Borwein - Pi and The AGM (Wiley, 1987)

Pi and the AGM
A Study in Analytic Number Theory

and Computational Complexity
JONATHAN M. BORWEIN
PETER B. BORWEIN
Department of Mathematics
Dalhousie Urtiversity
Halifm, Nova Scotia
A Wiley-Interscience Publication
JOHN WILEY & SONS
NewYork
Chichester
Brisbane
Toronto
Singapore
Copyright 01987 by John Wiley & Sons, Inc.

All rights reserved. Published simultaneously in Canada.
Reproduction or translation of any part of this work
beyond that permitted by Section 107 or 108 of the
1976 United States Copyright Act without the permission
of the copyright owner is unlawful. Requests for
permission or further information should be addressed to
the Permissions Department, John Wiley & Sons, Inc.
Library of Congress Cataloging-in-Publication Data:
Borwein, Jonathan M:
Pi and the AGM.
(Canadian Mathematical Society series of monographs

and advanced texts = Monographies et Btudes de la
SociktB mathematique du Canada)
"A Wiley-Interscience publication."
Includes index.
1. Numbers, Theory of. 2. Computational complexity.
3. Functions, Elliptic. 4. Pi. I. Borwein, Peter B.
11. Title. 111. Series: Canadian Mathematical Society
series of monographs and advanced texts.
QA241.B774 1986
512' .7
86-15811
ISBN 0-471-83138-7
Printed in the United States of America
To our mathematician father, David Borwein
Five thousand digits of pi.

3
1415926535
8214808651
4428810975
7245870066
3305727036
8979323846
3282306647
6659334461
0631558817
5759591953
2643383279
0938446095
2847564823
4881520920
0921861173
5028841971
5058223172
3786783165
9628292540
8193261179
6939937510
5359408128
2712019091
9171536436
3105118548
5820974944
4811174502
4564856692
7892590360
0744623799
5923078164
8410270193
3460348610
0113305305
6274956735
0628620899
8521105559
4543266482
4882046652
1885752724
8628034825 3421170679
6446229489 5493038196
1339360726 0249141273
1384146951.9415116094
8912279381 8301194912
9833673362
0005681271
4201995611
5024459455
5982534904
4406566430
4526356082
2129021960
3469083026
2875546873
8602139494
7785771342
8640344181
4252230825
1159562863
6395224737
7577896091
5981362977
3344685035
8823537875
1907021798
7363717872
4771309960
2619311881
9375195778
6094370277
1468440901
5187072113
7101000313
1857780532
0539217176
2249534301
4999999837
7838752886
1712268066
2931767523
4654958537
2978049951
5875332083
1300192787
8467481846
1050792279
059731733
8142061777
6611195909
7669405132
6892589235
1609631859
7669147303
2164201989
3809525720
5574857242
8583616035
9331367702
6782354781
1065485863
4541506959
6370766010
8989152104
6360093417
2788659361
5082953311
4710181942
7521620569
2164121992
5338182796
6861727855
9555961989
6602405803
4586315030
8230301952
8890750983
4676783744
8150193511
2861829745
0353018529
8175463746
9448255379
2533824300
5570674983
6899577362
4939319255
7747268471
3558764024
8505494588
2599413891
0604009277
0404753464
7496473263
5869269956
2497217752
0167113900
6208046684
9141992726
9092721079
8347913151
9848824012
2590694912
0426992279
7509302955
3211653449 8720275596 0236480665 4991198818 3479775356 6369807426 5425278625 5181841757 4672890977 7727938000
8164706001 6145249192 1732172147 7235014144 1973568548 1613611573 5255213347 5741849468 4385233239 0739414333
4547762416 8625189835 6948556209 9219222184 2725502542 5688767179 0494601653 4668049886 2723279178 6085784383
igTz$W$gtS
~ ~ $ 7&?$WtR$
~ ~ ~ ~RtLWW!$ $~~~~$P%$~
lb3$f$2$13 atJ3zi$3$$ ~ , 5 $ $3+:GWOb+
From the eleventh iteration of

x.+' := ( x : J ~+ ~ n ' 1 ~ ) / 2
Xo
:= 2'12
48$27%i;$$: $b8$$;6lW
Preface
When I was a student, abelian functions were, as an effect of the Jacobian

tradition, considered the uncontested summit of mathematics and each of us
was ambitious to make progress in this field. And now? The younger
generation hardly knows abelian functions.
How did this happen? In mathematics, as in other sciences, the same
processes can be observed again and again. First, new questions arise, for
internal or external reasons, and draw researchers away from the old
questions. And the old questions, just because they have been worked on so
much, need ever more comprehensive study for their mastery. This is
unpleasant, and so one is glad to turn to problems that have been less
developed and therefore require less foreknowledge-even if it is only a
matter of axiomatics, or set theory, or some such thing.
And so there is nothing for it but to collect together the old subjects in
good references.. . so that later developments may continue them, if fate
should so decree.
Felix Klein (1849- 1925) [79 p . 2941
A central thread of this book is the arithmetic-geometric mean iteration
of Gauss, Lagrange, and Legendre. A second thread is the caicuiation of T .
The two threads are intimately interwoven and provide a remarkable
example of the application to twentieth-century computational concerns of
the type of nineteenth-century analysis whose neglect Klein so deplores. The
calculation of digits of rr has had a fascination that has far exceeded
utilitarian concerns-a fascination that has driven some to dedicate their
lives to calculations we may now electronically effect in seconds. The
methods that make the computation of hundreds of millions of digits of rr or
any elementary function within our grasp are rooted in the AGM and this is
where our interest in the material began. Making sense of this material took
us in three directions and motivated our writing this book.
The first direction leads to nineteenth-century analysis and in particular
the transformation theory of elliptic integrais. This necessitates at least a
vii
Preface
ix
tions helped produce a better book. The cheerful technical assistance of P.

Hemming, of our departmental office, and of the staff at Wiley is also
gratefully acknowledged, as is the assistance of the Canadian Mathematical
Society and the support of the Natural Sciences and Engineering Research
Council of Canada.
Halifax, Nova Scotia

November 1986
-1
-6
Complete Elliptic
Integrals and the
Arithmetic-Geometric
Mean Iteration
The Complexity
of Calculating
Algebraic Functions
J.
J.
-2
7
The Complexity
of Pi and the
Elementary Functions
Theta Functions,
The AGM
and
Algorithms for Pi
r
J.
10
Other ~ G r o a c h e s
to the
&\
Jacobi's TripleProduct,
Theta Functions and
Number Theoretic
ADDli~ati~n~
Modular Equations
a
n
Transformation
Theory of
Elliptic lntegrals
5
Modular Equations
and Algebraic
Approximations
to Pi
--
General Means
Mean Iterations
11
Fu"l
Applications of
Theta Functions
The Story of Pi
Computation
and
Transcendence
---+
Necessary Reading Order
--
.
) Possible Reading Order
viii
Preface
brief discussion of a number of topics including elliptic integrals and

functions, theta functions, and modular functions. These attractive and once
central concerns of analysis have been dropped from the standard curriculum-and
much that is beautiful has become relatively inaccessible
except to the expert or the archivist. In presenting this material we have not
striven for generality. This is available in the specialty literature. At times
we have settled for giving only a taste of the material and a few pointers on
where it can be pursued.
We have found this excuse to consult the nineteenth-century masters a
pleasurable and rewarding bonus-as Hermann points out in his introduction to Klein [79], "We are so used to thinking in terms of the 'progress' of
science that it is hard for us to remember that certain matters were better
understood one hundred years ago."
The second direction takes us into the domain of analytic complexity.
How intrinsically difficult is it to calculate algebraic functions, elementary
functions and constants, and the familiar functions of mathematical physics?
Here part of the attraction is the surprising answers-the familiar methods
are often far from optimal.
Finally, an honest treatment invites exploration of applications and
ancillary material, particularly the rich and beautiful interconnections between the function theory and the number theory. Included, for example, are
the Rogers-Ramanujan identities; algebraic series for T; results on sums of
two and four squares; the transcendence of T and e ; and a discussion of
Madelung's constant, lattice sums, and elliptic invariants.
Our primary concern throughout has been the interplay of analysis and
mathematical application. We hope we have elucidated a variety of useful
and attractive analytic techniques. This book should be accessible to any
graduate student. Only rarely does it assume more than the content of
undergraduate courses in real and complex analysis. It is, however, at times
terse, at times computational, and some of the exercises are difficult. A fair
amount of the material, particularly on the approximation and computation
of T and the elementary functions, is new and only partially available in
research papers.
The accompanying flow-chart gives possible routes through the material.
Chapters 8 and 11 are largely self-contained. Chapter 8 is a treatment of
general mean iterations, while Chapter 11sketches some of the history of pi,
its calculation and its transcendence. There are numerous exercises (frequently with hints). The exercises often develop substantial additional
examples and bodies of theory, and even the casual reader is encouraged to
look at them.
The contributions of family, friends, students, and colleagues have been
many and varied and have greatly facilitated the production of this book. To
all these people we offer our thanks. A particular debt of gratitude is owed
to Professors R. Askey, B. Berndt, R. Brent, K. Dilcher, W. Gosper, Y.
Kanada, D. Shanks, and J. Zucker. Their thoughtful comments and sugges-
Biographical Information
Jonathan M. Borwein was born in St. Andrews Scotland in 1951. In 1971 he

obtained an Honours B.Sc. in mathematics from the University of Western
Ontario, where his father is still head of the mathematics department. He
was awarded an Ontario Rhodes Scholarship that year which he held at
Jesus College Oxford, where he was awarded a mathematics D.Phil. in 1974
under the supervision of Michael Dempster. Since then he ha%been on the
faculty of Dalhousie University and has been Professor of Mathematics since
1984. He has also been on faculty at Carnegie-Mellon University (19801982) and has spent visiting research periods at Cambridge, Limoges, and
the University of Montreal. His other research interests are in classical
analysis, functional analysis, and optimization theory.
Peter B. Borwein was born in St. Andrews Scotland in 1953. He obtained an
Honours B.Sc. in mathematics from the University of Western Ontario in
1974, and a Ph.D, under the supervision of David Boyd, from the University
of British Columbia in 1979. He spent 1979-1980 as a NATO research
fellow in Oxford. Since then he has been a faculty member at Dalhousie
University and is now Associate Professor of Mathematics. He has spent a
sabbatical year at the University of Toronto. His other research interests are
in approximation theory, classical analysis, and geometry.
Contents
Chapter 1
Complete Elliptic Integrals and the Arithmetic-Geometric

Mean Iteration
The Arithmetic-Geometric Mean Iteration 1

Gauss's Derivation of the Fundamental Limit Formula 5
Basic Properties of Complete Elliptic Integrals 7
Quadratic Transformations and Iterations and a Third
Proof of the Fundamental Limit Formula 11
1.5 Jacobi's Differential Equation and a Fourth Proof of the
Fundamental Limit Theorem 18
1.6 Legendre's Relation 23
1.7 Elliptic Functions 28
1.1
1.2
1.3
1.4
Chapter 2
Theta Functions and the Arithmetic-Geometric Mean

Iteration
A Theta Series Solution to the AGM 33

Poisson Summation 36
Poisson Summation and the AGM 40
The Derived Iteration and Some Convergence
Results 44
Two Algorithms for rr 46
General Theta Functions 52
The Landen Transformation 57
Chapter 3 Jacobi's Triple-Product and Some Number Theoretic
Applications
3.1
3.2
3.3
3.4
Jacobi's Triple-Product Identity 62

Some Further Theta Function Identities 67
A Combinatorial Approach to the Triple-Product Identity 76
Bressoud's "Easy Proof" of the Rogers-Ramanujan
Identities 78
xiii
33
xiv
Contents
3.5
3.6
3.7
Some Number Theoretic Applications 81

The Mellin Transform and the Zeta Function 87
Evaluation of Sums of Reciprocals of Fibonacci
Sequences 91
Chapter 4
Higher Order Transformations
4.1 A First Approach to Higher Order Transformations 102

4.2 An Elementary Transcendental Approach to Higher
Order Transformations 109
4.3 Elliptic Modular Functions 112
4.4 The Modular Equations for h and j 119
4.5 The Modular Equation in u-v Form 126
4.6 The Multiplier 136
4.7 Cubic Modular Identities 142
Chapter 5 Modular Equations and Algebraic Approximations to
~r
Singular Values of the Second Kind 152

Calculation of a 155
Further Formulae for a 164
Recursive Approximation to n 169
Generalized Elliptic Integrals and Rational and Algebraic
Series for l l n and 11K 177
5.6 Other Approximations 191
5.1
5.2
5.3
5.4
5.5
Chapter 6
6.1
6.2
6.3
6.4
The Complexity of Algebraic Functions
Complexity Concerns 200

The Fast Fourier Transform (FFT) 204
Fast Multiplication 209
Newton's Method and the Complexity of Algebraic
Functions 212
Chapter 7 Algorithms for the Elementary Functions

7.1 n and Log 219
7.2 Theta Function Algorithms for Log 224
7.3 The Complexity of Elementary and Elliptic
Functions 226
Chapter 8
8.1
8.2
8.3
8.4
General Means and Iterations
Abstract Means 230

Equivalence of Means 239
Compound Means 243
Convergence Rates and Some Examples 249
Contents
8.5
8.6
8.7
8.8
~arlson'sIntegrals and More Examples 256

Series Expansions of Certain Means 263
Multidimensional Means and Iterations 266
Algebraic Iterations and Functional Relations 273
Chapter 9
9.1
9.2
9.3
9.4
9.5
281
Some Additional Applications
Sums of Two Squares 281

(Chemical) Lattice Sums 288
Odd-Dimensional Sums and Benson's Formula 301
The Quintuple-Product Identity 306
Quintic and Septic Multipliers and Iterations 309
Chapter 10
Other Approaches to the Elementary Functions
10.1 Classical Approximations 316

10.2 Reduced Complexity Methods 326
Chapter 11
Pi
11.1 On the History of the Calculation of

11.2 On the Transcendence of rr 347
11.3 Irrationality Measures 362
Bibliography
Symbol List
Index
rr
337
316
Chapter One
Complete Elliptic Integrals and

the Arithmetic - Geometric
Mean Iteration
Abstract. The focus of this chapter is the arithmetic-geometric mean ( A G M )
iteration of Gauss, Lagrange, and Legendre and its relationship to elliptic
integrals. The iteration converges quadratically to a nonelementary transcendental function simply expressible in terms of complete elliptic integrals. This
result, which is fundamental to this monograph, is established in a variety of
ways. Some of the basic properties of elliptic integrals and functions are
discussed.
1.1
THE ARITHMETIC-GEOMETRIC MEAN ITERATION
One of the jewels of classical analysis is the arithmetic-geometric mean

(AGM) iteration of Gauss. It is the following two-term recursion:
It is customary and useful to introduce an auxiliary variable,
(1.1.3)
c , + ~ : =$(a,- b,).
If we assume that 0 < b, s a,, then from the arithmetic-geometric mean

inequality we have
The Arithmetic-Geometric Mean Iteration
b, r b,+l
5 a,+l 5 a,
and
Whence we observe that a, and b, converge to a common limit determined

uniquely by a, and 6,. This common limit will be denoted (on letting a := a,
and b := b,) by
M(a, b) := !trim a, = n-+m
lim b,
,
We will refer to this limiting process as the AGM and usually reserve the
symbols a,, b,, and c, for variables bound by the AGM relations. We also
reserve the symbol AG(a, b) for the common limit.
We say that a, + a ,with pth-order convergence if
-a
- a ) p = O(l)
an+1
(a,
where, as usual, a, = O(j3,) means that, for some constant c and for all n,
a,,r cp,. If the a, are functions defined for all x in a set K, and if the
implicit constant concealed by the 0 symbol in (1.1.6) is independent of x ,
then we say that the convergence is uniformly pth order. Roughly speaking,
quadratic (second-order) convergence doubles the number-of-digits agreement between successive iterates and the limit, cubic (third-order) convergence triples the agreement, and so on.
It is an easy observation from (1.1.4) that the AGM converges uniformly
quadratically for a,, b, restricted to compact subsets of (0, m). Very precise
estimates for the rate of convergence will be established later.
We also observe that M(a, b) is homogeneous, that is, for A > 0
(1.1.7)
AM(a, b) = M(Aa, Ab)
and thus there is little loss of generality, though often some loss of
symmetry, to setting a = 1.
The function M satisfies
1.1 The Arithmetic-Geometric Mean iteration
The analysis of the limit of the AGM rests on finding a two-variable function
M invariant under the transformation (1.1.8) or, equivalently, on finding a
function f satisfying the functional relation (1.1.9), namely,
If we set k, := x E ( 0 , l ) and
we have that k, 4 1 quadratically. In fact, the function g defined by
is the unique solution of (1.1.10) analytic in a neighbourhood of 1 that

satisfies g(1) = a. (See Exercise 4.) This form of the AGM, as a single
variable iteration, is usually called the Legendre form.
It is convenient and standard to define the complement x' of x by
x' :=
Differentiation will be denoted by f.
m.
Comments and Exercises

The early history of transformations of elliptic integrals, of which, as we
shall see in the next section, the AGM is an example, is laid out in an
entertaining article by G. N. Watson [33] entitled "The Marquis and the
Land-Agent: A Tale of the Eighteenth Century." The marquis is Fagnano
and the land-agent is Landen. Landen's transformation, published in 1775,
will be discussed later. The names of Euler and Lagrange should also be
associated with the early transformation theory. Lagrange uncovered the
AGM iteration sometime before 1785. Gauss rediscovered it independently
in the 1790s. He apparently first considered the iteration in 1791 at the age
of 14 (Almquist and Berndt [Pr]). It is, however, Gauss and Legendre who
develop the theory fully. As Watson [33] points out, "in the hands of
Legendre, the transformation became a most powerful method for computing elliptic integrals." Gauss is unique in having deduced the invariant
function from the functional equation rather than proceeding in the opposite
(and easier) direction.
1. Deduce that, for the AGM,
an=an+l+cn+l
bn=an+l-cn+l
c 2n = a2n - b : .
Hence the AGM is well defined for negative n. Show that
a _ , = 2"a:
b-, =2 " ~ :
C-,
= 2"b:
where a:, b: and c: are generated from the AGM commencing with
a: := a,, bg := c,, and c: := b,. Show that
Observe that this formula avoids the subtractive cancellation problems

inherent in calculating the very small number c, from cn+,= $ (an- b , ) .
Show that
2.
Consider the harmonic-geometric mean iteration
Show, for a,, Po E ( 0 , w), that the above iteration converges quadratically to H(ao, Po), where
H(a07
1
M ( l la,, 1 I&)
'
3. Consider the arithmetic-harmonic mean iteration
Show, for a,,

cally and that
Po E ( 0 , w), that
the above iteration converges quadrati-
1.2 Gauss's Derivation of the Fundamental Limit Formula
4.
Show that the AGM is a well-defined quadratically convergent iteration

for starting values a, := 1, b, : = z , where re(z) > 0. Likewise the function g of (1.1.12) is a single-valued analytic function on re(z) > 0. In
both cases the convergence is uniformly quadratic on compact subsets.
[The root must always be chosen to lie in re(z)>O.]. Show that
g(z) = M(l, z) for re(z) > 0.
1.2 GAUSS'S DERNATZON OF THE FUNDAMENTAL
LZMZT FORMULA
By May 30th, 1799, Gauss had observed, purely computationally, that
agreed to at least eleven decimal places. He commented in his diary that this
result "will surely open up a whole new field of analysisv-a claim vindicated by the subsequent directions of nineteenth-century mathematics. The
inverse of the above (indefinite) integral is the lemniscate sine, a function
Gauss studied in some detail. He had recognized it as a doubly periodic
function (see Section 1.7) by the year 1800 and hence had anticipated one of
the most important developments of Abel and Jacobi: the inversion of
algebraic integrals.
We now outline Gauss's derivation of the limit of the AGM (Gauss
[1866]). This is not the easiest development but it may be the most
motivated.
Theorem 1.1
First proof. Observe that [M(l + x, 1- x)]-', defined by (1.1.5), is

analytic and even in some neighbourhood of zero. (See Exercise 4, Section
1.) Thus we may suppose that
Upon' application of the AGM transformation we have
Comparing (1.2.2) and (1.2.3) gives
We leave it as an exercise to the reader to follow Gauss's footsteps by

solving the above equation for di.In fact,
If we observe, as in (1.2.3), that
we see by (1.2.2) that we are finished if we show that
This final equation requires expanding (1 - x%in2 o ) - ' / ~ and integrating

term by term.
It isperhaps possible to be guided to the limit of the AGM by the above
method. If, however, one has correctly guessed the limit, then proving it
correct is much more straightforward and only involves establishing the
invariance of the limit under the transformation.
Second proof.
Let
Then, as the substitution t := b tan 8 shows,
Now the substitution u := $(t - ablt) (and some care) yields
It follows that T(an, b,) is independent of n and hence, since (1.2.9)
1.3 Basic Properties of Complete Elliptic Integrals
evaluates as an arctan when a = b, we have on interchanging limit and

integration

An excellent account of the development and the importance of elliptic
function theory in the nineteenth century is to be found in Felix Klein's
classical work, "Development of Mathematics in the 19th Century" (Klein
[79]). Gauss's works are, of course, available in collected form (Gauss
[1866]). Before establishing the limit formula, Gauss produces partial
expansions of M(1, x) and M(1+ x, 1- x) plus a number of AGM calculations carried to as many as 20 decimals. It is apparent that both the
observation of the limit and the route to a proof were indicated by
prodigious numerical experimentation.
The second proof may be found in Carlson [71], Newman [82,85], Todd
[79], or Wimp [84]. Carlson [71] offers some interesting generalizations.
These are discussed in Section 8.5.
1. Fill in the details in the above proofs. In particular prove (1.2.5),
(1.2.7), and (1.2.10).
2. Show that [M(1+ x , 1- x)]-' and [M(l, x)]-' both solve the secondorder differential equation
Hint: For the second solution consider equation (1.2.6).

1.3 BASIC PROPERTIES OF COMPLETE ELLIPTIC INTEGRALS
The two basic integrals we will encounter are the complete elliptic integral of
the first kind,
and the complete elliptic integral of the second kind,
The complementary integrals E' and K' are the integrals in the complementary variable k' = I&-?,
As is traditional, we will use the notation fr(k) := f(kr) to indicate any

function in the complementary variable. The variable k is often called the
modulus, and kt is the complementary modulus.
The second integral arises in the rectification of ellipses. The arclength A of an ellipse with semiaxes a and b is given by
The first integral has the following physical interpretation. If p is the

period of a pendulum with amplitude a and length L, then
where g,is the gravitational constant. Note as a+0, K(sin(a/2))+ 1~12,

and we are left with a simple harmonic approximation.
The Gaussian hypergeometric series, discussed by Gauss in 1812 in what
is one of the first rigorous discussions of convergent series, is defined by
For the complete elliptic integrals we have the series expansions, for
lkl<l,
1.3 Basic Properties of Complete Elliptic Integrals
where (2i - I)!!:= 1 . 3 - 5 - (2i - 1). The derivations are left as Exercise 1.
We have, as in Exercise 2 of Section 1.2, that K and K t are both solutions
of
This equation has a regular singular point at zero (the roots of the indicia1
equation are both 0 ) which, as the reader familiar with the elementary
theory of second-order differential equations knows, says that
(1.3.9)
K 1 ( k )= a log k K ( k ) + f(k)
where f is analytic in a neighbourhood of zero. (See, for example, Birkhoff

and Rota [69].)
In fact,
This logarithmic asymptote at zero will be of considerable interest later in

the discussion of the complexity of log. Also,
(1.3.11)
E t ( k )= 1 +
[log
(i)
A]
-
k2
Once again the above verifications are left as exercises.

