÷
3
2
t −
1
2
t
2
.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
96 12 DI SCRETE TI ME OPTI MI ZATI ON
The initial conditions yield x
0
= 0 and x
1
= −y
0
− z
0
÷ 1 = 0, so we must have A ÷ B =
0, A −2B ÷
3
2
−
1
2
= 0, which implies, A = −
1
3
, B =
1
3
. Hence,
x
t
= −
1
3
÷
1
3
(−2)
t
÷
3
2
t −
1
2
t
2
Fromequation(i) we get y
t ÷1
−y
t
= x
t ÷1
−x
t
÷t −1 = −(−2)
t
withthe general solutiony
t
= A÷
1
3
(−2)
t
.
The initial condition y
0
= 0 yields A = −
1
3
, so
y
t
= −
1
3
÷
1
3
(−2)
t
Finally,
z
t
= −x
t ÷1
−y
t
÷1 =
2
3
÷
1
3
(−2)
t
−
1
2
t ÷
1
2
t
2
11.7
11.7.2 (a) Let f (x) = e
x
− 3. Then f (x) − x is convex, and it is easy to see from the intermediate
value theorem that the equation f (x) = x has two solutions, one in (−∞, 0) and one in (0, ∞). Since
[f
/
(x)[ = e
x
< 1 for all x < 0, the negative solution is a stable equilibrium of the difference equation
x
t ÷1
= f (x
t
). The solution of the difference equation starting at x
0
= −1 gives the values (rounded to
5 decimal places) x
1
= −2.63212, x
2
= −2.92807, x
3
= −2.94650, x
4
= −2.94748, x
5
= −2.94753,
x
6
= −2.94753, . . . , converging to the equilibrium value x
∗
≈ −2.94753.
(b) See the answer in the book.
12 Discrete Time Optimization
12.1
12.1.1 (a) To solve the problem by dynamical programming we ﬁrst ﬁnd
J
2
(x) = max
u
(1 −(x
2
÷2u
2
)) = 1 −x
2
, u
∗
2
(x) = 0
The fundamental equation (Theorem 12.1.1) then gives
J
1
(x) = max
u
(1 −(x
2
÷2u
2
) ÷J
2
(x −u))
= max
u
(1 −x
2
−2u
2
÷1 −(x −u)
2
)
= max
u
(2 −2x
2
÷2xu −3u
2
. .. .
g(u)
)
Let g(u) be the expression in the last parenthesis. Then g
/
(u) = 2x −6u, so g attains its maximum for
u = u
∗
1
(x) = x/3. That gives J
1
(x) = g(x/3) = 2 −
5
3
x
2
. We continue with
J
0
(x) = max
u
(1 −(x
2
÷2u
2
) ÷J
1
(x −u))
= · · · = max
u
(3 −
8
3
x
2
÷
10
3
xu −
11
3
u
2
. .. .
h(u)
)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
12 DI SCRETE TI ME OPTI MI ZATI ON 97
Here h
/
(u) = 10x/3 − 22u/3, which implies that u
∗
0
(x) = 5x/11, and therefore J
0
(x) = h(5x/11) =
3 −21x
2
/11.
Thus the desired maximum value is J
0
(x
0
) = J
0
(5) = −492/11 and, further,
u
∗
0
= u
0
(5) = 25/11,
u
∗
1
= u
∗
1
(x
∗
1
) = x
∗
1
/3 = 10/11,
x
∗
1
= x
0
−u
∗
0
= 30/11
x
∗
2
= x
∗
1
−u
∗
1
= 20/11
(b) We have x
1
= x
0
−u
0
= 5 −u
0
and x
2
= x
1
−u
1
= 5 −u
0
−u
1
. This gives
S(u
0
, u
1
, u
2
) = 1 −(x
2
0
÷2u
2
0
) ÷1 −(x
2
1
÷2u
2
1
) ÷1 −(x
2
2
÷2u
2
2
)
= 3 −5
2
−2u
2
0
−(5 −u
0
)
2
−2u
2
1
−(5 −u
0
−u
1
)
2
−2u
2
2
= · · · = −72 ÷20u
0
÷10u
1
−4u
2
0
−2u
0
u
1
−3u
2
1
−2u
2
2
It is clear from the second expression for S that S is a concave function. The ﬁrstorder partial derivatives
of S are
∂S/∂u
0
= 20 −8u
0
−2u
1
∂S/∂u
1
= 10 −2u
0
−6u
1
∂S/∂u
0
= −4u
2
and it is easily seen that the only stationary point is
(u
0
, u
1
, u
2
) = (25/11, 10/11, 0)
Since S is concave, this is a global maximum point for S. The maximum value is S
max
= −492/11,
which fortunately agrees with the result from part (a).
12.1.4 (a) J
T
(x) = 3x
2
with u
∗
T
(x) = 0, J
T −1
(x) = 5x
2
with u
∗
T −1
(x) = 1, J
T −2
(x) = 7x
2
with
u
∗
T −2
(x) = 1.
(b) We claim that J
T −n
(x) = (2n ÷ 3)x
2
with u
∗
T
(x) = 0 and u
∗
T −n
(x) = 1 for n = 1, . . . , T . The
formula is valid for n = 1. Suppose it is valid for n = k. Then J
T −(k÷1)
(x) = max
u∈[0,1]
[(3 − u)x
2
÷
J
T −k
(ux)] = max
u∈[0,1]
[(3 −u)x
2
÷(2k ÷3)(ux)
2
] = x
2
max
u∈[0,1]
[3 −u÷(2k ÷3)u
2
]. The function
g(u) = 3−u÷(2k÷3)u
2
is convexinuandhas its maximumat u = 1, andthenJ
T −(k÷1)
(x) = x
2
(5÷2k),
which is the proposed formula for n = k ÷1, so the formula follows by induction.
12.1.6 (a) J
T
(x) = max
u∈ޒ
(x −u
2
) = x for u
∗
T
(x) = 0. J
s
(x) = max
u∈ޒ
[x −u
2
÷J
s÷1
(2(x ÷u))] for
s = 0, 1, . . . , T − 1. In particular, J
T −1
(x) = max
u∈ޒ
[x − u
2
÷ J
T
(2(x ÷ u))] = max
u∈ޒ
[x − u
2
÷
2(x ÷u)] = 3x ÷1 for u
∗
T −1
(x) = 1.
(b) The formula is valid for n = 1. Suppose it is valid for n = k. Then J
T −(k÷1)
(x) = max
u∈ޒ
[x −u
2
÷
(2
k÷1
−1)(2x ÷2u) ÷
¸
k
j=0
(2
j
−1)
2
]. We see that the maximizer is u = u
∗
T −(k÷1)
(x) = 2
k÷1
−1, and
then J
T −(k÷1)
(x) = x −(2
k÷1
−1)
2
÷2(2
k÷1
−1)x ÷2(2
k÷1
−1)
2
÷
¸
k
j=0
(2
j
−1)
2
= (1 ÷2
k÷2
−
2)x ÷(2
k÷1
−1)
2
÷
¸
k
j=0
(2
j
−1)
2
= (2
(k÷1)÷1
−1)x ÷
¸
k÷1
j=0
(2
j
−1)
2
. This is the given formula for
n = k ÷1, so the formula follows by induction. Since u
∗
T −(k÷1)
(x) = 2
k÷1
−1, we get u
∗
t
(x) = 2
T −t
−1
for t = 0, 1, . . . , T , and V = J
0
(x
0
) = J
0
(0) =
¸
T
j=0
(2
j
−1)
2
.
12.1.7 (a) It is immediately clear that J
T
(x) = −αe
−γ x
T
. The result in part (b) shows immediately that
J
T −1
(x) = −2
√
αe
−γ x
and J
T −2
(x) = −2
2
√
αe
−γ x
= −2
3/2
α
1/4
e
−γ x
.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
98 12 DI SCRETE TI ME OPTI MI ZATI ON
(b) We know from part (a) that the formula J
t
(x) = −α
t
= −α
t
e
−γ x
holds for t = T with α
t
= α.
Suppose that it holds for a positive integer t ≤ T . The fundamental equation then gives J
t −1
(x) =
max
u∈ޒ
ϕ(u), where
ϕ(u) = −e
−γ u
÷J
t
(2x −u) = −e
−γ u
−α
t
e
−γ (2x−u)
The function ϕ is concave and
ϕ
/
(u) = 0 ⇐⇒ γ e
−γ u
−α
t
γ e
−γ (2x−u)
= 0 ⇐⇒ −γ u = ln α
t
−2γ x ÷γ u
⇐⇒ γ u = γ x −ln
√
α
t
This shows that ϕ has a unique stationary point u
∗
= x −(ln
√
α
t
)/γ , which is a maximum point for ϕ.
It follows that
J
t −1
(x) = ϕ(u
∗
) = −e
−γ u
∗
−α
t
e
−2γ x÷γ u
∗
= −e
ln
√
α
t
−γ x
−α
t
e
−γ x−ln
√
α
t
= −α
t −1
e
−γ x
where α
t −1
=
√
α
t
÷ (α
t
/
√
α
t
) = 2
√
α
t
. It follows by induction that the formula J
t
(x) = −α
t
e
−γ x
holds for t = T , T −1, . . . , 1, 0, with α
t
determined by the difference equation above and α
T
= α.
12.3
12.3.1 The equation α = 2
√
αβ ÷
1
2
can be written as (
√
α )
2
− 2
√
β
√
α −
1
2
= 0. This is a quadratic
equation for
√
α with the solution
√
α =
√
β ÷
√
β ÷1/2. (We cannot have
√
α =
√
β −
√
β ÷1/2,
because
√
α cannot be negative.) Hence, α = (
√
β ÷
√
β ÷1/2 )
2
.
To showoptimality, we can use Case Bin Note 12.7.3. We ﬁrst solve the corresponding ﬁnite horizon
problem
sup
u
t
T
¸
t =0
β
t
(−e
−u
t
−
1
2
e
−x
t
), x
t ÷1
= 2x
t
−u
t
, t = 0, 1, . . . , T −1, x
0
given
With the optimal value function J(t, x, T ) = sup
¸
T
s=t
β
s−t
(−e
−u
s
−
1
2
e
−x
s
) we get J(t, x, T ) =
−α
t
e
−x
, with α
T
=
1
2
and α
t
= 2
√
βα
t
÷
1
2
for t < T . For a ﬁxed T , α
t
→ α as t → −∞. Hence,
J(0, x, T ) →−αe
−x
as T →∞.
12.3.2 (a) The Bellman equation is J(x) = max
u∈ޒ
¸
−
2
3
x
2
− u
2
÷ βJ(x ÷ u)
¸
. If J(x) = −αx
2
is a
solution, then
−αx
2
= max
u∈ޒ
¸
−
2
3
x
2
−u
2
÷βJ(x ÷u)
¸
= max
u∈ޒ
ϕ(u) (∗)
where ϕ(u) = −
2
3
x
2
−u
2
−βα(x ÷u)
2
. The function ϕ is strictly concave, and has a unique maximum
point given by ϕ
/
(u) = −2u − 2βα(x ÷ u) = 0. Hence, u
∗
(x) = −βαx/(1 ÷ βα) and x ÷ u
∗
(x) =
x/(1 ÷βα). Equation (∗) now gives
−αx
2
= ϕ(u
∗
(x)) = −
2
3
x
2
−
β
2
α
2
(1 ÷βα)
2
x
2
−
βα
(1 ÷βα)
2
x
2
= −
2
3
x
2
−
β
2
α
2
÷βα
(1 ÷βα)
2
x
2
⇐⇒ α =
2
3
÷
βα(βα ÷1)
(1 ÷βα)
2
=
2
3
÷
βα
1 ÷βα
⇐⇒ 3α(1 ÷βα) = 2(1 ÷βα) ÷3βα ⇐⇒ 3βα
2
÷(3 −5β)α −2 = 0
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
12 DI SCRETE TI ME OPTI MI ZATI ON 99
The last equation has a exactly one positive solution, α =
5β −3 ÷
(5β −3)
2
÷24β
6β
, and the optimal
control is
u
∗
(x) = −
βα
1 ÷βα
x = −
5β −3 ÷
(5β −3)
2
÷24β
5β ÷3 ÷
(5β −3)
2
÷24β
x = · · · = −
β −3 ÷
(5β −3)
2
÷24β
6β
x
(b) It is clear that the value function J
0
(x) is ﬁnite: for any x
0
, let u
0
= −x
0
and u
t
= 0 for all t > 0.
Then
¸
∞
t =0
β
t
(−
2
3
x
2
t
−u
2
t
) = −
5
3
x
2
0
is ﬁnite. Also, for any sequence of controls u
0
, u
1
, . . . , the sum is
bounded above by 0. Hence, the value function exists, if not in the “max sense”, then at least in the “sup
sense”. It is also clear that J
0
(x) ≤ 0 for all x.
Further, let V
0
(x, π) be the sumthat results fromx = x
0
and the control sequence π = (u
0
, u
1
, . . . ).
Then for any number λ we get V
0
(λx, λπ) = λ
2
V
0
(x, π), and it follows that J
0
(λx) = λ
2
J
0
(x)—that is,
J = J
0
is homogeneous of degree 2, so we really do have J(x) = −αx
2
for a suitable α ≥ 0. It is also
clear that α ,= 0, and from the arguments above it follows that α <
5
3
.
Now let x be ﬁxed, and consider the problem of ﬁnding a u that maximizes
ϕ(u) = −
2
3
x
2
−u
2
÷J(x ÷u)
If [x ÷u[ > [x[, then J(x ÷u) < J(x) and
ϕ(u) = −
2
3
x
2
−u
2
÷βJ(x ÷u) < −
2
3
x
2
−u
2
÷βJ(x) < −
2
3
x
2
÷βJ(x) = ϕ(0)
Hence, such a u cannot be optimal. Also, if [u[ > [x[, then J(x ÷u) < 0 and
ϕ(u) = −
2
3
x
2
−u
2
÷βJ(x ÷u) < −
2
3
x
2
−u
2
< −
5
3
x
2
= ϕ(−x)
so this u cannot be optimal either. It follows that an optimal u must be such that [x ÷ u[ ≤ [x[ and
[u[ ≤ [x[. Then Note 12.3.2 applies with X(x
0
) = [−[x
0
[, [x
0
[].
12.4
12.4.2 (a) I =
¸
T
t =0
(u
2
t
−2x
2
t
) =
¸
T −1
t =0
u
2
t
÷u
2
T
−2x
2
0
−2
¸
T
t =1
x
2
t
= u
2
T
÷
¸
T −1
t =0
u
2
t
−2
¸
T −1
t =0
u
2
t
=
u
2
T
−
¸
T −1
t =0
u
2
t
. (Remember, x
0
= 0.) Hence, I is maximized when u
∗
0
= u
∗
1
= · · · = u
∗
T −1
= 0 and
u
∗
T
= ±1.
(b) (The reference in the problem should be to Theorem 12.4.1, not 12.4.2.) The Hamiltonian is
H(t, x, u, p) =
¸
u
2
−2x
2
÷pu for t < T
u
2
−2x
2
for t = T
and
H
/
u
(t, x, u, p) =
¸
2u ÷p for t < T
2u for t = T
, H
/
x
(t, x, u, p) = −4x for all t
Since x
∗
0
= x
∗
1
= · · · = x
∗
T
= 0 and u
∗
0
= u
∗
1
= · · · = u
∗
T −1
= 0, the difference equation (4) in
Theorem 12.4.1 implies p
t
= 0 for t = 0, 1, . . . , T − 1, and we already know that p
T
= 0. It follows
that H(t, x
∗
t
, u, p
t
) = u
2
− 2(x
∗
t
)
2
= u
2
. Hence, for t < T , u = u
∗
t
= 0 is not a maximum point of
H(t, x
∗
t
, u, p
t
) for u in [−1, 1], but actually a minimum point.
12.5
12.5.1 (c) H(t, x, u, p) = 1 ÷x −y −2u
2
−v
2
÷p
1
(x −u) ÷p
2
(y ÷v) for t = 0, 1, H(t, x, u, p) =
1 ÷x −y −2u
2
−v
2
for t = 2. Condition (3) yields for t = 0, 1: −4u
t
−p
1
t
= 0 and −2v
t
÷p
2
t
= 0.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
100 12 DI SCRETE TI ME OPTI MI ZATI ON
For t = 2 it yields −4u
2
= 0 and −2v
2
= 0. Hence, u
0
= −
1
4
p
1
0
, u
1
= −
1
4
p
1
1
, u
2
= 0, v
0
=
1
2
p
2
0
,
v
1
=
1
2
p
2
1
, v
2
= 0. From (4) and (6)(c
/
), p
1
0
= 1 ÷ p
1
1
, p
1
2
= 0, p
2
0
= −1 ÷ p
2
1
, p
2
2
= 0. Moreover,
from (5), p
1
1
= 1 ÷ p
1
2
, p
2
1
= −1 ÷ p
2
2
. Finally, x
1
= x
0
− u
0
= 5 − u
0
, x
2
= x
1
− u
1
, y
1
= y
0
÷ v
0
,
y
2
= y
1
÷v
1
. From these equations we ﬁnd the same solution as before.
12.6
12.6.6 There are two misprints in the objective function: the expression inside the square brackets should
be
T −1
¸
t =0
u
1/2
t
÷aX
1/2
T
. There is also a misprint in the answer in the book.
The optimality equation (12.6.5) boils down to
J
T
(x) = ax
1/2
and J
t
(x) = max
u
¸
u
1/2
÷
1
2
J
t ÷1
(0) ÷
1
2
J
t ÷1
(x −u)
¸
for t < T
With J
t
(x) = 2a
t
x
1/2
this gives a
T
= a/2 and
2a
t
x
1/2
= max
u
ϕ(u), where ϕ(u) = u
1/2
÷a
t ÷1
(x −u)
1/2
for t < T . The function ϕ is concave and
ϕ
/
(u) = 0 ⇐⇒
1
2u
1/2
=
a
t ÷1
2(x −u)
1/2
⇐⇒ u = u
t
(x) =
x
1 ÷a
2
t ÷1
Hence, max
u
ϕ(u) = ϕ
x/(1 ÷ a
2
t ÷1
)
= · · · = (1 ÷ a
2
t ÷1
)
1/2
x
1/2
, and so a
t
=
1
2
(1 ÷ a
2
t ÷1
)
1/2
. This
implies 1 ÷a
2
t ÷1
= 4a
2
t
, and therefore u
t
(x) = x/4a
2
t
.
12.6.8 The ﬁrst printing of the book contains a couple of embarrassing misprints in connection with this
problem and with the stochastic Euler equation. Formula (12.6.10) on page 454 should be
E
¸
F
/
2
(t, x
t
, x
t ÷1
(x
t
, V
t
), V
t
) [ v
t −1
¸
÷F
/
3
(t −1, x
t −1
, x
t
, v
t −1
) = 0 (∗)
and the objective function in the current problem should be
max E
¸
2
¸
t =0
¸
1 −(V
t
÷X
t ÷1
−X
t
)
2
¸
÷(1 ÷V
2
÷X
3
)
¸
Now deﬁne F by
F(2, x
2
, x
3
, v
2
) = 1 −(v
2
÷x
3
−x
2
)
2
÷1 ÷v
2
÷x
3
F(t, x
t
, x
t ÷1
, v
t
) = 1 −(v
t
÷x
t ÷1
−x
t
)
2
, t = 0, 1
For t = 1, 2, equation (∗) becomes
E
¸
2(V
t
÷x
t ÷1
(x
t
, V
t
) −x
t
) [ v
t −1
¸
−2(v
t −1
÷x
t
−x
t −1
) = 0
i.e.
1 ÷2E[x
t ÷1
(x
t
, V
t
)] −2x
t
= 2v
t −1
÷2x
t
−2x
t −1
(∗∗)
Equation (12.6.9) becomes
F
/
3
(2, x
2
, x
3
, v
2
) = −2(v
2
÷x
3
−x
2
) ÷1 = 0
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
12 DI SCRETE TI ME OPTI MI ZATI ON 101
which gives x
3
as a function of x
2
and v
2
:
x
3
= x
3
(x
2
, v
2
) = x
2
−v
2
÷
1
2
Now use (∗∗) for t = 2:
1 ÷2E[x
2
−V
2
÷
1
2
] −2x
2
= 2v
1
÷2x
2
−2x
1
⇐⇒ 1 ÷2(x
2
−
1
2
÷
1
2
) −2x
2
= 2v
1
÷2x
2
−2x
1
⇐⇒ x
2
= x
2
(x
1
, v
1
) = x
1
−v
1
÷
1
2
Then for t = 1:
1 ÷2E[x
1
−V
1
÷
1
2
] −2x
1
= 2v
0
÷2x
1
−2x
0
⇐⇒ 1 ÷2(x
1
−
1
2
÷
1
2
) −2x
1
= 2v
0
÷2x
1
−2x
0
⇐⇒ x
1
= x
0
−v
0
÷
1
2
Since x
0
= 0 is given, the ﬁnal answer is
x
1
=
1
2
−v
0
, x
2
= 1 −v
0
−v
1
, x
3
=
3
2
−v
0
−v
1
−v
2
12.7
12.7.1 (a) J(x) = ax
2
÷b, a = −[1 −2β −
1 ÷4β
2
]/2β, b = aβd/(1 −β). (With J(x) = ax
2
÷b,
the Bellman equation is
ax
2
÷b = max
u
¸
−u
2
−x
2
÷βE[a(x ÷u ÷V)
2
÷b]
¸
= max
u
¸
−u
2
−x
2
÷βa(x ÷u)
2
÷βad ÷βb
¸
Maximizing this concave function yields u = βax/(1−βa). Thus, ax
2
÷b = −β
2
a
2
x
2
/(1 −βa)
2
−x
2
÷
βax
2
/(1 −βa)
2
÷βad÷βb = x
2
(2βa−1)/(1−βa)÷βad÷βb for all x. Hence, a = (2βa−1)/(1−βa)
and b = βad ÷ βb. This quadratic equation gives a = [1 − 2β −
1 ÷4β
2
]/2β, and then b =
aβd/(1−β). We have to choose the negative solution for a, because J(x) = ax
2
÷b = a[x
2
÷βd/(1−β)]
must be negative.)
(b) J(t, x) = β
t
(a
t
x
2
÷b
t
), a
t −1
= −1 −β
2
a
2
t
/(1 −βa
t
)
2
÷βa
t
/(1 −βa
t
)
2
= −1 ÷βa
t
/(1 −βa
t
),
a
T
= −1, b
t −1
= βb
t
÷ βa
t
d, b
T
= 0. To ﬁnd lim
T →∞
J(0, x
0
, T ) we need to ﬁnd lim
T →∞
a
0
and lim
T →∞
b
0
(for any t , a
t
and b
t
depend on T ), write in particular a
0
= a
T
0
, b
0
= b
T
0
. Finding
these limits is the same as ﬁnding the limits lim
t →−∞
a
t
, lim
t →−∞
b
t
when T is ﬁxed. The function
ϕ(x) = −1 ÷ βx/(1 − βx) is increasing (calculate its derivative), and, since a
T −1
< a
T
and this
continues backwards, we get a
t −1
< a
t
for all t . Letting t → −∞in the difference equation for a
t
, we
ﬁnd that a = lim
t →−∞
a
t
satisﬁes a = −1 ÷βa/(1 −βa) = (2βa −1)/(1 −βa) (so a > −∞), in fact
a has the same value as in part (a). In a similar way, b
t
decreases when t decreases, and taking limits in
the equation for b
t −1
, we ﬁnd that b = lim
t →−∞
b
t
satisﬁes b = βb ÷ βad, i.e. b is also as in part (a).
Then, evidently, J(0, x, T ) = a
T
0
x
2
÷b
T
0
→ax
2
÷b = J(x) as T →∞.
12.7.2 The optimal control is u
t
(x) = x/(1 ÷ αa), and the value function is J(x) = a ln x ÷ b, where
a = 2/(1 −α), b = [αd ÷αa ln(αa) −(1 ÷αa) ln(1 ÷αa)](1 −α)
−1
, and d = E[ln V].
We can showoptimality at least in a restricted problem: Assume V ∈ [0, δ] for some (perhaps large) δ,
and restrict u to belong to [ε, x
t
] for some small positive ε. Note that X
t
≥ x
0
. Then [f [ = [(x −u)V[ ≤
δx. Choose b > 0 so small that αδ
b
< 1. For x ≥ x
0
, g = ln u ÷ ln X ∈ [ln ε ÷ ln x
0
, 2 ln x] ⊆
[ln ε ÷ ln x
0
, ln a ÷ x
b
], where a is chosen so large that ln x/x
b
≤ 1 when x ≥ a (by l’Hôpital’s rule
lim
x→∞
ln x/x
b
= 0). Now apply Note 12.7.2.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
102 13 TOPOL OGY AND SEPARATI ON
13 Topology and Separation
13.1
13.1.9 (a) If x ∈ int(S), then there is an open ball B around x such that B ⊆ S. But then B ⊆ T as well, so
x is an interior point of T . If y ∈ cl(S), then every open ball around y has a nonempty intersection with
S. Obviously any such ball also meets T , and so y belongs to the closure of T .
(b) Let y be a point in the closure of cl(S). We want to show that y ∈ cl(S). In order to show this, it is
sufﬁcient to show that every open ball around y intersects S. Thus, let B be an open ball around y. Since
y is in the closure of cl(S), the intersection B ∩cl(S) is nonempty. Let z be any point in this intersection.
Since z ∈ B and B is open, there is an open ball B
/
around z such that B
/
⊆ B. And since z ∈ cl(S),
there is at least one point w in B
/
∩ S. Then w ∈ B ∩ S, and so B ∩ S ,= ∅. Hence, cl(S) is closed.
13.1.15 (a) False. Since S ⊆ S, it is clear that int(S) ⊆ int(S ). But we do not always have equality, as
we can see from the following example: Let S = {x ∈ ޒ
n
: 0 < x < 1}. This set (a “punctured”
open ball) is obviously open, so int(S) = S. But its closure is S = {x ∈ ޒ
n
: x ≤ 1} whose interior,
int(S ) = {x ∈ ޒ
n
: x < 1}, also contains the centre of the ball, so int(S) ,= int(S). (Draw a picture
for the case n = 2!)
(b) True. Every set is contained in its closure, so S ⊆ S and T ⊆ T . Therefore, S ∪ T ⊆ S ∪ T . By
Theorem 13.1.2(c), the set S ∪T is closed. Therefore, S ∪ T ⊆ S ∪T (cf. Problem 10(b)). On the other
hand, S ⊆ S ∪ T , so by Problem 9(a), S ⊆ S ∪ T . Similarly, T ⊆ S ∪ T , and so S ∪ T ⊆ S ∪ T . Since
each of the sets S ∪ T and S ∪ T is a subset of the other, they must be equal.
(c) False. Consider, for example, the punctured open ball S in part (a). The boundary, ∂S, of that set
consists of the origin and all the points on the sphere x = 1, so ∂S is not a subset of S. In fact, for any
set T , we have ∂T ⊆ T if and only if T is closed.
(d) True. Let x ∈ S ∩T . We shall show that x ∈ S ∩ T . Let U be an arbitrary open ball centred at x. It is
enough to showthat U intersects S∩T . Since x ∈ S and S is open, there exists an open ball V centred at x
such that V ⊆ S. Then W = U ∩V is also an open ball centred at x and W ⊆ S. (In fact, W is the smaller
of the two balls U and V.) Now, W ∩T ,= ∅ since x ∈ T . Moreover, W ∩T = U ∩V ∩T ⊆ U ∩S ∩T ,
so it follows that U ∩ (S ∩ T ) is indeed nonempty.
13.2
13.2.5 Suppose for a contradiction that the sequence {x
k
} does not converge to x
0
. Then there exists an open
ball B = B
r
(x
0
) around x
0
such that x
k
/ ∈ B
r
(x
0
) for inﬁnitely many k. These x
k
form a subsequence
{x
k
j
} of the original sequence, and they all belong to the set A = X`B = X∩(ޒ
n
`B). Since X is closed
and B is open, A is closed. Because A is contained in the bounded set X, A is also bounded. Hence A is
compact. Therefore {x
k
j
} has a convergent subsequence, converging to a point y in A. But this convergent
subsequence is also a subsequence of the original sequence, and so it should converge to x
0
. But that is
impossible because y ,= x
0
. This contradiction shows that {x
k
} must converge to x
0
after all.
13.2.6 There is an obvious way to identify ޒ
m
ޒ
n
with ޒ
m÷n
: we let the point (x, y) = ((x
1
, . . . , x
m
),
(y
1
, . . . , y
n
)) in ޒ
m
ޒ
n
correspond to the point (x
1
, . . . , x
m
, y
1
, . . . , y
n
) in ޒ
m÷n
. Now let A and B be
compact subsets of ޒ
m
and ޒ
n
, respectively. We shall use the Bolzano–Weierstrass theorem (Theorem
13.2.5) to show that A B is compact.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
13 TOPOL OGY AND SEPARATI ON 103
Let {(a
k
, b
k
)}
k
be a sequence in A B. Since A is compact, the sequence {a
k
}
k
has a convergent
subsequence {a
k
j
}
j
, and since B is compact, {b
k
j
}
j
has a convergent subsequence {b
k
j
i
}
i
. Let a
/
i
= a
k
j
i
and b
/
i
= b
k
j
i
. Then {(a
/
i
, b
/
i
)}
i
is a subsequence of {(a
k
, b
k
}
k
, and the sequences {a
/
i
}
i
and {b
/
i
}
i
are
both convergent. Since a sequence in ޒ
p
converges if and only if it converges componentwise (Theorem
13.2.1), it follows that {(a
/
i
, b
/
i
)}
i
is convergent. Hence, A B is compact.
13.3
13.3.7 Let A ⊆ S ⊆ ޒ
n
, and consider the following two statements:
(a) A is relatively closed in S.
(b) Whever a sequence {x
k
} in A converges to a limit x
0
∈ S, then x
0
∈ A.
Claim: (a) ⇐⇒ (b).
Proof of ⇒: Since A is relatively closed in S we have A = S ∩ F, where F is closed in ޒ
n
. Let {x
k
} be
a sequence in A that converges to x
0
∈ S. Since all x
k
belong to A, they also belong to F, and because
F is closed, Theorem 3.2.3 tells us that x
0
∈ F. Hence, x
0
∈ S ∩ F = A.
Proof of ⇐: Suppose that (b) is satisﬁed. We want to show that A = S ∩F for some closed set F in ޒ
n
.
In fact, we shall show that A = S ∩
¨
A, where
¨
A is the closure of A in ޒ
n
. It is clear that A ⊆ S ∩
¨
A, so
what we need to show is that S ∩
¨
A ⊆ A. Let x
0
∈ S ∩
¨
A. Since x
0
∈
¨
A there exists a sequence {x
k
} of
points in A converging to x
0
. Since x
0
∈ S, we have x
0
∈ A by assumption. It follows that S ∩
¨
A ⊆ A.
Now that we have proved the equivalence of (a) and (b) above, let us use it to prove part (b) of Theorem
13.3.5: Let S ⊆ ޒ
n
and let f be a function from S to ޒ
m
. Then
f : S ⊆ ޒ
n
→ޒ
m
is continuous ⇐⇒ f
−1
(F) is relatively closed in S for each closed set F in ޒ
m
Proof of ⇒: Suppose that f is continuous and let F be a closed set in ޒ
m
. We want to prove that f
−1
(F)
is relatively closed in S. By the equivalence (a) ⇐⇒ (b) above it sufﬁces to show that, if a point x
0
in
S is the limit of a sequence {x
k
}
k
in f
−1
(F) then x
0
∈ f
−1
(F). If we have such a point x
0
and such a
sequence {x
k
}, then all f(x
k
) ∈ F, and because f is continuous, (x
0
) = lim
k
(x
k
). Thus, f(x
0
) the limit
of a sequence in F, and since F is closed, f(x
0
) ∈ F. Therefore x
0
∈ f
−1
(F).
Proof of ⇐: Suppose that f
−1
(F) is relatively closed in S for every closed F in ޒ
m
, and let x
0
be a point
in S. We want to show that f is continuous at x
0
. In other words, we want to show that f(x
k
) → f(x
0
)
for every sequence {x
k
} in S that converges to x
0
.
Suppose (∗) x
k
→ x
0
but x
k
,→ f(x
0
). Then there is an ε > 0 such that f(x
k
) − f(x
0
) ≥ ε for
inﬁnitely many k. Let F = {y ∈ ޒ
m
: y − f(x
0
) ≥ ε} be the complement in ޒ
m
of the open εball
around f(x
0
), and let {x
/
j
} = {x
k
j
} be the subsequence of {x
k
} where k
1
< k
2
· · · run through all those k
for which f(x
k
) −f(x
0
) ≥ ε. The set F is closed in ޒ
m
, and {x
/
j
} is a sequence in f
−1
(F), the inverse
image of F, with lim
j→∞
x
/
j
= x
0
∈ S. By assumption, f
−1
(F) is relatively closed in S, and by the
equivalence (a) ⇐⇒ (b) above we must have x
0
∈ f
−1
(F). But then f(x
0
) ∈ F, and by the deﬁnition
of F we must have f(x
0
) −f(x
0
) ≥ ε > 0, which is absurd. This shows that the assumption (∗) must
be false, and so f is indeed continuous.
13.3.8 Assume ﬁrst that f is continuous at x
0
. We shall prove that the defender can always win in this case.
Let the challenger choose ε > 0. Since f is continuous at x
0
the defender is able to choose a δ > 0 such
that f(x) − x
0
) < ε whenever x − x
0
 < δ, and then the challenger will be unable to ﬁnd an x that
lets him win. Thus the defender wins.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
104 13 TOPOL OGY AND SEPARATI ON
Now assume that f is discontinuous at x
0
. Then there will exist at least one ε > 0 that cannot be
matched by any δ. So let the challenger choose such an ε. Then no matter what δ > 0 the defender
chooses, the challenger will be able to ﬁnd an x with x −x
0
 < δ and f(x) −f(x
0
) ≥ ε. Thus, in this
case, the challenger wins.
13.4
13.4.3 For a ﬁxed x, the maximum of f (x, y) with respect to y for y in [−3, 3] must be attained at a point
where f
/
2
(x, y) = 0 or y = ±3. Since f
/
2
(x, y) = −12xy
3
−12(x−1)y
2
÷12y = −12xy(y−1/x)(y÷1),
f (x, y) is strictly increasing with respect to y when y ∈ (−∞, −1], strictly decreasing in [−1, 0], strictly
increasing again in [0, 1/x], and strictly decreasing in [1/x, ∞). Hence the only possible maximum
points are y = −1 and y = 1/x if x > 1/3 and y = −1 and y = 3 if 0 < x ≤ 1/3. Simple
calculations give f (x, −1) = x ÷ 2, f (x, 1/x) = (2x ÷ 1)/x
3
, and f (x, 3) = 162 − 351x. It follows
that f (x, 1/x) − f (x, −1) = (x − 1)(x ÷ 1)
3
/x
3
, so for x > 1 the maximum occurs for y = −1, if
x = 1 the maximum occurs for y = ±1, and if 1/3 < x ≤ 1 the maximum occurs for y = 1/x. Finally,
if 0 < x ≤ 1/3, the maximum occurs for y = 3. The value function V is given by
V(x) =
⎧
⎨
⎩
162 −351x if 0 < x ≤ 1/3
(2x ÷1)/x
3
if 1/3 < x ≤ 1
x ÷2 if x > 1
It is clear that V is continuous, because the onesided limits of V(x) at x = 1/3 are equal, and so are
the onesided limits at x = 1. Figures M13.4.3(a)–(c) shows the graph of the function y .→f (x, y) for
three different values of x, Fig. M13.4.3(d) shows the set M(x) of maximizers as a function of x. Except
at x = 1, the graph is the graph of a continuous function, as it should be because the maximizer is unique
for each x ,= 1.
z
5
10
15
y
−2 −1 1 2
z
5
10
15
y
−2 −1 1 2
z
5
10
15
y
−2 −1 1 2
y
−1
1
2
3
x
1 2
(a) z = f (1/2, y) (b) z = f (1, y) (c) z = f (2, y) (d) x .→M(x)
M13.4.3
13.5
13.5.7 Let S be a compact set in ޒ
n
. Carathéodory’s theorem (Theorem 13.5.1) tells us that every point in
co(S) can be written as a convex combination of at most n ÷1 points in S. We claim that every point in
co(S) can be written as a linear combination of exactly n÷1 points in S. Indeed, if x = λ
1
x
1
÷· · · λ
m
x
m
with m < n ÷1, we just add n ÷1 −m terms that are all equal to 0x
1
.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
14 CORRESPONDENCES AND F I XED POI NTS 105
Now let T =
n
S
n÷1
, where
n
= {(λ
1
, . . . , λ
n÷1
) ∈ ޒ
n÷1
: λ
i
≥ 0 for i = 1, . . . , n ÷1: λ
1
÷
· · · ÷λ
n÷1
= 1} and S
n÷1
= S · · · S is the Cartesian product of n ÷1 copies of S. Deﬁne a function
f : T →ޒ
n
by f (λ
1
, . . . , λ
n÷1
, x
1
, . . . , x
n÷1
) = λ
1
x
1
÷· · · ÷λ
n÷1
x
n÷1
. Then f maps every point in
T to a point in co(S). On the other hand, the argument above shows that every point in co(S) belongs to
the image f (T ) of T under f , so in fact co(S) = f (T ). The set
n
is obviously a compact (i.e. closed
and bounded) subset of ޒ
n÷1
, so if S is also compact, then so is T . (It follows from Problem 13.2.6 that,
if A and B are compact subsets of ޒ
m
and ޒ
p
, respectively, then A B is a compact subset of ޒ
m÷p
.
This immediately extends to the Cartesian product of a ﬁnite number of sets.) Since f is continuous,
Theorem 13.3.3 shows that co(S) = f (T ) is compact.
What we have shown here is that, if S is closed and bounded, then so is co(S). If S is closed but
unbounded, then co(S) need not even be closed. Consider, for example, the closed set S = {(x, y) ∈ ޒ
2
:
x > 0, y > 0, xy ≥ 1} ∪ {(0, 0)}. The convex hull of S is co(S) = {(x, y) : x > 0, y > 0} ∪ {(0, 0)},
i.e. the open ﬁrst quadrant together with the origin. This set is not closed. (Draw a picture!)
13.6
13.6.3 If x is not an interior point of the convex set S (⊆ ޒ
n
), then by Theorem 13.6.2 there exists a nonzero
vector a in ޒ
n
such that a · z ≤ a · x for every z in S. Then S ⊆ H
−
= {z : a · z ≤ a · x}. Since the half
space H
−
is closed, we also have
¨
S ⊆ H
−
. Every open ball around x contains points that do not belong
to H
−
, for instance points of the form x ÷t a for small positive numbers t . Hence, no open ball around x
is contained in
¨
S, and so S is not an interior point of
¨
S.
14 Correspondences and Fixed Points
Proof of Theorem 14.1.5(c): We will use the result in (14.1.8) to prove that H = G F is lower
hemicontinuous. Let z
0
∈ H(x
0
) and let U be a neighbourhood of z
0
. Then z
0
∈ G(y
0
) for some y
0
in
F(x
0
). Since G is l.h.c. at y
0
, there exists a neighbourhood V of y
0
such that U ∩ G(y) ,= ∅ for all y in
V. Moreover, since y
0
∈ F(x
0
), there is a neighbourhood N of x
0
such that V ∩F(x) ,= ∅ for all x ∈ N.
Let x ∈ N and y ∈ V ∩ F(x). Then U ∩ G(y) ,= ∅, and if z ∈ U ∩ G(y), then z ∈ G(y), y ∈ F(x), so
z ∈ H(x).
14.1
14.1.4 We shall use the characterization of lower hemicontinuity in (14.1.8) on page 506 to show that if F
and G are l.h.c. at x
0
, then so is H. Let U be an open neighbourhood of a point (y
0
, z
0
) in H(x
0
). Then
there are open neighbourhoods U
1
and U
2
of y
0
and z
0
in ޒ
l
and ޒ
m
, respectively, such that U
1
U
2
⊆ U.
Since F is l.h.c., there are neighbourhoods N
1
and N
2
of x
0
such that F(x) ∩U
1
,= ∅ for all x in N
1
∩X
and G(x) ∩ U
2
,= ∅ for all x in N
2
∩ X. Let N = N
1
∩ N
2
. Then H(x) ∩ U ⊇ H(x) ∩ (U
1
U
2
) =
(F(x) ∩ U
1
) (G(x) ∩ U
2
) ,= ∅ for all x in N ∩ X. It follows that H is l.h.c.
For the result about upper hemicontinuity we need to assume that F(x
0
) and G(x
0
) are compact.
Then H(x
0
) = F(x
0
) G(x
0
) is also compact. Suppose that F and G are u.h.c. at x
0
. Since F(x
0
)
is compact, it is closed, and Note 14.1.1 then says that F has the closed graph property at x
0
. Because
F(x
0
) is bounded there exists a bounded and open set U in ޒ
l
that contains F(x
0
), and since F is upper
hemicontinuous at x
0
there is a neighbourhood N
F
of x
0
such that F(x) ⊆ U for all x in N
F
. Similarly, G
has the closed graph property at x
0
and there exists a bounded and open set V in ޒ
m
and a neighbourhood
N
G
of x
0
such that G(x) ⊆ V for all x in N
G
.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
106 14 CORRESPONDENCES AND F I XED POI NTS
Since F and G have the closed graph property at x
0
, so has H. This is an easy consequence of
the deﬁnition (14.1.4). We are now all set to use Theorem 14.1.2 to prove that H is u.h.c at x
0
. Let
W = U V ⊆ ޒ
l÷m
and let N = N
F
∩ N
G
. Then W is bounded in ޒ
l÷m
, N is a neighbourhood of x
0
,
and H(x) = F(x) G(x) ⊆ W for all x in N, so H is locally bounded near x
0
. The desired conclusion
follows.
14.1.6 We shall use the characterization of lower hemicontinuity that is given in (14.1.8) to prove that G
is l.h.c. Let y
0
∈ G(x
0
). Then y
0
is a convex combination y
0
=
¸
k
i=1
λ
i
y
0
i
of points y
0
i
in F(x
0
). If
U is a neighbourhood of y
0
, then U contains an open ball B = B(y
0
: ε) for some ε > 0. For each
i = 1, . . . , k, there is a neighbourhood N
i
of x such that F(x) ∩ B(y
0
i
: ε) ,= ∅ for all x in N
i
∩ X. Let
N = N
1
∩· · · ∩N
k
be the intersection of these neighbourhoods. Then for any x in N and every i there is
at least one point y
i
in F(x) ∩ B(y
0
i
: ε). If we let y =
¸
i
λ
i
y
i
, then y ∈ co(F(x)) = G(x). Moreover,
d(y, y
0
) = y −y
0
 = 
¸
i
(λ
i
y
i
−λ
i
y
0
i
) ≤
¸
i
λ
i
y
i
−y
0
i
 <
¸
i
λ
i
ε = ε, so y ∈ B(y
0
: ε) ⊆ U. It
follows that G(x) ∩ U ,= ∅. Hence G is l.h.c.
14.1.10 Every constant correspondence is both l.h.c. and u.h.c. But even if a(x) and b(x) are constant
functions, complications may arise if one or both of the endpoints of the interval (a((x), b(x)) sometimes,
but not always, belong to F(x). Consider for example the correspondences C
1
and C
2
given by
C
1
(x) =
¸
[0, 1] if x ≤ 0,
[0, 1) if x > 0,
C
2
(x) =
¸
[0, 1] if x < 0,
[0, 1) if x ≥ 0.
C
1
is u.h.c. everywhere, while C
2
is not u.h.c. at x = 0. (But both of them are l.h.c. everywhere.)
Every correspondence F that satisﬁes the conditions in the problem is l.h.c., provided only that its
effective domain is open—i.e., if F(x
0
) ,= ∅, then F(x) ,= ∅ for all x sufﬁciently close to x
0
.
With nonconstant a(x) and b(x) many complications arise in connection with upper hemicontinuity.
A detailed study would take us too far aﬁeld, so let us concentrate on the four possibilities
F(x) = [a(x), b(x)], G(x) = [a(x), b(x)), H(x) = (a(x), b(x)], K(x) = (a(x), b(x))
(for all x). F has a closed graph and is locally bounded, so Theorem 14.1.2 implies that F is u.h.c.
The other three are usually not u.h.c., except in the constant case. For example, the correspondence
F(x) = (−1−x
2
, 1÷x
2
) is not u.h.c. at x = 0, since the open set U = (−1, 1) contains F(x) for x = 0
but not for any other value of x, no matter how close to 0 it may be. In fact, F is not u.h.c. at any other
point either.
14.1.11 The set F(x) is compact for each x in X. By Problem 13.5.7, G(x) = co(F(x)) is also compact for
each x. If G(x
0
) is contained in an open set U, there exists an α > 0 such that G(x
0
) ⊆ U
/
⊆ U, where
U
/
is the open “αneighbourhood” of G(x
0
) deﬁned as
U
/
= B(G(x
0
): α) =
¸
y ∈ ޒ
m
: there is a y
/
in G(x
0
) with y −y
/
 < α
¸
This follows from the technical result below, with S = U = ޒ
m
` U and K = F(x
0
). Since F is u.h.c.
and F(x
0
) ⊆ G(x
0
) ⊆ U
/
, there exists a δ > 0 such that F(x) ⊆ U
/
for every x in N = B(x
0
: δ). It is
not hard to see that U
/
is convex, and therefore G(x) = co(F(x)) is also contained in U
/
(and so in U)
for all x in N.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
A SETS, COMPL ETENESS, AND CONVERGENCE 107
A little technical result:
If S and K are disjoint closed subsets of ޒ
m
and K is compact, then there exists a positive number α
such that d(x, y) ≥ α for all x in S and all y in K. In other words, the distance between a point in S and
a point in K is never less than α.
Proof: Let y be a ﬁxed point in ޒ
n
and let x be a point in S. By Problem 13.3.5, h(x) = d(x, y)
attains a minimum at some point x
0
in S. Let us call h(x
0
) the distance between the point y and the set
S, and denote it by d(y, S). If y
/
is another point in ޒ
n
, then
d(y
/
, S) ≤ d(y
/
, x
0
) ≤ d(y
/
, y) ÷d(y, x
0
) = d(y
/
, y) ÷d(y, S)
so d(y
/
, S) − d(y, S) ≤ d(y
/
, y). By symmetry, d(y, S) − d(y
/
, S) ≤ d(y, y
/
) = d(y
/
, y). Hence,
[d(y
/
, S) − d(y, S)[ ≤ d(y
/
, y). It follows that g(y) = d(y, S) is a continuous function of y. (In the
deﬁnition of continuity in (13.3.1), every ε > 0 can be matched by δ = ε.) Since K is compact, g attains
a minimum value α over K. Then α = g(y
∗
) = d(y
∗
, x
∗
) for a point y
∗
in K and a point x
∗
in S, so
α > 0.
14.2
14.2.3 By Example 14.1.6, the budget correspondence B(p, m) is lower hemicontinuous and has the
closed graph property at any point (p
0
, m
0
) where m
0
> 0. It is also locally bounded near (p
0
, m
0
),
so by Theorem 14.1.2, B is also upper hemicontinuous at (p
0
, m
0
). What if m
0
= 0? In that case
B(p
0
, m
0
) = B(p
0
, 0) consists of a single point, namely the origin 0 in ޒ
n
. If U is an open set in ޒ
n
that contains 0, then it will obviously contain B(p, m) for any (p, m) close to (p
0
, 0), and it follows that
B is upper hemicontinuous at (p
0
, 0). Lower hemicontinuity at that point follows easily from the result
in (14.1.8).
Thus, B(p, m) is continuous at every point (p, m) in X = ޒ
n
÷÷
ޒ
÷
.
The maximum theorem (Theorem 14.2.1) then implies that the demand correspondence ξ(p, m) is
upper hemicontinuous and the indirect utility function V(p, m) is continuous. (Note that x . F(x) in
Theorem 14.2.1 corresponds to (p, m) . B(p, m) in this problem, while y and f (x, y) in the theorem
correspond to x and U(x) in the problem.) The demand correspondence will not always be lower hemi
continuous.
Suppose that U is quasiconcave. Then, if x
1
and x
2
are distinct points in ξ(p, m) (i.e. maximum
points for U over B(p, m)), Theorem2.5.1 implies that λx
1
÷(1−λ)x
2
also belongs to ξ(p, m) for every
λ in [0, 1]. In other words, ξ(p, m) is a convex set.
A Sets, Completeness, and Convergence
A.1
A.1.3 The answer in the book has been pared down to the bone, with a function deﬁned on a set with only
two elements in it. For an example with a little more ﬂesh on it, consider the following:
Let f (x) = x
2
for all x in ޒ, and deﬁne two intervals S
1
= [−4, 2] and S
2
= [−1, 3]. The
intersection of S
1
and S
2
is S
1
∩S
2
= [−1, 2]. The images of S
1
and S
2
under f are f (S
1
) = [0, 16] and
f (S
2
) = [0, 9], and the image of S
1
∩ S
2
is f (S
1
∩ S
2
) = [0, 4]. Here, then, is a case where f (S
1
∩ S
2
)
is a proper subset of f (S
1
) ∩ f (S
2
).
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
108 A SETS, COMPL ETENESS, AND CONVERGENCE
A.1.4 A relation is a linear ordering if and only if it is (i) reﬂexive, (ii) transitive, (iii) antisymmetric, and
(iv) complete. For each of these four properties it is easy to see that if a relation R has that property, then
so has R
−1
. Let us just see how to handle (ii):
We are assuming that R is transitive relation in a set S. Let x, y, and z be elements in S such that
xR
−1
y and yR
−1
z. We want to show that xR
−1
z. We have yRx and zRy, and since R is transitive,
zRx. Therefore xR
−1
z, as promised.
A.2
A.2.2 It is shown in the answer in the book that s < r, and since s is positive and s
2
> 2, we also have s > 2.
Thus the rational number s is closer to
√
2 than r is. But how did anyone come up with the expression
s = (2 ÷ r
2
)/2r in the ﬁrst place? The answer is Newton’s method, which is a famous procedure for
ﬁnding and improving approximate solutions to an equation f (x) = 0, where f is a differentiable
function.
Given one approximate solution, x
0
, we let x
1
= x
0
−f (x
0
)/f
/
(x
0
), and then we construct x
2
from
x
1
in the same fashion, and so on. If the initial approximation x
0
is good enough, the sequence x
1
, x
2
, . . .
will usually converge to a root of f (x) = 0. (A brief discussion of this method can be found in EMEA,
for instance.) Now apply this procedure to the equation f (x) = x
2
− 2 = 0. If x
0
= r is a positive
number, then x
1
= r −f (r)/f
/
(r) = r −(r
2
−2)/2r = (r
2
÷2)/2r is precisely the number s.
A.3
A.3.4 For the solution of this problem we need the fact that, if two convergent sequences have inﬁnitely
many terms in common, they must have the same limit. (This is an easy consequence of TheoremA.3.3.)
Consider the particular subsequences {w
k
} = {x
2k−1
} and {z
k
} = {x
2k
} of {x
k
}. They converge to 0
and 2, respectively. Together, these two subsequences contain all the terms of the original sequence, so
any subsequence of {x
k
} must have an inﬁnite number of terms in common with at least one of {w
k
} and
{z
k
}. Hence, any convergent subsequence of {x
k
} must converge to either 0 or 2.
The six subsequences {y
6k
}, {y
6k÷1
}, {y
6k÷2
}, {y
6k÷3
}, {y
6k÷4
}, and {y
6k÷5
} converge to sin 0 = 0,
sin(π/3) =
1
2
√
3, sin(2π/3) =
1
2
√
3, sin(3π/3) = 0, sin(4π/3) = −
1
2
√
3, and sin(5π/3) = −
1
2
√
3,
respectively, and they contain all but the ﬁrst ﬁve elements of {y
k
}. In the same way as for {x
k
}, it follows
that any convergent subsequence of {y
k
} must converge to 0,
1
2
√
3, or −
1
2
√
3.
Figures MA.3.4(a) and MA3.4(b) illustrate the behaviour of {x
k
} and {y
k
}.
x
k
10 20
2
1
−1
y
k
10 20
1
2
√
3
−
1
2
√
3
MA.3.4(a) MA.3.4(b)
A.3.5 (b) (i): For each natural number n, let M
n
= sup{x
k
: k ≥ n} and N
n
= sup{y
k
: k ≥ n}. Then for all
k ≥ n, we have x
k
≤ M
n
andy
k
≤ N
n
, andsox
k
÷y
k
≤ M
n
÷N
n
. Thus sup{x
k
÷y
k
: k ≥ n} ≤ M
n
÷N
n
≤
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
B TRI GONOMETRI C F UNCTI ONS 109
lim
n→∞
(M
n
÷N
n
) = lim
n→∞
M
n
÷lim
n→∞
N
n
, or lim
k→∞
(x
k
÷y
k
) ≤ lim
k→∞
x
k
÷lim
k→∞
y
k
.
The proof of (ii) is similar and is left to the reader.
A.3.6 Let n and m be natural numbers with n > m, and let p = n −m. Then
[x
n
−x
m
[ = [x
m÷p
−x
m
[
= [(x
m÷p
−x
m÷p−1
) ÷(x
m÷p−1
−x
m÷p−2
) ÷· · · ÷(x
m÷2
−x
m÷1
) ÷(x
m÷1
−x
m
)[
≤ [x
m÷p
−x
m÷p−1
[ ÷[x
m÷p−1
−x
m÷p−2
[ ÷· · · ÷[x
m÷2
−x
m÷1
[ ÷[x
m÷1
−x
m
[
<
1
2
m÷p−1
÷
1
2
m÷p−2
÷· · · ÷
1
2
m÷1
÷
1
2
m
=
1
2
m
1 ÷
1
2
÷· · · ÷
1
2
p−1
=
1
2
m
1 −(
1
2
)
p
1 −
1
2
<
1
2
m
1
1 −
1
2
=
1
2
m−1
which obviously becomes arbitrarily small for m sufﬁciently large.
B Trigonometric Functions
B.1
B.1.4 cos(y −π/2) = sin y follows directly from Problem 3. Then, from the hints in the question, as well
as (B.1.8) and the result of Problem 3 again, it follows that
sin(x ÷y) = cos(x ÷y −π/2) = cos x cos(y −π/2) −sin x sin(y −π/2)
= cos x sin y ÷sin x cos y = sin x cos y ÷cos x sin y
Substituting −y for y then yields sin(x −y) = sin x cos(−y) ÷cos x sin(−y) = sin x cos y −cos x sin y.
B.1.6 It is clear from the deﬁnitions of sin x and cos x that sin(x ÷π) = −sin x and cos(x ÷π) = −cos x
for all x. This also follows from Problem 4 and formula (B.1.8) together with the fact that sin π = 0 and
cos π = −1. The formula for sin(x −y) in Problem 4 also shows that sin(π −x) = sin x for all x. These
results come in handy in this problem.
(a) sin(π −π/6) = sin(π/6) = 1/2.
(b) (cos(π ÷π/6) = −cos(π/6) = −
1
2
√
3.
(c) By Problem 2(a), sin(−3π/4) = −sin(3π/4) = −
1
2
√
2.
(d) cos(5π/4) = cos(π/4 ÷π) = −cos(π4) = −
1
2
√
2.
(e) By formula (6), tan(7π/6) = tan(π/6) =
1
3
√
3.
(f) sin(π/12) = sin(π/3 −π/4) = sin(π/3) cos(π/4) −cos(π/3) sin(π/4) =
1
4
(
√
6 −
√
2).
B.1.7 (a) The formula for sin(x ÷y) in Problem 4 gives
√
2 sin(x ÷π/4) −cos x =
√
2(sin x cos(π/4) ÷
cos x sin(π/4)) −cos x =
√
2(sin x ·
1
2
√
2 ÷cos x ·
1
2
√
2) −cos x = sin x
(b) Since sin(π −x) = sin x and cos(2π −x) = cos(−x) = cos x, we have
sin[π −(α ÷β)]
cos[2π −(α ÷β)]
=
sin(α ÷β)
cos(α ÷β)
= tan(α ÷β)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
110 B TRI GONOMETRI C F UNCTI ONS
(c) Formula (B.1.8) and the formulas in Problems 3 and 4 give
sin(a ÷x) −sin(a −x)
cos(a ÷x) −cos(a −x)
=
sin a cos x ÷cos a sin x −sin a cos x ÷cos a sin x
cos a cos x −sin a sin x −cos a cos x −sin a sin x
=
2 cos a sin x
−2 sin a sin x
= −
cos a
sin a
= −cot a
B.1.8 With x =
1
2
(A ÷ B) and y =
1
2
(A − B) we get sin A − sin B = sin(x ÷ y) − sin(x − y) =
sin x cos y ÷cos x sin y −sin x cos y ÷cos x sin y = 2 cos x sin y = 2 cos
A ÷B
2
sin
A −B
2
.
B.1.12 (a) This is a sine curve y = Asin(ax) with A = 2 and a(8π) = 2π, i.e. a = 1/4.
(b) y = 2 ÷cos x. (A cosine curve with amplitude 1 and period 2π shifted 2 units upwards.)
(c) y = 2e
−x/π
cos x. (An exponentially damped cosine curve with amplitude 2e
−x/π
.)
B.1.13 Since the lengths of the line segments AC and BD are equal, we have (cos x −cos y)
2
÷(sin x ÷
sin y)
2
= (cos(x ÷y) −1)
2
÷sin
2
(x ÷y). The lefthand side is
LHS = cos
2
x −2 cos x cos y ÷cos
2
y ÷sin
2
x ÷2 sin x sin y ÷sin
2
y
= 2 −2 cos x cos y ÷2 sin x sin y
and the righthand side is
RHS = cos
2
(x ÷y) −2 cos(x ÷y) ÷1 ÷sin
2
(x ÷y) = 2 −2 cos(x ÷y)
where we have repeatedly used that sin
2
u ÷ cos
2
u = 1. The equation LHS = RHS implies that
cos(x ÷y) = cos x cos y −sin x sin y.
B.2
B.2.1 (b) (x cos x)
/
= x
/
cos x ÷x(cos x)
/
= cos x −x sin x.
(c) Let u = x
2
. Then
d
dx
(tan(x
2
)) =
d
du
(tan u)
du
dx
=
1
cos
2
u
2x =
2x
cos
2
(x
2
)
.
(d) (e
2x
cos x)
/
= (e
2x
)
/
cos x ÷e
2x
(cos x)
/
= 2e
2x
cos x −e
2x
sin x = e
2x
(2 cos x −sin x).
B.2.4 (c) All you need is the chain rule and a steady hand.
B.2.5 (c) lim
t →0
1 −cos t
t
2
=
“0
0
”
= lim
t →0
sin t
2t
=
1
2
lim
t →0
sin t
t
=
1
2
by (B.2.10) (or just use l’Hôpital’s rule
again).
B.2.6 The derivative of f
/
(x) is f
/
(x) = 3(sin x −x −1)
2
(cos x −1). It is easy to see that sin x < x ÷1 for
all x > 0, because sin x ≤ 1 < x ÷1. Moreover, cos x < 1 for all x in the open interval J = (0, 3π/2).
It follows that f
/
(x) < 0 for all x in J. Thus, f is strictly decreasing in the closed interval I = [0, 3π/2],
and attains its maximum value −1 at x = 0 and its minimum value −(2 ÷ 3π/2)
3
≈ −302.43 at
x = 3π/2.
