(International Series of Numerical Mathematics _ Internationale Schriftenreihe zur Numerischen Mathematik _ Série Internationale D’Analyse Numérique 14) John Todd (auth.)-Basic Numerical Mathematics_ .pdf

ISNM
INTERNATIONAL SERIES OF NUMERICAL MATHEMATICS

INTERNATIONALE SCHRIFfENREIHE ZUR NUMERISCHEN MATHEMATIK
SERlE INTERNATIONALE D'ANALYSE NUMERIQUE
Editors:
Ch. Blanc, Lausanne; A. Ghizzetti, Roma; P. Henrici, Zurich;
A. Ostrowski, Montagnola; J. Todd, Pasadena
VOL. 14
BasicNumerical
Mathematics
Vol. 1:
Numerical Analysis
by
John Todd
Professor of Mathematics
California Institute of Technology
1979
BIRKHAuSER VERLAG
BASEL AND STUTTGART
CIP-Kurztitelaufnahme der Deutschen Bibliothek

Todd, John:
Basic numerical mathematics/by John Todd.
- Basel, Stuttgart: Birkhiluser.
Vol. 1. Numerical analysis. - 1979.
(International series of numerical mathematics;
Vol. 14)
ISBN-13: 978-3-0348-7231-7
e-ISBN-13: 978-3-0348-7229-4
001: 10.1007/978-3-0348-7229-4
Licensed edition for North- and Southamerica, Academic
Press, Inc., New York, London/San Francisco.
A Subsidiary of Harcourt Brace Jovanovich, Publishers.
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any
form or by any means, electronic, mechanical, photocopying, recording or otherwise, without the prior permission
of the copyright owner.
Birkhiluser Verlag, Basel, 1979
Softcover reprint of the hardcover 1st edition 1979
"Numerical Analysis" is for DONALD H. SADLER

"Numerical Algebra" is for OLGA TAUSSKY TODD
Contents
Notations and Abbreviations

Preface.
Chapter
Chpater
Chapter
Chapter
Chapter
Chapter
Chapter
Chapter
Chapter
Chapter
8
10
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
The algorithms of Gauss, Borchardt and Carlson.

Orders of magnitude and rates of convergence
Recurrence relations for powers . . .
The solution of equations . . . . .
Uniform convergence and approximation
The acceleration processes of Aitken and Euler .
Asymptotic series
Interpolation. . . . . . . . . . . . .
Quadrature . . . . . . . . . . . . .
Difference equations, differentiation and differential
equations . .
13
24
33
46
55
66
73
84
98
. 118
Appendix: Bessel functions
. 141
Solutions to Selected Problems
. 160
Bibliographical Remarks
. 247
. .
Table of Contents, Vol. 2 Numerical Algebra.
. 250
Index.
.251
. . . . . . . . . . . . . .
We use the standard logical symbolism, e.g., E for belongs to, c for is
included in, .~. for implies.
R is the set of real numbers.
If Xc R is bounded above so that there is an MER such that x EX
.~. x ::::;; M, then there is a AE R such that x E X .~. x ::::;; A and such that if
e > 0 there is an Xo E X such that xo > A-e. We write A = lub X; it is the
least upper bound. If AE X; we write A = max X. Conventionally, if X is
not bounded above, e.g., in the case X = {1, 2, 3, ... } we write lub X = 00.
In the same way we define glb X, min X.
[a, b] denotes the set of real numbers x such that a::::;; x ::::;; b.
(a, b) denotes the set of real numbers x such that a < x < b.
These are, respectively, closed and open intervals. Half -open intervals
[a, b), (a, b], [a,oo), (-00, b] are defined in an obvious way.
The derivatives or differential coefficients of a real function of a real
variable, f, when they exist, are denoted by
f ', f", f'" , (4) , ... , f n ), ....

The set of f such that f n) is continuous in [a, b] is denoted by Cn[a, b].
We write, if f E Cn[a, b],
Mn =~[a, b]= max Itn)(x)l.
a:sxsb
The forward difference operators are defined as follows:

af(n) = fen + 1) - fen),
ar+lf(n) = a(arf(n),
r=1,2, .. ..
In the case when the interval of tabulation is h:

aJ(x)=f(x+h)-f(x)
and a,,+lf(x)=ah(ahf(x.
The promoter operator is defined by Ef(n) = fen + 1).

We use the standard order symbolism, e.g., f(x) = O'(g(x as
means that there exist xo, A such that
x~oo
x~xo.~.lf(x)I::::;;A Ig(x)1
while f(x)= O'(g(x as

that
x~oo
X
means that given any e>O there is an
~xl.~.lf(x)l<e Ig(x)l.
Xl
such
We use the standard notations, e.g., 9D, 7S to mean quantities given

correct to, nine decimal places, or to seven significant figures. Thus we write
1T=3.1416 to 4D
e =2.7183 to 5S
The symbol x = a(h)a + nh indicates that x runs through the set of

values a, a + h, a + 2h, ... , a + nh.
We use == to indicate approximate equality.
I' will be used to indicate (or emphasize) the omission of a special term
in a sum
n
L' ll;j = L l1jj

j
j=l
j+i
Factorials are denoted by n! = 1 x 2 x 3 x ... x nand semifactorials by

(2n + 1)!! = (2n + 1) x (2n -1) x ... x 3 x 1,
(2n)!!=(2n)X(2n-2)x ... x2.
Binomial coefficients are defined by
n!
=nx(n-1)x ... x(n-r+1)
( n)=
r
(n-r)!r!
1x2x ... xr
.
We use log or In to denote natural logarithms i.e., logarithms to the

base e.
Preface
There is no doubt nowadays that numerical mathematics is an essential

component of any educational program. It is probably more efficient to
present such material after a strong grasp of (at least) linear
algebra and calculus has already been attained - but at this stage those
not specializing in numerical mathematics are often interested in getting
more deeply into their chosen field than in developing skills for later use. An
alternative approach is to incorporate the numerical aspects of linear algebra
and calculus as these subjects are being developed. Long experience has
persuaded us that a third attack on this problem is the best and this is
developed in the present two volumes, which are, however, easily adaptable
to other circumstances.
The approach we prefer is to treat the numerical aspects separately, but
after some theoretical background. This is often desirable because of the
shortage of persons qualified to present the combined approach and also
because the numerical approach provides an often welcome change which,
however, in addition, can lead to better appreciation of the fundamental
concepts. For instance, in a 6-quarter course in Calculus and Linear
Algebra, the material in Volume 1 can be handled in the third quarter and
that in Volume 2 in the fifth or sixth quarter.
The two volumes are independent and can be used in either order - the
second requires a little more background in programming since the machine
problems involve the use of arrays (vectors and matrices) while the first is
mostly concerned with scalar computation.
In the first of these, subtitled "Numerical Analysis", we assume that the
fundamental ideas of calculus of one variable have been absorbed: in
particular, the ideas of convergence and continuity. We then take off with a
study of "rate of convergence" and follow this with accounts of "acceleration process" and of "asymptotic series" - these permit illumination and
consolidation of earlier concepts. After this we return to the more traditional topics of interpolation, quadrature and differential equations.
Throughout both volumes we emphasize the idea of "controlled computational experiments": we try to check our programs and get some idea of
errors by using them on problems of which we already know the solutionsuch experiments can in some way replace the error analyses which are not
appropriate in beginning courses. We also try to exhibit "bad examples"
Preface
11
which show some of the difficulties which are present in our subject and
which can curb reckless use of equipment.
In the Appendix we have included some relatively unfamiliar parts of
the theory of Bessel functions which are used in the construction of some of
our examples. This is at a markedly higher level than the rest of the volume.
In the second volume, subtitled "Numerical Algebra", we assume that
the fundamental ideas of linear algebra: vector space, basis, matrix, determinant, characteristic values and vectors, have been absorbed. We use
repeatedly the existence of an orthogonal matrix which diagonalizes a real
symmetric matrix; we make considerable use of partitioned or block matrices, but we need the Jordan normal form only incidentally. After an initial
chapter on the manipulation of vectors and matrices we study norms,
especially induced norms. Then the direct solution of the inversion problem
is taken up, first in the context of theoretical arithmetic (i.e., when round-off
is disregarded) and then in the context of practical computation. Various
methods of handling the characteristic value problems are then discussed.
Next, several iterative methods for the solution of systems of linear equations
are examined. It is then feasible to discuss two applications: the first, the
solution of a two-point boundary value problem, and the second, that of
least squares curve fitting. This volume concludes with an account of the
singular value decomposition and pseudo-inverses.
Here, as in Volume 1, the ideas of "controlled computational experiments" and "bad examples" are emphasized. There is, however, one
marked difference between the two volumes. In the first, on the whole, the
machine problems are to be done entirely by the students; in the second,
they are expected to use the subroutines provided by the computing
system - it is too much to expect a beginner to write efficient matrix
programs; instead we encourage him to compare and evaluate the various
library programs to which he has access.
The problems have been collected in connection with courses given
over a period of almost 30 years beginning at King's College, London, in
1946 when only a few desk machines were available. Since then such
machines as SEAC, various models of UNIVAC, Burroughs, and IBM
equipment and, most recently, PDP 10, have been used in conjunction with
the courses which have been given at New York University, and at the
California Institute of Technology.
We recommend the use of systems with "remote consoles" because, for
instance, on the one hand, the instantaneous detection of clerical slips and,
on the other, the sequential observation of convergents is especially valuable
to beginners. The programming language used is immaterial. However, most
12
Preface
of the problems in Volume 1 can be dealt with using simple programmable

hand calculators, but many of these in Volume 2 require the more sophisticated hand calculators (i.e., those with replaceable programs).
The machine problems have been chosen so that a beginning can be
made with very little programming knowledge, and competence in the use of
the various facilities available can be developed as the course proceeds. In
view of the variety of computing systems available, it is not possible to deal
with this aspect of the course explicitly - this has to be handled having
regard to local conditions.
We have not considered it necessary to give the machine programs
required in the solution of the problems: the programs are almost always
trivial and when they are not, the use of library subroutines is intended. A
typical problem later in Volume 2 will require, e.g., the generation of .a
special matrix, a call to the library for a subroutine to operate on the matrix
and then a program to evaluate the error in the alleged solution provided by
the machine.
Courses such as this cannot be taught properly, no matter how expert
the teaching assistants are, unless the instructor has genuine practical
experience in the use of computers and a minimum requirement for this is a
personal familiarity with a significant proportion of our problems, or of a
similar set of problems. Such an experience will be helpful in the selection of
problems to be assigned, and in the order in which they should be attacked.
Further benefit will accrue from an inquiry, explicit or implicit, as to the
objective of the problems: routine manipulation, preparation for future
work, alternative approaches to material already covered, etc.
We have occasionally left blanks for numerical values to be filled in by
the reader: for instance in cases where the result is sensitive to the machine
used, or where it is important that the reader has personally dealt with the
material.
Finally, I acknowledge gratefully, on the one hand the continuing
courtesy and assistance of the Birkhauser-Verlag, and, on the other hand,
the help of generations of devoted teaching assistants, some now distinguished mathematicians, who bore the brunt of passing on the knowledge I
acquired, and whose comments have been of great value.
CHAPTER 1
The Algorithms of Gauss, Borchardt and Carlson
We shall study several algorithms of the form
= f(Xn, Yn)
Yn+1 = g(Xn, Yn)
Xn+l
where f, g are given functions and xo, Yo are given, from the point of view of
the rates of convergence of the sequences {Xn}, {yn}. In most cases xo, Yo will
be non-negative, the sequences {Xn}, {yn} will be monotonic and bounded
and therefore convergent. The limits of {Xn}, {yn} will be the same in each
case but the rates of convergence to these limits will differ markedly from
case to case.
The generation of the sequences {Xn}, {yn} on a computer of any kind is
an exercise suited to a complete novice.
The theoretical discussions which establish the observed rates of convergence require only the basic facts of calculus.
1. THE ARITIIMETIC-GEOMETRIC MEAN OF GAUSS

Let non-negative a = ao, b = bo be given. Consider the sequences
defined for n = 0, 1, 2, ... by
~+1 =!(~ +bn ),
bn+1 = .J(~bn).
It is clear that if we interchange ao, bo we do not alter ~,bn for n ~ 1.

Consequently we may assume that 0:5; b :5; a. Note that ~+h bn +1 lie be-
tween
~, bn
From the identity

~+l-bn+l =![.Ja:_.JbJ2
it follows that ~+1 > bn + l unless ~ = bn This is the classical inequality of

the arithmetic and geometric means. If a = b = I, then ~ = bn = 1 for all n.
It follows, assuming 0 al > ... > ~ > .......... > bn > ... > bl > boo
If b = 0 then bn
= 0 for all n.
Thus the sequences {~}, {bn } are bounded monotone sequences and so
convergent with limits a, (3 such that a ~ (3. We show that a = (3.
14
Chapter 1
Clearly
r r
so that
a,.+1-bn +1= [./;;;'-,/bn < [a,. -bn

a,.+l +bn +1 .Ja,.+Jb..
a,. +bn
the last inequality resulting from the fact that if x > y
.J~-./Y x-y
-=-=<-.Jx+"/y x+y
which can be proved by dividing across by
It follows that
../X- JY and cross multiplying.
[a
a,. - bn < o - bO]2"

a,. +bn
ao+bo
which indicates that a = fJ since
a,. -bn < 2ao[(ao-bo)/(ao+bo)p" -+0.

Here we use the fact that if 101 < 1 then {on} is a null sequence. The common
limit is called the arithmetic-geometric mean of a, b.
The approach of a,., bn to their common limit is quite rapid. We
examine this in the case ao = 1, bo = 10-4. The estimate for a,. - bn is
2[(1-10-4 )1(1 + 1O-4 )P" :52[1-10-4 ]2".
Let us estimate for what value of n this quantity will be <1(10-8 ) and so
negligible to BD. We want to find n so that
[1_10-4 ]2" <1x 10-8
Now, from tables, e-20 = 2.06 ...
to choose n so that
10-9 <1 X 10-8 and so it will be sufficient
[1-10-4 ]2" :5e-20

We now use the fact that as x-+ oo, the function (1-x- 1 )X increases
steadily to e- 1 [Compare Problem 1.5]. This means that
[1_1O-4 poxlO" <e-20

and so
15
provided 2n > 20 X 104 = 200 X 103 , which is certainly true if n = 18 since

28 =256>200, 2 10 = 1024> 103
We now give the actual values of a,., bn in this case and we see that our
estimates are rather generous.
ao =l
bo =O.OOOl
bl =0.01
al =0.50005
b2 = 0.070714214
a2 = 0.25502 5
a3 =0.16286961
b3 = 0.13429033
a 4 =0.14857996
b4 =0.14789122
as =
bs =
b6 =
a6=
b7 =
a7 =
b8 =
a8 =
2.
THE ALGORITHM OF BORCHARDT
We shall now make a small modification of the Gaussian algorithm.

This makes a change in the rate of convergence and in this case the limit is
readily obtained as a function of Clo, f3o.
Let non-negative Clo, f30 be given. Define for n = 0, 1, 2, ...
Cln+l =!(a,. +f3n),

f3n+l = .J(a,.+1f3n)'
Suppose first that f3o> Clo. It follows by induction from the definitions that
f3n > f3n+1 > Cln+l > a,..

Hence the sequences {a,.}, {13n} converge. Since
f3n+1 - a,. +1 = { (.J
.J-;;::;
}
.J \ [/3n - a,.]
2 a,.+l + f3n1
and since the first factor on the right is less than ;i, the limits of the
sequences coincide. The convergence in this case is ultimately much slower
than that in the Gaussian case. The same results obtain when f30 < Clo.
We shall now determine the limit in closed form. Although the defining
relations are still homogeneous they are not symmetrical. Assume first that
f30 = 1 > ClO' Let (J be defined by cos (J = Clo.
Using the relation 1 + cos x = 2 cos2 (xI2) repeatedly we find
Cll = cos2 (J12,

f31 = cos (J12,
Cl2 = cos (J12 cos2 (J12 2 ,

f32 = cos (J12 cos (J12 2 ,
16
Chapter 1
and establish by induction that

f3n
= cos (2- 10) cos (2- 2 0) ... cos (2- nO).
If we multiply across by 2n sin (2- n O) and use repeatedly the relation

2 sin x cos x = sin 2x we find
2n sin (2- nO)f3n = 2n- 1 cos (2- 10) ... cos (2- n+10)[2 sin (T"O) cos (2- nO)]
= 2n - 2 cos (2- 1 0) ... cos (2- n +2 0)

x [2 sin (2- n + 1 0) cos (2- n + 1 0)]
=
= sin o.
Hence
f3n
= 0- 1 sin O[O.Jsin On]
where On = 2- nO. Clearly, since limx->o (x/sin x) = 1, we have

lim f3n = 0- 1 sin o.
If we have ao> f30 we may assume f30 = 1 and ao = cosh t. A similar
discussion using hyperbolic functions instead of trigonometrical ones shows
that the common limit of the sequence is
(sinh t)/t.
3.
THE CARLSON ALGORmIM
Let non-negative ao, bo be given. Define for n = 0, 1, 2, ...

~+1 =.J~(~ +bn)/2,
bn+1 =.Jb"(~ +bn)/2.
Suppoe that ao> boo Then we find from the definitions that
The sequences
{~}, {bn }
therefore converge and since
~+1-bn+1 =
+ b )/2}
{.J(~
~
h
(~-bn)
n
"~+,, bn
where the first factor on the right is always less than .Ji < 1 and approaches
it follows that the sequences have a common limit, say ,. The rate of
convergence is, however, ultimately much slower than "in the Gaussian case.
!,
17
We shall now determine the limit of the Carlson sequences in terms of

ao, boo We observe that
a~+1-b~+l
(a~-b~)/2
a~-b~
2 log JaJbn 2 log (aJbn)
2 log (a,.+1/bn+l)
Hence, for all n
a~-b~
a~-b~
2 log (aJbn) 2 log (ao/boY

By the Mean Value Theorem
where
a~
is between a,. and bn Hence, for all n

a,. + bn
a~- b~
2/a~
2 log (ao/boY
We now let n -i>CX) and then a,. ~ I, bn ~ 1,
F=
a~ ~ I.
This gives
b2
ao- 0
2 log (ao/b o)
so that
4.
HISTORICAL REMARKS
The arithmetic-geometric mean of Gauss is of the highest importance in

the history of mathematics. As a teenager in 1791 Gauss, without computers, made extensive calculations of arithmetic-geometric means. In particular he found that
M(J2, 1) = 1.19814 02347 35592 20744.
It seems clear that he was searching for a formula for M(a, b), of the kind
given in 1.2, 1.3 and some of the problems. It was not until 1799 that he
made progress. At that time he computed the definite integral
dt
= Jo J(1- t4)
He then recalled his value of M(J2, 1) given above and observed that the
product AM(J2, 1) coincided to many decimal places with !1T. In his diary,
18
Chapter 1
on 30 May 1799, Gauss wrote that if one could prove rigorously that
AM(../2, 1} =11T, then new fields of mathematics would open. In his diary, on
23 December 1799, Gauss noted that he had proved this result, and more;
in later years his prophesy was fulfilled.
Stirling had actually obtained A to 16D in 1730. In Chapter 9 below
we shall show how to evaluate A approximately, using however another
method due to Gauss, quite different from that of Stirling. It has been shown
that this integral, which can be interpreted geometrically as the quarterperimeter of the lemniscate of Bernoulli, is a transcendental number.
There is no obvious way to establish Gauss' result. All known methods
have the same character, which is similar to the developments in 1.3, in
Problem 1.8 and in Problem 1.12. The identity AM(../2, 1} = 11T is the
special case k = 1/../2 of the following theorem.
Theorem 1. If O:s bo = kao:S a o then

lim a.. = lim bn
= a o1T/[2K'(k 2 }]
where
a complete elliptic integral.

Proof. (B. C. Carlson). Write, when a f 0
Change the variables by writing
COS6=~
t+a 2
-adt
d6=----=
2(t+a 2 }.Ji
and we find
Change the variable again by writing
x(x+bD
x+ai '
t=--~
dt= (x+a1a}(x+a1b) dx
(x+aD 2
19
where
to get
1
R(a 2 , b 2 ) = -
foo
7T 0
dx
= R(ai, bi).
J[x(x+ai)(x+bi)]
Applying this result repeatedly we find (compare Problem 1.15)

R(a~, b~)
where M
= lim a,. = lim bn-
= R(ai, bi) = ... = R(~,~)

The common value is clearly
~
r'll"/2
7T.lo
dO = M- 1
Hence
M
2 2)]-1
= [R (ao, bo
7Ta o
= 2K'(k 2 )
where
where k = bol ao.

The algorithm which we have attributed to Borchardt, who discussed it
in 1880, was apparently known to Gauss who wrote about it to Pfaff in 1800.
For further historical remarks see Problem 1.20, Solution.
The Carlson algorithm was discussed in 1971.
For an alternative account of Theorem 1, see Problem 1.19. The three
algorithms of this chapter are discussed in a geometrical way by I. J.
Schoenberg (Delta, 7 (1978), 49-65).
Chapter 1, Problems
1.1. Read again a proof of the fundamental theorem on monotonic sequences: if x" :5x,,+1 and if x,,:5M for all n = 0,1,2, ... then there is an
x :5 M such that
lim x" =x.
1.2. Show that if x" ~ Yn for all n = 0, 1, 2, ... and if x" ~ x and Yn ~ Y then
x~y.
1.3. Show that if
101 < 1 then lim on = o.
20
Chapter 1
1.4. Read again the proof of the fact that

(1 +n-l)n.- e.
1.5. Show that as X'- OO the sequence (1-x- I )X increases to e- l and the
sequence (1- X-l)x-l decreases to the same limit.
1.6. Observe the behavior of the arithmetic-geometric sequences {a,,}, {bn }
e.g., when ao = 1, bo =0.2 and when ao =../2, bo = 1.
Specifically, compute enough terms of the sequences to find the
arithmetic-geometric means to 8 decimal places. At each step print out the
values of
a", bn , a" - bn , 2ao[(a o- bo)/(ao+ bo)]2".
1.7. H M(ao, bo) is the arithmetic-geometric mean of ao, bo, show that for
any t~O
M(tao, tbo) = tM(ao, bo),
i.e., M is homogeneous (of degree one).
Use the fact to determine, with the help of the result of Problem 1.6,
M(6000, 1200).
1.8. For ao, bo given (both different from zero) define
a,,+1 = ~(a" + bn ), the arithmetic mean,
bn +1 = 2/{(l/a,,) + (l/bn )}, the harmonic mean.
Show that ao~bo implies that {a,,} is monotone decreasing, that {bn } is
monotone increasing and that both sequences converge to .J aobo, the
geometric mean of ao, boo
Show that
a,,+1 - bn +1 = [a" - bn ]2/(4a,,+1)'
Observe the behavior of the sequences {a,,}, {bn } in special cases.
1.9. For ao, bo given (both different from zero) define
a,,+l = 2/{(1/a,,) + (llbn )},
bn +1 = Ja"bn
Discuss the convergence of the sequences {a,,}, {bn } either directly, or by
relating them to the Gaussian sequences for ao\ bOlo
1.10. a) Observe the behavior of the Borchardt sequences in the cases
21
when ao = 0.2, 130 = 1 and when ao = v'2, 130 = 1. Compare the rates of convergence with those in the Gaussian case. Check the limits you obtain from
tables.
b) Repeat a) for the Carlson sequences.
1.11. If A o, Bo are non-negative and if we define for n = 0, 1, 2, ...
Bn+1 =.J~Bn
An+1
= !(~ + Bn +1)
discuss the convergence of the sequences {~}, {Bn} either directly, or by

relating them to the Borchardt sequences of ao = A o, ~o = JAoBo.
1.12. Obtain the common limit of the Borchardt sequences by proving that
f(n)=2narccos(a.J~n) and g(n)=2nJ(~~-a~) do not depend on nand
then showing that [f(n)/g(n)]~ l-1 so that the limit 1 must be g(O)/f(O).
1.13. Read again proofs of the First Mean Value Theorem.
1.14. Using appropriate tables of elliptic integrals check the value obtained
for M(l, 0.2).
1.15. Justify the relation
R(a~, b~) = R(~;~)
in the proof of Theorem 1.

1.16. (B. C. Carlson)
a) Discuss the convergence of the algorithm
ao =!(1+x),
go=JX,
gn+1
= J ~+1gn>
n=O, 1, ... ,
where x>O.
Calculate 10g.Ji. Examine the convergence of the sequence {i( ~ +
2gn )}.
b) Discuss the convergence of the algorithm
ao=l,
~+1 =!(~ +g,.),
gn+l
=J~+1g,.,
where x>O.
Calculate arctan 1.
1.17. Let ao>bo>O. Define
~+1 =!(~ +J~bJ,
n=O, 1, ... ,
22
Chapter 1
Show that {a,.}, {bn } have a common limit. Determine this limit and discuss
the rate of convergence.
1.18. (B. C. Carlson)
Show that the sequences {x,,}, {Yn}, where 0:-:::; Yo < Xo and where for
n;z: 0,
Yn+l = .Jx"x,,+1,
converge to a common limit 1= l(xo, Yo). First of all print out {x,,}, {X~I/2} for
n = 0(1) 20 when Xo = 1, Yo = O. Is the convergence monotonic? Is the
convergence ultimately geometric, and if so, what is the COmmon ratio?
By means of the change of variable
t = s(s + x"Yn)/[s +~(x" + Yn?J

show that
f"
(t + x~t3/4(t + y~)-1!2 dt =
Conclude that
4[I(x o, YO)]-I/2 =
f'
f" (~+ X~+1t3/4(X+ Y~+1)-I/2
dx.
(t + X~)-3/4(t + y~)-1!2 dt.
Show that in the case Xo = 1, Yo = 0 the last integral, by an appropriate

change of variable, becomes
4
(1- T4)-1I2
dT.
Check your results with those of Gauss.

1.19. (J. Landen) If O:-:::;k:-:::;l, O:-:::;k':-:::;l, k 2+k'2=1, k 1=(1-k')/(1+k')
show, using the change of variables,
sin cp = (1 + k') sin 0 cos 0 (1- k 2 sin2 0)-1/2
that
r7T/2 ___d_O---:-_-=
Jo
(1- k 2sin2 0)112
1 + k'
1,"/2
dq,
(1- ki sin 2 cp )1/2
1.20. (J. Schwab) Find expressions for a,., the radius of the circle inscribed
in a regular n-gon with perimeter 1, and for bn> the radius of the circumscribing circle. Hence, or otherwise, show that
Show that the sequences
have a common limit. What is the limit?

Work out the case n = 6. Discuss the degenerate case n = 2.
23
CHAPTER 2
Orders of Magnitude and Rates of Convergence
1.
ORDERS OF MAGNTIUDE
In many areas of mathematics the use of the (J- and o-symbols of

Landau is a great convenience. These give some quantitative meaning to the
phrases "of the same order (of magnitude) as", "of smaller order than".
Calculations using these symbols correctly are useful preliminaries to numerical work but, we shall see, can be misleading if not interpreted properly.
Consider, for instance, the statements:
(1)
f(x) = (J(x 2) as
x~CXl.
(2)
g(x) = (J(1)
as
x~CXl.
(3)
h(x) = o(x)
as
x~O.
(4)
k(x)=o(1)
as
x~O.
The relation (1) means that for some constant A and some Xo
x ~xo~lf(x)I:5Ax2.
For instance, this is true for any Xo with A = 2 if f(x) = 2X2 and for Xo ~ 1
with A = 5 if f(x) = 2X2 + 3x. Similarly (2) means that for some constant A
and some Xo
x ~xo~lg(x)I:5A.
For instance, this is true for any Xo and A = 1 when g(x) = sin x. In other
words (2) means that g(x) is bounded as X~CXl.
The relation (3) means that given any e >0 there is 8 = 8(e) such that
h(x)1
0<lxl<8~ -x- <e
i.e., h(x)/x~O as x~O.

For instance this is true with 8=8(e)=e when h(x)=x 2. Similarly (4)
means that given e>O, there is a 8=8(e) such that
0<lxl<8~lk(x)l<e
i.e., that k(x)~O as x~O.

We now give a formal definition.
25
Definition S. The statements

(6)
f(x)
= O(g(x))
as
x~oo
(7)
h(x) = o(k(x))
as
x~oo
have the following meaning: there is an Xo and a constant A

we can clearly assume to be positive) such that
(8)
and, for any
:2:Xo~ If(x)1 :5Alg(x)1
> 0, however small, there is an
(9)
= A (xo) (which
Xo
= x o( e),
such that
:2:Xo~ Ih(x)l:5 elk(x)l.
REMARKs: (10) The definition is phrased for the case when x is a real
variable - a similar definition applies when we consider functions of a
positive integral variable n.
(11) Usually g(x) and k(x) are standard positive functions like x R or eX and
the absolute value signs are not needed on the right of (8) and (9).
The vagueness about the interdependence of A and Xo in (1) and e and
Xo in (2) can be disturbing to the numerical analyst. For instance, if he is
concerned about scaling quantities to avoid overflow, a statement like

f(x) = 0(1)
is useless for this is true for f(x) = 10100 sin x. In practice, a calculation
involving O's and o's, while useful in preliminary investigations, must be
reworked, getting explicit estimates for the constants involved, before being
employed in actual computation.
The meaning of statements like
f(x) = (J(g(x))
as
x~o
h(x)=O(k(x)) as
x~a
or
is clear. In the first case the conclusion of (8) must hold for
for some
'J}
> 0 and in the second the conclusion of (9) must hold for
O<lx-al:::;~
for some
~>O.
26
Chapter 2
It is easy to prove theorems of the following kind (all results being

understood as holding "as x~oo", for instance).
(12)
f(x) = O'(g(x
and
g(x) = O'(h(x
imply
f(x) = O'(g(x
and
hex) = O'(g(x
imply that
(13)
f(x) = O'(h(x
Ff(x) + Hh(x) = O'(g(x
for any constants F, Hand

(14)
f(x)h(x)
= O'g(x)f).
Observe that if f(x) = O'(g(x it does not follow that g(x) = O'(f(x nor
do the relations
fleX) = O'(g(x,
f2(X) = O'(g(x
imply anything about the relative orders of fl' fz, let alone their eqUality.
Before proving in detail a general result of use in Chapter 4 we discuss
two examples.
Examples
(a) To prove that g(n) = 2n 3 + 3n 2 +4n + 5 = 0'(n 3 ) as n~OO we have various possibilities which are exemplified by the following:
(15)
g(n):5(2+3+4+5)n 3 =14n 3 if n;:::1.

g(n):52n 3 +(3+4+5)n 2 =2n 3 +12n 2 if n;:::1
and hence
(16)
and
(17)
g(n):52.12n 3
n;::: 100.
if
Each of the three relations (15), (16), (17) justifies the statement g(n) =
0'(n 3 ) and the interdependence of the no and A is clear. The constant A can
be taken to be any number bigger than 2 but the no increases as we take A
closer to 2.
(b) To prove that xl! = o(e X ) as x~oo we can proceed as follows. Note that
xl!
-------:-::---<
Xl2
1 +x+ ... + 12! + ...
12!
X
x>o.
We show that given e > 0 we can choose Xo = x o( e) such that
x;:::xo
27
For instance, take

5 X 108 we shall have
xlle- x
= 10-6 Since 121 =4790 01600=4.79 ...
108 <
< 10-6 if x ~xo= 5 X 108 X 106 = 5 X 1014
This is not a very realistic value; indeed a more realistic value is xo = 60 for
60 11 =3.62 ... x 1019 < 10-6 e60 = 1.14 ... X 1020
and
(xne-
X ),
= (n -
x)e- x x n- 1 < 0 if
x> n.
We shall require the following result in our later discussion of the

Newton process:
Theorem 18. If, as
x~O,
f(x) = 1+ax+(J(x 2 ),
(19)
then, as
g(x)= 1+bx+(J(x 2 )
x~O,
(20)
Proof. This result can be obtained by combining the following results,
which follow from the same hypotheses (19), again as x~O,
(21)
(22)
f(x)g(x)
= 1 + (a + b)x + (J(x
2 ),
l/f(x) = 1- ax + (J(x 2 ).
We shall establish (21) and (22). The meaning of (19) is that there are
positive constants A, B, a, {3 such that
(23)
If(x)-1-axl:5Ax 2
if
(24)
Ig(x)-1-bxl:5Bx 2
if Ixl:5{3.
Ixl:5a,
If we write Al = max (A, B), al = min (1, a, (3) then (23), (24) both hold
with A, B replaced by Ai> and a, {3 by a1. We have to show that there are
positive constants C, 'Y such that
(25)
If(x)g(x)-[1+(a+b)x]I:5Cx 2
if Ixl:5'Y.
The following identity holds:

fg-[1 +(a + b)x]= (f-1-ax)(g-1- bx)
(26)
+ ax(g-1)+ bx(f-1)
+(f-1-ax)+(g-1- bx)
-abx 2
28
Chapter 2
Taking absolute values in
Ifg-[1 +(a +b)x]1
(26)
using
(23), (24),
::5A~x4+lallxl [lbllxl+ At
we find for
Ixl::5at:
IxI2]+lblIxl [Iallxl+ At Ix12]
+ Atlxl2+ At Ixl2
+ lallbllxl2
::5C Ixl2
if
C=Ai+2Iallbl+(lal+lbl)At +2At+lallbl
=Ai+(2+lal+lbi)At +3lallbl
since Ixl::5at::51. This gives (21) or (25) with y=at.
To establish (22) we use the identity
!-(1-ax)= 1-(f-1-ax)(1-ax)-(1 +ax)(1-ax)
(27)
a 2 x 2 -(f-1-ax)(1-ax)
=----~----~--~
Since f(x) = 1+ax +0'(x 2) as X-+O we can find an a2 such that If(x)I~! if
Ixl::5 a2. We shall find a2 explicitly. First we take at = min (1, a). Then (23)
gives
We want to have
i.e.,
which will follow, since we
lallxl + A Ixl2::5!,
are assuming Ixl::5 at ::51,
from
Ixl(lal + A) ::5!,
i.e., from
We therefore take
Observe that while a can be zero, we may always assume that A is not zero
so a2 is well defined. We now return to the identity (27) and, taking absolute
29
values, we find
17-(1- ax) 1::52[laI2IxI2+Alx1 (1 + lallxl)],

::5lxI 2[2(laI 2+ A(1 + laD]
2
Ixl ::5a2
which is (22).
2. RATES OF CONVERGENCE
In Chapter 1 we studied sequences which had differing rates of convergence. We shall now give a formal definition of order of convergence.
Definition. A sequence {x,,} with limit x is said to converge with order r

if, when en = x" - x,
(28)
where 0<111<00.
It is not difficult to construct sequences which converge to zero, e.g.,
with any prescribed order, A, say. Take any x, 0 < x < 1 and then consider
the sequence x" = X AR In this case en = X AR and en+1 = e~ so that
lim (en+1/e~ = 1, and the sequence converges to zero with order A.
Certain variants of this definition are in use. For instance, we may only
require that
0 <1 <len+1I<1 <00

1-
lenl' -
In connection with the latter we note that if en +1 = O(e~) then convergence

will be of order r at least.
Indeed, from a practical point of view, it is more convenient to consider
a process which gives a nest of intervals {[x", Ynn contracting to a limit 1 to
have, say, quadratic convergence where
Yn+1 - x,,+1 = OYn - x,,)2)

as n~oo. Clearly certain ambiguities can arise e.g., if we take x" i 0
quadratically and Yn t 0 linearly then the nest {[x", Ynn will contract linearly
to O.
We shall, on the whole, use the formal definition. We conclude with a
determination of the order of convergence of some of the sequences
discussed in Chapter 1.
30
Chapter 2
The Gauss sequence is quadratically convergent to its limit M =

M(ao, bo). Since
we have
as required.
Similarly, if B is the limit of the Borchardt sequence we have
{3n > {3n+l > an +1 = !(<Xn + (3n) > <Xn
and
B-<Xn+1{3n+l-<Xn+l)= ( , - ~\ (/3n-<Xn)
2 V<Xn+l+V{3n/
<i({3n -<Xn)
<!(B-<Xn).
Thus we have linear convergence with ratio :::1
Chapter 2, Problems
2.1. Read an account of the Integral Test for the convergence and divergence of a series with positive terms.
2.2. The following series are well-known to be divergent
L-,L-,L-,L-
In n
In n nOn n)
00
00
00
00
Estimate for each series a integer N such that the sum of the first N terms
exceeds 10. Where reasonable, obtain the least such N on your computer.
2.3. Obtain
rough
estimates
L n -2 , L n- 10, L n- 11/lO , L n-101/100.
for
the
sums
of
the
series
2.4. If 0 < a
31
< 1 show that

n 1- a
fa(n) =
(l-a)-
L r-
,=1
tends to a limit as n ~oo.

What is a corresponding result in the case a
Compute h(n) for n = 1(1) 10.
= 1?
2.5. If a>O then for all n>O,
L r- 1- a <c~n-a
00
C1n-a <
r=n
for some positive constants
Ch C2.
2.6. Use the results of Problem 2.5 to show that it is not profitable to try to
evaluate
'\'
-1
'- a" =
~ (2n-1)!!
1
t...
.-n=1 (2n)!!
4n + 1
directly on your computer.

Show that it is not profitable to evaluate
directly on your computer.

2.7. Show that the convergence of the Carlson sequence of Chapter 1 is
ultimately linear.
2.8. Show that if p,(x) is a polynomial of degree r in x and a >0 then
p,(x) = O'(exp ax) as x~oo.
2.9. Express the standard result for Taylor's Series with remainder, when
In+1)(x) is bounded, in terms of order symbols.
2.10. Obtain a precise version of the result (1+t+ ... +~)-IOgn~'Y by

following the procedure indicated:
a) If g has a continuous third derivative in [0, 1], show, by considering
G(x) = g(x)- g(O)-tx[g'(O)+ g'(x)]
- x 3 [g(l)- g(O)-t{g'(O) + g'(l)}],
32
Chapter 2
that there is a number 6, 0 < 6 < 1 such that

g(1) = g(O) +![g'(O) + g'(1)]- Ag'''(6).
b) Apply this to the case when
g(x) = log (n + x).
c) If
show that
and give a value for the constant A implicit in the (Y.

d) From c) deduce, using Problem 2.5, that for N> n
1
1 B
2n 2N n 2
1
1 B
2n 2N n 2
-----:s;Yn -YN:S;-+-+-
and hence, letting N ~OO,
2.11. Show that as
n~oo,
x being fixed,
CHAPTER 3
Recurrence Relations for Powers
In this chapter we shall study the solution of equations of the form

xP=N
where N is a given number, 0 < N < 1, and p is a real number, !,i, -1, for
instance. We shall discuss various sequences {x,,} whose limit is the solution
sought and whose terms are defined by recurrence relations, usually of the
form
x,,+1 = f(x,,)
where f is a polynomial, or the ratio of two polynomials. We shall be

interested mainly in the relative rates of convergence of the sequences. In
Chapter 4 we shall discuss the general equation
F(x)=O
and Newton's Method, discussed there in general terms, provides a single

source for the schemes studied in this chapter.
Convergence in each case is guaranteed by monotony and if x" ~ x we
shall denote the remainder by en = x" - x. We shall discuss sequences which
are linearly, quadratically or cubically convergent in the sense of Chapter 2.
We shall also discuss for what range of values of our initial "guess" Xo
convergence takes place.
1.
REcIPROCALS
Consider the sequence defined by

(1)
x,,+1 = (1- N)x" + 1.
This is illustrated graphically in Fig. 3.1. It is clear geometrically that if

xo<N- 1 then x" steadily increases to N- 1 while if xo> N- 1 , then x" steadily
decreases to N- 1 . If Xo = N- 1 then x" == N- 1 .
To establish these facts analytically we observe that the defining relation can be written in the forms:
x,,+1 - x"
= (1- Nx,,).
x,,+l - N- 1 = (1- N)(x" - N- 1 ).
34
Chapter 3
y.(1-N)x+l
Fig. 3.1. Linear convergence to 1/N.
The first shows the monotony of the sequence and the second its boundedness. Let I denote the limit. Passage to the limit in the defining relation gives
1= (1- N)l + 1, i.e., Nl = 1,
To discuss the rate of convergence we use the second variant of the

recurrence relation. It can be written as
which shows that convergence is linear, N being between 0 and 1.

Observe that in this case we get convergence to N- l no matter how Xo
is chosen.
Consider next the sequence defined by
(2)
x,.+1 = x,.(2- Nx,.).
This is illustrated graphically in Fig. 3.2. It is clear geometrically that if

Xo = 0 or Xo = 2N- l then Xl = X 2 = ... = 0 so that there is convergence to O.
Further, if O<xo<N- l then x,. increases steadily to N- l ; if xo=N- l then
x,. =N- l and if N- l <xo<2N- l then O<Xl <N- l and Xl <X2<'" <N- l .
These results can all be established analytically as in our discussion of
(1).
If x,. < 0 then x,.+l = 2x,. - Nx~ < 2x,. and so, if Xo < 0, x,. diverges to -00.
If x,.>2N- l then x,. +1 <0; hence if xo>2N- l we have Xl<O and x,.
diverges to -00.
35
Fig. 3.2. Quadratic convergence to liN.
In the case of convergence we have

Bn+l
= x,,+l - N- l = 2x" -
Nx~- N- l
= -N[x~-2N-lx" + N- 2 ]
=-NB~,
showing that convergence is quadratic.

Consider next the sequence defined by
(3)
We begin by drawing a graph of

y = x[3(1- Nx)+(Nx?]
in the case N =~. We take Xo = 1 and find Xl = i, X2 = ~~~, ...

We can write the recurrence relation in the form
(4)
If 0 < x" < N- l then x" - N- l is negative and so is its cube and so is
x,,+1- N- l ; further, since
we have
x"+1-x,, >0.
36
Chapter 3
Y-X
Y'x[3(1-Nx)+ (Nx)2]
Fig. 3.3. Cubic convergence to liN.
Thus if O<x.. <N-t, the sequence {x..} is strictly increasing. It is bounded by

N- 1 and so has a limit, " say which necessarily satisfies 0 < b; N-t. Passing
to the limit in the recurrence relation we find
i.e.,
I(Nl- 2)(Nl-l) = 0
so 1=0, 2N- 1 or N- 1 In view of the relation 0<l:5N- 1 we must have

l=N- 1
If xo=N- 1 then XO=Xl =x2 = ... =N- 1 . If N- 1 <xo<2N- 1 we can
prove in the same way that {x..} is now a strictly decreasing sequence with
limit N- 1
Convergence takes place, therefore, if 0<xo<2N- 1 and the relation (4)
shows that it is cubic. If Xo = 0 then x.. == 0 while if Xo = 2N- 1 then x.. == 2N- 1 .
If xo<O or xo>2N- 1 then we have divergence to -00 or to 00, respectively
because x.. <0 implies x..+1 <x.. and x.. >2N- 1 implies x.. + 1 > x.. and there is
no possible limit <0 or >2N- 1
The points 0, N-t, 2N- 1 are fixed points and 0 and 2N- 1 are repulsive,
while N- 1 is aUractive.
2.
37
SQUARE ROOTS
Consider the sequence defined by

x,,+1 = x" +!(N - x~.
(5)
The behavior of this sequence is illustrated graphically in Fig. 3.4. If

-IN<xo<2+JNthen x,,~JN. If xo=-JNor xo=2+JNthen Xl =X2=
... = -IN and x" ~ -IN. If Xo < -IN or xo> 2 + IN then -IN> Xl > X2 >
.... This sequence must converge to a limit I < -IN' which is impossible
since the only possible limits are IN' or diverge to -00.
Since
Bn+1 = Bn[1-!(JN+x,,)]
we have linear convergence when -IN< Xo < 2 + IN.
The following three sequences converge quadratically to
behavior is outlined in Problems 3.2, 3.3, 3.4.
(6)
IN;
their
x,,+1=!(x,,+NX;;l).
This is easily motivated: if x" ::i IN then N/ x" ~ IN and it would appear that
their average !(x" + (N/x,,)) would be a better approximation than either.
(7)
Yn+l = 2Y!I[3y~- N].
(8)
Zn+1 = zn[3N - z~/2N.