The first and second integrals are linked by the equations
dE
-=dk
(See Exercises 2 and 3.)
E-K
k
10

The functions K and E are nonelementary transcendental functions. This is
a result of Liouville's. In general an elliptic integral is an integral of the form
where R is a rational function of x and y and where y2 is a quartic

polynomial in x. Except in special cases, such as repeated factors, this is
always a nonelementary function of u. Incomplete elliptic integrals of the
third kind are of the form
The integral is complete when u = 1. (The complete third integral can be

expressed in terms of K and E.) Analogously we may define incomplete
elliptic integrals of the first and second kind. The basic result due to
Legendre is that any elliptic integral may be algebraically reduced to a linear
combination of elliptic integrals of the first, second, and third kind. (See
also Exercise 5 of Section 1.4 and Exercise 6 of Section 1.6.)
The formulae in the section may be found, in tabular form, in Abramowitz and Stegun [64], Gradshteyn and Ryzhik [go], and the Bateman
project (ErdClyi et al. [53]), so may transformation formulae for the
hypergeometric functions. Of course the indispensible companion volume is
Whittaker and Watson [27].
A number of the seminal nineteenth-century papers, including Gauss's on
the hypergeometric series and Jacobi's on theta functions, are available in
translation in Birkhoff [73].
1. Establish the series expansions (1.3.6) and (1.3.7) for K and E by
expanding the radical by the binomial theorem and integrating term by
term.
2. Establish, by differentiating the integral (1.3.2)' that
dE
-=dk
E-K
k
3. Verify, from the series expansions (1.3.6) and (1.3.7)' that
4.
From the integral representations show that
9%
1.4 Third Proof of the Fundamental Limit Formula
b) K1(k)= log(;)
+ 0(k2110gkt)
k J 0.
Hint:
d0
v1- (1 - k2)sin28
k' sin 0 d0
v k 2 + (kt)' cos2 8
+
[I2
~~
1+ kt sin 0
The second integral evaluates to log [(I + k') lk]. (See Borwein and
Borwein [84a].)
5.
6.
Verify the expansions (1.3.10) and (1.3.11). (See also Exercise 1,

Section 2.3.)
Establish the relation, for re(c) > re(b) > 0,
where C is a constant independent of z . [In fact C = r(c) lr(b)r(c - b)

where, as usual, I? is the gamma function. (See Section 1.6.)] Note that
this provides an analytic continuation of F to @ - [I, m).
7. Show that F(a, b; c; z) satisfies
dy - aby = 0 .
+ [C - (a + b + l)z] dz
d2y
i ( 1 - z) 7
dz
This is the hypergeometric differential equation.
1.4
QUADRATIC TRANSFORMATIONS AND ITERATIONS AND A

THIRD PROOF OF THE FUNDAMENTAL LIMIT FORMULA
The complete elliptic integrals satisfy the following functional relations

which we collect together as
12
Theorem 1.2
For k E (0,I),
(upward)
2
(b) K(k) = 1+ k'
1- k'
id
(downward)
(upward)
(d) E(k) = (1 + k') E(-)
1- k'
- klK(k)
(downward) .
Proof. Parts (a) and (b) are equivalent and follow from the previous
discussion. We give a direct proof of (b) based on Ivory [1796]. This is also a
third proof of Theorem 1.1. Let 1:= (1 - k')l(l + k'). Then we have
V? = kl(1 + k') and 1+ 1' = 2(1+ k") / ( I + k')'. Thus
on replacing sin2 8 by (1 - cos 28) 12. Then
Here we have used the binomial theorem twice. Since only the terms with
m = n are nonzero, we have
on using (1.3.6). This completes (b). If g(k) := 22/lt/(l+ k), then g-l =
(1 - k') l(1 + k') and [I + gl(k)] I2 = 1/ ( I + k). Now (a) follows by substituting g(k) for k in (b). To derive part (c), we differentiate (a) to get, for
K = dKldk,
This is coupled with the differential equation (1.3.13) in the forms
E(k) = kkt21C(k)+ k ' 2 ~ ( k )
(1.4.2)
and
) ) (1.4.3) and then employ

We now use (1.4.1) to eliminate ~ ( ~ ( kfrom
(1.4.2) and (a) to solve for E(g(k)) in terms of K ( k ) and E(k). The
algebraical details are left to the reader. Part ( d ) may be derived analogously from ( b ) or by substituting g - l ( k ) for k in (c).
The transformations are termed upward and downward because, for
k E ( 0 , I ) , iterating (a) leads to a sequence of k values increasing to 1 and
iterating ( b ) forms a sequence of k values decreasing to 0 .
It is convenient to introduce homogeneous forms of E and K and to
recast Theorem 1.2 in AGM terms.
Let
Theorem 1.3
Proof. Part (a) has been observed in the second proof of the fundamental limit theorem (Theorem 1.1).
To see part ( b ) , notice that if k , := c,la,, then k; = b,la,,
and
J(an,b , ) = a n E ( d l Recall that c: = a:
- b:.
5)
an
anE(kn).
The relationship between k,+, and k , is given by
14
Thus, establishing (b) is equivalent to establishing
which on setting a,+, = (a, + b,)/2 and dividing by a, may be seen to be

equivalent to part (d) of Theorem 1.2.
These transformations may be iterated to produce quadratically convergent algorithms for K and E. (See also Exercise 1.)
Algorithm 1.1
(a) ~ ' ( k , )=
2
qn*2 .=,
fi (1 + k')
1+ k,
"
(upward iteration)
=
,,
where
and
k, E (0, 11 .
2
-2
1+ k:,
n (1 + k,)
.=,
2fi
knfl := 1+ k,
(b) K(k,)
;.=n,
"
(downward iteration)
where
1- k:,
k,+, := 1+ k:,
and
k, E [O,1) .
The first product is the unique solution of Exercise la) analytic in a

neighbourhood of 1that takes the value */2 at 1, while the second product
is the unique solution of functional relation (b) of Theorem 1.2 that takes
the value 7~12at 0 and is analytic in a neighbourhood of 0. This observation
verifies that the above products are analytic in neighbourhoods of 0 and 1,
respectively, since they are uniformly convergent products of analytic functions. Also, specifying the value of an analytic solution of either functional
relation at any point in (0,l) in fact specifies the function at an infinite set of
points with limit point within the domain of analyticity and hence uniquely
defines the function up to a constant multiple of the value at the limit point.
(See Exercise 4.)
The algorithms for E and K in AGM form are particularly attractive.
Algorithm 1.2
For a,:=l, b,:= klE(O,l], and c,:= k,
1.4 Third Proof of the Fundamental Limit Fornula
( b ) E ( k ) = ( 1 - C 2"-lc:) ~ ( k ) .
n=O
Proof. Part (a) is Theorem 1.1. For part ( b ) we use Theorem 1.3,
2J(an.+1,b,+J
- J(a,,
b,) = a,b,I(a,, b,) = a,b,I(a,,
bo)
2
and since 4a,+,
- 2a,2 - 2a,bn = -cn,2
2n+1[~(an+17
b,+d - a:+14a,, boll - 2"[J(an7b,) - a:%,
n-1
=2
cnI(ao,b,)
b0)l
Thus on summing
m
J(a,, b,) = (a: - 2 2n-1c:) l(a0,b,)

n=O
which specializes to ( b ) . Here we must observe that

A, :=2"[a:1(an, b,) - J(an7b,)]
(a: - b:) sin28
.\/a: cos2 8 + b: sin28
d8
Thus 0 5 A, 5 2"c:I(a,, b,), and A, tends to zero as n+

zero quadratically. 0
since c: tends to
~0

The algorithms of the section provide remarkably efficient methods for the
calculation of E and K and related functions. The analysis of these algorithms will be reserved for later chapters. We have restricted our attention
in this section primarily to a real variable k. This is simplifying but entirely
unnecessary. All of the algorithms and functional relations extend naturally
to the complex domain. In fact, all the algorithms and functional equations
of this section hold at least for k E {re(z)>O) - [1, m). The interested
reader may readily establish the exact domains of validity for the various
relations. The analysis of the AGM iteration for complex starting values is
reasonably complicated. (See Cox [85].) The problem is to decide which
The Arithmetic-GeometricMean Iteration
16
a.
root is appropriate in the computation of b,+, =

The right choice is
made to ensure la,+l - b,+,lS la,+, + b,+,l [with im (b,+,la,+,) > O in the
case of equality]. The surprise is that no matter how the roots are chosen,
the AGM iteration converges (provided a, # -b,) though unless the right
choice is made all but finitely often, the limit will be zero.
Exercise 5 on the Landen transform provides an algorithm for calculating
incomplete elliptic integrals. We will revisit Landen's transform in Chapter
2. A wealth of formulae on the calculation of elliptic integrals may be found
in King [24].
1. Show that
and observe that K'IK satisfies a multiplication theorem, namely,
Show that for every integer n there is a unique algebraic x E ( 0 , l ) so

that
2.
Show that if k,+, := (1 - kh) l(1 + k;) and j, := K,then
and hence
Also

m
Show that j, converges quartically to zero.

3. (The quartic AGM) Let a,, b,, and c, satisfy the AGM relations. Set
an:= a:: and pn :=.:b:
Show that
and
Show also that
Show that the convergence is governed by
Derive the following quartic algorithms. For a, := 1 and

m11,
4.
Po:=
Consider a functional relation
where s and g are analytic in some complex neighbourhood U of zero.

Suppose that, for some p > 1, limn,, g(n)(z)+ 0 with pth order convergence uniformly on U.[g(nfdenotes g composed with itself n times.]
Show that the above relation has a unique nonzero analytic solution on
U if and only if s(0) = 1.
5. (The Landen transform) Consider the incomplete elliptic integral of
the first kind
18
Show, as in the second proof of Theorem 1.1, that if we set k,+, :=

(1 - k:) 1(1 + k:) and tan (+,+, - 4,) = k: tan 4 , then
and
Show that if k,+, := 2*/(1+

then
k,) and sin (24,+, - A) := k, sin +,,
and
d)
F(+,, k,) =
(n-1+
"
.=o 1 k,
log tan
($ + 51 lim 4,) .
n-+m
e) Establish the quadratic convergence.,

Similar methods for calculating incomplete second and third integrals
may be found in King [24]. (See also Section 2.7.)
6. Prove Euler's addition theorem. Let g(x) := (1 - x2)(1 - k2x2). Then
where c := [ b m + a m . -] I
This result, which dates
from 1753, is, according to Birkhoff [73], the "first notable theorem
about elliptic integrals."
1.5 JACOBI'S DIFFERENTIAL EQUATION AND A FOURTH PROOF
OF THE FUNDAMENTAL LZMZT THEOREM
The second-order linear differential equation (1.3.8) satisfied by K and Kt is
Equivalently
G(k) := k 1 ' 2 k t ~ ( k ) and
satisfy
G*(k) := k 1 ' 2 k t ~ t ( k )
1.5 Fourth Proof of the Fundamental Limit Theorem
19
Also, the functional relations [(Theorem 1.2(a) and Exercise la) of Section
1.41 become
2fi
(1.5.2)
and
G*(k) :=*(I
(1.5.3)
+k
2fi
) J Z G*(m)
If we set g(k) := 2 f i / ( l + k), these become

(1.5.4)
and
Theorem 1.4
Suppose that f, g, and a are all in c2(0, 1) and that g maps [O,1] into 10, I].
If, for x E (0, I),
and
(b) f(x) =
f( g(x))
c a constant
then
Proof.
Use (b) to change variables in (a). Then, on suppressing x,
d2f(g)
df( g)
f=(2$g+pg)-+$f(g)+g2p
7
ds
dg
where p :=
(remember that d! is the derivative with respect to x). Also,
from (a) and (b),
(1.5.6)
fff=
fi
(1.5.7)
f = af = apf(g)
Substituting (1.5.7) into (1.5.6) to eliminate f yields
20
which on comparison with (a) gives
since 21ig i-pg = 0. Finally we compute p, p and p in terms of g to get

(c>
The bracketed quantity on the right of (c) is often called the Schwarz
derivative of g.
Corollary 1.1 (Jacobi's Differential Equation)
Suppose G ( x ) := X " ~ X ' K ( X )If. p and g are algebraic functions that map
[O, 11 into [0, 11 and
then g satisfies an algebraic differential equation
where
-(7)
1 1+x2
.
4x2 l - x
Proof. We may assume that G ( x ) does not satisfy any equation of the
form
r(x):= -
with p algebraic since by (1.3.13), (1.3.10), and (1.3.11) GIG has a

logarithmic singularity at 1. Thus since G satisfies equation (1.5.1), as in
(1.5.6) we must have
The result now follows from Theorem 1.4 applied to (1.5.1).
1.5 Fourth Proof of the Fundamental L i t Theorem
21
This corollary has a converse which provides an algorithmic check on

whether g is an algebraic transform of K. (See Exercise 1.) Theorem 1.4 has
a partial converse.
Theorem 1.5
Suppose that f,gg,and a are all in ~ ' ( 0 ~ 1and

) that g maps [O,1] into [0, 11.
Suppose that
and
Then
is also a solution of (b).

Proof.
This is essentially the computation of (1.5.8). Replacing f by
flf(g) leaves (b) invariant provided (a) holds.
We now offer a proof of Theorem 1.1 based on Theorem 1.5.

Fourth proof of the fundamental limit theorem. For
and
2vT
g(x) := 1+ x
condition (a) of Theorem 1.5 holds. This is a tedious though (eventually)

entirely rational calculation. Thus if G is a solution of
then so is
However, the above second-order linear differential equation has a regular

singular point at zero and has fundamental solutions of the form
flg1W
and
where g, and g2 are analytic and nonzero. Now observe that
is analytic and nonzero in a neighbourhood of zero, and hence G(x) must be

a constant multiple of the first fundamental solution. Moreover,
cannot have a logarithmic singularity at zero, and hence ( ~ ~ ( x ) ) - " ~ G ( ~ ( x ) )

must also be a constant multiple of the first fundamental solution. In
particular for some c,
~/m
If we multiply by
and take the limit as x-+ 0, we deduce that c = 1.
We have thus proven that G satisfies equation (1.5.4). Equation (1.5.4)
transforms into
and we have shown that K, defined as an analytic solution of the differential

equation (1.3.8), satisfies the above functional relation, and hence we have
found an analytic invariant for the Legendre form of the AGM. [See
(l.l.lO).]
A form of Jacobi's highly nonlinear differential equation may be found in

Cayley [1895], which is an excellent account of nineteenth-century elliptic
function theory with particular emphasis on the transformation theory.
The general question of when a functional relation
has a closed-form solution in terms of familiar functions is difficult. We shall

consider it again later. Theorems 1.4 and 1.5 do, however, suggest how one
1.6 Legendre's Relation
23
might proceed to check whether a solution of the above functional relation

satisfies a second-order differential equation with rational coefficients. (See
Exercises 2 and 3.)
1. Suppose that g(x) is an algebraic function, g maps [0, 11 into [0, 11, and
g(0) = 0. Suppose that g satisfies the algebraic differential equation of
Corollary 1.1. Show, as the fourth proof of Theorem 1.1, that there
exists a constant c so that
and that
where F(x)lvX is analytic in a neighbourhood of zero, r is a rational

function, and g(x) := 2 f i l ( l + x). Without identifying F explicitly,
show that r has double poles at 0 and 1. (Considerations of this nature
can turn the fourth proof into less of a verification.)
3. Suppose that a , p , and g mapping [0, 11 into [O,1] are algebraic, that
and that
Suppose also that f does not satisfy a linear differential equation with
algebraic coefficients of order less than 3. Show that
ax
+b
id4 = ,d
and
p(x) = e(cx + d)2 .
1.6 LEGENDRE'S RELATION
The four quantities K, E, K', and E' are related by a remarkable relation.
24
Theorem 1.6 (Legendre's Relation)
For O < k < l ,
Proof. Let G := k ' I 2 k ' ~and G* := k 1 1 2 k ' ~as

r , in Section 1.5. Then,
by (1.5.1),
and hence there is a constant c so that
In other words, the Wronskian of G and G* is constant. On writing G and

G* in terms of K and K' this becomes
We now use equation (1.3.13),
to eliminate the derivatives in (1.6.2) and deduce that

(1.6.3)
E K r + E'K- K K r = c .
It remains to evaluate the constant c. This may be done directly from the
series expansion at zero (see Section 1.3, Exercise 5), or by using Exercise 4
of Section 1.3.
We define the gamma and beta functions by
(1.6.4)
T(x) :=
low
re(x) > 0
e-' tx-' dt
The function I' satisfies the functional relation

(1.6.6)
~ ( x ) ~-(x)l = -.
sin wx
1.6 Legendre's Relation
The relationship between I? and P is
These standard results will be discussed further in Section 3.6. Our present
objective is to evaluate K(l / f i ) and E(l l f i ) .
Theorem 1.7
and
Proof.
The change of variables x2 := t2/(2 - t2) gives
The above arclemniscate (giving the arclength of a lemniscate) can be

evaluated in terms of P. We set u := t4 and see that
Finally, by (1.6.6),
and
fin
The evaluation of E ( l /fi)

is left as Exercise 1. 17
26
Since ~ (I*) 1 = ~ ' ( I*1)

and ~ (I*) 1 = ~ ' ( l 1f i ) , the evaluation of
K(l lt/Z) and E ( l I*)
gives an alternate route to evaluating the constant in
Legendre's identity that avoids developing the asymptotic expansion of K'
at zero.
We observe that at I/*
Legendre's identity reduces to
LqowKIT and E l v can be calculated quadratically by Algorithm 1.2 using

only the operations of addition, multiplication, division, and square-root
extraction and commencing with a rational starting value. This provides an
excellent approach to calculating IT-an approach we will pursue in some
detail in Chapters 2 and 5.
Legendre derives the constant in (1.6.3) by evaluating ~ ( ( f+i1) / G ) .
I)/(See Whittaker and Watson [27] and Exercise 6.) Note that (fland (fi + 1 ) A 6 are complementary. It transpires that
Values of k for which K'IK is a rational surd are of considerable interest and
importance. (See Chapter 4.) Such k are called singular values, the simplest
of which is l l f i , where
.l. Show that
Start by writing
Then set u :=
and show that
1.6
2.
Legendre's Relation
Show that
3. Show that
Show first that

f ) :=
lo
1 e-x2(t2+1)
d t
and
g(x) :=
t2
dt
satisfy f ' + g' = 0 and f + g = ~ 1 4 .

b) Use a) to show that r($)
= ~7.
Use Theorem 1.2 to deduce that
- 1 is a
[In fact ~ ( f i - 1) = (fi
1)1'2,r($)~(~)/213'4~1'2.]
Thus
second singular value. We return to both singular values and r function
evaluations of K subsequently.
Establish Legendre's relation directly from (1.3.12) and (1.3.13) applied
to E, E', K, and K'. Use Exercise 4 of Section 1.3 to determine the
constant.
From the general theory of the Weierstrass function developed in the
next section, or more directly, one can show (Whittaker and Watson
[27, p. 5161) that for a, b > 0,
where e is the real root of the above cubic and
a) If g2 = 0, then
a2 + b2 = 2fi12g31113
a2 - b2 = -3(2g3)1'3
28
b) If g2 = 0 and g3 > 0, then
c)
Let e := g3 := 5 . Then aZ + b2 = 2V3, and bi) gives
Also, bii) gives
Now substitution of s := 1/t and s := (1 + t3)-'I2, respectively, leads

to
d) Thus when k := sin(*/ 12) = (fi

- 1)/ 2 f i
and
This evaluates K in terms of 'T functions and shows that sin(rrll2) is

a third singular value.
1.7 ELLIPTIC FUNCTIONS
The study of elliptic functions began in the 1820s when Abel and Jacobi
independently discovered that the inverses of elliptic integrals are doubly
periodic functions. As noted earlier, they had been anticipated by Gauss
who at the turn of the century had studied a particular elliptic function, the
lemniscate sine. This work, however, had not been published. This was one
of the critical discoveries of the era and was vital to the concomitant
development of complex analysis and the later development of modular and
automorphic functions.
1.7 Elliptic Functions
29
Except for minimal application in Sections 2.6 and 2.7, we have no great
need for this body of material and offer only a brief sketch of this attractive
theory, primarily in the exercises at the end of this section. Details may be
found in Whittaker and Watson [27] and, more recently, in Bowman [53],
Du Val [73], Eagle [58], and Lang [73].
An elliptic function is a meromorphic function with two periods w, and
w,, with im(w,lw,) # 0. That is, for all z, f(z) = f(z + w,) = f(z + w,). We
assume that w, and w, are minimal in the sense that they are not multiples
of any smaller period. The function f is completely determined by its
behaviour on any parallelogram that is a translate of the parallelogram with
vertices 0, w,, w,, and w, w,. Any such parallelogram is called fundamental. Associated with the periods w, and w, there is the lattice L := {nw, +
mw,)m, n integers). It transpires that given any lattice L there exist elliptic
functions with the appropriate associated periods. However, there are not
many such functions. As we shall indicate in the exercises, any two such
functions are connected by an algebraic equation. The circular functions
may be thought of as degenerate elliptic functions (one of the periods is
infinite), as may rational functions (both periods infinite).
A function f is said to have an algebraic addition theorem if there exists a
polynomial R in three variables with complex coefficients so that for
x, Y E@,
One of the many remarkable facts, due in part to Weierstrass, is that a

meromorphic function has an algebraic addition theorem if and only if it is
elliptic or degenerate elliptic. (See Exercises 11,12 and 13 or Hancock [09].)
The Jacobian elliptic function sn is defined by
The value k is thought of as a parameter and is often omitted. It is true,

though not obvious, that sn(u) is a meromorphic function with periods 4K
and 2iK1 and, hence, is elliptic. (See Exercise 2 of Section 2.7.) The two
other basic Jacobian elliptic functions are cn and dn, defined by
cn(u,k)
(1.7.2)
U=
dt
V(1- t2)(kf2+ k2t2)
and
Note that sin and tanh are limiting cases
30
sn(u, 0) = sin u
and
sn(u, 1) = tanh u
A "half-angle" formula for sn is
Again this is not obvious. It is worth observing that a half-angle formula is a

specialization (x = y) of an algebraic addition theorem. These matters are
pursued further in Sections 2.6 and 2.7. In particular, Exercise 6 of Section
2.7 presents an addition formula for sn.
The following exercises develop some of the elementary theory of elliptic

functions. The approach is via the Weierstrass p function and is standard.
See, for example, Erdklyi et al. [53], Du Val [73], or Lang [73].
1. Show that the elliptic functions (with respect to a fixed lattice L ) form
a field that is closed under differentiation.
2. Show that an entire elliptic function is constant.
3. Let f be elliptic with respect to L. Assume f is nonconstant. Let P be a
fundamental parallelogram whose boundary B contains no zeros or
poles off.
a) Show that the sum of the residues of f in P is 0 by using the

periodicity o f f to show that S, f = 0. Thus any nontrivial elliptic
function has at least two poles in P. The number of poles in P,
counted according to multiplicity, is called the order off.
b) Show that f has the same number of zeros as poles in P counted
according to multiplicity. Hint:Consider $, flf and observe that
f/f is elliptic.
c) Show that f assumes every complex value exactly order off times
in P.
4.
Show that the Weierstrass function
is meromorphic, even, and has double poles at each lattice point

[ L 1 : =L - (0, O)].
5. a) Show that
1.7 Elliptic Functions
Show that p is an odd elliptic function of order 3.

b) Use a) and the fact that p is even to deduce that p is elliptic.
Observe that p has order 2.
Show that every even elliptic function f is a rational function of p.
Hint: Observe that if z is a zero off in P (a fundamental parallelogram) then so is - z mod L. (By -z mod L we mean the unique point in
P, that is equivalent to - z with respect to L; that is, there exist m and
n so that mw, + nw, - z = - z mod L and - z mod L E P.) Observe
that there are exactly four points in P where z =
- z mod L (0, w,/2, w2/2, (w, + w2)/2) and that iff has a zero at one
of these four points, it must have even multiplicity. Consider
where the product is extended over the zeros off, choosing only one
representative from each pair (z, - z mod L ) and where Si is the
multiplicity of the zero (except in the case zi = -zi mod L, in which
case 6,is half the multiplicity). Now shqw that f and g have the same
zeros with the same multiplicities. Treat the poles similarly.
Show that any elliptic function f is a rational function of p and p.
Hint: Consider
which decomposes f into an even elliptic function and an even elliptic

function divided by p.
Show that there exist constants g2 and g, so that
Hint: Consider the order of the pole ofp2 - 4p3 at zero. Observe that
g2 and g, depend on L. Different lattices lead to uniformizations
(parametrizations) of different cubics.
Show that any two nonconstant elliptic functions f and g (with respect
to the same lattice) satisfy an algebraic equation
where Cl is a polynomial in two variables with complex coefficients.

