PART D

Complex
Analysis

Chap. 13 Complex Numbers and Functions.

Complex Differentiation
Complex numbers appeared in the textbook before in different topics. Solving linear homogeneous ODEs
led to characteristic equations, (3), p. 54 in Sec. 2.2, with complex numbers in Example 5, p. 57, and Case
III of the table on p. 58. Solving algebraic eigenvalue problems in Chap. 8 led to characteristic equations
of matrices whose roots, the eigenvalues, could also be complex as shown in Example 4, p. 328.
Whereas, in these type of problems, complex numbers appear almost naturally as complex roots of
polynomials (the simplest being x2 C1 D 0), it is much less immediate to consider complex analysis—the
systematic study of complex numbers, complex functions, and “complex” calculus. Indeed, complex
analysis will be the direction of study in Part D. The area has important engineering applications in
electrostatics, heat flow, and fluid flow. Further motivation for the study of complex analysis is given on
p. 607 of the textbook.
We start with the basics in Chap. 13 by reviewing complex numbers z D x C yi in Sec. 13.1 and
introducing complex integration in Sec.13.3. Those functions that are differentiable in the complex, on
some domain, are called analytic and will form the basis of complex analysis. Not all functions are
analytic. This leads to the most important topic of this chapter, the Cauchy–Riemann equations (1),
p. 625 in Sec. 13.4, which allow us to test whether a function is analytic. They are very short but you
have to remember them! The rest of the chapter (Secs. 13.5–13.7) is devoted to elementary complex
functions (exponential, trigonometric, hyperbolic, and logarithmic functions).
Your knowledge and understanding of real calculus will be useful. Concepts that you learned in real
calculus carry over to complex calculus; however, be aware that there are distinct differences between
real calculus and complex analysis that we clearly mark. For example, whereas the real equation e x D 1
has only one solution, its complex counterpart e z D 1 has infinitely many solutions.
Sec. 13.1 Complex Numbers and Their Geometric Representation
Much of the material may be familiar to you, but we start from scratch to assure everyone starts at the
same level. This section begins with the four basic algebraic operations of complex numbers (addition,
subtraction, multiplication, and division). Of these, the one that perhaps differs most from real numbers
258 Complex Analysis Part D

is division (or forming a quotient). Thus make sure that you remember how to calculate the quotient of
two complex numbers as given in equation (7), Example 2, p. 610, and Prob. 3. In (7) we take the number
z2 from the denominator and form its complex conjugate zN 2 and a new quotient zN2=zN2. We multiply
the given quotient by this new quotient zN 2=zN2 (which is equal to 1 and thus allowed):

z 1 ;
D z1 D z1 D z1 zNN2
z2 z2 z2 z2

which we multiply out, recalling that i2 D 1 [see (5), p. 609]. The final result is a complex number in a
form that allows us to separate its real (Rez/ and imaginary (Imz/ parts. Also remember that 1=i D i
(see Prob. 1), as it occurs frequently. We continue by defining the complex plane and use it to graph
complex numbers (note Fig. 318, p. 611, and Fig. 322, p. 612). We use equation (8), p. 612, to go from
complex to real.

Problem Set. 13.1. Page 612

1. Powers of i. We compute the various powers of i by the rules of addition, subtraction,

multiplication, and division given on pp. 609–610 of the textbook. We have formally that
i2 D ii
D .0;1/.0;1/ [by (1), p. 609]
(I1) D .0 0 1 1;0 1 C 1 0/ [by (3), p. 609]
D .0 1;0 C 0/ (arithmetic)

D . 1;0/ [by (1)],

D1
where in (3), that is, multiplication of complex numbers, we used x1 D 0, x2 D 0, y1 D 1, y2 D 1.

(I2) i3 D i2i D . 1/ i D i:

Here we used (I1) in the second equality. To get (I3), we apply (I2) twice:

(I3) i4 D i2i2 D . 1/ . 1/ D 1: (I4) i5 D i4i D 1 i D i;

and the pattern repeats itself as summarized in the table below.

We use (7), p. 610, in the following calculation:

1 D 1 iND 1 . i/ D .1 C 0i/.0 i/ D 1 0 CC 0 1 C 0 0C1 1 D D

(I5) i 0 i i: i i iN i . i/ .0 C i/.0 i/ 02 12 02 12
By (I5) and (I1) we get
Chap. 13 Complex Numbers and Functions. Complex Differentiation 259

1 1 1 2

(I6) D D . i/. i/ D . 1/i . 1/i D 1 i D 1;

i2 i i
2
11 1
. 1/. i/ i [from (I6) and (I5)];
i3 D i iD D
2 2
11 1
. 1/. 1/ 1,
i4 D i iD D

and the pattern repeats itself. Memorize that i 2 D 1 and 1=i D i as they will appear quite

frequently.

i8 i9 : .
i4 i5 i6 i7
Start! i0 i i2 i3
1 i 1 i

Start 1=i4 1=i3 1=i2 1=i

1=i8 1=i7 1=i6 1=i5
. . 1=i10 1=i9

Sec. 13.1. Prob. 1. Table of powers of i

3. Division of complex numbers

a. The calculations of (7), p. 610, in detail are

C
D z1 x1 iy1 z2 D x2 C

iy2

z(by definition of z1 and z2)

C
D x1 iy1 x2 iy2
(N.B. corresponds to multiplication by 1)
x2 C iy2 x2 iy2
260 Complex Analysis Part D

D .x1 C iy1/.x2 iy2/

.x2 C iy2/.x2 iy2/ x1x2 x1iy2 C

x2iy2 C iy2x2 iy2iy2 ix1y2 C (multiplying it out: (3) in notation (4), p. 609)
D x2x2
ix2y1 i2y1y2 (grouping terms, using commutativity)
x1x2
D x22 ix2y2 C ix2y2 i2y22 ix1y2 C
(using i2 D 1 and simplifying)
x1x2 ix2y1 C y1y2
D
x22 C y22

CC
iy1x2 iy1iy2 x1x2 y1y2 C x2y1

Cx1y2

i (breaking into real part and imaginary part).

D x22 C y22 x22 y22

b. A practical example using (7) is

26 18i D .26 18i/ .6 C 2i/ D 26 6 C 26 2i C18 6i 18 2i2

6 2i .6 2i/ .6 C 2i/ 62 22

D 156 C 52i C108i C 36 D 192 56i D

4:8 1:4i:
36 4 40
5. Pure imaginary number a. If z D x C iy is pure imaginary, then zND z:
Proof. Let z D x C iy be pure imaginary. Then x D 0, by definition on the bottom of p. 609.

Hence

(A) z D iy and (B) zND iy (by definition. of complex conjugate, p. 612).

If we multipy both sides of (A) by 1, we get

z D iy;

which is equal to zN, hence

Chap. 13 Complex Numbers and Functions. Complex Differentiation 261

z DNz:

b. If zND z then z D x C iy is pure imaginary.

Proof. Let z D x C iy so that zND x iy. We are given that zND z, so

zND x iy D z D .x C iy/ D x iy:

By the definition of equality (p. 609) we know that the real parts must be equal and that the
imaginary parts must be equal. Thus

RezND Re. z/; x D

x;

2x D 0; x

D 0;

and

ImzND Im. z/; y D

y;

which is true for any y. Thus

z D x C iy D iy:

But this means, by definition, that z is pure imaginary, as had to be shown.

11. Complex arithmetic

z1 z2 D . 2 C 11i/ .2 i/

D 2 C 11i 2CiD.2 2/ C .11 C 1/i D 4 C 12i

.z1 z2/2 D . 4 C 12i/. 4 C 12i/ D 16 48i 48i 144 D 128 96i

2
.z 1z 2/ D 128 96 8 16 25 4 3 D
16 16 16 D 16 2 i 8 6i:
Next consider
z1 z2 2
:
4 4
We have
z2 11 z
. 2 C 11i/ D C i; i:
4 4 4 4 4
262 Complex Analysis Part D

Their difference is
z1 z2 2 2 11 1
i 1 3i:
4 4D4 4C4C4DC
Hence
z1 z2 2 2

D . 1 C 3i/. 1 C 3i/ D 1 3i 3i C 9i D 1 6i 9D8 6i;

4 4
which is the same result as before.

z
19. Real part and imaginary part of z= N. For z D x C iy, we have by (7), p. 610,

z D z zNND zN z

zN zN zNN z z

since the conjugate of the conjugate of a complex number is the complex number itself (which you
may want to prove!). Then

z D z2 D .x C iy/2 D x2 C 2ixyC y2 D x2 Cy2 C 2Cxy

i :

zN zzN x2 C y2 x2 y2 x2 y2 x2 y2

Hence we get the result as shown on p. A34 of the textbook:

z x2 y2 z 2xy
Re D C I Im N D C : zN x2 y2 z x2 y2
Sec. 13.2 Polar Form of Complex Numbers. Powers and Roots
Polar coordinates, defined by (1) and (2) on p. 613, play a more important role in complex analysis than
in calculus. Their study gives a deeper understanding of multiplication and division of complex numbers
(pp. 615–616) and absolute values. More details are as follows.
The polar angle (taken counterclockwise, see Fig. 323, p. 614) of a complex number is determined
only up to integer multiples of 2. While often this is not essential, there are situations where it matters.
For this purpose, we introduce the concept of the principal value Arg z in (5), p. 614, and illustrate it in
Example 1, Probs. 9 and 13.
The triangle inequality defined in (6), p. 614, and illustrated in Example 2, p. 615, is very important
since it will be used frequently in establishing bounds such as in Chap. 15.
Often it will be used in its generalized form (6*), p. 615, which can be understood by the following
geometric reasoning. Draw several complex numbers as little arrows and let each tail coincide with the
preceding head. This gives you a zigzaging line of n parts, and the left side of (6*) equals the distance
Chap. 13 Complex Numbers and Functions. Complex Differentiation 263

from the tail of z1 to the head of zn. Can you “see” it? Now take your zigzag line and pull it taut; then you
have the right side as the length of the zigzag line straightened out.
In almost all cases when we use (6*) in establishing bounds, it will not matter whether or not the right
side of (6*) is much larger than the left. However, it will be essential that we have such an upper bound
for the absolute value of the sum on the left, so that in a limit process, the latter cannot go to infinity.

The last topic is roots of complex numbers, illustrated in Figs. 327–329, p. 617, and Prob. 21. Look at
these figures and see how, for different n, the roots of unity (16), p. 617, lie symmetrically on the unit
circle.

Problem Set 13.2. Page 618

1. Polar form. Sketch z D 1 C i to understand what is going on. Point z is the pointjjDp .1;1/ in the

complex plane. From this we see that the distance of z from the origin is z 2. This is the absolute

value of z. Furthermore, z lies on the bisecting line of the first quadrant, so that its argument (the

angle between the positive ray of the x-axis and the segment from 0 to z) is 45 ı or =4.

Now we show how the results follow from (3) and (4), p. 613. In the notation of (3) and (4) we
have z D x C iy D 1 C i: Hence the real part of z is x D 1 and the imaginary part of z is y D 1. From (3)
we obtain

p
jz jD 12 C 12 Dp2;
as before. From (4) we obtain
y
tan D D 1; D 45ı or :
x 4
Hence the polar form (2), p. 613, is

Dp C z 2 cos i
sin :
4 4
Note that here we have explained the first part of Example 1, p. 614, in great detail.
264 Complex Analysis Part D

1 1+i

o
45
x
1

Sec. 13.2 Prob. 1. Graph of z D 1 C i in the complex plane

5. Polar form. We use (7), p. 610, in Sec. 13.1, to obtain
p p8 C 232i p
pp2 C 1332i D p2 C 1332i
(A) :
8 i 8 i 8 C 3i

The numerator of (A) simplifies to

p 1 pC 2 D p C2p 1p 2 D 38 C2p 1 p D
2 C 3i 8 3i 16 3 2 3 8i 9 9 3 2 32 2 i:

The denominator of (A) is

2
:
p C 22 D 8 CDCD
8 3

Putting them together gives the simplification of (A), that is,

pp2 C 31i D 38 D 38 9 D 38 D 1:

8 i 9 76 76 2

1
Hence z D corresponds to 2;0 in the complex plane. Furthermore, by (3), p. 613,

jzjD r Dpx2 C y2 D 122 C 02 D

and by (4), p. 613,
y 0
tan D D 1 D 0I D 180ı D :
x 2

Hence by (2), p. 613, the desired polar form is

z D r.cos C i sin/ D .cos C i sin/:

Chap. 13 Complex Numbers and Functions. Complex Differentiation 265

o
180

x
–1
2

Sec. 13.2 Prob. 5. Graph of z D in the complex plane

7. Polar form. For the given z we have

jzjDq12 C 122 Dq C 1 2;
1 4

y 1 1
1
tan D D 2 D I D arctan
:x 1 2 2
The desired polar form of z is
q

z
Djzj.cos C i sin/ D 1 C 142 cos arctan 12C i sin
arctan 12:

9. Principal argument. The first and second quadrants correspond to 0 Argz . The third and fourth
quadrants correspond to <Arg z 0. Note that Arg z is continuous on the positive real semiaxis and
has a jump of 2 on the negative real semiaxis. This is a convenient convention. Points on the
negative real semiaxis, e.g., 4:7, have the principal argument Argz D : To find the principal
argument of z D 1 C i, we convert z to polar form:

p
jzjD . 1/2 C 12 Dp2;

y 1 tan D D D 1: x 1
Hence

D D 135ı:

Hence z, in polar form, is

z Dp2 cos 34 C i sin 34:

As explained near the end of p. 613, is called the argument of z and denoted by argz: Thus is
266 Complex Analysis Part D

D argz D ˙ 2n; n D 0;1;2; :

The reason is that sine and cosine are periodic with 2, so 135 ı looks the same as 135ı C 360ı, etc.

To avoid this concern, we define the principal argument Arg z [see (5), p. 614]. We have

Arg z D :

You should sketch the principal argument.

13. Principal argument. The complex number 1 C i in polar form is

p C

1 C i D 2 cos i sin by Prob. 1.

4 4

Then, using DeMoivre’s formula (13), p. 616, with r Dp2 and n D 20,

C 20 p 20 C .1 i/ D 2 cos 20 i sin 20 by Prob. 1:

4 4
D 210 .cos5 C i sin5/ D 210

.cos C i sin/:

Hence
argz D ˙ 2n; n D 0;1;2;I Argz D :

Furthermore, note that

.1 C i/20 D 210 .cos C i sin/ D 210. 1 C i 0/ D 210 D 1024:

Graph the prinicipal argument.

17. Conversion to x+iy. To convert from polar form to the form x C iy, we have to evaluate sin
and cos for the given : Here

!
p 1 C 1 Dp p2 C p2 Dp16 Cp16 DC
8 cos i sin 8 i i 2 2i:
4 4 2 2 2 2

21. Roots. From Prob. 1 and Example 1, p. 614 in this section, we know that 1 C i in polar form is
Chap. 13 Complex Numbers and Functions. Complex Differentiation 267

1 C i Dp2 cos 14 C i cos 14:

Hence by (15), p. 617,

p3 1 C i D .1 C i/1=3 Dp21=3 cos 41 C 2k C i cos 14 C 2k :

3 3

Now we can simplify

p 1=3 D 1=21=3 D 1=6

2 2 2

and
C 2k D =4 C 2k D C 8k D .1 C 8k/
:
3 3 3 12 12 12
Hence

.1 8k/ .1 8k/
p3 C D 1=6 CC C
1 i 2 cos i sin ;

12 12
where k D 0;1;2 (3 roots; thus 3 values of k). Written out we get
1=6
For k D 0 z0 D 2 cosC i sin:
12 12

1=6 9 9
For k1 D z1 D 2 cosC i sin:
12 12

17 17
1=6
D
For k 2 z2 D 2 cos C i sin :
12 12

The three roots are regularly spaced around a circle of radius 2 1=6 D 1:1225 with center 0.
268 Complex Analysis Part D

o
120

~1.12
~1.12 o
15
o x
120
~1.12

Sec. 13.2. Prob. 21. The three roots z0; z1;z2 of z Dp3 1 C i in the complex plane

29. Equations involving roots of complex numbers. Applying the usual formula for the solutions

b ˙pb2 4ac
zD
2a of a quadratic equation we first have
to

(Eq) z2 zC1 i D 0;

(A) zD
:

Now, in (A), we have to simplify p 3 C 4i : Let z D p C qi be a complex number where p;q are real.
Then
z2 D .p C qi/2 D p2 q2 C 2pqi D 3 C 4i:
We know that for two complex numbers to be equal, their real parts and imaginary parts must be
equal, respectively. Hence, from the imaginary part
2pq D 4;

(B) pq D 2;

2
p
D :q
This can then be used, in the real part,
Chap. 13 Complex Numbers and Functions. Complex Differentiation 269

p2 q2 D 3;

2 4
pC 2 D 3;

p
p4 C 4 D 3p2; p4

3p2 C 4 D 0:

To solve this quartic equation, we set h D p 2 and get the quadratic equation

h2 C 3h 4 D 0;

which factors into

.h 1/.h C 4/ D 0 so that hD1 and h D 4:

Hence
p2 D 1 and p2 D 4:

Since p must be real, p2 D 4 is of no interest. We are left with p 2 D 1 so

(C) (a) p D 1; (b) p D 1:

Substituting [C(a)] into (B) gives

pq D 1 q D 2 so q D 2:

Similarly, substituting [C(b)] into (B) gives

pq D . 1/ q D 2 so q D 2:

We have p D 1; q D 2 and p D 1, q D 2. Thus, for z D p C qi (see above), we get

1 C 2i and 1 2i D .1 C 2i/:

Hence (A) simplifies to

p
1 ˙p 3 C 4i D 1 ˙ .1 C 2i/2 D 1 ˙ .1 C 2i/ z D:

2 2 2
This gives us the desired solutions to (Eq), that is,

1 D 1 C .1 C 2i/ D 2i D z i
2 2
and
270 Complex Analysis Part D

2 1 .1 C 2i/ D 2 2i D
zD 1 i:
2 2
Verify the result by plugging the two values into equation (Eq) and see that you get zero.
Sec. 13.3 Derivative. Analytic Function

The material follows the calculus you are used to with certain differences due to working in the complex
plane with complex functions f.z/. In particular, the concept of limit is different as z may approach z0 from
any direction (see pp. 621–622 and Example 4). This also means that the derivative, which looks the
same as in calculus, is different in complex analysis. Open the textbook on p. 623 and take a look at
Example 4. We show from the definition of derivative (4), p. 622, which uses the concept of limit, that
f.z/ DNz is not differentiable. The essence of the example is that approaching z along path I in Fig. 334,
p. 623, gives a value different from that along path II. This is not allowed with limits (see pp. 621–622).
We call those functions that are differentiable in some domain analytic (p. 623). You can think of them
as the “good functions,” and they will form the preferred functions of complex analysis and its
applications. Note that f.z/ DNz is not analytic. (You may want to build a small list of nonanalytic
functions, as you encounter them. In Sec. 13.4 we shall learn a famous method for testing analyticity.)
The differentiation rules are the same as in real calculus (see Example 3, pp. 622–623 and Prob. 19).
Here are two examples
f.z/ D .1 z/16; f 0.z/ D 16.1 z/15.

1/ D 16.1 z/15;

where the factor . 1/ comes from the chain rule;

f.z/ D i; f 0.z/ D 0

since i is a constant.
Go over the material to see that many concepts from calculus carry over to complex analysis. Use this
section as a reference section for many of the concepts needed for Part D.

Problem Set 13.3. Page 624

1. Regions of practical interest. Closed circular disk. We want to write

jz C 1 5ij

in the form jz aj p

as suggested on p. 619. We can write

jz C 1 5ijDjz C .1 5i/j

Djz . .1 5i//j Djz . 1 C 5i/j:

Hence the desired region

Chap. 13 Complex Numbers and Functions. Complex Differentiation 271

jz . 1 C 5i/j

is a closed circular disk with center 1 C 5i (not 1 5i!) and radius .

7. Regions. Half-plane. Let z D x C yi. Then Rez D x as defined on p. 609. We are required to
determine what
Rez 1

means. By our reasoning we have Rez D x 1 so that the region of interest is

x 1:
y

x
1

1 – 5i

3
–5 2

Sec. 13.3. Prob. 1. Sketch of closed circular disk jz C 1 5ij

This is a closed right half-plane bounded by x D 1, that is, a half-plane to the right of x D 1 that
includes the boundary.
y

–1 x

Sec. 13.3. Prob. 7. Sketch of half-plane Rez 1

272 Complex Analysis Part D

11. Function values are obtained, as in calculus, by substitution of the given value into the given
function. Perhaps a quicker solution than the one shown on p. A35 of the textbook, and following the
approach of p. 621, is as follows. The function
1
f.z/ D evaluated at z D 1
i1 z
is
1 1 1
f.1 i/ D D C D D i;
1 .1 i/ 1 1 i i
with the last equality by (I5) in Prob. 1 of Sec. 13.1 on p. 258 of this Manual. Hence
Ref D Re.i/ D 0, Imf D Im.i/ D 1:

17. Continuity. Let us use polar coordinates (Sec. 13.2) to see whether the function defined by
8 z¤0
Re.z/
D < for zD0
for
f.z/1 jzj
:0
is continuous at z D 0. Then x D r cos, y D r sin by (1), p. 613, and, using the material on p. 613,
we get
Re.z/ x r cos
f.z/ D jjD jjD . 1 z 1 z 1 r
We note that as r ! 0,

1 r!1 and r cos ! 0

so that
r cos
!0 as r ! 0; for any value of :
1 r
By (3), p. 622, we can conclude that f is continuous at z D 0:

19. Differentiation. Note that differentiation in complex analysis is as in calculus. We have

f.z/ D .z 4i/8;

f 0.z/ D 8.z 4i/7;

f 0.3 C 4i/ D 8.3 C 4i 4i/7 D 8 37 D 8 2187 D 17;496:

Chap. 13 Complex Numbers and Functions. Complex Differentiation 273

Remark. Be aware of the chain rule. Thus if, for example, we want to differentiate

g.z/ D . 2z 4i/10; then

g0.z/ D 10. 2z 4i/9 . 2/ D 20. 2z 4i/9;

where the factor 2 comes in by the chain rule.

Sec. 13.4 Cauchy–Riemann Equations. Laplace’s Equation

We discussed the concept of analytic functions in Sec. 13.4 and we learned that these are the functions
that are differentiable in some domain and that operations of complex analysis can be applied to them.
Unfortunately, not all functions are analytic as we saw in Example 4, p. 623. How can we tell whether a
function is analytic? The Cauchy–Riemann equations (1), p. 625, allow us to test whether a complex
function is analytic. Details are as follows.
If a complex function f.z/ D u.x;y/ C iv.x;y/ is analytic in D, then u and v satisfy the Cauchy–Riemann
equations

(1) ux D vy; uy D vx

(Theorem 1, p. 625) as well as Laplace’s equations r2u D 0, r2v D 0 (Theorem 3, p. 628; see also
Example 4, p. 629, and Prob. 15). The converse of Theorem 1 is also true (Theorem 2, p. 627), provided
the derivatives in (1) are continuous. For these reasons the Cauchy–Riemann equations are most
important in complex analysis, which is the study of analytic functions.
Examples 1 and 2, p. 627, and Probs. 3 and 5 use the Cauchy–Riemann equations to test whether
the given functions are analytic. In particular, note that Prob. 5 gives complete details on how to use
the Cauchy–Riemann equations (1), p. 625, in complex form (7), p. 628, and even how to conclude
nonanlyticity by observing the given function. You have to memorize the Cauchy–Riemann equations
(1). Remember the minus sign in the second equation!
Problem Set 13.4. Page 629

3. Check of analyticity. Cauchy–Riemann equations (1), p. 625. From the given function

f.z/ D e 2x .cos2y i sin2y/

D e 2x cos2y C i e 2x sin2y

we see that the real part of f is

u D e 2x cos2y

and the imaginary part is

v D e 2x sin2y:
274 Complex Analysis Part D

To check whether f is analytic, we want to apply the important Cauchy–Riemann equations (1),
p. 625. To this end, we compute the following four partial (real) derivatives:

ux D 2e 2x cos2y; vy D e 2x .2cos2y/ D 2e 2x

cos2y; uy D e 2xŒ. sin2y/ 2 D 2e 2x

sin2y; vx D e 2x. 2/.sin2y/ D 2e 2x sin2y:

Note that the factor 2 in ux and the factor 2 in vy result from the chain rule. Can you identify
the use of the chain rule in the other two partial derivatives? We see that

ux D cos2y D vy

and
uy D 2e 2x sin2y D vx:

This shows that the Cauchy–Riemann equations are satisfied for all z D x Ciy and we conclude that f
is indeed analytic.
In Sec. 13.5 we will learn that the given function f defines the complex exponential function
ez; with z D 2x C i2y and that, in general, the complex exponential function is analytic.

5. Not analytic. We show that f.z/ D Re.z2/ i Im.z2/ is not analytic in three different ways.

Solution 1. Standard solution in x,y coordinates. We have that, if z D x C iy, then

z2 D .x C iy/.x C iy/ D x2 C 2ixy C i2y2 D .x2 y2/ C i.2xy/:

Thus we see that
Re.z2/ D x2 y2 and Im.z2/ D 2xy:

Thus the given function is

f.z/ D x2 y2 i2y:

Hence
u D x2 y2; v D 2xy:

To test whether f.z/ satisfies the Cauchy–Riemann equations (1), p. 625, we have to take four
partial derivatives
ux D 2x and vy D 2x
so that

(XCR1) u x ¤ v y:
Chap. 13 Complex Numbers and Functions. Complex Differentiation 275

(We could stop here and have a complete answer that the given function is not analytic! However, for
demonstration purposes we continue.)

uy D 2y and vx D 2y

so that

(CR2) uy D vx:

We see that although the given function f.z/ satisfies the second Cauchy–Riemann equation (1),
p. 625, as seen by (CR2), it does not satisfy the first Cauchy–Riemann equation (1) as seen by
(XCR1). We note that the functions u.x;y/, v.x;y/ are continuous and conclude by Theorems 1, p.
625, and 2, p. 627 that f.z/ is not analytic.

Solution 2.
In y D r sin. Hence polar
x2 y2 D r2 cos2 r2 sin2 D r2.cos2 sin2 /;

2xy D 2r2 cos sin:

Together, we get our given function f.z/ in polar coordinates

f.z/ D r2.cos2 cos2 / i2r2 cos sin D r2.cos2 cos2 2i cos sin/:
coordinates. We have z D r.cos C i sin/ by (2), p. 613, so that x D r cos,

We have

u D r2.cos2 sin2 /; v D 2r2

cos sin:

Then the partial derivatives are

ur D 2r.cos2 sin2 /
276 Complex Analysis Part D

and, by the product rule,

v D . 2r2/. sin/.sin/ C . 2r2/cos cos

D 2r2.sin2 cos2 /;

and

1 2 2
v D 2r.sin cos /; r
We see that ur D .1=r/v so that

1
ur ¤ v:
r
This means that f does not satisfy the first Cauchy–Riemann equation in polar coordinates (7),
p. 628, and f is not analytic. (Again we could stop here. However, for pedagogical reasons we
continue.)

vr D 4r cos sin;

u D r2.2sin cos/ 2cos. sin/ D 4r2.sin cos/

and
1 u D 4r.sin cos/: r
This shows that f.z/ satisfies the second Cauchy–Riemann equation in polar coordinates (7), that is,
1
vr D u:
r
However, this does not help, since the first Cauchy–Riemann equation is not satisfied. We
conclude that f.z/ is not analytic.

Solution 3. Observation about f(z). We note that

.zN/2 D .x iy/.x iy/ D x2 2ixy y2 D x 2 y2 2ixy:

We compare this with our given function and see that

f.z/ D .zN/2 DNz Nz:

Furthermore,

f.x/ D .zN/2 D .z2/:

Chap. 13 Complex Numbers and Functions. Complex Differentiation 277

From Example 4, p. 623 in Sec. 13.3, we know that zN is not differentiable so we conclude that
the

given f.x/ D .z2/ is also not differentiable. Hence f.z/ is not analytic (by definition on p. 623).

Remark. Solution 3 is the most elegant one. Solution 1 is the standard one where we stop when the
first Cauchy–Riemann equation is not satisfied. Solution 2 is included here to show how the Cauchy–
Riemann equations are calculated in polar coordinates. (Here Solution 2 is more difficult than
Solution 1 but sometimes conversion to polar makes calculating the partial derivatives simpler.)

15. Harmonic functions appear as real and imaginary parts of analytic functions.
First solution method. Identifying the function.
If you remember that the given function u D x=.x 2 C y2/ is the real part of 1=z, then you are done.
Indeed,
1 1
z D x C iy

1 x iy
D x C iy x iy

x iy

x2
Cy
i
D x2 C y2 C x2 C y2
so that clearly

1 x
Re D C ; z x2 y2
and hence the given function u is analytic. Moreover, our derivation also shows that a conjugate
harmonic of u is y=.x2 C y2/.
Second solution method. Direct calculation as in Example 4, p. 629.
If you don’t remember that, you have to work systematically by differentiation, beginning with
proving that u satisfies Laplace’s equation (8), p. 628. Such somewhat lengthy differentiations, as
well as other calculations, can often be simplified and made more reliable by introducing suitable
shorter notations for certain expressions. In the present case we can write
278 Complex Analysis Part D

x 22 uD ; where G D x C y :
G
Then

(A) Gx D 2x; Gy D 2y:

By applying the product rule of differentiation (and the chain rule), not the quotient rule, we obtain
the first partial derivative.

1 x.2x/
(B) ux D G G2 :

By differentiating this again, using the product and chain rules, we obtain the second partial
derivative:

2x 4x 8x3

(C) uxx D G2 G2 C G3 :

Similarly, the partial derivative of u with respect to y is obtained from (A) in the form

2xy
(D) uy D 2:

G
The partial derivative of this with respect to y is

2x 8xy2
(E) uyy D G2 C G3 :

Adding (C) and (E) and remembering that G D x 2 C y2 gives us

C D 8x C 8x.x2 C y2/ D 8x C 8x D uxx uyy G2 G3 G2 G2 0:

This shows that u D x=G D x=.x2 Cy2/ satisfies Laplace’s equation (8), p. 628, and thus is harmonic.
Next we want to determine a harmonic conjugate. From (D) and the second Cauchy–Riemann
equation (1), p. 625, we obtain
2xy
uy D 2 D vx:

G
2 2
Integration of 2x=G D Gx=G , with respect to x, gives 1=G, so that integration of vx, with
respect to x, gives
Chap. 13 Complex Numbers and Functions. Complex Differentiation 279

y y
(F) v D D C C h.y/:
G x2 y2

Now we show that h.y/ must be a constant. We obtain, by differentiating (F) with respect to y and
taking the common denominator G2, the following:

D 1 C 2y2 D x2 C y2 C vy G G2 G2 h0.y/:

On the other hand, we have from (B) that

1 2x2 y2 x2 ux D
G G2 D G2 :

By the first Cauchy–Riemann equation (1), p. 625, we have

vy D ux;

which means, written out, in our case

x2 C y2 C D y2 x2

2 h0.y/ G2 :
G

But this means that

h0.y/ D 0 and hence h.y/ D const;

as we claimed. Since this constant is arbitrary, we can choose h.y/ D 0 and obtain, from (F), the
desired conjugate harmonic

y y
vD C C h.y/ D C ; x2 y2
x 2 y2
which is the same answer as in our first solution method.

Sec. 13.5 Exponential Function

Equation (1), p. 630, defines the complex exponential function. Equations (2) and (3) on that page are as
in calculus. Note that equation (4), p. 631, is a special case of equation (3). The Euler formula (5), p. 631,
is very important and gives the polar form (6) of

z D x C iy D r .cos C i sin / D rei :

280 Complex Analysis Part D

It would be useful for you to remember equations (7), (8), and (9). The periodicity (12), p. 632, has no
counterpart in real. It motivates the fundamental region (13), p. 632, of e z.
Solving complex equations, such as Prob. 19, gives practice in the use of complex elementary
functions and illustrates the difference between these functions and their real counterparts. In
particular, Prob. 19 has infinitely many solutions in complex but only one solution in real!
Problem Set 13.5. Page 632

5. Function values. We note that e2C3i D ez D exCiy: Thus we use

(1), p. 630, with x D 2 and y D 3 and obtain e 2C3 i D e2.cos3 C i

sin3/

D e2.cos. C 2/ C i sin. C 2// since cos.3/ D cos. C 2/; same for sin.3/
D e2.cos C i sin/ (cos and sin both have periods of 2/

D e2. 1 C i 0/

D e2 7:389:

From (10), p. 631, we have the absolute value ˇe 2C3 iˇˇDjez jD ex D e2

7:389:

9. Polar form. We want to write z D 4 C 3i in exponential form (6), p. 631. This means expressing it in
the form z D rei:

We have

p p
r Djz jD x2 C y2 D 42 C 32 Dp25 D 5:
We know, by Sec. 13.2, pp. 613–619, that the principal argument of the given z is

y 3
Arg z D arctan D arctan D 0:643501:

x 4

Hence, by (6), p. 631, we get that z in polar form is

z
D 5ei arctan.3=4/ D 5e0:643501i:

Checking the answer. By (2), p. 613, in Sec. 13.2, we know that any complex number z D x Ciy has
polar form
Chap. 13 Complex Numbers and Functions. Complex Differentiation 281

z D r.cos C i sin/:

Thus, for z D 4 C 3i, we have

z D 5.cos0:643501 C i sin0:643501/ D

5.0:8 C 0:6i/ D 4 C 3i:

15. Real and imaginary parts. We want to find the real and imaginary parts of exp.z2/: From the
beginning of Sec. 13.5 of the textbook we know that the notation exp means
2
2 z
exp.z / D e :

Now for z D x C iy,

z2 D .x C iy/.x C iy/ D x2 y2 C i2xy:

Thus
ez2 D ex2 y2Ci2xy D ex2 y2ei2xy [by (3), p. 630].

Now
ei2xy D cos.2xy/ C i sin.2xy/ [by (1), p. 630; (5), p. 631].

Putting it together

ez2 D ex2 y2Œcos.2xy/ C i sin.2xy/

D ex2 y2 cos2xy C i.ex2 y2 sin2xy/:

Hence
Reexp.z2/D ex2 y2 cos2xyI Imexp.z2/D ex2 y2 sin2xy;

as given on p. A36 of the textbook.

19. Equation. To solve

(A) ez D 1

we set z D x C iy: Then

ez D exCiy D exeiy D ex.cosy C i siny/ [by (5), p. 631]

D ex cosy C iex siny

D1 [by (A)]
282 Complex Analysis Part D

D 1 C i 0:

Equate the real and imaginary parts on both sides to obtain

(B) Re.ez/ D ex cos y D 1; (C) Im.ez/

D ex sin y D 0:

Since ex > 0 but the product in (C) must equal zero requires that

siny D 0 which means that (D) y D 0;˙;˙2;˙3;::::

Since the product in (B) is positive, cosy has to be positive. If we look at (D), we know that cosy is
1 for y D˙;˙3;˙5;::: but C1 for y D 0;˙2;˙4;:::: Hence (B) and (D) give

(E) y D 0;˙2;˙4;::::

Since (B) requires that the product be equal to 1 and the cosine for the values of y in (E) is 1, we
have ex D 1: Hence

(F) x D 0:

Then (E) and (F) together yield

xD0 y D 0;˙2;˙4;::: ;

and the desired solution to (A) is

z D x C yi D˙2ni; n D 0;1;2;::::

Note that (A), being complex, has infinitely many solutions in contrast to the same equation in
real, which has only one solution.
Sec. 13.6 Trigonometric and Hyperbolic Functions. Euler’s Formula
In complex, the exponential, trigonometric, and hyperbolic functions are related by the definitions (1),
p. 633, and (11), p. 635, and by the Euler formula (5), p. 634, as well as by (14) and (15), p. 635. Thus we
can convert them back and forth. Formulas (6) and (7) are needed for computing values. Problem 9 uses
such a formula to compute function values.

Problem Set 13.6. Page 636

1. Formulas for hyperbolic functions. To show that

coshz D coshx cosy C i sinhx siny

we do the following. We start with the definition of coshz. Since we want to avoid carrying a factor
along, we multiply both sides of (11), p. 635, by 2 and get

2coshz D ez C e z
Chap. 13 Complex Numbers and Functions. Complex Differentiation 283

D exCiy C e x iy (setting z D x C iy)

x x
D
e .cosy C i siny/ C e .cosy i siny/ (by (1), p.

630) D cosy.ex C e x/ C i siny.ex e x/

D cosy.2coshx/ C i siny.2sinhx/ (by (17), p. A65 of Sec. A3.1 of App. 3) D 2coshx cosy C 2i

sinhx siny:

Division by 2 on both sides yields the desired result. Note that the formula just proven is useful
because it expresses coshz in terms of its real and imaginary parts.
The related formula for sinhz follows the same proof pattern, this time start with
2sinhz D ez e z. Fill in the details.

9. Function values. The strategy for Probs. 6–12 is to find formulas in this section or in the problem
set that allow us to get, as an answer, a real number or complex number. For example, the formulas
in Prob. 1 are of the type we want for this kind of problem.
In the present case, by Prob. 1 (just proved before!), we denote the first given complex
number by z1 D 1 C 2i so that x1 D 1 and y1 D 2 and use

coshz1 D coshx1 cosy1 C i sinhx1 siny1:

Then
coshz1 D cosh. 1 C 2i/ D cosh. 1/cos2 C i sinh. 1/sin2:

Now by (11), p. 635,

D D 1 1 1 1 C e2 I D D 1 e2 coshx1 cosh. 1/ .e C e / D sinhx1 sinh. 1/

:
2 2e 2e
Using a calculator (or CAS) to get the actual values we have

C D 1 C e2 C 1e2
cosh. 1 2i/ cos2 i sin2
2e 2e
D 1:543081 . 0:4161468/ C i. 1:752011/ .0:9092974/

D 0:642148 1:068607i;
which corresponds to the rounded answer on p. A36.
For the second function value z2 D 2 i we notice that, by (1), p. 633,

cosz D 12 eiz C e iz
284 Complex Analysis Part D

and, by (11), p. 635,

coshz D 21 .ez C e z/:

Now

(A) iz2 D i. 2 i/ D 2i i2 D 1 2i D z1:

Hence

cosz2 D 12 eiz2 C e iz2 1

D
z1

z1

.e e / [by (A)]
2C

D coshz1

D cosh.1 2i/

so we get the same value as before!

13. Equations. We want to show that the complex cosine function is even.
First solution directly from definition (1), p. 633.
We start with
cos. z/ D 21 ei. z/ C e i. z/:

We see that for any complex number z D x C iy:

i. z/ D iŒ .x C iy/ D i. x iy/ D ix i2y D y ix:

Similarly,
iz D i.x C iy/ D ix i2y D y ix D i. z/:

So we have
iz D i. z/:
Similarly,
i. z/ D i. x yi/ D y C xi

and
iz D i.x C iy/ D ix C i2y D i. z/

so that
iz D i. z/:
Putting these two boxed equations to good use, we have
Chap. 13 Complex Numbers and Functions. Complex Differentiation 285

cos. z/ D 21 ei. z/ C e i. z/D e iz C eizD 21 eiz C e izD cosz:

Thus cos. z/ D cosz, which means that the complex cosine function (like its real counterpart) is
even.
Second solution by using (6a), p. 634. From that formula we know

cosz D cosx coshy i sinx sinhy:

We consider
cos. z/ D cos. x iy/

D cos. x C i. y//
D cos. x/cosh. y/ i sin. x/sinh. y/
D cosx coshy i. sinx/. sinhy/

D coshx coshy i sinx sinhy D

cosz:

The fourth equality used that, for real x and y, both cos and cosh are even and sin and sinh are
odd,
that is,

cos. x/ D cosxI cosh. y/ D coshyI sin. x/ D sinx

sinh. x/ D sinhx:

Similary, show that the complex sine function is odd, that is, sin. z/ D sinz:

17. Equations. To solve the given complex equation, coshz D 0, we use that, by the first equality in
Prob. 1, p. 636, of Sec. 13.6, the given equation is equivalent to a pair of real equations:

Re.coshz/ D coshx cosy D 0; Im.coshz/ D

sinhx siny D 0:

Since coshx ¤ 0 for all x, we must have cosy D 0, hence y D˙.2n C 1/=2 where n D 0;1;2;:::: For
these y we have siny ¤ 0, noting that the real cos and sin have no common zeros! Hence sinhx
D 0 so that x D 0. Thus our reasoning gives the solution

z D .x;y/ D .0; ˙.2n C 1/=2/; that is, z D˙.2n C 1/i=2 where n D 0;1;2;::::
286 Complex Analysis Part D

Sec. 13.7 Logarithm. General Power. Principal Value

Work this section with extra care, so that you understand:

1. The meaning of formulas (1), (2), (3), p. 637.

2. The difference between the real logarithm lnx, which is a function defined for x > 0, and the
complex logarithm lnz, which is an infinitely many-valued relation, which, by formula (3), p.
637, “decomposes” into infinitely many functions.

Example 1, p. 637, and Probs. 5, 15, and 21 illustrate these formulas.

General powers zc are defined by (7), p. 639, and illustrated in Example 3 at the bottom of that
page.
Problem Set 13.7. Page 640

5. Principal value. Note that the real logarithm of a negative number is undefined. The principal
value Ln z of ln z is defined by (2), p. 637, that is,

Ln z D lnjzjC iArg z

where Arg z is the principal value of arg z. Now recall from (5), p. 614 of Sec. 13.2, that the
principal value of the argument of z is defined by

< Arg z :

In particular, for a negative real number we always have Arg z DC, as you should keep in mind.
From this, and (2), we obtain the answer

Ln. 11/ D lnj 11jC i D ln11 C i:

15. All values of a complex logarithm. We need the absolute value and the argument of e i because, by
(1) and (2), p. 637,

ln.ei/ D lnjeijC i arg.ei/

D lnjeijC i Arg.ei/ ˙ 2ni; where n D 0;1;2;::::

Now the absolute value of the exponential function e z with a pure imaginary exponent always equals
1, as you should memorize; the derivation is
q eiyjDjcosy C i sinyjD cos2 y C sin2 y D 1:
j

(Can you see where this calculation would break down if y were not real?) In our case,

(A) jeijD 1;hence lnjeijD ln.1/ D 0:

The argument of ei is obtained from (10), p. 631 in Sec. 13.5, that is,
Chap. 13 Complex Numbers and Functions. Complex Differentiation 287

arg.ez/ D Arg.ez/ ˙ 2n D y ˙ 2n where n D 0;1;2;::::

In our problem we have z D i D x C iy, hence y D 1. Thus

(B) arg.ei/ D 1 ˙ 2n; where n D 0;1;2;::::

From (A) and (B) we obtain the answer

ln.ei/ D lnjeijC i arg.ei/

D 0 C i.1 ˙ 2n/; where n D 0;1;2; :

21. Equation. We want to solve

lnz D 0:6 C 0:4i

D lnjzjC i argz [by (1), p. 637].

We equate the real parts and the imaginary parts:

0:6 D lnjzj; thus jzjD e0:6:

0:4 D argz:
Next we note that

z D elnz D elnjzjCi argz D e0:6e0:4i:

We consider

e0:4i D e0C0:4i D e0.cos0:4 C i sin0:4/ [by (1), p. 630, Sec. 13.5] D cos0:4 C i

sin0:4:

Putting it together, we get

z D e0:60:6e0:4i

D e .cos0:4 C i sin0:4/ D 1:822119 .

0:921061 C 0:389418i/ D 1:6783 C

0:70957i:

23. General powers. Principal value. We start with the given equation and use (8), p. 640, and the
definition of principal value to get
288 Complex Analysis Part D

.1 C i/1 1 D e.1 i/ Ln.1Ci/:

Now the principal value

Ln .1 C i/ D lnj1 C ijC iArg.1 C i/ [by (2), p. 637].

Also

p
j1 C ijD 12 C 12 Dp2

and

Arg.1 C i/ D [see (5) and Example 1, both on p.

614]. 4
Hence

Ln.1 C i/ D lnp2 C i
4
so that

.1 i/Ln.1 C i/ D .1 i/ lnp2 C i
4
D pC p 2 ln 2 i i ln 2 i
4 4

D lnp2 C C i lnp2 :
4 4
Thus

.1 C i/1 1 D exphlnp2 C C i lnp2i

4 4
D pC h p i exp ln 2 exp i2
4 4
D p h pC pi
exp ln 2 exp cos 2 i sin ln 2 [by (1), p. 630]
44

Dp2 e=4 hcos lnp2C i sin lnp2i:

Chap. 13 Complex Numbers and Functions. Complex Differentiation 289

4 4
Numerical values are

pD
ln 2 0:4388246;
4
pD D
cosln 2 cos.0:4388246/ 0:9052517;
4
pD D
sinln 2 sin.0:4388246/ 0:4248757;
=4
4 p2e D 3:1017664:

Hence .1 C i/1 1 evaluates to

.1 C i/1 1 D 2:8079 C 1:3179i:

Chap. 14 Complex Integration
The first method of integration (“indefinite integration and substitution of limits”) is a direct analog of
regular calculus and thus a good starting point for studying complex integration. The focal point of
Chap. 14 is the very important Cauchy integral theorem (p. 653) in Sec. 14.2. This leads to Cauchy’s
integral formula (1), p. 660 in Sec. 14.3, allowing us to evaluate certain complex integrals whose
integrand is of the form f.z/=.z z0/ with f being analytic. The chapter concludes with the surprising
result that all analytic functions have derivatives of all orders. Complex integration has a very distinct
flavor of its own and should therefore make an interesting study. The amount of theory in this chapter
is very manageable but powerful in that it allows us to solve many different integrals.
General orientation. Chapter 13 provides the background material for Chap. 14. We can broadly
classify the material in Chap. 14 as a first approach to complex integration based on Cauchy’s integral
theorem and his related integral formula. The groundwork to a second approach to complex integration
is given in Chap. 15 with the actual method of integration (“residue integration”) given in Chap. 16.
Prerequisite. You should remember the material of Chap. 13, including the concept of analytic
functions (Sec. 13.3), the important Cauchy–Riemann equations of Sec. 13.4, and Euler’s formula (5),
p. 634 in Sec. 13.6. We make use of some of the properties of elementary complex functions when
solving problems—so, if you forgot,—consult Chap. 13 in your textbook. You should recall how to solve
basic real integrals (see inside cover of textbook if needed). You should also have some knowledge of
roots of complex polynomials.

Sec. 14.1 Line Integral in the Complex Plane

The indefinite complex integrals are obtained from inverting, just as in regular calculus. Thus the starting
point for the theory of complex integration is the consideration of definite complex integrals, which are
defined as complex line integrals and explored on pp. 643–646. As an aside, the reader familiar with real
line integrals (Sec. 10.1, pp. 413–419 in the text, pp. 169–172 in Vol. 1 of this Manual) will notice a
similarity between Z Z Z Z Z v dy C i udy C v dx the two. Indeed (8), p.
646, can be used to f.z/ dz D udx C C C make the relationship
between complex C C line integrals and real
line integrals explicit, D Z C Z that is,

.udx v dy/i udy C v dx ;

CC

where C is the curve of integration and the resulting integrals are real.
(However, having not studied real line integrals is not a hindrance to learning and enjoying complex
analysis as we go in a systematic fashion with the only prerequisite for Part D being elementary calculus.)
The first practical method of complex integration involves indefinite integration and substitution of
limits and is directly inspired from elementary calculus. It requires that the function be analytic. The
details are given in Theorem 1, formula (9), p. 647, and illustrated below by Examples 1–4 and Probs. 23
and 27.
A prerequisite to understanding the second practical method of integration (use of a representation of
a path) is to understand parametrization of complex curves (Examples 1–4, p. 647, Probs. 1, 7, and 19).
Indeed, (10), p. 647, of Theorem 2 is a more general approach than (9) of Theorem 1, because Theorem
Chap. 14 Complex Integration 291

2 applies to any continuous complex function not just analytic functions. However, the price of generality
is a slight increase in difficulty.
Problem Set 14.1. Page 651

1. Path. We have to determine the path of

z.t/ D .1 C i/t .2 t 5/:

Since the parametric representation

z.t/ D x.t/ C iy.t/ D 1 C 12it

DtCi t

is linear in the parameter t, the representation is that of a straight line in the complex z-plane. Its
slope is positive, that is

1
y.t/ 2t 1
:
x.t/ D t D 2

The straight-line segment starts at t D 2, corresponding to

z0 D z.2/ D 2 C i 2D2Ci

and ends at t D 5:

z1 D z.5/ D 5 C i:

Sketch it.

7. Path. To identify what path is represented by

z.t/ D 2 C 4eit=2 with 0 t 2

it is best to derive the solution stepwise.

From Example 5, p. 648, we know that

z.t/ D eit with 0 t 2

represents a unit circle (i.e., radius 1, center 0) traveled in the counterclockwise direction. Hence

z.t/ D eit=2 with 0 t 4

also represents that unit circle. Then

z.t/ D 4eit=2 with 0 t 2

292 Complex Analysis Part D

represents a semicircle (half circle) of radius 4 with center 0 traversed in the counterclockwise
direction.
Finally

z.t/ D 2 C 4eit=2 with 0 t 2

is a shift of that semicircle to center 2, corresponding to the answer on p. A36 in App. 2 of the
textbook.
y

2 x

Sec. 14.1 Prob. 7. Semicircle

Remark. Our solution demonstrates a way of doing mathematics by going from a simpler problem,
whose answer we know, to more difficult problems whose answers we infer from the simple
problem.

19. Parametric representation. Parabola. We are given that

yD1 14 x2 where 2 x 2:

Hence we may set

x Dt so that y D 1 14 x2 D 1 14 t 2:

Now, for x D t D 2, we get

1 2
y D1 t D1
4 . 2/2 D 0
so that

z0 D 2 C 0i:

Similarly, z1 D 2 C 0i and corresponds to t D 2. Hence

z.t/ D x.t/ C iy.t/

1 2
DtCi1 t;
4 ( 2 t 2).
21. Integration. Before we solve the problem we should use the Cauchy–Riemann equations to
determine if the integrand Rez is analytic. The integrand

w D u C iv D f.z/ D Rez D x

is not analytic. Indeed, the first Cauchy–Riemann equation

uv D vy [by (1), p. 625 in Sec. 13.4]

is not satisfied because

Chap. 14 Complex Integration 293

ux D 1 but vD0 so that vy D 0:

(The second Cauchy–Riemann equation is satisfied, but, of course, that is not enough for
analyticity.) Hence we cannot apply the first method (9), p. 647, which would be more convenient,
but we must use the second method (10), p. 647.
The shortest path from z0 D 1 C i to z1 D 3 C 3i is a straight-line segment with these points as
endpoints. Sketch the path. The difference of these points is

(A) z1 z0 D .3 C 3i/ .1 C i/ D 2 C 2i:

We set

(B) z.t/ D z0 C .z1 z0/t:

Then, by taking the values t D 0 and t D 1, we have

z.0/ D z0 and z.1/ D z1

because z0 cancels when t D 1: Hence (B) is a general representation of a segment with given endpoints
z0 and z1, and t ranging from 0 to 1.
Now we start with Equation (B) and substitute (A) into (B) and, by use of z 0 D z.0/ D 1 C i, we obtain

z.t/ D x.t/ C iy.t/

D z0 C .z1 z0/t

(C) D 1 C i C .2 C 2i/t

D 1 C 2t C i.1 C 2t/:

We integrate by using (10), p. 647. In (10) we need

f.z.t// D x.t/ D 1 C 2t;

as well as the derivative of z.t/ with respect to t, that is,

dz
z.t/P D D 2 C 2i:
dt

Both of these expressions are obtained from (C).

We are now ready to integrate. From (10), p. 647, we obtain

Z Zb
f.z/ dz D f Œz.t/ z.t/dtP
C a

Z1
294 Complex Analysis Part D

D .1 C 2t/.2 C 2i/
dt
0

Z1
D .22i/ C .1 C 2t/ dt:
0

Now

Z Z Z t2 2

.1 C 2t/ dt D dt C 2 t dt D t C 2 DtCt;
2

so that

1 ˇ
Z f.z/ dz D .2 C 2i/ t C t2 ˇ0
C

D .2 C 2i/.1 C t2/

D 2.2 C 2i/

D 4 C 4i;
which is the final answer on p. A37 in App. 2 of the textbook (with a somewhat different
parametrization).

23. Integration by the first method (Theorem 1, p. 647). From (3), p. 630 of Sec. 13.5 of the text,
we know that ez is analytic. Hence we use indefinite integration and substitution of upper and
lower limits. We have
Z
ez dz D ez C const [by (2), p. 630].

Z 2i h zi2i 2i i
z
(I1) e dz D e De e :
i i

Euler’s formula (5), p. 634, states that

eiz D cosz C i sinz:

Hence

e2i D cos2 C i sin2 D 1 C i 0 D 1;

i
e D cos C i sin D 1C0D 1:

Hence the integral (I1) evaluates to 1 . 1/ D 2:

Chap. 14 Complex Integration 295

27. Integration by the first method (Theorem 1, p. 647). The integrand sec2 z is analytic except at
the points where cosz is 0 [see Example 2(b), pp. 634–635 of the textbook]. Since

.tanz/0 D sec2 z [by (4), p. 634],

2
Z i=4 h ii=4
(I2) sec z dz D tanz D tani tan:
=4 =4

This can be simplified because

1 sin 1 p1 tan D 41 D 12D 1:

4 cos 4 p2

Also

1 sin 1i i sinh 1 1
tan i D 4 D 4 D i tanh ; 4 cosi cosh 4

since, by (15), p. 635 of Sec. 13.6 of textbook,

siniz D i sinhz

and

cosiz D coshz

with z D .
A numeric value to six significant digits of the desired real hyperbolic tangent is 0:655794: Hence
(I2) evaluates to

i tanh 1 D 0:655794i 1:

Remember that the real hyperbolic tangent varies between 1 and 1, as can be inferred from
the behavior of the curves of sinhx and coshx in Fig. 551 and confirmed in Fig. 552, p. A65 (in Part
A3.1 of App. 3 of the textbook).