B.2.8 You can read all these values off from Table B.2.1.
B.2.9 (a) Let u = 2x. The chain rule yields
d
dx
(arcsin 2x) =
d
du
(arcsin u)
du
dx
=
2
√
1 −u
2
=
2
√
1 −4x
2
.
(b) (d/dx)(arctan v) with v = 1 ÷x
2
.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
B TRI GONOMETRI C F UNCTI ONS 111
(c) Let w =
√
x. Then
d
dx
(arccos
√
x ) =
d
dw
(arccos w)
dw
dx
= −
1
√
1 −w
2
1
2
√
x
= −
1
2
√
1 −x
√
x
.
B.2.10 (c) Integration by parts yields
I =
sin
2
x dx =
sin x(−cos x)
/
dx = sin x(−cos x) −
(sin x)
/
(−cos x) dx
= −sin x cos x ÷
cos
2
x dx = −sin x cos x ÷
(1 −sin
2
x) dx
Hence, I = −sin x cos x ÷ x − I ÷ C, and we get I =
1
2
(x − sin x cos x) ÷ C
1
, where C
1
=
1
2
C.
Note: When integrating trigonometric functions it is very easy to get wrong signs here and there, and
integration by parts also often leads to such mistakes, so it is a good rule to check the results by ﬁnding
the derivatives when that is possible.
(d) Integration by parts here too.
B.2.11 (a) Let u = cos x. Then du = −sin x dx and
tan x dx = −
du
u
= −ln [u[ ÷ C =
−ln [ cos x[ ÷C.
(b) With v = sin x, we get dv = cos x dx and
cos xe
sin x
dx =
e
v
dv = e
v
÷C = e
sin x
÷C.
(c) As in part (a), let u = cos x. Then
cos
5
x sin x dx =
−u
5
dx = −
1
6
u
6
÷C = −
1
6
cos
6
x ÷C.
B.3
B.3.3 To simplify a complex fraction (a ÷bi)/(c ÷di), where a, b, c, d are real, it is usually a good idea to
multiply both the numerator and the denominator by c − di, the conjugate of the original denominator.
This has the effect of making the denominator real (and positive) because (c−di)(c÷di) = c
2
−(di)
2
=
c
2
−d
2
i
2
= c
2
÷d
2
. Thus,
(a)
(3 ÷2i)(1 ÷i)
(1 −i)(1 ÷i)
=
3 ÷5i ÷2i
2
1 −i
2
=
1 ÷5i
2
=
1
2
÷
5
2
i, (b)
(4 −3i)(−i)
i(−i)
=
−3 −4i
1
= −3−4i.
(c) Simplify the numerator and denominator before making the denominator real:
(3 −2i)(2 −i)
(−1 −i)(3 ÷2i)
=
6 −7i ÷2i
2
−3 −5i −2i
2
=
4 −7i
−1 −5i
=
(4 −7i)(−1 ÷5i)
(−1)
2
−(5i)
2
=
−4 ÷27i −35i
2
1 −25i
2
=
31
26
÷
27
26
i.
(d)
1 −i
1 ÷i
=
(1 −i)
2
1 −i
2
=
1 −2i ÷i
2
2
=
−2i
2
= −i, so
1 −i
1 ÷i
3
= (−i)
3
= −i
3
= −i
2
i = i.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
112 B TRI GONOMETRI C F UNCTI ONS
Corrections to FMEA 2nd edn (2008), ﬁrst printing
(Spelling and grammatical mistakes are not included.)
Page 26, line 12 from the bottom. . . . (See Example B.3.2.)
Page 59, Theorem 2.3.5: In parts (b) and (c) the conclusion should be that U(x) = F(f (x)) is convex!
Page 117, Theorem 3.3.1(b): Add the assumption that S is convex.
Page 124, Problem 3.3.10: Replace the inequality in (∗) with equality.
Page 132, Theorem 3.5.1, line 1–2: Assume that f and g
1
, . . . , g
m
are deﬁned in a set S and that x
∗
is an
interior point of S.
Page 164, Problem 10, line 5: The function f must also be continuous.
Page 192, Problem 6: The differential equation has no solution deﬁned over the entire real line. We must be
satisﬁed with a function x that satisﬁes the conditions in the problem in an open interval around 0.
Page 199, Problem 6, line 2: . . . all t > 0.
Page 225, Problem 6, line 3: . . . , provided ˙ x ,= 0, we have
Page 352, Theorem 9.11.2, line 2: . . . that
∞
t
0
[f (t, x(t ), u(t ))[e
−t
dt < ∞for all . . .
Page 357, Problem 9.12.3(c), line 2: . . . with K(t ) > 0 for
Page 382: The entries in the ﬁrst column of the table should be “t ∈ [0, t
∗
]” and “t ∈ (t
∗
, T ]”.
Page 444, Problem 12.4.2(b): . . . in Theorem 12.4.1 are . . .
Page 454, formula (10): E
¸
F
/
2
(t, x
t
, x
t ÷1
(x
t
, V
t
), V
t
) [ v
t −1
¸
÷F
/
3
(t −1, x
t −1
, x
t
, v
t −1
) = 0
Page 455, line 5: . . . . Next, for t = T ,
Page 455, line 8:
2
3
v
T −1
should be
2
3
V
T −1
.
Page 457, Problem 12.6.6: The summation goes from t = 0, and the scrap value is X
1/2
T
.
Page 508: See SM (not Problem 11) for the proof of Theorem 14.1.5(c).
Page 566, Problem 2.6.1(c): f (x, y) ≈ x ÷2y −
1
2
x
2
−2xy −2y
2
Page 566, Problem 2.6.4: z ≈ 1 −x ÷y ÷
3
2
x
2
−2xy ÷
1
2
y
2
Page 566, Problem 2.7.3, line 2: . . . , w
x
= 1/2.
Page 570, Problem 3.7.3, line 2: f
∗
(r, s) =
1
4
r
2
. . . . to ﬁnd the squares of the largest
Page 570, Problem 3.8.3, line 2: Replace (c) by (b).
Page 570, Problem 3.8.5, line 1: Drop the labels (a) and (b).
Page 571: New Problem 3.11.1: See SM.
Page 574, Figure A5.2.2: The integral curve through (0, 2) is only the semicircle above the t axis.
Page 576, Problem 5.6.1: Delete “and x ≡ 0”. (We require x > 0.)
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
B TRI GONOMETRI C F UNCTI ONS 113
Page 577, Problem 5.7.3, line 2: . . . . See Fig. A5.7.3(b).
Page 582, Problem 6.9.6: Drop the references to (a) and (b).
Page 583, Problem 7.2.3: Remove (b) from line 2.
Page 586, Problem 9.4.2: . . . where A = 2e/(e
2
−1).
Page 588, Problem9.7.3: (a) . . . , x
∗
(t ) =
e
(α−2β)T ÷αt
e
(α−2β)T
−1
−
e
2(α−β)t
e
(α−2β)T
−1
=
e
(α−2β)T
−e
(α−2β)t
e
αt
e
(α−2β)T
−1
, . . .
Page 589, Problem 9.11.4: (x
∗
(t ), u
∗
(t )) =
¸
(e −e
−t
, 1) if t ∈ [−1, 0]
(e −1, 0) if t ∈ (0, ∞)
, p(t ) = e
−t
.
Page 589, Problem 9.12.3, line 1: . . . , and so
˙
C
∗
/C
∗
÷
˙
λ/λ = 0. . . .
Page 590, Problem 10.4.3, line 3: . . . . When u = 2, ˙ x = −3 and . . .
Page 591, Problem 10.6.3, line 2: Replace x
∗
by x
∗
and t
∗
by t
/
.
Page 593, Problem 11.7.2, line 1: (a) x
∗
≈ −2.94753.
Page 595, Problem 12.6.6: u
t
(x) = x
t
/4a
2
t
= x
t
/(1 ÷a
2
t ÷1
), a
T
= a/2, a
t
=
1
2
(1 ÷a
2
t ÷1
)
1/2
for t < T .
Page 595, Problem 12.6.8: x
1
=
1
2
−v
0
, x
2
= 1 −v
0
−v
1
, x
3
=
3
2
−v
0
−v
1
−v
2
Page 599, Problem 14.2.3: The demand correspondence ξ is upper hemicontinuous, but it need not be lower
hemicontinuous.
©Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
Preface
This student’s solutions manual accompanies Further Mathematics for Economic Analysis (2nd edition, FT SM ⊃ Prentice Hall, 2008). Its main purpose is to provide more detailed solutions to the problems marked with ⊂ in the text. The Manual should be used in conjunction with the answers in the book. In some few cases only a part of the problem is done in detail, because the rest follows the same pattern. At the end of this manual there is a list of misprints and other errors in the book, and even one or two in the errata list in the preliminary and incomplete version of this manual released in September this year. We would greatly appreciate suggestions for improvements from our readers as well as help in weeding out the inaccuracies and errors that probably remain. Oslo and Coventry, October 2008 Arne Strøm (arne.strom@econ.uio.no) Knut Sydsæter (knut.sydsater@econ.uio.no) Atle Seierstad (atle.seierstad@econ.uio.no) Peter Hammond (hammond@stanford.edu)
Contents
1 Topics in Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 2 3 5 6 7 8 9 10 11 12 14 A Multivariable Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 Static Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Differential Equations I: FirstOrder Equations in One Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 Differential Equations II: SecondOrder Equations and Systems in the Plane . . . . . . . . . . . . . . . . . . 49 Differential Equations III: HigherOrder Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 Calculus of Variations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 Control Theory: Basic Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 Control Theory with Many Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 Difference Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 Discrete Time Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 Correspondences and Fixed Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 Sets, Completeness, and Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 Corrections to the book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 Version 1.0 28102008 879
© Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
4 Topics in Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
13 Topology and Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
B Trigonometric Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
CHAPTER 1
TOPICS IN LINEAR ALGEBRA
1
Chapter 1 Topics in Linear Algebra
1.2
⎞ 1 2 0 1.2.3 Let A = ⎝ 0 1 1 ⎠ be a matrix with the three given vectors as columns. Cofactor expansion of 1 0 1 A along the ﬁrst row yields 1 0 1 2 1 0 0 1 1 =1 0 1 1 0 −2 1 1 1 + 0 = 1 − 2(−1) = 3 = 0 1 ⎛
By Theorem 1.2.1 this shows that the given vectors are linearly independent. 1.2.6 Part (a) is just a special case of part (b), so we will only prove (b). To show that a1 , a2 , . . . , an are linearly independent it sufﬁces to show that if c1 , c2 , . . . , cn are real numbers such that c1 a1 + c2 a2 + · · · + cn an = 0 then all the ci have to be zero. So suppose that we have such a set of real numbers. Then for each i = 1, 2, . . . , n, we have ai · (c1 a1 + c2 a2 + · · · + cn an ) = ai · 0 = 0 (1) Since ai ·aj = 0 when i = j , the lefthand side of (1) reduces to ai ·(ci ai ) = ci ai Because ai = 0 we have ai = 0, and it follows that ci = 0.
2.
Hence, ci ai
2
= 0.
1.3
1.3.1 (a) The rank is 1. See the answer in the book. (b) The minor formed from the ﬁrst two columns in the matrix is 1 3 = −6 = 0. Since this minor 2 0 is of order 2, the rank of the matrix must be at least 2, and since the matrix has only two rows, the rank cannot be greater than 2, so the rank equals 2. −1 3 = 5 = 0, so the rank (c) The ﬁrst two rows and last two columns of the matrix yield the minor −4 7 of the matrix is at least 2. On the other hand, all the four minors of order 3 are zero, so the rank is less than 3. Hence the rank is 2. (It can be shown that r2 = 3r1 + r3 , where r1 , r2 , and r3 are the rows of the matrix.) An alternative argument runs as follows: The rank of a matrix does not change if we add a multiple of one row to another row, so ⎛ 1 ⎝ 2 −1 2 4 −2 −1 −4 −1 ⎞ 3 −2 7⎠ ← −2 ← 1 1 ∼ ⎝0 0 ⎛ 2 0 0 −1 −2 −2 ⎞ 3 1⎠. 1
© Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
Knut Sydsæter. −1 x 2 if x = −1 or x = 2 Conclusion: r(A) = 3 otherwise (b) A little calculation shows that the determinant of the matrix is t 3 + 4t 2 − 4t − 16. and w. 1. we get the minor = 5t + 14. On the other hand. so r(A) can never be less than 2.2 CHAPTER 1 TOPICS IN LINEAR ALGEBRA Here ∼ means that the last matrix is obtained from the ﬁrst one by elementary row operations. ⎛ ⎞ x 0 x2 − 2 (a) By cofactor expansion along the ﬁrst row. it follows that the determinant is t 3 + 4t 2 − 4t − 16 = t 2 (t + 4) − 4(t + 4) = (t 2 − 4)(t + 4) = (t + 2)(t − 2)(t + 4) Thus. The rank is 3 if x = −1 and x = 2. © Arne Strøm.) (e) (f) The determinant of the whole matrix is zero. then A = 0 and r(A) ≤ 2. 1 −1 2 2 1 = 9 = 0. so the rank is at least 2. (The three rows. so the rank of A equals 3. y. All the four minors of order 3 are zero. and if we note that this expression has t + 4 as a factor. because r2 = −14r1 + 9r3 . the rank of the matrix is 3. 1 t +4 which is different from 0 for all the three special values of t that we found above. and therefore the original matrix also has rank 2. r1 . and thus the rank of the matrix is 2 if t = −4. 1 3 0 (d) The ﬁrst three columns of the matrix yield the minor 2 4 0 = −4 = 0. −2. so the rank must −1 4 be less than 3. the minor we get if we strike out the ﬁrst row and the third column in A is 0 1 = 1 = 0 for all x. or 2 3 otherwise (c) The ﬁrst and third rows are identical.2 (a) The determinant is (x + 1)(x − 2).3. if t does not equal any of the numbers −2. and −4. r2 . so the rank is 3. of the matrix are linearly dependent. It follows that the matrix has rank 3. the ﬁrst three rows and the ﬁrst three columns yield the minor 1 2 −1 −2 1 1 −1 1 = −7 = 0 −1 so the rank is at least 3. So the rank is 2 for all values of x. the determinant of the matrix A = ⎝ 0 1 1 ⎠ −1 x x − 1 is A = x · (−1) − 0 · 1 + (x 2 − 2) · 1 = x 2 − x − 2 = (x + 1)(x − 2) If x = −1 and x = 2. On the other hand. The rank is 2 if x = −1 or x = 2. and r3 . 2. Atle Seierstad. But the ﬁrst two rows are always linearly independent. 5 6 If we strike out the second row and the ﬁrst column of the matrix. The last matrix obviously has rank 2. so the rank must be less than 4. then A = 0. Hence the rank is 2. and Peter Hammond 2008 . z. as are the second and fourth. If x = −1 or x = 2.
4. the rank of At can never be less than 2. then the system has solutions if and only if b = 9/2.) If a = 0. Thus. x3 = −a. with a arbitrary. (The rank of the coefﬁcient matrix is 2 and there are 3 unknowns. Because the upper left 2 × 2 minor of At is −1 = 0. The determinant of the coefﬁcient matrix is 1 D= 1 2 −1 2 1 1 −1 = 9 = 0 3 Hence.e. the system has a unique solution for any value of b. 1.CHAPTER 1 TOPICS IN LINEAR ALGEBRA 3 1. so the rank of the coefﬁcient matrix is 3. too. and the system either has solutions with 1 degree of freedom or has no solutions at all.4.) 1. so r(A2 ) = 2.6 (a) At  = (t − 2)(t + 3). (We could get the number of degrees of freedom from Theorem 1.) (b) By Gaussian elimination (or other means) we ﬁnd that the solution is x1 = a. the system has a unique solution. then the system has solutions if and only if −9b + 30 = 0. The vector equation A−3 x = ⎝ 3 ⎠ is equivalent to the equation system x3 6 x1 + 3x2 + 2x3 = 11 2x1 + 5x2 − 3x3 = 3 4x1 + 10x2 − 6x3 = 6 © Arne Strøm. ⎛ ⎞ ⎛ ⎞ 11 x1 (b) Let x = ⎝ x2 ⎠. the system has 3 − 3 = 0 degrees of freedom. One way of ﬁnding out is by Gaussian elimination. r(A−3 ) = 2.2 in this case.2: since the rank of the coefﬁcient matrix is 3 and there are 3 unknowns. Thus. The minor formed from the ﬁrst three columns of the coefﬁcient matrix has determinant −3 = 0. and x4 = a. (This agrees with Theorem 1. and Peter Hammond 2008 (1) (2) (3) . Atle Seierstad. so r(At ) = 3 if t = 2 and t = −3. If a = 0 or a = 7. by using elementary row operations on the augmented coefﬁcient matrix: ⎛ ⎞ ⎛ ⎞ 1 2 3 1 1 −3 1 2 3 1 ⎝ −1 a −21 2 ⎠ ← ∼ ⎝ 0 a + 2 −18 3 ⎠← 3 7 a b ← −(a + 2) 0 1 a−9 b−3 ⎞ ⎛ 1 2 3 1 ∼ ⎝ 0 0 −a 2 + 7a −ab − 2b + 3a + 9 ⎠ ← 0 1 a−9 b−3 ← ⎛ ⎞ 1 2 3 1 ⎝0 1 ⎠ ∼ a−9 b−3 0 0 a(7 − a) −ab − 2b + 3a + 9 This conﬁrms that as long as a = 0 and a = 7. if a = 0 and a = 7. there is one degree of freedom. But if a = 0 or a = 7. by Cramer’s rule the solution is unique and the system has 0 degrees of freedom.4.4 1. i. x2 = −a. b = 10/3. depending on the value of b. If a = 7.4.3 The determinant of the coefﬁcient matrix is a 2 − 7a = a(a − 7).4. i.2 (a) It is clear that x1 = x2 = x3 = 0 is a solution. the system has solutions if and only if −ab − 2b + 3a + 9 = 0. and then it has solutions with 1 degree of freedom. the rank of the coefﬁcient matrix is 2 (why?).e. Knut Sydsæter. Since there are 4 unknowns. the system has 4 − 3 = 1 degree of freedom.
F denote the matrices given in (a).5. where s and t are arbitrary real numbers 1 (c) The characteristic polynomial of C is C−λI = λ2 −25.1 For convenience. and Peter Hammond 2008 . and then we are left with the two equations x1 + 3x2 + 2x3 = 11 2x1 + 5x2 − 3x3 = 3 We can consider these equations as an equation system with x1 and x2 as the unknowns: x1 + 3x2 = 11 − 2x3 2x1 + 5x2 = 3 + 3x3 (1 ) (2 ) (1) (2) For each value of x3 this system has the unique solution x1 = 19x3 − 46. so we can remove it. i. The eigenvectors corresponding to an eigenvalue λ x are the vectors x = = 0 that satisfy Ax = λx. and we see immediately that the eigenvalues are λ1 = 5 and λ2 = −5. 2−λ −2 complex eigenvalues. (f). y 2x − 7y = −x 3x − 8y = −y and 2x − 7y = −5x 3x − 8y = −5y This gives us the eigenvectors v1 = s (different from 0). (b). . B. Knut Sydsæter. .5 1. Thus the vector equation A3 x = (11. . (b) The characteristic equation of B is B − λI = 4 = λ2 − 8λ + 20 = 0. x2 = −7x3 + 19. 1.e. 6) has the solution x = (19s − 46. . B has two 6−λ 7 3 ⇐⇒ 7x = 7y for λ = −5 ⇐⇒ 3x = 7y for λ = −1 and v2 = t 1 .4 CHAPTER 1 TOPICS IN LINEAR ALGEBRA Equation (3) is obviously equivalent to (2). let A. s) where s runs through all real numbers. . . . . 4 ± 2i. 3. Atle Seierstad. −7s + 19. 2 −7 (a) The characteristic polynomial of the matrix A = is 3 −8 A − λI = 2−λ 3 −7 = λ2 + 6λ + 5 = (λ + 1)(λ + 5) −8 − λ so the eigenvalues of A are λ1 = −1 and λ2 = −5. The eigenvectors are determined by the equation systems x + 4y = 5x 6x − y = 5y ⇐⇒ x = y and x + 4y = −5x 6x − y = −5y 3 ⇐⇒ y = − x 2 © Arne Strøm. and no real eigenvalues. respectively.
and v3 = u ⎝ −2 ⎠. and λ3 = 3. (f) The characteristic polynomial of F is 1 − λ −1 0 −1 2 − λ −1 = −λ3 + 4λ2 − 3λ = −λ(λ2 − 4λ + 3) = −λ(λ − 1)(λ − 3) 0 −1 1 − λ The eigenvalues are therefore λ1 = 0. λ2 = 1. v2 = t ⎝ 1 ⎠ . v2 = t ⎝ 0 ⎠. An extreme case is the identity matrix In : all (nonzero) nvectors are eigenvectors for In . v3 = u ⎝ 0 ⎠ 0 0 1 where s. (Note: The eigenvalues of a diagonal matrix are always precisely the diagonal elements. and Peter Hammond 2008 . Knut Sydsæter. 1 1 1 © Arne Strøm. By the same method as above we ﬁnd that ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ 1 −1 1 the corresponding eigenvectors are v1 = s ⎝ 1 ⎠. λ2 = 3.) (e) The characteristic polynomial of E is 2−λ 1 −1 = −λ3 + λ2 + 2λ = −λ(λ2 − λ − 2) 0 1−λ 1 2 0 −2 − λ The eigenvalues are the roots of the equation −λ(λ2 − λ − 2) = 0. t. and (multiples of) the standard unit vectors will be eigenvectors. The eigenvectors corresponding to λ1 = −1 are solutions of ⎧ ⎪ 2x1 + x2 − x3 = −x1 1 ⎨ x1 = 2 x3 x2 = −x1 Ex = −x ⇐⇒ ⇐⇒ x2 + x3 = −x2 ⇐⇒ 1 ⎪ x2 = − 2 x3 x3 = 2x1 ⎩ 2x1 − 2x3 = −x2 ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ 1 1 2 so they are of the form v1 = s ⎝ −1 ⎠. Atle Seierstad. there will be other eigenvectors as well. u are arbitrary nonzero numbers. so the eigenvectors are v1 = s 1 1 and v2 = t −2 3 where s and t are arbitrary nonzero numbers. λ2 = 0 and λ3 = 2.CHAPTER 1 TOPICS IN LINEAR ALGEBRA 5 respectively. Similarly. namely λ1 = −1. v2 = t ⎝ −1 ⎠ and v3 = u ⎝ 1 ⎠ are the eigenvectors 2 1 1 corresponding to λ2 = 0 and λ3 = 2. But if two or more of the diagonal elements are equal. and the corresponding eigenvectors are ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ 1 0 0 v1 = s ⎝ 0 ⎠ . (d) The characteristic polynomial of D is 2−λ D − λI = 0 0 0 0 = (2 − λ)(3 − λ)(4 − λ) 3−λ 0 0 4−λ The eigenvalues are obviously λ1 = 2. λ3 = 4.
0 ⎛ ⎞ 1 ⎜ 0⎟ ⎟ x3 = ⎜ ⎝ 0⎠ −1 ⎛ = (3 − λ)3 (7 − λ) These three vectors are obviously linearly independent because if ⎞ c1 + c2 + c3 ⎟ ⎜ −c1 ⎟ c1 x1 + c2 x2 + c3 x3 = ⎜ ⎠ ⎝ −c2 −c3 ⎛ is the zero vector. it is easy to show that p(A) = −A3 + (2a + b)A2 − 2abA = 0. Knut Sydsæter. then Cn x = λn x for every natural number n. x3 . so the (b) The characteristic polynomial of A is p(λ) = a a−λ 0 0 0 b−λ eigenvalues of A are λ1 = 0.5. then λ3 x = λ2 x + λx. 1. One simple set of solutions is ⎞ 1 ⎜ −1 ⎟ ⎟ x1 = ⎜ ⎝ 0⎠. namely x1 + x2 + x3 + x4 = 0 The system has solutions with 4 − 1 = 3 degrees of freedom. λ3 = b. Atle Seierstad. © Arne Strøm. Then λ3 − λ2 − λ = 0. y.9. If C3 = C2 + C. C + In  = 0. (b) An eigenvector x = (x1 . λ4 = 7. which is not true for λ = −1.5. 1. 0 ⎛ ⎞ 1 ⎜ 0⎟ ⎟ x2 = ⎜ ⎝ −1 ⎠ .7(b) yields 4−λ 1 1 1 4 1 4−λ 1 1 = (3 − λ)4 1 + A − λI = 1 1 4−λ 1 3−λ 1 1 1 4−λ Hence. and consequently C + In has an inverse. x4 ) of A corresponding to the eigenvalue λ = 3 must satisfy the equation system (A − 3I)x = 0. x2 . z) ⎝ ax + ay ⎠ = (ax 2 + ay 2 + 2axy + bz2 ) bz ⎞ ⎞ ⎛ 2 ⎛ 3 2 2a 2a 4a 4a 3 0 0 A2 = ⎝ 2a 2 2a 2 0 ⎠.5 (c) If λ is an eigenvalue for C with an associated eigenvector x. (c) From (b) we get p(λ) = −λ3 + (2a + b)λ2 − 2abλ. λ2 = 2a. Using the expressions for A2 and A3 that we found in part (a). and Peter Hammond 2008 . Then λ = −1 would be an eigenvalue for C.4 (a) The formula in Problem 1. the eigenvalues of A are λ1 = λ2 = λ3 = 3. because x = 0. The 4 equations in this system are all the same.5. and so we would have λ3 − λ2 − λ = 0.6 CHAPTER 1 TOPICS IN LINEAR ALGEBRA ⎞ ax + ay 1. Hence −1 is not an eigenvalue for C. so (λ3 − λ2 − λ)x = 0. If C + In did not have an inverse. A3 = ⎝ 4a 3 4a 3 0 ⎠ 0 0 b2 0 0 b3 ⎛ (a 1 × 1 matrix). then c1 = c2 = c3 = 0.2 (a) X AX = X (AX) = (x. a−λ a 0 = (λ2 − 2aλ)(b − λ).
1 (a) Let A = 2 1 1 . x2 . Eigenvectors x = (x1 .e. and Peter Hammond 2008 0 1 . P−1 = P .e. ⎝0⎠ 2 2 0 0 1 Fortunately.6. Knut Sydsæter.CHAPTER 1 TOPICS IN LINEAR ALGEBRA 7 1. and so we have a suitable orthogonal matrix √ ⎞ ⎛ 1√ 1 2 2 2 2 0 √ ⎟ ⎜ 1√ 1 P = ⎝ −2 2 2 2 0 ⎠ 0 © Arne Strøm. x3 arbitrary. The characteristic polynomial of A is 2 2−λ 1 1 = (2 − λ)2 − 1 = λ2 − 4λ + 3 = (λ − 1)(λ − 3) 2−λ p(λ) = Thus. √ 1 .) We therefore have the diagonalization √ √ √ √ 1 1 1 1 2 1 1 0 2 2 −2 2 2 2 2 2 −1 √ √ √ P AP = 1 √ = 1 1 1 1 2 0 3 −2 2 2 2 2 2 2 2 ⎛ ⎞ 1 1 0 (b) Let B = ⎝ 1 1 0 ⎠. −x1 + x2 =0 x1 − x2 =0 0=0 This gives x1 = x2 . these three vectors are mutually orthogonal (this is not automatically true for two eigenvectors associated with the same eigenvalue). It is easily seen that one eigenvector associated with the eigenvalue λ1 = 0 is x1 = (1. The characteristic polynomial of B is 0 0 2 1−λ 1 0 = (2 − λ) (1 − λ)2 − 1 = (2 − λ)(λ2 − 2λ) = −λ(λ − 2)2 1 1−λ 0 0 0 2−λ (use cofactor expansion of the ﬁrst determinant along the last row or the last column).6 1. The associated eigenvectors with length 1 are uniquely determined up to sign as √ √ 1 1 2 2 2 2 √ and x2 = 1 √ x1 = 1 −2 2 2 2 This yields the orthogonal matrix P= 1 2 2 √ 1 −2 2 √ 1 2 2 √ 1 2 2 √ (It is easy to verify that P P = I. 0) . x3 ) associated with the eigenvalue 2 are given by (B − 2I)x = 0. −1. i. i. Atle Seierstad. the eigenvalues are λ1 = 1 and λ2 = 3. The eigenvalues are λ1 = 0 and λ2 = λ3 = 2. One set of linearly independent eigenvectors with length 1 is then ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ 1 1 0 1 ⎝ 1 ⎝ ⎠ √ −1 ⎠ .
⎛ ⎞ 5 0 1 (b) The symmetric coefﬁcient matrix of Q is ⎝ 0 2 1 ⎠. −4) = diag(λ1 . But it is not always easy to see how to rewrite a quadratic form in this fashion. λ3 ). x2 ) = 0 only if x1 = x2 = 0.) © Arne Strøm. (An alternative way to see this is to write Q as a sum of squares: Q(x1 .5 For the given A.1 that Q is positive deﬁnite. v = ⎝ 3 ⎠ . An eigenvector x = (x. λ2 = 6. Therefore A3 = A2 A = (5A−5I)A = 5A2 −5A = 20A−25I 50 75 . v = w = 5 2. it follows from Theorem 1. x2 . x3 ) = (x1 + x3 )2 + 2 2 2 (x2 + x3 )2 + 4x1 + x2 + 2x3 is obviously nonnegative and it is zero only if all the square terms are zero.7.7. we have A2 = 5A−5I. and A4 = 20A2 − 25A = 75A − 100I = 75 125 1. y.8 CHAPTER 1 TOPICS IN LINEAR ALGEBRA It is now easy to verify that P−1 = P and ⎞ 0 0 0 P−1 BP = ⎝ 0 2 0 ⎠ 0 0 2 ⎛ ⎞ 1 3 4 (c) The characteristic polynomial of C = ⎝ 3 1 0 ⎠ is 4 0 1 1−λ 3 4 3 4 3 =4 1−λ 0 4 0 1−λ 1−λ 1−λ + (1 − λ) 0 3 3 = (1 − λ) (1 − λ)2 − 25 1−λ ⎛ (cofactor expansion along the last column). z) corresponding to the eigenvalue λ must satisfy x + 3y + 4z = λx Cx = λx ⇐⇒ 3x + y 4x = λy + z = λz One set of unnormalized eigenvectors is ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ 0 5 −5 u = ⎝ −4 ⎠ . λ2 . so Q is positive deﬁnite. The leading principal minors are 1 1 4 5 0 1 5 0 D1 = 5. and λ3 = −4. 6. and a corresponding orthogonal matrix is √ √ ⎞ ⎛ 5 5 0 10 2 − 10 2 √ √ ⎟ ⎜ 3 3 P = ⎝ − 4 10 2 5 10 2 ⎠ √ √ 3 4 4 5 10 2 10 2 A straightforward calculation conﬁrms that P CP = diag(1. Knut Sydsæter. D3 = 0 2 1 = 33 0 2 1 1 4 Since all the leading principal minors are positive. w = ⎝ 3 ⎠ 3 4 4 √ with lengths u = 5. Atle Seierstad.6.7 1.5 (a) It is clear that Q(x1 . and Peter Hammond 2008 . x2 ) ≥ 0 for all x1 and x2 and that Q(x1 . and so the eigenvalues are λ1 = 1. 1. D2 = = 10.
101 and c2 = 7 + 2 6 ≈ 11. c2 ). respectively. D2 = 2 > 0.7. Knut Sydsæter. then Q(x) = x Ax = x (λx) = λx x = λ x 2 = λ. (3) Q is indeﬁnite if c < c1 or c > c2 .8. by Theorem 1.1: 1 2 1 © Arne Strøm.4 Positive deﬁnite subject to the constraint.e.CHAPTER 1 TOPICS IN LINEAR ALGEBRA 9 (c) Since Q(x1 . 0 1 .e. 1. because the constraint is simply x 2 = 1. −1).8 0 0 1. which are λ1 = 9 and λ2 = −5. x2 ) = −(x1 − x2 )2 ≤ 0 for all x1 and x2 . if c lies in the closed interval [c1 . (ii) positive semideﬁnite if a11 ≥ 0. i. c2 ]. 1. and A ≥ 0.10 (b) If x satisﬁes the Lagrange conditions. if c lies outside the closed interval [c1 . and the leading principal minors are 0 2 −8 D1 = −3 < 0. i.7 (a) The symmetric coefﬁcient matrix of Q is A= a11 a21 a12 a22 = 3 −(5 + c)/2 −(5 + c)/2 2c Since a11 > 0. The determinant of A is 1 1 1 A = 6c − 4 (5 + c)2 = − 4 (c2 − 14c + 25) = − 4 (c − c1 )(c − c2 ) √ √ where c1 = 7 − 2 6 ≈ 2. the maximum and minimum values of Q(x) are simply the largest and smallest eigenvalue of Q. Atle Seierstad. It is (i) positive deﬁnite if a11 > 0 and A > 0. c2 ]. if c lies in the open interval (c1 .3 Negative deﬁnite subject to the constraint. 1. a22 ≥ 0. i. 0 0 1. x2 ) = 0 whenever x1 = x2 .8. (iii) indeﬁnite if A < 0.7. the form can never be negative semideﬁnite. since Q(x1 . It follows that (1) Q is positive deﬁnite if c1 < c < c2 . and Peter Hammond 2008 0 0 4 −2 1 0 0 2 −1 −3 1 4 −5 1 2 1 2 1 1 0 1 −2 1 −1 0 1 1 2 = −180 < 0 −2 2 −1 1 1 0 1 −3 0 = 25 > 0. ⎛ ⎞ −3 1 0 (d) The symmetric coefﬁcient matrix of Q is ⎝ 1 −1 2 ⎠.7. (2) Q is positive semideﬁnite if c1 ≤ c ≤ c2 .8. But Q is not deﬁnite.e. The corresponding √ √ 1 1 maximum and minimum points (eigenvectors) are ± 2 2 (1. and D3 = −4 < 0. Q is negative semideﬁnite. by Theorem 1. Q is negative deﬁnite.899 are the roots of the quadratic equation c2 − 14c + 25 = 0.1.1: 1 1 1 0. Hence. By Theorem 1.8. 1) and ± 2 2 (1.
. where W = I2 0 .4) we get = A22 . 1. −1 1 −1 0 1 I3 W −1 −1 1 0 1 0 0 0 0 Further. ⎠ X A . Thus the upper right submatrix of A−1 must be the transpose of the lower left submatrix.9. Knut Sydsæter. Then 1 = I2 − WW = 2 0 0 1 1 . . where A and X . −xn ⎞ . Atle Seierstad.9. A12 and A21 will both be zero matrices. A−1 must also be symmetric. A bit of calculation shows that − −1 W = 2 W . 1. In either case we ﬁnd that . . so −1 = 2 0 −1 −2 ⎛ ⎞ 1 1 1 1 1 2I3 − WW I3 − 2 WW 2 W W 1 = 2 . where I4 and I1 are the The matrix in (c) can be partitioned as A = A21 A22 v I1 identity matrices of orders 4 and 1. we get = I1 − v v = (−2) = −2I1 . and Peter Hammond 2008 . 0 A22 −1 0 0 A11 A−1 11 whereas formula (1. . and in (b) it is natural to let A11 be the upper left 1×1 matrix. ⎟ 2 0⎠ 0 −1 0 1 0 1 −1 = 1 2 v.4 The matrix B = ⎜ . . We evaluate the determinant of B by each of the formulas (6) and (7): By formula (6).9. 0). .9 1. = 0 A22 0 A−1 22 and the answers to (a) and (b) are as given in the book. 1. so A−1 = 1 −1 W W 2 −1 2 1 1 1 (Note that because A is symmetric.9. In both cases the partitioned matrix will be of the form 0 A11 . Then A−1 + A−1 A12 −1 A21 A−1 = I4 − 2 vv = I4 − 2 ⎜ 11 11 11 ⎝ 1 1 1 0 ⎠ = 2 ⎝ −1 −1 1 0⎠ and − −1 A A−1 21 11 = 1 2 I1 v I4 = 1 2v A−1 = I4 − 1 2 vv 1 2v 1 2v 1 − 2 I1 1 ⎜ −1 ⎜ 1 = 2 ⎜ −1 ⎜ ⎝ 0 1 ⎛ = 1 2 (1. . respectively. . ⎝ . B = A · I1 − (−X )A−1 X = A · I1 + X A−1 X where the last factor is the determinant of a 1 × 1 matrix and therefore equal to the single element 1 + X A−1 X of that matrix. −A−1 A12 11 ⎞ 0 1 0 1⎟ ⎟ 0 1 ⎟.e.4) in the book. 0 0 0 2 so we ﬁnally get An alternative partitioning is A = I2 − 0 0 0 3 = 1 0 W . xn an1 . as before. −1 B = I1  · A − XI1 (−X ) = A + XX  and by formula (7). 0) is the transpose of the 4 × 1 matrix (column vector) v. . i.3 The obvious partitioning to use in (a) is with A11 as 2×2 matrix in the upper left corner of A. If we use formula (1.) ⎛1 ⎜ x1 1. © Arne Strøm. which helps us save a little work. 1.10 CHAPTER 1 TOPICS IN LINEAR ALGEBRA 1.5) gives ˜ = A11 . −x1 a11 . ann are as given in the problem. I4 v A11 A12 = . and WW = ⎝ 1 1 1 ⎠. ⎛ ⎞ ⎛ ⎞ 1 1 1 0 1 −1 −1 0 ⎜ 1 1 1 0 ⎟ 1 ⎜ −1 1 −1 0 ⎟ 1 1 ⎟ ⎜ ⎟. If we use formula (1. . a1n ⎟ I1 −X . . .9. and v = (1. . ⎟ can be partitioned as B = .
Each of these terms is the product of a sign factor and n elements from A. A11 A21 A − A12 A−1 A21 A12 22 = 11 A21 A22 −A Im 0 = A11 − A12 A−1 A21  A22  22 A22 1. A = A11  A22 .9. Atle Seierstad.) All such pairs of terms will occur exactly once and so A = 1 2 . and then operate on the last n − k rows of A such that A22 becomes upper triangular. then the second column.) A term in this sum will automatically be zero unless the factors from the ﬁrst k columns are taken from A11 and the last n − k factors from A22 . but we can make it so by means of elementary row operations. chosen so that no two elements are taken from the same row of A or from the same column. We will consider 0 A22 two ways to demonstrate that A = A11  A22 . where A11 is a k × k matrix and A22 is (n − k) × (n − k). all elements below the main diagonal are 0. i. We know that the determinant of a triangular matrix equals the product of the elements on the main diagonal. The sign factor ±1 is determined by the positions of the elements selected. the ﬁrst factor on the left has determinant Ik  In−k  = 1. By the formula we showed for the upper triangular case. We perform such operations on the ﬁrst k rows of A in such a way that A11 becomes upper triangular.7 (a) With D = In 0 A Im DE = and E = In B we get and ED = In B 0 Im + BA In + AB 0 B Im Cofactor expansion of the determinant of DE along each of the last m columns shows that DE = In + AB. cofactor expansion along each of the ﬁrst n rows shows that ED = Im + BA. simply look at the determinant 11 A22 0 A21 of the A22 (b) The equality follows by direct multiplication. by (a) again.9. Then A11 and A22 are also upper triangular. © Arne Strøm. The number σ of sign changes that A undergoes is then the sum of the numbers σ1 and σ2 of sign changes inﬂicted on A11  and A22 . that is.CHAPTER 1 TOPICS IN LINEAR ALGEBRA 11 1. If the factors are selected in this way. (See EMEA or almost any book on linear algebra for details.) In this case it is clear that A = A11  A22 .e. Of course. respectively. Alternatively. (a) Let A = (I) By deﬁnition. (Just think of cofactor expansion along the ﬁrst column. we could use formula (7) with A22 = Im to evaluate DE and formula (6) with A11 = In to evaluate ED. A11 A21 A 0 = A11  A22 . and Peter Hammond 2008 . (The sign factors will match. the term in question will be a product of one term in the sum 1 making up A11  and one from the sum 2 that makes up A22 . Operation (i) does not affect the value of the determinant. we get A = A11  A22  in the general case too. the determinant of the n × n matrix A is a sum A of n! terms. more speciﬁcally the operations of (i) adding a multiple of one row to another row and (ii) interchanging two rows. By the result in (a). Similarly. and since σ1 + σ2 = σ . Knut Sydsæter. etc. in the general case A need not be upper triangular at all.6 A11 A12 . and so. To show the formula transposed matrix. (II) Suppose that A is upper triangular. while operation (ii) multiplies the value by −1. (−1)σ A = (−1)σ1 A11  · (−1)σ2 A22 .
⎟ . −1. . Therefore a1 ⎜ a1 − 1 ⎜ ⎜ 1 ⎜ F = In + AB = ⎜ a2 − 1 ⎜ ⎜ . Formula (2. AB is an n × n matrix with every column equal to A. 1) = √ = 2 2 2 2 Note: It is pure coincidence that the gradient of f at (2.. . . .. 1) · (1. Atle Seierstad. −2z). −2) · (1.. ⎝ 1 an − 1 ⎛a 1 ⎛ and 1 a1 − 1 a2 a2 − 1 . 2. .1. .5 (a) The vector from (3..1) gives √ 1 2 2 3 gb (0.. . 1) · a = √ ∇f (2. ⎟ . 1) is a = the directional derivative of f in this direction at (2. 1 ⎞ a1 − 1 ⎟ ⎟ 1 ⎟ ⎟ a2 − 1 ⎟ ⎟ . 1 1 .. x 2 exy − x.1. .. 1) is 1 v v= 1 √ (1. . (b) The gradient of g is ∇g(x. 1) · b = √ (0.. . 2 + y 2 + z2 2 + z2 x 2 + y 2 + z2 x x +y © Arne Strøm. 1 an − 1 0 a2 − 1 . 0 From the result in (a) it follows that G = (a1 − 1)(a2 − 1) · · · (an − 1) In + AB = (a1 − 1)(a2 − 1) · · · (an − 1) I1 + BA n = (a1 − 1)(a2 − 1) · · · (an − 1) 1 + i=1 1 ai − 1 Chapter 2 Multivariable Calculus 2. 1. an − 1 ⎞ ⎟ ⎟ (In + AB) ⎠ ⎜1 G=⎜ .. 2) is (−1. 1) = ∇g(0. and the unit vector in this direction is a = √1 (−4. 1) = (−4. 1) = √ (2. . 1. y. 0 . . 1.1.8. 1. 1).. 2) − (3. 1) is b = √3 (1..1 2. . 1)... The gradient of f is 18 ∇f (x. . 1 1 a2 . z) = y ln(x 2 + y 2 + z2 ) + 2x 2 y 2xy 2 2xyz .1) equals (2. an ⎞ ⎛a − 1 1 ⎟ ⎜ 0 ⎟=⎜ . −1. 0.. . ⎠ an an − 1 0 0 . 1) to (−1. x ln(x 2 + y 2 + z2 ) + 2 . . 2 By formula (2. . ⎠ ⎝ . 2. and the unit vector in the 1 direction given by (1.12 CHAPTER 2 MULTIVARIABLE CALCULUS (b) With A and B as in the hint.1). 1).. ⎝ . 1.8) √ 1 3 2 1 3 fa (2. .3 (a) The unit vector in the direction given by v = (1.. z) = ((1 + xy)exy − y. 1. and Peter Hammond 2008 . . . 1) = − √ = − 3 3 3 2.. Knut Sydsæter. . . 1) · (1. . .. 1 .. 1). y. 1) = ∇f (2.. .. . ⎜ . 1.
2. 2. and Peter Hammond 2008 n = 1. . For example.9 (a) We know that y = −F1 (x. but we √ know that 2 is irrational. 0) = ∇f (0. The triangle inequality shows that S1 is convex. let t = r1 + (r2 − r1 )/ 2. but it is not convex.1. (b) S1 is the interior of the ball S. y) F F −F F F F F −F F = − 11 2 2 1 21 · 1 − 12 2 2 1 22 − 1 (F2 ) (F2 ) F2 = dy dx −F11 (F2 )2 + 2F12 F1 F2 − F22 (F1 )2 −(F11 F2 − F1 F21 )F2 + (F12 F2 − F1 F22 )F1 = (F2 )3 (F2 )3 (Remember that F21 = F12 . It follows that S is convex. 1. 1]. Let x1 and x2 be points in S.2. with x1 < x2 . −1. then x ≤ r. 1) = (ln 3+2/3.7 (a) Let us call a set S of numbers midpoint convex if 2 (x1 + x2 ) whenever x1 and x2 belong to S. . for between any two rational √ numbers r1 and r2 there are always irrational numbers.6 (a) If x and y are points in S and λ ∈ [0. . and by the triangle inequality. y). 2n 2 2 2n © Arne Strøm. (b) Suppose S is midpoint convex and closed.6 The unit vector in the direction of maximal increase of f at (0. Neither S2 nor S3 is convex. and let λ in (0.2. y = ∂ F1 (x. Then t √ lies between r1 and r2 .) Expanding the determinant in the problem yields precisely the numerator in the last fraction. 1) the direction of fastest growth for f is given by ∇f (1. 0) is a = √1 (1. On the other hand. 1. We shall prove that the point z = λx1 +(1−λ)x2 belongs to S. 1) in the direction a is √ fa (1. fa (0. λx + (1 − λ)y belongs to S. . −1. 1).8) the directional derivative of f at (1. y) dx (y ) = − − dx ∂x F2 (x. . while S3 consists of the shell and the part of n that lies outside S. 2n .CHAPTER 2 MULTIVARIABLE CALCULUS 13 By formula (2. 3. 3) and therefore 10 ∇f (0. 2/3). y)/F2 (x. t = 4 and √ √ 2 10 4 4 10 (1. k = 0.1. 2/3) · (−4. 2. 1) = (1/ 18 ) ∇f (1. then 2 = (r2 − r1 )/(t − r1 ) would also be rational. 1.2 2. Hence. The set S = of rational numbers is midpoint convex.1. 0) = 4a = √ (1. and we ﬁnd that S must contain all points of the form 2n − k k k k x1 + x2 = n x1 + 1 − n x2 . 3) ∇f (0. We also know that fa (0. If t were rational.e. it must contain 1 1 3 1 1 1 3 y1 = 2 (x1 + x2 ). 0) = ta for a number t ≥ 0. and then it contains y21 = 2 (x1 + y1 ) = 4 x1 + 4 x2 and y22 = 2 (y1 + x2 ) = 4 x1 + 4 x2 . 1. ln 3 + 2/3. 1. and by the chain rule for functions of several variables. Knut Sydsæter. Since S is midpoint convex. λx + (1 − λ)y ≤ λ x + (1 − λ) y ≤ λr + (1 − λ)r = r. We can continue in this fashion. y) dx ∂y F2 (x. 1) · (−4. 1) √ √ = (1/ 18 ) (ln 3 + 2/3. 0) · a = ta · a = t a 2 = t. The set S2 is the spherical shell we mentioned.1. . 3) = (1. 1) = −(5 ln 3 + 8/3)/ 18 (b) At (1. by (2. ln 3+2/3. 3) = 10 5 10 2. 1. 1 2. . Hence.8). constructing new midpoints between the points that have already been constructed. . what we get when we remove the spherical “shell” {x : x = r} from S. y ≤ r. y) d ∂ F1 (x. 0) = 4. i. Atle Seierstad.
because z11 = ex+y + ex−y > 0 and z11 z22 − (z12 )2 = 4e2x > 0. it is linear along each ray from the origin and therefore it cannot be strictly convex or strictly concave for any value of ρ.9 (a) This is mainly an exercise in manipulating determinants. y) = −ex+y . and therefore constant along each line joining two points in such a plane.6 on homogeneous functions in EMEA. for each n we ﬁnd 2n + 1 evenly spaced points in the interval [x1 .3. see e.6 (b) Let λ2 > λ1 > 0 and deﬁne μ = λ1 /λ2 . then f (λx)/λ is strictly decreasing as a function of λ. where u = x + 2y + 3z. i.14 CHAPTER 2 MULTIVARIABLE CALCULUS Thus.2.3. 2. however. the “direct” second derivatives f11 and f22 have the same sign as −(ρ + 1). but with care and patience you will ﬁnd that 2 f11 (v1 . Now let r be any positive number. Section 12. v2 ) = A(δ1 v1 + δ2 v2 ) is a linear function. f12 (v1 .3. f11 f22 − (f12 )2 = 0 everywhere. Atle Seierstad. b). In order to prove that S is an interval.1. 2. so z11 (x. Let x be a point in (a. there is an α in S with α < x. 3 2 f (x) = 2. Let n be so large that (x2 − x1 )/2n < 2r. x2 ] that all belong to S. 1 3 1 (c) Take any x = 0 in the domain of f . and consider the open rball (open interval) B = B(z. Therefore f cannot be strictly concave. z is a strictly concave function of x and y. 1) and by the concavity of f we have μf (λ2 x) + (1 − μ)f (0) ≤ f (μλ2 x + (1 − μ)0). It follows that z does indeed belong to cl(S) = S.8 Let S be a convex subset of containing more than one point. v2 ) = −P v1 −ρ −(1/ρ)−2 where P = (ρ + 1)δ1 δ2 A(v1 v2 )−ρ−2 δ1 v1 + δ2 v2 These formulas show that for all v1 and v2 .e. It also follows that if f is strictly concave. so B contains at least one point from S. It follows from Theorem 2. y) = −ex − ex+y . So w is a convex function of an afﬁne function. x belongs to S. v2 ) = P v1 v2 . Also.3 2. hence convex according to (2. Then μ ∈ (0. and since S is convex. and concave if ρ ≥ −1.3.1 that f is convex if ρ ≤ −1. −ρ 2 f22 (v1 . so f (λ2 x)/λ2 ≤ f (λ1 x)/λ1 . Since x < b = sup S. b). and let a = inf S.3. The distance between two neighbouring points in this collection is (x2 − x1 )/2n . The equation f11 f22 − (f12 )2 = 0 is a consequence of the fact that f is homogeneous of degree 1. By Theorem 2. Knut Sydsæter.5 (a) z11 (x. Then B must contain at least one of the points (k/2n )x1 + (1 − k/2n )x2 constructed above. If ρ = −1 then f (v1 . (λ1 /λ2 )f (λ2 x) ≤ f (λ1 x). it sufﬁces to show that S contains the open interval (a. z12 (x. Since f is homogeneous of degree 1. Then x is a convex combination of α and β. z + r) around z.8 The challenge here is mainly in getting the derivatives right.g. y) = −ex+y . Then x = 2x and f ( 2 x + 2 2x) = f ( 2 x) = 1 1 1 2 f (x) + f (x) = 2 f (x) + 2 f (2x). Then a < b. b = sup S (where a and b may be ﬁnite or inﬁnite). r) = (z − r. because it is constant on every plane of the form x + 2y + 3z = c.3. try to write them out in full for the case k = 3 (or k = 2).3.8). Similarly. 2. and Peter Hammond 2008 . there exists a β in S with x < β. 2. If you feel that the calculations below look frightening. Note that zij = ai aj z/xi xj © Arne Strøm. (c) w = u2 . It is not strictly convex. v2 ) = −P v2 . which indeed is both convex and concave. y) < 0 and z11 z22 − (z12 )2 = e2x+y > 0. (b) z is strictly convex. and z22 (x.
. Therefore Dk i=1 i=1 has the same sign as (−1)k . . xk )2 (c) By assumption. .20) tells us that a common factor in any column (or row) in a determinant can be “moved outside”. (b) More determinant calculations. . = (−1)k−1 (sk − 1) .. . . . then sk = k ai < 1 for all k. z1k z2k . · · · ak − 1 ··· ··· . . . . so 1 a2 . . . xk )2 = (sk − 1) 1 a2 − 1 . zkk .. . . . . ak xk ··· ··· . . . . zk1 z12 z22 . . ak ··· a1 ··· a2 . a1 a2 z x1 x2 a2 (a2 − 1) z 2 x2 . a k a2 z xk x2 ··· ··· . .. . . . ak · · · sk − 1 ··· a2 . and zii = ai (ai − 1)z/xi2 . ak ak − 1 xk where equality (1) holds because aj z/xj is a common factor in column j for each j and equality (2) holds because 1/xi is a common factor in row i for each i. . k to the ﬁrst row. zk2 . ak · · · ak − 1 Now subtract column 1 from all the other columns. . . . Atle Seierstad. .. ak (ak − 1) z 2 xk a1 a2 − 1 . . ... . Then each entry in the ﬁrst row becomes equal to sk − 1. ak xk ··· a1 x1 a1 − 1 a2 a2 (2) a1 a2 . . .. . . xk )2 .22) says that this does not change the value of the determinant. . sk − 1 a2 . ak k z (x1 x2 . Therefore. ai > 0 for all i. . .1. . . a 1 ak z x1 xk a2 ak z x2 xk . Dk = a1 a2 .. . . and Peter Hammond 2008 . . Knut Sydsæter. . . . . It follows from Theorem 2. ak sk − 1 a2 − 1 .. . . Rule (1. . a1 (a1 − 1) z 2 x1 a2 a1 z x2 x1 = . . . . . and add rows 2. We use the expression for Dk that i=1 we found in part (a). ak k z x1 x2 . ak a1 a2 . Let sk = k ai = a1 + · · · + ak .2(b) that f is strictly concave.. . ak a1 z xk x1 a1 x1 a2 − 1 x2 . . · · · ak − 1 z11 Dk = z21 . . 1 a2 . . . xk a1 − 1 x1 a2 x2 . . xk )2 1 a2 . . 0 · · · −1 Dk = (sk − 1) a1 a2 . © Arne Strøm. (x1 x2 . . . . 3. . . . . ··· (1) = a1 a2 .CHAPTER 2 MULTIVARIABLE CALCULUS 15 for i = j .. ak k z . . so if n ai < 1.. xk )2 . . ak 0 ··· 0 −1 · · · 0 a1 a2 . Afterwards we take the common factor sk − 1 in row 1 and move it outside. ak k z (x1 x2 . . ak x2 = zk . ak k z (x1 x2 . . .1. . .. . (x1 x2 .3. . Rule (1. . . .
and (λk /μ) + (λk+1 /μ) = 1. It is obvious that A(1) is true. . y0 ) © Arne Strøm. μf (y) = μf λk+1 λk xk + xk+1 μ μ ≥μ λk λk+1 f (xk ) + f (xk+1 ) = λk f (xk ) + λk+1 f (xk+1 ) μ μ (∗) and this inequality together with (∗) yields f (λ1 x1 + · · · + λk+1 xk+1 ) ≥ λ1 f (x1 ) + · · · + λk−1 f (xk−1 ) + λk f (xk ) + λk+1 f (xk+1 ) which shows that A(k + 1) is true. we have f (x. . It then follows by induction that A(m) is true for all m ≥ 2.1(c) tells us that f (x. Atle Seierstad. . Then μ = λk + λk+1 > 0 and we can deﬁne y = (1/μ)(λk xk + λk+1 xk+1 ). Then z = λx + (1 − λ)y belongs to S. it follows that λf (x) + (1 − λ)f (y) − f (z) = λ[f (x) − f (z)] + (1 − λ)[f (y) − f (z)] ≤ p · [λ(x − z) + (1 − λ)(y − z)] = p · 0 = 0 by the deﬁnition of z. y0 ). y − y0 ) = p(x. .4. Let f be a function deﬁned on a convex set S in n and for each natural number m let A(m) be the following statement: “The inequality f (λ1 x1 + · · · + λm xm ) ≥ λ1 f (x1 ) + · · · + λm f (xm ) holds for all x1 . We can assume that λk+1 > 0. since it just says that f (x) = f (x). and therefore f (x) − f (z) ≤ p · (x − z) and f (y) − f (z) ≤ p · (y − z) Since both λ and 1 − λ are nonnegative. Let x1 . λk+1 be nonnegative numbers with sum 1. . . since f is concave. x0 ) + p(y. 2.4. . y0 ) = 3 3 (4x0 .4. y belong to S and let λ ∈ [0.4. . By A(k) we have f (λ1 x1 + · · · + λk+1 xk+1 ) = f (λ1 x1 + · · · + λk−1 xk−1 + μy) ≥ λ1 f (x1 ) + · · · + λk−1 f (xk−1 ) + μf (y) Moreover. . y) = x 4 + y 4 is a strictly convex function of (x.” We shall prove that A(m) is true for every natural number m. y) if and only if f (x. . . To show that f is concave.4. 4y0 ). f has a supergradient p at z. Knut Sydsæter. it is sufﬁcient to show that λf (x) + (1 − λ)f (y) ≤ f (z). λm ≥ 0 with λ1 + · · · + λm = 1. .5 Let x. We shall prove that A(k + 1) is also true. To sum up. y0 ) − ∇f (x0 . 1].6 Theorem 2. and so y ∈ S. y) = (x0 . 2. Note that both (i) and (ii) are necessary for this conclusion. and A(2) is also true. y) − f (x0 .3 We shall show that Jensen’s inequality (2. By assumption. xm in S and all λ1 ≥ 0.4 2. and Peter Hammond 2008 . .2) holds for all natural numbers m. Now suppose that A(k) is true. since f is concave. we have shown that: (i) A(2) is true. y0 ) · (x − x0 . Since ∇f (x0 . xk+1 be points in S and let λ1 . y0 ) > ∇f (x0 . for otherwise we are really in the case m = k. y0 ) · (x − x0 . The point y is a convex combination of xk and xk+1 . where k is some natural number greater than 1. y − y0 ) whenever (x. y) − f (x0 . . (ii) A(k) ⇒ A(k + 1) for every natural number k ≥ 2. . . not just for m = 3.16 CHAPTER 2 MULTIVARIABLE CALCULUS 2.
so f is not quasiconcave. This argument also holds if a = −∞ or b = ∞. if f is decreasing to the left of c and increasing to the right. Hence. b) and that there is a point c in (a. y) ≥ −1 is P−1 = {(x. x0 ) + p(y. −2/3]. Then f is quasiconcave on the interval (a.5. © Arne Strøm. is not quasiconcave. Their sum.5 2.1(a) and note that if x and y are points in (a. The second factor here is 0 if t = t0 . Alternatively we can use Theorem 2. and Peter Hammond 2008 .5.5. and polynomial division yields 2 2 p(t. b such that f is increasing on (a. y) : y ≤ x −2/3 }. Now suppose f is a function deﬁned on an interval (a.5. b). and c < x < y. so it is divisible by t − t0 . f (y)} for all z between x and y. We just have to look at each of the three possibilities x < y ≤ c. so it follows that p(t.5. M2. and decreasing in [−2/3. where P−1 is the unshaded part of the plane). b).5. b). (c) The set of points for which f (x.2(c). p(x. Atle Seierstad. 2. 2. Alternatively.) Then the upper level sets must be intervals (or empty). y] in each case. p(t. and also in the case of a closed or halfopen interval. then f is quasiconvex. Similarly.2 (a) f is linear.6 Since f is decreasing and g is increasing. t0 ) = t 4 − t0 − 4t0 (t − t0 ) for all t and t0 .CHAPTER 2 MULTIVARIABLE CALCULUS 17 4 3 2 3 where p(t.5.2(c) Neither P a nor Pa y 3 f (x) ≤ −1 2 1 1 2 3 4 x 3 2 1 1 y y = x3 + x2 + 1 y = f (x) 1 2 3 x Figure M2. though. We could use this argument in Problem 2. ∞). and it follows that f is quasiconcave.2 that both of these functions are quasiconcave as well as quasiconvex.) (d) The polynomial x 3 + x 2 + 1 is increasing in the interval (−∞.2(d).2(d). is convex for a = −1. 0].2(d) The graph of f . which is not a convex set (see Fig. (See Fig. for instance. then f (z) ≥ min{f (x). t0 ) = (t − t0 )(t 3 + t 2 t0 + tt0 − 3t0 ).4. x ≤ c < y. we could use the result in the note below.5. and consider the behaviour of f over the interval [x. t0 ) > 0 whenever t = t0 .2 in the book that a function of one variable that is increasing or decreasing on a whole interval is both quasiconcave and quasiconvex on that interval. (It is quasiconvex in the ﬁrst quadrant. it follows from Example 2.5. c] and decreasing on [c. t0 ) = (t − t0 )(t − t0 )(t 2 + 2tt0 + 3t0 ) = (t − t0 )2 [(t + t0 )2 + 2t0 ] The expression in square brackets is strictly positive unless t = t0 = 0. A note on quasiconcave functions of one variable It is shown in Example 2. y0 ) > 0 unless both x = x0 and y = y0 . f (x) + g(x) = x 3 − x. So f is increasing in (−∞. so it is concave and therefore also quasiconcave. Now. 4 f (x) ≤ −1 3 2 1 4 3 2 1 1 2 Figure M2. M2. This follows because the upper level sets must be intervals. −2/3] and decreasing in [−2/3. Knut Sydsæter.