Suppose that f is meromorphic and singly periodic with period w.
The Arithmetic.Geornetric Mean Iteration
Suppose that f assumes no value infinitely many times in any period

strip. Show that f is a rational function of e2""'". (One can use this as
a definition of the class of trigonometric or circular functions.)
Prove that the function p satisfies the algebraic addition theorem
Hint: Show that, as a function of x,
has no singularities at the points 0, I+-yand, hence, is independent of x.

With Exercise 8, this is an addition theorem.
Show that every elliptic function f has an algebraic addition theorem.
Hint: Connect f to p algebraically and use the addition theorem for p.
Show that iff meromorphic and has an algebraic addition theorem SZ,
then either
a) f is a rational function,
b) f is a trigonometric function, or
c) f is an elliptic function.
Hint: Case a) occurs if f has a pole at infinity. So suppose to the
contrary that infinity is an essential singularity. Now for some w,
f(z) = w has infinitely many solutions z,, z2, . . . . Choose z* so that
z*, Z* + zl, Z* + z2, . . . are not poles off. Show that there exist m and
n, both less than a constant depending only on Q, so that f(z* + z,) =
f(z* + 2,). Show that in any neighbourhood there exist infinitely many
points and some n and rn so that f(z + z,) = f(z + 2,). Thus f has a
period z, - z,. Case b) occurs iff has only one period, in which case f
can assume any value only finitely many times within any period strip.
In that case Exercise 10 applies. Otherwise we have case c).
Chapter Two
Theta Functions and the

Arithmetic- Geometric Mean
Iteration
Abstract. In this chapter we solve the AGM iteration in theta function terms
and derive a variety of useful properties of theta functions. The central tool is
the Poisson summation formula. These results are then applied to produce
quadratically convergent products for TT and e" and quadratically convergent
sums for T.
Wefinish the chapter by discussing the Landen transform and the relationship between theta and elliptic functions.
2.1 A THETA SERIES SOLUTION TO THE AGM

The basic theta functions are defined for ( ql< 1 by
(2.1.1)
(2.1.2)
(2.1.3)
These series are more properly viewed as functions of two complex
variables one of which is presently set to zero. (See Section 2.6.) Theta
functions have various number theoretic connections. Note that 8, is a
34
generating function for squares, while as B4(q)= B3(-q), 6, also considers

the parity of the square. Thus
(2.1.4)
63(q) + 64(q) = 2 n C
qn2=283(q4>.
even
Also
where r2(n) counts the number of ways of writing n = j2 + k2. Here we

distinguish sign and permutation [so that, for example, r2(5) = 8 since
(+2)' + (21)' = (21)' + (+2)'] and set r2(0) := 1. Now it is elementary
that r2(2n) = r2(n). (See Exercise 2.) It follows that
Also, (2.1.4) and (2.1.6) allow us to solve for O3(q)8,(q). We have
which bears an obvious resemblance to the AGM. Similarly,
This last term may be rewritten as
k,m= -m
k+m odd
which, on setting k = i - j and m = i + j
+ 1, gives
2.1 A Theta Series Solution to the AGM
Hence
This combines with (2.1.7i) to produce
and these last two and (2.1.7ii) yield Jacobi's identity
Now set k:= k(q):=e;(q)lei(q). Then (2.1.10) shows that kr =

e:(q)/8;(q). If we return to (2.1.7) and set a, := 8:(q2") and b, := 8:(q2")
we observe that a, and b, satisfy the AGM iteration. Moreover, since
&(O) = 1, the limit is 1. Thus
We recast these last observations in:

Theorem 2.1
Let 0 < k < 1 be given. The AGM satisfies
(2.11)
~ ( 1k,t ) = e;'(q)
for
k t = e:(q)lei(q)
where q is the unique solution in ( 0 , l ) to k = 8;(q)
In particular,
Proof. This follows from the previous discussion and Theorem 1.1. The
uniqueness of q will be obvious from the results of Section 3.1, which will
show that 0, decreases and 0, increases on (0,l). Cl
The results of Theorem 2.1 remain true more generally in the complex
plane. This is discussed in Exercise 4.
The solution of the AGM in theta function terms can be found in Gauss
36
[1866]. The systematic investigation of theta functions in the context of

elliptic function theory originates in Jacobi's masterpiece Fundaments Nova
Theoriae [l829].
While we have given a number-theoretically motivated development, it is
possible to give a very elegant formal verification. (See Exercise 1.) This
technique, used frequently by Liouville, is discussed in detail in Bell [27].
I
1. a) Let f : Z x
++ 08. Establish the formal identity
valid whenever both terms on the right-hand side converge.

m2+n2
to
b) Apply (2.1.14) to qm2+"', ( - ~ ) ~ + " q , and (-l)mqm2+n2
derive (2.1.8), (2.1.9), and (2.1.7ii) respectively.
c) Hence rederive (2.1.7) and (2.1.10):
Prove that r2(n) = r2(2n). Hint: 2a2 + 2b2 = (a + b)2 (a - b)'.

3. a) Show that the downward transformation [Theorem 1.2(b)] sends
k(q) to A(q) := k(q2). Precisely,
2.
b) Show that the corresponding transformation for K [Theorem

1.2(b)] is
4. Show that (2.1.12) is valid for complex q with I ql < 1. Precisely,
for such q. [That 8Jq) does not vanish will be apparent from Section
3.1.1
2.2 POISSON SUMMATION
A most analytically accessible route to the behaviour of the AGM lies in the
Poisson summation formula, which we now describe. We then give some
examples of its use before returning in the next section to its relationship
with the AGM. The formula we need is:
2.2
Poisson Summation
37
Theorem 2.2
Let f be a nonnegative function, increasing on (-m, 0] and decreasing on

[0, m). Assume that $_", f(x) dx exists as an improper Riemann integral.
Then, for each x in IW,
each series being absolutely convergent.

Proof. A complete broof may be found in Apastol[74, pp. 332-3331. In
essence one considers the function F(x) := C;=-, f(n + x), which under the
given hypotheses is of bounded variation on compact intervals, and which is,
by construction, periodic. The left-hand side of (2.2.1) is merely the average
[F(x+) + F(x-)] 12, while the right-hand side is obtained by computing the
Fourier coefficients foi F and regrouping. El
EXAMPLE
2.1.
We apply the formula to f given by
The right-hand side becomes
and the left-hand side becomes
n>-x
If x := 0, we derive
which yields
on replacing y by 2nx.
If x:= 1, we derive
xZ
otherwise
'
38
By elementary analyticity considerations, (2.2.2) and (2.2.3) remain valid in

the complex plane and produce the classical formulae for cot, cosec, and so
on. (See Exercise 1.)
EXAMPLE
2.2. This time apply the formula with
Then (2.2.1) becomes
Now the integral on the right is

2
jOme -'""cos (27rkt) dt
F(+)
where
lo
m
(2.2.4)
F( y ) :=
z,-
e-x2cos (2xy) dx = - e
2
(See Exercise 2.) Thus we deduce
Again, analyticity considerations show that (2.2.5) holds for re(s) > 0. This
is a general form of the theta transformation formula. For future reference
we make the notational agreement that Oi(s):= Oi(e-"") and observe that for
x := 0 , (2.2.5) can be written as
Note that for large s the sum t/S03(s)converges much more rapidly than
03(s-I). For example, if s := 100, 03(0.01) and 10 + 20e-loo" coincide
through more than 500 digits.
The result of (2.2.1) was known to Poisson by 1827. With x := 0 he had
obtained the formula in 1823. Equation (2.2.5) was first obtained by Jacobi
using elliptic function theory in 1828.
2.2
Poisson Summation
1. Establish the formulae
c)
2.
n-
cot (
~ 2=
)
1 "
2z
-+2
z ,,=I z - n
Establish (2.2.4) by showing that
and using F(0) = fi/2. (See Exercise 3 of Section 1.6.)

3. Let I(s, y) := $_", e -s?rt2-2ryt dt. Evaluate I(s, y) for real s, y by completing the square. By analytic continuation I(s, ik) = s-lI2 e-"k21s.
4. Recall that the Laplace transform is defined by
Provided that f(t) = O(ebt) as t+ m, F will be analytic for re(y) > b.

Show that the Laplace transform of (2.2.5) with x := 0 produces (2.2.2).
This entails evaluating integrals of the form
This is a special case of a Bessel function transform which can be

e-2baby substituting s2:= t and
evaluated explicitly as F(a, b) = (*la)
v := b Is - as. (Various extensions and related matters are discussed in
Bellman [61].) In principle, therefore, one can derive (2.2.5) as an
inverse Laplace transform of the derivative of the product form of sin
given in Exercise Id).
5. Let f be nonnegative, continuous, decreasing, and Riemann integrable
on [0, m). Let
Show that
whenever a ,
> 0 satisfy a p = 2 ~ Deduce

.
that
(Exercises 2 and 5 follow Apostol [74].)
2.3 POISSON SUMMATION AND THE AGM
We commence by specializing (2.2.5). Setting x := 0 produces
while x := gives
(2.3.2)
d e 2 ( e - s " ) = e4(e-*'">
and dually on setting s = s-',

(2.3.3)
d e 4 ( e - " ) = e2(e-*Is)
where re@)>O. On dividing (2.3.2) by (2.3.1) we have

(2.3.4)
k(e-*s) = kt(e-*Is) .
From Theorem 2.1 we see that
Thus
on using Theorem 2.1 and (2.3.1) again. This produces the following
fundamental theorem.
Theorem 2.3
For all k in (0, I),
Proof. We .have established (a) and ( b ) follows from our identification

of M ( l , k t ) and .rrl(2K(k)). This second equation is often written as q =
and q is called the nome associated with k . In principleit solves the
version .problem for q in terms of k .
We know that k = 8 ~ ( q ) / 8 ~ =
( q4-) + 0 ( q ) as q tends to zero from
above. Theorem 2.3(a) and this information show that since ~ ( 1k')
, tends
to 1 as k tends to zero,
which reproduces the asymptotic of Exercise 4 of Section 1.3. As we shall

see (Exercise I ) , we can derive the exact asymptotic (1.3.10) from these
considerations.
Now consider the AGM iteration commencing with a, := 1 and b , := k'.
Then as we saw in Section 1, b,la, = k'(q2")and c,la, = k(q2n).Hence
In Exercise 2 one establishes that the differential identity
holds. It follows from (2.3.5) and (2.3.7) that
and from (2.3.8),

?r,
bho
b:
while
2 dk
a =- O
k h , '
Since b , tends to M ( l , k ' ) , we have established that
42
and since s = -IT-'
log q,
~ e w r i t i nk,~k', and K in (2.3.9) in theta terms produces
Differentiation of (2.3.1) and (2.3.3) yields

S
(2.3.12)
s2e3(s)+ 5 03(s) = -s-"'&(s-')
(2.3.13)
s ~ ~ ~ ( s ) + - ' ~ ~ ( s ) = - s - ~ ' ~ ~ ~ (w.r.t.s)

(s-~).
(w.r.t.s)
On using (2.3.1) and (2.3.2) again we deduce that

(2.3.14)
e3
s(s)
'33
+ s-'
'33
Now, since s2'3:(s) = '3:(s-'),
1=s 3' 4 (s)

(3-') = - 2
3' 4
+ s-l 4 (s-l) .
3' 2
this shows that
is equivalent to (2.3.11). Finally, adding (2.3.11) and (2.3.15) gives
on using (2.1.10).
We can now also express E in terms of theta functions. We have from
(1.3.13),
because of (2.3.9). Hence

*Differentiation with respect to s.
2.3 Poisson Summation and the AGM
and
Similarly
and
Observe that k = 4 G f ( q ) with f analytic for I q l < 1 and f(0) = 1.

Standard real reversion arguments show that for 0 < k < 1,
where g is (real) analytic with g(0) = 1. Show that
where h(0) = 0 and h is analytic.

Observe that the right-hand side of (1.3.10) and K r both solve
(1.3.8). Moreover, both can be expressed in the form of the
right-hand side of (2.3.19). Deduce that (1.3.10) holds.
Show that 2-"bi2d log (an/cn)is independent of n.
From a) deduce that 2-"ci2d log (bnlan)is independent of n.
Use a) and b) to show that 2-"ai2d log (b,lc,) is independent of n.
Show that all three coincide with bi2dlog (a,lc,). This again is due
to Gauss.
Establish (2.3.18) by using (2.3.16).
Observe that (2.3.18) and (2.3.17) immediately give another proof
of Legendre's relation.
Show that 02(e-") = 04(e-").
b) show that 4~ = (x;=-, e-n2p)l@i=-, n e -n2p)c) Let r,(n) give the number of distinct representations of n as a sum
of k squares. Show that
Hint: Take the kth power of (2.3.1) and differentiate.

All of results in this section can be derived without Poisson summation as
the following exercises show.
5. Derive (2.3.7) by combining Exercise le) of Section 1.4 and Exercise

4b) of Section 1.3. This can also be found in Gauss [1866], but not
apparently in later nineteenth-century authors. (See Borwein and Borwein [84a] and King [24].)
6. Use k - 4 f i to rederive Theorem 2.3 from Theorem 1.1.
7. Observe that Theorem 2.3(a) and Theorem 2.1 show (2.3.1) and
(2.3.4). Hence deduce (2.3.2) and (2.3.3). Thus we have established the
8 transformation formulae [(2.3.1) to (2.3.3)] directly from the AGM.
8. Suppose that in (2.3.7) the 4 is omitted. Show that the convergence,
which was quadratic, is now linear. Of course the limit is unchanged.
2.4
THE DERNED ITERATION AND SOME CONVERGENCE

RESULTS
It is generally the case with AGM related approximations that it is relatively

easy to establish the convergence rate and much harder to determine to
what the given iteration converges. We now consider various preparatory
convergence results. In the next section these will be used to produce two
surprising algorithms for T.
Motivated by the fact that f i ~ ( i 1 1 / Z ) R ( Ii f i ) = T (Exercise 2 of
Section 1.6), we commence by computing K. If we consider the AGM
sequence with a, := 1 and b, := k, it is apparent that an and bn viewed as
functions of k converge uniformly and analytically to M(1, k). It follows that
, Since M(1, k) =
the derived iterations a, and bn converge to ~ ( 1k).
rr/2Kt(k), we see that
on noting that (dKtldk)(k) = -(klkt)(dKldk)(kt). Equivalently,
Now the derived iteration is
2.4 The Derived Iteration and Some Convergence Results
Since b,la, converges quadratically to 1, it is easy to show directly that a,

and b, converge quadratically to ~ ( 1k), for 0 < k < 1. Moreover b, decreases and a, increases, at least eventually (Exercise 1). For our purposes
we will generally consider the Legendre forms x, := a,lb, and y, := b,/a, .
Then
where x, := k-', y := 6,
and yo is undefined. Moreover,
(See Exercise 2.) We also have
and
so that for n r 1
(2.4.6i)
x,+~- 1 5 $ (x, - 1I2
This establishes the quadratic convergence of x, and y, to 1. From (2.4.1),

(2.4.2), and the previous discussion it is now apparent that
and since both M and M are quadratically computable, so is

section we turn this identity into an explicit algorithm.
T. In
the next

1. Show that a, and 6, defined by (2.4.3) converge quadratically to
~ ( 1k),
, by showing that a, increases and b, decreases. Then show
b, - a, 5 (<
- 1) 12. For this to hold for all n, assume x, 5 3.
46
2.
Establish (2.4.4) and (2.4.5).
to show that X , 1 + 8e-2"z(K/~')(~")

3. a) Use X, = $~(q2")~$:(q2n)
as
n-+a.
Z ~ - ~ T ( K I K ' ) ( X ; ~.)b) Show that cn 5 4 a n e.- an, and that c, 8, as
n+a.
4. Show that with a, := 1 and b, := k,
Hence deduce that with x, and y,, as in (2.4.4),
.- 2114
In particular with x, := f i and yl .-
5. Show that
$)
~ ( 1 ,
= a-112~2(:)
2.5
and
&(I,
1
z)
=2 f i a -
512
F (4)
TWO ALGORITHMS FOR .sr
The systematic use of the derived AGM leads directly to quadratic algorithms for IT.
Algorithm 2.1
Let x, := fi, .rr, := 2 + f i , and y, := 2lI4. Define
+ 11Y, + 1
x, + 1
(iii) wn:= IT,,-*Yn + 1
Yn<
(ii) Y , + ~ : =
Then n;, decreases monotonically to
n21
nrl.
IT.
Moreover, for n r 0,
and, for n 2 2,
Proof. We first establish the limit. This follows from (2.3.7) (see
Exercise 1) or from (2.4.7) as we now indicate. Let nn:= ~ f i b ~ + ~ a ~ + ~ / a ~ + ~
where a, := 1 and bo := 1I*.
Then by (2.4.7) of the previous section,
%-, n. From (2.4.4) we have
Moreover, no= 2 f i and the algorithm converges to n as claimed. Since

yn 2 xn 2 1, it is obvious that nndecreases. Let us observe that
provided that
Next
Hence
and
48
Here we have used (2.5.4) and a geometric estimate. From the above
Thus, for n > 0, upon checking the early cases,
and with (2.5.4)'
Finally (2.5.3) follows from (2.5.2).

The first nine iterations give 1, 3, 8, 19'41, 83, 170, 345 and 694 digits of
,T.The 24th will produce more than 45 million digits at the expense of only a
few hundred arithmetic operations. A more exact asymptotic will be derived
in Chapter 5. (See Exercise 5.) Note also that the number of leading zeros
of y,+, gives the number of digits of agreement between v, and ,T to within
1.
The second algorithm, based on an identity of Gauss [1866], was discovered by Brent [76a] and Salamin [76] independently.
Algorithm 2.2
Let a, := 1 and b, := 1/ f i . Define

v,,:=
1-
24+1k 2
2 c,'
xi=,,
v m
where c, :=
= c ~ - , / 4 a nand a, and b, are computed by the AGM
iteration. Then v,, increases monotonically to ,T and satisfies
and
Proof. This algorithm is based on the use of the second integral E

rather than K. With k := 1/ f i Legendre's relation is (2E - K) K = v/2.
Combine this with Algorithm 1.2 of Section 1.4 to derive
2.5 Two Algorithms for
~r
which on truncation shows that T,, converges to

vergence estimates as Exercises 3 and 4. 17
IT.
We leave the con-
The first eight iterations produce 0, 3, 8, 19, 41, 84, 171 and 344 digits,
which agrees extraordinarily well with the asymptotic.
Both of these algorithms generalize in many ways. (See Chapter 5.) At
the moment we only exhibit two additional identities.
If we use the differential equation for K equation (1.3.13), we may
rewrite Legendre's identity as
[See equation (1.6.2).] With (2.4.2) we now derive

(2.11)
K1(k) ~ ( 1k')
,
K'(kf ) ~ ( 1k),
~ ( k )~ ~ (kl)
1 , kk'2 ~ ( k '~) ~ (k)1 ,
- = k'k2 IT
and, as in Algorithm 2.1, we can produce
ll
K(k)
K (k)
1 + ~ n
1+x,
( 1 - k lK(k)
) a ,=, (-)+(I-k)-
I+y;
rI (-)1 + x , *
.- k-I, y, := fi,y := fl,and the iterations are

Here x, := kt-', x: .given by (2.4.4). When k := 1lfi, (2.5.12) reduces exactly to Algorithm
2.1. Also whenever K1(k)1K(k) = d7 for rational r (a singular value of k),
we can in principle find k algebraically, as we will see in Chapter 4. We
already know of four such values of k. (Exercise 7.)
In a similar fashion we may use Algorithm 1.2 to substitute for both E
and E' in Legendre's relation. This as observed in Salamin [76] produces
v=
2M(1, k) M(1, k')

1- c;=, 2"-l(ci + c ; ~ ) '
Again, c, and c; are computed from complementary AGM iterations. When

k := 1l f i , this identity reduces to (2.5.9).
From (2.5.2) and (2.5.8) we observe explicitly that the corresponding
algorithms converge quadratically. We are primarily interested in the
50
cumulative error given by (2.5.3) and (2.5.7). We will say, informally, that
any such algorithm allows for quadratic or fast computation. In similar
fashion, we talk about mth-order computation if the cumulative error (i.e.,
digits correct) is of order mn after n steps.
Algorithm 2.1 is derived in Borwein and Borwein [84a] by the method of
Exercise 1.
1. a) Differentiate (2.3.7) and apply (2.3.8) to establish that with
a, := 1 and b, := k t ,
Here a,* as usual denotes the AGM iteration commencing with

b, := k, and all derivatives are with respect to k.
b) Let k := k' := 1/fiand observe that the previous formula reduces to (2.4.7).
2. a) Fill in the details in (2.5.3) to (2.5.6).
b) Show that
,,.
3-,,.- lim
"+-yn+,-l
2'
Note that while Algorithm 2.1 relies on evaluatini K,Algorithm

2.2 in fact relies on evaluating 8. Indeed (1.3.12) and Algorithm
1.2(b) combine to show that
3. a)
112"-1
Use (2.3.7) to show that (cn/4an)

increases monotonically
and quadratically to q. This gives a quadratic algorithm for emon
using a, := 1 and b, := 1 / a .
b) Show explicitly that, with a, := 1 and b, := k t ,
4. As in Algorithm 2.2, show that
Then deduce that
2.5 Two Algorithms for
lr
Now use Exercise 3a) to show (2.5.7) and (2.5.8).

5. As in Exercise 4), show that in Algorithm 2.1 .rr, - rr is of order
213e - n 2 n + 1 . Convergence proofs of this type can be found in detail in
Salamin [76] and Borwein and Borwein [84a] with discussion of the
asymptotics.
6. Prove (2.5.12).
7. Recall from Section 1.6 (see also Chapter 4) that
ii) ~ ' ( f-i 1) = *K(fi
- 1)
iv) K'(3 - 2 f i ) = f i K ( 3 - 2*).

Observe that for these values of k and kt (2.5.12) reduces to an
algebraic combination of two infinite products. Equally (2.5.13) simplifies.
8. Establish the general identity of Gauss, Brent, and Salamin given as
(2.5.13).
9. Use the quartic transformation to produce a quartically convergent
infinite product for n. (See Exercise 3e) of Section 1.4.)
10. Show that with x, and y, as in Algorithm 2.1,
and
11. There are actually eight natural products implicit in (2.4.7). One can
select either b, or a, (b, or a,) for each of the means. While

Algorithm 2.1 is the cleanest of these, it is not the best approximation.
a) Show that there are four decreasing and four increasing products.
b) Show that the two best approximations are given by 2fib:la, and
2 f i a 3 , l b, .
52
c) Show that with xn and yn as in Algorithm 2.1,
with both products converging monotonically to T.

d) Similarly, produce a decreasing analogue to Algorithm 2.2.
12; The identity (2.5.14) can be found in Gauss [1866]. Show similarly the
following identity due to Jacobi. (See King [24].)
x
m
K'(k) =log(F)
16k' + 3
v2-"log(?)
K(k)
n=l
a,:=l
b,:= k'.
2.6 GENERQL THETA FUNCTIONS
Theta functions are more properly considered as a function of two variables-a parameter q and an analytic variable z. So far we have considered
only special theta functions (z = 0). In this section we sketch some relevant
parts of the general theory. We write
&(z, q) := Ol(z, t) := 2qU4sin z - zq9l4sin 32
+ 2qZsf4sin 52 - . .
where q = emit and im (t) > 0.

When z = 0, then 01(0, q) = 0 and q(0, q) = q(q) for j = 2,3,4. When
the precise value of q is unimportant, one writes Oj(q) = q and q(z, q) =
Bj(z). When j is unimportant, one writes O(2). It IS straightforward to
establish the following functional identities (Exercise 1):
2.6
General Theta Functions
53
where the multiplier M is given by

M : = q 114eiz
(2.6.3)
From (2.6.2) or directly one has

8, ( z ) = -8, ( z + w )
(2.6.4)
8,(z) = 8,(z
+ w)
8, ( z ) = -qe2iz81( z + w t )
8,(z) = qeZiz8,(z+ w t )
These identities show us that any theta function is entirely determined by its
values on any fundamental parallelogram
P(zo) := { z l z = z ,
+ r l w + r,nt, 0 Ir,, r, I 1) .
We assume as we may that the given function has no zeros on the boundary
of P(z,). It is obvious that z = 0 is a zero of & ( z ) so that w / 2 , w / 2 + w t / 2 ,
and w t / 2 are zeros of 8&), 8,(z) and 8,(z), respectively. Moreover, (2.6.4)
shows that z , + mw + nwt (m, n integral) is a zero of a theta function
whenever z , is. We now show that each theta function has exactly one zero
in each fundamental parallelogram. Thus we will have specified all the zeros
above.
Consider the integral
which gives the number of zeros of 8 inside P(z,). Explicitly,
Now logarithmic differentiation of (2.6.4) shows that
Thus 2 d N = 25ri and N = 1 as required.