Sec. 14.2 Cauchy’s Integral Theorem

Cauchy’s integral theorem, p. 653, is the most important theorem in the whole chapter. It states that the
integral around a simple closed path (a contour integral) is zero, provided the integrand is an analytic
function. Expressing this in a formula
I
(1) f.z/ dz D 0
C where C is a simple closed path
296 Complex Analysis Part D

H
and C lives in a complex domain D that is simply connected. The little circle on the integral sign marks
a contour integral.
Take a look at Fig. 345, p. 652, for the meaning of simple closed path and Fig. 346, p. 653, for a simply
connected domain. In its basic form, Theorem 1 (Cauchy’s integral theorem) requires that the path not
touch itself (a circle, an ellipse, a rectangle, etc., but not a figure 8) and lies inside a domain D that has
no holes (see Fig. 347, p. 653).
You have to memorize Cauchy’s integral theorem. Not only is this theorem important by itself, as a
main instrument of complex integration, it also has important implications explored further in this
section as well as in Secs. 14.3 and 14.4.
Other highlights in Sec. 14.2 are path independence (Theorem 2, p. 655), deformation of path (p. 656,
Example 6, Prob. 11), and extending Cauchy’s theorem to multiply connected domains (pp. 658–659).
We show where we can use Cauchy’s integral theorem (Examples 1 and 2, p. 653, Probs. 9 and 13) and
where we cannot use the theorem (Examples 3 and 5, pp. 653–654, Probs. 11 and 23). Often the
decision hinges on the location of the points at which the integrand f.z/ is not analytic. If the points lie
inside C (Prob. 23) then we cannot use Theorem 1 but use integration methods of Sec. 14.1. If the points
lie outside C (Prob. 13) we can use Theorem 1.

Problem Set 14.2. Page 659

3. Deformation of path. In Example 4, p. 654, the integrand is not analytic at z D 0, but it is

everywhere else. Hence we can deform the contour (the unit circle) into any contour that
contains z D 0 in its interior. The contour (the square) in Prob. 1 is of this type. Hence the answer
is yes.

9. Cauchy’s integral theorem is applicable since f.z/ D e z2 is analytic for all z, and thus entire
(see p. 630 in Sec. 13.5 of the textbook). Hence, by Cauchy’s theorem (Theorem 1, p. 653),

I 2
z
e dz D 0 with C unit circle, counterclockwise.
C

More generally, the integral is 0 around any closed path of integration.

11. Cauchy’s integral theorem (Theorem 1, p. 653) is not applicable. Deformation of path. We
see that 2z 1 D 0 at z D . Hence, at this point, the function

f.z/ D
is not analytic. Since z D lies inside the contour of integration (the unit circle), Cauchy’s theorem is
not applicable. Hence we have to integrate by the use of path. However, we can choose a most
convenient path by applying the principle of deformation of path, described on p. 656 of the
textbook. This allows us to move the given unit circle e it by . We obtain the path C given by

z.t/ D C eit where 0 t 2:

Chap. 14 Complex Integration 297

Note that t is traversed counterclockwise as t increases from 0 to 2, as required in the problem.

Then
1 1 1

f.z/ D f.z.t// D 2z.t/ 1D2 21 C eit 1 D 2eit :

Differentiation gives

z.t/P D ieit; (chain rule!).

Using the second evaluation method (Theorem 2, p. 647, of Sec. 14.1) we get

Z Zb
f.x/ dz D f Œz.t/z.t/ dtP [by (10), p. 647]
C a

Z2 1 it

D ie dt
0 2eit

Z 2 eit
Di dt
0 2eit

Z 21
Di dt
0 2
t2
Di
20
D i:

Note that the answer also follows directly from (3), p. 656, with m D 1 and z0 D .

13. Nonanalytic outside the contour. To solve the problem, we consider z4 1:1 D 0, so that
4
z D 1:1: By (15), p. 617 of Sec. 13.2,
p 2k 2k k D 0;1;2;3;
;
4
zDr cos C4 C i sin C4

where r D p4 1:1 D 1:0241 and the four roots are

298 Complex Analysis Part D

z0 DD ppp444 1:1.cos0 CCCi sin0/ p4 1:1;

p4 1:1 i;
p4 1:1;
p4 1:1 i:
z1 1:1 cos i sin
2 2
z2 D 1:1.cos i sin/ D
D
3 3
z3 D p 1:1
4
cos C i sin D
2 2 D
y

–1 1 x
1.0241

–i C

Sec. 14.2 Prob. 13. Area of integration C versus location of

roots z0, z1, z2, z3 of denominator of integrand

Since z0, z1;z2;z3 all lie on the circle with center .0;0/ and radius r D p 4 1:1 D 1:0241 > 1, they are
outside the given unit circle C. Hence f.z/ is analytic on and inside the unit circle C. Hence Cauchy’s
integral theorem applies and gives us

I I 1
f.z/ dz D dz D 0:
4
C Cz 1:1

23. Contour integration. We want to evaluate the contour integral

I
2z 1 dz where C as given in the accompanying figure on p.
659.
2
Cz z

We use partial fractions (given hint) on the integrand. We note that the denominator of the
integrand factors into z2 z D z.z 1/ so that we write

2z 1 A B
:
z2 zDzCz 1
Chap. 14 Complex Integration 299

Multiplying the expression by z and then substituting z D 0 gives the value for A:

2z 1 Bz 1
D A C ; D A C 0; A D 1 : z 1 z 1 1

Similarly, multiplying z 1 and then substituting z D 1, gives the value for B:

2z 1 A.z 1/ 1
B; 0 B; B 1:
z D z C 1D C D
Hence
2z 1 1 1
:
D C
z.z 1/ z z 1

The integrand is not analytic at z D 0 and z D 1, which clearly lie inside C. Hence Cauchy’s integral
theorem, p. 653,
second. Hence we get
does not apply.
I
2z
Instead we use
1 I 1 I 1 dz D dz C
(3), C 2 z z Cz Cz 1 dz D 2i C 2i D 4i: p. 656, with m D
1 for the two integrands obtained by partial fractions. Note that z 0 D 0, in the first integral, and
then z0 D 1 in the

Sec. 14.3 Cauchy’s Integral Formula

Cauchy’s integral theorem leads to Cauchy’s integral formula (p. 660):

I
f.z/
(1) dz D 2if.z0/:
Cz z0

Formula (1) evaluates contour integrals

I
(A) g.z/dz
C

with an integrand

f.z/
g.z/ D with f.z/ analytic: z
z0
300 Complex Analysis Part D

Hence one must first find

f.z/ D .z z0/g.z/:

For instance, in Example 1, p. 661 of the text,

ez z

g.z/ D hence f.z/ D .z 2/g.z/ D e : z 2

The next task consists of identifying where the point z 0 lies with respect to the contour C of
integration. If z0 lies inside C (and the conditions of Theorem 1 are satisfied), then (1) is applied directly
(Examples 1 and 2, p. 661). If z0 lies outside C, then we use Cauchy’s integral theorem of Sec. 14.3 (Prob.
3). We extend our discussion to several points at which g.z/ is not analytic.
Example 3, pp. 661–662, and Probs. 1 and 11 illustrate that the evaluation of (A) depends on the
location of the points at which g.z/ is not analytic, relative to the contour of the integration. The section
ends with multiply connected domains (3), p. 662 (Prob. 19).

Problem Set 14.3. Page 663

1. Contour integration by Cauchy’s integral formula (1), p. 660. The contour jz C 1j D 1 can be
written as jz . 1/j D 1. Thus, it is a circle of radius 1 with center 1. The given function to be
integrated is

z2
g.z/ D : z2 1

Our first task is to find out where g.z/ is not analytic. We consider

z2 1D0 so that z2 D 1:

Hence the points at which g.z/ is not analytic are

zD1 and z D 1:
Our next task is to find out which of these two values lies inside the contour and make sure that
neither of them lies on the contour (a case we would not yet be able to handle). The value z D 1 lies
outside the circle (contour) and z D 1 lies inside the contour. We have

z2 z2
2
g.z/ D D C :z 1 .z
1/.z 1/

Also

z2 f.z/ f.z/
g.z/ D D D : z2 1 z z0
z . 1/

Together
Chap. 14 Complex Integration 301

f.z/ z2
:
z C 1 D .z C 1/.z 1/

Multiplying both sides by z C 1 gives

z2
f.z/ D ;z
1

which we use for (1), p. 660. Hence

I z2 I f.z/
dz dz [in the form (1), p. 660]
C z2 1 D Cz z0

I
z2=.z 1/
D dz [Note z0 D 1
C z . 1/
D 2i f.z0/ D 2i f. 1/

D 2i

D i:

y
C

x
–1 1

Sec. 14.3 Prob. 1.

Contour C of integration
3. Contour integration. Cauchy’s integral theorem, p. 653. The contour C3 W j z C i j D 1:4 is a circle
of radius 1:4 and center z 0 D i. Just as in Prob. 1, we have to see whether the points z 1 D 1 and z2 D
1 lie inside the contour C 3. The distance between the points z 0 D i and z1 D 1 is, by (3) and Fig. 324,
p. 614 in Sec. 13.2, as follows.

p p
jz0 z1j D ji 1j D j 1 C ij D x2 C y2 D . 1/2 C 12 D p2 > 1:4:
302 Complex Analysis Part D

Hence z1 lies outside the circle C3.

By symmetry z2 D 1 also lies outside the contour.
Hence g.z/ D z2=.z 1/ is analytic on and inside C3: We apply Cauchy’s integral theorem and
get, by (1) on p. 653 in Sec. 14.2,

I z2 z2
3 dz D 0 by setting f.z/ D in (1) :
2
C z2 1 z 1

11. Contour integral. The contour C is an ellipse

with focal points 0 and 2i. The given

integrand is 1

g.z/ D C
2
z 4

We consider z2 C 4 D 0 so that z D ˙2i. Hence the points at which g.z/ is not analytic are z D 2i and z
D 2i.
To see whether these points lie inside the contour C we calculate for z D 2i D x C yi so that
x D 0 and y D 2 and

4x2 C .y 2/2 D 4 02 C .2 2/2 D 0 < 4;

so that z D 2i lies inside the contour. Similarly, z D 2i corresponds to x D 0;y D 2 and

4x2 C .y 2/2 D . 2 2/2 D 16 > 4;

so that z D2i lies outside the ellipse. We

have

1 f.z/ f.z/
D g.z/D D : z2 C 4 z z z 2i
0

Together
f.z/ 1 1
2
z 2i D z C 4 D .z C 2i/.z
2i/

where
1
f.z/ D C:z
2i
Cauchy’s integral formula gives us
Chap. 14 Complex Integration 303

I I
dz f.z/
dz dz [by (1), p. 660]
2
C z C4 D C z C z0

D I 1=.z C 2i/
dz
C z 2i
D 2if.z0/

D 2if.2i/

1
D 2i C
2i 2i
1
D 2i
4i

D .

13. Contour integral. We use Cauchy’s integral formula. The integral is of the form (1), p. 660, with z
z0 D z 2, hence z0 D 2. Also, f.z/ D z C 2 is analytic, so that we can use (1) and calculate

2if.2/ D 8i.

19. Annulus. We have to find the points in the annulus 1 < j z j < 3 at which

ez2 ez2
g.z/ D D C
2 2
z .z 1 i/ z Œz .1 i/

is not analytic. We see that z D 1 C i is such a point in the annulus. Another point is z D 0, but this is
not in the annulus, that is, not between the circles, but in the “hole.” Hence we calculate

ez2
f.z/ D Œz .1 C i/g.z/ D :
z2

We evaluate it at z D 1 C i and also note that

(C) z2 D .1 C i/2 D 2i:

We obtain by Cauchy’s integral formula, p. 660,

304 Complex Analysis Part D

e.1Ci/2
2if.1 C i/ D 2i
2i
D e.1Ci/2

D e2i [by (C)]

D .cos2 C i sin2/ [by Euler’s formula]:
A numeric value is

. 0:416147 C 0:909297i/ D 1:30736 C 2:85664i:

Sec. 14.4 Derivatives of Analytic Functions

The main formula is (1), p. 664. It shows the surprising fact that complex analytic functions have
derivatives of all orders. Be aware that, in the formula, the power in the denominator is one degree
higher (n C 1) than the order of differentiation (n).

Problem Set 14.4. Page 667

1. Contour integration. Use of a third derivative. Using (1), p. 664, we see that the given function is
sinz f.z/
z4 D .z z0/nC1 with f.z/ D sinzI z0 D 0 and n C 1 D 4:

Thus n D 3: By Theorem 1, p. 664, we have

I f.z/ 2i .3/
4
(A) dz D 3Š f .z0/:
C .z z0/

f f 00
Since f.z/ D sinz, 0.z/ D cosz, .z/ D sinz, so that

f .3/ D . sinz/0 D cosz:

Furthermore z0 D 0 and

/ /
f .3/.z0 D cos.0 D 1:

Hence, by (A), we get the answer that

I
sinz 2i
dz . 1/
C z4 D 3Š
2
Chap. 14 Complex Integration 305

D i

3 2 1 D i:

5. Contour integration. This is similar to Prob. 1. Here the denominator of the function to be
1 4 1 4
integrated is z 2 ; and z 2 D 0 gives z0 D which lies inside the unit circle. To use
Theorem 1, p. 664, we need the third derivative of cosh2z. We have, by the chain rule,

f.z/ D cosh2z

f 0.z/ D 2sinh2z f

00
.z/ D 4sinh2z

f .3/.z/ D 8sinh2z:
We evaluate the last equality at z0 D and get

.3/
1 1
f D 8sinh 2
2 2
D 8sinh1

D 8 1 e1 e1 [by (17), p. A65 of Sec. A3.1 in App. 3]

2
1
D4 e
e

D 9:40161:

Thus

I
cosh2z 2i
dz 9:40161
14
C z z D 3Š

D i 9:40161

D 3:13387 i D 9:84534i:

9. First derivative. We have to solve

I tanz 2 2
306 Complex Analysis Part D

dz; with C the ellipse 16x y 1 traversed clockwise.

C z2 C D

The first derivative will occur because the given function is .tanz/=z2: Now

sinz tanz D is not analytic at the points ˙ .2n C 1/=2:

cosz
But all these infinitely many points lie outside the ellipse

x2 2

y 1
.14/2 C D

whose semiaxes are and 1. In addition,

tanz
z2 is not analytic at z D z0 D 0;

where it is of the form of the integrand in (1 0), p. 664. Accordingly, we calculate

f.z/ D z2g.z/ D tanz

and the derivative (chain rule)

f 0.z/ D 2 z:
cos

Hence (1), p. 664, gives you the value of the integral in the counterclockwise direction, that is,

2i
(B) 2if 0.0/ D D 22i:
1

Since the contour is to be traversed in the clockwise direction, we obtain a minus sign in result (B)
and get the final answer 22i:

13. First derivative. Logarithm. The question asks us to evaluate

I
Ln z
dz; C z 3 2 traversed counterclockwise.
2
C .z 2/ Wj jD

We see that the given integrand is Ln.z/=.z 2/2 and the contour of integration is a circle of radius 2
with center 3: At 0 and on the ray of the real axis, the function Lnz is not analytic, and it is essential
that these points lie outside the contour. Otherwise, that is, if that ray intersected or touched the
Chap. 14 Complex Integration 307

contour, we would not be able to integrate. Fortunately, in our problem, the circle is always to the
right of these points.
In view of the fact that the integrand is not analytic at z D z 0 D 2, which lies inside the contour,
then, according to (1), p. 664, with n C 1 D 2, hence n D 1, and z 0 D 2; the integral equals 2i times
the value of the first derivative of Ln z evaluated at at z 0 D 2. We have the derivative of Ln z is

1
.Ln z/0 D
z

which, evaluated at z D z0 D 2, is : This gives a factor to the result, so that the final answer is

2i D i:
Chap. 15 Power Series, Taylor Series
We shift our studies from complex functions to power series of complex functions, marking the beginning
of another distinct approach to complex integration. It is called “residue integration” and relies on
generalized Taylor series—topics to be covered in Chap. 16. However, to properly understand these topics,
we have to start with the topics of power series and Taylor series, which are the themes of Chap. 15.
The secondapproach to complex integration based on residues owes gratitude to Weierstrass (see
footnote 5, p. 703 in the textbook), Riemann (see footnote 4, p. 625 in Sec. 13.4), and others. Weierstrass,
in particular, championed the use of power series in complex analysis and left a distinct mark on the field
through teaching it to his students (who took good lecture notes for posterity; indeed we own such a
handwritten copy) and his relatively few but important publications during his lifetime. (His collected work
is much larger as it also contains unpublished material.)
The two approaches of complex integration coexist and should not be a source of confusion. (For more
on this topic turn to p. x of the Preface of the textbook and read the first paragraph.)
We start with convergence tests for complex series, which are quite similar to those for real series.
Indeed, if you have a good understanding of real series, Sec. 15.1 may be a review and you could move on
to the next section on power series and their radius of convergence. We learn that complex power series
represent analytic functions (Sec. 15.3) and that, conversely, every analytic function can be represented by
a power series in terms of a (complex) Taylor series (Sec. 15.4). Moreover, we can generate new power
series from old power series (of analytic functions) by termwise differentiation and termwise integration.
We conclude our study with uniform convergence.
From calculus, you want to review sequences and series and their convergence tests. You should
remember analytic functions and Cauchy’s integral formula (1), p. 660 in Sec. 14.3. A knowledge of how to
calculate real Taylor series is helpful for Sec. 15.4. The material is quite hands-on in that you will construct
power series and calculate their radii of convergence.

Sec. 15.1 Sequences, Series, Convergence Tests

This is similar to sequences and series in real calculus. Before you go on—test your knowledge of real
series and answer the following questions: What is the harmonic series? Does it converge or diverge? Can
you show that your answer is correct? Close the book and work on the problem. Compare your answer
with the answer on p. 314 at the end of this chapter in this Manual. If you got a correct answer, great! If
not, then you should definitely study Sec. 15.1 in the textbook.
Most important, from a practical point of view, is the ratio test (see Theorem 7, p. 676 and Theorem 8,
p. 677).
The harmonic series is used in the proof of Theorem 8 (p. 677) and in the Caution after Theorem 3,
p. 674. One difference between calculus and complex analysis is Theorem 2, p. 674, which treats the
convergence of a complex series as the convergence of its real part and its complex part, respectively.

Problem Set 15.1. Page 679

3. Sequence. The sequence to be characterized is

Chap. 15 Power Series, Taylor Series 309

n zn D
C
:
4 2ni
First solution method:
D
n zn
C
4 2ni
n 4 2ni
[by (7), p. 610 of Sec. 13.1]
D 4 C 2ni 4 2ni

n.4 2ni/
D 42 C .2n/2

4n
n2
i
:
D 42 C 4n2 C 8 C 2n2
We have just written zn in the form

zn D xn C iyn:

By Theorem 1, p. 672, we treat each of the sequences fx ng and fyng separately when characterizing the
behavior of fzng : Thus

4n
nlim!1 xn D nlim!1 4 C 4n2
n

D lim n2
2 (divide numerator and denominator by 4n2/ n!1 1Cn2n
n

D nlim!1 n12 C 1

D lim nlim!12 Cn lim 1

1

n
!1 n n!1

0
0:
D0C1D

Furthermore,
n2
lim yn D lim C
310 Complex Analysis Part D

2 n!1 n!1 8
2n
2
n

D lim n2 2
n!1 8C2n
n2
lim
D 8 n!1 n2
C2
D
:
0C2D 2
Hence the sequence converges to

0 i:
Ci 2 D
Second solution method (as given on p. A38):
D
n zn
C
4 2ni
n

D 2ni (division of numerator and denominator by 2ni)

4C2ni2ni

1 1
. i/ i
D D
2i D 22 C i D 22 C C2 2:
2

1
1 1 1
ni C ni ni ni

Now

1 lim 1i 1
nlim!1 zn D nlim!1 1 C2 ni2 D nlim!1n!11 C n2lim!1ni2 D 1 C2i0 Di: i
Since the sequence converges it is also bounded.

5. Sequence. The terms zn D . 1/n C 10i; n D 1;2;3; , are

z1 D 1 C 10i; z2 D 1 C 10i; z3 D 1 C 10i; z4 D 1 C 10i; :

The sequence is bounded because

j zn j D j. 1/n C 10ij
D pp C n2 C 102 Œ. 1/
DD p1 100
Chap. 15 Power Series, Taylor Series 311

101 <
11:

For odd subscripts the terms are 1 C 10i and for even subscripts 1 C 10i. The sequence has two
limit points 1 C 10i and 1 C 10i, but, by definition of convergence (p. 672), it can only have
one. Hence the sequence fzng diverges.

9. Sequence. Calculate
j zn j D j0:9 C 0:1ij2n D .j0:9 C

0:1ij2/n

n
.0:81 C 0:01/
D
n
D 0:82 ! 0 as n ! 0:

Conclude that the sequence converges absolutely to 0.

13. Bounded complex sequence. To verify the claim of this problem, we first have to show that:
(i) If a complex sequence is bounded, then the two corresponding sequences of real parts and
imaginary parts are also bounded.
Proof of (i). Let fzng be an arbitrary complex sequence that is bounded. This means that there is a
constant K such that

j zn j < K for all n (i.e., all terms of the sequence).

Set

zn D xn C i yn

as on p. 672 of the text. Then

q
jznj D xn2 C yn2 [by (3), p. 613 of Sec. 13.2]

and
jznj2 D xn2 C yn2:

Now
xn2 xn2 C yn2 D jznj2 since xn2 0; yn2 0:

Furthermore,
xn2 D jxnj2 since xn2 0:

Thus
jxnj2 jznj2
312 Complex Analysis Part D

jxnj jznj
so that
jxnj < K:

Similarly,
jynj2 yn2 jznj2 < K2

so that
jynj < K:

Since n was arbitrary, we have shown that fx ng and fyng are bounded by some constant K: Next we
have to show that:
(ii) If the two sequences of real parts and imaginary parts are bounded, then the complex sequence is also
bounded.
Proof of (ii). Let fxng and fyng be bounded sequences of the real parts and imaginary parts, respectively.
This means that there is a constant L such that
L L j xn j < p ; jyn j < p
:
2 2
Then
L2 L2
j 2 2

j x n< ; j yn j < ;
2 2
so that
j zn j2 D xn2 C yn2

L2 L2
< C2 2
< L2:

Hence fzng is bounded.

19. Series convergent? Comparison test.

ˇˇ in ˇˇ

jznj D ˇˇn 2 i ˇˇ

D jinj j

[by (10), p. 615 in Sec. 13.2]

jn2 i
n
j
i
Chap. 15 Power Series, Taylor Series 313

D jn2j ij

1 D p n4 C
1
[by (3), p. 613 in Sec. 13.2]

1 1
<pD : n 4 n2

Since

X1 1
converges [see p. 677 in the Proof of (c) of Theorem 8],
2
nD1 n

we conclude, by the comparison test, p. 675, that the series given in this problem also converges.

23. Series convergent? Ratio test. We apply Theorem 8, p. 677. First we form the ratio z nC1=zn and
simplify algebraically. Since
. 1/n .1 i/2nC1 zn D
C ;
.2n/Š
the test ratio is
1/nC1.1 C i/2.nC1/C1 =.2.n C 1//Š
znC1 D . zn
. 1/n.1 C i/2nC1 =.2n/Š
. 1/nC1.1 C i/2nC3 .2n/ŠC
D .2.n C 1//Š n
. 1/ .1 i/2nC1
.1 C i/2 .2n/Š

D . 1/
.2n C 2/Š 1

D. .1 C i/2
1/
.2n C 2/.2n C 1/
.2i/
D . 1/ : .2n
C 2/.2n C 1/
Then we take the absolute value of the ratio and simplify by (3), p. 613, of Sec. 13.2:

ˇˇznC1ˇˇˇ D ˇˇˇ. 1/ C 2i C ˇˇˇ

ˇ zn ˇ ˇ .2n 2/.2n 1/ˇ

1
D CC :
.n 1/.2n 1/
314 Complex Analysis Part D

Hence

lim ˇˇznC1ˇˇˇ D C 1 C D L D 0:
ˇ
n!1 ˇ zn ˇ .n 1/.2n 1/
because

1 1 1
lim C C D lim C lim C D 0 0 D 0 n!1 .n 1/.2n 1/ n!1 n 2
n!1 2n 1

Thus, by the ratio test (Theorem 8), the series converges absolutely and hence converges.

Sec. 15.2 Power Series

Since analytic functions can be represented by infinite power series (1), p. 680,

(1) a0 C a1.z z0/ C a2.z z0/2 C ;

such series are very important to complex analysis, much more so than in calculus. Here z 0;called the
center of the series, can take on any complex number (once chosen, it is fixed). When z 0 D 0, then we get
(2), p. 680. An example is

z z2
z
(E) eD1C C C :
1Š 2Š
More on this in Sec. 15.4. We want to know where (1) converges and use the Cauchy–Hadamard formula
(6), p. 683, in Theorem 2 to determine the radius of convergence R, that is,

ˇ
ˇ an ˇ ˇ

(6) R D lim ˇ ˇ [remember that the .n C 1/st term is in the denominator!]. n!1 ˇanC1ˇ

Formula (6) shows that the radius of convergence is the limit of the quotient D j anC1=an j in the ratio test
(Theorem 8, p. 677). This isj an=anC1 j (if it exists). This in turn is the reciprocal of the quotient L
understandable; if the limit of L is small, then its reciprocal, the radius of convergence R, will be large. The
following table characterizes (6).

Table. Area of convergence of power series (1)

Value of R Area of convergence of series (1) Illustrative examples

R D c (c a constant: real, positive) Convergence in disk jz z0j <c Ex. 5, p. 683, Prob. 13

R Convergence everywhere Ex. 2, p. 680, series (E), Prob. 7

D1

RD0 Convergence only at the center z D z0 Ex. 3, p. 681

Chap. 15 Power Series, Taylor Series 315

Remarks: R D 1 means the function is entire. R D 0 is the useless case.

Problem Set 15.2. Page 684

7. Radius of convergence. The given series

1 1 2n
X . 1/n 2
z
.2n/Š
nD0

is in powers of z , and its center is : We use the Cauchy–Hadamard formula (6), p. 683, to
determine the radius of convergence R: We have

an D . 1/n .2.n C 1//Š D . 1/n .2.n C 1//Š

:
anC1 .2n/Š . 1/nC1 . 1/nC1 .2n/Š
We simplify the two fractions in the last equality:
. 1/n . 1/n
1
nC1 n
. 1/ D . 1/ . 1/ D

and

.2.n C 1//Š D .2n C 2/Š D .2n C 2/.2n C 1/2n 1 D C C

.2n 2/.2n 1/:

.2n/Š .2n/Š 2n 1
Together, the desired ratio simplifies to
an
.2n 2/.2n 1/;
anC1 D C C

and its absolute value is

ˇˇ an ˇˇ

ˇ ˇ D .2n C 2/.2n C 1/:

ˇanC1ˇ

Now as n ! 1

ˇˇ an ˇˇ

ˇ ˇ D .2n C 2/.2n C 1/ ! 1:
ˇanC1ˇ

Hence
R D 1:
316 Complex Analysis Part D

This means that the series converges everywhere, see Example 2, p. 680, and the top of p. 683, of
the textbook.
We were fortunate that the radius of convergence was 1 because our series is of the form

X1 2n
anz :
nD0

p
Had R been finite, the radius of convergence would have been R (see the next problem).
Remark. Plausibility of result. From regular calculus you may recognize that the real series
X1 . 1/n 1 2n 1
x D cos x 2
.2n/Š nD0 2
is the Taylor series for cos x 1
2. The complex analog is cos z
1
2: Since the complex cosine
function is an entire function, its has an infinite radius of convergence.

13. Radius of convergence. The given series is

X1 n 4n
16 .z C i/ :
nD0

Since z C i D z . i/, the center of the series is i. We can write the series as

X1 16n .z C i/4n D X1 16n .z C i/4n D X1 16ntn

nD0 nD0 nD0

where

(A) t D .z C i/4:
We use the Cauchy–Hadamard formula (6), p. 683, to determine the radius of convergence R t [where
the subscript t refers to the substitution (A)]:

an 16n 16n 1
:
anC1 D 16nC1 D 16n 16 D 16

Hence by (6), p. 683,

11
Rt D lim D : n!1 16 16

This is the radius of convergence of the given series, regarded as a function of t: From (A) we have

z C i D t1=4:

Hence the radius of convergence Rz, for the given series in z, is

Chap. 15 Power Series, Taylor Series 317

Rz D .Rt/1=4 D 161 1=4 D q4 161 D :

We denote Rz by R to signify that it is the wanted radius of convergence for the given series. Hence
the series converges in the open disk

jz . i/j < with center i and radius R D :

15. Radius of convergence. Since the given series

X1 .2n/Š n
2
.z 2i/
4n .nŠ/ nD0

is in powers of z 2i, its center is 2i. We use (6), p. 683, to determine R

an D .2n/Š 4nC1..nCC 1/Š/2 anC1 4n .nŠ/2 .2.n

1//Š

which groups, conveniently, to

D .2n/Š 4nC1 ..n C 1/Š/2

:
.2.n C 1//Š 4n .nŠ/2

To avoid calculation errors, we simplify each fraction separately, that is,

.2Cn/Š D C2n.2n C1/ 1 D C 1 C

;
.2.n 1//Š .2n 2/.2n 1/2n 1 .2n 2/.2n 1/
4nC1
D 4;
4n
and

..n C 1/Š/2 .n C 1/n 12 D C 2

2
D .n 1/ :
.nŠ/ n 1
Hence, putting the fractions together and further simplification gives us

an 4.n C 1/2 D 4.nCC 1/.n CC1/ D 2.n CC 1/ D 2n CC 2

D C C ;
anC1 .2n 2/.2n 1/ 2.n 1/.2n 1/ 2n 1 2n 1
318 Complex Analysis Part D

so that the final result is

nli m!1 ˇˇˇˇaanCn1ˇˇˇˇ D 2n CC 2 D 2nnC2 D

lim 2 CC n21 lim lim

n!1 2n 1 n!1 2nnC1 n!1 2 n

D nnlimlim!1!1 22 CC nn21 D 22 CC 00 D 22 D 1 D R

Thus the series converges in the open disk jz 2i j < 1 of radius R D 1 and center 2i:

Sec. 15.3 Functions Given by Power Series

We now give some theoretical foundations for power series and show how we can develop a new power
series from an existing one. This can be done in two ways. We can differentiate a power series term by
term without changing the radius of convergence (Theorem 3, p. 687, Example 1, p. 688, Prob. 5).
Similarly, we can integrate (Theorem 4, p. 688, Prob. 9). Most importantly, Theorem 5, p. 688, gives the
reason why power series are of central importance in complex analysis since power series are analytic and
so are “differentiated” power series (with the radius of convergence preserved).

Problem Set 15.3. Page 689

5. Radius of convergence by differentiation: Theorem 3, p. 687. We start with the geometric series
1
Xz 2i n z 2i z 2i 2 z 2i 3
(A) g.z/ D D1C C C C :
2 2 2 2
nD0

Using Example 1, p. 680, of Sec. 15.2, we know that it converges for j z

2i j j j

<1 and thus for z 2i < 2:

2
Theorem 3, p. 687, allows us to differentiate the series in (A), termwise, with the radius of
convergence preserved. Hence we get

1 z 2i 1 z 2i 2 1
(B) g0.z/ D 0 C C2 C3 C
2 2 2 2 2

X1 n.z 2i/n 1
D where jz 2i j < 2:
2n nD1
Chap. 15 Power Series, Taylor Series 319

Note that we sum from n D 1 because the term for n D 0 is 0. Applying

Theorem 3 to (B) yields

X1 n.n 1/.z 2i/n 2

(C) g00.z/ D n where j z 2i j < 2:
2
nD2

From (C) it follows that

1 2i/n
(D) .z 2i/2g00.z/ D X n.n
1/.z2n
nD2
1 2i n
X z
D n.n 1/
2 where jz 2i j < 2:
nD2 But

(D) is precisely the given series.

Complete the problem by verifying the result by the Cauchy–Hadamard formula (6), p. 683, in
Sec. 15.2.

9. Radius of convergence by integration: Theorem 4, p. 688. We start with the geometric series (see
Example 1, p. 680) which has radius of convergence 1:
1X n 2 3

wD1CwCwCw j w j < 1:
nD0

Hence,

n 2 3
X1
. 2w/ D 1 2w C 4w 8w j w j <;
nD0

and then,

n 2 3
X1
. 2w/ D 2w C 4w 8w j w j <:
nD1

We substitute w D z2 into the last series and get

X1 n 2n 2 4 6
(E) . 2/ z D 2z C 4z 8z C
nD1
320 Complex Analysis Part D

which converges for

j 21
zj< and hence jzj<p :
2
Our aim is to produce the series given in the problem. We observe that the desired series has
factors n C 2, n C 1, and n in the denominator of its coefficients. This suggests that we should use
three integrations to determine the radius of convergence. We use Theorem 4, p. 688, to justify
termwise integration. We divide (E) by z

3 5 X n 2n 1
1

2z C 4z 8z C D . 2/ z :
nD1

We integrate termwise (omitting the constants of integration)

Z z2 Z 3 z4 Z 5 z6
2 z dz D 2 ; 4 z dz D 4 ; 8 z dz D 8 ; ,
2 4 6
which is

X1 z
n 2n 1
. 2/ where j z j < p : nD1 2n 2

However, we want to get the factor 1=n so we multiply the result by 2, that is, X

n z2n X1 n z2n

(F) 2 . 2/ D . 2/ :
2n n nD1 nD1

Next we aim for the factor 1=.n C 1/: We multiply the series obtained in (F) by z

X1 z
n 2nC1

. 2/ ;
n nD1

and integrate termwise 1

Z D. 2/n Z 2nC1 dz
X
2/n z2nC1
. dz z nD1

n n
. 2/n z2nC1C1

and get the series D C

C ; n 2n 1 1
Chap. 15 Power Series, Taylor Series 321

We multiply the result by 2 (to obtain precisely the factor 1=n) and get (G) We get
1 2/n z2nCC2 D X1 .2/n z2nC2
X : X
(G) 2 . 1
nD1 n.n C 1/
2n.n 1/ n
nD1
z2nC4
We multiply the right-hand side of (G) by z: . 2/
1
X n z2nC3
. 2/
:
n.n 1/
n.n
nD1 C
1/2.n
and integrate 2/
nD1
C
Z . 2/ z2nCC3 dz D . 2C/n Z C

z2nC3 dz D . 2C/n n z2nC3C1

:
n.n 1/ n.n 1/ n.n 1/
2n C 3 C 1
We have an unwanted factor 2 in the denominator but only wanted .n C 2/, so we multiply by 2 and
get
1

X n z2nC4
. 2/ :
n.n 1/.n 2/
nD1 C C

However, our desired series is in powers of z 2n instead of z2nC4: Thus we must divide by z4 and get

1 X1 n z2nC4 X1 n z2n
(H) . 2/ CC D . 2/ C C
:
z4 n.n 1/.n 2/ n.n 1/.n 2/ nD1 nD1

But this is precisely the desired series. Since our derivation from (E) to (H) did not change the radius
of convergence (Theorem 4), we conclude that the series given in this problem
has radius of
convergence jz j < 1=p2, that is, center 0 and radius 1=p2:
Do part (a) of the problem, that is, obtain the answer by (6), p. 683.
10, 13, 14, 18 Hint. For problems 10, 13, 14, and 18, the notation for the coefficients is
explained on pp. 1026–1028 of Sec. 24.4, of the textbook.
322 Complex Analysis Part D

17. Odd
implies
a2m. z/2m D a2m. 1/2mz2m D a2m . 1/2m z2m D a2m1mz2m D a2mz2m D a2mz2m
functions. The even-numbered coefficients in (2), p. 685, are zero because f. z/ D f.z/

But

a2mz2m D a2mz2m

means

a2m D a2m

so that

a2m C a2m D 0 hence a2m D 0;

Complete the problem by thinking of examples.

Sec. 15.4 Taylor and Maclaurin Series

Every analytic function f.z/ can be represented by a Taylor series (Theorem 1, p. 691) and a general way of
doing so is given by (1) and (2), p. 690. It would be useful if you knew some Taylor series, such as for e z
[see (12), p. 694], sinz, and cosz [(14), p. 695]. Also important is the geometric series (11) in Example 1
and Prob. 19. The section ends with practical methods to develop power series by substitution,
integration, geometric series, and binomial series with partial fractions (pp. 695–696, Examples 5–8, Prob.
3).
Example 2, p. 694, shows the Maclaurin series of the exponential function. Using it for defining e z would
have forced us to introduce series rather early. We tried this out several times with student groups of
different interests, but found the approach chosen in our book didactically superior.

Problem Set 15.4. Page 697

3. Maclaurin series. Sine function. To obtain the Maclaurin series for sin2z2 we start with (14),
p. 695, writing t instead of z
1

X n t2nC1 t3 t5
sint D . 1/ CDt C C :
.2n 1/Š 3Š 5Š
nD0

Then we set t D 2z2 and have

1 .2z2/2nC1
X 1/n
sin2z D . 2

nD0
.2n C 1/Š
1X 22nC1z4nC2
D . 1/n
nD0 .2n C 1/Š
2 23z6 25z10
Chap. 15 Power Series, Taylor Series 323

2z
D C C
3Š 5Š
4 4
2 6 10
D
2z zC z C :
3 15
The center of the series thus obtained is z 0 D 0 (i.e., z D z z0 D z 0) by definition of Maclaurin
series on p. 690. The radius of convergence is R D 1, since the series converges for all z:

15. Higher transcendental functions. Fresnel integral. It is defined by

Zz S.z/sint2 dt:
D 0
To find the Maclaurin series of S.z/ we start with the Maclaurin series for sinw; and set w
D t2. From Prob. 3 of this section we know that

2 X
1 n t4nC2 2 16 1 10
sint D . 1/ CDt tC t C :
.2n 1/Š 3Š 5Š
nD0

Theorem 4, p. 688, allows us to perform termwise integration of power series. Hence

Zz 2 Z z1 n t4nC2

X
sint dt D . 1/ C dt 0 0 nD0 .2n 1/Š
1X
ˇz t4nC3
D .
nD0
1/n
ˇˇ
.2n C 1/Š.4n C 3/ˇ
ˇtD0
1X z4nC3
D .1/n ; .2n
nD0 C 1/Š.4n C 3/
which we obtained by setting t D z as required by the upper limit of integration. The lower limit t D
0 contributed 0. Hence

X1 n z4nC3 1 3 1 7 1 11 S.z/ D . 1/
CCDzzCzC :
.2n 1/Š.4n 3/ 1Š3 3Š7 5Š11
nD0

Since the radius of convergence for the Maclaurin series of the sine function is R D 1, so is R for S.z/:

19. Geometric series.

324 Complex Analysis Part D

First solution: We want to find the Taylor series of 1=.1 z/ with center z0 D i. We know that we are
dealing with the geometric series

1 X1 n
D z [by (11), p. 694],
1 z
nD0

with z0 D 0.
Thus consider
11 1DCD:
1 z 1 z i i .1 i/ .z i/

Next we work on the 1 i in the denominator by removing it as a common factor. We get

1 1 1 1
:
.1 i/ .z i/ D .1 i/1 z1 ii D1 i 1 z1 ii

This looks attractive because

1 1 z i
is of the form with w
1 z1 i i 1 w D1 i

and

1 X1 n
Dw:
1 w
nD0
Thus we have the desired Taylor series, which is

1 X1 z in 1 X1 1 n
(S) D .z i/ :
1 i 1 i 1 i .1 i/n nD0 nD0

This can be further simplified by noting that

1 D1Ci
[by (7), p. 610]
1 i 2
so that (S) becomes

1 C i X1 1 C i n n

.z i/ :
2 2

nD0
Chap. 15 Power Series, Taylor Series 325

This is precisely the answer on p. A39 of the textbook with the terms written out.
The radius of convergence of the series is

ˇz i ˇˇ

ˇ jwj < 1,
that is, ˇ ˇ < 1:
ˇ1 i ˇ Now

ˇˇz i ˇˇ D jz ij D pjz Cij

ˇ ˇ :

ˇ1 iˇ j1 ij 1 1

Hence

jzp ij j j p
<1 and z i< 2
2

so that the radius of convergence is R D p2:

Second solution: Use Example 7, p. 696 with c D 1 and z0 D i.

Remark. The method of applying (1), p. 690, directly is a less attractive way as it involves
differentiating functions of the form 1=.1 i/n.

21. Taylor series. Sine function. For this problem, we develop the Taylor series directly with (1),
p. 690. This is like the method used in regular calculus. We have for f.z/ D sinz and z0 D =2:

f.z/ D sinz f.z0/ D sinD 1I

2

f 0.z/ D cosz f 0.z0/ D cos D 0I

2

f 00.z/ D sinz f 00.z0/ D sinD 1I

2

f 000.z/ D cosz f 000.z0/ D cosD 0I

2
.4/
.z/ D sinz f .4/.z0/ D sin D 1I
f
2
.5/
.z/ D cosz f .5/.z0/ D cos D 0I
f
326 Complex Analysis Part D

2
.6/
.z/ D sinz f .6/.z0/ D sin D 1:
f

2
Hence the Taylor series for sinz with z0 D =2 W

2 4 6
1 1 1
f.z/ D 1 z C zz C
2Š 2 4Š 2 6Š 2

X1 n 1 2n
D . 1/ z ;
.2n/Š 2
nD0

The radius of convergence is R D 1:

Sec. 15.5 Uniform Convergence. Optional

The material in this section is for general information about uniform convergence (defined on p. 698) of
arbitrary series with variable terms (functions of z). What you should know is the content of Theorem 1, p.
699. Example 4 and Prob. 13 illustrate the Weierstrass M-test, p. 703.

Problem Set 15.5. Page 704

3. Power series. By Theorem 1, p. 699, a power series in powers of z z0 converges uniformly in the
closed disk j z z0j 5 r, where r < R and R is the radius of convergence of the series. Hence, solving
Probs. 2–9 amounts to determining the radius of convergence. In Prob. 3 we have a power series in
powers of

(A) Z D .z C i/2

of the form

X1 n
(B) anZ
nD0

with coefficients an D 31n: Hence the Cauchy–Hadamard formula (6), p. 683 in Sec. 15.2, gives the

radius of convergence R of this series in Z in the form

an 3n
3;
anC1 D 3 .nC1/ D
Chap. 15 Power Series, Taylor Series 327

so the series (B) converges uniformly in every closed disk jZj 5 r < R D 3. Substituting (A) and taking
square roots, we see that this means uniform convergence of the given power series in powers of z C
i in every closed disk:

(C) jz C ij 5 r < R D p3:

We can also write this differently by setting

(D) ıDR r:

We know that
R > r:

Subtracting r on both sides of the inequality gives

R r>r r
and by (D) and simplifying

ıDR r>r rD0 thus ı > 0:

Furthermore, (D) also gives us

rDR ı:

Together,

jz C ij 5 R ı D p3 ı (ı > 0).

This is the form in which the answer is given on p. A39 in App. 2 of the textbook.

7. Power series. No uniform convergence. We have to calculate the radius of convergence for

X1 nŠ 1 n n2 C
2 zi :
nD1

We want to use the Cauchy–Hadamard formula (6), p. 683 of Sec. 15.2. We start with

an D nŠ .n CC1/2

;
2
anC1 n .n 1/Š

which is written out D n 1 .n CC 1/.n C 1/ n 2 .n 1/.n 1/

and, with cancellations, becomes

DnC1
: n2
328 Complex Analysis Part D

Thus

ˇ ˇ
an ˇ nC1D 1C1 D 1C 1D
ˇ

nlim>1 ˇˇanC1ˇˇ D nlim>1 n2 nlim>1 n n2 „nlim>ƒ‚1 n…n„lim>ƒ‚1 n…2 0:

0 0

Hence R D 0, which means that the given series converges only at the center:

z0 D i:

Hence it does not converge uniformly anywhere. Indeed, the result is not surprising since

nŠ >> n2;

and thus the coefficients of the series

2 6 24 120 720 nŠ
1; ; ; ; ; ; !1 as n ! 1:
4 9 16 25 36 n

13. Uniform convergence. Weierstrass M-test. We want to show that

X1 sinn jzj
n2
nD1

converges uniformly for all z:

p
Since jzj D r D x2 C y2 is a real number, sin jzj is a real number such that

1 sin jzj 1 and 1 sinm jzj 1,

where m is a natural number. Hence j sin jzj j 1 and j sinm jzj j 1:

Now

ˇˇˇˇsinm2jzjˇˇˇˇ D jsinjmm2jjzjj D j sinmm2jzj j m12 for any z: m

Since

X1 1
converges (see Sec. 15.1 in the proof of Theorem 8, p. 677),
m2
mD1

we know, by the Weierstrass M-test, p. 703, that the given series converges uniformly.
Chap. 15 Power Series, Taylor Series 329

Solution for the Harmonic Series Problem (see p. 298 of the Student Solutions Manual) The harmonic
series is

1
1 1 1 X 1
(HS) 1C2C3C4C D m :
mD1

The harmonic series diverges. One elementary way to show this is to consider particular partial sums of
the series.

s1 D 1 s2 D 1 C

s4 D 1 C C C D 1 C C C
„ ƒ‚ … 2 terms

>1C C C 14 D 1 C C D 1 C 2

s8 1 1
D C C C C C C C D C C C 41 C C C C
„ ƒ‚ … „ ƒ‚ … 2 terms 4
terms

>1C C C 14 C C C C 18 D 1 C C C D1C3 ;
„ ƒ‚ … „ ƒ‚ … „ ƒ‚ …
2 terms 4 terms 3 fractions of value

s16 1
D C C C C C C C C C C C C C C C
1
D C C C 14 C C C C 81 C C C C C C C C
„ ƒ‚ … „ ƒ‚ … „ ƒ‚ … 2 terms 4 terms 8 terms > 1C

C C 14 C C C C 18 C C C C C C C C
„ ƒ‚ … „ ƒ‚ … „ ƒ‚ … 2 terms 4 terms 8 terms

1 1 4 ;
D C C C C D C
„ ƒ‚…
4 fractions of value

s32 D 1 C C C D1C C C 14 C C C 81 C C C 161 C C C

„ ƒ‚ … „ ƒ‚ … „ ƒ‚ … „ ƒ‚ …
31 terms 4 terms 8 terms 16 terms

>1C C C 41 C C C 81 C C C 161 C C C
„ ƒ‚ … „ ƒ‚ … „ ƒ‚ … „ ƒ‚ … 2 terms 4 terms 8 terms 16 terms

D1 C C C 4 C8 C 16 D1C C C C C D1C5 :
„ ƒ‚…
330 Complex Analysis Part D

5 fractions of value

Thus in general
s n> 1 C n :
2
As n ! 1, then

1Cn!1 and hence s n ! 1:

2

This shows that the sequence of partial sums s n is unbounded, and hence the sequence of all partial
2
sums of the series is unbounded. Hence, the harmonic series diverges.
Another way to show that the hamonic series diverges is by the integral test from calculus
(which we can use since f.x/ is continuous, positive, and decreasing on the real interval Œ1; 1 )

Z 11 Z t1 t

(A) 1 x dx D tlim!1 1x dx D tlim!1 ŒlnxxD1 D tlim!1 lnt „ƒ‚…ln01 D tlim!1 lnt ! 1:

Since the integral in (A) does not exist (diverges), the related harmonic series (HS) [whose nth term equals
f.n/] diverges.

Remark. The name harmonic comes from overtones in music (harmony!). The harmonic series is so
important because, although its terms go to zero as m ! 1, it still diverges. Go back to p. 298.
Chap. 16 Laurent Series. Residue Integration
In Chap. 16, we solve complex integrals over simple closed paths C where the integrand f.z/ is analytic
except at a point z0 (or at several such points) inside C: In this scenario we cannot use Cauchy’s integral
theorem (1), p. 653, but need to continue our study of complex series, which we began in Chap. 15. We
generalize Taylor series to Laurent series which allow such singularities at z 0. Laurent series have both
positive and negative integer powers and have no significant counterpart in calculus. Their study provides
the background theory (Sec. 16.2) needed for these complex integrals with singularities. We shall use
residue integration, in Sec. 16.3, to solve them. Perhaps most amazing is that we can use residue
integration to even solve certain types of real definite integrals (Sec. 16.4) that would be difficult to solve
with regular calculus. This completes our study of the second approach to complex integration based on
residues that we began in Chap. 15.
Before you study this chapter you should know analytic functions (p. 625, in Sec. 13.4), Cauchy’s integral
theorem (p. 653, in Sec. 14.2), power series (Sec. 15.2, pp. 680–685), and Taylor series (1), p. 690. From
calculus, you should know how to integrate functions in the complex several times as well as know how to
factor quadratic polynomials and check whether their roots lie inside a circle or other simple closed paths.

Sec. 16.1 Laurent Series

Laurent series generalize Taylor series by allowing the development of a function f.z/ in powers of z z 0
when f.z/ is singular at z0 (for “singular,” see p. 693 of Sec. 15.4 in the textbook). A Laurent series (1), p.
709, consists of positive as well as negative integer powers of z z 0 and a constant. The Laurent series
converges in an annulus, a circular ring with center z 0 as shown in Fig. 370, p. 709 of the textbook.
The details are given in the important Theorem 1, p. 709, and expressed by (1) and (2), which can be
written in shortened form (10) and (20), p. 710.
Take a look at Example 4, p. 713, and Example 5, pp. 713–714. A function may have different Laurent
series in different annuli with the same center z 0. Of these series, the most important Laurent series is the
one that converges directly near the center z 0, at which the given function has a singularity. Its negative
powers form the so-called principal part of the singularity of f.z/ at z 0 (Example 4 with z0 D 0 and Probs. 1
and 8).