The following sequence, due to Dedekind, converges cubically to
(9)
IN:
Xn+1 = [x~ + 3Nx,,]/[3x~ + N].
See Problem 3.13.
Fig. 3.4. Linear and quadratic convergence to ../N.
Chapter 3
38
3.
ITERATION
The processes which we have been studying to solve equations f(x) = 0

e.g., X2- N = 0 have been iterative: we have replaced f(x) by x - g(x) and
have "iterated"
x,,+1 = g(x,,)
to get a fixed point of the mapping x~ g(x). In each case we have

established convergence in certain circumstances, by special methods. We
now give a general result which is often sufficient.
Theorem 10. Suppose that in an interval J: [xo - h, Xo + h] the following
conditions are satisfied: g is defined in J and
(a) Ig'(x)I~O,
0<0<1,
(b) Ig(xo)-xol~h(1-0).
Then (1) the relation x,,+l = g(x,,) makes sense for all n and all the points
(xo), Xl' x2, ... lie in J and (2) x" ~ x* E J, the only fixed point of g in J.
Proof. We use induction to prove (1). The case n = 1 is trivial. Suppose

Xo, Xl> ... x" have been defined and are in J and that Ix,. - x,.-ll ~ or-llxl - xol
for r = 1, 2, ... ,n. Then x,,+l = g(x,,) is defined and there is a Cn between
x", x,,-l such that
Ix,,+l - x" 1= Ig(x,,) - g(x,,-l)1 = Ig'('n)(x" - x,,-I)1
~
0 . on-ll xl - xol =
on IXI -
xol.
Now
and so
1x,,+1-Xol~(on +on-l+
... + 1)lxl- xol
~ 1- 0 IXI - xol
1
1-0
~-h(1-0)=h
i.e., x,,+l E J.
The inequality Ix" - x,,-ll ~ 0,.-1 IXI - xol shows that the sequence x,,+1 =
Xo + (Xl - xo) + ... + (x,,+l - x,,) consists of the partial sums of a convergent
series and is convergent to a limit, x*, necessarily in J.
39
The limit x* is the unique fixed point (in J). For if we had another fixed
point y we would have
g(y) = y,
g(x*)=x*
and so
x*-y
= g(x*)- g(y) = (x*- y)g'(z)
where Z is between x*, y but \g'(z)\:$;0<1 and we have a contradiction.

This completes the proof. We note that we can easily estimate the rate
of convergence. In fact, by the argument used above, if n > r,
Ix,. - x,. I:$; (on-l + ... + O')IXl - xol

= 0'(1 + ... + on-,-1)lx1- xol
:$; 0' . _1_ . h(1- 0) no matter what n is

1-0 .
=hO'.
Let n ~ 00 and we get
Ix,. - x*l:$; M'.
(11)
4.
ORIGIN OF TIlE RECURRENCE RELATIONS
The Newton process for solving an equation

f(x)=O
by choosing a first approximation Xo to the root sought and improving it

according to
x,.+l = x,. - [f(x,.)]Jf'(x,.)
(12)
is one of the fundamental algorithms in numerical mathematics and will be

discussed in Chapter 4. Here we indicate how some of the relations
discussed above appear as special cases of (12) if f is chosen appropriately.
For instance, taking
f(x)=N -x- 1
we get
f'(x) =x-2
and
x,.+1 = x,. -
N _X;;l
-2'
Xn
i.e.,
x,.+l = x,.[2- Nx,.].
40
Chapter 3
In the same way, taking

f(x)=N-x 2
we get
It is not obvious whether or not some of the other relations can be put
in the Newtonian form. This problem can be examined in the following way
(d. also Problem 3.12).
Consider the Dedekind recurrence (9) which we can write as
Xn+1 -
Xn = (- 2x~ + 2NXn)/(3x~ + N).
If this arises from the Newton process for a function
f it must be of the form
This means that f must satisfy the differential equation
f' 3x 2 + N
f 2x 3 -2Nx
We solve this by expressing the right hand side in partial fractions getting
[=-~+~
f
2x x 2 -N
which can be integrated at sight to give
log f(x) =
-~ log
x + log (x 2 - N) + constant.
Hence, taking the constant to be zero, we find
We now inquire whether any iteration process of the form
can be regarded as generated by a Newton process applied to a certain

equation f(x) = O. We proceed formally and have to solve
f(x)
x - f'(x) = g(x)
i.e.,
f'(x)
f(x)
1
x - g(x}"
Hence
logf(x)=
41
(t-g(t)t 1 dt.
The arbitrary constant is irrelevant since the correction is -f(x..)/f'(x..).

Let us take, to begin with, two examples where there is linear convergence:
g(x) = (1- N)x + 1
which gives
f(x) = (Nx _l)l/N,
f'(x) = (Nx _l)(1/N)-1.
which gives
IN- X)2/.JN,
f(x)= (~
"N+x
Note that in each case f'(l) = 0 so there is no contradiction with the fact that
the Newton process is quadratically convergent-this fact will be established
assuming that f'(l) =1= O.
In the case where
we have
f'(x) =2x.
Here f'(JN) =1= O.

In the Dedekind case we found
for which
and
f"(x) = iX- S/2 (X 2 - N).

Here we have f'(JN)=2JN and f"(JN)=O so that not only will the
argument of p. 48 apply giving quadratic convergence but the coefficient of
(x.. - IN"f, which is !f"(JN)If'(JN), vanishes so that we have (at least) cubic
convergence.
42
5.
Chapter 3
PRAcnCAL COMPUfATION AND THEORETICAL ARITHMETIC
It is important to distinguish between the two subjects in the title of this

section: this distinction can be made clear in the present context.
Take the square root recurrence
x,,+l =!(x" + (N/x,,)).
Suppose Xo ~ -IN; then we can never actually attain the correct square root
for
x"+l--IN=(x,, -IN)2/(2x,,)
so that if xo~-IN, Xl~-IN, .... Further, if xo>-IN
x,,+l - x" = (N - x~/(2x,,)
and so the sequence Xl> X2, .. is strictly decreasing. If Xo < -IN then Xl>-IN
and Xl> X 2 > .... This is from the point of view of theoretical arithmetic. In
contrast, from the point of view of practical computation, an infinite descent
is impossible: there are only a finite (but large) set of numbers available as
outputs of a computation. In practical computation it is necessary to choose
a starting value Xo and to choose a "stopping rule", to decide what x" to
accept as the square root of N. It is not appropriate here to discuss in detail
how these choices should be made. In connection with the choice of xo, note
that a choice e.g., of Xo = 1 would be simpler than allowing Xo to depend on
N, which would probably save time. Again, one could fix the number of
iterations used, to cover the worst case, but this would be time wasting in
many others; alternatively, one could decide to stop whenever two consecutive x,'s were the same or when the sequence turned back i.e., when
x,+1 > x,; or, one could check at each stage whether Ix~ - NI was less than a
prescribed tolerance.
Whenever a specific algorithm is chosen, it should then be examined
thoroughly so that it could be guaranteed that the output S, corresponding
to the input N, would be near to -IN either in the sense that
IS-JNl<Bl
or
IS2-NI<B2
where Bl> B2 are particular (small) numbers. These examinations depend
very much on the machine being used, especially on the arithmetic unit, so
that the E'S will differ from machine to machine.
Returning again to the distinction between theoretical arithmetic and
practical computation we note that while in the first we are not concerned
43
with the size of the numbers which occur in our computation in the second
we must be sure that all intermediate results of our computation remains
within scale i.e., do not cause overflow or underflow on the machine being
used. To guarantee this again requires a detailed examination of the
algorithm and of the machine.
Another example of the distinction we are making is the following:
Theoretically the harmonic series
l+!+i+l+ ...
is divergent. If one tries to sum this on a computer we will have a finite
series, for n -1 will become and remain zero from the point of view of the
machine, and apparently we have convergence.
Chapter 3, Problems
3.1. Demonstrate the relative speeds of the various algorithms for N- 1 and
for N 1/ 2 by printing out in parallel columns the corresponding values of x".
Observe how the choice of Xo affects the speed of convergence.
3.2. Discuss the behavior of the following sequence when

O<N<l,
x,,+1 =!(x" + NX;;I).
Xo
is given and
3.3. Discuss the behavior of the following sequence when Yo is given and
when O<N<1.
Yn+1 = 2y~(3y~- N).
Show that it arises from the solution of x 3 - Nx = 0 by the Newton method.

3.4. 0N. G. Hwang-J. Todd) Discuss the behavior of the following sequence when Xo is given and when O<N<l, x,,+1 =x,,(3N-x~)/(2N).
3.5. Obtain J3 to within 10-16 by the use of convenient recurrence relations
and a desk -calculator.
3.6. Discuss the convergence of the sequences defined by
x,,+l =~X(x,,)-!Y(x,,)
and by
x,,+1 =!Y(x,,)+!Z(x,,)
where
X(x)=![x+Nx- 1 ],
Z(x) = x[3N - x 2 ]/2N.
Chapter 3
44
3.7. Discuss the behavior of the following sequence when

when N is positive
Xo
is given and
x,,+1 = x,,[4- Nx~/3.

Compute [1.2345678]-113.
3.8. (H. S. Wall) Let
f(x, y) =
a + xy(x n- 2+ x n- 3 y + ... + yn-2)

n-l+ X n-2 Y+ + y n-l
X
where n is an integer 2:2. Show that if a > 0, xo> 0, Xl>
and if
p=o, 1,2, ... ,

then
3.9. What is the order of convergence of the sequence in Problem 3.8?

3.10. Discuss the convergence of rn = pJqn where
Pn = apn-l + Nqn-h
qn = Pn-l + aqn-l'
n=1,2, ...
where Po, qo are given, in the following cases

(i) N>O,
a>JN,
(ii)
N>O,
O<a<m.
3.11. Experiment on your computer (with numerical output or with display)

with the recurrence relation
Zn+l
=~[Zn + ~]
when A is real and positive and when Zo is complex. Can you conjecture
what sets of Zo give convergence?
Continue your experiments in the case when A = ae i9 is complex. For
what set of initial values Zo do you expect convergence?
Examine the case when (J = 7T and Zo is real.
Examine the case when A = - 3 and Zo = 1.
3.12. Show that the recurrence relation
x,,+l = x" (p + 1- Nxr:J/p
arises from a Newton process for N-1IP.
45
3.13. Discuss the recurrence relation of Dedekind

x,,+l
where N;::: 0 and
Xo
= [x~ +3Nx,,]J[3x~ +N]
is given.
3.14. Discuss the linear scheme for N- 1 in the light of the general theorem.
3.15. Discuss the convergence of the iteration scheme for a zero of x + x 3
3.16. Use one of the recurrence relations for mto make a table of mfor
N = 12451(1) 12500. Describe carefully your stopping rule. If V is the
50-dimensional vector with components VN =m =SQRT N, N =
12451(1) 12500 find the following three norms of V:
IIV1b= ~L v~,
Here SQRT N indicates the output of the machine subroutine for
./N.
CHAPTER 4
The Solution of Equations
1.
QuADRATIC EQUATIONS
It is known that there is no formula (involving radicals only) for the

roots of a general equation of degree exceeding 4. The familiar form for the
roots of the quadratic
is
x = -!p .J(p2/4) - q.
The formulas for the case of cubics and quartics are given in algebra texts
and the reader should try to use them, checking the results obtained from
tables.
The simple formula above need not be the answer to the problem of
solving a quadratic. H we use a typical floating point machine in the case
q = 1 and p large and negative, say, there will be severe cancellation in the
calculation of the smaller root from the above formula
e.g., p=
q=l
gives x_ =
which has a relative error of
to avoid this pitfall. We have
x+ = -!p +.J{(p2/4)-1}
%. It is easy
and
x_= l/x+
Another disadvantage of this formula is the following. For suitably

chosen large values of -p, q we may find that although the data and the
solution are well inside machine range, there is overflow during the calculation and no results. E.g., if we take
p=
q=
in our machine we fail since p2/4 is out of range. We can counter this by
calculating the radical as
P SORT [1 - (q/p )/p].
This matter has been discussed at length by G. E. Forsythe.
2. BAD
47
EXAMPLES
The following examples (due to A. M. Ostrowski, G. E. Forsythe and 1.

H. Wilkinson) show some of the troubles which can arise with equations of
higher degrees and indicates that the approximate solution of polynomial
equations, which would appear to be a comparatively simple problem, is far
from being so.
(1)
has four roots 1,1,1,1. If we change the right hand side to 49 x 10~8 Z2 i.e.,
change the coefficient of Z2 by less than 1 part in 107 , then (1) becomes
(z -1)4 = (7 X
{(z -1)2-7 X
1O~4
1O~4
Z)2.
z}{(z-1)2+7x 1O~4 z}= 0
i.e., with roots

1.02681,0.97389,0.99965 iO.026455,
i.e., changing by nearly 3 parts in 100.

(2)
has 10 zero roots. If we change the right hand side to 1O~lO the roots of the
new equation all have absolute value 1O~l.
20
(3)
I1(z-r)=O, i.e.,
r=l
z2-210z19 + ... +(20)!=0

has roots 1,2, ... ,20.
If we change the coefficient of
about
1O~9,
Z19 from 210 to 210+2~2\ a relative error of

the roots of the new equation are
1.000000000
2.000000000
3.000000000
4.000000000
4.999999928
6.00000 6944
6.999697234
8.007267603
8.91725 0249
20.84690 8101
10.095266145 0.64340 0904i

11.793633881 1.65232 9728i
13.992358137 2.51883 0070i
16.73073 74662.81262 4894i
19.50243 9700 1.94033 0347i
showing major discrepancies.

This example shows that it is not only in the case of multiple (or
clustered) roots that numerical trouble can arise.
48
Chapter 4
An important question now requires an answer: what do we mean by a

good approximation xa to a root Xo of
f(x) = O?
Do we mean that
IXa -
xol
is small
or that (the residual)
If(xa)1
is small?
Results of the first kind are called "forward" error estimates and those
of the second are "backward". Results of the second kind are very acceptable to the experimental scientist whose data is generally subject to experimental error and to the numerical analyst who cannot introduce most
numbers exactly into his machine. Let us quote two results about the simple
case of square roots i.e., when f(x) == X2 - N.
(4) Given an s-place fixed point binary machine, there is a square root
algorithm such that the alleged square root and the true square root never
differ by more than 0.76 x 2-' in absolute value.
(5) Given a floating point machine, there is an algorithm which produces
from an input ii = 2'1 X A, 1:5 A:51 an output whose square differs from ii
by at most
2'11-27
x 4.38.
Although in many cases it is true that when x - Xa is "small" then f(xa)

is "small", this is not always so. Actually both types of error analysis are
usually required in serious study. However they are necessarily long - the
details of the actual machine program (as realized by the compiler) and the
details of the arithmetic unit of the computer must be allowed for. We
cannot pursue this question further at this stage: All we can do is to advise
the user of packaged programs to be careful.
3. NEWfoN's PROCESS
The classical result of Newton, that if x" is an approximation to a root
of f(x) = 0, then
x,,+1
= x" -[f(x,,)]lf'(x,,)
is a better one, which is motivated in the diagram, is well suited for the
approximate determination of the roots of f(x) = 0, once they have been
separated. We outline a proof that the sequence of approximations has
quadratic convergence.
49

y
'(Xn)f---~
Fig. 4.1
Suppose
f(~)
= 0 and that we may expand f(x) in a Taylor series
f(x) = O+(x -~)f'(~)+(x -~ff"(~)/2! + ....
(6)
It follows that
f'(x) = f'(~)+(x-~)f"(~)+ (X;!~)2 f"'(~)+ ....

Hence, if f'(x) f: 0,
f(x) (x - ~)[f'(~) + (x - ~)f"(~)/2! + (x - ~)2f"'(~)/3! + ... ]
f'(x) =
f'(~)+(x-~)f"(~)+(x-~)2f"(~)/2!+ ...
f' (~) f: 0 we have

f(x) = ( _ ~)[1 +~(x - ~)f'(I;)/f'(I;) + Ox _1;)2)]
Assuming further that

f'(x)
l+(x-~)f'(~)/f'(~)+(Jx-l;f)
which, from the results of Chapter 2, gives
J,~~ = (x -I;)[l-~(x -1;){"(I;)/f'(I;) + (Jx -I;f)]
(7)
Now, using (7) for x = x" we find
x,,+1 -
~ = x" -
[f(x,,)/f'(x,,)] -
= (x" -I;)-(x" -I;)
x [l-!(x" -
~)f"(~)/f'(~) + (Jx" - ~)2)]
=![f'(~)/f'(~)](x" -~f+(Jx" _~)3)
as announced.
so
Chapter 4
To establish convergence properly we require more elaborate considerations. The main difficulties are that we do DDt know ~ nor how close Xo
must be to it to make sure x.. ~~.
We give here an account following Henrici. We shall introduce the
conditions as we require them in the proof - these conditions are ascribed to
Fourier. The convergence of Newton's Method can be established under
much weaker conditions and in much wider contexts.
We first want to be sure that there are zeros of f(x) and that we know
what zero we are going after, i.e., we want to separate the zeros. To deal
with this point we assume that in an interval [a, b], where a < b we have,
say,
(8)
f(a)<O,
f(bO.
Continuity of f in [a, b] then assures us of the existence of at least one zero

in [a, b]. Further, the Newton algorithm can only be used if f' exists and if
we assume that
(9)
f'(xO
in [a, b]
so that f is increasing then there can be at most one zero in [a, b]. We shall
denote this zero by ~.
We shall assume that
(10)
f"(x):s;O
in [a, b]
so that the graph of f(x) is convex downwards.

We want to be sure that the Newtonian construction never leads us
outside the interval [a, b], where e.g., the function might not be differentiable and so we could not continue.
We shall see that if Xo satisfies a:s; xo:S; ~ then a:s; xo:S; Xl :s; ~ and there
is no difficulty in continuing. However there could be trouble if ~:s; xo:S; b,
e.g., if the graph of y = f(x) had a nearly horizontal tangent at b, then the
correction, f(xo)/f'(xo), would be large and negative and Xl could be to the
left of a. A convenient sufficient condition to ensure that this circumstance
cannot occur is that
(11) the tangent to y = f(x) at
= b intersects the
axis between a, b.
That this suffices is almost clear geometrically. The reader is advised to solve
Problems 4.5, 4.10 before continuing.
We now proceed with our proof. Take an initial guess Xo such that
a :s; xo:S; b. Let ~ be the unique zero of f(x). We distinguish two cases
a:s;xo:S;~, ~<xo<b. [If xo=~ we have ~=XO=XI .... ]
51
Case 1. a::; Xo ::; ,
We shall prove that x,.::;, for all n and that x,. t . It follows that this
bounded monotone sequence converges to an 'I'J::; f Since f, f' are continuous, and f' f:. 0, passing to the limit in
gives
so that
But , is the only zero of f in [a, b J. Hence ,= lim x,..
Assume that x,.::;f The First Mean Value Theorem gives f(')-f(x,.)=
-f(x,.) = (' - x,.)f'(') where x,. ::;,::; g. Since ["(x)::; 0 it follows that f' is not
increasing and so f'(')::; f'(xn) giving
so that
Since, by hypothesis xo::; g, we have an induction proof that x,.::; g.

We have to show that Xn t . Since x,.::; ~ we have f(x,.)::; 0 and since
f'(x,.) > 0 we have [-f(x,.)/f'(x,.)J ~ 0 so that x,.+l ~ x,..
This completes the discussion of Case 1.
Case 2. g::; Xo ::; b
This is not similar to Case 1. Indeed we show that under our hypotheses
a::; Xl ::; ~ and we get back to the first case.
Again using the First Mean Value Theorem, we have
where
~::;,::; Xo.
Since
f'
is not increasing we have
so that
Xl = Xo -
[f(xo)/f'(xo)]
::; Xo- (xo- g) = g.
The fact that a::; Xl follows from the condition (8). To justify our
diagram we have to show that the point (xo, f(xo)) is below the tangent at b.
52
Chapter 4
The tangent has equation

y -f(b) = f'(b)(x-b)
and its ordinate at Xo is

(xo- b )f'(b) + f(b)
which is greater than f(xo) since the difference

f(b) - f(xo) + (xo - b )f'(b) = (xo - b )[f'(b) - f'(C)] 2: 0
by the First Mean Value Theorem, where Xo < C< b, and because
f'(b)-f'(C)~O. The fact that the tangent at Xo has a greater slope than that
at b follows because f'(x) is not increasing.
This completes the proof of convergence of the Newton Process.
Provided that we can "separate" the roots, the Newton process is
reasonably satisfactory. We shall not discuss the problem of separation
here - Sturm's Theorem can be used in the important case when all the roots
are known to be real.
4.
NEWfON'S PROCESS FOR POLYNOMIAL EQUATIONS
To use the Newton formula

conveniently when
f(x)
= aoxn + a1x n- 1+ ... + ~
requires that we would be able to compute the polynomials f(x), f'(x)

efficiently. This is a problem in itself and we restrict ourselves to the Homer
method: if we compute
fr+! = xfr + a.+ b
= 0,1, ... , n-l
then we find after n multiplications and n additions
fn = f(x).
Is this the best possible algorithm? It is certainly better, e.g., than direct
calculation of each term and then summation. Consider, e.g., 3x 2 + 2x + 1. If
we proceed thus:
we use 3 multiplications and 2 additions as compared to the 2 multiplications and 2 additions required if we use the Homer scheme
3x, (3x + 2), x(3x + 2), x(3x + 2) + 1.
53
We shall accept this scheme and we now observe that it is possible to

compute both f(x) and f'(x) conveniently as follows, as a development of the
idea of synthetic division. In the case when f(x) = 4x 3 + 3x 2 + 2x + 1 and
x = 2 we proceed as follows
4
/' 8 22 48
ul . r . b 2
4 11"" 24.1' 49 = f(2))"m tip lcation y
4
/'!.)"38
19 62 =1'(2)
In the general case the formulas are, when

f(x) = aoxn + alx n - 1 + ... + ~-lX +~,
f'(x) = naox n - 1 + ... +~-1>
as follows
fo=ao,
fn = f(x),
go=O,
gn = f'(x).
It should be clear that precise computations of f(Xn) are essential for

proper applications of the Newton process: when we are near the desired
root, !(Xn) will be small and round-off errors in the calculations may
completely obscure the value of f(Xn) and the value of our results. Allowance for this must be made in a good program.
Chapter 4, Problems
4.1. Find the root of 3x 4 +4.6x 3 -O.7x 2 +9.2x+ 10 = 0, near x = -1, using a
desk calculator.
4.2. Write a program for the Newton process, applied to a polynomial of
degree s10, using the scheme outlined in the text to evaluate the function
and its derivative. In the first place this program may be written in the case
of real coefficients and for real roots and tested on the quartic equation of
the preceding problem. A complex program should then be written and used
on the quartic, to find roots near 0.7 1.2i.
54
Chapter 4
4.3. (L. Collatz) Write a program to obtain the quotient q(x) and remainder r(x) when a polynomial
aOxR+alx R- 1+ ... +~-lX+~
is divided by a quadratic x 2 + r:x + s. Apply this in special cases.
4.4. Discuss the behavior of the Newton process, when the root sought is a
multiple root.
4.5. Discuss the Newton method for the determination of the solution of
H(x)=3x 3 -3x+28 =0
between -1 and -1- ~8 where 15 8 < $ .
4.6. Justify the scheme given for the calculation of {(a), f(a) when
{(x) = aoxR + a1x R- 1+ ... +~.
4.7. Repeat Problems 4.1, 4.2 in the case of the quartic

x 4 -2x 3 +2x 2-6x +4= 0,
given that it has roots near 0.75, -0.4 l.5i.
4.8. (E. Halley, 1694) Show that the iteration formula

x,.+l = x,. -
2{(x,.)f(x,.)
]
2{!'(x,.)F - !(x,.)f'(x,.)
has cubic convergence to a simple zero of {(x).

What is its behavior in the neighborhood of a double zero of {(x)?
4.9. Find the two least positive roots of the equation tan x = x. Find an
approximation to the root near (n +!hr in the form
x
= (n+!)7r-an +!hr)-l- bn +!)1T)-2- cn +!)1T)-5 ....
4.10. lliustrate the Fourler-Henrlci proof of Newton's Method in the cases of

{(x) = 1-2x2
with a = -1, b = -0.5 where
Xo
is taken to be -0.8 and -0.6.
CHAPTER 5
Uniform Convergence and Approximations
Many different kinds of convergence require discussion in mathematics.

The original kind is the following:
"The series L u, is convergent to a sum U if given any B > 0 there is an
no = no( B) such that the remainder after n terms
n-l
rn=U-
Lo u,=Un+Un+1+
is less in absolute value than B for all n ~ no".

This condition of convergence is sufficient to ensure that we can
multiply a series by any constant, or add two series term by term and get the
expected results:
(au,) = a u,
L
L u,+ L = L(u,+v
Vr
r )
The next variety of convergence to be studied is absolute (or unconditional) convergence.

"A series L u, is absolutely convergent if L lu,I is convergent".
When a series is absolutely convergent we can rearrange it at will
without destroying its convergence, or altering its sum. A series such as
which is convergent but not absolutely convergent can be rearranged so that

it is convergent to any (real) number or divergent to +00 or to -00. Absolute
convergence of L u, and L Vr (indeed of one of these series) is sufficient to

ensure that multiplication (in the Cauchy sense) leaos to a correct result
where
1.
Wr
UOVr
+ UIVr-l + ... + u,vo.
UNIFORM CONVERGENCE
Operations on series whose terms depend on a variable, such as passage

to a limit, integration or differentiation with respect to the variable arise
naturally. We need to know conditions sufficient to ensure the truth of
56
Chapter 5
relations of the form
~ L u,(x) = L [!~ u,(x)]
(1)
or
(2)
it being assumed for instance, that each u,(x) is continuous and that L u,(x)
is convergent for x near a in the first case, and for x between a and b in the
second.
It will appear that the concept of uniform convergence is the appropriate one in the present context. This concept is particularly natural in the
numerical study of convergence. We discuss, for two examples, the dependence of the no( E) on the argument x.
Consider first the series
co
Since
n~o (n + x + 1)(n +x +2)'

1
-------------=-----(n+x+1)(n+x+2) (n+x+1)
1
(n+x+2)
we have
n-1
L= 1+x
,=0
1
n+1+x
so that the sum is (1 + X)-1 and the remainder after n terms is (n + 1 + X)-1.
Our second example is more complicated. It was studied by G. G.
Stokes in 1847. Since
x(x+2)n 2 +x(4-x)n+1-x
Un == n(n + 1)[(n -1)x + 1](nx + 1)
[1~- n 1]
[2
2]
+ 1 + (n -1)x + 1 nx + 1
we have
,t1 u,=[i- n~1]+[i- nx2+1]=3- n~1-nx2+1

so that if x ~ 0 the sum is 3 and the remainder after n terms is
1
2
---+----.
n+1 nx+1'
however, if x = 0, the sum is 1 and the remainder after n terms is (n + 1)-1.
57
"o(e.x)
x
Fig. 5.1
There is a discontinuity in the sum-function at x = 0, although the individual

terms are continuous there.
Let no(e, x) be the least no such that rn(x), the remainder after n terms
in the series under discussion, satisfies
Irn(x)l<e
if n~no(e,x).
A graph of no( e, x) in the case of the first series is of the form in Fig.
5.1 while that in the second series is of the form in Fig. 5.2.
These examples suggest the following definition.
Definition 3. A series L u,(x), whose terms depend on a variable x,
which is convergent to a sum u(x) for x E X, is said to be uniformly convergent
"o(e.x)
100 L_--=::======~
Fig. 5.2
__x
58
Chapter 5
to u(x) in X, if, given any e >0, there is an no= no(e), independent of x, such
that
ifn~no(e).
It is clear that our first series is uniformly convergent to its sum in the
interval X =[0,1] and that the second is not uniformly convergent in the
interval X = [0, 1].
It can be shown that the results (1) and (2) are true if the series is
uniformly convergent in an interval including a or in the interval [a, b].
An equivalent approach to the definition of uniform convergence and
one which is equally appropriate in the numerical context is the following.
Take the series I u,. (x) convergent to u (x) for x E X. Fix n and consider
rn(x) as a function of x: it will have an upper bound, M n, say. This means
that the error committed by truncating our series after n terms does not
exceed M,. for any x EX. H we now let n vary, we can ask whether Mn ~O
as n~oo, or not. In the first case, we can get an arbitrarily good approximation to u(x), for all x E X, by taking an appropriate (fixed) number of terms
and in the second, this is not possible. The first case is exactly that of
uniform convergence. [0. Problem 5.4].
2.
WEIERSTRASS THEOREM. BERNSTEIN PROOF
Since polynomials (or the ratios of polynomials) are the quantities

which we can compute, the question arises as to whether we can approximate the functions with which we deal by polynomials. We are therefore
interested in knowing whether, given a function f(x) continuous in an
interval [a, b], there is a series of polynomials uniformly convergent to f(x)
in [a, b] i.e., if
f(x) =
L fr(x)
where each f,. is a polynomial (not necessarily of degree r). H this were so we
could then calculate f(x) approximately throughout [a, b] by truncating the
series at an appropriate point. This, fortunately, is true. The result was
proved by Weierstrass in 1885 and there have been many proofs given since
then. One of these was given by S. N. Bernstein in 1912. It has apparently
an enormous advantage over most of the others in so far as that the
polynomial approximations to an arbitrary f(x) can be written down in terms
of the values of f(x) at the rational points.
59
Definition 4. The Bernstein polynomials Bn(f, x) for a function f{x)

continuous in [0,1] are defined by
Theorem
s.
Bn(f, x)~f(x) uniformly in [0,1].
Take the case f(x) = x 2 Then an easy calculation with binomial coefficients (cf. Problem 5.5) gives
Bn(f, x) = x 2 +[x(1- x)/n].

Since for 0:5x:51, we have Ix(1-x)I:5i, it is clear that
IBn(f, x)- x 2 1:5 (4n)-1

for all x in [0, 1]; hence the convergence of Bn to x 2 is uniform, and we have
verified the theorem in this special case.
A proof of the theorem in the general case can be motivated through
elementary probabilistic considerations. We give this motivation. Suppose
we consider repeated independent trials of an event with probability x of
success. What is the probability of getting exactly one success in three trials?
The possible outcomes and the probabilities are:
S F F
F S F
F F S
x(1-x)(1-x)
(1- x)x(1- x)
(1-x)(1-x)x
Hence the probability is
More generally, the probability of getting exactly r successes in n trials is
A check on this is given by the following reasoning: in n trials we must have

0, 1, 2, ... or n successes and hence we ought to have L~=o Pn,r = 1 and this is
so since L(~)xr(1-x)n-r is just the binomial expansion of (x+(1-x)t.
When n is large the probability of getting r successes will be large when
r is near nx and small elsewhere. This means that
r near nx
Pn,r =l=
1,
r not near nx
Pn,r =l=0.
60
Chapter 5
Hence the significant part of the sum for Bn (f, x) occurs for r near nx but,
since f is continuous, f(rln) is then near f(x). Thus
Bn(f,x)=';=f(x).
We have noted the apparent advantage of the Bn(f, x), that they can be
written down as a mean of the values of f at the points rln, r = 0, 1, 2, ... , n.
However, they have a disadvantage in so far as that the convergence is slow:
thus even for f(x) = x 2 , the error is O(n- 1), according to Problem 5.5. This
order result is true at any point Xo when f(x) has a second derivative at Xo.
This means that in order to get a 6 decimal approximation to sin x we have
to take the Bernstein polynomial of order about a million. Much better
results are possible - see NBS Handbook, p. 76.
3.
BEST APPROXIMATION: CHEBYSHEV POLYNOMIAlS
The problem of best approximation to continuous functions was first

studied by Chebyshev in 1853 but there have been rapid developments,
both theoretical and practical, since 1950. Most practical work has been
concerned with the approximation of functions f(x) continuous in [a, b] by
polynomials, or by rational functions. Chebyshev showed that for any
positive integer n there was a unique polynomial of best approximation, in
the sense that there is a Pn(x) such that
En = max If(x)-Pn(x)l:s max If(x)-Pn(x)1

asxsb
asxsb
where Pn(x) is any polynomial of degree nand Pn(x) is characterized by an

equal-ripple behavior of the error. Note that E,. i O. It is only a very few
cases that Pn(x) can be written down. The classical case is that in which
f(x) == 0 - here we must normalize the polynomials in order to make our
problem sensible e.g., by requiring that
Pn(x)=x n +alx n- 1 + ... +~-lX+~.
The problem can be restated: find that polynomial of degree n -1 which
gives the best approximation to xn. The solution is given by
Theorem 6. No matter what the coefficients at. a 2 ,
-l:sx:sl
are
-l~x:=s;l
and there is equality if and only if Pn(x) = fn(x) where fn(x) =

2- n + 1 Tn (x) = 2- n + 1 cos (n arccos x).
61
Proof. The proof depends on the equal-ripple behavior of Tn(x). First,

max-1SxSlITn(x)1 = 21- n. Suppose
max Ixn + b1 x n - 1 + ... + bnl = fJ. < 21- n.
-lsxsl
Consider
e(x)= Tn(x)-(xn+blXn-l+ ... +bn).

This is a polynomial of degree n -1 at most. Its values at the points
x". = cos m7T/n, m = 0, 1, 2, ... , n have the same sign as those of Tn (x".) i.e.,
alternately plus and minus. Hence e(x) has at least n zeros, but this means
that it vanishes identically which implies fJ. = i- n , a contradiction. This
shows that
The proof of the equality statement is left to the reader.
Another case in which the polynomials of best approximation can be
given explicitly is that of f(x) = (1 +xr 1 in the interval [0,1]. This is
discussed in Problem 5.7.
Algorithms which generate a sequence of polynomials which converge
to the polynomial of best approximation have been constructed. For
practical purposes it is sufficient to choose a polynomial of "very good"
approximation and lists of these are available in many places e.g., the NBS
Handbook. For an informal description of such processes see C. Hastings,
Approximations for Digital Computers, Princeton 1955.
In many of these algorithms it is desirable to start off with a "good"
approximation. This can often be done by expanding the function in
question in a Chebyshev series
f(x) =!aO +a 1 T 1 (x) + a2 T 2(x) + ...

and taking an appropriate number of terms of this series. The coefficients in
these Chebyshev series are Fourier coefficients and can be evaluated as they
are: since
+l
-1
dx
Tm(x)Tn(x) ~=
1 x
I. .
cos mO cos nO dO
=!
1. .
{o
=!7T
[cos (m + n)O + cos (m - n)O] dO

if m;en
if m=n
62
Chapter 5
we can find the a's by evaluating the integrals
+1
-1
Tm(x)f(x) .J
dx
(1- x)
r. . f(cos 0) cos mOdO.
1
0
Tables of Chebyshev coefficients for various functions have been given

by C. W. Clenshaw and by C. W. Clenshaw and S. K. Picken in NPL
Mathematical Tables 5(1962), 8(1966).
For examples see Problem 5.11.
Chapter 5, Problems
5.1. Let no(e,x) be the least no such that (n+l+x)-1<e for
Tabulate no(e, x) for say
n~no(e,x).
x = 1, 0.1, 0.01, 0.001, 0

e = 3,1,0.1,0.01,0.001.
5.2. Repeat Problem 1 in the case when the inequality is

(n + 1)-1 + 2(nx + 1)-1 < e.
5.3. Calculate
M,. = max ITn(x)1
Osxs1
when
1
T =--n
n+l+x
and when
Tn (X) =-1+--1.
n+
nx+
5.4. Suppose that u(x) = I u,.(x), the series being convergent in the interval
[0,1] and that u(x)=uO(X)+U1(X)+ ... +Un_1(X)+Tn(X). Let M,.=
maxosxs1ITn(x)l. Let no= no(e, x) be the least integer such that if n ~ no then
ITn(x)l:5:e. Let no(e)=lubosxs1 no(e, x). Prove that
no(e)<oo for all
e>O
is equivalent to limMn =0.
5.5. Compute Bn(f, x) for f= x\ k = 0,1,2,3.

5.6. Write lIe(x)11 =max..-1sxS1Ie(x)l.
63
If e;(x)=lxl-qi(X) evaluate Ilei(x)11 when
ql(X) = _kX4+~X2+~,
q2(X) = -1.06553 7x 4+ 1.93029 7x 2+O.067621,
q3(X) = Ns(-7x 4+ 14x 2+ 1),
qix) = 2(-16x 4+36x 2+3)/(157T).
Which qi(X) gives the best unifonn approximation to Ixl in [0, I]?
5.7. a) Sum the series
.J2{i-e cos 0 +e 2 cos 20 -e 3 cos 30+ ... }
where e = 3 - 2.J2, 0 = arccos (2x -1) and O:s; x:s; 1.
b) Show that for a certain 1>
(l+x)-l-7Tn (x)=
(-It-len
4
cos (nO +1
where
en
7Tn (X) = J2m-eTTex) + ... + (-It- l e n - l T:','_l(x)]+ ( - I t - l2 T:','(x)}
-e
where T:','(x) = cos (n arccos (2x -1)).

c) Evaluate
Ei= max 1(I+x)-l-ti (x)1
Qsxsl
when
and when
t2(X) = !+~J2-~x = 1Tl(X),
d) Write down the Bernstein polynomial Bl = Bl(f(x), x) for f(x) =

l and evaluate
(1 +
max 1(I+x)-l-B l l.
xr
Osxsl
e) Use the computer to find the least number of tenns of the series
(1 + x
= 1 - x + x 2- x 3+ ...
which must be taken to ensure a remainder less than 1x 10-5 in absolute

value when x =0.1(0.1)0.9.
5.8. Write down the Bernstein polynomial B 4 (f, x) in full. [Do not use the
summation sign.] If this is
B 4(f, x) = ao(x)f(O) + a l (x)f(1!4) + a2(x)f@+a 3(x)f(3/4)+aix)f(l)
64
Chapter 5
by changing the variable to y = 2x -1 get a Bernstein polynomial B*(f*, y)

appropriate for the interval [-1, 1].
Evaluate B* when f* = Iyl and draw a rough graph of B* and of
e*(y)=lyl-B*. Find
max le*(y)l.
-t:s;y:s;t
5.9. The Pade fractions r....v(x) where IL, v = 0,1,2, ... , corresponding to a
power series C(x) = I c,.xn, are the unique rational functions
r....v(x) = N ....v(x)/D....v(x)
where N is a polynomial of degree :5v and D a polynomial of degree :51L

such that, fOrnlally,
consists of powers of x greater than IL + v.

Compute the first 9 entries in the Pade table for eX i.e., r....v for
IL,V = 0,1,2.
5.10. Is r2.t(x) of Problem 5.9 a good approximation to eX for small x?
Compare this approximation with the truncated Chebyshev expansion
of eX, i.e.,
f(x) =
3,
L a,T,(x)
r=O
where
aO = 2.53213
176, a 1 = 1.13031821, a2 =0.27149536, a3 =0.04433685
and where the prime indicates that the first ternl of f(x) is !ao. Confine your
attention to the interval [-1, 1].
5.11. Obtain the fOrnlal expansions:
(a)
2 4 co (_l)n+1
Ixl =-+4 2 1 T2n(X).
7r 7r n=1 n-
(b) sin (!7rX) = 2Jt (!7r)Tt (x)-2J3 (!7r)T3 (x)+ 2JsG7r)Ts(x)- ....
5.12. Suppose an experiment has a probability 0.6 of success. What is the
probability P20.k(0.6) of exactly k successes in 20 independent repetitions of
this experiment? Give the computed values of
P20.k
(0.6) for
k = 0(1)20
and draw a rough graph of these values. Repeat in the case of 50 repetitions.
65
5.13. For n=20, x=0.7, compute for r=0(1)20
Pn,T(X) = (;)xr(l- x)n-r.

Evaluate L Pn,r(x) and L2 Pn,r(x) where LI is over r such that I(rln) and where L2 is over the residual set.
xl s 0.1
5.14. Find theoretically and practically estimates for the error incurred in
truncating the power series for the Bessel function Jo(x), specifically in
taking as an approximation,
where y =x2/4, in the range Osxs1.

Write a subroutine for computing Jo(x) approximately in the range
Os x s 1, which gives an error signal if an x outside this range is assigned.
Find maxI,,;r,,;lO le(x.)1 where X = r/lO and
T
eT(x) = Jo(x) -[1- a l (x/3?+ az{x/3)4

- aix/3)6 + a4(x/3)8 - as(x/3)10 + a6(x/3)12]
where
al = 2.24999 97
a2 = 1.26562 08
a3 = 0,3163866
a5 = 0.00394 44
a4 = 0.04444 79
a6 = 0.00021 00.
5.15. Discuss the uniform convergence of the series

=
L x 2(1 + x2tn.
n=O
5.16. Discuss the uniform convergence of the sequence
sn(x) = n 2xe- nx
and the truth of the relation limS~sn(x)dx=SHlimsn(x)]dx.
CHAPTER 6
The Acceleration Processes of Aitken and Euler
We have seen that there is a great variation in the speed of convergence

of series or sequences. Some are too slowly convergent to be summed
directly on the fastest computer. This is the case of the logarithmic series
log 2 = 1-~+~-~+ ...
in which a crude estimate for the remainder after n terms is n- 1 so that
2 x 106 terms would be needed to get a value of log 2 correct to 6D.
Various devices are available to transform cheaply a series or sequence
into one which is more rapidly convergent and we shall study some of the
more famous: one due to Euler which is linear and one to Aitken which is
non-linear. These must be used with care. The new series may converge
slower than the original - this often happens if the old series is fairly rapidly
convergent and one is being greedy. Also, in practice, one must not try to
get too great accuracy: the theory will apply if the terms have a certain
behavior and in practice this behavior will be destroyed by the changes
wrought e.g., by round-off errors in the calculators.
In Chapters 9, 10 we will discuss another type of acceleration process:
the h 2 - extrapolation of L. F. Richardson.
1. THE.AITKEN TRANSFORM OF A
SEQUENCE
Suppose x" ~ x and that the convergence is approximately geometric
IAI<1.
(1)
Then
x,,+l-X:
x" -x
A.
Equating the left hand sides of the last two equations

x" +2 - X
x,,+1 - X
Xn +1- X
x,,-X
and solving for the limit we get

(2)
This is correct if we have x" - x = A,\ n: e.g., if x"

Xo = 1, Xl = 1.5, Xz = 1.75 and we find
x=1.75
(1.75 -1.5?
1.75-2(1.5)+1
= 2-2-n .
67
In this case
1.75 +0.25 = 2.
However, if the relation (1) is approximate, it may be expected that the

value of x given by (2) is a better approximation to the limit than those
used. Hence, if we define
(x,,+2 -
x,,+1)2
x" +2 - 2 x" +1 + x"
we may expect that the sequence {i,.} converges faster than {x,,}. This is
indeed the case. We have the following qualitative result: if
\,\\<1
then
Xn-x=o(,\n).
(3)
Recalling the meaning of o( ) we see that this result shows that as

provided A:fO,
n~co,
i,. - x
--~O
x" -x
so that the new sequence converges faster than the original one.
If the original sequence does not have the approximate geometric
behavior, even if it is more rapidly convergent, the new sequence can have
slower convergence. As an example, consider the sequence of approximations to N- 1 given by Zo = 1, Zn+1 = zn(2- Nzn) which have quadratic con-
vergence. The behavior of the differences of this sequence and the sequence
zn is given below in the case when N = 1.
Zn
6.zn = Zn+1 -
Zn
0.5
-0.125
1.5
0.375
-0.2578
1.875
3.000
0.1172
1.9922
2.0455
68
Chapter 6
We note that it does not pay to go too far with this improvement. The
correction term -(x,,+2 - x,,+1)2/(x,,+2 - 2x,,+1 + x,,) is the ratio of two quantities which tend to zero as n~oo and will in practice be poorly determined.
We have stated that we want to have cheap acceleration devices. The
Aitken process is such a one. The main effort involved is in the determination of the terms of the sequence - the computation of the transformed
sequence only requires a few operations.
2. THE EULER TRANSFORM OF AN ALTERNATING SERIES

We introduce the operator !l. and its iterates by the equations:
!l.x" = x,,+1 - x", !l.2x" = !l.(!l.x,,) = (x" +2 - x,,+1) - (x,,+1 - x,,)
= x,,+2- 2x,,+1 + x",
and generally, !l.r+1x" = !l.(!l.rx,,). It is also convenient to define

Clearly
E
= 1 +!l..
These, and other operators, will be discussed further in Chapters 8, 9, .10.