The identities in (2.6.4) show that each Oi/t$,i # j , is doubly periodic. In
combination with our knowledge of the zeros of each t$ we can painlessly
apply Liouville's principle (bounded entire functions are constant) to establish many identities. We illustrate this with the following.
Proposition 2.1
(2.6.6)
Proof.
ez(z)ez = e;(z)e;
ef (z)ef .
Consider
By (2.6.4) f(z) is doubly periodic with periods r and ~ t Moreover

.
8z(z)
has a double zero at d 2 in P(0). Also (2.6.2) shows that 19;(~t/2)0;=
8;(7rt/2)0f. Thus f is elliptic with at most one simple pole in P(O), and by
Exercise 3a) of Section 1.7, f can have no poles. Hence f is constant, being
bounded and analytic. Since 8,(.rr/2) = 0 and 03(v/2) = 04, 04(rr/2)= 03, the
constant must be 8; and (2.6.6) follows. 0
Note that on letting z = 0 in (2.6.6), we recover (2.1.10): 8; = 8; + 8:.
It is a simple matter to recast Example 2.2 as the classical theta
transformation formula.
Theorem 2.4
For z an arbitrary complex number and im (t) > 0,
Here one takes the principal square root. The proof is left as Exercise 3.
It is equally simple to use the Jacobi triple-product of Chapter 3 to
2.6 General Theta Functions
produce product expressions for q(z) (Exercise 4). Various other relationships are indicated in the exercises.
There is a proliferation of notations for theta functions. We follow the most

usual notations (used in Bellman [61]) and note that both Dickson
[71, vol. 3, p. 931 and Whittaker and Watson [27] give tables of alternate
notations. The abuse of functional notation, in particular the distinctions
among 5 , q(q), q(t), and q(z), necessitates some caution. Both Bellman
[61] and Whittaker and Watson [27] provide an accessible introduction to a
considerable amount of material on theta functions.
1. Verify the identities of (2.6.2) and (2.6.4). Note that (2.6.2) shows that
we can restrict attention to one theta function (say, 8,) and lose no
information.
2. Show using Liouville's principle that,
Results of this kind are discussed in detail by Whittaker and Watson.

Many are given in tabular form in ErdClyi et al. [53].
3. Establish Theorem 2.4.
m
4.
Show that with Q, := r]:(1 - q2n),

n=l
i) 03(z) = Q,
r]:(1 + 2q2"-' cos (22) + q4"-')
n=l
ii) B,(z) = 2 q 1 ' 4 ~sin

o z
I1 (1 - 2q2" cos (22) + q4") .
n=l
Establish similar identities for 02(z) and 04(z). Many variations are listed
in ErdClyi et al. [53]. (This relies on Section 3.1.)
5. Use Exercise 4 to show that
(2)
i) -2_=4C
8&)
" (- l)"qn sin (2nz)

n=l
1- q2"
[-I
sin2(nz)
ii) log 63( 4 = 4 " (-l)"qn
2n
63(0)
.=I 1- q
n
6.
56
for lim (z)l< f log lql.

One of the easy but important properties of theta functions is that they
solve the one-dimensional heat or diffusion equation
This provides a significant connection between the analytic properties of

theta functions as functions of z and their number theoretic properties
which revolve around the variable q. Various of these matters are
pursued in Bellman [61] and Rademacher [73].
7. BelIman [61] gives an interesting functional equation approach to
Theorem 2.4. Consider entire solutions f to
i) f(z+.rr)=f(z)
ii) f(z + .rrt) = be-""f(z)
where b is an unspecified function of t independent of z. Suppose,
temporarily, that f possesses an absolutely convergent Fourier expansion valid for all z:
a) Show that c,,, = b-lq2"cn. Hence deduce that for I ql = leTitl< 1,
is an analytic solution to i ) and ii) and is unique up to choice of

constant. When q = b-', we recover 8,(z).
b) Show that any entire solution to i) and ii) has an absolutely
convergent Fourier series. Hint: Use contour integration to show
that for any integral k,
IJ-T,2
?rl2
f(z) e-2inzdz/ 5 e-'*"a
max
-m/2sw~m/2
If(w - &)I
c) Use the uniqueness of solutions of i) and ii) to rederive Theorem

2.4. Hint: Show that both sides satisfy i) and ii). Then use Exercise
6 to normalize the equation.
There is also a considerable literature on multidimensional theta functions. (See Bellman [61].) Since they do not impinge on our main considerations, we say no more.
2.7 The Landen Transformation
2.7
57
THE LANDEN TRANSFORMATION
We finish this chapter by deriving an expression for sn in terms of theta

quotients and by relating this expression to the transformation of incomplete
elliptic integrals. We begin with the classical Landen transformation in theta
form.
Theorem 2.5
For all z and im (t) > 0,
Proof. As a function of z, 8,(2z72t) has zeros when 22 = n t + mn +

2nnt or when z = mnl2 + (2n + l ) ~ t / 2 This
.
is exactly where 03(z7t) or
04(z7t) is zero. Again (2.6.4) shows that f,(z) := 8,(z7 t)04(z, t) /fI4(2z72t) is
doubly periodic with periods n and nt. Thus Liouville7stheorem shows that
f,(z) is a constant with respect to z. The second equality follows on
substituting z + ntl2 for z and using (2.6.2).
To establish the existence and analyticity of sn in theta terms we begin
with
(See Exercise 1.) Now let p := 8,(z)/84(z) and observe that
(This relies on Exercises 2a) and 2b) of the previous section.) Then
replacing z by ~'3;' and p by y = p'3,/82, we observe that, since k2 = 8;/8:,
This is solved by
58
and then, in agreement with (1.7.1),
which solves the inversion problem (at least for real k) for the given
integral. We wish to have cn2 sn2 = 1 and k2 sn2 + dn2 = 1, and it is
appropriate to define
(See Exercise 2.)

Finally we wish to recast Theorem 2.5 in elliptic function terms.
Theorem 2.6 (The Descending Landen Transform)
Let 0 < J/ and 0 <'k < 1 be given. If
and 0 I r h is given by
sin
:=
(1 + k t ) sin t,b cos
VGT-aj
then
(2.7.8)
(1 + k')
Io*
1- k2 sin2 o
dB
I/IT~Z%
is the incomplete elliptic integral of the first kind.
Proof. From Theorem 2.5 we deduce
In. terms of sn, cn, and dn this becomes
where u = z ~ ; ( q )and u, = 2z8;(q2) since k, = 82(q2)/8,(q2).

Thus k, = (1 - k r ) l ( l + kt) and u, = 2 ~ 8 ; ( ~ ~ ) 1 8 =
; ((1
~ )+ kr)u. Collecting information, we have
(2.7.9)
u, = (1 + kt)u
1-k'
k, = 1+kt
sn cn
sn(u,, k,) = (1 + k r ) dn
since kkyl" = 1+ kr. The change of variables sin 8 = sn(u, k) and sin 8, =
sn(u,, k,) now produces (2.7.8). 0
A profusion of information on the numerical use and derivation of various
Landen transforms is given in King [24]. More bibliographic information is
available in Watson [33] and in Whittaker and Watson [27].
1. Prove the differential equation (2.7.2) by showing that
is doubly periodic with periods r and rt12 and that +(z) has simple
poles possibly only at r12, r 1 2 + rt12, and translated points. Then
show that +(z + rt12) = +(z), and hence relative to the periods r and
r t / 2 , 4 is doubly periodic with a single pole. Thus 4 is constant.
Establish (2.7.5).
With cn and dn defined as in (2.7.6) and (2.7.7) show that
cn2 + sn2 = 1, dn2 + k2 9n2 = 1, and that cn(0) = dn(0) = 1.
Show that (dldu) sn(u, k) = cn(u, k) dn(u, k).
Show that sn is strictly increasing on (0, k) with sn(0, k) = 0 and
sn(K, k) = 1. [You may find it convenient to use the product
formulae for O,(z) and 8,(z).]
Show that dn(K) = kr.
Establish the double periodicity of sn(u, k), cn(u, k), and dn(u, k)
with respect to u and verify the following table:
sn
cn
dn
Periods
Zeros
4K,2iK1
4K, 2(K iK')
2K,4iK1
2mK + 2niK'
(2m 1)K + 2niK'
(2m l ) K + (2n 1)iK'
Poles
+
+
2mK
2mK
2mK
Residues
+ (2n + 1)iK'
+ (2n + l ) i K 1
+ (2n + 1)iK'
(-1)"lk
(-l)"+'ilk
(-l)"+li
3. The ascending Landen transform is given by reversing the roles of k and

k,, $ and $,. Thus if
and
then
('
k, F($, k) = F($,, k,)
These transforms and their analogues for incomplete second and third
integrals clearly lead to quadratic iterations which are studied in detail
in King [24] without reference to theta functions. Note that if $ = -12,
we recover the quadratic transformation for K.
a) Show that, in the notation of Theorem 2.6,
cos $, =
cos2 $ - k' sin2I!,I
v m $
b) Hence show that
1- k'
sin (2$ - $, ) = -sin
1+ k'
= k,
sin
c) Thus show that, if k = 2 f i / ( l + k,) and sin (2$ - $,) = kl sin

then
Compare Exercise 5 of Section 1.4.

4.
a) Show that if u := 2 K x l ~ then

,
sin (x)
sn(u, k) = 2q114k-112
n=l
1- 2qZncos ( 2 4
- 2q2n-1 cos ( 2 4
b) Find similar expressions for cn and dn.
+ q4"
+ q4n-2
61
Combine Exercise 4a) and Theorem 2.6 to produce a quadratically

converging approximation to sn(u, k) given one for sin x.
, Use (2.7.5) and Exercise 2 of Section 2.6 to show "the addition
theorem"
Obtain similar expressions for cn(u + v) and dn(u + u). (See Whittaker
and Watson [27] and Exercise 11 of Section 1.7.)
Chapter Three
Jacobi' s Triple Product and Some

Number Theoretic Applications
Abstract. We establish Jacobi's triple-product identity and apply it quite
variously. We first use it to derive the fundamental product identities for the
theta functions. We then rederive the triple product via Cauchy7s q-binomial
theorem and present Bressoud's beautiful elementary proof of the celebrated
Rogers-Ramanujan identities. After this we derive Jacobi's formula for r,(k)
(the representation of k as a sum of four squares) and two partition results
due to Ramanujan. We also establish the Gaussian sum formula bnd indicate
another proof of the theta transformation formula. Then we briefly discuss the
Mellin transform and use it to give the classical reflection formula for the
Riemann zeta function. Finally, we show how certain reciprocal series can be
evaluated in terms of theta functions. In particular, we give a result due to
Landau on the Fibonacci numbers. We also sum the squares of the reciprocals
of the Fibonacci numbers.
3.1 JACOBI'S TRIPLE-PRODUCT IDENTITY
Our first proof of the triple-product identity is:

Theorem 3.1
For each pair of complex numbers x and q , with x # 0 and I ql < 1,
Proof.
Let F(x, q ) := IIz=, ( 1 + ~ q ' " - ~ ) ( lx-lqZn-').

+
NOWF(., q ) is
3.1 Jacobi's Triple-Product Identity
63
and has a Laurent expansion at zero. Observe that

F(xq2; q ) = ( x q ) - ' ~ ( x ,q). Thus if
m
t:
F(x, 9 ) = n = - m en( 9 ) ~ "
; then en( q ) = c-,(q) and cn(q) = q2"-'c,-,(q)
for n r 0. It follows that
It remains to evaluate co(q). Letting x := 1 in (3.1.2) gives c,(q)O,(q) =

(1 + qZn-1)2and letting x := -1 gives co(q)04(q)= Ii;=,(1 - q 211-1 )2.
sin;
d83(4)8,0= 04(q2),we deduce that
m
co(q)~4(q2
=)
(1 - q4n-2).
n=l
ut on replacing q by q2,
Hence
Since c,(0) = 1 and e, is analytic at zero,
11 (1 - q 2 m ) .
m=l
This establishes (3.1.1).

2
Cl
It is convenient to make the following notational abbreviations.
Some Number Theoretic Applications
64
From these definitions one easily verifies Euler7s identity Q1Q2Q3= 1,

which may also be written
Also
We gather the first three specializations of the triple-product identity into:

Corollary 3.1
.For Iq( < 1, one has

m
(3.1.6)
e3(q) = Q,Q; =
n=l
(1 - qZn)(1+ q2n-1)2
Proof. These follow on using x := 1, -1, and q , respectively, in (3.1.1)

(Exercise 2).
More generally, let k and 1 be real numbers, and let q := qk and x := +q'
in (3.1.1). Then
When k := 8 and 1:=
4, this gives
This is Euler's pentagonal number theorem, which he found empirically and

which affords a combinatorial interpretation [Exercise 3b)], as most of these
identities do.
When k := 1:= 4, we have
3.1 Jacobi's Triple-Product Identity
Finally, k := $ and 1:= 3 ' 4 give two formulae which play a central role in
the Rogers-Ramanujan identities (Section 3.4):
n
m
(3.1.12a)
(1 - q5"+')(l - q5n+4)(1
- qSn+')=
n =O
n=-m
(-l)nq(5n+3)n12
We finish the section with a sli htly less immediate corollary of the triple
product. If q := q1I2 and x := ql/'w, then (3.1.1) becomes
The right-hand side is
If we now let w tend to -1 from above this gives
Thus with (3.1.10),

The proof of (3.1.2) is due to Gauss [1866]. The AGM identity (2.1.7) was
also in his possession, but Jacobi, by elliptic function techniques, was the
first to publish a proof of the triple-product identity. There are many proofs
in the literature. Given the AGM, none perhaps is as simple as the one
given here.
1. Verify the identities (3.1.4) and (3.1.5).
2.
a) Prove Corollary 3.1.

b) Show that
66
3. a) Establish (3.1.9), (3.1.10), and (3.1.11).

b) A pentagonal number is a number of the form n(3n + 1)/2. Show
that (3.1.10) implies that every nonpentagonal number can be
partitioned into an even number of distinct parts as often as into an
odd number of distinct parts. Show that for pentagonal numbers
there is a surplus or deficit of 1, depending on whether n is odd or
even. This was first observed by Legendre in 1830. (See Dickson
[71, vol. 21.)
4.
Establish (3.1.12) and give an interpretation in terms of partitions.

Ewe11 [81] observes that Euler's pentagonal number formula (3.1.10)
allows one to establish Jacobi's triple product, given (3.1.2). Thus one
can base the identity on a combinatonal proof of (3.1.10) such as is
given in Hardy and Wright [60].
5. Use (3.1.10) to show that c,(q3)
6.
IT:=,
(1 - q6") = 1.
A complex analytic approach to Jacobi's triple-product is as follows.

a) Show that (3.1.1) is equivalent to
where q := emit.
From the discussion of Section 2.6 show that both sides of (3.1.17)
are analytic with zeros at z = (n + $.)?rt + m?r (n, m integral). By
Liouville's theorem they differ only by a multiplicative constant.
Use Exercise 5 to show that this constant is 1.
Justify taking the limit in (3.1.14).
Prove that (3.1.15) is equivalent to
Observe that (3.1.18) has the following number theoretic interpretation due to Catalan (Dickson [71, vol. 21). The excess in the
number of even values of x + y + z in
over the number of odd values of x + y + z is (2n I)(-1)". In

particular, any number of the form 3(2n + 1)2 must have at least
(2n + 1)/6 decompositions as a sum of three squares.
Show that 0, and 0, never vanish (I ql < 1).
Show that 0, and 0, increase monotonically on ( 0 , l ) and 0, decreases monotonically on (0,l).
3.2 Some Further Theta Function Identities
67
9. Let p(n) denote the number of partitions of a natural number n into

positive integral parts (the order being irrelevant). Thus p(4) = 5 and
p(5) = 7.
a) Show that
b)
For the second equality see also Proposition 3.4 of Section 3.3.
Use
to write (and implement) a recursive formula for p(n). This is how

MacMahon (1918) computed p(n), 1r n 5 200; p(200) =
3972999029388. For more information on partition theory the
reader is referred to Hardy and Wright [60] or Andrews [76].
c) (Euler) Establish that
where a(m) := m(3m + 1) 12 is the mth pentagonal number.
3.2 SOME FURTHER THETA FUNCTION IDENTITIES
In this section we collect a number of definitions and relations, some of

intrinsic interest and some for future reference.
Let r E (0, m) and define hx(r) := k(e-dr). As in Chapter 2, we consider
k as a function of q (=e-&). Then
since hx(r) = Oi(e-dr)/O:(e-d)
From Corollary 3.1 we have
(See Exercise 1.). Also
68
when q := e - n K ' ( k ) l K ( k )
We next derive another beautiful identity due to Jacobi, in which 8:
denotes the derivative of 8, with respect to z at zero. [See equation (2.6.1).]
Proof
Identity (3.1.14) can be rewritten as
For various applications to lattice sums it is natural to augment our theta

definitions by
and
m
By similar arguments [see (3.1.14)] one can show that
and
We leave these identities as Exercise 6 and note that (3.2.4) and (3.2.8)
have number theoretic interpretations like that of (3.1.15). These three
identities and their remanipulations are among the very few known reductions of three theta terms to one known theta expression. (See Glasser and
Zucker [80].)
Following Weber 1081 and others, it is usual to identify the following
quantities. With q and r as above and r := i<r,
7 = q(v=7) = ~
( 7:=
) q1112~o
=A(*)
=A(T):= q - 1 1 2 4 ~=~ (4ktZ/k)1112
= f2(*)
= f2(r) := 2
=f(*)
= f(r) := q
112 1112
q Ql = (4k2/k')1112
Q2 = (4/kk')1'12.
-1124
K = -IT$ f 4 = - *
2
2kr
2 4
'
f*=-7)
2k
2 4
f 2 .
As will be discussed in Chapter 4, whenever r is rational, fl, f2, and f

c equations (whose degree is determined by the number of
with determinant -4r). Thus once q and one other off,, f2,
k , k t is known, all six are determined.
A significant identity which follows from (3.2.9iv) is
3.2.12)
5'
(-l)m+n -4IT
=
log [f(-)l=
n,m=-mm + rn
IT
Here and hereafter, the prime over a summation indicates that the term
m = n = 0 is omitted and summation is over expanding rectangles. One
establishes (3.2.12) by writing
= IT
(-l)*
" "
2" cosech ( M k ) = 2 2
V7k
k=l
k = l m=-w
(_,)*+,
rk
[using (2.2.3)], and so
since C:=, (-l)"+' lm2 = 7r2/12. This is the desired result. One may observe that whenever one can evaluate the double sum, one can also evaluate
kk'.
In some future work we will be following Ramanujan rather than Weber.
Ramanujan studied
Gn:= (2kk')-1112
= 2-l14f(-)
For some purposes these give slightly cleaner results. For example,
For the moment we denote k ( q n ) by 6 and K ( 6 ) by A. We have
(233
kk' - = 4-12;
so that
We also have
and
These formulae lie considerably deeper, either in elliptic function theory, or

as direct computations. Indeed each formula entails formulae for .powers of
theta functions. We prove only (3.2.16) and leave (3.2.17) as an exercise.
We need an identity whose proof we also leave as an exercise.
Lemma 3.1
Proposition 3.2
For Iql< 1,
Proof. From (2.3.16), on differentiating with respect to q = e-"', and

Corollary 3.1 we have
I;:(
d
e ; ( q ) = -4q - - 3 = 8 q - log dq
),
+1
which yields (3.2.20). The other two follow similarly.

Theorem 3.2
For 141< 1,
(3.2.23)
and
Proof. Apply Lemma 3.1 to (3.2.20) to establish (3.2.23). To prove

(3.2.24) one uses the lemma with q and - q . 0
Now (3.2.16) is immediate from (3.2.24). Formula (3.2.17) is similarly
derived, if one knows in addition (see Rademacher [73]) that
Finally, (3.2.18) follows from (see Rademacher [73])
Of course, implicit in (3.2.23) is a formula for the number of representations of n as a sum of four squares. We discuss this more fully later.
72

1. Establish the asymptotic of (3.2.1).
Proposition 3.1 is often established by analytic arguments like those of

Exercise 6 of Section 3.1. Proposition 3.1 may then be used, as it is in
Whittaker and Watson, to establish the triple-product identity. (See Exercise 6.)
2. a) Show that
b) Show that
(This was set in Trinity College, Cambridge, in 1881.)

- 1)l a , then kk' = and Kt(
Recall
that if k := (fi
c)
f i ~ ( k ) Deduce
.
that
3. Establish the identities in (3.2.9), (3.2.10), (3.2.11), (3.2.14), and

(3.2.15).
4. a) Justify the derivation and convergence of (3.2.12).

b) Evaluate the left-hand sum when r := 1, 2, 3, 4, 7.
c ) Prove that as r tends to infinity, the sum converges to --rr2/6. It is
already "close" by the time r = 7.
d) Prove that
73
5. Show that
g4n = 2 1 ' 4 g n ~ n
Gn = GI,, and g,l = g4,n
4 ( ~ , G , ) ~ ( G-I ~gjf) =
6. a) Prove Lemma 3.1 by expanding both sides.
b) Prove (3.2.7) and (3.2.8).
c) Establish Theorem 3.2.
d) Prove. (3.2.16) and (3.2.17). Try to prove (3.2.18). Exercise 9 will
help.
The next two exercises sketch out some facts about Jacobi's imaginary
quadratic transformations.
a)
b)
a.
7. a) The transformation k, := k-l is given in theta terms by

= ($)2(e--(i+s-1))
b) Kt(k1) = kK1 and K(k,) = k(K + iKt).

c) The transformation k, := kt-' gives K1(k1)= klK and K(k,) =
kt(K' + iK).
d) The transformation k, := iklk' gives kl = -kt-', K(k,) = klK,
and K1(k1)= kt(K' - iK), since re@)> 0.
The
transformation kl = k'lik gives ki = k-l, K(kl) = kK', and
e)
K1(k,) = k(K + iK').
Z
8. a)
Use sl := i + s-' to derive from (2.3.14) that
b) Now use Exercise 7a) and (2.3.17) and (2.3.18) to show that when
k, := k-l, kE(kl) = E(k) - ( k ~ ) ~ ~ ( k ) .
Derive
similar formulae for E(k,) for the other three imaginary
c)
transformations.
l/Z.
d) Evaluate K'IK when k:= fl+
9.
a) For r := 1, 2, 3, 4, show that 03(irq)= 03(q4)+ ir8,(q4).
c)
Evaluate 0:-~8,k for k := 1, 2, 3, 4.
74
10. Show that
,,=--
(- 1)" cosech (3n + 1)7 ~ . 1

= - log [8(2 - fi)]
3n + 1
9
[Let r := 9, 2kkr = ( 2 - f i ) 2in (3.2.12). See (4.6.10).]

The formula for sums of two squares is derived from Jacobi's identity
This will be shown in Chapter 9.