Problem Set 16.1. Page 714

Hint. To obtain the Laurent series for probs. 1–8 use either a familiar Maclaurin series of Chap. 15 or a
series in powers of 1=z:

1. Laurent series near a singularity at 0. To solve this problem we start with the Maclaurin series for
cosz, that is,
z2 z4 z6 [by (14), p. 695].
(A) cosz D 1 C
2Š 4Š 6Š C

Next, since we want

cosz
;
z4
we divide (A) by z4, that is,
cosz 1 z4 D z2 z4 z6
z4 2Š C 4Š 6Š C
332 Complex Analysis Part D

1 1 1 z2
D z4 :
24 720 C
The principal part consists of
1 1
:
z4 2z2

Furthermore, the series converges for all z ¤ 0. 7. Laurent series

near a singularity at 0. We start with

w2 w4 coshw D 1 C C C [by (15),

p. 695].
2Š 4Š
We set w D 1=z: Then

1 11 11
cosh D 1 C C C . z 2Š z2 4Š z4

Multiplication by z3 yields

1 11 11
3 3 3 3
z cosh DzC zC z C z 2Š z2 4Š z4
3 z z
11111D C CC
2Š 4Š z 6Š z2
3 1 1 1 2
DzC zC z C z C :
2 24
We see that the principal part is

z 1C z 2C :

Furthermore, the series converges for all z ¤ 0; or equivalently the region of convergence is
0 < jzj < 1:

15. Laurent series. Singularity at z0 D. We use (6), p. A64, of Sec. A3.1 in App. 3 of the textbook and
simplify by noting that cos D 1 and sin D 0:
cosz D cos..z /C/
 Chap. 16 Laurent Series. Residue Integration 333

(B) D cos.z /cos C sin.z /sin D cos.z /:

Now
w2 w4 w6
cosw D 1 CC :
2Š 4Š 6Š
We set

wDz
and get
cos.z .z /2 .z /4 .z /6 :
/D1 C C
2Š 4Š 6Š
Then
.z /2 .z /4 .z /6
cos.z /D 1C C .
2Š 4Š 6Š C
We multiply by

.z /2
and get
cos.z / 1 1 .z /2 .z /4
/2 D C .
2 C
.z .z / C 2Š 4Š 6Š
Hence by (B)
cosz 2 1 1 2 1 4

D .z / C .z /C .z /C .
2
.z / 2 24 720

The principal part is

all z ¤ ). .z
2
/ and the radius of
19. Taylor and Laurent Series. The geometric series is
convergence is 0 < jz j<1
1 X1 n D w [by (11), p.
jwj<1 (converges for
694].
1 w nD0

We need

so we set w D z2:
334 Complex Analysis Part D

Then we get the Taylor series

1 X1 2n ˇ 2ˇ
2
D .z / ˇz ˇ < 1
1 z
nD0

X 2n ˇ 2ˇ 2

D z or ˇz ˇ D jzj < 1 so that j z j < 1

nD0

D 1 C z2 C z4 C z6 C :

Similarly, we obtain the Laurent series converging for j z j > 1 by the following trick, which you should
remember:
1 1 1

1 z2 1 z

1 X1 12n
D z2 z
nD0

z
24 C 6C
D

D z2 z4 z6 z8
1

X 1
D z2nC2 j
j z > 1:
nD0

23. Taylor and Laurent series. We want all Taylor and Laurent series for
z8

1
with z0 D 0:
We start with

1 X1 n
D w jwj<1 [by (11), p. 694].
1 w
Chap. 16 Laurent Series. Residue Integration 335

nD0

We set w D z4 and get

1 X1 4n D z j z j < 1:
4
1 z
nD0

We multiply this by z8 to obtain the desired Taylor series:

z8 4 D z8 X1 z4n D X1 z4nC8 jzj<1

1 z nD0 nD0

D z8 C z12 C z16 C :

From Prob. 19 we know that the Laurent series for

1 X1 1
w > 1:
1 w2 D nD0 w2nC2 j j
2
We set w D z
1 1 so that j z j > 1:
X 1 X 1 ˇ 2 ˇ ˇz ˇ
>1
D .z 2/2nC2 D
nD0 z4nC4
nD0

Multiply the result by z8:

1 1
z8 8 X 1 X z8
z D :
1
z
nD0 4nC4 z
nD0 4nC4
Now

4zn8C4 D 8 .4nC4/ D z4 4n:

z

z
Hence the desired Laurent series for
z8
4
with center z0 D 0
1 z
is
z8

1 z4
 336 Complex Analysis Part D

so that the principal part is

and j z j > 1:
1 X 4 4n z4 z4
D z D 1 z8
nD0
z4 z8
Note that we could have developed the Laurent series without using the result by Prob. 19 (but in
the same vein as Prob. 19) by starting with

1 1
;etc.
1 z4 D z4 1 1
z4

Sec. 16.2 Singularities and Zeros. Infinity

Major points of this section are as follows. We have to distinguish between the concepts of singularity and
pole. A function f.z/ has a singularity at z0 if f.z/ is not analytic at z D z0, but every neighborhood of z D z0
contains points at which f.z/ is analytic.
Furthermore, if there is at least one such neighborhood that does not contain any other singularity,
then z D z0 is called an isolated singularity. For isolated singularities we can develop a Laurent series that
converges in the immediate neighborhood of z D z 0. We look at the principal part of that series. If it is of
the form
b1
(with b1 ¤ 0), z
z0

then the isolated singularity at z D z 0 is a simple pole (Example 1, pp. 715–716). However, if the principal
part is of the form

b1 b2 bm
;
2 m
z z0 C .z z0/ C .z z0/

then we have a pole of order m. It can also happen that the principal part has infinitely many terms; then
f.z/ has an isolated essential singularity at z D z 0 (see Example 1, pp. 715–716, Prob. 17).
A third concept is that of a zero, which follows our intuition. A function f.z/ has a zero at z D z 0 if f.z0/ D 0:
Just as poles have orders so do zeros. If f.z0/ D 0 but the derivative f 0.z/ ¤ 0, then the zero is a simple
zero (i.e., a first-order zero). If f.z0/ D 0; f 0.z0/ D 0, but f 00.z/ ¤ 0, then we have a second-order zero, and
so on (see Prob. 3 for fourth-order zero). This relates to Taylor series because, when developing Taylor
series, we calculate f.z0/; f 0.z0/; f 00.z0/; :f .n/.z0/ by (4), p. 691, in Sec. 15.4. In the case of a second-order
Chap. 16 Laurent Series. Residue Integration 337

zero, the first two coefficients of the Taylor series are zero. Thus zeros can be classified by Taylor series
as shown by (3), p. 717.
Make sure that you understand the material of this section, in particular the concepts of pole and order
of pole as you will need these for residue integration. Theorem 4, p. 717, relates poles and zeros and will
be frequently used in Sec. 16.3. Problem Set 16.2. Page 719

3. Zeros. We claim that f.z/ D .z C81i/4 has a fourth-order zero at z D 81i. We show this directly:

f.z/ D .z C 81i/4 D 0 gives z D z0 D 81i:

To determine the order of that zero we differentiate until f .n/.z0/ ¤ 0: We have

f.z/ D .z C 81i/4;3 f. f 81i/ D f.z0/ D 0I
0
. f 00.
f 0.z/ D 4.z C 81i/ ;2 81i/ D 0I
f 000. f
iv. 81i/ D 0I
f 00.z/ D 12.z C 81i/ ;
81i/ D 0I 81i/ ¤
f iv000.z/ D 24.z C 81i/; 0:

f .z/ D 24;
Hence, by definition of order of a zero, p. 717, we conclude that the order at z 0 is 4. Note that we
demonstrated a special case of the theorem that states that if g has a zero of first order (simple zero)
at z0, then gn (n a positive integer) has a zero of nth order at z 0.
5. Zeros. Cancellation. The point of this, and similar problems, is that we have to be cautious. In the
present case, z D 0 is not a zero of the given function because

z 2 sin2 z D z 2 .z/2 C D 2C :

11. Zeros. Show that the assumption, in terms of a formula, is

(A) f.z/ D .z z0/ng.z/ with g.z0/ ¤ 0;

so that
f.z0/ D 0; f 0.z0/ D 0; ; f .n 1/.z0/ D 0;

as it should be for an nth-order zero.

Show that (A) implies
h.z/ D f 2.z/ D .z z0/2ng2.z/;

so that, by successive product differentiation, the derivatives of h.z/ will be zero at z 0 as long as a
factor of z z0 is present in each term. If n D 1, this happens for h and h 0, giving a second-order zero z0
of h. If n D 2, we have .z z0/4 and obtain f , f 0, f 00, f 000 equal to zero at z0, giving a fourth-order zero
z0 of h. And so on.

17. Singularities. We start with cotz: By definition,

1 cosz cotz D
D : tanz sinz
338 Complex Analysis Part D

By definition on p. 715, cotz is singular where cotz is not analytic. This occurs where sinz D 0, hence
for

(B) z D 0; ˙; ˙2;::: D ˙n, where n D 0;1;2;::::

Since cosz and sinz share no common zeros, we conclude that cotz is singular where sinz is 0; as
given in (B). The zeros are simple poles. Next we consider
cos4 z
4

cot z D 4:
sin z

Now sin4 z D 0 for z as given in (B). But, since sin4 z is the sine function to the fourth power and sinz
has simple zeros, the zeros of sin4 z are of order 4: Hence, by Theorem 4, p. 717, cot4 z has poles of
order 4 at (B).
But we are not finished yet. Inspired by Example 5, p. 718, we see that cosz also has an essential
singularity at 1. We claim that cos4 z also has an essential singularity at 1. To show this we would
have to develop the Maclaurin series of cos4 z. One way to do this is to develop the first few terms of
that series by (1), p. 690, of Sec. 15.4. We get (using calculus: product rule, chain rule)
1 1
4 2 4
(C) cos w D 1 4w C 40w C :
2Š 4Š
The odd powers are zero because in the derivation of (C) these terms contain sine terms (chain rule!)
that are zero at w0 D 0.
We set w D 1=z and multiply out the coefficients in (C):

1 5
4 2 4
(D) cos D 1 2z C z C :z 3
We see that the principal part of the Laurent series (D) is (D) without the constant term 1. It is
infinite and thus cot4 z has an essential singularity at 1 by p. 718. Since multiplication of the series by
1=sin4 z does not change the type of singularity, we conclude that cot4 z also has an essential
singularity at 1:

Sec. 16.3 Residue Integration Method

This section deals with evaluating complex integrals (1), p. 720, taken over a simple closed path C. The
important concept is that of a residue, which is the coefficient b 1 of a Laurent series that converges for all
points near a singularity z D z0 inside C, as explained on p. 720. Examples 1 and 2 show how to evaluate
integrals that have only one singularity within C.
A systematic study of residue integration requires us to consider simple poles (i.e., of order 1) and poles
of higher order. For simple poles, we use (3) or (4), on p. 721, to compute residues. This is shown in
Example 3, p. 722, and Prob. 5. The discussion extends to higher order poles (of order m) and leads to (5),
p. 722, and Example 4, p. 722. It is critical that you determine the order of the poles inside C correctly. In
Chap. 16 Laurent Series. Residue Integration 339

many cases we can use Theorem 4, on p. 717 of Sec. 16.2, to determine m. However, when h.z/ in
Theorem 4 is also zero at z0; the theorem cannot be applied. This is illustrated in Prob. 3.
Having determined the residues correctly, it is fairly straightforward to use the residue theorem
(Theorem 1, p. 723) to evaluate integrals (1), p. 720, as shown in Examples 5 and 6, p. 724, and Prob. 17.

Problem Set 16.3. Page 725

3. Use of the Laurent series. The function

sin2z
f.z/ D 6 has a singularity at z D z0 D 0:
z

However, since both sin2z and z6 are 0 for z0 D 0, we cannot use Theorem 4 of Sec. 16.2, p. 717, to
determine the order of that zero. Hence we cannot apply (5), p. 722, directly as we do not know the
value of m.
We develop the first few terms of the Laurent series for f.z/: From (14) in Sec. 15.4, p. 695, we
know that
w3 w5 w7
D 3Š C 5Š 7Š C
sinw w :
We set w D 2z and get

.2z/3 .2z/5 .2z/7

(A) sin2z D 2z CC :
3Š 5Š 7Š

Since we need the Laurent series of sin2z=z6 we multiply (A) by z 6 and get
6 6 .2z/3 .2z/5 .2z/7
(B) z sin2z D z 2z CC
3Š 5Š 7Š
2 81 32 1 128
D z5
3Š z3 C 5Š
z 7Š C
z :
The principal part of (B) is (see definition on p. 709)
2 81 32 1
:
5 3
z 3Š z C 5Š z

We see that
sin2z
f.z/ D 6 has a pole of fifth order at z D z0 D 0 [by (2), p. 715].
z

Note that the pole of f is only of fifth order and not of sixth order because sin2z has a simple zero at
z D 0:
Using the first line in the proof of (5), p. 722, we see that the coefficient of z 1 in the Laurent
series (C) is
340 Complex Analysis Part D

32 32 4
b1 D D D :
5Š 54321 15

Hence the desired residue at 0 is :

Checking our result by (5), p. 722. Having determined that the order of the singularity at z D z 0 D
0 is 5 and that we have a pole of order 5 at z 0 D 0, we can use (5), p. 722, with m D 5. We have

Res0 sin2z D 1 lim d6 1 .z 0/6f.z/ zDz D0 z6 .6 1/Š z!0 dz6 1

1 d5 6 sin2z
D lim z
5
5Š z!0 dz z6

1 d5
D lim sin2z :
5Š z!0 dz5

We need

g.z/ D sin2zI

g0.z/ D 2cos2zI

g00.z/ D 4sin2zI

g000.z/ D 8cos2zI

g.4/.z/ D 16sin2zI

g.5/.z/ D 32cos2z:

Then

lim f32cos2zg D 32 lim fcos2zg D 32 1: z!0 z!

0

Hence
sin2z 1 4
Res D 32 1 D ; as before. z D z0 D 0 z6 5Š 15

Remark. In certain problems, developing a few terms of the Laurent series may be easier than using
(5), p. 722, if the differentiation is labor intensive such as requiring several applications of the
quotient rule of calculus (see p. 623, of Sec. 13.3).

5. Residues. Use of formulas (3) and (4), p. 721. Step 1.

Find the singularities of f.z/. From

8 2 2
Chap. 16 Laurent Series. Residue Integration 341

f.z/ D C we see that 1Cz D0 implies z D 1, hence z D i and z D i: 1

z2

Hence we have singularities at z0 D i and z0 D i:

Step 2. Determine the order of the singularities and determine whether they are poles. Since the
numerator of f is 8 D h.z/ ¤ 0 (in Theorem 4), we see that the singularities in step 1 are simple, i.e., of
order 1. Furthermore, by Theorem 4, p. 717, we have two poles of order 1 at i and i,
respectively.
Step 3. Compute the value of the residues. We can do this in two ways.
Solution 1. By (3), p. 721, we have

8
Res f.z/ D lim .z i/ C
zDi z!i 1 z2

8
D lim .z i/ C z!i .z i/.z i/ 8

D lim z!i z i
8 4
D D D 4i:
2i i
Also

8
Res f.z/ D lim .z . i// C

zD i z! i .z i/.z i/
8
Res f.z/ D 4i and Res f.z/ D
D limi .z C i/
4i:
z! .z i/.z C i/
8
zDi zD i
D D 4i:
Res f.z/ D Res0 D D
2i
z = z0 z = z q.z/
Hence the two residues are q0.z0/
Solution 2. By (4), p. 721, we have
8 ˇˇ 8 ˇˇ 8
p.z/ p.z0/ .1 C z2/0 ˇˇz = z0 D 2zˇˇz = z0 D 2z0 :
For z0 D i we have

8
Res0 f.z/ D D 4i; z Di 2i

and for z0 D i

8
342 Complex Analysis Part D

Res f.z/ D D 4i; z0D i 2i

as before.

15. Residue theorem. We note that

sin2z
f.z/ D tan 2z D cos2z
is singular where cos2z D 0: This occurs at

3 5
2z D ˙ ;˙ ;˙ ;
2 2 2
and hence at
(A) z D ˙; ˙; ˙ :
Since sin2z ¤ 0 at these points, we can use Theorem 4, p. 717, to conclude that we have infinitely
many poles at (A).
Consider the path of integration C W jz 0:2j D 0:2. It is a circle in the complex plane with
center 0:2 and radius 0:2: We need to be only concerned with those poles that lie inside C: There is
only one pole of interest, that is,

z D D 0:25 (i.e., j0:25 0:2j D 0:05 < 0:2):

We use (4) p. 721, to evaluate the residue of f at z 0 D : We have

1
p.z/ D sin2z; p D sin D 1;
4 2

q.z/ D cos2z; q0 .z/ D 2 sin2z (chain rule!); q0 1

4 D 2 sin2 D 2:

Hence

p.1/ 1 1

Res1f.z/ D 41 D D : z0 = 4 q0.4/ 2 2

Thus, by (6) of Theorem 1, p. 723,

I I
f.z/dz D tan 2z dz
C CWjz 0:2jD0:2W

2i Res1f.z/ z0 = 4
D

1
Chap. 16 Laurent Series. Residue Integration 343

D 2i
2
D i:

17. Residue integration. We use the same approach as in Prob. 15. We note that

3 5
cosz D 0 at zD˙ ;˙ ;˙ :
2 2 2

Also ez is entire, see p. 631 of Sec. 13.5.

From Theorem 4, p. 717, we conclude that we have infinitely many simple poles at

3 5
D˙2˙2˙2 z;;.
Here the closed path is a circle:
ˇ i ˇˇ
and only zD and zD lie within C:
ˇ
2 2
C W ˇz ˇ D 4:5
ˇ 2ˇ

This can be seen because for

r 2 r22 p2
ˇ
2:2214 < 4:5:
i
D W ˇˇ i ˇˇ D ˇˇ z ˇ ˇ ˇ 2 ˇ DC 2 D 4 D 2 D
ˇ
2 ˇ2 2ˇ 2 2
Same for z D =2: Hence, by (4) p. 721,

=2
e=2
Res= f.z/ D D e ; z =2 sin=2
and

e =2 e =2 =2

Resf.z/ D D D e : z = 2 sin. =2/ sin=2

Using (6), p. 723,

I I ez
f.z/ dz D dz
C CWjz i=2jD4:5 cosz

D 2i

Res f.z/ C Res f.z/ z = =2 z = =2

344 Complex Analysis Part D

D 2i e=2 C e =2

D2i 2sinh[by (17), p. A65, App. 3]

D function]
2i

D
2sinh [since sinh is an odd

4i sinh
2 D
28:919i:

Sec. 16.4 Residue Integration of Real Integrals

It is surprising that residue integration, a method of complex analysis, can also be used to evaluate certain
kinds of complicated real integrals. The key ideas in this section are as follows. To apply residue
integration, we need a closed path, that is, a contour. Take a look at the different real integrals in the
textbook, pp. 725–732. For real integrals (1), p. 726, we obtain a contour by the transformation (2), p. 726.
This is illustrated in Example 1 and Prob. 7.
For real integrals (4), p. 726, and (10), p. 729 (real “Fourier integrals”), we start from a finite interval
from R to R on the real axis (the x-axis) and close it in complex by a semicircle S as shown in Fig. 374, p.
727. Then we “blow up” this contour and make an assumption (degree of the denominator degree of the
numerator C2) under which the integral over the blown-up semicircle will be 0. Note that we only take
those poles that are in the upper half-plane of the complex plane and ignore the others. Example 2,
p. 728, and Prob. 11 solve integrals of the kind given by (4). Real Fourier integrals (10) are solved in
Example 3, pp. 729–730, and Prob. 21.
Finally, we solve real integrals (11) whose integrand becomes infinite at some point a in the interval of
integration (Fig. 377, p. 731; Example 4, p. 732; Prob. 25) and requires the concept of Cauchy principal
value (13), p. 730. The pole a lies on the real axis of the complex plane.
Problem Set 16.4. Page 733

Z2 a Z2 1
a d:
dDa0 sin
a
0
sin
Using (2), p. 726, we get
1 1
a z
sin D a 2i z
7. Integral involving sine. Here the given integral is

and
dz
d D [see textbook after (2)]. iz
Hence

Z2 1 I i dz
Chap. 16 Laurent Series. Residue Integration 345

0 a sin D C iz a 21i z z1

d ;
where C is the unit circle.
Now

1 1 iz 1 iz
iz a z D iza zC
2i z 2i 2i z
z2 1
D iza C
2 2

D 1 z2 2aiz 12
so that the last integral is equal to

I
dz
2 :
2
Cz 2aiz 1

We need to find the roots of z2 2aiz 1: Using the familiar formula for finding roots of a quadratic
equation,
b pb2 4ac
2
az C bz C c D 0, z1;z2 D ˙
2a

with a D 1;b D 2ai;c D 1 we obtain

p p a2:
1;2 2ai ˙ . 2ai/2 4 1 . 1/ D 2ai ˙ 4.1 a2/ D ˙ p z Dai 1

2 2

By Theorem 4 of Sec. 16.2 on p. 717, we have two simple poles at p p

z1 D ai C 1 a2 and at z2 D ai 1 a2:
However, z1 is outside the unit circle and thus of no interest (see p. 726). Hence, by (3), p. 721, in Sec. 16.3
of the textbook, we compute the residue at z2: 1
Res f.z/ D Res
z = z2 z = zz1/.z z2/
1
D lim
.z z! z2/ z2
z2 .z z1/.z /
1
D lim2
z!z z z1
z 2 1 3 ai 1 a2
D4 p 5
D ai p1 a2 ai p1 a2
346 Complex Analysis Part D

1
:
D 2p1 a2
Thus by Theorem 1, p. 723, (Residue Theorem),

Z2 1 I dz
dD 2a
0 a sin C z2 2aiz 1
D
2a 2i Res f.z/
z = z2

1
D 4ai p
2 1 a2
2ai
:
D p1 a2

We can get rid of the i in the numerator by

p
p1 a2 D . 1/.a2 1/ D ipa2
1

so that the answer becomes

2a
p (as on p. A41 of the text). a2 1

11. Improper integral: Infinite interval of integration. Use of (7), p. 728. The integrand, considered
as a function of complex z, is
1
f.z/ D C 2/2 : .
1z

We factor the denominator and set it to 0

1 C z22 D .z2 C 1/.z2 C 1/ D .z i/.z C i/.z i/.z C i/ D .z i/2.z C i/2 D 0:

This shows that there are singularities (p. 715) at z D i and z D i; respectively.
We have to consider only z D i since it lies in the upper half-plane (defined on p. 619, Sec. 13.3)

and ignore z D i since it lies in the lower half-plane. This is as in Example 2, p. 728 (where only z 1

and z2 are used and z3 and z4 are ignored).

Furthermore, since the numerator of f.z/ is not zero for z D i, we have a pole of order 2 at z D i by
Theorem 4 of Sec. 16.2 on p 717. (The ignored singularity at z D i also leads to a pole of order 2.)
Chap. 16 Laurent Series. Residue Integration 347

The degree of the numerator of f.z/ is 1 and the degree of the denominator is 4, so that we are
allowed to apply (7), p. 728. We have by (5*), p. 722,

Res f.z/ D 1 lim d2 1 .z i/2 f.z/:

zDi .2 1/Š z!i dz2 1

Now we first do the differentiation

d 2 1 d 1
.z i/ C D C
dz .z i/2.z i/2 dz .z i/2

D .z C i/ 20

D 2.z C i/ 3
and then find the residue

Res f.z/ D lim 2.z C i/ 3 zDi z!i

2 2
D .z C i/3 zDi D .i C i/3

2 1 i
:
D 23i3 D 4i3 D 4

Hence by (7), p. 728, the real infinite integral is

Z1 1
dx D 2i Resf.z/
2
1 .x2 C 1/ z=i

i
D 2i
4
2
i
:
D D
2 2

21. Improper integral: Infinite interval of integration. Simple pole in upper-half plane. Simple
pole on real axis. Fourier integral. We note that the given integral is a Fourier integral of the form

Z1 1
f.x/sinsx dx with f.x/ D C and sD1 [see (8), p. 729].
348 Complex Analysis Part D

1 .x 1/.x2 4/

The denominator of the integrand, expressed in z factors, is

.z 1/.z2 C 4/ D .z 1/.z 2i/.z C 2i/ D 0:

This gives singularities of z D 1;2i; 2i, respectively. The pole at z D 2i lies in the upper half-plane
(defined in Sec. 13.3 on p. 619), while the pole at z D 1 lies on the contour. Because of the pole on
the contour, we need to find the principal value by (14) in Theorem 1 on pp. 731–2 rather than using
(10) on p. 729. (The simple pole z D 2i lies in the lower half-plane and, thus, is not wanted.)
y

2i

x
1

–2i

Sec. 16.4 Prob. 21. Fourier integral. Only the poles at z D 1 and 2i that

lie in the upper half-plane (shaded area) are used in the residue integration We

compute the residues

f.z/eiz (s D 1/
as discussed on p. 729, and in Example 3. Using (4), p. 721, we get

1 p.z/
iz 2 0
Res f.z/e D Res C eiz D zD1 zD1 .z 1/.z 4/ q .z/

zD1

where

p.z/ D eizI q.z/ D .z 1/.z2 C 4/ D z3 z2 C 4z 4I q0.z/ D 3z2 2z C 4:

Hence

p.1/ ei ei
iz
Res f.z/e D D C D : zD1 q0.1/ 3 2 4 5

Now by (5), p. 634, in Sec. 13.6 (“Euler’s formula in the complex”),

ei D cos1 C i sin1

so that
ei 1 1 1
Chap. 16 Laurent Series. Residue Integration 349

D .cos1 C i sin1/ D cos1 C i sin1:

5 5 5 5
Hence

1 iz
Re Res f.z/e D Re cos1 C i sin1 D : zD1 5

Also
p.z/
iz
Res f.z/e D
zD2i q0.z/ zD2i

ei2i

D 3.2i/2 2.2i/ C 4

e2

D 8 4i

D e 2. 8 C 4i/

82 C 4 2

D 8e 2 C 4e 2i

80
e2 e 2i
:
D C
10 20
Hence

e 2 iz
Re Res f.z/e D : zD2i 10

Using (14), p. 732, the solution to the desired real Fourier integral (with s D 0) is

Z 1 sinx X isz X isz pr. v C D Re Resf.z/e C 2 Re Res f.z/e

1 .x 1/.x2 4/ zD1 zD2i

D cos1 2e 2 D cos1 e 2 D 0:254448:

5 10 5
Note that we wrote pr.v., that is, Cauchy principal value (p. 730) on account of the pole on the
contour (x-axis) and the behavior of the integrand.
350 Complex Analysis Part D

25. Improper integrals. Poles on the real axis. We use (14) and the approach of Example 4, p. 732.
The denominator of the integrand is

x3 x D x x2 1 D x.x 1/.x C 1/:

Considering this in the complex domain, we have

(A) z3 z D z.z 1/.z C 1/

so that there are singularities at z D 0;1; 1:

Since the numerator is x C 5 and in the complex domain z C 5, we see that z C 5 is not zero for z
D 0;1; 1: Hence by Theorem 4 of Sec. 16.2 on p. 717,

zC5
f.z/ D has three simple poles at 0;1; 1: z3 z

We compute the residues as in Sec. 16.3, by using (4), p. 721,

p.z/ D z C 5I q.z/ D z3 z so that q0.z/ D 3z2 1:

Hence at z D 0

p.0/ 5
Resf.z/ 5:
0 2
z=0 D q .0/ D 3 0 1D

At z D 1

p.1/ 6
Resf.z/ 3:
z=1 D q0.1/ D 3 1D

Finally at z D 1

D p. 1/ D 1C5 D4D
Res f.z/ 2: z = -1 q0. 1/ 3. 1/2 1 2

We are ready to use (14), p. 732. Note that there are no poles in the upper half-plane as (A) does
not contain factors with nonzero imaginary parts. This means that the first summation in (14) is
zero. Hence

Z 1x C 5 D C C D D
pr:v: dx i.5 3 2/ i0 0:
3
1x x
Chap. 17 Conformal Mapping
We shift gears and introduce a third approach to problem solving in complex analysis. Recall that so far we
covered two approaches of complex analysis. The first method concerned evaluating complex integrals by
Cauchy’s integral formula (Sec. 14.3, p. 660 of the textbook and p. 291 in this Manual). Specific
background material needed was Cauchy’s integral theorem (Sec. 14.2) and, in general, Chaps. 13 and 14.
The second method dealt with residue integration, which we applied to both complex integrals in Sec.
16.3 (p. 719 in the textbook and p. 291 in this Manual) and real integrals in Sec. 16.4 (p. 725,
p. 326 in this Manual). The background material was general power series, Taylor series (Chap. 15), and,
most importantly, Laurent series which admitted negative powers (Sec. 16.1, p. 708) and thus lead to the
study of poles (Sec. 16.2, p. 715).
The new method is a geometric approach to complex analysis and involves the use of conformal
mappings. We need to explain two terms: (a) mapping and (b) conformal. For (a), recall from p. 621 in Sec.
13.3 that any complex function f.z/, where z D x C iy is a complex variable, can be written in the form

(1) w D f.z/ D u.x;y/ C iv.x;y/ (see also p. 737 in Sec. 17.1).

We want to study the geometry of complex functions f.z/ and consider (1).
In basic (real) calculus we graphed continuous real functions y D f.x/ of a real variable x as curves in the
Cartesian xy-plane. This required one (real) plane. If you look at (1), you may notice that we need to
represent geometrically both the variable z and the variable w as points in the complex plane. The idea is
to use two separate complex planes for the two variables: one for the z-plane and one for the w-plane.
And this is indeed what we shall do. So if we graph the points z D x C iy in the z-plane (as we have done
many times in Chap. 13) and, in addition, graph the corresponding w D u C iv (points obtained from
plugging in z into f ) in the w-plane (with uv-axes), then the function w D f.z/ defines a correspondence
(mapping) between the points of these two planes (for details, see p. 737). In practice, the graphs are
usually not obtained pointwise, as suggested by the definition, but from mappings of sectors, rays, lines,
circles, etc.
We don’t just take any function f.z/ but we prefer analytic functions. In comes the concept of (b)
conformality. The mapping (1) is conformal if it preserves angles between oriented curves both in
magnitude as well as in sense. Theorem 1, on p. 738 in Sec. 17.1, links the concepts of analyticity with
conformality: An analytic function w D f.z/ is conformal except at points z 0 (critical points) where its
derivative f 0.z0/ D 0:
The rest of the chapter discusses important conformal mappings and creates their graphs. Sections 17.1
(p. 737) and 17.4 (p. 750) examine conformal mappings of the major analytic functions from Chap. 13.
Sections 17.3 (p. 746) and 17.4 deal with the novel linear fractional transformation, a transformation that
is a fraction (see p. 746). The chapter concludes with Riemann surfaces, which allow multivalued relations
of Sec. 13.7 (p. 636) to become single-valued and hence functions in the usual sense. We will see the
astonishing versatility of conformal mapping in Chapter 18 where we apply it to practical problems in
potential theory.
You might have to allocate more study time for this chapter than you did for Chaps. 15 and 16.
You should study this chapter diligently so that you will be well prepared for the applications in Chap. 18.
352 Complex Analysis Part D

As background material for Chap. 17 you should remember Chap. 13, including how to graph complex
numbers (Sec. 13.1, p. 608), polar form of complex numbers (Sec. 13.2, p. 613), complex functions (pp.
620–621), ez (Sec. 13.5, p. 630), Euler’s formula (5), p. 634, sinz;cosz;sinhz;coshz, and their various
formulas (Sec. 13.6, p. 633), and the multivalued relations of Sec. 13.7, p. 636 (for optional Sec. 17.5).
Furthermore, you should know how to find roots of polynomials and know how to algebraically
manipulate fractions (in Secs. 17.2 and 17.3).
Sec. 17.1 Geometry of Analytic Functions: Conformal Mapping
We discussed mappings and conformal mappings in detail in the opening to Chap. 17 of this Manual.
Related material in the textbook is: mapping (1), p. 737, and illustrated by Example 1; conformal, p. 738;
conformality and analyticity in Theorem 1, p. 738. The section continues with four more examples of
conformal mappings and their graphs. They are w D z n (Example 2, p. 739), w D z C 1=z (Joukowski airfoil,
Example 3, pp. 739–740), w D ez (Example 4, p. 740), and w D Ln z (Example 5, p. 741). The last topic is the
magnification ratio, which is illustrated in Prob. 33, p. 742.
In the examples in the text and the exercises, we consider how sectors, rays, lines, circles, etc. are
mapped from the z-plane onto the w-plane by the specific given mapping. We use polar coordinates and
Cartesian coordinates. Since there is no general rule that fits all problems, you have to look over,
understand, and remember the specific mappings discussed in the examples in the text and
supplemented by those from the problem sets. To fully understand specific mappings, make graphs or
sketches. Finally you may want to build a table of conformal mappings:
Mapping Region to be Mapped Image of Region Reference

wDzn Upper half plane v0 Example 2, p. 739

Sector 0 n
Graph in z-plane Graph in w-plane
y v

π/n
x u
Put in more mappings and graphs or sketches. The table does not have to be complete, it is just to help
you remember the most important examples for exams and for solving problems.
Illustration of mapping. Turn to p. 621 of Sec. 13.3 and look at Example 1. Note that this example defines a
mapping w D f.z/ D z2 C 3z: It then shows how the point z0 D 1 C 3i (from the z-plane) is being mapped onto
w0 D f.z0/ D f.1 C 3i/ D .1 C 3i/2 C 3.1 C 3i/ D 1 C 6i C 9i2 C 3 C 9i D
5 C 15i (of the w-plane). A second such example is Example 2, p. 621.

More details on Example 1, p. 737. Turn to p. 737 and take a look at the example and Fig. 738. We
remember that the function f.z/ D z2 is analytic (see pp. 622–624 of Sec. 13.3, Example 1, p. 627 of Sec.
13.4). The mapping of this function is

w D f.z/ D z2:

It has a critical point where the derivative of its underlying function f is zero, that is, where
Chap. 17 Conformal Mapping 353

f 0.z/ D 0 here f 0.z/ D .z2/0 D 2z D 0 hence z D 0:

Thus the critical point is at z D 0. By Theorem 1, p. 738, f.z/ is conformal except at z D 0. Indeed, at z D 0
conformality is violated in that the angles are doubled, as clearly shown in Fig. 378, p. 737. The same
reasoning is used in Example 2, p. 739.

Problem Set 17.1. Page 741

3. Mapping. To obtain a figure similar to Fig. 378, p. 737, we follow Example 1 on that page. Using
polar forms [see (6), p. 631 in Sec. 13.5]

z D rei and w D Rei

we have, under the given mapping w D z3,

Rei D w D f.z/ D f.rei/ D .rei/3 D r3ei3:

We compare the moduli and arguments (for definition, see p. 613) and get

R D r3 and D 3:

Hence circles r D r0 are mapped onto circles R D r03 and rays D 0 are mapped onto rays D 30. Note that
the resulting circle R D r03 is a circle bigger than r D r0 when r0 > 1 and a smaller one when r0 < 1.
Furthermore, the process of mapping a ray D 0 onto a ray D 30 corresponds to a rotation.
We are ready to draw the desired figure and consider the region
2
1 r 1:3 with :
9 9
It gets mapped onto the region 13 R .1:3/3 with 3 =9 3 2=9. This simplifies to

2
1 R 2:197 with :
3 3
354 Complex Analysis Part D

y
1.6
1.3
1

x x
11.3 1.6 –4.096 –2.197 –1 0 1 2.197 4.096

Sec. 17.1 Prob. 3. Given region and its image under the mapping w D z 3

7. Mapping of curves. Rotation. First we want to show that the given mapping w D iz is indeed a
rotation. To do this, we express z in polar coordinates [by (6), p. 631], that is,

(A) z D rei where r > 0:

Then we obtain the image of (A) under the given mapping by substituting (A) directly into the
mapping and simplifying:

w D f.z/ D f.rei/ D izjz r ei

D ei=2rei [using i D ei=2 by (8), p. 631]

D rei.C=2/

Q
D reiQ where D C =2:

This shows that this mapping, w D iz, is indeed a rotation about 0 through an angle of =2 in the
positive sense, that is, in a counterclockwise direction.
We want to determine the images of x D 1;2;3;4, and so we consider the more general problem
of determining the image of x D c, where c is a constant. Then for x D c, z becomes

z D x C iy D c C iy

so that under the mapping

(B) w D f.z/ D izjz c iy D i.c C iy/ D ic C i2y D y C ic:

Chap. 17 Conformal Mapping 355

DC

This means that the image of points on a line x D c is w D y C ic. Thus x D 1 is mapped onto w D y C
i; x D 2 onto w D y C 2i, etc. Furthermore, z D x D c on the real axis and is mapped by (B) onto the
imaginary axis w D ic. So z D x D 1 is mapped onto w D i, and z D x D 2 is mapped onto w D 2i; etc.
Similar steps for horizontal lines y D k D const give us

z D x C ik so that w D i.x C ik/ D k C ix:

Hence y D 1 is mapped onto w D 1 C ix, and y D 2 onto w D 2 C ix: (Do you see a
counterclockwise rotation by =2?). Furthermore, z D y D k is mapped onto w D k and z D y D 1 onto w
D 1; z D y D 2 onto w D 2. Complete the problem by sketching or graphing the desired images.

11. Mapping of regions.i To examine the given mapping, w D z2, we express z in polar coordinates,

that is, z D re and substitute it into the mapping

Rei D w D f.z/ D f.rei/ D z2ˇˇz rei D .rei/2 D r2.ei/2 D r2ei2:

D

This shows that the w D z2 doubles angles at z D 0 and that it squares the moduli. Hence for our
problem, under the given mapping,
< < becomes < <
8 8 4 4
or equivalently (since D Arg z and D Arg w/

< Arg z < becomes < Arg w < [for definition of Arg, see (5), p. 614]:
8 8 4 4
Furthermore, r D maps onto R D : Since

jzj D r [by (3), p. 613 in Sec. 13.2]

we get jzj ; which becomes jwj ;

which corresponds to the answer on p. A41 in Appendix 2 of the textbook. Together we

obtain the figures below.
356 Complex Analysis Part D

y v

π π
8 4

–π 1 x –π 1 u
8 2 4 4

z-plane w-plane

Sec. 17.1 Prob. 11. Given region and its image under the mapping w D f.z/ D z 2

15. Mapping of regions. The given region (see p. 619 in Sec. 13.3)
ˇ 1ˇ
ˇz 2ˇ is a closed circular disk of radiuswith center at x D:

The corresponding circle can be expressed as

x 122 C y2 D 122 :

Written out

x2 x C C y2 D

and rearranged is

x2 C y2 xC D :

Subtract from both sides of the equation and get

x2 C y 2 x D 0:

But

x2 C y2 D jzj2 D zz [by (3), p. 613 of Sec. 13.2].

Furthermore,

z C z x D:
2
Chap. 17 Conformal Mapping 357

Substituting these last two relations into our equation yields z C z D

(*) zz 0:
2
Now we are ready to consider the given mapping, w D 1=z, so that z D 1=w and obtain

C 1
z zD11 w C w1
zz
2 ww 2

11

1 1 D 2 w w
2 ww
11
.2D 2 ww C Œu iv Œu iv/
[by (1), p. 737 and definition of w]
11

D .2 2u/ w w

D0 [from the l.h.s of (*)].

For the last equality to hold, we see that

2 2u D 0 so that u D 1:

This shows that, for the given mapping, the circle maps onto u D 1: The center
1
of the circle ;0 maps onto
2

1ˇˇ 1
f.z/ D ˇD 1D 2 D u C iv so that u D 2:
zˇzD 2

This is > 1 so that the inside of the circle maps to u 1:

17. Mapping of regions. Exponential function. We take the exponents of

Ln2 x Ln4; which is equivalent to Ln(2 1/ x Ln4

358 Complex Analysis Part D

and, because the logarithm Ln x is monotone increasing with x, we obtain

ex 4:
Now

ex D j e z j D j w j [by (10) in Sec. 13.5, p. 631].

Hence the given region gets mapped onto

j w j 4:

21. Failure of conformality. Cubic polynomial. The general cubic polynomial (CP) is

(CP) a3z3 C a2z2 C a1z C a0:

Conformality fails at the critical points. These are the points at which the derivative of the cubic
polynomial is zero. We differentiate (CP) and set the derivative to zero:

.a3z3 C a2z2 C a1z C a0/0 D 3a3z2 C 2a2z C a1 D 0:

We factor by the well-known quadratic formula for a general second-order polynomial az 2 C bz

C c D 0 and obtain

p p
D b ˙ pb2 4ac D 2a2 ˙ 4a 22 4 3a3 a1 D a2 ˙ a22 3 a 3 a1

z1;2:
2a 2 3a3 3a3

Thus the mapping, w D f .z/, is not conformal if f 0 .z/ D 0. This happens when

D a2 ˙ pa22 3 a3 a1 z:
3a3

Remark. You may want to verify that our answer corresponds to the answer on p. A41 in
Appendix 2 of the textbook. Set

a3 D 1, a2 D a, a1 D b, a0 D c:

Note that we can set a3 D 1 in (CP) without loss of generality as we can always divide the cubic
polynomial by a3 if 0 < ja3j ¤ 1:

33. Magnification ratio. By (4), p. 741, we need

Chap. 17 Conformal Mapping 359

ˇ ˇ
ˇ.ez/0 ˇ D j ez j D ex (by Sec. 13.5, p. 630):

Hence M D 1 when x D 0, which is at every point on the y-axis.

Also, M < 1 everywhere in the left half-plane because e x < 1 when x < 0, and M > 1 everywhere
in the right half-plane.
By (5), p. 741, we show that the Jacobian is

J D ˇˇf 0 .z/ˇˇ2 D j .ez/0 j2 D j ez j2 D .ex/2 D e2x:

Confirm this by using partial derivatives in (5).

Sec. 17.2 Linear Fractional Transformations. (Mobius Transformations)¨
This new function on p. 473

D az C b ¤
(1) LFT w (where ad bc 0)
cz C d

is useful in modeling and solving boundary value problems in potential theory [as in Example 2 of Sec.
18.2, where the first function on p. 765, of the textbook, is a linear fractional transformation (LFT)
.1/ with a D b D d D 1 and c D 1].
LFTs are versatile because—with different constants—LFTs can model translations, rotations, linear
transformations, and inversions of circles as shown in (3), p. 743. They also have attractive properties
(Theorem 1, p. 744). Problem 3 (in a matrix setting) and Prob. 5 (in a general setting) explore the
relationship between LFT (1) and its inverse (4), p. 745. Fixed points are defined on p. 745 and illustrated in
Probs. 13 and 17.

Problem Set 17.2. Page 745

3. Matrices. a. Using 2 2 matrices, prove that the coefficient matrices of (1), p. 743, and (4), p. 745,
are inverses of each other, provided that

ad bc D 1:

Solution. We start with

az b
(1) wD CC (where ad bc ¤ 0),
cz d
and note that its coefficient matrix is
a b
(M1) AD :
c d
Using formula (4*) in Sec. 7.8
on p. 304, we have that the inverse of matrix A is
360 Complex Analysis Part D

1
1 d b
(M2) A D ; (wheredet A D ad bc/: det A c a

We are given that the inverse mapping of (1) is

dw b
(4) zD C
cw a
so that its coefficient matrix is

d b
(M3) BD :
c a

Looking at (M2) and (M3) we see that if the only way for A 1 D B is for

1 1
D D 1; that is, ad bc D 1:
det A ad bc
Conversely, if ad bc D 1, then

1 11
D DD 1 so that A D 1 B D B [by (M2), (M3)]. det A ad bc
This proves a.
b. The composition of LFTs corresponds to the multiplication of coefficient matrices.
Hint: Start by defining two general LFTs of the form (1) that are different from each other.

5. Inverse. a. Derive (4), p. 745, from (1), p. 743. We

start with

D az C b ¤
(1) w (where ad bc 0)
cz C d

and multiply both sides of (1) by cz C d, thereby obtaining

.cz C d/w D az C b:

Next we group the z-terms together on the left and the other terms on the right:

czw az D b dw
Chap. 17 Conformal Mapping 361

so that

(A) z .cw a/ D b dw:

We divide both sides of (A) by .cw a/ and get

(A0) zDb
cw a

This is not quite (4) yet. To obtain (4), we multiply (A 0) by (which we can always do) and get

D .b dw/ D b CCdw D dw Cb z :
.cw a/ cw a cw a

But this is precisely (4)! (Note that the result is determined only up to a common factor in the
numerator and the denominator).
b. Derive (1) from (4).
This follows the same approach as in a, this time starting with (4) and deriving (1). For practice you
should fill in the steps.

7. Inverse mapping. The given mapping is a linear fractional transformation. Using (1), p. 743, we have

D i D 0 z C i D az CC b
w

2z 1 2z 1 cz d
so that, by comparison,

a D 0; b D i; c D 2; d D 1:
We now use (4), p. 745, with the values of a; b; c; d just determined and get that the inverse
mapping of (1) is

dw Cb 1 w i D w i z D z.w/ D D C : cw a

2 w 0 2w

This compares with the answer in the textbook on p. A41 since

w iD .w C i/ D w C i
:
2w .2w/ 2w

To check that our answer is correct, we solve z.w/ for w and have

w i
362 Complex Analysis Part D

zD and hence z. 2w/ D w i;

2w
or

2wz D w i:

Adding w gives us

2wz C w D i so that w. 2z C 1/ D i:

We solve the last equation for w and then factor out a minus sign both in the numerator and
denominator to get

i .i/ i
D 2z C 1 D .2z 1/ D 2z 1 w :
The last fraction is precisely the given mapping with which we started, which validates our answer.

13. Fixed points. The fixed points of the mapping are those points z that are mapped onto themselves
as explained on p. 475. This means for our given mapping we consider

from given mapping

‚ …„ ƒ
16z5 D w D f.z/ D z.
„ ƒ‚ …
by definition of
fixed point

Hence our task is to solve

16z5 D z or equivalently 16z5 zD0 so that z 16z4 1 D 0:

The first root (“fixed point”) is immediate, that is, z D 0 : We then have to solve

(B) 16z4 1 D 0:

For the next fixed point, from basic elementary algebra, we use that

(C) x2 a2 D .x a/.x C a/:

In (B) we set

16z4 D .4z/2 D 2 to obtain 2

1D0 and .2 1/.2 C 1/ D 0 [by (C)].

Written out we have

(D) 4z2 1 4z2 C 1 D 0:

Chap. 17 Conformal Mapping 363

For the first factor in (D) we use (C) again and, setting to zero, we obtain

4z2 1 D .2z 1/.2z 1/ D 0

so that two more fixed points are

2z 1D0 giving: z D 12 and similarly zD 12

Considering the second factor in (D) r p

2C D 0 gives z2 D 1 and zD˙ 1 D ˙ p 1 D ˙i : 4z

1
4 4 4 2

This means we have two more fixed points

: zD i and zD i

We have found five fixed points, and, since a quintic polynomial has five roots (not necessarily distinct),
we know we have found all fixed points of w:
Remark. We wanted to show how to solve this problem step by step. However, we could have solved
the problem more elegantly by factoring the given polynomial immediately in three steps:

16z5 z D z 16z4 1 D z 4z2 1 4z2 C 1 D z .2z 1/.2z C 1/.2z C i/.2z i/ D 0:

Another way is to solve the problem in polar coordinates with (15), p. 617 (whose usage is illustrated
in Prob. 21 of Sec. 13.2 on p. 264, in this Manual).

17. Linear fractional transformations (LFTs) with fixed points. In general, fixed points of mappings w
D f.z/ are defined by

(E) w D f.z/ D z:

LFTs are given by (1), p. 743,

D az C b
(1) w :
cz C d

Taking (E) and (1) together gives the starting point of the general problem of finding fixed points for
LFTs, that is, az C b D w D C z. cz d
This corresponds to (5), p. 745. We obtain

az C b D C C D z 0 so that az b z.cz d/ 0:
364 Complex Analysis Part D

cz C d

The last equation can be written as

(F) cz2 C .d a/z bD0 [see formula in (5), p. 745].

For our problem, we have to find all LFTs with fixed point z D 0. This means that (F) must have a
root z D 0: We have from (F) that

(F*) z.cz C d a/ D b

and, with the desired fixed point, makes the left-hand side (F*) equal to 0 so that the right-hand side
of (F*) must be 0, hence

b D 0:

Substitute this into (1) gives the answer

D az CC b D azCC 0 D azC
(G) w :
cz d cz d cz d
To check our answer, let us find the fixed points of (G). We have
az
CDz so that az D z.cz C d/:
cz d
This gives

z.cz C d a/ D 0

which clearly has z D 0 as one of its roots (“fixed point”).

Sec. 17.3 Special Linear Fractional Transformations

The important formula is (2), p. 746 (also stated in Prob. 5 below). It shows that by showing how three
points z1, z2, z3 are mapped onto w1, w2, w3 we can derive an LFT in the form (1) of Sec. 17.2. In the
textbook, we give six examples of LFTs that are useful in Chap. 18. Note that we allow points to take on
the value of 1 (infinity), see Examples 2, 3, and 4 on p. 748 and Probs. 5 and 13. Moreover, the approach
of Sec. 17.3, as detailed in Theorem 1, p. 746, assures us that we obtain a unique transformation.
Remark on standard domain. By “standard domains,” on p. 747, we mean domains that occur frequently in
applications, either directly or after some conformal mapping to another given domain. For instance, this
Chap. 17 Conformal Mapping 365

happens in connection with boundary value problems for PDEs in two space variables. The term is not a
technical term.

Problem Set 17.3. Page 750

3. Fixed points. To show that a transformation and its inverse have the same fixed points we proceed as
follows. If a function w D f.z/ maps z 1 onto w1, we have w1 D f.z1/, and, by definition of the inverse f 1,
we also have z1 D f 1.w1/. Now for a fixed point z D w D z 1, we have z1 D f.z1/, hence z1 D f 1.z1/, as
claimed.
5. Filling in the details of Example 2, p. 748, by formula (2), p. 746. We want to derive the mapping
in Example 2, p. 748, from (2), p. 746. As required in Example 2, we set

z1 D 0, z2 D 1;z3 D 1I w1 D 1, w2 D i, w3 D 1

in
w w1 w2 w3 z z1 z2 z3
(2) D
w w3 w2 w1 z z3 z2 z1

and get

w 1 i 1 z 0 1
(A) C :
w 1 i C 1 D z 1 110

On the left-hand side, we can simplify by (7), p. 610, of Sec. 13.1, and obtain

iC1D 1 iD 1 i 1 CC i D 1 C2i C 1 D 2i D
i:
i 1 1 i 1 i 1 i 12 12 2
On the right-hand side, as indicated by Theorem 1, p. 746, we replace
11z1
by 1. Together we obtain, from (A), w C 1 D z 0

. i/ 1
w110
w 1
so that C D

. i/ z: w 1
This gives us the intermediate result w C 1 D z D
366 Complex Analysis Part D

iz.
w 1 i
Note that we used 1=i D i (by Prob. 1, p. 612, and solved on p. 258 in this Manual). We solve
for w and get
w C 1 D iz .w 1/I w C 1 D izw izI w izw D iz 1I w .1 i z/ D iz 1;

so that
iz 1 1 1
iz 1w
D iz C i D C iD C iD
1D C;z z z
or, alternatively,
i
izCi 1 z 11 z 1i z i
iz 1 i. iz 1/ z i
w ;
D iz C 1 D i. iz C 1/ D z C i
both of which lead to the desired result.
13. LFT for given points. Our task is to determine which LFT maps 0, 1, 1 into 1, 1, 0:

First solution by Theorem 1, p. 746. By (2), p. 746, with

z1 D 0, z2 D 1;z3 D 1I w1 D 1, w2 D 1, w3 D 0

we have
w
(B) 1 1110 D zz10 1110 :

w 0
As required by Theorem 1, we have to replace, on the left-hand side,
w 1 by 1
11

and also, on the right-hand side,

11z1 by 1:

This simplifies (B)

1 0 so that
1 00 1 zw
z Dw 1
0 1 D 1 D z:

Hence
1wDz

is the desired LFT.

Second solution by inspection. We know that
Chap. 17 Conformal Mapping 367

z 7! w D f.z/ [read z gets mapped onto w D f.z/]

and here

1 7! 1;
0:
2 7! 1; 1 7!
Looking at how these three points are mapped, we would conjecture that w D 1=z and see that this

mapping does fulfill the three requirements. 17. zwhich leads us to Example 4, pp. 748–749. We

setMapping of a disk onto a disk.D i=2 is mapped onto w D 0: From p. 619, we know thatWe have to

find an LFT that mapsjzj 1 andjzj wj 1onto1 represent disks,jwj 1 so that

i i
z0 D 2 and c D z0 D 2 D2;
in (3), p. 749, and obtain
i 2z i
z z0 z 2z i 2 2z i
wD D i 2 D iz2 2 D D :
cz 1 2 z 1 2 2 iz 2 iz 2

Complete the answer by sketching the images of the lines x D const and y D const.

19. Mapping of an angular region onto a unit disk. Our task is to find an analytic function, w D f.z/,
that maps the region 0 argz =4 onto the unit disk j w j 1: We follow Example 6, p. 749, which
combines a linear fractional transformation with another transformation. We know, from Example 2,
p. 739 of Sec. 17.1, that t D z4 maps the given angular region 0 argz =4 onto the upper t-half-plane.
(Make a sketch, similar to Fig. 382, p. 739.) (Note that the transformation D t 8 would map the given
region onto the full -plane, but this would of no help in obtaining the desired unit disk in the next
step.)
Next we use (2) in Theorem 1, p. 746, to map that t-half-plane onto the unit disk j w j 1 in the
w-plane. We note that this is the inverse problem of the problem solved in Example 3 on p. 748 of
the text.
Clearly, the real t-axis (boundary of the half-plane) must be mapped onto the unit circle j w j D
1. Since no specific points on the real t-axis and their images on the unit circle j w j D 1 are
prescribed, we can obtain infinitely many solutions (mapping functions).
For instance, if we map t1 D 1, t2 D 0, t3 D 1 onto w1 D 1, w2 D i, w3 D 1,
368 Complex Analysis Part D

respectively—a rather natural choice under which 1 and 1 are fixed points—we obtain, with these
values inserted into (2) in modified form (2*), that is, inserted into

w w1 w2 w3 t t1 t2 t3
(2*) D
w w3 w2 w1 t t3 t2 t1

equation (C):

wC1 iC1DtC1 0C1DtC1 D t 1

(C) . 1/ :w 1 i 1 t 1 0
1 t 1 t 1

We want to solve (C) for w: Cross-multiplication and equating leads to

.w C 1/. i 1/.t 1/ D .w 1/. i C 1/. t 1/:

This gives us
w. i 1/.t 1/ C . i 1/.t 1/ D w. i C 1/. t 1/ . i C 1/. t 1/

and w. i 1/.t 1/ w. i C 1/. t 1/ D .i 1/.t 1/ . i C 1/. t 1/:

We get w. it C i t C 1 it i C t C 1/ D it i C t 1 it i C t C 1;

which simplifies to w. 2it C 2/ D 2i C 2t and 2w. it C 1/ D 2. i C t/:

Solving for w gives us

i t t i
(D) wD CC D C :
it 1 it 1

From above we know that the mapping t D z 4, which substituted into (D). gives us

t i z4 i
(E) wD C D C ;
4
it 1 iz 1

which is the answer given on p. A42. Note that the mapping defined by (E) maps t D i onto w D 0, the
center of the disk.
Chap. 17 Conformal Mapping 369

B →A C← –1 B 1 →A

z-plane t-plane

C A

–1 1

w-plane

Sec. 17.3 Prob. 19. z-, t-, and w-planes and regions for the given LFT

Sec. 17.4 Conformal Mapping by Other Functions

We continue our discussion of conformal mappings of the basic analytic functions (from Chap. 13) that we
started in Sec. 17.1. It includes w D sinz (pp. 750–751, Prob. 11), w D cosz (p. 752, Prob. 21), w D sinhz
and w D coshz (p. 752), and w D tanz (pp. 752–753). Take your time to study the examples, as they are
quite tricky.
We expand the concept of transformation by introducing the composition of transformations (see Fig.
394, p. 753, three transformations). It allows us to break a more difficult conformal mapping problem into
two or three intermediate conformal mapping problems—one following the other. (Aside: We
encountered this concept before, but in a different setting, that is, in Sec. 7.9, Composition of Linear
Transformations, pp. 316–317 of the textbook, as the more theoretically inclined reader may note.)
Problem Set 17.4. Page 754

3. Mapping w = ez. The given region to be mapped by w D e z is a (solid)

rectangle R defined by x and y :

Since

j w j D j e z j D ex [by (10), p. 631 in Sec. 13.5];

370 Complex Analysis Part D

we have that the inequality

x implies that e 1=2 D 0:607 j w j e1=2 D 1:649:

This is an annulus in the w-plane with center 0. The inequality

gives no further restriction, since y ranges between and . Indeed, the side x D of R is

mapped onto the circle e 1=2 in the w-plane and the side x D onto the circle of radius e1=2. The
images of the two horizontal sides of R lie on the real w-axis, extending from e 1=2 to e1=2 and
coinciding.
Remark. Take another look at Example 4 in Sec. 17.1 on p. 740 to see how other rectangles are
mapped by the complex exponential function.
y v
C π B

B C
–1/2 1/2 x –1/2 1/2 u
A D e e

D –π A

z-plane w-plane

Sec. 17.4 Prob. 3. Region R and its image

11. Mapping w D sinz. The region to be mapped is

rectangle R given by 0<x< and 0 < y < 2:

2
We use the approach of pp. 750–751 of the textbook, which discusses mapping by the complex sine
function. We pay attention to Fig. 391 on p. 751. We use

(1) w D sinz D sinx coshy C i cosx sinhy [p. 750 or (6b), p. 634, in Sec. 13.6].
Since

0<x< we have sinx > 0

2
Chap. 17 Conformal Mapping 371

and, because

coshy > 0; we obtain u D sinhx coshy > 0:

This means that the entire image of R lies in the right half-plane of the w-plane. The

origin z D 0 maps onto the origin w D 0:

On the bottom edge of the rectangle (z A D 0 to zB D =2)

w D sinx cosh0 C cosx sinh0 D sinx

so it goes from w D 0 to 1. On the vertical right edge (z B D x D =2 to zC D =2 C 2i)

w D sin=2coshy C i cos=2sinhy D coshy

so it is mapped from

wD1 to w D cosh2 3:76:

The upper horizontal side y D 2, =2 > x > 0 is mapped onto the upper right part of the ellipse

u2 v 2 2 C 2
D
1 .u > 0/; .v < 0/:
cosh 2 sinh 2

Finally, on the left edge of R (from z D D 0 C 2i to zA D 0/,

w D sin0coshy C i cos0sinhy D i sinhy;

so it is mapped into the v-axis u D 0 from i sinh2 to 0. Note that, since the region to be mapped consists of
the interior of a rectangle but not its boundary, the graphs also consist of the interior of the regions
without the boundary.

v
D
3

D C

A B A B C
372 Complex Analysis Part D

y
22

11

0 π/2 x 0 1 2 3 4 u z-plane w-plane

Sec. 17.4 Prob. 11. Rectangle and its image under w D sinz
21. Mapping w D cosz. We note that the rectangle to be mapped is the same as in Prob. 11. We can
solve this problem in two ways.
Method 1. Expressing cosine in terms of sine. We relate the present problem to Prob. 11 by using

cosz D sin.z C /:

We set

tDzC :

Then the image of the given rectangle [x in .0;=2/; y in .0;2/] in the t-plane is bounded by Re t in
;x C 12 or ;; and Im t in .0;2/; i.e. shifted =2 to the right. Now

w D sint D sin x C 21coshy C i cos x C 12sinhy:

Now proceed as in Prob. 11.