We can usually manipulate these operators freely with impunity. For
instance:
UO-U1 +U2-'" =(1- E+ E2- ... )uo
=(1+E)-1 uo
= (2 + !l.)-1 UO
=!(1 +!!l.)-1Uo
=![uo-!!l.uO+i!l.2uo- ... ]
This is the Euler transform.

In some cases the differences of a sequence are smaller than the terms of the
sequence itself; moreover, in the last series the r-th difference has a factor
(_1)'2-(r+1) and so the terms in it may decrease rapidly and it may have
faster convergence than the original series. Let us examine this in the
classical case of the series
In 2 = l-!+i-i+i- ....
It is possible to apply the Euler transformation directly to this series,
but it is more convenient to apply it to the tail. We obtain
1-!+ ... -i=0.63452 3809,
69
and we difference the sequence ~, l~' -h, -b., ... thus:

0.111111111
-0.011111111
0.100000000
+0.002020202
-0.009090909
-0.000505051
0.090909091
+0.001515151
-0.007575758
+0.000155402
-0.000349649
-0.00005 5505
0.083333333
The next leading differences are +0.000022212, -0.000009740, ....

Using all of these, we find
L (-It-lin = 0.055555556+0.00277 7778+0.00025 2525 +0.000031566

00
+0.000004856+0.000000867 +0.00000 0174+0.00000 0038.
giving
In 2 = 0.69314 7169,
which we compare with the true value
In 2 = 0.69314 7181.
It can be proved that the Euler transform of a convergent series always
converges (and to the same sum). This is an exercise in the manipulation of
binomial coefficients. [Problem 6.12].
We conclude our account of the Euler transform by showing that the
following conditions are sufficient to ensure that the transformed series
converges faster than the original one.
(a) u" is a null sequence such -that for all m, n 2::: 0
(- Ll)mu" 2::: 0
and
(b) (u,,+l/u,,) 2::: k >!.

Condition (a) is called "complete monotony" and condition (b) requires
that the terms do not decrease faster than a geometric progression with
common ratio !.
Let rm r~ be the remainders after n + 1 terms in the original series and
in its transform. Then
(-ltrn = -u,,+l + u,,+2 - u,,+3 + u,,+4 - ...
= Llu,,+l + Llu,,+3 + ...

Since {Llu,,} is a monotone null sequence we have
Irnl2:::-~[Llu,,+l + Llu,,+2 + Llu,,+3 + Llu,,+4+ ... ]
= ~u,,+l
>!uok n + 1
70
Chapter 6
The transformed series is a positive series and

r~=
:5
IA"+1uol
Uo [
1 1
]
Uo
2"+2 + ... 2"+2 1 +2+ 22+ ... =2"+1
Hence
I~:I:5~ (2~r <
1.
Thus if Un+l/Un --+ 1 then the Euler transform converges practically like
IZ-".
Chapter 6, Problems
6.1. H x,. - x = AA n + o(A ") where IAI < 1 show that, in the notation of the
text,
x = o(A ").
x.. -
6.2. Write a program to compute the Aitken transform of a given sequence,

incorporating such safeguards as you see fit. Apply it to good and bad
examples, and in particular to the Borchardt and Carlson sequences of
Chapter 1.
Alternatively, thoroughly check the Aitken subroutine available in your
system.
6.3. Write a program for iterating the Aitken transform: i.e., going from x,.
to x,. to
to i,., say. Construct an example for which this is effective.
x..
6.4. Apply the Aitken process twice to the sequence {On}, checking your
results with those given
0.734347668
2.146907221
1.180755792
1.832664021
1.39566 2355
1.68771 8271
1.49279 9008
1.62280 9187
1.536116743
1.59391 7337
1.55538 1939
1.57317 3615
1.57078 8289
1.57079 6499
1.57079 6327
71
6.5. Write a program to compute and sum the Euler transform of an

alternating series.
This should be able to handle the case when the general term of the
original series is given by a formula e.g.,
Un = n- 1 as in our example, or
Un
= (1n n)/n,
or when the terms are stored as numbers. The program should also allow for
a choice of the division of the series into a head and tail.
Alternatively, thoroughly check the Euler subroutine available in your
system.
Apply this program to compute
L (-1)n(1n n)/n =-1(1n 2f+-y In 2, L(-1)"(ln n)2/n,

00
00
and
L (-1)"(1n n)3/n.
00
6.6 Obtain formally the relation
1
L Unxn =-1L (~nao)
x
00
00
"=0
n=O
x )"
-1- .
X
[This gives the Euler transform on putting x = -1.]
6.7. Show that
~nao= (-1)n[ ao- (~)a1 + (~)a2-'" +(-1)"Un ].

Hence, or otherwise, calculate
Then write down the Euler transform of 1-!+i- ... and sum it.
6.8. Extend the results of Problem 6.7 to the case where there is a delay of r
in starting the Euler transform.
6.9. Evaluate the Euler transforms of the three series, each summed from
s =0 to s =00,
Verify that each converges to the proper sum and state whether or not the
speed of convergence has been improved.
72
Chapter 6
6.10. Apply the Euler transform to Gregory's series for 1f/4,
Lo (-1t/(2n +1),
00
(1f/4) = arctan 1 =
both practically and theoretically, with and without delay.
6.11. (A. van Wijngaarden). If

Euler transform of the series
Mf(n)=~[f(n)+f(n+1)] show that the
can be written as
~[vo+Mvo+Wvo+ ... ].
6.12. Show that if Sn is the sum of the first n terms of a series (so = 0) and if
Sn is the sum of the first n terms of the Euler transform of the series then
Using this relation show that if
Sn ~ s
then
Sn ~ s.
6.13. Work out the sums for the series
1-~+t-i when we use the Euler

transform with delay v, for v = 0, 1, 2, 3,4. What is the optimal delay?
6.14. Experiment with the Euler transform in the evaluation of

=
B1(x)=-y-Iogx+ k~l
(_l)k-lxk
k(k!)
for moderate values of x.
6.15. Apply the Euler transform to determine the sums of the series
00
(-1t
00
(-1t
00
(-1t
00
(-1t
~ 3n+1'~4n+1'~5n+1'~6n+1
6.16. Discuss the behavior of the Aitken transform {zn} of the sequence {zn}
defined by
0<N<1,
Zo= 1,
In particular show that although {zn} converges to N- 1 it does so

infinitely slower than {zn}, i.e.,
-1
Zn+Z- N
0
-l~
zn+Z- N
CHAPTER 7
Asymptotic Series
1. A
CLASSICAL EXAMPLE
The concept of asymptotic series is of great importance in much of

mathematics. We shall discuss the case of the infinite integral
f(x) =
f'
t-1e x - t dt.
This is certainly convergent for x> O. Indeed, think of x as fixed and

consider the incomplete integral
g(X) = IX t-1e x - t dt.
x
Since the integrand is positive g(X) increases with X, and, from the
fundamental theorem on monotone functions, limx->oo g(X) will exist i.e., the
infinite integral f(x) will converge, if we can show that g(X) is bounded as
X --+00. Clearly, for t 2! x.
and so integrating
g(X):::;;x-1e IX e- t dt = x-1eX[e- -e-X]:::;;x- 1
X
for X 2! x, i.e., g(X) is bounded.

It is permissible to integrate the integral for f(x) by parts repeatedly.
We get
[ x-t]'" + 1xr'" ex- t . [1]

~2 dt
f(x) = -et
=
1 1
(-l)n-l(n -1)'
=--2+ ... +
n
+(-l)nn!
x x
x
Consider the corresponding infinite series

1
2!
3!
--2+3-4+
.
x x x x
I'" t-n-1ex-tdt.
x
74
Chapter 7
This series is never convergent for its terms do not tend to zero. Nevertheless
it can be properly used to estimate [(x). To see this let us consider
:5n!x- n - 1
I"" ex-tdt
x
Suppose x is fixed. Then since
our estimate for Irn(x)1 decreases as n increases from 1 to [x], the integral
part of x, and then increases to 00. Thus for a fixed value of x, there is a limit
to the accuracy to which we can approximate [(x) by taking a number of
terms of the series - if this accuracy is good enough for our purposes, we
have a very convenient way of finding [(x) but if it is not, we have to
consider other means. Note that for fixed n, the estimate for rn(x) can be
made arbitrarily small by increasing x. Thus this method is likely to be
successful for large x.
We examine the case when x = 15 numerically. We note that in our case
but not in general the error estimate is just the first term omitted. It is
convenient to write a program to print out consecutive terms as well as the
partial sums. We find
= Ul +U2+"
Un
1
2
0.06666 66667
-0.00444 44444
0.0666666667
0.0622222223
14
15
16
17
-0.0000002133
0.0000001991
-0.0000001991
0.0000002124
0.06272 01779
0.06272 03770
0.06272 01779
0.06272 03903
sn
+Un
75
Asymptotic Series
We can therefore assert that

{(IS) =0.06272 03770
with an error estimate
0.0000001991.
The value correct to 8D is
f(15)
= 0.06272 028,
so that the actual error is about 0.00000 008, i.e. about half the estimate.
We shall now indicate an alternative approach to this problem. We
change the variable in the infinite integral from t to T where t = x + T. then
dt = dT and the limits become T = 0, T = 00. We find
f(x)
= roo e- dT.
T
.lo
X+T
From the sum formula for a geometric series we have
Multiplying across by e-T and integrating with respect to

we find
f (X)=--2++
x x
between 0 and 00
()n-l (n -I)!
-1
n
+rn
x
where
which is the result already obtained and which establishes the asymptotic
expansion for f(x).
Let us emphasize the difference between convergent series and asymptotic series such as that discussed. In the first case, for any x, limn-->oo rn(x) =
0; in the second, for any n, limx-->oo rn(x) = O.
The formal definition of an asymptotic series is due to Poincare:
F(x)~Ao+AlX-l+A2X-2+ ...
as
x---+oo
if
lim [F(x)-(A o+ . .. +Anx-n)]x n = 0 for
x-->oo
n = 0,1,2, ....
76
Chapter 7
It is clear from our error estimates that
J"
x
X-
I I I ! 2! 3!
t- dt----+---+ ...
X
x 2 x 3 X4
according to this definition.

Using one of the fundamental properties of the exponential function we
can prove that
x-+oo.
This means that the same asymptotic series can represent many functions.
It can be shown that asymptotic series can be manipulated fairly freely
e.g., it is clear that we can add and subtract asymptotic series. Formal
multiplication is generally legitimate and asymptotic series can usually be
integrated term by term. We shall not develop this theory but instead shall
discuss several more examples.
2. THE FRESNEL INIEGRALS
These are defined for x > 0 by
U(x) =
f"" cosJtt dt,
V(x) =
r"" sin t dt.
Jx Jt
This is not the standard notation but is convenient in this section: see
Problem 7.7. It is convenient to handle these together by writing w =
U + iV. We shall actually discuss
w(a)=
J"
x
eit
dt
to.
where a> O. Integration by parts gives

w(a) = ix-o.e ix - iaw(a + 1).
Repeating this process we find

( 1)
()=ieiX[I+~+a(a+l)
a(a+l) ... (a+n-l)]+
(.)2 + ... +
(.IX )n
rn+l
x o.
IX
IX
wa
where
r n+l = i- n- 1 a(a + 1) ... (a + n)w(a + n + 1).
Now, t being real, we have
Irn+1l:::5 a(a + 1) ... (a + n)
1"
x
dt
to.+n+1 =
a(a+l) ... (a+n-l)

xo.+n
,
Asymptotic Series
77
so that
Irn+ll ~ lu n +
11
where u,,+1 is the (n + l)-st term on the right of (1). Now

rn =U n +l+ rn+l
and so
Thus we justify the asymptotic series
ie iX
w(x)~-[X-iY]
Jx
where
and
1 1.3.5 1.3.5.7.9
Y=----+--(2X)3
2x
(2X)5
Thus
u ~ (l/YX)[- X sin x + Y cos x],

V ~ (l/YX)[X cos x + Y sin xl
[Notice that here we have rather extended our definition of the symbol
3.
~.]
THE ERROR FUNCTIONS
The error function and the complementary error function are defined
by
erf x =2-
.[;
I" e0
t2
dt
'
i=
2 " e- t2 dt,
erfc x = 1-erf x = .[;
the normalizing factor (21);) being chosen so that erf x~ 1, as x ~OO.

[Compare Problem 9.26.]
The asymptotic expansion
Chapter 7
78
can be obtained in several ways. The most straightforward is by integration

by parts beginning thus:
g(X)=2xeX2loo e- t2 dt=2XeX2loo (-2te- t2 )(-2t)-1 dt
= [2xe X2 . e-t2 . (-2t)-1];+2xe x2 foo e-t2(_2t 2)-1 dt

x
= 1 + [2xe X2 . e-t2 . (4t3)-1];+2xeX2l00 e- t2(3(4t 4 )-1) dt

2
_ 1 -1- +1.3
x2 foo
6)-1) d t.
(
-2)2-2xe
(
e -t ( 3.58t
2
2x
2x
x
The general result can be established formally by induction. It is clear that

the general term t,. in the series above does not tend to zero for any x, and
so the series cannot be convergent. To prove that it is asymptotic in the
sense explained above we have to show that x2nTn ~O as x~oo for any n
where
Tn = 2xe
X2 (-1t2n + 1)1) loo e- t2 dt

22n+1(n I)
t2n+2 .
We shall actually prove that ITnl :5ltn+11, i.e., the absolute value of the error
committed by truncating the series I t. at any stage is less than the absolute
value of the first term omitted.
To do this we change the variable in the integral in Tn from t to y where
t=.J y +x 2
The limits of integration in y are 0 and
00.
Hence
x(-1t- 12n+1)1) roo

e-Ydy
Tn =
n !22n +1
1 (y + x 2)(2n+3)/2'
Since y is positive the denominator of the integrand has a lower bound X2n + 3
and
so that certainly
x~oo.
To illustrate the limited accuracy possible with asymptotic expansions

we note that for x=4 since ITnl~ITn+11 according as 2n+1~2x2=32, the
Asymptotic Series
79
best result is obtained with n = 15 for which

(31)!
'15
1 .
-7
=(15)! . 295:::;: 1. x 10
so that we cannot obtain a value for g(4) which can be guaranteed to be

good to within 10-7
A second way of obtaining the asymptotic expansion of g(x) is to
change the variable from t to s, where s = t 2 getting
and then integrating by parts.

A third method consists in making the change of variable used in the
estimation phase of our first method. We find
r'"
xe- Y dy
g(x)=.10 (Y+X 2?/2;
we now use the binomial expansion of (y + X 2)-1I2 and integrate term-byterm. See Problem 7.9.
4.
THE GAMMA FUNCITON
One of the most useful asymptotic series is due to Stirling:

f(x) =
t x - 1 e- t dt - e- xx x -l/2(2'7T)1I2 [ 1 + l~x + 28~X2 - ...
J.
This will often be used in the form
n!-(;)" ~2'7Tn
and we shall establish this in Chapter 9.
5.
THE COMPUTATION OF FUNCITONS
It is appropriate to discuss this problem briefly here. It is clear from our

examples that asymptotic expansions are often a convenient way of evaluating a function for "large" arguments. It is also clear that if the function can
be represented as a power series, then that may be an appropriate way of
evaluating it for "small" arguments. In many cases, neither representation
will be convenient for the "intermediate" range. Among the possibilities are
differential equations or continued fractions.
80
Chapter 7
We shall discuss here, in the text, our basic example, up to the

derivation of the differential equation - the solution of differential equations
is discussed in Chapter 10. Further examples occur in our Problems.
Observe that while the derivation of the asymptotic series is comparatively
simple, that for the power series is rather complicated.
From the usual rule for differentiating an integral with respect to a limit
we find
i.e.,
y'=y-x- I
so that y = f(x) satisfies the differential equation

xy' - xy + 1 =
o.
In order to get a "power" series for f(x) we need the following lemma.
Lemma.
'Y =
(1-e- t )t- 1 dt-l>O t-Ie- t dt.
Proof. Both integrals on the right hand side are convergent. We use the
definition
'Y = lim
and the theorem
(1 +-i+ .. . +';-IOg n)
Iun (1 --;;x)n =e
-x
By changing the variable from t to 1- T we find
e1-(1-tt
r1-Tn
1
1
_"":"""t-:.-. dt = .10 -1---T dT=1+2:+ .. . +-;;.
Jo
Hence
. {L
'Y =Iun
in
1- (1- tt dt- -dt}

tIt
81
Asymptotic Series
subject to justification of passages to the limit.

We now rewrite the integral for El(X):::: e-Xf(x) in the form
where all the integrals are meaningful. Our lemma shows that the first two
integrals give -y, the third gives -log x and the last can be integrated
term by term after expanding the exponential. We get
00
El(X)::::-y-logX+k~l
(_1)k-l Xk
k(k!)
This function is discussed in the NBS Handbook, p. 228.
6. THE TABULATION OF FUNCfIONS

It is convenient to discuss this matter here. In many cases the most
efficient way to find f(x), where the "special function" f and the argument x
are given is to "look it up in a book". The value of f(x) may be needed by a
customer or it may be needed to check a program. How do we know to
which book to refer? We assume the reader to be familiar with the NBS
Handbook, which is his first resource. If f is not tabulated in it, or if the
precision is inadequate, the next step is to look at
An index of mathematical tables, 2 vols., 1962, A. Fletcher, J. C. P.
Miller, L. Rosenhead and L. J. Comrie, Blackwell, Oxford, England and
Addison-Wesley, Reading, Mass. U.S.A.
These volumes, in addition to listing the contents of tables (from 1596
onwards), indicate what aids to interpolation are provided, give recommended tables in BOLD TYPE, list errors and contain much of human
interest about the table-makers. We shall refer to this as FMRC.
For information on more recent tables reference must be made to the
periodical Math. Compo and to Math. Reviews.
Tables of functions arising in mathematical statistics are covered in
Guide to tables in mathematical statistics, 1962, J. A. Greenwood and
H. O. Hartley, Princeton, N.J.
Chapter 7
82
There is also a Russian index

Spravochnik po mathematiceskim tablicam, 1956, A. V. Lebedevand
R. M. Federova, Moscow. English trans. 1960, Oxford.
Chapter 7, Problems
=.r:
7.1. Using the asymptotic series developed for f(x)

t-1e x - t dt obtain the
values at x = 20, 10, 5 as accurately as you can, estimating the errors in your
results. By reference to tables find the actual values of f(20), f(1O) and f(5)
and compare the actual errors with the estimates.
7.2. Repeat Problem 7.1 for the Fresnel integrals, i.e., evaluate U(x), V(x)
for x = 5(5)20.
7.3. Using the asymptotic series developed for g(x) = .firxe X2 erfc x evaluate
g(x) as accurately as you can for x = 3(1)6, estimating the errors in your
result.
Obtain the correct values of erfc x and g(x) from tables and give the
actual errors.
7.4. (E. T. Goodwin-J. Staton) Obtain an asymptotic expansion for
r"" e-
F(x) =
du,
u+x
u2
as x~oo. How many terms of the series are needed to give F(x) correct to
4D for x;;:: 10? Compute, in particular, F(10).
7.5. Show that the function F(x) of Problem 7.4 satisfies the differential
equation
,
x>O.
Y +2xy =V7r--,
7.6. Obtain the following representation for the function F(x) of Problem
7.4, in terms of the auxiliary functions e -x 2 and log x (which are well
tabulated), and power series:
[
F(x)=e- x2 .fir
t
""
X 2n + 1
n!(2n+1)
Compute, in particular, F(1) to 7D.
""
x2n
~ n!(2n) -logx-h .
Asymptotic Series
83
7.7. (L. Fox-D. H. Sadler) Discuss computationally convenient representations of the function
{(x)
= f= sin t dt.
o
rx+i
In particular: express {(x) in terms of the Fresnel integrals; obtain a

power series for {(x) and derive an asymptotic expression for {(x) and
indicate for what range of x these are convenient for tabulation to 6D;
finally, find a differential equation satisfied by {(x).
7.8. Show that as x ~oo
7.9. Justify the third method suggested for the determination of the asymptotic expansion for
g(x) = 2xe x2
e- t2 dt.
CHAPTER 8
Interpolation
Interpolation originally meant reading between the lines of a table. It is

a basic problem in numerical mathematics, and a solution can be formulated
as follows:
If distinct abscissas xo, x1> . , x,. and corresponding values fo, f1> ... , fn
are given, find a function f(x) such that f(X;)=h and use f(p) as the value
corresponding to the abscissa p.
1.
LAGRANGIAN INTERPOLATION
The simplest case is when n = 1. Linear interpolation then assigns to

f(p) the valQe
We assert that f(x) is the unique polynomial of degree 1 agreeing with the
data at xo, Xl' Compare Fig. 8.1, where p is taken between xo, Xl only for
convenience.
This is the simplest example of polynomial interpolation - interpolation
by rational functions (the ratio of two polynomials) or by splines (functions
which are "piece-wise" polynomials) or by trigonometric or exponential
polynomials is often considered, but cannot be discussed here. [See, however, Problem 8.17]
In order to discuss polynomial interpolation in general we need the
"Fundamental Theorem of Algebra" which is originally due to Gauss.
Theorem 1. Every (polynomial) equation
has a root.
Many proofs of this are available. We note that in order that it be true
we must admit complex roots even though we restrict the coefficients to be
real- the equation Z2 + 1 = 0 shows this. However it is remarkable that if we
admit complex numbers as solutions, then the result is true even if we allow
complex numbers as coefficients.
Interpolation
85
"
fo~-+-----I
Fig. 8.1
If f(a)
= 0 then
f(z)
= f(z)- f(a) = (z -
a)[aoz n- l + a~zn-2+ . .. + a~-l]
because
It follows from this remark (sometimes called the Remainder Theorem) and
the Fundamental Theorem of Algebra that we can factorize any f(z) in the
form
(2)
where the ai are not necessarily distinct. From this fact we can conclude that
if a polynomial vanishes for more than n distinct values of x it vanishes
identically. Suppose that f(z) vanishes for z = al>' .. , lXn and for z = ao. The
relation (2) must hold. If we put z = ao in this we get
0= a O(aO-al)(aO-a2)' .. (ao-lXn).
None of the factors ao - a; can vanish. Hence a o = O.
Hence we have a polynomial of degree n -1,
fl(Z) = alz n - 1+ ... + ~
which certainly vanishes for more than n - 1 distinct values of z. The same
argument shows that a l = O.
Proceeding we see that f must be identically zero.
It is a corollary to this that we need to establish uniqueness of
polynomial interpolation. In fact, if two polynomials of degree n at most
agree for more than n distinct values of the variable, they must coincideindeed, their difference, of degree n at most, vanishes for more than n
distinct values of the variable.
86
Chapter 8
Suppose that p(x) is a polynomial of degree n at most. Then it has n + 1

unknown coefficients and we may expect it to be determined by n + 1 pieces
of information e.g., the data of the problem formulated above. It is possible
to write down a polynomial p(x) which interpolates f at the nodes x". To do
this we consider
Ii (x) =
(3)
n (x - Xj)/(X; - Xj).
n
i=O
iji
It is clear that each
~(x)
is a polynomial of degree n and that

j =/= i,
Hence
(4)
L(f, x) =
L t~(x)
n
i=O
is a polynomial of degree n at most and its value at X; is t, i = 0, 1, ... , n.

Thus L(f, x) is a solution to our interpolation problem. It is the unique
solution for any other solution Ll (f, x) would agree with the data and
therefore with L(f, x) at n + 1 distinct points - in view of the preceding
remark, Ll == L.
This L(f, x) is called the Lagrangian interpolation polynomial.
2. AITKEN's ALGORITHM
A method for establishing the existence of such a polynomial is outlined

in Problem 8.13. We shall now discuss an elegant algorithm due to A. C.
Aitken for the evaluation of Ln at a particular point x. It reduces the general
(n + i)-point interpolation to a succession of ~n(n + 1) linear interpolations.
An important advantage of this scheme is that it does not require the
use of auxiliary interpolation tables. Cf. Problem 9.7. We shall describe the
4-point case for simplicity; it will be convenient to change 0Jlf notation.
Given f(a)=A, f(b)=B, f(c)=C, f(d)=D, we show how to find f(p).
Interpolate linearly between (a, A), (b, B) to find (p, B 1 ); then interpolate
linearly between (a, A), (c, C) to find (p, C1) and then between (a, A),
(d, D) to find (p, Dl). The next stage is to interpolate linearly between
(b, B 1), (c, C 1) to find (p, C2 ) and then between (b, B 1), (d, D 1) to find
(p, D~. Finally, interpolate linearly between (c, C2 ) and (d, D 2 ) to find
(p,P).
The scheme is illust:l\ated graphically in Fig. 8.2. We note that there is

no assumption that the a, b, c, d are equally spaced, and p can be anywhere,
and so the method can be applied to the case of inverse interpolation.
87
Interpolation
y
,,y-t (x)
,,
,,
a
Fig. 8.2
We give, without comment, two examples which show how the scheme
can be carried out and how labor can be saved by dropping common initial
figures. Many extensions of the method have been given by Aitken and
Neville.
Given /(1) = 1, /(2) = 125, /(3) = 729, /(4) = 2197, find /(2.5).
l=a
2=b
3=c
4=d
l=A
125=B
729=C
2197=D
187=B1
547=C1 367=C2
1099=D1 415=D2
343=P.
Here /(x)=(4x-3? and the interpolation is exact, as it should be.

Given /(0) = 47.434165, /(1) = 47.539457., /(2) = 47.64451 7, /(3) =
47.749346, find f(1.4321).
o
1
2
3
47.434165
539437
644517
749346
0.584954
682
517
837
60
-1.4321
-0.4311
+0.5689
24 + 1.5689
We find f(1.4321) = 47.584824, which agrees with the fact that

f(x) = .J225 0 + lOx.
We shall now show generally that the Aitken algorithm leads to the
Lagrangian interpolant. We follow a proof of Feller. We want to evaluate
f(p), where f is a polynomial of degree n determined by its values at the
Chapter 8
88
distinct points Xo,
xl>
,x". Consider
fl)(X) =
p]
[f(xo) Xo et f(x) x - p ,
X-Xo
x-xo
We observe that fl)(x) is a polynomial of degree n -1 and that fl)(p) = f(p).

Hence our problem is equivalent to that of evaluating fl)(p), where fl) is
determined by its values at Xl> X2' ... , x". Repetition of this process according to the scheme
Xo
f(xo)
Xl
f(Xl)
X2
f(x 2) fl)(X2)
x"
f(x,,) (I)(x,,) (2)(x,,)
fl)(Xl)
f 2)(X2)
... fn)(x,,) = f(p)
leads to the determination of f(p).

For clarity we write out explicitly the general term
(5)
where j=i+1, i+2, ... , n.

3.
INvERSE INTERPOLATION
We have noted that the Aitken process does not require the Xi to be
equally spaced. It can be used to solve equations: for to solve
f(x)=O
is just the evaluation of
f-l(O).
We give an example. If f(O) = -342, f(1) = -218, f(2) = 386, f(3) =

1854 find the zero of f between 1, 2. We use the Aitken method, recording
all the numbers used:
-342
-218
386
1854
0
1 2.7581
2 0.9396 2.1018
3 0.4672 0.5171
124
728
604
1.9926 2196 2072
1468
89
Interpolation
This is a very bad result. Indeed f(x) = (4x+ 1)3_343 and the zero is 1.5 not
1.9926.
Inverse interpolation - except linear inverse interpolation, which is the
same as direct - is tricky and great care is required.
4.
ERRORS IN INfERPOLATION
We have not yet discussed the error in interpolation:

(6)
e(x)= \f(x)-Jo 4(X)f(XJ\.
This discussion cannot begin unless we know something qualitatively about

the behavior of f(x) between the nodes X;. A natural assumption is to
assume that the X; lie in an interval [a, b] and that f(x) is continuously
differentiable k times in [a, b] and, say,
Ifk)(x)I$;Mk
in [a, b].
It is clear that if f is a polynomial of degree n or less interpolation is exact:

f(x)=L,,(f, x). This suggests that we might expect error estimates of the
form
(7)
for some constant K. This is so and an actual result in the case of n = 1 is

(8)
this, and a little more, can be obtained by a slightly sophisticated application

of the Mean Value Theorem.
We write b - a = h, x = a + ph, 0 $; p $; 1. Then Taylor's Theorem gives
p 2 h2
f(a+ph) = f(a)+phf'(a)+---z! f'(c),
c=c(p),
f(b) = f(a)+hf'(a)- ~; f'(d),
a$;c$;b,
a$;d$;b.
We have, for the error in linear interpolation for f(x) between f(a) and f(b):
f(a + ph)-[f(a) + p{f(b) - f(a)}] =!h 2{p2f'(C)- pf'(d)}
which can be put in the form (7).

We can get the same result with c
Consider
=d =, by the following trick.
F(p) = f(a + ph)-[f(a)+p{f(b)- f(a)}]- Kp(p -1)
Chapter 8
90
and choose a Po, 0 < Po < 1 and then choose K so that F(po} = O. Then since
F(p} also vanishes for p = 0, p = 1 we conclude that F' (p) vanishes twice in
[0,1] and
F'(p} == h 2f'(a
+ ph} - 2K
must vanish once in [0,1], say for p = (), so that

K
= ih 2f'(a + ()h).
Thus we have, since F(po} = 0,

f(a + Poh} = f(a}+ Po{f(b} - f(a)}+ih2(p~- po}f'(a + ()h).
Clearly () depends on Po. We can, however, drop the subscript 0 and

conclude that for each p, 0 $ P $1, there is a {= {(p), a $ {$ b such that
f(a + ph} = f(a} + p{f(b}- f(a)}+ih 2(p2- p}f'({}.
Thus the error in linear interpolation is

ih 2(p2- p}f'({}
and this is bounded by

since Ip(p -1)1 . , x.., p then
(9)
where
5.
= ~(p) E J.
HERMITE INIERPOLATION
There are clearly many possible extensions of the Lagrange problem.

For instance we may ask whether given two sets of n + 1 constants not
necessarily distinct
and one set of distinct abscissas
there is a polynomial H of degree 2n + 1 such that
H(x;) = /;,
H'(x;) = f{,
i = 0, 1, ... , n.
91
Interpolation
This is the problem of Hermite interpolation. The answer is that it is

solvable and uniquely. In fact with the notation in use
(10)
H(x) =
L{4(x)y{1-21r(xJ(x - X;)}{; + L{4(x)y(x - x;)/:.
It is possible to give an error estimate similar to (9). H {;
n= f'(X;) where f
(11)
2n + 2)(X)
= f(x;)
and
exists then
_
f2n+2)(~)
f(x)-H(x)- (2n+2)!
(x-X;)
where ~ is in the interval including xo, Xl, ... , Xn, x. This can be done by
repeated applications of Rolle's Theorem to
(12)
E(z) = f(z) - H(z) -[f(x)- H(x)]
n [(z - X;)2/(X - X;)],
where x is distinct from the x;.

An extreme case is when we ask for a polynomial T(x) of degree n for
which T(r)(xo) = f r) for r = 0, 1, 2, ... , n. The solution is
T(x) =
L fr}(x-xoYlr!.
r=O
We note that the process of counting the coefficients which works in the
Lagrangian, Hermitian and Taylor case does not always work. For instance
it is not possible to find a quadratic q(x) for which, say, q(1) = 0, q'(O) = 1.
The so-called Hermite-Birkhoff problem arises in this area.
6.
INTERPOLATION IN PRACTICE; ERRORS
We have not been very explicit about the degree of interpolation

appropriate in a table.
In the first place, let us consider the error in the interpolant L.. (f, x) =
Lii{;4(x) due to rounding errors in the data. Assuming the data rounded
correctly so that the individual errors are at most half a unit in the last place
we need to know an upper bound A.. for A.. (x) = Lii \4(x)\. This number,
known as the Lebesgue constant for the particular distribution of abscissas
Xo, Xl>' .. , Xn, can be determined (or estimated) once for all. In the simplest
case n = 1, xo=O, Xl = 1 we have
'0(x)=1-x,
When n = 2 with Xo = -1, Xl = 0, X2 = 1 we find A2 = 1.25. Indeed

'l(x) =!x(x 1)
92
Chapter 8
and if 0:5x:51 we have

A2(X) =!x(l- x)+(1-x 2)+!x(1 + x) = 1+ x- x 2
which has as its maximum 1.25 at x =!. [The behavior in -1:5 x :5 0 is

symmetrical.] Generally it can be shown that A,. = O(Iog n).
The above analysis assumes that we use the exact values of the
Lagrangian coefficients, which will in general not be the case. It is also a
worst case analysis; if we are content with statistical estimates much smaller
errors can be given.
It is also clear that this argument can be adapted to deal with the case
when the data are subject to error.
We have also got to consider the error in our interpolant, in the context
of theoretical arithmetic. On an assumption that f(X)E cn+l[a, b] we have
seen that
If(x) - L,. (f, x)1 :5 CnM n+1
where Mn+l is a bound of f n+1)(x) in [a, b] and Cn a constant. Granted the

assumption that fE cn+l[a, b], we have to estimate M n+1. It will be shown
in Chapter 10 that
an+1f(x) = hn+1fh+1)(~)
and so an estimate for f n+1) is given by h-n- 1an+1f(x) which can be
obtained from tabulated values of f(x) only.
There remains the choice of n. It is easy to see how errors in a table
propagate and to observe that the (n + l)st differences are subject to a
rounding error of at most 2n. Variations up to this amount can be regarded
as noise. Probabilistic considerations reduce the maxima considerably.
It is reasonable to arrange to have the rounding error and interpolation
error about equal.
Practically one often prints out the Aitken triangle and observes when
the interpolants become stationary. However in this area, experience is
necessary.
With regard to published tables of functions, we note that often
information about interpolation errors is included in the table, or the
introductory material in the volume. E.g., most tables in the NBS Handbook
have a symbol such as
at the bottom of a column. This indicates that linear interpolation gives

an error not exceeding 4 x 10-5 in absolute value and the full accuracy is
Interpolation
93
obtained by using 7 point interpolation. Let us work an example from NBS

Handbook, p. 400.
J 3 (x)
10.0 0.05858
10.2 0.10400
lOA
0.14497
10.6 0.17992
10.8 0.20768
Linear interpolation gives, for instance,

10.1
0.08129
10.7
0.19380.
Values to 8D, from the British Association Tables, Vol. 10, p. 40, are
10.1
0.08167273
10.7
0.19475993
indicating errors of 38 x 10-5 and 96 x 10-5 within the range announced,

1 x 10-3
Using 2-, 3-, 4-, and 5-point interpolation for x = 10.5 gives
0.16244, 0.16320, 0.16327, 0.16328
as compared with the SD value of 0.16328016.

Chapter 8, Problems
8.1. Read again material about the Fundamental Theorem of Algebra.

S.2. Write
l(x)=I1j~o(x-Xj).
Show that
l(x)
i;(x) = (x - xJl'(xJ .
94
Chapter 8
8.3. Write a program which when given a, b, f(a), feb), p as input gives as
output the linear interpolant of f at p.
8.4. Write a program to implement the Aitken algorithm, say for up to 11
points. Arrange to have it give as output the appropriate interpolant only, or
the whole triangular array, according to a choice of an entry or a parameter.
Alternatively, thoroughly check the Aitken subroutine available in your
system.
8.5. No specific problems on direct and inverse interpolation will be set.
The reader should set up his own problems and at the same time familiarize
himself with the special functions by using the NBS Handbook as follows:
Take a function tabulated, say at interval 0.001. Use values at interval
0.05 as data and interpolate using various values of n and varying positions
of x relative to Xo, Xl> . ,x,.. Compare the results obtained with the
tabulated value.
Take values of x near where a function, e.g., Jr(x), changes sign and
find its zero by inverse interpolation. Check the result from the tables, if the
zeros are tabulated.
8.6. Consider the interpolation of f(x) = X4 by a cubic, Lix), the chosen
abscissas being -1, 0, 1, 2. In particular discuss the behavior of the error
E(x) = X4- L 3 (x) in the range [-1,2].
8.7. Write down the fundamental Lagrangian polynomials li(x) in the case
of four-point interpolation at the nodes -1, 0, 1, 2. Tabulate these for
x = -1(0.1)2 and show how to use these tables to facilitate interpolation.
8.8. What is a bound for the error incurred in linear interpolation in a table
of sin x, at interval 0.2?
8.9. Evaluate
L xl~ (0)
n
Sj =
for j
0, 1, 2, ... ,n, n + 1.
i=O
Check your results in a special case, e.g.,
3,
Xo
1,
Xl =
2,
X2 =
3,
X3
4.
8.10. Write down the Lagrangian polynomial Ln(x) of degree n which

interpolates Ixl at n + 1 equally spaced points x~n) in (-1,1):
2k
n
xln ) = -1 +-, k
0, 1, ... , n.
Evaluate Ln@ for n = 1(1)12 using the computer. Evaluate 4(0), 4(1).
Interpolation
8.11. Find the monic polynomial f(x)

gree n which is such that
95
= xn + a 1 Xn~ 1 + ... + an~l x + ~ of de-
IL = max (\f(O)\,
\f(l)J, ... , \f(n)J)
is least.
8.12. Establish the relations

H(x) =
L {Ii (xW{l- 21(x;)(x - X;)}fi + L {l;(xW(x - x;)t:
and
t2n+2)(~)
f(x)-H(x) = (2n+2)!
(x-xY
where the notation is that of the text.

8.13. Establish the existence of a unique polynomial of interpolation by
showing that, provided xo, Xl>' .. ,Xn are distinct, the Vandermonde determinant
1 Xo
x~
1
Vn
V(xo,
Xl> ... ,
Xl
x~
xn) = det
#0
1 Xn
xnn
so that the system of equations

i = 0,1,2, ... , n.
has a unique solution.

Deduce that the 10, 11, ... ,In corresponding to any set of real and
distinct xo, Xl> ... , Xn form a basis for the vector space Vn of all polynomials
of degree at most n, with real coefficients.
8.14. Interpolate the following table in the neighborhood of the zero of

f(x), first at interval 0.1, then at interval 0.01. Is linear interpolation
justifiable in the last table? Find an estimate for the zero of f(x).
X
0
1
2
3
4
5
f(x)
-855
-513
-172
+167
+507
+846
17443
73755
86901
39929
03555
00809
96
Chapter 8
8.15. (a) If f(x) is a real function of the real variable x, O~x ~ 1, and
f(O) = 0, f(l) = 1, what can you say about fm?
(b) If, in addition, f is positive and convex, what can you say about
f(!)?
(c) If, in addition, f is a quadratic in x, what can you say about f(!)?
(d) If, alternatively, f'(x) exists in [0, 1] and is bounded by 1 in
absolute value, what can you say about fm?
(e) If, alternatively, all the derivatives of f at 0 exist, and are bounded
by 1 in absolute value and if the Maclaurin series for f converges to f in
[0, 1], what can you say about fm?
8.16. Discuss the existence of Hermite interpolants, e.g., in the two-point
case, along the lines of Problem 8.13.
8.17. Establish the existence of a (cubic) spline in the two panel case.
Specifically, show that there is an f~ such that, no matter what fo, f!> f'cr.!>
are, the Hermite interpolants in [x-I> xo] and [xo, Xl] fit smoothly together at
Xo i.e., that the second derivatives there coincide.
Work out the case when
Xl
=1,
xo=O,
fo= 1,
fl =0,
f'cr.l =0.
Generalize.
8.18. Show how to determine the turning point i and the extreme value
1= f(i) of a function f(x) given its values at three points near i, by
assuming that f(x) = a +bx +cx 2
8.19. Show that if fr = f(r), r = 0,1,2, ... , n and if A has its usual meaning
then
Compute the differences of the function f for which f(O) = 789,

f(1) = 1356, f(2) = 2268, f(3) = 3648, {(4) = 5819, f(5) = 9304. Show that f
coincides with a quartic at x = 0(1)5. Experiment with interpolation of
various orders, to evaluate f(2.5).
CHAPTER 9
Quadrature
We are now concerned with the approximate evaluation of definite

integrals, i.e., areas under curves,
1=
f(x)dx.
We shall deal sometimes with weighted integrals

1=
f(x)w(x) dx
where the weight function w(x) is usually nonnegative and continuous in

[a, b].
We use a very simple principle:

Q, an approximate value of I, is given by the integral of an approximation to f(x).
It is natural to approximate f(x) by a polynomial, since these are easy to
integrate. If we take
f(x)=L..(f, x)= L ~(x)f(x;)
(1)
then we have the so-called Lagrangian quadrature:

1= Q =
(2)
~(x)f(x;) dx = Lf(x;) ~(x) dx = L wJ(X;)
where the coefficients Wi are independent of f, only depending on the nodes

xo, ... , x.. and a, b. Clearly this is exact if f(x) is a polynomial of degree
::5;n.
1.
TRAPEZOIDAL QUADRATURE
We consider first the very simple trapezoidal case corresponding to

linear interpolation when we take n = 1 and a = xo, Xl = b and assume
f E C 2 [ a, b]. The area under the curve is approximated by that of the
trapezium which is
(3)
(b - a) . Bf(a) +!f(b)].
98
Chapter 9
If 1f"(x)1 :::;M2 in [a, b], using the result about the error in linear interpolation, we can conclude that
With a little more care we can improve this to

1- 0
(4)
= --b(b - a)3f"(c),
for some C E [a, b]. [Compare Problem 9.3.]

Supposing that this accuracy is not enough, what can we do? We could
approximate f(x) by polynomials of higher degree or, and this is the
approach we want to discuss, we can divide the area I into panels and apply
the trapezoidal rule (3) to each one. Let us take b = a + nh and take n
panels each of width h. Then we have
(5)
II-OI:::;n '112 (b:arM2=-b(b-a)M2' h 2
so that we gain a factor of h 2 in our error estimate at the expense of about

1n times as much work: we have to evaluate f at (n + 1) points instead of at
2 points.
With a little more trouble (Problem 9.3) we can find
1- Q = -lib - a)f"(c)h 2
(6)
for some C E [a, b].

The corresponding result for an n + I-point Lagrangian quadrature is
got by integrating the relation (Chapter 8),
f(x)-L,.(f,x)=
r+1)(C(x)) n
(n+l)! n(x-xJ
to get
II-OI:::;max Itn+ 1)(x)1
(7)
(n+l)!
I Ii Ix-~I
b
dx.
i=O
REMARK: Note first that (7) implies that any n + I-point quadrature of a
polynomial of degree n is exact. Note also that this estimate is a very crude
one: we cannot apply the Mean Value Theorem of the Integral Calculus
<f>(x)I/I(x) dx = <f>(c)
I/I(x) dx,
unless we know that 1/1 is of one sign in [a, b]. [Contrast this with Problem
9.12]'
99
Quadrature
Let US note what (7) gives in the case n = 3,

x3=2. We find (for details see Problem 9.1).
Xo =
1,
Xl
= 0,
X2 =
-1,
o =U( -1)+U(0)+U(1)+U(2),
with an error estimate
compared with the result

1-0= -ior>(c),
-1~c~2,
which is found by a more complicated argument.