11. A comprehensive list of hyperbolic identities can be derived, as was
(3.2.12). We have, for q := e-"",
1
k2
-E l o g ( m )
= - n-s
iT
(3.2.28iii)
1
-log Q 2 = 2
z
1
(-1)" cosech (mn)

n
7
m
02
1
log Q,Q: = 1% --m= n ( e z ~ n+~
2q
= 21- C1 1 - tanh ( r n s )
(3.2.28~)
(3.2.28vi)
2 log Q , Q ~= 2 log 0,
m
2K
= C (2n - ~ ) [ e " ( ~ "+- 11~ )=~-4 log (--)

1
(3.2.28vii)
1
- log Q,Q:
2
=-
51 1% 04
C tanh-l
1
(3.2.28viii) log
(F)
Q1(qZ)
"
(-I)n
=C
Qz(q )
1 n(e2"n"
l)
1
(
l
)
"
[
l
- tanh ( r n s ) ]
=-C
2 1
n
rs
--
(3.2.28ix)
1
e-""s = - 4 log
1 - k'
" cosech [ ~ ( 2-nl ) s ]

1
logQ,Q~=~
- - - log k'
1
2n-1
4
These are taken from Zucker [84]. Many more identities follow by
differentiation or as below.
a) Show that
Hint: Use (3.2.28ix). Then replace q by ql'Z so that k' is replaced

by ( 1 - k ) l ( l + k ) . Then replace q by -q so that k is replaced by
iklk' (as in Exercise 7).
b ) Show that
.=,
2n + 1
~ / 8
~ / 2 4
p:=l
p:=3
'
3.3 A COMBZNATORZAL APPROACH TO THE TRIPLE-PRODUt

IDENTITY
The Gaussian or q-binomial coefficients are polynomials defined by
and by
when 0 < m < n. Here ( q ) , is defined by
This allows for any complex value of s , but we will consider only integral
values. Thus for n in N, ( q ) , = fly=, (1 - q') and (q)7f, = 0. For m < 0 or
( )
m > n we either define

to be zero, or observe that it is implicit in our
m
definition of ( q ) , (for integral m ) . The next proposition gathers up some
easy facts.
Proposition 3.3
For lql< 1,
(c)
(i)q
is a polynomial in q.
Cauchy's binomial theorem is, for n = 1 , 2 , 3 ,
3.3 A Combinatorial Approach to the Triple-Product Identity
77
This is easily established inductively by showing that both sides of (3.3.3)

agree for n = 1 and satisfy
(3.3.4)
F ~ ( Yq), = F ~ - I ( Yq)
, + yqnFn-l(y, q ) .
One can also argue, combinatorially, that the coefficient of ysq' on each side
gives the number of partitions of t into s distinct parts not exceeding n.
If we now let y := x ~ and- n ~
:= 2 N , we obtain, after some manipulation,
a result also due to Cauchy.

This bears a striking resemblance to the triple product in lthe form of
2N
(3.1.13). Indeed on letting N tend to m one has ( N - m ) tending to
Q(q) := II;,, (1 - qm), and we are left with (3.1.13) if we can justify the
exchange of limit and summation (Exercise 3). Thus (3.3.5) deserves to be
considered as a finite form of the triple product.
The following Eulerian result will also be used in the next section. For
notational simplicity the empty product equals 1.
Proposition 3.4
For k = 1, ,2,
...,
Proof. Inductively one establishes that each side satisfies
Obviously the result is true for k = 1. Alternatively one can argue that the
coefficient of x t q h n each side counts partitions of s into t parts not
exceeding k.
Here and in the next section we follow Bressoud [83].

1. Prove Proposition 3.3 and (3.3.3).
78
2.
Verify (3.3.5).
3.
a) Prove that for 1 ql < 1,
b) Justify the exchange of limit and summation in (3.3.5) by observing

that the terms possess uniform majorants that are summable.
4.
a) Prove Proposition 3.4.

b) Deduce that
and compare with (3.1.19).

3.4
BRESSOUD'S 'EASY PROOF' OF THE

ROGERS-MANUJAN IDENTITIES
The Rogers-Ramanujan identities are
They afford a remarkable combinatorial ,interpretation (Exercise 1). We

use equations (3.1.12~)and (3.1.12b) and the notation of the last section to
rewrite the identities as
It is this form that we establish. The key observation is the following.

Lemma 3.2
For n positive and integral and any complex a,
3.4 Rogers-Ramanujan Identities
Proof. Note that the above sums are all finite. Set k := n - m and
x := q2m in Proposition 3.4. Then multiply each side by ( q ) i A . We get
Make this substitution for ( q ) i : m in the left-hand side of (3.4.5). We have
We now sum over m and s := m + j and exhibit the right-hand side of

(3.4.5).
The effect of Lemma 3.2 is to reduce the power of qm2 by 1. Repeated

use of the lemma results in an expression which can be handled by the finite
triple-product. As we will see, the k = 2 case of the next result contains the
desired identities.
Theorem 3.3
Given positive integers k and N, we have
Proof. Commence with applying the lemma to the right-hand side of

(3.4.6). We have
80
Some Number Theoretie Applications
We continue by applying the lemma k - 1 more times and arrive at
We use the finite triple-product (3.3.5) to write
This completes the proof.

The theorem is applied by specifying k and/or x. When x := -1, we
observe that unless sk = 0, the products on the left-hand side are zero. Thus
If we let N tend to
m,
we arrive at
For k : = 1, the left-hand side is 1 and the formula recaptures Euler's

pentagonal number theorem. For k : = 2, (3.4.8) coincides with the first
Rogers-Ramanujan identity (3.4.3). The second identity is derived similarly
by specifying x := - q and k := 2 (Exercise 3).
The Rogers-Ramanujan identities were discovered by Rogers in 1894 and

rediscovered by Ramanujan in 1913 (a letter to Hardy) and Schur in 1917.
They have continued to receive a great deal of interest. As recently as 1979
Baxter rediscovered them in a physical context.
Hardy and Wright say "no proof is really easy (and it would perhaps be
unreasonable to expect an easy proof)." Bressoud's proof certainly comes
close to contradicting this. Much of the early history can be found in
Hardy's footnotes to Ramanujan [62].
3.5
Some Number Theoretie Apptications
81
Show that (3.4.1) says that the number of partitions of n into parts with
minimal difference 2 is the number of partitions into parts congruent to
1 or 4 modulo 5. Equally, (3.4.2) says that the number of partitions of n
into parts with minimal difference 2 and minimal part 2 is the number of
partitions into parts congruent to 2 or 3 modulo 5.
Verify the equivalence of (3.4.1) to (3.4.3) and (3.4.2) to (3.4.4).
Derive the second identity (3.4.4) from Theorem 3.3. There is an
interesting continued fraction associated with the identities.
When x := e M A ( n rational), this is in principle evaluable in closed

form. Ramanujan gives
In a recent paper, Bhargava and Chandrashekar Adiga [84] detail

(3.4.9) and other continued fraction identities. The closed form (3.4.10)
and its extensions are elaborated on in Ramanathan [84].
Investigate the analogue of Theorem 3.3 in which 2 k replaces 2 k + 1.
3.5
SOME NUMBER THEORETIC APPLICATIONS
Our first application is a proof of Jacobi's formula for r4(n), the number of
representations of n as a sum of four squares [including sign and permutation so r4(2)= 24 and r 4 ( l )= 81
Theorem 3.4
For each positive integer n ,
In particular every positive integer is the sum of four or fewer squares.

Proof. Expand the right-hand side of formula (3.2.23) for B:(&
obtain
to
0:(q)=1+8):
3r n q m k = 1 + 8 2( C d ) q n
m=O k = l
n=l
dln
Now compare coefficients with o:(~) written as
The subsidiary conclusion is Lagrange's famous result. Our second application is an analytic proof of Fermat's theorem that any prime of the form
4k + 1 is the sum of two squares. It is convenient to define
and
(3.5.2)
w(n) := a,(n) + a,(odd n)
where odd(n) is the odd part of n (that is, the largest odd divisor of n). We
set w(n) := 0 for n 5 0. The proof of the following lemma is left as
Exercise 3.
Lemma 3.3
The value w(n) is divisible by 4 unless n is an odd square.

Theorem 3.5
An odd prime p is the sum of two integral squares if and only if it is

congruent to 1 modulo 4.
Proof. The 'only if' is immediate on consideration of residues mod 4.
Now argue as in Proposition 3.2 and Theorem 3.4 to show that
(since QoQ, = Q). Hence, on expanding these expressions (as in Exercise 12

of Section 3.7)
On multiplying by 0, and comparing terms we have
3.5 Some Number Theoretic Applications

r+l 2
(3.5.3) w(n) + 2
(-lliw(n
83
2
n =r
otherwise
- j2) = (2(i1)
jzl
'
Suppose now that n = p = 4m + 1 with p prime. Then w(p) = 2u,(p) =

8m + 4 and we see that
Thus some wxp - j2) is not divisible by 4 and Lemma 3.3 implies that
p = j2 + k2 for some integers j and k.
Our third application is to establish the following reciprocity result for
Gaussian sums. Let
where p and q are nonzero integers.
Theorem 3.6
For positive integers p and q with pq even,
S(P, q ) =
63
q l-i-
S(q, P)
(Here the bar represents complex conjugation.)

We will prove the result from the transformation formula for 0, (2.3.1).
This needs the following:
Lemma 3.4
For pq even and q positive,
Proof. Write
@,(E
+ iplq) := C;=-,
. ~h en
e-mn2(~+ip/q)
on taking the periodicity of e-n"ip/4 into account. (Here we have used

e-"iq2p1q
= 1, which entails pq even.) Thus
on using the integral test. We now take the limit as E+
Since S( p , q ) = Z:=,
o f and calculate that
e-"ir2p'q, this gives the desired result.
Proof of theorem. We now use the theta transform (2.3.1) to write
Thus the lemma shows that
~im
&'+OC * e 3 ( E '
- 3 + 0(&'2)) .
By Lemma 3.4 applied to - q and p we deduce that
Our final application is a partition result of Ramanujan's which relies on

the triple-product identity. (See Exercise 9 of Section 3.1 for the definition
of P . )
Theorem 3.7
( a ) p(5n
( b ) p(7n
+ 4 ) is divisible by 5 .
+ 5 ) is divisible by 7 .
Proof. With Q ( q ) = II;=, ( 1 - q n ) as before, we write
where k := 1 + ( 3 n + l ) n / 2 + m(m + 1 ) / 2 . This uses the triple-product identities (3.1.10) and (3.1.14) multiplied together. One now considers when k is
3.5 Some Number Theoretic Applications
85
divisible by 5. Since 2(n + 1)2+ (2m + 1)' = 8k(mod 5), we must have
2(n + 1)' + (2m 1)' divisible by 5. An inspection of residues shows that
this can only happen if 2(n + 1)2 and (2m + 1)' are both divisible by 5.
Hence, 2m + 1is divisible by 5 and so is the coefficient of q5m+5 in qQ4(q).
From the binomial theorem one deduces that
in the sense that all coefficients are congruent. It then follows that Q(q5)/
5m+5
in
Q5(q)= l(mod 5), and hence the coefficient of q
is divisible by 5. But
so that the coefficient of q5m+5 in q ~ - l ( q )is divisible by 5. However, by

Exercise 9 of Section 3.1,
Case (b) is similar, but uses the square of (3.1.14) instead of the product
with Euler's series. (See Exercise 5.) 17
The identity (3.5.1) was discovered by Jacobi on April 24th, 1828. He also
subsequently observed similar number theoretic interpretations of the formulae for 0; and 0:, and he gave an arithmetic proof of his theorem. The
identities can also be found in Gauss's unpublished work. A wealth of this
and similar information can be found in Dickson 171, vol. 21.
An analysis of the components of our proof of (3.5.1) shows that it is
rather simpler than that in Hardy and Wright [60], which proceeds from
r,(n), or in Rademacher [73], which uses elliptic function arguments. We
use only the triple-product identify and the AGM.
1. Show that r4(n) is 8 times the sum of the odd divisors when n is odd,
and r4(n) is 24 times the sum of the odd divisors when n is even.
and deduce that the number of representations of n as a sum of 4 odd

d.
squares is 16 C,,,.=,
d.df odd
The argument of Theorem 3.5 is due to Ewe11 [83]. We give the general
formula for r,(n) in Chapter 9.
3. Prove Lemma 3.3. Consider odd and even cases and use the multiplicativity of a,.
Gaussian sums arise naturally in the study of cyclotomic polynomials and
hence occurred to Gauss while studying constructible polygons. Apparently
this led Gauss to the lemniscate and thence to elliptic functions. The
formula (3.5.4) (with p :=2) plays a key role in establishing the class
number formula for binary forms. Landau [58] is sufficiently taken by the
result that he presents three proofs. The result we give is due to Dirichlet.
The case with p := 2 is due to Gauss save for the "detail" of determining the
sign of the complex square root. (See Landau for the significance of the
sign.) Our proof follows Bellman [61].
4.
Show Gauss's result. For q 2 1,
This, generalized, leads quickly to a proof of quadratic reciprocity. (See

Apostol [76a].)
There is a host of more recondite modular results on partitions. (See
Andrews [76] or Hardy and Wright [60].) The rule of thumb that additive
number theory is generally harder than multiplicative theory is born out by
the relative paucity of partition information.
5.
a) Show that (1- q)-P 5 (1 - qP)-'(mod p) holds for any prime p.

b) Prove Theorem 3.7(6).
There is a combinatorial proof of the theta transformation formula which

is suggestive of the arguments used in the sections on the Rogers-Ramanujan identities, in that it produces a 'finite theta transform' and moves to the
limit. The proof is due to Polya. Again we follow Bellman [61].
m
6.
2m )zk
a) ( z
k=-m
for any r and integral m.
3.6
fi
The M e l l i Transform and the Zeta Function
87
b) Let w := e2""', I a positive integer. Then
c)
(Here 1x1 is the greatest integer less than or equal to x.)

Fix s and t with t real and positive. Let 1 := [(mt)1'2] and z = es".
Then
e(s+2?rrk)121
-I12skd12
+ e -(s+2wrk)l21
2
d) Now let I go to infinity and use
e)
3.6
n1I2
7
(-)n +2nr ,
e -x2
lim 'n
= x.
V7-i
Deduce from c) and d) the following form of the general theta
transformation:
ii)
lim
n--
=-
if
n+m
THE MELLZN TRANSFORM AND THE ZETA FUNCTION
We continue our tour through theta function theory with a discussion of the
Riemann zeta function. This also allows us to catalogue a few useful
properties of the Mellin transform for future use.
The Mellin transform is a specialized Laplace transform defined by
For integrable functions with suitable behaviour at zero and infinity, Ms is

analytic in a strip a < re(s) < b. For example, the gamma function,
r(s) := J," e-"xS-' dx, is analytic in re(s) > 0. A most useful identity is
(3.6.2)
lom
f(xy)xs-' dx = y-"
/om
f(x)xs-' dx
y >0 .
88
Under mild conditions, the Mellin transform is invertible and one can
identify two functions whose transforms agree for re(s) > 0. (Certainly this is
true if the functions are transformable and continuous.) This is an appropriate place to state the following functional characterization of the gamma
function, whose proof is left as a guided exercise.
Theorem 3.8
The gamma function is the unique function f : (0, a)-t [O, w) such that
( 1 ) f(1) = 1
( 2 ) f(x + 1) = xf(x) for x > 0, and
(3) log f(x) is convex.
Many otherwise tedious facts are easy consequence of this functional
characterization of the gamma function. We list three whose proofs are left
as exercises. These are the functional relation r ( s ) r ( l - s) = rlsin r s
(1.6.6), the beta function formula p(s, t) = r(s)r(t) / r ( s + t) (1.6.7), and
the duplication formula
We now derive the functional equation for the Riemann zeta function
c(s) := Cz=l n-" re($ > 1. We observe in passing that 5 has an immediate
analytic continuation to re@) > 0 simply by writing
so that
(3.6.4)
l(s) =
-g( - y + l n - s
n=l
More interestingly, consider g(t) := [8,(t)
Thus
re(s) > 0 .
- 11/2.
For re(s) > 4 we have
3.6
The Mellin Transform and the Zeta Function
Here we have used the theta transform for 8, (2.3.1) to substitute for g on
[I, w). This leads to
if we evaluate the third integral.and replace t by l l t in the second integral.

Since Ig(t)l= O(e-"') as t d w , we see that the integral is an entire
function of s. Thus r(s/2)L(s) is analytic except for simple poles at s = 0, 1.
Since r(s) has a simple pole at 0, we see that l(s) is analytic except for a
simple pole at s = 1. In particular (3.6.5) gives an analytic continuation of
l(s) to the entire complex plane. Moreover, as Riemann discovered, the
right-hand side of (3.6.5) is invariant under the change of variable s := 1- s.
Thus we have
This is the celebrated functional equation for the zeta function.

It is also possible to deduce the Poisson summation formula from the

functional equation for the zeta function. (See Bellman [61].)
1. a) Use Holder's inequality to show that I? satisfies theorem 3.8.
b) Conversely, g(x) := log f(x) satisfies g(n + 1) = log (n!) and
Thus
0 5 g(x) - log
n!nX
5 x l o g (I
X(X+ 1)-- .(x + n)
+ );
and
r(x) = lim
n+-
n!nx
x(x + 1) - - . (x + n)
=fW .
This argument is due to Bohr and Mollerup. Interesting extensions

can be found in Askey [80] and in Denninger [84].
2.
a) Prove formulae (1.6.6), (1.6.7), and (3.6.3). In the latter case write
the ostensible identity in the form r(x) = f(x) and verify that f
satisfies Theorem 3.8.
b) Establish that I? has an analytic continuation to the entire plane
with simple poles at the negative integers and zero, and with no
zeros.
fi
90
3. Let
-(t+x2/41) f 1 1 2
dt.
e-I have the same Mellin transforms by using

Verify that f(x) and
the duplication formula. Hence reprove the result of Exercise 4 of
Section 2.2. Alternatively, establish the duplication formula from that
exercise.
4.
a) Use (3.6.4) and the fact that [(s) has a pole with residue 1 at 1to
show that
b) Show that [(-2n) = 0 for positive integral n. The reflection formula

(3.6.6) at least hints of the centrality of the line re(s) = in the
behaviour of the zeta function. The factorization, due to Euler,
l ( ~ =)
(1 -p-s)-l
re(s) > 1
p prime
shows the connection between prime distribution and the zeta

function. The Riemann hypothesis is that all the nontrivial zeros of
[(s) lie on re(s) = $. The asymptotic distribution of the primes is
inextricably tied up with this famous conjecture. (See Rademacher
[731.)
5. Use Theorem 3.8 to establish Gauss's multiplication formula,
6.
(Stirling's formula)
The formula is
Outline: Substitute x = s(1 + u) in Ms(ex) and write
Now replace u by t
m and obtain
for a (dominated) kernel &(t) which approaches e-t2 uniformly for

bounded t.
3.7 Reciprocals of Fibonacci Sequences
3.7
91
EVALUATION OF SUMS OF RECIPROCALS OF FIBONACCI

SEQUENCES
Since the theta functions provide quadratic analogues for the geometric
series, it is natural to ask about their relationship to sums of the form
m
(3.7.1)
C a,'
where
a,,, := Ma,
+ Nan-,
n=l
with a, and a, given (M, N # 0). The one-term (N = 0) recursion leads to

the geometric series. The two-term recursion leads to theta series and their
relatives. If a # p are the roots of x2 = Mx + N, then
(Exercise 1) and
a+p=M
ap=-N.
We will consider only the case in which N = + 1 and M is real. We may write
a,,, = (2c)a, + &an-,, Is t = 1, and to assure that a , p are real, we assume
c>max (0, -6). Then ap = + 1 and
We must consider summing series of the form
The following proposition provides the key. We define the Lambert series
Proposition 3.5
For
and
O < p < a with ap = l ,
92
Similarly,
and
Proof. ( i ) The first equality follows on multiplying top and bottom by

13:. ( i i ) Now
observe that a 2 and p2 satisfy the hypotheses of ( i ) . Then (iii) follows on
subtraction. We leave (iv), ( v ) , and ( v i ) as Exercise 2.
pn and the second now follows from (3.2.27), the formula for
Proposition 3.5 will allow us to sum C:=, a,' for certain starting values.
Specifically, we can handIe E = +-Iand A = + B in (3.7.3) and (3.7.4).
1 ( A = - B) Let a, := 0 and a , := 1. [This normalization comes by

setting A := (2-)-'.]
CASE
(i)
(E=
-1)
Then a p = 1 and c > l . Now
as follows from (3.7.2) and Proposition 3.5 (iv). Similarly, one can
evaluate
and
m
C a;,:,
n=O
=2
V T i [ L ( p ) - 2 L ( p 2 )+ L ( p 4 ) ].
(ii) (E = 1) Then
ap = -1.
Now
C a:;
E2GT-i
93
P < 0 and
1
= 2 V 2 z [ L ( p 2 ) - L(p4)]
a2" - p2*
n=O
,,=I
c. Q;,:,=2GTiC
-- rn
2
aZn+l lplZnf1
ce:(l~I) - G(P"I
-.
CASE
2 (A = B) Let a, := 1 and a, := c. [This normalization comes from
(i) (E = -1)
Then
(ii) (E = 1) Then
ap = 1 and c > 1. Now
p < 0 and a 1 /.? 1 = 1. It follows that
and
In certain cases the theta series involved above are particularly simple to
evaluate. For example, in Case 2(i), we have p = c - -.
Thus c =
( p p -l) 12. If p := lo-", the series 8,(P) and 8,(p2) in (3.7.8) can be
evaluated entirely by writing down sequences of 1's and 0's. Thus for
2c = 10" + lo-", the sums in (3.7.8) can be evaluated at the same speed as
one can multiply n-digit numbers. (See Chapter 6.) Moreover, since theta
functions can be fast computed for any algebraic P, the series are always fast
computable. (See Chapter 7 and Exercise 6.) Hence the series of Exercises
3, 4, and 5 are quadratically computable.
Finally, consider (3.7.7) for c := sinh ( m ) and with s > 0. Then
an+,:=2sinh(ns)an+a,-,, ao:=O, a l : = l , and
m
(3.7.10)
S(s) := C a Z 1 =
n=O
cosh ( m )
[O,(e-?" - ~ ~ ( e - " ~.) ]
2
If we use Theorem 2.1, we have
where
on using
For singular values of k (see Section 4.6), this formula becomes particularly
pretty. [See Exercise 6b).]
1. a) Show that when x2 = Mx + N has distinct roots a and p, an+,=
Ma, + Nan-, is solved by (3.7.2).
b) If a = p, show that
2. Establish (iv), (v), and (vi) of Proposition 3.5.
3. a) Use (3.7.7) to show that for the Fibonacci numbers
Here F, :=O, Fl := 1, and Fn+,:= Fn+ Fn-,. This result is due to

Landau [1899], as is
This is also discussed in Ribenboim [85].