Method 2. Direct solution. To solve directly, we recall that

w D cosz D cosx coshy i sinx sinhy

and use zA;zB;zC; and zD as the four corners of the rectangle as in Prob. 11. Now,

zA maps to cos0cosh0 i sin0sinh0 D 1;

zB maps to coscosh0 i sinsinh0 D 0;

2 2

zC maps to cos cosh2i sin sinh2 D i sinh2,

2 2
zD maps to cos0cosh2y i sin0sinh2 D cosh2.
On the bottom edge of the rectangle (zA D 0 to zB D =2) w D cosx cosh0 i sinx sinh0 D

cosx so it goes from w D 1 to 0:

Chap. 17 Conformal Mapping 373

On the vertical right edge (zB D x D =2 to zC D =2 C 2i)

w D cos coshy i sin sinhy D i sinhy so it is mapped from

w D 0 to w D i sinh2:
2 2

The upper horizontal side y D 2, =2 > x > 0 is mapped onto the lower right part of the ellipse:

u2 v2 2 C 2
D
1 .u > 0/:
cosh 2 sinh 2

Finally, on the left edge of R (from z D D 0 C 2i to zA D 0/

w D cos0coshy i sin0sinhy D coshy;

so it is mapped into the v-axis u D 0 from cosh2 to 0.

As in Prob. 11, our solution consists only of the interior of the regions depicted.

y
2 2
D C D C

1 1

A B A B
0 x 0
π/2 π/2 π z-plane t-plane
1 2 3 4
0
B A D

–1

–2

–3
C

w-plane

Sec. 17.4 Prob. 20. Given region in the z-plane and its images in the t- and
w-planes for the mapping of w D cosz

Sec. 17.5 Riemann Surfaces. Optional

374 Complex Analysis Part D

Problem Set 17.5. Page 756

Riemann surfaces (Fig. 395, p. 755) contain an ingenious idea that allows multivalued relations, such as w
D pz and w D lnz (defined in Sec. 13.7, pp. 636–640) to become single-valued. The Riemann surfaces (see
Fig. 395 on p. 755) consist of several sheets that are connected at certain points (“branch points”). On
these sheets, the multivalued relations become single-valued. Thus, for the complex square root being
double-valued, the Riemann surface needs two sheets with branch point 0.

1. Square root. We are given thatp z moves from z D twice around the circle jzj D and want to know

what w D z does.

We use polar coordinates. We set

z D rei [by (6), p. 631 in Sec. 13.5]:

On the given circle,

jzj D r D [see p. 619 and (3), p. 613],

so that we actually have

(A) z D 41ei:
Hence the given mapping w D pz

D
1 i1=2 e [by (A)]:
4

D ei=2

Since z moves twice around the circle jzj D ,

increases by 2 2 D 4:
Hence
4 D 2:
increases by 2
2
This means that w goes once around the circle jwj D , that is, the circle of radius centered at 0 in the w-
plane.
Chap. 18 Complex Analysis and Potential Theory
We recall that potential theory is the area that deals with finding solutions (that have continuous second
partial derivatives)—so-called harmonic functions—to Laplace’s equation. The question that arises is how
do we apply complex analysis and conformal mapping to potential theory. First, the main idea which links
potential theory to complex analysis is to associate with the real potential ˆ in the two–dimensional
Laplace’s equation
r2
ˆ D ˆxx C ˆyy D 0

a complex potential F

(2) F.z/ D ˆ.x; y/ C i ‰.x; y/.

This idea is so powerful because (2) allows us to model problems in distinct areas such as in electrostatic
fields (Secs. 18.1, p. 759, 18.2, p. 763, 18.5, p. 777), heat conduction (Sec. 18.3, p. 767), and fluid flow
(Sec. 18.4, p. 771). The main adjustment needed, in each different area, is the interpretation of ˆ and ‰ in
(2), specifically the meaning of ˆ D const and its associated conjugate potential ‰ D const. In electrostatic
fields, ˆ D const are the electrostatic equipotential lines and ‰ D const are the lines of electrical force—
the two types of lines intersecting at right angles. For heat flow, they are isotherms and heat flow lines,
respectively. And finally, for fluid flow, they are equipotential lines and streamlines.
Second, we can apply conformal mapping to potential theory because Theorem 1, p. 763 in Sec. 18.2,
asserts “closure” of harmonic functions under conformal mapping in the sense that harmonic functions
remain harmonic under conformal mapping.
Potential theory is arguably the most important reason for the importance of complex analysis in
applied mathematics. Here, in Chap. 18, the third approach to solving problems in complex analysis—the
geometric approach of conformal mapping applied to solving boundary value problems in two–
dimensional potential theory—comes to full fruition.
As background, it is very important that you remember conformal mapping of basic analytic functions
(power function, exponential function in Sec. 17.1, p. 737, trigonometric and hyperbolic functions in Sec.
17.4, p. 750), and linear fractional transformations [(1), p. 743, and (2), p. 746]. For Sec. 18.1, you may
also want to review Laplace’s equation and Coulomb’s law (pp. 400–401 in Sec. 9.7), for Sec. 18.5,
Cauchy’s integral formula (Theorem 1, p. 660 in Sec. 14.3), and the basics of how to construct Fourier
series (see pp. 476–479, pp. 486–487 in Secs. 11.1 and 11.2, respectively). The chapter ends with a brief
review of complex analysis in part D on p. 371 of this Manual.

Sec. 18.1 Electrostatic Fields

We know from electrostatics that the force of attraction between two particles of opposite or the same
charge is governed by Coulomb’s law (12) in Sec. 9.7, p. 401. Furthermore, this force is the gradient of a
function ˆ known as the electrostatic potential. Here we are interested in the electrostatic potential ˆ
because, at any points in the electrostatic field that are free of charge, ˆ is the solution of Laplace’s
equation in 3D:
r2
376 Complex Analysis Part D

ˆ D ˆxx C ˆyy C ˆzz D 0 (see Sec. 12.11, pp. 593–594, pp. 596–598):

Laplace’s equation is so important that the study of its solutions is called potential theory.
Since we want to apply complex analysis to potential theory, we restrict our studies to two dimensions
throughout the entire chapter. Laplace’s equation in 2D becomes

(1) r2ˆ D ˆxx C ˆyy D 0 (see p. 759 in Sec. 18.1 of textbook).

Then the equipotential surfaces ˆ.x;y;z/ D const (from the 3D case) appear as equipotential lines in the xy-
plane (Examples 1–3, pp. 759–760).
The next part of Sec. 18.1 introduces the key idea that it is advantageous to work with complex
potentials instead of just real potentials. The underlying formula for this bold step is

(2) F.z/ D ˆ.x; y/ C i ‰.x; y/,

where F is the complex potential (corresponding to the real potential ˆ) and ‰ is the complex conjugate
potential (uniquely determined except for an additive constant, see p. 629 of Sec. 13.4). The advantages
for using complex potentials F are:

1. It is mathematically easier to solve problems with F in complex analysis because we can use conformal
mappings.

2. Formula (2) has a physical meaning. The curves ‰ D const (“lines of force”) intersect the curves ˆ D
const (“equipotential lines”) at right angles in the xy-plane because of conformality (p. 738).
Illustrations of (2) are given in Examples 4–6 on p. 761 and in Probs. 3 and 15. The section concludes
with the method of superposition (Example 7, pp. 761–762, Prob. 11).

Problem Set 18.1. Page 762

3. Potential between two coaxial cylinders. The first cylinder has radius r 1 D 10 [cm] and potential U1 D
10 [kV]. The second cylinder has radius r2 D 1 [m] D 100 [cm] and potential
U2 D 10 [kV]. From Example 2, p. 759 in Sec. 18.1, we know that the potential ˆ.r/ between two
coaxial cylinders is given by

ˆ.r/ D aln r C b where a and b are to be determined from given boundary conditions.

In our problem we have from the first cylinder with r 1 D 10 and U1 D 10

ˆ.r1/ D ˆ.10/ D aln 10 C b D U1 D 10;

so that
Chap. 18 Complex Analysis and Potential Theory 377

(C1) ˆ.10/ D aln 10 C b D 10:

Similarly, from the second cylinder we have

(C2) ˆ.100/ D aln 100 C b D 10:

We determine a and b: We subtract (C2) from (C1) and use that

(L) aln100 D aln.102/ D 2aln10

to get

ˆ.10/ ˆ.100/ D aln 10 C b .aln 100 C b/

D aln 10 aln100

D aln10 2aln10 [by (L)]

D aln10 [from the r.h.s. of (C1) and (C2)]

D 10 . 10/ D 20 [from the l.h.s. of (C1) and (C2)].

Solving this for a gives

aln10 D 20;

We substitute this into (C1) and get

20 20 C b D 10 or b D 30,
aln 10 C b D ln 10 C b D
ln10

and the real potential is

20
ln r C 30:
ˆ.r/ D aln r C b D ln10
Thus, by Example 5, p. 761, the associated complex potential is

20
F .z/ D 30 Lnz where ˆ.r/ D ReF .z/:
ln10

11. Two source lines. Verification of Example 7, pp. 761–762. The equipotential lines in Example 7, p. 761,
are
378 Complex Analysis Part D

ˇˇz c ˇˇ
ˇ ˇ D k D const .k and c real/: ˇz C cˇ

Hence

jz c j D k j z C c j:

We square both sides and get

(A) jz c j2 D K j z C c j2 where K is a constant (and equal to k2/.

We note that, by (3), p. 613,

jz c j2 D j x Ciy c j2 D j .x c/Ciy j2 D .x c/2 Cy2 and j z Cc j2 D .x Cc/2 Cy2.

Using this, and writing (A) in terms of the real and imaginary parts and taking all the terms to the
left, we obtain

.x c/2 C y2 K Œ.x C c/2 C y2 D 0:

Writing out the squares gives

(B) x2 2c x C c2 C y2 K .x2 C 2c x C c2 C y2/ D 0:

We consider two cases. First, consider k D 1, hence K D 1, most terms in (B) cancel, and we are left
with

4c x D 0 hence x D 0 (because c ¤ 0).

This is the y-axis. Then

j j2 D j C j2 D 2 C 2 jjz C cjj D D z c z c y c ; 1; Ln 1 0:
z c

This shows that the y-axis has potential 0.

We can now continue with (B), assuming that K ¤ 1. Collecting terms in (B), we have

.1 K/.x2 C y2 C c2/ 2c x .1 C K/ D 0:

Division by 1 K (¤ 0 because K ¤ 1) gives

c .1 K/
2 2 2
x Cy Cc 2Lx D 0 where LD C :
Chap. 18 Complex Analysis and Potential Theory 379

1 K

Completing the square in x, we finally obtain

.x L/2 C y2 D L2 c2:

This is a circle with center at L on the real axis and radius pL2 c2. We simplify pL2

c2 as follows. First, we consider

2 2D c .1 C K/2 2

L c c (by inserting L) 1 K

D c 2.1 C K/2 2

c
2
.1 K/
.1 K/2
D c2 C 1
2
.1 K/
D 2 .1 C K/2 .1 K/2 c

.1 K/2 .1 K/2
2
C C K2/
2
1 2K K .1 2K
Dc
.1 K/2
c24K2
:
2
Hence D .1 K/
s 2
2 2
p c 4K c2K 2ck 2
L2 c2 D D D (using K D k ).
.1 K/2 1 K 1 k2

Thus the radius equals 2ck2=.1 k2/:

15. Potential in a sector. To solve the given problem, we note that

z2 D .x C iy/2 D x2 y2 C 2ixy
gives the potential in sectors of opening =2 bounded by the bisecting straight lines of the quadrants
because

x2 y2 D 0 when y D ˙x:
380 Complex Analysis Part D

Similarly, higher powers of z give potentials in sectors of smaller openings on whose boundaries the
potential is zero. For

z3 D .x C iy/3 D x3 C 3ix2y 3xy2 iy3

the real potential is

ˆ0 D Re z3 D x3 3xy2 D x.x2 3y2/

and
x
ˆD0 when yD˙p I
3

these are the boundaries given in the problem, the opening of the sector being =3, that is, 60 ı. To
satisfy the other boundary condition, multiply ˆ0 by 220 [V] and get

ˆ D 220 .x3 3xy2/ D Re.220z3/ [V].

Sec. 18.2 Use of Conformal Mapping. Modeling

Here we experience, for the first time, the full power of applying the geometric approach of conformal
mappings to boundary value problems (“Dirichlet problems,” p. 564, p. 763) in two-dimensional potential
theory. Indeed, we continue to solve problems of electrostatic potentials in a complex setting (2), p. 760
(see Example 1, p. 764, Example 2, p. 765; Probs. 7 and 17). However, now we apply conformal mappings
(defined on p. 738 in Sec. 17.1) with the purpose of simplifying the problem by mapping a given domain
onto one for which the solution is known or can be found more easily. This solution, thus obtained, is
mapped back to the given domain.
Our approach of using conformal mappings is theoretically sound and, if applied properly, will give us
correct answers. Indeed, Theorem 1, p. 763, assures us that if we apply any conformal mapping to a given
harmonic function then the resulting function is still harmonic. [Recall that harmonic functions (p. 460 in
Sec. 10.8) are those functions that are solutions to Laplace’s equation (from Sec. 18.1) and have
continuous second-order partial derivatives.]

Problem Set 18.2. Page 766

7. Mapping by w = sin z. Look at Sec. 17.4, pp. 750–751 (also Prob. 11, p. 754 of textbook and solved on
p. 348 of this Manual) for the conformal mapping by

w D u C iv D sinz D sinx coshy C i cosx sinhy:

Chap. 18 Complex Analysis and Potential Theory 381

We conclude that the lower side (z A to zB) 0 < x < =2 .y D 0/ of the given rectangle D maps onto 0 <
u < 1 .v D 0/ because cosh0 D 1 and sinh0 D 0. The right side (zB to zC) 0 < y < 1 .x D =2/ maps onto
1 < u < cosh=2 .v D 0/. The upper side (zC to zD) maps onto a quarter of the ellipse

u2 v2
1
2 C 2 D
cosh 1 sinh 1

in the first quadrant of the w-plane. Finally, the left side (z D to zA) maps onto sinh1 > v > 0 .u D
0/.

Now the given potential is

ˆ .u;v/ D u2 v2

D sin2 x cosh2 y cos2 x sinh2 y:

Hence ˆ D sin2 x on the lower side .y D 0/, and grows from 0 to 1. On the right side, ˆ D cosh2 y, which

increases from 1 to cosh2 1.

On the upper side we have the potential

ˆ D sin2 x cosh2 1 cos2 x sinh2 1;

which begins with the value cosh2 1 and decreases to sinh2 1. Finally, on the left side it begins with
sinh2 1 and returns to its value 0 at the origin. Note

that any y D c maps onto an ellipse

u2 v2
1
cosh c C sinh c D
2 2

and any x D k maps onto an hyperbola

u2 v2
1:
sin2 k cos2 k D
382 Complex Analysis Part D

v
D
1

y
1
D C

A B 0 u
1
0 x
π/2 A B C z-plane w-plane

Sec. 18.2 Prob. 7. Given region and image under conformal mapping w D sinz

17. Linear fractional transformation. We want to find a linear fractional transformation

(LTF) z D g.Z/ that maps jZj 1 onto jzj 1 with Z D i=2 being mapped onto z D 0: Our task is to find an
LTF with such properties. The candidate is the LTF defined by (3) on p. 749 in Sec. 17.3. Here Z plays
the role of z in (3), and z plays the role of w. Thus,

(A) zD Zi 2i :

2Z 1
We can multiply both the numerator and denominator in (A) by 2 and get the answer on p. A43 in
App. 2:

2Z i
(A2) zD :
iZ 2

To complete the problem, we evaluate (A2) with Z D 0:6 C 0:8i and 0:6 C 0:8i, respectively.
We get for Z D 0:6 C 0:8i

D 2Z i D 2.0:6 CC0:8i/ i D 1:2 CC1:6i i D 1:2 CC0:6i D

(B) z1,
iZ 2 i.0:6 0:8i/ 2 0:6i 0:8 2 .0:6i 1:2/

which is the desired value. Similarly, you can show that for Z D 0:6 0:8i, one gets z D 1: Thus

p
jZj D j0:6 ˙ 0:8ij D 0:62 C 0:82 D 1 1;
Chap. 18 Complex Analysis and Potential Theory 383

which means that jZj 1. And (B), with a similar calculation, shows that our chosen Z’s get mapped by
(A2) onto z D ˙1, so that indeed jzj 1: Together, this shows that (A2) is the desired LTF as described in
Prob. 17 and illustrated in Fig. 407, p. 766. Convince yourself that Fig. 407 is correct.

Sec. 18.3 Heat Problems

Complex analysis can model two-dimensional heat problems that are independent of time. From the top
of p. 564 in Sec.12.6, we know that the heat equation is

(H) Tt D c2r2T .

We assume that the heat flow is independent of time (“steady”), which means that T t D 0: Hence (H)
reduces to Laplace’s equation

(1) r2T D Txx C Tyy D 0:

This allows us to introduce methods of complex analysis because T [or T.x;y/] is the real part of the
complex heat potential

F.z/ D T.x;y/ C i ‰.x;y/:

[Terminology: T.x;y/ is called the heat potential, ‰.x;y/ D const are called heat flow lines, and T.x;y/ D
const are called isotherms.]
It follows that we can reinterpret all the examples of Secs. 18.1 and 18.2 in electrostatics as problems of
heat flow (p. 767). This is another great illustration of Underlying Theme 3 on p. ix of the textbook of the
powerful unifying principles of engineering mathematics.

Problem Set 18.3. Page 769

7. Temperature in thin metal plate. A potential in a sector (in an angular region) whose sides are kept at
constant temperatures is of the form

T.x;y/ D a C b
y
(A) D aarctan Cb
x
D a Argz C b (see similar Example 3 on pp. 768–769):

Here we use the fact that

Argz D D Im.Lnz/ is a harmonic function.

384 Complex Analysis Part D

The two constants, a and b; can be determined from the given values on the two sides Argz D 0 and
Argz D =2. Namely, for Argz D 0 (the x-axis) we have

T D b D T1:

Then for Argz D =2 we have

TDa C T1 D T2:
2

Solving for a gives

2.T2 T1/ a D
:

Hence a potential giving the required values on the two sides is

2.T2 T1/
T.x;y/ D Argz C T1:

Complete the problem by finding the associated complex potential F.z/ obeying ReF.z/ D
T.x;y/ and check on p. A43 in App. 2 of the textbook.

15. Temperature in thin metal plate with portion of boundary insulated. Mixed boundary value problem.
We start as in Prob. 7 by noting that a potential in an angular region whose sides are kept at
constant temperatures is of the form

(B) T.x;y/ D a Argz C b;

and using the fact that Argz D D Im.Lnz/ is a harmonic function. We determine the values for the
two constants a and b from the given values on the two sides Argz D 0 and Argz D =4. For Argz D 0
(the x-axis) we have T D b D 20 and for Argz D =4 we have

320
TDa 20 D 60 so that aD :
4
Hence a potential that satisfies the conditions of the problem is

320
(C) TD Argz 20:
Chap. 18 Complex Analysis and Potential Theory 385

Now comes an important observation. The curved portion of the boundary (a circular arc) is
insulated. Hence, on this arc, the normal derivative of the temperature T must be zero. But the
normal direction is the radial direction; so the partial derivative with respect to r must vanish. Now
formula (C) shows that T is independent of r, that is, the condition under discussion is automatically
satisfied. (If this were not the case, the whole solution would not be valid.) Finally we derive the
complex potential F. From Sec. 13.7 we recall that

(D) Lnz D ln j z j C i Argz [by (2), p. 637]:

Hence for Argz to become the real part (as it must be the case because F D T C i ‰), we must
multiply both sides of (D) by i. Indeed, then

i Lnz D i ln j z j C Argz:

Hence from this and (C) we see that the desired complex heat potential is

320
(E) F.z/ D 20 C. i Lnz/

320
D 20i Lnz;
which, by (C) and (E), leads to the answer given on p. A43 in App. 2 of the textbook.

Sec. 18.4 Fluid Flow

The central formula of Sec. 18.4 is on p. 771:

(3) V D V1 C i V2 D F 0.z/:

It derives its importance from relating the complex velocity vector of the fluid flow

(1) V D V1 C iV2

to the complex potential of the fluid flow

(2) F.z/ D ˆ.x; y/ C i ‰.x; y/,

whose imaginary part ‰ gives the streamlines of the flow in the form

‰.x; y/ D const:

Similarly, the real part ˆ gives the equipotential lines of the flow:

ˆ.x; y/ D const:
386 Complex Analysis Part D

The use of (3), p. 771, is illustrated in different flows in Example 1 (“flow around a corner,” p. 772), Prob.
7 (“parallel flow”) and in Example 2, and Prob. 15 (“flow around a cylinder”).
Flows may be compressible or incompressible, rotational or irrotational, or may differ by other general
properties. We reach the connection to complex analysis, that is, Laplace’s equation (5) applied to ˆ and
‰ of (2), written out

(5) r2ˆ D ˆxx C ˆyy D 0; r2‰ D ‰xx C ‰yy D 0 [on p. 772]

by first assuming the flow to be incompressible and irrotational (see Theorem 1, p. 773).

Rotational flows can be modeled to some extent by complex logarithms, as shown in the textbook on
pp. 776–777 in the context of a Team Project.

We encounter a third illustration in Chap. 18 of Underlying Theme 3 of the textbook on p. ix because

the model (2), p. 760, developed for electrostatic potentials is now the model for fluid flow. More details
are given in the paragraph on “basic comment on modeling” on p. 766 in Sec. 18.2.

Problem Set 18.4. Page 776

7. Parallel flow. Our task is to interpret the flow with complex potential F.z/ D z:

We start by noting that a flow is completely determined by its complex potential

(2) F.z/ D ˆ.x; y/ C i ‰.x; y/ (p. 771).

The stream function ‰ gives the streamlines ‰ D const and is generally more important than the
velocity potential ˆ, which gives the equipotential lines ˆ D const. The flow can best be visualized in
terms of the velocity vector V , which is obtained from the complex potential in the form (3), p. 771,

(3) V D V1 C i V2 D F 0.z/:

(We need a special vector notation, in this case, because a complex function V can always be
regarded as a vector function with components V 1 and V2.) Hence,
for the given complex potential

(A) F.z/ D z D x C iy;

we have

F 0.z/ D 1 and F 0.z/ D 1 C 0iI

thus,

(B) V D V1 D 1 and V2 D 0:
Chap. 18 Complex Analysis and Potential Theory 387

The velocity vector in (B) is parallel to the x-axis and is positive, i.e., V D V 1 points to the right (in the
positive x-direction).
Hence we are dealing with a uniform flow (a flow of constant velocity) that is parallel (the
streamlines are straight lines parallel to the x-axis) and is flowing to the right (because V is positive).
From (A) we see that the equipotential lines are vertical parallel straight lines; indeed,

ˆ.x;y/ D Re F.z/ D x D constI hence x D const:

Using our discussion, sketch the flow.

15. Flow around a cylinder. Here we are asked to change F.z/ in Example 2, p. 772, slightly to obtain a
flow around a cylinder of radius r0 that gives the flow in Example 2 if r0 ! 1:

Solution. Since a cylinder of radius r0 is obtained from a cylinder of radius 1 by a dilatation (a

uniform stretch or contraction in all directions in the complex plane), it is natural to replace z by az
with a real constant a because this corresponds to such a stretch. That is, we replace the complex
potential 1 z C
z
in Example 2, p. 772, by

F.z/ D ˆ.r;/ C i ‰.r;/ 1

D az C
az
i i
1
D are C e [by (6), p. 631 applied to both terms]:
ar
The stream function ‰ is the imaginary part of F. Since, by Euler’s formula,

e˙i D cos ˙ i sin [by (5), p. 631 in Sec. 13.5]

we obtain

‰.r;/ D Im.F/

i 1 i D Im are C e ar

1
D Im ar .cos C i sin / C .cos i sin/ (by Euler’s formula applied twice) ar

1 1
D Im ar cos C cos C ar i sin i sin (regrouping for imaginary part) ar
ar
388 Complex Analysis Part D

1
D ar sin sin ar
1
D ar sin:
ar

The streamlines are the curves D const. As in Example 2 of the text, the streamline ‰ D 0 consists of
the x-axis . D 0 and ), where sin D 0; and of the locus where the other factor of ‰
is zero, that is,

1 2 1 ar D 0; thus .ar/ D 1 or a D : ar r

Since we were given that the cylinder has radius r D r 0, we must have

1aD
: r0

With this, we obtain the answer

1 z r0
F.z/ D az C D C : az r0 z
Sec. 18.5 Poisson’s Integral Formula for Potentials
The beauty of this section is that it brings together various material from complex analysis and Fourier
analysis. The section applies Cauchy’s integral formula (1), p. 778 (see Theorem 1, p. 660 in Sec. 14.3), to
a complex potential F.z/ and uses it on p. 778 to derive Poisson’s integral formula (5), p. 779.
Take a look at pp. 779–780. Formula (5) yields the potential in a disk D. Ordinarily such a disk has a
continuous boundary jzj D R, which is a circle. However, this requirement can be loosened: (5) is
applicable even if the boundary is only piecewise continuous, such as in Figs. 405 and 406 of a typical
example of a potential between two semicircular plates (Example 2 on p. 765).
From (5) we obtain the potential in a region R by mapping R conformally onto D, solving the problem in
D, and then using the mapping to obtain the potential in R. The latter is given by the important formula
(7), p. 780,

X1 r n
(7) ˆ.r; / D a0 C .an cosnx C bn sinnx/
R nD1

where, typically, we consider the potential over the disk r < R:

In Example 1, p. 780, and Probs. 5–18 (with Probs. 7 and 13 solved below) the potential ˆ.r; / in the unit
disk is calculated. In particular note, for the unit disk r < 1 and given boundary function ˆ.1; /, we have
that

r n r n r D R .D 1/ so
that in (7) D D1
Chap. 18 Complex Analysis and Potential Theory 389

R r
and (7) simplifies to a genuine Fourier series:

X
1
0
(7 ) ˆ.r; / D a0 C .an cos n C bn sin n/:
nD1

To determine (70) requires that we compute the Fourier coefficients of (7) by (8), p. 780, under the
simplification of r D R. Hence the techniques of calculating Fourier series explained in Sec. 11.1, pp.
474–483 of the textbook and pp. 202–208 of Vol. 1 of this Manual, and furthermore, in Sec. 11.2, pp.
483–491 of the textbook and pp. 208–211 of Vol. 1 of this Manual come into play. This is illustrated in
Prob. 13 and Example 1.

Problem Set 18.5. Page 781

7–19. Harmonic functions in a disk. In each of 7–19 Harmonic functions in a disk. In each of Probs. 7–19
we are given a boundary function ˆ.1; /: Then, using (7), p. 780, and related formula (8), we want
to find the potential ˆ.r; / in the open unit disk r < 1 and compute some values of ˆ.r; / as well as
sketch the equipotential lines. We note that, typically, these problems are solved by Fourier series
as explained above.

7. Sinusoidal boundary values lead to a series (7) that, in this problem, reduces to finitely many terms
(a “trigonometric polynomial”). The given boundary function

ˆ.1; / D a cos2 4

is not immediately one of the terms in (7), but we can express it in terms of a cosine function of
multiple angle as follows. Indeed, in App. 3, p. A64 of the textbook, we read

(10) cos2 x D C cos2x:

In (10), we set

xD4

and get

cos2 4 D C cos8:

Hence we can write the boundary function as

ˆ.1; / D a cos2 4

1 1 a a
390 Complex Analysis Part D

Da C cos8 D C cos8:
2 2 2 2

From (7) we now see immedately that the potential in the unit disk satisfying the given boundary
condition is
a a8
ˆ.r; / D C r cos8:
2 2
Note that the answer is already in the desired form so we do not need to calculate the Fourier
coefficients by (8)!

13. Piecewise linear boundary values given by

ˆ.1; /
D < <
2 2
0
otherwise
lead to a series (7) whose coefficients are given by (8).
We follow Example 1, p. 780, and calculate the Fourier coefficients (8). Because the function
ˆ.1; / is an odd function (see p. 486), we know that its Fourier series reduce to a Fourier sine series
so that all the an D 0: From (8) we obtain

2 Z =2
bn D sinn d n D 1;2;3;:::
0

2 sinn n cosn =2

D n2 0

2 sin n n cos n sinn0 n0cosn0

D n2 n2

1 1
D 2sin 2n n cos 2n :
2
n

For n D 1 this simplifies to

2sin 1 cos 1 2 1 0 2 b1 D 2 2 D D :

For n D 2; 3; 4;:::, we get the following values for the Fourier coefficients:
Chap. 18 Complex Analysis and Potential Theory 391

2sin
1/ D 2 D 1

DD

Observe that in computing bn for n odd, the cos terms are zero, while for n even, the sin terms are zero.
Hence putting it together

2 1 2 1
ˆ.1; / D sin C sin2 sin3 sin4 C C :
2 9 4

From this, we obtain the potential (7) in the disk (R D 1) in the form

2 12 2 3 14
(A) ˆ.r; / D r sin C r sin2 r sin3 r sin4 C C :
2 9 4
The following figure shows the given boundary potential (straight line), an approximation of it [the sum
of the first, first two, first three, and first four terms (dot dash) of the series (A) with r D 1] along with an
approximation of the potential on the circle of radius r D (the sum of those four terms for r D drawn
with a long dash). Make a sketch of the disk (a circle) and indicate the boundary values around the circle.
392 Complex Analysis Part D

y
1.5

0.5

x
–π –3 π –π –π 0 π π 3π π
2 8 4 8 8 4 8 2

–0.5

–1

–1.5

Sec. 18.5 Prob. 13. Boundary potential and approximations for r D 1 and r D

Sec. 18.6 General Properties of Harmonic Functions.

Uniqueness Theorem for the Dirchlet Problem

Recall three concepts (needed in this section): analytic functions (p. 623) are functions that are defined
and differentiable at every point in a domain D: Furthermore, one is able to test whether a function is
analytic by the two very important Cauchy–Riemann equations on p. 625. Harmonic functions (p. 460) are
functions that are solutions to Laplace’s equation r2ˆ D 0 and their second-order partial derivatives are
continuous. Finally, a Dirichlet problem (p. 564) is a boundary value problem where the values of the
function are prescribed (given) along the boundary.
The material is very accessible and needs some understanding of how to evaluate double integrals and
also apply Cauchy’s integral formula (Sec. 14.3, p. 660). We derive general properties of harmonic
functions from analytic functions. Indeed, the first two mean value theorems go together, in that Theorem
1, (p. 781; Prob. 3) is for analytic functions and leads directly to Theorem 2 (p. 782; Prob. 7) for harmonic
functions. Similarly, Theorems 3 and 4 are related to each other. Of the general properties of harmonic
functions, the maximum principle of Theorem 4, p. 783, is quite important. The chapter ends on a high
note with Theorem 5, p. 784, which states that an existing solution to a Dirichlet problem for the 2D
Laplace equation must be unique.

Orientation. We have reached the end of Part D on complex analysis, a field whose diversity of topics
and richness of ideas may represent a challenge to the student. Thus we include, for study purposes, a
brief review of complex analysis on p. 371 of this Manual.
Chap. 18 Complex Analysis and Potential Theory 393

Problem Set 18.6. Page 784

3. Mean value of an analytic function. Verification of Theorem 1, p. 781, for given problem. The
problem is to verify that Theorem 1, p. 781, holds for

(A) F.z/ D .3z 2/2; z0 D 4; jz 4 j D 1:

Solution. We have to verify that

1Z2 i˛

(2) F.z0/ D F.z0 C re /d˛

2 0

holds for (A). Here we integrate F.z/ D .3z 2/2 around the circle, j z 4 j D 1; of radius r D 1 and center

z0 D 4, and hence we have to verify (2) with these values. This means we have to show

that

1Z2 i˛ 1Z2 i˛

(2*) F.z0/ D F.4/ D F.z0 C re /d˛ D F.4 C 1 e /d˛:

2 0 2 0

Since

F.z0/ D F.4/ D .3 4 2/2 D .10/2 D 100;

we have to show that the integral on the right-hand side of (2*) takes on that value of 100, that is,
we must show that

1Z2
(2**) F.4 C 1 ei ˛/d˛ D 100:
2 0
We go in a stepwise fashion. The path of integration is the circle, j z 4 j D 1, so that

z D z0 C rei ˛ D 4 C 1 ei ˛ D 4 C ei ˛:

Hence, on this path, the integrand is

F.z0 C ei ˛/ D 34 C ei ˛ 22 D 12 C 3ei ˛ 22

D 10 C 3ei ˛2

D 100 C 60ei ˛ C 9e2i ˛:

394 Complex Analysis Part D

Indefinite integration over ˛ gives

Z Z
i˛ i˛ 2i ˛
F.4 C 1 e /d˛ D .100 C 60e C 9e /d˛

Z Z Z
i˛
D 100 d˛ C 60 e d˛ C 9 e2i ˛ d˛

1 i˛ 9 2i ˛
D 100˛ C 60 e C e
:i 2i
Next we consider the definite integral

Z2 i˛ 1 i˛ 9 2i ˛˛D2
F.4 C 1 e /d˛ D 100˛ C 60 e C e :
1 i 2i ˛D0

At the upper limit of 2 this integral evaluates to

2 i2 9 2i2 60 9
129
100 .2/ C 60 eC e D 200 C C D 200 C :i 2i i
2i 2i

At the lower limit of 0 it evaluates to

10 9 0 129
0 C 60 e C e D : i 2i 2i
Hence the difference between the value at the upper limit and the value at the lower limit is

Z2
i˛ 129 129
F.4 C 1 e /d˛ D 200 C D 200:
1 2i 2i

The integral in (2**) has a factor 1=.2/ in front, so that we put that factor in front of the last
integral and obtain

2 Z2 i˛ 1
F.4 C 1 e /d˛ D 200 D 100 where 100 D F.4/:
2 0 2
Chap. 18 Complex Analysis and Potential Theory 395

Thus we have shown that (2**) holds and thereby verified Theorem 1 for (A).

7. Mean values of harmonic functions. Verification of Theorem 2, p. 782. Our problem is similar in spirit
to that of Prob. 3 in that it requires us to verify another mean value theorem for a given example—
here for a harmonic function. Turn to p. 782 and look at the two formulas [one with no number, one
numbered (3)] in the proof of Theorem 2. We shall verify them for given function ˆ defined on a point
.x ;y
0 / and a circle. To get a better familiarity of the material, you may want to write down all the
0

details of the solution with the integrals, as we did in Prob. 3. We verify Theorem 2 for

(B) ˆ.x; y/ D .x 1/.y 1/; .x0; y0/ D .2; 2/; z D 2

i˛
2i C e :

The function ˆ.x;y/ is indeed harmonic (for definition, see pp. 628 and 758–759). You should verify this
by differentiation, that is, by showing that ˆ is a solution of
r2
ˆ D ˆxx C ˆyy D 0 [(1), p. 759].

We continue.
We note that z0 D x0 C iy0 D 2 2i is the center of the circle in (B).

In terms of the real and imaginary parts of the path, 2 2i C ei ˛; is then [by Euler’s formula (5), p. 634 in
Sec. 13.6]

(C) x D 2 C cos ˛; y D 2 C sin ˛:

This is the representation we need, since ˆ is a real function of the two real variables x and y. We see that

ˆ.z0; y0/ D ˆ.2; 2/

D .x0 1/.y0 1/x0D2; y0D 2

D .2 1/. 2 1/ D 3:

Hence we have to show that each of the two mean values equals 3.
Substituting (C) into (B) (which is a completely schematic process) gives

ˆ.2 C cos ˛; 2 C sin ˛/ D .2 C cos ˛ 1/. 2 C sin ˛ 1/

(D) D .1 C cos ˛/. 3 C sin ˛/

D 3 C 1 sin ˛ 2 cos ˛ C cos ˛ sin ˛:

Consider the mean value over the circle. Now

396 Complex Analysis Part D

Z2
. 3 C 1 sin ˛ 2 cos ˛ C cos ˛ sin ˛/ d˛
0
2
3˛ cos ˛ 2sin ˛ C sin2 ˛
D
0
!
6 D cos 22sin 2 Csin2 2 0 cos0 2 sin0 Csin2 0

„ƒ‚… „ƒ‚…„ƒ‚… „ƒ‚… „ƒ‚…„ƒ‚…

1 00 1 00

D . 6 1/ . 1/ D 6:

We have to multiply this result by a factor 1=.2/: (This is the factor in front of the unnumbered
formula of the first integral in the proof of Theorem 2.) Doing so we get

1
. 6/ D 3:
2

This is the mean value of the given harmonic function over the circle considered and completes the
verification of the first part of the theorem for our given data.
Next we work on (3), p. 782. Now calculate the mean value over the disk of radius 1 and center
(2; 2). The integrand of the double integral in formula (3) in the proof of Theorem 2 is similar
to that in (D). However, in (D) we had r D 1 (the circle over which we integrated), whereas now we
have r being variable and we integrate over it from 0 to 1: In addition we have a factor r resulting
from the element of area in polar coordinates, which is r dr d. Hence, instead of .1 C cos˛/. 3 C sin˛/
in (D), we now have

.1 C r cos˛/. 3 C r sin˛/r D 3r C 1r2 sin˛ 2r2 cos˛ C r3 cos˛ sin˛:

The factors of r have no influence on the integration over ˛ from 0 to 2 so

Z2
3r C 1r2 sin˛ 2r2 cos˛ C r3 cos˛ sin˛ d˛
1 ˛D2
0 2r2 sin˛ C r3 sin2 ˛

D 3r˛ r2 cos˛ 2 ˛D0

6r 0 !
1 2r sin2 C
2
1 2r sin0
2

D r3sin2 2 C r3sin2 0
r2cos2 r2 cos0
„ƒ‚… „ƒ‚… 2 „ƒ‚… „ƒ‚… „ƒ‚… 2 „ƒ‚…
1 0 0 1 0 0
Chap. 18 Complex Analysis and Potential Theory 397

D 6r r2 r2

D 6r:
Hence
Z1 Z1 hr i 1 1
6r dr D 6 r dr D 6 D 6 D 3:
0
0 0 2 2

In front of the double integral we have the factor 1= r 02 D 1= because the circle of integration has

radius 1. Hence our second result is 3= D 3. This completes the verification.

Remark. The problem requires you to only verify (3). We also verified the first formula in the
proof of Theorem 2 to give you a more complete illustration of the theorem.

19. Location of maxima of a harmonic function and its conjugate: The question is whether a harmonic
function ˆ and a harmonic conjugate ‰ in a region R have their maximum at the same point of R.
The answer is “not in general.” We look for a counterexample that is as simple as possible. For
example, a simple case would be the conjugate harmonics ‰:
x D Re z and y D Im z in the square 0 x 1; 0 y 1:

Then we have

max x D 1 at all points on the right boundary

and

max y D 1 at all points of the upper boundary.

Hence in this case there is a point

.1;1/, that is, z D 1 C i;

where both functions ˆ and ‰ have a maximum. But, if we leave out that point .1;1/ in the square
and consider only

the region RW 0 x < 1; 0 y < 1;

then

max x and max y cannot occur at the same point.

You may want to investigate the question further. What about a triangle, a square with vertices ˙1;
˙i, and so on?
398 Complex Analysis Part D

Brief Review of Part D on Complex Analysis

Since complex analysis is a rather diverse area, we include this brief review of the essential ideas of
complex analysis. Our main point is that to get a good grasp of the field, keep the three approaches
(methods) of complex analysis apart and firmly planted in your mind. This is in tune with Underlying
Theme 4 of “Clearly identifying the conceptual structure of subject matter” on p. x of the textbook. The
three approaches were [with particularly important sections marked in boldface, page references given
for the Textbook (T) and this Manual (M)]:

1. Evaluating integrals by Cauchy’s integral formula [see Sec. 14.3, p. 660 (T), p. 291 (M); general
background Chap. 13, p. 608 (T), p. 257 (M), and Chap. 14, p. 643 (T), p. 283 (M)]. The method required
a basic understanding of analytic functions [p. 623 (T), p. 267 (M)], the Cauchy–Riemann equations [p.
625 (T), p. 269 (M)], and Cauchy’s integral theorem [p. 653 (T), p. 288 (M)].

2. Residue integration [applied to complex integrals see Sec. 16.3, p. 719 (T), p. 322 (M); applied to real
integrals see Sec. 16.4, p. 725 (T), p. 326 (M); general background Chap. 15, p. 671 (T), p. 298 (M), and
Chap. 16, p. 708 (T), p. 316 (M)]. The method needed a basic understanding of radius of convergence of
power series and the Cauchy–Hadamard formula [p. 683 (T), p. 303 (M)] and Taylor series p. 690 (T), p.
309 (M). This led to the very important Laurent series [which allowed negative powers, p. 709 (T), p.
316 (M)] and gave us order of singularities, poles, and zeros [p. 717 (T), p. 320 (M)].

3. Geometric approach of conformal mapping applied to potential theory [in electrostatic fields Sec. 18.1,
p. 759 (T), p. 353 (M); Sec. 18.2, p. 763 (T), p. 357 (M); Sec. 18.5, p. 777 (T), p. 364 (M), in heat
conduction, Sec. 18.3, p. 767 (T), p. 359 (M), in fluid flow in Sec. 18.4, p. 771 (T), p. 361 (M); general
background in Chap. 17, p. 736 (T), p. 332 (M)]. The method required an understanding of conformal
mapping [p. 738 (T), p. 333 (M)], linear fractional transformations [p. 743 (T), p. 339 (M)], and their
fixed points [pp. 745, 746 (T), pp. 339, 341 (M)], and a practical understanding of how to apply
conformal mappings to basic complex functions.
399 Complex Analysis Part D

In general, just like in regular calculus, you have to know basic complex functions (sine, cosine,
exponential, logarithm, power function, etc.) and know how they are different from their real
counterparts. You have to know Euler’s formula [(5), p. 634 (T), p. 277 (M)] and Laplace’s equation
[Theorem 3, p. 628 (T), p. 269 (M)].
400 Numeric Analysis Part E

PART E
Numeric
Analysis

Numeric analysis in Part E (also known as numerics or numerical analysis) is an area rich in applications
that include modeling chemical or biological processes, planning ecologically sound heating systems,
determining trajectories of satellites and spacecrafts, and many others. Indeed, in your career as an
engineer, physicist, applied mathematician, or in another field, it is likely that you will encounter projects
that will require the use of some numerical methods, with the help of some software or CAS (computer
algebra system), to solve a problem by generating results in terms of tables of numbers or figures.
The study of numeric analysis completes your prior studies in the sense that a lot of the material you
learned before from a more algebraic perspective is now presented again from a numeric perspective. At
first, we familiarize you with general concepts needed throughout numerics (floating point, roundoff,
stability, algorithm, errors, etc.) and with general tasks (solution of equations, interpolation, numeric
integration and differentiation) in Chap. 19. Then we continue with numerics for linear systems of
equations and eigenvalue problems for matrices in Chap. 20—material previously presented in an
algebraic fashion in Chaps. 7 and 8. Finally, in Chap. 21 we discuss numerical methods for differential
equations (ODEs and PDEs) and thus related to Part A and Chap. 12.
Use of Technology. We have listed on pp. 788–789 software, computer algebra systems (CASs),
programmable graphic calculators, computer guides, etc. In particular, note the Mathematica Computer
Guide and Maple Computer Guide (for stepwise guidance on how to solve problems by writing programs
for these two CASs) by Kreyszig and Norminton that accompany the textbook (see p. 789). However, the
problems in the problem sets in the textbook can be solved by a simple calculator, perhaps with some
graphing capabilities, except for the CAS projects, CAS experiments, or CAS problems (see Remark on
Software Use on p. 788 of textbook).

Chap. 19 Numerics in General

This chapter has five sections, the first on general concepts needed throughout numerics and the other
four on three basic areas, namely, solution of equations (Sec. 19.2), interpolation (Secs. 19.3 and 19.4),
and numeric integration and differentiation (Sec. 19.5).
 In this chapter you should also obtain a feel for the spirit, viewpoint, and nature of numerics. You will
notice that numeric analysis has a flavor distinct from that of calculus.
A convenient framework on how to solve numeric problems consists of five steps:

1. Modeling the problem

2. Selecting a numeric method

3. Programming

4. Doing the computation 5. Interpreting the results

as shown on p. 791. Solving a single equation of the form f.x/ D 0; as shown in Sec. 19.2, may serve as one
of many illustrations.
From calculus, you should review Taylor series [in formula (1), p. 690, replace complex z with real x],
limits and convergence (see pp. 671–672), and, for Sec. 19.5, review, from calculus, the basics of how one
developed, geometrically, the Riemann integral.

Sec. 19.1 Introduction

This section introduces some of the general ideas and concepts that underlie all of numerics. As such it
touches upon a fair amount of material in a concise fashion. Upon reading it for the first time, the
material of Sec. 19.1 may seem rather abstract to you, however, with further studies of numerics, it will
take on concrete meaning. For example, the concepts of algorithm and stability (p. 796 of textbook) are
explained here in Sec. 19.1 but illustrated in subsequent sections. Overall, Sec. 19.1 can be thought of as
a reference section for Part E. Hence, once in a while, it may be useful for you to refer back to this section.
Your numeric calculations require that you do computations to a certain amount of precision. It is here
that rounding (p. 792 of the textbook) comes into play. Take a look at the roundoff rule at the bottom of p.
792 and at Example 1 at the top of p. 793. The concept of rounding uses the definition of decimals on
p. 791 in a fixed-point system. Note that when counting decimals, only the numbers after the decimal
point are counted, that is,

78:94599; 0:98700; 10000:00000 all have 5 decimals, abbreviated 5D.

Make sure that you understand Example 1. Here is a self-test. (a) Round the number 1:23454621 to seven
decimals, abbreviated (7D). (b) Round the number 398:723555 to four decimals (4D). Please close
this Student Solutions Manual (!). Check the answer on p. 27 of the Manual. If your answer is correct,
great.
If not, go over your answers and study Example 1 again.
The standard decimal system is not very useful for scientific computation and so we introduce the
floating-point system on p. 791. We have

D 624:70:6247103I 0:00000000000001735 D 1735

10 17 D 0:1735 10 13
I
402 Numeric Analysis Part E

„ ƒ‚ …
13 zeros

0:02000 D 0:2000 10 1 ,

where the underscored number is in floating-point form.

Each of these floating-point numbers above has four significant digits, also denoted by 4S. The digits
are “significant” in the sense that they convey numerical information and are not just placeholders of
zeros that fix the position of the decimal points, whose positions could also be achieved by multiplication
of suitable powers of 10n, respectively. This leads to our next topic of rounding with significant digits.
Chap. 19 Numerics in General 3

The roundoff rule for significant digits is as follows. To round a number x to k significant digits, do the
following three steps:

1. Express the given number as a floating-point number:

x D ˙m 10n; 0:1 jmj < 1; where n is an integer [see also (1), p. 792].

Note that here m can have any number of digits.

2. For now, ignore the factor 10n. Apply the roundoff rule (for decimals) on p. 792 to m only.

3. Take the result from step 2 and multiply it by 10 n: This gives us the desired number x rounded to k
significant digits.

Self-test: Apply the roundoff rule for significant digits to round 102:89565 to six significant digits (6S).
Check your result on p. 27.
The computations in numerics of unknown quantities are approximations, that is, they are not exact but
involve errors (p. 794). Rounding, as just discussed by the roundoff rule, produces roundoff errors
bounded by (3), p. 793. To gain accuracy in calculations that involve rounding, one may carry extra digits
called guarding digits (p. 793). Severe problems in calculations may involve the loss of significant digits
that can occur when we subtract two numbers of about the same size as shown in Example 2 on pp. 793–
794 and in Problem 9.
We also distinguish between error, defined by (6) and (6*) and relative error (7) and (7 0), p. 794,
respectively. The error is defined in the relationship

True value D Approximation C Error:

The relative error is defined by

Error
Relative error D (where True value ¤ 0).
True value

As one continues to compute over many steps, errors tend to get worse, that is, they propagate. In
particular, bounds for errors add under addition and subtraction and bounds for relative errors add under
multiplication and division (see Theorem 1, p. 795).
Other concepts to consider are underflow, overflow (p. 792), basic error principle, and algorithm (p.
796). Most important is the concept of stability because we want algorithms to be stable in that small
changes in our initial data should only cause small changes in our final results.

Remark on calculations and exam. Your answers may vary slightly in some later digits from the answers
given here and those in App. 2 of the textbook. You may have used different order of calculations,
rounding, technology, etc. Also, for the exam, ask your professor what technology is allowed and be
familiar with the use and the capabilities of that technology as it may save you valuable time on the exam
and give you a better grade. It may also be a good idea, for practice, to use the same technology for your
homework.
4 Numeric Analysis Part E

Problem Set 19.1. Page 796

9. Quadratic equation. We want to solve the quadratic equation x 2 30x C 1 D 0 in two different
ways—first with 4S accuracy and then with 2S accuracy. (a) 4S.
First, we use the well-known quadratic formula
1 p 1 p
(4) x1 D bC b2 4ac ; x2 D b b2
4ac
2a 2a
with

a D 1; bD 30; c D 1:

We get, for the square root term calculated with 4S (“significant digits,” see pp. 791–792),

p
. 30/2 4 D p900 4 D p896 D 29:933 D 29:93:

Hence, computing x1 and x2 rounded to four significant digits, i.e., 4S,

x1 D .30 C 29:93/ D 59:93 D 29:965 D 29:97

and

x2 D .30 29:93/ D 0:07 D 0:035:

It is important to notice that x2, obtained from 4S values, is just 2S—i.e., we have lost two digits. As an
alternative method of solution for x2, use (5), p. 794,

1 D1 Cp2 2 D c
(5) x b b 4ac ; x :
2a ax1

The root x1 (where the similar size numbers are added) equals 29:97, as before. For x 2, you now obtain
c 1 x2 D 1 D D 0:0333667 D 0:03337 (to four significant digits): ax 29:97

(b) 2S. With 2S the calculations are as follows. We have to calculate the square root as

p
. 30/2 4 D p900 4 D p899:6 D p900 D 30 (to two significant digits, i.e., 2S).