We observe that in the case f(x)=x 4 , when r(x) =24, we have
while 1=[x 5 /3II=so that 1-0=-9/10 which is exactly (-3/80)x24.

So the error estimate of Problem 9.1 is a best possible one.
We also note that (7) is not necessarily even optimal in form. In fact,
when we have an odd number of points, 2n + 1, say, we can get an
(2n + 2)nd derivative.
Let us examine the simplest case when the nodes are -1,0,1. Then,
because S~~ x 3 dx = 0, the integral of a cubic coincides with that of a
quadratic. If Q2(X) is the quadratic interpolating f(x) then
o =1[f( -1)+4f(0)+f(1)]=
fl
[Q2(X)+cx(x 2-1)] dx
and by choice of c we can improve our error estimate from
M3l1 Ix(x 2 -1) I dx -12

_M3
3!
to M4
90 .
For details see Problem 9.4.

To compare the 3-point and 4-point quadrature we must refer them to
the same interval, say [a, b]. The errors are
The numerical factors are -1/6480 and -1/2880. To discuss costeffectiveness we must remember that we have to make an extra function
evaluation in the first case; this discussion is included in Problem 9.20.
100
2.
Chapter 9
RICHARDSON - EXTRAPOLATION - ROMBERG QUADRATURE
H this sort of error is still not satisfactory what can we do? There is
available a very elegant technique due to W. Romberg, which is based on
ideas popularized by L. F. Richardson. We shall begin by discussing the
Richardson device, which can also be applied with profit in many other areas
of numerical analysis. [See Problem 9.22 and Chapter 10, p. 126.]
Suppose we are trying to evaluate a quantity cp, or a function cp(x),
which is the solution to a continuous problem e.g., the fundamental frequency of a vibrating string (length I, density p, under tension T) or the
circumference of a circle of radius unity. We approximate the continuous
problem by a discrete one, characterized by a mesh size h. In the first case
we replace the continuous string by beads of mass ~m, m, ... , m, ~m, where
m = pl/(n + 1), placed at distance h = I/(n + 1), the beads of mass ~m being at
the endpoints. In the second we replace the circle by a polygon with n = h- 1
sides. H this discretization is done reasonably we often find that if cp(x, h) is
the solution to the discrete problem then
(8)
for some cpix), where R2 is small. H we neglect R2 and observe or calculate
cp(x, h) for two values of h, hb h2 say, we have, approximately,
cp(x, hI) = cp(x) + hiCP2(X),

cp(x, h2) = cp(x) + h~CP2(X).
From these two equations we can eliminate CP2(X) to get
( ) _ Mcp(x, hI) - hicp(x, h2)

cp x h~-hi
'
expressing cp(x) as a mean of the two observed values. This process is called
h 2 -extrapolation or deferred approach to the limit. Compare Fig. 9.1.
Fig. 9.1
101
Quadrature
Let US carry this out in the two cases mentioned. In each case we shall
be trying to find a constant, not a function.
We know from the theory of vibrations that the frequency in the
discrete case, with n interior beads is u(h) where
u(h) =
2~': sin ;~
while that in the continuous case is
u=T~~'
The effect of extrapolation can be best seen if we consider the following
table
lh- l =(n+1)
2
u(h)/u
0.9003
0.9549 0.9745 0.9836

10
20
40
lh- l =(n+1)
00
1
u(h)/u
0.9959 0.9990 0.9997
If we extrapolate from 2 and 4 we get 0.9992 and from 5 and 10 we get
1.000.
We know that the perimeter of a square inscribed in the unit circle is
4J2 while that of a regular hexagon is 6. Extrapolating we get as an estimate
for 7T
~(4J2)-k(6) 108-32J2 31373
2(~-l~)
20
..
In each of these cases the extrapolation is legitimate because the
existence of a relation of the form
cp(h) = cp + h 2cp2 + R2
follows from elementary trigonometry. [Cf. Problem 9.7J

It is clear that if instead of the expansion (8) we had
(9)
we could eliminate the h4 term. In general, if we have an expansion of the

form,
cp(x + h) = cp(x) + h 2cp2(X) + h 4cp4(X) + h 6 cp6(X) + ... ,
we can eliminate successively the terms in h 2, h4, h 6, .. ..

With this device in mind we return to the quadrature problem. We
know from the theory of integration that if ToO), Tol), To2 ),... denote the
102
Chapter 9
trapezoidal approximations to I with 1,2,4, ... panels then
We choose this successive halving of the panels so that we can always make
use of functional values already computed. [A more detailed investigation
shows that, taking more things into consideration, this scheme is not the
most economical one.] We now refer back to the result (6) and see that
h 2 -extrapolation is indicated. This means we compute a new column T)k) by
'T't'k) _
Ii -
4n + )-n
k
0
k)
0
and we can expect this to be more rapidly convergent than the first column.
This is indeed the case for we have really obtained the values given by
Simpson's Rule for which
for some CE [a, b]. (See Problem 9.4.)

The next thing to do is an h 4 -extrapolation getting a new column
which we may expect to be more rapidly convergent than the second

column.
Then we do an h 6 -extrapolation to get a fourth column
k )-
64T2k + 1) - T2k )
63
And so on, until satisfactory convergence of the columns, or the

diagonal is obtained. We give, following Bauer, Rutishauser and Stiefel, the
results for the case of I = Ii dx/x = In 2 = 0.69314 7180.
0.750000000
0.708333333
0.697023809
0.694121851
0.693391202
0.694444444
0.693253967 0.693 174603
0.693154532 0.693147901 0.693147479
0.693 147652 0.693 147 193 0.693 147 182 0.693 147 181
Quadrature
103
The corresponding errors are, in units of the ninth decimal,

56852820
15186153 1297264
3876629 106787 27423
7352
721 299
974671
2 1
244022
472
13
The reduction by factors of approximately 4, 16, 64 in the first, second, third
columns is evident.
As we have pointed out before, the main labor in numerical quadrature
is the computation of the values of the integrand. In the preparation of the
first column of the above table we note that all values obtained are used in
succeeding entries. The determination of the entries in the second and later
columns is done by the use of the generic recurrence relation at a negligible
additional expense.
3.
GAUSSIAN QuADRATURE
There is another approach to better quadrature formulas, which goes

back to Gauss. We noted at the beginning of this chapter that given any set
of nodes Xo, Xl> ,x.. we could obtain a quadrature formula by integrating
the Lagrangian polynomial which interpolated the integrand at the X;. Is
there any specially desirable way of choosing the X;?
To answer this we need some of the theory of "orthogonal polynomials". If w(x) is a positive weight function in an interval [a, b], which may
be finite or infinite, we can, by the Gram-Schmidt process, orthogonalize the
sequence
1,
X,
x2,
with respect to the inner product

(f, g) =
f(x)g(x)w(x) dx.
[We must assume that all the moments ILn = S! xnw(x) dx are finite.] That is,
we can find a sequence of polynomials 'lTn(x) of exact degree n, such that
'lTn(x)'lTm(x)w(x)dx=O
if
m~n.
In the case
a= -1,
b=1,
w(x)=1
Chapter 9
104
we get the Legendre polynomials Pn(x). In the case

b=oo,
a=O,
w(x)=e-
we get the Laguerre polynomials L,.(x). In the case

a=-1,
b=1,
we get the Chebyshev polynomials,
Tn (x) = COS (n arccos x),

as can be verified by elementary trigonometry.
It can be proved in general that the 1Tn(x) have all their zeros distinct,
real and inside [a, b]-this is obvious in the Chebyshev case for
T (x) = 0
if x = cos (2r -1)7T

2n '
r= 1, 2, ... , n.
We have noted that an (n + 1) point Lagrangian quadrature is exact if

the integrand is a polynomial of degree s n. This is also true in the case of a
weighted quadrature (where w(x)~O):
1=
f(x)w(x) dx = Q =
=
L,. (f, x) . w(x) dx
ita f(x;)
~(x)w(x) dx
Consider now weighted quadratures with the nodes chosen to be the zeros
of the polynomial1Tn(x) orthogonal with respect to w(x) in [a, b]. Let f be a
polynomial of degree 2n -1 at most and write
f(x) = q(X)7Tn (x) + r(x)
(10)
r
r
where q, r are polynomials of degree n -1 at most. ,Then

1=
f(x)w(x) dx =
q(X)7Tn (X)W(x) dx +
r(x)w(x) dx.
By orthogonality, the first of the two integrals on the right is zero and so
1=
r(x)w(x) dx.
Since r(x) is of degree n -1 at most the weighted quadrature

r(x)w(x) dx =
L r(x;) ~
105
Quadrature
is exact for any nodes x;. If the

that
7rn (x)
are the zeros of
X;
f(x;) = r(x;),
it follows from (10)
= 0,1, ... , n-1.
Hence
1= Lf(x;)~
when f is a polynomial of degree 2n -1 at most. This is the so-called

Gaussian (Mechanical) Quadrature.
It is reasonable to suppose that error estimates of the form
11- 01 :s c..M2n
could be established for integrands in C 2 n[a, b], where Cn is a constant
depending on the weight function. See Problem 9.12.
Tables of the x;, ~ and the Cn are available in all the most important
cases. See e.g., the NBS Handbook, Chapters 22, 25.
4.
GAUSS - CHEBYSHEV QuADRATURE
We shall discuss only the Chebyshev case when the weight-function is

1N1- x 2 , which we can apply to the historically interesting problem discussed in Chapter 1. The nodes in this case are
cos 2r-1)7r/2n), r = 1,2, ... , n.
We have next to calculate the weights
[+1
A". = t1
Tn (x)
(x -
dx
x.n)T~(x.n) . .J1- x2= T~(x.n)
r'7l"
cos n8 d8
cos 6 -cos 6m
where 6m = (2m -1)7r/2n. We shall show that

1
r.
n)
1'71"
cos n8 d8
cos 8 -cos 8m
7r
- T~(x.n)
n
so that the weights are all equal:
A". =7r/n.
Actually T~(x) = -sin (n arccos x) x n x (-l/J(l- x 2)) giving
T' (
n
x.n
)::: n sin 2m -1)7r/2)

sin 8m
-lr- n
1
sin 8m
There does not appear to be any easy way to evaluate

shall indicate three alternatives.
I". directly. We
106
Chapter 9
Method 1
We first notice that the integral 1m is "improper" for the integrand
becomes infinite for 0 = Om. We must interpret r,::) as a "Cauchy Principal
Value Integral":
1~)=cpvl"' ... do=
lim
s --+0+
[1
9m
-"
do+l'"
... dO].
8", +s
The theory of Principal Value Integrals is important but rather delicate. The
prototype is, where 0 < a < 1,
CPV
11 ~=
lim [1
x-a
,,--+0
-"
1 ~]
+r
Ja +" x-a
= lim {[log Ix - a 1]0-" + [log Ix - a IJ!+,,}

,,--+0
= lim [log lei-log a +log (1-a)-log lel}=log{(1-a)/a}.

,,--+0
Note that this is the result obtained by "disregarding" the singularity.

Returning to the case of 1m, a straightforward computation shows that if
<1>(0) =
[~IOg s~ !(q> + 0)]

smq>
sm!(q>-O)
then
d:u <1>(0)
Hence, putting
r!) =
1
cos O-cos ~ .
= Om, and n = 0, we find, for all m,
lim [((Om - e)-(O)}+{('IT)-(Om + e)}]
s--+O+
We next take n = 1. Then

1(1)
I'"
cos 0 dO
COS
O-cos Om
I'" [1 +
COS
cos Om
O-cos Om
] dO
= 'IT + cos Omr!)

= 'IT.
For
n> 1 we obtain a recurrence relation which can be solved explicitly.
107
Quadrature
[Compare Chapter 10]. In fact

[<n+1)
m
+ r.n-1) = r"ll" cos (n + 1)0 + cos (n -1)8 dO

m
cos 8 -cos 8m
= 2{"II" cos n8 cos 8 d8
cos 8 - COS Om
= 2r
"l " [cos n8 + cos 8
cos n8 ]
0
0 dO
cos -cos m
Thus we have
a difference equation, or recurrence relation, with initial conditions [<~) = 0,

I~) = 7r. It is easy to check that the solution to this is
I(n) _
m
-7r
sin n8m
8
sm m
as announced.
Method 2
The integrand can be factorized in the form
cos n8
cos 8 -cos 8m
2n -
[cos 8-cos 8r ]
r=O
rf'm
-this because Tn (x)=2n - 1 [t'.:J (x-cos 8r ). We can reconstitute this product as a sum of cosines of multiples of 8, say,
(12)
cos n8
cos 8 - cos 8m
ao+alcos8+"'+~-lCOS
If we integrate this between 0 and

contributes is the first and it gives
7r
(n-1) 8.
the only term on the right which
We have therefore to evaluate ao. This can be done as follows. We write
Chapter 9
108
down (11) for 8 = 8m, 8m + 271'/n, ... , 8m + (n -1)271'/n getting
ao+at cos 8m + ...
+ ll,,-t cos (n -1)8m =
ao+ at cos ( 8m + 2n71') + ...
+ ll,,-t cos {(n -1) ( 8m + 2n71') } = 0
T~(x".)
ao+at cos (8m +(n-1) 2n71')+ .. . +ll,,-t cos {(n-1)(8m +(n-1)
2;) }=o
We have to check the entries on the right hand side.

The first entry has to be evaluated as
lim cos n8 - cos n8m
8 -8
COS 8 - COS 8m
m
(we have inserted the zero term cos n8.n in the numerator). With the usual
notation
cos n8-cos n8m =Tn(x)- Tn (x".) = l".(x)T~(x".)
cos 8 - cos 8m
X - x".
and, as 8 -+ 8m ,
-+ x". and
l". (x)n(x".) -+ l". (x".)T~(x".) = T~(x".).

The remaining entries are all zeros since
while
cos (8m + 2;:) -
COS 8m
0 if r = 1, 2, ... , n-l.
We now add up the n equations to find, as required,

1m = 71'ao = 71'T~(x".)/n
since the coefficients of all the other a's vanish - the formula for the sum of
the cosines of angles in an arithmetic progression gives
cos r8m + cos r( 8m + 2n71') + ... + cos r( 8m +(n -1) 2n71')
cos ( rOm + 71'(nn- 1)sin r7r
. r71'
n
SIn-
=0.
For another evaluation of all the coefficients a;, see Problem 9.24.
109
Quadrature
Method 3
We begin with a special case of the Christoffel-Darboux formula.
Lemma 12. If x -:j:. y then:
I +2
r=l
cos rx cos ry =
cos (n + l)x cos ny -cos nx cos (n + l)y

.
cos X -cos y
Proof. Write
2 cos rx cos ry = cos r(x+y)+cos r(x-y)
and use the formula
_sin !nO cos !(n + 1)0
~cosrO-
r=l
.1
SlD'ZO
Apply the lemma in the case when y = Om = (2m -1)'IT/2n, x = 0 to get

-cos nO cos [(2m-1)+Om]
1 +2
L cos rO cos rOm = - - - ------cos O-cos Om
r=l
( _1)m-1 cos nO sin Om

cos O-cos Om
We now integrate the last relation with respect to 0 from 0 to
'IT
'IT
to get
= {( _l)m-l sin Om}Im
so that
which is the result required.

The evaluation of the constant in the error bound is discussed in
Problem 9.12: it is found that in this case
Unfortunately there are few cases when f 2n ) can be obtained conveniently.

5.
THE LEMNISCATE CONSTANTS
We now return to the discussion of the length of a quadrant of the

lemniscate of Bernoulli given in polar coordinates by the equation
r2 = cos 20.
Chapter 9
110
Fig. 9.2
This curve has the shape of a figure-of-eight as indicated in Fig. 9.2

Elementary calculations give
( dS)2 = 1 + (rdo)2 =_14 .

dr
dr
1-r
Hence the perimeter of a quadrant of the curve is
dt
o J1- t 4
1
this quantity is known as the lemniscate constant and is usually denoted by

!w or A. It, together with
B=~=Il~4
2w
J1-t
'
has been calculated to 164 decimals. We shall obtain approximations to A, B

using the Chebyshev quadrature
2A =
fl
-1
dt
J1- t4 =
fl
-1
J1 + t2 . J1- t2
and
where
Xr =
dt
(2r - 1}7r/2n.
Quadrature
111
We find the following values:
n=1
2
3
On
Rn
7r =3.14159
.J2/37r=
7r
"3 {I + [4/.J7n=
.J2/37r/2=
2.62207554 = 2A
1.1981402347 = 2B
7r/.J7 =
4
5
6
7
8
9
10
20
00
4AB=7r
As already mentioned Gauss established the representation of the

arithmetic-geometric mean in terms of complete elliptic integrals, after
observing the identity to eleven decimals of the quantity
2B = 7r/2A = 1.19814 02347 = lim R
with M(1, ../2) which he had computed by quite different methods.
The first evaluations of A were made by ,James Bernoulli. James
Stirling in 1730 found A by representing the integrand by a binomial
expansion, integrating term by term and finally applying certain special
devices to accelerate the convergence of the series. For further discussion of
these constants see John Todd, Comm. ACM, 18 (1975), 16--19.
6.
STIRLING'S FORMULA
We shall now use the Trapezoidal Formula to establish

(13)
n! ~ (;) n J27rn.
The basic idea is that log n! = g'log r and to approximate this first by
J~ log t dt and then approximate the integral by trapezoidal sums.
Since (xlnx-x),=lnx we have
In lnxdx =nlnn-n+1.
Chapter 9
112
We now write
I" 12 1
1= 1+
+ ... +
Lr+
1"
+ ... + -1
and the general term

1'+1In X dx =![In r+In(r+ 1)]+e,.
where, from Problem 9.3, we have for r = 1, 2, ...
1
1
1
e. = -12 (In xk+&, = 12 . (r+1't.?'
(14)
Hence we have
n In n-n+l =![(In 1 +In 2) + (In 2+In 3)+ ... +(In (n-l)+In n)]+
L e,.
"-1
.=1
i.e.,
n-l
L e,.-1
(n+!)Inn-n=In(n!)+
.=1
or
(15)
In(n!)=(n+!)lnn-n+
"-1
L e,.-I.
.=1
It is clear from (14) that L e,. is a series of positive terms which is

convergent by comparison with L r-2 Denote its sum by l-u and (15) gives
(16)
In (n!) = (n +!) In n - n -u-p..
where
00
Pn=Le,.
r=n
We now have two problems, the first to estimate p.. and the second to
evaluate u. It is clear that
A L (r+l)-2:5p..:5A
00
r=n
and the result of Problem 2.5 shows that
L r-2
00
r=n
Quadrature
113
With a little more effort we can obtain inequalities

1
---<p,.<
,n2:2.
12(n + 1)
12(n -1)
The evaluation of u is not so easy and we get it by use of Wallis'

Formula (cf. Problem 9.16):
.
24n x(n!t
bm 2n) !f(2n + 1)
(17)
~1T.
Take logarithms in (17) and use (16) twice (once as is, and once with n
replaced by 2n) to get
4n In 2+4(n +~) In n -4n -4u- 4p,.
-2[(2n +~) In 2n -2n -u-P2n]-ln (2n + 1) ~ In ~1T.
Simplifying and noting that
In (2n + I)-In 2-ln n =In (1 +(2n)-1)~0
we get, remembering that u is a constant,
u
-!In 21T.
Hence we have
In n! = (n+~) In n -n +In ..[i;-Pn
which gives (13).
Chapter 9, Problems
9.1. Find the coefficients of the Lagrangian quadrature on the interval
( -1,2), based on the nodes -1, 0, 1, 2. Find an error estimate.
9.2. Find the coefficients of the Lagrangian quadrature, on the interval
( - 2, 2), based on the nodes 0, 1. Find an error estimate.
9.3. Derive the error estimate for n-panel trapezoidal formula
I-Q= -A{b-a)3f'(c)/n 2 ,
a:5c :5b,
where it is assumed that 1E C 2 [a, b].

9.4. Obtain the Simpson quadrature, with error estimate, assuming
C 4 [ -1,1],
fl
I(x) dx =1[f( -1)+4/(0)+1(1)]+ E,
1E
Chapter 9
114
where, M4 being an upper bound of r)(x) on [ -1,1],
IElsioM 4.
Extend this to the many-panel case.
9.5. Estimate the difference between the Milne quadrature (Problem 9.2)
and the Simpson quadrature (Problem 9.4.)
9.6. Prove that the second column of the Romberg array gives the Simpson
quadratures,
9.7. Show that the perimeter of a regular n-gon inscribed in the unit circle
IS
2n sin 7Tln
and that the ratio of this to the circumference of the circle is
7T 2
1- 6n 2+O(n-4).
9.8. Write a program for carrying out a Romberg quadrature for J:f(x) dx,
where f(x) is specified by a subroutine. Allow for up to 10 steps and
arrange for printing out the whole Romberg triangle. Alternatively,
familiarize yourself with the Romberg subroutine available to you, and
check it thoroughly.
9.9. Complete the table of Qn> R,..
9.10. Prove that
~=
(2n)!
o J1-t 4 n=o4n(n!?(4n+1)
X 4n + 1
Ixl<1.
Assuming that it is permissible to let x ~ 1 in this we find
fl
A dx - oJ1-x 4 -
(2n)!
4n(n!)2(4n+1)
By writing (1- X4)-1/2= (1- x 2)-1/2(1 + X2)-1/2 and using the fact that
J.o t2n(1 -t
l
2)-112
d t= 1.3.5 .... .(2n -1) .7T

2.4.6 .... .(2n)
2
and the binomial theorem for exponent

tion
A =~
-1 as before obtain the representa-
{1 + n=1f (_I t [1.3.5

.... (2n -1)]2}.
2.4.6 .... (2n)
Quadrature
115
Experiment with the evaluation of A from these two series, applying

the Euler transform to the second one.
9.11. Show that the polynomials 7Tn (X), orthonormal with respect to a
positive weight function w(x) in [a, b] are unique (if we insist that the
coefficient of x n in 7Tn (x) is positive).
9.12. Use the result of Problem S.12, to obtain the following error estimate
for Gaussian quadrature: if f 2n )(x) is continuous in [a, b] and {7Tn (X)} is the
system of monic orthogonal polynomials associated with the weight function
w(x) then for some~, a~~~b.
Evaluate the integral occurring in this estimate in the case of the

Chebyshev quadrature and in the case of the Legendre quadrature.
9.13. (a) Evaluate S6 x dx/(1 + x 3 ) by (1) finding the indefinite integral and
evaluating it; (2) many panel trapezoidal quadrature and (3) Romberg
quadrature. Where possible give a (theoretical) error estimate and compare
it with the actual (observed) error.
(b) The same as (a) but for S6 (1 + XTI dx, for r = 3, 4, 5, 6.
9.14. Discuss the evaluation of the Sievert integral

Sex, y) =
exp (-y sec t)dt.
In particular discuss the following set of estimates Sh for S(0.5,0.5)

obtained by quadratures at the intervals h indicated:
h = 0.125
Sh: 0.29665 65245
h = 0.05
0.2966574757
h = 0.0125
h = 0.005
0.2966575010 0.2966575011.
9.15. Evaluate the following integrals, all between 0, 1 using, if necessary,

subroutines for the Bessel functions involved.
f
=f
11 =
dx/Jo(x),
12 =
14
{xJO(x)/J1(x)} dx,
Is =
17 = flOg {x/2J l (x))} dx,
f
f
x dx/Jl(x),
13 = flOg [JO(X)]-l dx,
Jl(x) dx/Jo(x),
h=
Is = flOg {X 2/SJ2(X)} dx.
J2 (x) dx/Jl(x),
Chapter 9
116
9.16. Establish Wallis' Fonnula

lim
24n X (n !)4
2n!)f(2n + 1)
'TT
2
Is this a practical method for computing 'TT?
rJ1d~t4X rJ~~:4=~
9.17. Establish the following result of Euler
9.18. Check the expression derived for the error in the "~" Rule (Problem
9.1) in the case of a function f(x) which is such that fEC 4 [a,b] by
evaluating the quadrature error in the case of X4 dx.
Suppose the interval [a, b) is divided into N equal panels and the "~
rule" is applied to each. Give an estimate for the total error in the general
case. How many panels are required if the total error is to be less than
E (b - a)5? How many times is it necessary to evaluate the integrand?
S:
9.19. Let {~(x)} be the fundamental Lagrangian polynomials based on the

zeros of 'TTn(x), the polynomial of degree n of the orthogonal system defined
by the weight function w(x) on the interval [a, b]. Show that
r
a
w(x)~(x) dx = w(x)[~(X)]2 dx,
i = 1, 2, ... , n.
9.20. Compare the efficiencies of Simpson's Rule, the

2-point Gauss-Legendre quadrature for the evaluation of
and the
f(x) dx, where
~-Rule
J~
(4)
If(x)1 ::51,
9.21. Draw a rough graph of Io(x) in the range O::5x ::55. Using a library
subroutine (which you should check (and say how)), or your own subroutine,
for Io(x) evaluate
f(x) =
Io(t) dt
for x = 1.2(0.2)3.
9.22. Experiment with the Richardson idea in connection with the evaluation of Euler's constant "I lim "In lim {1 +!+ ... + (l/n)-log n} =
0.57721566 along the following lines.
From Problem 2.10, in the first place, we observe that
I'n = I'+0'(n- 1 ).
Quadrature
117
Evaluating 'Yn for two values of n and tentatively replacing the 0(n- 1 ) by
An -1 enables an estimate of l' to be obtained.
From the final result of Problem 2.10 we observe that
'Yn ='Y+~n-l+0(n-2).
Replacing 0(n- 2 ) by Bn- 2 enables a better estimate of l' to be obtained.

9.23. Evaluate S6 2 f(x) dx by various methods, e.g. Simpson's Rule,
(Milne)3, (i)4, and discuss the results, comparing the observed and theoretical errors, where f(x) is given by the table:
105 f(x)
lOx
105 f(x)
37500
-41206
33794
-23300
23200
+20794
7294
10
+100000
-11300
11
+224294
-28906
12
+404700
-40800
lOx
[Note f(x) is got by rounding a polynomial.]

9.24. Evaluate the Fourier Coefficients ao, at. ... , ~-1 in (10) by first
establishing the relation
cos nx - cos na
cos x-cos a
cosec a{2 sin a cos (n -l)x +2 sin 2a cos (n -2)x + ...
+ 2 sin (n -l)a cos x + sin na}.

9.25. Find in your calculus text books, or elsewhere, methods of
establishing
(" e -x dx
2
= i.J;.
Discuss the numerical evaluation of this integral by various methods.
CHAPTER 10
Difference Equations, Differentiation and Differential

Equations
1.
DIFFERENCE EQUATIONS
Difference equations occur in many branches of science directly and

also indirectly when we approximate a differential equation so that we can
obtain a numerical "solution" on a digital computer.
We begin with a brief account of linear difference equations with
constant coefficients - these have been encountered already as recurrence
relations.
The unique solution of the first order system
u,,+1 = au",
(1)
Uo given
a;eO,
is manifestly
(2)
Any second order equation of the form

can be written as
(3)
where a, (3 are the roots of the quadratic

02-aO-b=O.
We reduce the solution of (3) to that of (1) by rewriting it in two ways as

(u,,+2- (3u,,+1) = a(u,,+1 - (3u,,)
and as
(u,,+2- a u,,+1)
= (3(u,,+l- a u,,).
The solution (2) gives us

(4)
Un+l - (3u"
= (Ul -
u"+1-au,,
= (Ul- auO)(3n.
(3uo)a n
If a;e (3 we can eliminate u,,+1 from (4) to find

u" =Aa n + B(3n
Difference Equations. Differentiation and Differential Equations
119
where A, B are certain constants. If a = 13 we find, instead,
where C, D are certain constants.

These results extend to the case of a linear difference equation of order
n with constant coefficients. Observe that, in general, unless lal < 1, 1131 < 1
the solutions will be unbounded.
Very few non-linear difference equations have been discussed. The
Arithmetic-Geometric mean algorithm of Gauss (Chapter 1) can be regarded as difference equations for a,. or bn Many of the special functions
satisfy linear difference equations with variable coefficients e.g., Tn(x) and
Pn(x). (See Problem 10.2). As an example we discuss the Bessel functions
which were originally defined by:
r"
1
1n(x) = 1T.Io
cos (x sin O-nO) dO.
This function satisfies
1n+1(x) = 2nx- 11n(x)- 1n- 1 (x).
[See Appendix, p.144, for a proof.]

It would appear that this equation would be a convenient way to
compute 1n(x), given 10(x), 1 1(x). Let us see what happens if we try to find
120(1) given 10D values of 10 (1), 1 1 (1). We underline the digits which are
incorrect!
10=0.76519 76866
II =0.44005 05857
12 =0.11490 34848
13 = 0.01956 33535
14 = 0.00247 66362
Is = 0.00024 97381
h = 0.00002 07248
17= -0.00000 10385
It is clear that this seemingly attractive approach is completely unsuitable.
The trouble is caused by the "numerical instability" of the process - the
unavoidable rounding errors in the data are amplified by the factor 2nx- 1
and there is also some cancellation.
However something can be saved: the fact that the ascending recurrence is unstable suggests that the descending one (i.e., solving for 1n - 1 in
120
Chapter 10
terms of In and I n+1) might be stable! There is a difficulty here since we have
no initial values but it turns out that arbitrary initial values will do!
For instance if we take i 40 = 0, i 39 = 1 X 10-8 and use
i n- 1 = 2nin - i n+ 1
we obtain
i 40 =0
i39 = 0.00000
i38 = 0.00000
i37 = 0.00005
i 36 =0.00438
i 35 =0.31567
i 20 = 0.43716
001
078
927
520
512
256 x 1026
This, of course, is not likely to be the correct result since the difference
equation is homogeneous and we can multiply by any scale factor. There are
several ways of finding the appropriate mUltiplier: the simplest is to carry on
the above recurrence to obtain
io = 0.86360016 x 1050
Comparing this with the correct value of 10 we see that the appropriate scale
factor is
k =lo/io = 0.88605 550x 10-50
Applying this to
i 20 we find our estimate

120 = 0 20 = 0.38735029 x 10-24
which is correct to within a unit in the last place!

The choice of the starting points 40, 39 is largely dictated by experience. If instead we start at
1- 24 = 10-8
we obtain
io = 0.80486703 x 1023 ,
k = 0.95071315 X 10-23 ,
120 = 0.38735029 x 10-24
Another way to obtain the normalizing factor k is to use the fact that
Difference Equations, Differentiation and Differential Equations
121
or the fact that

These facts are established in the Appendix.
This method is applicable to compute many of the important special
functions which arise in various branches of science. Among those who have
contributed recently to the analysis of this method are Gautschi, Olver and
Thacher.
2.
DIFFERENCING
We have already introduced the forward difference operator a when

the interval is 1 by the equation af(n)=f(n+1)-f(n) and by af(x)=
f(x + h) - f(x) when the interval is h. In general we define, for r = 1, 2, ... ,
ar+1f(n) = a{arf(n)}
so that, for instance,

a 2 f(n)
= fen + 2) - 2f(n + 1) + fen)
and, generally,
(5)
arf(n) = fen + r)- G)f(n + r-1)+ (;)f(n + r-2)- . .. + (-lYf(n).
It is easy to check that if
fEe then
If we consider differences at interval h so that

fen) =F(nh)
then
a'(F(nh
= hrpr)(p)
It follows from the last equation, that as h ~ 0, arF(nh) ~
when F
is a smooth function. The differences of order r + 1 or higher of a polynomial of degree r are exactly zero, no matter what the interval is. Indeed
ahx r = (x + hY - xr = rx r- 1 h + ...
so that differencing a polynomial produces a polynomial of degree one

smaller; since the differences of a constant function are zero, the result
follows.
This statement about polynomials depend on exact arithmetic and will
in general, fail in the presence of round-off. The effect of a round-off error
or a mistake in a table is obvious from (5): it propagates out in the shape of
122
Chapter 10
a triangle with a vertex at the wrong entry and binomial coefficients along
the base thus
o
o
o
1
-1
-2
o
o
o
-3
3
This phenomenon permits the checking of tables. See Problem 10.15.

3.
DIFFERENTIATION
Before beginning our discussion of differential equations we need to

discuss the approximate representation of differential operators, in particular, D2 = (d/dx?
We do this symbolically. A very simple treatment starts by considering
the Maclaurin expansion
f(h) = f(O) + hf'(O) + (h 2 /2!)f'(0) + ...
which we can write as
(1 + a)f(O) = ehDf(O)
where a indicates, as before, the forward difference operator

af(nh) = fn + l)h)-f(nh).
We now write
e hD
= l+a,
and we can obtain h 2D2 by formally squaring the logarithmic series to get
h 2 D2 =a 2 -a3 +(11/12)a4 -
More efficient representations can be obtained in terms of central

differences:
For instance, we find

(6)
~h(f(O =
f@.h)-f(-!h),
~~(f(0 =
f(h)-2f(0) +f( - h).
123
Note that, as compared with the previous representation, there are no odd
terms present and the coefficient of the even ones are smaller. [Collections
of useful representations are available, e.g., in Interpolation and Allied
Tables.]
We shall make use of the first two terms of (6). Let us see what sort of
approximation they give. If f E C4
f(x h) = f(x) hf'(x) +(h 2/2!)f"(x) (h 3 /3!)f"(x)+ O(h4)
and we see that

f"(x) f(x+h)-2~~)+f(x-h) =O(h2).
In the same way, if
f E C6 , we find
f"(x)- [f(X + h)-2~~)+ f(x - h)

_f(x +2h)-4f(x + h)+6f(x)-4f(x - h)+f(X-2h)] = O(h4)
12h2
.
If we write f(nh)
= fn then we have two approximations to f"(nh):

fn+1 - 2fn + fn-l
h2
(7)
- fn+2 + 16fn+l - 30fn + 16fn-l - fn-2

12h2
The comparative strength of these, in approximating 1"(1) = - f(l), is indicated by the following table, which also indicates the effect of varying h.
4.
sin x = f(x)
0.6
0.8
1.0
1.2
1.4
0.564642
0.717356
0.841471
0.932039
0.985450
f'(1)
3-point
interval 0.4 -0.8303125
interval 0.2 - 0.83867 5
interval 0.1 -0.84077 0
interval0.05 -0.8412
interval 0.01 -0.84
5-point
-0.841562
DIFFERENTIAL EQUATIONS - TYPES OF PROBLEMS
The question of the numerical determination of solutions to such

differential equations as the initial value problem
(8)
y'=f(x, y),
y(a) = b
124
Chapter 10
or the non-trivial solutions of the two point boundary value problem
y"+(Af(x)- g(XY =0,
(9)
y(O) = 0,
y(l) =0
is far from trivial! Even if an explicit solution is found, e.g., in E .. Kamke,

Differentialgleichungen, wsungenmethoden and wsungen, I, II, 1944, 1950.
its actual computation may be laborious. To illustrate the last point we note
that the solution of
y(O) = 1
(10)
is (see Appendix for details)
f(i)L 3Ii!x 2 ) + 2f(i)I3/4(!x 2 )
(11)
y = x f(i)IlIih 2) + 2f(i)Il/4(h 2 )
where
Other examples for study are, for appropriate initial conditions:

(12)
y' =..Ix+JY
(13)
y"=xy
(14)
y"=6y2+X
(15)
y"= -yo
No formulas for the solutions to (12) or to (14) are available. The solution to
(15) is, of course, y = a cos x + b sin x. The solution for (13) which satisfies
the initial conditions
y'(O) = r
2/3
/fG) = 0.35502 8 .. .
y'(O) = 3- 1/3/fG) = 0.258819 .. .

is the Airy integral
(16)
Ai(x) =
1'" cos Gt + xt) dt.

3
[The reader is invited to consider the numerical evaluation of this integral!

See Appendix.]
125
The equation (15) can be written in the form (8) if we interpret it

vectorially
Y'=F
where
= [Y']
Y
and F
[01 -1]O
As an example of the Sturm-Liouville system (9) we take the problem

of finding values of A. for which
(17)
y"+A.xy =0,
y(O) = 0 = y(l),
has nontrivial solutions.

5.
NUMERICAL SOLUTION: PICARD, EULER, HEUN
Some of the proofs of the existence of solutions to (8), say, have a

constructive character and can lead to algorithms for computing the numerical solution. For instance there is the Picard method in which a sequence of
functions is defined by quadratures which, subject to mild assumptions on
{(x, y), converge to a solution. Specifically if we write Yo(x)=b and then
define
(18)
Yn+1(x)=b+
{(t, Yn(t)) dt
then the sequence Yn(x) converges to the required solution. Since we know
how to integrate numerically this is, in principle, a solution to our problem.
A few experiments, in which the integration is done analytically, shows that
convergence is likely to be slow. This method, however, is usable for small
ranges of x.
Another scheme, due to Euler, consists in approximating the integral
curve by polygons: we write
x(n)=a+nh,
y(n) = y(x(n))
and then put Yo = b and, for n = 0, 1, ...

(19)
y(n + 1) = y(n) + h{(x(n), y(n)).
This gives us an approximate solution which depends on the mesh size h. If

we refine the mesh, e.g., by continual halving, we will get a sequence of
appropriate solutions. It can be proved, again subject to mild restrictions on
{(x, y), that these approximate solutions converge to the solution of (8).
126
Chapter 10
y
ok---~------~----~
Fig. 10.1
It is easy to see that the Euler method is not promising numerically.

Suppose the solution can be represented as a power series:
(20)
y(x + h) = y(X) + hy'(x) + (h 2/2!)y"(x) + ....
Our approximate solution (19) only agrees with (20) as far as the linear term
since y'(x)=f(x, y). The local error is therefore O(h2). Since there will be
O(h- 1 ) steps required to cover a finite interval we might expect the total
error to be O(h- 1 )O(h 2) = O(h) so that we would have to take a step size of
10-6 to get a 6D solution. By a slight modification of this, due to Heun, we
can improve things by an order of magnitude.
We define, for n ~ 0,
y*(n + 1) = yen) + hf(a + nh, yen)),
(21)
y**(n + 1) = yen) + hf(a + (n + l)h, y*(n + 1)),

yen + 1) = t{y*(n + 1) + y**(n + 1)}.
Compare diagram.
If we assume that y can be expanded in a power series we find, after a
little manipulation, that the local error is 0 (h 3 ); in fact, with some care, we
find it does not exceed in absolute value
3
h
12
Y"(~)_3y,,(~)(af)
ay (~,'11)
I.
This means that a total error of O{h2) may be expected. An important

feature of this method is that we can change h without any trouble. Another
is that we can often use Richardson extrapolation.
We illustrate the use of Richardson extrapolation to improve the results
of Heun integration in a simple case. The following values were obtained by
integrating the differential equation
y'=y,
yeO) = 1
127
at interval 0.1 and at interval 0.05 from 0 to 1,
YO.! = 0.27141 x 10,

YO.05 = 0.27172 x 10.
The difference is 31 x 10 and the Richardson extrapolant is
4Yo.o~-
YO.! = YO.05 +~(0.00031 x 10) = 0.27182 x 10
which is correct.
Continuing the integration we get at x = 2(1)5:
0.73662 x 10
x = 2: 0.73831 x 10 (0.73831 +~(0.00169)) x 10 = 0.73888 x 10
0.19993 X 102
x = 3: 0.20061 x 102
(0.20061 +~(0.00068)) x 102 = 0.20085 X 102
0.54261 X 102
x = 4: 0.54511 x 102
(0.54511 +~(0.00250)) x 102 = 0.54595 X 102
= 5' 0.14727 X 1033 (0.14812+-31 (0.00085)) x 103 = 0.14840 X 103

. 0.14812 X 10
Thus the errors are -1, -1, -1, 1 in the last place.
The popular Runge-Kutta method is a development of the Heun
algorithm: in it 4 values are averaged compared with 2 in the Heun case.
The local error in the Runge-Kutta method is O(h 5 ) and the global error is
O(h4) so that h 4 -extrapolation can be applied.
Among other methods are some in which extrapolation is built-in. We
shall not discuss these but now turn to two less sophisticated methods.
6.
MILNE-SIMPSON PREDICfOR-CORRECfOR
We shall discuss now a primitive Predictor-Corrector Method. These

have the disadvantage that special starting procedures are required which
makes reduction of mesh size awkward.
We use two quadrature formula (Chapter 9, Problems 2, 4). The Milne
formula is:
f2
f(x) dx ~1[2f(-1) - f(O) + 2f(1)]
with error
14h 5 f4({} )
45
! ,
128
Chapter 10
The Simpson formula is
[+1
L1 f(x) dx :H(f( -1)+4f(0)+f(1)]

with error
The application to the solution of y' = f(x, y), yeO) = b is as follows.

Assume a mesh size h and assume that by some method (e.g., Picard's) that
we have obtained adequate approximations to Yl = y(h), Y2, Y3 and consequently to fl = f(h, y(h)), f2' f3. Since
we can "predict"
(22)
by Milne's quadrature. We then compute f4 and apply Simpson's quadrature
to get a better value of Y4 from
as
(23)
If there is satisfactory agreement between the two values of Y4 we can
proceed. Should the agreement be unsatisfactory, a reduction in h may be

appropriate.
We note that the first value of f4 is multiplied by h/3 in (23) and
uncertainty in it is not important but, should a correction or revision in
(22) be required, the consequential change in f4 should be made. The error
in the Simpson quadrature is about (-1/28) times that in the Milne
quadrature; hence if the observed difference is not more than 15 units in the
last place, it may be assumed that the Simpson value is correct to within a
round-off.
7.
129
LOCAL TAYLOR SERIES
The following method is convenient when it is possible to get a

reasonable recurrence relation for the derivatives. The formulas used are
(24)
where the "reduced derivatives" Tnf are defined by
Tnf(x) = hnfn)(x)/n!
We choose the Airy equation to illustrate the method:

(25)
y"=xy.
For a discussion of this equation see Appendix. The Airy integral given by
(16) is one solution. We differentiate this n times by Leibniz' Theorem and
find, in the T-notation,
(n + 1)(n +2)Tn+2 = h 2 xT n
+h3 T n- 1 ,
n>O,
2T2y = h2 xy.
We recall that Ai(x) is defined as the solution of (25) for which
(28)
We proceed as follows, choosing h = 0.1 and working to 6 places:
x=O
y
ry
T2y
0.355028
0
0.000059
T4y
-0.000002
0.329203
0.380849
I'+ -0.025713
I~
0.329203
-0.025882 -0.025713
T3 y
I+
I_
x=0.1
-0.025697
0.000165
x=0.2 x =0.3 x =0.4
x=0.5
0.231694
-0.022491
130
Chapter 10
The numbers in the first column are obtained as follows. The first two
come from the initial conditions (28), the third from (27). The fourth comes
from (26 1) and the fifth from (262 ). The entries opposite L come from (24)
and give y(O.l): the first gives us a new value to be entered at the top of
the second column and the second would enable us to check y( -0.1) had
this been available. The entries opposite L~ are
T
2T2 + 3T3 4T4 + ...
and give the values of T(O.l): the first is entered in the second column and
the second would enable us to check T( -0.1) had this been available. We
then compute T2y(0.1) from (27) and we are ready to proceed.
This method has the advantage that we check our values as we go along
and do not have to wait until, for instance we can check our numerical
solution from an asymptotic representation of the solution.
8.
INSTABILITY
We want to discuss the equation (15) further in the case when y(O) = 0,
y'(O) = 1. We replace it by
y(n + 1) = (2- h 2 )y(n) - y(n -1),
(29)
where we are using

". y(n+1)-2y(n)+y(n-1)
y =:=
h2
Using h = 0.1 and the correct value for y(l) we get the solutions given in
column (2) - the correct 10D values of sin x are given in column (1). The
values obtained are good to about 3D. If we want better results we seem to
have two possibilities: take a smaller h, or take a better approximation to
the differential equation. We choose the latter and use the better 5 point
approximation to y" which we have derived earlier. We get, for h = 0.1
(30)
y(n +2) = 16y(n + 1)-29.88y(n)+ 16y(n -1)- y(n -2)
If we use the correct values for y(l), y(2) and y(3) we get, according to
whether we work to 5D or 10D, the results given in columns (4) and (3).
How is this to be explained?
The general theory of linear difference equations with constant coefficients which was outlined earlier indicates that the solution to (30) is of the
form
131
where AI> A2, A3, A4 are the roots of the quartic

A4-16A 3+29.88A 2-16A + 1 = O.
We find that Al = 13.94, A2 = All, A3 = A4 = exp (i xO.1000). Thus
y(n) = AA~+ BAln+C cos O.ln + D sin O.ln.
The solution we are looking for is given by A = B = C = 0, D = 1. It is clear

that if any component of the solution A~ enters then it will swamp all the
rest. This is indeed what happened: on the determination of y(4) in one
case, (3), we get a component with A >0 and, in the other, (4), one with
A<O.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
b.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
1.6
9. Two
(1)
(2)
(3)
0.0000000000
0.0998334166
0.1986693308
0.2955202067
0.38941 83423
0.4794255386
0.56464 24734
0.64421 76872
0.7173560909
0.7833269096
0.8414709848
0.8912073601
0.9320390860
0.9635581854
0.9854497300
0.9974949866
0.99957 36030
0.00000
0.09983
0.19866
0.29550
0.38939
0.47939
0.56460
0.64416
0.71728
0.78323
0.84135
0.89106
0.93186
0.96334
0.98519
0.99719
0.99922
(4)
0.0000000000 0.00000
0.0998334166 0.09983
0.1986693308 0.19867
0.2955202067 0.29552
0.3894183865 0.38934
0.4794259960 0.47819
0.56464 90616 0.54721
0.6443099144 0.40096
0.7186422373 -2.67357
0.8012545441
1.09135 22239
4.37411 56871
POINT PROBLEMS
The following results concerning the simple equation

(31)
y"=-AY,
y(O) = 0 = y(l)
are, on the whole, representative of the behavior of the general case (9). The
general solution of (31) is
y = A sin ./Ax + B cos ./Ax
and the condition y(O) = 0 gives B = O. For non-trivial solutions we must
132
Chapter 10
have A;c 0 and in order to satisfy y(1) = 0 we must have

sin J)..=O,
so that A must be positive and of the form n 2 1T2 for some n = 1,2, .... The
solution for A = n 2 1T2 is a multiple of
y = sin n-lTX
and has n oscillations in [0,1].