The Lucas numbers are defined by the same recurrence,
Ln+,:= Ln + Ln-, but with Lo :=2 and L, := 1. (See Hardy and
Wright [60].) Show that
satisfies xn+,:= 3xn - xn-I , xo := 1, and

Note that xn := FZn+,
x,:=2. Similarly, x n : = Lzn satisfies xn+,:=3xn-xn-,, xo:=2,
and xl := 3.
Verify (3.7.8) and (3.7.9).
Show that for an+,:= 4an - an-, , a, := 1, a, := 2, one has
and
5.
L e t a o : = 0 , a l : = 1 , a n d a n + l : = 9 . 9 a n + a n ~Showthat
l.
6.
a) Recall that o : ( ~ )= (2/r)K(k), where q = e-"K"K(k)

. As shown in
Chapter 7, it is possible to quadratically compute k given q. Since
K is quadratically computable given k, we can fast compute B,(q)
given q.
b) Show that with S(s) given by (3.7.10), we have
96
C) Show that S(s) = cosh ( m ) E;=, sech [(2n + l ) m ] , so that

m
Og(e-") - O:(e-"')
sech [(2n + l ) m ] =
'7
and
O;(e-"")
co
sech ( n m ) =
+1
'7
d) Show that
x
a
sech [(2n + l ) ~=]
(2 - f i ) r 2 ( $ )
n=O
7.
1 6 ~ 7
Equations (3.2.28) and (3.2.29) can be used to derive closed forms for
a host of other reciprocal sums. (Fn and Ln are as in Exercise 3.)
a) Use (3.2.29) to show that
where p := (3 - ~) 12.
b) Use (3.2.28iv) to show that
1
5--.,, 2nFzn
v3 log
12
where p := (fi
- 1) 12.
Use (3,2.28i) and transformation formulae to show that
1 2
l+k'2K2-KE
cosech2(nas) = - + 7
n=l
6 T
x
m
i)
[(?)
ii)
n=l
m
iii)
n=O
y)
4 [(q)
[(
1+ 7
cosech2 (2nm) = 6 T
2 coseeh2[(2n + l)?rs] =
i)
2EK 1
2 sech2( n m ) = -m2
2
n=l
K2
K2 - KE].
d) Similarly, use (3.2.28~)to show that

m

m
ii)
C sech2(2nm) = E K +7T k ' ~ '- -21
n=l
m
iii)
2 sech2 [(2n + l ) r s ] = EK - k ' ~ '

2
77
n=O
Show that
ii)
C sech2
n =O
27T
'
Combine results of c) and d) to show that, with P := (3 - fi)

12,
and
and deduce similar formulae for Cz=, (-1)"~;:

(-1)"Lin2.
Show that the Lucas numbers satisfy
and for C:=,
and show that a similar formula holds for all recursions of the form
an+,=(2c)an+ an-,, ao:=2c, and a;:=l.
Show that
Show that
Hint: Differentiate (3.2.28iii) and compare the result to Exercise

7f), equation (3.7.12).
Show that
-2
2 (-1)"m
cosech (2mm) =
m=l
2 sech2 [(2m + l ) m ]
m =O
and hence that
9. Show that
and
10. a)
Show (by expanding both sides) that
b) Hence show that all the series considered in this section can be
computed with at most O ( f i ) operations for n digits. (See
Chapter 6.)
A remarkable elementary result is
This is a specialization of identities established in Backstrom [81]. (Related

results can be found in Carlitz [71].) Thus
A related formula is
;i
ai" ,
Show that
where q := (fi
- 1)/2.
Now k is close to 1, so that K E / ~ IisTclose

~ to
2
K
-=-=
2-2
47r
~~/lo~[(V7-1)/2]
I
63[e
4 log [(fi
+ 1) 121
where the last equality follows from the theta transform. One sees that
K E / ~ T 3~ is~ close
+
to
which explains, in some part, Backstrom's formal manipulation in

Backstrom [sl]; and which evaluates (3.7.14) as requested therein.
Alrnqvist [Pr] treats this sum and some relatives. Note that
e m2110g[(V7-1)121
loF9SO that, as in all these Fibonacci series, transformation considerably speeds convergence.
Lambert series occur naturally in multiplicative number theory, as the

.owing exercise shows.
12. a)
Show formally that for any real valued function f,

m
xn
Zf@) iTp nt]= l F(n)xn

n=l
-
where
i)
%--1-x"x n
C,
r(n)xn
n=l
where r(n) is the number of divisors of n.
where u,(n) is the sum of the kth powers of divisors of n.
if n is square
if n is nonsquare.
1=
Show that f(n) = (-1)'"'
where
rn
n = n[ pfi
(in prime decomposition) .
i=l
Hint: Use Mobius inversion. Since F is multiplicative, so is f. Thus

using Liouville7sfunction e(n) := (- 1)Eglei
c)
Show that
where p := (3 - f i ) / 2 .
Zucker [79] gives general formulae for sums of powers of hyperbolic
functions (in which the coefficients are defined recursively and have been
computed extensively by Ramanujan and Zucker). Using these one can
~
of 8 when
evaluate C:=, (- I ) " + ' F , ~ ~ and C;=, F , ~ ~in+ terms
k := 172,3,. . . . There are similar Lucas number results, and if K and E are
used, many more sums are expressible. We give two examples:
and
Here, as before, P := (3 - v 3 ) / 2 .
13. L e t u , : = O , u , : = 1 , a n d u n + l : = a u n + u n ~ l . L e t v o : = 2 , v l : = a , a n d
v , + ~:= av, + vn-I .
a) Establish that
Hint: The second sum in a) can be made to telescope.
b) In particular,
c)
Compare
and
ii)
x"
n=1
d) ~ h o kthat
k2
= 6(3 - 4 log 2).
Chapter Four

Abstract. We develop algebraic transformations of prime order for the
elliptic integrals. For small numbers this can be managed purely algebraically.
However, the development of modular equations for arbitrary primes is
most comfortably effected via transcendental rnethodr. This requires iome
rudimentary modular function theory. The cubic equation is studied in
particular detail.
4.1 A FIRST APPROACH TO HIGHER ORDER TRANSFORMATIONS

The fundamental relation from Theorem 1.2
is remarkable for a number of reasons. One notable consequence is the ab

initio unlikely observation that when k is algebraic and
the values of the transcendental function K at 1 and k are algebraically

connected. Equation (4.1.2) is one form of the quadratic modular equation.
It can be rewritten as
We will develop a class of algebraic equations (modular equations) that

induce algebraic transformations on K in a similar fashion.
We commence by sketching, ii la Cayley [1895], a purely algebraic
102
4.1 A First Approach to Higher Order Transformations
103
ch to the modular equation. It transpires that this approach becomes

complicated for all but a few simple cases and is hard to use as a
gorous basis for the general theory. Thus in Sections 4.3 and 4.4 we will
erive the general theory using function theoretic techniques.
We are looking for a relation of the form
M(6 k ) dy
dx
- k2x2)
v
(
1
x2)(1
-y2)(l - l Z 2 )
here 1 and k satisfy a polynomial equation 0 in two variables. Such an

gebraic connection between the two moduli (1 and k ) is called a modular
uation or modular transformation for k . Specific but equivalent modular
uations for algebraically related functions are defined in (4.4.2), (4.4.6),
d (4.5.1). The function M will be an algebraic function of k and 1 and is
lled the multiplier. With this in mind, let P and Q be polynomials in x2 so
at deg(P + x ~ ) ~ ( xl )+=- n ( n odd) and write
1-y
l+y
-=
( P - X Q ) ~-.
I-x
( P + x Q ) ~1 + x
important condition to impose on P, Q , and 1 := l ( k ) is that (4.1.5)

t be invariant when ( x , y ) is replaced by ( l l k x , l l l y ) . (See Exercise 1.)
now try to solve for P and Q . Set
Also, from the invariance of (4.1.5),
(4.1.7b)
l - l y = ( l - k x ) ~ ~ / ~1 + l y = ( 1 + k x ) ~ 2 1 ~
where C and D are polynomials of the same form as P and Q . Thus we

deduce that
t can now be computed (Exercise 1) that
104
1 - = riv- vu
Q(0)
-=1+2M(1, k)
ABCD
P(0)
is independent of x. Let us return to (4.1.5). For n = 4p + 1 we must have

(4.1.10)
degP=2p
degQ=2(p-1).
For n = 4 p + 3 ,
(4.1.11)
deg P = 2p
deg Q = 2p
Since P and Q are even functions, we have in either case 4 (n + 1) degrees of

freedom. It remains to use the invariance in (4.1.5) as (x, y) -+ (1 lkx, 1/ly).
This leads immediately to the equation
since y = U/V. Thus we are reduced to solving
where the * operation is defined as follows. If
then
Equating coefficients in (4.1.13) leads to a system of $(n + 1) nonlinear

equations. There are $(n + 1) degrees of freedom in the coefficients of P
and Q. Since the equation is homogeneous in P and Q, one of the
coefficients may be assumed to be 1. This leaves one additional condition to
be satisfied by k and 1 and leads to the desired unique algebraic relation
between k and 1. The pitfaIIs of this approach are now, of course, apparent.
We end up with a large system of nonlinear equations that are virtually
impossible to solve or analyse directly.
We illustrate with the cases 3 and 5.
CUBIC TRANSFORMATION
(4.1.13) becomes
(n = 3). We have P = 1 and Q = a. Equation
: This leads to two equations,
This is solved parametrically by
and (l')', namely,
with similar expressions for
.1.15b)
k f 2=
(1 - a ) ( l +
2a+1
and
lf2=
( 1 + a ) ( l - a)3
( 2 f f+ I ) ~ *
From this one can deduce that
(4.117)
(k2- 12)4 = 128k212(1- k 2 ) ( 1- 12)(2- k2 - l2 + 2k212).
In the associated variables u := k1I4 and v := ill4 this has a simpler form,
u4 - v 4 + 2uv(l - u2v2)= 0 .
See Section 4.5.) The muItiplier M has any of the following forms:
- .
zv3-
(See Exercise 2 of this section and Exercise 2 of Section 4.6.)
( n = 5). We have P = 1 + px2 and Q = a. Equation (4.1.13) leads to the three equations (h2:= k 5 / l )
QUINTIC TRANSFORMATION
'
This eventually solves, with u := k1I4and v := 1 ' I 4 , as
106
(4.1.21)
u6 - u6
+ 5u2u2(u2- u2) + 4uv(1-
u4u4)= 0 .
The multiplier M has any of the forms
The transformations are called cubic, quintic, and so on, because of the
underlying order of the transformation (and so of convergence).
The very classical approach of the section to cubic and quintic modular
equations is due to Jacobi [1829]. This algebraic approach was extended to
the septic (n = 7) case by Cayley [1874], who also treated the endecadic
(n = 11) case partially. The calculations are formidable. We have followed
Cayley closely in this discussion. The associated variables u := k1I4 and
v := l1I4 of JacobiYsconsiderably simplify the calculations, as we will see in
Section 4.5.
In order to find the relationship between K(1) and K(k) implicit in (4.1.4)
we must show that the underlying transformation (4.1.5) is one to one and
onto on the interval [0,1]. This can be done directly for n = 3 (Exercise 3)
and other small n. However, as with most of the details, it is easier to use
the general transcendental approach of Sections 4.3, 4.4, and 4.5.
1. a) Show that an equation of the form (4.1.5) that is invariant under
the change of variables (x, y)+ (1 lkx, llly) exists. Hint: Set
Replacing (x, y) by (1 lkx, 1Ily) gives
which holds, provided that
[See also (4.1.13) .]

b) Establish (4.1.7b) and exhibit C and D.
c) Establish (4.1.9).
Hint: Let U and V be as in (4.1.6). Let y := UIV and let
Y(a, b) := (a2 - b2)(a2- 12b2).Then
Y(V, U) = ( v 2 - u 2 ) ( v 2- Z ~ T J ~ )
and
Now observe that any square factor (x - a)' of Y(V,U ) is a linear

factor of vu - uI? Since
and
the fact that ( m - VU)/ABCD is independent of x now follows.

The explicit form of the multiplier M(1, k) is derived by setting
x = 0 in ( m - VU)/ABCD.
Establish the four forms of the cubic modular equations (4.1.15),
(4.1.16), (4.1.17), and (4.1.18).
Complete the calculation of the quintic modular equation (4.1.21).
Show that the n = 3 relation between x and y underlying the cubic
transformation is one to one and onto on [O,l], and hence (4.1.4) can
be integrated over [0, 11.
(An explicit cubic algorithm for K) The cubic modular equation
(4.1.18) is of degree 4 and, hence, can be solved explicitly for u in terms
of v.
a) Show, for v E (0, I), that u E (0, v) is a solution of (4.1.18), where
u = - v3
+2
D-R
2
and
b) Show, for u E (0, I), that v E (u, 1) is a solution of (4.1.18), where
108
and
c) Show, for v E (0, I), that there is a unique u E (0, v) so that

(4.1.18) is solved by (u, v). . We can define an iteration as follows.
For vi E (0, I), let vi+, E (0, vi) be such that u := vi+, and v := vi
satisfy the modular equation (4.1.18). Show, using a) and b), that
where v, E ( 0 , l ) and vl E (0, v,) is computed from v, by Exercise

4a).
d) Show that, for v, E (0, I),
(Use Exercise 3.)

e) Show that vi tends to zero cubically. (For further details see
Borwein and Borwein [84b].)
5.
a) Let u E ( 0 , l ) and v E (u, 1) satisfy the cubic modular equation

(4.1.18). Show that
Hint:
b) Let u E ( 0 , l ) and v E (u, 1) satisfy the quintic modular equation

(4.1.21). Show that
4.2 An Elementary Transcendental Approach
K'(u4)
K'(v4)
-- 5 -.
(4.1.26)
K(u~)
Hint: Show that if
K(v4)
(u, v ) := ( k114, I 114 ) satisfies (4.1.21), then
so
does ( u , v ) = ( z " / ~ , k' 1 4 ) . NOWExercise 1 of Section 1.5 shows that
the two sides of the above equation differ by a constant. Use the
logarithmic asymptotic at 0 and the relationship between u and v as
v +0 to evaluate this constant.
These important identities will be revisited in Section 4.4.

6. a) Verify Schlafli's form of the modular equation of degree 5 ,
(4.1.27)
where u := 2-lI4f(7) and v := 2-'I4f(57), f as in (3.2.9).

b) Compute the corresponding equation for f l .
4.2 . AN ELEMENTARY TRANSCENDENTAL APPROACH TO
HIGHER ORDER TRANSFORMATIONS
In terms of the nome q we have, by Theorem 2.3, the identification
k(q):=k=%l)
kyq):= k'= G(q)
q = e-nK'(k)lK(k)
From Exercise l e ) of Section 1.4 and Exercise 5 of Section 4.1 we see that
the quadratic modular equation (4.1.3) is satisfied by 1 := k(q1'2) and
k := k( q), while the cubic equation is solved by 1 := k ( q 1 1 3and
) k := k ( q )
and the quintic equation is solved by 1 := k ( q l 1 5 ) and k := k( q). [To see this
just observe that (4.2.4) uniquely determines q.] In general we will see in
the next sections that the pth-order modular equation for k is a polynomial
in two variables with integer coefficients that is satisfied by k ( q P )and k ( q ) .
We observe from (4.2.4) that for these algebraically connected moduli
110
Before turning to the general theory we wish to give an elementary

derivation of the cubic transformation in theta function terms. From
(4.1.16) the cubic modular equation for k is
From the preceding discussion this is seen to be equivalent to:

Theorem 4.1
Proof. From the definitions and Exercise 1 of Section 2.1 applied to

qn2+3m2
and (-l)n+mqn2+
3m2 we have
and
Now
(h+j+1)2+3(h-j)2=(2h-j+ $)'+3(j+4)'.
Thus subtraction of the two theta identities produces
m
63(qP3(q3)- 4(q)64(q3)= 2
IC
m,n=-m
(m+ f)'+3(n
+J ) 2
m+n even
(as replacing rn by 1 - rn shows). 0

This is as far as we wish to pursue the transformation theory on an ad hoc

basis. The next section introduces enough of the theory of modular functions to provide a general framework for the development of modular
equations. We have only considered modular equations for p a prime. If we
view the modular equation as the algebraic relation between k ( q n )and k ( q )
and can find this relation for p a prime, then for composite n a relationship
4.2 An Elementary Transcendental Approach
111
constructed out of the modular equations corresponding to the prime

of n. Rational n are treated similarly.
Construct modular equations of order 4, 6, and 8 [that is, construct
algebraic relationships that are satisfied by k(q) and k(q2), k(q6), and
k( q8), respectively].
Show, for
that
O 5 k(qpn)5 4eWCp" 4 E
and that
Many modular identities follow from:
3.
(Schroter's formula) Consider a general theta function written as
where x Z 0, q ( < 1 (as in Section 3.1). Let a and b be positive integers.

a) Show that
Hint: Write
Lets bechosenso that n = m + ( a + b ) s + k ( 0 r k < a + b) andlet

u := m bs. Then u and s range over Z as m and n do. Also
xmyn= ( x ~ ) " ( x - ~ ~ ~ ) " ~ ~

and
am2 + bn2 = (a
+ b)u2 + 2bku + ab(a + b)s2 + 2abks + bk2 .
Now rearrange. (See .Tannery and Molk [1893].)
112
b) A form of the seventh-order modular equation is
(See (4.5.4).)
Use part a) with a := 7 and b := 1 to establish this formula.
Hint: Set x:= y : = + l to find a formula for o , ( ~ ) B , ( ~ ~ ) +
B ~ ( ~ ) B ~ ( ~Now
: ) . set x := y7 .- k Y 7 to similarly write
62(q)ez(q )f82(-q)ez(-q7)=ez(q)ez(q )- On making
rearrangements this yields
Thus an application of equation (2.1.7ii) gives
c)
as required.
Establish the cubic modular equation (4.2.7) as above.
4.3 ELLIPTIC MODULAR FUNCTIONS
The theory of elliptic modular functions and more general automorphic

functions is, in part, a natural extension of the theory of elliptic functions.
The basic defining property of elliptic functions is their invariance under a
group of linear transformations. Automorphic functions are functions
meromorphic in the upper half-plane X := {im(t) > 0) that are invariant
under a group of linear fractional transformations. We will, of necessity,
explore only the rudiments of this remarkable and difficult theory.
4.3 Elliptic Modular Functions
113
The (inhomogeneous) modular group T (T-group) is the set of all

transformations of the form
W=-
at + b
ct + d
a, b, c, d integers,
) The A-group is the subgroup A of
ad - bc = 1.
r with a, d odd and b, c even.
That both of the above are groups (under composition) is straightforward.

e transformation
w=- at + b
ct + d
n be represented as either of the two matrices
oduct becomes matrix multiplication (See Exercise 1.). The

modular group SL(2, Z)distinguishes these matrices. Note
t of the modular group fixes the real axis and maps X onto
(a) The set Fr of t E %* := {im(t) > 0) U {im)U {Q) is defined by

Fr := {lre(t)l< and It/ > 1) U {re(t) = - and It1 r 1)
U {It1 = 1 and -
4 5 re(t) 5 0) .
( b ) The set FAof t E X* is defined by
FA:= {lre(t)l< 1and 121+ 11> 1) U {re(t) = -1)
u 1121+ 11= 1) .
These two sets described in a) and b) are fundamental sets for the T- and
A-groups. The interiors of these two sets (F! and F:) are fundamental
regions in the following sense.
114
Theorem 4.2
(a) Every point in X is the image under some element of

of Fr.
of some point
If A E I? is not the identity, then A(F;) fl F: = 0.

(b) Every point in X is the image under some element of the A-group of
exactly one point of FA.
If a E A is not the identity, then A(F:) fl F: = 0.
The proof of this theorem is elementary though not entirely straightforward. (See Exercise 2.) Any set F (with interior F O )which satisfies (a) or (b)
of the above theorem is also a fundamental set.
Definition 4.3
(a) A r-modular function is a function f which satisfies:

(i) f is meromorphic in X.
(ii) f (A(t)) = f(t) for all t E X * and A E I'.
(iii) f(t) tends to a limit [possibly infinite in the sense that 1lf(t) -t 0]
as t tends to the vertices of the fundamental region Fr where the
approach is from within the fundamental region F:. [In the case
of iw the convergence is uniform in re(x + iy) as y-tw.1 The
vertices of the fundamental region are (0, I), (-112, f l 1 2 ) and
iw. Since f is meromorphic in X, this condition is automatically
satisfied at ( 0 , l ) and (-112, V?f/2) and need only be checked at
iw.
( b ) A A-modular function is a function f which satisfies (i), (ii), and (iii)
above with the r-group replaced by the A-group. For condition (iii)
the vertices of the fundamental region F: are (-1, O), (0, O), and iw.
Our notation is not entirely standard. What we have termed r-modular is
often just called modular or automorphic with respect to the I?-group, while
what we have labelled as A-modular is often referred to as automorphic or
modular with respect to the A-group.
The existence of a A-modular function is the content of the following
theorem.
Theorem 4.3
The function
is a A-modular function.
Proof. From Corollary 3.1 we have
d it is clear that h is meromorphic in 32'. For the invariance of h under the

-group it suffices, by Exercise le), to show that
h(t + 2) = h(t)
A(-)
= h(t)
st equation follows since e"" = e"i('+2). The second equation is a

uence of (2.3.1), (2.3.3), and (2.1.10), which combine to yield
hat t = is.) Hence
Finally we observe that, in a limiting sense,

.3.5)
h(im)=O
h(O)=l
h(+l)=m.
The first value is immediate from (4.3.1), while the value at zero can be
alculated from (4.3.4). (Observe that as t-0 in FA, llt-+a in FA.)The
alue at 1 is computed from Jacobi's imaginary transformation
- (See Exercisc
'
Some additional properties of A are established in Exercises 4, 5, and 10.

From (4.3.4) and (4.3.6) one can prove, as in Exercise 6, the following
theorem.
Theorem 4.4
The function
is r-modular. J is called Klein's absolute invariant.
116
The basic result we need is a version of Liouville's theorem.

Theorem 4.5
A r-modular function that is bounded on F, is constant. Similarly, a

A-modular function that is bounded on FAis constant.
Proof. Suppose f is I?-modular and is bounded and nonconstant on F,.
By Theorem 4.2 this implies that f is analytic on X. Consider f(t) - f(iw).
This function has no poles interior to F,, and so achieves its maximum
modulus at some finite point on the boundary of F,. By the invariance of
f - f(iw) under I? this is a global maximum at an interior point of X, which is
impossible.
For the second part consider the A-modular function
This is sufficient theory for our discussion of modular equations in the

next section.
This is only the very tip of the iceberg. We have restricted our attention to
two particular groups where we can directly establish the existence of
modular functions. In general this restriction is unnecessary. Only slightly
further into the theory are results such as: any modular function takes each
complex value the same number of times in the fundamentaI region. An
important consequence of this is that h takes every value exactly once in FA
and h has a well-defined inverse that has branch points only at 0, 1, and
(Exercise 10). J has similar properties on F,. One can now prove, much as
for elliptic functions, that two nonconstant functions which are modular with
respect to the same group are algebraically connected. Furthermore, if one
of these functions is univalent on the fundamental region, then the other is a
rational function of it.
This wide-ranging and difficult body of theory that is intimately tied in to
many questions in number theory and algebraic geometry may be pursued in
any number of texts, such as Apostol [76b], Chandrasekharan [85], Lang
[73], Lehner [66], Rankin [77], or Schoeneberg [76].
Two of the seminal papers of the subject, both dating from 1882, are due
to Klein and PoincarC (selections of which may be found in Birkhoff [73]).
Poincare was interested in studying linear differential equations with algebraic coefficients. Klein, who considered this his main field of work, in
keeping with his Erlanger Programme had more algebraic and geometric
interests (Klein and Fricke [1892]).
4.3 Elliptic Modular Functions
117
a) Verify that T and A are groups.

b) Verify that composition of transformations is equivalent to multiplication of the associated matrices.
c) Show that two transformations represent the same function if and
only if they have the same coefficients (associated matrix up to
sign).
d) Show that T is generated by
S
and
T~:=(!
-:).
e) Show that A is generated by

1 2
SA:=(o
and
1 0
TA:=(2
Prove Theorem 4.2.

a) Hint: To prove part (a), fix z E R a n d let A =
that
(: f;) E T.Show
Pick any element of that minimizes Icz + dl and let w be the

image of z under this transformation. By considering T,(w) show
that Iwl? 1 and by considering s ~ ) ( w )show that for some k,
I r e [ ~ ( ~ ) ( w ) ]1.
l s For this k, (sfk)(w)(
? 1. Thus every element of
R is the image of some element of F,.
b) Show that no two elements of F: map to each other under an
element of r. Examine the image of F, under S, and T,.
c) Deduce part (6) of Theorem 4.2 from part (a).
The upper half-plane can be tesselated by images of the fundamental
region under the generating transformations. Sketch pictures of the
tesselations associated with the r-group and the A-group.
a) Show that A(A) is transformed into one of A, 1- A, 1/A, 1/ ( I - A),
A/(A - I), 1- 1/A by any A E I?. Hint: Examine A(t + 1) and
A(-llt) using similar arguments to those of Theorem 4.3,. Use
Exercise Id).
b) Show that A(mln) = 0, gcd(m, n) = 1, if and only if m is odd and n
is even.
Show that with q := eiw:
where the b, are integers. Show that llA(t) is finite at every point of
F, - {ia).
a) Prove Theorem 4.4 from (4.3.4) and (4.3.6).
b) Let q := e2"': Show that
where the c, are integers. [In fact, qj(t) = ~ y ( 2 5 6 q+ Q,

C) Show that f 24(t), -f y(t), and -f y(t) are the roots of
-24
) .]
[See (3.2.9) for definitions.]

a) Show that iff is r-modular then for some integer k and nonzero
constant c
as t+ i a .
Hint: Let f(q) := f(t) where q : = e2"" for t in F,. Show, by
modularity off, that f is meromorphic in q in a neighbourhood of
zero. Show also that f has a pole at zero. Thus f has a convergent
+expansion at i a with finite principal part.
b) Show that the r-modular functions form a field. Likewise the
A-modular functions.
Establish that f y , f y , and
modular.
f24
[see (3.2.9) for definitions] are A-
The Schwarz derivative of any function f is
Show that iff is modular, then so is ~ ( f ) l ( f ) ~Show

.
that f is not in
general modular.
(On the inverse of A)
a) Show that A maps the set A := {re(z) = -1,0 < im z) one to one
onto (--a, 0). Hint: A(t f 1) = A(t)l[A(t) - 11. Now consider A on
the imaginary axis.
b) Show that A maps the semicircle B := {lz + I = 1, im(z) > 0) one
to one onto (1, a). Hint: A(-1 It) = 1- A(t).
Show
that A maps the interior of FA one to one onto C c)
4.4 The Modular Equations for A and j
{(-a, 01 U [I, 00)). Hint: The number of zeros of h - c is
where y is a contour of the form seen in the accompanying figure.