2. y) = exy + xyexy . and in such cases f cannot be strictly concave. y) = xexy .5. Let x = y and let λ ∈ (0. 1). and (see the original proof) f (λx + (1 − λ)y) = f (μx + (1 − μ)y ) > f (x ) = f (y ) = μf (x ) + (1 − μ)f (y ) = (μβ/α)f (x) + (β − βμ)f (y) = λf (x) + (1 − λ)f (y) with strict inequality because f is strictly quasiconcave.3.5. or in parts of one or both of these intervals. We want to prove that f is strictly quasiconcave. we have f (z) > f (x). which is not a convex set of . It follows that h(λx + (1 − λ)y) = h (λ + (1 − λ)t)x = λ + (1 − λ)t h(x) > λ + (1 − λ)t q h(x) = λh(x) + (1 − λ)h(tx) = λh(x) + (1 − λ)h(y) q 2.6 2. 0) + f12 (0. we would like to change the name of the function from f to h. (B) Suppose x ∗ < z. © Arne Strøm. Similarly. We can assume that x = 0 and y = tx for some nonnegative number t = 1. y]. We get x = y and f (x ) = f (y ). it is clear from Fig. Since 0 < q < 1. 0)xy = 1 + xy. f (y)}. f11 (x. Then x ∗ < z < y. In both cases we get f (z) > min{f (x). y) ≈ f (0. for example. and Peter Hammond 2008 . and therefore h(λx + (1 − λ)y) = f (λx + (1 − λ)y) q > λf (x) + (1 − λ)f (y) q > λf (x)q + (1 − λ)f (y)q = λh(x) + (1 − λ)h(y) It remains to show that h(λx + (1 − λ)y) > λh(x) + (1 − λ)h(y) in the case where x and y lie on the same ray from the origin. f (y)}. and so f is strictly quasiconcave (b) No. and we want to prove that h is strictly concave in the convex cone K. x . A2.5. and y as in the proof of Theorem 2.11 With apologies to the reader. it may. the qth power function t → t q is strictly increasing and strictly concave. Since the qth power function is strictly concave. Even if f is concave. f12 (x. so we are left with the quadratic approximation f (x. β. x ∗ ] and [x ∗ . ∞). be linear in each of the intervals (−∞. Atle Seierstad. the lower level set P 0 = (−∞. and so f + g is not quasiconvex either.7 (a) Let f be singlepeaked with a peak at x ∗ . Assume ﬁrst that x and y do not lie on the same ray from the origin in n . we want to prove that if x < y and z is a point strictly between x and y.6. x ∗ ]. these derivatives all vanish at the origin. With the exception of f12 . f2 (x. Deﬁne a new function f by f (x) = h(x)1/q . ∞). Then both f (x) and f (y) are strictly positive and we can deﬁne α. 1] is not convex. y) = yexy . 0] ∪ [1. with h(x) > 0 for all x = 0 in K and h(0) = 0.18 CHAPTER 2 MULTIVARIABLE CALCULUS however. That is.6 in the answer section of the book that the upper level set P0 = {x : f (x) + g(x) ≥ 0} = [−1. and since f is strictly decreasing in [x ∗ . −1] ∪ [0. So h(x) is strictly quasiconcave and homogeneous of degree q ∈ (0. For instance. (λ + (1 − λ)t)q > λ1q + (1 − λ)t q = λ + (1 − λ)t q . we get f (z) > f (y). y) = y 2 exy .y . f22 (x. Since x < z ≤ x ∗ and f is strictly increasing in the interval [x. then f (z) > min{f (x). μ.5. Then f (x) is also strictly quasiconcave (use the deﬁnition) and satisﬁes the conditions of Theorem 2. y) = x 2 ex. There are two cases to consider: (A) Suppose z ≤ x ∗ .3. 2. Knut Sydsæter.1 (a) f1 (x. 1).5.
That procedure leads to some rather nasty fractions with a good chance of going wrong. f2 (0.4 z is deﬁned implicitly as a function of x and y by the equation F (x.and secondorder derivatives are: f1 (x. 0) = 0. f11 (x. y) = . z) = (0. y) ≈ f (0. and Peter Hammond 2008 . f2 (x. and zyy = 1. f1 (0. y) = − . f22 (0. f11 (0. y) = −2yex −y .2 (a) F (x. In order to ﬁnd the Taylor polynomial we need the ﬁrst. Knut Sydsæter. z is a C 1 function of x and y around this point. f12 (x. f2 (0. z) = x 3 + y 3 + z3 − xyz − 1 is obviously C 1 everywhere. z) and the corresponding formula for zy . 0) is therefore 1 3 z ≈ 1 − x + y + 2 x 2 − 2xy + 2 y 2 (The ﬁrst printing of the book has a misprint in the answer to this problem.7 2. y. 0)x + f2 (0. f22 (0. One possibility is to use the formula zx = −F1 (x. 1 2 . 0)xy + 2 f22 (0. 0) we get f (0. y) around (0. too—the coefﬁcient of x 2 is wrong. y) = −4xyex −y . 1) = 3 = 0. f12 (x.and secondorder derivatives of z with respect to x and y. f2 (x. 2 2 f22 (x. f1 (0. 0) = −4. 0) is 1 1 f (x. 0) = −2. y) = (2 + 4x 2 )ex −y . where F (x. 0) = f2 (0. f11 (0. We get zx /z = 3x 2 y − z − xzx zy /z = x − xzy + 1 3 ⇒ ⇒ (1 + xz)zx = 3x 2 yz − z2 (1 + xz)zy = x z + z 3 (1) (2) Taking derivatives with respect to x and y in (1). y) = − 2 2 (1 + x + 2y) (1 + x + 2y) (1 + x + 2y)2 At (0. so instead we shall differentiate the equation ln z = x 3 y − xz + y. Atle Seierstad. so by the implicit function theorem the equation deﬁnes z as a C 1 function g(x. 0)y 2 1 = 1 + x + 2y − 2 x 2 − 2xy − 2y 2 (There is a misprint in the answer in the ﬁrst printing of the book. 0) = 2. 0) = 1. z) = ln z − x 3 y + xz − y.CHAPTER 2 − 2 2 2 MULTIVARIABLE CALCULUS 2 2 2 19 2 (b) f1 (x. keeping in mind that z is a function of x and y. we get xzy zy + (1 + xz)zyy = x 3 zy + zy Now that we have ﬁnished taking derivatives. f12 (0. f22 (x. y. (c) The ﬁrst. y. and the quadratic approximation to f (x. Since F3 (x. and F3 (0. z) = 0. y. zxy = −2. Hence. The quadratic approximation to z around (0. y) = 1 + x + 2y 1 + x + 2y 2 4 1 . 0) = 2. 0. and then take derivatives of these expressions to ﬁnd the secondorder derivatives. 0) = 1.7. Equations (1) and (2) give zx = −1 and zy = 1 (at the particular point we are interested in). 0)x 2 + 2 f22 (0. y) = 2xex y . 0)y 2 = 1 + x 2 − y 2 . and then (3)–(5) give zxx = 3. 0)x 2 + f12 (0. 0) + 2 f11 (0. 1). 0)y + 2 f11 (0.6. z) = 1/z + x = 1 = 0 at (x. we can let x = y = 0 and z = 1 in the equations we have found. y. With x = y = 0 we get ln z = 0.) 2. 0) = 0. y) ≈ f (0.) 2. 1 1 Quadratic approximation: f (x. y) = (−2 + 4y 2 )ex −y . 0) = −2. 0) + f1 (0. so z = 1. y) in a neighbourhood © Arne Strøm. y) = − . f11 (x. y. y. z)/F3 (x. 0. we get (z + xzx )zx + (1 + xz)zxx = 6xyz + 3x 2 yzx − 2zzx xzy zx + (1 + xz)zxy = 3x z + 3x yzy − 2zzy 2 2 (3) (4) (5) and differentiating (2) with respect to y. 0) = −1.
20
CHAPTER 2
MULTIVARIABLE CALCULUS
of (x0 , y0 , z0 ) = (0, 0, 1). To ﬁnd the partial derivatives g1 and g2 there is no need to use the matrix formulation in Theorem 2.7.2. After all, g is just a single realvalued function, and the partial derivative g1 (x, y) is just what we get if we treat y as a constant and take the derivative of g with respect to x. Thus, g1 (x, y) = −F1 (x, y, z)/F3 (x, y, z) = −(3x 2 − yz)/(3z2 − xy) and, in particular, g1 (0, 0) = −F1 (0, 0, 1)/F3 (0, 0, 1) = 0. Likewise, g2 (0, 0) = 0. (b) As in part (a), F3 is C 1 everywhere and F3 (x, y, z) = ez − 2z = 0 for z = 0, so the equation F = 0 deﬁnes z as a C 1 function g(x, y) around (x0 , y0 , z0 ) = (1, 0, 0). We get g1 (x, y) = −F1 (x, y, z)/F3 (x, y, z) = 2x/(ez − 2z) and g2 (x, y, z) = −F2 /F3 = 2y/(ez − 2z), so g1 (1, 0) = 2 and g2 (1, 0) = 0. 2.7.3 The given equation system can be written as f(x, y, z, u, v, w) = 0, where f = (f1 , f2 , f3 ) is the function 3 × 3 → 3 given by f1 (x, y, z, u, v, w) = y 2 − z + u − v − w 3 + 1, f2 (x, y, z, u, v, w) = −2x + y − z2 + u + v 3 − w + 3, and f3 (x, y, z, u, v, w) = x 2 + z − u − v + w3 − 3. The Jacobian determinant of f with respect to u, v, w is ∂(f1 , f2 , f3 ) = ∂(u, v, w) 1 −1 1 3v 2 −1 −1 −3w2 −1 = 6w2 − 2 3w 2
This determinant is different from 0 at P , so according to Theorem 2.7.2 the equation system does deﬁne u, v, and w as C 1 functions of x, y, and z. The easiest way to ﬁnd the partial derivatives of these functions with respect to x is probably to take the derivatives with respect to x in each of the three given equations, remembering that u, v, w are functions of x, y, z. We get ⎫ ⎧ ux − vx − 3w 2 wx = 0 ⎪ ⎪ ux − vx − 3wx = 0 ⎬ ⎨ −2 + ux + 3v 2 vx − wx = 0 , so at P we get u − wx = 2 ⎪ ⎪ x ⎭ ⎩ 2 2x − ux − vx + 3w wx = 0 −ux − vx + 3wx = −2 The unique solution of this system is ux = 5/2, vx = 1, wx = 1/2. (In the ﬁrst printing of the book wx was incorrectly given as 5/2.) 2.7.4 The Jacobian determinant is is nonzero for all u (and v). fu gu fv eu cos v = u e sin v gv −eu sin v = e2u (cos2 v + sin2 v) = e2u , which eu cos v
(a) eu = 0, so the equations imply cos v = sin v = 0, but that is impossible because cos2 v + sin2 v = 1. (b) We must have cos v = sin v = √−u , and since cos2 v + sin2 v = 1 we get 2 cos2 v = 1 and e √ 1 therefore sin v = cos v = 1/2 = 2 2. (Remember that cos v = e−u cannot be negative.) The only 1 values of v that satisfy these equations are v = ( 4 + 2k)π , where k runs through all integers. Further, √ √ 1 eu = 1/ cos v = 2 gives u = 2 2. 2.7.6 The Jacobian is x1 . We ﬁnd that x1 = y1 + y2 , x2 = y2 /(y1 + y2 ) (provided y1 + y2 = 0). The transformation maps the given rectangle onto the set S in the y1 y2 plane given by the inequalities (i) 1 ≤ y1 + y2 ≤ 2, (ii) y2 1 2 ≤ ≤ 2 y1 + y2 3
The inequalities (i) show that y1 + y2 > 0, and if we multiply by 6(y1 + y2 ) in (ii) we get the equivalent inequalities (iii) 3(y1 + y2 ) ≤ 6y2 ≤ 4(y1 + y2 ) ⇐⇒ y1 ≤ y2 ≤ 2y1
© Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 2
MULTIVARIABLE CALCULUS
21
It follows that S is a quadrilateral with corners at (1/2, 1/2), (1, 1), (2/3, 4/3), and (1/3, 2/3). See ﬁgure M2.7.6.
y2 2 y2 = 2y1 y2 = y1 1 S y1 + y 2 = 2 1 2 y1 + y2 = 1
Figure M2.7.6
y1
∂ ∂ cos θ −r sin θ ∂r (r cos θ ) ∂θ (r cos θ ) = =r 2.7.8 (a) J = ∂ (r sin θ ) ∂ (r sin θ ) sin θ r cos θ ∂r ∂θ (b) J = 0 everywhere except at the origin in the rθplane, so T is locally onetoone in A. But it is not globally onetoone in A, since we have, for example, T (r, 0) = T (r, 2π). 2.7.10 (a) Taking differentials in the equation system 1 + (x + y)u − (2 + u)1+v = 0 2u − (1 + xy)eu(x−1) = 0 we get u(dx + dy) + (x + y) du − e(1+v) ln(2+u) ln(2 + u) dv + 1+v du = 0 2+u (1)
2 du − eu(x−1) (y dx + x dy) − (1 + xy)eu(x−1) (x − 1) du + u dx = 0 If we now let x = y = u = 1 and v = 0, we get 2 du + dx + dy − 3 ln 3 dv − du = 0 and 2 du − dx − dy − 2 dx = 0 Rearranging this system gives du − 3 ln 3 dv = −dx − dy 2 du with the solutions du = Hence, ux (1, 1) = 3 ∂u (1, 1) = ∂x 2 and vx (1, 1) = 5 ∂v (1, 1) = ∂x 6 ln 3 3 1 dx + dy 2 2 and dv = 1 3 ln 3 5 3 dx + dy 2 2 = 3 dx + dy
© Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
22
CHAPTER 2
MULTIVARIABLE CALCULUS
(Alternatively we could have used the formula for the derivatives of an implicit function, but it is still a good idea to substitute the values of x, y, u, and v after ﬁnding the derivatives.) (b) Deﬁne a function f by f (u) = u − aeu(b−1) . Then f (0) = −a and f (1) = 1 − aeb−1 . Since b ≤ 1, we have eb−1 ≤ e0 = 1. It follows that aeb−1 ≤ a ≤ 1, so f (1) ≥ 0. On the other hand, f (0) ≤ 0, so the intermediate value theorem ensures that f has at least one zero in [0, 1]. Further, f (u) = 1 − a(b − 1)eu(b−1) ≥ 1, because a(b − 1) ≤ 0. Therefore f is strictly increasing and cannot have more than one zero, so the solution of f (u) = 0 is unique. (c) For given values of x and y, let a = (1 + xy)/2 and b = x. The equation in part (b) is then equivalent to the second equation in system (1). Thus we get a uniquely determined u, and this u belongs to [0, 1]. (Note that, when x and y lie in [0, 1], then the values of a and b that we have chosen also lie in [0, 1].) The ﬁrst equation in (1) now determines v uniquely, as v = −1 + ln(1 + (x + y)u) ln(2 + u)
2.8
2.8.1 (a) The system has 7 variables, Y , C, I , G, T , r, and M, and 4 equations, so the counting rule says that the system has 7 − 4 = 3 degrees of freedom. (b) We can write the system as f1 (M, T , G, Y, C, I, r) = Y − C − I − G = 0 f2 (M, T , G, Y, C, I, r) = C − f (Y − T ) = 0 f3 (M, T , G, Y, C, I, r) = I − h(r) f4 (M, T , G, Y, C, I, r) = r − m(M) =0 =0 (∗)
Suppose that f , h, and m are C 1 functions and that the system has an equilibrium point (i.e. a solution of the equations), (M0 , T0 , G0 , Y0 , C0 , I0 , r0 ). The Jacobian determinant of f = (f1 , f2 , f3 , f4 ) with respect to (Y, C, I, r) is 1 ∂(f1 , f2 , f3 , f4 ) −f (Y − T ) = 0 ∂(Y, C, I, r) 0 −1 1 0 0 −1 0 1 0 0 0 = 1 − f (Y − T ) −h (r) 1
so by the implicit function theorem the system (∗) deﬁnes Y , C, I , and r as C 1 functions of M, T , and G around the equilibrium point if f (Y − T ) = 1. Note that the functions h and m have no inﬂuence on the Jacobian determinant. The reason is that once M, T , and G are given, the last equation in (∗) immediately determines r, and the next to last equation then determines I . The problem therefore reduces to the question whether the ﬁrst two equations can determine Y and T when the values of the other variables are given. The implicit function theorem tells us that the answer will certainly be yes if the Jacobian determinant ∂(f1 , f2 ) 1 = −f (Y − T ) ∂(Y, C) is nonzero, i.e. if f (Y − T ) = 1.
© Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
−1 = 1 − f (Y − T ) 1
u) = ∂x/∂y = −(∂G/∂y)/(∂G/∂x) = −f2 (x. the display (2. y). Moreover. so v = 2(ln u)2 − ln u. q2 ) = p(v1 )1/3 (v2 )1/2 − q1 v1 − q2 v2 = 1 6 −2 −3 p q1 q2 432 © Arne Strøm. so v= y − uy 1−u = y + uy 1+u 2. u) = g(ϕ(y. y.2 (a) The Jacobian determinants are ux ∂(u. then ψ1 (y. v = g(x.3 We need to assume that the functions f and g are C 1 in an open ball around (x0 . v is functionally dependent on u. y) = g(ϕ(y. y)/f1 (x. u) in an open ball around (y0 . u) is a function of u alone. y). v) ∂x = ∂v ∂(x. pv 3 1 with the unique solution ∗ v1 = 1/3 1/2 1 1/3 −1/2 pv v − q2 = 0 2 1 2 ∗ v2 = 1 6 −3 −3 p q1 q2 . y0 ). y) g2 (x. v2 ) is a maximum point. v) = ∂(x. y) vx uy vy uy vy 1 y −2y (y + x)2 = ex+y 4x + 4y − 1 −x y2 2x (x + y)2 ex+y 4x + 4y − 1 =0 (i) = = 2x 2x − =0 2 y(y + x) y(y + x)2 (ii) (b) (i) It is not hard to see that v = 2(x + y)2 − (x + y) and x + y = ln u. u) only depends on u. then F satisﬁes the requirements of deﬁnition (2. y) =0 so f2 (x. u). y0 ).1 3.8. y) = 0 deﬁnes x as a function x = ϕ(y. (If we let F (u. and v = ψ(y. and Peter Hammond 2008 .8. y)g2 (x.) Chapter 3 Static Optimization 3. Since f1 (x0 . where u0 = f (x0 .8. Thus.6) on Cobb–Douglas ∗ ∗ functions. u) = g1 ϕ1 + g2 = −f2 g1 /f1 + g2 = −f1 g2 /f1 + g2 = 0. (ii) We have x = uy. v) = v − ψ(y0 . Within this open ball we have ϕ1 (y. u) = u − f (x. Hence. y). ψ(y. u0 ). Knut Sydsæter. y) = f1 (x. y) ∂x ∂u f1 (x.1.5). cf. y) in A the Jacobian determinant is ∂u ∂(u. y) vx ux ∂(u. y) ∂y = ∂v g1 (x. q1 . y0 ) = 0. If we let ψ(y. the equation G(x. For all (x. u). 216 1 6 −2 −4 p q1 q2 144 The objective function pv1 v2 − q1 v1 − q2 v2 is concave. Atle Seierstad. u). y) ∂y f2 (x. and therefore (v1 .5. v) = ∂(x.3 (a) The ﬁrstorder conditions for maximum are 1 −2/3 1/2 v2 − q1 = 0. (b) The value function is ∗ ∗ ∗ ∗ π ∗ (p. y).CHAPTER 3 STATIC OPTIMIZATION 23 2. y)g1 (x.
f2 (x. .y)=(x ∗ . an ). s) = f (x ∗ . s) 9 ∂f ∗ (r. y. Since ∂π/∂vi = ∗ ∗ pai /(vi + 1) − qi . vn for given values of p.6 We want to maximize n π(v. = 6sy ∂r ∂s so 9 ∂f (x.3). q = (q1 . y).3).1. . = s3 ∂r ∂s 8 On the other hand. . s) = 2 r 2 and y ∗ (r.1. . the only stationary point is v∗ = v∗ (p. s) = r 3. an ) = i=1 (pai ln(vi + 1) − qi vi ) with respect to v1 . p q1 q2 = −v2 =− =− ∂q1 216 ∂q2 144 3.1. this is a maximum point. qn ). . . q. . Knut Sydsæter. . = 6sy ∗ = s 3 ∂r ∂s 8 (x. . a) ∂p v=v∗ (p.y ∗ ) (x. 3. Since f (x. qn . s) is concave with respect to (x. p.y ∗ ) in accordance with the envelope result (3. p. y ∗ . y. r. and Peter Hammond 2008 . . a) ∂ai = p ln v=v∗ (p. . r. q. y ∗ ) is a maximum point.24 CHAPTER 3 STATIC OPTIMIZATION and it follows that ∂π ∗ (p. ∂f (x. q. . a) = π(v1 . . . an ) = −1 pai ln − qi (pai ln(vi + 1) − qi vi ) = qi qi i=1 n i=1 = Easy calculations now yield ∂π ∗ (p. r. s) = 2rx.a) pai ∂π(v. s) = 2rx ∗ = r 3 . .5 The ﬁrstorder conditions for maximum are f1 (x. y. vn . . .q. a) = (v1 . vn ). q1 .a) pai ∂π(v. where vi∗ = pai /qi − 1. q. Moreover. s) = r 2 − 2x = 0. s) = r 4 + s 4 4 32 so ∂f ∗ (r. q1 .q. the point (x ∗ . a) =− + 1 = −vi∗ = qi ∂qi ∂qi ∂π ∗ (p. q. vn . p. r. s) ∂f (x. q1 . . Atle Seierstad. . . . . r. Since π is concave with respect to v1 . p. y. q2 ) 1 6 −2 −4 ∗ ∗ p q1 q2 = −v1 . . . The corresponding maximum value is n n pai pai ∗ ∗ ∗ π (p. a) = ∂p n i=1 (pai ln p + pai ln ai − pai ln qi − pai + qi ) n i=1 pai ai ln = qi ai ln(vi∗ + 1) = i=1 ∂π(v. qn . . q2 ) 1 5 −2 −3 ∗ ∗ = p q1 q2 = (v1 )1/3 (v2 )1/2 ∂p 72 ∂π ∗ (p.y)=(x ∗ . . y. . . y. . y. a1 . q. p. .1. . . q1 . 1 9 f ∗ (r. s) = 3s 2 − 16y = 0 1 3 with the solutions x ∗ (r.q. q. a) = p ln(vi∗ + 1) = qi ∂ai v=v∗ (p. r. . . q2 ) 1 6 −3 −3 ∂π ∗ (p. a) ∂π ∗ (p. . . s) ∂f (x. r. © Arne Strøm. . . s) = 16 s 2 . .a) in accordance with formula (3. q1 . a1 . and a = (a1 . . r. q.
The values of the leading principal minors at the stationary points are given in the following table: D1 (0. −1. 3/2). and D3 = H = 4. and Peter Hammond 2008 . (b) The stationary points are the solutions of the equation system f1 (x1 .3 (a) The ﬁrstorder conditions are f1 (x. so (0. x2 . 0. z) = 2x + 2xy = 0 f2 (x. −1. The Hessian matrix is (at every point) ⎛ ⎞ 2 −1 2 H = f (x1 . This solution is of course (x1 . −2. x3 ) = (0. 0) (0. 0) ( 3. 0. 2) is a local minimum point. y. z) = x 2 + 2yz = 0 f3 (x. x3 . x3 . x2 . x3 . x2 . x4 ) = 20 + 8x1 − 12x2 = 0 f3 (x1 . x4 ) = 6 − 2x4 © Arne Strøm. x3 ) = −x1 + 2x2 + x3 = 0 f3 (x1 . D2 = minimum point by Theorem 3. x3 ) = 2x1 − x2 + 2x3 = 0 f2 (x1 .1 that (0. 3/2) √ (− 3. x4 ) = 8x2 − 8x1 = 0 2 f2 (x1 . Atle Seierstad. (0. y.2 3. −1. 3. z) = y 2 + 2z − 4 = 0 √ This system gives ﬁve stationary points: (0.2. (± 3. D2 = 4(1 + y)z − 4x 2 . y. and all the other stationary points are saddle points. x2 . x3 . x2 .2. 0. x2 . 2. 0). z) = ⎝ 2x 2z 2y ⎠ 0 2y 2 with leading principal minors D1 = 2 + 2y. 0. y. 2). The Hessian matrix is ⎛ ⎞ 2 + 2y 2x 0 f (x. x4 ) = 48 − 24x3 = 0 f4 (x1 . 0). 2) (0.1(a). Knut Sydsæter. ±2.2. so by Cramer’s rule it has a unique solution. x3 ) = 2x1 + x2 + 6x3 = 0 The determinant of this linear equation system is 4. 0) is a local 2 It follows from Theorem 3. and D3 = 8(1 + y)(z − y 2 ) − 8x 2 .CHAPTER 3 STATIC OPTIMIZATION 25 3. 3/2) √ 2 6 −2 0 0 D2 8 0 0 −12 −12 D3 16 −96 32 −24 −24 2 −1 −1 = 3. x2 .2. x3 ) = ⎝ −1 2 1⎠ 2 1 6 with leading principal minors D1 = 2.1 The ﬁrstorder conditions are f1 (x1 . 0. x2 . x2 .
009. 7 7. and the ﬁrstorder conditions are: (i) 1 − 2λ1 x − λ2 = 0. −1. (5/3. so x 2 = 216/378 = 4/7. 3). and by the extreme value theorem the objective function does attain a maximum over the admissible set. The Hessian matrix is −8 ⎜ 8 f (x1 . The last two equations determine x3 and x4 . √ 1 (b) Equation (3. D4 = 48D2 At (−1.10) shows that f ∗ ≈ λ1 · (−1) + λ2 · 0. 22 7 7 7 7 √ 7 at x1 . so this point is a local maximum point. 3. 2. 7 © Arne Strøm. 5/3. Hence there are two 7 points that satisfy the ﬁrstorder conditions: x1 = 2 7 √ 32 √ √ 7. so y = 16x and z = −11x. y. D3 = −24D2 . (The point x2 is the minimum point and fmin = − 108 7. −1.2 (a) The admissible set is the intersection of the sphere x 2 +y 2 +z2 = 216 and the plane x+2y+3z = 0.3. The Lagrangian is L(x. The constraint x 2 + y 2 + z2 = 216 √ then yields (1 + 256 + 121)x 2 = 216. y. ⎛ 8 −24x2 0 0 0 0 −24 0 ⎞ 0 0⎟ ⎟ 0⎠ −2 D2 = 192x2 − 64 = 64(3x2 − 1). Conditions (i) 3 and (iii) yield 1 − 2λ1 x = 1 − 2 λ1 z = 0. 2. 3) and (−1.6 ≈ −0. λ2 = 6 . (iii) 1 − 2λ1 z − 3λ2 = 0 From (i) and (ii) we get λ2 = 1 − 2λ1 x = 2 − λ1 y. − 32 7.3. x4 ) = ⎜ ⎝ 0 0 and the leading principal minors of the Hessian are D1 = −8. There are two stationary points. so z = 5x − y. f (x. so this point is a saddle point. 7 √ √ √ x2 = − 2 7.3 3.3. Therefore part (b) of Theorem 3. It follows that y − 2x = 3x − z. z) = x + 4y + z − λ1 (x 2 + y 2 + z2 − 216) − λ2 (x + 2y + 3z). and then the second equation gives 12x1 − 8x1 − 20 = 0 with the two solutions x1 = 5/3 and x1 = −1. √ 1 The multipliers are then λ1 = 1/(y − 2x) = 1/(14x) = ± 28 7 and λ2 = 1 − 2λ1 x = 6 . which implies λ1 ( 2 z − 2x) = − 2 . D2 > 0. Multiply by − 2 to get 3 3 3 3 λ1 (3x − z) = 1. 3).1 is sufﬁcient to show that x1 is a maximum point and x2 is a minimum point in this problem. x2 . z) = x + 4y + z. x3 . 2. (ii) 4 − 2λ1 y − 2λ2 = 0. 5/3. so we did not need to use the extreme value theorem. 3) we get D2 < 0. Inserting this expression for z in the constraint x + 2y + 3z = 0 yields 16x − y = 0. which implies λ1 (y − 2x) = 1. Atle Seierstad. This set (a circle) is closed and bounded (and nonempty!). 2. we get D1 < 0.26 CHAPTER 3 STATIC OPTIMIZATION 2 The ﬁrst equation gives x2 = x1 .1 = − 28 7 + 0. which passes through the center of the sphere. and x = ± 2 7. attains it maximum value fmax = √ √ 1 with λ1 = 28 7. D3 < 0. 7 The objective function. (5/3.) 7 7 108 7 Comment: It is clear that the Lagrangian is concave if λ1 > 0 and convex if λ1 < 0. and D4 > 0. Knut Sydsæter. The other stationary point. and Peter Hammond 2008 . − 22 7 .
z) = ± 10 10 . attains its maximum value 1 at the points (∗). then λ2 = 0 and (iii) implies z = 0. (iii) 2z − 8λ1 z − 2λ2 = 0 From (i) and (ii) we get (iv) λ2 = 2(1 − λ1 )x = 2 (1 − λ1 )y. x 2 + y 2 + z2 . z) = ± 110 110.3.) (b) f ∗ ≈ 1 · 0. © Arne Strøm.CHAPTER 3 STATIC OPTIMIZATION 27 1 1 3. and we get the two candidates √ √ 1 3 (x.05 = 0. z) = x 2 + y 2 + z2 − λ1 (x 2 + y 2 + 4z2 − 1) − λ2 (x + 3y + 2z). 1 1 ∗ ∗ (b) U ∗ (m) = 2 ln(1+x1 (m))+ 4 ln(1+x2 (m)) = 3 so dU ∗ /dm = 4 (m + 5)−1 = λ. namely an ellipse. y. x2 = x2 (m) = 1 (m − 4). in fact). y) = 2 ln(1 + x1 ) + 4 ln(1 + x2 ) − λ(2x1 + 3x2 − m). 11 The objective function.1(b) shows that the points in (∗) are maximum points. even if we do not check the rank condition in Theorem 3.3. 3. the ﬁrstorder conditions are 1 1 (i) − 2λ = 0. (ii) 2y − 2λ1 y − 3λ2 = 0. y. and the second constraint gives z = −5x.3.05 + 0 · 0. 3 (A) If λ1 = 1. ∓ 10 10 .05. which obviously has a maximum for z = 0. 1 2 1 ln( 1 (m+5))+ 4 ln( 1 (m+5)) = 3 9 3 4 ln(m+5)−ln 3.1 has rank 2 at all admissible points. Firstorder conditions: (i) 2x − 2λ1 x − λ2 = 0. Atle Seierstad. so the usual ﬁrstorder conditions are necessary. Knut Sydsæter. 3 9 3 and then λ = 4 (m + 5)−1 . We then just have to solve the equations x 2 + y 2 = 1 and x + 3y = 0 for x and y. (There is a much simpler way to ﬁnd the maximum points in this problem: Because of the ﬁrst constraint. It is worth noting that with λ1 = 1 the Lagrangian is concave (linear. y. It is not very hard to show that the matrix g (x) in Theorem 3. The constraints reduce to x 2 +y 2 = 1 and x +3y = 0. The ﬁrst constraint then yields x 2 + 9x 2 + 100x 2 = 1 which leads to the two candidates √ √ √ 1 3 5 (x. and Peter Hammond 2008 . the objective function x 2 + y 2 + z2 equals 1 − 3z2 .6 (a) The two constraints determine an ellipsoid centred at the origin and a plane through the origin. while the points (∗∗) give the minimum value 7/22. ± 110 110 . so Theorem 3. ∓ 110 110 (∗∗) 7 22 In this case the multipliers λ1 and λ2 can be determined from equations (i) and (iii). Lagrangian: L(x. This curve is closed and bounded (and nonempty). The admissible set is the curve of intersection of these two surfaces.3. and we get λ1 = and λ2 = 15 x.1(a) (but without that rank condition we cannot be sure that the ﬁrstorder conditions are necessary. 0 (∗) (B) If λ1 = 1. so the extreme value theorem guarantees the existence of both maximum and minimum points. then (iv) implies y = 3x. respectively.3. (ii) − 3λ = 0 2(1 + x1 ) 4(1 + x2 ) Equations (i) and (ii) yield λ= 1 1 = 4(1 + x1 ) 12(1 + x2 ) ⇒ 1 + x1 = 3(1 + x2 ) ⇐⇒ x1 − 3x2 = 2 ∗ ∗ Together with the constraint 2x1 + 3x2 = m this gives x1 = x1 (m) = 1 (m + 2).4 (a) With the Lagrangian L(x. so there might be other maximum points beside the two that we have found).
. 0. z2 ) is a local minimum point. r) ∂f (x. . then (i) and (ii) yield 1 − λ2 = 0 and 1 + λ2 = 0. . Knut Sydsæter. r) ∂L(x. so we must have λ1 = 0. r) = f (x. . y2 . y1 . r) subject to (∗) m The Lagrangian for problem (3. . − 2 . and Peter Hammond 2008 . Atle Seierstad. z2 ) = (− 1 . Hence (x1 . and get the two solutions (x1 . r) = 0. namely the problem in (3. so please correct gj (x. . so all we need check is 0 0 B3 (x. r) ≤ 0 to gj (x. r) = 0 for j = 1. . r) − λm+i = 0. j = 1. . Now consider two problems.3. and so ∂f ∗ (r) ∂f (x. This is clearly impossible. r) − j =1 λj gj (x. k max f (x. r) = ∂ri ∂ri 3.4 3. . (iii) 1 − 2λ1 z + λ2 = 0 Equations (ii) and (iii) give 2λ1 y = 2λ1 z.10 There was a misprint in this problem in the ﬁrst printing of the book: The constraints should all be equality constraints. r) = − ∂ri ∂ri m j =1 ∂gj (x. We solve this equation together with the two constraints. (x2 . (ii) 1 − 2λ1 y + λ2 = 0.3. y. z1 ) = (1. y1 . y. © Arne Strøm. z1 ) is a local maximum point and (x2 . ri = bm+i . y1 . . Firstorder conditions: (i) 1 − 2λ1 x − λ2 = 0.13) and the problem gj (x. k Equation (3. y2 . . r) = λm+i = − ∂ri ∂ri m j =1 ∂gj (x. The ﬁrstorder conditions i=1 for maximum in problem (∗) imply that ∂ L(x. y2 . . . λ2 = −1 λ2 = 1 3 In the secondderivative test (Theorem 3. and therefore y = z. m. r) and the Lagrangian for m problem (∗) is L(x.9) implies that ∂f ∗ (r)/∂ri = λm+i .3. − 2 ).1) we now have n = 3 and m = 2. z2 ) = 16. r) − j =1 λj gj (x. m i = 1. . λ1 = −1. z) = x +y +z −λ1 (x 2 +y 2 +z2 −1)−λ2 (x −y −z −1). 0). r) − k λm+i (ri − bm+i ).3 Lagrangian: L(x. ∂ri i = 1. .4. z) = 2x 2y 2z 0 0 1 −1 −1 2x 1 −2λ1 0 0 2y −1 0 −2λ1 0 2z −1 0 0 −2λ1 A little tedious computation yields B3 (x1 .13) is L(x. z1 ) = −16 and B3 (x2 . 3 3 3 λ1 = 1. r) = f (x. If λ1 = 0.28 CHAPTER 3 STATIC OPTIMIZATION 3.4. .3.
Moreover. We consider the four possible combinations of λ1 = 0. so no candidates. y) must also satisfy the constraints x + 2y ≤ c x+ y≤2 (5) (6) (b) Let c = 5/2. and the necessary Kuhn–Tucker conditions for (x. by eliminating λ1 from (1) and (2) we get x + 1 = 2y + 2. (IV) λ1 > 0. y = 1/2. and so x = 4 − c and y = c − 2. Atle Seierstad. But then x + y = 17/8 > 2. x + y = 2. 1/2) solves the problem. contradicting (6).) For such c. and λ2 = 0 if x + y < 2 (1) (2) (3) (4) Of course. (II) λ1 > 0.5. and Peter Hammond 2008 . then formula (3. x+2y = 5/2. and λ1 = 0 if x + 2y < c L1 (x. y) = λ2 ≥ 0. contradicting (5). Then x + y = 2 and x + 2y = 5/2. λ2 > 0. A direct argument can run as follows: For all values of c such that λ1 and λ2 are positive. We ﬁnd that λ1 = 4/15 and λ2 = 2/15.3 We reformulate the problem as a standard maximization problem: maximize −4 ln(x 2 + 2) − y 2 subject to −x 2 − y ≤ −2 −x ≤ −1 © Arne Strøm. and so x = y = 1. λ2 = 0. the derivative of the value function is V (c) = ∂V dy 1 1 ∂V dx + =− + = λ1 ∂x dc ∂y dc 1+x 1+y 3. But then x + 2y = 3 > 5/2. This contradicts (1).5 3. (III) λ1 = 0. y) = (3/2. and λ2 > 0. so x = 3/2. λ2 = 0. Moreover. so this is a candidate. (I) λ1 = λ2 = 0.5.CHAPTER 3 STATIC OPTIMIZATION 29 3. y) = − 2λ1 − λ2 = 0 y+1 λ1 ≥ 0. they are both positive if and only if 7/3 < c < 3. so (x. y) = ln(x + 1) + ln(y + 1) − λ1 (x + 2y − c) − λ2 (x + y − 2). λ2 > 0. (x.2 (a) The Lagrangian is L(x. No candidates. Knut Sydsæter. y) to solve the problem are: 1 − λ 1 − λ2 = 0 x+1 1 L2 (x. From (3) and (5). (More precisely.6) yields V (5/2) = λ1 = 4/15. y = 3/8. The last two equations have the solution x = 7/4. The Lagrangian is obviously concave in x and y.5. and the only one. x and y must satisfy the constraints with equality. (c) If we assume that V (c) is a differentiable function of c. No candidates. λ1 > 0. Then equations (1) and (2) yield λ1 = 1 1 1 1 − = − y+1 x+1 c−1 5−c and λ2 = 1 2 1 2 − = − x+1 y+1 5−c c−1 It is clear that these expressions remain positive for c in an open interval around c = 5/2. From (4) and (6). eliminating λ2 from (1) and (2) we get x = y.
Inserted into x√+ y = 2. λ2 ) = ( 4 2. y) = 4 ln(x 2 +2)+ value √ √ √ √ √ y 2 under the given constraints is f ( 4 2. λ2 = 0. 1}. y) to be a maximum point: L1 (x. and λ1 = 4 − 2 2. (i) (ii) (iii) (iv) (D) λ1 > 0. x +y ≥2 2 (i) (ii) and λ1 = 0 if x + y > 2 2 (iii) (iv) (v) (vi) and λ2 = 0 if x > 1 x≥1 We try the four possible combinations of zero or positive multipliers: (A) λ1 = 0. (B) λ1 = 0. x = 1. λ2 > 0. which contradicts x ≥ 1. and λ1 = 0 if x < 1 λ2 ≥ 0. and λ2 > 0. and b ≤ 2. and λ2 = 0 if y < 2 Of course. (D) λ1 > 0. Atle Seierstad. this gives x 4 =√ or x = 4 2. λ2 ≥ 0. The solution can be summarized as x ∗ = min{a.5. Because of the constraints this is possible only if a ≤ 1 and b ≤ 2. a = 1 + 2 λ1 > 1. so the necessary Kuhn–Tucker conditions together with the constraints are: L1 = − x2 8x + 2λ1 x + λ2 = 0 +2 L2 = −2y + λ1 = 0 λ1 ≥ 0. y) must also satisfy the constraints (v) x ≤ 1 and (vi) y ≤ 2. This gives x = 1. © Arne Strøm. y) = −(x −a)2 −(y −b)2 −λ1 (x −1)−λ2 (y −2) we get the Kuhn–Tucker necessary conditions for (x. y = 2. 1 (C) λ1 > 0. (C) λ1 > 0. y = b. λ2 = 0. So x = y = 1. It follows + 2. λ2 = 0. λ2 = 0.30 CHAPTER 3 STATIC OPTIMIZATION The Lagrangian is L(x. λ1 > 0. y) = −2(y − b) − λ2 = 0 λ1 ≥ 0. λ2 > 0. Then from (i). and the minimum √ of f (x. Then (iii)–(vi) imply x 2 + y = 2 and x = 1. 3. From (iii) and (v) we get x 2 + y = 2. Equation (i) gives λ1 = 4/(x 2 + 2). 2}. Thus the only possible solution is the one given in (C). λ2 = 0. a maximum point (x. Which case applies depends on the values of a and b. 2 − 2. 4 − 2 2. We get x = a and y = 2. Contradiction. This contradicts (v). From (i) we see that x = 0. We try the four possible combinations of λ1 = 0. 0) is a candidate. Knut Sydsæter. (A) λ1 = 0. y) = −2(x − a) − λ1 = 0 L2 (x. So (x. λ1 . Constraint (v) implies a ≤ 1. (B) λ1 = 0. 2− 2 ) = 4 ln( 2+2)+(2− 2 )2 = 4 ln( 2+2)+6−4 2. both constraints must be satisﬁed with equality: 1 1 x = 1. and since the objective function is concave. λ2 > 0. and equation (ii) yields 1 b = y + 2 λ2 > y = 2. But then λ2 = 8/3 − 4 < 0. and then √ 2 (ii) gives y = λ1 /2 = 2/(x 2 √ 2). y ∗ = min{b. these points are maximum points. and (ii) gives y = 0.4 With the Lagrangian L(x. y. We also get a = 1 + 2 λ1 > 1 and b = 2 + 2 λ2 > 2. With both multipliers positive. √ √ √ that y = 2/( 2 + 2) = 2 − 2. From (iv) and (vi). λ2 + 2λ1 = 8/3 and (ii) gives λ1 = 2. We observe that in each of the four cases (A)–(D) there is exactly one point that satisﬁes the Kuhn– Tucker conditions. and Peter Hammond 2008 . y) = −4 ln(x 2 + 2) − y 2 − λ1 (−x 2 − y + 2) − λ2 (−x + 1). λ2 > 0. Equations (i) and (ii) give x = a and y = b.
Atle Seierstad. It is readily seen that the point (x ∗ . λ2 > 0. (x2 . Since h(x. and Peter Hammond 2008 h(x1 . b) itself is admissible and is also the optimal point. x) = x 5 − x 3 . b) as possible. 0) is a candidate for optimum.e. y4 ) = (1. y) = x 5 − y 3 − λ1 (x − 1) − λ2 (x − y) the Kuhn–Tucker conditions and the constraints are L1 (x. 5 5 (C) λ1 > 0. y2 ) = 125 15 (For the last evaluation we used the fact that h is an odd function.6 (a) With the Lagrangian L(x. y) = x 5 − y 3 at each of the four maximum candidates we have found shows that √ 5 3 5 3 2 3 3 6 h(x2 . and thus 5y 4 = √ 2 . and the maximization problem is equivalent to ﬁnding an admissible point as close to (a. In particular. The complementary slackness conditions show that in this case both constraints must be satisﬁed with equality. (D) λ1 > 0. y). y1 ) = (0. λ2 = 0. 1 15 ) ± 5 5 5 √ and (x3 . −y) = −h(x. and therefore 5y 2 = 3. y1 ) = h(x4 . But this violates constraint (vi). the complementary slackness condition (iv) tells us that we cannot have x < y. 1) Evaluating the objective function h(x. while constraint (vi) says x ≤ y. ≤1 ≤y (i) (ii) (iii) (iv) (v) (vi) and λ1 = 0 if x < 1 and λ2 = 0 if x < y Consider the four possible combinations of zero or positive multipliers: (A) λ1 = 0. Now (ii) gives y = 0. so it is not obvious that there is a maximum. λ2 > 0. so f (x) = h(x. y) = 5x 4 − λ1 − λ2 −3y 2 + λ2 L2 (x. Therefore x = y. But part (b) of this problem will show that a maximum point does exist. y) is strictly decreasing with respect to y. (b) Let h(x. Note that although the feasible set is closed. y) = λ1 λ2 x x =0 =0 ≥ 0. y). in case (A) the point (a. y4 ) = 0 . y ∗ ) given above is the optimal point. λ2 = 0. so √ √ √ √ x = y = ± 3/5 =√ 15/25 = ± 1 15. y3 ) = h(−x2 . To ﬁnd the maximum of f (x) for x ≤ 1 we take a look at f (x) = 5x 2 − 3x 2 = 5x 2 (x 2 − 3 ) = 5x 2 (x + x2 )(x − x2 ) 5 © Arne Strøm. while (iii) and (v) give x = 1. −y2 ) = −h(x2 . so we get no new candidates for optimum here. (B) λ1 = 0. − 1 15 ). then (x3 .) Hence. (x4 . if there is a maximum point for h in the feasible set. Since λ2 = 0 we have y = 0. 3. i.5. Knut Sydsæter. it is not bounded.1. We get two new candidates. y3 ) is that maximum point. Since λ2 > 0. y2 ) = ( 1 15.CHAPTER 3 STATIC OPTIMIZATION 31 The admissible set in this problem is the same as in Example 3. so we get one new candidate point. h(−x. y2 ) = x2 − y2 = x2 − x2 = (x2 − 1)x2 = − 2 x2 = − 125 15 5 √ 6 h(x3 .5. it is clear that we get the maximum when y = x. Equations (i) and (ii) then give 3y 5x 4 = 0 + λ2 = 3y 2 . ≥ 0. y3 ) = (− 1 15. Thus (x1 . Equations (i) and (ii) give x = 0 and y = 0. y) = x 5 − y 3 and deﬁne f (x) = maxy≥x h(x.
1/2) shows that (1/2. (The extreme value theorem guarantees that there really is a maximum point. we must have x+y = 1. (B) λ1 = 0. and strictly increasing again in [x2 . −1) in case (A) above. If x = y = −1. −1) and (1/2. Atle Seierstad. 0) or x ∈ (0.32 CHAPTER 3 STATIC OPTIMIZATION where x2 = √ √ 3/5 = 1 15 (as in part (a)). ∞) f (x) < 0 if x ∈ (−x2 . this implies y = x for any point that satisﬁes the ﬁrstorder conditions. y3 ) that we found in part (a) really is a maximum point in that problem? The reason is that for every point (x. Since x = y. and λ1 = 0 if x 2 + y 2 < 2 and λ2 = 0 if x + y < 1 (i) (ii) (iii) (iv) It is usually a good idea to exploit similarities and symmetries in the ﬁrstorder conditions. So no candidates in this case either. λ2 = 0. Since λ2 > 0. The point (1.2 Lagrangian: L(x. −x2 ) or x ∈ (x2 . Firstorder conditions: L1 (x.e. λ2 > 0. Now x 2 + y 2 = 2 because λ1 > 0. and Peter Hammond 2008 . we get (x. 1/2) must be the maximum point. −1). Knut Sydsæter. and since x = y we get x = y = ±1. In this case both constraints must be active. Comparing the values of the objective function xy + x + y at the two points (−1. y) = xy + x + y − λ1 (x 2 + y 2 − 2) − λ2 (x + y − 1).) © Arne Strøm. we can eliminate λ2 from (i) and (ii) to get y − 2λ1 x = x − 2λ1 y ⇐⇒ (1 + 2λ1 )(y − x) = 0 Since 1 + 2λ1 ≥ 1. λ2 = 0. y) = x + 1 − 2λ1 y − λ2 = 0 λ1 ≥ 0. (C) λ1 > 0. It is clear that 5 f (x) > 0 if x ∈ (−∞. x2 ) Therefore f is strictly increasing in the interval (−∞. 1/2). Equations (i) and (ii) give (x. y) = (1/2. then equation (i) yields 2λ1 = λ2 = 0. In the present case. (We did get the point (−1. λ2 ≥ 0.6 3. contradicting the assumption λ1 > 0. y) with x ≤ 1 and x ≤ y we have h(x. the maximum point is −x2 (= x3 ). strictly decreasing in [−x2 . (Draw a graph!) Since f (−x2 ) = 125 and f (1) = 0. 1) violates the constraint x + y ≤ 1.6. the ﬁrst constraint yields x = y = 1/2. x 2 + y 2 = 1 and x + y = 1. y3 ) 3. and since x = y. x3 ) = h(x3 . So why does this show that the point (x3 . Thus there are no candidate points in this case. y) = y + 1 − 2λ1 x − λ2 = 0 L2 (x. but then x 2 + y 2 = 2. λ2 > 0.) (D) λ1 > 0. −x2 ]. x2 ]. i. y) = (−1. Now consider the various combinations of zero or positive values of λ1 and λ2 : (A) λ1 = 0. It follows that f (x) will reach its highest value when x = −x2 or when 6 x = 1. ∞). y) ≤ f (x) ≤ f (x3 ) = h(x3 .
b)/∂b. Therefore we must have λ1 > 0 and 2x 2 + 4y 2 = s 2 . λ1 > 0. then (i) and (ii) would yield (2 + 4λ2 )x = 0 ⇒ x = 0. and get the Kuhn–Tucker conditions L1 = 2x − 4λ1 x + 4λ2 x = 0 L2 = 2y − 8λ1 y + 8λ2 y = 0 λ1 ≥ 0. Strictly speaking. we must have −b > −(a−b)/2. and (i) yields λ2 = e−b −e−(a−b)/2 .CHAPTER 3 STATIC OPTIMIZATION 33 3. Then (i) and (ii) imply λ1 = e−x and x = y.7. b)/∂a and ∂f ∗ (a. y) = 100 − e−x − e−y − e−z − λ1 (x + y + z − a) − λ2 (x − b) and the Kuhn–Tucker conditions are e−x − λ1 − λ2 = 0 e −y (i) (ii) (iii) (iv) (v) − λ1 =0 =0 e−z − λ1 λ1 ≥ 0. and therefore a/3 ≤ b. so x = a/3. and Peter Hammond 2008 . i. a > 3b. so x + 2y = a. Knut Sydsæter. Moreover. we cannot have λ2 > 0. λ2 = 0. Equations (i) and (ii) reduce to (i ) (2 − 4λ1 )x = 0 and (ii ) (2 − 8λ1 )y = 0 © Arne Strøm. Atle Seierstad. s) = x 2 + y 2 − λ1 (2x 2 + 4y 2 − s 2 ) + λ2 (2x 2 + 4y 2 − r 2 ). and so y = (a − x)/2 = (a − b)/2. The Lagrangian is concave. 3. (B) Suppose λ2 > 0. y) = x 2 + y 2 subject to 2x 2 + 4y 2 ≤ s 2 −2x − 4y ≤ −r 2 2 2 (∗) (∗∗) We use the Lagrangian L(x.e.e. Thus. From (iv) we get x + y + z = a. so we know that the points we have found really are optimal. Then λ1 = e−y = e−(a−b)/2 . if a = 3b. (2 + 8λ2 )y = 0 ⇒ y=0 and and λ1 = 0 if 2x + 4y < s 2 2 2 (i) (ii) (iii) (iv) λ2 = 0 if 2x 2 + 4y 2 > r 2 which contradicts the constraint 2x 2 + 4y 2 ≥ r 2 . λ2 ≥ 0.3 (a) Consider the maximization problem and write it in standard form as maximize f (x. you must consider the onesided derivatives and show that the left and right derivatives are the same. but things work out all right in that case too. (A) Suppose λ2 = 0. and λ1 = 0 if x + y + z < a and λ2 = 0 if x < b Equations (ii) and (iii) imply that λ1 > 0 and y = z. Hence x + y + z = 3x = a. (b) See the answer in the book for the evaluation of ∂f ∗ (a. i. Condition (v) now implies x = b. which contradicts 2x 2 + 4y 2 = s 2 . λ2 ≥ 0.7 3.7. If λ1 = 0.1 (a) The Lagrangian is L(x. y. the Kuhn–Tucker conditions have a unique solution in each of the two cases a ≤ 3b and a > 3b. r. a ≤ 3b. for that would imply 2x 2 + 4y 2 = r 2 < s 2 . Since λ2 > 0. Hence.
34
CHAPTER 3
STATIC OPTIMIZATION
If x = 0, then y = 0 (because 2x 2 + 4y 2 = s 2 > 0), and (ii ) implies λ1 = 1/4. If x = 0, then (i ) implies λ1 = 1/2. We are left with the two possibilities √ 1 1 (A) λ1 = 1/2, y = 0, x = ± 2 2 s (B) λ1 = 1/4, x = 0, y = ± 2 s √ 1 1 Case (A) gives the optimum points, (x ∗ , y ∗ ) = (± 2 2 s, 0), and f ∗ (r, s) = f (x ∗ , y ∗ ) = 2 s 2 . To verify the envelope result (3.7.5), note that ∂L(x, y, r, s)/∂r = −2λ2 r, ∂L(x, y, r, s)/∂s = 2λ1 s
If we insert the optimal values of x ∗ and y ∗ and the corresponding values λ1 = 1/2 and λ2 = 0 of the multipliers, we get ∂L(x ∗ , y ∗ , r, s)/∂r = 0 = ∂f ∗ (r, s)/∂r, in accordance with (3.7.5). (b) The minimization problem is equivalent to maximizing g(x, y) = −f (x, y) subject to the constraints (∗) and (∗∗) in part (a). We get a new Lagrangian L(x, y, r, s) = −x 2 − y 2 − λ1 (2x 2 + 4y 2 − s 2 ) + λ2 (2x 2 + 4y 2 − r 2 ), and the ﬁrstorder conditions are as in (a), except that (i) and (ii) are replaced by L1 = −2x − 4λ1 x + 4λ2 x = 0 L2 = −2y − 8λ1 y + 8λ2 y = 0 (i ) (ii ) ∂L(x ∗ , y ∗ , r, s)/∂s = s = ∂f ∗ (r, s)/∂s
The solution proceeds along the same lines as in (a), but this time λ2 = 0 is impossible, so we get λ2 > 0 1 and λ1 = 0. The optimum points are (x ∗ , y ∗ ) = (0, ± 2 r), with λ2 = 1/4, and g ∗ (r, s) = g(x ∗ , y ∗ ) = 1 1 −(x ∗ )2 − (y ∗ )2 = − 4 r 2 , and the minimum value of f is f ∗ (r, s) = −g ∗ (r, s) = 4 r 2 . The equations in (3.7.5) now become
1 Lr (x ∗ , y ∗ , r, s) = −2λ2 r = − 2 r =
∂g ∗ ∂f ∗ =− , ∂r ∂r
Ls (x ∗ , y ∗ , r, s) = 2λ1 s = 0 =
∂g ∗ ∂f ∗ =− ∂s ∂s
(c) The admissible set is the area between two ellipses, and the problems in (a) and (b) are equivalent to ﬁnding the largest and the smallest distance from the origin to a point in this admissible set. 3.7.4 Let r and s be points in the domain of f ∗ , let λ ∈ [0, 1], and put t = λr+(1−λ)s. We want to prove that f ∗ (t) ≥ λf ∗ (r) + (1 − λ)f ∗ (s). There are points x and y such that f ∗ (r) = f (x, r) and f ∗ (s) = f (y, s). Let w = λx + (1 − λ)y. Since g is convex, we have g(w, t) λg(x, r) + (1 − λ)g(y, s) 0, so (w, t) is admissible. And since f is concave, we have f ∗ (t) ≥ f (w, t) = f (λ(x, r) + (1 − λ)(y, s)) ≥ λf (x, r) + (1 − λ)f (y, s) = λf ∗ (r) + (1 − λ)f ∗ (s)
3.8
3.8.2 (a) Lagrangian: L(x, y) = xy − λ(x + 2y − 2). Kuhn–Tucker conditions: ∂L/∂x = y − λ ≤ 0 ∂L/∂y = x − 2λ ≤ 0 λ≥0 (= 0 if x > 0) (= 0 if y > 0) (= 0 if x + 2y < 2) (i) (ii) (iii)
© Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
CHAPTER 3
STATIC OPTIMIZATION
35
There are admissible points where xy > 0, so we cannot have x = 0 or y = 0 at the optimum point or points. It follows that (i) and (ii) must be satisﬁed with equality, and λ must be positive. Hence, x = 2λ = 2y and x + 2y = 2, so x = 1, y = 1/2, λ = 1/2. (The extreme value theorem guarantees that there is a maximum, since the admissible set is closed and bounded. It is the closed line segment between (2, 0) and (0, 1).) (b) As in part (a), there are admissible points where x α y β > 0, so we may just as well accept (2, 0) and (0, 1) as admissible points and replace the constraints x > 0 and y > 0 by x ≥ 0 and y ≥ 0. Then the extreme value theorem guarantees that there is a maximum point in this case too, and it is clear that both x and y must be positive at the optimum. With the Lagrangian L(x, y) = x α y β − λ(x + 2y − 2) we get the Kuhn–Tucker conditions ∂L/∂x = αx α−1 y β − λ ≤ 0 ∂L/∂y = βx y
α β−1
(= 0 if x > 0) (= 0 if y > 0) (= 0 if x + 2y < 2)
(i) (ii) (iii)
− 2λ ≤ 0 λ≥0
It is clear that (i), (ii), and (iii) must all be satisﬁed with equality, and that λ > 0. From (i) and (ii) we get βx α y β−1 = 2λ = 2αx α−1 y β , so βx = 2αy. This equation, combined with (iii) yields the solution: x = 2α/(α + β), y = β/(α + β). (If we had not extended the admissible set to include the end points, then we could not have used the extreme value theorem to guarantee a maximum, but with the conditions on α and β the Lagrangian is concave, so we could still be certain that the point we have found is a maximum point. But the argument above, with a closed and bounded admissible set, works for all positive values of α and β, even if L is not concave.) 3.8.3 With the Lagrangian L(x, y, c) = cx + y − λ(x 2 + 3y 2 − 2) we get the Kuhn–Tucker conditions: L1 = c − 2λx ≤ 0 L2 = 1 − 6λy ≤ 0 λ≥0 (= 0 if x > 0) (= 0 if y > 0) (= 0 if x + 3y < 2)
2 2
(i) (ii) (iii)
If λ = 0, then (ii) implies 1 ≤ 0, but that is impossible. Hence, λ > 0 and x 2 + 3y 2 = 2. Further, (ii) implies 6λy ≥ 1, so y > 0. Therefore (ii) is an equality and y = 1/6λ. √ √ 2 (A) √ If x = 0, then (i) implies c ≤ 2λx = 0. Further, 3y = 2, so y = 2/3 = 6/3, and λ = 1/6y = 6/12. (B) If x > 0, then (i) is satisﬁed with equality and c = 2λx > 0, and x = c/2λ. The equation x 2 + 3y 2 = 2 then leads to λ = 6(3c2 + 1)/12, x = 6c/ 6(3c2 + 1), and y = 2/ 6(3c2 + 1). Since the admissible set is closed and bounded and the objective function f (x, y) = cx +y is continuous, the extreme value theorem guarantees that there is a maximum point for every value of c. The cases (A) and (B) studied above show that the Kuhn–Tucker conditions have exactly one solution in each case, so the solutions we found above are the optimal ones. √ If c ≤ 0, then we are in case (A) and f ∗ (c) = cx ∗ + y ∗ = 6/3. √ The value function f ∗ (c) is obviously continuous for c = 0, and because limc→0+ f ∗ (c) = 6/3 = f ∗ (0) = limc→0− f ∗ (c), it is continuous at c = 0 too. The value function is differentiable at all c = 0,
© Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
If c > 0, then we are in case (B) and f ∗ (c) = cx ∗ + y ∗ =
6(3c2 + 1)/3.