We shall now consider the numerical solution of (31). First since the
problem is homogeneous, we must normalize the solution; since y(O) = 0, we
choose to require y'(O) = 1. [If y'(O) = 0 then we would have a trivial
solution.] Then the solution is uniquely determined by a choice of A, and we
can determine the corresponding value of y(1) by integrating the equation
numerically, and our problem is to find the values of A, in particular the
smallest, for which y(1) = O. Our proposal is to find y(1) for various A'S and
to interpolate among them to get the one required.
Our first problem is to get an approximate value of A. We do this by
discretizing our problem to the following, using the simple approximation to
a second derivative. The equations
YO -2Y1 +Y2
(1/16)
-AY1
Yl -2Y2 +Y3 =_AY.

(1/16)
Y 2 -2Y3 + Y4 =-AY.
(1/16)
assert that the equation y" = - AY is approximately satisfied at
by values
This homogeneous system of equations which becomes
ILY + Y
1
=O}
Y 1 +ILY2 + Y 3 =0
Y 2 +ILY3 =0
133
when we write IL = (N16) - 2, has a non-trivial solution provided
det[~ ~ ~] =0,
o
1 IL
i.e. if IL = 0 or J2. Thus the approximation to the smallest A is

16(2-J2) = 9.3726.
We can now proceed to do the numerical integration. In view of the fact

that we know the exact solution
Y = A-1I2 sin A1I2x
we shall use this to get the exact values of y(l) which the reader should
compare with those got on the computer:
A =9
A112 =3,
y(l) =0.14112 00l/3
A = 9.61,
A1/2 = 3.1,
y(l) = 0.04158066/(3.1) = 0.001341,
A = 10.24,
A1/2 = 3.2,
y(l) = -0.05837 414/(3.2) = -0.001824.
=0.047040,
Inverse interpolation between these values gives

47040
1341
61
62.79
45699
-1824 124 119.37 86.76 48864 3165

so that our estimate for 1T2 = 9.86960 is 9.8676.
For a further examples of this type see Problems 10.12-13-14.
We re-emphasize the remark that the special case just examined is
representative of the general non-singular Sturm-Liouville problem (9).
There are, however, many practically important "singular" cases. In some of
these the interval is infinite and the boundary conditions may be that
y(x) ~ 0 as x ~ oo, or that f"'oo [Y(X)]2 dx is finite; further, the coefficients
f(x), g(x) may be discontinuous. In such cases the "spectrum" - the set of
characteristic values, A - can be quite different from the simple case.
Chapter 10, Problems

10.1. Find an explicit expression for the Fibonacci number Un defined by
Un +1
= Un + Un-l>
Uo= U1 =
1.
n ~ 1,
Chapter 10
134
Compute as many of these numbers as can be done conveniently on

your machine.
10.2. Obtain the three term recurrence relations of the form
TIn+1 (x) = (A,.x + Bn) TIn (x)- en IIn-l (x)

satisfied by (a) the Chebyshev polynomials, Tn(x), (b) the Legendre polynomials, Pn(x), (c) the Laguerre polynomials L,,(x) and (d) the Hermite
polynomials, Hn(x). In each case use the definition of the polynomials as
those obtained by orthogonalization of 1, X, x 2 , with respect to appro.:.
priate weight functions.
10.3. Discuss the behavior of the recurrence relation
u,,+1 = au~ + 2bu" + c
when
Uo
is given and
a,
b, c are real.
10.4. Show that the integral, with a positive integrand
In =
x n ex - 1 dx
satisfies the recurrence relation

Starting from
I o =1-e- 1 =0.632121 to 6D
calculate II> 12 , 13 , ,110 using the above recurrence relation.
Is there anything obviously wrong with your results so far?
Calculate, correct to 6D, the value of 110= 1334961 - 3628800e -1.
10.5. Experiment further with the numerical differentiation of a table of
sin x, using various intervals and various formulas. Discuss the balancing of
the round-off and truncation errors.
10.6. Examine theoretically the Picard and Euler methods in the case of the
differential equation
Y' = y, y(O) = 1
which has the solution y = eX. Show that in the first case
Yn(x)
= 1 + x + (x 2 /2!) + ... + (xn/n!)
135
while in the second, if the mesh size is h = n- 1 , the approximate solution is

Yn(x) =
(1 +~) n.
lO.7. (Hautus-Veltkamp). Examine theoretically the approximate solution

of the differential equation
Y' = y2, y(O) = 1
by the Euler method, say in the interval [0,1].

lO.8. Write a program for the Heun method and apply it to equation (10)
and equation (12) with initial conditions y(O) = O. The program should allow
for computing at an interval h and printing the results only at certain
mUltiples of h.
You should observe the behavior of the error, say at x = 1, as a
function of h. In the second case the solution cannot be expanded as a
power series at the origin and the results are not as good as in the first.
10.9. Integrate the differential equation
xy'-xy+1=0
from x = 1 to x = lO with the initial condition
y(l) = 0.59634 7361.
(Print out values at integer arguments only.) Compare the result at x = 10
with f(lO), where
f(x) = i~ t-1e x - t dt,
x
computing {(10) from its asymptotic series (Problem 7.1).
10.10. Integrate the differential equation

y' +2xy + x-1_J;=0
from x = 1 to x = 10 with the initial condition
y(l) =0.60513 36.
(Print out values at integer arguments only.) Compare the result at x = 10
with F(10) where
~ e-t2 dt
F(x) =
l --,
x+t
computing F(10) from its asymptotic series (Problems 7.4-6).

10.11. Write a 2-dimensional version of the Heun scheme and apply it to
equations (13), (14), (15). Appropriate initial conditions for (13) have
Chapter 10
136
already been given (using these you can conveniently check your computations from standard tables). In (14) you may take y(O) = 1 and y'(O) = ~m3.
You should observe that in this case the solution tends to infinity as
x ~ X(A) where X(A) depends on the value A of y'(O). In order to estimate
X(A) it will be necessary to decrease h as x increases.
10.12. Discuss the differential equation
y(O) = 0 = y(l)
y"= -AXY,
in the same way as (31) was handled in the text.

10.13. Discuss the solution of the equation
I/I"(x)+(E- V(X))I/I(X) =0
where E is a parameter and V(x) is given by

V(x)=-5,
Ixl<l,
V(x)=O,
Ixl~1.
In particular find a value of E for which a non-trivial odd solution

exists which tends to zero as Ixl ~ 00.
10.14. Discuss the Problem (13) in the case where
Ixl<l,
V(x) = 0,
Show that there is exactly one value of E for which there is an even or odd
(which?) solution with I/I(x) ~ 0 as Ixl ~ 00. Find E to 2D and the corresponding I/I(x) for x =0(0.1)l.
10.15. Write a program to print out the differences of a set of say 12 values
fi> ... ,f12 Check this when the data is
(a) fT = r3 and (b) 16 = 1, all other f's = O.

(c) Use the program and the information you have just developed to
indicate possible errors in the following sets of values:
213
214
215
216
217
-98912726
-77220413
-55444184
-33583877
-1639330
218
219
220
221
222
10389609
32530132
54701371
76984498
99352675
137
10.16. Use the two-dimensional Heun program to solve the differential

equation
y"+y =X-1!2
with initial conditions
y(1) = 0.809525,
y'(l) = -0.232199.
Find y(1.1), y(1.2), ... , y(2) and y'(1.1), y'(1.2), ... , y'(2) using h = 0.01,
h =0.0025.
10.17. Solve the difference equations

(a) Xn+l =(an+b)Xn, a;i:O, Xo given.
(b) Xn+1 =(an+b)Xn +CXn-h a;i:O, Xo,
Xl
given.
10.18. Complete the following difference table:

X
f(x)
1 771
91
2
48
36
5
32
-41
6
7
10.19. Is it likely that the following sequences are the values of a polynomial at equally spaced arguments?
(a)
2, 3, 5, 7,11,13,17,19,23, ...
(b)
2, 5, 4, 12, 6, 9, 23, 11, 27, 34, ...
(c) 41, 43, 47, 53, 61, 71, 83, 97, ...
10.20. Evaluate numerically the derivatives of Ai
at
= 1.4 given the
138
Chapter 10
following set of values:
Aix
1.1
1.2
1.3
1.4
1.5
1.6
1.7
0.12004943
0.10612576
0.09347467
0.08203805
0.07174950
0.06253691
0.05432479
10.21. Integrate the differential equation Y" = xy from

x=1,
y = 0.13529242,
y' = -0.15914744
using the local Taylor series method:

y(xh) = T(O) T(1)+ T(2) T(3) +...
where
h2 x
h3
T(n + 2) = (n + 1)(n + 2) T(n) + (n + 1)(n + 2) T(n -1).
10.22. (D. R. Rartree) Discuss the evaluation of

X(x) = L~ exp [ - (x - t 2Y] dt
along the following lines. Express x(O), x'(O) in terms of r -functions. Show
that
y(x) = exp (!x 2 )X(x)
satisfies the differential equation
Obtain two solutions Yl> Y2 to this for x = -2(0.1)8 where

Yl(O) = 1,
yHO) =0,
Y2(0)=0,
y~(O)
= 1.
Evaluate X(x) for x = -2(1)8.

10.23. Discuss the solutions of the Emden equation
139
for various values of n, e.g., n = 1(0.5)5, which have y(O) = 1, y'(O) = O. In

particular, find the first positive zeros of the solutions.
10.24. Study the solutions of
y"+2x- 1 y'-y +yk =0,
k>l.
which are finite at x = 0 and vanish at x = 00.

10.25. Usethe predictor-corrector method to solve the equation y'=x+y,
y(O) = 1, finding in particular y(1).
APPENDIX
Bessel Functions
It is important to be able to carry out controlled computational experi-
ments to evaluate computing procedures. For this purpose it is necessary to

have reliable numerical data. The various tables of Bessel Functions, e.g.,
BAAS, Bessel Functions, I (1937), II (1952); The Airy Integral (1946),
Cambridge Univ. Press.
RSMT, Bessel Functions, III (1960), IV (1964), Cambridge Univ. Press.
Harvard, Annals of Computation Lab., 2-14, 1945-1951, Harvard
Univ. Press.
NBS, Tables of Bessel functions of fractional order, I (1948), II (1949),
Columbia Univ. Press.
are ideal for this purpose, supplementing the less extensive tables and
graphs in
E. Jahnke, F. Erode, F. LOsch, Tables of higher functions, McGrawHill, 1960.
M. Abramowitz-I. A. Stegun, Handbook of mathematical functions,
NBS Applied Math. Series, 55, 1964.
We therefore give an account of some aspects of the theory of Bessel
Functions which are relevant in the present context. The definitive treatise is
G. N. Watson, Bessel Functions, Cambridge Univ. Press, 1922-1941.
We shall confine our attention largely to In{x)-the Bessel function of
the first kind of integral order n. Two definitions are:

(1)
(2)
11
In(x) = 21T
....
11. .
cos (-x sin 0 +nO) dO = 1T

In(x) =
cos (nO -x sin O)dO
G)n+2r
L,r .(n(-1)'
,+ r).
00
r=O
Our first objective is to establish the equiValence of (1), (2) and we begin by
finding some properties of the In's. We indicate by subscripts on the
numbers of our formulas the definitions on which they are based.
Using the fact that cos A -cos B = 2 sin !(A + B) sin !(B - A) we obtain
142
Appendix
from (1), with n ~ 1:
! r'7l"
I n - 1(X)-J n + 1(X) =
sin (-x sin O+nO) sin OdO
=2J~(x)
assuming it permissible to differentiate (1) with respect to x, i.e., to interchange the operations d/ dx and I ... dO on the right.
We note that when n = 1 we have
1 2'71"
J1 (x) = 2'IT.Io cos (0 - x sin 0) dO
=
1'71" cos 0 cos (x sin 0) dO +
1'71" sin 0 sin (x sin 0) dO.
The first integral vanishes because we can write it as

1 1'71"/2 +1 1'71" =1 +1 ;
-1 L'7I" =1
2
'IT
'IT
'IT ",/2
if we change the variable from 0 to </J = 'IT - 0 in 12 we see that 12 = -11

Thus
J 1 (x) =
L'7I" sin 0 sin (x sin 0) dO.
Assuming that differentiation under the sign of integration is permissible we

find
Jo(x) =
1'71" -sin (-x sin O)(-sin 0) dO
= ~1 1'71" sin 0 sin (x sin 0) dO

and so we have
1.
THE DIFFERENTIAL EQUATION
We first note that the series (2) is convergent for all x: the ratio of the
r+1-st term to the r-th is -x 2 /(4r(n + r)) and this tends to zero as r~oo for
143
Bessel Functions
any x, n. We can therefore differentiate term by term to find

,
_
co (-1Y{n+2r) (~)n+2r-l
In(x)- r~o r!(n+r)!2 2
'
"
In(x) -
(-1Y{n+2r)(n+2r-1) (x)n+2r-2
-2
.
r=O
r. n+r)'2
. 2
'(
co
Consider the coefficient of x n+2r in

(-i)'/[r!(n + r)!2n+2r] times
x2J~(x)+xJ~(x)+(x2-n2)Jn(x):
it is
(n +2r)(n +2r-1)+(n+2r)-4r(n + r)- n 2
which is zero. This establishes (52).

It is easy to check that for general values of v
(2')
G) r~o
(x)2r
-
(-1)'
co
J (x)= -
r!r(r+ 1 + v) 2
also satisfies the equation

(8)
Indeed, if v is not an integer, Lv(x) also satisfies this equation and J", Lv
form a pair of independent solutions to (8). We note from (2') that, for
n =0, 1,2, ...
Ln(x)=
( x)-n
2:
(X)2r
(-1)'
co
r~or!r(r+1-n) 2:
and the coefficients of the terms with index 0, 1, ... , n -ion the right
vanish because of the behavior of the r(t) when t = 1- n, 2- n, . .. ,0.
Hence, putting s = r - n,
Ln(x)=
( x)-n
2:
{~)-n
= \2
=(-1t
co
-1)'
(x)2r
r~nr!r(r+1-n) 2:
co
.~o
(-1t+.
(x)2'+2n
r(s+ 1 + n)r(s + 1) 2:
( x)n
co
(-1)'
(X)
2: .~os!r(s+l+n) 2:
2.
= (_1)RJn (X).
Therefore, in the case when v is an integer we have to seek elsewhere for

another solution of (8). This leads to the introduction of the Bessel functions
of the second kind - we shall not be concerned with these here.
144
Appendix
2. MIxED RECURRENCE RELATIONS
These are simple consequences of (2). We deal only with the first
relation. The coefficient of x2v+2r-1 on the right is
1
(-1)'
1
1
(-1)'
1
--'---'---. _ . 2r+2v =_.
.
-.
(r+ v)
2v r!f(r+l+v) 22r
2v- 1 r!(r+v)f(r+v) 22r
which after cancelling the factor (r + v), is exactly the coefficient of x2v+2r-1
on the left.
H we differentiate out the left hand sides of (92 ) and multiply across by
x-v and -XV we obtain
v
x
-J~(x)+- J)x)
= J v+1(x).
From these, by adding and subtracting, we obtain

2v
- Jv(x) = J v- 1(X)+Jv+1(x);
2J~(x)
= J v- 1x)-Jv+1(x).
We now show how to derive the first part of (1l~ on the basis of the
original definition (1). [The second part of (11 2 ) is just (3 1),]
Elementary trigonometry gives, if x -F 0,
cos n + 1)8- x sin 8)+cos n -1)8- x sin 8)
= 2 cos 8 cos (n8 -
= 2n cos (n8 x
x sin 8)
x sin 8) -~ (n - x cos 8) cos (n8 - x sin 8).

x
H we integrate across with respect to 8 between 0,21T we find
2 2mr cos 'P d8

21TJn +1(X) + 21TJn _ 1(X) =2n
- 21TJn (X)-x
x
where 'P = n8 - x sin 8. The last integral vanishes and so
We are now able to identify (1), (2). We do this by checking the results
first for n = 0, and then using the fact the recurrence relations are satisfied
on the basis of (1), (2).
145
Bessel Functions
Since the series for cos t is convergent for all t we can expand the
integrand in (1) as a power series and integrate using the fact that for all
even r
(/2)'71"
smrOdO=
(r-1)(r-3) ... (1)

r(r-2) . .. (2)
'IT
.2
to find
1
-
r'7l"
'IT
.
1 r'7l"
cos (x sm 0) dO ='IT
2r 0
(-1)'x2r sin
L
(2 )'
r=O
r .
00
=r=O
L
00
[(
dO
_1)'x2r 1 r'7l" .
]
(2 ) , . sm2r OdO =
r.
'IT
(-1)' (x)2r
,)2 -2 .
r=O r.
L -(
00
This completes the identification for the case n = o.

We now note that a special case of the second relation of (92 ) is
J6(x) = -J1(x). Combining this with (4 1) and the identification already
established for n = 0, we establish identification for n = 1.
To complete our program we use (3 1) and (11 2 ) to proceed to the case
n + 1 from nand n-1.
3.
SOLUTION OF THE DIFFERENTIAL EQUATION
y"+ AXY = 0, y(O) = 0
We write J = J 1/3(~A 1/2X3/2) and shall show that y = x 1/2J is the solution.
Differentiating and substituting we find
y" + AXY = (AX3/2_h-3/2)J +~A 112J' + AX 3/2J".

By definition J 1/3(x) satisfies
X2 !1/3(X) + xJi/3(x) + (x2-b)J 1/ix) = o.
We now replace x by ~A 1I2X3/2 in this equation and solve for h 3/2J" getting
AX3/2J" = -~A 1/2J' - (h 3/2 -ix-3/2)J
which when substituted in the expression for Y" + AXY cancels out everything.
4.
SoLUTION OF THE DIFFERENTIAL EQUATION
y' =X2_y2, yeO) = 1
It has been found convenient to introduce the modified Bessel function.

For n integral
(12)
satisfies the differential equation

(13)
146
Appendix
The properties of the I's can be obtained by translation of those of the 1's
just as we obtain properties of the hyperbolic functions from those of the
trigonometric ones, or by a parallel development. Some care has to be taken
in the determination of the multiplier on the right in (12) when we deal with
cases when n is not an integer.
We show that the complete solution of the differential equation u"x 2 u =0 is
(14)
Differentiating we find, omitting the argument 1X2 of the I's,
u' =1X-1/2(CIIl/4 + ... )+ XlI2 . X(clI~/4 + ... )
and
u" = -iX-1/2(CIIl/4 + .. .)+hl/2(clI~/4 + ...)

+~xlI2(cII~/4 + ... )+ XS/2 (clI1I4 + ... )
so that
u" - x 2u = -iX- 3/2 [(1 + 4X4) (clIlI4 + .. .)-8x2(clI~/4 + ...)-4x 4(clIlI4 + ...)].
Now IlIit) satisfies
t 2y"+ ty'-<l6+t2)y = 0
and replacing t by 1X2, this becomes
The same equation is satisfied by LlI4 and so we see that u" - x 2 u = O. We

have already noted that when v is not an integer, lv are independent
solutions of (8) and Iv are consequently independent solutions of (13). This
establishes our assertion about (14).
The equation y' = X2 - y2 is of the Riccati type and can be transformed
to the linear equation u" - x 2 u = 0 by the substitution y = u' /u. From what
we have just seen the solution is
y=
Cl(Xl/2IlIih2' +Ci XlI2L l/4(1X2)'

CIXlI2Il/i1X2) + C2Xl/2 LlIix2)
where the c's will be determined by the initial condition y(O) = 1.

We now use the following transforms of (7 2):
Bessel Functions
147
in the case when n = ~ to get

y=
For x near zero
C1X 3/2 L3/4(~X2) + C2X3/2 13/4(~x2)

Cl X 1/2 Il /ih 2) + C2X1/2 L 1/4(h 2)'
Iv(x)~(x/2t/f(v+1)
so that
y ~2clf(3/4)/c2f(1/4).
Hence (Cl/C2) = f(1/4)/f(3/4) and the solution is

rmI_3/i~x2) + 2f(~)I3/4(~X2)
y = X f(~)Il/4(h2)+2f(~)Il/4(h2) .
From the NBS tables of Bessel functions of fractional order, II we

find
y
(1) = (3.625609908)(0.9800767696) + 2(1.22541 6702)(0.39858 50517)

(3.625609908)(0.81967 59660)+2(1.225416702)(1.25197 01940)
1.848449429 0 75001 5703
2.464547638'
.
5.
RELATIONS OF JACOBI AND HANSEN
We shall now establish the fact that

Jo(z) + 2J2(z) + 2Jiz) + ... = 1,
which was mentioned in Chapter 10. We do this on the basis of the series
definition (2).
For s"?m the coefficient of (iz)2s in J 2m (z) is (-l)m-s[((s-m)!)
s+m)!)]-l. Hence the coefficient of (~Z)2S in Jo(z)+2I:~lJ2m(Z) is
(
-1
s[-s!s!1 - (s-l)!(s+l)!
2
2 ]
+ + -1
... ( ) 0!(2s)!
s __
= ~;:;;
[(~S)-2C ~1) + ... +(_1)S2(~S)].
The expression [... ] can be readily identified with the expansion of

(-W[1-1]2S by replacing, for r = 1, 2, ... , s,
2(s-r
2S) b
y
(2S) (2S)
s-r + s+r .
Hence all the coefficients in Jo(z)+2I:~lJ2m(Z) vanish, except the first,

which is 1. This is what we want, but so far our work has been formal.
148
Appendix
To justify our manipulation, which is essentially the transformation of

a sum by rows of a double series into one by columns, it will be enough to
establish absolute convergence of the double series. This is easy, for changing signs in the argument of the preceding paragraph shows that the double
series is dominated by
"~ (zZ
11 1)2S (2s)!
1 (
)2s _ " IzI2s _
II
1+1 - ~(2s)!-cosh Z <00.
We shall next establish the following relation of P. A. Hansen, which
can also be used for normalization in backward recurrences:
J~(x)+2Ji(x)+2Pz(x)+ . .. = 1.
We note that for x real and r = 1, 2, ... this gives the inequalities
We begin by determining the coefficient of (-1)S(!xfs in Pn(x). It is
(-1)n
Stn r!(r + n)!(s - 1
n - r)!(s - r)!
r=O
( -1)n s-n
(S) (
= (S!)2
r~o
= (-1)n
(.!.)2(
2s )
s!
s-n
s-n-r
by a standard result, ascribed to Vandermonde. [To get a combination of

(s - n) objects from 2s objects we take r from one fixed set of s objects and
(s - n - r) from the remaining set of s objects, and do this for r =
0, 1, ... , s - n.]
To get the coefficient of (-1)S (!x )25 in the sum
J~+2Ji+ ...
we have to add, and we get, as before,
s;e 1
s=1
6.
BOUNDS ON DERIVATIVES OF
Jo(X)
It is easy to show that for x real, r = 0, 1, 2, ...
Bessel Functions
149
In fact from the representation (1), if we can differentiate under the integral
SIgn,
(-1)'
1~2r)(x) =-:;;:-
and
l~r+1)(X) =
f"
.10 sin2n 0 cos (x sin 0) dO
(_1)'+1 f"
7T .10 Sin2r + 1 0 sin (x sin 0) dO.
The stated result follows since the integrands do not exceed 1 in absolute
value.
7.
ZEROS OF BESSEL FUNCTIONS
We have made use on several occasions of the zeros of Bessel functions.

We shall state some of the results:
(15) Iv (x) has an infinite number of real zeros for any real v and, if v> -1,
all the zeros are real.
(16) Except possibly for x = 0, the zeros are simple.
(17) Denoting by iv,I> iv,2' ... the positive zeros of lv(x), arranged in ascending order, for v> -1 we have
0<iv,1 <iv+1,1 <iv,2<iv+1,2<iv,3<'"
i.e., the zeros of IV+1 are interlaced with those of Iv.
(18) Denoting by iv, i~ the least positive zeros of lv, l~ we have, when v>O,
iv>v,i~>v.
We shall establish part of (15) and (16), (17), (18). Bessel and Lommel
showed that if -!<1I:5! then lv(x) has the following signs in intervals m7T,
(m+!)7T:
!~
~ ~11'
211'
~11'
311'
rlT
4~
+.
-.
2n7T (2n + 1)7T(2n + 2)7T
It follows that lv(x), -!<V:5!, has an odd number of zeros in G7T,7T),

(~7T,
27T), .. .. We shall establish the sign pattern when v = O. We have

2 r<1I2)"
10(X)=;.Io
cos (x sint) dt.
Suppose x = (m +!0)7T where 0:5 0:51. Then if we put x sin t =!7TU, dt =

du/.J(2m+0)2- u2 we get
lo(x) =~ f 2m + 1I cos (!7TU) du .
7T Jo
J(2m+0)2- u 2
150
If we write
Appendix
2m+8
= L2 +
14 + ... +1
2m
m-2
2m 8
+
m
= -VI +V2+" . + (-l)mvm + (-l)mv:"

then it is easy to show (draw a graph of the integrand) that
and so
sign Io(x) = (-l)m.
The step from - ~:s; v:s;~ to general v is carried out as follows: Rolle's
theorem and the recurrence relations
show that between any consecutive pair of zeros of x-VIv(x) there is at least
one zero of x- vI v+1(x) and between any consecutive pair of zeros of
x v+II v+1(x) there is at least one zero of xv+IIv(x). Thus we can move up and
down from - ~:s; v :s;~.
To establish (16) we use the differential equation
Z2 Y"+zy'+(Z2_ V2)y =0
to conclude that if there was a double zero at a point x'" 0 then y"(O) and all
higher derivatives would vanish there and so y would be identically zero.
To establish (17) we use, ~ before, the recurrence relations.
To establish (18) we proceed as follows:
When v> 0 the power series of Iv and I~ show that they are positive for
small positive x. The differential equation, written in the form
shows that if x < v and Iv > 0 then xJ~ is positive and increasing and so Iv is
increasing. That is, so long as 0 < x < v both Iv and xJ~ are positive and
increasing and so cannot vanish.
8. THE Amy INTEGRAL

[Detailed references to the sources of this discussion will be found at
the end of this section.]
This is defined as
(19)
1
Ai(x) = 'Tf'
Loo cos (lt 3 + xt) dt
Bessel Functions
151
and satisfies the differential equation
y"=xy.
(20)
A direct proof of this fact, according to Hardy, "is not a particularly simple
matter". It, together with
Bi(x) =
(21)
{exp (_~t3+ xt)+sin (~t3+ xt)} dt,
has been tabulated extensively [British Association, Math. Tables, Partvolume B, 1946]. The pair Ai(x), Bi(x) are an independent pair of solutions
of the equation (20) and were expressed in terms of the Bessel functions
I1/3' J1/3 with argument ~X3/2 by Wirtinger and Nicholson. (See (25)
below.)
First we have to discuss the convergence of (19). This follows from the
fact which can be proved by integration by parts, that,
LX> Q(x)eiP(x) dx
where P, Q are polynomials, P having real coefficients, is convergent if and
only if q ~ p - 2 where q = degree Q, p = degree P. We shall discuss the
problem directly by a method which we use again.
Since 1 = (t 2+ x - x)/t 2 we have
I(b) =
f"
cos Gt 3+ xt) dt = I~ (t2 + x) cos (~t3 + xt) dt/t 2
-x
fO cos(~t3+xt)dt/f=I1+12'
say.
We apply the Second Mean Value TheoreIl}. to 11 to get, for some

B~b:
so that
152
Appendix
Clearly
Thus
and so the integral (19) is convergent and indeed uniformly convergent in

any interval [-X, Xl We shall use the uniformity at an essential stage later.
The next step is to differentiate (19) formally to get
-1T- 1
(22)
1"0 t sin (1t 3+ xt) dt
and then to show that this integral is uniformly convergent with respect to x,
so that we can apply a standard theorem on differentiation under the
integral sign. [See e.g., Titchmarsh, 59.]
Since
we have
where
11 =
1= {(t2+ x) sin (1t 3+ xt)}t-
I 2 =-x
i=
dt,
{(f+x)sin(1t 3+xt)}t- 3 dt,
13 = x31= {sin Gt3+ xt)}t-3 dt.

The Second Mean Value Theorem can be applied to the first two integrals
giving
11 = b- 1 sin GBi+ xB 1 )
12 = -xb- 3 sin GBi +xB 2)
where b :5B1> b :5B 2. The third integral is estimated trivially as
1131 :5~x3b-2.
153
Bessel Functions
Hence
and so the integral I(b) is convergent and indeed uniformly convergent with
respect to x in any interval [-X, X]. It follows, from the theorem cited,
that
Ai'(x) = -1T- 1
1
00
sin (~t3+ xt) dt.
If we differentiate formally again we are led to
tOO t2 sin (~t3 + xt) dt:

but this integrand does not converge. This follows from the fact stated
above (p. 151), or it can be shown directly as follows. If we change the variable
from t to y = ~t3 + xt then dy = (f + x) dt and since for large t, y ~ ~t3 the
integral becomes approximately
foo sin y dy
which is divergent.
To get the result we require we follow Stolz and Gibson in using a
technique of de la Vallee Poussin. We describe it formally in a general case
and then deal with the special case rigorously.
Suppose that
F(x) =
f(x, t) dt
and that fx(x, t) is continuous. Then

f(X, t)- f(O, t) =
IX fx(x, t) dx.
If we integrate across with respect to t between 0, T and let T
(23)
00
we get
[IT dt IX fAx, t) dX]

= ~ [IX dx IT fx(x, t) dt]
F(X)- F(O) = ;~
The easy case is when we can interchange limT->oo and S. .. dx to get

F(X)- F(O) =
IX dx 1fx(x, t) dt
00
Appendix
154
which when differentiated gives

F'(x) =
fx(x, t) dt.
However we are interested in the case when this inversion is not

possible and we proceed as follows. Break up the inner integral in (23) into
IT
iT I{I(x, t) dt
fx(x, t) dt = (f)(X, T)+
where
while
IT I{I(x, t) dt
is uniformly convergent with respect to x so that
~~ i
dx
iT I{I(x, t) dt = i dx IX>I{I(x, t) dt.

X
In these circumstances
F(X) - F(O) =
and, as before
F'(x) =
LX dx L I{I(x, t) dt
oo
L I{I(x, t) dt.
oo
In the special case we have f(x, t) = t sin (it 3 + xt) and we want to find
appropriate (f), 1/1 so that
IT t cosGt +xt)dt=(f)(x, T)+ iT I/I(x,t)dt.

2
We write
so that
IT t cos (it +xt) dt= iT (t +x) cos (it +xt) dt

-x iT cos (it +xt) dt = sin (iT +xT)-x iT cos (it +xt) dt.
2
155
Bessel Functions
Now as T --+ 00,
3
3
+ ILT)d IL = cos(iT )-cos(iT +x1)--+ 0
(lT3
SID 3
and the uniform convergence of the integral
IT cos
(it 3 + xt) dt
has already been established. We are therefore able to conclude that
d~ [-1T-1
t sin Gt 3 +xt) dt] = +1T-1X
cos Gt 3 +xt) dt
i.e.
Ai"(x) = xAi(x)
as required.
REFERENCES
G. H. HARDy, On certain definite integrals considered by Airy and Stokes,

Quart. J. Math. 41 (1910), 226-240 = Coll. Papers 4 (1969), 460-474.
O. STOLZ, Grundzuge der Differential- und Integral-Rechnung, B3 (1899),
Leipzig.
G. A. GmsoN, Advanced Calculus, Macmillan, London 1931,439, 452.
C. J. de la VALLEE POUSSIN, Etude des integrales a limites infinies pour
lesquelles la function sous Ie signe est continue, Ann. Soc. Scient. de
Bruxelles 16B (1892), 150-180.
E. C. TrrcHMARsH, Theory of functions, 1939. Oxford.
9.
REPRESENTATION OF
Ai(x)
AS A POWER SERIES
Let
Next, differentiating n times the equation y" - xy

(24)
= 0 we get:
y(n+2) _ xy(n) - ny(n-l) = 0
and then putting x = 0 we get

y(n+2)(0) - ny(n-l)(O) = O.
156
Appendix
We therefore find a Maclaurin series for Ai(x) of the form
(l+ ;, x +...)
ao
Ai(x) =
where
-a1(x+ 41 x 4+ ... )
3-2 / 3 /[(2/3) = 0.35503

al = r 1/ 3 /f(1/3) = 0.25882'
ao =
A real variable derivation of the expressions for ao and al follows.

Write
where n > 0, a> 0, b > O. Put a = r cos 8, b = r sin 8, rx = y, ern = u, sr n = v,

to get
u(8)== U =
v(8) == v =
1
1
00
00
e-YOOS 6yn-l cos (y sin 8)dy,
e-YOOS 6yn-l sin (y sin 8) dy.
We differentiate formally to get
du = roo e-YOOS6yn sin 8 cos (y sin 8) dy ~t
troo e-YOOS 6yn cos 8 sin (y sin8) dy.
Both these integrals are uniformly convergent with respect to 8 in -7T/2 <
-8o !5:.8 !5:.80 < 7T/2 because they are dominated by
1"
o
e-COS6oYyR dy =
r(n +1)
.
(cos 8ot+1
Hence the differentiation is legitimate.

We can write
du =
d8
roo yR ~ {e-Y oos6 sin (y sin 8)} dy
dy
so that, integrating by parts,
du = -n roo yn-l . e-YCOS 6sin(y sin 8) dy

d8
t
In the same way
dv
d8=nu
= -nv
157
Bessel Functions
and, combining, we get
so that
u(O) = A cos nO + B sin nO.
When 0=0
v (0) = 0,
so that
u(O) = r(n) cos nO,
v(O)
= r(n) sin nO.
It is clear that c, s are continuous functions of a at a

from
oo
-ax
xn
-1
= 0 and we can deduce
cos b d
r(n) cos
X x=-nO
sin
rn sin
that
n7T
coscos x dx = ____2
o x 1- n
bn
'
oo
r(n)
ooSinbX
_ r(n)sinT
~dxo X
10.
REpRESENTATION OF
Ai(x)
bn
O<l-n<l
0<1-n<2.
IN TERMS OF BESSEL FUNCfIONS
From the power series representations of J1/3(X) and Il/ix) we easily

check that if x > 0, ~ = ~X3/2, then
(25)
Ai(x) =iXl/2{Ll/i~) - 11/i~)},

Ai(-x) =ix 1/2{J -1/3(~)+ J1/3(~)}.
Chapterl
1.3. Solution
Take 8> 0 and choose a so that 8(1 + a) = 1. This implies a> 0 and so
Hence for any 8> 0, 8n < 8 if n> no(8) = (a8 tl. Thus 8n -+ O. In case 8 -+ 0 we
observe that 18 nl = 181n.
1.5. Solution
Differentiating we find
~ [lOg (l_1.)X] =~ [x log (1-1.)] = 10g(1-1.) +_1 . ~.

x
2
dx
dx
1-1. x
1
2
=-2+-3+ ... >0.
2x 3x
We use the logarithmic series and the binomial series (in the geometric case)
at *.
Similarly
1 ) +-=
1 -1 -1 - ... <0.
-d [ log (l)X-l]
1-=log (1-dx
x
x
X
2X2 3x 3
1.6. Solution
See Problem 1.13 below for the value of M(1,0.2). The results of
Gauss in the case ao =./2, bo = 1 are:
ao = 1.414213562373095048802
at = 1.20710 6781186547524401
a2 = 1.19815 6948094634295559
a3 = 1.198140234793877209083
a4 = 1.198140234735592207441
bo = 1.000000000000000 00000 0
bt = 1.18920711500272106671 7
b2 = 1.1981235214 03120122607
b3 = 1.198140234677307205798
b4 = 1.198140234735592207439
162
Note that the observed values of a,. - bn are much smaller than the
estimated values of a,. - bn
1.7. Solution
M(6000, 1200) = 6000 M(l, 0.2) = 3.12480 9828 x 103

1.B. Solution
It is well known that if A, G, H are the arithmetic, geometric and
harmonic means of two positive numbers then A > G > H with equalities if
and only if the given numbers are equal. This implies the monotony statements. Boundedness is obvious. Hence a,.-+a, bn -+(3. Since a,.+1 =
!(a,. + bn ), we have a = !(a + (3) so that (3 = a = I, say. It is also clear that
so that a,.bn = aob o and, passing to the limit, [2 = aobo. Since I> 0, it follows
that I is the geometric mean of ao, boo
To discuss the rate of convergence we note that
a,.+1 - bn +1= !(a,. + bn ) - [2a,.b.J(a,. + bn )]
= ![(a,. + bn )2-4a,.bn ]/(a,. + bn )
= (a,. - bn )2/(4a,.+I).
Thus the convergence is of the same type as that of the Gaussian
algorithm.
Observe that this algorithm suggests a method for finding square roots.
If we want to find IN then we start with
bo=N/xo.
A moment's thought reveals that this is really the familiar algorithm
a,.+1 =
~(a,. + ~)
which will be discussed in Chapter 3.

If we take a o =./2, bo = 1, then 1= 2114 = 1.18920711.
1.9. Solution
Let ao = a l , (30 = bOI and form the Gaussian sequences for ao, (30. We
assert that ~ = a;;\ (3n = b;;l. This is true for n = O. Assume that it is true
for n = r then
Chapter 1
163
and
Hence lim a,. = lim bn = [M(ao 1, bo1}r1.

1.10. Solution
[We use numerical values from the NBS Handbook.]
a} Borchardt
ao=0.2,
(30 = 1,
cos 6 =0.2,
sin 6 = ../0.96 = 0.97979 590.
From NBS Handbook p. 169:

cos 1.369 = 0.20042 95292,
sin 1.369 = 0.9797081218.
If
0.2= cos (1.369+e) = cos 1.369- e sin 1.369-!e 2 cos 1.369+ ...
then a first approximation to e is
e1
= 0.0004295292 . 0 00043 8
sin 1.369
..
.
We now compute
cos (1.36943 8) = cos 1.369
- e1 sin
= 0.2004295292
1.369 - 0.0004291122
-!eicos1.369-
191
= 0.20000 03979.
The next approximation is given by
e2
= 0.000000 3979 = 0 00000 041

sin 1.36943 8 '
.
We take
6 = 1.36943 841
and then
I = (sin 6/6) = 0.715472776.
(30 = 1,
cosh t =..J2,
From NBS Handbook p. 22: t = 0.881373587 and
I = sinh t = 1.13459266.
t
sinh t= 1.
164
b) Carlson.
(Q.96 _ ~0.48 _ I
0.48
- V2lOg5 - log 5 - V1.60943 79124
1-
bo =0.2:
= 0.546114246.
bo = 1:
ao=..fi,
I = V 210: J2 =
VIO~ 2 = VO.6931~ 71806

= 1.20112240.
[The logarithms are taken from the NBS Handbook, p. 2.]

1.11. Solution
The direct treatment is the following. Take the case when Ao = cos2 8,
Bo = 1. Then induction gives
A..+l = cos 8 cos (2- 1 8) ... cos (Tn8) cos2 (2- n- 1 8),
Bn+1 = cos 8 cos (2- 1 8) ... cos (2- n8).
As before
2n+ 1 sin (2- n 8)Bn + 1 = sin 28
and so
lim Bn = sin 28/[lim 2n+ 1 sin (2- n8)] = sin 28/28.
1.12. Solution
Since (a,J f3n) = cos 6/2n we have
f(n) = 2n arc cos (a,J~n) = 8
and
g(n) = 2n(~~-a~l/2 = 2n cos (2- 1 6) ... (cos 2-n8)[1-cos2 2-n8]112
=
2n~n
sin 2-n6
= sin 8
so that !(n), g(n) are independent of n.
Next we note that
f(n) arcsin ~~-a~)112/~n)
g(n) = ~n~~-a~)112/~n)
Since an ~ land
~n ~ l
so that I = g(O)/f(O).
we have
Chapter 1
165
1.14. Solution
Since K' depends on k through k 2 some tables have m = k 2 as the
argument. Then
K'(m) = 1'71"/2 (1-(1-m)sin 2
0t 1/2 dO
{(1-t 2)(1-(1-m)f)}-1/2dt.
Then k = 0.2 corresponds to m = 0.04 and we find from NBS Handbook, p.