Use the invariance A(t) = A(t + 2) to estimate the integral on the
sides of the contour. Use the relation A(-llt) = 1- A(t) to relate
the integral on BC to the integral on ED. Use the relation
A(t Ir: 1) = A(t)l[A(t) - 11 to estimate the integral on CD and DC
in terms of BAB and then (by t-+ -1lt) in terms of EE. Finally
take limits.
d) Thus with respect to FryA has a well-defined analytic inverse with
branch points at O,1 and a.
11. (Picard's theorem) Show that a nonconstant entire function assumes
every complex value except possibly one.
Hint: Suppose F does not assume either a or P, then G(z) :=
[F(z) - a l l ( @- a ) never assumes 0 or 1. If w := A-l, then o(G(z))
is entire (by analytic continuation). Show that w(G(z)) is constant
since w(G(z)) E {im(z) r 0) and hence e"(G(z))is a bounded entire
function.
It is worth observing that the apparently special case analysis of the
function A leads directly to the celebrated general theorem of Picard.
THE MODULAR EQUATIONS FOR A AND j
4.4
A transformation of order p is a matrix

(4-4.1)
i)
a, b, c, d integers,
ad - bc = p
or the associated linear fractional transformation. We assume throughout
120
that p is an odd prime, though for much of the development this is

unnecessary. We will denote the set 'of all such transformations by T,. We
say that M is equivalent to N mod G ( M = N mod G ) for a group of
transformations G if there is an S E G so that M = SN. We need the
following purely algebraic result.
Lemma 4.1
(a)
Every M E Tp is equivalent mod

of the set d , where
to one of the p + 1transformations'
( b ) The p 1 elements of d are pairwise inequivalent mod r.

(c) Every B of the form B := BiC, where C is in the A-group and Bi E 93,
is equivalent mod A to some element of 93, where
(d) The p + 1 elements of 93 are painvise inequivalent mod A.

The proof is left as Exercise 1.
Theorem 4.6
(a) The p
+ 1 functions
are permuted by any element of the r-group.

(b) The p 1 functions
are permuted by any element of the A-group.

Proof. For part (a) we must show that { J o A i 0 SIP = { J o
for any S E T. We first observe by the lemma that if
I', then since
AiS E T p ,
AiS=Ajmodr
Thus by the modularity of J ,
forsomej
121
which with part (b) of the lemma finishes the proof. The second part is
identical via parts (c) and (d) of the lemma. 0
The modular equation for A of order p is the polynomial
P
wp(x,~):=n(x-Ai)
ni:=no~i.
i=O
This is obviously of degree p + 1 in x and has a root at each A,. Note that
Ap(t) := A( pt) and Ai(t) = A((t + 2i) lp), i <p . Thus as functions of q,
iqliP), where a = 1. We now show that
Ap(q) = A( qP) and hi( q) = A(CY
(independent of t) Wp is also a polynomial in A. This relies on two basic
facts. First, any symmetric polynomial in the Ai is A-modular and second,
any A-modular function is a rational function of A.
Theorem 4.7
Wp(x, A) is a polynomial of degree p + 1in x and A with integer coefficients.
The coefficients of x p f l and AP'l are both 1.
Proof. Consider q ( x , A) := II;=, ( y - xi) , where Ai := 16/Ai and
y := 16lx. This is a convenient form to work with. Observe that Wp and WP
are connected by
n
P
A) .
16p*1~p(x,
A)= [ x ~ + l
i-0
Now by Theorem 4.6 any symmetric polynomial in A,', A;', . . . , h i 1 is

left invariant by any element of the A-group. (See Exercise 6 of Section
11.2.) It follows that any such polynomial is A-modular, and by Exercise 4b)
of the last section is finite valued in the fundamental set except possibly at
t:= im. In particular if si is the coefficient of y' in Wp (viewed as a
polynomial in y), then si is A-modular. From Corollary 3.1,
xi
A(~)=
n=l
q,n-l
)8
and we have integers a, such that

-
"
A:= - = - +
anqn
A
q n-o
ii=
1
aq
A
+ C ananiqnip
n=O
= qP
-l+ Cmand"'
n=O
i<p
122
where a is a primitive pth root of unity. It can now be established that there
are integers ci so that
It is a consequence of the symmetry that the nonintegral powers of q vanish.

(See Exercise 2.) Next there is a polynomial P of degree at most p + 1 with
integer coefficients so that
and si- P( h) has zero principal part. This is easily proved. First remove the
term by considering
q-(p+l)
and then proceed inductively. Observe from (4.4.4) and (4.3.5) that the
only candidate for a pole of si- ~ ( h is) q = 0 but P has been chosen so that
si- ~ ( h )is finite at q = 0. Thus we see that si- ~ ( h is
) a bounded
A-modular function and is hence, by Theorem 4.5, constant. Since
we have that si is a polynomial of degree at most p + 1 in h := 16lA with

integer coefficients. Hence w,(x, A) is a polynomial of degree p + 1 in 161x
and 16lA with integer coefficients. We can prove directly (see Exercise 3)
that
P
Exercise 7 shows that 16,+' divides every coefficient of $(x,

viewed as a polynomial in l l x and 1/A. Thus with (4.4.3),
A) when
is of the required form. 0

Analogously we have a modular equation for j := 1728J. (See Exercise
4.)
Theorem 4.8
The modular equation for j of order p
P
(4.4.6)
FP(x,j):=(-j)
ji:=j0Ai
i=o
is a polynomial with integer coefficients of degree p

coefficients of xP+' and jP+'are both 1.
+ 1 in x
and j. The
The modular equation (4.4.2) is irreducible over C(A) (the rational

functions in A) since any root can be transformed into any other by an
appropriate transformation. (See Exercise 5.) Likewise (4.4.6) is irreducible
over the rational functions in j. The Galois group of F, over Q ( ') is a
p!
group of order p(p2 - I), which is nonsolvable for p r 5. (See Exerc~se10.)
Here Qp is Q adjoin the pth roots of unity. For nonprime n, modular
equations can be constructed from the modular equations corresponding to
the prime factors of n. (See Exercise 8.)
Further properties of modular equations are chronicled in Lang [73],
Schoeneberg [76], and particularly in Weber [08]. We have chosen a path of
limited generality focusing on the modular equations for A and j. It should
however be fairly clear that analogous equations hold for other modular
functions.
1. Prove Lemma 4.1.
Hint: For (a) prove that M is equivalent to an upper triangular matrix
m o d r . Then write out the system of equations required for two
triangular matrices to be equivalent. For part (c) show that BiC is
equivalent to a triangular matrix mod A. Note that B,C has determinant p, and hence the diagonal entries of this equivalent triangular
matrix are + I , +p.
2. a)
Suppose that
l
m
f(q) :=
cnqn
cn real.
n=-h
Show, for a := e2"i1Pand p prime, that

D
[f(anqnlp)]m
m integer
n=l
has no fractional powers of q in its expansion. (See Exercise 4 of

Section 6.2.)
b) Prove that (4.4.5) holds by applying Newton's formulae to express
siin terms of powers of the roots. (See Exercise 6 of Section 11.2.)
Higher. Order Transformations
Show directly from (4.4.4) that
Prove Theorem 4.8 by modifying the proof of Theorem 4.7. Use

Exercise 6b) of section 4.3. The proof is somewhat easier since we can
consider IlT,, (x - ji) directly.
Show that the modular equations (4.4.2) and (4.4.6) are irreducible in
x, over C(A) and C( j ) , respectively, by elaborating on the comments
following Theorem 4.8. This requires a minimal knowledge of Galois
theory. Note that the transformations of Lemma 4.1 act transitively on
the roots.
Show that
a) wP(x, 1)= (x - 1)'"
b) w,(x, 0) = xP+'.
Hint:Consider the orbits of 0 and im under the A-group.

a) Show that
and
and if ciPjis the coefficient of x h i in Wp, then
Thus there is a fourfold symmetry in the coefficients of Wp.

Hint: W,( A( qP), A( q)) = 0 and Wp(A( q), A( qp)) = 0 and by Exercise 5, Wp(x, A) is irreducible in x over C(A). Also, Wp(x, A) and
Wp(A, x ) have a common root x = A(qP) and are of the same
degree in x. Hence,
where R is a rational function of A. Show that this implies that

R = 1. For the second symmetry use A(tl(t - 1)) = 1lA(t).
b) Consider wP(x, A), as in the proof of Theorem 4.7, as a polynomial in l l x and l l h . Let
be the coefficient of x-'A-'. Show
directly that
and, by part a), that
c)
Show that the coefficient of xiAi in Wp(x, A) is divisible by

161~+l-fi+j)l
In general, for n not necessarily prime, the modular equation (4.4.2)

has degree $(n) := n n P r n ( +
l 1Jp). Let n := p, - - .p, be a product of
distinct primes. Prove that there exists a two-variable polynomial
Wn(x, y) of degree at most +(n) that satisfies
Hint: Let n = p, p,. Since
we deduce that A(qPIPZ)is algebraic over Q(A(q)) and is of degree

+(plp2). Now proceed inductively.
Show that the coefficients of W,(x, A) are
Note that by Exercises 6 and 7 is suffices to determine c,,, and c,,,.

The c,,, coefficient is always -16~-'. (See Exercise 6 of the next
section.)
(On the Galois group of F,) If A E r then A induces an automorphism
on the jithat fixes j. Hence A is an element of the splitting field for Fp
over Q( j ) . Any two elements S and T of I? induce the same automorphism exactly when
for some a. Equivalently MST-'M E r for all M E Tp. One can show
that I?, equivalenced as above, is the Galois group for Fp over Q p ( j )

and that it is a group of order p(p2 - 1) which is nonsolvable for p 2 5.
(See Klein and Fricke [1892], Schoeneberg [76], and Exercise 8 of
Section 4.5.)
4.5
THE MODULAR EQUATION IN u
- v FORM
The actual calculation of the modular relation for h is most readily carried
out in the associated variables u := k1I4:= A 'I8 and v := 1'I4 := x 118. Analogously to Theorem 4.7, though with a few substantial additional details (see
Exercises 1 and 2), we have the following theorem.
Theorem 4.9
If p is an odd prime, then the modular equation in u-v form is given by
where
and a is a primitive pth root of unity. This modular equation is a polynomial

in u and v of degree p -f- 1 (independent of t ) with integer coefficients.
For p = rl (mod 8),
The coefficients of up+' and up+ are 1.

For p = 2 3 (mod 8),
a, (v, 1) = (v f
- 1)
flp(v, U)= -.np(-u, v) = Op(-v, -u) .
The coefficient of up+' is -1 and the coefficient of up+' is 1.
The most striking property of the modular equation in u and v is the
"octicity." Because ui(qP) is of the form aiq118f(a8iq),where f is analytic in
q, only every eighth coefficient of the modular equation is nonzero. (See
Exercise 3.) We illustrate with some examples which give the nonzero
coefficients and the column sums.
4.5 The Modular Equation in u - v Form

u4
u3
uZ
4.5
The Modular Equation in u - v Form
132
The exhibited portion of a,, is sufficient to easily calculate the remainder

of the coefficients because of the symmetries. For p = +1 mod 8 the table is
symmetric through both diagonals. Forp = +3 mod 8 the reflections through
the diagonals change the sign of the entries according to ci, = (-l)i+lcj,i and
Ci . = (-1)Ic ,+l-j,p+l-i, where ciYjis the coefficient of viu' in a. (See
.I
Exercise 4.)
The numerical calculation of these modular equations is fairly straightforward. From the q expansions for the 8 functions we can compute
u,, u,, . . . , up for a variety of q values (for p := 23 one must use three
vaIues of q). We then use (4.5.1) to calculate the coefficient of vi at these
values. However, we know the form of the coefficient (for p := 23 and
i :=23 one has, for example, that the coefficient of vi is of the form
au2, + bu15 + cu7), and we can easily calculate u at the same q values we
'
4.5 The Modular Equation in u
- v Form
133
to calculate the coefficient of vi. This leads to a system of linear

solve. We know that the system has an integral solution. We
know what the column sums are in each case. One can use this
rmation to reduce the size of the system by 1 or, perhaps more
asonably, as a check on the solution.
The main limitation is that the size of the entries grows exponentially
th modular equation will have entries of size roughly 2', so for
e examples one must work to a high degree of precision. Note that the
of the linear system only grows as pl8.
fairly complete account of modular equations up to 1928 is given in

na [28]. This includes equations of degree 103, 107, 127, 167, 191, and
modular equations up to degree 20 are presented in Cayley
as we have presented them. We easily computed a,,by the method
d in the section. The others were originally calculated by Sohnke
, whose method roughly parallels the one we have described, except,
he computed an expansion for u(q)lqUs and computed sufficient
cients of urn,to calculate the elementary symmetric functions directly.
en instead of solving a linear system, he compares coefficients. As Cayley
ut, "The process is a laborious one (although less so than
eforehand have been imagined)."
rly simple form of the modular equation for p = 23, due to
= 1.
(k1)'I4 + (krlr)lI4+ 22f3(klk11r)1f12
For many theoretical purposes modular equations for j are preferable.

However, for calculations the modular equation for u is usually simpler. The
extent of the numerical simplification is quite remarkable. Du Val 1731
exhibits low-order modular equations for the I?-modular function I : = J l
( J - 1). The cubic modular equation for I tabulates as a 5 X 5 matrix where
all but one of the 25 entries are either 15- or 16-digit integers. Modular
equations for j up to order 11 have been calculated. (See Kaltofen and Yui
[84].) For the 11th-order modular equation the coefficients are enormous.
For example, the coefficient of j6 is
Various of the h and j modular equations are presented in Greenhill [1892].

In particular, clean forms for p = 29, 31, 47, and 71 are given for W,.
134
Exercises 1 and 2 outline the proof of Theorem 4.9. Some of the details
are rather complicated. The flavour should come through.
1. a)
Observe using (3.2.3) that
b) Analyze the action of A-group on u. Show that

U O S ~ ) = U
and
uoTA=u
Observe that u o S, = p u where p is an eight root of unity.

Analyze the action that S, induces on the functions A,, . . . , A, and
induces
on the functions u,, . . . , up. What is the permutation
on u,, . . . , up?
d) Analyze the permutation that T , induces on the functions
A,, . . . , A, and on the functions u,, . . . , up. Identify Ap(TA)and
A,(T
in particular.
c)
s?)
2.
Prove Theorem 4.9.

Hint:Use Theorem 4.7 and Exercise 1. Consider how wP(v8,us) splits
over Q(u). (Note that u is invariant with respect to group generated by
Sf' and T,.)
3.
Prove the "octicity" of the u-v modular equation. That is, for p = + 1
mod 8 the nonzero entries of the table associated with a, lie only on the
main diagonal and every eighth sub and super diagonal, while for
p = 2 3 mod 8 the nonzero entries are only in every second entry of
every fourth diagonal.
4.
a) Prove the symmetries (or antisymmetries) of the u-v modular

equation with respect to reflection through both diagonals.
b) Evaluate the row sums explicitly from Theorem 4.9.
c) Observe that with the aid of the "octicity" one can read off the
modular equations of degrees 3, 5, and 7. Verify the modular
equations of degrees 11 and 13. This requires either calculating the
ui at a single value of q, or using Exercise 6.
5. Show that
a, can be written as
4.5 The Modular Equation in u
- v Form
From our analysis we know that
a)
(p2-1)18
(p-1)12
Show that c,,, = -(-I)

2
Hint: flp(uo,U)= 0 and u = f i q 1 I 8
~ ( u i ) Thus
.
+ 0(u2) while
uo = fi/Zq118p
+
and
C1,l
- -(-1)(~~-1)18lim up .
u-0
Uo
b) Show that if 0 < u < u < 1 and Rp(u, u) = 0, then
c)
> 0 ,has
Hence if wP(y:, y:+,)=0 with 1>y , > ~ ~ + ~ one
In 1858, Hermite and Kronecker separately gave solutions of a general

quintic using quintic modular equations. Hermite's method is outlined below.
a) Let
where i + j is chosen mod 5. Then
are the five roots of the quintic
b) The quintic modular equation (4.1.21)
136
can equivalently be transformed into

512 3
~ ~ -4 23 45 ~ ( 1 - ~6 ~
u) ( ~i - ~
u 8-) 22( i5+ u 8 ) = ~ .
The details are formidable.

c) Any quintic can be algebraically reduced (via solution of a quartic
equation) to the Bring form
and hence a) and c) provide a solution of the quintic in terms of the

roots of the modular equation. This requires solving
and leads to a quartic equation in u4.

The amount of calculation required above is prohibitive. However, that some combination of the ui solves a quintic is not overly
surprising since the Galois group of the quintic modular equation
(4.1.21) is S5. The reduction to Bring form in c) is effected via the
Tschirnhaus substitution.
,
8.
(On the Galois groups for W, and W7 over Q ( h ) )

a) For p = 5, S, and T, induce the following permutations of the roots
(i = hi)
S,:
(0,1,2,3,4,5)-,(1,2,3,4,0,5)
TA: (0,1, 27 37 4, 5)+(0, 5, 3 , 1 , 2 , 4 ) .
b) For p = 7
s,:
m i , 2,3,4,5,6,7)-, ( 1 , 2 , 3 , 4 , 5 , 6 , 0 , 7 )
T,: ( 0 , 1 , 2 , 3 , 4 , 5 , 6 , 7 ) - , ( 0 , 3 , 1 , 4 , 6 , 7 , 5 , 2 ) .
Show, using a), that the Galois group of W, contains A , and is not
solvable.
d) Show, using b), that the Galois group for W7 is not solvable.
In both cases these permutations actually generate the Galois
group. (See Exercise 10 of 4.4, and Exercise 1 of 4.5.)
c)
4.6
THE MULTIPLIER
As we saw in Section 4.4, the pth-order transformation can be considered as

determined by k : = k(q) and 1:= k(q'lP). We define M,, the multiplier of
order p, by
4.6
The Multiplier
Mp(l, k ) :=
'fa4
K(k)
B ; ( ~ " ~- )
-
KO
in the future will denote K(1) := L.
=k(q)
and 1 := k(q"P), then wP(l2,k 2 ) = 0 and
11t2 ,dk
p"p=---k k t 2 dl
v ( 1 - u s ) du
~ ( 1 u- s ) dv
'
particular, M, is an algebraic function of k and 1.
Proof.
Since rrK1lK= -log q [by Theorem 2.3(b)], we have

1
Tr
using (2.3.10). Similarly,

1 7 7
Lt)=-Zm.
dl ( L
K'IK [again by Theorem 2.3(b)], we can write the 1, k

n dividing ( 4 . 6 . 3 ~ )by (4.6.33). We now verify the u-v
n by v ( 1 - u s )l [ u ( l - u s ) ] (duldv) as a function of u and
, u). In this form we see that M ; is rational. In fact

1. (See also Exercise 1.)
s use (4.6.2) to compute M2 and M,. When p : = 2 , we have
( 1 + k ) and k = ( 1 - 1') / ( I + l ' ) , by equation (2.1.15). Thus
dk~ - 1r2 G(l+
kl2 2
2 ~ : = 1- ~ '1-k
( 1+ k)2
k k t 2 dl - kk'2
1
l+lt
M2(l, k ) = -= l+k
2
ich corresponds with Theorem 2.6.
138
+
= 1, by equation (4.2.6). This proWhen p := 3, we have
duces, on differentiating implicitly and using (4.6.2),
Let a 4:= k3/1. Then 13/k= [(2 + a ) / ( I + 2a)14 and
(See Exercise 2 of Section 4.1.) But a = u3/v so that
because
For p := 5 or 7, similar, but more elaborate, calculation produces
(4.6.7)
M7 =
v(1- uu)[l - uv
7
v -u
+ ( u v ) ~ =] -
U-U
7 u ( l - uu)[l
- uv
+ (uv)']
'
(See Cayley [1895].) Many other multipliers have been calculated and can
be found in Ramanujan's collected works, Cayley [1874], Tannery and Molk
[1893], in Weber [08] and elsewhere. The main technique for larger p is via
manipulation of theta series. Thus one has Ramanujan's form of the
multiplier for 13:
We list also
(4.6.9)
17Ml7(l, k) =
(f)
(k) + (g)
112
112
( )""
11'
((:)lI4
-2 7 l+ kk
112
)"'I
+ ((f
7
1
M17(k7 '1
'
4.6 The Multiplier
139
conclude this section by touching on the matter of singular values, k p 7

ch for us are defined to be the solutions in ( 0 7 1 )of w p ( k f Zk72 ) = 0.
se are often called singular moduli for the function A. Corresponding
es for J are discussed in Exercise 6.
en, since K f ( k ) l K ( k )is isotone, Theorem 2.3(b) shows that this is the
ue solution to
the notation of equation (3.2.1), k, = h * ( p ) and k i = A*(l Ip), so that

= k(e-*)
and 1, := k; = k(e-"I*).
Sophisticated number-theoretic
hniques are available for computing k p for large p , without knowledge of
. This is discussed briefly in Exercise 5. For small p one can solve directly
=-
2k,11 = 1
.\/z
l,=vZ
' Fz
- 3 l I 4 ) ( v 3- 1 )
I =
fi(v3i1)
1,
m+m2kSl,
I,
2k-1-= -
= fi- 2
'
(v7+ 3 l I 4 ) ( v 3- 1)
2
A more comprehensive list is given in the next chapter. A profusion of

modular equations of degrees 3, 5, and 7 are given in Chapter 19 of
amanujan's Second ~ o t e b o o k .
1. From the results of Section 1.5 (in particular Theorem 1.5 and Exercise
1) we know that G ( k ):= k 1 I 2 k ' ~ ( ksatisfies
)
G ( k )=
q%
G(1)
140
where k and 1 are solutions of the pth-order modular equation W,. Use
this to show that
where c is a constant. This provides an easy alternate derivation of

Theorem 4.10 up to the evaluation of the constant c.
2.
a) Verify the computation of M, in (4.6.5).

b) Compute M, in (4.6.6).
Cayley [I8741 discusses algebraic methods for computing Wp and M, at

length. These seem only to be entirely reasonable for 2, 3, 5 , 7, and in part,
11. The discussion therein also illuminates the rational nature of M,(u, v).
3. a) Show that pMp(l, k)MP(k1,1') = 1.
b) Use Theorem 4.10 to show that for all u and v,
Hint: Consider the similar symmetry of (n,.

4.
a) Cayley observes that given any polynomial identity F(u, v) = 0

which satisfies F(u, v) = F(-u, -v), one can produce a similar
identity G(u', v2) = 0, with G of the same degree. One uses
b) Use this techni ue to develop modular equations and multipliers in

vZn( n : = l , 2 , 4 ) for p : = 2 , 3 , 5 , 7 . (See Cayley
terms of the
[I8741.)
3,
5. Verify the singular values in (4.6.10). In each odd case one verifies
t, = 2k,kA first and uses
The invariants of (3.2.9) to (3.2.13) lie at the heart of calculating

singular values. Armed with these and either Ramanujan's insight or
some knowledge of group theory, singular values can be calculated in
profusion. Watson, in a long series of papers commencing with Watson
[32], has recreated what he believes to be Ramanujan's procedure,
while Weber [08] explains the classical theory and lists many examples.
Zucker [77] indicates an attractive way of calculating many large
singular values such as Ramanujan's celebrated
4.6 The Multiplier
141
(4.6.12) kzl0= (fi

- 1)'(2 - fi)(fi
-6)'(8 -3 f i )
X
(VIO - 3)'(4
- dE)'(fl-
a ) ( 6-
m).
Zucker's methods are described in Section 9.2.