36
CHAPTER 3
STATIC OPTIMIZATION
and it is not hard to show that both onesided derivatives of f ∗ (c) at c = 0 are 0, so f ∗ is differentiable there too. For c ≤ 0 we get (f ∗ ) (c) = 0 = x ∗ , and a little calculation shows that (f ∗ ) (c) = x ∗ for all c > 0 as well. Thus (f ∗ ) (c) = L3 (x ∗ , y ∗ , c) for all c in accordance with equation (3.7.5). 3.8.5 See Fig. A3.8.5 in the answer section of the book. Since the Lagrangian
1 1 L(x, y) = x + y − 2 (x + y)2 − 4 x − 1 y − λ1 (x − 5) − λ2 (y − 3) − λ3 (−x + 2y − 2) 3
is concave, a point that satisﬁes Kuhn–Tucker conditions must be a maximum point. The objective function has no stationary points, so any maximum points must lie on the boundary of S. The Kuhn– Tucker conditions are: L1 (x, y) = 1 − (x + y) − L2 (x, y) = 1 − (x + y) −
1 4 1 3
− λ 1 + λ3 ≤ 0 − λ2 − 2λ3 ≤ 0 λ1 ≥ 0 λ2 ≥ 0 λ3 ≥ 0
(= 0 if x > 0) (= 0 if y > 0) (= 0 if x < 5) (= 0 if y < 3) (= 0 if −x + 2y < 2)
(i) (ii) (iii) (iv) (v)
The solution is x = 3/4, y = 0, with λ1 = λ2 = λ3 = 0. Once we have found or been told about this point it is easy to check that it satisﬁes (i)–(v), but otherwise it can be a very tedious job to go through all possible combinations of zero or positive multipliers as well as x = 0 or x > 0 and y = 0 or y > 0. In this problem there are 32 different combinations to check if we do not see any shortcuts. In fact it would probably be more efﬁcient to check each of the ﬁve straight line segments that form the boundary of S. But it would hardly be practical to do that in a problem with more than two variables, because it quickly becomes difﬁcult to visualize the geometry of the admissible set. 3.8.6 (a) The last inequality in (∗) gives
m ∗ λj (gj (x∗ ) − bj ) ≥ j =1 j =1 m
λj (gj (x∗ ) − bj )
for all λ
0
(∗∗)
m If gk (x∗ ) > bk for some k, then j =1 λj (gj (x∗ ) − bj ) can be made arbitrary large by choosing λk large and λj = 0 for all j = k. Hence, gj (x∗ ) ≤ bj , j = 1, . . . , m. By choosing all λj equal to 0 in (∗∗), m ∗ ∗ ∗ we get j =1 λj (gj (x∗ ) − bj ) ≥ 0. Now, λj ≥ 0 and gj (x∗ ) ≤ bj for every j , so each λj (gj (x∗ ) − bj ) m ∗ ∗ ) − b ) = 0. Finally, whenever x is admissible, the inequality must be zero, and then j =1 λj (gj (x j m m ∗ ∗ L(x, λ∗ ) ≤ L(x∗ , λ∗ ) implies that f (x)−f (x∗ ) ≤ j =1 λj [gj (x)−gj (x∗ )] = j =1 λj [gj (x)−bj ] ≤ 0. Therefore f (x) ≤ f (x∗ ), so x∗ solves problem (1).
(b) Proof of the second inequality in (∗): Under the given assumptions, L(x∗ , λ∗ ) − L(x∗ , λ) = m m m ∗ ∗ ∗ ∗ j =1 λj [gj (x ) − bj ] − j =1 λj [gj (x ) − bj ] = j =1 λj [gj (x ) − bj ] ≤ 0 when λ ≥ 0. Since the ﬁrst inequality in (∗) is assumed, we have shown that (x∗ , λ∗ ) is a saddle point for L.
3.9
m m 3.9.1 (A) implies that π(x∗ ) ≥ π(ˆ ), i.e. f (x∗ ) − j =1 λj gj (x∗ ) ≥ f (ˆ ) − j =1 λj gj (ˆ ). But, because x x x x ˆ m m x also solves (3.9.1), f (ˆ ) = f (x∗ ) and then j =1 λj gj (ˆ ) ≥ j =1 λj gj (x∗ ). Thus, because λj ≥ 0 and x
© Arne Strøm, Knut Sydsæter, Atle Seierstad, and Peter Hammond 2008
11. If λ > 0. z∗ ) = (0.10 3. m m m Therefore f (ˆ ) − j =1 λj gj (ˆ ) = f (x∗ ) − j =1 λj gj (x∗ ) ≥ f (x) − j =1 λj gj (x) for all x x x 0. y ∗ . m proving (A). 3.5). . y ∗ . z∗ . There are two possibilities: λ = 1 and λ = 1. z) = (± 1 10 a. (b) f ∗ (a) = a 2 . ˆ 3. y. ∓ 10 10 a. with λ = 1. and Peter Hammond 2008 . we have m m m m λj bj ≥ j =1 j =1 λj gj (ˆ ) ≥ x j =1 λj gj (x ) = j =1 ∗ λj bj (∗) Here the two middle terms. (A) If λ = 1 then (ii) and (iii)√ imply y = z. ∓ 2 2 a). y. z) = (0. Since x = −2y. being squeezed between two equal numbers. then f (x) > 0 for all x in an open interval U around x0 (because f is continuous). 0) is obviously not a maximum point but a global minimum point in this problem. But the point (x. then μ = 0 and x = √ so y + √= 0 and y 2 + z2 = a 2 . y. There are (I) If f (x0 ) > 0. if gk (ˆ ) < bk and λk > 0 for any k. for all x in U we have f + (x)2 = f (x)2 . If λ = 0. ∓ 10 10 a). so λ = 3/5. 0. and let x0 be any point in A ⊆ three cases to consider: . We use Theorem 3. a)/∂a = 2λa = 2a. j = 1.CHAPTER 3 STATIC OPTIMIZATION 37 x gj (ˆ ) ≤ bj . a) = x 2 + y 2 + z2 − λ(2x 2 + y 2 + z2 − a 2 ) − μ(x + y + z). so x = y = z = μ = 0. then 2x 2 + y 2 + z2 = a 2 because of complementary slackness. which implies 3x = 0. then j =1 λj (gj (ˆ ) − bj ) < 0. Also. If we evaluate the objective function x 2 + y 2 + z2 at each of the points we have found. which contradicts x x (∗). ± 2 2 a. the necessary conditions are: ∂L/∂x = 2x − 4λx − μ = 0 ⇐⇒ 2(1 − 2λ)x = μ ∂L/∂y = 2y − 2λy − μ = 0 ⇐⇒ 2(1 − λ)y = μ ∂L/∂z = 2z − 2λz − μ = 0 ⇐⇒ 2(1 − λ)z = μ λ ≥ 0.1 Assume ﬁrst that f is a function of just one variable. m. z) = (0. since the corresponding constraint is an equality.1 (a) Since the admissible set is closed and bounded. z. with y = −z = a 2 /2.9. .11 3.8. ± 2 2 a. μ = 0 both solve the problem. equations (i) and (ii) 5 yield −4(1 − 2λ)y = μ = 2(1 − λ)y.10. . (B) If λ = 1. Atle Seierstad. y. ∓ 2 2 a). With L(x. Thus x satisﬁes (A)–(C). z 1 1 gives the two points (x. we ﬁnd that √ √ 1 1 the two points (x ∗ . Knut Sydsæter. and λ = 0 if 2x + y + z < a 2 2 2 2 (i) (ii) (iii) (iv) There is no sign restriction on μ. and ∂L(x ∗ . Such is often the case when the Lagrangian is not concave. . and also because of (3. We have found several other points that satisfy the necessary conditions for a maximum but are not maximum points. not an inequality. and so (d/dx)(f + (x)2 ) = 2f (x)f (x) = 2f + (x)f (x). √ √ 1 1 with the solutions (x. Then. we get x = y = z = μ/2. This 0. so df ∗ (a)/da = 2a.3 (necessary ﬁrstorder conditions for problems with mixed constraints). there is at least one maximum point in this problem. so the constraints yield x + 2y = 0 and 2x 2 + 2y 2 = a 2 . must themselves be equal. © Arne Strøm.
dx = 2(u − 1) du. . There is then an open interval W around x0 such that for all x = x0 in W . f (x) = 2f + (x)∇f (x) ∂x1 ∂xn as x → x0 which implies f (x) − f (x0 ) < f (x0 ) + 1 = K + 1 x − x0 Note that f need not really be C 1 . This result immediately carries over to the partial derivatives of f + (x)2 if f is a function n → : ∂ ∂ f (x) (f + (x))2 = 2f + (x) ∂xi ∂xi and the gradient of f + (x)2 is ∇(f + (x)2 ) = ∂ ∂ (f + (x))2 . Thus. .1. then there is an open interval V around x0 such that for all x in V we have f (x) < 0 and f + (x) = 0. let K = f (x0 ). . Chapter 4 Topics in Integration 4. f (x) − f (x0 ) − f (x0 ) < 1. We get 9 4 √ ( x − 1)2 dx = x 9 4 √ x−2 x+1 dx = x 9 4 2 1 dx = 1− √ + x x √ 9 (x − 4 x + ln x) = 1 + ln 4 4 9 (b) With u = 1 + 1 0 √ x we get x = (u − 1)2 .) If f (x0 ) = 0. . . we get f + (x)2 f (x)2 f + (x)2 − f + (x0 )2 = ≤ < (K + 1)2 x − x0  → 0 x − x0 x − x0 x − x0 and so ((f + )2 ) (x0 ) = 0 = 2f + (x0 )f (x0 ). (f + (x))2 ∂x1 ∂xn ∂ ∂ = 2f + (x) f (x). Knut Sydsæter. in all three cases we have shown that (d/dx)(f + (x)2 ) = 2f + (x)f (x). All that is needed is that all the ﬁrstorder partial derivatives of f exist. .1 4. and then (d/dx)(f + (x)2 ) = 0 = 2f + (x)f (x). x − x0 and therefore f (x) = f (x) − f (x0 ) < (K + 1)x − x0  Then for all x = x0 in W . and Peter Hammond 2008 .38 CHAPTER 4 TOPICS IN INTEGRATION (II) If f (x0 ) < 0. (III) (The most interesting case.5 (a) Expand the integrand. and √ x ) dx = 1 2 ln(1 + 2(u − 1) ln u du = 2 1 1 2 2u 2 1 (u − 1)2 ln u − 1 1 2 2 (u − 1)2 du u = ln 2 − − 2u + ln u = © Arne Strøm. . . Atle Seierstad.
3.2). 1 1 1 1 1 (k + 1 + 2 ) = ((k + 2 ) + 1) = (k + 2 ) (k + 2 ) = (k + 2 ) 1√ 2 π. it must be a stationary point—that is. Therefore c − p + (h + p)F (Q∗ ) = 0. which implies F (Q∗ ) = (p − c)/(p + h).10 (a) g (Q) = c + h a 0 a Q Q 0 f (D) dD −p a Q f (D) dD. the inequality for Gt boils down to k(τ )f (t − τ ) ≤ q(τ ) for all τ ≤ t.2 4.3.1) and then introduce u = αx 2 as a new variable. By mathematical induction it is true for all natural numbers n. we need to know that F (0) = 0 x dx = 1/2.3. Q (b) Since f (D) dD = 1. Knut Sydsæter.2 in this case) we need t t to know that there exist functions p(τ ) and q(τ ) such that −∞ p(τ ) dτ and −∞ q(τ ) dτ converge. ( 2 ) = ( 2 + 1) = 2 ( 2 ) = correct for n = 1.2. 4. then formula (4.2. which gives F (α) = 2α 2α 2 α ueu du = 1 2α 2 α (ueu − eu ) = 1 F (α) − F (0) eα − 1 − α “0” = lim = ··· = = 2 α→0 α→0 α 2α 0 4 −δe−δ(t−τ ) y(τ ) dτ = y(t) − δx(t) 4. 4.2. Since G (τ.2 We use formula (4. 22k−1 (k © Arne Strøm. assuming α = 0. The given formula is thus (2k − 1)! √ π − 1)! (2k − 1)! 2k(2k + 1) √ (2k + 1)! √ 2k + 1 (2k − 1)! √ π= π = 2k+1 π = 2k−1 (k − 1)! 2k−1 (k − 1)! k! 2 2 2 · 2k · 2 2 Thus the proposed formula is valid also for n = k + 1. using (4.2. This yields F (α) = 0 1 ∂ 2 (xeαx ) dx = ∂α 1 0 x 3 eαx dx = 1 2α 1 2 1 2α 2 eαx 2 Direct calculation of F yields F (α) = conﬁrming the result above.1) yields F (0) = 0 x 3 dx = 1/4.8 See the answer in the book. Suppose it is correct for n = k. and g (Q) = (h + p)f (Q) ≥ 0.2. dx = 3(u − 1)2 du. Then.2 From the functional equation (4.2.1). t) ≤ q(τ ) for all τ ≤ t. Atle Seierstad. and 27 0 x 1/3 dx = 1 + x 1/3 4 1 u−1 3(u − 1)2 du = u 4 1 45 3u3 − 9u2 + 9u − 3 du = · · · = − 3 ln 4 u 2 4.2. 1 We assumed above that α = 0.CHAPTER 4 TOPICS IN INTEGRATION 39 (c) Let u = 1 + x 1/3 . and Peter Hammond 2008 . To 1 get the answer by differentiating F directly. . Then F (0) = lim as it should be. x(t) = e−δ(t−τ ) y(t) + ˙ t −∞ 0 αeα − eα + 1 2α 2 0 0 α −1 α − eα + 1 e αe = . and such that f (t − τ )k(τ ) ≤ p(τ ) and Gt (τ. If α = 0. Then x = (u − 1)3 . t) = −k(τ )f (t − τ ). 4.3 3 1 1 1 4.2. g (Q∗ ) = 0.6 By Leibniz’s formula (Theorem 4. we have f (D) dD = 0 a f (D) dD − 0 Q f (D) dD = 1 − 0 Q f (D) dD and therefore g (Q) = c − p + (h + p) 0 f (D) dD Since Q∗ is the minimum point of g(Q).2). In order to use Leibniz’s formula (Theorem 4.
1 0 2 2 2 ( 1 + x2 + x3 + · · · + xn ) dx2 = 3 3 2 2 2 [ 1 x2 + 1 x2 + x2 (x3 + x4 + · · · + xn )] = 3 3 2 3 2 2 2 + x3 + x4 + · · · + xn etc.5 (a) Introduce u = λx as a new variable. © Arne Strøm. With u = (λ − t)x as a new variable we α M(t) = (α) uα−1 e−u λα λ du = = α α (λ − t) (α) (λ − t) λ−t = λα (λ − t)−α Differentiation gives M (t) = αλα (λ − t)−α−1 and in general M (n) (t) = α(α + 1) · · · (α + n − 1)λα (λ − t)−α−n = Hence. we get x = (w2 ebw − w2 ) − 2 dw = and dx = − 2 dw. Therefore I = w w x w 1 1/a 1 b 1 (−ebw + 1) dw = (e − eb/a ) + − 1. x y=0 x x x2 x 0 x 1 1/a 1 1 1 1 With w = . so I = e − 2 dx. Obviously.4.5. b a 1 1 x2 b 4. By induction. I = n/3. Then ∞ −∞ f (x) dx = ∞ −∞ λα (α) ∞ 0 x α−1 e−λx dx = λα (α) ∞ 0 λα (α) ∞ 0 (u/λ)α−1 e−u 1 du = λ 1 (α) ∞ 0 uα−1 e−u du = 1 (b) M(t) = get etx f (x) dx = λα (α) ∞ 0 x α−1 e−(λ−t)x dx. 2 + 3a a a + 3a 1 0 1 0 3 2 2 2 [ 1 x1 + x1 (x2 + x3 + · · · + xn )] = 3 1 3 2 2 2 + x2 + x3 + · · · + xn 2 2 2 (x1 + x2 + · · · + xn ) dx1 = Next.4.4 4. Knut Sydsæter. y) dx dy is 1 a 1 x=a 2k k dy dx dy = − 3 (x + y + 1)2 0 0 (x + y + 1) 0 x=0 1 1 k k k k − dy = − = (y + 1)2 (y + a + 1)2 y+1 y+a+1 0 0 k k k(a 2 + 3a) k +k− = =− + 2 a+2 a+1 2(a 2 + 3a + 2) The integral equals 1 if k = ka = 4.3. M (0) = α λ and M (n) (0) = α(α + 1) · · · (α + n − 1) = λn ∞ n −∞ x f (x) dx λα (α + n) (α) (λ − t)α+n (α + n) (α)λn (Alternatively we could have used the formula M (n) (0) = from Problem 4. Atle Seierstad.) 4. ka > 2 if a > 0.5 The innermost integral is 1 0 4 2(a 2 + 3a + 2) =2+ 2 .3 The integral I = I= R f (x.4.2 The inner integral is a 1 y=b y/x 1 1 1 b/x 1 1 y/x e dy = 2 e = 2 eb/x − 2 .2.40 CHAPTER 4 TOPICS IN INTEGRATION 4. and Peter Hammond 2008 .
Atle Seierstad.5.7. 2n − 1. .6. 1 1 1 1 1 1 = 2 3 n(n − 1)2n − 3 (2n − 1)2nn + 2 n2n = 2 − → 2 n 2 n n n 2 4. 2]. Then (2xi∗ − yj∗ + 1) xi and 2n−1 n−1 yj = 2 j 11 i i j 1 − +1 =2 3 − 3 + 2 n n nn n n n 2n−1 n−1 2 j =0 i=0 j 1 i 1 − 3 + 2 =2 3 3 n n n n =2 1 n3 i − j =0 2n−1 j =0 i=0 1 n3 n−1 2n−1 j + i=0 j =0 n−1 1 n2 2n−1 n−1 1 j =0 i=0 2n−1 1 1 n(n − 1) − 3 2 n i=0 1 1 (2n − 1)2n + 2 2 n n j =0 as n → ∞.1 (a) Use the same subdivision as in Example 4. and Peter Hammond 2008 . 1]. y3 67 dy = 2 120 xy 2 dx dy + 1 2 1 √ y−1 xy 2 dx dy = 0 1 y4 dy + 2 2 1 y2 − 4. . then x runs from 0 to y.6. Thus. Knut Sydsæter. Hence (see Fig.2 Q2 Q3 © Arne Strøm. then x must run from V = 0 1 0 y √ y − 1 to 1.2 If y ∈ [0.6 x − y = x − y if x ≥ y and y − x if x < y.6(a) in the book). and if y ∈ [1.5 4. 1 0 0 1 0 0 1 0 1 x − y dx dy = y x − y dx dy + A 1 y 1 0 x − y dx dy B 1 0 x=y x=0 = = (y − x) dx + (x − y) dx dy = 1 yx − x 2 + 2 1 dy = 2 1 0 x=1 x=y 1 2 x − yx 2 dy 1 1 1 y 2 − y 2 + − y − y 2 + y 2 dy = 2 2 2 y2 − y + 1 1 3 1 2 1 y − y + y = 3 2 2 3 4. . .5. A4.1.6 4.CHAPTER 4 TOPICS IN INTEGRATION 41 4. except that j = 0.5.7 y P3 P1 P2 x Q1 Figure M4.
1 (a) Use polar coordinates and let An = {(x. y) : 1 ≤ x 2 + y 2 ≤ n2 }. Finally. and that x1 ≤ x2 ≤ x3 . 1 x1 y1 1 and an easy computation shows that this equals A = 2 1 x2 y2 . whose area is 1 Tij = 2 (xj − xi )(yi + yj ). and Peter Hammond 2008 . In either case the area equals A. y = 1 + r sin θ. That may change the sign of the determinant but not its absolute value.7.3 (b) In this problem it is convenient to use polar coordinates centred at (0. (b) With polar coordinates and with An as in part (a). a trapezoid in the US). Qj . If P2 lies above P1 P3 .7. then In → ∞.7.7. 1). y) 1 = − 5 . so let x = r cos θ.5 (b) Introduce new variables u = y − 2x. y = 5 5 ∂(x. For each pair (i. then In → π/(p − 1) as n → ∞. we can move it there by a parallel translation. yi ).8 4. v) (x + y) dx dy = A2 4 8 1 −1 + 2 v and 5 2 3 u + v J  du dv = 5 5 8 4 1 −1 2 3 1 u + v du dv = 144/25 5 5 5 4. v = 3x + y. 2. y)/∂(r. Then (x 2 +y 2 )−3 dx dy = An ∗ 0 2π 1 n −6 1 r r n n 1 r −6 r dr dθ = 2π ∗ 1 1 r −5 dr = 2 π(1−n−4 ) → 2 π as n → ∞ About the equality =: The integral Jn = dr is independent of θ. Then x = − 1 u + 1 v. θ) = r in this situation too. all lie in the ﬁrst quadrant.8. nor will it change the value of the determinant.4 Note that the integrand takes both positive and negative values in the domain of integration. and Pj form a quadrilateral with two parallel sides (called a trapezium in Britain.8. but if p ≤ 1. 3.2 shows the triangle P1 P2 P3 together with the normals from the Pi to the corresponding points Qi on the xaxis. Qi . j ) with i < j . 4. if the triangle does not lie in the ﬁrst quadrant. If P2 lies below the line P1 P3 .42 CHAPTER 4 TOPICS IN INTEGRATION 4. It follows that J = ∂(u. The Jacobian is ∂(x. we can renumber them. Knut Sydsæter. then the area 1 x3 y3 of the triangle is T12 + T23 − T13 = −A. then the area of the triangle is T13 − T12 − T23 . Such a translation will not change the area of the triangle. i = 1. 4. Figure M4. we get In = An (x + y ) 2 2 −p dx dy = 2π 1 n r 1−2p n2−2p − 1 dr = 2 − 2p ⎩ 2π ln n 2π if p = 1 if p = 1 If p > 1. therefore ⎧ ⎨ 2π 0 Jn dθ = 2πJn . Atle Seierstad. and x 2 dx dy = A 0 2π 0 1/2 r 3 cos2 θ dr dθ = 0 2π r=1/2 r=0 r4 1 cos2 θ dθ = 4 64 2π 0 cos2 θ dθ = 3 5u π 64 4.2 Assume for convenience that the points Pi = (xi . If the xi are in a different order from what we assumed above. In the calculations of the two iterated integrals the positive and the negative parts will more or less balance © Arne Strøm. the points Pi . since we are just adding multiples of the ﬁrst column to the other two columns.
or by solving the separable © Arne Strøm.) 4. The 1 1 y(y + x) trouble arises when we try to take the difference. Knut Sydsæter.2 The solution curves are semicircles (not full circles) of the form t 2 + x 2 = C.2 x x 2 + y 2 = 4.3. Then x = 2 (u + v). and so do the integrals of x(y + x)−3 . (Here we used e−v dv · arctan n. v) = −1/2.1.8.2. with the √ √ 1 initial condition x(0) = 0. The square Bn = [−n. 0). Instead of “for all t” it should have said “for all t in some open interval I around 0”. with C an arbitrary positive constant. With that modiﬁcation the answer in the book is quite correct. (−2n. x = 0.8. One can show that. x > 0 1 1 t Figure M5. 5. n] × [−n. the equation has the solution x = tan( 2 t 2 ) over (− π. 2 2 n n n n e−(x−y) e−v 1 1 1 2 dx dy = e−v dv du = du dv = and In = 2 2 2 2 −n −n 1 + u −n −n 2 1 + u Bn 1 + (x + y) n ∞ √ π 3/2 π π 2 2 = π· = .7 (b) With polar coordinates and with the sets An as in Example 3.3). (Actually.1 5. the given differential equation has no solution deﬁned on the entire real line. Then In → e−v dv · 2 2 2 −n −∞ Poisson’s integral formula (4. That leads us into an ∞ − ∞ situation that does not lead to any deﬁnite value.2. and the Jacobian determinant is ∂(x. The two iterated integrals 1 1 y(y + x)−3 dx dy and b d −3 dy dx both tend to ∞ as b and d tend to ∞. 1 1 4.CHAPTER 5 DIFFERENTIAL EQUATIONS I: FIRSTORDER EQUATIONS IN ONE VARIABLE d b 43 each other. y = 2 (u − v). (0. − ln(x 2 + y 2 ) A x2 + y2 dx dy = lim π/2 1 1/n n→∞ 0 − ln r 2 r dr dθ = −π lim n→∞ r 1 1/n ln r dr = · · · = π Chapter 5 Differential Equations I: Firstorder Equations in One Variable 5. but this solution cannot be extended to any larger interval because x(t) runs off to inﬁnity as t approaches either endpoint. π). Let n → ∞. but not in exactly the same way. 0). Atle Seierstad. 2n).6 (b) Introduce new variables u = x + y. and Peter Hammond 2008 . and (0. y)/∂(u. (This can be shown by direct differentiation. n] in the uvplane corresponds to the square Bn with corners (2n. −2n) in the xyplane. v = x − y.2 5.6 The statement of the problem is slightly inaccurate.
x0 ) = (1.3. x0 ) = ( 2. x ˙ t2 = x 1 + t3 ⇒ ln x = 1 3 ln 1 + t 3  + C1 ⇒ x = C 1 + t3 3 Integral curve through (t0 . and we get (a) The equation x dx = x a dt t ⇒ ln x = a ln t + C1 = ln t a + C1 ⇒ x = eln x = t a · eC1 = Ct a where C = eC1 . = at + b. (c) No constant solutions. so x = 1 + e−2t 5.3. 2). (b) One constant solution. so x = t 2 − 1. The nonconstant ˙ solutions are found by separating the variables: dx = (x + 1)2 e−2t dt ⇒ − 1 1 = − 2 e−2t + C1 x+1 ⇒ x= 1 − e−2t + C 1 + e−2t − C where C = 1 + 2C1 .3 5. One constant solution. x = eln t−t+C1 = eln t e−t eC1 = C2 te−t . (The ﬁgure in answer in the book is misleading since it shows a full circle. x ≡ 0. The integral curve through (t0 . x0 ) we must have 1 − e0 + C = 0. 1) must have C = −1 and x > 0. x > 0. where C = ±C2 = ±eC1 . x(t) = − 4 − t 2 .3 (a) One constant solution.6 For convenience we shall suppose that x > 0 and t > 0. but not the one you were asked to ﬁnd. which is also a solution of the differential equation. Knut Sydsæter. Otherwise. 1 − e−2t . where C = 2C1 .2 is the graph of another function. in other words it is the graph of the function x(t) = 4 − t 2 over the interval √ (−2. M5. The lower semicircle shown in Fig. separating the variables.2.) 5. 2) is ˙ √ given by t 2 +x 2 = 4. dx = x 1−t dt = t 1 − 1 dt t ⇒ ln x = ln t − t + C1 Hence. tx ˙ (b) Here. 1/e) is x = te−t . and Peter Hammond 2008 .) The integral curve through (0. t dt 1 2 2x x dx = ⇒ 1 = 2 t 2 + C1 (d) The equation is equivalent to x = e−2t (x + 1)2 . i.e. 2) for C = 2. Otherwise. x ≡ −1. ⇒ x 2 = t 2 + C. That gives x dx = x at + b dt = t a+ b dt t ⇒ ln x = at + b ln t + C1 ⇒ x = Ceat t b © Arne Strøm. tx ˙ = a is separable. C = 0. x ≡ 0. Atle Seierstad. For the solution to pass through (t0 . so x = Cte−t . x0 ) = (0.44 CHAPTER 5 DIFFERENTIAL EQUATIONS I: FIRSTORDER EQUATIONS IN ONE VARIABLE equation x = −t/x using the method set out in the next section. √ √ An integral curve through (t0 .
x dx = dt. K = (1 − b + c)eQt + C 1−b+c Q Q C = C1 (1 − b + c).CHAPTER 5 DIFFERENTIAL EQUATIONS I: FIRSTORDER EQUATIONS IN ONE VARIABLE 45 (c) The equation tx ˙ x ˙ 1 1 1 1 a = ax + b gives = . Hence. and so K0 − pα + 1 α 1/α α K = K0 + sαAbα (t + a)pα+1 − a pα+1 /(pα + 1) It follows that K = L α K0 + sαAbα (t + a)pα+1 − a pα+1 pα + 1 bα (t + α)pα 1/α 1/α α K0 a pα+1 sαA t +a− + = α b (t + α)pα pα + 1 (t + α)pα 1/α → ∞ as t → ∞ © Arne Strøm. Knut Sydsæter. 0 α K0 + (sA/λ)Lα (eαλt − 1) K 0 = L [Lα eαλt ]1/α 0 1/α = 1−α α K0 sA (1 − e−αλt ) α αλt + λ L0 e 1/α → (1−α)/α sA λ 1/α and AK 1−α Lα K X = =A L L L →A sA λ (1−α)/α = A1/α s λ (b) The equation is separable.9 (a) Note ﬁrst that L = L0 eλt = [Lα eαλt ]1/α . where = e + C1 . so as t → ∞. 1 − Cat b K c−b dK = P eQt dt. 5. We get K α−1 dK = sAbα The initial condition gives C = (t + a)pα dt ⇒ 1 α sAbα K = (t + a)pα+1 + C α pα + 1 sAbα pα+1 1 α a . we get x x(ax + b) t x(ax + b) b x ax + b b dt t ⇒ ln x − ln ax + b = b ln t + C1 ⇒ x = Ct b ax + b 1 a − dx = x ax + b where C = ±eC1 . Atle Seierstad. Separating the variables we get 0 1/(1−b+c) P K 1−b+c P Qt .8 Cbt b .3. and ﬁnally x = 5. The hint in the problem gives (b) We separate the variables and get (β − αx)(x − a) Integration yields β dx + β − αx ⇐⇒ a dx = (β − αa) x−a dt ⇐⇒ − β ln β − αx + a ln x − a = (β − αa)t + C1 α β ln β − αx − a ln x − a = ln β − αxβ/α + ln x − a−a = −(β − αa)t − C1 α ⇐⇒ β − αxβ/α x − a−a = e−(β−αa)t−C1 = e−(β−αa)t e−C1 = Ce(αa−β)t where C = e−C1 . (a) Let P = Anα a b and Q = αv + ε. and Peter Hammond 2008 . dK/dt = sAK 1−α Lα = sAK 1−α bα (t + a)α . We have the same answer as in the book (since β − αx = αx − β).3. Since = − .
2 With f (t. ˙ 1−r = x −1 = 1/x. x(t) is positive and increasing for large t as long as μ > ρ. then x(t) → β/α. x + (t + 2x)x = 0 is exact. Then x = −z−2 z and the differential equation becomes ˙ ˙ Let z = x ˙ ˙ −z−2 z + (2/t)z−1 = z−2 ⇐⇒ z − (2/t)z = −1 whose solution is z = Ct 2 + t. We assumed that t > 0 the differential equation is obtained from tx + x 1 and x > 0.4. Knut Sydsæter. Then x = 2z˙ and the given equation becomes ˙ 2z˙ = 4z2 + 2et z ⇐⇒ z − 2z = et z (Recall that x is assumed to be positive. the given equation is equivalent to the Bernoulli equation x + (2/t)x = x r with r = 2. (d) Here a(t) = −2/t. Atle Seierstad. by logarithmic differentiation.11) that we can choose β(x) = exp( (1/x) dx) = exp(ln x) = x as an integrating factor. x/x = X/X − N/N. we have (gt − fx )/f = 1/x. 5. and therefore z > 0. It follows that x = (a −1)αx −(a −1)β.4. √ 1 (c) In this case. (b) Here a(t) dt = − (1/t) dt = − ln t.1 (a) With t > 0.6. and Peter Hammond 2008 . b(t) = −2a 2 /t 2 . (ii) implies that ˙ ˙ ˙ ˙ X/X = a N/N.5 5. X(t) = A[N(t)]a . If 0 < a < 1. so we are in Case (II). x) = tx + x 2 − t0 x0 − x0 . and (5. 5.6) yields the solution x = Ct + t 2 .4.4. a(t) dt = −2 ln t. 2 Hence. ˙ ˙ 5. N(t) → (β/αA)1/(a−1) . so that x = 1/z. so C will be positive. i. which leads to x = Ct 2 + 2a 2 /3t. Thus we substitute z = x 1−1/2 = ˙ z x 1/2 . (c) Using equation (∗∗).4.6) yields the solution x = C t 2 − 1 + t 2 − 1. and X(t) → A(β/αA)a/(a−1) as t → ∞. where C is a constant.9 From x = X/N.4. ˙ ˙ The solution is x(t) = [x(0) − β/α]eα(a−1)t + β/α. Thus x = (Ct 2 + t)−1 . so x/x = (a −1)N/N = (a −1)[α −β(1/x)]. x = z2 . we get x(t) = x(0) + ασ − μ N0 μ − ασ N0 Note that. and (5.4 5. and the solution of tx + x 2 = C is x = − 2 t + 1 2 4t + C.e.4) yields z = e2t (C + e−2t et dt) = e2t (C + e−t dt) = e2t (C − e−t ) = Ce2t − et . x) = 1 and g(t.6 5. (b) The equation is a Bernoulli equation as in (5. Then (ii) and x = X/N together imply that N(t) = [x(t)/A]1/(a−1) . x) = t/x + 2. Moreover. © Arne Strøm.4.46 CHAPTER 5 DIFFERENTIAL EQUATIONS I: FIRSTORDER EQUATIONS IN ONE VARIABLE 5. Faster growth per capita is to be expected because foreign aid contributes positively.6 Use formula (5. Then apply this with ξ = ασ − μ.1). and (8) easily yields h(t. 5.5. It follows from (5.6. The solution of ˙ 2 = C. H0 −(ρ−ασ )t H0 (μ−ρ)t σ σ e + e . a(t) dt = − 2 ln(t 2 − 1).) Formula (5.6) to solve these equations.5.10 (b) It sufﬁces to note that (1 − e−ξ t )/ξ > 0 whenever ξ = 0 (look at ξ > 0 and ξ < 0 separately). even if ασ < ρ. with r = 1/2. So foreign aid must grow faster than the population.
Knut Sydsæter. we ﬁnd that (see (5. By (5.4) and calculate like this: z − 2z = et ⇐⇒ (˙ − 2z)e−2t = e−t ˙ z d (ze−2t ) = e−t ⇐⇒ ze−2t = −e−t + C. where A = ±e2C .3).7 The equation is of the form x = g(x/t) with g(z) = 1 + z − z2 .3 Introducing z = K 1−b as a new variable.6. ⇐⇒ dt The solution of the given equation is therefore x = z2 = (Ce2t − et )2 (c) As in part (a). the solution of this linear differential equation is z = Ce−P t + Qe−P t = Ce−P t + Qe−P t eP t e(av+ε)t dt = Ce−P t + Qe−P t e(av+ε+P )t dt Qe(av+ε)t 1 e(av+ε+P )t = Ce−P t + av + ε + P av + ε + P Insert the values of P and Q.6) yields the solution z = eln t C − e− ln t ln t dt = t C − t ln t dt = Ct + ln t + 1 t2 ⇒ x = (Ct + ln t + 1)−1 etc. we separate the variables and get 1 1 1 dt dt dz . By a wellknown identity. we substitute x = 1/z. Atle Seierstad.4.3). Integration = 2 1−z t 2 1+z 1−z t 1 1 yields 2 ln 1 + z − 2 ln 1 − z = ln t + C. Solving for z gives z = © Arne Strøm. ˙ According to (5.6.3)) z + P z = Qe(av+ε)t . To ﬁnd the other solutions. and Peter Hammond 2008 .6. According to Problem 5. but the problem explicitly calls for solutions with x > 0.) 5. ˙ the substitution z = x/t leads to the separable equation t z = g(z) − z = 1 − z2 . the solution is z = Ceγ2 (1−α)t − γ1 b/γ2 .6. 5.) (The answer in the book also lists x ≡ 0 as a solution. 5. and ﬁnally x = tz = 2 . This has the two ˙ constant solutions z = −1 and z = 1. This leads to the differential equation ˙ z − (1/t)z = −(ln t)/t ˙ Formula (5. According to (5.4.4).6.CHAPTER 5 DIFFERENTIAL EQUATIONS I: FIRSTORDER EQUATIONS IN ONE VARIABLE 47 (Alternatively we could go back to the method that was used to deduce (5.6.6. and it certainly satisﬁes the equation. In addition we have At 2 + 1 At + 1 the two solutions x = −t and x = t.4 Introduce z = K 1−α as a new variable. Then K = z1/(1−b) gives the answer in the book.4. we get the equation z − γ2 (1 − α)z = γ1 b(1 − α). with x = −z−2 dz.4. so ln 1+z = 2 ln t + 2C = ln t 2 + 2C 1−z ⇒ 1+z = At 2 1−z At 3 − t At 2 − 1 . corresponding to z = ±1. where ˙ a P = αδ(1 − b) and Q = αAn0 (1 − b). Then K = z1/(1−α) gives the answer in the book. + dz = + C.
From F (K. 5. (c) See the answer in the book. M5. 5. The necessary ﬁrstorder condition for this to be maximized w.8. For x0 = 1. But then 1 − x ≤ a(t) + b(t)/x. The argument ˙ √ for t > b is similar. so 1 − x ≤ a(t) + b(t)/x 2 . A5. and the solution through (t0 . See Fig.4 The equation is separable. L) = Lf (k)(1/L) = f (k).3 were satisﬁed.7. and Peter Hammond 2008 .r. x 3 2 1 −4 −3 −2 −1 −1 −2 −3 Figure M5. and √ ϕ(a) = limt→a (ϕ(t) − ϕ(a))/(t − a) = 0 = 2 ϕ(a). we obtain FK (K.8. Similarly.5 See Fig. For t < a.8. so when k = k ∗ we have c = f (k ∗ )−λk ∗ . For t < a. In the same way we show that the differential ˙ equation is satisﬁed at t = b. and decreases as the growth rate of the work force increases.3(a) in the answer section of the book for some integral curves. capital per worker increases as the savings rate increases. and for t slightly larger than a.48 CHAPTER 5 DIFFERENTIAL EQUATIONS I: FIRSTORDER EQUATIONS IN ONE VARIABLE 5. For t in (a. there would exist numbers a(t) and b(t) such that x(1 − x) ≤ a(t)x + b(t) for all x. −(t − a) It follows that when t is near a. Thus ∂F /∂K = λ. If x0 > 1.8 1 5. b) we have ϕ(t) = 0 = 2 ϕ(t). then x(t) → ∞ as t → (− ln A)− . L) = Lf (k) with k = K/L. But F (K. Atle Seierstad.8. (ϕ(t) − ϕ(a))/(t − a) = ˙ 2 /(t − a) = −(t − a) = a − t. x(t) → −1 as t → ∞. which clearly is impossible when x is sufﬁciently large. √ 5.4 (a) ∂k ∗ /∂s = f (k ∗ )/[λ − sf (k ∗ )] > 0 and ∂k ∗ /∂λ = −k ∗ /[λ − sf (k ∗ )] < 0 when λ > sf (k ∗ ).7.t. L) = Lf (k) and so FK = Lf (k)dk/dK = f (k) because k = K/L with L ﬁxed.3 (a) This is a separable equation with solution x(t) = (1+Aet )/(1−Aet ) where A = (x0 −1)/(x0 +1) for x0 = −1. so ϕ is differentiable at a. In equilibrium. and x(t) → −∞ as t → (− ln A)+ .7. which occurs when 0 < A < 1. k ∗ is that f (k ∗ ) = λ.8. For x0 = −1 we get the constant solution x(t) ≡ −1. 2 ) is x = 1/(1 + e−t ). (ϕ(t) − ϕ(a))/(t − a) ≤ t − a.5. (4) implies x 2 (1 − x) ≤ a(t)x 2 + b(t).5 x = ϕ(t) t 1 2 3 4 © Arne Strøm. c = (X − K)/L = (1 − s)X/L = (1 − s)f (k). But sf (k ∗ ) = λk ∗ . ˙ (b) From equations (i) to (iv). Knut Sydsæter. (ϕ(t) − ϕ(a))/(t − a) = 0. then for every t. x0 ) = (0. If condition (3) in Theorem 5. which becomes impossible as x → −∞.7 5. ϕ(t) = −2(t − a) = 2(a − t) = 2 (a − t)2 = 2 ϕ(t).
1/t ) (b) (i) Obviously.1. and so u = Aet − t 2 − 2t − 2. we get u + 2u = 8. and further © Arne Strøm. The equation for t in (a) becomes 1 t = −(t )3 (−x)(1/t 3 ) = x. s ). We know that if x = ϕ(t) = 0. which has the solution (see (5. 1/t ) (t )3 ⇐⇒ t = −(t )3 F (x. Then u − u = t 2 .4. with A = − 2 C. or d 2 t/dx 2 = x.CHAPTER 6 / DIFFERENTIAL EQUATIONS II: SECONDORDER EQUATIONS 49 6 Differential Equations II: SecondOrder Equations and Systems in the Plane 6. with the solution (see (5. x) becomes ¨ ˙ − t = F (x. then the equation ¨ ˙ ˙ ˙ x − ϕ(t) = 0 deﬁnes t as a function of x. Knut Sydsæter.4 In each of these three equations.1. where t = dt/dx.6 (a) Suppose x = ϕ(t) is a solution of x = F (x. Integration yields dt/dx = 2 x 2 + A. (The prime denotes differentiation with respect to x.3)) u = Ce−2t + 4. let u = x.4. But then ˙ ˙ x= 1 1 (Ce−2t + 4) dt = − 2 Ce−2t + 4t + B = Ae−2t + 4t + B. and Peter Hammond 2008 . (Aet − t 2 − 2t − 2) dt = Aet − 1 t 3 − t 2 − 2t + B. which has the solution (see (5. (c) Let u = x.1 6.4)) u = Aet + et ˙ ˙ integration by parts twice gives Then x = e−t t 2 dt. Using e−t t 2 dt = −t 2 e−t − 2te−t − 2e−t . and afterwards ﬁnd x as x = u dt. x(t) ≡ C is a solution for any constant C. (a) Putting u = x.) Now the ˙ chain rule gives d d 1 t 1 t d 1 dx x = (x) = ¨ ˙ =− 2 =− 3 = dt dt t dx t dt (t ) t (t ) and the differential equation x = F (x. 3 6. That will give simple ﬁrstorder equations for u which ˙ you can solve. Atle Seierstad.4)) ˙ ˙ u = Ce2t + e2t e−2t 2e2t dt = Ce2t + e2t 2 dt = Ce2t + e2t 2t = (C + 2t)e2t Integration by parts then yields x= 1 (C + 2t)e2t dt = 2 (C + 2t)e2t − 1 2 2e2t dt 1 1 1 = 2 (C + 2t)e2t − 2 e2t + B = 2 (C − 1)e2t + te2t + B = Ae2t + te2t + B 1 with A = 2 (C − 1). (b) With u = x. we get u − 2u = 2e2t .4. with ∂ (x − ϕ(t)) dt 1 (∂/∂x)(x − ϕ(t)) 1 = =− = − ∂x = ∂ dx (∂/∂t)(x − ϕ(t)) ϕ(t) ˙ x ˙ ∂t (x − ϕ(t)) Therefore x = 1/t .
To ﬁnd a particular solution of the given equation we try u∗ (t) = p sin t + q cos t. Thus the general solution of the given equation is x = Ae2t + Be−3t − t − 1/6. A nonconstant solution 6 of the equation x = −x x 3 is therefore given implicitly by the equation x 3 + A1 x + B1 = 6t. and Peter Hammond 2008 . For solutions with x = 0 we use the ˙ transformation in (a) and get t = −(t )3 1/(t )2 t =− x x ⇐⇒ 1 t =− t x ⇐⇒ ln t  = − ln x + A1 This yields t = A2 /x.3 6. which inserted into the given equation yields −2pe−t = e−t . Inserting this into the given equation yields ˙ −2p sin t − 2q cos t = sin t. x = Aet + Be−t . ¨ Then u∗ = p cos t − q sin t.2. Since a = b. Atle Seierstad. where A2 = ±eA1 .2 6.5 ˙ ¨ Let x = (t + k)−1 . Since the righthand side of the equation is a polynomial of degree 1. The general solution of x − x = sin t is ¨ 1 therefore x = Aet + Be−t − 2 sin t. and the general solution of the given equation is x = Au1 + Bu2 = a(t + a)−1 + B(t + b)−1 . Knut Sydsæter. We get u∗ +u∗ −6u∗ = C−6(Ct +D) = −6Ct +(C−6D). and (t + a)(t + b)x + 2(2t + a + b)x + 2x ¨ ˙ = (t + k)−3 2(t + 2)(t + b) − 2(2t + a + b)(t + k) + 2(t + k)2 = (t + k)−3 2 k 2 − (a + b)k + ab = (t + k)−3 2(k − a)(k − b) Thus x = (t + k)−1 solves the given differential equation if and only if k = a or k = b.1(b) now tells us that the general solution is Au1 + Bu2 + u∗ . the functions u1 = (t + a)−1 and u2 = (t + b)−1 are not proportional. and u∗ = −p sin t + q cos t. Finally.2. in other words if and only if C = −1 and so u D = −1/6.2.1).2. (Note that the constants A and B here are different from 0. where A and B are arbitrary constants. where u∗ is any particular solution of the equation. and then t = t dx = A2 ln x + B1 . x(t) ≡ C is a solution for every constant C = 0. we recover the constant solutions mentioned at the beginning. © Arne Strøm.3. ∗ is a solution if and only if −6C = 6 and C = 6D.1(a) says that the general solution of x + x − 6x = 0 is x = Au1 + Bu2 = ¨ ˙ 2t + Be−3t .2 (a) The characteristic equation is r 2 − 1 = 0. we try to ﬁnd ¨ ˙ a particular solution of the form u∗ = Ct +D.3. Theorem 6. where A = 1/A2 and B2 = −B1 /A2 . with roots r = 1 and r = −1. Thus p = −1/2 and q = 0. so p = −1/2. Since e−t is a solution ˙ ¨ of the homogeneous equation. 6. we try u∗ (t) = pte−t . 6. and the pe 1 general solution is x = Aet + Be−t − 2 te−t . Then x = −(t + k)−2 .3 (a) Direct calculations show that u1 + u1 − 6u1 = 0 and u2 + u2 − 6u2 = 0. The general solution of the corresponding homogeneous differential equation is therefore (Case (I) in Theorem 6. x = 2(t + k)−3 . x = ±eB2 eAt = BeAt . This yields ln x = At + B2 .) 6. If we let A = 0 in Beat . ¨ ˙ (ii) In this case. (b) The general solution of the homogeneous equation is again x = Aet + Be−t . Since u1 and u2 are not ¨ ˙ ¨ ˙ proportional. Ae (b) Theorem 6.50 CHAPTER 6 / DIFFERENTIAL EQUATIONS II: SECONDORDER EQUATIONS integration results in t = 1 x 3 + Ax + B. Then u∗ (t) = pe−t − pte−t and u∗ (t) = −pe−t − −t + pte−t .
) One particular solution of the nonhomogeneous equation is therefore u∗ = −t 2 . The general solution of the homogeneous equation is therefore (Case (II) in Theorem 6. so we must have ˙ ¨ γ [β + α(1 − β)]P − γ δ ∗ P t − γ δ ∗ Q = −γ δ ∗ kt − γ δ ∗ L0 It follows that P = k and Q = L0 + [β + α(1 − β)]k/δ ∗ .3. and ﬁnd p = 2/75. The condition x(0) = 0 yields A + 6 = 0.3. To ﬁnd a particular solution we try u∗ (t) = pt + q. the equation ˙ 1 1 1 1 x(π/2) = 0 gives −2A + 1 = 0. Therefore A = 1/2. we get x = rt r−1 . Knut Sydsæter. To ﬁnd a particular solution we try u∗ (t) = Ct + D. ˙ ¨ Inserting this into the given equation and cancelling t r gives the equation r 2 + 4r + 3 = 0. with roots 1 r = − 2 γ [β + α(1 − β)] ± 1 2 γ 2 [β + α(1 − β)]2 + 4γ δ ∗ 1 Oscillations occur if the characteristic roots are complex. and so ˙ t − tet + t 2 − 4t + 6. with roots r = −1 and r = −3. The constants A and B are determined by the initial conditions. and so the general solution is x = At + Bt 3 − t 2 . we ﬁnd B = π/2 + 1/4. so A = −6. 6. The general solution of the homogeneous equation is therefore (Case (II) in Theorem 6. The general solution is therefore x = At −1 + Bt −3 .1). x = Ae−t + Bte−t .3.3. the condition x(0) = 1 implies −A + B − 4 = 1. ˙ 6. the general 2 3 solution of the given equation is x = Ae5t + Bte5t + 75 t + 125 . Since ˙ x(t) = −Ae−t + Be−t − Bte−t + 2t − 4. Then ¨ ˙ ∗ = 2P t + Q.3 (a) The characteristic equation is r 2 + 2r + 1 = (r + 1)2 = 0. so r = 5 is a double root. and x = r(r − 1)t r−2 . x = A sin 2t + B cos 2t. The initial condition x(π/2) = 0 gives A sin π + B cos π + π/2 + 1/4 = 0. The required solution is x = −6e (b) The characteristic equation is r 2 + 4 = 0. and Peter Hammond 2008 . © Arne Strøm.CHAPTER 6 / DIFFERENTIAL EQUATIONS II: SECONDORDER EQUATIONS 51 (c) The characteristic equation is r 2 − 10r + 25 = (r − 5)2 = 0. The characteristic equation is r 2 + γ [β + α(1 − β)]r − γ δ ∗ = 0. Hence. and ﬁnd C = 1 and D = 1/4. That is because Qt is a solution of the homogeneous equation. B = A + 5 = −1.3. we try u∗ (t) = P t 2 + Qt + R. and 2C + 2D + E = 0. we try to ﬁnd a particular solution of the form u∗ = P t + Q. Since sin π = 0 and cos π = −1. or Ct 2 +(4C+D)t +2C+2D+E = t 2 . if and only if 4 γ 2 [β +α(1−β)]2 +γ δ ∗ < 0. Then u∗ = 2Ct + D. (b) Substituting x = t r into the homogeneous equation yields the equation r 2 − 4r + 3 = (r − 1)(r − 3). Since x = 2A cos 2t − 2B sin 2t + 1. Equating like powers of t yields C = 1.8 (a) This is an Euler differential equation.1). To ﬁnd a particular solution we try u∗ (t) = Ct 2 + Dt + E. It follows that the general solution is x = A sin 2t +B cos 2t +t +1/4. and E = 6. q = 3/125. Inserted into the given equation this yields ¨ u ˙ 2P t 2 − 3t (2P t + Q) + 3P t 2 + 3Qt + 3R = t 2 ⇐⇒ −P t 2 + 3R = t 2 This holds for all t if and only if P = −1 and R = 0. and u∗ = 2P .e. 4C + D = 0. With x = t r . Inserted into ˙ ¨ the given equation this yields 2C+4Ct +2D+Ct 2 +Dt +E = t 2 . Atle Seierstad. The general solution of the homogeneous equation is therefore x = At + Bt 3 . x = Ae5t + Bte5t . 6.3. It follows that D = −4. The general solution of the homogeneous equation is therefore (Case (III) in Theorem 6. So the general solution is x = Ae−t + Bte−t + t 2 − 4t + 6.4 Since the righthand side is a linear function of t. (Note that Q did not appear in the last equation. To ﬁnd the solution with the given initial conditions we must determine A and B. t 2 x − 3t x + 3x = t 2 . so r = −1 is a double root. To ﬁnd a particular solution of the nonhomogeneous equation.1). We get u∗ = P and u∗ = 0. so r = ±2i are the complex roots. and so x = 2 sin 2t + ( 4 + 2 π) cos t + t + 4 . i. and u∗ = 2C.
˙ ˙ we get x = t + 1. and therefore ˙ ˙ x= 1 2 + at. then the solution is ¨ √ 1 rt + Be−rt − k/r 2 . Therefore x = 1. then the solution is p(t) = At + B + 2 kt 2 . The equation is not stable for any values of the constants.2 (a) If we add the two equations we get x + y = (a + b)(x + y). Note that a(s1 − d1 ) > 0. We can also see the instability from the criterion in (6. the characteristic roots are r1 = r2 = 0.10 By Leibniz’s rule. one positive and one negative. if λ = 0. then ae(a+b)t be(a+b)t x= + A. where λ = γ (a − α).4. 6. If λ > 0. p must ¨ satisfy p + a(s1 − d1 )p = a(d0 − s0 ). Because of ˙ ˙ 1 1 ˙ the initial conditions x(0) = 2 . then x = a and y = b = −a. Knut Sydsæter. From the ﬁrst equation ˙ ˙ and the initial conditions we get x(0) = 0 + y(0) = 1. From the second equation we then get ˙ 1 1 ¨ − 2 x + 2 cos t = 2x + 1 − cos t ⇐⇒ x + 4x = −2 + 3 cos t ¨ 1 (∗) © Arne Strøm. This is obvious from the form of the solutions—if λ ≥ 0. The initial conditions yield A = x(0) − and so x(t) = 2ae(a+b)t + b − a . y= 1 2 − at (b) The ﬁrst equation gives y = x − 2tx. 2(a + b) y(t) = 2be(a+b)t + a − b 2(a + b) a 1 a b−a = − = . ˙ 2 1 (c) The ﬁrst equation yields y = − 2 x + 2 sin t. and if λ < 0 the solutions oscillate with a ﬁxed amplitude. If λ > 0. where r = p(t) = Ae λ.3. we must have C = 1. a+b 2 a+b 2(a + b) B = y(0) − b 1 b a−b = − = a+b 2 a+b 2(a + b) If a + b = 0.3 Write the equation as p(t) − λp(t) = k. and Peter Hammond 2008 . Hence x + y = Ce(a+b)t . Because x(0) = 1. and y = x − 2tx = 1 − 2t (t + 1) = −2t 2 − 2t + 1. ¨ 6. y(0) = 2 . In other words. p = a[D(p(t)) − S(p(t))] = a(d1 − s1 )p + a(d0 − s0 ). the corresponding homogeneous equation has solutions that run off to inﬁnity. Atle Seierstad. then the solution is p(t) = A cos −λ t + B sin −λ t − k/λ.52 CHAPTER 6 / DIFFERENTIAL EQUATIONS II: SECONDORDER EQUATIONS 6.2): The characteristic equation is r 2 − λ = 0. and y = be ˙ If a + b = 0. y= +B a+b a+b for suitable constants A and B. this equation has two real solutions. if √ √ λ < 0.5. and if we use this in the second equation we get ˙ x − 2x − 2t x = −2t − 2x ⇐⇒ x − 2t x = −2t ¨ ˙ ¨ ˙ This is a ﬁrstorder linear equation for x with general solution x = Ce−t + 1. so C = 0. This implies x = a(x + y) = ae(a+b)t (a+b)t .5 6.4.4 6. If λ < 0. and see the answer in the book. If λ = 0. the equation has complex roots with real part equal to 0.