608
K' = 3.01611 24925
and
1.57079 63268
M(1, 0.2) = 3.0161124925
0.52080 1638.
1.15. Solution
What is involved here is essentially the relation
lim 1'71"/2 In (0) dO =
r/\imfn(t~) d6
when fn(6)
=(a~ cos2 0 + b~ sin2 Ot 1/2 .
Although this relation is valid in the present case it is not in general:

e.g., if for n = 0, 1, 2, ...
gn(O)=O for
gn (2- n - 2 '7T)
6=0, andfor 2- n - 1'77":50:52- 1'77"
= 2n + 2 '7T- 1
gn linear in 0, 2- n - 2 '77" and in 2- n - 2 '77", 2- n - 1 '77".
See diagram. In this case
t'IT/2Iim gn(O) dO = O.
For further remarks on similar problems see Chapter 5. Since
it is clear that
an-1<{.
- n (1I)<b-1
11 n
166
b~t-----==;;;;;;o;o-...,
M-l~
a~l~V____----l
O'--------!---
ln
9
2
Also
It is also clear (see diagram) that for any n
Suppose the common value of the integrals S;;/2 In (6) d6 was different from
Eo. This supposition leads to a
contradiction in the following way:
Since a..l M, bn t M we have Ib;;-1-a;;-11'IT12a.. -bJ'lT12b~. Since
a,. - bn ~ 0 we can find an no such that for n;a.: no,
S;;/2 d61M and that they differ by, say,
(a,. -bnhT/2b~<IEol.
The last two displayed formulae are in contradiction. Thus we have established the relation required.
1.16. Solution
a) If A is the common limit of {a,.} and {gn} then A1JX is the common
limit of the Borchardt sequences with initial values (1 + x)/2JX, 1. The limit
in this case is sinh tIt where cosh t = (1 + x)/2JX. It is clear that sinh t =
v'(cosh2 t-l) = (1-x)/2JX and that
exp t = cosh t + sinh t = 1!JX,
i.e., x = exp (- 2t)
so that - 2t = log x. Thus

A
i-x
-2
Jx= 2Jx 'logx'
i.e.,
log ..fi =! log 2 = 0.346573 5903.
x-l
x-l
A = - or logx=--.
log x
A
167
Chapter 1
The sequence {a,. - g,.} converges to zero geometrically, with approximate ratio 4 whereas the sequence {:l(a,. +2gn )} converges to its limit
geometrically with approximate ratio 16.
b) If B is the common limit of {a,.} and {gn} then BJ../(1 + x 2 ) is the
common limit of the Borchardt sequences with initial values l/.J(1 + x 2 ), 1.
Hence
B
so that
B=
x
x
.
arctan x' I.e., arctan x = B'
arctan 1 =i1T = 0.78539816.

[B. C. Carlson, Math. Comp. 26 (1972), 543-549 gives techniques to
accelerate the convergence of these sequences. See also his paper, Amer.
Math. Monthly 79 (1972), 615-618.]
1.17. Solution
Clearly
These give
a,.+1 - bn+1 _ !(a,. - bn ) = a,. - bn

log (a,.+1/bn+l) -! log (a,./bn) log (a,./bnr
Hence
a,. -bn
ao-b o
log (a,./bn) log (ao/bor
Now, by the Mean Value Theorem,

log (a,./bn) = log a,. -log bn = (a,. - bn)/c,., for some c,. between a,., bn
Hence
Let
n~ao
and we have
168
1.18. Solution
Clearly x..+l lies between
also between x.., Yn. Further
x.., Yn
and Yn+l lies between
x.., x..+l
and so
Hence if x.. ~ Yn then x..+1 s.: Yn+l. It follows that the odd terms and the
even terms of each sequence are monotonic and bounded and therefore
have limits. The relation established above ensures the equality of all the
limits and that as n ~oo
The argument used in Problem 1.13 shows that
1
00
(t+ X~)-3/4(t+ y~)-1I2 dt =
L+
oo
(s
L+
oo
(s
Ft 3/4(S + F)-1/2 ds
12 )-5/4 ds
= [-4(s + F)-1I4J;' = 41- 1/2.
Changing the variable from T to T by T = (1 +
1
00
(T+1)-3/4y--1I2 dT=4
n- 1I4 shows that
(1-T 4)-1I2 dT.
Hence 1(1,0) =A -2 where A = J~ (1_T 4)-1I2 dT = 1.31102877714605987.

1.19. Solution
We find:
cos qJ = [1- (1 + k') sin2 8](1- k 2 sin2 (J}-I12

(1-k~ sin2 cp)1I2=[1-(1-k') sin2 8](1-Psin2 8)-112.
Differentiating the expression for sin cp gives
cos qJ(dqJ/d8)
= (1 + k')[(cos2 8 -sin2 8)(1- k 2 sin2 8) + k 2 sin2 8 cos2 8](1- P

Combining these three relations gives
dqJ(1- k~ sin2 qJ )-112 = (1 + k') d8(1- P sin2 8)-112.
sin2 8)-3/2.
Chapter 1
169
As 0 increases from 0 to !?T the numerator in sin <p increases from 0 to

!(1 + k') and then decreases to zero while the denominator decreases from 1
to k'. Hence <p increases from 0 to ?T. Since
{1' d<p(I- ki sin2 <p t
1l2
= 211'/2 d<p(I- ki sin2 <p )-112
the result follows.

This result is due to J. Landen (1719-1790). From it we can recover
Theorem 1 by taking P=I-(b~/a~) which gives k'2=bMa~ and k 1 =
(ao-bo)/(ao+b o) so that
i.e.,
r1'/2 dO(a~ cos2 (I + b~ sin2 0)-112 =.10r 2d<p(ai cos2 <p + bi sin2 <p )-112.
l '/
Jo
Repeating this indefinitely we find
1.20. Solution
This problem derives from J. Schwab, Elemens de geometrie, [ere

partie, Geometrie plane. Nancy, 1813.
I am indebted to I. J. Schoenberg for sending me copies of parts of
Schwab's book and for suggesting that I record the fact that the so-called
"Borchardt" - Algorithm was known to Schwab before Borchardt (18171880) was born. One sector of a regular n-gon is indicated below. It is
clear that
an = (2n)-1 cot (?TIn),
bn = (2n)-1 cosec (?TIn).

The relations (1) follow by elementary trigonometry.
170
The convergence of the sequences (2) follow from our discussion of

"Borchardt's Algorithm" and the limit is
(2n)-1 cosec (1T/n){[sin (1T/n)]J( 1T/n)} = 1/(21T).
It is clear "geometrically" that the common limit is the radius of a circle

with circumference 1.
Schwab gives 7D values in the case of polygons of 6, 12,24, ... ,6144
sides. With our normalization
a6144 =
b6144 = 0.15915 4950
0.159154931,
while
(21T)-1 = 0.159154943.
In the case n = 2, a2 = 0, b2 = 0.25.

Compare also Problem 9.7.
Chapter 2
2.1. Solution
Let f be a positive decreasing function defined for all x ~ 1. Then the
sequence {sn}, where Sn = L~=l f(r), and the integral j f(x) dx converge or
diverge together. The "infinite integral" .17 f(x) dx is said to converge or
diverge according as limx.-oo F(X) exists or not, where
F(X) =
Since
f(r)~f(x)
IX f(x) dx.
when r:5x:5r+1 we have, for
e(n) = f(l) + f(2) + ... + f(n -1) -
f(x) dx = ~t:
n~2,
1'+1 (f(r) - f(x
dx ~ O.
171
Chapter 2
Rewriting this as
f(1)+f(2)+ . +f(n - 1);:::
f(x)dx
and passing to the limit, in the convergence case we find
f f(r)
1
2!
f oo f(x) dx.
1
Moreover, e(n+1)-e(n)=S~ + I(f(n)-f(x))dX2!O, so that the sequence {e(n)} is an increasing one. Since
e(n) = f(1) -
J: 1'+1
(f(x) - f(r + 1)) dx -
I~1
f(x) dx
we see that e(n)::5f(1). Hence en--,;l, 0::5I::5f(1). See diagram.

y
n +1
The sum of the shaded areas to the right of the ordinate through 1 is
manifestly positive; translating them to the left as indicated, the sum is
manifestly less than f(1).
It follows in the case of convergence, that
rf(n)::5f(l)+f f(x)dx.
OO
172
We have, therefore, obtained both upper and lower bounds for the
infinite series in terms of the infinite integral.
2.2. Solution
We discuss
we have
rn-
From graphical considerations, as in Integral Test,
1/ 2
Also,
SN-l:51
X- 1 / 2 dx=2N 1/ 2-2,
i.e.,
~:52N-l/2_1.
These two relations show that SN = (J(Nl/ 2 ) exactly. [Compare Problem 2.5
below.]
The first of these relations gives
SN> 10 if 2(N + 1)1/2> 12, i.e., if (N + 1) > 36, i.e., if N> 35.
Actually S32=9.94, S33= 10.11 so that the least number of terms

required is 33.
The solution to the first part is 20, that to the third is 12367 and that to
the fourth is approximately 1.6 x 104321 .
For further examples of this type see:
G. H. Hardy, Orders of infinity, Cambridge, 1924.
R. P. Boas, Jr., and J. W. Wrench, Jr., Partial sums of the harmonic
series, Amer. Math. Monthly 78 (1971) 864-870.
R. P. Boas, Jr., Growth of partial sums of divergent series, Math.
Compo 31 (1977), 257-264.
R. P. Boas, Jr., Partial sums of infinite series and how they grow, Amer.
Math. Monthly 84 (1977), 237-258.
2.3. Solution
All the series converge. We can use results obtained in the solution of
Problem 2.1. For instance, in the case of n- 101/l(lO, we have
100 =
r X-101/100 dx:5 f n- 101/100:51 + Jr X- 10l/1OO dx = 1 +[-100x-1I1oor

J
cc
cc
=101.
Similarly
cc
10:5L n- 11/ 1O :511,

1
cc
1/9:5L n- 1:510/9,
1
Chapter 2
173
Actually
Ln00
Ln00
lO
= 1.00099 4575,
= (1/6)'1T 2 = 1.64493 40668.
2.4. Solution
In the notation of Problem 2.2 we have
1
fa(n) =-e(n)+--n-a
I-a
and, letting
n~oo,
we find
limfa(n)=-I+-11 .
-a
When a
= 1 we have
{log n -(1 + 1/2+ ... + l/n)}~-'Y = -0.5772156649.

Generally, lim fa (n) = -'(a), , being the Riemann Zeta-function.
Here are the first 4 values of fl/2(n).
n
2h- (1 + ~) = 1.12132034
2J3-(1+
4- (1 +
~+ ~)=1.1796446
J"z+ ~+~) = 1.21554295
The limit is -,(1/2) = 1.46035 45088.

See T. M. Apostol, Calculus II, p. 618-9, Exercises 15.23 #3.
2.5. Solution
This result states that the remainder after n terms of a series I ,-i-a is
exactly of order a. We use again facts established in the solution to Problem
2.1.
Let f(x) = (n -1 + x)-l-a. Then f(x) decreases as x increases. Further,
as X~OO.
x
[(n-l+X)-a]x n-a
F(X)= 1 f(x)dx=
(-a)
1 ~~,
174
so that I~ f(r) is convergent and lies between

h f(x) dx. Thus
n-a
a
-::5
L f(r) = L r00
00
r=1
r=n
1- a
n-a
a
h f(x) dx
and n- 1 - a +
(1 1)
+- = n-a -+a
so that we can take
2.6. Solution
The first series is obtained by expanding the integrand in the lemniscate
integral S~ (1- x4t1l2 dx by the binomial theorem and integrating term-byterm.
From Stirling's Formula [po 111] we find that the general term an
is approximately (l/47T)n- 3/2 It follows that the error after n-terms is
0'(n- 1I2 ) so that direct summation is unfeasible.
The second series is also obtained from the lemniscate integral by
writing the integrand as (1- X2)-1I2(1 + x 2 tl/2 and expanding the second
factor by the binomial theorem and the integrating term-by-term using the
fact that
1 2n(1_ 2)-1I2d =(2n-1)!!~

ott
t
(2n)!! 2
Again using Stirling's Formula, we find that bn is approximately (2n)-I.

The series I (-1tb n is an alternating one ({bn } is monotone) and the error in
such a series is less in absolute value than the first term omitted.
See Problem 9.10.
2.7. Solution
Denote the limit by C. Then we are concerned with the ultimate
behavior of the ratio (a,. - C)/(a,.+1 - C).
We may assume a,. = 1 and bn = x = 1 - y, where y > 0 is small. Then
2y_y2
=
1-(y/2)
C 2 = 1-x 2 =
2 log X-I 2y +y2+(2y3/3)+... 1 +(y/2)+(y3/3)+ .. .
= 1-y +(y2/6)+ ... .

Hence
C = 1- (y/2) - (y2/24) + ....

Since
an +l = J(1 + x)/2 = J1- y/2 = 1-(y/4)-(y2/32)+ ...
175
Chapter 2
we have
a.. - C
(y/2) + (y2/24) + .... 1/2

(y/4) + (y2/96) + ....
.
a..+1- C
2.8. Solution
This follows by induction from the fact that if f(x) = o(h(x)) and
g(x)=o(h(x)) then f(x)+g(x)=o(h(x)) and the fact that xr=o(e as
X )
x~oo.
Alternatively, if Pr(x)=aO+a1x+ ... +a.x r, note that if
x~1
and use the fact that if f(x) = o(h(x)) then Af(x) = o(h(x)), for any constant A.
2.9. Solution
f(x)=
L fk)(a)(x-a)k/k!+E,.(x)
n
k=O
where
1E,.(x)l<
Mlx-al
n +1
(n+1)!
This can be written as
or as
2.11. Solution
Since
d {(
t)n}
dt e t 1-~
te
=--;
t
('
t)n-l
1-~
if t increases from 0 to x (where x:::; n) we have
176
SolutiQns to Selected Problems
Integrating across with respect to t between 0, x give
Dividing across by eX gives
This is result required: if n ~ x then
Chapter 3
3.2. Solution
If {x,,} converges to a limit l;t: 0 then I =!(l + Nil) so that 1= IN.
Graphical considerations suggest that there will be convergence to IN if Xo
is positive and to -IN if Xo is negative. It will be enough to discuss the first
case. Further, Bn+1 = !B~X~1 so that convergence is quadratic. If Xo < IN then
x1>JNand we find X1>X2> ... >IN. The sequence {x,,}n;;,,1 is therefore a
bounded decreasing one and its limit is necessarily IN.
3.3. Solution
Deal with this as in Problem 3.4 below, or transform it algebraically
into Problem 3.4. Quadratic convergence to N 1/ 2 takes place for appropriately chosen Yo since
3.4. Solution
For further details see W. G. Hwang and John Todd, J. Approx. Theory
9 (1973), 299-306. See diagram opposite.
When 0 < Xo < (3N)1/2 the sequence converges quadratically to N. When
xo> (SN)1/2 the sequence oscillates infinitely. There is an increasing sequence 13r with 13-1 = (3N)1I2 which converges to y = (SN)1I2 and is such that
when 13r<XO<f3r+1 the sequence {x,,} converges to (-1YN1/2. For Xo=
0, 13-10 130, ... the sequence converges to O. For Xo = Y the sequence oscillates: x" = (-1)ny. The behavior for negative Xo is obtained by symmetry.
Chapter 3
177
3.5. Solution
If we use the relation (8) of 3.2 the only division we need to do is by 6
which can be done mentally. We have
X1-m=
-(xo-!N'f(xo+2JN)/(2N)
and so if we choose Xo so that
Ixo-.J31 < 10~8 we have
IX 1 - J3I:::; 1O- 16(3J3+ 10-8 )/6< 10- 16 .

Preliminary calculations e.g., the use of the above relation with Xo = 1.732,
which is the usual approximation to J3 which one remembers, suggests a
choice of
Xo = 1.73205 081.
We compute (exactly)
x~ = 3.00000 00084 21656 1
[xo-1O- 8 ]Z= 3-0.00000 002621936
which guarantees xo-J3 < 10-8 We can write the recurrence relation in the
178
form
N-X~]
Xl=XO [ 1+---zN
and complete our calculation as follows:
From our computation of x~ we find
(3- x~)/6 = -0.00000 00014 03609 3

and multiplying by
Xo
we get
xo(3-x~)/6=-0.00000 00024 31122 62.
Adding
Xo
to the last line gives
J3 =
1.73205 08075 68877 38.
O. Emersleben gives 29(353) in place of the last two figures 38 so that our
result appears just good to 15D. NBS Handbook (p. 2) gives 29(35).
3.6. Solution
We find, by elementary algebra, when x,.+1 =~X(x,.)-!Y(x,.) that
x,.+l-vW=~B<x,. + NX~1)_JN]-M{2x!,(3x~- N)}-JN]
3
1
= 4x,. (x,. _IN)2 2(3x~- N) (x,. -JNf(2x,. +JM
= (x,. - IN)3(5x,. + 3JN)/4x,. (3x~ - N)
m.
so that this sequence has cubic convergence to

In a similar way, when x,.+1 =!Y(x,.)+!Z(x,.) we obtain
x,.+1- vW=![{2x!'(3x 2- N)}-JN] +![x,. (3N - x~/2N}-JN]
(x,. -IN)2(2x,. +IN) (x,. _IN)2(x,. +2JN)

2(3x~-N)
4N
= -(x,. -IN)3(3x~+9x,. vW+4N)/(4N(3x~- N),
again showing cubic convergence to /N.
These examples illustrate how cubically convergent schemes can be
obtained by combining two quadratic schemes. Similarly, by combining
these two cubically convergent schemes we can get one with quartic convergence.
Chapter 3
179
3.7. Solution
Draw the graph of y = x[4- Nx 3 ]/3. It is obvious geometrically that if
0<XO<4 113 N- 1I3 the sequence will converge quadratically to N- 1/3 . To
prove this we note the identity
x,,+1- N- 1I3 =
_(3N)-1[x,. -
N-1/3]2[X~+2xN-1I3+3N-2/3].
If xo=O or xo=4113N-1/3 there will be convergence to zero. If xo<O or

Xo > 4 1/3N- l {3 there will be divergence to -00.
[1.2345678]-113 = 0.93216 97773.

3.B. Solution
If lim Xp = I exists then [(I, l) = I and this gives In = a, I = a lIn. Observe
that [(x, y) = [(y, x). Since
(1)
it follows that if x = I, [(x, y) = I for any y. It follows that if Xo or Xl is I then
X2 = X3 = ... = I. We have to discuss separately the cases when xo:5 Xl < I,
Xo I < xo:5 Xl. For instance, take the case when l < xo:5 Xl. It is clear
from (1) that if x> I then [(x, y) increases with y, X being fixed: since
[(x, l) = I we have [(x, y) > l if y> ,. Hence X2> ,. It also follows from (1)
that X2 < Xo. Proceeding we see that {x,.} is a decreasing sequence, bounded
below by ,. It must converge to I.
For further details see Amer. Math. Monthly, 55 (1948) Problem
E 790, p. 367.
3.9. Solution
If 1= a lln then
[(x, y)-l = (x-l)(y -l)g(x, y)
where g(x, y) ~ 0 for x = I, y = I. Let 0 = 1.618 be the positive root of the

quadratic q2 - q -1 = O. Then 0 is the order of convergence. If
xp -l=O'(e r. )
then the above equation shows that Tp+2 = Tp+1 + Tp and the solution to this
Fibonacci difference equation (d. Chapter 10) is Tp = AOP + BO-P ~ AOP.
3.10. Solution
In case i) show that Tn decreases steadily to ../N.
In case ii) show that T2n decreases steadily to ../N - this means that we
can obtain upper and lower bounds for ../N.
For further information on this problem, and extensions to deal with
the solution of quadratics and the determination of cube and other roots, see
180
A. N. Khovanskii. The application of continued fractions and their generalizations to problems in approximation theory, Noordhoff, 1963.
3.11. Solution
This problem is fully discussed theoretically by K. E. Hirst, J. London
Math. Soc. {2} 6 (1972), 56-60 and earlier in E. T. Whittaker and G.
Robinson, The Calculus of Obseroations, Blackie, London {4}, 1944, p. 79.
3.13. Solution
Suppose xo> O. Then
x.. is clearly positive for all nand

x..+l- x.. = 2x..(N - x~)/(3x~+ N) ~O
according as
x.. ~.JN.
Also
x..+l -./N= (x.. -JF:ry3/(3x 2 + N) ~O

according as x.. ~./N. Consequently if 0 < xo.JN we have x.. t and x.. :5.JN
while if .IN< Xo we have x..! and x.. ::::.IN. Convergence takes place for any
xo> 0 and the limit satisfies
(31 2 + N)l = x 3 + 3Nl i.e.,
P = Nl i.e., 1= 0, .IN
and so 1=.JN. We have x..+l-.JN=(x.._JN)3/(3x~+N) so that convergence is cubic.

The behavior of this sequence can be illustrated graphically in the usual
way. Observe that if y = (x 3 + 3Nx )/(3x 2 + N) then y - 3x as x - 0 and
y - xl3 as x~oo. Also y' = 3[(x 2 - N)/(3x 2 + N)Y::::O for all x and y' =0 for
x = N1I2
y
Y'x
y-3x
yt x
3 x
3.16. Solution
This is not entirely trivial. See e.g., J. L. Blue, ACM Trans. Math.
Software 4 (1978), 15-23.
181
Chapter 4
Chapter 4
4.1. Solution
-0.882027570.
4.2. Solution
0.67434714 i 1.1978487. the remaining root is -2.
4.3. Solution
If q(x)=qox n- 2+ ... +qn-3X+qn-2 and r(x)=qn-1X+qn then
= 0, 1, 2, ... , n -
2.
4.4. Solution
We take the case of a double root. Then f'(~) = 0 but f'(~) ' O. We find
Xn+1 - ~ = !(Xn -~) + O'(Xn - ~f
and we have linear convergence, not quadratic. We can, however, restore

the quadratic convergence by changing the Newton formula to
in the case of a double root and to
in the case of a root of multiplicity r.

4.5. Solution
We show that the conditions (8), (9), (10), (11) are satisfied. First
(8)
H(-I) = 26>0
and H( -1-~6)= -6-~63<0.
Next,
(9)
H'(x)=9x 2-3>0 if
Ixl>l/.J3.
Then
(10)
H"(x) = 18x <0 if x <0.
Finally the tangent to the curve y = H(x) at (-1, 28) has equation
y-28
x+l =6, i.e.,
y =6x+6+26
and cuts the x-axis at -1- 6/3, so that (11) is satisfied.
182
Thus the Newton process converges. For instance taking 0 = 1 we

begin:
xo=-1
-(4/3)
-1.3333
X2 = -(148/117)
-1.2650
-1.2406
Xl =
lim x"
= -1.2600
See Problem 3.4 and reference there given.

4.6. Solution
Suppose we have
f(x)
= (x -a)q(x)+ r.
Then f(a)=r and since f'(x)=q(x)+(x-a)q'(x) we have f'(a)=q(a).

We have to show that q(x)=foxn-l+ ... +fn_l and that r=fn- We
multiply to get
(x -a)q(x) = foxn +(fl -afo)x n- 1+ . .. + (fn-l -afn-2)X +(fn -afn-l)- fn
and then the recurrence relations defining the
t give
(x -a)q(x) = aoxn +aIX n- 1+ ... +an-1x +~ - fn
i.e.,
(x - a)q(x)+ fn = f(x).
4.7. Solution
0.77091 70,
-0.3854585 1.56388 45i
4.8. Solution
See M. Davies and B. Dawson, Numer. Math. 24 (1975), 133-135,
George H. Brown, Jr., Amer. Math. Monthly 84 (1977), 726-727.
We may assume the zero at ~ = 0 and we use Maclaurin series for
convenience instead of the Taylor series used on page 49. We assume f4}(X)
is bounded in the neighborhood of the origin. Then
f(x)
= xf'(O) + x 2 f'(0)/2! + x 3 f"'(0)/3! +0'(x 4 )
f'(x) = f'(0) + xr(O) + x 2 f"'(0)/2! + 0'(x 3 )

r(x) = r(O) + xf"'(O) + 0'(x 2)
183
Chapter 5
and the correction term is, suppressing the arguments,
x [21' + xf' +lx 2f'" + O(x 3 )][f' + xf' +!x 2f'" + O(x 3 )]
2[f' + xf' +!x 2f'" + O(X 3 )]2- x[f' +!xf' +ix 2f'" + O(x 3 )][f' + xf'" + O(x 2)]
X[2/,2 + 3x/,f' + x2(f12+1/'n + O(x 3 )]
21'2 + 3xf'f' + X2(~f'I2+ f'f'~ + O(x 3 )
The new estimate is therefore
showing that convergence is cubic.

If 1'(0) = 0 then the new estimate is of the form
~x +O(X2)
showing that convergence is linear.

4.9. Solution
4.4934094579, 7.7252518369.
a = 1, b =~, c = U.
4.10. Solution
f(a)=-l,
f'(x)
f(b)=!
= -4x ~O
in [a, b]
f"(x) = -4::;;0 in [a, b]
The tangent to the curve at x = b has equation y = 2x +~ and cuts the x-axis
at x = -1, between a and b.
Thus all the conditions (8) (9) (10) (11) are satisfied.
If xo=-0.8 then Xl =-0.7125<~=-2-l/2=-0.7071 while if Xo=
-0.6>~ then Xl = -0.7142<~.
Chapter 5
5.4 Solution
Suppose lim M,. = O. Take any E> 0 and then no = no( E) such that
Mn ::;; E if n ~ no; this implies that Irn (x)\::;; E for all x and all n ~ no. Hence
no(E, x)::;; no and so no(E) is finite.
On the other hand, suppose no( E) < 00 for all E > O.
If Mn ~ 0 is false there is an E > 0 such that for every N (however large)
there is an n" > N such that M... > s. Choose N = no(E). Then \rn(x)\ < E for
n > N and so
M.... ::;; E,
a contradiction.
184
5.5. Solution
The solutions are 1; x; x 2 +[x(1-x)]/n and x 3 +[3x 2 (1-x)]/n+
[x(1- x)(2- x)]/n 2 We shall establish the result in the cases k = 1 and k = 2.
In the case k = 1 we have
L (;)xr(1- x)n-r (;) = L (; =Dx r(1- xt- r

=
xL (nr-1
-1)x r- (1_ x)(n-l)-(r-l)
1
x[x + (1- x)]n-l
=x.
The general term in the case k = 2 which is
( n) xr(1- x)n-r ~2
r
can be split into two by cancelling out an r and an n as in the case k = 1 and
then writing r=(r-1)+1. We find
L (;) x r(1_x)n-l
(;Y n:
=
L (;=~) x r(1-xt- r
+1.
n
n-1
=-
x2
L (nr-1
-1) x r(1- xt-r
L (n-2) x r- 2 (1- x)n-r

r-2
+-x L (n-1) x
n
n-1
n
r-1
r-
1 (1-
xt- r
x
n
=--x 2 +2
x(1-x)
=x +--'--n
as stated.
5.6. Solution
ql(X) is the Bernstein polynomial, adjusted to the interval [-1,1] and
obtained in Problem 5.8 below.
qz(x) is (approximately) the polynomial of best approximation, obtained
by Remez.
q3(X) is the truncated Legendre expansion.
qix) is obtained by truncating the Chebyshev expansion obtained in
Problem 5.11 below.
185
Chapter 5
In the evaluation of lie; (x )11 we can confine our attention to O:s; x :s; 1.
The value of \iel(X)\\ is shown to be 0.375 in Problem 5.8. It is less easy to find
the other norms exactly and they can be estimated as
max \e;(x)\,
x =0(0.01)1
which gives
l\eix)1\ = 0.067621,
l\eix)l\=
Remez gives Ile2(x)II+0.067621 and this is assumed positively at x =

1, x90.28 and negatively at 0 and at x90.78 exemplifying the equalripple behavior characteristic of the polynomial of best approximation.
5.7. Solution
a) S =J292{ -!+(1-ee i6 +e 2 e 2i6 -
)}
= J292{ -!+ (1 + eeilltl}

=J2{!-!e 2 }{1 +2e cos 0 +e 2}-1
2
3+cos 0
1
1+x
b) e,.(x) =
{1!X
-1Tn
(X)}
J2
3+cosO
{!(1- e
2) -
(-l)e"[cos n6 + e cos (n -1)6]

1+e 2+2ecos6
+(-1)" en COS nO}
(1-e 2 )
(-l)"J2e n
= 2e(3+cos 0) [cos nO+e cos (n -1)6]-(-1)n
e"-l cos n6
4
(-1)"e"-1
J2
[4 cos nO+4e cos (n-1)6-3J2cos nO
4 2(3+cosO)
- J2 cos 0 COS nO]
= (-1)n- 1 e" x {COS (n + 1)6 +2e cos nO +e 2 COS (n -1)6}.

4
1 + 2e COS 0 + e2
186
The result we need can be established by elementary trigonometry. For

it is easy to verify that
cos (n + 1)6 + 2e cos n6 + e 2 cos (n -1)6
1+2ecos6+e 2
cos (n6+c/
where
cosc/>=3cos6+1
3 + cos 6 '
..I.. _2.J2sin 6
sm ' f I - 3 + cos 6
It is now clear that

max 1(1 +xt l -1Tn(x)1 =kn.
The fact that the error e,. (x) has n + 2 extrema, alternately len
implies, because of the fundamental equal-ripple theorem of Chebyshev,
that 1Tn(x) is the polynomial of degree n of best approximation to (1 + X)-l
in [0,1]. Note that it is obtained by taking the truncated Chebyshev
expansion of (1 + X)-l and dividing the term involving Tn(x) by (1-e 2 ).
c) tl consists of the first two terms of the series in a) and the maximum
error is
5
t2 is 1Tl
-W98 * 0.0502.
and the maximum error is ie * 0.0429 which is assumed, with
alternating signs at 0, J2-l.

When n = 6, the error in using 1Tn{x) is ie 6 =l= 6.4 x 10-6 ; if we did not
modify T6 the error is
and this is in absolute value less than
d) B l =ll(l-x)+!lx=l-tx.
max 1(1 +X)-l- Bli =~-h*0.0858.

e) 6, 8, 10, 13, 18, 23, 33, 53, 110.
187
Chapter 5
5.8. Solution
Since
B4 = (1- x 4)f(0) +4x(1- x)3f(1/4) +6x 2(1- x)2f(1/2)

+4x 3(1- x)f(3/4) + x 4f(1)
we have, using x = (1/2)(1 + y), 1- x = (1/2)(1- y),
16B! = (1- y)4f(0) +4(1- y2)(1- y)2f(1/4) +6(1- y2ff(1/2)

+ 4(1 + Y)2(1- y2)f(3/4) + (1 + y)4f(1)
so that
16B! = [[(0) -4f(1/4) + 6f(1/2) -4f(3/4) + f(1)]y4

+ [-4f(0) + 8f(1/4) - 8[(3/4) + 4[(l)] y3
+[6[(0) -12f(1/2) +6[(I)]y2

+ [-4[(0) - 8[(1/4) + 8[(3/4) + 4f(l)]y
+ [[(0) + 4[(1/4) + 6[(1/2) + 4[(3/4) + [(1)].
In the special case of Iyl we have [(1/2) = 0, [(1/4) = [(3/4) = 1/2, [(0) =
[(1) = 1 which gives
B~(IYI, y)= _h4+~y2+i.
Consider the difference e *(y ) = 1y 1- B ~ in the range [0, 1]. Its derivative
is
h 3 -h + 1 =~(y -1)2(y +2)
which is positive in [0,1]. The error e*(O) =-3/8, increasing to e*(1)=0.
5.9. Solution
We deal with
r2,1
first. Equating coefficients of 1, x, x 2, x 3 in
we find
ao= (30
a1 = (31 + (30
o=~(30+ (31 + (32
!
.
0= t(3o + ~(31 + (32

The last two equations are homogeneous in (30, (31) (32 and therefore
have a non-trivial solution:
188
Substituting in the first two we get
and so we find
r2,l(x) = (6+2x)/(6-4x+ x 2),
The general case is handled in the same way. The IL equations arising
from tV +1 , ,t"'+v involve the IL + 1 coefficients 130,"" 13", and we can
find a non-trivial solution. Putting these in the equations arising from
1, t, ... ,tV gives the <xo, ,<Xv directly.
Observe that if XX divides D so that fJO=fJl= ... =fJX-l=O then XX
divides N for <Xo = ... = <XX-l = O. Thus we can divide out by XX and
normalize with the constant term in the denominator unity.
Observe also that the rational function NID is necessarily unique. To
see this suppose that NID and NID are IL, v entries in the Pade table for
f(x). Then Df-N and Dt-N each involve powers of X larger than IL+V.
Hence E = ND - ND involves powers of x larger than IL + v but E is of
degree IL+V at most: hence E=O and so (NID) = (NrD).
We can therefore assume NID irreducible and, say, D(O) = 1. This will
give uniqueness.
In some cases the general Pade fraction can be written down explicitly.
see e.g., Y. L. Luke (The special functions and their approximations, I, II
Academic Press, New York, 1969) who gives, e.g., information on (1 +
Z-1)1I2, (1 + Z-1)1/3, Z 10 (1 + Z-l), Z-l arctan z.
1,
1+x,
x2
1+x+ Z '
1
,
1-x
2+x
,
2-x
6+4x+x 2
,
6-2x
2
6+2x
2-2x+X2' 6-4x+X2'
12+6x+x 2
12-6x+x 2'
5.11. Solution
a) Suppose lcos 01 = tao + L a.. cos nO. If we multiply across by cos mO
and integrate between -7T, 7T we get, on the right, 7Ta",. Similarly, on the left,
we get
1m =
t:
COS
mOlcos 01 dO = 2
=2
fT
COS
,"/2
mOlcos 01 dO
COS
mO COS 0 dO - 2
I'"
,"/2
COS
mO COS 0 dO.
Chapter 5
189
Changing the variable to cp = 7T - 6 in the last integral gives
C'II"/2
1m =2 ~
[1 +(-lr] cos m6 cos 6d6.
If m is odd, 1m = O. If m = 0, 10 = 4 S~/2'11" cos 6 d6 = 4. If m = 2n
12n = 2 {'II"/2 [cos (2n + 1)6 + cos (2n -1)6] d6
= 2[(sin (2n + 1)7T/2)(2n + 1)-1 + (sin (2n -1)7T/2)(2n -1)-1]
= 2( -1)n[(2n + 1)-1- (2n -1)-1]

= 4(-I)n+l(4n2-1tl.
This gives the result required.

b) To establish the second result we use an integral representation for
the Bessel function In(z). [See Appendix]. Suppose sin @7T cos 6) =~ao+
L ~ cos n6. We have to evaluate
In =
By putting 6 = 7T Since the cosine is
tegral as SO/2 + S:/2
part, we find when
be shown that if
t:
cos m6 sin (~7T cos 6) d6.
'" we find 1m = -(- l r1m so that 1m = 0 if m is even.

even, 1m = 2 So COS m6 sin (~7T cos 6) d6. Writing this inand changing the variable to '" = 7T - 6 in the second
m is odd, that 1m = 4 So/2 cos m6 sin (~7T cos 6) d6. It can
In(z) = (27T)-1
(1)
t:
cos (n6 -
sin 6) d6
then, if n is odd,
(2)
I n(z)=;;-(-I)(n-l)/2
kC'II"/2 cos nTJ sin(z cos TJ) dTJ.
Hence, if m is odd, 1m =27T(-I)(m-1)/2Im(~7T)'as required.

To establish (2) we use the addition formula for the cosine in (1) to
get
In (z) =
! I'll"
COS
! {'II"
n6 cos (z sin 6) d6 +
sin n6 sin (z sin 6) d6.
Putting 6 = 7T - cp shows that the first integral vanishes for n odd so that in
this case
In(z) =
!r
sin n6 sin (z sin 6) d6,
n odd.
190
Writing j1T0-- j1T/2

0 + j1T
1T/2 and changing variable in the second integral to
1/1 = 7T - 0 gives
2 I1T!2
I n (z) = 7T
sin nO sin (z sin 0) dO.
If we now put 0 =17T-TJ we obtain (2).
5.14. Solution
The last approximation to Jo(x) is due to E. E. Allen (see NBS
Handbook, p. 369) and it is asserted that for -3:5x:53,.!e,.(x)!:55xlO- s .
5.15. Solution
For x = 0, every term is zero and so the sum is zero. For x;e 0 the series
is a geometric one with first term x 2 and common ratio (1 + X 2)-1 < 1; it is
therefore convergent with sum
x 2 J(1-(1 + X2)-1) = 1 + x 2
The sum function is clearly discontinuous at x = 0 and so convergence

cannot be uniform in any interval including this point.
Consider rn(x). For x=O, rn(O)=O, For x;eO, rn(x)=(1+x 2 )-n and
!rn(x)!<s if (1+x 2 )">s-1 i.e., if no (s,xlogs-lJlog(1+x 2 )). As x~O,
noes, x) is clearly unbounded for log (1 + x 2 ) ~ log 1 = O. This establishes
non-uniform convergence in the sense of Definition 3.
Using the second approach we observe that
and so {Mn} is certainly not a null sequence.
5.16. Solution
Since Sn (0) == 0, lim Sn (0) = O. For x;e 0, by Problem 2.8, lim Sn (x) = O.
Hence s(x)=limsn(x) is identically zero in [0,1]. Hence gs(x) dx=O. On
the other hand, integrating by parts,
r r
sn(x)dx=
(nx)e-nx(ndx) =
te-tdt
e- t dt
[-te-t];j+
-en + 1)e- n + 1 ~ 1, as
n ~OO.
We show that there is non-uniform convergence. It is more convenient to use the Mn - definition. Since Sn (x) - S(x) = Sn (x) we need to
191
Chapter 6
evaluate
Since sn(x)=n 2[I-nx]e- nX, sn(x) increases from 0, at x=O to n/e at
x = n-t, and decreases to .n2/e n at x = 1. Clearly Mn = n/e and this is
certainly not a null sequence.
Chapter 6
6.3. Solution
Try, for instance, Xn = 2(1!2t + 3(1!3t + 4(1/4t.
6.4. Solution
See John Todd and S. E. Warschawski, On the solution of the
Lichtenstein - Gerschgorin integral equation in conformal mapping, II, NBS
Applied Math. Series 42 (1955), 31-44, esp. 41.
6.5. Solution
See J. Liang and John Todd, The Stieltjes Constants. J. Research Nat.
Bureau Stand. 76B (1972), 161-178. The first sum is 0.1598689037, the
second 0.0653725926 and the third 0.0094139502.
6.6. Solution
L a.. xn = L xn(1 + ;1)nao = (1- x(1 + ;1))-lao

= (1- x)-I(I- x(l- x )-1;1)-1 ao
=1~xL(l~Xr;1nao
6.7. Solution
By induction, beginning with ;1ao = (-I)O[aoJ. If ;1nao is in the given
form we have
;1n+l a o = ;1(;1nao) = ;1nal -;1 nao
= (_l)n+l
[ao-(~) al+" . + (-Ita..
- a l + ... + (-It+l
_
n+l [
-(-1)
a o-
(~) a.. + (-I t
+1 a n+1]
(n +1 1) a 1+ .. +(-1 )n (n +1 1) a..
+(-It+ 1a n+l] .
192
r r
Integrating tenn by tenn,

1=
(l-x)ndx=
(1-(~)X+ ... +(-1)nxn)dx
n)x2
xn]1
= [ x- ( 1 "2+ ... +(-1)n n+1 0
= 1_(n)!+ +(_1)n_1_
1 2 ...
n+1
where a,,=(n+1)-1. On the other hand, I=[-(1-x)n+1/(n+1)~=

1/(n + 1). Hence the Euler transfonn of the logarithmic series is
L (-1)n2- n- 1/(n + 1).

That this has the same sum as the original follows from the relation
log(1-x)-1=x+(x 2 /2)+(x 3 /3)+ ... ,
Ixl<1
by putting x = 1/2.
6.S. Solution
Observe that
anU, =(-1)"
= (-1)n
= (-1)n
[r!1-(~) r!2 +(~) r!3 - ... J
xr(l-x)ndx
r!n!
(n+r+1)!
Hence the Euler transfonn of the tail is
6.9. Solution
We use the result of Problem 6.7 above to compute amf(O) when
f(s) = n-S In fact
a mf(O)=(-1r
[1- (7) n-1+ (~) n-2- .. . + (-1)mn-mJ
r.
=( -1)m[1-n- 1
The Euler transfonn of I (-n)-S (which has sum n/(n+1) when n>1)
Chapter 6
193
is therefore
(1/2)
f (-1)"2-8(-1)8(1- n- )8 =(1/2) f (n2n-1)8,

1
8=0
8=0
a geometric series, again with sum n/(n + 1).

Thus from a geometric progression with common ratio - n -1 we obtain
one with common ratio (n -1)/(2n) so that we get acceleration of rate of
convergence when n = 2, retention when n = 3 and deceleration when n = 4.
6.10. Solution
We consider using a head of r terms
1 1
1--+-+
3 5 ... +(-1) - 2r-1
and applying the Euler transform to the tail

(-1)'
[2r~1- .. .].
The n-th difference of the sequence (2r+ 1t\ (2r+3t\ . .. is

lln(U r+l) = (-It
= (-l)n
= (-1)"
[2r~ 1- (;) 2r~3 + ... +(-It 2r+;n+ 1]
x2r(1- x2)n dx
"'/2
sin2r 6 COS2 n+ 1 6 d6
=(-lt2 n n!(2r-1)!!
r=O.
(2n +2r+ I)!!'
These results are easily established using the reduction formulas for
l(m, n) =
112",
sinm x cosn x dx.
With this notation we have

(1)
(m + n)l(m, n) = (m -l)l(m -2, n) = (n -l)l(m, n -2),
l(m, n) =l(n, m),
1(0, 1) =
",12
cos xdx = 1.
194
We establish (1) by integrating by parts: we find
I(m, n) =
= [
",/2
(sinm x cos x) cosn- 1
Sinm + 1 x
] ",/2
1 cosn- 1 X
+
m+
X dx
1"'/2 Sin
- (n -1) cosnm+l
m+1
2 X
sin x dx
so that
(m + 1)I(m, n) = (n -1)
=(n -1)
1"'/2 sinm+
f.
",/2
x cosn- 2 X dx
(1-cos2 x) sinm x cosn-2
X dx
= (n -1)I(m, n - 2) - (n -l)I(m, n).
Hence
(m +n)I(m, n) = (n -1)I(m, n -2).
We therefore obtain, if we start the transform from the beginning,
Observe that this series is much more rapidly convergent than the Gregory
series. In fact
1 n!2 n!
f7r
u..=2 (2n+1)!-V;;2
n
-n-l
We can check that the new series actually has the proper sum, e.g., as
follows. Consider the Maclaurin series for
f(x) = (arcsin x)/.J1- x 2
This can be found by differentiating to get
so that
(1-x 2 )f'-xf-1 =0.

If we differentiate this n times by Leibniz' theorem we get
195
Chapter 6
If we put x = 0 then
This gives
22
2 .4
3+ _ _ 5+
f( x ) -- x +3! x
3!5! x ....
Putting x = 7T/.J2 we find
~/ fI=~[1+!+1.2+ ... ]
4
'J2: J2
3 3.5
which is the result required.

6.11. Solution
VO-V1 +V2+'" = (1- E)-lvo
= (1/2)(1- m)-lvo
= (1/2)[vo+ Mv o+ Wvo+ ... ].
6.12. Solution
We take so=O, Sn =I~=l (-ly-1a" for n~l so that (-1)n- 1a.. =

Sn - Sn-1 for n ~ 1. By definition Sn = I~=l 2- -1)v-1av-1a1'
We use induction to prove that Sn = 2- n I~=o (~)ST' This is trivial for
n = 1 since Sl = (1/2)a b Sl = a 1
Assume the result established for a particular value, n. Then
V (
Sn+1 = Sn + 2-(n+1)(-1)" anal

= Sn +T(n+l) nf (_ly-l ( n ) a"
v=l
v-1
= 2-(n+1) [2
from Problem 6.7
(n) Sv + nf (Sv - Sv-1) x ( : 1)]

v=o v
v=l
V
by induction hypothesis
This completes the induction proof.