Hence, by (4),

x1 D .30 C 30/ D 60 D 30

and
Chap. 19 Numerics in General 5

x2 D .30 30/ D 0:

In contrast, from (5), you obtain better results for the second root. We still have x 1 D 30 but

1 1 x2 D 1 D D 0:033333 D 0:033 (to two significant digits). x

30

Purpose of Prob. 9. The point of this and similar examples and problems is not to show that calculations
with fewer significant digits generally give inferior results (this is fairly plain, although not always the
case). The point is to show, in terms of simple numbers, what will happen in principle, regardless of the
number of digits used in a calculation. Here, formula (4) illustrates the loss of significant digits, easily
recognizable when we work with pencil (or calculator) and paper, but difficult to spot in a long
computation in which only a few (if any) intermediate results are printed out. This explains the necessity
of developing programs that are virtually free of possible cancellation effects.
19. We obtain the Maclaurin series for the exponential function by (12), p. 694, of the textbook where
we replace z, a complex number, by u a real number. [For those familiar with complex numbers, note
that (12) holds for any complex number z D x C iy and so in particular for z D x C iy D x C i 0 D x D
Rez, thereby justifying the use of (12)! Or consult your old calculus book. Or compute it yourself.]
Anyhow, we have

x x2 x3 x4 x5 x10
0) f.x/ D ex D 1 C C C C C C C C :
(12
1Š 2Š 3Š 4Š 5Š 10Š

[All computations to six digits (6S).] We are given that the exact 6S value of 1=e is

1
(6S) D 0:36787
e 9

(a) For e 1; the Maclaurin series (120) with x D 1 becomes

1 . 1/ . 1/2 . 1/3 . 1/4 . 1/5 . 1/10

(B) f. 1/ D e D1C C C C C C C C
1Š 2Š 3Š 4Š 5Š 10Š
1 1 1 1 1 1
1 :
D C C C C C
1Š 2Š 3Š 4Š 5Š 10Š

Now, using (B) with five terms, we get

1 1 1 1 1
1 C C D 0:366667 (6S)
1Š 2Š 3Š 4Š 5Š
6 Numeric Analysis Part E

(B5) Error diff W (A) (B5) D 0:367879 0:366667 D 0:001212 I

with eight terms,

1 1 1 1 1 1 1 1
(B8) 1 C C C C D 0:367882
1Š 2Š 3Š 4Š 5Š 6Š 7Š 8Š
Error diff (A) (B8) D 0:367879 0:367882 D 0:000003

while, with ten terms,

1 1 1 1 1 1 1 1 1
1 (B10) 1 C C C C C D 0:367879
1Š 2Š 3Š 4Š 5Š 6Š 7Š 8Š 9Š 10Š
Error diff (A) (B10) D 0:367879 0:367879 D 0:

(b) For the 1=e1 method, that is, computing ex with x D 1 and then taking the reciprocal, we get

1 1 1 1 1 1 1
(C) f.1/ D eD1C C C C C C C C ;:
1Š 2Š 3Š 4Š 5Š 10Š
so (C), with five terms is

1 1 1 1 1 1
(C5) eD1C C C C C D 2:71667
1Š 2Š 3Š 4Š 5Š
giving the reciprocal

1 1
(C5) 1D D 0:368098 [using the result of (C5)] e
2:71667
Error diff W (A) C5* D 0:367879 0:368098 D 0:000219:

This is much better than the corresponding result (B5) in (a). With seven terms we obtain

1 1 1 1 1 1 1
(C7) 1C C C C C C C D 2:71825
1Š 2Š 3Š 4Š 5Š 6Š 7Š

(C7*) D 0:367884

Error diff W (A) (C7*) D 0:367879 0:367884 D 0:000005:

This result is almost as good as (B8) in (a), that is, the one with eight terms. With ten terms we get
Chap. 19 Numerics in General 7

1 1 1 1 1 1 1 1 1 1
(C10) 1C C C C C C C C C C D 2:71828
1Š 2Š 3Š 4Š 5Š 6Š 7Š 8Š 9Š 10Š

Sec. 19.1. Prob. 19. Table. Computation of e 1 and 1=e1 for the MacLaurin series as a
computer would do it
No. of Terms Decimal (a) e1 Exact is (b) e1 1=e1 Exact is
Factorial Terms Result 0:367879 Result 0:367879
Terms via Error: via Error:
in (12 )0
(B) Exact (B) (C) Exact (C)

1 1 C 1 0:632121 C 1 1 0:632121
1
1 1 0 0:367879 2 0:5 0:132121
1Š C
1
2 0:5 0:5 0:132121 2:5 0:4 0:032121
2Š C C
1
2:6666
3 0:166667 0:333333 0:034546 0:375000 0:007121
7
3Š C
1
2:7083
4 0:0416667 0:375000 0:007121 0:369230 0:001351
4
4Š C C
1
2:7166
5 0:00833333 0:366667 0:001212 0:368098 0:000219
7
5Š C
1
2:7180
6 0:00138889 0:368056 0:000177 0:367909 0:000030
6
6Š C C
1
2:7182
7 0:000198413 0:367858 0:000021 0:367882 0:000003
6
7Š C
1
2:7182
8 0:0000248016 0:367883 0:000004 0:367880 0:000001
8
8Š C C
1
2:7182
9 0:00000275573 0:367880 0:000001 0:367880 0:000001
8
9Š C
1
2:7182
10 0:000000275573 0:367880 0:000001 0:367880 0:000001
8
10Š C C
giving (C10*)
8 Numeric Analysis Part E

D 0:367879

Error diff W (A) (C10*) D 0:367879 0:367879 D 0

1
the same as (a). With the 1=e method, we get more accuracy for the same number of terms or we
get the same accuracy with fewer terms. The 1=9Š and 1=10Š terms are so small that they have no
effect on the result. The effect will be much greater in Prob. 20.
In the table, on the previous page, all computations are done to 6S accuracy. That means that each
term is rounded to 6S, “added” to the previous sum, and the result is then rounded before the next
term is added. For example, 1=4Š D 0:0416666667 (becomes 0:0416667/; is added to 0:333333 to
give 0:3749997, which becomes 0:375000: This is the way a computer does it, and it will produce a
different result from that obtained by adding the first four terms and then rounding. The signs in the
(a) and (b) columns indicate that the corresponding term should be added to or subtracted from the
current sum.

Sec. 19.2 Solution of Equations by Iteration

The problem of finding solutions to a single equation (p. 798 of textbook)

(1) f.x/ D 0

appears in many applications in engineering. This problem appeared, for example, in the context of
characteristic equations (Chaps. 2, 4, 8), finding eigenvalues (Chap. 8), and finding zeros of Bessel
functions (Chap. 12). We distinguish between algebraic equations, that is, when (1) is a polynomial such as

f.x/ D x3 5x C 3 D 0 [see Probs. 21, 27]

or a transcendental equation such as

f.x/ D tanx x D 0:

In the former case, the solutions to (1) are called roots and the problem of finding them is called finding
roots.
Since, in general, there are no direct formulas for solving (1), except in a few simple cases, the task of
solving (1) is made for numerics.
The first method described is a fixed-point iteration on pp. 798–801 in the text and illustrated by
Example 1, pp. 799–800, and Example 2, pp. 800–801. The main idea is to transform equation (1) from
above by some algebraic process into the form

(2) x D g.x/:

This in turn leads us to choose an x0 and compute x1 D g.x0/; x2 D g.x1/; and in general

(3) xnC1 D g.xn/ where n D 0;1;2; :

Chap. 19 Numerics in General 9

We have set up an iteration because we substitute x0 into g and get g.x0/ D x1, the next value for the

iteration. Then we substitute x1 into g and get g.x1/ D x2 and so forth.

A solution to (2) is called a fixed point as motivated on top of p. 799. Furthermore, Example 1
demonstrates the method and shows that the “algebraic process” that we use to transform (1) to (2) is
not unique. Indeed, the quadratic equation in Example 1 is written in two ways (4a) and (4b) and the
corresponding iterations illustrated in Fig. 426 at the bottom of p. 799. Making the “best” choice for g.x/
can pose a significant challenge. More on this method is given in Theorem 1 (sufficient condition for
convergence), Example 2, and Prob. 1.
Most important in this section is the famous Newton method. The method is defined recursively by

f.xn/
(5) xnC1 D xn f 0.xn/ where n D 0;1;2; ;N 1:

The details are given in Fig. 428 on p. 801 and in the algorithm in Table 19.1 on p. 802. Newton’s
method can either be derived by a geometric argument or by Taylor’s formula (5*), p. 801. Examples 3,
4, 5, and 6 show the versatility of Newton’s method in that it can be applied to transcendental and
algebraic equations. Problem 21 gives complete details on how to use the method. Newton’s method
converges of second order (Theorem 2, p. 804). Example 7, p. 805, shows when Newton’s method runs
into difficulties due to the problem of ill-conditioning when the denominator of (5) is very small in
absolute value near a solution of (1).
Newton’s method can be modified if we replace the derivative f 0.x/ in (5) by the difference quotient

f 0.xn/ D f.x n/

xn xn 1

and simplify algebraically. The result is the secant method given by (10), p. 806, which is illustrated by
Example 8 and Prob. 27. Its convergence is superlinear (nearly as fast as Newton’s method). The method
may be advantageous over Newton’s method when the derivative is difficult to compute and
computationally expensive to evaluate.

Problem Set 19.2. Page 807

1. Monotonicity and Nonmonotonicity. We consider the case of nonmonotonicity, as in Example 2 in

the book, Fig. 427, p. 801. Nonmonotonicity occurs if a sequence g.x/ is monotone decreasing, that
is,

(A) g.x1/ g.x2/ if x1 < x2.

(Make a sketch to better understand the reasoning.) Then

(B) g.x/ g.s/ if and only if x s;

where s is such that g.s/ D s [the intersection of y D x and y D g 1.x/ in Fig. 427] and
10 Numeric Analysis Part E

(C) g.x/ g.s/ if and only if x s:

Suppose we start with x1 > s. Then g.x1/ g.s/ by (C). If g.x1/ D g.s/ [which could happen if
g.x/ is constant between s and x1], then x1 is a solution of f.x/ D 0, and we are done. Otherwise
g.x1/ < g.s/, and by the definition of x2 [formula (3), p. 798 in the text] and since s is a fixed point [s D
g.s/], we obtain

x2 D g.x1/ < g.s/ D s so that x2 < s:

Hence by (B),

g.x2/ g.s/:
The equality sign would give a solution, as before. Strict inequality, and the use of (3) in the text, give

x3 D g.x2/ > g.s/ D s; so that x3 > s;

and so on. This gives a sequence of values that are alternatingly larger and smaller than s, as
illustrated in Fig. 427 of the text.
Complete the problem by considering monotonicty, as in Example 1, p. 799.

21. Newton’s method. The equation is f.x/ D x3 5x C 3 D 0 with x0 D 2;0; 2: The derivative of f.x/ is

f 0.x/ D 3x2 5:

Newton’s method (5), p. 802,

D f.xn/ D xn3 5xn C 3 xnC1 xn f 0.xn/ xn

3xn2 5 :

We have nothing to compute for the interation n D 0: For the iteration n D 1 we have
f.x0/
x1 D x0
f 0.x0/
23 52 2 C 3

D2 3 2 5
8 10 C 3
D2 12 5

D2
Chap. 19 Numerics in General 11

D2 0:1428571429

D2 0:142857

D 1:857143 D 1.85714 (6S).

From the next iteration (iteration, n D 2) we obtain

f.x1/
x2 D x1
f 0.x1/
.1:85714/3 5 1:85714 3
1:85714
D C
2
3 .1:85714/ 5

D 0.119518
D 1:85714 1:85714

5.34691
D 1:85714 0:02235272 D 1:85714 0:0223527

D 1:8347873 D 1.83479 (6S):

The iteration n D 3 gives us

f.x2/
x3 D x 2
f 0.x2/
D 1:83479 .1:83479/3 5 1:83479 C 3
3 .1:83479/2 5

D 0.00278677
D 1:83479 1:83479
5.09936
D 1:83479 0:0005464940698 D 1:83479 0:000546494 D 1:834243506 D 1.83424

(6S):

For n D 4 we obtain

f.x3/
x4 D x3
f 0.x3/
12 Numeric Analysis Part E

.1:83424/3 5 1:83424 3
D 1:83424 C

3 .1:83424/2 5
D 1:83424D 1:83424

D 1:83424 . 3:184373227 10 6/ D 1:83424 . 3:18437 10 6/

D 1:834243184 D 1.83424 (6S).

Because we have the same value for the root (6S) as we had in the previous iteration, we are
finished.
Hence the iterative sequence converges to x4 D1.83424 (6S), which is the first root of the given
cubic polynomial.
The next set of iterations starts with x0 D 0 and converges to x4 D0.656620 (6S), which is the second
root of the given cubic polynomial. Finally starting with x 0 D –2 yields x4 D–2.49086 (6S).
The details are given in the three-part table on the next page. Note that your answer might vary
slightly in the last digits, depending on what CAS or software or calculator you are using.

27. Secant method. The equation is as in Prob. 21, that is,

(P) x3 5x C 3 D 0:

This time we are looking for only one root between the given values x 0 D 1:0 and x1 D 2:0:

Solution. We use (10), p. 806, and get

xn xn 1
xnC1 D xn f.xn/f.xn/ f.xn 1/

D xn xn3 5xn C 3 3 5xn C 3xn

xxn3n11 5xn 1 C 3:
xn

The 3 in the denominator of the second equality cancels, and we get the following formula for our
iteration:

xnC1 D xn xn3 5xn C 3 xn3 5xnxn xxn3n11 5xn 1:

Sec. 19.2. Prob. 21. Table. Newton’s method with 6S accuracy
n xn f.xn/ f 0.xn/
0 2 1 7
Chap. 19 Numerics in General 13

1 1:85714 0:119518 5:34691

2 1:83479 0:00278677 5:09936
3 1:83424 0:000016219 5:09331
4 1:83424
n xn f.xn/ f 0.xn/
0 0 3 5
1 0:6 0:216 3:92
2 0:655102 0:00563268 3:7125
2
3 0:656619 0:00000530425 3:7065
5
4 0:656620 0:00000159770 3:7065
5
5 0:656620
n xn f.xn/ f 0.xn/
0 2 5 7
1 2:7142 3:42573 17:1021
9
2 2:5139 0:318694 13:9603
8
3 2:4911 0:00389923 13:6175
5
4 2:4908 0:000049216 13:6132
6 6
5 2:4908
6
x13 5x1 C 3
x2
D
5x11 xx003 x 5x
x1
x 0
D 2:0
Œ.2:0/3 5 2:0 C 3
2: 5
0 .2:0/3
2:0
1:0
For x0 D 1:0 and x1 D 2:0 we have

1:0
Œ.1:0/3 5 1:0
D 2:0 1:0 2:0 Œ 4:0 D 2:0 0:50 D 1:5 (exact).
 Next we use x1 D 2:0 and x2 D 1:5 to get
x
x1 5x1
x3 D x2 x23 5x2 C 3 x13
14 Numeric Analysis Part E
5x22
D x2 C .1:5/3 5 1:5
3 1:5 2:0 Œ.2:0/ 3
5 2:0
1:5 Œ.1:5/ 5 1:5 3

D 1:5 . 1:125/ D 1:76471 (6S):

The next iteration uses x2 D 1:5 and x3 D 1:76471 to get x4 D 1:87360 (6S). We obtain convergence at step
n D 8 and obtain x8 D 1:83424, which is one of the roots of (P). The following table shows all the steps.
Note that only after we computed x8 and found it equal (6S) to x7 did we conclude convergence.
Sec. 19.2 Prob. 27.
Table A. Secant method with 6S accuracy
n xn
2 1:5
3 1:76471
4 1:87360
5 1:83121
6 1:83412
7 1:83424
8 1.83424
For 12S values convergence occurs when n D 10.

Sec. 19.2 Prob. 27.Table B. Secant method with 12S accuracy

n xn
2 1:5
3 1:7647058823
5
4 1:8735995403
6
5 1:8312058339
1
6 1:8341181270
8
7 1:8342435958
6
8 1:8342431842
6
9 1:8342431843
1
1 1.8342431843
0 1
Note further that, for the given starting values, the convergence is monotone and is somewhat
slower than that for Newton’s method in Prob. 21. These properties are not typical but depend on
the kind of function we are dealing with. Note that Table A, by itself, represents a complete answer.
Chap. 19 Numerics in General 15

Sec. 19.3 Interpolation

Here is an overview of the rather long Sec. 19.3. The three main topics are the problem of interpolation
(pp. 808–809), Lagrange interpolation (pp. 809–812), and Newton’s form of interpolation (pp. 812–819).
Perhaps the main challenge of this section is to understand and get used to the (standard) notation of the
formulas, particularly those of Newton’s form of interpolation. Just write them out by hand and practice.
The problem of interpolation. We are given values of a function f.x/ as ordered pairs, say

(A) .x0;f0/; .x1;f1/;.x2;f2/; ;.xn;fn/ where fj D f.xj /; j D 0;1;2;:::;n:

The function may be a “mathematical” function, such as a Bessel function, or a “measured” function, say
air resistance of an airplane at different speeds. In interpolation, we want to find approximate values of
f.x/ for new x that lie between those given in (A). The idea in interpolation (p. 808) is to find a polynomial
pn.x/ of degree n or less—the so called “interpolation polynomial”—that goes through the values in (A),
that is,

(1) pn.x0/ D f0; pn.x1/ D f1; pn.x2/ D f2; ; pn.xn/ D fn:

We call pn.x/ a polynomial approximation of f.x/ and use it to get those new f.x/’s mentioned before. When
they lie within the interval Œx0;xn; then we call this interpolation and, if they lie outside the interval,
extrapolation.
Lagrange interpolation. The problem of finding an interpolation polynomial p n satisfying (1) for given
data exists and is unique (see p. 809) but may be expressed in different forms. The first type of
interpolation is the Lagrange interpolation, discussed on pp. 809–812. Take a careful look at the linear
case (2), p. 809, which is illustrated in Fig. 431. Example 1 on the next page applies linear Lagrange
interpolation to the natural logarithm to 3D accuracy. If you understand this example well, then the rest of
the material follows the same idea, except for details and more involved (but standard) notation. Example
2, pp. 810–811, does the same calculations for quadratic Lagrange interpolation [formulas (3a), (3b), p.
810] and obtains 4D accuracy. Further illustration of the (quadratic) technique applied to the sine function
and error function is shown in Probs. 7 and 9, respectively. This all can be generalized by (4a), (4b) on p.
811. Various error estimates are discussed on pp. 811–812. Example 3(B) illustrates the basic error
principle from Sec. 19.1 on p. 796.
Newton’s form of interpolation. We owe the greatest contribution to polynomial interpolation to Sir
Isaac Newton (on his life cf. footnote 3, p. 15, of the textbook), whose forms of interpolation have three
advantages over those of Lagrange:

1. If we want a higher degree of accuracy, then, in Newton’s form, we can use all previous work and just
add another term. This flexibility is not possible with Lagrange’s form of interpolation.

2. Newton’s form needs fewer arithmetic calculations than Lagrange’s form.

3. Finally, it is easier to use the basic error principle from Sec. 19.1 for Newton’s forms of interpolation.

The first interpolation of Newton is Newton’s divided difference interpolation (10), p. 814, with the kth
divided difference defined recursively by (8), p. 813. The corresponding algorithm is given in Table 19.2, p.
814, and the method illustrated by Example 4, p. 815, Probs. 13 and 15. The computation requires that we
16 Numeric Analysis Part E

set up a divided difference table, as shown on the top of p. 815. To understand this table, it may be useful
to write out the formulas for the terms, using (7), (8), and the unnumbered equations between them on p.
813.
If the nodes are equally spaced apart by a distance h, then we obtain Newton’s forward difference
interpolation (14), p. 816, with the kth forward difference defined by (13), p. 816. [This corresponds to
(10) and (8) for the arbitrarily spaced case.] An error analysis is given by (16) and the method is illustrated
by Example 5, pp. 817–818.
If we run the subscripts of the nodes backwards (see second column in table on top of p. 819), then we
obtain Newton’s backward difference interpolation (18), p. 818, and illustrated in Example 6.

Problem Set 19.3. Page 819

7. Interpolation and extrapolation. We use quadratic interpolation through three points. From (3a),
(3b), p. 810, we know that p2.x/ D L0.x/f0 C L1.x/f1 C

L2.x/f2

.x x1/.x x2/ .x x0/.x x2/ .x x0/.x x1/

D f0 C f1 C f2 .x0 x1/.x0 x2/ .x1 x0/.x1 x2/ .x2 x1/.x2 x1/

is the quadratic polynomial needed for interpolation, which goes through the three given
points .x0;f0/, .x1;f1/; and .x1;f1/. For our problem
k xk fk
x0 D 0
0 f0 D sin0
f
1 x1 D 1 D sin
4 4
f
2 x2 D 2 D sin
2 2
so that the desired quadratic polynomial for interpolating sin x with nodes at x D 0, =4, and =2 is
.x
p2.x/ D /.x / .x 0/.x / .x 0/.x /
.0 4 2 sin0 C 2 sin C 2 sin

4 /.0 2/ .4 0/. 4 2/ 4 . 2 0/. 2 4/ 2

x2 3x C 182 sin0 C x2 122x sin C x2 214x sin 2
D 4
2
4
8 16 8
3 1 28 2 1 16 2 1 8
(A) D x2 x C 0C x x 0:707107 C x x 1
2 2 2
4 8 2 4
2
D 0:3357x C 1:164x:

We use (A) to compute sinx for x D (“extrapolation” since x D lies outside the interval
Chap. 19 Numerics in General 17

0 x 2 ), x D (“interpolation” since x D lies inside the interval 0 x 2 ), x D

(interpolation), and 58 (extrapolation) and get, by (A), the following results:

x p2.x/ sinx error D sinx p2 .x/
0:382
0:5089 0:1262
7
0:382
0:4053 0:0226
7
0:923
0:9054 0:0185
9
0:923
0:9913 0:0674
9
We observe that the values obtained by interpolation have smaller errors than the ones obtained by
extrapolation. This tends to be true and the reason can be seen in the following figure. Once we are out of
the interval over which the interpolationg polynomial has been produced, the polynomial will no longer
be “close” to the function—in fact, it will begin to become very large. Extrapolation should not be used
without careful examination of the polynomial.
Note that a different order of computation can change the last digit, which explains a slight difference
between our result and that given on p. A46 of the textbook.

2 4 6 x
–2
–2

–4

–6

–8

Sec. 19.3 Prob. 7. Interpolation and extrapolation

9. Lagrange polynomial for the error function erfx. The error function [defined by (35) on p. A67 in
App. 3, Sec. A3.1, and graphed in Fig. 554, p. A68, in the textbook] given by

2Zx t2

erfx D p e dt
0

cannot be solved by elementary calculus and, thus, is an example where numerical methods are
appropriate.
Our problem is similar in spirit to Prob. 7. From (3a), (3b), and the given data for the error function
erfx, we obtain the Lagrange polynomial and simplify

.x 0:5/.x 1:0/ .x 0:25/.x 1:0/

2 D C
18 Numeric Analysis Part E

p .x/ 0:27633 0:52050

. 0:25/. 0:75/ 0:25. 0:5/
.x 0:25/.x 0:5/
(B) C 0:84270
0:75 0:5
D 0:44304x2 C 1:30896x 0:023220:

We use (B) to calculate

p2.0:75/ D 0:44304 .0:75/2 C 1:30896 0:75 0:023220 D 0:70929:

This approximate value p2.0:75/ D 0:70929 is not very accurate. The exact 5S value is erf
0:75 D 0:71116 so that the error is

error D erf 0:75 p2.0:75/ [by (6), p. 794] D 0:71116

0:70929 D 0:00187:

y
1.0

0.5

–1.0 1.0 2.0 3.0 4.0 x

–0.5

–1.0

–1.5

Sec. 19.3 Prob. 9. The functions erf x and Lagrange polynomial p 2.x/:
See also Fig. 554 on p. A68 in App. A of the textbook

13. Lower degree. Newton’s divided difference interpolation. We need, from pp. 813–814,

ajC1 D f xj;xjC1 D xfjjCC11 fxjj D f.xxjjCC11/ f.xxj j /

and
fx
C
ajC2 D f xj;xjC1;xjC2 D jC1;xxjj C22 xfjxj;xjC1:

Then the desired polynomial is

Chap. 19 Numerics in General 19

(C) pjC2.x/ D fj C .x xj /f xj;xjC1 C .x xj /.x xjC1/f xj;xjC1;xjC2:

From the five given points .xj ;fj / we construct a table similar to the one in Example 4, p. 815. We get
j xj fj D f.xj / ajC1 D f xj;xjC1 ajC2 D f xj;xjC1;xjC2

0 4 50 18 50
16:0
2 . 4/ D
1 2 18 2 18
8 C 16 D 2:0 0
8:0
C4
0 . 2/ D

2 0 2 0 8
C D 2:0
2C2
D0
3 2 2
D 2:0
D 8:0
4 4 18
From the table and (C), with j D 0, we get the following interpolation polynomial. Note that,
because all the ajD2 differences are equal, we do not need to compute the remaining differences and
the polynomial is of degree 2:

p2.x/ D f0 C .x x0/f Œx0;x1 C .x x0/.x x1/f Œx0;x1;x2 (see formula on top of p. 814)

D 50 C Œx . 4/. 16:0/ C Œx . 4/Œx . 2/.2:0/

D50 C .x C 4/. 16:0/ C .x C 4/.x C 2/.2:0/

2
D 50 16x 64 C 2x C 12x C 16 D 2x2 C . 16 C

12/x C .50 64 C 16/ D 2x2 4x C 2:

15. Newton’s divided difference formula (10), p. 814. Using the data from Example 2, p. 810, we build
the following table:
j xj fj f xj;xjC1 f xj;xjC1;xjC2

0 9:0 2:1972

D 0:1082
1 9:5 2:2513
D 0:005235
D 0:09773
2 11:0 2:3979
Then, using the table with the values needed for (10) underscored, the desired polynomial is
20 Numeric Analysis Part E

p2.x/ D 2:1972 C .x 9/ 0:1082 C .x 9/.x 9:5/ 0:005235 D

1:6709925 C 0:0113525x C 0:005235x 2:

Sec. 19.4 Spline Interpolation

We continue our study of interpolation started in Sec. 19.3. Since, for large n, the interpolation polynomial
Pn.x/ may oscillate wildly between the nodes x 0;x1;x2;:::;xn, the approach of Newton’s interpolation with
one polynomial of Sec. 19.3 may not be good enough, Indeed, this is illustrated in Fig. 434, p. 821, for n D
10, and it was shown by reknown numerical analyst Carl Runge that, in general, this example exhibits
numeric instability. Also look at Fig. 435, p. 821.
The new approach is to use n low-degree polynomials involving two or three nodes instead of one
high-degree polynomial connecting all the nodes! This method of spline interpolation, initiated by I. J.
Schoenberg is used widely in applications and forms the basis for CAD (computer-aided design), for
example, in car design (Bezier curves named after French engineer P. Bezier of the Renault Automobile
Company, see p. 827 in Problem Set 19.4).
Here we concentrate on cubic splines as they are the most important ones in applications because they
are smooth (continuous first derivative) and also have smooth first derivatives. Theorem 1 guarantees
their existence and uniqueness. The proof and its completion (Prob. 3) suggest the approach for
determining splines. The best way to understand Sec. 19.4 is to study Example 1, p. 824. It uses (12), (13),
and (14) (equidistant nodes) on pp. 823–824. A second illustration is Prob. 13. Figure 437 of the Shrine of
the Book in Jerusalem in Example 2 (p. 825) shows the interpolation polynomial of degree 12, which
oscillates (reminiscent of Runge’s example in Fig. 434), whereas the spline follows the contour of the
building quite accurately.

Problem Set 19.4. Page 826

3. Existence and uniqueness of cubic splines. Derivation of (7) and (8) from (6), p. 822, from the Proof
of Theorem 1. Formula (6), p. 822, of the unique cubic polynomial is quite involved:

qj .x/ D f.xj /cj2.x xjC1/2Œ1 C 2cj .x xj /

C f.xjC1/cj2.x xj /2Œ1 C 2cj .x xjC1/

(6) C kj cj2.x xj /.x xjC1/2

C kjC1cj2.x xj /2.x xjC1/:

We need to differentiate (6) twice to get (7) and (8), and one might make some errors in the (paper-
and-pencil) derivation. The point of the problem then is that we can minimize our chance of making
errors by introducing suitable short notations.
For instance, for the expressions involving x; we may set

Xj D x xj , XjC1 D x xjC1;
Chap. 19 Numerics in General 21

and, for the constant quantities occurring in (6), we may choose the short notations:

A D f.xj /cj2; B D 2cj ; C D f.xjC1/cj2; D D kj cj2; E D kjC1 cj2:

Then formula (6) becomes simply qj .x/ D AXj2C1.1 C BXj / C CXj2 .1 BXjC1/ C DXj Xj2C1 C EXj2XjC1:

Differentiate this twice with respect to x, applying the product rule for the second derivative, that is,

.uv/00 D u00v C 2u0v0 C uv00;

and noting that the first derivative ofthe differentiations in two steps if one wants to.) We obtainX j is
simply 1, and so is that of XjC1. (Of course, one may do

(I) qj00.x/ D AŒ2.1 C BXj / C 4XjC1B C 0 C CŒ2.1 BXjC1/ C 4Xj . B/ C 0

C D.0 C 4XjC1 C 2Xj / C E.2XjC1 C 4Xj C 0/;

where 4 D 2 2 with one 2 resulting from the product rule and the other from differentiating a
square. And the zeros arise from factors whose second derivative is zero. Now calculate q j00 at x D xj .
Since

Xj D x xj , we see that Xj D 0 at x D xj :

Hence, in each line, the term containing X j disappears. This gives

qj00.xj / D A.2 C 4BXjC1/ C C.2 2BXjC1/ C 4DXjC1 C 2EXjC1:

Also, when x D xj , then

1
XjC1 D xj xjC1 D c j [see (6*), p. 822, which defines c j ].

Inserting this, as well as the expressions for A;B;:::;E, we obtain (7) on p. 822. Indeed,

4c

qj00.xj / D f.xj /cj2 2C2 cjj C f.xjC1/cj2 2 2 2ccjj C 4kjccjj2 C 2kjCcj1cj2 :

c
Cancellation of some of the factors involving j gives
22 Numeric Analysis Part E

(7) qj00.xj / D 6f.xj /cj2 C 6f.xjC1/cj2 4kj cj 2kjC1cj :

The derivation of (8), p. 822, is similar.

For x D xjC1, we have

XjC1 D xjC1 xjC1 D 0;

so that (I) simplifies to

qj00.xjC1/ D A.2 C 2BXj / C C.2 4BXj / C 2DXj C 4EXj :

Furthermore, for x D xjC1, we have, by (6*), p. 822,

1
Xj D xjC1 xj D c j ;

and, by substituting A;:::;E into the last equation, we obtain

! ! 2
4k

2
qj00.xjC1/ D f.xj /cj2 2C4 ccjj C f.x C1/cj2 2 8ccjj C 2kcjjcj C jcCj1cj :
j
Again, cancellation of some factors cj and simplification finally gives (8), that is,

(8) qj00.xjC1/ D 6cj2f.xj / 6cj2f.xjC1/ C 2cj kj C 4cj kjC1:

For practice and obtaining familiarity with cubic splines, you may want to work out all the details
of the derivation.

13. Determination of a spline. We proceed as in Example 1, p. 824. Arrange the given data in a table for
easier work:
j xj f.xj / kj
0 0 1 0
1 1 0
2 2 1
3 3 0 6
Since there are four nodes, the spline will consist of three polynomials, q 0.x/; q1.x/; and q2.x/. The
polynomial q0.x/ gives the spline for x from 0 to 1, q 1.x/ gives the spline for x from 1 to 2; and q 2.x/
gives the spline for x from 2 to 3, respectively.

Step 1. Since n D 3 and h D 1; (14), p. 824, has two equations:

3
Chap. 19 Numerics in General 23

k0 C 4k1 C k2 D 0 C 4k1 C k2 D.f2 f0/ D 6; j D 1;

h
3
k1 C 4k2 C k3 D k1 C 4k2 6 D.f3 f1/ D 0; j D 2:
h

It is easy to show, by direct substitution, that k 1 D 2 and k2 D 2 satisfy these equations.

Step 2 for q0.x/ Determine the coefficients of the spline from (13), p. 823. We see that, in general, j

D 0;:::;n 1, so that, in the present case, we have j D 0 (this will give the spline from 0 to 1), j D 1

(which will give the spline from 1 to 2), and j D 2 (which will give the spline from 2 to 3). Take j D 0.

Then (13) gives a00 D q0.p0/ D f0 D 1; a01 D q00 .x0/ D k0 D 0;

1
a02 Dq000.x0/ D.f1 f0/.k1 2k0/ D 3 .0 1/ .2 0/ D 1;
2
1
a03 Dq0000.x0/ D.f0 f1/ C.k1 C k0/ D 2 .1 0/ C . 2 C 0/ D 0:
6
With these Taylor coefficients we obtain, from (12), p. 823, the first part of the spline in the form

q0.x/ D 1 C 0.x x0/ 1.x x0/2 C 0.x x0/3

D1C0 1.x 0/2 C 0.x0/3 D 1

x2:

Step 2 for q1.x/ a10 D q1.x1/ D f1 D 0; a11 D q10 .x1/

D k1 D 2;

1
a12 Dq100.x1/ D.f2 f1/.k2 C 2k1/ D 3 . 1 0/ .2 4/ D 1;
2
1
a13 Dq1000.x1/ D.f1 f2/ C.k2 C k1/ D 2 .0 C 1/ C .2 2/ D 2:
6

With these coefficients and x1 D 0 we obtain from (12), p. 823, with j D 1 the polynomial

q1.x/ D 0 2.x x1/ 1.x x1/2 C 2.x x1/3

D 2.x 1/ .x 1/2 C 2.x 1/3

24 Numeric Analysis Part E

D 1 C 6x 7x2 C 2x3;

which gives the spline on the interval from 1 to 2.

Step 3 for q2.x/

a20 D q2.x2/ D f2 D 1; a21 D q20 .x2/ D k2 D 2;

1
a22 Dq200.x2/ D.f3 f2/.k3 C 2k2/ D 3 .0 C 1/ . 6 C 4/ D 5;
2
1
a23 Dq2000.x2/ D.f2 f3/ C.k3 C k2/ D 2 . 1 1/ C . 6 C 2/ D 6:
6

With these coefficients and x1 D 0 we obtain, from (12), p. 823, with j D 1; the polynomial

q2.x/ D 1 C 2.x x2/ C 5.x x2/2 6.x x2/3

D 1 C 2.x 2/ C 5.x 2/2 6.x 2/3

D 63 90x C 41x2 6x3;

which gives the spline on the interval from 2 to 3.

To check the answer, you should verify that the spline gives the function values f.x j / and the values kj of
the derivatives in the table at the beginning. Also make sure that the first and second derivatives of the
spline at x D 1 are continuous by verifying that

q00 .1/ D q10 .1/ D 2 and q000.1/ D q100.1/ D 2:

The third derivative is no longer continuous,

q0000.1/ D 0 but q1000.1/ D 12:

(Otherwise you would have a single cubic polynomial from 0 to 1.)

Do the same for x D 2:
Chap. 19 Numerics in General 25
y
2

0
1 2 3 4 x

–1

–2

Sec. 19.4 Prob. 13. Spline

We see that in the graph the curve q 0 is represented by the dashed line ( ), q1 by the dotted line
( ), and q2 by the dot-dash line ( ).

Sec. 19.5 Numeric Integration and Differentiation

The essential idea of numeric integration is to approximate the integral by a sum that can be easily
evaluated. There are different ways to do this approximation and the best way to understand them is to
look at the diagrams.
The simplest numeric integration is the rectangular rule where we approximate the area under the
curve by rectangles of given (often equal) width and height by a constant value (usually the value at an
endpoint or the midpoint) over that width as shown in Fig. 441 on p. 828. This gives us formula (1) and is
illustrated in Prob. 1.
We usually get more accuracy if we replace the rectangles by trapezoids in Fig. 442. p. 828, and we
obtain the trapezoidal rule (2) as illustrated in Example 1, p. 829, and Prob. 5. We discuss various error
estimates of the trapezoidal rule (see pp. 829–831) in equations (3), (4), and (5) and apply them in
Example 2 and Prob. 5.
Most important in this section is Simpson’s rule on p. 832:

Zb h
(7) f.x/dx .f0 C 4f1 C 2f2 C 4f3 C C 2f2m 2 C 4f2m 2 C f2m/; a 3

where

b ahD and fj stands for

f.xj /:
2m
Simpson’s rule is sufficiently accurate for most problems but still sufficiently simple to compute and is
stable with respect to rounding. Errors are given by (9), p. 833, and (10), p. 834, Examples 3, 4, 5, 6
(“adaptive integration”) (see pp. 833–835 of the textbook), and Prob. 17 give various illustrations of this
26 Numeric Analysis Part E

important practical method. The discussion on numeric integration ends with Gauss integration (11), p.
837, with Table 19.7 listing nodes and coefficients for n D 2;3;4;5 (see Examples 7 and 8, pp. 837–838,
Prob. 25).
Whereas integration is a process of “smoothing,” numeric differentiation “makes things rough” (tends
to enlarge errors) and should be avoided as much as possible by changing models—but we shall need it in
Chap. 21 on the numeric solution of partial differential equations (PDEs).
Problem Set 19.5. Page 839

1. Rectangular rule (1), p. 828. This rule is generally too inaccurate in practice. Our task is to evaluate
the integral of Example 1, p. 829,

Z 1 2 Je x dx
D
0

by means of the rectangular rule (1) with intervals of size 0:1: The integral cannot be evaluated by
elementary calculus, but leads to the error function erfx, defined by (35), p. A67, in Sec. A3.1, of
App. 3 of the textbook.
Since, in (1), we take the midpoints 0:05, 0:15, :::, we calculate
j xj xj2 f.xj/ D exp xj2
1 0:05 0:0025 0:997503
2 0:15 0:0225 0:977751
3 0:25 0:0625 0:939413
4 0:35 0:1225 0:884706
5 0:45 0:2025 0:816686
6 0:55 0:3025 0:738968
7 0:65 0:4225 0:655406
8 0:75 0:5625 0:569783
9 0:85 0:7225 0:485537
10 0:95 0:9025 0:405555
Sum 7:471308 D P10 f.x/
jD1 j
Since the upper limit of integration is b D 1, the lower limit a D 0, and the number of subintervals n
D 10, we get

b a 1 0 1
h D D D D 0:1: n 10 10
Hence by (1), p. 828,

Z1 2 10

X
Rectangular rule: J D e x dx h f.xj/ D 0:1 7:471308 D 0:7471308 D 0:747131 (6S).
0
jD1
Chap. 19 Numerics in General 27

We compare this with the exact 6S value of 0:746824 and obtain

Error for rectangular rule D True Value Approximation

D 0:746824 0:747131 D 0:000307 [by (6), p. 794].

We compare our result with the one obtained in Example 1, p. 829, by the trapezoidal rule (2) on
that page, that is,

Error for trapezoidal rule D True Value Approximation

D 0:746824 0:746211 D 0:0000613 [by (6), p. 794].

This shows that the trapezoidal rule gave a more accurate answer, as was expected.
Here are some questions worth pondering about related to the rectangular rule in our calculations.
When using the rectangular rule, the approximate value was larger than the true value. Why?
(Answer: The curve of the integrand is concave.)
What would you get if you took the left endpoint of each subinterval? (Answer: An upper bound for
the value of the integral.)
If you took the right endpoint? (Answer: A lower bound.)

5. Trapezoidal rule: Error estimation by halving. The question asks us to evaluate the integral

Z 1x
J D sin dx 0 2

by the trapezoidal rule (2), p. 829, with h D 1; 0:5; 0:25 and estimate its error for h D 0:5 and h D
0:25 by halving, defined by (5), p. 830.

Step 1. Obtain the true value of J. The purpose of such problems (that can readily be solved by
calculus) is to demonstrate a numeric method and its quality—by allowing us to calculate errors (6),
p.794, and error estimates [here (5), p. 830]. We solve the indefinite integral by substitution

x du 2
uD ; D ; dx D du 2 dx 2

and 2

Z x Z 2 2Z 22x cosu D cos :

sin dx D .sin u/ du D sin udu D 2
Hence the definite integral evaluates to

Z1
x 2 h xi1
JD sin dx D cos
28 Numeric Analysis Part E

0 2 2 0

2 2 2
D
h i
(A)coscos 0D .01/ D D 0:63662:
2

Step 2a. Evaluate the integral by the trapezoidal rule (2), p. 821, with h D 1: In the trapezoidal rule (2) we
subdivide the interval of integration a x b into n subintervals of equal length h, so that

b a h D
:n

We also approximate f by a broken line of segments as in Fig. 442, p. 828, and obtain

Zb 1
(2) Jh D f.x/dx D h f.a/ C f.x1/ C f.x2/ C C f.x n 1/ Cf.b/ : a 2

From (A), we know that the limits of integration are a D 0; b D 1. With h D 1 we get

ba10
n D D D 1 interval; that is, interval Œa;b  D Œ0;1 : h 1

Hence (2) simplifies to

Zb
J1: D 0f.x/dx
a
1
(B) Dh f.a/ C f.b/
2
1
D 1:0 f.0/ C f.1/ 2

1 1 1 1 1
D 1:0 sin0 C sin D 1:0 0C 1 D D 0:50000:
2 2 2 2 2 2

From (A) and (B) we see that the error is

Error D Truevalue approximation D 0:63662 0:50000 D 0:13662 [by (6), p. 794]

Step 2b. Evaluate the integral by the trapezoidal rule (2) with h D 0:5. We get

b a 1 0 n D D D 2 intervals. h 0:5
Chap. 19 Numerics in General 29

The whole interval extends from 0 to 1, so that two equally spaced subintervals would be Œ0;  and
Œ ;1: Hence

1 1 1
J0:5 D h .f.a/ C f.x1// C .f.x1/ C f.b// D 0:5 f.0/ C f C f.1/
22 2
D 0:51 sin0 C sin 12 C 1 sin
(C)
2 2 2 2
p p " p#
2
1 2 1 2 1 using sin D
0:5" 0 C C 1# D 4C D 0:60355
4 2
D 2 2 2
with an error of 0:63662 0:60355 D 0:03307:

Step 2c. Evaluate by (2) with h D 0:25. We get

b a 1 0 n D D D 4 intervals. h
0:25

They each have a length of and so are Œ0; , Œ ; , Œ ; , and Œ ;1.

1
J0:25 D h f.a/ C f.x1/ C f.x2/ C f.x3/ C f.b/ 2
1 1 1 3
(D) D 0:25 f.0/ C f Cf Cf C f.1/
2 4 2 4

1 3 1
D 0:25 sin0 C sin C sin C sin C sin
2 8 4 8 2 2
D 0:25 .0 C 0:38268 C 0:70711 C 0:92388 C 0:50000/ D 0:62842:

The error is 0:63662 0:62842 D 0:00820:

Step 3a. Estimate the error by halving, that is, calculate 0:5 by (5), p. 830. Turn to pp. 829–830 of the
textbook. Note that the error (3), p. 830, contains the factor h 2. Hence, in halving, we can expect the
error to be multiplied by about 122 D . Indeed, this property is nicely reflected by the numerical
values (B)–(D). Now we turn to error estimating (5), that is,

(5) h=2 13.Jh=2 Jh/:

Here we obtain
30 Numeric Analysis Part E

1
0:5 .J
3 0:5 J1:0/ D .0:60355 0:50000/ D 0:03452:

The agreement of this estimate 0:03452 with the actual value of the error 0:03307 is good.
Step 3b. Estimate the error by halving, that is, calculate 0:25: We get, using (5),

1
0:25 .J
3 0:25 J0:5/ D .0:62842 0:60355/ D 0:00829;

which compares very well with the actual error, that is,

0:00820 0:00829 D 0:00009:

Although, in other cases, the difference between estimate and actual value may be larger, estimation
will still serve its purpose, namely, to give an impression of the order of magnitude of the error.
Remark. Note that since we calculated the integral by (2), p. 829, for three choices of h D 1, 0:5, 0:25
in Steps 2a–2c, we were able to make two error estimates (5), p. 830, in steps 3a, 3b.
y 1

0.8

0.6

0.4

0.2

0
0.2 0.4 0.6 0.8 1
x

Sec. 19.5 Prob. 5. Given sine curve and approximating polygons in the three trapezoidal rules used.
The agreement of these estimates with the actual value of the errors is very good

17. Simpson’s rule for a nonelementary integral. Simpson’s rule (7), p. 832, is

Zb h
(7) f.x/dx .f0 C 4f1 C 2f2 C 4f3 C C 2f2m 2 C 4f2m 1 C f2m/; a 3

where

b ahD and fj stands for f.xj /:

2m
The nonelementary integral is the sine integral (40) in Sec. A3.1 of App. 3 on p. A68 of the textbook:

Si.x/ D Z x sin x dx:

Chap. 19 Numerics in General 31

0 x

Being nonelementary means that we cannot solve the integral by calculus. For x D 1, its exact value
(by your CAS or Table A4 on p. A98 in App. 5) is

Z1
sin x
Si.1/ D dx D 0:9460831:
0 x

We construct a table with both 2m D 2 and 2m D 4, with values of the integrand accurate to seven
digits
sin xj sin xj
j1 xj fj D f.xj / D j xj fj D f.xj / D
xj xj
0 0 1:0000000 0 0 1:0000000
1 0:25 0:9896158
1 0:5 0:9588511 2 0:5 0:9588511
3 0:75 0:9088517
2 1:0 0:8414710 4 1:0 0:8414710
Simpson’s rule, with m D 1; i.e., h D 0:5, is

h 0:5
Si.1/ D .f0 C 4f1 C f2/ D .1 C 4 0:9588511 C 0:8414710/ D 0:9461459:
3 3

With m D 2; i.e. h D 0:25, h

Si.1/ D .f0 C 4f1 C 2f2 C 4f3 C f4/ 3

D .1 C 4 0:9896158 C 2 0:9588511 C 4 0:9088517 C 0:8414710/ D 0:9460870:

25. Gauss integration for the error function. n D 2 is required. The transformation must convert the
interval to Œ 1;1:

11

The exact 9S value is 0:746824133:

The nodes and coefficients are shown in Table 19.7, on p. 837, in the textbook with n D 5. Using
them, we compute
2
32 Numeric Analysis Part E

We can do this with x D at C b so that 0 D a. 1/ C b and 1 D a.1/ C b: We see that a D b D satisfies

this so x D 12.t C 1/.

Since dx D 12 dt, our integral takes the form

Z 1 e 2 dx D 1 Z 1 e 1=4.tC1/2 dt:
x
Note the high accuracy achieved with a rather modest amount of work.

Multiply this by 2=p to obtain an approximation to the error function erf 1.D 0:842700793 with .

9S// given by (35) on p. A67 in App. 3.1

2
p 0:746824127 D 0:842700786:

Solution to Self Test on Rounding Problem to Decimals (see p. 2 of this Solutions Manual and Study Guide)

(a) 1:23454621 C 5 10 .7C1/ D 1:23454621 C 0:00000005 D 1:23454626:

„ ƒ‚ … 7 zeros

Then we chop off the eigth digit “6” and obtain the rounded number to seven decimals (7D) 1:2345462:

(b) 398:723555 C 5 10 .4C1/ D 398:723555 C 5 10 5

D
398:723555 C 0: 0000 5
„ƒ‚… 4 zeros
D 398:723605

Next we chop off from the fifth digit onward, that is, “05” and obtain the rounded number to four
decimals (4D) 398:7236:

Solution to Self Test on Rounding Problem to Significant Digits (see p. 3 of this Solutions Manual and Study
Guide)

We follow the three steps.

1. 102:89565 D 0:10289565 103I

1 J D XAj e 1=4.tjC1/2 D 0:746824127:

2 jD1
Chap. 19 Numerics in General 33

2. We ignore the factor 103. Then we apply the roundoff rule for decimals to the number 0:10289565 to
get

0:10289565 C 5 10 .6C1/ D 0:10289615 .6D/:

3. Finally we have to reintroduce the factor 10 3 to obtain our final answer, that is,

0:102896 103 D 102:896 .6S/:

Chap. 20 Numeric Linear Algebra
Chapter 20 contains two main topics: solving systems of linear equations numerically (Secs. 20.1–20.5, pp.
844–876) and solving eigenvalue problems numerically (Secs. 20.6–20.9, pp. 876–898). Highlights are as
follows.
Section 20.1 starts with the familiar Gauss elimination method (from Sec. 7.3), now in the context of
numerics with partial pivoting, row scaling, and operation count and the method itself expressed in
algorithmic form. This is followed by methods that are more efficient than Gauss (Doolittle, Crout, Cholesky)
in Sec. 20.2 and iterative methods (Gauss–Seidel, Jacobi) in Sec. 20.3. We study the behavior of linear
systems in detail in Sec. 20.4 and introduce the concept of a condition number that will help us to
determine whether a system is good (“well-conditioned”) or bad (“ill-conditioned”). The first part of Chap.
20 closes with the least squares method, an application in curve fitting, which has important uses in
statistics (see Sec. 25.9).
Although we can find the roots of the characteristic equations in eigenvalue problems by methods from
Sec. 19.2, such as Newton’s method, there are other ways in numerics concerned with eigenvalue
problems. Quite surprising is Gerschgorin’s theorem in Sec. 20.7 because it allows us to obtain information
directly, i.e., without iteration, from the elements of a square matrix, about the range in which the
eigenvalues of that matrix lie. This is, of course, not as good as obtaining actual numbers for those
eigenvalues but is sufficient in some problems.
Other approaches are an iterative method to determine an approximation of a dominant eigenvalue in a
square matrix (the power method, Sec. 20.8) and a two-stage method to compute all the eigenvalues of a
real symmetric matrix in Sec. 20.9.
The chapter has both easier and more involved sections. Sections 20.2, 20.3, 20.7, 20.9 are more involved
and may require more study time. You should remember formulas (4) and (8) in Sec. 20.5 and their use.
In terms of prior knowledge, you should be familiar with matrices (Secs. 7.1, 7.2), and it would be helpful
if you had some prior knowledge of Gauss elimination with back substitution (see Sec. 7.3, pp. 272–280).
Section 20.1 moves faster than Sec. 7.3 and does not contain some details such as the three types of
solutions that occur in linear systems. For the second main topic of Chap. 20 you should be familiar with the
material contained in Sec. 20.6, pp. 876–879. Thus you should know what a matrix eigenvalue problem is
(pp. 323–324), remember how to find eigenvectors and eigenvalues of matrices (pp. 324–328 in Sec. 8.1),
know that similar matrices have the same eigenvalues (see Theorem 2, p. 878, also Theorem 3, p. 340 in
Sec. 8.4), and refresh your knowledge of special matrices in Theorem 5, p. 879.

Sec. 20.1 Linear Systems: Gauss Elimination

Gauss elimination with back substitution is a systematic way of solving systems of linear equations (1), p.
845. We discussed this method before in Sec. 7.3 (pp. 272–282) in the context of linear algebra. This time
the context is numerics and the current discussion is kept independent of Chap. 7, except for an occasional
reference to that chapter. Pay close attention to the partial pivoting introduced here, as it is the main
difference between the Gauss elimination presented in Sec. 20.1 and that of Sec. 7.3. The reason that we
need pivoting in numerics is that we have only a finite number of digits available. With many systems, this
can result in a severe loss of accuracy. Here (p. 846), to pivot a kk, we choose as our pivoting equation the
one that has the absolutely largest coefficients ajk in column k on or below the main diagonal. The details
are explained carefully in a completely worked out Example 1, pp. 846–847. The importance of this
particular partial pivoting strategy is demonstrated in Example 3, pp. 848–849. In (a) the “absolutely
largest” partial pivoting strategy is not followed and leads to a bad value for x 1. This corresponds to the
method of Sec. 7.3. In (b) it is followed and a good value for x 1 is obtained!
2 Numeric Analysis Part E

Table 20.1, p. 849, presents Gauss elimination with back substitution in algorithmic form. The section
ends with an operation count of 2n3=3 for Gauss elimination (p. 850) and n2 C n for back substitution (p.
851). Operation count is one way to judge the quality of a numeric method.
The solved problems show that a system of linear equations may have no solution (Prob. 3), a unique
solution (Prob. 9), or infinitely many solutions (Prob. 11). This was also explained in detail on pp. 277–280
in Sec. 7.3. You may want to solve a few problems by hand until you feel reasonably comfortable with the
Gaussian algorithm and the particular type of pivoting.

Problem Set 20.1. Page 851

3. System without a solution. We are given a system of two linear equations

[Eq. (1)] 7:2x1 3:5x2 D 16:0; [Eq. (2)] 14:4x1 C 7:0x2 D 31:0:

We multiply the first equation [Eq. (1)] by 2 to get

14:4x1 7:0x2 D 32:0:

If we add this equation to the second equation [Eq. (2)] of the given system, we get

0x1 C 0x2 D 63:0:

This last equation has no solution because the x1, x2 are each multiplied by 0, added, and equated to
63:0! Or looking at it in another way, we get the false statement that 0 D 63: [A solution would exist if
the right sides of Eq. (1) and Eq. (2) were related in the same fashion, for instance, 16:0 and 32:0
instead of 31:0.] Of course, for most systems with more than two equations, one cannot immediately
see whether there will be solutions, but the Gauss elimination with partial pivoting will work in each
case, giving the solution(s) or indicating that there is none. Geometrically, the result means that these
equations represent two lines with the same slope of

7:2 14:4 3:5 D

7:0 D
2:057143

but different x2-intercepts, that is,

D 4:571429 for Eq. (1),

and D 4:428571 for Eq. (2).