1 y = − 2 sin 2t + sin t ˙ ˙ ˙ ¨ ˙ ¨ ˙ 6. Thus and u ¨ 1 1 1 the general solution of (∗) is x = A cos 2t + B sin 2t − 2 + cos t. It is clear that if A = 0. and E = 0. Since the trace of A is 0. If α + 1/β > 2. If we insert these expressions for p and p into the second equation. we get ˙ απ − π = π − ˙ ¨ 1 α π+ π ˙ β β ⇐⇒ π+ ¨ 1 α −α π + 1− ˙ π =0 β β (∗) which is a secondorder differential equation for π with constant coefﬁcients. Inserting this into the second equation.2 = 2 (α − 1/β) ± 2 (α + 1/β)2 − 4. the solution of (∗) is = Aet −Be−t 0 −1 z solution of the given system is therefore x = z − 5 = Aet + Be−t − 5 and y = w + 2 = −Be−t + 2. the same solution as in (i). y ∗ ) = (−5. with roots r1. stable. has the general solution x = A cos 2t + B sin 2t. σ = απ − π. x + 4x = 0. It follows that σ = απ − π = (α − r1 )Aer1 t + (α − r2 )Ber2 t . Then p = −2e−2t (x −x)+e−2t (x − x) = e−2t (x −3x +2x). The given equation system is equivalent to the system w = w + 2z ˙ z = −z ˙ with coefﬁcient matrix A = 1 0 ⇐⇒ w ˙ z ˙ =A w z (∗) 2 . The According to (6. Knut Sydsæter. where 1± 2 are the roots of the √ √ ˙ characteristic equation. Then u∗ = −D cos t − E sin t ∗ + 4u∗ = 4C + 3D cos t + 3E sin t. then x does not converge to the equilibrium value as t → ∞. r 2 − 2r − 1 = 0. so the solutions we are looking for are 1 1 x = − 2 cos 2t + cos t − 2 . we get ˙ ¨ ˙ ˙ ˙ e−2t (x − 3x + 2x) = 2e−2t x − e−2t (x − x) = e−2t (3x − x) ⇒ x − 2x − x = 0 ¨ ˙ √ √ √ The general solution of the last equation is x = Ae(1+ 2 )t + Be(1− 2 )t . ¨ ¨ To ﬁnd a particular solution of (∗) we try u∗ = C + D cos t + E sin t. which conﬁrms that the system is not asymptotically stable.4). y = −y.3 The ﬁrst equation gives p = e−2t (x −x).5. the system is not globally asymptotically −1 1 1 and . ˙ 6. and the general solution of (∗) is π = Aer1 t + Ber2 t . The characteristic equation 1 1 is r 2 + (1/β − α)r + (1 − α/β) = 0. ˙ 6. and we get y = − 2 x + 2 sin t = ˙ 1 −A cos 2t + B sin 2t + sin t. is (x ∗ . 2).6 6.5. A straightforward calculation gives x − x = A 2e(1+ 2 )t − √ √ √ √ √ √ B 2e(1− 2 )t and then p = e−2t (x − x) = A 2e( 2−1)t − B 2e(− 2−1)t . with corresponding eigenvectors 0 −1 1 1 Aet + Be−t w + Be−t = .6.9). The initial conditions x(0) = y(0) = 0 yield B − 2 + 1 = 0 and −A = 0.5. then r1 and r2 are real and different. Let ˙ ˙ z = x + 5 and w = y − 2 (cf. example 6. and Peter Hammond 2008 .4 From the ﬁrst equation.5. D = 1.3 (ii) The equilibrium point of the linear system x = x + 2y. The eigenvalues of A are λ1 = 1 and λ2 = −1. © Arne Strøm. Atle Seierstad.CHAPTER 6 / DIFFERENTIAL EQUATIONS II: SECONDORDER EQUATIONS 53 The corresponding homogeneous equation. It follows that C = −1/2.
y ∗ ) is given by the equation system ax ∗ + 2y ∗ + α = 0. The second equation then yields ˙ 1 ¨ 2 (x 1 − a x) = 2x + 2 a(x − ax − α) + β ˙ ˙ ⇐⇒ x − 2a x + (a 2 − 4)x = 2β − αa ¨ ˙ (♦) The characteristic equation is r 2 − 2ar + a 2 − 4 = 0. Since the righthand side of the equation is a constant. which has the roots r1. so x = −e−t . the general solution of (♦) is x = Ae(a−2)t + Be(a+2)t + (2β − αa)/(a 2 − 4). The second equation is the system the becomes y = e−t y − y 2 . It follows that for t > 0 we have −t t −e−s ds ≥ e−1 t.7 6. i. y = 2 (x − ax − α). we look for a suitable constant u∗ . and β = −1. It follows that the homogeneous equation associated with (♦) has the general solution xH = Ae(a−2)t + Be(a+2)t . and the initial condition x(0) = −1 implies ˙ A = −1. A6.e.4(a) and (b) in the answer section of the book. i. and Peter Hammond 2008 .4 (a) See Figs.e. y = −˙ /z2 . Indeed. (c) With a = −1. Atle Seierstad. Thus. which is a Bernoulli equation.1 says that the equilibrium point 2 a is globally asymptotically stable if and only if 2a < 0 and a 2 − 4 > 0. and so do all the convergent solutions. With z = y 1−2 = y −1 = 1/y we get y = 1/z. the solutions with B = 0.7. 1 ˙ Then the equation y = 2 (x − ax − α) yields y = −Ae(a−2)t + Be(a+2)t + (2α − aβ)/(a 2 − 4). the equations show that any solution through a point on the yaxis below the origin will move straight downwards. this is just the stationary solution of the system in part (a). We get tr(A) = 2a and A = a 2 − 4. which is equivalent to a < −2. 2x ∗ + ay ∗ + β = 0 Easy calculations show that x ∗ = (2β − αa)/(a 2 − 4) and y ∗ = (2α − aβ)/(a 2 − 4). and ˙ z −˙ /z2 = e−t (1/z) − 1/z2 z ⇒ z + e−t z = 1 ˙ t s If we use formula (5.6. (b) The ﬁrst equation has the general solution x = Ae−t . y = −e−3t + 3. Knut Sydsæter.4 (a) From the ﬁrst equation.4. given by A = B = 0. Of course. The equilibrium point is globally asymptotically stable if and only if e(a−2)t and e(a+2)t both tend to 0 as t → ∞. the general solution of the system is x = Ae−3t + Bet + 2. which turns out to be u∗ = (2β −αa)/(a 2 −4). 6. The system has a single equilibrium point. An alternative way to check stability is to consider the trace and determinant of the coefﬁcient matrix a 2 A= .6.7. we get − and z = ee −t −1 a(ξ ) dξ = e−t − e−s + 0 t ee −t −e−s ds = ee −t e−e −t e−1 + 0 t e−e ds −s y= e−1 + t 0 e−e−s ds −s For all s ≥ 0 we have e−s ≤ 1. namely (0. b(t) = 1. (b) The equilibrium point (x ∗ . 3) along the line x + y = 5. so −e−s ≥ −1 and e−e ≥ e−1 . the general solution of equation (♦) itself is xH + u∗ . t0 = 1. y = −Ae−3t + Bet + 3.7) with a(t) = e−t . where u∗ is any particular solution of (♦).54 CHAPTER 6 / DIFFERENTIAL EQUATIONS II: SECONDORDER EQUATIONS 1 6. if and only if a < −2. and this happens if and only if both a − 2 and a + 2 are negative. It seems clear from the phase diagram that this is not a stable equilibrium. One such solution is then x = e−3t + 2. 0). 0 e © Arne Strøm. 0). α = −4. Hence. and z0 = 1. 3) if and only if B = 0. It is clear that this will converge to the equilibrium point (2. e−e ≤ e0 = 1 and y(t) ≤ 1/(e−1 + te−1 ) → 0 as t∞. and Theorem 6.2 = a ± 2. This solution moves towards (2. away from (0.
(K0 − s/δ)e−δ(1−α)t + s/δ 6. A6. with γ1 b = s and γ2 = −δ. but not the one the authors had in mind. 0) is an equilibrium point.8. and Peter Hammond 2008 . The initial condition x(0) = 1 then implies ˙ x(t) = e−t .9. fY gK = α(IK − SK )IY < 0 everywhere. As t → ∞.8. This is a separable equation and we get ˙ dy =− y e−2t dt ⇒ ln y = 1 −2t e +C 2 −2t −1)/2 1 The initial condition y(0) = 1 yields C = − 2 . K) = α I (Y. K) = I (Y.5 Let f (Y. c∗ ) with c∗ > 0 we have c∗ = f (k ∗ ) − δk ∗ and f (k ∗ ) = r + δ. c) is A(k.8.6. K) = α(IY − SY ) + IK < 0 = A(Y. c) = (0. c) = f (k) − δk − c. c∗ ) = cf (k ∗ ) < 0. 8/3). so (k ∗ . e−1/2 ). tr A(Y. c) = −c(r + δ − f (k)) ˙ The origin (k. Now. Finally. K) = IY gY gK IK and A(Y. y(t)) tends to (0. The matrix A in Theorem 6. with determinant A = −2. the solution K(t) = 1/(1−α) 1−α tends to (s/δ)1/(1−α) = K ∗ as t → ∞. the point 6.6 See the answer in the book for the stability question. c∗ ) = f (k ∗ ) − δ cf (k ∗ ) −1 0 = r cf (k ∗ ) −1 0 has determinant A(k ∗ . Then A(k ∗ . ˙ K must satisfy the Bernoulli equation K = sK α − δK. At an equilibrium point (k ∗ .6 The equation x = −x has the general solution x = Ae−t .2 is α(IY − SY ) α(IK − SK ) fY fK . Knut Sydsæter. This corresponds to the equation in Problem 1/(1−α) 5. It is a saddle point because the Jacobian at (4/3.2 Write the system as ˙ k = F (k.CHAPTER 6 / DIFFERENTIAL EQUATIONS II: SECONDORDER EQUATIONS 55 6.3 (a) (x0 . K) − S(Y. y0 ) = (4/3. Since δ(1 − α) > 0. c∗ ) is a saddle point. Atle Seierstad. c) = Fk Gk Fc Gc = f (k) − δ cf (k) −1 f (k) − r − δ c = G(k.9 6. K) = α(IY − SY )IK − α(IK − SK )IY = α(IY SK − IK SY ) > 0 by the assumptions in the problem. y 2 /2x 2 1 − y/x 2 −1 (b) See Fig. The general solution is therefore K(t) = Ce−δ(1−α)t + s/δ . and we get y(t) = e(e (x(t).8 6. 1−α The initial condition K(0) = K0 yields C = K0 − s/δ.9.4. . K) and g(Y. so an equilibrium point for the system must be globally asymptotically stable according to Olech’s theorem. 6. © Arne Strøm.7.9. 6.3 in the answer section of the book. The second equation becomes y = −e−2t y. 8/3) is y−x−2 x −2/3 4/3 A= = . K). The Jacobian matrix at (k.
u∗ = (2F −2E+(E−4F )t +F t 2 )e−t . the general solution of the associated homogeneous equation is xH = (A + Bt)et + Ce−t . We must therefore increase the degree of the polynomial factor and try ˙ with a quadratic polynomial instead. −3/2) = A2 = −4 3 tr(A1 ) = −4. A(9/4.2).2.4 (a) The equilibrium points are the solutions of the equation system (i) y 2 − x = 0. (E+(2F −E)t −F t 2 )e−t . Knut Sydsæter. Let u∗ = (Et + F t 2 )e−t . (r 3 −r 2 −r +1)÷(r −1) = (r 2 −1).4 in the answer section of the book. Atle Seierstad.1. −1 2y −1 3 The Jacobian at (x. It follows that (9/4. 3/2) = A1 = and −2x + 1/2. 3/2) is locally asymptotically stable. and the determinant of the system is sin t cos t = − sin2 t − cos2 t = −1.9. Since x = y 2 must be nonnegative. −3/2) is a saddle point.56 7 DIFFERENTIAL EQUATIONS III: HIGHERORDER EQUATIONS 6.9. y) = .1. These conclusions are conﬁrmed by the solution to part (b). A6. whereas (9/4. we get x = 9/4 and y = ±3/2. A2  = −15.3 The homogeneous equation x + x = 0 has the two linearly independent solutions u1 = sin t and ¨ ¨ u2 = cos t. ˙ ˙ Let x = C1 (t) sin t + C2 (t) cos t. t It follows that u∗ (t) = sin t sin t ˙ C2 (t) = − t cos t cos t dt − cos t dt is the general solution of the given equation t t (provided we include an arbitrary constant of integration in each of the two integrals). it might seem natural to try u∗ = (D + Et)e−t as a particular solution. so r 3 −r 2 −r +1 = (r −1)(r 2 −1) = (r −1)2 (r +1). so the only possibilities are ±1. −2y −4 −3 −1 −3 . According to the general method for ﬁnding linearly independent solutions to linear homogeneous equations with constant coefﬁcients. 7 Differential Equations III: HigherOrder Equations 7. A bit of tedious algebra gives u∗ = .) Cramer’s rule gives cos t − sin t cos t ˙ C1 (t) = . The two equations to determine C1 (t) and C2 (t) are ˙ ˙ C1 (t) sin t + C2 (t) cos t = 0 ˙ ˙ C1 (t) cos t − C2 (t) sin t = 1/t ˙ ˙ This is a linear equation system in the unknowns C1 (t) and C2 (t). which has the solutions x = (−1 ± 10)/4. and tr(A2 ) = 2. (See Section B. Note 4. (ii) 25/4 − y 2 − (x − 1/4)2 = 0 Substituting x for y 2 in (ii) leads to the quadratic equation x 2 + x/2 − 99/16 = 0. so A(9/4.1 7.2 7. and we see that both r = −1 and r = 1 are roots.. y) is A(x. The determinants and traces of these matrices are A1  = 15. Moreover. ¨ © Arne Strøm.7. u ∗ = (3E−6F +(6F −E)t −F t 2 )e−t .2 Integer roots of the characteristic equation r 3 −r 2 −r +1 = 0 must divide the constant term 1 (EMEA. To ﬁnd the solution of the equation x +x = 1/t we use the method of variation of parameters. 7. But that does not work because r = −1 is a root in the characteristic equation. Looking at the righthand side of the given nonhomogeneous equation.. and Peter Hammond 2008 . See Fig.
Requiring this solution to satisfy the initial conditions gives A = −1.1. (γ1 σ + γ3 )μ0 eμt 0 e−μτ K(τ )dτ = K − (γ1 κ + γ2 )K. and different if p 2 − 4q > 0 and q = 0. We then ﬁnd x2 from (a): x2 = x1 + x1 − x3 = ¨ ˙ ˙ −t − C e−2t + C e2t ... The eigenvalues of ˙ (ii) We write the system as x = Ax. (a) again. This equals 8te−t if F = 1 and E = 2. (b) x2 = x1 −x2 +x3 .4 7. If these conditions are satisﬁed.t.. and λ3 = 2.3 7. the general solution is x1 = C1 e−t + C2 e−2t + C3 e2t . Inserting this into (∗) . Differentiating ˙ ˙ ˙ (a) w. nonzero.2 The roots of the characteristic equation r 3 + 4r 2 + 5r + 2 = 0 are r1 = r2 = −1 and r3 = −2. The other two are the roots r and r of r 2 − pr + q = 0..1 (i) The system is: (a) x1 = −x1 +x2 +x3 .3 Differentiating the equation w.3. x1 + x2 + 2x3 = 0 x1 + x2 + x3 = 0 x 1 + x 2 + x3 = 0 .7 DIFFERENTIAL EQUATIONS III: HIGHERORDER EQUATIONS 57 .5)) .. Knut Sydsæter.3.. and so x = (t − 1)et + (t + 1)2 e−t .4. ¨ ˙ ˙ ˙ t and inserting from (c) gives x 1 + x1 − 2x1 = 2x3 = 2x1 + 2(x2 + x3 ) = 2x1 + 2(x1 + x1 ). Atle Seierstad. t gives (using the product rule and (4. which are all negative.) The eigenvectors associated with the eigenvalues are determined by the three systems x2 + x3 = 0 x1 + x3 = 0 . C1 e 2 3 ⎛ ⎞ ⎛ ⎞ −1 1 1 x1 ⎝ 1 −1 1 ⎠ and x = ⎝ x2 ⎠. C = 1. From (d) 1 we ﬁnd x3 = 2 (x1 + x1 − 2x1 ) = −C1 e−t + 2C3 e2t . t ˙ ˙ e−μτ K(τ )dτ + (γ1 σ + γ3 )μ0 eμt e−μt K(t) (∗) 7. ¨ ˙ . One root of the characteristic 3 − pr 2 + qr = 0 is r = 0. is a solution of the homogeneous equation for every value of D. 7. equation r 3 1 2 and they are real. and we ﬁnd them to be λ1 = −1. using .2.t. Differentiating once more w. ¨ ˙ and ﬁnally u ∗ − u∗ − u∗ + u∗ = (4E − 8F + 8F t)e−t . (These are the same as the solutions of the characteristic equation of the differential equation (e). See the remark above Theorem 6. This is no coincidence.. so global asymptotic stability also follows from Theorem 7. Why did we not include a term of the form De−t in the tentative solution u∗ ? The answer is that De−t .r.6..r. Since the characteristic ¨ ˙ polynomial is (r + 1)(r + 2)(r − 2). t and inserting from (b) and (c) gives (d) x1 + x1 − 2x1 = 2x3 .1. B = 1. so the general solution of the given equation is x = xH + u∗ = (A + Bt)et + (C + 2t + t 2 )e−t . ¨ ˙ 7. Thus the differential equation for x1 is (e) x 1 + x1 − 4x1 − 4x1 = 0. x1 + x2 + 3x3 = 0 −3x1 + x2 + x3 = 0 x1 − 3x2 + x3 = 0 x 1 + x2 − x 3 = 0 © Arne Strøm. ¨ ˙ yields the equation K − p K + q K = 0 given in the answer in the book. ¨ K = (γ1 κ + γ2 )K + (γ1 σ + γ3 )μ0 μeμt 0 t ˙ ˙ ¨ From the given equation..t. so D would not appear in u ∗ − u∗ − u∗ +u∗ . (c) x3 = x1 +x2 +x3 .r. it follows from the theory in this section that the general solution of the differential equation is of the form K(t) = C1 er1 t + C2 er2 t + C3 er3 t . 1 −1 − λ 1 1 1 1−λ λ2 = −2. where A = x3 1 1 1 −1 − λ 1 1 A are the solutions of the equation = 0..1 concerning secondorder systems... and Peter Hammond 2008 .
0) is a vector (x1 (t). We claim that the function L(x. y0 = (bk − hε)/(ab + δε). and Peter Hammond 2008 . P(0. x3 (0)) is the ith standard unit vector ei . x2 (0). and Cramer’s rule gives the solution as x0 = (ah + kδ)/(ab + δε).4 (a) Equation (∗) is separable. so it has a unique solution. −1 ⎛ ⎞ 1 v2 = ⎝ −1 ⎠ . Atle Seierstad. 0).58 7 DIFFERENTIAL EQUATIONS III: HIGHERORDER EQUATIONS The following vectors are solutions of these systems: ⎞ 1 v1 = ⎝ 1 ⎠ . Therefore. f1 (x) g2 (y) the function H (x. 0) = (d/dt)P(t. 7. y) = b(x − x0 ln x) + a(y − y0 ln y) − b(x0 − x0 ln x0 ) + a(y0 − y0 ln y0 ) © Arne Strøm.5. where x0 = h/b and y0 = k/a. 0 ⎛ ⎛ ⎞ 1 v3 = ⎝ 1 ⎠ 2 The solution of the given system of differential equations is then ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ x1 1 1 1 ⎝ x2 ⎠ = C1 e−t ⎝ 1 ⎠ + C2 e−2t ⎝ −1 ⎠ + C3 e2t ⎝ 1 ⎠ x3 −1 0 2 We see that we have arrived at the same solutions as above. y0 ) with x0 = 0 and y0 = 0 is an equilibrium point if and only if bx0 − δy0 = h εx0 + ay0 = k (∗) The determinant of this system is ab + δε = 0.5 7. 0) = ⎝ −t 3e 1 −t 3e 1 + 2 e−2t + 1 e2t 6 1 − 2 e−2t + 1 e2t 6 1 −t 3e 1 −t 3e 1 − 2 e−2t + 1 e2t 6 1 + 2 e−2t + 1 e2t 6 1 − 1 e−t + 3 e2t 3 1 −t 3e ⎞ ⎟ 1 − 1 e−t + 3 e2t ⎠ 3 + 2 e2t 3 − 1 e−t + 1 e2t 3 3 − 1 e−t + 1 e2t 3 3 The ith column of P(t. Knut Sydsæter. (iii) The resolvent. (b) H (x.5. It is not hard to verify that AP(t. as given in the problem. In particular. is ⎛1 ⎜ P(t. y) is constant along each solution curve for the system. x3 (t)) of particular solutions of the system such that (x1 (0). ˙ y = y(−h + bx − δy) ˙ It is clear that a point (x0 . 7. with t0 = 0. x2 (t). so f2 (y) g1 (x) dx = dy + C for some constant C. y) = k − ay h k bx − h dx − dy = b− dx − − a dy = x y x y bx − h ln x − (k ln y − ay) + C = b(x − x0 ln x) + a(y − y0 ln y) + C.5 We can write the system as x = x(k − ay − εx). 0) = I3 .
Solving this equation for the ﬁrst variable yields z = x + ϕ(xy). v2 )). We have proved that L is a strong Liapunov function for the system. or f (x) = ln v1 + ϕ(v2 /v1 ). and has an inverse f −1 that is also strictly increasing. y) = (x0 . Lyy = ay0 /y 2 > 0. Thus ϕ(x) = −x + x 2 for all x. y) = a(1 − y0 /y) are both 0 at (x0 .3) are both separable. f (x)−ln v1 ) = 0. with the solutions v2 /v1 = C1 . v2 )) is homothetic. The latter equation gives z = x +C1 . dy/dx = y 2 /x 2 and dz/dx = z2 /x 2 . y0 ) = 0. The general solution is therefore (1/x − 1/y. Hence. where ϕ is an arbitrary differentiable function.3) are dv2 /dv1 = v2 /v1 and dx/dv1 = xε(x)/v1 .1 (a) By integration. these solutions are just two ways of writing the same thing.7.7. The functions ϕ and ψ are related by the equation ψ(u) = ϕ(−2u) + 3u. Then g is homogeneous of degree 1. and Lxy = 0. Knut Sydsæter. The general solution of (∗) is given by (z − x. y0 ). 1/x − 1/z) x . y0 ). z = 4 x 4 + 2 x 2 y 2 − ex y + ϕ(y).7. © Arne Strøm. Actually. (c) The equations in (7.2) leads to the solution z = 3x + ϕ(y − 2x). and hence z = 1 − xϕ(1/x − 1/y) differentiable function.7 DIFFERENTIAL EQUATIONS III: HIGHERORDER EQUATIONS 59 is a strong Liapunov function for the system.3 (a) The equations in (7. so L(x. v2 ) = eln v1 +ϕ(v2 /v1 ) = v1 eϕ(v2 /v1 ) . and we see that x = f −1 (ln(g(v1 . y) = x + ϕ(xy) = x − xy + x 2 y 2 . If we use the recipe with y instead of x as the independent variable in equations (7. f (x) − ln v1 = C2 . y) is strictly convex and has a unique minimum at (x0 . where ϕ is an arbitrary = 0. It looks as if x has a special role here. where ϕ(y) is an arbitrary function and plays the role of a constant of integration when we integrate with respect to x. We used the equations k = εx0 + ay0 and h = −δy0 + bx0 from (∗).7. Atle Seierstad. The solutions are −1/y = −1/x + C1 . 7. y0 ) as its minimum point. 7. Since f (x) = 1/xε(x) > 0. where y seems to be singled out. we are led to z = 3y/2 + ψ(x − y/2).7.7 1 1 7. 7. Moreover. First note that L(x0 . y0 ) is locally asymptotically stable.3). The composition F of the two increasing functions f −1 and ln is increasing. (b) The recipe for solving equations of the form (7. but that is an illusion. where f (x) = (1/xε(x)) dx. −1/z = −1/x + C2 . where ϕ is an arbitrary differentiable function.7. y) = b(1 − x0 /x) and Ly (x. and Peter Hammond 2008 . or 1/z = 1/x − ϕ(1/x − 1/y). (b) The condition f (x. It follows that x = F (g(v1 . Finally. Lx (x.7. f is strictly increasing. The ﬁrst equation is separable with solution xy = C2 . xy) = 0. so (x0 . Deﬁne g(v1 . and hence f (x. x = f −1 (ln v1 + ϕ(v2 /v1 )).3) are dy/dx = −y/x and dz/dx = 1.7. with (x0 . and so z − x = C1 . ˙ ˙ ˙ L = b 1 − x0 /x x + a 1 − y0 /y y = b(k − ay − εx)(x − x0 ) + a(−h + bx − δy)(y − y0 ) = b(εx0 + ay0 − ay − εx)(x − x0 ) + a(δy0 − bx0 + bx − δy)(y − y0 ) = −εb(x − x0 )2 − aδ(y − y0 )2 which is negative for (x.7 The equations in (7. The general solution is (v2 /v1 . 1) = x 2 implies that x + ϕ(x) = x 2 . y0 ) and Lxx = bx0 /x 2 > 0.
x.6 Let F (t. x. x.1. and so x = 4 t + C ln t + A. x) is convex with respect to (x. x) = −ex−ax we get Fx = aex−ax and Fx = −ex−ax . √ x ˙ d ˙ .3 (a) With F (t.2.2 8. x) = 1 + x 2 . 8. x.2. and the Euler equation ˙ ˙ ˙ ˙ becomes d ˙ 2x + 2et − (2x) = 0 ⇐⇒ 2x + 2et − 2x = 0 ⇐⇒ x − x = et ¨ ¨ dt ˙ ˙ ˙ (b) With F (t. so Fx = [2(x − x) + 2x]e−at and Fx = −2(x − x)e−at . and Peter Hammond 2008 x ˙ √ 1 + x2 ˙ = 0. where C and A are constants. x) = [(x − x)2 + x 2 ]e−at . The ˙ ˙ ˙ Euler equation becomes [2(x − x) + 2x]e−at + ˙ d [2(x − x)e−at ] = 0 ˙ dt ˙ ¨ ˙ ⇐⇒ [2(x − x) + 2x]e−at + 2(x − x)e−at − 2a(x − x)e−at = 0 ˙ ⇐⇒ 2(x − x) + 2x + 2(x − x) − 2a(x − x) = 0 ˙ ˙ ¨ ˙ ⇐⇒ x − a x + (a − 2)x = 0 ¨ ˙ ˙ ˙ (d) With F (t. We have Fxx = 2 > 0. by ˙ ˙ Theorem 8. x. and the Euler equation ˙ ˙ ˙ ˙ becomes 2t + 3x − ˙ d (3x + 2t x) = 0 ⇐⇒ 2t + 3x − 3x − 2x − 2t x = 0 ⇐⇒ t x + x = t ˙ ˙ ˙ ˙ ¨ ¨ ˙ dt It is worth noting that this is an exact equation for x because it says that (d/dt)(t x) = t. x = t + 1 is the optimal solution of the problem. 1]. ˙ ˙ ˙ ˙ ˙ ˙ 8. Fx x = 2t. √ x ˙ 1 + x2 ˙ =C (constant) . and the boundary conditions x(0) = 1 and x(1) = 2 yield ¨ A = B = 1. ˙ ˙ 1 2 1 1 2 t x = 2 t + C.2) is 2x + 2t x − ˙ d (2tx + 2x) = 0 ⇐⇒ 2x + 2t x − (2x + 2t x + 2x) = 0 ⇐⇒ x = 0 ˙ ˙ ˙ ¨ ¨ dt The general solution of x = 0 is x = At + B. Then Fx = 0 and Fx = √ ˙ ˙ 2 dt ˙ 1+x ˙ therefore reduces to d dt This implies © Arne Strøm. 1].e. and Fx x = 2 > 0.60 8 CALCULUS OF VARIATIONS 8 Calculus of Variations 8. Hence. Atle Seierstad. Hence. x. Thus the only admissible function that satisﬁes the Euler equation is x = t + 1. x) = x 2 + 2tx x + x 2 we get Fx = 2x + 2t x and Fx = 2tx + 2x. it follows that F (t. x.2. Fx − Fx = 0. The Euler equation is ˙ ˙ ˙ aex−ax + d x−ax ˙ ˙ e˙ = 0 ⇐⇒ aex−ax + ex−ax (x − a x) = 0 ⇐⇒ x − a x + a = 0 ¨ ˙ ¨ ˙ dt ˙ ˙ (c) Here F (t.4 With F (t. x) = 2tx + 3x x + t x 2 we get Fx = 2t + 3x and Fx = 3x + 2t x. x) = x 2 + x 2 + 2xet we get Fx = 2x + 2et and Fx = 2x. Knut Sydsæter. and since Fxx Fx x − (Fx x )2 = 4 − 4t 2 ≥ 0 for ˙ ˙˙ ˙˙ ˙ all t in [0. x) as long as t ∈ [0. which implies x = 2 t + C/t. i. The Euler equation.2.3. so the Euler equation ˙ ˙ (8.
and Theorem 8. at least among the admissible C 2 functions. the Euler equation reduces to d −U (c − xert )ert = 0.1 in ˙˙ Section 8. Of course. Knut Sydsæter. namely the straight line through the points (t0 .3 8. Atle Seierstad. 1 + x2 ˙ so x = C1 ˙ C (= √ ) 1 − C2 Since x is a constant. Thus the graph of x is a straight line. so U (c − xert )ert = K (a constant) ¯ ˙ ¯ ˙ dt d (Evaluating the derivative ¯ ˙ −U (c − xert )ert above leads to the equation dt −U (c − xert )(−xert − r xert )ert − rU (c − xert )ert = 0 ¨ ˙ ¯ ˙ ¯ ˙ This can be simpliﬁed to x + rx = ¨ ˙ which will most likely be harder to solve. x). x is a linear function of t: ˙ x = C1 t + C2 . this is precisely what we would expect: the shortest curve between two points is the straight line segment between them. and we get exp(−v c + v xert ) = Ke−rt ¯ ˙ −v c + v xert = ln K − rt ¯ ˙ xert = C − rt/v ˙ x = (C − rt/v)e ˙ © Arne Strøm.2 (a) The objective function is F (t. and Peter Hammond 2008 (C = c + (ln K)/v) ¯ −rt . then U (c) = e−vc .3.8 CALCULUS OF VARIATIONS 61 x2 ˙ = C2. x) = U (c − xert ). and ˙ ¯ ˙ ∂F = 0. ∂x As ∂F /∂x = 0. ∂x ˙ If U (c) = −e−vc /v.3 shows that the function we have found really does give the minimum.) (b) It follows from part (a) that ¯ ˙ U (c − xert ) = Ke−rt rU (c − xert ) −rt ¯ ˙ e ¯ ˙ U (c − xert ) ∂F ¯ ˙ = −U (c − xert )ert . x0 ) and (t1 .3. The function x is given by the equation x(t) = x0 + x1 − x 0 (t − t0 ) t1 − t0 ˙ ˙ Note also that Fx x = 1/(1 + x 2 )3/2 > 0. x. x1 ). 8. so F is convex with respect to (x.
8. To ﬁnd a particular integral of (∗) we try with a function u∗ = Ceαt . A + (B + T /v)e−rT = x(T ) = 0 U is concave because U (c) = −ve−vc < 0. and Peter Hammond 2008 .62 8 CALCULUS OF VARIATIONS Integration by parts yields x = x(t) = (C − rt/v)e−rt dt = A + (B + t/v)e−rt where B = (1 − vC)/rv.1 shows that the solution we have found is optimal. y.3. and the general solution of (∗) is ¯ y = yH + u∗ = Aet/σ + Btet/σ + z l0 ¯ eαt 1 − ασ (If ασ = 1. We get u∗ − ¨ 2 ∗ 1 2α αt 1 (1 − ασ )2 αt u + 2 u∗ = α 2 Ceαt − ˙ Ce + 2 Ceαt = C e σ σ σ σ σ2 It follows that u∗ is a solution of (∗) if and only if C = zl0 /(1 − ασ ). Hence. K. Atle Seierstad.4.3. Finally.) 8. then (∗) is homogeneous and the general solution is yH .1 With F (t. ˙ The Euler equation is e−t/4 ˙ ¨ d 2e−t/4 e−t/4 e−t/4 (2K − K) 2 1 − = 0 ⇐⇒ − − −e−t/4 =0 ˙ ˙ ˙ ˙ ˙ dt 2K − K 2K − K 2K − K 4(2K − K) (2K − K)2 e−t/4 ˙ ˙ ˙ ¨ 8(2K − K) − (2K − K) − 4(2K − K) = 0 ⇐⇒ ˙ 4(2K − K)2 ¨ ˙ ⇐⇒ 4K − 15K + 14K = 0 © Arne Strøm. Knut Sydsæter. the constants A and B are determined by the equations A + B = x(0) = x0 .4 ˙ ˙ ˙ ˙ 8. Theorem 8. K) = e−t/4 ln(2K − K) we get FK = 2e−t/4 /(2K − K) and FK = −e−t/4 /(2K − K). so the corresponding homogeneous differential equation has the general solution yH = Aet/σ + Btet/σ . y) = ln y−σ y−¯ l(t) we get Fy = 1/[y−σ y−¯ l(t)] and Fy = −σ/[y−σ y−¯ l(t)]. ˙ ˙ z ˙ z ˙ z ˙ The Euler equation is then 1 d −σ 1 σ (y − σ y − zl(t)) ˙ ¨ ¯˙ − = 0 ⇐⇒ − =0 y − σ y − zl(t) dt y − σ y − zl(t) ˙ ¯ ˙ ¯ y − σ y − zl(t) ˙ ¯ [y − σ y − zl(t)]2 ˙ ¯ 1 2 z ¯ ˙ ⇐⇒ y − σ y − zl(t) − σ y + σ 2 y + σ zl(t) = 0 ⇐⇒ y − y + 2 y = 2 l(t) − σ l(t) ˙ ¯ ˙ ˙ ¨ ¯˙ ¨ σ σ σ (b) With l(t) = l0 eαt the Euler equation becomes y− ¨ 2 1 z(1 − ασ )l0 αt ¯ y + 2y = ˙ e σ σ σ2 (∗) The characteristic equation r 2 − (2/σ )r + 1/σ 2 = 0 has the double root r1 = r2 = 1/σ .4 (a) With F (t.
K.3 With F (t.4. K(T ) = Ae2T + Be7T /4 = KT By means of Cramer’s rule or by other methods you will ﬁnd that A= 8. Hence the general solution of the Euler equation is K = Ae2t + Be7t/4 . x.4. ˙ ˙ (b) With T = 10 and S = 20 we get B = −5/2 + 5 − 20/(e − 1) = 25/2(e − 1). x) is concave with respect to (x. Then Fx = te−t/10 /100 and Fx = −2xe−t/10 . and the ˙ ˙ ˙ d 2 t −t/10 t −t/10 −2xe−t/10 = 0 ⇐⇒ ˙ e xe−t/10 = 0 ˙ − + 2xe−t/10 − ¨ e dt 100 10 100 1 1 ⇐⇒ x − ¨ x=− ˙ t 10 200 The general solution of the corresponding homogeneous equation is xH = A + Bet/10 (∗) To ﬁnd a particular solution u∗ of (∗) we try with u = P t 2 + Qt. and Peter Hammond 2008 . Knut Sydsæter. e2T − e7T /4 B= e2T K0 − KT e2T − e7T /4 − x 2 e−t/10 . which has the roots r1 = 2 and r2 = 7/4. The ˙ Euler equation is UC (fK − δ) + d ˙ U = 0 ⇐⇒ UC (fK − δ) + UCC C + UCt = 0 dt C © Arne Strøm. x) = ˙ Euler equation becomes 1 100 tx KT − e7T /4 K0 . this is indeed an optimal solution. The boundary conditions yield the equations K(0) = A + B = K0 . x.8 CALCULUS OF VARIATIONS 63 The characteristic equation of the Euler equation is 4r 2 − 15r + 14 = 0. x). Atle Seierstad. so the general solution of the Euler equation is x = A + Bet/10 + 1 2 1 t + t 40 2 The boundary conditions x(0) = 0 and x(T ) = S yield the equations A + B = 0. K) = U (C. and ˙ ¨ if we insert this into (∗) we get Q 1 P =− t 2P − t − 5 10 200 This yields P = 5/200 = 1/40 and Q = 20P = 1/2. t) we get FK = UC CK = UC (fk − δ) and FK = UC · (−1) = −UC .2 (a) Let F (t. Then u = 2P t + Q and u = 2P . and the optimal solution is 1 2 1 25(et/10 − 1) + t + t x= 2(e − 1) 40 2 ˙ 8. with the solution A = −B = A + BeT /10 + T 2 /40 + T /2 = S T 2 /40 + T /2 − S eT /10 − 1 Since F (t.
we get b (x) = 2αx +β. ˙ Rearranging the terms we can write the Euler equation as p− ¨ βA + 2αAC − C A2 − A/α p= 2 B 2αB 2 and it is easy to see that we get the answer given in the book. p. © Arne Strøm. The integrand F (t.5. Atle Seierstad. and therefore ˙ U + UC (fK − δ) 1 UCt C = − Ct =− + fK − δ C CUCC ω UC ˇ where ω = CUCC /UC is the elasticity of marginal utility with respect to consumption. p) = Ap +B p +C. p) − b(D(p. ˇ 8. 8.4 (a) F (t. Knut Sydsæter.4. and Peter Hammond 2008 . so the Euler equation is ˙ ˙ Fp − d d (Fp ) = 0 ⇐⇒ D + p − b (D) Dp − [p − b (D)]Dp = 0 ˙ ˙ dt dt ˙ ˙ (b) With b(x) = αx 2 +βx +γ and x = D(p.2 (a) The Euler equation is (−2x − 10x)e−t − ˙ d [(−2x − 2x)e−t ] = 0 ⇐⇒ e−t (−2x − 10x + 2x + 2x − 2x − 2x) = 0 ˙ ˙ ¨ ˙ ˙ dt ⇐⇒ x + x − 6x = 0 ¨ ˙ The general solution of this equation is x = Ae3t +Be−2t . p) = pD(p. p)) yields Fp = D + pDp − b (D)Dp = D + [p − b (D)]Dp ˙ ˙ ˙ and Fp = [p − b (D)]Dp .5 8.5. is concave with respect to (x. we get the transversality condition Fx ˙ t=1 = 0. x(1) = Ae3 + Be−2 = 1 which have the solution A = −B = 1/(e3 − e−2 ). so the optimal solution is x ≡ 0. and the boundary conditions yield the equations x(0) = A + B = 0. and ∂D/∂ p = B.1). this yields A = B = 0. Insertion into the Euler equation gives ˙ Ap + B p + C + [p − 2α(Ap + B p + C) − β]A − ˙ ˙ d dt p − 2α(Ap + B p + C) − β B = 0 ˙ ˙ ˙ ˙ ¨ which yields Ap + B p + C + Ap − 2αA2 p − 2αAB p − 2αAC − βA − B p + 2αAB p + 2αB 2 p = 0. x. which implies x(1) + x(1) = 0 ⇐⇒ 3Ae3 − 2Be−2 + Ae3 + Be−2 = 0 ⇐⇒ 4Ae3 − Be−2 = 0 ˙ Together with the equation A + B = x(0) = 0.64 8 CALCULUS OF VARIATIONS ˙ The Euler equation implies C = −[UCt + UC (fK − δ)]/UCC . x) = (10 − x 2 − 2x x − 5x 2 )e−t ˙ ˙ ˙ 2 − 2x x − 5x 2 = −(x − x)2 − 4x 2 ). x) (look at the Hessian or note that −x ˙ ˙ ˙ ˙ so the solution we have found is optimal (Theorem 8. (b) (i) With x(1) free. ∂D/∂p = A.
3 reduces to A − r A = (ρ − r)/b. 9 Control Theory: Basic Techniques 9.4 The Hamiltonian H (t. It remains to determine x ∗ (t). x. x) = g (x) = g (A) > 0. x)e−rt ] = 0 and Cx (5.2. we get z − rz = (ρ − r)/b. The Hamiltonian H (t. Putting z = A.5. x ∗ (0) = 0. From (i) 1 1 1 1 ˙ we have u∗ (t) = − 2 p(t) = 2 (1−t).4 With U (C) = a − e−bC . With x ∗ (0) = 0. p(10) = 0.3 We transform the problem to a maximization problem by maximizing 0 [−x(t) − u(t)2 ] dt. we get −A(4e3 + e−2 ) ≤ 0.2 1 9. u∗ (t). x. x ∗ (t) = −u∗ (t) = 2 t − 2 . Atle Seierstad. x) = K must be constant. This is a linear ˙ ﬁrstorder differential equation with constant coefﬁcients. u) = g(u) it ˙ ˙ ˙ means that g (x) must be constant. p(t)) = −4u∗ (t) + p(t) = 0. ˙ © Arne Strøm. the following conditions are sufﬁcient for optimality (using (9. p(t) = t + A for some constant A. ˙ (iii) x ∗ (t) = u∗ (t). ˙ Since x(t) = Ae3t − Ae−2t . u). the following conditions are sufﬁcient for optimality: (i) Hu (t. u). These constants are determined by the boundary conditions A(0) = A0 and A(T ) = At . ˙ From (ii). u. x ∗ (t). so according to Theorem 9. u∗ (t).9 CONTROL THEORY: BASIC TECHNIQUES 65 =0 (ii) With the terminal condition x(1) ≥ 2 the transversality condition is Fx ˙ if x(1) > 2.2. ˙ (iii) x ∗ (t) = −u∗ (t). so A = 300 and x = 300t. We must have 5A ≥ 1500. and t=1 ≤ 0. so according to Theorem 9. Thus planting will take place at the constant rate of 300 hectares per year. x ∗ (0) = 0. Since Cx (t. Therefore A(e3 − e−2 ) = 2 and x = e3 − e−2 8.5.1 (using (9.5. ∗ (t) = 1 t 2 − 1 t. Then A = z dt = Kert − (ρ − r)t/br + L. so x = At. x ∗ (t). Since g is strictly increasing. x 4 2 9. x) = Ke−rt . x(5)) ≥ 0 (= 0 if x(5) > 1500). 2(e3t − e−2 ) . With C(t. 1 1 1 and so x ∗ (t) = 4 t 2 − 2 t + B. Knut Sydsæter. p) = 1−4x−2u2 +pu is concave in (x.1. and p(1) = 0 gives A = −1 and so p(t) = t − 1.7)): (i) Hu (t. u∗ (t). where K is a constant. x must also be a constant. (ii) p(t) = −Hx (t. where K = P /r and L are constants. p(t)) = 1. It ˙ ˙ ˙ follows that Cx (t.2.2. we get B = 0. In our case this reduces to −x(1) − x(1) ≤ 0. p(t)) = −2u∗ (t) − p(t) = 0. with Fx ˙ t=1 and −x(1) − x(1) = 0 if x(1) > 2 ˙ −A(4e3 + e−2 ) = 0 if x(1) = A(e3 − e−2 ) > 2 From the last condition it is clear that we cannot have A(e3 − e−2 ) > 2. p(1) = 0. x ∗ (t). p(t)) = 4. So the optimal solution is u∗ (t) = 2 (1 − t). we get U (C) = be−bC and U (C) = −b2 e−bC . p) = −x − u2 − pu is concave in (x. From (iii). (ii) p(t) = −Hx (t.7). ˙ ˙ 8. and the solution is z = P ert − (ρ − r)/br. ˙ ˙ (b) It follows from part (a) that if r = 0 then Cx (t. and that is impossible. x ∗ (t).5 (a) The conditions are −(d/dt)[Cx (t. with p(t) = t − 1.2. so equation (∗) in Example ¨ ˙ ˙ 8. u∗ (t).2. the transversality condition shows ˙ ˙ ˙ that x(5) = 1500. u. for then we must have A = 0. and Peter Hammond 2008 .
If ˙ x ∗ (1) > 1. x ∗ (t). Note that Hu = −2u + p(t) = 1 −2u− 2 (T 2 −t 2 ) < 0 for all t < T .4 9. 1]. p) = x −u2 +p(x +u) is concave in (x. Inserting these results into (iv) and ˙ ¨ − 0. u∗ (t) = x ∗ (t) = 2 A(et + e−t ). Furthermore. x ∗ (0) = 0. with general solution x ∗ (t) = Bet + 2 et ˙ e−t (eT −t − 1) dt = 1 1 1 1 1 1 1 1 Bet − 4 eT −t + 2 . but this time u∗ (t) is the value of u that maximizes H (t. (The value of A here is twice the value of A given in the book.1p = −1 + 0. we get x 2 9. ˙ ˙ ¨ ˙ From (iii) we get p = 2K + 0. and this nonpositive number is an interior point of [−1.) 1 9. the following conditions are sufﬁcient for optimality: (i) (ii) (iii) (iv) Hu (t.4. we have x ∗ (t) = 2 A(et − e−t ). but that would give x ∗ (t) = 0 for all t. Note that 1 Hu = 0 when u = − 4 (T 2 − t 2 ). u. (ii) p(t) − 0. p(t) = 4t + A for some constant A and p(10) = 40 + A = 0. 1] provided © Arne Strøm. so A = −40 and p(t) = 4t − 40. Atle Seierstad. u∗ (t).6 (b) Conditions (7) and (5) reduce to 1 (i) I ∗ (t) = 2 p(t). u∗ (t). From (i) we have u∗ (t) = 2 (eT −t − 1). p(t)) = −2u∗ (t) + p(t) = 0. p(t)) for u ∈ [−1.2. The general solution is p(t) = Ae−t −1. x ∗ (t). p(t) = −Hx (t. p(t)) = 2x ∗ (t).2. p(t)) = −2u∗ (t) + p(t) = 0. ˙ p(1) ≥ 0. Thus (K ∗ (t). u) = 1 − x 2 − u2 + pu is concave in (x. simplifying yields K 1 Moreover. u. we have p(t) = − 2 (T 2 − t 2 ). Knut Sydsæter. x ∗ (t). x ∗ (t) = 4 eT +t − 4 eT −t − 2 et + 2 . x. so according to Theorem 9. and with x ∗ (0) = 0.4. (ii) p(t) = −Hx (t. Therefore we must have x ∗ (1) = 1.2K. x ∗ (t) = x ∗ (t) + 2 (eT −t − 1).1p(t) = −1 + 0. From (iii). x ∗ (t) = u∗ (t) = t − 10. It follows that p(t) = Aet + Be−t for ˙ ¨ ˙ 1 1 ˙ suitable constants A and B.2. 9. and then x ∗ (t) ≡ x0 . x ∗ (1) ≥ 1.2 The Hamiltonian H (t. ˙ From (ii) we get the linear differential equation p(t)+p(t) = −1.04K = −0. and since x ∗ (0) = 0 gives 1 1 B = A. u∗ (t).5 The Hamiltonian H (t. ˙ ∗ (t) = 1 t 2 − 10t. x ∗ (0) = 0. Finally. and 2 9. ˙ 1 and p(T ) = 0 gives A = eT . p.66 9 CONTROL THEORY: BASIC TECHNIQUES From (ii).1 (using (9. we must have p(1) = 0.1K. ˙ (iv) p − 0.2.2.1K ∗ (t). so p(t) = eT −t − 1. which contradicts x ∗ (1) ≥ 1. p(T ) = 0.1K ∗ (t) = 2 p(t) − 0. x ∗ (t) = u∗ (t).2 (using (9.3 (a) As in Example 9. so according to Theorem 9. and therefore A = 0. the optimal u∗ (t) must maximize the concave function H (t. 1 1 from (iii). and then p = 2K + 0.7). x. which gives A = 2/(e − e−1 ) = 2e/(e2 − 1).2. and Peter Hammond 2008 .5.06K. x ∗ (t). p(t)) must satisfy ˙ (iii) K = 1 p − 0. the following conditions are sufﬁcient for optimality: (i) Hu (t. and p(t) = A(et + e−t ). Thus.4. Thus the optimal choice of u must be u∗ (t) ≡ 0. x ∗ (t) = 2 p = 2 (Aet − Be−t ). 1]. ˙ From p(t) = 2x ∗ (t) we get p(t) = 2x ∗ (t) = 2u∗ (t) = p(t). x ∗ (t). and p(1) = 0 if x ∗ (1) > 1. So for each t in [0. The condition x ∗ (0) = 0 gives B = 4 eT − 2 . u). K ∗ (t) = I ∗ (t) − 0.2K.1. u). T ]. p(t)) = 1 − tx ∗ (t) − u2 + p(t)u for u ∈ [0. ˙ (iii) x ∗ (t) = x ∗ (t) + u∗ (t). p(t)) = −1 − p(t). x ∗ (t). 1 (b) Also in this case Hu = −2u + p(t) = −2u − 2 (T 2 − t 2 ) and Huu = −2. u.06K ∗ (t).7). u∗ (t). 1 From (i) we have u∗ (t) = 4 (4t − 40) = t − 10.
Here x ∗ (0) = 0 yields C = 0. ˙ From (ii) we get p(t) = A − t for some constant A. and then x ∗ (t) = −5. To check the maximum condition (9.4.4. if u∗ (t) > 0. equivalent to Hu (t. 10) we have p(t) > 0. u.) For the rest see the answer in the book. With x ∗ (0) = 10. and thus u∗ (t) = 1. and the solution is as given in the book. x ∗ (5) ≥ 0. 5]. u∗ (t) = 1 in [0.2. x ∗ (t). (Choosing u∗ (t) = 1 cannot be optimal because Hu is negative for u = 1. then (Hu )∗ = −2bu∗ (t)+p(t) = 0. u∗ (t). Thus u∗ (t) = a(2t − T )/4b + B/T and x ∗ (t) = at (t − T )/4b + Bt/T Note that u∗ (t) is increasing in t. we conclude that t ∗ = 2. ˙ p(5) ≥ 0. Then (i) is ¯ ¯ ∗ (t). x. we conclude that for some t ∗ in x (0. and then x ∗ (t) = −u∗ (t) with x ∗ (0) = 10 ˙ that x 2 ¯ ¯ gives x ∗ (t) = 5p(e0.1t − pu is concave in (x. Moreover. Hence p > 0 and from (iii) (and x ∗ (5) ≥ 0) we conclude ¯ ∗ (5) = 0. t ∗ ]. p) = −(ax + bu2 ) + pu is concave in (x. and u∗ (0) = B/T − aT /4b ≥ 0 if and only if B ≥ aT 2 /4b. and (i) implies that u∗ (t) = 1 if p(t) > 0 and u∗ (t) = 0 if p(t) < 0. t ∗ ] and u∗ (t) = 0 for t in (t ∗ .4. The condition x ∗ (5) = 0 gives p = 3/(e0. 9. From (v) we have u∗ (t) = 5 − 1 pe0. and p(5) = 0 if x ∗ (5) > 0. then (v) implies that u∗ (t) ≡ 5. (ii) p = −1. 9.7 (b) The Hamiltonian H (t. x ∗ (10) = 2. Then u∗ (t) = p(t)/2b = (1/2b)(at + A).5) in Theorem 9. ¯ If p = 0. T ]. ¯ ˙ which gives x ∗ (5) = −15. we conclude that t ∗ = x1 − x0 . u) = x + pu is concave in (x. ˙ (iii) x ∗ (t) = u∗ (t). x ∗ (t) = −u∗ (t). x. conditions (i)–(iv) are satisﬁed by the pair (x 9. and p(t) < 0 with u∗ (t) = 0 for t in (t ∗ . and the solution is as given in the book.1t = p.5 − 1). p(t) = −Hx (t. the following conditions are sufﬁcient for optimality: (i) u = u∗ (t) maximizes p(t)u for u ∈ [0. a contradiction. If t ≤ T 2 − 4. Since p(t) is strictly decreasing. So in this case we have found an optimal solution. for t in [0. 10]. p. 10].9 CONTROL THEORY: BASIC TECHNIQUES 67 √ √ 1 − 4 (T 2 − t 2 ) > −1. u. i. 1]. T ]. Atle Seierstad. because it is easy to check that all the conditions in Theorem © Arne Strøm. We see that x ∗ (t) = t for t in [0.1.1t − 1) − 5t + 10. Now u∗ (t) ≡ 0 and u∗ (t) ≡ 1 are both impossible because x ∗ (t) = u∗ (t) and ˙ ∗ (0) = 0 contradict x ∗ (10) = 2. Suppose u∗ (t) > 0 for all t in [0.4 (a) Since H (t. u. (b) As in (a) we ﬁnd that for some t ∗ in (0.e. t ∗ ] and x ∗ (t) = t ∗ for t in (t ∗ . we have found an optimal solution. and if u∗ (t) = 0. and Peter Hammond 2008 . T ). x ˙ 2 and x ∗ (T ) = B gives A = 2bB/T − aT /2.6 The Hamiltonian H (t. x ∗ (0) = 10. Then x ∗ (t) = t + x0 for t in [0. u). p(t)) = 0. At t ∗ we have p(t ∗ ) = 0. so x ∗ (t) = (1/2b)( 1 at 2 + At) + C.1t − p(t)u for u ≥ 0.4. then (Hu )∗ = −2bu∗ (t) + p(t) ≤ 0. u). x (v) [10 − 2u∗ (t)]e−0. so according to Theorem 9. then u = −1 is optimal. Also p(t) = at + A for some constant A.4. p(t)) = [10u − u2 − 2]e−0. u∗ (t)). u∗ (t). Knut Sydsæter. x ∗ (0) = 0. the following conditions are sufﬁcient for optimality: (i) (ii) (iii) (iv) u∗ (t) maximizes H (t. x ∗ (t). T ]. p(t)) = 0. so A = t ∗ and p(t) = t ∗ − t. we get x ∗ (t) = 10−5t.1t . Since all the ¯ ∗ (t). ∗ (t) = u∗ (t) = (1/2b)(at + A). t > T 2 − 4. it sufﬁces to check that. that is. Suppose u∗ (t) > 0 for all t in [0. u). t ∗ ] and x ∗ (t) = t ∗ + x0 for t in (t ∗ .4. p) = [10u − u2 − 2]e−0. x. Since x ∗ (10) = 2. ˙ From (ii) we have p(t) ≡ p for some constant p. Since x ∗ (T ) = x1 .
and ˙ ∗ (t) = −2 − 2t ∗ . x For t ≥ t ∗ . See Kamien and Schwartz (1991) for further economic interpretations. So we have found an optimal solution.2 are satisﬁed. The ˙ pair (x ∗ . t ∗ ]. 2) with p(t ∗ ) = 2. T ] we have u∗ (t) > 0 and we see that (Hu )∗ = −2bu∗ (t) + p(t) = 0.68 9 CONTROL THEORY: BASIC TECHNIQUES 9. x ∗ (t) = (a/4b)(t−t ∗ )2 . The minus √ sign would make t ∗ > T . In (t ∗ . The function x ∗ must be increasing. So u∗ (t ∗+ ) = 0. 1]. Note that t ∗ > 0 ⇐⇒ B < aT 2 /4b. t ∗ ) for t ∈ (t ∗ .4. p(t)) = −2x ∗ (t).) Suppose B < aT 2 /4b. so x ∗ (t) remains constant. For t in [0. x ∗ (0) = 1. p. We cannot have u∗ (t) ≡ 0 because then x ∗ (t) ≡ 0. T ] ⎩ ⎩ if t ∈ (t ∗ . and p(t) >2 <2 for t ∈ [0. and Peter Hammond 2008 . we have u∗ (t) = 0 and (Hu )∗ = −2bu∗ (t) + p(t) = a(t − t ∗ ) ≤ 0. so p(0) ≥ p(2) + 4 = 4. we have x ∗ (t) = u∗ (t) = 0. −2x(t) = −2 − 2t. we have A = −at ∗ . p(t ∗ ) must equal 2. t ∗ ]. which gives p(t) = −2t − t 0 0 ∗ (t ∗ ) = 1 + t ∗ (the limit of x ∗ (t) as t approaches t ∗ from the left). It remains to check that the proposed solution satisﬁes all the conditions in Theorem 9. p(t) = 2 + C for a suitable constant C . x ∗ (t) = x ∗ (t ∗ ) = 1 + t ∗ . where C = p(2) + 2(1 + t ∗ )2 = 4(1 + t ∗ ). and with x ∗ (t ∗ ) = 0. or T − t ∗ = ± 4bB/a. and p = a = −(Hx )∗ . (This solution is valid if the required total production B is large relative to the time period T available. t ∗ ] ⎨0 ⎨0 ∗ ∗ u (t) = a(t − t ∗ ) . with t close to t ∗ . Hence p(t) = −2x ∗ (t) ≤ −2x ∗ (0) = −2 < 0.4. u∗ (t). Consequently p is ˙ ˙ strictly decreasing. T ] we have as before u∗ (t) = (1/2b)(at +A). For t in (t ∗ . x ∗ (t ∗ ) = 0. Because p ≤ −2. p(t) = −2x ˙ 1 1 Now. (Since x(2) is free. There is therefore a unique t ∗ in (0. p(t) = a(t − t ∗ ) . so −2(1 + t ∗ )t ∗ + 4(1 + t ∗ ) = 2.8 The Hamiltonian is H (t. the unique maximizer of H would be > 0 for some t < t ∗ . Atle Seierstad. as long as we really check that all the conditions in the sufﬁciency theorem are satisﬁed. 2). we put p0 = 1. 2] ˙ ˙ For t ≤ t ∗ . t ∗ ] for t ∈ (t ∗ . so p(t) = −2(1 + t ∗ )t + C . which gives (a/4b)(T − t ∗ )2 = B.4. as it should be.1) gives the following necessary conditions: (i) u = u∗ (t) maximizes (p(t) − 2)u for u ∈ [0. x ∗ (t). 9. In particular. Moreover. Then x ∗ (t) = u∗ (t) = (a/2b)(t−t ∗ ). we have p(t) > 0 for t in [0.) The maximum principle (Theorem 9. ˙ From (i) we see that p(t) > 2 ⇒ u∗ (t) = 1 and p(t) < 2 ⇒ u∗ (t) = 0. or (t ∗ )2 − t ∗ − 1 = 0. We end up with the following suggestion for an optimal solution: ⎧ ⎧ if t ∈ [0. and u∗ (t) = (a/2b)(t −t ∗ ). then by continuity of p(t). (ii) p = −Hx (t. so we must have t ∗ = T − 2 bB/a. ˙ (iii) x ∗ (t) = u∗ (t). Knut Sydsæter. u) = x 2 − 2u + pu = x 2 + (p − 2)u. Note: When we use sufﬁcient conditions. 2] ⇒ u∗ (t) = 1 0 for t ∈ [0. we have x ∗ (t) = u∗ (t) = 1. and the storage cost a is sufﬁciently small relative to the unit production cost b. p(2) = 0. Then we guess that production is postponed in an initial time period. Now t ∗ is determined ˙ √ by the condition x ∗ (T ) = B. so x ∗ (t) = x ∗ (0) + t = 1 + t. and u∗ (t) = x ∗ (t) ≡ 0 in [0. because x ∗ (t) ≥ 0. This quadratic equation © Arne Strøm.4. u∗ ) is admissible and satisﬁes the terminal condition x ∗ (T ) = B. and since u∗ (t ∗+ ) = 0. Since x ∗ is continuous.2. we can use “wild” guesses about the optimal control and the optimal path. It remains only to check the maximum condition. T ] if t ∈ (t 2b 4b √ with t ∗ = T − 2 bB/a. If u∗ (t ∗+ ) > 0. x (t) = a(t − t ∗ )2 ∗. t ∗ ] if t ∈ [0. which contradicts x ∗ (T ) = B > 0 (assumed implicitly). x. we see that ˙ 2 2 ˙ p(2) − p(0) = 0 p(t) dt ≤ 0 (−2) dt = −4. Since p(2) = 0.
5. The associated control problem is 1 max 0 (2xe−t − 2xu − u2 ) dt. p) = 2xe−t − 2xu − u2 + pu. ˙ ˙ t=2 with −4x ∗ (2) = 0 ˙ if x ∗ (2) > 4 √ √ √ 1 1 1 ˙ ˙ Equivalently. The function F is (for t ﬁxed) concave in (x. so (i) u∗ (t) = 2 p(t) − x ∗ (t). x(0) = 0. and Peter Hammond 2008 . If p(t)u has a maximum for some u in (−∞. so we have found the solution. Fx = −2x and Fx = −4x. The only possible solution is spelled out in the answer in the book.3): (Fx )∗ = −4x ∗ (2) ≤ 0. x ∗ (2) ≥ 0. We know that for t ≥ t ∗ we have √ p(t) = −2(1 + t ∗ )t + C1 = −2(1 + t ∗ )t + 4(1 + t ∗ ) = 2(1 + t ∗ )(2 − t) = (3 + 2 5 )(2 − t). x ˙ ˙ ˙ ∗ (t) = u∗ (t) = 1 p − x ∗ (t) = −e−t . then the Hamiltonian would simply be pu. using (i) and (ii). 1 Since the control region is open. We conclude that x ∗ (2) = 4 and thus A = (4e 2 − 1)/(e2 2 − 1). p(t)) = 0. C 9 3 we use the equation p(t ∗ ) = 2. contra˙ √ √ dicting (∗). As we found in (a). u. We also know that there is a constant C0 such that p(t) = −2t − t 2 + C0 for t ≤ t ∗ . which gives C0 − 2t ∗ − (t ∗ )2 = 2. From the boundary follows that x = e ˙ conditions x(0) = 0 and x(1) = 1 we get A = e−1 and B = 1. x ∗ (2) = Ae 2 +(1−A)e− 2 ≥ 4 requires √ √ √ √ √ (∗) A ≥ (4 − e− 2 )/(e 2 − e− 2 ) = (4e 2 − 1)/(e2 2 − 1) ≈ 0. Atle Seierstad. Hu (t. so the Euler equation Fx − (d/dt)Fx √ 0 ˙ ˙ ˙ = √ 1 1 1 reduces to x − 2 x = 0.5.5. (Incidentally. √ √ Since x ∗ (0) = 1. x) = 2xe−t − 2x x − x 2 we get Fx = 2e−t − 2x and Fx = −2x − 2x. ∞).97 1 √ √ We now invoke the transversality condition (8. or C0 = 2 + 2t ∗ + (t ∗ )2 = 2 + 2 5.06. moreover. Since x ∗ (t) = 2 2[Ae 2 2 t − (1 − A)e− 2 2 t ]. x(2) ≥ 4 © Arne Strøm. x(1) = 1. x. x) = 3 − x 2 − 2x 2 . ¨ 1 1 So if x ∗ (t) solves the problem. u∗ (t). if p0 = 0. The characteristic equation r 2 = 2 has solutions r1 = 2 2 and r2 = − 2 2. x. x). then p(t) = 0. with x ∗ (2) = 0 if x ∗ (2) > 4. Knut Sydsæter. x ∗ (t). which reduces to x = −e−t . Now that we know x ∗ (t). Moreover ∗ (t). The Euler equation ˙ ˙ ˙ ˙ ˙ ˙ d is then 2e−t − 2x − dt (−2x − 2x) = 0. there is precisely one solution that also satisﬁes the boundary conditions on x ∗ (t). x. ˙ The control problem is 2 max 0 (3 − x 2 − 2u2 ) dt.5 9. ∞) ˙ The Hamiltonian (with p0 = 1) is then H (t. ˙ √ √ √ √ √ √ 1 we have x ∗ (2) = 2 2[Ae 2 − (1 − A)e− 2 ] = 0 provided A = e− 2 /(e 2 + e− 2 ) ≈ 0. x = u. Hence x ∗ (t) must be a solution of the ˙ ˙ get x ¨ 2 ˙ Euler equation that we found in (a). the control function u∗ (t) and the adjoint function p(t) are given by the equations above. x = u. From x ∗ (t) = u∗ (t) we (ii) p(t) = −Hx (t.2 With F (t. we have A+B = 1. so the only admissible solution of the Euler equation is x ∗ (t) = −e−t + e−1 t + 1. 9. or B = 1−A. u ∈ (−∞.) We have Hu = −2x − 2u + p and Hx = 2e−t − 2u. then we must have x ∗ (t) = Ae 2 2 t + Be− 2 2 t . u∗ (t). t ∗ = (1 + 5 )/2. or 2e−t − 2x + 2x + 2x = 0. It ˙ ˙ ˙ ˙ ¨ ¨ −t + A and x = −e−t + At + B for suitable constants A and B.1 With F (t. To determine the constant√ 0 . ˙ x(0) = 1. and. but p(t) = 0 is impossible when p0 = 0. p(t)) = −2e−t + 2u∗ (t) = −2e−t + 2x ∗ (t).9 CONTROL THEORY: BASIC TECHNIQUES 69 √ √ gives t ∗ = (1 ± 5 )/2. ˙ ˙ 9. and since t ∗ must be positive.