The fact that Sn ~ I implies Sn ~ I now follows from a general theorem
of Toeplitz. [See e.g., K. Knopp, Theory of Infinite Series, p. 72]. We now
196
give a direct proof in the present special case. Because (0) + (~) + ... + (;:) =
(1 + 1t = 2n we have
and therefore it will be sufficient to establish the result when {s,.} is a null
sequence. For convenience put a",r = 2-n(~) so that 1 ~ an,r> 0, L~=o an,r = 1.
Notice that for r fixed an,r is a polynomial in n (of degree r) divided by
an exponential 2n = e n10g2 - hence an,r ~ 0 as n ~ 00. [Compare Problem
2.8].
Given any e > 0, we can find no = no( e) such that ISn I< ~e if n ~ no.
Then
ISnl < lan,oso+ ... + an,nos...,1 + lan,no+1sno+l + ... + an,ns,.1

< lan,oso + ... + an,nos...,1 + ~e
We can now choose nl~nO such that lan,oso+" .+an,noSnaI<~e. This is
because, no being fixed, {an,oso + ... + an,nos..J is a null sequence. Hence, if

n ~ nl> we have 15..1 < e i.e., {Sn} is a null sequence. This completes the
proof.
6.13. Solution
1
1
-2
1
2:
'3
'3
-4
-6
1
12
r=4:
111
1-2:+"3-4= 112/192
192 log., 2 = 133.08
197
Chapter 6
6.15. Solution
dx
- 1r=
o +X
1
i1
(-l)n
[1-x r +x 2r - ... ]dx=L1 -
00
Those integrals can be evaluated explicitly, e.g.,

(-l)n
~ 1+3n
00
--=
11
0
dx
-~=
1+x 3
[1-log
3
+m
x+1
2X-1]1
+-arctan-v'(x 2-x+1) J3
J3 0
1
7r
=-log2+ ~=0.83564885.
3
3,,3
For the other integrals see Problem 9.13. The results are
fo 1+4n
(-It =0.86697299,
(-l)n
Lo -+1
5 =0.88831357,
n
00
(-It
L
6 =0.90377177.
o 1+ n
00
6.16. Solution
We have seen in Chapter 2 that if En = Xn - N- 1 then En+1 = -NB~.

Hence Bn+2=-NE~+1 =-N(-NB~2=-N3B~. Clearly
This gives
N2B~(1- ~e~f
= Bn+2- en (1 +2Nen - N3B~r
The second term on the right is O(E~) and swamps the first which is O'(e~).
Hence En+2=0(E~~0. However
En+2=1+ l+O'(E~
Bn+2
NBn(l +O'(En))
= +1
NBn
--+-00.
+0(1)
198
Chapter 7
7.1. Solution
In the usual notation [(x) = eXEt(x) and we find from NBS Handbook,
pp.242,243
[(20) = 0.04771 8545

[(10) = 0.09156 33339,
[(5) = 0.17042 2176
7.2. Solution
We have
U(x) =
iOOCOjidt_ rCOS,frdt= .Ji7T[1-2CiX)]
and, similarly, V(x) =.J(7T/2)7T[1- 2S2(x)].

From Watson, Bessel Functions, p. 744, we find
C 2(5) = 0.32845 7, so that
U(5) = 0.42999 5
S2(5) = 0.46594 2
V(5) = 0.08537 1
C2(10) = 0.43696 4
U(10) = 0.15800 8
S2(10) = 0.60843 6
V(10) = -0.27180 9
Ci15) = 0.56933 5
U(15) = -0.17379 7
Si15) = 0.57580 3
V(15) = -0.19001 0
Ci20) = 0.580389
U(20) = -0.20150 5
S2(20) = 0.46164 6
V(20) = 0.961392
The required values of the trigonometrical functions can be found in NBS

Handbook, p. 175. From p. 3 we have
.J!7T= 1.2533141373.
7.3. Solution
Computation for x = 3
1.0000000 -0.05555 56
0.00925 93 -0.00257 20
0.0010002 -0.0005001
0.00030 56 -0.0002207
0.00018 34 -0.00017 37 ~ least term
0.0001834
Chapter 7
199
Sum of first nine terms 0.9519037; error bound 0.00017 37.

Correct value of g(3): 0.951814, actual error 0.00009 O.
Correct value of erfc 3: 2.20905 x 10-5
X
= 4: error estimate 1.6 x 10-7 , g(4) = 0.971304, erfc4 = 1.54173 x 10-8 .
x=5:
g(5) = 0.98109 4, erfc 5 = 1.53746 X 10- 12
x=6:
g(6)=0.986651, erfc6= 2.15197 x 10- 17
The correct values of erfc x were obtained from pp. 242, 295 - of
National Bureau of Standards Applied Math. Series #41, Tables of the
error function and its derivatives. Washington, D.C. 1954. From these the
correct values of g can then be obtained from
g(x) = ';;xe x2 erfc x
using, e.g., the tables of exp x in the NBS Handbook, p. 138.

7.4. Solution
For more details about Problems 7.4-6 see E. T. Goodwin and J.

Staton, Quart. J. Mech. Appl. Math. 1 (1948), 319-326; a brief version of
the table in this paper is in NBS Handbook, p. 1004.
Write
1
1[
U]-l
u+x=~ 1+~
1 n-l (-u)r
=~r~o ~
(-u)n
+xn(u+x)"
Multiplying across by e- t and integrating we see
so that
Since x > 0 and u ~ 0 we have u + x > u and so

Irn(x)l:sx- n
ro
u n- 1e- u2 duo
Hence Irn(x)1 does not exceed the absolute value of the first term omitted.
The integrals occurring can be expressed in terms of the Gamma
function (by using the change of variables from u 2 to t).
[0
e-u2ur-1 du =!f(!r).
200
Using the facts that f(s + 1) = sf(s) and that
rm = J;. we find
The error being bounded above by the absolute value of the first term
omitted we see that, correct to 4 decimals for x ~ 0,
.;; {1 I} -2X2-1
F(x)=- -+2 x 2x 3
since
1 1
-4
Ir3 (x) I<-4:5-X
10 .
2x
Note that the optimal value of n is about 2x 2. For instance for x = 5, we

find f(50/2H= 6.2x 1023 , f(49/2) = 1.26 x 1023 (NBS Handbook, p. 272).
It can be shown that the Euler process can be successfully applied to the
asymptotic series.
7.5. Solution
Integrating by parts we find
oo ue-u2 du
u+x
1
2x
1 roo e- u2 du
21
(u+xY
Write u/(u+x) in the left hand side of the above relation as 1-[x/(u+x)]
and we find
foo
o
ue-u2 du = J;. _ xF(x).

u+x
2
Differentiating the definition of F(x) with respect to x we get

,
F(x)=
foo e- u2 du
-1
(U+X)2'
Combining the three displayed relations gives the result required.

7.6. Solution
We want to get a "power series" representation for F(x).
There is difficulty because F(x) is only defined for x> 0 and the
differential equation obtained in Problem 7.5 suggests that F(x) behaves like
-log x near x = o.
201
Chapter 7
In order to get more information about F(x) for x -0 we write
F(x) =
roo t2tt )
.10 e x+t
and note that for t - 0, et2 -1 + t 2 We therefore suspect that F(x) and
roo
dt
Fo(x) = .10 (1 + t 2 )(x + t)

should behave similarly for x - O. We can evaluate Fo(x) analytically: using
partial fractions we find
and, integrating,
_
1
[x+t
Fo(x) - (1 + x2) log (1 + t2)1/2 + x arctan t
00
1
1
- (1 + x2) [-log x +z1TX].
_
We therefore have
1]
dt .
F(x)-Fo(x) = l oo [ e- t2 _
-2 1 +t
x+t
Letting x --+ 0 we find, subject to justification,

lim {F(x) + log x}= loo
Changing the variable from t to
[e- t2 - 1 ~t2] ~t.
= t2 in the last integral gives
lim{F(x)+logx}="2.1oroo [ e-
1] d
1+T
TT,
Easy manipulations of the integrals in the lemma (p. 80) show that
roo
Jo
[e-
1 ] dT = -'Y,
1 +T T
T __
so that
lim{F(x)+logx}= -!-y.
(1)
To get the representation given we multiply across the differential

equation by the integrating factor e and rewrite it as
x2
(e y + log x)' = .[;e

X2
x2
x- 1 (e X2 -1).
202
Integrating the power series on the right we find

X2
<X>
X2n + 1
e F+logx=v''lTL
'(2
1)
o n. n+
X2n
L----'--(2
)+const.
n. n
<X>
By letting x ~ 0, using (1), we see that the constant must be

we obtained the required result.
We find F(l) = 0.6501.
-h and hence
7.7. Solution
The standard definitions of the Fresnel integrals are
C(y) =
cos G'lTf) dt;
1 IX sin t dt 1 IX
-r=11/2(t) dt,
'V2'lT 0 'Vt
2 0
Six)= , -
1 IX cos t dt 1 IX
---r=1 1/2(t)dt.
v2'lT 0 vt
2 0
C 2(x)= r;::-
C(x), Sex) and C 2(!'lTX 2), Si!'lTx 2) are tabulated in the NBS Handbook.
In the present context it is more convenient to have Cix), S2(X). If we
write u = v'(x + t) we find
Using the fact that

(1)
.ro sin u 2 du = .ro cos u 2du = v' 'IT/8 we find
[(x) = v'!'lT(cos x -sin x)-2 cos x
f'x
sin u 2 du +2 sin x
rx
cos u 2 duo
If we put t = u 2 in the integrals on the right we get
= v'!'7l{cos x(1-2Six+sin x(2C2(x) -1)].

Reference to FMRC, p. 461 tells us the Six), C 2(x) are tabulated in
203
Chapter 7
Watson, Bessel Functions. For
In particular we can use

S2(1) = 0.24755 8,
sin 1 = 0.84147 1
Ci1) = 0.721706,
cos 1 = 0.54030 2,
to get
f(1) = 0.80952 55.
In order to derive the power series for f(x) we rewrite (1) as
rJi
f(x) = v'!17(cos x -sin x)-2.1o sin (u 2- x) du
and contemplate expanding the integrand as a power series in u 2 - x and

then integrating. For this purpose we need to find In = S~x (u 2 - x)" duo We
have
= -2nIn -2nxIn - 1
so that
Also
Hence
12 ,+1
22(2r + 1)(2r)x 2
= (4r+ 1)(4r+3) 12 ,-1
so that
I
_ -22,+I2r+ 1)!)X(4'+3)/2
2,+1 -
(4r+3)!!
204
and
f(x)=~!17(cosx-sinx)+(-2)
(-lYI2r+I/(2r+l)!
r=O
'" (-1Y2r+ 1)!)24r +3x(4r+3)/2

= ~!17(cos x -sin x) +
(4 3)'
r=O
r+ .
_
Checking the ratio of terms in the last series, or estimating the general
term, we see that convergence is reasonable for, e.g., 0:5 x:5 3.
An asymptotic representation of f(x) can be found by repeated integration by parts. Thus
f(x) = [ -(x + tr I/2 cos {1;'+ [' cos t( -!(x + t?/2) dt
= x- I12 + [ -!(x + t)-312 sin t]+
f'" !-~(x
+ t)-3/2 sin t dt
= x-l/2_ ~:~ [ ' (x + t)-312 sin t dt.
Generally we find
This will be convenient for say, x;;:: 10.

Finally, to get a differential equation satisfied by f(x) we differentiate
under the integral sign to get
,
f (x) =
r=
-2.10
sin t dt
(x + t)3/2'
dt
f '(x) = ~4.10r= (xsin+ tt)5/2
.
If we integrate by parts we find
r=
1 1
cos tdt
[ -cos t]= 11= cos tdt
f(x)= (X+t)1/2 0 - 2 0 (x + t)312 = xl/ 2-2.1o (X+t)3/ 2
If we integrate the expression for f"(x) by parts we get
" [3
-2
sin t ]
= 32
f (x)= 43 (x + t?/2 0 +43"
r=
1= cos t dt 1
cos t dt
0 (X+t)3/2=2.1o (x + t)3/ 2
From the last two displayed formulas we find

f"(x) + f(x) = x- 1I2 ,
the required equation which can be used in the intermediate range 3-10.
205
Chapter 7
We give some values of f(x)
o
1
2
3
4
5
1.2533141
0.8095255
0.6429039
0.5468906
0.4828942
0.436552
10 0.314027
15 0.257368
20 0.223196
7.B. Solution
We have
Also, if x > 0, we have
Hence
so that
Hence
x 2n - 1 IG(x)-{ .. .}I=(2n)!/x2~0
as
x~oo.
Also
x 2n IG(x)-{ .. .}+O x- 2n l = (2n)!/x ~ 0 as
The last two relations show that we indeed have
1 2!
4!
G(x)----+x x 3 x 5 '"
according to the strict letter of our definition.
x ~ 00.
206
For x = 5, two terms give G(x) to within 10-2 ; for x = 10 five terms give
G(x) to within 4x 10-5 and for x = 15 seven terms give G(x) to within
2x 10-7
7.9. Solution
The only trouble is finding the error in the binomial expansion and from
it to show the true asymptotic character of the series derived formally.
Integrating the relation
d
- (1 + t)v = v(1 + t)v-l
dt
from 0 to x we get
(1+x)V-1=v
=
r
r
(1+t)v-l dt
(1 + x - T)v-l dT,
the last line being obtained by changing the variable from t to T = x - t.

Thus we have
(1+x)v=1+v
(1+x-tt- 1 dt.
Integrating by parts gives

(1 +xt = 1 +vx+v(v-1)
(1 +x-t)V-2t dt.
Repeating this operation we find

(1 + xt ={1 + vx+ v(v-1)x 2/2! + ... + v(v-1) .. . (v- n + 1)x n /n!}+ rn+l(x)
where
rn +l(x)=v(v-1) ... (v-n)
(1+x-tt- n - 1 (t n /n!)dt.
Assuming x>O it is clear that when O::5;t::5;x when 1 +x~(1 +x-t)~ 1

and so, provided n> v -1, the first factor in the integrand is less than 1 and
( )\ <\v(v-1) ... (v-n)\ n+l
\
rn+1 x (n+1)!
x
i.e., the remainder is less in absolute value than the first omitted term.
207
Chapter 8
In our application we have
where
=(-1)n 1.3.... .(2n -1) . .r:..
tn
2nn!
x2n
and where
.(2n + 1) yn+1
2n+1(n + 1)!
x 2n '
-< 1.3 ....
rn+1
Both infinite integrals exist. Integrating we get

g (x )
(1)
=f
1...
(-1)r 1.3 .... .(2r+ 1) + R
n=O
2r
2r
n+1
where
The asymptotic character of (1) is clear since for each n
Chapter 8
8.2. Solution
Differentiating the product l(x) by Leibniz' Rule we find

l'(x) =
(x - x2)(x - x 3) ... (x - Xn-1)(X - Xn)

+(x - Xl)
(x - x 3) ... (x - Xn-1)(X - Xn)
+ ...
+(x - X1)(X - x2)(x - X3) ... (x - Xn-1)'
Putting x = ~ all terms vanish except the i-th so that

product being over j = 0,1, ... , n, j;i: i.
l'(~)
=n (~- Xj),
the
208
8.6. Solution
We can write down L 3 (x) in the general form and simplify it to
L 3 (x)=x(2x 2 +x-2). Draw a graph otL3 (x) using its values at the nodes
and the fact that it has turning points at x = (-1 m)/6 with values
(19=t= 13m)/54.
The error can be written down directly by noting that it is a polynomial
of degree 4 with leading coefficient 1 which vanishes at 1, 0, 2. Thus
E(x) = x(x 2 -1)(x -2).

It is clear that E(x) is symmetric about x =! and
E'(x)=4x 3 -6x 2 -2x +2=4(x -!)(x 2 - x -1).

It follows that E(x) has turning points at !, !(1v'5) with values 9/16 and
-1. At -!, ~ the value is -15/16.
The results just obtained indicate how the error in interpolation varies
with the position at which we interpolate - as common sense suggests, it is
better to interpolate near the center of a table, than at its ends.
8.7. Solution
We have L 1(x) = -x(x -l)(x -2)/6, etc. giving
L1
x = 0.4 -0.064
10
11
12
0.672
0.448
-0.056
x = 0.5 -0.0625 0.5625 0.5625 -0.0625
To find the interpolant 1(0.4) given 1-1> 10' 11> 12 we have just to form the scalar
product of these numbers with the numbers in the row labelled x = 0.4
above.
For further information on interpolation, and much good advice on
computation, see Interpolation and Allied Tables.
8.8. Solution
Since (sin x)" = -sin x and since Isin xl:s 1 the error does not exceed
i x (0.2)2 xl = 0.005. This means that it is not reasonable to use more than
two decimals in the table since the round-off error would then be much
209
Chapter 8
smaller than the interpolation error. These conclusions can be checked by

actual calculations, for instance from the values
sin 1.4 0.98545
sin 1.5 0.99749
sin 1.6 0.99957
Linear interpolation between 1.4, 1.6 gives 0.99251 with an error of
0.00498.
8.9. Solution
If f(x) is a polynomial of degree n at most then
n
f(x)=
L f(xJHx).
i=O
If we take f(x)
= Xi, j = 0,1,2, ... , n in this we find

j= 1, 2, ... , n.
SO= 1,
To deal with the case j
'7Tn(x)
= (x -
= n + 1 we note that if
xo)(x - Xl) (X - Xn) = x n+ 1 + a1 xn + ... + a n+1
then for i = 0, 1, 2, ... , n we have
'7Tn (X;)
= 0,
Multiplying across by HO) and summing from i = 0 to i = n we get
Special case:
,o{x)
(x - 2)(x - 3)(x - 4)
-6
12(x)
, 0 (0)
(x -l)(x - 2)(x - 4)
-2
(x -l)(x - 3)(x - 4)
2
'
1 ( ) = (x -l)(x - 2)(x - 3)
3 X
'
= 4,
So = 1,
'7T3(X)
I ()
Sl = S2 = S3 = 0,
S4 = - 24.
= (x -l)(x -2)(x-3)(x -4) = x 4 -10x 3 +35x 2-50x +24.
210
8.10. Solution
Lemma.
If
then
r
r
(n -i)!
co -c1+c2-c3 + ... +(-1)c,.=(-1) r!(n-r-l)1"
Proof. The sum required is clearly the coefficient of xr in
(1 + xt[x r - x r- 1+ x r- 2 _ . .. +(-1)']
i.e., in
(-1)'[1- x + ... + (-l)'xr][l + x]n
i.e., in
i.e., in
i.e., in
(-1)'[1 + x]n-1
which is the result given.
L,.(1) is clearly +1 since 1 are always nodes. Hence L,.(1)~ 1.
If n is odd it is clear from symmetry that L,. (0) = O. We deal now with
L 2n (0). We write
k=1,2, ... , n,
and note that

(2n -l)(xk - Xj) = 2(k - j).
Using the explicit form of the Lagrangian interpolant we see that

L 2n (0) = ~ \Xk \ (-It[(2n -1)(2n - 3) ... (3)(1)]2/( -(2n -l)xk)
k=l
2(k-l) 2(k-2) ... (2)(-2)(-4) ... (-2(2n-k
=(-1)n-1[(2n-3)!!]2
2 2n - 1
=(-1)
(signxk)(2n-l)
k=l (k -1)!2n - k)!)(-l)k
n-1 [(2n-3)!!]2
k.
22n-1(2n-2)! k~l (-1) SlgnXk
(2n-l)
k-l .
Chapter 8
211
Remembering that Xl> .. , x" are negative and Xn+l' ... , X2n positive and
that the coefficients in a binomial series are symmetric we see that the sum
in the last expression is just
2[co -
where
Co,
Cl
+ ... + (-It- l cn _l]
C1o ... are the coefficients in (1 + x?n-l. From the lemma this is
(2n-2)!
2(-I)n-l ((n -1)!)2
Hence
L 2n (0)
J2
[ 2(2n-3)!!
n- l (n -I)! .
Application of Stirling's Theorem gives

Theorem is
L2n(0)~(1Tn)-1.
Indeed Stirling's
so that
It follows that 4 (0) ~ o. It is remarkable that it is only at the points 0,

1 that we have 4(X) ~ Ix!. For a proof of this result of S. N. Bernstein
see, e.g., 1. P. Natanson, Constructive Function Theory (in Russian, 1951, in
German translation by K. Bogel, Akademie-Verlag, Berlin, 1955.)
8.11. Solution
If f(i) = fb i
= 0, 1, ... , n then we can obtain a representation of f(x) as

a Lagrangian polynomial f(x) = L~~o tl;(x) where
Since f(x) is monic and each l;(x) is monic we must have
212
It follows, from this and the definition of IL, that
IL
L Z..'( n _.)'~1.
I.
i=O
However
L i!(n-i)!
(n)
n! L i
(1 + It
n!
2n
nr
Hence IL ~ n !/2n but since we may take {; = (-1 Yn !/2n we can have equality
and uniqueness, apart from sign.
8.12. Solution
We use the facts that

Clearly
H'(x) =
4(~)
= 8(i, j). It is obvious that H(x;) = {;, all i.
L [24(x)l[(x){I-21[(x;)(x - .x;)} +{li (xW{-21[(x;)}]{;

+ L [24(x)I[(x)(x -:x:;) +{4(x)FJfr
and again
H'(.x;) = [21[(x;) -21[(.x;)]{; + {r = {r.
With E(z) defined by (12) it is clear that E(x) = 0 and E(.x;) = 0 for
i = 0,1, ... , n. Hence E'(z) vanishes for (n + 1) values of z distinct from the
.x;. But it is clear that E'(x) = 0 for i = 0, 1, ... , n. Thus E'(z) vanishes for
2n + 2 distinct values of z. Repeated applications of Rolle's Theorem shows
that E(2n+2)(z) must vanish somewhere in the interval including x and the .x;
say at ~. Since H is a polynomial of degree 2n + 1 at most we have
o= f2n+2)(~)_[f(x)- H(x)][(2n +2)!/0 (x -
xY]
which gives (11).

8.13. Solution
We use induction to prove that
n~i>j
For n = 1,
1
V l =det [ 1
xo] =xl-XO
Xl
Suppose we have established the result for n = r. Consider
Chapter 8
213
and expand it in terms of minors of the first row. Clearly V r +1 will be a

polynomial of degree at most r+ 1 in x. Also V r + 1 will vanish for x =
Xl> X2, ... ,Xr+l so that
Vr+l
= k(x -
Xl)(X - X2) ... (x -
Xr+l),
where k is a constant. Now k is the coefficient of x r +1 which is also evidently

(-ly+1V(xl> x 2, ... , x;.). The induction hypothesis gives
k=TI ..
r~t>J~l
(Xi-X.)
J
so that
(_1)'+1 = (x - X1)(X - x 2) ... (x - x r+1)
TI
(Xi - Xj).
r::2!i>j
and so
V(xo, Xl> ... , Xr+1)
TI
(Xi - Xj).
r+l~i>j
This completes the proof.

It is clear that 1, x, ... ,x n form a basis for 'Vn and since these are
independent (by an application of the Fundamental Theorem of Algebra),
the space has dimension n + 1.
Since, for r = 0,1, ... , n, L,. = I~=o x~li(x) is a polynomial of degree at
most n, coinciding with xr at xo, Xl> ... ,Xn it follows that L,. must be
identical with xr. In other words
xC:
x~
x~
1n(x)
Since the Vandermonde matrix is non-singular, it follows that 'o(x),

11 (x), ... , 1n (x) form a basis for 'Vn"
8.14. Solution
This is a table of 1010 Jo(5.495 + 10- 3 x). To the zero
io,2 = 5.5200781103
of Jo(x) corresponds a zero of f(x) at
x = 2.50781103.
214
There is a zero of [(x) between 2 and 3. Using all six points for
Lagrangian interpolation to subtabulate indicates a zero between 2.5 and
2.6 since
[(2.5) = -:2 65784
[(2.6) = +3136604.
Estimating the position of the zero by linear interpolation gives 2.5078. We
therefore subtabulate again getting
f(2.507) =
[(2.508) =
f(2.509) =
and observe that the second difference is

. Linear (inverse) interpolation is permissible and we find the value given.
8.15. Solution
For discussions of wide generalizations of parts of this problem see, e.g.
P. W. Gaffney, J. lost. Math. Appl. 21 (1978), 211-226.
C. A. Micchelli, T. J. Rivlin, S. Winograd, Numer. Math. 26 (1978),
191-200.
For instance, if [(0) = 0, [(1) = 1 and 1[,(x)I::::;2 in [0, 1] then the graph
of [(x) must be within the left hand parallelogram. If [(0) = 0, [(1) = 1 and
1f'(x)I::::;2 in [0,1] the graph of [(x) must lie within the lens shaped region on
the right.
y-2x-x 2
215
Chapter 8
8.16. Solution
We want to show that there is a cubic H(x)=a+bx+cx 2 +dx 3 such
that
(1)
H(xJ=t,
H'(xJ=ir,
i=0,1,
n,
where Xo f= Xl and fo, flo

f~ are arbitrary.
The relations (1) when written out in full give a set of four linear
equations for the unknowns a, b, c, d. The determinant of this system is
x~ x~
1 Xo
[ '0 1 2xo 3x~
h = det 1 Xl xi xf
o 1 2XI 3xi
We evaluate h by row operations as follows. Take row 3 -row l and divide

the third row through by Xl - Xo. We find
Taking row2 -row l and row3 -rOWI we find, on dividing through the second
and third rows by (Xl - xo),
Hence the system is non-singular and a, b, c, d can be found.

For information on the use of Hermite interpolation see, e.g., H. E.
Salzer, J. Res. Nat. Bur. Standards 52 (1956), 211-216.
8.17. Solution
From Problem 8.16, Solution it follows that the second derivatives at Xo
in the left and right panels are, if h+ = Xl - XO,
h_ = Xo - X-I'
and
-6foh~2 + 6flh~2 - 4nh~l- 2f~h~l.
Equating these gives a linear equation for f~ with coefficient 4h=1 +4h~1 f= O.
Hence f~ is determined uniquely.
216
n,
In the general case, when, fo, f1> ... ,fn and

f~ are given, quantities
f~, ... ,f~-l can be determined uniquely so that all the abutting cubics fit
smoothly, because the system of equations is non-singular, having a dominant diagonal.
In the special case = 0 and the spline is given by
-2x 3 -3x 2 +1 in [-1,0]

2x 3 -
3X2
+1
in [0, 1].
Observe that there is a jump in the third derivative from -12 to 12 at x = O.

These results can also be obtained by using the results of Problem 8.12.
8.18. Solution
For simplicity suppose the three points are equally spaced and, without
loss, take them to be Xl = -1, Xo = 0, Xl = 1. Then we have fo = a, fl =
ab+c so that b=!(fl-f-l) and c=!(f-1-2fo+fl). It is clear that x=
-b/(2c) and {= a - b2 /(4c). For instance, given f-l = -0.3992, fo = -0.4026,
fl = -0.4018 we estimate
x = (13/42) = 0.3095,
{=0.4028.
For developments of this method see e.g., H. E. Salzer, Formulas for

finding the argument for which a function has a given derivative, MfAC 5
(1951), 213-215.
8.19. Solution
This is Newton's interpolation formula. It is, by uniqueness, necessarily
a rearrangement of the Lagrangian expression. It can be obtained formally
by writing
and truncating the binomial expansion: if f is a polynomial of degree n, then

0= !J.n+lf = !J.n+lf = ....
To establish the formula we can use induction and the basic recurrence
relation between binomial coefficients:
The quartic is
so that q(2.5) = 2884.
217
Chapter 9
The difference table for q(x) is:

0
789
567
1 1356
345
912
2 2268
123
468
1380
791
3 3648
2171
4
200
323
200
523
1314
5819
3485
5 9304
The constant term in the quartic is clearly q(O) = 789 and the leading
term must be (200/24)x 4 to give a constant fourth difference of 200. If we
subtract off these terms and divide through by x we are left with a quadratic
whose coefficients are those of the three middle terms of the quartic.
The Newton interpolation formula is:
[(2.5) = 789 + (2.5 x 567) + (2.5 x 1.5)345/2!
+ (2.5 x 1.5 x 0.5) 123/3!

+ (2.5 x 1.5 x 0.5 x (-0.5))200/4!
Truncating this we get successively
789
2206.5,
2853.375,
2891.8125,
2884.
Note that 2206.5 is got by linear extrapolation from q(O), q(1); if we

interpolate between q(2), q(3) we get 2958.
Chapter 9
9.1. Solution
Write down the fundamental polynomials and integrate them. For

instance
I (x) -1
x(x-1)(x-2)
(-1)(-2)(-3)
1[X 3 --6
3x 2 + 2x],
and
1
2
-1
L1(X)dx=i,
(2 lix) dx = i.
L1
218
This gives the so called
'i Rule'
Q = i/( -1) +U(O) +U(1) +U(2).
The fact that the sum of the weights is 3, which is the integral of I(x) = 1
between -1, 2, is a check.
We shall show that if Ie C 4 [-1,2] then
t:
I(x) dx -[U(-1)+~/(0)+U(1)+U(2)] = -iof4)(~)
where -1~~~~. In order that our computations have more symmetry we

shall prove, equivalently,
p(h)=
1
3k
-3k
I(x) dx - 34h [f(-3h) + 3/(-h) + 3/(h)+/(3h)] =
_~h5/(4)(~)
where -3h~~:53h and we assume fe C 4 [-3h, 3h]. We differentiate E(h)

four times with respect to h obtaining:
i.e.,
~E'(h) = [f-3 - 1-1 - 11 +13]+ h[f~3 +1~1 - I~ - I~];
~E"(h) = -2[f~3 - 1~1 + I~ - I~]- h[3f~3 + 1"--1 +
aE"'(h) = 3[["--3 - f"--l -
n+ 3/~];
fl + [3] + h[9f~~ + f~l- fl- 9f3'];
~E(4)(h) =4[f"--'I-fn-27h[f~j +1\4)]- h[f~I + ft4)].
Hence, since f4) is continuous, we have

~E(4)(h)
= -8hr)(~l) -
54hr)(~2) - 2hr)(~3)
= -64hr)(~4)
where -h~~l~h, -3h~~2~3h, -h~~3~h and so -3h~~4~3h.

We also note that E(O) = E'(O) = E"(O) = E"'(O) = O. Hence, integrating
the relation
we get
E"'(h) = -144
hr)(~it)) dt = -72f4)(~s)h2;
integrating again
E"(h) = -24hy4)(~6)'
E'(h) = -6h4f4)(~7)'
Chapter 9
219
and, finally,
where
-3h~~~3h.
9.2. Solution
We may assume the interpolating quadratic to be of the form q(x) =

[(0) + ax + bx 2 Since q(1) = [(1) we have two linear equations for a, b.
Actually we only need b since
Q
t:
q(x) dx
=4[(0)+e~)b =~[2[(-1)-[(0)+2t(1)].
We note that this quadrature is of "open" type; it does not involve the
values at the end points and can therefore be used as a "predictor" in the
solution of differential equations e.g., in Milne's method.
We also note that it is to be expected that the error incurred in open
formulas is larger than that in closed ones.
It does not seem possible to obtain an error estimate in this case by the
method used in the previous problem. We use an essentially general method
based on a modification of Steffensen's classical account given by D. R.
Hayes and L. Rubin (Amer. Math. Monthly, 77 (1970), 1065-1072).
Another general method for this type of problem is due to Peano (cf. A.
Ghizzetti and A. Ossicini, Quadrature Formulae, Birkhiiuser and Academic
Press, 1970); see also B. Wendroff, Theoretical Numerical Analysis,
Academic Press, 1966).
Let L(t, x) denote the Lagrangian polynomial based on the nodes
-h, 0, h and let 7T(X) = (x + h)x(x - h). Then we put
(1)
[(x) - L(f, x) = 7T(x)R(x)
and get
(2)
1- Q =
2h
7T(x)R(x) dx.
-2h
This defines R(x) except at the nodes where we define it by continuity. With
this convention it is easy to verify that R(x) is continuously differentiable.
If we write
l(x) =
rx
12h
7T(t) dt
then l(x)=Hx 2 (x 2 -2h 2 )-8h 4 ] and l(2h)=0. Hence, integrating (2) by
220
parts, we find
2h
t2h
1- Q = [1(t)R(t)f-~h -
l(t)R'(t) dt
2h
= -
l(t)R'(t) dt.
-2h
Now as l(x)::50 in (-2h, 2h) we may apply the Mean Value Theorem to get
1
=-R'(~)x -1;:h
1- Q = - R'(~)
2h
l(t) dt,
-2h
We shall now show that R'(~)=f4)(C)/24 where -2h::5C::52h, which

gives
1- Q =Sh5(4)(~).
(3)
Take a fixed x* in [-2h, 2h] and consider

~(x) = f(x)- L(f, x)-'7T(x)[R(x*)+(x - x*)R'(x*)].
Since L,
are of degree 3 at most we have
11'
(4)
~(4)(X) =
(4)(X)-4!R'(x*).
We shall show that there is a

(5)
~'(x*)=O
for any x*,
Csuch that
~"(x*) =
~(4)(C) = O.
We observe that
0 if x* is a node.
The first of these results follows from

~'(x)
= f'(x) -
L'(x)-1T(x)R'(x*)-1T'(X)[R(x*) + (x - x*)R'(x*)]
using the fact that f'(x)-L'(X)-1T(X)R'(x)=1T'(x)R(x) which comes by

differentiating (1).
The second of these follows because differentiating (1) again we get
~"(x)
= f'(x)-.L"(x)-21T'(x)R'(x*)-1T"(X)[R(x*)+(x -
x*)R'(x*)J
and using f'(x)- L"(x)-21T'(x)R(x) = 1T"(x)R(x) we find

~"(x*)
= 1T(x*)R"(x*)
which vanishes if x* is a node.

When x* is in general position, ~(x) has (in the special case under
discussion) 4 zeros: x* and the three nodes. Hence ~'(x) has three zeros by
Rolle's Theorem, and an additional distinct one at x* by (5). Hence ~"(x)
has three zeros.
221
Chapter 9
When x* is a node, ~(x) has three zeros and ~'(x) has two zeros by
Rolle's Theorem and an additional distinct one by (5). Hence ~"(x) has two
zeros by Rolle's Theorem and an additional distinct one at x* by the second
part of (5). Again ~"(x) has three zeros.
In both cases, by Rolle's Theorem, ~(4)(X) has a zero, say at ~. It then
follows from (4) that
t<;:~) = R'(x*)
where
depends on x*, but x* was arbitrary. This completes the proof of (3).
9.3. Solution
Assume that m::5 f'(x)::5 M in [a, b]. By the error estimate for linear
interpolation applied to f(x) in the subinterval [a,., a,.+1] where a,. =
a+rh, h =(b-a)/n, we have
f(x) - L(x) = (x - a,.)(x - a,.+1)f'(c,.)/2!
where a,.::5 c,.::5 a,.+1' This is true for r = 0, 1,2, ... , n -1. Since for a,.::5 x::5
a,.+1 we have
we can integrate these ineqUalities between a,., a,.+1 to get
!Mf'+l
(x - a,.)(x - a,.+1) dx::5
f'+1
[f(x) - L(x)] dx
::5!m
f'+l
(x - a,.)(x - a,.+1) dx
which, after a little algebra, reduces to
-Mh 3/12::5
f'+l
[f(x)- L(x)] dx ::5-mh 3/12.
We now sum the last inequalities with respect to r and get

m::5
123 [b [L(x)-f(x)]dx::5M.
Ja
nh
Now for a fixed h, the middle term above is a constant between m and M;
since f'(x) is continuous in [a, b] and bounded there by m, M, it must
assume this value (at least once), say at c. Hence
[L(x)-f(x)]dx=T-I=nh3 f'(c)=(b-a)3f'(c)
'
12
12n 2
222
9.4. Solution
We give two rather similar proofs.
(1) We verify by integration by parts, that
uv'" dx = uv"-u'v'+u"v- u"'vdx
if u, veach have continuous third derivatives. Taking u.=ix(1- X)2, v =

f(x) + f(-x) the above relation gives
t:
Since
f(x) dx = i[f(-1) +4f(0)+f(1)]-
uv", dx.
If4)(t)1 ::5M4 and since, from the Mean Value Theorem,

v.'" = f"'(x) - f"(- x) = 2x(4)(~);
we have
IEI::5 2M4l1 x 2(1- xf dx/6 = MJ90.

(2) Let L 2(x) be a quadratic interpolating f(x) at -h, 0, h. Then repeated
integration by parts gives for F(x) = f(x)- L 2(x), since F(h) = F(O) = 0,
t:
(x + h)3(3x - h)P4)(X) dx +
(x - h)(3x + h)3p4)(X) dx = 72
t:
F(x) dx.
Now, L 2(x) being a quadratic, P(4)(X) = ( 4)(x) and so 1P(4)(x)1 ::5M4. Hence
72IEI::52M4i" (h-X)3(3x+h) dx =4M4h s/5.

This gives the result required. For a.derivation of this result under weaker
conditions see Anon, Amer. Math. Monthly 76 (1969), 929-930.
9.5. Solution
We have
M=(4/3)[2f(-1)-f(0)+2f(1)]
and the adjusted
S=
(2/3)[f(-2)+4f(0)+f(2)] so that
S - M = (2/3)[f(-2)-4f(-1) + 6f(0)-4f(1) + f(2)]
= (2/3)114f(-2)
which is bounded by (2/3)M4.

9.6. Solution
It would be enough to do this in the two-point case but the characteristic pattern 1 424 1 first turns up in the 3 point case. We have, when
Chapter 9
223
b-a = 1,
T~l) = U(O) +!tG) +U(1)
T~2) = U(O) +um +um +tf(~) +U(1)
so that
= fi[f(O) + 4fm + 2fm + 4f(~) + f(1)].

9.11. Solution
Suppose 1Tn(x) = k"xn + .... Then, by orthogonality,
1Tn(x)k"xnw(x) dx
r(
1Tn(x)fw(x) dx
so that
1Tn(x)xnw(x) dx = k;:l.
If there was another orthonormal system {1T~l)(X)} where 1T~l)(X) =

k~l)xn + ... we would have
f 1T~l)(x)xnw(x)
b
dx
= (k~l)-l.
This implies that
1Tn (X)1T~l)(X)W(x) dx
= k,Jk~l) = k~l)/k"
so that
k~
= (k~1)2
and, both being positive, this gives kn = k~l).

It follows that
(1)
r
a
[1Tn(X)-1T~1)(x)]2w(x) dx = 1-2+ 1 = O.
The result (1) implies that 1Tn(X)=n~l)(x) if we assume that

w(x) is not zero in any subinteroal of [a, b].
This condition is satisfied in all the classical cases. To establish uniqueness

we proceed as follows.
224
If 1Tn (X) ~ 1T~l)(X) then, these being polynomials and so continuous,

there will be an interval (c, d) included in (a, b) such that in it
l1Tn(x) -1T~l)(x)l;::: 8 > O.
Hence
a contradiction.
9.12. Solution
Let Xl, . .. , Xn be the zeros of 'lTn(x). Let H(x) be the Hermite interpolant introduced in Problem 8.12. If we mUltiply the error estimate (11),
(p. 91), across by w(x) and integrate between [a, b] we get
fb
a
(f(x)-H(x))w(x)dx=
1b f2n)(~(x))
(2n)!
n
!1(x-xY w(x)dx.
Since H(x) is of degree 2n -1, and H(X;) = !(X;), we have

H(x)w(x) dx
= LAiH(X;) = LAJ(X;).
Thus
1- Q
= rb f 2n )(g(x))
Ja
f(2n)(g)
= (2n)!
[fIi~l (x - :x;f]w(x) dx
1 !1
(2n)!
b [
]
(X - X;? w(x) dx
using the Mean Value Theorem, since the last integrand is positive.
In the Chebyshev case we have
= (2n)!22n- 2
1T
(2n)!2 -
=---....,.
2n 1
r'IT cos2 nO dO
To deal with the Legendre case we note that it can be verified (e.g., by
integration by parts, Apostol II, 179) that if
Pn(x) = [1/(2 n(n !]Dn{(x 2-1)"}
Chapter 9
225
then
Also
Hence
and so the coefficient of
f 2n )(e) is
9.13. Solution
(a) The indefinite integral is
1
x 2-x+1 1
xJ3
-log
+-arctan-6
(x+1)2 J3
2-x
when -1 < x < 2. Hence the definite integral is
-log 4 + arctan J3
6
J3
_log 2 + 1T~ = 0.60459 9788 _ 0.230149060

3
3,,3
= 0.374450728.
(b) The indefinite integrals are, respectively:
+-log
5
(1
) J(10+2v'5)
+x +
10
arctan
4x+v'S-1
4x-v'S-1
arctan
;
J(10+2v'5)
J(10-2,J5)
226
The definite integrals, between 0 and 1, are respectively:

../2
[log (3 +../2) + Tr] = 0.86697299;
J"S1 3+J"S 11 2 J10+2J5

3+J"S
- og - - + 5 og +
arctan -;::::==::::::
10
3-J"S
10
J10+2.j5
JlO-2.j5
10
3-J"S
arctan -;::::====
JlO-2.j5
J"S
TrJ"S
= Slog !(J"S + 1)+! log 2+50 J(10+2.j5) = 0.88831357;
IOg(~J3) +~=0.90377177.
236
The general results of which these are special cases are due to A. F.
Timofeyev (1933)-see, e.g., formula 2.142 in I. S. Gradshteyn and I. M.
Ryzhik, Table of integrals, series and products, Academic Press, 1965. See
also I. J. Schwatt, An introduction to the operations with series. Philadelphia,
1924.
9.14. Solution
A short table of this integral in which the argument of y is in degrees is
given in NBS Handbook p. 1001. When y =0.5 in radians we find
S(0.5, 0.5) = 0.296657503.
The differences between the given estimates are 9312, 253, 1 in units of
the tenth decimal. The ratio 9312/253 is about the fourth power of the ratio
of the intervals. We may suspect that Simpson's Rule, or some other of the
same accuracy was used. We can use Richardson extrapolation to get a
correction of
16
(625 -16) x 9312 = 245
leading to an estimate of 0.2966575002.
9.15. Solution
These integrals were evaluated by E. Brixy (ZAMM 20 (1940), 236238); unfortunately some of his results seem wrong. It is desirable to draw
rough graphs of the integrands and to check that the integrals all exist. Note
Chapter 9
227
that, e.g., the integrand in 16 is of the form 010 at x = 0 and so some minor
adjustment in the program may be necessary.
Since JMx)=-J1 (x) we have
Is = -
[Jb(x)IJo(x)] dx
= [log [JO(X)]-I]~ = log [lIJo(1)] = 0.267621.
Since xl2 (x) = J 1(x)-xlHx) we have
I6.10
= r1J (x)-xlHx) d = rl!!:... [1 {_x }] d
xll(X)
x.lo dx og J 1(x)
x
1
= [log
{J ;x)}I
1
= log [1IJ1 (1)]-10g 2 = 0.127718.

Since xlO(x)=2J1(x)-xlz{x) and since, as we have just seen in the
discussion of h, [Jz{x)IJ1(x)] = [log {xIJ1(x)}]' we have, on integration by
parts,
14 =
(2- x [log {xIJ1(x)}],) dx
= 2-[x
10g{xIJl(x)}~+
log {xIJ1(x)} dx
= 2+ log J1(1) + log 2+ 17

This result can be used as a check on the values of 14 , 17
11 = 1.09407,
12 = 2.08773,
13 = 0.08673 4,
14 = 1.91448 5,
17 = 0.04220 4,
Is = 0.02795 4.
9.16. Solution
If 0:5 8 :5~1T then 0 :5 cos 8:51 and so
which implies I 2n - 1 2:: I 2n 2:: I 2n + 1 where
COS2 n- 1
82:: cos2n 8 2:: COS2n+ 1 8
It is easy to find a reduction formula for this integral: in fact

nIn
= (n -1)1,.-2,
n 2:: 2,
and 10 = ~1T, II = 1. These give
2n + 1
(2n(n!))2
= (2n + 1)1"
228
Inserting these values in the inequalities
we get
[2n- 1 n-1)1)]2
------'->
(2n-1)!
'1T
[2n(n!)]2
22n(n!)2 2
(2n+1)!
(2n)!
. ->"---~
which gives
1 ~> (2n)!(2n + 1)! '1T> 1
+2n
24n(n!)4 2which is the result required.
9.17. Solution
This is established in a way similar to that used to get Wallis' Formula.
We write
In =
tn dt
(1- t4 )1/2
and observe that In decreases as n increases. It is easy to show that 13 = ~ for

[-~(1- t 4)1I2]' = t 3(1- t 4)-1I2.
We next obtain the reduction formula
In+3 = [n/(n + 2)]In -
We then use this, and the facts that
to conclude that
and hence that
Use of the reduction formula gives
I
4.
1.5 ..... (4r-3) L

3.7 ..... (4r-1) 0,
- 3.7 ..... (4r-1) I

4.+2 - 59
...... (4r+1 ) 2
and
1
=!. 2.4 ..... (2r)
4.+3 2 3.5 ..... (2r+ 1)"
Chapter 9
229
We now observe that

1012 = (4r+ 1)14,14r+2 = (4r+ 1)(14r +3)2 X(14,/14r +3) X(14r+~14r+3)'
We let r~oo and the last two factors tend to unity and the first, because
of Wallis' Formula, tends to ~1T.
9.18. Solution
1=
X4 dx = (1/5)(b 5-a 5) = (l/5)(b - a)[a 4+a 3b +a 2b 2+ ab 3+ b4].
Q = 3~:3a) [a 4+3e a3+ br

=
+3(a~2br +b 4]
(b -a) x[27a4+27b4+(16a4+32a3b +26a2b2+8ab3+ b4)

8x27
+(a 4+8a 3b +24a 2b 2+32ab 3+ 16b 4)]
(b-a)
2x27
= - - [lla 4+ 10a 3b + 12a 2b 2+ 10ab 3+ llb 4].
Hence the actual error is:

1-Q= -(b-a)[ 4-4 3b+6 2b2-4 b3+b4]= -(b-a)5
270
a
a
a
a
270 .
The estimate is
In the case of N panels, the estimate will be
where M4 is maxasxs" If(x)l.