Hence Eq. (1) and Eq. (2) are parallel lines, as show in the figure on the next page.

9. System with a unique solution. Pivoting. ALGORITHM GAUSS, p. 849. Open your textbook to
p. 849 and consider Table 20.1, which contains the algorithm for the Gauss elimination. To follow the
discussion, control it for Prob. 9 in terms of matrices with paper and pencil. In each case, write down
Chap. 20 Numeric Linear Algebra 3

all three rows of a matrix, not just one or two rows, as is done below to save some space and to avoid
copying the same numbers several times.
At the beginning, k D 1. Since a11 D 0, we must pivot. Between lines 1 and 2 in Table 20.1 we search
for the absolutely greatest aj1: This is a31.D 13/. According to the algorithm, we have to interchange
Eqs. (1) (current row) and (3) (row with the maximum), that is, Rows 1 and 3 of the augmented matrix.
This gives
x2
10
Eq. (2)

5
Eq. (1)

4x
1
–4 –3 –2 –1 0 1 2 3

–5

–10

–15

Sec. 20.1 Prob. 3. Graphic solution of a system of two parallel equations

[Eq. (1) and Eq. (2)]

23
13 8 0178:54
67
(A) 66 0 885:887:
67
45
0 6 13137:86

Don’t forget to interchange the entries on the right side (that is, in the last column of the augmented
matrix).

To get 0 as the first entry of Row 2, subtract times Row 1 from Row 2. The new Row 2 is
hi
(A2) 0 3:692308 8168:28308 :

This was k D 1 and j D 2 in lines 4 and 5 in the table.

Now comes k D 1 and j D n D 3 in line 4. The calculation is

a31 0
m31 D D D 0: a11 13
4 Numeric Analysis Part E

Hence, the operations in line 4 simply have no effect; they merely reproduce Row 3 of the matrix in (A). This
completes k D 1.

Next is k D 2. In the loop between lines 1 and 2 in Table 20.1, we have the following: Since
6 > 3:692308, the maximum is in Row 3 so we interchange Row 2 (A2) and Row 3 in (A). This gives the
matrix
2 8 3
0178:54
13 6 13137:86
6 3:692308 7
(B) 60 7: 5
4 8168:28308
0
In line 4 of the table with k D 2 and j D k C 1 D 3 we calculate

a32 3:692308 m32 D D D 0:615385:

a22 6

Performing the operations in line 5 of the table for p D 3; 4, we obtain the new Row 3

hi
(B3) 0 0 16253:12 :

The system and its matrix have now reached triangular form. We
begin back substitution with line 6 of the table:

a34 253:12 x3 D D D 15:82: a33

16

(Remember that, in the table, the right sides b 1; b2; b3 are denoted by a14; a24; a34, respectively.)
Line 7 of the table with i D 2;1 gives

x2 D .137:86 13 15:82/ D 11:3 .i D 2/

and

x1 D .178:54 . 8 . 11:3// D 6:78 .i D 1/:

Note that, depending on the number of digits you use in your calculation, your values may be
slightly affected by roundoff.

11. System with more than one solution. Homogeneous system. A homogeneous system always has the
trivial solution x1 D 0, x2 D 0, :::, xn D 0: Say the coefficient matrix of the homogeneous system has rank
r. The homogeneous system has a nontrivial solution if and only if

r<n where n is the number of unknowns.

The details are given in Theorem 2, p. 290 in Sec. 7.5, and related Theorem 3, p. 291.
Chap. 20 Numeric Linear Algebra 5

In the present problem, we have a homogenous system with n D 3 equations. For such a system, we
may have r D 3 (the trivial solution only), r D 2 [one (suitable) unknown remains arbitrary— infinitely
many solutions], and r D 1 [two (suitable) variables remain arbitrary, infinitely many solutions]. Note
that r D 0 is impossible unless the matrices are zero matrices. In most cases we have choices as to
which of the variables we want to leave arbitrary; the present result will show this. To avoid
misunderstandings: we need not determine those ranks, but the Gauss elimination will automatically
give all solutions. Your CAS may give only some solutions (for example, those obtained by equating
arbitrary unknowns to zero); so be careful.

The augmented matrix of the given system is

2 3:4 6:12 2:72
3
6 1:0 1:80 0:80 0
6 2:7 4:86 2:16 7
6 07:
4 7
5
0
Because 3:4 is the largest entry in column 1, we add 1=3:4 Row 1 (the pivot row) to Row 2 to obtain
the new Row 2:
hi
0 0 00 :
Add 2:7=3:4 Row 1 (the pivot row) to Row 3 of the given matrix to obtain
hi
0 0 4:320 :

We end up with a “triangular” system of the form (after interchanging rows 2 and 3)

3:4x1 6:12x2 2:72x3 D 0;

4:32x3 D 0; 0 D 0:

Note that the last equation contains no information. From this, we get

4:32x3 D 0 which implies that (S1) x3 D 0:

We substitute this into the first equation and get

(S2) 3:4x1 6:12x2 D 0:

Since the system reduced to two equations (S1) and (S2) in three unknowns, we have the choice of
one parameter t.

If we set
(S3) then (S2) x1 D t (arbitrary);
becomes
6 Numeric Analysis Part E

3:4t 6:12x2 D 0
(S2*)
so that

(S2**) x2 D t D 0:556t:

Then the solution consists of equations (S3), (S2**), and (S1). This corresponds to the solution on p.
A48 in App. 2 of the textbook.

If we set

(S4) x2 D tQ (arbitrary, we call it tQ instead of t to show its independence from t/; then

we solve for x1 and get

6:12
(S5) x1 D tQ D 1:8 t;Q
3:4
and the solution consists of (S5), (S4), and (S1). The two solutions are equivalent.

Sec. 20.2 Linear Systems: LU-Factorization, Matrix Inversion

The inspiration for this section is the observation that an n n invertible matrix can be written in the form

(2) A D LU;

where L is a lower triangular and U an upper triangular matrix, respectively.

In Doolittle’s method, we set up a decomposition in the form (2), where m jk in the matrix L are the
multipliers of the Gauss elimination with the main diagonal 1;1; :::;1 as shown in Example 1 at the bottom
of p. 853. The LU-decomposition (2), when substituted into (1), on p. 852, leads to

AxLUxL. Ux/ D Ly D b; D D
|{z} y

which means we have written

(3) (a) Ly D b where (b) Ux D y:

This means we can solve first (3a) for y and then (3b) for x: Both systems (3a), (3b) are triangular, so we can
solve them as in the back substitution for the Gauss elimination. Indeed, this is our approach with
Doolittle’s method on p. 854. The example is the same as Example 1, on p. 846 in Sec. 20.1. However,
Doolittle requires only about half as many operations as Gauss elimination.
If we assign 1;1;:::;1 to the main diagonal of the matrix U (instead of L) we get Crout’s method.
A third method based on (2) is Cholesky’s method, where the n n matrix A is symmetric, positive definite.
This means

.symmetric/ A D AT;
Chap. 20 Numeric Linear Algebra 7

and

.PD/ xTAx >0 for all x ¤ 0:

Under Cholesky’s method, we get formulas (6), p. 855, for factorization. The method is illustrated by
Example 2, pp. 855–856, and Prob. 7. Cholesky’s method is attractive because it is numerically stable
(Theorem 1, p. 856).
Matrix inversion by the Gauss–Jordan elimination method is discussed on pp. 856–857 and shown in
Prob. 17.

More Details on Example 1. Doolittle’s Method, pp. 853–854. In the calculation of the entries of L and U (or
LT in Cholesky’s method) in the factorization A D LU with given A, we employ the usual matrix
multiplication

Row times Column.

In all three methods in this section, the point is that the calculation can proceed in an order such that we
solve only one equation at a time. This is possible because we are dealing with triangular matrices, so that
the sums of n D 3 products often reduce to sums of two products or even to a single product, as we will see.
This will be a discussion of the steps of the calculation, on p. 853, in terms of the matrix equation A D LU,
written out

2 3 2 32 3
3 5 2 1 0 0 u11 u12 u13

A D 660 8 277 D LU D 66m21 1 07766 0 u22 u2377:

4 5 4 54 5
6 2 8 m31 m32 1 0 0 u33

Remember that, in Doolittle’s method, the main diagonal of L is 1,1, 1. Also, the notation m jk suggests
multiplier, because, in Doolittle’s method, the matrix L is the matrix of the multipliers in the Gauss
elimination. Begin with Row 1
column of U; thus,
of A. The entry a11 D 3 is the
3D1 0 0 u11 0 0T D 1 u11; dot product of the first row of
L and the first where 1 is
prescribed. Thus, u11 D 3: Similarly, a12 D 5 D 1 u12 C 0 u22 C 0 0 D u12; thus u12 D 5: Finally, a13 D 2 D u13. This
takes care of the first row of A. In connection with the second row of A we have to consider the second row
of L, which involves m21 and 1. We obtain
a21 D 0 D m21 u11 C 1 0 C0 D m21 3; hence m21 D 0; a22 D 8 D m21 u12 C 1 u22 C 0
D u22; hence u22 D 8; a23 D 2 D m21 u13 C 1 u23 C 0 D u23; hence u23 D 2:

In connection with the third row of A we have to consider the third row of L, consisting of m 31; m32; 1:
We obtain

a31 D 6 D m31u11 C 0 C0 D m31 3; hence m31 D 2;

a32 D 2 D m31 u12 C m32 u22 C 0 D 2 5 C m32 8; hence m32 D 1; a33 D 8 D m31u13 C m32u23 C 1
u33 D 2 2 1 2 C u33; hence u33 D 6:
8 Numeric Analysis Part E

In (4), on p. 854, the first line concerns the first row of A and the second line concerns the first column of A;
hence in that respect the order of calculation is slightly different from that in Example 1.

Problem Set 20.2. Page 857

7. Cholesky’s method. The coefficient matrix A of the given system of linear equations is given by

2 3
9 6 12
6 7
AD 66 13 11 7 (as explained in Sec. 7.3, pp. 272–273).
4 5
12 11 26

We clearly see that the given matrix A is symmetric, since the entries off the main diagonal are mirror
images of each other (see definition of symmetric on p. 335 in Sec. 8.3). The Cholesky factorization of
A (see top of p. 856 in Example 1) is

2 3 2 32 3
9 6 12 l11 0 0 l11 l21 l31
6 7 6 76 7
66 13 117 D 6l21 l22 0760 l22 l327:
4 5 4 54 5
12 11 26 l31 l32 l33 0 0 l33

We do not have to check whether A is also positive definite because, if it is not, all that would happen
is that we would obtain a complex triangular matrix L and would then probably choose another
method. We continue.
Going through A row by row and applying matrix multiplication (Row times Column) as just before
we calculate the following.

D l112 ; l11 D pa21a11 D6p9 D 3;

hence
a11 D 9 D l112 C 0 C 0 a12 D 6 hence
D l11l21 C 0 C 0 D 3l21; l12 D l11 D 3 D 2;
a13 D 12 D l11l31 C 0 C 0 D 3l31; hence l31 D 4:
In the second row of A we have a21 D a12 (symmetry!) and need only two calculations:
a22 D 13 D l212 C l222 C 0 a23 D D .2/2 C l222 ; D hence l22 D 3;
11 D l21l31 C l22l32 C 0 2 4 C 3l32; hence l32 D 1:
In the third row of A we have a31 D a13 and a32 D a23 and need only one calculation:

a33 D 26 D l312 C l322 C l332 D .4/2 C 1 C l332 ; hence l33 D 3:

Now solve Ax D b, where b D Œ17:4 23:6 30:8T. We first use L and solve Ly D b, where y D Œy 1 y2
y3T. Since L is triangular, we only do back substitution as in the Gauss algorithm. Now since L is lower
triangular, whereas the Gauss elimination produces an upper triangular matrix, begin with the first equation
Chap. 20 Numeric Linear Algebra 9

and obtain y1. Then obtain y2 and finally y3. This simple calculation is written to the right of the
corresponding equations:

2 32 3 2 3 y1 D 17:4 D 5:8;
3 0 0 y1 17:4
6 76 7 6 7
62 3 07 6y27 D 623:67 y2 D.23:6 2y1/ D 4;
4 54 5 4 5
4 1 3 y3 30:8 y3 D .30:8 4y1 y2/ D 1:2:

In the second part of the procedure you solve L Tx D y for x. This is another back substitution. Since L T is
upper triangular, just as in the Gauss method after the elimination has been completed, the present back
substitution is exactly as in the Gauss method, beginning with the last equation, which gives x 3, then using
the second equation to get x2, and finally the first equation to obtain x 1.
Details on the back substitution are as follows:

2 32 3 2 3
3 2 4 x1 5:8 (S1) 3x1 C 2x2 C 4x3 D 5:8;

6 76 7 6 7
60 3 17 6x27 D 6 4 7 written out is (S2) 3x2 C x3 D 4;
4 54 5 4 5
0 0 3 x3 1:2 (S3) 3x3 D 1:2:

Hence Eq. (S3)

3x3 D 1:2 gives (S4) x3 D 1:2 D 0:4:

Substituting (S4) into (S2) gives

3x2 C x3 D 3x2 C 1:2 D 4 so that (S5) x2 D .4 1:2/ D 1:2:

Substituting (S4) and (S5) into (S1) yields

3x1C2x2C4x3 D 3x1C21:2C40:4 D 5:8 so that (S6) x1 D .5:8 2:4 1:6/ D 1:8 D 0:4:

Hence the solution is (S6), (S5), (S4):

x1 D 0:4; x2 D 1:2; x3 D 0:4:

We check the solution by substituting it into the given linear system written as a matrix equation.
Indeed,

2 32 3 2 32 3
9 6 12 x1 9 6 12 0:6
10 Numeric Analysis Part E

Ax D 66 6 13 117766x277 D 66 6 13 1177661:277
4 54 5 4 54 5
12 11 26 x3 12 11 26 0:4
2 3 2 3 2 3
9 0:6 C 6 1:2 C 12 0:4 5:4 C 7:2 C 4:8 17:4

6 D 7 6 7 6 7
6 6 0:6 C 13 1:2 C 11 0:4 7 D 6 3:6 C 15:6 C 4:4 7 D 623:67 D b;
4 5 4 5 4 5
12 0:6 C 11 1:2 C 26 0:4 7:2 C 13:2 C 10:4 30:8

which is correct.
Discussion. We want to show that A is positive definite, that is, by definition on p. 346 in Prob. 24 in
Sec. 8.4, and also on p. 855:

.PD/ xTAx >0 for all x ¤ 0:

We calculate
0 2 312 3
9 6 12 x1

xTAx D BBx1 x2 x366 6 13 1177CC66x277

@ 4 5A4 5
12 11 26 x3
2 3 x1
6 7
D 9x1 C 6x2 C 12x3 6x1 C 13x2 C 11x3 12x1 C 11x2 C 26x3 6x2 7
4 5
x3
D .9x1 C 6x2 C 12x3/ x1 C .6x1 C 13x2 C 11x3/ x2 C .12x1 C 11x2 C 26x3/ x3
2 2 2
D
9x1 C 12x1x2 C 24x1x3 C 22x2x3 C 13x2 C 26x3:

We get the quadratic form Q and want to show that (A) is true for Q:

(A) Q D 9x12 C 12x1x2 C 24x1x3 C 22x2x3 C 13x22 C 26x32 > 0 for all x1;x2;x3 ¤ 0:

Since Q cannot be written into a form . / 2; it is not trivial to show that (A) is true. Thus we look for
other ways to verify (A). One such way is to use a mathematical result given in Prob. 25, p. 346.
It states that positive definiteness (PD) holds if and only if all the principal minors of A are positive.
This result is also known as Sylvester’s criterion.
Chap. 20 Numeric Linear Algebra 11

For the given matrix A, we have three principal minors. They are:

ˇ ˇ ˇ ˇ
ˇa11 a12ˇ ˇ9 6ˇ
a119 > 0; D ˇ ˇDˇ ˇ D 9 13 6 6 D 101 > 0;
ˇ ˇ ˇ ˇ
ˇa21 a22ˇ ˇ6 13ˇ

ˇ ˇ ˇ ˇ ˇ ˇ
ˇ13 11ˇ ˇ 6 11ˇ ˇ 6 13ˇ detA9ˇ ˇ 6ˇ ˇ C 12ˇ ˇ
D
ˇ ˇ ˇ ˇ ˇ ˇ
ˇ11 26ˇ ˇ12 26ˇ ˇ12 11ˇ
D 9 217 6 24 C 12 . 90/ D 729 > 0:
Since all principal minors of A are positive, we conclude, by Sylvester’s criterion, that A is indeed
positive definite.
The moral of the story is that, for large A; showing positive definiteness is not trivial, although in
some cases it may be concluded from the kind of physical (or other) application.

17. Matrix inversion. Gauss–Jordan nethod. The method suggested in this section is illustrated in detail
in Sec. 7.8 by Example 1, on pp. 303–304, in the textbook, as well as in Prob. 1 on pp. 123–124 in
Volume I of the Student Solutions Manual. It may be useful to look at one or both examples. In
your answer, you may want to write down the matrix operations stated here in our solution to
Prob. 17 to the right of the matrix as is done in Example 1, p. 303, of the textbook.
The matrix to be inverted is
2 3

1 4 2

G D 66 4 25 477:

4 5

2 4 24

We start by appending the given 3

3 6 matrix : 3 matrix G by the 3 3 unit matrix I
2 4 2 1 0 03 to obtain the following
1 25 4 0 1
4 24 077:
G1D 66 4 0 0
5
4
1
2

Thus the left 3 3 submatrix is the given matrix and the right 3 3 submatrix is the 3 3 unit matrix I. We apply
the Gauss–Jordan method to G1 to obtain the desired inverse matrix. At the end of the process, the left 3 3
submatrix will be the 3 3 unit matrix, and the right 3 3 submatrix will be the inverse of the given matrix.

The 4 in Row 2 of G1 is the largest value in Column 1 so we interchange Row 2 and Row 1 and
12 Numeric Analysis Part E

get
23
4 25 40 1 0

G2D 66 1 4 21 0 077:
45
2 4 240 0 1

Next we replace Row 2 by Row 2 C Row 1 and replace Row 3 by Row 3 C Row 1. This gives us the new
matrix
23
4 25 40 1 0
D
67
G36 03107:
67
45
0 26 0 1

Now, because > ; we swap Rows 2 and 3 of G3 to obtain

23
4 25 40 1 0
D
67
G46 026017:
67
9 33
Replace Row 3 by Row 3 = 2 Row 2. The new matrix is 45
4
0 3 1
2 03
4 25 4 0 1 0
D
67
G56 02601 7:
67
45
0 0 1
This was the Gauss part. The given matrix is triangularized. Now comes the Jordan part that
diagonalizes it. We know that we need 1’s along the diagonal in the left-hand matrix, so we multiply
Row 1 by : In addition, we also multiply Row 2 by ; and Row 3 by to get
23
1100
67
G6 660 1 0 332 77:
D
45
1 0 1
52
Eliminate the entries in Rows 1 and 2 (Col. 3) by replacing Row 2 by Row 2 33Row 3 and Row
Chap. 20 Numeric Linear Algebra 13

2 by Row 1C Row 3. This gives the matrix

23
10
67
G7 660 1 0 3177:
D
45
1 0 1

Finally, we eliminate in the second column of G7. We do this by replacing Row 1 of G 7 by Row 1
C Row 2. The final matrix is

23
2 0 0
67
G8 660 1 0 13 77:
D
45
0 0 1
The last three columns constitute the inverse of the given matrix, that is,

2 3

6 D
G1 6 261 77:

6 93 7
45

You may want to check the result by showing that

GG 1 D I and G 1G D I:

Sec. 20.3 Linear Systems: Solution by Iteration

We distinguish between direct methods and indirect methods (p. 858). Direct methods are those methods
for which we can specify in advance how many numeric computations it will take to get a solution. The
Gauss elimination and its variants (Secs. 20.1, 20.2) are examples of direct methods. Indirect or iterative
methods are those methods where we start from an approximation to the true solution and, if successful,
obtain better and better approximations from a computational cycle repeated as often as may be necessary
for achieving a required accuracy. Such methods are useful for solving linear systems that involve large
sparse systems (p. 858).
The first indirect method, the Gauss–Seidel iteration method (Example 1, Prob. 9) requires that we take a
given linear system (1) and write it in the form (2). You see that the variables have been separated and
appear on the left-hand side of the equal sign with coefficient 1. The system (2) is now prepared for
iteration. Next one chooses a starting value, here x1.0/ D 100; x2.0/ D 100; etc. (follow the textbook on
14 Numeric Analysis Part E

p. 859). Equation (3) shows how Gauss–Seidel continues with these starting values. And here comes a
crucial point that is particular to the method, that is, Gauss–Seidel always uses (where possible) the most
recent and therefore “most up to date” approximation for each unknown (“successive corrections”). This is
shown in the darker shaded blue area in (3) and explained in detail in the textbook as well as in Prob. 9.
The second method, Jacobi iteration (13), p. 862 (Prob. 17), is very similar to Gauss–Seidel but avoids
using the most recent approximation of each unknown within an iteration cycle. Instead, as is much more
common with iteration methods, all values are updated at once (“simultaneous corrections”).
For these methods to converge, we require “diagonal dominance,” that is, the largest (in absolute value)
element in each row must be on the diagonal.
Other aspects of Gauss–Seidel include a more formal discussion [precise formulas (4), (5), (6)],
ALGORITHM GAUSS–SEIDEL (see p. 860), convergence criteria (p. 861, Example 2, p. 862), and residual (12).
Pay close attention to formulas (9), (10), (11) for matrix norms (Prob. 19) on p. 861, as they will play an
important role in Sec. 20.4.

Problem Set 20.3. Page 863

9. Gauss–Seidel iteration. We write down the augmented matrix of the given system of linear equations
(see p. 273 of Sec. 7.3 in the textbook):
23
5 1 219
D
67
A61 4227:
67
45
2 3 839

This is a case in which we do not need to reorder the given linear equations, since we note that the
large entries 5, 4, 8 of the coefficient part of the augmented matrix stand on the main diagonal. Hence
we can expect convergence.
Remark. If, say, instead the augmented matrix had been
23
5 1 219
67
62 3 8397
67
45
1 4 22

meaning that 5, 3, 2 would be the entries of the main diagonal so that 8 and 4 would be larger entries
outside the main diagonal, then we would have had to reorder the equations, that is, exchange the
second and third equations. This would have led to a system corresponding to augmented matrix A
above and expected convergence.
We continue. We divide the equations so that their main diagonal entries equal 1 and keep these
terms on the left while moving the other terms to the right of the equal sign. In detail, this means that
we multiply the first given equation of the problem by , the second one by , and the third one by :
We get
Chap. 20 Numeric Linear Algebra 15

x1 C x2 C x3 D ;

x1 C x2 x3 D ; x1 C

x2 C x3 D ; and then

moving the off-diagonal

entries to the right:

x1 D x2 x 3;

(GS) x2 D x1 C x3;

x3 D x1 x2:

We start from x1.0/ D 1; x2.0/ D 1; x3.0/ D 1 (or any reasonable choice) and get

.1/ 19 1 .0/ 2 .0/

x1 D x2 x3 5 5 5 D 1
1

19 1 2 16
DD 3:2 .exact/;

x2.1/ D 1x1

x3.0/ 4 4D 3:2 C

D 0:8 C D 0:8 .exact/;

.1/ 39 2 .1/ 3 .1/

x3 D x1 x2
8 8 8

D 3:2 . 0:8/

D 4:375 .exact/:
16 Numeric Analysis Part E

Note that we always use the latest possible value in the iteration, that is, for example, in computing x 2.1/ we
use x1.1/ (new! and not x1.0/) and x3.0/ (no newer value available). In computing x3.1/ we use x1.1/ (new!) and x2.1/
(new!) (see also p. 859 of the textbook).
Then we substitute x1.1/ D 3:2; x2.1/ D 0:8; x3.1/ D 4:375 into system (GS) and get

x1.2/ D 2:210000; x2.2/ D 1:135000; x3.2/ D 3:89688:

The results are summarized in the following table. The values were computed to 6S with two guard digits
for accuracy.
Prob. 9. Gauss–Seidel Iteration Method. Table of Iterations. Five Steps.

Step x1 x2 x3
mD1 3:2 0:8 4:375
mD2 2:21000 1:13500 3:89688
mD3 2:01425 0:944875 4:01711
mD4 2:00418 1:00751 3:99614
D m 5 2:00004 0:998059 4:00072

The exact solution is 2;1;4.

11. Effect of starting values. The point of this problem is to show that there is surprisingly little difference
between corresponding values, as the answer on p. A49 in App. 2 shows, although the starting values
differ considerably. Hence it is hardly necessary to search extensively for “good” starting values.

17. Jacobi Iteration. Convergence related to eigenvalues. An outline of the solution is as follows. You may
want to work out some more of the details. We are asked to consider the matrix of the system of
linear equations in Prob. 10 on p. 863, that is,
2 3

4 0 5

AQ D 661 6 277:

4 5

8 2 1

We note that aQ13 D 8 is a large entry outside the main diagonal (see Remark in Prob. 9 above). To
obtain convergence, we reorder the rows as shown, that is, we exchange Row 3 with Row 1, and get,

2 3
8 2 1
6 7
61 6 27:
4 5
4 0 5
Chap. 20 Numeric Linear Algebra 17

Then we divide the rows by the diagonal entries 8, 6, and 5, respectively, as required in (13), p. 862
(see ajj D 1 at the end of the formula). (Equivalently, this means we take Row1, Row2,
1
5 Row3):
23
1
6 7 D
A61 1 1377:
66
45
0 1
As described in the problem, we now have to consider
23
0
6 7
B I A 66 16 0 1377:
D D
45
0 0
The eigenvalues are obtained as the solutions of the characteristic equation (see pp. 326–327)
ˇ 1ˇ
ˇ 8ˇ
ˇˇ
ˇ ˇ D det.B I/ˇˇ 1
13ˇˇ
6
ˇ ˇ
ˇ4 ˇ
ˇ5 0 ˇ
3
D C D 0:

A sketch, as given below, shows that there is a real root near 0:5, but there are no further real
3
roots because, for large jj; the curve comes closer and closer to the curve of . Hence the other
eigenvalues must be complex conjugates. A root-finding method (see Sec. 19.2, pp. 801–806, also
Prob. 21 in the Student Solutions Manual on p. YY) gives a more accurate value of 0:5196. Division of
the characteristic equation by C 0:5196 gives the quadratic equation

2
C 0:5196 0:1283 D 0:

The roots are 0:2598 ˙ 0:2466i [by the well-known root-finding formula (4) for quadratic equations on
p. 54 of the textbook or on p. 15 in Volume I of the Student Solutions Manual]. Since all three roots
are less than 1 in absolute value, that is,
q
18 Numeric Analysis Part E

j 0:2598 ˙ 0:2466ij D .0:2598/2 C .˙0:2466/2 [by (3), p. 613]

D 0:3582 < 1 j

0:5196j D 0:5196 < 1;

the spectral radius is less than 1, by definition. This is necessary and sufficient for convergence (see at
the end of the section at the top of p. 863).

y
0.8

0.4

–1.2 –0.8 –0.4 0 0.4 0.8 1.2 x

–0.4

–0.8

Sec. 20.3 Prob. 17. Curve of the characteristic polynomial

19. Matrix norms. The given matrix is

2 3

10 1 1

C D 66 1 10 177:

4 5

1 1 10

All the norms are given on p. 861. The Frobenius norm is

vu 3 3

(9) kCk D utXXcjk2

jD1 kD1

p
D 102 C 1 C 1 C 1 C 102 C 1 C 1 C 1 C 102 D p303
D 17:49:
The column “sum” norm is
3

Xˇ ˇ
(10) kCk D max ˇcjk ˇ D 12:
k jD1
Chap. 20 Numeric Linear Algebra 19

Note that, to compute (10), we took the absolute value of each entry in each column and added
them up. Each column gave the value of 12. So the maximum over the three columns was 12.
Similarly, by (11), p. 861, the row “sum” norm is 12:
Together this problem illustrates that the three norms usually tend to give values of a similar order
of magnitude. Hence, one often chooses the norm that is most convenient from a computational point
of view. However, a matrix norm often results from the choice of a vector norm. When this happens,
we are not completely free to choose the norm. This new aspect will be introduced in the next section
of this chapter.

Sec. 20.4 Linear Systems: Ill-Conditioning, Norms

A computational problem is called ill-conditioned (p. 864) if small changes in the data cause large changes
in the solution. The desirable counterpart, where small changes in data cause only small changes in the
solution, is labeled well-conditioned. Take a look at Fig. 445 at the bottom of p. 864. The system in (a) is
well-conditioned. The system shown in part (b) is ill-conditioned because, if we raise or lower one of the
lines just a little bit, the the point of intersection (the solution) will move substantially, signifying ill-
conditioning. Example 1, p. 865, expresses the same idea in an algebraic example.
Keeping these examples in mind, we move to the central concept of this section, the condition number
.A/ of a square matrix on p. 868:

(13) .A/ D kAk jjAjj 1 :

Here is the Greek letter kappa (see back inside cover of textbook), kAk denotes the norm of matrix A, and
A 1 denotes the norm of its inverse. We need to backtrack and look at the concept of norms, which is of
general interest in numerics.
Vector norms kxk for column vectors x D xj with n components (n fixed), p. 866, are generalized
concepts of length or distance and are defined by four properties (3). Most common are the l 1-norm (5),
“Euclidean” or l2-norm (6), and l1-norm (7)—all illustrated in Example 3, p. 866. Matrix norms, p. 867, build
on vector norms and are defined by k D kAxk ¤

(9) kA max k k .x 0/.

x¤0 x

We use the l1-norm (5) for matrices—obtaining the column “sum” norm (10)—and the l 1-norm (7) for
matrices—obtaining the row “sum” norm (11)—both on p. 861 of Sec. 20.3. Example 4, pp. 866–867,
illustrates this. We continue our discussion of the condition number.
We take the coefficient matrix A of a linear system Ax D b and calculate .A/: If .A/ is small, then the
linear system is well-conditioned (Theorem 1, Example 5, p. 868).
We look at the proof of Theorem 1. We see the role of .A/ from (15), p. 868, is that a small condition
number gives a small difference in the norm of xxQ between an approximate solution xQ and the unknown
exact solution x of a linear system Ax D b:
Problem 9 gives a complete example on how to compute the condition number .A/ for the well-
conditioned case. Contrast this with Prob. 19, which solves an ill-conditioned system by Gauss elimination
20 Numeric Analysis Part E

with partial pivoting and also computes the very large condition number .A/. See also Example 1, p. 865,
and Example 6, p. 869.
Finally, the topic of residual [see (1), p. 865] is explored in Example 2, p. 865, and Prob. 21.
There is no sharp dividing line between well-conditioned and ill-conditioned as discussed in “Further
Comments on Condition Numbers” at the bottom of p. 870.
Problem Set 20.4. Page 871

9. Matrix norms and condition numbers. From the given matrix

" #

2 1

AD

1 4

we compute its inverse by (4*), p. 304, in Sec. 7.8:

2 "4 1# "1 #
1

A D D :
24 10 0 2 0

We want the matrix norms for A and A 1, that is, kAk and A 1 . We begin with the l1-vector norm,
which is defined by (5), p. 866. We have to remember that the l 1-vector norm gives, for matrices, the
column “sum” norm (the “sum” indicating that we take sums of absolute values) as explained in the
blue box in the middle of p. 867. This gives, under the l1-norm [summing over the absolute
k
values of the entries of each column i (here i D 1;2) and then selecting the maximum], Ak D max
fj2j C j0j ; j1j C j4jg D max fj2j ; j5jg D 5;
i i

and

A 1 D max˚ˇ1ˇ C j0j ;ˇˇ 18ˇˇ C ˇˇ14ˇˇ D maxi ˚ˇˇ12ˇˇ;ˇˇ38ˇˇ D:

ˇ 2ˇ i

Thus the condition number is

.A/ D kAk A 1 D 5 D 2:5 [by definition, see (13), p. 868].

Now we turn to the l1-vector norm, defined by (7), p. 866. We have to remember that this vector
norm gives for matrices the row “sum” norm. This gives, under the l 1-norm [summing over the
absolute values of the entries of each row j .in our situation j D 1; 2/ and then selecting the
k
maximum], Ak D max fj2j C j1j ; j0j C j4jg D max fj2j ; j4jg D 4;
j j
Chap. 20 Numeric Linear Algebra 21

and

1 D ˚ˇ1ˇ C ˇ 1ˇ; j0j C ˇˇ14ˇˇ D maxj˚ˇˇ58ˇˇ;ˇˇ14ˇˇ D:

A max ˇ2ˇ ˇ 8ˇ j

Thus the condition number is

.A/ D kAk A 1 D 4 D 2:5 [by definition, see (13), p. 868].

Since the value of the condition number is not large, we conclude that the matrix A is not ill-
conditioned.
19. An ill-conditioned system

1. Solving Ax D b1. The linear system written out is

(1) 4:50x1 C 3:55x2 D 5:2; (2)3:55x1 C 2:80x2 D 4:1:

The coefficient matrix A; given in the problem, is

" # " #
4:50 3:55 5:2
AD and b1 D :
3:55 2:80 4:1

We use Gauss elimination with partial pivoting (p. 846) to obtain a solution to the linear system.
We form the augmented matrix (pp. 845, 847):
"#
4:50 3:555:2
ŒAjb1 D:
3:55 2:804:1

We pivot 4:5 in Row 1 and use it to eliminate 3:55 in Row 2, that is,

Row 2 Row 1; which is, Row 2 0:7888888889 Row 1;

and get
"#
4:5 3:555:2
:
0 0:0005555555950:00222222228

Back substitution (p. 847) gives us by Eq. (2)

x2 D D 3:999999992 4:

Substituting this into (1) yields

22 Numeric Analysis Part E

x1 D .5:2 3:55 x2/ D .5:2 3:55 4/ D 2:

2. Solving Ax D b2. The slightly modified system is

(1) 4:50x1 C 3:55x2 D 5:2; (3)3:55x1 C 2:80x2 D 4:0:

The coefficient matrix A is as before with b2 slightly different from b1; that is,

" #
5:2 b2 D
:
4:0
We form the augmented matrix

"4:50 3:555:2#
ŒAjb2 D
3:55 2:804:0

and use Gauss elimination with partial pivoting with exactly the same row operation but startlingly
different numbers!

"4:5 3:555:2 #

0 0:000555950:1022228Row 2 0:788889 Row 1.

(There will be a small, nonzero, value in the a21 position due to using a finite number of digits.) Back

substitution now gives, by (3),

x2 D D 183:87 184

and hence, by (1),

x1 D .5:2 3:55 x2/ D .5:2 3:55 184/ D 144:

3. Computing the condition number of A. First, we need the inverse of A: By (4*), p. 304, we have
1 1 " 2:80 #
3:55
A D 3:55
3:55/ . 3:55/ 4:50
2:80 4:50 .
" 2:80 # "1120 #
3:55 1420
D 400
4:50 D 1420 :
3:55
1800
The l1-norm for matrix A; which we obtain by summing over the absolute values of the entries of each

column i .here i D 1;2/ and then selecting the maximum kAk D maxi fj2:80j C j3:55j ; j3:55j C j4:50jg D

maxi fj6:35j ; j8:05jg D 8:05;

Chap. 20 Numeric Linear Algebra 23

and similarly for A 1

A 1 D maxi fj1120j C j1420j ; j1420j C j1800jg D maxi fj2540j ; j3220jg D 3220:

Then by (13), p. 868, the condition number is

.A/ D kAk A 1 D 8:05 3220 D 25921.

Furthermore, because matrix A is symmetric (and, consequently, so is its inverse A 1), the values of the
l1-norm, i.e., the row “sum” norm, for both matrices A and A 1 are equal to their corresponding values of
the l1-norm, respectively. Hence the computation of .A/ would yield the same value.

4. Interpretation and discussion of result. The condition number .A/ D 25921 is very large, signifying that
the given system is indeed very ill-conditioned. This was confirmed by direct calculations in steps 1 and 2
by Gauss elimination with partial pivoting, where a small change by 0:1 in the second component from b 1
to b2 causes the solution to change from Œ 2;4T to Œ 144;184T, a change of about 1,000 times that of
that component! Note that we used 10 decimals in our first set of calculations to get satisfactory results.
You may want to experiment with a small number of decimals and see how you get nonsensical results.
Furthermore, note that the two rows of A are almost proportional.

21. Small residuals for very poor solutions. Use (2), p. 865, defining the residual of the “approximate
solution” Œ 10:0 14:1T of the actual solution Œ 2 4T; to obtain
" # " #" #
5:2 r D 3:55 10:0
4:50
4:1 2:80 14:1
3:55
" # " #
5:2 5:055
D 4:1 3:980
" # 0:145
D :
0:120

While the residual is not very large, the approximate solution has a first component that is 5 times
that of the true solution and a second component that is 3:5 times as great. For ill-conditioned
matrices, a small residue does not mean a good approximation.

Sec. 20.5 Least Squares Method

We may describe the underlying problem as follows. We obtained several points in the xy-plane, say by
some experiment, through which we want to fit a straight line. We could do this visually by fitting a line in
such a way that the absolute vertical distance of the points from the line would be as short as possible, as
suggested by Fig. 447, p. 873. Now, to obtain an attractive algebraic model, if the absolute value of a point
to a line is the smallest, then so is the square of the vertical distance of the point to the line. (The reason we
do not want to use absolute value is that it is not differentiable throughout its domain.) Thus we want to fit
24 Numeric Analysis Part E

a straight line in such a way that the sum of the squares of the distances of all those points from the line is
minimal, i.e., “least”—giving us the name “least squares method.”
The formal description of fitting a straight line by the least squares method is given in (2), p. 873, and
solved by two normal equations (4). While these equations are not particularly difficult, you need some
practice, such as Prob. 1, in order to remember how to correctly set up and solve such problems on the
exam.
The least squares method also plays an important role in regression analysis in statistics. Indeed, the
normal equations (4) show up again in Sec. 25.9, as (10) on p. 1105.
We extend the method to fitting a parabola by the least squares method and obtain three normal
equations (8), p. 874. This generalization is illustrated in Example 2, p. 874, with Fig. 448 on p. 875, and in
complete detail in Prob. 9.
Finally, the most general case is (5) and (6), p. 874.

Problem Set 20.5. Page 875

1. Fitting by a straight line. Method of least squares. We are given four points .0;2/; .2;0/; .3; 2/; .5; 3/
through which we should fit algebraically (instead of geometrically or sketching approximately) a
straight line. We use the method of least squares of Example 1, on p. 873 in the textbook. This
requires that we solve the normal auxiliary quantities needed in Eqs. (4), p. 873 in the textbook. When
using paper and pencil or if you use your computer as a typesetting tool, you may organize the
auxiliary quantities needed in (4) in a table as follows:
xj yj xj2 x j yj
0 2 0 0
2 0 4 0
3 2 9 6
5 3 25 15
Sum 10 3 38 21

From the last line of the table we see that the sums are

X X X2 X
xj D 10; yj D 3; xj D 38; xj yj D 21;

and n D 4, since we used four pairs of values. This determines the following coefficients for the variables of
(4), p. 873:
(1) 4a C 10b D 3;
(2) and gives the augmented 10a C 38b D 21;
matrix

" 3#
4 10
:
10 38 21
This would be a nice candidate for Cramer’s rule. Indeed, we shall solve the system by Cramer’s rule (2),
(3), Example 1, p. 292 in Sec. 7.6. Following that page, we have

ˇ ˇ
Chap. 20 Numeric Linear Algebra 25

ˇ4 10ˇ
DdetAˇ D ˇ D D 4 38 10 10 D 152 100 D 52:
ˇ ˇ
ˇ10 38ˇ
Furthermore
ˇ ˇ
ˇ 3 4ˇ
ˇ ˇ
ˇ ˇ
D ˇ 21 10 ˇ D 3 10 . 21/ 4 D 114 C 210 D 96 D 24 D
a 1:846;
D D D 52 13

ˇ ˇ
ˇ4 3ˇ
ˇ ˇ
ˇ ˇ
D ˇ10 21 ˇ D 4 . 21/ . 3/ 10 D 84 . 30/ D 27 D
b 1:038: D D D 26
From this we immediately get our desired straight line:

y D a C bx

D 1:846 1:038x:
y
2

0
1 2 3 4 5 x
–1

–2

–3

Sec. 20.5 Prob. 1. Given data and straight line fitted by least squares. (Note that the
axes have equal scales)

9. Fitting by a quadratic parabola. A quadratic parabola is uniquely determined by three given points. In
this problem, five points are given. We can fit a quadratic parabola by solving the normal equations (8),
p. 874. We arrange the data and auxiliary quantities in (8) again in a table:
x y x2 x3 x4 xy x2y
26 Numeric Analysis Part E

2 3 4 8 16 6 12
3 0 9 27 81 0 0
5 1 25 125 625 5 25
6 0 36 216 1296 0 0
7 2 49 343 2401 14 98
Sum 23 4 85
123 719
4419 15

The last line of the table gives us

the following information:
X X
X3
X2
xD
x D 23; yD
719;
4; x D 123;
X4 X 85;
X2
x D 4419; xy D
15; xyD
with the number of points being n D 5: Hence, looking at (8) on p. 874, and using the sums just
obtained, we can carefully construct the augmented matrix of the system of normal equations:
23
5 23 1234
67
6 23 123 719157:
45
123 719 441985

The system of normal equations is

5b0 C 23b1 C 123b2 D 4; 23b0 C 123b1 C

719b2 D 15; 123b0 C 719b1 C 4419b2 D 85:

We use Gauss elimination but, noting that the largest numbers are in the third row, we swap the first and
third rows,
23 Row 3
123 719 441985
67
Row 1
6 23 123 719157
45
5 23 1234

Then we perform the following row reduction operations:

23
123 719 441985
Chap. 20 Numeric Linear Algebra 27

6 7
6 0 11:4472 107:317 0:89431 7 Row 2 Row 1
45
0 6:22764 56:6342 0:544715 Row 3 Row 1

23
123 719 441985
6 7
6 0 11:4472 107:317 0:89431 7 Row 2 Row 1
45
0 6:22764 56:6342 0:544715 Row 3 Row 1

23
123 719 441985
67
6 0 11:4472 107:3170:894317
4 5 Row 3 Row 2
0 0 1:749651:03125

Back substitution gives us, from the last row of the last matrix,

b2 D D 0:589404:

The equation in the second row of the last matrix is

11:4472b1 107:317b2 D 0:89431:

We use it to obtain a value for b1:

11:4472b1 D 0:89431 C 107:317b2

D 0:89431 C 107:317 . 0:589404/ D 62:3588

so that

b1 D D 5:44752:

Finally, from the first equation,

123b0 C 719b1 C 4419b2 D 85;

we get
123b0 D 85 719b1 4419b2
D 85 719 .5:44752/ 4419 .0:589404/
D 85 3916:73 C 2604:56
28 Numeric Analysis Part E

D 1397:19:

Hence

b0 D D 11:3592:

Rounding our answer, to 4S, we have

b0 D 11:36; b1 D 5:448; b2 D 0:5894:

Hence the desired quadratic parabola that fits the data by the least squares principle is

yD 11:36 C 5:447x 0:5894x2:

y
1

0 1 2 3 4 5 6 7 8 x

–1

–2

–3

–4

Sec. 20.5 Prob. 9. Given points and quadratic parabola fitted by least squares

11. Comparison of linear and quadratic fit. The figure on the next page shows that a straight line obviously
is not sufficient. The quadratic parabola gives a much better fit. It depends on the physical or other
law underlying the data whether the fit by a quadratic polynomial is satisfactory and whether the
remaining discrepancies can be attributed to chance variations, such as inaccuracy of measurement.
Calculation shows that the augmented matrix of the normal equations for the straight line is

" #
5 10 8:3
10 30 17:5

and gives y D 1:48 C 0:09x: The augmented matrix for the quadratic polynomial is
3
2

5 10 30 8:30
7
6

610 30 100 17:507

4 5

30 100 354 56:31

Chap. 20 Numeric Linear Algebra 29

and gives y D 1:896 0:741x C 0:208x2: For practice, you should fill in the details.
y
3

0 1 2 3 4 5 x

Sec. 20.5 Prob. 11. Fit by a straight line and by a quadratic parabola

Sec. 20.6 Matrix Eigenvalue Problems: Introduction

This section gives you the general facts on eigenvalues necessary for the understanding of the special
numeric methods to be discussed, so that you need not consult Chap. 8. Theorem 2 on similarity of
matrices is particularly important.

Sec. 20.7 Inclusion of Matrix Eigenvalues

The central issue in finding eigenvalues of an n n matrix is to determine the roots of the corresponding
characteristic polynomial of degree n. This is usually quite difficult and requires the use of an iterative
numerical method, say from Sec. 19.2, or from Secs. 20.8 and 20.9 for matrices with additional properties.
However, sometimes we may only want some rough approximation of one or more eigenvalues of the
matrix, thereby avoiding costly computations. This leads to our main topic of eigenvalue inclusion.
Gerschgorin was only 30 years old when he published his beautiful and imaginative theorem, Theorem 1, p.
879. Take a look at Gerschgorin’s theorem at the bottom of that page. Formula (1) says that the eigenvalues
of an n n matrix lie in the complex plane in closed circular disks. The centers of these disks are the
elements of the diagonal of the matrix, and the size of these disks are determined by the sum of the
elements off the diagonal in each corresponding row, respectively. Turn over to p. 880 and look at Example
1, which applies Gerschgorin’s theorem to a 3 3 matrix and gets three disks, so called Gerschgorin disks,
two of which overlap as shown in Fig. 449. The centers of these disks can serve as crude approximations of
the eigenvalues of the matrix and the radii of the disks as the corresponding error bounds.
Problems 1 and 5 are further illustrations of Gerschgorin’s theorem for real and complex matrices,
respectively.
Gerschgorin’s theorem (Theorem 1) and its extension (Theorem 2, p. 881) are types of theorems know as
inclusion theorems. Inclusion theorems (p. 882) are theorems that give point sets in the complex plane that
“include,” i.e., contain one or several eigenvalues of a given matrix. Other such theorems are Schur’s
theorem (Theorem 4, p. 882), Perron’s theorem (Theorem 5, p. 882) for real or complex square matrices,
and Collatz inclusion theorem (Theorem 6, p. 883), which applies only to real square matrices whose
elements are all positive. Be aware that, throughout Secs. 20.7–20.9, some theorems can only be applied to
certain types of matrices.
Finally, Probs. 7, 11, and 13 are of a more theoretical nature.

Problem Set 20.7. Page 884

1. Gerschgorin disks. Real matrix.

1. Determination of the Gerschgorin disks. The diagonal entries of the given real matrix (which we shall
denote by A)
30 Numeric Analysis Part E

2 2 43

A D 66 2 0 277

4 5

2 4 7

are 5; 0; and 7: By Gerschgorin’s theorem (Theorem 1, p. 879), these are the centers of the three desired
Gerschgorin disks D1, D2; and D3; respectively. For the first disk, we have the radius by
(1), p. 879,

ja11 j D j5 j ja12j C ja13j D j2j C j4j D 6;

so that

D1 W j5 j 6

or equivalently,

D1 W center 5; radius 6.

This means, to obtain the radius of a Gerschgorin disk, we add up the absolute value of the entries in the
same row as the diagonal entry (except for the value of the diagonal entry itself). Thus for the other two
Gerschgorin disks we have

D2 W center 0; radius 4 D .j 2j C j2j/;

D3 W center 7; radius 9 D .j2j C j7j/:

Below is a sketch of the three Gerschgorin disks. Note that they intersect in the closed interval
4 13.
y
6

–4 –2 0 2 4 6 8 10 12 14 x
–2

–4

–6

Sec. 20.7 Prob. 1. Gerschgorin disks. The disks have centers 5; 0; 7 and radii 6; 4; 6;
respectively
Chap. 20 Numeric Linear Algebra 31

2. Determination of actual eigenvalues. We compute the characteristic polynomial p./:

ˇ ˇ
ˇ5 2 4ˇ
ˇ ˇ p./det.A I/ˇˇ 2 2 ˇˇ
D D
ˇ ˇ
ˇ ˇ
ˇ2 4 7 ˇ
ˇ ˇ ˇ ˇ ˇ ˇ
ˇ 2 ˇ ˇ 2 2 ˇ ˇ 2 ˇ D .5 /ˇ ˇ 2ˇ ˇ C 4ˇ ˇ
ˇ ˇ ˇ ˇ ˇ ˇ
ˇ4 7 ˇ ˇ2 7 ˇ ˇ2 4ˇ
D .5 /Œ . 7 / 8 2Œ 2.7 / 4 C 4Œ 8 C 2  D 3 122 C 23 C 36 D 0:

We want to find the roots of the characteristic polynomial. We know the following observations:

F1. The product of the eigenvalues of a characteristic polynomial is equal to the constant term of that
polynomial.

F2. The sum of the eigenvalues is equal to . 1/n 1 times the coefficient of the second highest term of
the characteristic polynomial. (Another example is discussed on pp. 129–130, in Volume 1, of the
Student Solutions Manual).

Using these two facts, we factor the constant term 36 and get 36 D 1 2 2 3 3 3. We calculate,2
starting with the smallest factors (both positive as given) and negative: p.1/ D 1 12 1 C 23
1 C 36 D 48 ¤ 0;p. 1/ D 0: We found an eigenvalue! Thus a factor is . C 1/ and we could use long
division and apply the well-known quadratic formula for finding roots. Or we can continue: p.2/ D
42; p. 2/ D 66; p.4/ D 0: We found another eigenvalue. From F2, we know that the 3 1 sum of the
three eigenvalues must equal . 1/ 12 D 12: Hence 1 C 4 C D 12 so the other eigenvalue must
be equal to D 9: Hence the three eigenvalues (or the spectrum) are 1; 4; 9:

3. Discussion. The inclusion interval obtained from Gerschgorin’s theorem is larger; this is typical. But
the interval is the best possible in the sense that we can find, for a set of disks (with real or complex
centers), a corresponding matrix such that its spectrum cannot be included in a set of smaller closed
disks with the main diagonal entries of that matrix as centers.

5. Gerschgorin disks. Complex matrix. To obtain the radii of the Geschgorin disks, we compute by
(1), p. 879,

p
ja12j C ja13j D jij C j1 C ij D p12 C 12 C 12 D 1 C p2 [by (3), p. 613], ja21j C ja23j j ij C j0j D

1;
32 Numeric Analysis Part E

p
ja31j C ja32j D j1 ij C j0j D 12 C . 1/2 C 0 D p2:

The diagonal elements, and hence centers of the Gerschgorin disks, are

a11 D 2; a22 D 3; a33 D 8:

Putting it all together: The disks are D1 W center in Prob. 1. You may want to sketch the Gerschgorin
disks and determine in which closed interval they intersect.
The determination of the actual eigenvalues is as follows. Developing the determinant along the last
row, with the usual checkerboard pattern in mind giving the correct plus and minus signs of the
cofactors (see bottom of p. 294), we obtain

ˇˇ2 i 1 C iˇˇ
ˇ
ˇ ˇ p./ D det.A I/ D ˇˇ i
ˇ
3 0ˇ

ˇ
ˇˇ1 i 0 8 ˇˇ

D ˇˇˇˇ i 1 C iˇˇˇˇ C ˇˇˇˇˇ2 i ˇˇˇˇ

.1 i/ 0 .8 /
ˇ3 0ˇ i 3 ˇ

D .1 i/Œ0 .1 C i/.3 / C .8 /Œ.2 /.3 / 1 D 3 C 132 43 C 34 D 0:

The constant term of the characteristic polynomial is 34 and factors as follows:

34 D 1 2 17:

However, none of its positive and negative factors, when substituted into the characteristic polynomial,
yields p./ equal to zero. Hence we would have to resort to a root-finding method from Sec. 19.2, p. 802,
such as Newton’s method. A starting value, as suggested by Geschgorin’s theorem, would be D 1:0000:
However, the problem suggests the use of a CAS (if available). Using a CAS (here Mathematica), the
spectrum f1;2; 3g is

1 D 1:16308;

2 D 3:51108; 3 D

8:32584:

Comment. We initially tried to use the approach of Prob. 1 when we determined the characteristic
polyomial, factored the constant term, and then tried to determined whether any of these factors
yielded zeros. This was to show that we first try a simpler approach and then go to more involved
methods.
Chap. 20 Numeric Linear Algebra 33

7. Similarity transformation. The matrix in Prob. 2 shows a typical situation. It may have resulted from a
numeric method of diagonalization that left off-diagonal entries of various sizes but not exceeding 10 2
in absolute value. Gerschgorin’s theorem then gives circles of radius 2 10 2. These furnish bounds for
the deviation of the eigenvalues from the main diagonal entries. This describes the starting situation for
the present problem. Now, in various applications, one is often interested in the eigenvalue of largest or
smallest absolute value. In our matrix, the smallest eigenvalue is about 5, with a maximum possible
deviation of 2 10 2, as given by Gerschgorin’s theorem. We now wish to decrease the size of this
Gerschgorin disk as much as possible. Example 2, on p. 881 in the text, shows us how we should
proceed. The entry 5 stands in the first row and column. Hence we should apply, to A; a similarity
transformation involving a diagonal matrix T with main diagonal a;1;1, where a is as large as possible.
The inverse of T is the diagonal matrix with main diagonal 1=a;1;1. Leave a arbitrary and first determine
the result of the similarity transformation (as in Example 2).