For all t in [0. x). p(t)) = 2(x ∗ (t))2 e−2t + (p(t) − 1)uet . ˙ ∗ = (−2 − 2u∗ (t))e−t/10 + p(t) = 0.5. u). x ∗ (t). (Hu ) ¯ 1 pet/10 − 1. T ) = 0 T 1 x ∗ (t) − 2 u∗ (t)2 dt = t=T T 0 1 1 T t − 2 t 2 + x0 − 2 (T − t)2 dt (∗) 1 1 Integrating we ﬁnd V (x0 . Since 2(x ∗ (t))2 e−2t does not depend on u. The rest x ¨ ˙ 4 ˙ 4 2 is easy. If (x ∗ (t). p(t)) = −u∗ (t) + p(t) = 0.2 (see (9. x ∗ (t). (ii) p(t) = 2x ∗ (t). (b) The optimal value function is V (x0 . u maximizes H (t. T ]. we have x ∗ (t) = u∗ (t)et ≥ 0. (Recall that x(T ) is free. p) = x − 2 u2 + pu is concave in (x. u∗ (t) = 1 if p(t) > 1. x ∗ (0) = x0 . x. The boundary conditions ¯ ˙ ¯ 2 yield p = 0 and A = 1. 1] that strictly decreasing.4 (a) The Hamiltonian is H (t. and so p(t) = 2e−2t + C. so the Euler equation is ˙ ˙ ˙ ˙ ˙ d −t/10 ˙ = 0. From the maximum principle. (ii) p(t) = −Hx (t. u∗ (t). u∗ (t). u∗ (t). integration gives x ∗ (t) = 5pet/10 − t + A. T ) = t=0 ( 2 T t 2 − 1 t 3 + x0 t + 6 (T − t)3 ) = 1 T 3 + x0 T . (If we were asked only to ﬁnd ∂V /∂x0 . Here p(t) = −(Hx )∗ = 0. u∗ (t)) is 1 optimal then (i) u∗ (t) = 4 p(t). so ∂V /∂x0 = T = 6 6 p(0). u∗ (t) = 0 if p(t) < 1. Since x ∗ (t) = u∗ (t) = T − t and x ∗ (0) = x0 .2. Fx = 0 and Fx = (−2 − x)e−t/10 .70 9 CONTROL THEORY: BASIC TECHNIQUES The Hamiltonian is H = 3 − x 2 − 2u2 + pu. see that ∂V /∂T = 2 T + x0 = H 9. we have p(t) < 1 for t close to T . It follows that x ∗ (t) = u∗ (t) = 0 and x ∗ (t) ≡ x ∗ (0) = 1 for all t in [0. u. u∗ (t) is the value of u in [0. x. and Peter Hammond 2008 . for every t in [0. the differential equation for p(t) is p(t) = −4e−2t x ∗ (t). we get 1 x ∗ (t) = T t − 2 t 2 + x0 . The Hamiltonian for the control problem is H = (−2u − u2 )e−t/10 + pu. Since x ∗ (t) = u∗ (t). Therefore. it would be easier to differentiate (∗) with respect to x0 T under the integral sign: ∂V /∂x0 = 0 dt = T . p(t)) = −1 with p(T ) = 0. T ]. ¯ 9. Since ˙ Hx = 4xe−2t . There are two possibilities: Case A: Suppose p(0) ≤ 1.1 (a) The Hamiltonian H (t. the following conditions are sufﬁcient for optimality: (i) Hu (t. ˙ − (−2 − x)e dt so we have found the solution.2. ˙ © Arne Strøm.) The value of the Hamiltonian “along the optimal path” is H ∗ (t) = H (t. so p(t) is ˙ ˙ ∗ (t) is the value of u in [0.3 With F (t.6. and hence x ∗ (t) ≥ x ∗ (0) = 1 for all t in [0.) Suppose that (x ∗ (t). u. T ]. We have p(t) = −4e−2t x ∗ (t) = −4e−2t . 9. x) = (−2x − x 2 )e−t/10 . x ∗ (t). Thus. and so H ∗ (T ) = 2 T 2 + x0 . 1] that maximizes (p(t) − 1)u. Differentiating (iii) yields ˙ ˙ ∗ (t) = u∗ (t) = 1 p(t) = 1 2x ∗ (t) = 1 x ∗ (t). so Hx = −2x and Hu = −4u + p. We 1 2 ∗ (T ). u∗ (t)) is an admissible pair with adjoint function p(t). which is the same differential equation as before. u. p(t)) = 1 1 1 1 x ∗ (t) − 2 u∗ (t)2 + p(t)u∗ (t) = T t − 2 t 2 + x0 − 2 (T − t)2 + (T − t)2 . we get p(t) < 1. Since p is strictly decreasing. ˙ ˙ From (i) and (ii) we have u∗ (t) = p(t) = T − t.6. so according to Theorem 9. See the answer in the book. T ]. Since p(T ) = 0. and hence u∗ (t) = 0 for all t > 0. so we get the same solution. x. (iii) x ∗ (t) = u∗ (t). Knut Sydsæter. Moreover. ˙ It remains to determine p(t). ˙ (iii) x ∗ (t) = u∗ (t). Atle Seierstad.7)). x ∗ (t). The function F is (for t ﬁxed) concave in (x. and thus u∗ (t) = so p(t) = p.6 1 9. p) = 2x 2 e−2t − uet + puet = 2x 2 e−2t + (p − 1)uet . p(t) < 0. a constant.
˙ ∗ p(t) = −4e−2t x ∗ (t) = −4e−t if 0 ≤ t < t ∗ . Atle Seierstad. This is a second degree equation ∗ for determining et . and T 0 V (T ) = 2(x ∗ (t))2 e−2t − u∗ (t)et dt = 0 T 2e−2t dt = T 0 −e−2t = 1 − e−2T © Arne Strøm. if and only if 1 ( 4 e2T + 1)2 = 1 4T 16 e + 1 4T 16 e 1 2T + 1. Integration gives ˙ ˙ p(t) = 4e−t + C1 2e t ∗ −2t if 0 ≤ t < t ∗ if t ∗ < t ≤ 1 + C2 Since p is continuous. 1 This gives e−2T ≥ 1/2. x ∗ (T ). and Peter Hammond 2008 . This gives u∗ (t) = 1 0 if 0 ≤ t ≤ t ∗ if t ∗ < t ≤ 1 ∗ Since x ∗ (t) = u∗ (t) and x ∗ (0) = 1. we get x ∗ (t) = et if 0 ≤ t ≤ t ∗ and x ∗ (t) = et if t ∗ ≤ t ≤ 1. and hence C = −2e−2T . Case B: Suppose p(0) > 1. and so H ∗ (T ) = H (T . and so −2T ≥ ln(1/2) = − ln 2. u∗ (T ). (b) We have to consider two cases A and B. Since we have assumed p(0) ≤ 1. Then there exists a point t ∗ ∈ (0. H ∗ (T ) = 2e−2t . Knut Sydsæter. and we ﬁnd ∗ 1 et = − 4 e2T + ∗ 1 4T 16 e + e2T and so t ∗ = ln 1 4T 16 e 1 + e2T − 4 e2T (Since et is positive.9 CONTROL THEORY: BASIC TECHNIQUES 71 where the constant C is determined by p(T ) = 0. if 0 ≤ t ≤ t ∗ if t ∗ ≤ t ≤ 1 ∗ ∗ −2T p(t) = 4(e−t − 4e−t ) + 1 2et (e−2t − e−2T ) ∗ It remains to determine t ∗ .e. From (ii) and (iii) we get 1 = 2e−t − 2et ∗ et yields ∗ ∗ et = 2 − 2e2t −2T ∗ ∗ . p(T ) = 0. we must have 2(1−e−2T ) ≤ 1. It follows that p(t) = 2(e−2t −e−2T ) = 2e−2t (1−e−2(T −t) ). or T ≤ 2 ln 2. Thus. both expressions for p(t) are valid for t = t ∗ . Note that in both cases u∗ (T ) = 0 (and p(T ) = 0).) This solution makes sense if and only if t ∗ > 0. In this case p(t) > 1 for t < t ∗ and p(t) < 1 for t > t ∗ . This inequality is equivalent to 1 4T 16 e + e2T > and so e2T > 2 ⇐⇒ T > 1 2 ln 2. then x ∗ (T ) = 1. and p(t) = −4et −2t if t ∗ < t ≤ 1. Moreover. p(T )) = 2x ∗ (T )2 e−2T . 2e 1 + e2T − 4 e2T > 1. For the summing up see the book. i. Case A: If T ≤ 1 2 ln 2. T ) with p(t ∗ ) = 1. so we get the equations (i) p(t ∗ ) = 1 = 4e−t + C1 ∗ (ii) p(t ∗ ) = 1 = 2e−t + C2 ∗ ∗ −2T ∗ (iii) p(T ) = 0 = 2et ∗ −2T + C2 From (i) and (iii) we get C1 = 1 − 4e−t and C2 = −2et ∗ . we must take the positive square root. Hence. Multiplying this equation by (iv) 1 1 Multiplying with 2 e2T and rearranging gives (et )2 + 2 et e2T −e2T = 0.
s ∗ (t). ˙ 2 k (t) © Arne Strøm. Differentiating w.7. so the following conditions are sufﬁcient for optimality: (i) (Hu )∗ = −u∗ (t) + p(t) = 0. √ √ √ √ 9.1 (a) The Hamiltonian H = 100 − x − 2 u2 + pu is concave in (x. 1 (ii) p(t) = −Hk (t. ∞). ˙ (iii) x ∗ (t) = u∗ (t). x0 under the integral sign yields ∂V /∂x0 = x1 − x0 − 1 0 1 2 1 (t − 1 + t + x1 − x0 − 2 ) dt = 1 0 (2t − 3 2 + x1 − x0 ) dt = 1 0 3 (t 2 − 2 t + (x1 − x0 )t) = = p(0). and (i) and (iii) give x ∗ (t) = u∗ (t) = p(t) = t + A. (b) V = V (x0 .7 1 9.) The maximum principle (Theorem 9.72 9 CONTROL THEORY: BASIC TECHNIQUES This implies V (T ) = 2e−2T = H ∗ (T ). Integrating and ˙ 1 using the two boundary conditions gives B = x0 .4. as it should be. u) and u ∈ (−∞. x ∗ (0) = x0 . (Alternatively we could have inserted the expressions found for t ∗ and et into the expression for V (T ). x1 ) = 0 1 1 (100 − x ∗ (t) − 2 u∗ (t)2 ) dt = 0 1 0 1 1 1 100 − 2 t 2 − (x1 − x0 − 2 )t − x0 − 1 2 t + x 1 − x0 − 1 0 1 2 2 dt.r. A = x1 − x0 − 2 . Hence we get V (T ) = 2 ∗ dt ∗ dt ∗ dt ∗ ∗ dt ∗ ∗ ∗ ∗ − et − e2t −2T (2 − 2) = (2 − et − 2e2t −2T ) + 2e2t −2T dT dT dT dT ∗ ∗ −2T Equation (iv) shows that 2 − et − 2e2t = 0.2 (a) The Hamiltonian is H = (1 − s) k + ps k = k + k(p − 1)s. p(t)) = − √ ∗ [1 + s ∗ (t)(p(t) − 1)].1) gives the conditions: √ √ (i) s = s ∗ (t) maximizes k ∗ (t) + k ∗ (t)(p(t) − 1)s for s ∈ [0. V (T ) = = 0 2(x ∗ (t))2 e−2t − u∗ (t)et dt (2 − et ) dt + T t∗ t∗ 2e2t ∗ −2t dt = 2t ∗ − et + 1 − e2t ∗ ∗ −2T +1 Note that t ∗ depends of T . 1]. and found V (T ) in terms of T . ∂V /∂x1 = 1 2 ) dt 1 [−t − t + x1 − x0 − 2 )] dt = 1 2 (−2t − x1 + x0 + = 1 0 (−t − x1 t + x0 t + 2 1 2 t) = −x1 + x0 − = −p(1). and x ∗ (1) = x1 . The algebra required is heavy. so the solution is as given in the book. so we assume p0 = 1. and therefore ∗ −2T V (T ) = 0 + 2e2t = H ∗ (T ) ∗ as expected. This gives H ∗ (T ) = 2e2t T 0 ∗ ∗ −2T . Knut Sydsæter. (ii) p(t) = 1. In the same way.t. we have x ∗ (T ) = x ∗ (t ∗ ) = et . ˙ From (ii) we get p(t) = t + A. Furthermore. and Peter Hammond 2008 .) 9. Atle Seierstad. Case B: For T > 1 2 ln 2. p(10) = 0. (In (b) we use the Arrow theorem.7. as it should be. k ∗ (t).
t ∗ ]. so k ∗ (t) = ( 1 t + 1)2 . 10) such that p(t ∗ ) = 1. u. and we see that g (u) > 0 when u is slightly larger than 0. and p(t) can be found in the answer in the book. and so p(t) = −t/(t ∗ +2)+D. ˙ √ ∗ p(t) = −p(t)/2 k (t) < 0. t ∗ ] if (t ∗ . (If p(t) > 1. It remains only to ﬁnd p(t) on [0.3 (a) The Hamiltonian H (t. ˙ √ From (ii) we get p(t) = Ae−αt . for some constant E. a contradiction. But p(t ∗ ) = 1. so (10−t)/(t ∗ +2) = 1. x ∗ (t). from which it follows that t ∗ = 4.7. for any t. ˙ and with p(10) = 0 we get p(t) = 5−t/2. x ∗ (t). See the answer in the book. Thus. 10]. so e−βt = Ae−αt √ ∗ 2 u (t) or u∗ (t) = 1 2(α−β)t e 4A2 (b) See the answer in the book. t ∗ ]. We conclude that there must exist a t ∗ in (0. we ﬁnd 2 k ∗ (t) = t + C. this implies that p(t) = (10−t)/(t ∗ +2) on (t ∗ . and Peter Hammond 2008 . t ∗ ] and k ∗ (t) = 0 on [0. so E = 6. Knut Sydsæter. From (ii) this gives p(t) = −1/2. k ∗ (0) = 1. s ∗ (t) = 1 0 if [0. ˙ (iii) x ∗ (t) = αx ∗ (t) − u∗ (t). Atle Seierstad. 10]. or the Hamiltonian.7.1. On this interval p(t) = ˙ √ −p(t)/2 k ∗ (t) = −p(t)/(t + 2). 10] . p(t)) = e−βt u + p(t)(αx ∗ (t) − u) for u ≥ 0. k ∗ (t) = 1 ( 2 t + 1)2 1 ( 2 t ∗ + 1)2 if [0. u∗ (t). We have found the only possible solution. With k ∗ (0) = 1 we get C = 2. Then s ∗ (t) ≡ 0 and k ∗ (t) ≡ 1. t ∗ ]. ˙ Suppose p(t) < 1 for all t in [0. 10] we get from (ii) that p(t) = −1/2 k ∗ (t) = −1/(t ∗ +2). p(t)) = −αp(t). Then we see from (i) that s ∗ (t) = 1 0 if p(t) > 1 if p(t) < 1 (iv) By studying (ii) and (iv) we see that p(t) < 0 for all t. √ 9. The solution of this separable equation is p(t) = E/(t + 2). t ∗ ] if (t ∗ . x. This means that u∗ (t) = 0 cannot maximize g. x ∗ (0) = 1. t ∗ ] and s ∗ (t) = 0 on (t ∗ . u). 10] (v) √ ˙ On (t ∗ . then s ∗ (t) = 1 and from (ii). the following conditions are sufﬁcient for optimality: √ (i) u = u∗ (t) maximizes H (t. The explicit expressions for x ∗ (t).) © Arne Strøm. Deﬁne the function g by g(u) = e−βt u − Ae−αt u. By integrating the differential equation for √ ∗ (t) on [0. it follows from (iii) that k ∗ (t) ≥ 1 for all t.9 CONTROL THEORY: BASIC TECHNIQUES 73 √ ˙ (iii) k ∗ (t) = s ∗ (t) k ∗ (t). √ ˙ ˙ But then k ∗ (t) = k ∗ (t) on [0. p) = e−βt u + p(αx(t) − u) is concave in (x. Since s ∗ (t) ≥ 0. But p(4) = 1. u. Since p(10) = 0. 10] we have k ∗ (t) = ( 2 t ∗ + 1)2 since k ∗ (t) is continuous. so according to Theorem 9. But then p(t) > 1 for t < 8. u∗ (t). x ∗ (T ) = 0. Hence u∗ (t) > 0 and g (u∗ (t)) = 0. (There is a misprint in the formula for x ∗ (t): the ﬁrst fraction must be multiplied by eαt . Then s ∗ (t) = 1 on [0. (ii) p(t) = −Hx (t. Then g (u) = √ e−βt (1/2 u) − Ae−αt . (What if p(t) < 1?) Thus p(t) is strictly decreasing. On k 2 1 (t ∗ . 10]. From (iii) we get the following linear differential equation for x ∗ (t): x ∗ (t) = αx ∗ (t) − ˙ 1 2(α−β)t e 4A2 with solution x ∗ (t) = Ceαt − 1 e2(α−β)t 4A2 (α − 2β) The two constants A and C are determined by the boundary conditions in (iii).
t ∗ ].8. 5]. x ∗ (t). Thenp0 = 0 implies p(t) = 0.1. ˙ x ∗ (t) = u∗ (t). u). Then there exists a continuous function p(t) and a number p0 . and so u∗ (t) = 2p. p(t)) = p0 (−9 − 4 u2 ) + p(t)u for u ∈ . This reduces to aB/b = (u∗ (T ∗ ))2 . Hence p0 = 1.1t = 0. the only change in the conditions (i)–(iii) is that (ii) is augmented by p(T ) ≥ 0. Hence x ∗ (T ) = 0. (See Note 9.1λ = −(H c )x (t. x ∗ (t). ˙ 1 H (t ∗ . λ(t)) = 10u − u2 − 2 − λ(t)u for u ≥ 0. we ﬁnd that the equation H ∗ (T ∗ ) = 0 does not determine T ∗ . u∗ (t) = 6. 0) and (i) (ii) (iii) (iv) 1 u = u∗ (t) maximizes H (t.1t . such that for all t in [0.) Suppose (x ∗ (t). λ(t)) = 10 − 2u∗ (t) − Ae0. which has no maximum. u.2 The current value Hamiltonian H c = 10u − u2 − 2 − λu is concave in (x.1t for some constant A. p(t)) = (0.8. and Peter Hammond 2008 . Finally. and x ∗ (t) = 6t.9. and x ∗ (t ∗ ) = 16 yields p = 8/t ∗ . and integration gives x ∗ (t) = −5t + 5Ae0. Integrating ¯ ¯ ¯ ˙ ¯ ¯ ¯ and using x ∗ (0) = 0 this gives x ∗ (t) = 2pt. t ∗ ] we have (p0 . or ax ∗ (T ∗ ) + b(u∗ (T ∗ ))2 = p(T ∗ )u∗ (T ∗ ) = 2b(u∗ (T ∗ ))2 .2 Consider the case B ≥ aT 2 /4b. So the only possible solution ¯ ¯ ¯ ¯ ¯ ¯ is p = 3. p(t) = −Hx (t. ¯ 9. 1 Then x ∗ (t) = −5 + 2 Ae0. x ∗ (t). u∗ (t). © Arne Strøm. Then from Problem 9. (To save on storage √ costs. x ∗ (t) = −u∗ (t). x ∗ (5) ≥ 0. λ(5) ≥ 0. u). u. and then produces at an optimal rate until delivery time T . Here p = −3 gives t ∗ = 8/p < 0.) 9. u∗ (t). To determine the optimal T ∗ . 9. Knut Sydsæter. and u∗ (t) = 2p(t). so A = 0 and p(t) ≡ 0.1t − 1) − 5t + 10. ˙ From (ii) we have λ(t) = Ae0. u∗ (t ∗ ). but this does not guarantee that a solution to the necessary conditions is optimal. Then (i) is equivalent to (H c )u (t. x ∗ (0) = 10. that is. p(t)) = (0. p(t)) = 0. x ∗ (t). λ(t)) = 0. 1 Since H is concave in (x. x ∗ (T ) = 16. Then t ∗ = 8/3. Solving this equation for T ∗ gives the unique √ solution T ∗ = 2 bB/a.4. so waiting costs nothing.74 9 CONTROL THEORY: BASIC TECHNIQUES (b) If x(T ) = 0 is replaced by x(T ) ≥ 0. u). Again p(t) = Ae−αt . p(t ∗ )) = p0 (−9 − 4 u∗ (t ∗ )2 ) + p(t)u∗ (t ∗ ) = 0.1t + B.1t . The maximum condition (i) would then imply that u∗ (t) for u ≥ 0 √ would maximize e−βt u. u∗ (t)) is an optimal pair deﬁned on [0.1 (b) In this problem we use Theorem 9. Moreover. so p(T ) = 0 would imply Ae−αT = 0. Suppose u∗ (t) > 0 for all t in [0. u∗ (t) = a(2t − T )/4b + B/T and x ∗ (t) = at (t − T )/4b + Bt/T and p(t) = 2bu∗ (t). condition (i) is equivalent to (Hu )∗ = −p0 2 u∗ (t) + p(t) = 0.7. so x ∗ (t) = 5A(e0. which is either 0 or 1.8 9.) If B < aT 2 /4b. which contradicts (p0 . and the solution is as in (a). Note that no discounting is assumed. x ∗ (0) = 0. with p(T ) = 0 if x ∗ (T ) > 0. or p2 = 9.1 and we try to ﬁnd the only possible solution to the problem. (Note that this is the positive solution T of B = aT 2 /4b. so the following conditions are sufﬁcient for optimality: (i) (ii) (iii) (iv) u = u∗ (t) maximizes H c (t. ˙ λ(t) − 0. that is aT ∗ /4b + B/T ∗ = aB/b.9 9. the condition is H (T ∗ ) = 0. (iv) implies 1 −9 − 4 (2p)2 + p2p = 0. x ∗ (t) = u∗ (t) = 2p. or √ √ u∗ (T ∗ ) = aB/b.8. x ∗ (t ∗ ). From (ii) we have p(t) = p for some constant p. 1 (v) u∗ (t) = 5 − 2 Ae0. with λ(5) = 0 if x ∗ (5) > 0. x ∗ (t). 0). 1 The Hamiltonian H = p0 (−9 − 4 u2 ) + p(t)u is concave in (x. where B is a new ˙ constant. The initial condition x ∗ (0) = 10 yields B = 10 − 5A. Atle Seierstad. u∗ (t). the ﬁrm waits until t = T − 2 Bb/a to start production.8.
t. and p(t) < 1 for all t in [0.5.1t − t + B. p = −H2 (t. Then (ii) gives p(10) = 5/6 < 1 a contradiction. u.3 Compared with Problem 9. A) = e−ρT ϕ(A). we ∗ (T )) = U (rA∗ (T ) + w − u∗ (T )). w is a constant. (We assume that λ0 = 1. or k ∗ (t) = ( 2 t + 1)2 . With the assumptions in Example 8. and u ∈ . Using u∗ (t) = A∗ (t). Since ϕ(A) is also concave.5 − 1). the utility function U has U > 0 and U < 0. with A = 0 if x ∗ (5) > 0.) Then √ √ 1 ˙ s ∗ (t) ≡ 1 and k ∗ (t) = k ∗ (t). 1 Then x ∗ (t) = 2 Ae0.7.10.7. u). with k ∗ (0) = 1.1t − 1.3 The current value Hamiltonian H c = −2u − u2 + λu is concave in (x.6 are found.1λ = −(H c )x (t. A∗ (T )) = e−ρT ϕ (A∗ (T )). which for t = 5 yields the contradiction x ∗ (5) < 0. Suppose p(t) > 1 for all t in [0. In the answer in the book we go a little further. A∗ (t). u∗ (t)). A. Thus A > 0 and x ∗ (5) = 0. The scrap value function is S(k) = 10 k. x ∗ (t). p) = U (rA(t) + w − u(t))e−ρt + pu. A. 10) such that p(t ∗ ) = 1. ˙ A∗ (t) = u∗ (t). A∗ (t). 10] is again easily seen to be impossible. Atle Seierstad. p(t)) = 0.1t for some constant A. p) is concave in (A. From (ii) we have u∗ (t) = 2 λ(t)−1 = 2 Ae0. with solution x ∗ (t) = 5Ae0. λ(t)) = 0.9 CONTROL THEORY: BASIC TECHNIQUES 75 Next look at condition (iii) which reduces to A ≥ 0. we have found the solution. u. t and ˙ equate it to the expression for p(t) in (ii). x ∗ (0) = 1. u∗ (t).2 are still valid except that p(10) = 0 in (ii) is now replaced by 5 p(10) = √ ∗ (ii) k (10) Again p(t) is strictly decreasing.9. x ∗ (t) = u∗ (t). and we get the same solution as in Problem 9. 9. with the given λ(t). this gives ˙ ˙ ˙ ¨ ˙ ˙ U (rA∗ + w − A∗ )(r A∗ − A∗ )e−ρt − ρU (rA∗ + w − A∗ )e−ρt = −rU (rA∗ + w − A∗ )e−ρt ¨ ˙ Multiplying by −eρt and rearranging gives A∗ − r A∗ + (ρ − r)U /U = 0. so the following conditions are sufﬁcient for optimality: (i) (ii) (iii) (H c )u = −2 − 2u∗ (t) + λ = 0. This implies that H (t. the Hamiltonian is H (t. ˙ λ(t) − 0.r. The constants A and B are determined ˙ from the boundary conditions. 10]. Combining (i) and (iii). or p(t) = U (rA∗ (t) + w − u∗ (t))e−ρt .5 − 1) − 15. We differentiate the expression for p(t) in (i) w. A∗ (0) = A0 . k ∗ (10) = 36.1t −1.) Moreover. ˙ p(T ) = SA (T . p(t)) = −rU (rA∗ (t) + w − u∗ (t))e−ρt . u∗ (t). and (v) in the answer to Problem 9.5.2 the only difference is that instead of the condition “x(10) free”.2 is still valid (although © Arne Strøm. (This was not an option when we required p(10) = 0. ˙ 1 1 From (ii) we have λ(t) = Ae0. also see that ϕ (A 9. and the same solutions for x ∗ (t) and u∗ (t) as in Problem 9. We conclude that there must exist a t ∗ in (0.4. But A = 0 would imply x ∗ (t) = −5t + 10. 9. with √ S (k) = 5/ k.7.10.10 9.3. and Peter Hammond 2008 . Conditions (i)–(iv) in the answer to Problem 9.2 With A as the state variable. Since all the conditions (i)–(iv) are satisﬁed by the admissible pair (x ∗ (t). It follows that 2 k ∗ (t) = t + 2. x ∗ (1) = 0. a set of sufﬁcient conditions for optimality is then (assuming interior solution in (i)): (i) (ii) (iii) (iv) H3 (t. so A = 3/(e0.3. But then 0 = 5A(e0. u∗ (t). and the scrap value function is S(T . we √ have now included a scrap value in the objective function. u). In particular. Knut Sydsæter.
and it is valid also in this case. Hence u∗ (t) = Since x ∗ (t) = u∗ (t) and x(0) = 1/2. x. we have 1 p(t) = −t + 2 − 2 x ∗ (1) (iii) Since x ∗ (0) = 1/2 and 0 ≤ x ∗ ≤ 1. In particular. then according to (B) in Theorem 9. x. Knut Sydsæter. we have p(t) = p(0) − t. p(t) = − ˙ ∂H ∗ = −1 and ∂x p(1) = S (x ∗ (1)) = 1 2 (i) Moreover. 1] 1 0 if t ∈ [0. We now use (ii) to determine t ∗ : 2t ∗ − 8 = t∗ + 2 1 ∗ 2t 10 5 = ∗ t +2 +1 from which it follows that t ∗ = 9. which implies that D = (2t ∗ + 2)/(t ∗ + 2).) 1 9. see the answer in the book. Condition (i) 4 is replaced by p(t) = −1 ˙ 1 1 ¯ and p(1) = S (x ∗ (1)) = − 2 (x ∗ (1) − 2) = 1 − 2 x ∗ (1) (ii) 1 From (ii) we see that p(t) = −t + C and since p(1) = 1 − 2 x ∗ (1). The rest is routine. We use Theorem 9. 10] we again get p(t) = −t/(t ∗ + 2) + D and 1 = p(t ∗ ) = −t ∗ /(t ∗ + 2). we get ˙ x ∗ (t) = t + 1/2 1 if t ∈ [0.) If (x ∗ (t). Condition (iii) therefore gives ˙ 1/4 ≤ p(1) ≤ 3/4 Since p(t) = −1 for all t. 1/2] if t ∈ (1/2. and Peter Hammond 2008 . and p(t) is the adjoint function. With p0 = 1.4 (a) This is a problem with scrap value function S(x) = 2 x. u. From (i) we get p(t) = 3 2 −t We see that the strictly decreasing function p(t) is equal to 1 at t = 1/2.1. it is clear that 1/2 ≤ x ∗ (1) ≤ 3/2. 1/2] if t ∈ (1/2. p) = x −u+pu = x +(p −1)u. (Again the Arrow condition is satisﬁed. u). Atle Seierstad. u∗ (t) = 1 if p(t) > 1 and u∗ (t) = 0 if p(t) < 1. u∗ (t)) be an admissible pair.10. the Hamiltonian is H (t. but the Hamiltonian is as in (a). (Note that the scrap value function is not to be included in the Hamiltonian. u. and thus p(t) = −t/(t ∗ + 2) + (2t ∗ + 2)/(t ∗ + 2). On (t ∗ . 1] We have found the optimal solution because H (t.10. p(t)) and S(x) = x/2 are concave (in fact linear) functions of (x. Hence.1.10. Let (x ∗ (t). p(0) = p(1) + 1 and ˙ 5/4 ≤ p(0) ≤ 7/4 © Arne Strøm. u∗ (t)) is an optimal pair in the problem.76 9 CONTROL THEORY: BASIC TECHNIQUES with a different t ∗ ). ¯ (b) The scrap value function is now S(x) = − 1 (x − 2)2 . p(10) = (2t ∗ − 8)/(t ∗ + 2).
so u∗ (t) = 2 p(t). x ∗ (t). 9. u∗ (t)) is therefore exactly as in (a). x ∗ (t) = . p) = −2u∗ (t) + p(t) = 0. so there exists a unique number t ∗ between 0 and 1 such that p(t ∗ ) = 1. T ) = − 0 T (u∗ (t))2 dt − (x ∗ (T ))2 = − T 0 4x 2 = − 2T 0 2 (3e − 1) It follows that 1 2t e + e2T 2 2 4x0 (3e2T − 1)2 4x 2 = − 2T 0 2 (3e − 1) T 0 e2t dt + e2T 2x 2 3 2T 1 e − = − 2T 0 2 2 3e − 1 ∂V 4x0 = p(0) and = − 2T ∂x0 3e − 1 2 12x0 e2T ∂V = ∂T (3e2T − 1)2 We see from the solutions in (a) that p(T ) = −2x ∗ (T ) and u∗ (t) = −x ∗ (T ). 1]. p) = −u2 − px + pu is concave in (x. we get 1 p(t) = −2x ∗ (T )et−T and also u∗ (t) = 2 p(t) = −x ∗ (T )et−T . x ∗ (0) = x0 . Since p(t ∗ ) = 1. and Peter Hammond 2008 . 1] To determine t ∗ we use the transversality condition. p(t) = − 2T 3e2T − 1 3e − 1 3e − 1 x ∗ (T ) = (b) We have V (x0 . (ii) p(t) = −∂H ˙ (iii) x ∗ (t) = −x ∗ (t) + u∗ (t). ˙ From (ii) we get p(t) = Cet for an appropriate constant C. u. Knut Sydsæter. A different scrap value function would usually have given another solution. Atle Seierstad. ∗ /∂x = p(t) and p(T ) = −S (x ∗ (T )) = −2x ∗ (T ). The differential equation x ∗ (t) = −x ∗ (t) + u∗ (t) = −x ∗ (t) − x ∗ (T )et−T ˙ has the general solution 1 x ∗ (t) = De−t − 2 x ∗ (T )et−T 1 3 From x ∗ (T ) = De−T − 2 x ∗ (T ) we get D = 2 x ∗ (T )eT . t ∗ ]. u∗ (t). which means that t ∗ = 2 .9 CONTROL THEORY: BASIC TECHNIQUES 77 Condition (ii) is still valid. The initial condition x ∗ (0) = x0 gives 1 3 1 1 D − 2 x ∗ (T )e−T = 2 x ∗ (T )eT − 2 x ∗ (T )e−T = 2 x ∗ (T )(3eT − e−T ) = x0 so 3x0 e2T 3 2x0 2x0 eT and D = x ∗ (T )eT = 2T = 2T 3e − 1 2 3e − 1 3eT − e−T Thus the optimal solution is 4x0 et x0 (3e2T −t − et ) 2x0 et u∗ (t) = − 2T .5 (a) The Hamiltonian H (t. but this is accidental. x ∗ (t) = 1/2 + t 1/2 + t ∗ if t ∈ [0. The transversality condition then 1 1 1 3 1 1 gives t ∗ = p(1) = 1 − 2 x ∗ (1) = 1 − 2 2 + t ∗ = 4 − 2 t ∗ . if t ∈ (t ∗ . and u∗ (t) = 1 0 if t ∈ [0. The optimal pair (x ∗ (t). so H ∗ (T ) = −u∗ (T )2 + p(T )(−x ∗ (T ) + u∗ (T )) = 3x ∗ (T )2 = ∂V /∂T . so p(1) = p(t) + t − 1 for all t. u) and the scrap value function S(x) = −x 2 is concave in x. we get p(1) = t ∗ . © Arne Strøm. The following conditions are therefore sufﬁcient for optimality: 1 (i) H3 (t. and since p(T ) = −2x ∗ (T ). We know that p(t) = p(0) − t = p(1) + 1 − t. x. t ∗ ] if t ∈ (t ∗ .10.
so λ(t) = Aert . The solution of this linear u ˙ ∗ (0) = 10. so (i) (ii) (iii) x ∗ (t) = A rt (e − 1) + at 2r (iv) 1 1 1 From (ii). and u > 0. λ(t) = 2 e0. ˙ The initial condition x ∗ (0) = 0 gives B = −A/2r. λ(t)) = −0. λ) = −(x − u)2 + λ(u − x + a) is concave in (x. ˙ ˙ From (ii) we get λ(t) = 0. Thus we must have A = 1/2. x ∗ (T ) = − 2 λ(T ) = − 2 AerT . ˙ ˙ From (i) and (ii) it follows that λ(t) = rλ.2λ = −(H c )x (t.1t 2e © Arne Strøm. x. 1 −0. with λ(t) = 2 e0.1t . so x ∗ (t) = (A/2r)ert +at +B.1x − u) is concave in (x. and Peter Hammond 2008 .1t /A. u∗ . x ∗ .11. Solving for A yields 2arT A = − rT e (1 + r) − 1 The expressions for λ(t).e.1x ∗ (t) − e−0.1λ. differential equation.78 9 CONTROL THEORY: BASIC TECHNIQUES 9. u. From (iii)(c) we see that we must have A ≥ 0. x ∗ (t) = 0. The following conditions are therefore sufﬁcient for an admissible pair (x ∗ . and so A ≥ 1/2.2t  ≤ M for all t ≥ 0.1x ∗ (t) − u∗ (t). and u∗ (t) follow easily.1t + 5e−0.1t ≥ 0 for all t ≥ 0.11 9. See the answer in the book.1t ) = −5 limt→∞ e−0. for some constant A. i. x ∗ (0) = 10.1t /A) = −(10A − 5) − 5e−0. But then x ∗ (t) = u∗ (t)−x ∗ (t)+a = 2 Aert +a.2t (−x ∗ (t)) = Ae0. ˙ λ(t) − rλ = −∂(H c )∗ /∂x = 2(x ∗ (t) − u∗ (t)) + λ(t). ˙ λ(t) − 0. Condition (i) yields ∗ (t) = 1/λ(t) = e−0.2t  =  2 e0.6 The current value Hamiltonian H c (t. 9. Knut Sydsæter. x ∗ .1t e−0. However.1t /A. The scrap value function S(x) = −x 2 is concave in x.10. i. is easily seen to be x ∗ (t) = (10 − 5/A)e0. Atle Seierstad. (iv) x ∗ (t) = 0. u∗ . (i) (ii) (iii) λ(t)e−0.2t The limit of this expression can be ≥ 0 only if 10A − 5 ≤ 0. First. (iii)(b) is also satisﬁed. where A is a constant. (iii)(a) is satisﬁed because we 1 1 have limt→∞ 2 e0. (b) There exists a number M such that λ(t)e−0. with x Condition (iv) requires limt→∞ x ∗ (t) = limt→∞ [(10 − 5/A)e0. limt→∞ x ∗ (t) ≥ 0.1t .2t (−(10 − 5/A)e0. λ(t)) = 1/u∗ (t) + λ = 0.1t . Then from (iv). u). and λ(T ) = −2x ∗ (T ). Finally. x ∗ (0) = 0. x ∗ (t) = 10e−0. A ≤ 1/2. 1 Then u∗ (t) = 2e−0.1t /A] ≥ 0. u).2t (−10e−0.1t e−0. referring to (iii)(a). (c) There exists a number t such that λ(t) ≥ 0 for all t ≥ t .2t  = 1 1 ≤ 2 for all t ≥ 0.e. x ∗ (t) = u∗ (t) − x ∗ (t) + a.1t + 5e−0.2t (−x ∗ (t)) ≥ 0. This obviously requires 10 − 5/A ≥ 0.1t e−0. We have therefore found the solution.2t = 0. Then (i) yields 1 1 x ∗ (t)−u∗ (t) = − 2 Aert . It remains to check that the conditions in (iii) are satisﬁed. λ(t) = −2(x ∗ (t) − u∗ (t)). Since λ(t)e−0.1t − 5e−0.1t /A. x ∗ (t). and thus λ(t) = Ae0.1 The current value Hamiltonian H c = ln u − λ(0.1λ(t). u∗ ) to solve the problem: ∂(H c )∗ /∂u = 0 . and so (iv) with t = T yields − 2 AerT = (A/2r)(erT −1)+aT . so the following conditions are sufﬁcient for optimality: (H c )u (t. (a) limt→∞ λ(t)e−0.1t .
From (i) we see that u∗ (t) = 1 if et < 1. ln 2] we have λ(t) + (e−t − 1)λ(t) = −1. ˙ On this interval u∗ (t) = 0 and λ(t) − λ(t) = −2.t. ln 2].2 The current value Hamiltonian (with λ0 = 1) is H c = x(2 − u) + λuxe−t . Note that λ(t)e−t = 2e−t = 1 when t ∗ = ln 2. (The answer in the book is wrong. Then from (iv). x ∗ (t). and so λ(t) = Cet + 2.t. ˙ From (ii) it follows that λ(t) = Aet + 1. (c) There exists a number t s.1 (a) The current value Hamiltonian H c = ax − 2 u2 + λ(−bx + u) is concave in (x. u∗ (t) = 0 on (ln 2. ∞) we have x ∗ (t) = e1−e = e1−1/2 = e1/2 . It does not change the problem if x(∞) ≥ 0 is added as a restriction in the problem.) 9. λ(t)e−t  ≤ M for all t ≥ 0. ∗ (t)/x ∗ (t) = −t dt. ∞). ˙ On [0. so we have found the optimal solution. 1]. ∞).11. x ∗ (t) ≥ 1 for all t. The conditions in (iii) are obviously satisﬁed. 0] and x ∗ (t) = e − 1 in (0. 0] and u∗ (t) = 0 in (0.4 The current value Hamiltonian H c = (x − u) + λue−t is concave in (x. λ(t)) = −1. so the following conditions are sufﬁcient for optimality: © Arne Strøm. We use formula (5. u). x ∗ (0) = 1. Then x ∗ (t) = x ∗ (t)e−t on [0. and with x ˙ dx e −t − ln 2 x ∗ (0) = 1 this gives x ∗ (t) = e1−e . (a) limt→∞ λ(t)e−t (−x ∗ (t)) ≥ 0. ˙ λ(t) − λ(t) = −∂(H c )∗ /∂x = −(2 − u∗ (t)) − λ(t)u∗ (t)e−t . We have obtained the same answers as in the book. keeping in mind that p(t) = λ(t)e−t . ∞). so ln x ∗ (t) = −e−t + A. (b) There exists a number M s. or λ(t)e−t = C + 2e−t .12.9 CONTROL THEORY: BASIC TECHNIQUES 79 9. u). ˙ ∗ (t) = 0 on (ln 2.3 are trivially satisﬁed.12 1 9. and the expression in (A) reduces to limt→∞ 2e−t (0 − e1/2 ). Then λ(t)e−t < 1 on some maximal interval (t ∗ . Then λ(t) = 2. Then (B) and (C) in Note 9. x ∗ . Looking at (i) we see that ˙ u (t) = ∗ We guess that p(t) = λ(t)e−t → 0 as t → ∞. The problem is not affected by introducing the requirement that limt→∞ x ∗ (t) ≥ 0. λ(t)) = 2x ∗ (t) + x ∗ (t)(e−t λ(t) − 1)u for u ∈ [0. On (t ∗ . ∞). λ(t) ≥ 0 for all t ≥ t . so A = 0.11. So the following conditions are sufﬁcient for optimality: (i) (ii) (iii) u = u∗ (t) maximizes x ∗ (t) + e−t (λ(t) − et )u for u ∈ [0. The Arrow concavity condition holds in this problem and this yields optimality also in the inﬁnite horizon case. ∞). and the following conditions must be satisﬁed for (x ∗ (t). u∗ (t)) to be optimal: (i) (ii) (iii) u = u∗ (t) maximizes H c (t. 1].4. u.6) with a(t) = e−t − 1 and b(t) = −1. x ∗ (t) = u∗ (t)x ∗ (t)e−t . x ∗ (−1) = 0. ˙ λ(t) − λ(t) = −(H c )x (t. It follows that u∗ (t) = 1 in [−1. 9. Atle Seierstad. ln 2]. which is clearly 0. Then a(t) dt = (e−t − 1) dt = −e−t − t and λ(t) = Cee −t +t + ee −t +t e−e −t −t (−1) dt = Cee −t −t +t − ee −t +t e−e e−t dt = Cee −t −t +t − et because the last integral is obviously equal to e−e . We guess that p(t) = e−t λ(t) = e−t (Aet + 1) = A + e−t → 0 as t → ∞.11. Since λ(ln 2) = 2 we ﬁnd that C = 2e−1/2 . ˙ 1 0 if λ(t)e−t > 1 if λ(t)e−t < 1 Since x ∗ (t) ≥ 0 and x ∗ (0) = 1. which ∗ ∗ tends to 0 as t → ∞ only if C = 0. ln 2]. On [0. Knut Sydsæter. and we propose that u∗ (t) = 1 on [0. we get x ∗ (t) = e − e−t in [−1. and u∗ (t) = 0 if et > 1. (iv) x ∗ (t) = u∗ (t)e−t . u∗ . and Peter Hammond 2008 .
and x = x ∗ and λ must satisfy (∗) The equilibrium point is (x. −1) is a saddle point. condition (iii) is evidently satisﬁed. λ) = 2x − 2 The Jacobian matrix of (∗) is (∗) Fx Fλ 1 1 = . 40) is J = and J  = −1/400 < 0. Since Gx Gλ 0 b+r they are real and of opposite signs. λ) ˙ From (i) we get u∗ (t) = λ(t). x ∗ (t) = x ∗ (t) − u∗ (t). ∂(H c )∗ /∂C = 0 implies 1/C ∗ − λ = 0. λ) = x − u = x + λ ˙ ˙ λ = G(x.) 1/20 −1 (b) The Jacobian matrix evaluated at (400. But 0 then ∂V /∂x0 = 0 ae −(b+r)t dt = a/(b + r) = λ(0).12. (The ﬁrst equation is part of the problem. because if the limit is < 0. and thus x ∗ (t) = −bx ∗ (t) + λ(t). limt→∞ λ(t)e−rt (x(t) − x ∗ (t)) ≥ 0 for all admissible x(t). © Arne Strøm.3 (a) With H c = ln C + λ(AK α − C).80 9 CONTROL THEORY: BASIC TECHNIQUES (i) (ii) (iii) (iv) (H c )u (t.3. λ(t)) = −u∗ (t) + λ(t) = 0. u∗ (t). a/(b + r). The differential equation the maximum condition reduces to (i) (H u ˙ ˙ for λ is (ii) λ(t) − λ(t) = −(H c )x (t. c ) (t. For sufﬁcient conditions. Taking ˙ ˙ ˙ ln of each side and differentiating w. Of course. x ∗ (t) = −bx ∗ (t) + u∗ (t). u∗ . √ √ 9.12. then the second equation in (∗) holds. u). the equilibrium point is a saddle point. To if and only if A = 0. λ) = (a/b(b + r). x ∗ (0) = 10. and Peter Hammond 2008 . so the −1/400 0 equilibrium point is a saddle point. ˙ ˙ λ = (b + r)λ − a = G(x. we can restrict attention to admissible x(t) satisfying limt→∞ x(t)e−rt ≥ 0. Also. we ﬁnd x ∗ = Ae(1+ 2 )t + (1 − A)e(1− 2 )t . Then condition (iii) is satisﬁed and we have found the optimal solution.12. x ∗ (t). λ(t)) = −u∗ (t)−λ = 0. (c) If K0 = 100 and T = ∞ the solution curve converges towards the equilibrium point. (c) V = 0 ∞ 1 [ax ∗ (t) − 2 (u∗ (t))2 ]e−rt dt = ∞ ∞ ax0 e−(b+r)t dt + terms that do not depend on x0 . u∗ .2 From the answer to Problem 9. note that if limt→∞ x 2 (t)e−2t = limt→∞ (x(t)e−t )2 = 0. λ(t)) = −a + bλ(t). equivalently. It follows that if K = K ∗ (t) > 0 and C = ˙ −λαA(K C ∗ (t) > 0 solve the problem. 9. ∗ (t). then the objective function is −∞.4 (a) The current value Hamiltonian H c = −(x−1)2 − 2 u2 +λ(x−u) is concave in (x.9. Since u ∈ . x ∗ . x ∗ (t). ˙ λ(t) − rλ(t) = −(H c )x (t. We can therefore restrict attention to admissible solutions for which limt→∞ x(t)e−t = 0. then the value of the criterion is −∞.11. with eigenvalues −b and b + r. Knut Sydsæter. u∗ (t). For such x(t).1√with x(0) = 1. 1 9. Atle Seierstad. λ(t)) = 2x ∗ (t) − 2 − λ. The adjoint variable √ √ √ (1− p(t) = λ(t)e √ 2 )t (with p(t) = − 2 e(−1− 2 )t ). λ − rλ = −∂(H c )∗ /∂K = ∗ )α−1 or.t. ˙ x = −bx + λ = F (x. λ). λ/λ = r − αA(K ∗ )α−1 . see Note 9. Then x ∗ = e(1− 2 )t and λ(t) = − 2 e prove sufﬁciency. or C ∗ λ = 1. To prove sufﬁciency. The determinant is −2. t yields C ∗ /C ∗ + λ/λ = 0. x ∗ . and √ √ √ −2t tends to 0 as t → ∞ λ = A 2 e( 2+1)t − (1 − A) 2 e(1−√ 2 )t .r. so the equilibrium point 2 0 Gx Gλ (1. ¯ ¯ Fx Fλ −b 1 (b) The Jacobian matrix of (∗) is = . The eigenvalues are −1 and 2. u∗ (t)) must satisfy It follows that the optimal pair x x = F (x.
then u∗ (t) = 0 and x1 (t) = 0 with x1 (T −2/a) = x1 eaT −2 . because if lim −t < 0 the objective function is −∞. Thus p1 (t) strictly decreases from a level higher than 1 at t = 0 to 0 at t = T . T − 2/a]. It remains to prove that condition (d) in Theorem 9. g(x2 ) = F (x2 . Looking 1 at (i) we see that u∗ (t) = 1 for p1 (t) > 1 and 0 for p1 (t) < 1. u∗ ) 1 2 to be optimal: (i) (u1 . Note that from x(t) = −(1/2)e−t + 1 and λ(t) = e−t − 1 we ﬁnd that (x. In the same way.2 Suppose T < 2/a. 1) and choose u1 . Note that the integrand in the 1 objective function can be written (−x 2 + 2x − 1 − 2 u2 )e−t . The general solution of this equation is x(t) = Ae−t + Be2t + 1.3 To prove that g is concave. ˙∗ ∗ 0 ∗ 0 If follows that x1 (t) = x1 eat .10 CONTROL THEORY WITH MANY VARIABLES 81 ˙ (b) Solving the ﬁrst equation in (∗) for λ yields λ = x − x. 1]. so we need only consider admissible x(t) −t ≥ 0. u2 ) = λg(x1 ) + (1 − λ)g(x2 ). 1]. The following conditions are sufﬁcient for the admissible quadruple (x1 . 1 ∗ = T − 2. we ﬁnd the corresponding solution for λ.2. and then λ = x − x. u1 ).11.2. we get x2 (t) = x2 + ax1 eaT −2 (t − (T − 2/a)). u2 ∈ [0. and so x ∗ (t) = x 0 . λ ∈ (0. Note that p1 (0) = 2 T > 1 5 since T > 5.1 10. Note that λu1 + (1 − λ)u2 ∈ Uλx1 +(1−λ)x2 . u2 ) = (u∗ (t). so x1 (t) = x1 eaT −2 . so integration gives x2 (t) = ax1 eaT −2 t + B.1 is satisﬁed. u2 in Ux such that g(x1 ) = F (x1 . Since p1 (t ∗ ) = 1 when 2 (T − t ∗ ) = 1.3 (a) There are two state variables x1 and x2 . ˙ 1 1 We see immediately from (ii) that p1 (t) = 2 (T − t) and p2 (t) = 1 (T − t). u1 ) + (1 − λ)F (x2 . Moreover. so we introduce two adjoint variables p1 and p2 . λu1 + (1 − λ)u2 ) ≥ λF (x1 . In fact. p (0) = 1 T > 1 since T > 5. ˙ λ(t) = −2Ae−t − (A + 1/2)e2t − 1 If A = −1/2 we see that x(t) and λ(t) both diverge as t → ∞. In particular. If t ∈ [0. There are 1 also two control variables u1 and u2 . p2 (t) = −(Hx2 )∗ = −1/5. then u∗ (t) = 1 and then x1 (t) = ax1 (t) with x1 (0) = x1 . with the boundary condition ˙∗ ∗ 0 ∗ 0 0 x2 (T − 2/a) = x2 determining B. and Peter Hammond 2008 . a(1 − u 1 2 2 ∗ 0 ∗ 0 If t ∈ (T −2/a. λ(t)) satisﬁes λ = −2x + 1. u2 ). p1 (T ) = 0. for which limt→∞ x(t)e t→∞ x(t)e 10 Control Theory with Many Variables 10. 10. x2 . Atle Seierstad. T ]. p2 (T ) = 0. ˙∗ 0 ∗ 0 Moreover. Thus p (t) strictly decreases from we get t 2 2 5 © Arne Strøm. it follows that x(t) = −(1/2)e−t + 1 → 1 and λ(t) = e−t − 1 → −1 as t → ∞. x2 . u∗ . x1 (T − 2/a) = x1 eaT −2 .2 ∗ ∗ 0 10. λ) = (x(t). Knut Sydsæter. Then g(λx1 + (1 − λ)x2 ) ≥ F (λx1 + (1 − λ)x2 . let x1 . The Hamiltonian H = 2 x1 + 1 x2 − u1 − u2 + p1 u1 + p2 u2 is linear 5 ∗ ∗ and hence concave. Inserted into the second ˙ ¨ ˙ equation in (∗) yields x − x − 2x = −2. ¨ ˙ and x(0) = 1/2 yields B = −A − 1/2. 1 2 ˙ (ii) p1 (t) = −(Hx1 )∗ = −1/2.1. so x(t) = Ae−t − (A + 1/2)e2t + 1 Using λ = x − x. 10. u∗ (t)) maximizes (p1 (t) − 1)u1 + (p2 (t) − 1)u2 for u1 ∈ [0. we have x2 (t) = ˙∗ ∗ (t))x ∗ (t) = 0. For A = −1/2. x2 (t) = ax1 eaT −2 .
u1 + u2 ≤ 1}. The control region is U = {(u1 . p2 (t) = −(Hx2 )∗ = −1. u∗ ) to be admissible and optimal: 1 2 (i) (u1 .) Note that 1 ϕ(0. 1). 0) ≷ ϕ(0. x2 . we ﬁnd that a set of optimal controls are ⎧ ⎪ (1. and (0. we have p1 (T ) = S1 (x1 (t). The Hamiltonian and the differential equations for p1 and p2 are the same as in (a). Integrating and using p1 (T ) = 0. ϕ(1. p1 (T ) = 0. and Peter Hammond 2008 . u2 ) = max 1 2 ab(T − t)2 − c u1 + a(T − t) − c u2 s. T − c/a] (∗∗) ⎪ ⎩ (0. (b) The scrap value function is S(x1 . u2 (t)) = (0. t = T − c/a. x2 (t)) = 2. u2 ) ∈ U .5 (C)(c’). (In some cases there may be more than one maximum point. and condition (i) in (a) implies that u∗ (t) = u∗ (t) = 1 for all t ∈ [0. Since ϕ is a linear function it will attain its maximum over U at one of those corners. The following conditions are ∗ ∗ sufﬁcient for a quadruple (x1 . Knut Sydsæter. We conclude that 5 u∗ (t) = 1 1 0 if t ∈ [0. In fact. x2 (t)) = 3 and p2 (T ) = S2 (x1 (t). 0) ⇐⇒ a(T − t) − c ≷ 0 ⇐⇒ t ≶ T − c/a ϕ(1. 1 we get p1 (t) = 2 b(T − t)2 . according to ∗ ∗ Theorem 10. Therefore p1 (t) = −b(T − t). the optimal values of u1 and u2 are uniquely determined. for t = T − 2/b. 0) if t ∈ [0. x2 ) = 3x1 + 2x2 . but even then at least one corner will be a maximum point. √ and then c/a < 2c/ab < 2/b). 0). 0 ≤ u2 . u2 ) ∈ U (∗) The control region U is the closed triangular region with corners at (0. and ϕ(0.4 The Hamiltonian H = x2 + c(1 − u1 − u2 ) + p1 au1 + p2 (au2 + bx1 ) is linear and hence concave. p2 (T ) = 0. 1) ⇐⇒ 1 2 ab(T − t)2 − c ≷ a(T − t) − c ⇐⇒ t ≶ T − 2/b 2c/ab ⇐⇒ t ≶ T − ϕ(0. 0) ≷ ϕ(0. 1) ≷ ϕ(0. T ]. T − 2/b] ⎨ ∗ ∗ (u1 (t). 1) = a(T − t) − c. For t ∈ [0. (u1 . u2 ) = (u∗ (t). To ﬁnd the optimal controls we must for each t in [0. u2 ) : 0 ≤ u1 . 1 2 (ii) p1 (t) = −(Hx1 )∗ = −bp2 (t). See the answer in the book. We see that (with t < T ) ϕ(1. u∗ . T ] solve the problem max ϕ(u1 . The transversality conditions are changed.2. 0). T ] we see that p1 (t) and p2 (t) are both greater than 1. Atle Seierstad. x2 (t)) are easily worked out. 0) if t ∈ (T − c/a. T − 5] if t ∈ (T − 5. ˙ ˙ ∗ 0 ∗ ∗ 0 (iii) x1 (t) = au∗ (t). T ] ∗ ∗ The corresponding values for (x1 (t). 1 1 Then p1 (t) = 3 + 2 (T − t) and p2 (t) = 2 + 5 (T − t).t. (1. 0) = 2 ab(T − t)2 − c. x1 (0) = x1 . x2 (0) = x2 .1. T − 2] if t ∈ (T − 2. 10. Looking at (i) we see that u∗ (t) = 1 for p(t) > 1 and 0 for 2 p(t) < 1. 1) if t ∈ (T − 2/b.82 10 CONTROL THEORY WITH MANY VARIABLES a level higher than 1 at t = 0 to 0 at t = T . ˙∗ ˙∗ 1 2 ˙ From (ii) we see that p2 (t) = T − t. T ] (For values of t in the corresponding open intervals. Since p2 (t ∗∗ ) = 1 when 1 (T − t ∗∗ ) = 1. and thus 2 1 ∗ ∗ x1 (t) = x2 (t) = t. 0) = 0. x2 (t) = au∗ (t) + bx1 (t). T ] u∗ (t) = 2 1 0 if t ∈ [0. 0) ⇐⇒ T − t ≷ 2c/ab Putting all this together (note that the assumption T − c/a > T − 2/b in the problem implies c/a < 2/b. In the chains of equivalences below it is understood that either the top inequality holds all the way or the bottom inequality holds all the way. u∗ (t)) maximizes (ap1 (t) − c)u1 + (ap2 (t) − c) for (u1 . and t = T we choose the values so that the control functions © Arne Strøm. we get t ∗∗ = T − 5.
u). 1 ˙ p2 (t) = c(t −T ). (ii) p1 (t) = −∂H ∗ /∂x1 = −1 − bp2 (t).) 1 We then get p1 (0) = T (1 − 2 bcT ) < 0. Moreover. and Peter Hammond 2008 for t in [0.5 There are two state variables x1 and x2 . since p1 (T ) = p2 (T ) = 0. Note that p1 (t) is a quadratic polynomial in t.5 There will be two points t∗ and t∗∗ . x2 . T − 2/b) get very messy.2 the ∗ ∗ admissible triple (x1 . T ] This gives ∗ The expression for x2 (t) is somewhat messy: ⎧ 0 0 ⎨ bx1 t + x2 ∗ 1 1 0 0 x2 (t) = 2 bt 2 + b(x1 − t∗ )t + x2 + 2 b(t∗ )2 ⎩ 1 0 0 b(t∗∗ − t∗ )t + bx1 t + 2 b[(t∗ )2 − (t∗∗ )2 ] + x2 © Arne Strøm. T ]. 10.5. The Hamiltonian is H (t. T ] ⎧ 0 ⎨ x1 ∗ 0 x1 (t) = t − t∗ + x1 ⎩ 0 t∗∗ − t∗ + x1 for t in [0. The maximum value of ˙ p1 is p1 (t1 ) = 1/2bc. which gives p1 (t) = T −t − 2 bc(t −T )2 = 1 1 − 2 bct 2 + (bcT − 1)t + T − 2 bcT 2 . for t in (t∗ .2.1. x1 (t). t∗ ]. p1 (T ) = 0 ˙ and p2 (t) = −∂H ∗ /∂x2 = c. t∗∗ ] for t in (t∗∗ . t∗ ]. such that 0 < t∗ < t∗∗ < T and p1 (t∗ ) = p1 (t∗∗ ) = 1. (The other cases are much simpler. u0 ] (where does not depend on u).10 CONTROL THEORY WITH MANY VARIABLES 83 ∗ ∗ become leftcontinuous. p 1 t t∗ t∗∗ T Figure 10. t∗∗ ] . t∗∗ ] for t in (t∗∗ . p1 (t). Now. for t in (t∗ . so according to Theorem 10.2. x2 . p2 ) = x1 − cx2 + u0 − u + p1 u + p2 bx1 . x1 . p2 (T ) = 0 ˙ From (i) we see that p1 (t) > 1 ⇒ u∗ (t) = u0 . Atle Seierstad. ∗ ∗ ∗ ∗ (i) u = u∗ (t) maximizes H (t. p2 (t)) = x1 (t) − cx2 (t) + u0 − u + p1 (t)u + ∗ ∗ p2 (t)bx1 (t) = + p2 (t)bx1 (t) + (p1 (t) − 1)u for u in [0. x2 (t). The graph of p is shown in Figure 10.) The Hamiltonian is concave in (x1 . and p1 (t) = −1−bp2 (t) = −1−bc(t −T ). t∗ ].2. p1 (T ) = 0. (We have p0 = 1. T ] . but the expressions on the interval (T − c/a. 1 with p1 (0) = T (1 − 2 bcT ) and maximum at t1 = T − 1/bc (since p1 (t1 ) = 0). t∗∗ = T − 1 − 2bc bc bc bc bc ∗ The optimal control u∗ and the corresponding x1 (t) are given by t∗ = T − ⎧ ⎨0 u∗ (t) = 1 ⎩ 0 for t in [0. Knut Sydsæter. u. x2 . from (ii). since p1 (T ) = 0. for t in (t∗∗ . We restrict our attention to the main case bcT > 2 and 2bc < 1. so we introduce two adjoint variables p1 and p2 . p1 (t1 ) = 1/2bc > 1. u∗ ) is optimal provided for each t in [0. u. for t in (t∗ . 1 1 p1 (t) = 2 bct 2 − (bcT − 1)t − T + 2 bcT 2 + 1 = 0 ⇐⇒ t = T − 1√ 1 ± 1 − 2bc bc bc 1√ 1√ 1 1 − + 1 − 2bc . p1 . p1 (t) < 1 ⇒ u∗ (t) = 0.) The corresponding values of x1 (t) and x2 (t) follow from (iii) and (∗).