If we take M4 to be 1 then we have to choose N to make
5 x64 xN4> e- 1
i.e., N>(6480 e)-1/4. E.g., with e = 10-8 we must take
N> (6480)-114 X 102 == 102 /9 == 11.
and so about 34 evaluations of the integrand. Note that if we change M4 to
10 or to 0.1 we only change N by a factor of 101I4~ 1.78 or 0.60.
230
9.19. Solution
We have to show that
r
r
J=
~(x)(1-li(x))w(x) dx = o.
Since l;(.xi ) = 1 it follows that (x - x;) divides 1- 4(x), the quotient being a
polynomial of Qn-2(X) of degree at most n - 2. Thus
J=
{(x-x;)~(X)}Qn_2(X)W(X) dx
and, as the quantity in braces is a multiple of '1Tn(x), it follows by orthogonality that J = O.

9.20. Solution
Since If4 )(x)1 $1, the absolute values of the errors in Simpson's Rule
and in the i-Rule are bounded by
h 5 /90
and 3h 5 /80,
h being the interval. Suppose we use N panels. Then the total error in the
Simpson case is
N. (b-a)5.~
2N
90
and that in the i-case is
If the error is to be less than e(b -
a)5 we must have
N=~1/2880e
and
N = ~1/6480e.
The cost, in terms of evaluations of the function [, is
The basic error estimate in the Gauss-Legendre case is (Problem 9.12)
when we use the interval (-1, 1). For a general interval (a, (3) we have to
231
Chapter 9
f:
write
f(x) dx =
t:l f(i(3
-a)t+!(a + (3)). !(3 -a) dt =!(3 -a)
t:1
F(t) dt
where t=(2x-a-(3)/(3-a). As pr)(t)=[!(3-a)]rfr)(x) our error estimate should be mulitplied by [!(3 - a) fn+l. For n = 2 we find
Returning to our main problem, if we again use N panels, the total

error will be
N
(b
-a)5 4320
1
and the relevant value of N is

N =~1/4320B.
There are two evaluations per subinterval (at "awkward" abscissas) and so
the corresponding cost is
2N=~1/270B.
The relative efficiencies are therefore about

1.36 G-L.;
1.22 S; 1(i).
9.21. Solution
See NBS Handbook, p. 492. In particular
f(3)
= 1.3875672520-0.9197304101 = 0.4678368419.
9.22. Solution
We find
'YlO = 0.62638316;
'Y~O = 'YlO -
(1/20) = 0.57638 32;
'Y30 = 0.59378975;
'Y~o = 'Y30 -
(1/60) = 0.57712 ~O.
At the first stage
At the second stage
232
9.23. Solution
This is a table of P4 (x) = (35x 4 -30x 2+3)/8; the endings 4,6 are
exactly 375,625. The exact value of the integral is
[(7x 5 -10x 3 + 3x)/8]A2 = 0.46728.

Simpson:
(Milne)3:
(3/8)4:
0.467352
0.466304
0.467439
We shall examine these results more closely. The errors in the first three
methods are estimated as multiples of a mean of the fourth derivative of the
integrand i.e., 105. The multiples are obtained in Problems 9.4, 9.2, 9.1. We
have to be careful in scaling. The observed errors are respectively
-72x10-6 ,
976 x 10-6 ,
-159x10-6
The theoretical errors are

3 x!~x 10-5 x 105,
i.e.,
-7X 10-5 ,
-15.75 x 10-5
in close agreement.
9.24. Solution
The relation given is ascribed to C. Hermite (E. W. Hobson, Plane

Trigonometry (1925), 378). Write for r = 1, 2, .. , n,
2 sin ra cos (nx -rx) = sin (nx +r(a - x + sin (-nx+r(a + x.
We then use the formula
sinO+sin(0+cp)+ ... +sin(0+(n-1)cp)=
cos (0 _lcp)-COS (0 +1(2n -l)cp)

2
2.
SID2CP
to get
{ }=cos(nx+!(a-x-cos(oo+!(a-x
...
2sin!(a-x)
+
cos (nx-!(a + x-cos (00 +!(a + x

sin 00.
2sin!(a+x)
Using the relation 2 cos A sin B = sin (A + B)-sin (A - B) four times and
the fact that
4 sin !(a - x) sin !(a + x) = 2(cos x -cos a)
Chapter 9
233
we find
2(cos x -cos a){ . .. }=sin (nx +a)-sin (n -1)x -sin (n+ 1)a + sin (00 - x)
+ sin (n -1)x -sin (nx - a) -sin (n + 1)a + sin (00 + x)
-2 sin na cos x +2 sin 00 cos a
= sin (nx +a)-sin (nx -a)-2 sin (n + 1)a + sin (00 + x)
+ sin (oo-x)
-2 sin 00 cos x +2sin 00 cos a
= 2 cos nx sin a -2 sin 00 cos a -2 sin a cos 00
+2sinoocosx
-2sin 00 cos x +2 sin 00 cos a
= 2 sin a (cos nx -cos 00),
which gives the result required.
If we take a = 8m = !(2m -1)7r/n, x = 8 we get
cosn8
1 {
.
8
8 = -.-8- (_1)m-l + 2 sm (m -1}Om cos 8
cos -cos m sm m
+ 2 sin (m - 2) 8m cos 28 + ... + 2 sin 8m cos (m -1)8}
so that
(_1)m-l
a,. =
ao = sm
. 8m '
2 sin (m - r}Om
r = 1, 2, ... , m -1.
9.25. Solution
(a) For a popular method using double integrals see T. M. Apostol,
Calculus, II, 11.28, Exercise 16.
(b) The following method is given by A. M. Ostrowski, Aufgaben
sammlung . .. III, p. 51, 257.
We prove that
(1)
[f
e-t2 dtT = 11T -
e- x2(1+f2) dt/(1 + t 2 )
by observing first that both sides vanish for x = 0 since SA dt/( 1 + t 2 ) = 11T and
then noting that the derivatives of the two sides are identical: the derivative
of the right hand side is
-( -2x)
e- x 2(l+f2) dt = 2
e-
x2
u2
= 2e- LX ex2
du,
u2
duo
if u = xt
234
If we now let x~oo in (1) the left hand side has limit [io e- t2 dtf. The
integral on the right, since t ~ 0, is bounded by J~ e-x 2 dt = e-x 2 and this

tends to zero as x~oo. Hence the limit of the right hand side is hr.
(c) The following method is due to A. J. Bosch, Nieuw Tijdschr. Wisk.
64 (1977), 210-211.
We use the results of Problem 9.16. Since
12n +1 = (Tr/2(2n + 1))IZ"~

and since
12n+2<12n+l <12n
we have
2n+ 1
Tr
1
2n +2 12n <2(2n + 1) lZ"n <12n
which gives
2n+l
nTr
2n + 2 (n11J < 2(2n + 1) < (n11J
so that
For O:sx:sl we have
l-x 2 <e-x2 1 +X 2)-1
and taking n-th powers and multiplying across by In and integrating we get
.,/r"
(1- x2)n dx <.,/r"
e- nx2 dx =
L""e-
t2
dt <.,/r"
(1 + x 2 )-n dx.
Changing the variable in the integral on the left by x = sin (J shows that
r
nTr
lZ"~ Tr 2
"nI2n + 1 = 2(2n + 1) J;;~4
1 I
.;; = 2" Tr.
The integral on the right is less than
(where we use a change of variable x = tan 8). Hence the infinite integral
The numerical integration shows up some interesting points. Namely
Chapter 10
235
one can estimate the tail
by the formula of Chapter 7. This tail will certainly be less, e.g., than
10-6 if x ~ 3.5. We can therefore restrict ourselves to J~.5 e-x2 dx which
can be handled by any of the usual formulas.
If we use the trapezoidal rule it is found experimentally that remarkable
accuracy is obtained for quite large steps. For further discussion of this see,
e.g.,
E. T. Goodwin, Proc. Cambridge Phil. Soc., 45 (1949),241-245.
D. R. Hartree, Numerical Analysis, Oxford, 1952, 111.
~x
Chapter 10
10.1. Solution
The general solution is Un
= Aa n + B{3n
where a, {3 are the roots of
x 2 -x-1=0
and we obtain A, B from
1=A+B }
1=Aa+B{3
getting
Un
= 1 +v's (1 +v's)n + (-1 +v's)(1-v's)n.

2v's
2J5
The following values can be used as a check

U34
= 9227465
14930352
U36= 24157817
U37 = 39088169
U 38 = 63245986
U 39 = 1023 34155
U 35
(c. A. Laisant, L'enseignement math. 21 (1920), 52-56).
10.2. Solution
We deal with the general case. Suppose we have obtained
IIn (x) = k"x n + k~xn-l + ....
236
Then, if
A.. = kn+1/k..,
is a polynomial of degree at most n and can be expressed as a linear
combination of 110 , 1110 ... I1n. The coefficient of I1n is
The coefficient of I1 n- 1 is
I1n+ 1 I1n- 1 w dx -
A..
xl1nl1 n- 1 dx
= 0- A..
I1n(xl1 n- 1) dx
= -Ank..-1/k.. = -k..+1k..-1/k~.
The coefficients of the earlier l1's vanish by orthogonality. Hence we have
The results for special cases can be obtained from NBS Handbook, p.
782 and include
(a)
(n + 1)Pn+1 = (2n + 1)xPn - nPn-l>
(b)
(c)
(d)
(n + 1)L,. = 2n + 1)- x)L,. -
nL..-1o
10.3. Solution
The relation is equivalent to
Vn
aUn +b,
k(1-k) = ac-b 2 +b.
Convergence can only take place to k or 1- k. We may assume that k~!

when it is real. The following gives a complete answer:
(1) k not real, Vn ~ 00 monotonically
(2) IVol> k, Vn ~ 00 monotonically
(3) IVol> k, !::; k ::;~, Vn ~ 1- k and monotonically if !::; k::;1

(4) Ivol < k, ~ < k ::; 2, Vn oscillates finitely
(5) IVol<k, k>2, Vn ~oo except when Vo belongs to a certain set (cardinal
c, measure zero) in which case it oscillates finitely.
231
Chapter 10
For further details see:

T. W. Chaundy and Eric Phillips, Quarterly Journal (Oxford), 7 (1936),
74-80.
10.4. Solution.
For further details see E. Hansen, Signum Newsletter 3 (1968), #3.
19 = 0.09161 23,
110 = 0.0838771.
10.7. Solution
The exact solution to this equation is y = (1- x)-t, with a pole at x = 1.
Consider integrating at an interval h = N- 1 and write x,. = rh, y, = y(rh)
where
Y,+1 = y, + hy;.
We have
1
1
h
1
h2 y,
-=----=--h+-y r+1 y, 1 + hy, y,
1 + hy,'
= 0, 1, ... , n -1.
Summing we get
1 1
r-l
y
-=--rh+h 2
--s-=1-rh+hR, say.
y, Yo
s=O 1 + hys
'
We shall now estimate the R" which obviously increase steadily from Ro = 0
to RN Oearly
hy,
1
1
R,+1-R,=1+hy, =h-1 -r+R,+1 N+1+R,-r
We use the method of the "Integral Test". First, neglecting the positive
quantities R, we have
RN<
N-lL
r=O
1
<
N+1-r
iN
dx
N+1-x
log (N+1).
Next,
RN>
~1
,~o N +1 + log (N+ 1)-r >
f~l
1
N +1 + log (N+1)-x
=10 N +10g(N+1)
g log (N +1)+2'
We have therefore shown that

RN = log N +O(log log N)
= log h- 1 + O(log log (h- 1 ))
238
and hence
1
{ 1+ ~loglog(h-l)} .
h log h- 1
log (h- 1 )
For further details see

M. L. 1. Hautus and G. W. Veltkamp, Nieuw Arch. Wisk. {3}, 17 (1969),
79-80.
10.8. Solution
In practice there will be round-off so that, e.g., 100 times 0.001 will not
be 0.1 exactly. A refined program will allow for this and take a small step
(backwards or forwards) before printing so as to get the argument 0.1 (or, as
near this as the machine allows!).
Library programs will often have an automatic choice of an appropriate
interval incorporated.
The solution of y' = x 2 - y2, y(O) = 1 has the value 0.750015703 at x = 1.
See Appendix for details.
The results from a Univac computation for the equation
y' = !"'~ +
!"'y,
y(O) =0
are given below.

x
h =0.00025
h =0.00125
h =0.0025
h =0.0125
0
0.0125
0.0250
0.0375
0.0500
00000000000
00054793700
00159912174
00299888889
00469064935
00000000000
00054084766
00159108463
00299020417
00468145066
00000000000
00052730731
00157549032
00297323965
00466341119
00000000000
00034938562
00135795112
00273099005
00440237794
0.0625
0.0750
0.0875
0.1000
0.1125
00664121481
00882794453
01123406696
01384648727
01665459906
00663158054
00881792703
01122370425
01383580841
01664362708
00661263636
00879819092
01120325726
01381471217
01662193028
00633609674
00850825638
01090141626
01350207833
01629937288
0.1250
0.1375
0.1500
0.1625
0.1750
01964957501
02282391333
02617113183
02968555308
03336214844
01963832876
02281240853
02615938188
02967356957
03334994148
01961607102
02278962286
02613609625
02964980810
03332572516
01928429044
02244919744
02578751306
02929348394
03296202153
Chapter 10
239
h =0.00025
h =0.00125
h =0.0025
h =0.0125
0.1875
0.2000
0.2125
0.2250
0.2375
03719642145
04118431874
04532216083
04960658706
054034511"39
03718399993
04117169062
04530933327
04959356649
05402130364
03715934729
04114661815
04528385574
04956769724
05399505479
03678858123
04076907044
04489977709
04917731319
05359856951
0.2500
05860308623
05858969672
05856307932
05816067873
This is a classical problem studied by Moigno in 1844. Awkwardnesses

arise because the expansion of the solution around the origin is of the form:
10.9. Solution
The solution is y = e E 1 (x) and the following values are obtained from
NBS Handbook, p. 243:
x= 1 0.596347361
2 0.361328617
3 0.262083740
4 0.20634 5650
5 0.177297535
10 0.0915633339
20 0.04771 85455
X
10.11. Solution
Writing z = y' we replace our scalar differential equation by a vector
differential equation
where
[~]
is given for x = 0:
[;,~~J.
We have
Y*(1) = yeO) + hz(O)

z*(1) = z(O)+ hf(O, yeO), z(O))
and
Y**(1) = y(O)+ h[z(O)+ hf(O, yeO), z(O))]

z**(1) = z(O)+ hf(h, y(O)+ hz(O), z(O)+hf(O, yeO), z(O)))
giving
Y(1) = yeO) + hz(O) +1h 2 f(0, yeO), z(O))

z(1) = z(O) +1hf(0, yeO), z(O)) +1hf(h, yeO) + hz(O), z(O)+ hf(O, yeO), z(O)))
240
Clearly y(1), z(1) only depend on y(O), z(O), hand f. We have to evaluate f
twice, once with arguments 0, y(O), z(O) (which we use three times), and
once with arguments
h,
z(O) + hf(O, y(O), z(O)).
y(O)+ hz(O),
We are free to change h whenever it is desirable without change in the

formulas.
Equation (14) has been studied by P. Painleve - the solution becomes
singular, as in the simple case of
y(O)=A
which has a solution y = (A - X 2)-1. Thus the problem is one of "polar

exploration". The Painleve equation in the case y'(O) = 1 and on the real
axis has a series of poles of order 2 of which the first two are at
1.70 ... ,
2.97 ....
10.12. Solution
Discretizing using a mesh size h = 1, we get a system of three
homogeneous linear equations for y(l), y@, y(~):
(y(0))-2y(1)+y@ = -l~ . lAY (1)
y(1)-2y@+ y(~) = -l~ ~AY(~)
y@-2y()+(y(1))=
--h. AY()
Note that y(O) = y(1) = O.

In order that we should have a nontrivial solution the determinant of
this system must vanish, i.e.,
det [
(A -128)
64
32
(A -64)
64
3"
o ] =0.
32
(A _1~8)
This reduces to
A3_ A2[11 ;64] + A[128 X364 X 5] _ [64X6;X 128] = O.
Writing A = 64A we get
i.e.,
241
Chapter 10
The last equation has an obvious root A = 1 and the residual equation is
3A2 -8A+2=0
which has roots A = (4JiO)/3. Thus the roots are
Al = 64(4 - JiO)/3 = 17.8688,
A3 = 64(4+ JiO)/3 = 152.794.
The corresponding values of y(~), y@), y(i) are:
2+~J' [~], [2-~J.
[ 2+M
-1
-2+M
-2+M
2+M
These have sign patterns (+++), (++-), (+-+).

Integrating the differential equation with A = 17, 18, 19 enables us to
get a better estimate for the smallest eigenValue.
In order to check our integration subroutines we proceed as follows.
Integrate the differential equation with
A = 18.6624,
18.9225,
19.1844,
19.4481
The corresponding values of the solution at x = 1 should be

0.01066,
0.00122,
-0.00816,
-0.01747
and we interpolate inversely by the Aitken method

1066 6624
122 9225 9561
-816 11844 581 564
-1747 14481 601 564
giving A = 18.9564 which is the correct result.
The motivation of our choice of the four A'S is the following: more
details are given in the Appendix. The solution to
y"-AXY =0,
is a multiple of
which has the value
y(O) = 0,
y'(O) = 1
242
when x = 1. We want to choose A so that this vanishes. The zeros of J 1/3

have been tabulated in the NBS volume cited in the Appendix and are, to
4D,
2.9026,
6.0327,
9.1705, ....
These correspond to the following values of A:

18.9564,
81.8844,
189.219, ...
(which we can compare with the roots of the cubic). It is Al = 18.9564 with
which we are concerned.
The assigned values of A correspond to the following values of ~A 4
~(18.6624)1/2 =~(4.32) =
2.88,
2.90,
2.92,
2.94
The values of J 1/3 (X), x = 2.88(0.02)2.94 are 0.01066, 0.00122, -0.00816,

-0.01747.
10.13. Solution
[For a more theoretical discussion of this problem see, e.g., p. 279 of G.
Birkhoff and G. C. Rota, Ordinary differential equations, Ginn, Boston,
1962.]
This is a simple form of the Schrodinger equation. Consider the case
when V is even, vanishing outside [-1, 1]. We ask whether there are values
of E for which there are nontrivial solutions (Le., I/I 0) which are even, or
odd and which are such that I/I(x) ~ 0 as x ~ oo.
Since 1/1 is required to be even or odd we can restrict our attention to
x ~ O. When x ~ 1 the equation reduces to
I/I"(x) + EI/I(x) = 0
with general solution

l/I(x)=A exp(vCE(x-1+B exp(J.=E(1-x.
In order that cp(x) ~ 0 as x ~ 00 we must have E < 0, A = O.

When 0::S;X:51, the solution depends on the form of V(x). However it
will involve two arbitrary constants, say 1/1(0), 1/1'(0). Now if 1/1 is even,
1/1'(0) = 0 and if 1/1 is odd, 1/1(0) = O. Hence one constant is determined and the
solution is of the form
I/I(x) = Ccp(x)
where C is arbitrary and cp involves E.

To match the solutions at x = 1 we must have
Ccp(1) = 1/1(1) = B,
Ccp'(1) = 1/1'(1) = -r-EB
Chapter 10
243
which can be satisfied if and only if

T(E) == J=Bp(1) + cp'(1) = O.
This is the equation which determines the characteristic values (if any)
of E. We now return to the special V(x) of the problem and note that an
odd solution of
I/I"(x) +(E +5)I/I(x) = 0
is
cp(x) = sin.JE + 5x
for which cp(1) = sin.JE + 5, cp'(1) =.JE + 5 cos.JE + 5.

The characteristic values are determined by
Hsin .JE+5+.JE+5 cos.JE+5 =0
i.e.,
tan.JE+5= -
~E+5
E.
Remembering that E is negative we consider the behavior of the two

sides of the last equation in the range -5:5 E :5 0 where they are real. The
left side increases from 0 to 00 in (-5, -5 +hr2) and from -00 to tan J5 in
(-5 +i1T2, 0). The right side decreases from 0 at E = -5 to -00 at E = O.
There is exactly one intersection near E = -1. This can be estimated by
Newton's method. If
f(E) = tan.JE + 5 +.J(E + 5)/E
then
f(-1) = tan 2+2,
and the next approximation is

-1- 4(2+ tan 2) ~-1 + 0.74016 . -09 1
5 +sec2 2
10.7744 .
. 3 .
10.14. Solution
In general it is not possible to obtain I/I(x) explicitly as in Problem
10.13. We must proceed as follows. According as to whether we want an
even or odd solution we integrate the equation
I/I"(x) = (V(x) - E)I/I(x)
244
numerically from 0 to 1, beginning with 1/1(0) = 0, 1/1'(0) = 1 or 1/1(0) = 1,

1/1'(0) = 0, for some value of E. We then check whether
T(E) = HI/I(I) + 1/1'(1)
is or is not zero. If it is not zero we try to get a better guess for E and repeat
the process until we get T(E) close enough to zero.
Various artifices can be used to determine E, e.g., inverse interpolation
as on p. 133.
The result is that there is an even solution corresponding to E ~
0.337.
10.15. Solution
The table is meant to be one of (3x+912)3-38299 99877.
There were errors in the entries for arguments 217, 218, 219: the
corrected values are respectively-116 39330; 10389619; 325 03132
10.16. Solution
Compare Problem 7.7.
y(1.5) = 0.7127926,
y(2.0) = 0.6429039.
10.17. Solution
_anf(n+(b/a
)
a . x,. f(b/a)
Xo
b). x,. = (Ain + Bin+l)/C
where
A = {cio + (a + b )it}xo - ilXl>
and
J.n = (_c)1/2nJ n+(b/a) (2a- 1(-c)I/2)
See D. H. Lehmer, Proc. Conference Numerical Mathematics, Winnipeg

(1971), 15-30.
10.18. Solution
f(7) =2983
10.19. Solution
a) These are the odd prime numbers.
b) If Pn is the nth prime number of the form 4r+l so that PI =5,
P2 = 13, P3 = 17, . .. then the sequence is that of the least integer N for
Chapter 10
245
which Pn is a factor of W+1. Thus 22+1=5, 52+1=2.13, 42+1=17,

122+ 1 = 5.29, ...
c) These are the values of n 2 - n +41 for n = 0,1,2, ... and are all
prime numbers until n = 41.
10.20. Solution
Ai'(l.4) = -0.10850 959

Ai"(l.4) = 1.4Ai(1.4) = 0.11485327
10.21. Solution
See preceding solution, or British Assoc. Math. Tables, Part Vol. A.

10.22. Solution
1= e-t4 dt =!1= e-T . T- 3/4 dT = !f(!) = 0.906402479.
x(O) =
2t2e-t4dt=~r e-Tl1!4dT=~f(~)=0.612708351.
x'(O) =
Differentiating we find
y"=x 2y+
t(4t3-3xt)exp(-~x2+2xt2_t4)dt
-1=
exp (-h 2+2xt2- t4) dt.
Integrating by parts the second term on the right we find
{= ... dt
[-t exp {!x 2+2xt 2- t4 )Jo+
exp (-~x2+2xt2- t4 ) dt
The first term on the right vanishes at both limits and the second cancels the
third term on the right in the previous display. Hence
The solution is
and for x = -2(1) 8 this has values

0.0078, 0.2059, 0.9064, 0.9869, 0.6702, 0.5241, 0.4487,
0.3637, 0.3363, 0.3143.
246
10.23. Solution
See, British Assoc. Math. Tables, Vol. 2. For recent theoretical work
see papers by E. Hille.
10.24. Solution
See, J. L. Synge, On a certain non-linear equation, Proc. Royal Irish
Academy 62A (1961), 17-41. Z. Nehari, On a non-linear differential
equation arising in nuclear physics, Proc. Royal Irish Acadymy 62A (1963),
117-135.
10.25. Solution
We use an interval h = 0.1 and work to 4D. From the power-series
y(x) = 1 + x + x 2 + (x 3 /3) + (x4/12) + ...
we compute y(0.1), y(0.2), y(O.3) and [(0.1), [(0.2), [(0.3) and we then
predict
YP (0.4) = 1 + (4/30)[2.4206 -1.4428 + 3.3994] = 1.5836
We then compute [(0.4) = 1.9836 and check by

Yc (0.4) = 1.2428 + (1.30)[1.4428 + 6.7988 + 8.9836] = 1.5836.
We can accept this value and proceed, or alternatively, try a larger h = 0.2.
x
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
YP
1.0000
1.1103
1.2428
1.3997
1.5836
1.7974
Yc
y'=[(x,y)=x+y
1.5836
1.7974
1.0000
1.2103
1.4428
1.6997
1.9836
2.2974
The correct value at x = 1 is 2e. - 2 = 3.43656 ...
RECOMMENDED LITERATURE
T. M. APoSTOL, Calculus, I, II (Wiley, 1967-9).

A. M. OsTROWSKI, Vorlesungen uber Differential- and Integralrechnung, 3
vols., Aufgabensammlung zur Infinitesimalrechnung, 3 vols.
(Birkhauser, 1965-72).
TESTS AND
MONOGRAPHS
F. S. ACTON, Numerical methods that work (Harper and Row, 1970).

E. K. BLUM, Numerical analysis and computation; theory and practice
(Addison-Wesley, 1972)
C. BREZINSKI, Acceleration de la convergence en analyse numerique
(Springer, 1971).
S. D. CONTE and C. DE BOOR, Elementary numerical analysis (McGraw-Hill,
1973).
G. DAHLQUIST and A. BJORCK, tr. N. ANDERSON, Numerical methods
(Prentice-Hall, 1976).
J. W. DANIEL and R. E. MOORE, Computation and theory in ordinary
differential equations (Freeman, 1970).
P. J. DAVIS and P. RABINOWITZ, Methods of numerical integration (Academic
Press, 1975).
C.-E. FROBERG, Introduction to numerical analysis (Addison-Wesley, 1965).
R. W. HAMMING, Introduction to applied numerical analysis (McGraw-Hill,
1971).
H. M. NAUTICAL ALMANAC OFFlCE, Interpolation and allied tables (H. M.
Stationery Office, 1956).
P. HENRlCI, Elements of numerical analysis (Wiley, 1964).
P. HENRlCI, Discrete variable methods in ordinary differential equations
(Wiley, 1962).
P. HENRlCI, Computational analysis with HP 25 pocket calculator (Wiley,
1977).
248
F. H. Hn.DEBRAND, Introduction to numerical analysis, 2nd ed. (McGrawHill, 1976).

A. S. HOUSEHOLDER, Principles of numerical analysis (McGraw-Hill, 1953;
Dover, 1974).
A. S. HOUSEHOLDER, The numerical treatment of a single non-linear equation
(McGraw-Hill, 1970).
E. ISAACSON and H. B. KELLER, Analysis of numerical methods (Wiley,
1966).
L. M. Mn.NE-THOMSON, The calculus of finite differences (Macmillan, 1933).
Modem Computing Methods (H. M. Stationery Office, 1961).
B. R. MORTON, Numerical approximation (Routledge and Kegan Paul,
1961).
B. NOBLE, Numerical methods, 2 vols. (Oliver and Boyd, 1964).
A. M. OsTROWSKI, Solution of equations, 3rd ed. (Academic Press, 1973).
A. RALsTON and P. RABINOWITZ, A first course in numerical analysis, 2nd ed.
(McGraw-Hill, 1978).
T. J. RIvuN, The Chebyshev polynomials (Wiley, 1976).
H. RUTISHAUSER, Vorlesungen uber numerische Mathematik, 2 vols.
(Birkhauser, 1976).
H. RUTISHAUSER, Numerische Prozeduren - ed. W. Gander, L. Molinari and
H. SvecovR (Birkhauser, 1977).
F. ScHEID, Theory and problems of numerical analysis, (Schaum-McGrawHill, 1968).
L. F. SHAMPINE and R. C. ALLEN, Numerical computing: an introduction
(Saunders, 1973).
L. F. SHAMPINE and M. K. GORDON, Computer solution of ordinary differential
equations. The initial value problem (Freeman, 1975).
E. STIEFEL, tr. W. C. and C. J. RHEINsolDT, An introduction to numerical
mathematics (Academic Press, 1963).
J. STOER, Einfuhrung in die numerische Mathematik, I 2nd ed. (Springer,
1976).
J. STOER and R. BUURSCH, Einfuhrung in die numerische Mathematik, II
2nd ed. (Springer, 1978).
JOHN TODD, ed., Survey of numerical analysis (McGraw-Hill, 1962).
JOHN TODD, Numerical analysis, Chapter 7, Part 1 of E. U. Condon-H.
Odishaw, Handbook of Physics, 2nd ed. (McGraw-Hill, 1967).
JOHN TODD, Introduction to the constructive theory of functions (Birkhauser,
1963).
J. F. TRAUB, Iterative methods for the solution of equations (Prentice-Hall,
1966).
249
B. WENDROFF, Theoretical numerical analysis (Academic Press, 1966).

E. T. WHfITAKER and G. ROBINSON, The calculus of obseroations, 4th ed.
(Blackie, 1966).
J. H. WILKINSON, Rounding errors in algebraic processes (Prentice-Hall,
1963).
D. M. YOUNG and R. T. GREGORY, A suroey of numerical mathematics, 2
vols. (Addison-Wesley, 1972-73).
TABULAR MATERIAL
M. ABRAMOWITZ and I. A. STEGUN, ~s. Handbook of mathematical functions, National Bureau of Standards, Applied Math. Series, 55 (U.S.
Government Printing Office, 1964).
BARLOW'S TABLES, ed. L. J. Comrie (Spon, 1961).
L. J. COMRIE, Chamber's Shorter six-figure mathematical tables, (Chambers,
1950).
A. FLETCHER, J. C. P. MILLER, L. ROSENHEAD and L. J. COMRIE, An index of
mathematical tables, 2nd ed., 2 vols. (Addison-Wesley, 1962).
In addition to the literature mentioned above, there are many useful
expository articles available, and for up-to-date surveys of special topics,
reference can be made to Symposia Proceedings, such as those which appear
in the ISNM series.
There is a developing interest in the history of numerical mathematics
and computing machines. For the classical material see, respectively.
H. H. GOLDSTINE, A history of numerical analysis from the 16th through the
19th century (Spnnger, 1978),
B. RANDELL, The origins of digital computers, 2nd ed. (Springer, 1975).
For more recent history see the obituaries of the founders and the
periodical Annals of the history of computing, 1979-.
Contents
Vol. 2, Numerical Algebra
Preface .
Chapter
Chapter
Chapter
Chapter
Chapter
Chapter
Chapter
Chapter
Chapter
Chapter
Chapter
Chapter
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
Manipulation of Vectors and Matrices

Norms of Vectors and Matrices . . .
Induced Norms . . . . . . .
The Inversion Problem I: Theoretical Arithmetic
The Inversion Problem II: Practical Computation.
The Characteristic Value Problem - Generalities.
The Power Method, Deflation, Inverse Iteration. .
Characteristic Values. . . . . . . . . . . . . .
Iterative Methods for the Solution of Systems Ax = b
Application: Solution of a Boundary Value Problem.
Application: Least Squares Curve Fitting. . . . . .
Singular Value Decomposition and Pseudo-Inverses
.13
.16
.19
.29
.44
.53
.65
.71
.83
.99
105
110
. 117
.212
Index . . . . . . . . .
.214
INDEX
M. Abramowitz 141, 249

Acceleration methods 66, 68
A. C. Aitken
algorithm 86, 87, 241
transform 67
G. B. Airy integral 124, 129, 137, 150, ISS,
157,245
Algorithms
Borchardt 15, 166, 170
Carlson 16
Gauss 13
E. E. Allen 190
Arithmetic-geometric mean 13, 161
Asymptotic series 75
Bad examples 47, 88, 119, 130
F. L. Bauer 102
S. N. Bernstein 58, 211
polynomials 59, 63, 184, 187
F. W. Bessel 149
functions 64, 65, 115, 116, 119, 189,
227, 241
modified functions 145
Garrett Birkhoff 242
G. D. Birkhoff 91
J. L. Blue 180
R. P. Boas, Jr. 172
C. W. Borchardt 15, 163, 166
A. J. Bosch 234
E. Brixy 226
G. H. Brown, JI. 182
B. C. Carlson 16, 18,21,22,164,167,168
T. W. Chaundy 237
P. L. Chebyshev 186
coefficients 61
polynomials 60, 64, 186
E. B. Christoffel- J. G. Darboux formula
109
C. W. Clenshaw 62
L. Collatz 54, 181
Complementary error function 77
L. J. Comrie 81, 249
M. Davies 182
B. Dawson 182
R. Dedekind 37, 40
Deferred approach 100
Difference equation 118
Differencing 121
Elliptic integral 18, 21, 165
F. Emde 141
R. Emden equation 138
O. Emersleben 178
Error analysis
backward 48
forward 48
interpolation 89, 91
quadrature 98, 99, 105, 109, 113, 218,
229
Error function 77, 117
L. Euler 116
constant 31, 116, 231
method 125
transform 68, 69, 71, 72, 192, 193
Exponential integral 73, 81
R. M. Federova 82
W. Feller 87
Fibonacci 133, 179, 235
Fixed point 36, 38
A. Fletcher 81
FMRC 81, 202
G. E. Forsythe 46, 47
J. B. J. Fourier
coefficients 61
method 50
series 64
L. Fox 83, 202
A. J. Fresnel integrals 76, 202
Fundamental theorem of algebra 84, 213
P. W. Gaffney 214
Gamma function 79
K. F. Gauss arithmetic-geometric mean 161
K. F. Gauss - P. L. Chebyshev quadrature
105, 224
K. F. Gauss - A. M. Legendre quadrature
116,224,225,230
252
Walter Gautschi 121
A. Ghizzetti 219
G. A. Gibson 153, 155
E. T. Goodwin 82, 199, 235
I. S. Gradshteyn 226
J. P. Gram - E. Schmidt process 103
J. A. Greenwood 81
E. Halley 54, 182
E. Hansen 134, 237
P. A. Hansen 148
G. H. Hardy 151, 155, 172
H. O. Hartley 81
D. R. Hartree 138,135,245
C. Hastings 61
M. L. J. Hautus 135, 237
D. R. Hayes 219
P. Henrici 50
C. Hermite 232
interpolation 90, 224
K. Heun method 125, 135, 238, 239
E. Hille 246
K. E. Hirst 180
W. G. Horner's method 52
W. G. Hwang 43, 176
Instability 119, 130
Interpolation
Aitken algorithm 86
errors 89, 91
Hermite 90, 224
inverse 88, 241
Lagrange 84, 210
Newton 96, 216
spline 96, 215
Inverse interpolation 88, 241
Iteration 38
C. G. J. Jacobi relation 147
E. Jahnke 141
E. Kamke 124
A. N. Khovanski 180
K. Knopp 195
J. L. Lagrange interpolation 84
C. A. Laisant 235
E. Landau 24
J. Landen 22, 168
A. V. Lebedev 82
H. Lebesgue constants 91
A. M. Legendre
Index
expansion 184
polynomial 184, 236
D. H. Lehmer 244
G. W. Leibniz' Theorem 129, 207
Lemniscate constants 17, 109, 114
J. Liang 191
L. Lichtenstein - S. A. Gerschgorin equation 191
Local Taylor series 129
F. LOsch 141
E. C. J. von Lommel 149
Y. L. Luke 188
Mean value theorems
first 51, 89, 98, 99, 167, 224
second 151, 152
C. A. Micchelli 214
J. C. P. Miller 81
W. E. Milne quadrature 113
I. P. Natanson 211
NBS Handbook 76, 81, 92, 93, 94, 163,
164, 165, 178, 190, 198, 199, 200, 202,
231, 236, 239
Z. Nehari 246
E. H. Neville 87
I. Newton
process 39, 48
interpolation formula 216
J. W. Nicholson 151, 157
F. W. J. Olver 121
Order of convergence 29
Order symbols 24
Orthogonal polynomials 103
Chebyshev 104, 134, 236
Hermite 134, 236
Laguerre 104, 134, 236
Legendre 104, 134, 236
A. Ossicini 219
A. M. Ostrowski 47, 233
H. Pade fractions 64, 188
P. Painleve 240
J. F. Pfaff 19
Eric Phillips 237
E. Picard method 125
S. K. Picken 62
H. Poincare 75
Practical computation 42
Principal value integral 106
253
Index
Quadrature
Gaussian 103, 115
Lagrangian 98
Milne 113, 219, 232
Romberg 100
~ Rule 99, 113, 218, 229, 232
Simpson 99, 113, 222, 232
Trapezoidal 97, 221
Predictor - corrector 127
Recurrence relations 33
for reciprocals 33
for square root 37
Remainder Theorem 85
E. Ja. Remez 184
L. F. Richardson 66, 100, 126
T. J. Rivlin 214
G. Robinson 180
M. Rolle's Theorem 90, 212, 220
W. Romberg quadrature 100
L. Rosenhead 81
G. C. Rota 242
L. Rubin 219
H. Rutishauser 102
I. M. Ryzhik 226
D. H. Sadler 3, 83, 202
H. E. Salzer 215, 216
I. J. Schoenberg 19, 169
E. SchrOdinger 136, 242
J. Schwab 22, 169
I. J. Schwatt 226
Second mean value theorem 151, 152
R. M. Sievert integral 115
T. Simpson's rule 102, 113, 222, 230
Spline 96
J. Staton 82, 199
I. A. Stegun 141, 269
E. Stiefel 102
J. Stirling 18
formula 79, 111, 174, 211
G. G. Stokes 56
O. Stolz 153, 155
Stopping rule 42
J. C. F. Sturm - J. Liouville problem 124,
131
Subtabulation 95, 214
J. L. Synge 246
Synthetic division 53
Olga Taussky Todd 3
H. C. Thacher 121
Theoretical arithmetic 42
~ quadrature 99, 116, 229, 230
A. F. Timofeyev 226
E. C. Titchmarsh 152, 155
O. Toeplitz 195
UNIVAC 238
C. J. de la Vallee Poussin 153, 155
A. Vandermonde 95, 148, 213
G. W. Veltkamp 135, 237
H. S. Wall 44, 179

J. Wallis' formula 116, 229
S. E. Warschawski 191
G. N. Watson 141, 198, 203
B. Wendroff 219
E. T. Whittaker 180
K. T. W. Weierstrass' Theorem 58
A. van Wijngaarden 72
J. H. Wilkinson 47
S. Winograd 214
W. Wirtinger 151, 157
J. W. Wrench, Jr. 172

(International Series of Numerical Mathematics _ Internationale Schriftenreihe zur Numerischen Mathematik _ Série Internationale D’Analyse Numérique 14) John Todd (auth.)-Basic Numerical Mathematics_ .pdf

Transféré par

Informations du document

Description originale:

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

(International Series of Numerical Mathematics _ Internationale Schriftenreihe zur Numerischen Mathematik _ Série Internationale D’Analyse Numérique 14) John Todd (auth.)-Basic Numerical Mathematics_ .pdf

Transféré par

Droits d'auteur :

Formats disponibles

ISNM

INTERNATIONAL SERIES OF NUMERICAL MATHEMATICS

CIP-Kurztitelaufnahme der Deutschen Bibliothek

"Numerical Analysis" is for DONALD H. SADLER

Notations and Abbreviations

The algorithms of Gauss, Borchardt and Carlson.

Appendix: Bessel functions

Solutions to Selected Problems

Table of Contents, Vol. 2 Numerical Algebra.

Notations and Abbreviations

f ', f", f'" , (4) , ... , f n ), ....

The forward difference operators are defined as follows:

In the case when the interval of tabulation is h:

The promoter operator is defined by Ef(n) = fen + 1).

while f(x)= O'(g(x as

means that given any e>O there is an

Notations and Abbreviations

We use the standard notations, e.g., 9D, 7S to mean quantities given

The symbol x = a(h)a + nh indicates that x runs through the set of

L' ll;j = L l1jj

Factorials are denoted by n! = 1 x 2 x 3 x ... x nand semifactorials by

We use log or In to denote natural logarithms i.e., logarithms to the

There is no doubt nowadays that numerical mathematics is an essential

of the problems in Volume 1 can be dealt with using simple programmable

The Algorithms of Gauss, Borchardt and Carlson

We shall study several algorithms of the form

1. THE ARITIIMETIC-GEOMETRIC MEAN OF GAUSS

It is clear that if we interchange ao, bo we do not alter ~,bn for n ~ 1.

From the identity

it follows that ~+1 > bn + l unless ~ = bn This is the classical inequality of

a,.+1-bn +1= [./;;;'-,/bn < [a,. -bn

../X- JY and cross multiplying.

a,. - bn < o - bO]2"

a,. -bn < 2ao[(ao-bo)/(ao+bo)p" -+0.

10-9 <1 X 10-8 and so it will be sufficient

[1-10-4 ]2" :5e-20

[1_1O-4 poxlO" <e-20

The Algorithms of Gauss, Borchardt and Carlson

provided 2n > 20 X 104 = 200 X 103 , which is certainly true if n = 18 since

THE ALGORITHM OF BORCHARDT

We shall now make a small modification of the Gaussian algorithm.

Cln+l =!(a,. +f3n),

f3n > f3n+1 > Cln+l > a,..

f3n+1 - a,. +1 = { (.J

Cll = cos2 (J12,

Cl2 = cos (J12 cos2 (J12 2 ,

and establish by induction that

= cos (2- 10) cos (2- 2 0) ... cos (2- nO).

If we multiply across by 2n sin (2- n O) and use repeatedly the relation

= 2n - 2 cos (2- 1 0) ... cos (2- n +2 0)

= 0- 1 sin O[O.Jsin On]

where On = 2- nO. Clearly, since limx->o (x/sin x) = 1, we have

THE CARLSON ALGORmIM

Let non-negative ao, bo be given. Define for n = 0, 1, 2, ...

therefore converge and since

The Algorithms of Gauss, Borchardt and Carlson

We shall now determine the limit of the Carlson sequences in terms of

2 log JaJbn 2 log (aJbn)

2 log (aJbn) 2 log (ao/boY

is between a,. and bn Hence, for all n

We now let n -i>CX) and then a,. ~ I, bn ~ 1,

The arithmetic-geometric mean of Gauss is of the highest importance in

Theorem 1. If O:s bo = kao:S a o then