32 32
2 0 5 0:01 0:01 a 0 03

1=a 0
6
B D T 1AT D 6 0 1 077 660:01 8 0:0177 660 1 077

4 54 54 5

0 0 1 0:01 0:01 9 0 0 1
3
2 0:01=a

5 0:01=a

D 6 7
60:01a 8 0:01 7:
4 5
0:01a 0:01 9
We see that the Gerschgorin disks of the transformed matrix B; by Gerschgorin’s theorem, p. 879, are

Center Radius
5 0:02=a
8 0:01.a C 1/

9 0:01.a C 1/

The last two disks must be small enough so that they do not touch or even overlap the first disk. Since 8
5 D 3, the radius of the second disk, after the transformation, must be less than
3 0:02=a; that is,
0:01.a C 1/ < 3 0:02=a:

Multiplication by 100a (> 0) gives

a2 C a < 300a 2:
34 Numeric Analysis Part E

If we replace the inequality sign by an equality sign, we obtain the quadratic equation

a2 299a C 2 D 0:

Hence a must be less than the larger root 298:9933 of this equation, say, for convenience, a D 298.
Then the radius of the second disk is 0:01.a C 1/ D 2:99, so that the disk will not touch the first one,
and neither will the third, which is farther away from the first. The first disk is substantially reduced

in size, by a factor of almost 300, the radius of the reduced disk being D 0:000067114:

The choice of a D 100 would give a reduction by a factor 100, as requested in the problem. Our
systematic approach shows that we can do better.

For a D 100 the computation is

2 32 32 3 2 3
0:01 0 0 5:00 0:01 0:01 100 0 0 5 0:0001 0:0001
6 76 76 7 6 7
60 1 07 60:01 8:00 0:017 6 0 1 07 D 61 8 0:01 7:
4 54 54 5 4 5
0 0 1 0:01 0:01 9:00 0 0 1 1 0:01 9

Remark. In general, the error bounds of the Gerschgorin disk are quite poor unless the off-diagonal
entries are very small. However, for an eigenvalue in an isolated Gerschgorin disk, as in Fig. 449,
p. 880, it can be meaningful to make an error bound smaller by choosing an appropriate similarity
transformation

B D T 1AT;

where T is a diagonal matrix. Do you know why this is possible? Answer: This is allowed by Theorem 2,
p. 878, which ensures that similarity transformations preserve eigenvalues. So here we picked the
smallest eigenvalue and made the error bound smaller by a factor 1=100 as requested.

11. Spectral radius. By definition (see p. 324), the spectral radius of a square matrix A is the absolute value
of an eigenvalue of A that is largest in absolute value. Since every eigenvalue of A lies in a Gerschgorin
disk, for every eigenvalue of A we must have (make a sketch)

(I) j ajj j C Xˇˇajk ˇˇ ˇˇj ˇˇ

where we sum over all off-diagonal entries in Row j (and the eigenvalues of A are numbered suitably).
Since (I) is true for all eigenvalues of A, it must be true for the eigenvalue of A that is largest in
ˇ ˇ
absolute value, that is, the largest ˇj ˇ: But this is, by definition, the spectral radius .A/: The left-hand
side of (I) is precisely the row “sum” norm of A: Hence, we have proven that
Chap. 20 Numeric Linear Algebra 35

the row “sum” norm ofA .A/:

13. Spectral radius. The row norm was used in Prob. 11, but we could also use the Frobenius norm
s ˇ ˇXX 2
ˇj ˇcjk [see (9), p. 861]
j k

to find the upper bound. In this case, we would get (calling the elements a jk, since we called the matrix
in Prob. 1 A)

vu 3 3

ˇj ˇˇ utXX 2

ˇ ajk
jD1 kD1

D p52 C 22 C 42 C . 2/2 C 02 C 22 C 22 C 42 C 72 D p122

D 11:05:

Sec. 20.8 Power Method for Eigenvalues

The main attraction of the power method is its simplicity. For an n n matrix A with a dominant
eigenvalue (“dominant” means “largest in absolute value”) the method gives us an approximation (1), p.
885, usually of that eigenvalue. Furthermore, if matrix A is symmetric, that is, a jk D akj [by (1),
p. 335], then we also get an error bound (2) for approximation (1). Convergence may be slow but can be
improved by a spectral shift (Example 2, p. 887). Another use for a spectral shift is to make the method
converge to the smallest eigenvalue as shown in Prob. 11. Scaling can provide a convergent sequence of
eigenvectors (for more information, see Example 1, p. 886). The power method is explained in great detail
in Prob. 5.

More details on Example 1, pp. 886–887. Application of Power Method, Error Bound (Theorem 1, p. 885).
Scaling. We take a closer look at the six vectors listed at the beginning of the example:

23 2 3 2 3
1 0:890244 0:890244
6 7 D
x0 D 66177; x1 D 60:609756 7; x2660:60975677;
45 4 54 5
1 1 1

2 3 2 3 2 3

0:990663 0:999707 0:999991

36 Numeric Analysis Part E

6 7
x5 D 60:504682 7; x10 D 660:50014677; x15 D 660:50000577:

4 5 4 5 4 5

1 1 1

Vector x0 was scaled. The others were obtained by multiplication by the given matrix A and subsequent
scaling. We can use any of these vectors for obtaining a corresponding Rayleigh quotient q as an
approximate value of an (unknown) eigenvalue of A and a corresponding error bound ı for q. Hence we
have six possibilities using one of the given vectors, and indeed many more if we want to compute further
vectors. Note that we must not use two of the given vectors because of the scaling, but just one vector. For
instance, if we use x1, and then its product Ax1 we get

3
2 2 3 2 3

0:49 0:02 0:22 0:890244 0:668415

6 7 6 7 6 7
A D 60:02 0:28 0:20 7; x1 D 60:609756 7 ; Ax1 D 60:388537 7:

4 5 4 5 4 5

0:22 0:20 0:40 1 0:717805

From these data we calculate the inner products by Theorem 1, p. 885,

m0 D x1TTx1 D 2:164337; m1 D x1Ax1T D

1:549770; m2 D .Ax1/ Ax1 D

1:112983:

These now give the Rayleigh quotient q and error bound ı of q by (1), (2) p. 885:

q D m1=m0 D 0:716048;

p
ı D m2=m0 q2 D 0:038887;

where q approximates the eigenvalue 0.72 of A, so that the error of q is

D 0:72 q D 0:003952:

These values agree with those for j D 2 in the table for Example 1 on p. 887 of the textbook.

Problem Set 20.8. Page 887

5. Power method with scaling. The given matrix is

3
2 1
Chap. 20 Numeric Linear Algebra 37

2 1

A D 66 1 3 277:

4 5

1 2 3

Use the same notation as in Example 1 in the text. From x 0 D Œ1 1 1T calculate Ax0 and then scale it as
indicated in the problem, calling the resulting vector x 1. This is the first step. In the second step
calculate Ax1 and then scale it, calling the resulting vector x 2. And so on. More details are as follows:
Iteration 1: We start with

2 3
1

x0 D 66177:
45
1
Multiplication by the given matrix A gives us

2 32 3 23

2 1 1 1 2

Ax0 D 66 1 3 277661 77 D 664 77:

4 54 5 45

1 2 3 1 6

The calculations that give approximations q (Rayleigh quotients) and error bounds are as follows.
For m0; m1; and m2

2 3
1
6 7
m0 D xT0x0 D 1 1 1 61 7 D 1 1 C 1 1 C 1 1 D 3;

45

23

2
6 7
m1 D xT0Ax0 D 1 1 1 64 7 D 1 2 C 1 4 C 1 6 D 12;

45
38 Numeric Analysis Part E

23

2
6 7
m2 D .Ax0/T Ax0 D 2 4 6 64 7 D 2 2 C 4 4 C 6 6 D 56;

45

6
m1 12 q D D D 4:
m0 3

We know that ı2 D m2=m0 q2; so

2 m2 2 56 2
ıD q D 4 D 18:66667 16 D 2:666667; m0 3

ı D p2:666667 D 1:632993;

q ı D 4 1:632993 D 2:367007; q C ı D 4 C 1:632993 D

5:632993:

Iteration 2: If this is not sufficient, we iterate by using a scaling factor. We chose the absolute largest
component of Ax0: This is 6, so we get
23
2 3
0:3333333
67
D
x1664677 D 660:666666777:

67 4 5
45 1

Again, we multiply this vector by the given matrix A:

32 3 2 3
2
2 1 1 0:3333333 1
7 6 7 6 7
Ax1 D 66 1 3 2 7 60:6666667 7 D 63:666667 7:
4 54 5 4 5
1 2 3 1 4:666667
Chap. 20 Numeric Linear Algebra 39

As before, we compute the values required to obtain our next approximation to q and ı:

2 3
0:3333333
6 7
m0 D xT1x1 D 0:3333333 0:6666667 1 60:6666667 7 D 1:555556;
4 5
1
2 3
1
6 7
m1 D xT1Ax1 D 0:3333333 0:6666667 1 63:666667 7 D 7:444445;

4 5

4:666667

2 3

1
6 7
m2 D .Ax1/T Ax1 D 1 3:666667 4:6666676 3:666667 7 D 36:22223;
4 5
4:666667
m1 7:444445
q D D D 4:785713; m0 1:555556
2 2 2 36:22223 2m

ıD qD .4:785713/ ; m0 1:555556
D 23:28571 22:90305 D 0:38266:

It is important to notice that we have a loss of significant digits (subtractive cancelation) in the computation
of ı: The two terms that are used in the subtraction are similar and we go from seven digits to five. This
suggests that, for more than three iterations, we might require our numbers to have more digits.

ı D p0:38266 D 0:6185952; q ı D 4:785713

0:6185952 D 4:167118; q C ı D 4:785713 C 0:6185952 D

5:404308:

Iteration 3: Again, if the result is not good enough, we need to move to the next iteration by using the
largest value of Ax1 as our scaling factor. This is 4:666665 so we get for x 2

2 3 2 3
1=4:666667 0:2142857
6 7 6 7
x2 D 63:666667=4:666667 7 D 60:7857143 7;
4 5 4 5
40 Numeric Analysis Part E

4:666667=4:666667 1
from which

32 3 2 3
2

2 1 1 0:2142857 0:6428571
7 6 7 6 7
Ax2 D 66 1 3 2 7 60:78571437 D 64:142857 7:

4 54 5 4 5

1 2 3 1 4:785714

This is followed by one more scaling step for the final result of x 3:

2 3 2 3
0:6428571=4:785714 0:1343284
6 7 6 7
x3 D 64:142857=4:785714 7 D 60:8656717 7;
4 5 4 5
4:785714=4:785714 1
2 3
0:2142857
6 7
m0 D xT2x2 D 0:2142857 0:7857143 1 60:7857143 7 D 1:663265;

4 5

2 3

0:6428571
6 7
m1 D xT2Ax2 D 0:2142857 0:7857143 1 64:142857 7 D 8:178571;

4 5

4:785714

2 3

0:6428571
6 7
m2 D .Ax2/T Ax2 D 0:6428571 4:142857 4:785714 64:142857 7
4 5
4:785714 D
40:47959;
Chap. 20 Numeric Linear Algebra 41

m1 8:178571
q D D D 4:917179; m0 1:663265

2 m2 2 40:47959 2

ıD qD .4:917179/ m0 1:663265
D 24:33743 24:17865 D 0:1587774;

ı D p0:1587774 D 0:3984688;

q ı D 4:917179 0:3984688 D 4:51871; q C ı D 4:917179

C 0:3984688 D 5:315648:

The results are summarized and rounded in the following table. Note how the value of the ı gets smaller so
that we have a smaller error bound on q.
m0 xT0x0 D 3 xT1x1 D 1:55556 xT2x2 D 1:663
m1 xT0Ax0 D 12 xT1Ax1 D 7:44444 xT2Ax2 D 8:179
m2 .Ax0/T Ax0 D 56 Ax1T Ax1 D 36:22 .Ax2/T Ax2 D 40:48
m2 q
D 4:786
4 4:917
m0
m
ı2 D 2 q2 2:667 0:3826 0:1588
m0
ı 1:633 0:6186 0:3985
q ı 2:367 4:167 4:519
qCı 5:633 5:404 5:316
Solving the characteristic equation x3 C 8x2 15x shows that the matrix has the eigenvalues

0;3, and 5. Corresponding eigenvectors are

z1 D Œ0 1 1T; z2 D Œ 1 1 1T; z3 D Œ 2 1 1T;

respectively. We see that the interval obtained in the first step includes the eigenvalues 3 and 5. Only in the
second step and third step of the iteration did we obtain intervals that include only the largest eigenvalue,
as is usually the case from the beginning on. The reason for this interesting observation is the fact that x 0 is
a linear combination of all three eigenvectors,

x0 D z1 13.z2 C z3/;

as can be easily verified, and it needs several iterations until the powers of the largest eigenvalue
make the iterate xj come close to z1, the eigenvector corresponding to D 5. This situation occurs
quite frequently, and one needs more steps for obtaining satisfactory results the closer in absolute
value the other eigenvalues are to the absolutely largest one.

11. Spectral shift, smallest eigenvalue. In Prob. 3,

2 3
42 Numeric Analysis Part E

1 1 1

BDA 3I D 66 1 0 277:

4 5

1 2 0

Now the power method converges to the eigenvalue max of largest absolute value. (Here we assume
that the matrix does not have max as another eigenvalue.) Accordingly, to obtain convergence to
the smallest eigenvalue, make a shift to A C kI with a negative k. Choose k by trial and error,
reasoning about as follows. The given matrix has trace A D 2 C 3 C 3 D 8. This is the sum of the
eigenvalues. From Prob. 5 we know that the absolutely largest eigenvalue is about 5. Hence the
sum of the other eigenvalues equals about 3. Hence k D 3 suggested in the problem seems to be
a
reasonable choice. Our computation of the Rayleigh quotients and error bounds gives for the first
step x0 D Œ1 1 1T; x1 D Œ 1 1 3T; m0 D 3; m1 D 3; m2 D 11; q D 1;
q11 q8
ıD 3 1D 3, and so on, namely,

q 1 0:63636 0:28814 1:2749 2:0515 2:5288 2:7790 2:8993 2:9547 ı1:6323

2:2268 2:4910 2:3770 1:9603 1:4608 1:0277 0:70234 0:47355

We see that the Rayleigh quotients seem to converge to 3, which corresponds to the eigenvalue
0 of the given matrix. It is interesting that the sequence of the ı is not monotone; ı first increases
and starts decreasing when q gets closer to the limit 3. This is typical. Also, note that the
error bounds are much larger than the actual errors of q. This is also typical.

Sec. 20.9 Tridiagonalization and QR-Factorization

Somewhat more recent developments in numerics provided us with a widely used method of computing
all the eigenvalues of an n n real symmetric matrix A. Recall that, in such a special matrix, its entries off
the main diagonal are mirror images, that is, a jk D akj [by (1), p. 335].
In the first stage, we use Householder’s tridiagonalization method (pp. 889–892) to transform the matrix
A into a tridiagonal matrix B (“tri” D “three”), that is, a matrix having all its nonzero entries on the main
diagonal, in the position immediately below the main diagonal, or immediately above the main diagonal
(Fig. 450, matrix in Third Step, p. 889). In the second stage, we apply the QR-factorization method (pp.
892–896) to the tridiagonal matrix B to obtain a matrix B sC1 whose real diagonal entries are approximations
of the desired eigenvalues of A (whereby the nonzero entries are sufficiently small in absolute value). The
purpose of the first stage is to produce many zeros in the matrix and thus speed up the convergence for
the QR method in the second stage.
Perhaps the easiest way to understand Householder’s tridiagonalization method is to go through Example
1, pp. 890–891. A further illustration of the method is given in Prob. 3. Similarly, another good way to
Chap. 20 Numeric Linear Algebra 43

understand the QR-factorization method is to work through Example 2, pp. 894–896 with a further
demonstration of the method in Prob. 7. Both examples and both problems are each concerned with the
same real symmetric matrices, respectively.
An outline of this section is as follows: Discussion of the problem and biographic reference to

Householder’s tridiagonalization method, p. 888.

Householder’s tridiagonalization method (pp. 889–892).

Formula (1), on p. 889, is the general set of formulas for the similarity transformations P r to obtain, in
stages, the tridiagonal matrix B:
Figure 450 illustrates, visually, how a 5 5 matrix A gets transformed into A 1;A2;A3 so that at the end
B D A3:

Formulas (2) and (3), p. 889, show the general form of the similarity transformations P r and associated unit
vectors vr.
The important formula (4), on the top of p. 890, defines the components of the unit vectors v r of (2) and (3).
Notice, in 4(b), sgn a21 is the sign function. It extracts the sign from a number, here a 21. This function gives
“plus one” when a number is zero or positive and “minus one” when a number is negative. Thus, for
example,

sgn 8 D C1, sgn 0 D C1; sgn . 55/ D 1:

For each interation in formula (4) we increase, by 1, all subscripts of the components of the column
vector(s) vr .r D 2 for step 2/: We iterate n 2 times for an n n matrix.

Example 1, on p. 890, illustrates the method in detail.

Proof, p. 891, of Formula (1)
QR-factorization method (pp. 892–896).

Biographic references to the QR-factorization method, p. 892.

Assuming that Householder’s Tridiagonalization Method has been applied first to matrix A; we start with
tridiagonal matrix B D B0. Two different kind of matrices in Step 1, p. 892: orthogonal matrix Q0 (means
Q01 D QT0/ and upper triangular matrix R0: Step consists first of factorization (“QR-factorization”) and then
computation.
Formula (5) gives General Step with matrices Qs and Rs with 5(a) factorization (QR-factorization) and 5(b)
computation.
Proof, p. 892, of Formula (5).
Detailed outline on how to get the 5(a) factorization (QR-factorization), p. 892. The method needs
orthogonal matrices Cj that contain 2 2 plane rotation submatrices, which for n D 4 can be determined
by (11).
How to get 5(b) computation from 5(a), p. 892.
44 Numeric Analysis Part E

Example 2, on p. 894, illustrates the method in detail.

More Details on Example 2, p. 894. QR-Factorization Method. The tridiagonalized matrix is (p. 895)
p
23
618 0
6 7
D
B6 p18 7 p277:
6
4 p 5
0 2 6
We use the abbreviations c2;s2, and t2 for cos2, sin2, and tan2, respectively. We multiply B from the left by

2 3 c2 s2 0

C2 D 66 s2 c2 077:
4 5
0 0 1

The purpose of this multiplication is to obtain a matrix C 2B D Œ bjk.2/ for which the off-diagonal entry b21.2/

is zero. Now this entry is the inner product of Row 2 of C 2 times Column 1 of B, that is,

qq

s2 6 C c2. p18/ D 0; thus t2 D D :

From this and the formulas that express cos and sin in terms of tan we obtain
qq
c2 D 1= 1 C t22 D 2
3 D 0:816496581;

q q s2 D t2= 1 C t22 D D
0:577350269:

b .3/ Œb .3/
3 is determined similarly, with the purpose of obtaining 32 D 0 in C3C2B D jk .
Problem Set 20.9. Page 896
3. Tridiagonalization. The given matrix

2 3

7 2 3

A D 662 10 677

4 5
Chap. 20 Numeric Linear Algebra 45

3 6 7

is symmetric. Hence we can apply Householder’s method for obtaining a tridiagonal matrix (which will
have two zeros in the location of the entries 3). Proceed as in Example 1 of the text. Since A is of size n
D 3, we have to perform n 2 D 1 step. (In Example 1 we had n D 4 and needed n 2 D 2 steps.)
Calculate the vector v1 from (4), p. 890. Denote it simply by v and its components by v 1.D 0/, v2; v3
because we do only one step. Similarly, denote S1 in (4c) by S. Compute

q p2
S D a212 C 2D C 32 D p13 D 3:60555: a31 2
If we compute, using, say, six digits, we may expect that, instead of those two zeros in the
tridiagonalized matrix, we obtain entries of the order 10 6 or even larger in absolute value. We always
have v1 D 0. From (4a) we obtain the second component

r r
1 C a21=S D 1 C 2=3:60555 D v2
D0:881675:
2 2

From (4b) with j D 3 and sgn a21 D C1 (because a21 is positive) we obtain the third component

a31 3 v3 D D D 0:471858:
2v2 S 2 0:881675 3:60555
With these values we now compute Pr from (2), where r D 1; ::: ;n 2, so that we have only r D 1 and
can denote P1 simply by P. Note well that vTv would be the dot product of the vector by itself (thus the
square of its length), whereas vvT is a 3 3 matrix because of the usual matrix multiplication. We thus
obtain from (2), p. 889,
PDI 2vvT

D
2 2 v1v2 v1
6
I 2 6v2v1 v22 v1v33
7
6 7 v 2v 3 7
D 4 5
v3v1 v3v2 v23
2 2 2v1v2
1 2v1 2v1v33
6 7
7
6 2v2v1 1 2v22 2v2v3 7
6 5
4 1 2v23
2v3v1 2v3v2
46 Numeric Analysis Part E

2 0 3
1:0 0
6 D 7
0:832051 7:
60 0:554702 5
4
0:554700
0 0:832051
1
Finally use P, and its inverse P D P, for the similarity transformation that will produce the
tridiagonal matrix

2 3
7:0 3:605556 0:000001
6 7
B D PAP D P 62:0 10:539321 4:992308 7
4 5
3:0 9:152565 1:109404
2 3
7:0 3:605556 0:000001
7
6 D

6 3:605556 13:461578 3:6923227:

4 5
0:000001 3:692322 3:538467

The point of the use of similarity transformations is that they preserve the spectrum of A, consisting of
the eigenvalues

2; 5; 16;

which can be found, for instance, by graphing the characteristic polynomial of A and applying Newton’s
method for improving the values obtained from the graph.

7. QR-factorization. The purpose of this factorization is the determination of approximate values of all the
eigenvalues of a given matrix. To save work, one usually begins by tridiagonalizing the matrix, which
must be symmetric. This was done in Prob. 3. The matrix at the end of that problem

2 3
7:0 3:605551275 0
6 7
B0 D Œbjk D 6 3:605551275 13:46153846 3:692307692 7
4 5
0 3:692307692 3:538461538
is tridiagonal (note that greater accuracy is being used). Hence QR can begin. We proceed as in Example 2,
on p. 894, of the textbook. To save writing, we write c 2; s2; t2 for cos 2, sin 2, tan 2, respectively.

Step 1. Consider the matrix

2 3 c2 s2 0

C2 D 66 s2 c2 077
Chap. 20 Numeric Linear Algebra 47

4 5
0 0 1

Œw .0/ w .0/
with the angle of rotation 2 determined so that, in the product W 0 D C2B0 D jk ; the entry 21

is zero. By the usual matrix multiplication (row times column) w .210/ is the inner product of Row 2 of
C2 times Column 1 of B0, that is,

s2 b11.0/ C c2 b21.0/ D 0; hence t2 D s2=c2 D b21.0/=b11.0/:

From this, and the formulas for cos and sin in terms of tan (usually discussed in calculus), we obtain

r 2

c2 D 1= 1 C b21.0/=b11.0/ D 0:889000889;

(I/1)
b21.0/ r .0/ .0/2

s2 D .0/ = 1 C b21 =b11 D 0:4579054698: b11

Use these values in C2 and calculate C2B0 D W0 D Œwjk.0/. Thus

3
2

7:874007873 9:369450382 1:690727888

7
W0 D Œwjk.0/ D C2B0 D 660 10:31631801 3:282464821 7:

4 5

0 3:692307692 3:538461538

C2 has served its purpose: instead of b21.0/ D 3:605551276 we now have w21.0/ D 0. (Instead of w21.0/ D 0;
on the computer we may get 10 10 or another very small entry—the use of more digits in B 0 ensured the
0.) Now use the abbreviations c3; s3; t3 for cos 3, sin 3, tan 3. Consider the matrix
2 3

1 0 0

C3 D 660 c3 s377

4 5

0 s3 c3

with the angle of rotation 3 such that, in the product matrix R 0 D Œrjk D C3W0 D C3C2B0; the entry r32 is zero.
This entry is the inner product of Row 3 of C 3 times Column 2 of W0. Hence

s3 w22.0/ C c3 w32.0/ D 0; so that t3 D s3=c3 D w32.0/=w22.0/ D 0:357909449:

This gives, for c3 and s3;
48 Numeric Analysis Part E

q q
(II/1) c3 D 1= 1 C t32 D 0:9415130836; s3 D t3= 1 C t32 D 0:3369764287:

Using this, we obtain

2 3
7:874007873 9:369450382 1:690727888
6 7
R0 D C3W0 D C3C2B0 D 60 10:95716904 4:282861708 7:
4 5
0 0 2:225394561

(Again, instead of 0; you might obtain 10 10 or another very small term—similarly in the further calculations.)
Finally, we multiply R0 from the right by CT2CT3. This gives
B1 D R0CT2CT3 D C3C2B0CT2CT3

2 5:017347637 0 3

11:29032258

6 D 7
6 5:017347637 10:61443933 0:74990551287:
4 5
0 0:7499055119 2:095238095
The given matrix B0 (and, thus, also the matrix B1) has the eigenvalues 16, 6, 2. We see that the main
diagonal entries of B1 are approximations that are not very accurate, a fact that we could have concluded
from the relatively large size of the off-diagonal entries of B 1. In practice, one would perform further steps
of the iteration until all off-diagonal entries have decreased in absolute value to less than a given bound.
The answer, on p. A51 in App. 2, gives the results of two more steps, which are obtained by the following
calculations.

Œb .0/ Œb .1/
Step 2. The calculations are the same as before, with B 0 D jk  replaced by B1 D jk . Hence, instead

of (I/1), we now have

.1/ .1/ 2
D qC D 0:9138287756;
c2 1= 1 .b21 =b11 /

q
(I/2) s2 D .b21.1/=b11.1//= 1 C .b21.1/=b11.1//2 D 0:4060997031:

We can now write the matrix C2, which has the same general form as before, and calculate the product
W1 D Œwjk.1/ D C2B1
3
2

12:35496505 8:895517309 0:3045364061

Chap. 20 Numeric Linear Algebra 49

D 6 7
60 7:662236711 0:6852852366 7:
4 5
0 0:7499055119 2:095238095

Now calculate the entries of C3 from (II/1) with t3 D w32.0/=w22.0/ replaced by t3 D w32.1/=w22.1/, that is,

c3 D 1=p1 C .t3/2 D 0:9952448346;

(II/2) s3 D t3=p1 C .t3/2 D 0:09740492434:

We can now write C3, which has the same general form as in step 1, and calculate
R1 D C3W1 D C3C2B1
3
2

12:35496505 8:895517309 0:3045364061

D 6 7
60 7:698845998 0:88611310017:
4 5
0 0 2:018524735

This gives the next result

B2 D Œb.2jk/ D R1CT2CT3 D C3C2B1CT2CT3

2 3:126499072 0 3

14:90278952

D 6 7
6 3:126499074 7:088284172 0:19661424997:
4 5
0 0:1966142491 2:008926316

The approximations of the eigenvalues have improved. The off-diagonal entries are smaller than in B 1.
Nevertheless, in practice, the accuracy would still not be sufficient, so that one would do several more
steps. We do one more step, whose result is also given on p. A51 in App. 2 of the textbook.

Œb .1/ Œb .2/
Step 3. The calculations are the same as in step 2, with B 1 D jk  replaced by B2 D jk . Hence we

calculate the entries of C2 from

D q
c 21= 1 C .b21.2/=b11.2/ 2 D 0:9786942487;
/
.2/ .2/
50 Numeric Analysis Part E

(I/3) s2 D .b21 q1 C .b.2/ .2/ 2 D 0:2053230812:

=b11 /= 21 =b11 /

We can now write the matrix C2 and calculate the product

7
D6 W2 D Œwjk.2/ D C2B2
60 6:295320529 2 4:515275007 3 0:1924252356 7:
0:04036944359
4 15:22721682 5
0 0:1966142491 2:008926316
.2/ .2/
Now calculate the entries of C3 from (II/2) with t2 replaced by t3 D w22 =w32 ; that is;

p
c3 D 1= 1 C .t3/2 D 0:9995126436;
(II/3) p
s3 D t3= 1 C .t3/2 D 0:03121658809:

Write C3 and calculate

R2 D C3W2 D C3C2B2

2 3
15:22721682 4:515275007 0:04036944359
D 6 7
60 6:298390090 0:2550432812 7
4 5
0 0 2:001940393
and, finally,
B3 D R2CT2CT3 D C3C2B2CT2CT3

2 1:293204857 0 3

15:82987970

D 6 7
6 1:293204856 6:169155576 0:062493749427:
4 5
0 0:06249374864 2:000964734

This is a substantial improvement over the result of step 2.

Further steps would show convergence to 16, 6, 2, with roundoff errors in the last digits. Rounding effects
are also shown in small deviations of B2 and B3 from symmetry. Note that, for simplicity in displaying the
process, some very small numbers were set equal to zero.
 PART F
Optimization,
Graphs

The purpose of Part F is to introduce the main ideas and methods of unconstrained and constrained
optimization (Chap. 22) and graphs and combinatorial optimization (Chap. 23). These topics of discrete
mathematics are particularly well suited for modeling large-scale real-world problems and have many
applications as described on p. 949 of the textbook.

Chap. 22 Unconstrained Optimization. Linear Programming

Optimization is concerned with problems and solution techniques on how to “best” (optimally) allocate
limited resources in projects. Optimization methods can be applied to a wide variety of problems such as
efficiently running power plants, easing traffic congestions, making optimal production plans, and others.
Its methods are also applied to the latest fields of green logistics and green manufacturing.
Chapter 22 deals with two main topics: unconstrained optimization (Sec. 22.1) and a particular type of
constrained optimization, that is, linear programming (Secs. 22.2–22.4). We show how to solve linear
programming problems by the important simplex method in Secs. 22.3 (pp. 958–962) and 22.4 (pp. 962–
968).
Some prior knowledge of augmented matrix, pivoting, and row operation—concepts that occur in the
Gauss elimination method in Sec. 7.3, pp. 272–282—would be helpful, since the simplex method uses
these concepts. (However, the simplex method is different from the Gauss elimination method.)

Sec. 22.1 Basic Concepts. Unconstrained Optimization:

Method of Steepest Descent
The purpose of this section is twofold. First, we learn about what an optimization problem is (p. 951) and,
second, what unconstrained optimization is (pp. 951–952), which we illustrate by the method of steepest
descent.
In an optimization problem we want to optimize, that is, maximize or minimize some function f. This
function f is called the objective function and consists of several variables
2 Optimization, Graphs Part F

x1, x2, x3, ...; xn;

whose values we can choose, that is, control. Hence these variables are called control variables. This idea
of “control” can be immediately understood if we think of an application such as the yield of a chemical
process that depends on pressure x1 and temperature x2:
In most optimization problems, the control variables are restricted, that is, they are subject to some
constraints, as shall be illustrated in Secs. 22.2–22.4.
However, certain types of optimization problems have no restrictions and thus fall into the category of
unconstrained optimization. The theoretical details of such problems are explained on the bottom third
of p. 951 and continued on p. 952. Within unconstrained optimization the textbook selected a particular
way of solving such problems, that is, the method of steepest descent or gradient method. It is illustrated
in Example 1, pp. 952–953 and in great details in Prob. 3.

Problem Set 22.1. Page 953

3. Cauchy’s method of steepest descent. We are given the function

(A) f.x/ D 2x12 C x22 4x1 C 4x2

with the starting value (expressed as a column vector) x 0 D Œ0 0T: We proceed as in Example 1,
p. 952, beginning with the general formulas and using the starting value later. To simplify notations,
let us denote the components of the gradient of f by f 1 and f2. Then, the gradient of f is [see also (1),
p. 396]

rf.x/ D Œf1 f2T D Œ4x1 4 2x2 C 4T:

In terms of components,
f1 D 4x1 4; f2 D 2x2 C 4:
(B)

Furthermore,
trf.x/ D Œx1 tf1 x2 tf2T
z.t/ D Œz1 z2T D x
;
which, in terms of components, is
(C) z1.t/ D x1 tf1; z2.t/ D x2 tf2:

Now obtain g.t/ D f.z.t// from f.x/ in (A) by replacing x 1 with z1 and x2 with z2. This gives

g.t/ D 2z12 C z22 4z1 C 4z2:

We calculate the derivative of g.t/ with respect to t, obtaining

g0.t/ D 4z1z10 C 2z2z20 4z10 C 4z20 :

Chap. 22 Unconstrained Optimization. Linear Programming 3

From (C) we see that z10 D f1 and z20 D f2 with respect to t. We substitute this and z 1 and z2 from

(C) into g0.t/ and obtain

g0.t/ D 4.x1 tf1/. f1/ C 2.x2 tf2/. f2/ C 4f1 4f2:

Order the terms as follows: Collect the terms containing t and denote their sum by D (suggesting
“denominator” in what follows). This gives

(D) tD D t.4f12 C 2f 22/:

We denote the sum of the other terms by N (suggesting “numerator”) and get

(E) ND 4x1f1 2x2f2 C 4f1 4f2:

With these notations we have g0.t/ D tD C N. Solving g0.t/ D 0 for t gives

NtD :
D
Next we start the iteration process.

Step 1. For the given x D x0 D Œ0 0T we have x1 D 0, x2 D 0 and from (B)

f1 D 4 0 4D 4; f2 D 2 0 C 4 D 4;

tD D t .4 . 4/2 C 2 42/ D 96t

ND 4 0 . 6/ 2 0 4 C 4 . 4/ 4 4

D 16 16 D 32

so that
N 32 1 t D t0 D D
D D 0:3333333:
D 96 3

From this and (B) and (C) we obtain the next approximation x 1 of the desired solution in the form

x1 D z.t0/ D Œ0 t0. 4/ 0 t0 4T D Œ4t0 4t0T D 4 1

3 4 1 T
3 D 34 43T D Œ1:3333333 1:3333333T :

Also from (A) we find that f.x1/ is

42 42 4 4
f.x1/ D 2 C 4 C4
3 3 3 3
4 Optimization, Graphs Part F

D 32 C 16 C 16 16 DD
5:333333:
9 3
This completes the first step.

Step 2. Instead of x0 we now use x1; which is, in terms of components,

x1 D ; x2 D :
Then from (B) we get

16 12 4
f1 D 4 4 D D ;
3 3 3
4 8 12 4
f2 D 2 C4D C D ;
3 3 3 3

!
42 42 64 32
tD D t 4 C2 Dt C Dt ;
3 3 9 9

4 4 4 4ND
4 2 C44
3 3 3 3

D 64 C 32 C 0 D 32
;
9 9
so that
32
N 32 9 1 t D t1 D DD
0:3333333:
D 9 96 3

From this and (B) and (C) we obtain the next approximation x 2 of the desired solution in the form
8 16T
D
T x2 D z.t1/ D Œx1 t1f1; x2 t1f2T
Œ0:8888889 1:777778
4 1 4 1 T 12 4 12 4T
9 D 3 4 9 D
3 3 9
Also from (A) we find 3 D 9 that f.x / is 2

82 162 8 16
Chap. 22 Unconstrained Optimization. Linear Programming 5

f.x2/ D 2 C 4C 4
9 9 9 9

D 128 C 256 C 32 64 D 384 864 DDD

5:925926:
81 9 81
This completes the second step.

Step 3. Instead of x1 we now use x2, which is, in terms of components,

x1 D; x2 D:
Then from (B) we get

f1 D 4 4D D ;

16 32 36 4
f2 D 2 C4D C D ;
9 9 9 9

D 42 C 42! D 16 C 16 D 64 C 32 D
(1) tD t 4 2 t 4 2 t t; 9 9 81 81 81

8 4 16 4 4 N D 4 2 C4 4
9 9 9 9 9
D 128 C 128 144 144 D
;
81
so that
32
N 32 81 1
t D t2 D DD 0:3333333:
D 81 96 3

From this and (B) and (C) we obtain the next approximation x 3 of the desired solution in the form
x3 D z.t2/ D Œx1 t 2f 1; x2 t2f2T
8 1 4 16 1 4T 24 C 4
T

D 9
9 9 3 9
27

28 52T T
Œ1:037037 1:925926 :
6 Optimization, Graphs Part F

D 27 27 D

From (A) we find that f.x3/ is

282 522 28
f.x3/ D 2 C 4 C4
27 27 27
2 2
2 28 C 52 4 27 28 4 27 52

D 272
1568 C 2704 3024 5616

D 729

D D 5:991770:

This completes the third step.

The results for the first seven steps, with six significant digit accuracy, are as follows.

Discussion. Table I gives a more accurate answer in more steps than is required by the problem. Table II
gives the same answer—this time as fractions—thereby ensuring total accuracy. With the help of your
computer algebra system (CAS) or calculator, you can readily convert the fractions of Table II to the
desired number of decimals of your final answer and check your result. Thus any variation in your answer
from the given answer due to rounding errors or technology used can be Sec. 22.1. Prob. 3. Table I.
Method of steepest descent. Seven steps with 6S accuracy and one guarding digit
n x f
0 0:000000 0:000000 0:000000
1 1:333333 1:333333 5:333333
2 0:8888889 1:777778 5:925925
3 1:0370370 1:925926 5:991770
4 0:9876543 1:975309 5:999056
5 1:004115 1:991769 5:999894
6 0:9986283 1:997256 5:999998
7 1:000457 1:999086 5:999999
Sec. 22.1. Prob. 3. Table II. Method of steepest descent.
Seven steps expressed as fractions to ensure complete accuracy

n x f
Chap. 22 Unconstrained Optimization. Linear Programming 7

0 0 0
0

6
7

checked with Tables I and II. Furthermore, the last column in each table shows that the values of f
converge toward a minimum value of approximately minus 6. We can readily see this and other
information from the given function (A) by completing the square, as follows. Recall that, for a quadratic
equation,

ax2 C bx C c D 0
completing the square amounts to writing the equation in the form

b b2
2
C
a.x d/ C e D 0 where dD and eDc :
2a 4a

We apply to our given function f this method twice, that is, first to the x 1-terms 2x12 4x1, and then to
the x2-terms x22 C 4x2: For the x1-terms we note that a D 2, b D 4, c D 0 so that

b 4b2 16 d D D D 1 and e D c D0 D 2:
2a 22 4a 8
This gives us

(F) 2x12 4x1 D 2 .x1 1/2 2:

Using the same approach yields

(G) x22 C 4x2 D 1 .x1 C 2/2 4:

8 Optimization, Graphs Part F

Adding (F) and (G) together, we see that by completing the square, f can be written as

(H) f.x/ D 2 .x1 1/2 C 1 .x2 C 2/2 6:

Equation (H) explains the numeric results. It shows that f.x/ D 6 occurs at x1 D 1 and x2 D

2; which is in reasonably good agreement with the corresponding entries for n D 7 in the tables.

Furthermore, we see, geometrically, that the level curves f D const are ellipses with

principal axes in the directions of the coordinate axes (the function has no termp p x1x2)

and semiaxes of length

proportional to 2 and 1:
Remark. Your answer requires only three steps. We give seven steps for a better illustration of the
method. Also note that in our calculation we used fractions, thereby maintaining higher accuracy,
and converted these fractions into decimals only when needed.

Sec. 22.2 Linear Programming

The remaining sections of this chapter deal with constrained optimization which differs from
unconstrained optimization in that, in addition to the objective function, there are also some constraints.
We are only considering problems that have a linear objective function and whose constraints are linear.
Methods that solve such problems are called linear programming (or linear optimization, p. 954). A
typical example is as follows.
Consider a linear objective function, such as

z D f.x/ D 40x1 C 88x2;

subject to some constraints, consisting of linear inequalities, such as

(1) 2x1 C 8x2 60

(2) 5x1 C 2x2 60

with the usual additional constraints on the variables x 1 0; x2 0, as given in Example 1, p. 954, where the
goal is to find maximum x D .x1;x2/ to maximize revenue z in the objective function:
The inequality .1/ can be converted into an equality by introducing a variable x 3 (where x3 0), thus
obtaining

2x1 C 8x2 C x3 D 60:

Chap. 22 Unconstrained Optimization. Linear Programming 9

The variable x3 has taken up the slack or difference between the two sides of the inequality. Thus x 3 is
called a slack variable (see p. 956). We also introduce a slack variable x 4 for equation .2/ as shown in
Example 2, p. 956. This leads to the normal form of a linear optimization problem. This is an important
concept because any problem has to be first converted to a normal form before a systematic method of
solution (as shown in the next section) can be applied.
Problems 3, 21, and Fig. 474 of Example 1 on p. 955 explore the geometric aspects of linear
programming problems.

Problem Set 22.2. Page 957

3. Region, constraints. Perhaps the easiest way to do this problem is to denote x 1 by x and x2 by y. Then
our axes are labeled in a more familiar way and we can rewrite the problem as

(A0) 0:5x C y 2;
(B0) x C y 2;
(C0) x C 5y 5:
0
Consider inequality (A ). This is also equivalent to

(A00) y 0:5x C 2:

Now, if we consider the corresponding equality,

y D 0:5x C 2;

we get line 1 in Fig. A. Since (A00) is an inequality of the kind , the region determined by (A 00) and hence

(A0) must lie below 1 . We shade this in Fig. A.

The same reasoning applies to (B0).

(B0) H) (B00) y x C 2:

We consider y Dx C 2 and get line 2 in Fig. A. Since B 00 is an inequality ; we have that (A00) and (B0) lie

above line 2 as shaded.

Also (C0) H) (C00) y x C 1, which, as an equality, gives line 3 in Fig. A. Since we have , the corresponding
shaded region lies above line 3 as shaded in Fig. A.

Taking (A00), (B00), (C00) together gives the intersection of all three shaded regions. This is precisely the
region below 1 , to the right of 2 , and above 3 . It extends from .0;2/ below 1 , from .0;2/ to
10 Optimization, Graphs Part F

5 ; 7
6 6 above 2 , and from .1;1:2/ above 3 . Together we have the infinite region with boundaries as
marked in Fig. B, with the notation x1 (for x) and x2 (for y). Note that the region lies entirely in the first
quadrant of the x1x2-plane, so that the conditions x1 0; x2 0 (often imposed by the kind of application, for
instance, number of items produced, time or quantity of raw material needed, etc.) are automatically
satisfied.
y
5
1
4

3
2

–1 1 2 3 4 5 6 x

–1
2

Sec. 22.2 Prob. 3. Fig. A Graphical development of solution

x2
5
1
4

3
(2, 0) 3
2
( 2
1
(
5, 7
6 6
1 2 3 4 5 x1

Sec. 22.2 Prob. 3. Fig. B Final solution: region determined by the three inequalities given in
the problem statement

7. Location of maximum. Consider what happens as we move the straight line

z D c D const;

beginning its position when c D 0 (which is shown in Fig. 474, p. 955) and increase c continuously.

21. Maximum profit. The profit per lamp L1 is $150 and that per lamp L2 is $100. Hence the total profit
for producing x1 lamps L1 and x2 lamps L2 is

f.x1;x2/ D 150x1 C 100x2:

We want to determine x1 and x2 such that the profit f.x1;x2/ is as large as possible.
Chap. 22 Unconstrained Optimization. Linear Programming 11

Limitations arise due to the available workforce. For the sake of simplicity the problem is talking
about two workers W1 and W 2, but it is clear how the corresponding constraints could be made into
a larger problem if teams of workers were involved or if additional constraints arose from raw
material. The assumption is that, for this kind of high-quality work, W 1 is available 100 hours per
month and that he or she assembles three lamps L 1 per hour or two lamps L2 per hour. Hence W1
needs hour for assembling lamp L2 and hour for assembling lamp L2. For a production of x1
lampsL1 and x2 lamps L2, this gives the restriction (constraint)

(A) 13x1 C 12x2 100:

(As in other applications, it is essential to measure time or other physical quantities by the same
units throughout a calculation.) (A) with equality sign gives a straight line that intersects the x 1-axis
at 300 (put x2 D 0) and the x2-axis at 200 (put x1 D 0) as seen in Fig. C. If we put both x 1 D 0 and x2 D 0,
the inequality becomes 0 C 0 100; which is true. This means that the region to be determined
extends from that straight line downward.
Worker W2 paints the lamps, namely, 3 lamps L1 per hour or 6 lamps L2 per hour. Hence painting a
lamp L1 takes hour, and painting lamp L2 takes hour. W2 is available 80 hours per month. Hence if
x1 lamps L1 and x2 lamps L2 are produced per month, his or her availability gives the
constraint

(B) 13x1 C 16x2 80:

(B) with the equality sign gives a straight line that intersects the x 1-axis at 240 (put x2 D 0) and the x2-
axis at 480 (put x1 D 0); see Fig. C. If we put x1 D 0 and x2 D 0, the inequality (B) becomes 0 C 0 80,
which is true. Hence the region to be determined extends from that line downward. And the region
must lie in the first quadrant because we must have x 1 0 and x2 0.

The intersection of those two lines is at .210; 60/. This gives the maximum profit

f.210;60/ D 150 210 C 100 60 D $37;500:

Next we reason graphically that .210; 60/ does give the maximum profit. The straight line

f D 37;500

(the middle of the three lines in the figure) is given by

x2 D 375 1:5x1:

And by varying c in the line

f D const;

that is, in
12 Optimization, Graphs Part F

x2 D c 1:5x1;

which corresponds to moving the line up and down, it becomes clear that .210; 60/ does give the
maximum profit. We indicate the solution by a small circle in Fig. C.

Sec. 22.3 Simplex Method

This section forms the heart of Chap. 22 and explains the very important simplex method, which can
briefly be described as follows. The given optimization problem has to be expressed in normal form (1),
(2), p. 958, a concept explained in Sec. 22.2. Our discussion follows the example in the textbook which
first appeared as Example 1, p. 954, and continued as Example 2, p. 956, both in Sec. 22.2. Now here, in
Sec. 22.3, one constructs an augmented matrix as in (4), p. 959. Here z is the variable to be maximized, x 1,
x2 are the nonbasic variables, x3, x4 the basic variables, and b comes from the right-hand sides of the
equalities of the equations of the constraints of the normal form. Basic variables are the slack variables
and are characterized by the fact that their columns have only one nonzero entry (see p. 960).
500

400

300
x2

200

100 A

B
0
0 50 100 150 200 250 300
x1

Sec. 22.2 Prob. 21. Fig. C Constraints (A) (lower line) and (B)

From the initial simplex table, we select the column of the pivot by finding the first negative entry in
Row 1. Then we want to find the row of the pivot, which we obtain by dividing the right-hand sides by the
corresponding entries of the column just selected and take the smallest quotient. This will give us the
desired pivot row. Finally use this pivot row to eliminate entries above and below the pivot, just like in
the Gauss–Jordan method. This will lead to the second simplex table (5), p. 960. Repeat these steps until
there are no more negative entries in the nonbasic variables, that is, the nonbasic variables become basic
variables. We set the nonbasic variables to zero and read off the solution (p. 961).
Go over the details of this example with paper and pencil so that you get a firm grasp of this important
method. The advantage of this method over a geometric approach is that it allows us to solve large
problems in a systematic fashion.
Chap. 22 Unconstrained Optimization. Linear Programming 13

Further detailed illustrations of the simplex method are given in Prob. 3 (maximization) and Prob. 7
(minimization).

Problem Set 22.3. Page 961

3. Maximization by the simplex method. The objective function to be maximized is

(A) z D f.x1;x2/ D 3x1 C 2x2:

The constraints are

3x1 C 4x2 60;

(B) 4x1 C 3x2 60;

10x1 C 2x2 120:

Begin by writing this in normal form, see (1) and (2), p. 958. The inequalities are converted to
equations by introducing slack variables, one slack variable per inequality. In (A) and (B) we have
the variables x1 and x2. Hence we denote the slack variables by x3 [for the first inequality in (B)], x4
[for the second inequality in (B)], and x 5 (for the third). This gives the normal form (with the
objective function written as an equation)
z 3x1 2x2 D 0;

3x1 C 4x2 C x3 D 60;

(C) 4x1 C 3x2 C x4 D 60;

10x1 C 2x2 C x5 D 120:

This is a linear system of equations. The corresponding augmented matrix (a concept you should know!—
see Sec. 7.3, p. 273) is called the initial simplex table and is denoted by T0. It is

x x x x
x

Take a look at (3) on p. 963, which has an extra line on top showing z, the variables, and b [denoting the
terms on the right side in (C)]. We also added such a line in (D) and also drew the dashed lines, which
separate the first row of T0 from the others as well as the columns corresponding to z, to the given
variables, to the slack variables, and to the right sides.
14 Optimization, Graphs Part F

Perform Operation O1. The first column with a negative entry in Row 1 is Column 2, the entry being
3. This is the column of the first pivot. Perform Operation O 2. We divide the right sides by the
corresponding entries of the column just selected. This gives

D 20; D 15; D 12:

The smallest positive of these quotients is 12. It corresponds to Row 4. Hence select Row 4 as the row of
the pivot. Perform Operation O3, that is, create zeros in Column 2 by the row operations

Row1 C Row4;

Row2 Row4; Row3

Row4:
This gives the new simplex table (with Row 4 as before), where we mark the row operations next to the
augmented matrix with the understanding that these operations were applied to the prior augmented
matrix T0;

2 3

D Row1 C Row4

Row2 Row4
Row3 Row4:

This was the first step. (Note that the extra line on top of the augmented matrix showing z, the variables
and b as well as the dashed lines is optional but is put in for better understanding.) Now comes the
second step, which is necessary because of the negative entry in Row 1 of T1. Hence the column of
the pivot is Column 3 of T1. We compute

24 120 12 60
7:06; 5:45;60
17 D D 11 D DD
17 11
5 5

and compare. The second of these is the smallest. Hence the pivot row is Row 3. To create zeros in
Column 3 we have to do the row operations

Row1 C Row3;

Row2 Row3;

Row4 Row3;

5
Chap. 22 Unconstrained Optimization. Linear Programming 15

leaving Row 3 unchanged. This gives the simplex table

2 1 3
Row1 C Row3

Row2 Row3
D

Row4 Row3

Since no more negative entries appear in Row 1, we are finished. From Row 1 we see that

fmax D D 43:64:

In Row 4 we divide the entry in Column 7 by the entry in Column 2 and obtain the corresponding
1200
11
D 1200 1 D:
x1 value
10 11 10
Similarly, in Row 3 we divide the entry in Column 7 by the entry in Column 3 and obtain the
corresponding

12 12 5 x2 value 11 D
D:
5 1 11

You may want to convince yourself that the maximum is taken at one of the vertices of the polygon
determined by the constraints. This vertex is marked by a small circle in Fig. D.
20

15

x2 10

0
0 5 10 15 20
x1

Sec. 22.3 Prob. 3. Fig. D Region determined by the constraints

16 Optimization, Graphs Part F

7. Minimization by the simplex method. The given problem, in normal form [with z D f.x 1;x2/ written as
an equation], is

z 5x1 C 20x2 D 0;

2x1 C 10x2 C x3 D 5;

2x1 C 5x2 C x4 D 10:

From this we see that the initial simplex table is

Since we minimize (instead of maximizing), we consider the columns whose first entry is positive
(instead of negative). There is only one such column, namely, Column 3. The quotients are

D (from Row 2) and D 2 (from Row 3).

The smaller of these is . Hence we have to choose Row 2 as pivot row and 10 as the pivot. We
create zeros by the row operations Row 1 2 Row 2 (this gives the new Row 1) and Row 3
Row 2 (this gives the new Row 3), leaving Row 2 unchanged. The result is

Since there are no further positive

entries in the first row, we are
done. From Row 1 of T1 we see that

f min D 10:
From Row 2; with Columns 3 and 6; we see that

x2 D D :

Furthermore, from Row 3; with Columns 5 and 6; we obtain

15
2
D D 15:
x4
1 2

x
Now 4 appears in the second constraint, written as equation, that is,
2x1 C 5x2 C x4 D 10:

Inserting x2 D and x4 D gives

Chap. 22 Unconstrained Optimization. Linear Programming 17

2x1 C 10 D 10; hence x1 D 0:

Hence

the minimum 10 of z D f.x1;x2/ occurs at the point 0; 12:

Since this problem involves only two variables (not counting the slack variables), as a control and to
better understand the problem, you may want to graph the constraints. You will notice that they
determine a quadrangle. When you calculate the values of f at the four vertices of the quadrangle,
you should obtain

0 at .0;0/, 25 at .5;0/, 7:5 at .2:5;1/, and 10 at 0; 21:

This would confirm our result.

Sec. 22.4 Simplex Method. Difficulties

Of lesser importance are two types of difficulties that are encountered with the simplex method:
degeneracy, illustrated in Example 1 (pp. 962–965), Problem 1 and difficulties in starting, illustrated in
Example 2 (pp. 965–967).