1]. u∗ ) is an admissible triple which solves the problem. ¯ Condition (i) gives 1 if t < t ∗ u∗ (t) = 0 if t > t ∗ If t ∗ = 2. x ∗ (t).3.3. x ∗ (0) = 1. if t > 1 x ∗ (t) − 1 = y ∗ (t) = t if t ≤ 1 .84 10 CONTROL THEORY WITH MANY VARIABLES 10. contradicting t ∗ = 0. Note that since x ≥ 0 for all t ≥ 0 and x(0) ≥ 0.7 There are two state variables and one control variable. so p(t) = (1/r)e−rt .3 10. y ∗ (0) = 0. so we have found a solution: u∗ (t) = 1 0 if t ≤ 1 . ∞). Thus. p1 . In fact. u∗ (t)) is optimal. There is no restriction on x(t) as t → ∞. y ∗ (2) = t ∗ = 3/2 + p2 . and since p1 (2) = 0. u). y ∗ (t). 1 p2 (t) = − 2 10. and then we see that (draw a picture!) u∗ (t) = 1 0 if t ∈ [0. which is impossible. because of (iii). we see that u = u∗ (t) maximizes (1/r)e−rt (e−t − r)u for u ∈ [0. Then: (i) For each t in [0. Since y ∗ (2) ≤ 1.2 it remains only to verify (10. − ln r].2. Hence. ˙ (iii) x ∗ (t) = u∗ (t)e−t . Then we must have C = 0. (b) p2 (2) ≤ 0. Atle Seierstad. x ∗ (t). p1 (t). ¯ ¯ we must have 3/2 + p2 ≤ 1. and so p2 = −1/2 and t ∗ = 1. x ∗ (0) = x0 ≥ 0. From (iii). p2 (t)) = x ∗ (t) + (p1 (t) + p2 (t) − 2 )u for u in [0. x. 1]. y ∗ . then u∗ (t) ≡ 1 and y ∗ (t) ≡ t. Hence. y ∗ (t) = u∗ (t). In this case p2 = p2 (2) = 0 ¯ ∗ = 3/2. u. In particular.e. ∗ /∂y = 0. y. p2 (t) = p2 . 1 if t > 1 p1 (t) = 2 − t. we see that limt→∞ p(t)(x0 − x ∗ (t)) = limt→∞ (e−rt /r)(x0 − x0 − 1 + r) = 0.1. (ii) p1 (t) = −∂H ˙ p2 (t) = −∂H ˙ (iii) (a) p1 (2) = 0. (ii) p(t) = −Hx (t. p1 = −1. Knut Sydsæter. ∗ /∂x = −1. ¯ ¯ ¯ The Hamiltonian is concave in (x. The Hamiltonian 1 is H (t. y ∗ (2) = 1. so according to Note 10. y.10a). where ˙ ˙ ¯ 1 1 ¯ ¯ p2 is a constant. 2]. we have x(t) ≥ x0 for all t ≥ 0. We have e−t = r when t ∗ = − ln r. © Arne Strøm. u∗ (t) = 1 if e−t > r and u∗ (t) = 0 ∗ if e−t < r. t Thus we must have 0 < t ∗ < 2. then u∗ (t) ≡ 0 and y ∗ (t) ≡ 0. − ln r] if t ∈ (− ln r. ˙ We have p(t) = e−rt /r ≥ 0. ∞) From (iii) we ﬁnd that x ∗ (t) = x0 + 1 − e−t in [0. (iv) x ∗ (t) = u∗ (t). p(t)) = −e−rt . If (x ∗ (t). i. we have p1 (t) = 2 − t. p1 (t) + p2 (t) − 2 = 2 − t + p2 − 2 = t ∗ − t. Because p2 = 0. It remains to determine p2 and t ∗ .3. ˙ From (ii) we get p(t) = (1/r)e−rt + C. and then y ∗ (2) > 1. With this choice of p(t). where t ∗ = 3/2 + p2 is a constant. Suppose (x ∗ . u = u∗ (t) maximizes 1 H (t. and Peter Hammond 2008 . and then x ∗ (t) − 1 = y ∗ (t) = t for t ≤ t ∗ and x ∗ (t) − 1 = y ∗ (t) = t ∗ for t > t ∗ . 1]. p2 ) = x + (p1 + p2 − 2 )u. while x ∗ (t) = x0 + 1 − r in (− ln r. u∗ (t). which gives y ∗ (2) = 0 < 1. then (i) u = u∗ (t) maximizes (e−t p(t) − e−rt )u for u ∈ [0.1 The Hamiltonian H = (x − u)e−rt + p(t)ue−t is concave in (x. ˙ ˙ From (ii). We shall use Theorem 10. p2 ≤ −1/2 < 0. so we guess that p(t) → 0 as t → ∞.2. If t ∗ = 0. u. and p2 (2) = 0 if y ∗ (2) < 1. u) (in fact linear).
˙ From (iii) we see that x ∗ (t) ≥ 0 so x ∗ (t) ≥ x0 > 0.3. ˙ (iii) x ∗ (t) = u∗ (t)x ∗ (t). 10. Thus (i) implies that ˙ u (t) = ∗ 1 0 if p(t) > x ∗ (t) if p(t) < x ∗ (t) © Arne Strøm. x ∗ (0) = x0 > 0. Note that using the interpretation in Example 10. x ∗ (0) = x 0 > 0. which leads to consumption increasing at a constant rate. If (x1 (t). Then u∗ (t) = 0 for t close to 0. using (10. 1]. ˙2 (ii) p1 (t) = −ap1 (t)u ˙ 2 ∗ ∗ (t)x ∗ (t).4 10. u∗ (t)) is optimal. u ∈ [0. and Peter Hammond 2008 . Then p1 (t) = −ap2 (t) = −(a/r)e−rt and so p1 (t) = (a/r 2 )e−rt + D. Suppose p1 (0) < 1/r. Atle Seierstad. so p2 (t) = e−rt /r. Finally. Then u∗ (t) ≡ 1 and from (iii) we have x2 (t) ≡ 0.2 the total discounted consumption is 0.2 (a) The model is closely related to the model in Example 10.2. so we have found the optimal solution.2 First we apply Theorem 10.2.4. and again we see that p1 (t) > p2 (t) for all t ≥ 0. p(1) = 0. Let us see if we can have u∗ (t) = 0 for all t. ˙ 0 0 Then x2 (t) = (a − b)x1 ebt . From x1 (t) = a(b/a)x1 (t) = bx1 (t). so p1 (t) = p1 (0)e−at > (1/r)e−at > (1/r)e−rt .2 is satisﬁed because ux ≤ x since u ≤ 1.10(a)). and if (x ∗ (t). since r > a.2.1. with 0 < r < b < a. As in Example 10.2.4. (b) Choose the control u(t) = b/a. (ii) p(t) = −(Hx )∗ = −2(1 − u∗ (t))x ∗ (t) − p(t)u∗ (t).) ˙ ˙ If p1 (0) = p2 (0). Again we must ˙ have D = 0 and then p1 (t) = (a/r 2 )e−rt < (1/r)e−rt = p2 (t). Then the objective function is 0 x2 (t)e−rt dt = ∞ ∞ 0 (b−r)t 0 − e−rt ] dt = 0 [(a − b)(1/b)x1 [(1/(b − r)]e(b−r)t + e−rt /r]. with x1 (0) = x1 . x2 (t). From (i) we see that 1 if p1 (t) > e−rt /r u∗ (t) = 0 if p1 (t) < e−rt /r ˙ Suppose p1 (0) > p2 (0) = 1/r. Then u(t) ∈ [0. we ﬁnd that C = ˙ ∞ 0 0 bt −(a − b)(1/b)x1 .1.1]} is convex (a rectangle) by the same argument as in Example 10. we ﬁnd x1 (t) = x1 ebt . x) = {((1 − u)x 2 + γ . Then u∗ (t) = 1 to the immediate right of t = 0 and p1 (t) = −ap1 (t). ux) : γ ≤ 0. The control region U = [0. then ∗ (i) u = u∗ (t) maximizes ax1 (t)(p1 (t) − p2 (t))u for u ∈ [0. then p1 (0) = −ap2 (0) = −a/r > −1 = p2 (0). (In the terminology of Example 10. p (t) = −e−rt . The Hamiltonian is H = x2 e−rt + ∗ ∗ p1 aux1 + p2 a(1 − u)x1 . ∗ (t) − ap (t)(1 − u∗ (t)). x ∗ (0) = 0. There is no restriction on x2 (t) as t → ∞. 1] and we seek corresponding 0 0 admissible state variables. By using 0 [(a − b)(1/b)x1 e 0 < r < b < a we see that the integral diverges.10 CONTROL THEORY WITH MANY VARIABLES 85 10. u∗ (t)) is optimal then: (i) u = u∗ (t) maximizes x ∗ (t)(p(t) − x ∗ (t))u for u ∈ [0.4.4. Then we must have C = 0. The Hamiltonian is H = (1 − u)x 2 + pux = x 2 + ux(p − x). The set N(t.2 the Arrow condition is satisﬁed. x ∗ (t) = a(1 − u∗ (t))x ∗ (t).2.3. so x2 (t) = (a − b)x1 (1/b)ebt + C. ˙2 (iii) x1 (t) = au ˙ 1 1 1 1 2 From (ii) we get p2 (t) = e−rt /r + C. With x2 (0) = 0. and we see that we must have p1 (t) > p2 (t) ∗ for all t ≥ 0. so x2 (t) = (a − b)(1/b)x1 (e − 1).2. Knut Sydsæter. Thus there exists an optimal control. This is not the optimal solution. 1]. we have ∗ ∗ 0 0 p1 (t)(x1 − x1 (t)) = 0 and p2 (t)(0 − x2 (t)) = (1/r)e−rt (−ax1 ) → 0 as t → ∞. so we guess that p2 (t) → 0 as t → ∞. 1] is convex and the condition in Note 10. and the objective function is 0. in this case the discount factor r is so high that it is optimal with no further investment in the investment sector.
and Peter Hammond 2008 . p(t) = p = 0. so t ∗ = 3/4. x = −3 . Then we see that if p(t) > 0. from which it follows that t ∗ = 1/2. x ∗ (t) = −3t + 4. so it means that x 0 ˙ 0 ˙ x ∗ (t) ≥ x ∗ (0) − 3t = 4 − 3t. so p(t) is strictly decreasing. (ii) p(t) = −Hx (t. so t ∗ + 4 = −3t ∗ + 7. D = 141/16. In (t ∗ . so with x ∗ (0) = 4. such that (p0 . . x ∗ (t) − x ∗ (0) = t x ∗ (τ ) dτ ≥ t (−3) dτ = −3t. Since ˙ ˙ ∗ (t) is increasing from the level x > 0 and p(t) is strictly decreasing and is 0 at t = 1. ˙ (iii) x ∗ (t) = 1 − (u∗ (t))2 . But at t ∗ we ˙ ∗ ) = x ∗ (t ∗ ). 1] we have p(t) = −2x ∗ (t) = −2x0 et and p(1) = 0. Then u∗ (t) = 0 2 if t ∈ [0. we have x ∗ (t) = −3t + 7. ˙ ¯ ¯ ¯ Suppose p0 = 0. ˙ then u∗ (t) ≡ 0 and x ∗ (t) ≡ 1.86 10 CONTROL THEORY WITH MANY VARIABLES We claim that p(t) is strictly decreasing. x ∗ (t) = t + 4. 2].4. 3t (Note the misprint in line 3 of the answer to Problem 10. We ﬁnd that have p(t 0 0 p(t) = x0 e1−t in [0. 1]. If t ∗ = 0. and since x ∗ (1) = 4. and then x ∗ (1) = 4. Then u∗ (t) = 0 in this interval. (Formally. Knut Sydsæter. 1] ˙ ˙ In [0. 1] we have x ∗ (t) = −3. Now p(t) is continuous at t ∗ . contradicting p(0) ≤ x(0) = x0 . . if (x ∗ (t). then u∗ (t) ≡ 2 and x ∗ (t) ≡ −3. one should use u∗ (t) = 0. Then u∗ (t) ≡ 2 and we get a contradiction to x ∗ (1) = 4. Atle Seierstad. ∗ (t) is always ≥ 1. In fact. p(t)) = −2p0 x ∗ (t). so 2x et ∗ (1 − t ∗ ) = x et ∗ . so with x ∗ (0) = 4. x ∗ (t) = t + 4. and so p(t) = −t 2 −8t +C. It remains to ﬁnd p(t). then one should use u∗ (t) = 2. An optimal control u∗ (t) must maximize p(t)(1 − u2 ) for u ∈ [−1. there exist a continuous and piecewise differentiable function p(t) and a constant p0 . ¯ ˙ Thus. On the other hand. either 0 or 1. If p(t) > 0 for all t ∈ [0. 0) and (i) u = u∗ (t) maximizes p0 (x ∗ )2 + p(t)(1 − u2 ) for u ∈ [−1. The answer is summed up in the answer in the book. t ∗ ] if t ∈ (t ∗ . we have to have x 0 p(t) < x ∗ (t) in some interval (t ∗ . C = 105/16. x ∗ (0) = x ∗ (1) = 4. 1/2]. and then x ∗ (1) = 4. In (3/4. Since x ∗ (0) = 4. and so p(t) = ˙ 2 − 14t + D. . With t ∗ > 0. and p(t) ≡ −2x0 . ˙ Since p(3/4) = 0. u∗ (t).4. t ∗ ] we have x ∗ (t) = 1 and since x ∗ (0) = 4. 1]. We conclude that the strictly decreasing function p(t) is > 0 in some interval [0. Suppose p(t) < 0 for all t ∈ [0. then u∗ (t) ≡ 0. 1]. 1] we have p(t) = −2x ∗ (t) = 6t − 14.3 The Hamiltonian is H = p0 x 2 + p(1 − u2 ). x ∗ (t) ≡ x0 . 3/4] we have p(t) = −2x ∗ (t) = −2t −8. 1]. 2]. 10. which is ≥ 1 in [0. ˙ But then p(0) = 2x0 . For those t where p(t) < x ∗ (t) we have u∗ (t) = 0 and p(t) = −2x ∗ (t) < 0. for those t where p(t) > x ∗ (t) we have u∗ (t) = 1 and p(t) = −p(t) < 0. We have found the optimal solution. According to the maximum principle. (In neither case should one ˙ use u∗ (t) = −1!) From (iii) and u ∈ [−1. u∗ (t)) is an optimal pair. 2] we see that x ∗ (t) is never less than −3. so p(t) = 2x0 et (1 − t). When u = 2. t ∗ ] and < 0 in (t ∗ . so p0 = 1. . Since p(1) = 0 we get p(t) = 2x0 (1 − t). If p > 0. ) ˙ © Arne Strøm. Then from (ii).3 in the book: . p(t)) = (0. x ∗ (t). and if p(t) < 0. On [0. if p < 0.) From (ii) it follows that p(t) = −2x ∗ (t) < 0. Since p(3/4) = 0. 2]. then u∗ (t) ≡ 0 and again we get a contradiction to x ∗ (1) = 4. and u∗ (t) maximizes −pu2 for u ∈ [−1. assuming p0 = 0 leads to contradictions. 1]. u (t) = ∗ 1 0 if t < t ∗ if t > t ∗ ∗ and x (t) = ∗ x0 et x0 et ∗ if t < t ∗ if t > t ∗ ∗ In (t ∗ .
This is the case because we ﬁnd that q(t) < 0 and q(t ∗ ) = 0. The Lagrangian is L = H + q1 (u − c) + q2 (ax − u) = u + p(ax − u) + q1 (u − c) + q2 (ax − u). Then from (i) u∗ (t) = p(t) − q(t) = 2 − t. Then from (i). we get Be−t ∗ + 1 et ∗ − 1 = 2 − t ∗ . x. ˙ ˙ We guess that u∗ (t) = x ∗ (t) on some interval [0.6.1 (a) H = − 2 u2 − x − pu and L = H + q(x − u). or p(t) + p(t) = −1 + et .6. with x ∗ (0) = 1. Thus q(t ∗ − ) = −e−t − p(t ∗ ) = 0. and (iii) gives p(t) = −1 − q(t) = ˙ so x t . p(t) = −∂L∗ /∂x = −1 − q(t). Atle Seierstad.6 1 10. with q(t) = 0 if x ∗ (t) > u∗ (t).6. so A = et + 2 (t ∗ )2 − 2t ∗ . H is concave in (x.6. we have t ∗ − 2 = Aet − 1 − 2 e−t = 1 − 2 e−t − 1 − 2 e−t = −e−t . x ∗ (0) = 1.1. q(t) ≥ 0. and (iv) gives x ∗ (t) = 2 − t. 10.6. u). ∗ ∗ 1 or Aet = 1 − 2 e−t .5) is here L = H +q(x −u) = 1 1 x − 2 u2 + pu + q(x − u). u) and h1 and h2 are quasiconcave (in fact linear) in (x. since x ∗ (t) u ˙ 2 ∗ ∗ 1 1 is continuous at t ∗ .3 Note that there are two constraints.4). (ii) q(t) ≥ 0 (= 0 if x ∗ (t) > u∗ (t)).44. t ∗ ]. Finally. The Hamiltonian is concave (in fact linear) in (x. h1 = u − c ≥ 0 and h2 = ax − u ≥ 0.1 are (i) ∂L∗ /∂u = −u∗ (t) − p(t) − q(t) = 0. From the last ∗ two equalities we get et = 2 − t ∗ . From (i) we have q(t) = −u∗ (t) − p(t) = −e−t − p(t). u). According to Theorem 10. so x ∗ (t) = −u∗ (t) = −x ∗ (t).1)–(10. t ∗ ≈ 0. we have − 2 (t ∗ )2 + 2t ∗ + A = et . t ∗ ]. From the argument 1+e 2 ∗ ∗ ∗ ∗ 1 in the problem. Note that the Hamiltonian H = x − 2 x 2 + pu is concave (in fact linear) in (x. and so x ∗ (t) = t − 2 and then x ∗ (t) = 1 t 2 − 2t + B. Since ∗ . In (t ˙ ∗ (t) = −p(t) = 2 − t. i. Finally. To conﬁrm that all the conditions (i)–(iv) are satisﬁed. so Be−t ∗ = − 1 et ∗ + 3 − t ∗ . so x ∗ (t) = − 2 t 2 + 2t + A. u) and that h = x − u is quasiconcave (in fact linear) in (x.10 CONTROL THEORY WITH MANY VARIABLES 87 10. u) = x − u is linear and therefore quasiconcave.6.2 This is a problem of the form (10. and since p(2) = 0. This gives B = e−t − 2 (t ∗ )2 + 2t ∗ . ∗ so e−t = 2 − t ∗ . so (iii) gives p(t) = 1 − q(t) = ˙ −t +p(t). u) and h(t. and so (iii) gives p(t) = 1 2 − t. the following conditions are sufﬁcient for (x ∗ (t). ∗ (t) = et = u∗ (t). ∗ ∗ 1 1 we should verify that q(t) = ( 2 e2t + et )e−t − 2 et − 1 is nonnegative in [0. and with x ∗ (0) = 1 we get x ∗ (t) = u∗ (t) = e−t . This linear differential equation has the solution p(t) = Aet −1− 1 e−t . the following conditions are sufﬁcient for an admissible pair (x ∗ (t). and Peter Hammond 2008 . u∗ (t)) to be optimal: There exist a continuous function p(t) and a piecewise continuous function q(t) such that (i) (ii) (iii) (iv) ∂L∗ /∂u = −u∗ (t) + p(t) − q(t) = 0. Knut Sydsæter. we get Be−t = 2 et + 1. (iii) p(t) = 1 − q(t). and then q(t) = Be−t + 2 et − 1 − et = Be−t − 2 et − 1. 1] we have q(t) = 0 and p(t) = 1. u∗ (t)) to be optimal: There exist a continuous function p(t) and piecewise continuous functions q1 (t) and q2 (t) such that © Arne Strøm. In particular. ˙ ∗ ∗ 1 1 Since x ∗ (t) is continuous at t ∗ . p(2) = 0. so p(t ∗ ) = −e−t = Aet − 1 − 2 e−t . and then (iv) gives x ∗ (t) = x ∗ (t). Then from (ii) we have q(t) = 0.6. The Lagrangian (10. By using this relationship you will see that the values of A and B are the same as in the answer in the book. On (t ∗ . Then from (i) we have q(t) = p(t) − et .e. According to Theorem 10. ˙ 10.1. with p(2) = 0. ˙ ˙ (b) In [0. 2] we guess that x ∗ (t) > u∗ (t). q(t ∗ − ) = 0. p(t) is also continuous at t 2 2 ∗ 1 ∗ since q(t ∗− ) = 0 (see the argument in the previous problem). ∗ . The solution of this linear differential equation is p(t) = −1 − p(t) + e ˙ 1 1 1 Be−t + 2 et − 1. we get p(t) = t − 2. The conditions in Theorem 10. x ∗ (t ∗ ) = 2 (t ∗ )2 − 2t ∗ + B = e−t . ∗ ∗ ∗ ∗ ∗ 1 1 1 since p(t) is continuous at t ∗ . ˙ x ∗ (t) = u∗ (t).6. The solution is summed up in the answer in the book.6. t ∗ ] we have x ∗ (t) = u∗ (t).
Knut Sydsæter. t ] if t ∈ (t . p(t) = −∂L∗ /∂x = −ap(t) − aq2 (t). (x) Case A: x ∗ (T ) > xT .6. x ∗ (0) = x 0 . t ] if t ∈ (t . Then from (viii) we get (x 0 − c/a)eat + c/a = xT . t ] if t ∈ (t . The formula for p(t) in (ix) gives a(tb − T ) + 1 ≥ 0. replace x∗ by x ∗ and t ∗ by t . We must also check that x ∗ (T ) > xT . T ] . ˙ c ≤ u∗ (t) ≤ ax ∗ (t) In the similar model in Example 10. T ] q2 (t) = 0 a(t − t ) if t ∈ [0. all with t replaced by T − 1/a.2 it was optimal to start out with u∗ (t) = c in an initial interval. ˙ p(T ) ≥ 0 with p(T ) = 0 if x ∗ (T ) > xT . The solution with t = tB is valid if the last inequality is satisﬁed. The ﬁnal conclusion is given in the answer in the book. p(t) by (ix). In particular. Then we get p(t) = ˙ −ap −a if t ∈ [0. or tB ≥ tA . with q1 (t) = 0 if u∗ (t) > c. so let us try the same here: u∗ (t) = c ax ∗ (t) if t ∈ [0. It is a useful exercise to check that all the conditions in (i)–(vii) are satisﬁed. check that q1 (t) and q2 (t) are both ≥ 0. which solved for t gives t = tB . t ] if t ∈ (t . (viii) In the interval [0. Atle Seierstad. with q2 (t) = 0 if ax ∗ (t) > u∗ (t).88 10 CONTROL THEORY WITH MANY VARIABLES (i) (ii) (iii) (iv) (v) (vi) (vii) ∂L∗ /∂u = 1 − p(t) + q1 (t) − q2 (t) = 0. T ] and p(t ) = 1 ⇒ p(t) = e−a(t−t ) a(t − t) + 1 if t ∈ [0. In the interval (t . and then keep x ∗ (t) constant. We see from (viii) that this is equivalent to (x 0 − c/a)ea(T −1/a) + c/a > xT . Then (i) gives q2 (t) = 1 − p(t). t ] if t ∈ (t . which inserted into (iv) gives p(t) = −a. T ] we have u∗ (t) = ax ∗ (t) = (ax 0 − c)eat + c > c. In fact from (i). since q2 (t −q1 (t − ) ≤ 0 and since q1 (t + ) = 0. where xT − c/a 1 (xi) tB = ln a x 0 − c/a Note that from (v) we need to have p(T ) ≥ 0. (In line 2 of the answer to Problem 10. and q1 (t) and q2 (t) by (x). Because p(t) is continuous. q1 (t) ≥ 0. x ∗ (t) = ax ∗ (t) − u∗ (t). t ] if t ∈ (t . so we have p(t) = −ap(t). and then (iii) gives q2 (t) = 0. Case B: x ∗ (T ) = xT . we have 1 − p(t − ) = We claim moreover that p(t ) must be 1. p(t ) = 1.) © Arne Strøm. x ∗ (T ) ≥ xT . or tA > tB . T ] . so ˙ according to (ii). where tA = T − 1/a. so t = tA . t ] we have u∗ (t) < ax ∗ (t) because c < (ax 0 − c)eat + c.3. with tB deﬁned in (xi) below. T ] (ix) The corresponding values of q1 (t) and q2 (t) are q1 (t) = e−a(t−t ) − 1 0 if t ∈ [0. ˙ − ) = 0. The optimal solution is given by (viii). T ] then x ∗ (t) = (x 0 − c/a)eat + c/a (x 0 − c/a)eat + c/a if t ∈ [0. q2 (t) ≥ 0.6. q1 (t) = 0. Then p(T ) = 0 and thus a(t − T ) + 1 = 0. we have 1 − p(t + ) = q2 (t + ) ≥ 0. and Peter Hammond 2008 .
Atle Seierstad. p(t) = −∂L∗ /∂x = −1 − p(t) + q3 (t).7. ln 2] in the present problem too.7. with q2 (t) = 0 if u∗ (t) > −1. u) and h1 = 1 − u.4 The Lagrangian is L = H + q1 (1 − u) + q2 (1 + u) + q3 (2 − x − u) = x + p(x + u) + q1 (1 − u) + q2 (1 + u) + q3 (2 − x − u). p(ln 2) = 1 − ln 2. q1 (t) = p(t). 1). u∗ (t)) to be optimal: © Arne Strøm. i. 1). ln 2]. and h3 = 2 − x − u are quasiconcave (in fact linear) in (x. Here H is concave in (x. ˙ If we disregard the constraint x + u ≤ 2. 1) on [0.10 CONTROL THEORY WITH MANY VARIABLES 89 10.4. whose solution was (x(t). q2 (t). In fact. We know that the suggested solution is admissible. so p(t) = 1 − t.6.7 10.1 We maximize 0 (−u − x) dt. with x ∗ (t) = 2 on [ln 2. u∗ (t)) = (et − 1. (x ∗ (t). h2 = 1 + u. q3 (t) ≥ 0. ln 2]. (x ∗ . We guess that this is the optimal solution on the interval [0. u∗ (t)) satisﬁes all the conditions (i)–(vi). and Peter Hammond 2008 5 . u∗ (t)) is optimal. 1 + 2 ln 2 − 2t). p(t) = −1 with p(1) = 0. looking at the objective function. ˙ In the interval [0. 1]. i. (x ∗ (t). Then (i) gives p(t) = q3 (t) and from (v). p(1) = 0. The Hamiltonian H = x + p(x + u) is concave (in fact linear) in (x. with q1 (t) = 0 if u∗ (t) < 1. Then from (iii) and (iv). Now. 1] et − 1 2t + 1 − 2 ln 2 x ∗ (t) 1 1 + 2 ln 2 − 2t p(t) (4 − 2 ln 2)e−t − 1 1−t q1 (t) (4 − 2 ln 2)e−t − 1 0 q2 (t) 0 0 q3 (t) 0 1−t Having checked that (x ∗ (t). so the Lagrangian is L = H + qx = −u − x + p(u − t) + qx. 1) when t ∈ (ln 2. for t > ln 2 it seems optimal to increase x(t) as fast as the constraint x + u ≤ 2 allows. In particular. q1 (t) ≥ 0. Then (v) gives p(t) = −1 − p(t). At t = ln 2 we have x ∗ (ln 2) = eln 2 − 1 = 1. 1]. It remains to ﬁnd appropriate multipliers satisfying (i)–(v). x ∗ (0) = 0. and it is easy to verify that u∗ (t) = 1 + 2 ln 2 − 2t takes values in (−1.1. The complete suggestion for an optimal solution is therefore: u∗ (t) t ∈ [0. ln 2). The suggestion we have for an optimal solution is therefore: In [0.e. we get x ∗ (t) = 2t + 1 − 2 ln 2. ln 2] ˙ with p(ln 2) = 1 − ln 2 gives p(t) = (4 − 2 ln 2)e−t − 1. According to Theorem 10. with q3 (t) = 0 if u∗ (t) + x ∗ (t) < 2. So we start out with (x ∗ (t). 1]. Then from (i).e. u). q2 (t) ≥ 0. Note that q1 (t) and q3 (t) have jump discontinuities at t = ln 2. so by Theorem 10.1. x) = x is quasiconcave. this is the problem solved in Example 9. 1). such that (i) (ii) (iii) (iv) (v) (vi) ∂L∗ /∂u = p(t) − q1 (t) + q2 (t) − q3 (t) = 0. 10. q2 (t) = q3 (t) = 0. u(t)) = (et − 1. q1 (t) = q2 (t) = 0. in (ln 2. u∗ (t)) to be an optimal pair: there exist a continuous function p(t) and piecewise continuous functions q1 (t). so from (ii) and (iii). ln 2] t ∈ (ln 2. and q3 (t). 1]. ˙ x ∗ (t) = x ∗ (t) + u∗ (t). In the interval (ln 2. putting x ∗ (t) = u∗ (t) + x ∗ (t) = 2. u∗ ) is admissible. we will try to guess the optimal solution and then verify its optimality by checking that all the conditions (i)–(vi) are satisﬁed. u∗ (t)) = (2t + 1 − 2 ln 2. u∗ (t)) = (et − 1. u) and h(t.6. Note that in this case x(t) + u(t) = et ≤ 2 as long as t ≤ ln 2. u∗ (t) = 1 > −1 and x ∗ (t) + u∗ (t) = et < 2. and. u∗ (t) ∈ (−1. and x ∗ (ln 2) = 1.1. Solving the linear differential equation on [0. the following conditions are sufﬁcient for (x ∗ (t). Then ˙ u∗ (t) = 2 − x ∗ (t) = 1 + 2 ln 2 − 2t. as long as the h1 and h2 constraints are not ˙ violated. the conditions (i)–(vi) below are sufﬁcient for (x ∗ (t). we conclude that (x ∗ (t). Knut Sydsæter.
˙ p(10− ) − p(10) = β. Then x ∗ (t) = 0 and thus x ∗ (t) = √ ∗ ( 2) = 0. we try u∗ (t) = t in ( 2. 0. 1] if t ∈ (1. x ∗ (t) = u∗ (t). In [0. Finally. Then from (ii). We see that x ∗ (t) is decreasing and is 0 at t ∗ = 2. Finally. so A = −1. (i) is equivalent to (Hu )∗ = p(t) − 2u∗ (t) = 0. 2). with q(t) = 0 if x ∗ (t) > 0. ˙ (1 − t. and thus p(t) = t + A. (iv) gives β = 1. q(t) ≥ 0. u∗ (t)) is optimal. so the optimal solution is: (x ∗ (t). with β = 0 if x ∗ (5) > 0. one has to have p(t) = 0. p(5) = 0. with β = 0 if x ∗ (2) > 0. we get from (iv) that β = 0. so (x ∗ (t). we have 2 + A = 1. For In order still to keep x u∗ (t) = 0 to be the maximizer in (i) for t ∈ (1. 1] if t ∈ (1. so u∗ (t) = 2 p(t). t ∗ ]. t − 1) (0. For u∗ (t) = t to be the maximizer in (i) one has to have p(t) = 1. In [0. 2]. in particular p(2− ) = 0. We have x © Arne Strøm. since p(5) = 0. β ≥ 0. Here H is concave in (x. From (iii) we get q(t) = 1. 2] . p(10) = 0. x) = x is quasiconcave. q(t) ≥ 0. Knut Sydsæter. p(2) = 0.3 The Lagrangian is L = H + qx = −u2 − x + pu + qx. p(t) = −∂L∗ /∂x = 1 − q(t). 5]. ˙ √ ˙ In order still to keep x ∗ (t) down. ˙ p(2− ) − p(2) = β. β ≥ 0. p(1) = 1 + A = 0. Then (iii) gives p(t) = 1. u) and h(t.7. ˙ We start by putting u∗ (t) = −1. 1) we have by (ii). 10. and Peter Hammond 2008 . ˙ Since we want to keep x ∗ (t) down. Then with x ∗ (1) = 0 we get x ∗ (t) = 0. Now all the conditions (i)–(vi) are satisﬁed. β ≥ 0. q(t) = 0. u∗ (t). q(t) = 0 1 if t ∈ [0. x √ √ √ Since p(t) is continuous at t = 2. x ∗ (0) = 1. t ∗ ] be the maximal interval where x ∗ (t) > 0. so x ∗ (t) = − 2 t 2 + 1. from (iii) we get q(t) = 1.2 The Lagrangian is L = H + qx = 1 − x + pu + qx. in particular p(5− ) = 1. Then (iii) gives p(t) = 1. q(t) = 0. 2) we have by (ii). Here H is concave in (x. p(t)) = with β = 0. −1. Now all the conditions (i)–(vi) are satisﬁed. so the conditions (i)–(vi) below are therefore sufﬁcient for (x ∗ (t). we put u∗ (t) = 0 in some interval [0. Then x ∗ (t) = −t with √˙ √ 1 x ∗ (0) = 1. 10. q(t) = 0. p(t) = −∂L∗ /∂x = 1 − q(t). x ∗ (0) = 1. q(t) ≥ 0. 0) if t ∈ [0. x ∗ (t) = u∗ (t) − t. u∗ (t)) to be optimal: (i) (ii) (iii) (iv) (v) (vi) u = u∗ (t) maximizes 1 − x ∗ (t) + p(t)u for u ∈ [−1. so A = 1 − 2. with β = 0 if x ∗ (10) > 0. x) = x is quasiconcave.7. ∗ (0) = 1. and the conditions (i)–(vi) below are therefore sufﬁcient for (x ∗ (t). ˙ ∗ (t) down put u∗ (t) = 0 in (1. with q(t) = 0 if x ∗ (t) > 0. x ∗ (t) = u∗ (t). Then since p(2) = 0. u∗ (t)) to be optimal: (i) (ii) (iii) (iv) (v) (vi) u = u∗ (t) maximizes −x ∗ (t) + p(t)u − u2 for u ∈ . with q(t) = 0 if x ∗ (t) > 0. x ∗ (0) = 1. ˙ p(5− ) − p(5) = β. 1]. for some constant A. for some constant A. Then x ∗ (t) = 1 − t is decreasing and is 0 at t ∗ = 1. u) and h(t. Since p(t) is continuous at t = 1. Let [0.90 10 CONTROL THEORY WITH MANY VARIABLES (i) (ii) (iii) (iv) (v) (vi) u = u∗ (t) maximizes −u − x ∗ (t) + p(t)(u − t) = −x ∗ (t) − tp(t) + u(p(t) − 1) for u ≥ 0. Atle Seierstad. p(t) = −∂L∗ /∂x = 1 − q(t). and thus p(t) = t + A. 2] 1 Since H is concave in u and u ∈ .
q(t) = 1 and (iv) gives β = p(10− ) = 0 since p(10) = 0. ˙ (iii) x ∗ (t) = u∗ (t). and then x ∗ (t) = t + 1 in some interval [0. ˙ From (ii) we get p(t) = p for some constant p. Note that according to (i). contradicting u∗ (t) ≥ 0. we must have t ∗ ≤ 2. Then u∗ (t) = 2 (t + A) and x ∗ (t) = 2 (t + A). So u∗ (t) = x ∗ (t) = 0. t ∗ ]. u∗ (t)) to be optimal. Here H is concave in (x. It follows that h(t. (b) The Lagrangian is L = H + q(t + 1 − x) = (4 − t)u + pu + q(t + 1 − x). u). t ∗ ] and u∗ (t) = 0 in (t ∗ . contradicting x ∗ (3) = 3. u) and h(t. 3]. Knut Sydsæter. Since ¯ ¯ x ∗ (3) = 2t ∗ + 1 = 3. and since x ∗ (3) = 3 and x ∗ (t) ≥ 0. 3]. with t ∗ − 4 = p. 2]. so A = −2. and from (iii) ¯ we have x ∗ (t) ≡ 1. and so A = −t ∗ < 0. 10]. 2 (t − 2). With ˙ ∗ (t) = 1 (t − 2).10 CONTROL THEORY WITH MANY VARIABLES 91 1 1 ˙ and (iii) gives p(t) = 1. p(t) = t − 2 we have u 2 4 Looking at the objective function. The function x ∗ (t) must have a minimum at x 4 4 1 1 t ∗ so x ∗ (t ∗ ) = 2 (t ∗ + A) = 0. Then (iii) yields p(t) = p for some constant p. ∗ = 2 and then u∗ (t) = 0 in (2. 0. x ∗ (3) = 3. with q(t) = 0 if t + 1 > x ∗ (t). and thus p(t) = t + A. Hence x ∗ (t ∗ ) = 1 − 4 A2 = 0. 3]. u∗ (t) = 1 can only maximize the Hamiltonian in [0. B = − 4 A2 + 1.4 (a) The Hamiltonian H = (4 − t)u + pu is concave in (x. it is obvious that we need to keep u∗ (t) = 0 on (2. and so x ∗ (t ∗ ) = 1 (t ∗ + A)2 − 1 A2 + 1 = 0. Since p(t) is continuous at t = 2. 2] if t ∈ (2. 0) . t ∗ ]. We see that h(3. are sufﬁcient for (x ∗ (t). t ∗ ] provided p(t) = t − 4. From (iii) we further get q(t) = 1 ˙ in [0. It remains to determine β. x ∗ (0) = 1. x ∗ (3)) = 3 + 1 − 3 = 1 > 0. and u∗ (t) = 2 (t − 10) < 0 for t < 10. with β = 0 if 4 − x ∗ (3) > 0. The objective function indicates that we should keep u∗ (t) as large as possible. especially at the beginning. p(t) = −∂L∗ /∂x = q(t). u∗ (t). 10] (0. If p < t − 4 for all t in [0. 2]. (ii) p(t) = −∂H ∗ /∂x = 0. so we suggest u∗ (t) = 1. if t ∈ [0. 10] 10. so ¯ p(2 from (v) we get β = 0. when x ∗ (t) has become 0. It is clear that we have found the optimal solution. and (iv) gives p(3− ) = p(3) = −2. we get x ∗ (t) ≤ x ∗ (3) = 3 and In fact. t − 2) (x ∗ (t). Hence we have to choose u∗ (t) = 2 in some interval [0. With x ∗ (t) = t + 1 we get x ∗ (2) = 3. Condition (i) implies that we must have u∗ (t) = 2 if ¯ ¯ ¯ ¯ p > t − 4 and u∗ (t) = 0 if p < t − 4. 2] if t ∈ (2. so the optimal solution is 1 1 ( 4 (t − 2)2 . Thus t ∗ < 10 and ∗ (t ∗ ) = 0. t ∗ ].7. and x ∗ (t) = 2t ∗ + 1 in (t ∗ . u∗ (t)) to be optimal: (i) u = u∗ (t) maximizes p(t) − (t − 4) u for u ∈ [0. Atle Seierstad. β ≥ 0. x ∗ (t)) > 0 for t in (2. and x ∗ (t) = 1 (t − 2)2 in [0. and then p(t) = p = t ∗ − 4 = −3. Then from (iii). then u∗ (t) ≡ 0. Since x ∗ (0) = 1. and then p(t) = 0. we suggest t ˙ ∗ (t)) = t +1−x ∗ (t) ≥ t −2. we see that t ∗ = 1. and Peter Hammond 2008 . q(t) = 0 1 if t ∈ [0. then p(t) = t − 10 since 1 p(10) = 0. p(t)) = with β = 0. since all the conditions (i)–(iii) are satisﬁed. © Arne Strøm. so the following conditions are sufﬁcient for (x ∗ (t). ¯ ¯ − ) = 2 − 4 = −2 = p. Now all the conditions (i)–(vi) are satisﬁed. But having u∗ (t) larger than 1 will cause x ∗ (t) to violate the constraint x ≤ t + 1. In the same way we see that p > t − 4 for all t in [0. But by (ii) we have then h(t. From x ∗ (t) ≥ 0. 3]. so ˙ 1 1 x ∗ (t) = 4 (t + A)2 + B. 3]. ˙ p(3− ) − p(3) = −β. so the conditions (i) and (iii) in (a) in addition to (ii) (iii) (iv) (v) q(t) ≥ 0. 3]. Now from (iii) we have x ∗ (t) = 2t + 1 in [0. If t ∗ = 10. x) = t + 1 − x is quasiconcave. x q(t) = 0 in (2. 3] ¯ is impossible.
2. and it can be shown by induction that bt = s=0 (1 + rs )K − s=1 k=s (1 + rk )as − at © Arne Strøm.5).3 (a) Let the remaining debt on 1 January in year n be Ln . Then bt+1 = (1 + rt )bt − at+1 .2 11. Since the payment on the 1 principal in year n is Ln−1 − Ln and the interest is rLn−1 . (g) xt = x0 + tb increases (linearly) towards ∞. The loan will never be completely paid since Ln > 0 for all n (but it does tend to 0 in the limit as n → ∞). cf. . at the repayment. 3] the expression (p(t) − (t − 4))u = (−2 − (t − 4))u = (2 − t)u is maximized by u = u∗ (t) = 0. (c) xt increases monotonically and faster and faster towards ∞. where b0 = K. . the powers a t alternate between negative and positive values. We get damped oscillations around the limit x ∗ . 3].1 11. 11.1. and Peter Hammond 2008 .1. the power a t decreases monotonically towards 0 as a limit as t → ∞.4 Let rt be the interest rate in period t. 2. n = 1. Thus we get explosive oscillations about x ∗ . (a) Since 0 < a < 1. (iii) .2 In parts (a)–(f) the solution is given by xt = a t (x0 − x ∗ ) + x ∗ .92 11 DIFFERENCE EQUATIONS The optimal solution is spelled out the answer in the book. formula (11. . (f) Oscillations around x ∗ with constant amplitude. we have Ln−1 − Ln = 2 rLn−1 . since 2 − t is negative in (2. and (v) are now satisﬁed. We get b1 = (1 + r0 )b0 − a1 = (1 + r0 )K − a1 b2 = (1 + r1 )b1 − a2 = (1 + r1 )(1 + r0 )K − (1 + r1 )a1 − a2 b3 = (1 + r2 )b2 − a3 = (1 + r2 )(1 + r1 )(1 + r0 )K − (1 + r2 )(1 + r1 )a1 − (1 + r2 )a2 − a3 ··· ··· ··· t−1 t−1 t−1 The pattern is clear.133934 3 3 1 (c) The payment in year n will be Ln−1 − Ln + rLn−1 = 2 rLn−1 = 2 r(1 − 2 r)n−1 L. (d) Since a < −1. Because x0 − x ∗ < 0 it follows that xt will increase monotonically towards the limit x ∗ . (b) a t alternates between negative and positive values and tends to 0. Note in particular that for t in (2. 1 The solution is Ln = (1 − 2 r)n L. (ii) . Atle Seierstad. xt = x ∗ for all t. It is a useful exercise to check carefully that all the conditions (i). In (g)–(i) the solution is xt = x0 + tb.2. . 11 Difference Equations 11. (e) The solution is constant. while a t  tends to ∞ as t → ∞. (iv). and bt the outstanding balance. 1 1 (b) (1 − 2 r)10 L = 2 L implies that r = 2 − 2 · 2−1/10 ≈ 0. 11. (i) xt = x0 for all t. (h) The solution is monotonically (linearly) decreasing towards −∞. Knut Sydsæter. Then L0 = L.
. then there exist constants c1 and c2 . Proof of ⇐: If the determinant is zero.3 11. not both 0. (Section 11. Then c1 u(1) + c2 u(2) = −at c1 u(1) + c2 u(2) − bt c1 u(1) + c2 u(2) = 0 T T T −1 T −1 T −2 T −2 so (∗) holds for t = T also. and xt+2 − 7xt+1 + 12xt = 9A3t + 16B4t − 21A3t − 28B4t + 12A3t + 12B4t = 0.1 (a) xt+1 = A + B 2t+1 = A + 2B 2t and xt+2 = A + B 2t+2 = A + 4B 2t . Atle Seierstad.3. . if at+1 < rt bt . then the outstanding debt will increase. The general solution is then xt = A + Bt + u∗ . xt+2 = 9A3t + 16B4t . easily proved by induction.5 We shall prove that u(1) t and u(2) t are linearly dependent ⇐⇒ u(1) 0 u(1) 1 u(2) 0 u(2) 1 =0 Proof of ⇒: If the solutions u(1) and u(2) are linearly dependent. Knut Sydsæter. T − 1.11 DIFFERENCE EQUATIONS 93 It is reasonable to suppose that if the interest rate changes. in the form xt = A1t + B 2t . . and Peter Hammond 2008 . not t t both equal to 0. If the repayment in period t + 1 is less than the interest accrued during period t. for t = 0 and t = 1. the loan will never be paid off. so there exist constants c1 and c2 . the particular solution u∗ in part (a) becomes t t t u∗ = − t k=1 k(k − 1) + t k=1 1 1 (k − 1) = − 1 (t − 1)t (t + 1) + 2 t 2 (t − 1) = 6 t (t − 1)(t − 2) 3 1 The necessary summation formulas t k(k − 1) = 1 (t − 1)t (t + 1) and t (k − 1) = 2 t (t − 1) are k=1 k=1 3 ∗ really is a solution of the difference equation.e. then the repayments also change to ensure that bt+1 < bt . such that (∗) c1 u(1) + c2 u(1) = 0 t t for t = 0 and for t = 1. and we get t t t t u∗ t = −u(1) t k=1 ck−1 u(2) k + u(2) t k=1 ck−1 u(1) k =− k=1 kck−1 + t k=1 ck−1 as a particular solution of xt+2 − 2xt+1 + xt = ct . so xt+2 − 3xt+1 + 2xt = A + 4B 2t − 3A − 6B 2t + 2A + 2B 2t = 0 for all t. It is also easy to check that ut © Arne Strøm.3. 11.) (b) With xt = A3t + B4t we get xt+1 = 3A3t + 4B4t . at+1 > rt bt . 11. Then Dt = u(1) u(2) − u(1) u(2) = (t + 1) − t = 1 for t t+1 t+1 t all t.3. Now suppose that (∗) holds for t = 0. t (b) With ct = t. and t t so the columns of the determinant above are linearly dependent. u(1) and u(2) are t t linearly dependent. the columns are linearly dependent. and the determinant must be 0.4 shows how to ﬁnd this solution. It follows by induction that (∗) holds for all t ≥ 0. where T is some integer greater than 1.e. in particular. . and if this situation continues. i. Hence. such that c1 u(1) + c2 u(2) = 0 for all t. i. This holds. 11. 1.6 (a) From Problem 2 we can ﬁnd the linearly independent solution u(1) = 1 and u(2) = t of the t t homogeneous equation xt+2 − 2xt+2 + xt = 0.
Atle Seierstad.4.4. and f (1) > 0. (1 + g)2 − (b + k)(1 + g) + k (b) The equation m2 − (b + k)m + k = 0 has two complex roots if and only if (b + k)2 − 4k < 0. (b) By using the method of undetermined coefﬁcients to determine the constants P . A constant function will be a solution of the corresponding homogeneous function. The general solution of this equation is ut = A + B t. we have b = −1 − a. This yields P = 1. In this case we get t u∗ − 2u∗ + u∗ = D(t + 2)2 − 2D(t + 1)2 + Dt 2 = 2D t+2 t+2 t and the desired value of D is D = c/2. Q = 10 . Let us try a function of the form u∗ = Dt. √ (c) Part (III) of Theorem 11. and a + 2 > 0.4. and so the general t solution of the given equation is xt = (A + Bt)(−1)t + 2t . and we look for a particular solution of the form u∗ = Dt 2 .1. Q. So the t 2 2 1 3 1 general solution to the given equation is xt = A + B2t + 4 5t + 10 cos π t + 10 sin π t. which in turn are equivalent to a < 1 + b and a < 2. If a = −2. The oscillations are damped if and only if r < 1. 1 If a < 2.1 shows that the growth factor of the oscillations is r = k.4.4 11.4 Since 1 + a + b = 0. then both roots of the equation f (m) = m2 + am + b = 0 lie in the interval (−1. We ﬁnd a particular solution of the nonhomogeneous equation by inserting u∗ = P 2t . We get t u∗ + au∗ − (a + 1)u∗ = D(t + 2) + aD(t + 1) − (a + 1)Dt = D(a + 2) t+2 t+1 t Thus.9 (a) It seems natural to try a function of the form Yt∗ = C(1 + g)t . if a < 1 + b and b < 1. and R in the 1 3 1 particular solution u∗ = P 5t + Q cos π t + R sin π t. 1) if and only if a < 1 + b and b < 1.4. we obtain P = 4 . (Draw a picture!) These four inequalities are equivalent to 1 − a + b > 0.6 1 If b = 4 a 2 and xt = ut (−a/2)t . 1 11.e. so that will not work (unless c = 0). D = c/(a+2) can be used unless a = −2. then the lefthand side of equation (11. i.4.94 11 DIFFERENCE EQUATIONS 11. Knut Sydsæter.4. which is the result claimed for case II in Theorem 11. We get ∗ Yt+2 −(b+k)Yt+1 +kYt∗ = C(1+g)t [(1+g)2 −(b+k)(1+g)+k] = C(1+g)t [(1+g)2 −b(1+g)−kg] This shows that Yt∗ = a(1 + g)t (if the denominator of this fraction is nonzero). 1 + a + b > 0. so the general solution of the associated homogeneous equation is xtH = (A + Bt)(−1)t . so we are looking for a particular solution of the equation xt+2 + axt+1 − (a + 1)xt = c. 2 2 11. 11. so xt = ut (−a/2)t = (A + B t)(−a/2)t . On the other hand. then a < 2. then b ≤ 4 a 2 < 1. and R = 10 . f (1) > 0.1) becomes 1 1 xt+2 + axt+1 + 4 a 2 xt = ut+2 (−a/2)t+2 + aut+1 (−a/2)t+1 + 4 a 2 ut (−a/2)t 1 = 4 a 2 (−a/2)t (ut+2 − 2ut+1 + ut ) which is 0 if ut+2 − 2ut+1 + ut = 0. −2 + a < 0.2 (a) The characteristic equation m2 + 2m + 1 = (m + 1)2 = 0 has the double root m = −1. if and only if k < 1. f (−1) < 0. 1) if and only if f (−1) > 0. and Peter Hammond 2008 .4. 11. the difference equation is xt+2 −2xt+1 +xt = c.11 Claim: If 4 a 2 ≥ b. Proof: Both roots belong to (−1. © Arne Strøm.
5 11.3 The points (a1 .11 DIFFERENCE EQUATIONS 95 11. The general solution of xt+2 −xt = 0 is xt = A+B(−1)t . The characteristic equation of (iv) is m2 + m − 2 with the roots m1 = 1 and m2 = −2. and then (i) and (iii) imply −xt + xt+1 + yt − 1 + t = −xt+2 + 1 + xt + yt − 2t. β.6 11. We get u∗ = 2 t − 2 t 2 . which has the roots m1 = 1. so the general solution of (iv) is xt = xtH + u∗ = t t 3 1 A + B(−2)t + 2 t − 2 t 2 . 11. so A = 2 and B = − 2 . These points are precisely the points lying in the interior of the triangle formed by those three lines. gives yt = 2 xt+1 = 4 + 4 (−1)t .5. Atle Seierstad.e.) 11. and (2. and the initial conditions imply A+B = 1. 1). This. we get (i) yt+1 = −xt + xt+1 + yt − 1 + t and (ii) − xt+2 − yt+1 + 1 = −xt − yt + 2t Equation (ii) implies (iii) yt+1 = −xt+2 + 1 + xt + yt − 2t.6. © Arne Strøm. i.5 are satisﬁed when a1 = σβ/α − 2 and a2 = 1 − σβ. x1 = 2y0 = 2. 1). The characteristic equation of xt+2 − xt = 0 is m2 − 1 = 0. −1). in turn. and σ are all positive. The ﬁrst equation yields zt = −xt+1 − yt + 1.5. (b) We ﬁrst eliminate z. For a particular solution u∗ of (iv) itself we try with a quadratic t 3 1 polynomial u∗ = Dt + Et 2 . (0. a2 ) that satisfy the three inequalities in (∗∗) are the points that lie above each of the lines given by a2 = −1 − a1 and a2 = −1 + a1 and below the line a2 = 1. Knut Sydsæter.1 (a) From the given equations we get xt+2 = 2yt+1 = xt . A necessary and sufﬁcient condition for the roots of this equation to be complex (more precisely: not real) is σβ −2 α 2 < 4(1 − σβ) ⇐⇒ σ 2β 2 4σβ − + 4 < 4 − 4σβ ⇐⇒ σ 2 β 2 < 4ασβ − 4α 2 σβ 2 α α ⇐⇒ σβ < 4α − 4α 2 = 4α(1 − α) The difference equation is globally asymptotically stable if and only if the inequalities (∗∗) in Section 11. and Peter Hammond 2008 .5 The characteristic equation is m2 + (σβ/α − 2)m + (1 − σβ) = 0. and the initial conditions give x0 = 1. m2 = −1. and so (iv) xt+2 + xt+1 − 2xt = 2 − 3t By a stroke of good luck y does not appear in (iv). and so the homogeneous equation corresponding to (iv) has the general solution xtH = A + B(−2)t . 3 1 1 3 1 3 1 A−B = 2. Using this in the second and third equations. Thus. the triangle with corners at (−2. xt = 2 − 2 (−1)t . This gives the conditions 1+ σβ σβ − 2 + 1 − σβ > 0 and 1 − − 2 + 1 − σβ > 0 α α σβ σβ ⇐⇒ − σβ > 0 and 4 > + σβ α α ⇐⇒ α < 1 and (1 + α)σβ < 4α and σβ > 0 (Remember that α.
. ∞). so we must have A + B = 3 1 1 0. x4 = −2. which implies. u u∗ (x) = 0 2 The fundamental equation (Theorem 12. .1. converging to the equilibrium value x ∗ ≈ −2. one in (−∞. That gives J1 (x) = g(x/3) = 2 − 5 x 2 . Then f (x) − x is convex. .1 12.63212. (b) See the answer in the book. 12 Discrete Time Optimization 12. 0) and one in (0.94650.7 11. Knut Sydsæter. Hence.94753. and it is easy to see from the intermediate value theorem that the equation f (x) = x has two solutions.92807. x6 = −2. We continue with 1 3 J0 (x) = max(1 − (x 2 + 2u2 ) + J1 (x − u)) u 11 8 10 = · · · = max(3 − x 2 + xu − u2 ) u 3 3 3 h(u) © Arne Strøm. . so 3 yt = − 1 + 1 (−2)t 3 3 Finally. 3 The initial condition y0 = 0 yields A = − 1 . and Peter Hammond 2008 .1 (a) To solve the problem by dynamical programming we ﬁrst ﬁnd J2 (x) = max(1 − (x 2 + 2u2 )) = 1 − x 2 . Then g (u) = 2x − 6u.94753.2 (a) Let f (x) = ex − 3. Since f (x) = ex < 1 for all x < 0. x5 = −2. 3 3 1 xt = − 1 + 1 (−2)t + 2 t − 2 t 2 3 3 From equation (i) we get yt+1 −yt = xt+1 −xt +t−1 = −(−2)t with the general solution yt = A+ 1 (−2)t . x2 = −2.94753.1) then gives J1 (x) = max(1 − (x 2 + 2u2 ) + J2 (x − u)) u u u = max(1 − x 2 − 2u2 + 1 − (x − u)2 ) = max(2 − 2x 2 + 2xu − 3u2 ) g(u) Let g(u) be the expression in the last parenthesis. zt = −xt+1 − yt + 1 = 2 3 1 1 + 1 (−2)t − 2 t + 2 t 2 3 11. x3 = −2.94748. A = − 1 . Atle Seierstad. so g attains its maximum for u = u∗ (x) = x/3. B = 3 .1.96 12 DISCRETE TIME OPTIMIZATION The initial conditions yield x0 = 0 and x1 = −y0 − z0 + 1 = 0. The solution of the difference equation starting at x0 = −1 gives the values (rounded to 5 decimal places) x1 = −2. A − 2B + 2 − 2 = 0. the negative solution is a stable equilibrium of the difference equation xt+1 = f (xt ).7.