Problem Set 22.4. Page 968

1. Degeneracy. Choice of pivot. Undefined quotient. The given problem is

z D f1.x/ D 7x1 C 14x2

subject to
0 x1 6;

0 x2 3;

7x1 C 14x2 84:

Its normal form [with z D f.x1;x2/ written as an equation] is

z 7x1 14x2 D 0; x1 C x3 D 6;

x2 C x4 D 3;

7x1 C 14x2 C x5 D 84:

From this we see that the initial simplex table is

D
18 Optimization, Graphs Part F

The first pivot must be in Column 2 because of the entry 7 in this column. We
the row of the first pivot by calculating determine
D6 .from Row 2/

ratio undefined (we cannot divide 3 by 0/ .from Row 3/

D7 .from Row 4/:

Since 6 is smallest, Row 2 is the pivot row. With this the next simplex table becomes

z x1 x2 x3 x4 x5 b
23
1 0 14 7 0 0 42 Row1 C 7Row2
67
D 60 1 0 100 67
T167
60 0 1 010 37 Row3
45
0 0 14 7 0 1 42 Row4 7Row2

We have reached a point at which z D 42. To find the point, we calculate

x1 D 6 .from Row 2 and Column 2/; x4 D 3

.from Row 3 and Column 4/:

From this and the first constraint we obtain

x2 C x4 D x2 C 3 D 3; hence x2 D 0:

(More simply: x1;x4;x5 are basic. x2;x3 are nonbasic. Equating the latter to zero gives x 2 D 0, x3 D 0.) Thus z D
42 at the point .42;0/ on the x1-axis.
Column 3 of T1 contains the negative entry 14. Hence this column is the column of the next pivot.
To obtain the row of the pivot, we calculate

ratio undefined (we cannot divide 3 by 0/ .from Row 2/

D3 .from Row 3/; D3

.from Row 4/:

Since both ratios gave 3 we have a choice of using Row 3 or using Row 4 as a pivot. We pick Row 3
as a pivot. We obtain
Chap. 22 Unconstrained Optimization. Linear Programming 19

Row1 C 7Row2

Row4 7Row2

There are no more negative entries in Row 1. Hence we have reached the maximum z max D 84. We see
that x1;x2;x5 are basic, and x3, x4 are nonbasic variables. zmax occurs at .6;3/ because x1 D 6 (from Row 2 and
Column 2) and x2 D 3 (from Row 3 and Column 3). Point .6;3/ corresponds to a degenerate solution
because x5 D 0=1 D 0 from Row 4 and Column 6, in addition to x 3 D 0 and x4 D 0.

Geometrically, this means that the straight line

7x1 C 14x2 C x5 D 84

resulting from the third constraint, also passes through .x 1;x2/ D .6;3/; with x5 D 0 because

7 6 C 14 3 C 0 D 84:

Observation. In Example 1, p. 962, we reached a degenerate solution before we reached the maximum
(the optimal solution), and, for this reason, we had to do an additional step, that is, Step 2, on p. 964. In
contrast, in the present problem we reached the maximum when we reached a degenerate solution.
Hence no additional work was necessary.
Chap. 23 Graphs. Combinatorial Optimization
The field of combinatorial optimization deals with problems that are discrete [in contrast to functions in
vector calculus (Chaps. 9 and 10) which are continuous and differentiable] and whose solutions are often
difficult to obtain due to an extremely large number of cases that underlie the solution. Indeed, the
“combinatorial nature” of the field gives us difficulties because, even for relatively small n; nŠ D 1 2 3 n
(for nŠ read “n factorial,” see p. 1025 in Sec. 24.4 of the textbook) is very large. For example,
convince yourself, that

10Š D 1 2 3 4 5 6 7 8 9 10 D 24 30 56 90 D 3;628;800:

We look for optimal or suboptimal solutions to discrete problems, with a typical example being the
traveling salesman problem on p. 976 of the textbook (turn to that page and read the description). In that
problem, even for 10 cities, there are already

10Š 3;628;800
D D 1;814;251 possible routes.
2 2
Logistics dictates that the salesman needs some software tool for identifying an optimal or suboptimal
(but acceptable) route that he or she should take!
We start gently by discussing graphs and digraphs in Sec. 23.1, p. 970, as they are useful for modeling
combinatorial problems. A chapter orientation table summarizes the content of Chap. 23.

Table of main topics for Chap. 23 on graphs and combinatorial

optimization
Section Main topic Algorithm
Section 23.1, pp. 970–975 Introduction to graphs and
digraphs
Section 23.2, pp. 975–980 Shortest path problem Moore, p. 977
Section 23.3, pp. 980–984 Shortest path problem Dijkstra, p. 982
Section 23.4, pp. 984–988 Shortest spanning trees Kruskal, p. 985
Section 23.5, pp. 988–991 Shortest spanning trees Prim, p. 989
Section 23.6, pp. 991–997 Flow problems in networks
Section 23.7, pp. 998–1001 Flow problems in networks Ford–Fulkerson, p. 998
Section 23.8, pp. 1001–1006 Assignment problems
Applications of this chapter abound in electrical engineering, civil engineering, computer science,
operations research, industrial engineering, logistics, and others. Specifics include navigation systems for
cars, computer network designs and assignment problems of jobs to machines (ships to piers, etc.),
among others.
The material is intuitively appealing but requires that you remember the terminology (e.g., a point in a
graph is called vertex, the connecting lines are called edges, etc.).
2 Optimization, Graphs Part F

Sec. 23.1 Graphs and Digraphs

This section discusses important concepts that are used in this chapter. A graph G consists of points and
the lines that connect these points, as shown in Fig. 477, p. 971. We call the points vertices and the
connecting lines edges. This allows us to define the graph G as two finite sets, that is, G D .V;E/ where V is
a set of vertices and E a set of edges. Also, we do not allow isolated vertices, loops, and multiple edges, as
shown in Fig. 478, p. 971.
If, in addition, each of the edges has a direction, then graph G is called a directed graph or digraph (p.
972 and Fig. 479).
Another concept is degree of a vertex (p. 971), which measures how many edges are incident with that
vertex. For example, in Fig. 477, vertex 1 has degree 3 because there are three edges that are “involved
with” (i.e., end or start at) that vertex. These edges are denoted by e 1 D .1;4/ (connecting vertex 1 with
vertex 4), e2 D .1;2/ (vertex 1 with 2), and e5 D .1;3/ (vertex 1 with 3). Continuing with our example, e 1 D .
1;4/ indicates that vertex 1 is adjacent to vertex 4. Also vertex 1 is adjacent to vertex 2 and vertex 3,
respectively.
Whereas in a digraph we can only traverse in the direction of each edge, in a graph (being always
undirected), we can travel each edge in both directions.
While it is visually indispensable to draw graphs when discussing specific applications (routes of airlines,
networks of computers, organizational charts of companies, and others; see p. 971), when using
computers, it is preferable to represent graphs and digraphs by adjacency matrices (Examples 1, 2, p. 973,
Prob. 11) or incidence lists of vertices and edges (Example 3). These matrices contain only zeroes and
ones. They indicate whether pairs of vertices are connected, if “yes” by a 1 and “no” by a 0. (Since loops
are not allowed in graph G, the entries in the main diagonal of these matrices are always 0.)

Problem Set 23.1. Page 974

11. Adjacency matrix. Digraph. The four vertices of the figure are denoted 1, 2, 3, 4, and its four edges
by e1; e2; e3; e4. We observe that each edge has a direction, indicated by an arrow head, which
e
means that the given figure is a digraph. Edge 1 goes from vertex 1 to vertex 2, edge e2 goes from
vertex 1 to vertex 3, and so on. There are two edges connecting vertices 1 and 3. They have opposite
e e
directions ( 2 goes from vertex 1 to vertex 3, and 3 from vertex 3 to vertex 1, respectively).
Note that, in a graph, there cannot be two edges connecting the same pair of vertices.
An adjacency matrix has entries 1 and 0 and indicates whether any two vertices in the graph are
connected by an edge. If “yes,” the two edges are connected, then the corresponding entry is a “1,”
and if no a “0.” For n vertices, such an indexing scheme requires a square, n n matrix.

Our digraph has n D 4 vertices so that A is a 4 4 matrix. Its entry a 12 D 1 because the digraph has
an edge (namely, e1) that goes from vertex 1 to vertex 2. Now comes an important point worth
taking some time to think about: Entry a 12 is the entry in Row 1 and Column 2. Since e 12 goes from 1
to 2, by definition, the row number is the number of the vertex at which an edge begins, and the
column number is the number of the vertex at which the edge ends. Think this over and look at the
matrix in Example 2 on p. 973. Since there are three edges that begin at 1 and end at 2, 3, 4, and
since there is no edge that begins at 1 and ends at 1 (no loop), the first row of A is
Chap. 23 Graphs. Combinatorial Optimization 3

0 1 1 1:

Since the digraph has four edges, the matrix A must have four 1’s, the three we have just listed and a
fourth resulting from the edge that goes from 3 to 1. Obviously, this gives the entry a 31 D 1.
Continuing in this way we obtain the matrix
20 13
60
D 077
6 1 1
7;
A6 61 0 0 07
4 0 0 5
0 0 0 0
which is the answer on p. A55 of the book. Note that the second and fourth row of A contains all
zeroes since there are no directed edges that begin at vertex 2 and 4, respectively. In other words,
there are no edges with initial points 2 and 4!
15. Deriving the graph for a given adjacency matrix. Since the given matrix, say M of the wanted graph
GM, is

2 3
0 1 0 0

61 0 0 077

D6
M6 7;
60 0 0 17
4 5
0 0 1 0

which is a 4 4 matrix, the corresponding graph G M has four vertices. Since the matrix has four 1’s and
each edge contributes two 1’s, the graph GM has two edges. Since m12 D 1, the graph has the edge .
1;2/; here we have numbered the four vertices by 1, 2, 3, 4, and 1 and 2 are the endpoints of this
edge. Similarly, m34 D 1 implies that GM has the edge .3;4/ with endpoints 3 and 4. An adjacency
matrix of a graph is always symmetric. Hence we must have m 21 D 1 because m12 D 1, and similarly,
m43 D 1 since m34 D 1. Differently formulated, the vertices 1 and 2 are adjacent, they are connected
by an edge in GM, namely, by .1;2/. This results in a12 D 1 as well as a21 D 1.
Similarly for .3;4/. Together, this gives a graph that has two disjointed segments as shown below.
1 3
1 2

3 4
4 2

Sec. 23.1. Prob. 15. Graph GM obtained from adjacency matrix

M. Note that both sketches represent the same graph.
4 Optimization, Graphs Part F

19. Incidence matrix eB of a digraph. The incidence matrix of a graph or digraph is an n m matrix, where
n is the number of vertices and m is the number of edges. Each row corresponds to one of the
vertices and each column to one of the edges. Hence, in the case of a graph, each column contains
two 1’s. In the case of a digraph each column contains a 1 and a 1.
In this problem, we looked at the graph from Prob. 11. Since, for that graph, the number of
vertices = number of edges = 4, the incidence matrix is square (which is not the most general case)
and of dimension 4 4. The first column corresponds to edge e 1, which goes from vertex 1 to vertex 2.
Hence by definition, bQ11 D 1 and bQ21 D 1: The second column corresponds to edge e2, which
goes from vertex 1 to vertex 3. Hence bQ12 D 1 and bQ32 D 1. Proceeding in this way we get

3
21 1 1 1

7
61 0 0 0
D
4 5
0
0 0 1

Sec. 23.2 Shortest Path Problems. Complexity

6 7 eB6 7:
60 1 1 07

We distinguish between walk, trail, path, and cycle as shown in Fig. 481, p. 976. A path requires that each
vertex is visited at most once. A cycle is a path that ends at the same vertex from which it started. We also
call such a path closed. Thus a cycle is a closed path.
A weighted graph G D .V;E/ is one in which each edge has a given weight or length that is positive. For
example, in a graph that shows the routes of an airline, the vertices represent the cities, an edge between
two cities shows that the airline flies directly between those two cities, and the weight of an edge
indicates the (flight) distance in miles between such two cities.
A shortest path is a path such that the sum of the length of its edges is minimum; see p. 976. A shortest
path problem means finding a shortest path in a weighted graph G. A Hamiltonian cycle (Prob. 11) is a
cycle that contains all the vertices of a graph. An example of a shortest path problem is the traveling
salesman problem; which requires the determination of a shortest Hamiltonian cycle. For more details on
this important problem in combinatorial optimization, see the last paragraph on p. 976 or our opening
discussion of this chapter.
Moore’s BFS algorithm, p. 977 (with a backtracking rule in Prob. 1), is a systematic way for determining
a shortest path in a connected graph, whose vertices all have length 1. The algorithm uses a breadth first
search (BFS), that is, at each step, the algorithm visits all neighboring (i.e., adjacent) vertices of a vertex
reached. This is in contrast to a depth first search (DFS), which makes a long trail as in a maze.
Chap. 23 Graphs. Combinatorial Optimization 5

Finally we discuss the complexity of an algorithm (see pp. 978–979) and the order O, suggesting
“order.” In this “big O” notation, an algorithm of complexity

am C b D O.m/I am2 C bm C d D O.m2/I a2m C bm2 C dm C k D O.2m/

where a;b;d; and k are constant. This means that order O denotes the fastest growing term of the given
expression. Indeed, for constant k

2m >> m2 >> m >> k for large m.

A more formal definition of O is given and used in Prob. 19. Note that, Moore’s BFS algorithm is of
complexity O.m/: (In the last equation the symbol “>>” means “much greater than.”)

Problem Set 23.2. Page 979

1. Shortest path. Moore’s algorithm. We want to find the shortest path from s to t and its length, using
Moore’s algorithm (p. 977) and Example 1, p. 978. We numbered the vertices arbitrarily. This means
we picked a vertex and numbered it 1 and then numbered the other vertices consecutively 2 , 3 , :::.
We note that s ( 9 ) is a vertex that belongs to a hexagon ( 2 , 7 , 8 , 9 , 10 , 11 ). According to step 1 in
Moore’s algorithm, s gets a label 0. s has two adjacent vertices ( 8 and 10 ), which get the label 1.
Each of the latter has one adjacent vertex ( 7 and 11 , respectively), which gets the label 2. These two
vertices now labeled 2 are adjacent to the last still unlabeled vertex of the hexagon ( 2 ), which thus
gets the label 3. This leaves five vertices still unlabeled ( 1 , 3 , 4 , 5 ,
6 ). Two ( 1 , 3 ) of these five vertices are adjacent to the vertex ( 2 ) labeled 3 and thus get the label
4. Vertex 1 , labeled 4, is adjacent to the vertex t ( 6 ), which thus gets labeled 5, provided that there
is no shorter way for reaching t.
There is no shorter way. We could reach t ( 6 ) from the right, but the other vertex adjacent to t,
i.e., ( 5 ), gets the label 4 because the vertex ( 4 ) adjacent to it is labeled 3 since it is adjacent to a
vertex of the hexagon ( 7 ) labeled 2. This gives the label 5 for t ( 6 ), as before.
Hence, by Moore’s algorithm, the length of the shortest path from s to t is 5: The shortest path
goes through nodes 0, 1, 2, 3, 4, 5, as shown in the diagram on the next page in heavier lines.

11. Hamiltonian cycle. For the definition of a Hamiltonian cycle, see our brief discussion before or turn
to p. 976 of the textbook. Sketch the following Hamiltonian cycle (of the graph of Prob. 1), which we
describe as follows. We start at s downward and take the next three vertices on the hexagon H, then
the vertex outside H labeled 4 ( 3 ), then the vertex inside H; then t; then the vertex to the right of t (
5 ), and then the vertex below it (3). Then we return to H; taking the remaining two vertices of H and
return to s.

13. Postman problem. This problem models the typical behavior of a letter carrier. Naively stated, the
postman starts at his post office, picks up his bags of mail, delivers the mail to all the houses, and
comes back to the post office from which he/she started. (We assume that every house gets mail.)
6 Optimization, Graphs Part F

t
6
5
5 4
1
8

2 4 3
4 1 7
s 9 0

3 4
10
2
1
11 3
2

Sec. 23.2. Prob. 1. Shortest path by Moore’s algorithm

Thus the postman goes through all the streets “edges,” visits each house “vertex” at least once, and
returns to the vertex, which is the post office from where he/she came. Naturally, the postman
wants to travel the shortest distance possible.
We solve the problem by inspection. In the present situation—with the post office s located at
vertex 1—the postman can travel in four different ways:

First Route: 1—2—3—4—5—6—4—3—1

Second Route: 1—2—3—4—6—5—4—3—1
Third Route: 1—3—4—5—6—4—3—2—1
Fourth Route: 1—3—4—6—5—4—3—2—1

Each route contains 3—4 and 4—3, that is, vertices 3 and 4 are each traversed twice. The length of
the first route is (with the brackets related to the different parts of the trail)

.l12 C l23/ C l34 C .l45 C l56 C l64/ C l43 C .l31/

D .2 C 1/ C 4 C .3 C 4 C 5/ C 4 C .2/

D 3 C 4 C 12 C 4 C 2 D 25;

and so is that of all other three routes. Each route is optimal and represents a walk of minimum
length 25:

19. Order. We can formalize the discussion of order O (pp. 978–979 in the textbook) as follows. We say
that a function g.m/ is of the order h.m/, that is,

g.m/ D O.h.m//

if we can find some positive constants m0 and k such that

Chap. 23 Graphs. Combinatorial Optimization 7

0 g.m/ kh.m/ for all m m0:

This means that, from a point m0 onward, the curve of kh.m/ always lies above g.m/.
(a). To show that

2
pC D O.m/
(O1) 1 m

we do the following:

0 m2 C 1 m2 C 2m C 1 for all m 1:

So here m0 D 1 throughout our derivation. Next follows

0 m2 C 1 .m C 1/2 for all m 1:

Taking square roots gives us

p
0 1 C m2 m C 1 for all m 1:

Also the right-hand side of the last inequality can be bounded by

0m1 m C m C for all m 1

„ƒ‚…
2m

so that together

p
0 1 C m2 2m for all m 1;

from which, by definition of order, equation (O1) follows directly where k D 2: Another,
more elegant, solution can be obtained by noting that

r
1
p
1 C m2 D m C 1 < 2m for all m 1:
m2

(b). To show that

(O2) 0:02em C 100m2 D O.em/

one wants to find a positive integer m0 such that

8 Optimization, Graphs Part F

100m2 < em for all m m0:

Complete the derivation.

Sec. 23.3 Bellman’s Principle. Dijkstra’s Algorithm

In this section we consider connected graphs G (p. 981) with edges of positive length. Connectivity allows
us to traverse from any edge of G to any other edge of G, as say in Figs. 487 and 488, on p. 983. (Figure
478, p. 971 is not connected.) Then, if we take a shortest path in a connected graph, that extends through
several edges, and remove the last edge, that new (shortened) path is also a shortest path (to the prior
vertex). This is the essence of Bellman’s minimality principle (Theorem 1, Fig. 486, p. 981) and leads to the
Bellman equations (1), p. 981. These equations in turn suggest a method to compute the length of
shortest paths in G and form the heart of Dijkstra’s algorithm.
Dijkstra’s algorithm, p. 982, partitions the vertices of G into two sets PL of permanent labels and TL of
temporary labels, respectively. At each iteration (Steps 2 and 3), it selects a temporarily labeled vertex k
with the minimum distance label Lek from TL, removes vertex k from TL, and places it into PL:
Furthermore Lek becomes Lk: This signifies that we have found a shortest path from vertex 1 to vertex k:
Then, using the idea of Bellman’s equations, it updates the temporary labels in Step 3. The iterations
continue until all nodes become permanently labeled, that is, until TL D ∅ and PL D the set of all edges in
G: Then the algorithm returns the lengths L k .k D 2;:::;n/ of shortest paths from the given vertex (denoted
by 1) to any other vertex in G. There is one more idea to consider: those vertices that were not adjacent
to vertex 1, got a label of 1 in Step 1 (an initialization step). This is illustrated in Prob. 5.
Note that, in Step 2, the algorithm looks for the shortest edge among all edges that originate from a
node and selects it. Furthermore, the algorithm solves a more general problem than the one in Sec. 23.3,
where the length of the edges were all equal to 1. To completely understand this algorithm requires you to
follow its steps when going through Example 1, p. 982, with a sketch of Fig. 487, p. 983, at hand.
The problem of finding the shortest (“optimal”) distance in a graph has many applications in various
networks, such as networks of roads, railroad tracks, airline routes, as well as computer networks, the
Internet, and others (see opening paragraph of Sec. 23.2, p. 975). Thus Dijkstra’s algorithm is a very
important algorithm as it forms a theoretical basis for solving problems in different network settings. In
particular, it forms a basis for GPS navigation systems in cars, where we need directions on how to travel
between two points on a map.

Problem Set 23.3. Page 983

1. Shortest path.

(a). By inspection:
We drop 40 because 12 C 28 D 40 does the same.
We drop 36 because 12 C 16 D 28 is shorter.
We drop 28 because 16 C 8 D 24 is shorter.
Chap. 23 Graphs. Combinatorial Optimization 9

(b). By Dijkstra’s algorithm.

Dijkstra’s algorithm runs as follows. (Sketch the figure yourself and keep it handy while you are
working.)

Step 1

1. L1 D 0; Le2 D 12;Le3 D 40;Le4 D 36. Hence PL D f1g, TL D f2;3;4g. No 1 appears

because each of the vertices 2, 3, 4 is adjacent to 1, that is, is connected to vertex 1 by a single
edge.

2. L2 D min.Le2;Le3;Le4/ D min.12;40;36/ D 12. Hence k D 2, PL D f1;2g, TL D f3;4g.

Thus we started from vertex 1, as always, and added to the set PL the vertex which is closest
to vertex 1, namely vertex 2. This leaves 3 and 4 with temporary labels. These must now be
updated. This is Operation 3 of the algorithm (see Table 23.2 on p. 982).

3. Update the temporary label Le3 of vertex 3,

Le3 D min.40;12 C l23/ D min.40;12 C 28/ D 40;

where 40 is the old temporary label of vertex 3, and 28 is the distance from vertex 2 to
vertex 3, to which we have to add the distance 12 from vertex 1 to vertex 2, which is the
permanent label of vertex 2.
Update the temporary label Le4 of vertex 4,

Le4 D min.36;12 C l24/ D min.36;12 C 16/ D 28;

where 36 is the old temporary label of vertex 4, and 16 is the distance from vertex 2 to vertex
4. Vertex 2 belongs to the set of permanently labeled vertices, and 28 shows that vertex 4 is
now closer to this set PL than it had been before. This is the end of Step 1.

Step 2

1. Extend the set PL by including that vertex of TL that is closest to a vertex in PL, that is, add to
PL the vertex with the smallest temporary label. Now vertex 3 has the temporary label 40,
and vertex 4 has the temporary label 28. Accordingly, include vertex 4 in PL. Its permanent
label is

L4 D min.Le3;Le4/ D min.40;28/ D 28:

Hence we now have k D 4, so that PL D f1;2;4g and TL D f3g.

2. Update the temporary label Le3 of vertex 3,

10 Optimization, Graphs Part F

Le3 D min.40;28 C l43/ D min.40;28 C 8/ D 36;

where 40 is the old temporary label of vertex 3, and 8 is the distance from vertex 4 (which
already belongs to PL) to vertex 3.

Step 3
Since only a single vertex, 3, is left in TL, we finally assign the temporary label 36 as the
permanent label to vertex 3. Hence the remaining roads are

from vertex 1 to vertex 2 Length 12,

from vertex 2 to vertex 4 Length 16,
from vertex 4 to vertex 3 Length 8.

The total length of the remaining roads is 36 and these roads satisfy the condition that they
connect all four communities.

Since Dijkstra’s algorithm gives a shortest path from vertex 1 to each other vertex, it follows that
these shortest paths also provide paths from any of these vertices to every other vertex, as required
in the present problem. The solution agrees with the above solution by inspection.

5. Dijkstra’s algorithm. Use of label Lej D lij D 1: The procedure is the same as in Example 1,
p. 982, and as in Prob. 1 just considered. You should make a sketch of the graph and use it to follow
the steps.

Step 1

1. Vertex 1 gets permanent label 0. The other vertices get the temporary labels 2 (vertex 2), 1
(vertex 3), 5 (vertex 4), and 1 (vertex 5).
The further work is an application of Operation 2 [assigning a permanent label to the (or a)
vertex closest to PL and Operation 3 (updating the temporary labels of the vertices that are still
in the set TL of the temporarily labeled vertices], in alternating order.

2. L2 D 2 (the minimum of 2, 5, and 1).

3. Le3 D min.1;2 C 3/ D 5:

Le4 D min.5;2 C 1/ D 3:

Le5 D min.1; 1/ D 1:

Step 2

1. L4 D min.5;3; 1/ D 3: Thus PL D f1;2;4g;TL D f3;5g: Two vertices are left in TLI hence we have to
make two updates.
Chap. 23 Graphs. Combinatorial Optimization 11

2. Le3 D min.5;3 C 1/ D 4.

Le5 D min.1;3 C 4/ D 7.

Step 3

1. L3 D min.4;7/ D 4:

2. Le5 D min.7;4 C 2/ D 6.

Step 4

1. L5 D Le5 D 6:

Our result is as follows:

Vertex Permanent Edge added Length
Step added to PL label to the path of
edge
1 1;2 0;2 .1;2/ 2
2 4 3 .2;4/ 1
3 3 4 .4;3/ 1
4 5 6 .3;5/ 2
The permanent label of a vertex is the length of the shortest path from vertex 1 to that vertex. Mark
the shortest path from vertex 1 to vertex 5 in your sketch and convince yourself that we have omitted
three edges of length 3, 4, and 5 and retained the edges that are shorter.

Sec. 23.4 Shortest Spanning Trees: Greedy Algorithm

A tree is a graph that is connected and has no cycles (for definition of “connected,” see p. 977; for “cycle,”
p. 976). A spanning tree [see Fig. 489(b), p. 984], in a connected graph G; is a tree that contains all the
vertices of G. A shortest spanning tree T in a connected graph G, whose vertices have positive length, is a
spanning tree whose sum of the length of all edges of T is minimum compared to the sum of the length of
all edges for any other spanning tree in G:
Sections 23.4 (p. 984) and 23.5 (p. 988) are both devoted to finding the shortest spanning tree, a
problem also know as the minimum spanning tree (MST) problem.
Kruskal’s greedy algorithm (p. 985; see also Example 1 and Prob. 5) is a systematic method for finding a
shortest spanning tree. The efficiency of the algorithm is improved by using double labeling of vertices
(look at Table 23.5 on p. 986, which is related to Example 1). Complexity considerations (p. 987) make this
algorithm attractive for sparse graphs, that is, graphs with very few edges.
A greedy algorithm makes, at any instance, a decision that is locally optimal, that is, looks optimal at the
moment, and hopes that, in the end, this strategy will lead to the desired global (or overall) optimum. Do
you see that Kruskal uses such a strategy? Is Dijkstra’s algorithm a greedy algorithm? (For answer see p.
20).
12 Optimization, Graphs Part F

More details on Example 1, p. 985. Application of Kruskal’s algorithm with double labeling of vertices
(Table 23.3, p. 985). We reproduce the list of double labels, that is, Table 23.5, p. 986, and give some
further explanations to it. Note that this table was obtained from the rather simple Table 23.4, p. 985.
Choice 1 Choice 2 Choice 3 Choice 4 Choice 5
Vertex (3, 6) (1, 2) (1, 3) (4, 5) (3, 4)
1 (1,0)
2 (1,1)
3 (3,0) (1,1)
4 (4,0) (1,3)
5 (4,4) (1,4)
6 (3,3) (1,3)
By going line by line through our table, we can see what the shortest spanning tree looks like. Follow
our discussion and sketch our findings, obtaining a shortest spanning tree.

Line 1. (1, 0) shows that 1 is a root.

Line 2. (1, 1) shows that 2 is in a subtree with root 1 and is preceded by 1. [This tree consists of the
single edge (1, 2).]

Line 3. (3, 0) means that 3 first is a root, and (1, 1) shows that later it is in a subtree with root 1, and
then is preceded by 1, that is, joined to the root by a single edge (1, 3).

Line 4. (4, 0) shows that 4 first is a root, and (1, 3) shows that later it is in a subtree with root 1 and is
preceded by 3.

Line 5. (4, 4) shows that 5 first belongs to a subtree with root 4 and is preceded by 4, and (1, 4) shows
that later 5 is in a (larger) subtree with root 1 and is still preceded by 4. This subtree actually is the
whole tree to be found because we are now dealing with Choice 5.

Line 6. (3, 3) shows that 6 is first in a subtree with root 3 and is preceded by 3, and then later is in a
subtree with root 1 and is still preceded by 3.

Problem Set 23.4. Page 987

5. Kruskal’s algorithm. Trees constitute a very important type of graph. Kruskal’s algorithm is
straightforward. It begins by ordering the edges of a given graph G in ascending order of length. The
length of an edge .i; j/ is denoted by lij. Arrange the result in a table similar to Table 23.4 on p. 985.
The given graph G has n D 5 vertices. Hence a spanning tree in G has n 1 D 4 edges, so that
you can terminate your table when four edges have been chosen. Pick edges of the spanning tree to
be obtained in order of length, rejecting when a cycle would be created. This gives the following
table. (Look at the given graph!)
Edge Length Choice
(1, 4) 2 1st
(3, 4) 2 2nd
(4, 5) 3 3rd
Chap. 23 Graphs. Combinatorial Optimization 13

(3, 5) 4 (Reject)
(1, 2) 5 4th
We see that the spanning tree is the one in the answer on p. A56 and has the length L D 12.
In the case of the present small graph we would not gain much by double labeling. Nevertheless,
to understand the process as such (and also for a better understanding of the table on p. 986) do the
following for the present graph and tree. Graph the growing tree as on p. 986. Double label the
vertices, but attach a label only if it is new or if it changes in a step.
(1, 1)
2
4
1
4

4 1 3 5
(1, 1) (1, 0) (1, 4) (1, 4)

First Second Third Fourth

From these graphs we can now see what a corresponding table looks like. This table is simpler
than that in the book because the root of the growing tree (subtree of the answer) does not change;
it remains vertex 1.
Choice 1 Choice 2 Choice 3 Choice 4
Vertex (1, 4) (3, 4) (4, 5) (1, 2)
1 (1, 0)
2 (1, 1)
3 (1, 4)
4 (1, 1)
5 (1, 4)
We see that vertex 1 is the root of every tree in the graph. Vertex 2 gets the label (1, 1) because
vertex 1 is its root as well as its predecessor. In the label (1, 4) of vertex 3 the 1 is the root and 4 the
predecessor. Label (1, 1) of vertex 4 shows that the root as well as the predecessor is 1. Finally,
vertex 5 has the root 1 and the predecessor 4.

17. Trees that are paths. Let T be a tree with exactly two vertices of degree 1. Suppose that T is not a
path. Then it must have at least one vertex v of degree d 3. Each of the d edges, incident with v, will
eventually lead to a vertex of degree 1 (at least one such vertex) because T is a tree, so it cannot
have cycles (definition on p. 976 in Sec. 23.2). This contradicts the assumption that T has but two
vertices of degree 1.
Sec. 23.5 Shortest Spanning Trees: Prim’s Algorithm
From the previous section, recall that a spanning tree is a tree in a connected graph that contains all
vertices of the graph. Comparison over all such trees may give a shortest one, that is, one whose sum of
the length of edges is the shortest. We assume that all the lengths are positive (p. 984 of the textbook).
Another popular method to find a shortest spanning tree is by Prim’s algorithm. This algorithm is more
involved than Kruskal’s algorithm and should be used when the graph has more edges and branches.
14 Optimization, Graphs Part F

Prim’s algorithm shares similarities with Dijkstra’s algorithm. Both share a similar structure of three
steps. They are an initialization step, a middle step where most of the action takes place, and an updating
(final) step. Thus, if you studied and understood Dijkstra’s algorithm, you will readily appreciate Prim’s
algorithm. Instead of fixing a permanent label in Dijkstra, Prim’s adds an edge to a tree T in the second
step. Prim’s algorithm is illustrated in Example 1, p. 990. (For comparison, Dijkstra’s algorithm was
illustrated in Example 1, p. 982).
Here are two simple questions (open book) to test your understanding of the material. Can Prim’s
algorithm be applied to the graph of Example 1, p. 983? Can Dijkstra’s algorithm be applied to the graph
of Example 1, p. 990? Give an answer (Yes or No) and give a reason. Then turn to p. 20 to check your
answer.

Problem Set 23.5. Page 990

9. Shortest spanning tree obtained by Prim’s algorithm. In each step, U is the set of vertices of the tree
T to be grown, and S is the set of edges of T . The beginning is at vertex 1, as always. The table is
similar to that in Example 1 on p. 990. It contains the initial labels and then, in each column, the
effect of relabeling. Explanations follow after the table.
Relabeling
Vertex Initial (I) (II) (III)
l24 D 4 l34 l24 D 4 –
2
– –
3 D2
l12 D 16 l13 – l – l35
4 – l45
35
D 8 l14 D 4 D 10 D 10
5
l15 D 1 D 14
1. i.k/ D 1, U D f1g, S D ;. Vertices 2, 3, 4 are adjacent to vertex 1. This gives their initial labels
equal to the length of the edges connecting them with vertex 1 (see the table). Vertex 5 gets
the initial label 1 because the graph has no edge (1,5); that is, vertex 5 is not adjacent to vertex
1.

2. 4 D l14 D 4 is the smallest of the initial labels. Hence include vertex 4 in U and edge (1, 4) as the
first edge of the growing tree T . Thus, U D f1;4g, S D f.1;4/g:

3. Each time we include a vertex in U (and the corresponding edge in S) we have to update labels.
This gives the three numbers in column (I) because vertex 2 is adjacent to vertex 4, with l 24 D 4
[the length of edge (2, 4)], and so is vertex 3, with l 34 D 2 [the length of edge (3,4)]. Vertex 5 is
also adjacent to vertex 4, so that 1 is now gone and replaced by l 45 D 14 [the length of edge (4,
5)].

2. 3 D l34 D 2 is the smallest of the labels in (I). Hence include vertex 3 in U and edge (3, 4) in S. We
now have U D f1;3;4g and S D f.1;4/;.3;4/g.
Chap. 23 Graphs. Combinatorial Optimization 15

3. Column (II) shows the next updating. l24 D 4 remains because vertex 2 is not closer to the new
vertex 3 than to vertex 4. Vertex 5 is closer to vertex 3 than to vertex 4, hence the update is l 35 D
10, replacing 14.
2. The end of the procedure is now quite simple. l 24 is smaller than l35 in column (II), so that we set
2 D l24 D 4 and include vertex 2 in U and edge (2, 4) in S. We thus have U D f1;2;3;4g and S D
f.1;4/;.3;4/;.2;4/g.

3. Updating gives no change because vertex 5 is closer to vertex 3, whereas it is not even adjacent
to vertex 2.

2. 5 D l35 D 10. U D f1;2;3;4;5g, so that our spanning tree T consists of the edges S D
f.1;4/;.3;4/;.2;4/;.3;5/g.
The length of the shortest spanning tree is

X
L.T/ D lij D l14 C l34 C l24 C l35 D 4 C 2 C 4 C 10 D 20:

Sec. 23.6 Flows in Networks

Overview of Sec. 23.6
We can conveniently divide this long section into the following subtopics:

1. Theme. Sections 23.6 and 23.7 cover the third major topic of flow problems in networks. They have
many applications in electrical networks, water pipes, communication networks, traffic flow in
highways, and others. A typical example is the trucking problem. A trucking company wants to
transport crates, by truck, from a factory (“the source”) located in one city to a warehouse (“target”)
located far away in another city over a network of roads. There are certain constraints. The roads,
due to their construction (major highway, two-lane road), have a certain capacity, that is, they allow
a certain number of trucks and cars. They are also affected by the traffic flow, that is, the number of
trucks and cars on the road at different times. The company wants to determine the maximum
number of crates they can ship under the given constraints.
Section 23.6 covers the terminology and theory needed to analyze such problems and illustrates them
by examples. Section 23.7 gives a systematic way to determine maximum flow in a network.

2. Network, pp. 991–992

We consider digraphs G D .V;E/ (definition, p. 972) in this section and define a network in which each
edge .i;j/ has assigned to it a capacity c ij > 0. The capacity measures the maximum possible flow
along .i;j/: One vertex in the network is the source s and another the target t (or sink). We denote
a flow along a directed edge .i;j/ by fij: The flow is produced and flows naturally from the source to
the target (sink), where it disappears. See p. 991.
The edge condition means that the flow cannot exceed the capacity, that is,

0 fij cij:
16 Optimization, Graphs Part F

The vertex condition (Kirchhoff’s law) applies to each vertex i that is not s or t. It is given by

Inflow D Outflow:

More precisely we get (2), p. 992.

3. Paths, p. 992
Definition of path P W v1 ! vk in a digraph G as a sequence of edges

.v1;v2/;.v2;v3/;:::;.vk 1;vk/;

regardless of their directions in G, that forms a path as a graph.

Related concepts of forward edge and backward edge of a path, p. 992 and Figs. 494 and 495.
4. Flow Augmenting Paths, pp. 992–993

Our goal is to maximize the flow and thus we look for a path P W s ! t from the source to the sink,

whose edges are not fully used so that we can push additional flow through P: This leads to flow

augmenting path (in a network) in which

(i) no forward edge is used to capacity (ii)

no backward edge has flow 0;

see definition on top of p. 993. Do you see that Conditions (i) and (ii) mean f ij < cij and fij > 0 for
related edges, respectively?

5. Cut Sets, pp. 994–996

We introduce the concept of cut set .S;T/ because we want to know what is flowing from s to t. So
we cut the network somewhere between s and t and see what is flowing through the edges hit by
the cut. The cut set is precisely that set of edges that were hit by the cut; see upper half of p. 994.
On the cut set we define capacity cap(S, T ) to be the sum of all forward edges from source S to
target T: Write it out in a formula and compare your answer with (3), p. 994.

6. Four Theorems, pp. 995–996 The section discusses the following theorems about cut sets and flows.
They are:

Theorem 1. Net flow in cut sets. It states that any given flow in a network G is the net flow through
any cut set .S;T/ of G:
Theorem 2. Upper bound for flows. A flow f in a network G cannot exceed the capacity of any cut
set .S;T/ in G:
Chap. 23 Graphs. Combinatorial Optimization 17

Theorem 3. Main Theorem. Augmenting path theorem for flows. It states that a flow from s to t in a
network G is maximum if and only if there does not exist a flow augmenting path s ! t in G: The
last theorem is by Ford and Fulkerson. It is
Theorem 4. Max-Flow Min-Cut Theorem. It states that the maximum flow in any network G is equal
to the capacity of a cut set of minimum capacity (“minimum cut set”) in G.

7. Illustrations of Concepts.
An example of a network is given in Fig. 493, p. 992. Forward edge and backward edge are
illustrated in Figs. 494 and 495 on the same page. Example 1, p. 993, and Prob. 15 determine flow
augmenting paths. Figure 498 and explanation, p. 994, as well as Probs. 3 and 5 illustrate cut sets
and capacity. Note that, in the network in Fig. 498, the first number on each edge denotes capacity
and the second number flow. Intuitively, if you think of edges as roads, then capacity of the road
means how many cars can actually be on the road and flow denotes how many cars actually are on
the road. Finally, Prob. 17 finds maximum flow.

Problem Set 23.6. Page 997

3. Cut sets, capacity. We are given that S D f1;2;3g. T consists of the other vertices that are not in S.
Looking at Fig. 498, p. 994, we see that T D f4;5;6g: First draw Fig. 498 (without any cut) and then
draw a line that separates S from T . This is the cut. Then we see that the curve cuts the edge .1;4/
whose capacity is 10, the edge .5;2/, which is a backward edge, the edge .3;5/, whose capacity is 5,
and the edge .3:6/, whose capacity is 13. By definition (3), p. 994, the capacity cap .S;T/ is the sum
of the capacities of the forward edges from S to T . Here we have three forward edges and hence

cap .S;T/ D 10 C 5 C 13 D 28:

The edge .5;2/ goes from vertex 5, which belongs to T , to vertex 3, which belongs to S: This shows
that edge .5;2/ is indeed a backward edge, as noted above. And backward edges are not included in
the capacity of a cut set, by definition.

5. Cut sets, capacity. Here S D f1;2;4;5g. Looking at the graph in Fig. 499, p. 997, we see that T D
f3;6;7g: We draw Fig. 499 and insert the cut, that is a curve that separates S from T . We see that the
curve cuts edges .2;3/, .5;3/, and .5;6/: These edges are all forward edges and thus contribute to cap
.S;T/. The capacities of these edges are 8, 4, and 4, respectively. Using (3), p. 994, we have

cap .S;T/ D 8 C 4 C 4 D 16:

15. Flow augmenting paths. The given answer is

1 2 5; f D 2
18 Optimization, Graphs Part F

1 4 2 5; f D 2; etc.

From this, we see that the path 1 2 5 is flow augmenting and admits an additional flow:

 D min.4 2; 8 5/ D min.2; 3/ D 2:

Here 2 D 4 2 comes from edge .1; 2/ and 3 D 8 5 from edge .2; 5/.
Furthermore, we see that another flow augmenting path is 1 4 2 5 and admits an
increase of the given flow:

 D min.10 3; 5 3; 8 5/ D min.7; 2; 3/ D 2:

And so on. Of course, if we increased the flow on 12 5 by 2, then we have on edge .2; 5/
instead of .8; 5/ the new values .8; 7/ and can now increase the flow on 1 4 2 5
only by 8 7 D 1, the edge .2; 5/ now being the bottleneck edge.
For such a small network we can find flow augmenting paths (if they exist) by trial and error. For
large networks we need an algorithm, such as that of Ford and Fulkerson in Sec. 23.7, pp. 998–1000.

17. Maximum flow. The given flow in the network depicted in this problem on p. 997 is 10. We can see
this by looking at the two edges .4;6/ and .5;6/ that go into target t (the sink 6) and get the flow 1 C
9 D 10. Another way is to look at the three edges .1;3/, .1;4/, and .1;2/ that are leaving vertex 1 (the
source s) and get the flow 5 C 3 C 2 D 10.
To find the maximum flow by inspection we note the following. Each of the three edges going out
from vertex 1 could carry additional flow of 3. This is computed by the difference of capacity (the
first number) and flow (the second number on the edge), which, for the three edges, are

13 D 8 5 D 3, 14 D 6 3 D 3; 12 D 5 2 D 3:

Since the additional flow is 3, we may augment the given flow by 3 by using path 1 4 5 6. Then the
edges .1;4/ and .5;6/ are used to capacity. This increases the given flow from 10 to 10 C 3 D 13. Next
we can use the path 1 2 4 6: Its capacity is

 D min .5 2;4 2;4 1/ D 2:

This increases the flow from 13 to 13 C 2 D 15: For this new increased flow the capacity of the path
1 3 5 6 is

 D min .3;4;13 12/ D 1

because the first increase of 3 increased the flow in edge .5;6/ from 9 to 12. Hence we can increase
our flow from 15 to 15 C 1 D 16:
Finally, consider the path 1 3 4 6. The edge .4;3/ is a backward edge in this path.
By decreasing the existing flow in edge .4;3/ from 2 to 1, we can push a flow 1 through this path.
Chap. 23 Graphs. Combinatorial Optimization 19

Then edge .4;6/ is used to capacity, whereas edge .1;3/ is still not fully used. But since both edges are
going to vertex 6, that is, edges .4;6/ and .5;6/ are now used to capacity, we cannot augment the
flow further, so that we have reached the maximum flow

f D 16 C 1 D 17:

For our solution of maximum flow f D 17, the flows in the edges are
f12 D 4 f13 D (instead of
7 f14 D 6 f24 2)
D 4 f35 D 8 (instead of
f43 D 1 f45 D 5)
5 f46 D 4 f56 (instead of
D 13 3)
(instead of
2)
(instead of
7)
(instead of
2)
(instead of
2)
(instead of
1)
(instead of
9)
You should sketch the network with the new flow and check that Kirchhoff’s law

Inflow = Outflow for each vertex i that is not a source s or sink t

is satisfied at every vertex.

The answer on p. A57 presents a slightly different solution with the same final result of maximum flow f
D 17. In that solution (although not stated) f 43 D 0: For practice you may want to quickly go through that
solution and show that it satisfies Kirchhoff’s law at every vertex.

Sec. 23.7 Maximum Flow: Ford–Fulkerson Algorithm

We continue our discussion of flow problems in networks. Important is the Ford–Fulkerson algorithm for
maximum flow given in Table 23.8, pp. 998–999 and illustrated in detail in Example 1, pp. 999–1000 and
Prob. 7. For optimal learning, go through Example 1 line by line and see how the algorithm applies.
Ford–Fulkerson uses augmented paths to increase a given flow in a given network until the flow is
maximum. It accomplishes this goal by constructing stepwise flow augmenting paths, one at a time, until
no further paths can be constructed. This happens exactly when the flow is maximum.
20 Optimization, Graphs Part F

Problem Set 23.7. Page 1000

7. Maximum flow. Example 1 in the text on pp. 999–1000 shows how we can proceed in applying the
Ford–Fulkerson algorithm for obtaining flow augmenting paths until the maximum flow is reached.
No algorithms would be needed for the modest problems in our problem sets. Hence the point of
this, and similar problems, is to obtain familiarity with the most important algorithms for basic tasks
in this chapter, as they will be needed for solving large-scale real-life problems. Keep this in mind to
avoid misunderstandings. From time to time look at Example 1 in the text, which is similar and may
help you to see what to do next.

1. The given initial flow is f D 6. This can be seen by looking at flows 2 in edge .1; 2/, 1 in edge .1; 3/,
and 3 in edge .1; 4/, that begin at s and whose sum is 6, or, more simply, by looking at flows 5 and 1
in the two edges .2; 5/ and .3; 5/, respectively, that end at vertex 5 (the target t).

2. Label s (D 1) by ;. Mark the other edges 2, 3, 4, 5 as “unlabeled.”

3. Scan 1. This means labeling vertices 2, 3, and 4 adjacent to vertex 1 as explained in Step 3 of Table
23.8 (the table of the Ford–Fulkerson algorithm), which, in the present case, amounts to the
following. j D 2 is the first unlabeled vertex in this process, which corresponds to the first part of Step
3 in Table 23.8. We have c12 > f12 and compute

12 D c12 f12 D 4 2D2 and 2 D 12 D 2:

We label 2 with the forward label .1 C;2/ D .1C;2/.

j D 3 is the second unlabeled vertex adjacent to 1, and we compute

13 D c13 f13 D 3 1D2 and 3 D 13 D 2:

We label 3 with the forward label .1 C;3/ D .1C;2/.

j D 4 is the third unlabeled vertex adjacent to 1, and we compute

14 D c14 f14 D 10 3D7 and 4 D 14 D 7:

We label 4 with the forward label .1 C;4/ D .1C;7/.

4. Scan 2. This is necessary since we have not yet reached t (vertex 5), that is, we have not yet obtained
a flow augmenting path. Adjacent to vertex 2 are the vertices 1, 4, and 5. Vertices 1 and 4 are
labeled. Hence the only vertex to be considered is vertex 5. We compute

25 D c25 f25 D 8 5 D 3:

The calculation of 5 differs from the corresponding previous ones. From the table we see that
Chap. 23 Graphs. Combinatorial Optimization 21

5 D min.2; 25/ D min.2;3/ D 2:

The idea here is that 25 D 3 is of no help because in the previous edge (1, 2) you can increase the

flow only by 2. Label 5 with the forward label .2 C;5/ D .2C;2/.

5. We have obtained a first flow augmenting path P: 1 – 2 – 5.

6. We augment the flow by 5 = 2 and set f D 6 C 2 D 8:

7. Remove the labels from 2, 3, 4, 5, and go to Step 3. Sketch the given network, with the new flows f 12
D 4 and f25 D 7. The other flows remain the same as before. We will now obtain a second flow
augmenting path.

3. We scan 1. Adjacent are 2, 3, 4. We have c12 D f12; edge (1, 2) is used to capacity and is no longer to
be considered. For vertex 3 we compute

13 D c13 f13 D 3 1D2 and 3 D 13 D 2:

Label 3 with the forward label .1C;2/. For vertex 4 we compute

14 D c14 f14 D 10 3 D 7 and 4 D 14 D 7:
Label 4 with the forward label .1C;7/.

3. We need not scan 2 because we now have f 12 D 4 so that c12 f12 D 0; .1;2/ is used to
capacity; the condition c12 > f12 in the algorithm is not satisfied. Scan 3. Adjacent to 3 are the
vertices 4 and 5. For vertex 4 we have c43 D 6 but f43 D 0, so that the condition f43 > 0 is violated.
Similarly, for vertex 5 we have c35 D f35 D 1, so that the condition c35 > f35 is violated and we must
go on to vertex 4.

3. Scan 4. The only unlabeled vertex adjacent to 4 is 2, for which we compute

42 D c42 f42 D 5 3D2

and

2 D min.4; 42/ D min.7;2/ D 2:

Label 2 with the forward label .4C;2/.

4. Scan 2. Unlabeled adjacent to 2 is vertex 5. Compute

25 D c25 f25 D 8 7D1

22 Optimization, Graphs Part F

and

5 D min.2; 25/ D min.2;1/ D 1:

Label 5 with the forward label .2C;1/.

5. We have obtained a second flow augmenting path P: 1 – 4 – 2 – 5.

6. Augment the existing flow 8 by 5 = 1 and set f D 8 C 1 D 9:

7. Remove the labels from 2, 3, 4, 5 and go to Step 3. Sketch the given network with the new
flows, write the capacities and flows in each edge, obtaining edge .1; 2/: .4; 4/, edge .1; 3/: .3;
1/, edge .1; 4/: .10; 4/, edge .2; 5/: .8; 8/, edge .3; 5/: .1; 1/, edge .4; 2/: .5; 4/, and edge .4; 3/: .
6; 0/. We see that the two edges going into vertex 5 are used to capacity; hence the flow f D 9
is maximum. Indeed, the algorithm shows that vertex 5 can no longer be reached.

Sec. 23.8 Bipartite Graphs. Assignment Problems

We consider graphs. A bipartite graph G D .V;E/ allows us to partition (“partite”) a vertex set V into two
(“bi”) sets S and T , where S and T share no elements in common. This requirement of S \ T D ∅ by the
nature of a partition.
Other concepts that follow are matching and maximum cardinality matching (p. 1001 of the textbook),
exposed vertex, complete matching, alternating path, and augmenting path (p. 1002).
A matching M in G D .S;T I E/ is a set M of edges of graph G such that no two of those edges have a
vertex in common. In the special case, where the set M consists of the greatest possible number of edges,
M is called a maximum cardinality matching in G: Matchings are shown in Fig. 503 at the bottom of p.
1001.
A vertex is exposed or not covered by M if the vertex is not an endpoint of an edge in M. If, in addition,
the matching leaves no vertices exposed, then M is known as a complete matching. Can you see that this
exists only if S and T have the same number of vertices?
An alternating path consists alternately of edges that are in M and not in M, as shown below. Closely
related is an augmenting path, whereby, in the alternating path, both endpoints a and b are exposed. This
leads to Theorem 1, the augmenting path theorem for bipartite matching. It states that the matching in a
bipartite graph is of maximum cardinality , there does not exist an augmenting path with respect to the
matching.
The theorem forms the basis for algorithm matching, pp. 1003–1004, and is illustrated in Example 1. Go
through the algorithm and example to convince yourself how the algorithm works. In addition to the label
of the vertex, the method also requires a label that keeps track of backtracking paths.

(A) Alternating path

b
Chap. 23 Graphs. Combinatorial Optimization 23

a
(B) Augmenting path P

Sec. 23.8. Alternating path and augmenting path P. Heavy edges are
those belonging to a matching M

We augment a given matching by an edge by dropping from matching M the edges that are not an
augmenting path P (two edges in the figure above) and adding to M the other edges of P (three in the
figure, do you see it?).

Problem Set 23.8. Page 1005

1. A graph that is not bipartite. We proceed in the order of the numbers of the vertices. We put vertex
1 into S and its adjacent vertices 2, 3 into T . Then we consider 2, which is now in T . Hence, for the
graph to be bipartite, its adjacent vertices 1 and 3 should be in S. But vertex 3 has just been put into
T . This contradicts the definition of a bipartite graph on p. 1001 and shows that the graph is not
bipartite.

7. Bipartite graph. Since graphs can be graphed in different ways, one cannot see immediately whether
a graph is bipartite. Hence in the present problem we have to proceed systematically.

1. We put vertex 1 into S and all its adjacent vertices 2, 4, 6 into T . Thus

S D f1g; T D f2;4;6g:

2. Since vertex 2 is now in T , we put its adjacent vertices 1, 3, 5 into S: Thus

(P) S D f1;3;5g; T D f2;4;6g:

3. Next consider vertex 3, which is in S. For the graph to be bipartite, its adjacent vertices 2, 4, 6
should be in T , as is the case by (P).

4. Vertex 4 is in T: Its adjacent vertices 1, 3, 5 are in S which is true by (P).

5. Vertex 5 is in S. Hence for the graph to be bipartite, its adjacent vertices 2, 4, 6 should be in T .
This is indeed true by (P).
6. Vertex 6 is in T and its adjacent vertices 1, 3, 5 are in S.

Since none of the six steps gave us any contradiction, we conclude that the given graph in this problem
is bipartite. Take another look at the figure of our graph on p. 1005 to realize that, although the number of
vertices and edges is small, the present problem is not completely trivial. We can sketch the graph in such
a way that we can immediately see that it is bipartite.

17. K4 is planar because we can graph it as a square A, B, C, D, then add one diagonal, say, A, C, inside,
and then join B, D not by a diagonal inside (which would cross) but by a curve outside the square.
24 Optimization, Graphs Part F

Answer to question on greedy algorithm (see p. 10 in Sec. 23.4 of this Student Solutions Manual
and Study Guide). Yes, definitely, Dijkstra’s algorithm is an example of a greedy algorithm, as in Steps
2 and 3 it looks for the shortest path between the current vertex and the next vertex.
Answer to self-test on Prim’s and Dijkstra’s algorithms (see p. 12 of Sec. 23.5). Yes, since both trees
are spanning trees.