1. Then JT −(k+1) (x) = maxu∈[0. so the formula follows by induction. and V = J0 (x0 ) = J0 (0) = T j j =0 (2 − 1)2 . and then JT −(k+1) (x) = x 2 (5+2k).4 (a) JT (x) = 3x 2 with u∗ (x) = 0. 10/11. This gives 2 2 2 S(u0 . JT −1 (x) = maxu∈ [x − u2 + JT (2(x + u))] = maxu∈ [x − u2 + 2(x + u)] = 3x + 1 for u∗ −1 (x) = 1. . In particular. . which is the proposed formula for n = k + 1. . u1 . 12.1] [(3 − u)x 2 + (2k + 3)(ux)2 ] = x 2 maxu∈[0. Atle Seierstad. and Peter Hammond 2008 .1. Js (x) = maxu∈ [x − u2 + Js+1 (2(x + u))] for T s = 0. Suppose it is valid for n = k. 1. 0 ∗ ∗ u∗ = u∗ (x1 ) = x1 /3 = 10/11. and T k then JT −(k+1) (x) = x − (2k+1 − 1)2 + 2(2k+1 − 1)x + 2(2k+1 − 1)2 + j =0 (2j − 1)2 = (1 + 2k+2 − k k+1 2)x + (2k+1 − 1)2 + j =0 (2j − 1)2 = (2(k+1)+1 − 1)x + j =0 (2j − 1)2 . which fortunately agrees with the result from part (a). 12. T (b) We claim that JT −n (x) = (2n + 3)x 2 with u∗ (x) = 0 and u∗ −n (x) = 1 for n = 1. 1 1 ∗ x1 = x0 − u∗ = 30/11 0 ∗ ∗ x2 = x1 − u∗ = 20/11 1 (b) We have x1 = x0 − u0 = 5 − u0 and x2 = x1 − u1 = 5 − u0 − u1 . Then JT −(k+1) (x) = maxu∈ [x − u2 + k (2k+1 − 1)(2x + 2u) + j =0 (2j − 1)2 ]. This is the given formula for n = k + 1. 0) Since S is concave. . . this is a global maximum point for S. u1 . u2 ) = 1 − (x0 + 2u2 ) + 1 − (x1 + 2u2 ) + 1 − (x2 + 2u2 ) 0 1 2 = 3 − 52 − 2u2 − (5 − u0 )2 − 2u2 − (5 − u0 − u1 )2 − 2u2 0 1 2 = · · · = −72 + 20u0 + 10u1 − 4u2 − 2u0 u1 − 3u2 − 2u2 0 1 2 It is clear from the second expression for S that S is a concave function. u2 ) = (25/11.12 DISCRETE TIME OPTIMIZATION 97 Here h (u) = 10x/3 − 22u/3. Suppose it is valid for n = k. T (b) The formula is valid for n = 1. T − 1.1] [3 − u + (2k + 3)u2 ]. . Since u∗ −(k+1) (x) = 2k+1 − 1. T . The ﬁrstorder partial derivatives of S are ∂S/∂u0 = 20 − 8u0 − 2u1 ∂S/∂u1 = 10 − 2u0 − 6u1 ∂S/∂u0 = −4u2 and it is easily seen that the only stationary point is (u0 . The maximum value is Smax = −492/11. which implies that u∗ (x) = 5x/11. The function g(u) = 3−u+(2k+3)u2 is convex in u and has its maximum at u = 1. we get u∗ (x) = 2T −t − 1 t T for t = 0. JT −2 (x) = 7x 2 with T T u∗ −2 (x) = 1. The T T formula is valid for n = 1. The result in part (b) shows immediately that √ √ JT −1 (x) = −2 αe−γ x and JT −2 (x) = −2 2 αe−γ x = −23/2 α 1/4 e−γ x . . . . and therefore J0 (x) = h(5x/11) = 0 3 − 21x 2 /11. further.7 (a) It is immediately clear that JT (x) = −αe−γ xT . .1.1. © Arne Strøm. . Knut Sydsæter. so the formula follows by induction. We see that the maximizer is u = u∗ −(k+1) (x) = 2k+1 − 1. u∗ = u0 (5) = 25/11. 12. JT −1 (x) = 5x 2 with u∗ −1 (x) = 1.6 (a) JT (x) = maxu∈ (x − u2 ) = x for u∗ (x) = 0. T . Thus the desired maximum value is J0 (x0 ) = J0 (5) = −492/11 and. .1] [(3 − u)x 2 + JT −k (ux)] = maxu∈[0.
T ) = sup T β s−t (−e−us − 2 e−xs ) we get J (t. The fundamental equation then gives Jt−1 (x) = maxu∈ ϕ(u). For a ﬁxed T . 1. with αt determined by the difference equation above and αT = α. We ﬁrst solve the corresponding ﬁnite horizon problem T sup ut t=0 1 β t (−e−ut − 2 e−xt ). . and has a unique maximum 3 point given by ϕ (u) = −2u − 2βα(x + u) = 0.98 12 DISCRETE TIME OPTIMIZATION (b) We know from part (a) that the formula Jt (x) = −αt = −αt e−γ x holds for t = T with αt = α. √ √ √ because α cannot be negative. and Peter Hammond 2008 . x0 given 1 With the optimal value function J (t. This is a quadratic √ √ √ √ √ √ √ equation for α with the solution α = β + β + 1/2. −x as T → ∞. (We cannot have α = β − β + 1/2. αt → α as t → −∞.) Hence. u∗ (x) = −βαx/(1 + βα) and x + u∗ (x) = x/(1 + βα).2 (a) The Bellman equation is J (x) = maxu∈ − 2 x 2 − u2 + βJ (x + u) . x. . Equation (∗) now gives 2 β 2 α2 βα 2 β 2 α 2 + βα 2 −αx 2 = ϕ(u∗ (x)) = − x 2 − x2 − x2 = − x2 − x 3 (1 + βα)2 (1 + βα)2 3 (1 + βα)2 βα 2 βα(βα + 1) 2 ⇐⇒ α = + = + 2 3 (1 + βα) 3 1 + βα ⇐⇒ 3α(1 + βα) = 2(1 + βα) + 3βα ⇐⇒ 3βα 2 + (3 − 5β)α − 2 = 0 © Arne Strøm. t = 0. Hence.7. xt+1 = 2xt − ut . . Knut Sydsæter. . To show optimality. . where ϕ(u) = −e−γ u + Jt (2x − u) = −e−γ u − αt e−γ (2x−u) The function ϕ is concave and ϕ (u) = 0 ⇐⇒ γ e−γ u − αt γ e−γ (2x−u) = 0 ⇐⇒ −γ u = ln αt − 2γ x + γ u √ ⇐⇒ γ u = γ x − ln αt This shows that ϕ has a unique stationary point u∗ = x − (ln It follows that Jt−1 (x) = ϕ(u∗ ) = −e−γ u − αt e−2γ x+γ u = −eln ∗ ∗ √ αt )/γ . T − 1. 12. Atle Seierstad. then −αx 2 = max − 2 x 2 − u2 + βJ (x + u) = max ϕ(u) (∗) 3 u∈ u∈ where ϕ(u) = − 2 x 2 − u2 − βα(x + u)2 .1 The equation α = 2 αβ + 2 can be written as ( α )2 − 2 β α − 2 = 0. we can use Case B in Note 12. J (0. It follows by induction that the formula Jt (x) = −αt e−γ x holds for t = T . x. . T ) → −αe 12. − αt e−γ x−ln √ αt √ αt −γ x = −αt−1 e−γ x √ √ √ where αt−1 = αt + (αt / αt ) = 2 αt .3 √ √ √ √ 1 1 12. Hence. 1. x. T − 1.3. . 0. If J (x) = −αx 2 is a 3 solution. . which is a maximum point for ϕ.3. T ) = s=t √ 1 1 −αt e−x . The function ϕ is strictly concave.3. α = ( β + β + 1/2 )2 . with αT = 2 and αt = 2 βαt + 2 for t < T . Suppose that it holds for a positive integer t ≤ T .
let u0 = −x0 and ut = 0 for all t > 0. p) = and u2 − 2x 2 + pu u2 − 2x 2 for t < T for t = T 2u + p for t < T . and from the arguments above it follows that α < 5 . 2 Then ∞ β t (− 2 xt2 − u2 ) = − 5 x0 is ﬁnite. u. u. and the optimal 6β β −3+ (5β − 3)2 + 24β x 6β (5β − 3)2 + 24β (5β − 3)2 + 24β x = ··· = − (b) It is clear that the value function J0 (x) is ﬁnite: for any x0 . Hu (t.4.) Hence.12 DISCRETE TIME OPTIMIZATION 99 The last equation has a exactly one positive solution. T (b) (The reference in the problem should be to Theorem 12. p) = −4x for all t 2u for t = T ∗ ∗ ∗ Since x0 = x1 = · · · = xT = 0 and u∗ = u∗ = · · · = u∗ −1 = 0. Then Note 12. u. T − 1. x0 ]. Further. Hence. such a u cannot be optimal. xt∗ . . x. 3 Now let x be ﬁxed.2 applies with X(x0 ) = [−x0 . and we already know that pT = 0. the value function exists. the difference equation (4) in 0 1 T Theorem 12. . π) be the sum that results from x = x0 and the control sequence π = (u0 . Knut Sydsæter.4. Also. x. not 12. π). . It follows that H (t. Condition (3) yields for t = 0. for t < T .5 12. then J (x + u) < J (x) and ϕ(u) = − 2 x 2 − u2 + βJ (x + u) < − 2 x 2 − u2 + βJ (x) < − 2 x 2 + βJ (x) = ϕ(0) 3 3 3 Hence. Atle Seierstad. 12. I is maximized when u∗ = u∗ = · · · = u∗ −1 = 0 and t=0 t 0 1 T T u∗ = ±1. pt ) = u2 − 2(xt∗ )2 = u2 . (Remember.1 implies pt = 0 for t = 0. x. It follows that an optimal u must be such that x + u ≤ x and u ≤ x. let V0 (x. pt ) for u in [−1. and it follows that J0 (λx) = λ2 J0 (x)—that is. λπ) = λ2 V0 (x. .3. ). Hx (t. so we really do have J (x) = −αx 2 for a suitable α ≥ 0. x. . for any sequence of controls u0 . . 1. 1: −4ut − pt1 = 0 and −2vt + pt2 = 0. u. Also. H (t. J = J0 is homogeneous of degree 2. . if u > x.4 2 12.4. p) = 1 + x − y − 2u2 − v 2 for t = 2. © Arne Strøm.1.5. α = control is u∗ (x) = − 5β − 3 + βα x=− 1 + βα 5β + 3 + 5β − 3 + (5β − 3)2 + 24β . xt∗ . 1. It is also clear that J0 (x) ≤ 0 for all x. u = u∗ = 0 is not a maximum point of t H (t. then J (x + u) < 0 and ϕ(u) = − 2 x 2 − u2 + βJ (x + u) < − 2 x 2 − u2 < − 5 x 2 = ϕ(−x) 3 3 3 so this u cannot be optimal either. Hence. u1 . .2 (a) I = T (u2 − 2xt2 ) = T −1 u2 + u2 − 2x0 − 2 T xt2 = u2 + T −1 u2 − 2 T −1 u2 = t=0 t t=0 t t=1 t=0 t t=0 t T T u2 − T −1 u2 . if not in the “max sense”. Then for any number λ we get V0 (λx.1 (c) H (t. and Peter Hammond 2008 . It is also clear that α = 0.) The Hamiltonian is H (t. . p) = 12. p) = 1 + x − y − 2u2 − v 2 + p 1 (x − u) + p 2 (y + v) for t = 0. then at least in the “sup sense”.4. u1 . but actually a minimum point. x. u. u. x0 = 0. the sum is t t=0 3 3 bounded above by 0. 1]. . and consider the problem of ﬁnding a u that maximizes ϕ(u) = − 2 x 2 − u2 + J (x + u) 3 If x + u > x.2. u. .
12. For t = 1. From these equations we ﬁnd the same solution as before.5) boils down to JT (x) = ax 1/2 and 1 1 Jt (x) = max u1/2 + 2 Jt+1 (0) + 2 Jt+1 (x − u) u for t <T With Jt (x) = 2at x 1/2 this gives aT = a/2 and 2at x 1/2 = max ϕ(u). u0 = − 4 p0 . y1 = y0 + v0 . Finally.6 12. p1 = −1 + p2 . 1 + 2E[xt+1 (xt . x2 . xt .6.6. Atle Seierstad. Vt ). 1/2 The optimality equation (12. xt+1 . Vt ) − xt )  vt−1 − 2(vt−1 + xt − xt−1 ) = 0 i. v2 ) = −2(v2 + x3 − x2 ) + 1 = 0 © Arne Strøm.9) becomes F3 (2. u2 = 0. Moreover. v0 = 2 p0 . x1 = x0 − u0 = 5 − u0 . Vt )] − 2xt = 2vt−1 + 2xt − 2xt−1 Equation (12. u where ϕ(u) = u1/2 + at+1 (x − u)1/2 for t < T . xt−1 .8 The ﬁrst printing of the book contains a couple of embarrassing misprints in connection with this problem and with the stochastic Euler equation. x3 . maxu ϕ(u) = ϕ x/(1 + at+1 ) = · · · = (1 + at+1 )1/2 x 1/2 . This 2 implies 1 + at+1 = 4at2 . equation (∗) becomes E 2(Vt + xt+1 (xt . vt ) = 1 − (vt + xt+1 − xt )2 . 1 (∗∗) . x2 .6. p0 = −1 + p1 . u1 = − 4 p1 . Hence. Vt )  vt−1 + F3 (t − 1. x2 = x1 − u1 . p1 = 1 + p2 . 1 1 2 2 from (5).6. x3 . and so at = 2 (1 + at+1 )1/2 . 2.6 There are two misprints in the objective function: the expression inside the square brackets should T −1 be t=0 ut 1/2 + aXT . and Peter Hammond 2008 t = 0. xt . Formula (12. y2 = y1 + v1 . The function ϕ is concave and ϕ (u) = 0 ⇐⇒ x 1 at+1 ⇐⇒ u = ut (x) = = 2 1/2 1/2 2(x − u) 2u 1 + at+1 1 2 2 2 Hence.10) on page 454 should be E F2 (t.6. p0 = 1 + p1 . There is also a misprint in the answer in the book. Knut Sydsæter. 12. v2 ) = 1 − (v2 + x3 − x2 )2 + 1 + v2 + x3 F (t. v2 = 0. From (4) and (6)(c ). p2 = 0.100 12 DISCRETE TIME OPTIMIZATION 1 1 1 1 1 2 For t = 2 it yields −4u2 = 0 and −2v2 = 0. vt−1 ) = 0 and the objective function in the current problem should be 2 (∗) max E t=0 1 − (Vt + Xt+1 − Xt )2 + (1 + V2 + X3 ) Now deﬁne F by F (2. 1 2 1 1 1 2 2 2 v1 = 2 p1 . and therefore ut (x) = x/4at2 . xt+1 (xt . xt . p2 = 0.e.
Knut Sydsæter. because J (x) = ax 2 +b = a[x 2 +βd/(1−β)] must be negative. This quadratic equation gives a = [1 − 2β − 1 + 4β 2 ]/2β.7. in fact a has the same value as in part (a). and restrict u to belong to [ε. aT = −1. (With J (x) = ax 2 + b.2. The function ϕ(x) = −1 + βx/(1 − βx) is increasing (calculate its derivative). a = −[1 − 2β − the Bellman equation is u 1 + 4β 2 ]/2β.7. b is also as in part (a). bt−1 = βbt + βat d. we get at−1 < at for all t. xt ] for some small positive ε. ax 2 +b = −β 2 a 2 x 2 /(1 − βa)2 −x 2 + βax 2 /(1 − βa)2 +βad+βb = x 2 (2βa−1)/(1−βa)+βad+βb for all x. Atle Seierstad. ln a + x b ]. v1 ) = x1 − v1 + Then for t = 1: 1 1 + 2E[x1 − V1 + 2 ] − 2x1 = 2v0 + 2x1 − 2x0 ⇐⇒ 1 + 2(x1 − 1 2 1 + 2 ) − 2x1 = 2v0 + 2x1 − 2x0 1 2 ⇐⇒ x1 = x0 − v0 + Since x0 = 0 is given.12 DISCRETE TIME OPTIMIZATION 101 which gives x3 as a function of x2 and v2 : x3 = x3 (x2 . Choose b > 0 so small that αδ b < 1. limt→−∞ bt when T is ﬁxed. x2 = 1 − v0 − v1 . For x ≥ x0 . x0 .e. x3 = 3 2 − v0 − v1 − v2 12. and Peter Hammond 2008 . bt decreases when t decreases. T ) = a0 x 2 + b0 → ax 2 + b = J (x) as T → ∞. We can show optimality at least in a restricted problem: Assume V ∈ [0. Finding these limits is the same as ﬁnding the limits limt→−∞ at . we ﬁnd that a = limt→−∞ at satisﬁes a = −1 + βa/(1 − βa) = (2βa − 1)/(1 − βa) (so a > −∞). g = ln u + ln X ∈ [ln ε + ln x0 . δ] for some (perhaps large) δ. Then f  = (x − u)V  ≤ δx. x. since aT −1 < aT and this continues backwards. v2 ) = x2 − v2 + Now use (∗∗) for t = 2: 1 2 1 1 + 2E[x2 − V2 + 2 ] − 2x2 = 2v1 + 2x2 − 2x1 ⇐⇒ 1 + 2(x2 − 1 2 1 + 2 ) − 2x2 = 2v1 + 2x2 − 2x1 1 2 ⇐⇒ x2 = x2 (x1 . and taking limits in the equation for bt−1 . Now apply Note 12. b = aβd/(1 − β). evidently. ax 2 + b = max −u2 − x 2 + βE[a(x + u + V )2 + b] = max −u2 − x 2 + βa(x + u)2 + βad + βb u Maximizing this concave function yields u = βax/(1−βa).1 (a) J (x) = ax 2 + b. Note that Xt ≥ x0 . To ﬁnd limT →∞ J (0. b = [αd + αa ln(αa) − (1 + αa) ln(1 + αa)](1 − α)−1 . and. T T Then. at−1 = −1 − β 2 at2 /(1 − βat )2 + βat /(1 − βat )2 = −1 + βat /(1 − βat ). Letting t → −∞ in the difference equation for at . a = (2βa−1)/(1−βa) and b = βad + βb. and then b = aβd/(1−β). J (0. © Arne Strøm. In a similar way. i. write in particular a0 = a0 . Hence.7. and d = E[ln V ]. We have to choose the negative solution for a.7 12. where a is chosen so large that ln x/x b ≤ 1 when x ≥ a (by l’Hôpital’s rule limx→∞ ln x/x b = 0). at and bt depend on T ). we ﬁnd that b = limt→−∞ bt satisﬁes b = βb + βad. x) = β t (at x 2 + bt ). and the value function is J (x) = a ln x + b.) (b) J (t. T ) we need to ﬁnd limT →∞ a0 T T and limT →∞ b0 (for any t. bT = 0. 2 ln x] ⊆ [ln ε + ln x0 . where a = 2/(1 − α).2 The optimal control is ut (x) = x/(1 + αa). b0 = b0 . 12. the ﬁnal answer is x1 = 1 2 − v0 . Thus.
(c) False. Consider. . Then there exists an open ball B = Br (x0 ) around x0 such that xk ∈ Br (x0 ) for inﬁnitely many k. Obviously any such ball also meets T . yn )) in m × n correspond to the point (x1 . Since X is closed and B is open. int(S ) = { x ∈ n : x < 1}. Then w ∈ B ∩ S. the intersection B ∩ cl(S) is nonempty. Since z ∈ B and B is open. . Now let A and B be compact subsets of m and n .) Now. Then W = U ∩ V is also an open ball centred at x and W ⊆ S. T ⊆ S ∪ T . We shall show that x ∈ S ∩ T . This contradiction shows that {xk } must converge to x0 after all. and so S ∪ T ⊆ S ∪ T . cl(S) is closed.2. © Arne Strøm. and so it should converge to x0 . so S ⊆ S and T ⊆ T . By Theorem 13. It is enough to show that U intersects S ∩T . We want to show that y ∈ cl(S). . so by Problem 9(a).1. respectively. S ∪ T ⊆ S ∪ T . . there is an open ball B around z such that B ⊆ B. for any set T . so ∂S is not a subset of S. so x is an interior point of T . 13. also contains the centre of the ball. xm ). Moreover. And since z ∈ cl(S). W is the smaller of the two balls U and V . it is clear that int(S) ⊆ int(S ). 13. But this convergent subsequence is also a subsequence of the original sequence. . Let U be an arbitrary open ball centred at x. (y1 . W ∩ T = U ∩ V ∩ T ⊆ U ∩ S ∩ T .2 13. xm .102 13 TOPOLOGY AND SEPARATION 13 Topology and Separation 13. Therefore.5) to show that A × B is compact. . the punctured open ball S in part (a). But its closure is S = {x ∈ n : x ≤ 1} whose interior. and Peter Hammond 2008 . The boundary. In order to show this. there exists an open ball V centred at x such that V ⊆ S. Since y is in the closure of cl(S).5 Suppose for a contradiction that the sequence {xk } does not converge to x0 . it is sufﬁcient to show that every open ball around y intersects S. A is also bounded. we have ∂T ⊆ T if and only if T is closed.1. . of that set consists of the origin and all the points on the sphere x = 1. the set S ∪ T is closed. This set (a “punctured” open ball) is obviously open. . they must be equal.6 There is an obvious way to identify m × n with m+n : we let the point (x. S ⊆ S ∪ T . Every set is contained in its closure. Hence. W ∩ T = ∅ since x ∈ T . Since x ∈ S and S is open. But then B ⊆ T as well. . and so B ∩ S = ∅. and they all belong to the set A = X \ B = X ∩ ( n \ B).1 13. then there is an open ball B around x such that B ⊆ S. converging to a point y in A. Thus. y) = ((x1 .2. On the other hand.2. (In fact. let B be an open ball around y. y1 . Let z be any point in this intersection. We shall use the Bolzano–Weierstrass theorem (Theorem 13. Knut Sydsæter. Atle Seierstad. . Therefore. . . as we can see from the following example: Let S = {x ∈ n : 0 < x < 1}. ∂S. so int(S) = int(S). for example. then every open ball around y has a nonempty intersection with S. .1. These xk form a subsequence / {xkj } of the original sequence. . A is closed. (Draw a picture for the case n = 2!) (b) True. there is at least one point w in B ∩ S. Since each of the sets S ∪ T and S ∪ T is a subset of the other. 13. Therefore {xkj } has a convergent subsequence. . Hence A is compact. But that is impossible because y = x0 . In fact. . and so y belongs to the closure of T .9 (a) If x ∈ int(S). so it follows that U ∩ (S ∩ T ) is indeed nonempty. (b) Let y be a point in the closure of cl(S).15 (a) False. Because A is contained in the bounded set X. S ⊆ S ∪ T . yn ) in m+n . But we do not always have equality. If y ∈ cl(S). so int(S) = S. Let x ∈ S ∩ T . S ∪ T ⊆ S ∪ T (cf. Since S ⊆ S. Problem 10(b)). (d) True.2(c). Similarly.
Let F = {y ∈ m : y − f(x0 ) ≥ ε} be the complement in m of the open εball around f(x0 ). Let x0 ∈ S ∩ A. and so f is indeed continuous. it follows that {(ai . x0 ∈ S ∩ F = A. with limj →∞ xj = x0 ∈ S. and because F is closed. Claim: (a) ⇐⇒ (b). Let the challenger choose ε > 0. Now that we have proved the equivalence of (a) and (b) above. Since x0 ∈ A there exists a sequence {xk } of ¯ points in A converging to x0 . Knut Sydsæter. Then there is an ε > 0 such that f(xk ) − f(x0 ) ≥ ε for inﬁnitely many k. and by the deﬁnition of F we must have f(x0 ) − f(x0 ) ≥ ε > 0. and consider the following two statements: (a) A is relatively closed in S. We want to show that A = S ∩ F for some closed set F in n . the inverse image of F . Hence. Therefore x0 ∈ f −1 (F ). bi )}i is a subsequence of {(ak . which is absurd. By the equivalence (a) ⇐⇒ (b) above it sufﬁces to show that. 13. Let ai = akji and bi = bkji .1). Since x0 ∈ S. and because f is continuous.3 tells us that x0 ∈ F . and since B is compact. We want to show that f is continuous at x0 . Thus the defender wins. This shows that the assumption (∗) must be false. we want to show that f(xk ) → f(x0 ) for every sequence {xk } in S that converges to x0 . Then {(ai .2.3 13. Since f is continuous at x0 the defender is able to choose a δ > 0 such that f(x) − x0 ) < ε whenever x − x0 < δ. and then the challenger will be unable to ﬁnd an x that lets him win. The set F is closed in m . Since a sequence in p converges if and only if it converges componentwise (Theorem 13.3.5: Let S ⊆ n and let f be a function from S to m . we shall show that A = S ∩ A.2. let us use it to prove part (b) of Theorem 13. f(x0 ) the limit of a sequence in F . where F is closed in n . Proof of ⇐ : Suppose that f −1 (F ) is relatively closed in S for every closed F in m . Proof of ⇒: Since A is relatively closed in S we have A = S ∩ F . Hence. In other words. 13. Theorem 3. they also belong to F . where A is the closure of A in n . bi )}i is convergent. and let x0 be a point in S. the sequence {ak }k has a convergent subsequence {akj }j . (b) Whever a sequence {xk } in A converges to a limit x0 ∈ S.13 TOPOLOGY AND SEPARATION 103 Let {(ak .7 Let A ⊆ S ⊆ n. We want to prove that f −1 (F ) is relatively closed in S. and by the equivalence (a) ⇐⇒ (b) above we must have x0 ∈ f −1 (F ). and let {xj } = {xkj } be the subsequence of {xk } where k1 < k2 · · · run through all those k for which f(xk ) − f(x0 ) ≥ ε. Since A is compact. © Arne Strøm. But then f(x0 ) ∈ F . It is clear that A ⊆ S ∩ A. f −1 (F ) is relatively closed in S. and Peter Hammond 2008 . we have x0 ∈ A by assumption. By assumption. {bkj }j has a convergent subsequence {bkji }i . and since F is closed. and the sequences {ai }i and {bi }i are both convergent. then all f(xk ) ∈ F . Since all xk belong to A. so ¯ ¯ ¯ what we need to show is that S ∩ A ⊆ A. Thus. ¯ ¯ ¯ In fact. f(x0 ) ∈ F .8 Assume ﬁrst that f is continuous at x0 . We shall prove that the defender can always win in this case. (x0 ) = limk (xk ). and {xj } is a sequence in f −1 (F ). bk }k . Atle Seierstad. Let {xk } be a sequence in A that converges to x0 ∈ S.3. bk )}k be a sequence in A × B. If we have such a point x0 and such a sequence {xk }. Then f :S⊆ n → m is continuous ⇐⇒ f −1 (F ) is relatively closed in S for each closed set F in m Proof of ⇒: Suppose that f is continuous and let F be a closed set in m . A × B is compact. Proof of ⇐ : Suppose that (b) is satisﬁed. Suppose (∗) xk → x0 but xk → f(x0 ). if a point x0 in S is the limit of a sequence {xk }k in f −1 (F ) then x0 ∈ f −1 (F ). It follows that S ∩ A ⊆ A.3. then x0 ∈ A.
−1) = (x − 1)(x + 1)3 /x 3 . y) = 0 or y = ±3. strictly decreasing in [−1. −1) = x + 2. 13.7 Let S be a compact set in n . Thus. so for x > 1 the maximum occurs for y = −1.5. y) = −12xy 3 −12(x−1)y 2 +12y = −12xy(y−1/x)(y+1). and f (x.104 13 TOPOLOGY AND SEPARATION Now assume that f is discontinuous at x0 .3(d) shows the set M(x) of maximizers as a function of x. 3) = 162 − 351x. we just add n + 1 − m terms that are all equal to 0x1 . Knut Sydsæter. M13.5 13. because the onesided limits of V (x) at x = 1/3 are equal. y) (b) z = f (1. −1]. y) M13. f (x. Except at x = 1.1) tells us that every point in co(S) can be written as a convex combination of at most n + 1 points in S. and Peter Hammond 2008 . Carathéodory’s theorem (Theorem 13. Fig. y) for three different values of x. We claim that every point in co(S) can be written as a linear combination of exactly n + 1 points in S. the challenger will be able to ﬁnd an x with x − x0 < δ and f(x) − f(x0 ) ≥ ε.3 For a ﬁxed x. ∞). It follows that f (x. 1/x) − f (x. The value function V is given by ⎧ ⎨ 162 − 351x V (x) = (2x + 1)/x 3 ⎩ x+2 if 0 < x ≤ 1/3 if 1/3 < x ≤ 1 if x > 1 It is clear that V is continuous.3(a)–(c) shows the graph of the function y → f (x. Indeed. 1/x]. Atle Seierstad. and if 1/3 < x ≤ 1 the maximum occurs for y = 1/x.4. Finally. strictly increasing again in [0. Then there will exist at least one ε > 0 that cannot be matched by any δ.4.3 (c) z = f (2. if x = 1 the maximum occurs for y = ±1. if 0 < x ≤ 1/3. the maximum of f (x. the challenger wins. as it should be because the maximizer is unique for each x = 1. f (x. z 15 10 5 y −2 −1 1 2 −2 −1 1 2 15 10 5 y −2 −1 1 2 z 15 10 5 y −1 (d) x → M(x) z 3 2 1 y x 1 2 (a) z = f (1/2. 3] must be attained at a point where f2 (x. 0]. Then no matter what δ > 0 the defender chooses. the graph is the graph of a continuous function.4. 1/x) = (2x + 1)/x 3 . © Arne Strøm.4. y) is strictly increasing with respect to y when y ∈ (−∞. in this case.4 13. So let the challenger choose such an ε. the maximum occurs for y = 3. if x = λ1 x1 + · · · λm xm with m < n + 1. and so are the onesided limits at x = 1. Simple calculations give f (x.5. Figures M13. and strictly decreasing in [1/x. Since f2 (x. Hence the only possible maximum points are y = −1 and y = 1/x if x > 1/3 and y = −1 and y = 3 if 0 < x ≤ 1/3. y) with respect to y for y in [−3. y) 13.
Hence. Let N = N1 ∩ N2 . then so is H .6 that. it is closed. . such that U1 ×U2 ⊆ U .c. Consider. . λ1 + · · · + λn+1 = 1} and S n+1 = S × · · · × S is the Cartesian product of n + 1 copies of S.1 14.e.1 then says that F has the closed graph property at x0 .c.3 If x is not an interior point of the convex set S (⊆ n ). Let z0 ∈ H (x0 ) and let U be a neighbourhood of z0 . Similarly. y > 0} ∪ {(0. . . Then z0 ∈ G(y0 ) for some y0 in F (x0 ). then so is co(S). Since the half ¯ space H− is closed.3.c. 14. x1 . Deﬁne a function f : T → n by f (λ1 . and Note 14. The convex hull of S is co(S) = {(x. Let x ∈ N and y ∈ V ∩ F (x). y ∈ F (x). λn+1 .c. If S is closed but unbounded. y > 0. . then so is T .1. .5(c): We will use the result in (14. . For the result about upper hemicontinuity we need to assume that F (x0 ) and G(x0 ) are compact.1. On the other hand. the closed set S = {(x. What we have shown here is that.4 We shall use the characterization of lower hemicontinuity in (14. . . there exists a neighbourhood V of y0 such that U ∩ G(y) = ∅ for all y in V .h.8) to prove that H = G F is lower hemicontinuous. for instance points of the form x + ta for small positive numbers t. It follows that H is l. (It follows from Problem 13. λn+1 ) ∈ n+1 : λi ≥ 0 for i = 1.) Since f is continuous. Moreover. i. . Then H (x) ∩ U ⊇ H (x) ∩ (U1 × U2 ) = (F (x) ∩ U1 ) × (G(x) ∩ U2 ) = ∅ for all x in N ∩ X. the argument above shows that every point in co(S) belongs to the image f (T ) of T under f . Then S ⊆ H− = {z : a · z ≤ a · x}. then A × B is a compact subset of m+p .8) on page 506 to show that if F and G are l.h. Since F (x0 ) is compact.h. since y0 ∈ F (x0 ). no open ball around x ¯ ¯ is contained in S. respectively.h. Every open ball around x contains points that do not belong to H− . the open ﬁrst quadrant together with the origin. . This set is not closed. for example. then co(S) need not even be closed. so z ∈ H (x). Then H (x0 ) = F (x0 ) × G(x0 ) is also compact.h. n + 1. and so S is not an interior point of S. (Draw a picture!) 13.6. so if S is also compact.c. Then U ∩ G(y) = ∅. z0 ) in H (x0 ).1. Theorem 13. xy ≥ 1} ∪ {(0.e.. 0)}. Because F (x0 ) is bounded there exists a bounded and open set U in l that contains F (x0 ). Atle Seierstad. at x0 . at x0 . we also have S ⊆ H− . Knut Sydsæter.3 shows that co(S) = f (T ) is compact. y) ∈ 2 : x > 0.6. 0)}. if A and B are compact subsets of m and p . G has the closed graph property at x0 and there exists a bounded and open set V in m and a neighbourhood NG of x0 such that G(x) ⊆ V for all x in NG . where n = {(λ1 . .1. y) : x > 0. xn+1 ) = λ1 x1 + · · · + λn+1 xn+1 . then z ∈ G(y). Since G is l. 14 Correspondences and Fixed Points Proof of Theorem 14. Let U be an open neighbourhood of a point (y0 . Suppose that F and G are u. respectively. then by Theorem 13.2. so in fact co(S) = f (T ).6 13. . . at y0 .2 there exists a nonzero vector a in n such that a · z ≤ a · x for every z in S. Then f maps every point in T to a point in co(S). and since F is upper hemicontinuous at x0 there is a neighbourhood NF of x0 such that F (x) ⊆ U for all x in NF .14 CORRESPONDENCES AND FIXED POINTS 105 Now let T = n × S n+1 . and if z ∈ U ∩ G(y). if S is closed and bounded. Since F is l. closed and bounded) subset of n+1 . This immediately extends to the Cartesian product of a ﬁnite number of sets. there are neighbourhoods N1 and N2 of x0 such that F (x) ∩ U1 = ∅ for all x in N1 ∩ X and G(x) ∩ U2 = ∅ for all x in N2 ∩ X.1. Then there are open neighbourhoods U1 and U2 of y0 and z0 in l and m . . and Peter Hammond 2008 . there is a neighbourhood N of x0 such that V ∩ F (x) = ∅ for all x ∈ N. The set n is obviously a compact (i. © Arne Strøm. .
2 to prove that H is u. Then for any x in N and every i there is 0 at least one point yi in F (x) ∩ B(yi . b(x)). if x ≥ 0.2 implies that F is u. b(x)]. 1) if x < 0. Then y0 is a convex combination y0 = k λi yi of points yi in F (x0 ). with S = U = m \ U and K = F (x0 ). if F (x0 ) = ∅.4). C2 (x) = [0. The other three are usually not u. Knut Sydsæter.h. and Peter Hammond 2008 . the correspondence F (x) = (−1 − x 2 .106 14 CORRESPONDENCES AND FIXED POINTS Since F and G have the closed graph property at x0 .c. Since F is u. b(x)].h. K(x) = (a(x). while C2 is not u. By Problem 13. Moreover.1. . G(x) = co(F (x)) is also compact for each x.h. If we let y = i λi yi .h. ε) = ∅ for all x in Ni ∩ X. In fact. everywhere. H (x) = (a(x).c.h. 1] [0. b(x)) sometimes. .1.e. 14.h. where U is the open “αneighbourhood” of G(x0 ) deﬁned as U = B(G(x0 ). 1) contains F (x) for x = 0 but not for any other value of x.c.1. .7. C1 is u.1. Let y0 ∈ G(x0 ). G(x) = [a(x). at x = 0. 14. . ε) ⊆ U . But even if a(x) and b(x) are constant functions. If i=1 U is a neighbourhood of y0 . 0 0 d(y. k.10 Every constant correspondence is both l. Let N = N1 ∩ · · · ∩ Nk be the intersection of these neighbourhoods. so let us concentrate on the four possibilities F (x) = [a(x). This is an easy consequence of the deﬁnition (14. y0 ) = y − y0 = i (λi yi − λi yi ) ≤ i λi yi − yi < i λi ε = ε. then y ∈ co(F (x)) = G(x). (But both of them are l. since the open set U = (−1.c. and u.c at x0 . at x = 0.c. then U contains an open ball B = B(y0 . © Arne Strøm.h.h. complications may arise if one or both of the endpoints of the interval (a((x). there exists a δ > 0 such that F (x) ⊆ U for every x in N = B(x0 .c. at any other point either.) Every correspondence F that satisﬁes the conditions in the problem is l. If G(x0 ) is contained in an open set U . It is not hard to see that U is convex. The desired conclusion follows.h. A detailed study would take us too far aﬁeld. F has a closed graph and is locally bounded.. and H (x) = F (x) × G(x) ⊆ W for all x in N. provided only that its effective domain is open—i. so has H . With nonconstant a(x) and b(x) many complications arise in connection with upper hemicontinuity.c. It follows that G(x) ∩ U = ∅. [0.c.h. F is not u..6 We shall use the characterization of lower hemicontinuity that is given in (14. and F (x0 ) ⊆ G(x0 ) ⊆ U .1..c.11 The set F (x) is compact for each x in X. there is a neighbourhood Ni of x such that F (x) ∩ B(yi . so H is locally bounded near x0 . Hence G is l.1. no matter how close to 0 it may be. Consider for example the correspondences C1 and C2 given by C1 (x) = [0. Then W is bounded in l+m .5.h. For each 0 i = 1. belong to F (x).1.h. and therefore G(x) = co(F (x)) is also contained in U (and so in U ) for all x in N. b(x)) (for all x).h.c. α) = y ∈ m : there is a y in G(x0 ) with y − y <α This follows from the technical result below. everywhere. so Theorem 14. For example. N is a neighbourhood of x0 . then F (x) = ∅ for all x sufﬁciently close to x0 .c. Let W = U × V ⊆ l+m and let N = NF ∩ NG . 1] if x ≤ 0. We are now all set to use Theorem 14. 1 + x 2 ) is not u. 14.h.8) to prove that G 0 0 is l.c. ε). except in the constant case.c. there exists an α > 0 such that G(x0 ) ⊆ U ⊆ U . ε) for some ε > 0. but not always. so y ∈ B(y0 . Atle Seierstad. 1) if x > 0. δ).
S) − d(y .3 By Example 14. every ε > 0 can be matched by δ = ε.) Since K is compact. is a case where f (S1 ∩ S2 ) is a proper subset of f (S1 ) ∩ f (S2 ). The maximum theorem (Theorem 14. m) is continuous. Suppose that U is quasiconcave. The intersection of S1 and S2 is S1 ∩ S2 = [−1. m) is upper hemicontinuous and the indirect utility function V (p.2. y). y ) = d(y . m) in this problem. and it follows that B is upper hemicontinuous at (p0 . (Note that x F (x) in Theorem 14.e. m) is continuous at every point (p. S) is a continuous function of y. x0 ) = d(y . m) for every λ in [0.A SETS. © Arne Strøm. the distance between a point in S and a point in K is never less than α. d(y. 0) consists of a single point.5. y). m0 ) where m0 > 0. m) (i. Here.3. Then. The images of S1 and S2 under f are f (S1 ) = [0.8).1) then implies that the demand correspondence ξ(p. S) ≤ d(y . It follows that g(y) = d(y. consider the following: Let f (x) = x 2 for all x in . Lower hemicontinuity at that point follows easily from the result in (14. A Sets. the budget correspondence B(p. so α > 0.1 implies that λx1 + (1 − λ)x2 also belongs to ξ(p. 16] and f (S2 ) = [0. What if m0 = 0? In that case B(p0 . y) ≥ α for all x in S and all y in K.1.6. g attains a minimum value α over K. Then α = g(y∗ ) = d(y∗ . 1]. then.1. (In the deﬁnition of continuity in (13. COMPLETENESS. and the image of S1 ∩ S2 is f (S1 ∩ S2 ) = [0. then there exists a positive number α such that d(x. ξ(p. In other words. 14. m) for any (p. 2] and S2 = [−1. x∗ ) for a point y∗ in K and a point x∗ in S. y) attains a minimum at some point x0 in S. then d(y . By Problem 13. h(x) = d(x. so by Theorem 14. If U is an open set in n that contains 0. S) − d(y. n Thus. then it will obviously contain B(p. m) B(p. By symmetry. B(p.3 The answer in the book has been pared down to the bone. Atle Seierstad. m) close to (p0 . namely the origin 0 in n . 3]. and deﬁne two intervals S1 = [−4. S) ≤ d(y .1).2. d(y . m)). 9].1. if x1 and x2 are distinct points in ξ(p. m0 ). m) in X = ++ × + . 0). B is also upper hemicontinuous at (p0 . 0). Let us call h(x0 ) the distance between the point y and the set S.1 A. m) is a convex set.2. 2]. and Convergence A. Theorem 2. maximum points for U over B(p. m0 ) = B(p0 . For an example with a little more ﬂesh on it.1 corresponds to (p. y) + d(y. m0 ).3. If y is another point in n . S) so d(y .2 14. Hence. 4]. y).1. S). In other words. and denote it by d(y. with a function deﬁned on a set with only two elements in it. It is also locally bounded near (p0 . while y and f (x. S) ≤ d(y . m) is lower hemicontinuous and has the closed graph property at any point (p0 . x0 ) ≤ d(y . S) − d(y. Knut Sydsæter.5. y) in the theorem correspond to x and U (x) in the problem. Proof: Let y be a ﬁxed point in n and let x be a point in S.2. AND CONVERGENCE 107 A little technical result: If S and K are disjoint closed subsets of m and K is compact. y) + d(y. and Peter Hammond 2008 . Completeness.) The demand correspondence will not always be lower hemicontinuous. S) ≤ d(y.
Knut Sydsæter. If the initial approximation x0 is good enough. Hence. x2 .3. the sequence x1 . √ 1 1 that any convergent subsequence of {yk } must converge to 0. as promised.) Now apply this procedure to the equation f (x) = x 2 − 2 = 0. Figures MA. x 2 √ 3 k 10 √ 1 −2 3 MA. √ Thus the rational number s is closer to 2 than r is.2 A.1. (ii) transitive.3. But how did anyone come up with the expression s = (2 + r 2 )/2r in the ﬁrst place? The answer is Newton’s method. then x1 = r − f (r)/f (r) = r − (r 2 − 2)/2r = (r 2 + 2)/2r is precisely the number s.4 A relation is a linear ordering if and only if it is (i) reﬂexive. and {y6k+5 } converge to sin 0 = 0.2 It is shown in the answer in the book that s < r.4 For the solution of this problem we need the fact that.4(b) y 1 1 2 k 10 −1 MA.108 A SETS. respectively.3. if two convergent sequences have inﬁnitely many terms in common. A. it follows k }. AND CONVERGENCE A. let Mn = sup{xk : k ≥ n} and Nn = sup{yk : k ≥ n}. and (iv) complete.5 (b) (i): For each natural number n. {y6k+1 }. and then we construct x2 from x1 in the same fashion. and so xk +yk ≤ Mn +Nn .3. we also have s > 2. .4(a) and MA3. Let us just see how to handle (ii): We are assuming that R is transitive relation in a set S. COMPLETENESS.2. The six subsequences {y6k }. they must have the same limit. respectively. and so on. or − 2 3. xR zR x. {y6k+2 }.4(b) illustrate the behaviour of {xk } and {yk }. and they contain all but the ﬁrst ﬁve elements of {y√ In the same way as for {xk }. Therefore xR −1 z. these two subsequences contain all the terms of the original sequence. sin(4π/3) = − 2 3. Atle Seierstad. Given one approximate solution. (iii) antisymmetric. and sin(5π/3) = − 2 3. For each of these four properties it is easy to see that if a relation R has that property. Then for all k ≥ n. y.) Consider the particular subsequences {wk } = {x2k−1 } and {zk } = {x2k } of {xk }.3. We want to show that xR −1 z. x0 . sin(2π/3) = 2 3. Thus sup{xk +yk : k ≥ n} ≤ Mn +Nn ≤ © Arne Strøm. Together. . will usually converge to a root of f (x) = 0. 2 3. A. We have yR x and zR y. If x0 = r is a positive number. . (This is an easy consequence of Theorem A. we have xk ≤ Mn and yk ≤ Nn . and since R is transitive. then so has R −1 . (A brief discussion of this method can be found in EMEA.3. Let x. so any subsequence of {xk } must have an inﬁnite number of terms in common with at least one of {wk } and {zk }. we let x1 = x0 − f (x0 )/f (x0 ).3 A.4(a) 20 20 A. √ √ √ √ 1 1 1 1 sin(π/3) = 2 3.3. any convergent subsequence of {xk } must converge to either 0 or 2. {y6k+3 }. for instance. where f is a differentiable function. sin(3π/3) = 0. {y6k+4 }. They converge to 0 and 2. and since s is positive and s 2 > 2. and z be elements in S such that −1 y and yR −1 z. and Peter Hammond 2008 . which is a famous procedure for ﬁnding and improving approximate solutions to an equation f (x) = 0.
A. The formula for sin(x − y) in Problem 4 also shows that sin(π − x) = sin x for all x. (a) sin(π − π/6) = sin(π/6) = 1/2.6 Let n and m be natural numbers with n > m.6 It is clear from the deﬁnitions of sin x and cos x that sin(x + π) = − sin x and cos(x + π) = − cos x for all x. and let p = n − m. Knut Sydsæter. √ 1 (b) (cos(π + π/6) = − cos(π/6) = − 2 3. √ (e) By formula (6). from the hints in the question. These results come in handy in this problem. B Trigonometric Functions B. Then. √ 1 (d) cos(5π/4) = cos(π/4 + π) = − cos(π4) = − 2 2. we have sin[π − (α + β)] sin(α + β) = = tan(α + β) cos[2π − (α + β)] cos(α + β) © Arne Strøm.1. Then xn − xm  = xm+p − xm  = (xm+p − xm+p−1 ) + (xm+p−1 − xm+p−2 ) + · · · + (xm+2 − xm+1 ) + (xm+1 − xm ) ≤ xm+p − xm+p−1  + xm+p−1 − xm+p−2  + · · · + xm+2 − xm+1  + xm+1 − xm  1 1 1 1 < m+p−1 + m+p−2 + · · · + m+1 + m 2 2 2 2 1 1 1 1 1 − ( 2 )p 1 1 1 1 = m 1 + + · · · + p−1 = m < m = m−1 1 1 2 2 2 2 1− 2 2 1− 2 2 which obviously becomes arbitrarily small for m sufﬁciently large. tan(7π/6) = tan(π/6) = 1 3. 3 .1.4 cos(y − π/2) = sin y follows directly from Problem 3.8) together with the fact that sin π = 0 and cos π = −1. The proof of (ii) is similar and is left to the reader.1. √ √ B. as well as (B. sin(−3π/4) = − sin(3π/4) = − 2 2. or limk→∞ (xk + yk ) ≤ limk→∞ xk + limk→∞ yk .B TRIGONOMETRIC FUNCTIONS 109 limn→∞ (Mn + Nn ) = limn→∞ Mn + limn→∞ Nn .1 B. This also follows from Problem 4 and formula (B.3.1.1. and Peter Hammond 2008 √ 1 (c) By Problem 2(a). Atle Seierstad.7 (a) The formula for sin(x + y) in Problem 4 gives √2 sin(x + π/4) − cos x = 2(sin x cos(π/4) + √ √ 1 1 cos x sin(π/4)) − cos x = 2(sin x · 2 2 + cos x · 2 2) − cos x = sin x (b) Since sin(π − x) = sin x and cos(2π − x) = cos(−x) = cos x. it follows that sin(x + y) = cos(x + y − π/2) = cos x cos(y − π/2) − sin x sin(y − π/2) = cos x sin y + sin x cos y = sin x cos y + cos x sin y Substituting −y for y then yields sin(x − y) = sin x cos(−y) + cos x sin(−y) = sin x cos y − cos x sin y. √ √ 1 (f) sin(π/12) = sin(π/3 − π/4) = sin(π/3) cos(π/4) − cos(π/3) sin(π/4) = 4 ( 6 − 2).8) and the result of Problem 3 again. B.
(arcsin 2x) = (arcsin u) =√ dx du dx 1 − u2 1 − 4x 2 (b) (d/dx)(arctan v) with v = 1 + x 2 .2.e. The equation LHS = RHS implies that cos(x + y) = cos x cos y − sin x sin y. It follows that f (x) < 0 for all x in J .6 The derivative of f (x) is f (x) = 3(sin x − x − 1)2 (cos x − 1).2. a = 1/4. cos x < 1 for all x in the open interval J = (0. 3π/2]. It is easy to see that sin x < x + 1 for all x > 0. we have (cos x − cos y)2 + (sin x + sin y)2 = (cos(x + y) − 1)2 + sin2 (x + y). The lefthand side is LHS = cos2 x − 2 cos x cos y + cos2 y + sin2 x + 2 sin x sin y + sin2 y = 2 − 2 cos x cos y + 2 sin x sin y and the righthand side is RHS = cos2 (x + y) − 2 cos(x + y) + 1 + sin2 (x + y) = 2 − 2 cos(x + y) where we have repeatedly used that sin2 u + cos2 u = 1.1.2. (An exponentially damped cosine curve with amplitude 2e−x/π .2.) (c) y = 2e−x/π cos x. Thus. 3π/2).) B.8 You can read all these values off from Table B. Moreover. because sin x ≤ 1 < x + 1. i.2. B.110 B TRIGONOMETRIC FUNCTIONS (c) Formula (B.10) (or just use l’Hôpital’s rule = t→0 2t t2 0 2 t→0 t 2 B.2.43 at x = 3π/2.8) and the formulas in Problems 3 and 4 give sin(a + x) − sin(a − x) sin a cos x + cos a sin x − sin a cos x + cos a sin x = cos(a + x) − cos(a − x) cos a cos x − sin a sin x − cos a cos x − sin a sin x cos a 2 cos a sin x =− = − cot a = −2 sin a sin x sin a − B) we get sin A − sin B = sin(x + y) − sin(x − y) = sin x cos y + cos x sin y − sin x cos y + cos x sin y = 2 cos x sin y = 2 cos A + B sin A − B . Atle Seierstad. Then 2u dx du dx cos cos2 (x 2 ) (d) (e2x cos x) = (e2x ) cos x + e2x (cos x) = 2e2x cos x − e2x sin x = e2x (2 cos x − sin x).1.1 (b) (x cos x) = x cos x + x(cos x) = cos x − x sin x.1. and Peter Hammond 2008 . 2 2 B.2. d 2 d du 2 B.8 With x = 1 2 (A 1 2 (A + B) and y = (b) y = 2 + cos x. The chain rule yields =√ .9 (a) Let u = 2x. du 1 2x d d (tan(x 2 )) = (tan u) = 2x = . (A cosine curve with amplitude 1 and period 2π shifted 2 units upwards. Knut Sydsæter. B. (c) Let u = x 2 . B. f is strictly decreasing in the closed interval I = [0.1.5 (c) lim again). 1 1 1 − cos t sin t sin t “0” = lim = lim = by (B. B. © Arne Strøm.12 (a) This is a sine curve y = A sin(ax) with A = 2 and a(8π) = 2π .2.1. t→0 B.4 (c) All you need is the chain rule and a steady hand. and attains its maximum value −1 at x = 0 and its minimum value −(2 + 3π/2)3 ≈ −302.2 B.13 Since the lengths of the line segments AC and BD are equal.
it is usually a good idea to multiply both the numerator and the denominator by c − di. where C1 = 2 C.2.11 (a) Let u = cos x. Then √ . B.2. Knut Sydsæter. b. cos5 x sin x dx = −u5 dx = − 1 u6 + C = − 1 cos6 x + C. and we get I = 2 (x − sin x cos x) + C1 .B TRIGONOMETRIC FUNCTIONS 111 (c) Let w = B. √ =− √ 2 2 x dx dw dx 2 1−x x 1−w (c) Integration by parts yields I= sin2 x dx = sin x(− cos x) dx = sin x(− cos x) − cos2 x dx = − sin x cos x + (sin x) (− cos x) dx = − sin x cos x + (1 − sin2 x) dx 1 1 Hence. Note: When integrating trigonometric functions it is very easy to get wrong signs here and there.10 √ √ dw 1 1 d d 1 (arccos x ) = (arccos w) = −√ x. so = = 2 1−i 2 2 1+i 1−i 1+i 3 = (−i)3 = −i 3 = −i 2 i = i.3 To simplify a complex fraction (a + bi)/(c + di). © Arne Strøm. Thus. Then du = − sin x dx and − ln  cos x + C. (d) Integration by parts here too. = = = 2 − (5i)2 2 2 −1 − 5i (−1) 1 − 25i 26 26 −3 − 5i − 2i (d) −2i 1 − 2i + i 2 (1 − i)2 1−i = = −i. and integration by parts also often leads to such mistakes. = = + i. we get dv = cos x dx and (c) As in part (a). the conjugate of the original denominator. let u = cos x.3 B. I = − sin x cos x + x − I + C. This has the effect of making the denominator real (and positive) because (c −di)(c +di) = c2 −(di)2 = c2 − d 2 i 2 = c2 + d 2 . d are real. Atle Seierstad. and Peter Hammond 2008 . (b) With v = sin x. so it is a good rule to check the results by ﬁnding the derivatives when that is possible. c. Then cos xesin x dx = tan x dx = − du = − ln u + C = u ev dv = ev + C = esin x + C. (b) = (a) 2 2 2 i(−i) 1 (1 − i)(1 + i) 1−i 2 (3 − 2i)(2 − i) (c) Simplify the numerator and denominator before making the denominator real: = (−1 − i)(3 + 2i) (4 − 7i)(−1 + 5i) 4 − 7i −4 + 27i − 35i 2 31 27 6 − 7i + 2i 2 = + i. where a. (4 − 3i)(−i) −3 − 4i 3 + 5i + 2i 2 1 5 1 + 5i (3 + 2i)(1 + i) = = −3 − 4i. 6 6 B.3.
we have ˙ Page 352. Page 357.3. Theorem 3.1(c): f (x.2(b): . . Page 192.4: z ≈ 1 − x + y + 2 x 2 − 2xy + 2 y 2 1/2 Page 566.3. . 2) is only the semicircle above the taxis. . gm are deﬁned in a set S and that x∗ is an interior point of S. . Problem 12.10: Replace the inequality in (∗) with equality.12. Page 124.112 B TRIGONOMETRIC FUNCTIONS Corrections to FMEA 2nd edn (2008). Vt )  vt−1 + F3 (t − 1. Problem 9. Figure A5. line 8: 2 vT −1 should be 2 VT −1 . line 2: .5(c). xt .2.1: See SM. . Page 454. . Problem 12. Page 164.7.5. line 2: . . 3 3 Page 457. . . line 2: . .3. line 2: . . line 1–2: Assume that f and g1 . line 3: .5: In parts (b) and (c) the conclusion should be that U (x) = F (f (x)) is convex! Page 117.3. . line 2: f ∗ (r. line 2: Replace (c) by (b). Problem 10.) Page 59.3.1(b): Add the assumption that S is convex. with K(t) > 0 for Page 382: The entries in the ﬁrst column of the table should be “t ∈ [0.1. provided x = 0.8. Problem 6.6. formula (10): E F2 (t. Next.4. .3. Page 570. (See Example B.7.6: The summation goes from t = 0. . . ﬁrst printing (Spelling and grammatical mistakes are not included. all t > 0. t ∗ ]” and “t ∈ (t ∗ . Knut Sydsæter. . Problem 2.11. y) ≈ x + 2y − 2 x 2 − 2xy − 2y 2 3 1 Page 566. that ∞ −t t0 f (t. . . . . and the scrap value is XT . Problem 2. Problem 3. . . for t = T . and Peter Hammond 2008 . s) = 4 r 2 . Page 508: See SM (not Problem 11) for the proof of Theorem 14.4.1 are .11. line 12 from the bottom. line 5: . 1 Page 570.6.1: Delete “and x ≡ 0”. Theorem 9. Problem 3. xt . We must be satisﬁed with a function x that satisﬁes the conditions in the problem in an open interval around 0. in Theorem 12. Problem 3. T ]”. . Problem 2. . Problem 6: The differential equation has no solution deﬁned over the entire real line. .3. line 5: The function f must also be continuous. Problem 3.6. .3(c). Theorem 2. 1 Page 566. Page 455. Vt ). . .6. wx = 1/2. Page 444. Page 225. Problem 6. Atle Seierstad. Page 576. Page 199. Page 574. u(t))e dt < ∞ for all .8. Page 571: New Problem 3. to ﬁnd the squares of the largest Page 570.) © Arne Strøm.2. . (We require x > 0.1. xt+1 (xt . Problem 5.2. .) Page 26. Page 132. vt−1 ) = 0 Page 455. Theorem 3. xt−1 .5. x(t). line 1: Drop the labels (a) and (b).2: The integral curve through (0. .
6: Drop the references to (a) and (b). © Arne Strøm. ∞) p(t) = e−t .2. . x2 = 1 − v0 − v1 . Problem 11.6. line 1: (a) x ∗ ≈ −2.6. . Problem 5.3: (a) . but it need not be lower hemicontinuous. and Peter Hammond 2008 .7. Problem 6. Page 593.7.3: Remove (b) from line 2.3.3. at = 2 (1 + at+1 )1/2 for t < T .. line 1: .6: ut (x) = xt /4at2 = xt /(1 + at+1 ). line 2: Replace x∗ by x ∗ and t ∗ by t . Problem 10.3. Problem 12. .9.4. . x = −3 and . Problem 9. . Page 582. x3 = 3 2 − v0 − v 1 − v 2 Page 599. .B TRIGONOMETRIC FUNCTIONS 113 Page 577. . aT = a/2.2: . Problem 10. Page 583. When u = 2. Page 588. A5.7. Atle Seierstad. Page 595. ˙ Page 591. . .94753. Knut Sydsæter. . .8: x1 = 1 2 − v0 . . . where A = 2e/(e2 − 1). line 2: . .3(b).2.4: (x ∗ (t). 1) if t ∈ [−1.11. ˙ ˙ Page 589. u∗ (t)) = e(α−2β)T − e(α−2β)t eαt e(α−2β)T +αt e2(α−β)t − (α−2β)T = . . Problem 7. .3. 0] . 1 2 2 Page 595. . line 3: . Problem 12. Page 586.2.6..12. Problem 9. Page 590.4. 0) if t ∈ (0. . (e − 1. Problem 9. . x ∗ (t) = Page 589. e(α−2β)T − 1 e −1 e(α−2β)T − 1 (e − e−t . Problem 9. and so C ∗ /C ∗ + λ/λ = 0.7. See Fig.3: The demand correspondence ξ is upper hemicontinuous. Problem 14. .