In this exercise we sketch the relationship between binary quadratic
forms and singular invariants or values for Fp [solutions of Fp( j, j) = 01.
Let
where /SI = ad - bc = p > 0 and a, b, c, d are integral, and assume

with no loss of generality that c > 0, b # 0. We say t with im(t) > 0
satisfies complex multiplication by p if ( p t , p ) = S(t, 1) so that
t = (at + b) l(ct + d). In particular, t is the solution with im(t) > 0 to
an integral equation AX' + Bx + C = 0 with - D := B' - 4AC < 0.
[This corresponds to asking for which p one can solve
p(pz, wl, w2) = R(p(z, wl, w2)) for some rational function R. Here
P(Z):= p(z, wl, w2) is the Weierstrass p function of Section 1.7. Or,
in other words, for which p does the lattice L generated by w, and
w2 contain the lattice p L ? Hence, the multiplication.]
Suppose that
t=- a t + b
'ct + d
(4.6.13)
and
a*t + b*
c*t + d*
ad- b c = p , a*d*-b*c*=l.
t* =
Then j(t*) = j(t). Also t* solves a quadratic with the same discriminant as that for t, and t* possesses a multiplication by p if and
only if t does. Now (4.4.6) implies Fp(x, j) = 0 is solved by j(t),
because S = A, m o d r for some i and hence j(t) = ji(t). Thus
F,(j, j) = 0 is solved by j(t) exactly when t possesses complex
multiplication. Note that F,(j, j) will have lower degree than
Fp(x7j)Since j is one to one on the fundamental region, j(t) = j(t*) if and
only if the associated primitive binary quadratic forms are properly
equivalent. A binary form ax2 + bxy + cy2 = 0 is primitive if
gcd(a, b, c) = 1. Two forms are properly equivalent if there is an
integral linear transformation of determinant 1 converting the one
into the other. (See Dickson [71, vol. 31.)
Thus the study of the degree of F,(j, j) becomes the study of
h(- D): the class number of primitive forms of negative discriminant -D. More of this is sketched in Hardy [40, chap. 101, and in
Tannery and Molk [1893].
142
7. For fixed m and p let yo := k, be the mth singular value and let y, be
the sequence of solutions of w,(~:, y:+,) = 0 as defined in Exercise 6 of
Section 4.5. Use Theorem 2.3 and (3.2.3) to show that
When
This provides an algorithm of order p for computing e-&.
p := 2, this reduces to the algorithm in Exercise 3 of Section 2.5.
By the same process as in Section 4.5 one can produce polynomials in
x := Mil(l, k) and k. Cayley [1874], following Joubert, gives
and
He gives a similar expression when p := 11.

8.
a) Verify (4.6.14) and observe that 3~,(1/*, k,) = m

f
l
.
b) Explicitly solve (4.6.14) to produce M i 1 as a function of k.
c) Use (4.6.15) to determine k,.
4.7
CUBZC MODULAR ZDENTZTZES
Ramanujan in his notebooks gives the following remarkable identity:
[See (24.28) and (24.29) of Chapter 18 of his second notebook in Berndt

[Prl .I
An equivalent form and some variants of this formula are established in
the next theorem.
Theorem 4.11
With
4.7 Cubic Modular Identities
one has
(c)
82( 4')
[3,0-1]3=-4(-;;7;T)
kr3/k6
114
Proof. We establish only (a). Then ( b ) and ( c ) follow (Exercise 2). It is

most convenient to use the quintuple-product extension of Jacobi's triple
product, which is developed in Section 9.4. Equation (9.4.3)with j := 6 and
k := 1 may be remanipulated to show that
Hence, with (3.1.6),
and Euler's identity (3.1.4) reduces this to
Now (3.2.9iv) shows
,
m
( 1 + q2n+1)3
= flq118(kk1)-114
n=O
and
while (3.1.4), (3.1.7), and (3.1.8) combine to yield
These identities result in
144
This establishes the first equality for part (a). The second is most easily
seen by using equation (4.1.15), with y := k and k := I, to write k3/y =
[(2 + a ) l ( 2 a + 1)14 k'3/r1= [ ( I - a ) / ( 2 a + l ) I 4 and using (4.1.19) or
(4.6.5) to write 0;(q3)10:(q)= 1 / ( 2 a + I ) ~ Then
.
both sides of the second
equality become 4(2 + a ) ( l - a ) / ( 2 a + I ) ~ .
These identities have many remarkable consequences some of which we
leave as exercises. Two, however, are worthy of more explicit analysis.
AN ITERATION FOR THE CUBIC MULTIPLIER. In decreasing form, as above, we
consider the multiplier M, := K,, ,IKn where K, := ~ ( q ~ "Then
) . part a)
of Theorem 4.11 shows that, with m , := 3M,, we have
where
/0:(q3"),so that
Alternatively, part ( b ) shows that if we set rn := 38:(q3n+1)
m , = (r, + 3 ) / ( r , - 1) (see Exercise 8 ) . We have
where
This latter is more suitable when we begin with an even singular value.
We have written (4.7.4) and (4.7.5) in a form consistent with Ramanujan7s
invariants Gn and g,. [See (3.2.13) and Exercise 5 , Section 3.2.1 In these
terms, the iterations are initialized by
and
We also write M(n) := m(n) 13.
q := <-&
[,
CUBIC RECURSIONS FOR
G,
AND
gn. We have
$>i and, using the theta inversion formula in part (a),
'
Thus
(4.7.9)
Similarly
9 = ( 1 + 2 f i GG) n( 1 + 2 f i $ )
Ggn
146
Correspondingly
From these two formulae, and other singular values given by Ramanujan,
one can give explicit equations for some very large invariants (G;,,, , G$,, ,
3
g,,,,,
for example). We illustrate with g,,,,. Ramanujan gives
Then manipulation of (4.7.13) yields
(See also Exercise 13.) Analogous results for quintic and septic multipliers
are treated in Section 9.5.
The quintuple-product identity was certainly known in essence to Ramanujan so that our derivation of Theorem 4.11 is in all likelihood similar to that
which he had in mind. The proof we give in Section 9.4 is self-contained and
can be read with ease now. Biagioli [Pr] has given a modular function proof
of (4.7.1).
1. a)
Show that (4.7.1) is equivalent to
(either by modular considerations or by direct inversion of the

underlying quartic polynomial).
b) Show that the AGM iterations in theta form can be written as
4.7
Cubic Modular Identities
(Note that in this case the replacement of q by - q does not give a

formula in 0,.)
a) Show that Theorem 4.11, parts (b) and (c) are equivalent to part
(a)
b) Derive part (c) directly from (9.4.3).

Use (4.7.9) and (4.7.10) to verify that
i) G ; = ( f i + l ) / f i
ii) & = V 3 + f i
iii) G;, = 21'4/(2113- 1).
a) Use Exercise 3 and (4.7.12) and (4.7.13) to verify that
b) Calculate GZ4,and g,,.

Show that, in the notation of Proposition 3.1,
Show that
Use Schlafli's equation (Exercise 6 of Section 4.1) to establish that
ii)
2
f i + l
g 1 , = 2
fi+1
iii) G,, = 2
148
8. In the notation of (4.7.7) and (4.7.8) show using (4.1.15) that
and
Also (4.7.9) gives
9. a) Given that
verify that
3 - 4 7 . vi+d 312
GW6=(-)(
7/3
and
b) Use the modular equations of degrees 3 and 7 to verify the value

of G,,.
10. In each case, given the first invariant verify the following ones.
and
and
g~: = (f-i
149
g;$3=(V26+5)(fl-3)3
2
3)(fi
- 2),
g;0613
=(
and
f i - 3 ) ( f i + 2) and
Verify that m ( 4 ) = f i and r(5) = 6+ fi.

Hence verify that m( 5 ) = 3 f i 2 f i - 6- 3 and that r( $ ) =
l,lqEx5j.
m.
and r(4) = q 3 + fi +
Verify that m(1) =
Hence
Verify that r ( 2 ) = f i + f i and r n ( $ ) = f i - f i + l .
m(&)=3(36+7-4fi-5fi.
Verify that m($) = (3 - 6
+ 1) and m(&) = ( 6 G + 9) ( 4 6+6fi).
Show that (4.7.9) can be written as
where x := G,3 and y := G,,.3 Hence compute G3 and G,.

Find the parallel expression for g;, g:N. [Compare (4.1.27).]
13. One can explicitly solve (4.7.9) and (4.7.10) to obtain the following
formulae.
a) Show that G:,
and Gal, are the two solutions to
X ~ - ~ ~ G , [ G > + ~ G :G+> + ~ ] x +
G~,[G>+I+~G:+ G ~ + I ]
=o.
b) Show that ,g:,
and
x2 - fig,[&
-&,,
are the two solutions to
+ 1y~
-
- g',[& - 1]+-/I
=o.
c)
Given that G2, = ( f i + 1) 12, show that
G2,, = (2 + f i \ / j ) " 3 ( q ) [ v G m
+1 5 y
and
G25 19 =
(2 +1?)1/3=
m
i([f)-
151141.
150
d) Given that g& = (5 + m ) / 2 , show that
and
14. a) Entry 23 in Chapter 18 of Ramanujan's second notebook (Berndt

[Pr]) gives
i)
a2
=(
(z)
(=)
m + x)"'
x~
ii)
exp
n2.rry
n=-a
n=-a
= (/,I
x +Y
cos
n=-m
x +Y
e-"'"'
cos (n2.rry)
e-n2mx
x +Y
,,==-a
sin (n2.rry)
--n 2.rrx
n2.rry
exp x + y sin (
x + Y n)
(-)
- x)"'
n=-m
- ( p ' T j 7+ x)lI2
e-"'"
cos (n2.rry)
e-n2?rx
sin (n2.rry).
n=-m
Use Exercise 5 of Section 2.2 to prove these when x and y are

complex with re(x + iy) > 0.
b) Deduce that if re(s) > 0,
2
m
n=-m
e-n2?rs
cos (n27Ts1)=
c) Use ai) to deduce that

6,(e-") =
d) Use b) with s := f i / 3 to deduce that
t$(e-'")
e)
f)
Use d) to conclude that
i) Use b) with s := V3/2 to obtain

k12= ( f i - .\/2)'(.\/2 - 1)'
and
gi;
=f
i ( f i+I ) ~ .
ii) Use b) with s := f i / 4 to obtain

kL.=(4+fi)(2+\/j)2
k15
iii) Use b) with s := */4
g)
and
~::=4(7+3fi).
to obtain
Use c) to show that
As in Ewe11 [86] show that
and so develop a formula for r3(n).

Hint: Use the quintuple-product of Section 9.4 and mimic the
derivation of equation (9.1.14).
Evaluate
Chapter Five
Modular Equations and

Algebraic Approximations
to rr
Abstract. In this chapter we study the algebraic relationships between elliptic
integrals of the first and second kind. This study is applied. to produce nthorder iteratons for T, rapid series for T-', and assorted other algebraic
approximations to T.
5.1
SINGULAR VALUES OF THE SECOND KIND
In Section 4.6 singular values were introduced. We will call A* the singular
value function (of the first kind) where A*(r) := k(e-dr) as in Section 3.2.
We introduce the singular value function (of the second kind) a which we
define for positive r by
Since A*(r) tends to 0 as r tends to m, we have a(r) converging to T-' as r

increases. Indeed, as we shall see, the convergence is exponential which
allows us to use a(r) to approximate l / a effectively.
Using Legendre's identity and the fact that K1(A*(r))= <r( A*(r)), we
have
5.1 Singular Values of the Second Kind
and
153
k := h*(r)
If in (5.1.2)we use the differential equation for K , equation (1.3.13), we

may also write
Now since h * ( l l r ) = h*'(r), we discover that
In particular a ( 1 ) = 3. We may rearrange (5.1.2) and (5.1.3) as follows:
and
which may be viewed as one-sided forms of Legendre's identity. In the next

sections we will give these identities substance by showing that a ( r ) is
algebraic for rational r and by computing many values. Another useful
equivalent form is
where
Theorem 5.1
The function a is montonically decreasing for r 2 1.
Proof. Since A* is decreasing, it suffices to show for k < 1l f i that

f := ( E ' K - T / ~ ) / K is
' an increasing function of k. This we establish by
computing f and using the differential equations for E' and K . (See
Modular Equations and Algebraic Approximations to m
154
Exercises 2 and 3 of Section 1.3.) We deduce from Legendre's identity that

f(k) =
v / 2 ( E I K - 1 ) + k Z ~ ( K-' E ' ) + ( k l ) ' ~ ' ( K

-E)
k(k')Z~Z
~ E ) ( E - v l ( 2 K ) ) E(K' - E ' ) - E(K - E ) ] + ( k ' ) ' ( ~- E ) ( E 1- ~ l ( 2 K ) )

-k'[(k(k'y~'
Since, for k < 1ll/Z,
we finish by observing that, for k < 1 / f i ,
and so substitution into the last equality completes the proof. CI

We next provide a theta function expression for a. We combine (2.1.13)
and (2.3.17) with (5.1.2) to write
where q := e -4
. Expanding this gives
(5.1.11)
O < a(r) - rr-' = 8(v7 - rr-l) e-dr

5 16.\rre-&
+ re-'^^)
r r l
and
One should compare (3.2.1).

The function a is implicit in Ramanujan's work. In the next section we

indicate how. Zucker [79] computes a(n) for n := 1'2, 3,4, 5'7, while 4 3 )
was known to Legendre. Formula (5.1.10) allows one to numerically
compute a very easily.
1. a) Show that
where q := e -d
5.2
Calculation of a
b) Use a) to reprove Exercise 4b) of Section 2.3:
By Exercise 4 of Section 1.6, A*(2) = fi- 1. Use Theorem 1.2 to show

that-4 2 ) =
- 1. Hence a ( $ )= $ .
Prove (5.1.11) and (5.1.12).
5.2
CALCULATION OF a
We begin with an appropriate generalization of the quadratic transformation

formula for E given in Theorem 1.2.
Let p > 0, k := k ( q ) , and 1 := k ( q l ' p ) be given. Then
1zr2 dMp(L k )
MP(W
dl
:[
(1)-PI.
Note: Herein (dMp(l,k ) / d l ) is the full derivative of Mp(l, k ) with respect to

Proof. We have Mp(l, k ) K ( l ) = K ( k ) . We differentiate both sides with
respect to k and use (4.6.2) to write
dK
dMP
( k ) = MPU7k ) ;51(1) + K(') 7
('7 k ) Next we use the differential equation for K to write
E(k) - kr2K(k)
dK
dl
- (1) =
E(1) - lr2K(1)
11r2
and we substitute these two identities into (5.2.2). We then have
dMp(17 k )
PM;(E,k ) [ E ( k )- k r 2 K ( k ) =
] Mp(l, k ) [ E ( l )- l r 2 ~ ( 1+) ]K(1)11r2
dl
On dividing each side by K(k)we have (5.2.1).
We now derive the key identity for a .
156
Theorem 5.2
For p and r > 0 let 1 := h*(r) and k := ~ * ( ~ Then
~ r ) .
=M
~ ~ k( ) za (, r ) - v'-r[M;'(l, k)12 - pk2 +
(5.2.3.)
pkk''
dM,(l, k ) ]
Mp(L k )
dk
Proof. We suppress variables in the multiplier when convenient. From

(5.1.2) and (5.2.1) we have
This gives
on using (5.1.2) again. Another application of (4.6.2) produces the desired

formula.
In particular, if we let r := l l p above, then 1 = k', and with (5.1.5) we
derive that
where k := h * ( p ) .
Let us observe that, since h * ( p ) and Mp are algebraic for rational p, this
shows that a ( p ) is algebraic for rational p.
EXAMPLE 5.1. When p := 2, M,(l, k ) = 11(1+ k ) and k = A1(2) = f i - 1.
Then dM,(k1, k ) ldk = - and 4 2 ) = f i ( f i - 1)' + (\/Z - 1)2 = f i - 1.
(Compare Exercise 2 of Section 1.) Similarly 4 3 ) = (fi
- 1 ) / 2 and also
a(7)= (fi
- 2 ) 12. (See Exercise 3.)
5.2
Calculation of a
157
Before continuing to evaluate a it is necessary to connect it to Ramanujan's multiplier (of the second kind).
where
Proposition 5.2
Let p , k , and 1 be as in Theorem 5.2. Then
Proof. We start with (3.2.15), which gives 77(q)lq(q"P)in the form
We differentiate logarithmically and obtain
Now we use (4.6.2) for dlldk and (2.3.10) for qdkldq and obtain
This gives (5.2.9). 0

It is convenient to introduce two additional cpantihes:
and
(5.2.12)
~ ( p:=) R, ( k t , k )
k:=e-M.
Modular Equations and Algebraic Approximations to
158
TI
In these terms we have:

Theorem 5.3
With r, p, k , and 1 as above, we have
(a)
(b) With k(e-4)
= h * ( p ) , we
have
Proof. We deduce (5.2.13) and (5.2.14) by comparing (5.2.9) and

(5.2.3). We obtain (5.2.15) on comparing (5.2.9) (with 1 = k ' ) to
(5.2.6).
Ramanujan has computed R, for p := 2, 3 , 4 , 5 , 7 , 11, 15, 17, 19, 23, 31,
35. From these, many values of a are obtainable. The following table is
taken from Ramanujan [14] with the entry for R, corrected.
The verification that R, has the given form is tedious but straightforwara
for small p. [See Exercise l c ).] For larger p we rely on Ramanujan.
5.2. For p := 7 , the modular equation in the form (k1)'I4+
shows that 2k7k; = i , k , := h*(7). Then k , = ( 3 - fi)
/4*,
(k I )
and so 4 7 ) = R,(k;, k,) = 3(1 -t- 2k,k;) = 2718, 4 7 ) = (fi
- 2 ) / 2 , and
6(7) = 2.
EXAMPLE
I
114
-1
We now make (5.2.14) explicit for p := 2, 3, 4.

Proposition 5.3
If I : = h*(r) and k : = h*(4r), then
(i)
1 - I'
k= 1 + 1'
and
(5.2.16)
(ii) a ( 4 r ) = ( 1 + k)'a(r)-24Tk.
Calculation of a
5.2
1 + kl+ k'l'
3(1+ kl + k'l')
15
[1+ (kl)lf4+ (kf1')1'4]4

- (1 + kl + k r l r )
17
[44(1+ k212+ kr21r2)

+ 168(kl+ k'l' - kk'll')
- 102(1- kl - k'1')(4kk'11')1'3- 192(4kk'll')2'3]112
19
1+
6[(1+kl + k'l') + a
31
3{3(1+kl + k'l') + 4 ( a +
+
- 4 ( k k ' Z ~ ' ) ' ~ ~(kl)lJ4
[l
+ (k'Z')1'4])
35
2 [ V m + r n - W ]
+ (4kk'll')-''6[1-
m- W ]
Proof. Since R2(l,k ) = I'
+k
rn m)
-m
13
and ~ i ' ( 1 k, ) = 2 / ( 1 + 1') = 1
3e2(1,k ) = ( 1 + k ) ' ( l + 1') - 2(1+ k 2 )+ ( 1
+ k ) ( l l+ k )
and
since I'
+ k = ( 1 + 11')1(1+ 1'). Thus
For example,
h*(4)=
(a- 1)'
and
a ( 4 ) = 2 ( f i - 1)'
+ k,
we
160
TI
Proposition 5.4
If 1 := A*(r) and k := h*(9r), then
JT-
(i> M,'(z, k ) = 1 + 4 -=: s(r)

and
(5.2.17)
(ii) a ( 9 r ) = s 2 ( r ) a ( r ) -
v7[s2(r)+ 2s(r) - 31
2
where m ( r ) := 3 /s(r) satisfies
(iii) m(9r) =
Proof.
([m2(r)- 1 1 ~ ' +~ 1)'

m(r)
( i ) and (iii) were established in Section 4.7. For ( i i ) write

3.5,(1, k ) = s2(r)(l+ 1')
+ s(r)(l + k l + k ' l ' ) - 3(1 + k 2 ) .
Substitution in terms of a , as in Section 4.6, and some easy algebra yield

s(r) = 2 a + 1 and
~ ~ (k 1) =, 2 a ( a + 2 ) = [s(r)- l][s(r)+ 31 12
as claimed. O
A convenient variant of (5.2.17) is.

(5.2.18)
S(9r) = s2(r)S(r)+ 2 f i s ( r )
where, as before, S(r) = fi- 2a(r).
5.3. If r := $ , then as in Section 4.6, m(r) = s(r) = f i . Then

(5.2.17) gives a ( 3 ) = 3 4 4 ) - 1. But (5.1.5) shows that a ( 3 ) = f i - 3 a ( $ ) .
Thus a ( $ )= ( f i + 1 ) / 6 and 4 3 ) = ( f i - 1 ) / 2 . Now we have m ( 3 ) =
(2lI3+ l ) * l f i SO that 4 3 ) = (4'13 - 1 ) f i and S(27) = 3(24'3 - 1). Thus
a(27) = 3 [ ( f i + 1 )12 - 2ll3].
EXAMPLE
Proposition 5.5
If 1 := h*(r) and k := h*(16r), then
( i ) fi=
1-C - F
1+i c 7
5.2
(ii)
Calculation of a
a(l6r) = (1 + ~ ) ~ u (-r 4) f i y ( l + y
+ y2)
Proof. This may be derived similarly but is easily deduced from Proposition 5.2 and the quartic iteration of Exercise 3 of Section 1.4. 0
EXAMPLE
5.4. For r := 4 we have a(4) = 2(.\/2 - 1) and A*(4) = (fi

- 1)'.
vPo = (1 - 2 5 1 8 ~/(I )+ 2 5 1 8 V i E T i )= ( V X T i - 2518)/

(m
+ 25'8) and u(64) = 8[2(V% - 1) - (2lI4 - I)'] I(+ PI8)'
Similarly, qh*(16) = (2lI4 - 1) /(2'l4 + 1)
which gives eight digits of
and a(16) = 4(.\/8 - 1) /(2lI4 + I),.
T~US
T-l.
Combining some of these calculations with (5.1.6) we have established
E..
;
2
(T)
K)
k := ---v3-1
2fi
and
with a similar identity whenever u(r) and h*(r) are known.

Computation of h*(r) will be discussed further in Section 9.2. In many of the

following numerical exercises the algebra is not entirely straightforward.
1. a) Verify that the quadratic case of (5.2.1) coincides with the formula
given in Theorem 1.2.
b) Verify (5.2.6).
c) Verify that R,, R,, and R , are as claimed.
2.
a) Show that in terms of Ramanujan's G, and g, of (3.2.13) one can

write
i)
ii)
A*(n) = $G,~[V-d
m
A*(n)= $ [ d m - - V m ]
iii) A*(n) = g
m - gi].
162
TI
b) Find similar identities for A*'(n).

c) Verify the singular values given in (4.6.10).
d) In each following case, given G, or g, verify k,. Then verify G, or
gn.
i) G:, = 2 - ' I 4 ( f i + I),
kl, = (3 - .1/5)(fi
- *)(2
-f
i )l(8fi)
ii) G% = -\/5 + 2, k,, = (G- 2)(3 - 2 . 5'14) /fi

iii) g : = f i + l ,
k, = ( f i + l ) ( G - f i - 1 )
= (2-*)(*-d)
iv) g!,=.1/5+2,
klo=(fi+2)(3fi-I.6-2)
= (fi
- 3 ) ( v 2 - 11,
V) g:8 = 5 + 2 f i , k18= (5 + 2 f i ) ( 7 f i - 5 - 2 f i ) .
3. Use (5.2.6) with p := 3 to compute a(3).
4. Use (5.2.15) to obtain the following values of a (or 6).
i) s(~)=Vii) ~ ( 1 1 =
) [2xZ- (x - 3) +- Y
i
i
4
3
Note: G;," = x - 3 where x - x = 2.
~13 ~
(a
-1
.
-6
1) 12.
Use (4.6.8) and (5.2.6) to compute 6 ( 1 3 ) = ( 7 + 3 f i ) ~ ; , 6 where
G;:=
(a3)/2.
iii) a(15) =
5.
6. Show that Rp(l, k) = R,(kr, 1'), that Rp-~(k,I) = -p-'~,(l, k), and
that ( ~ ( p - ' )= - a ( p ) lp.
7. Use (5.2.14) to show that
) (810 + 3 0 6 f i ) f i 7314
ii) 2A*(49)A*'(49) = (1863 + 7 0 4 ~ 7= (7lI4 -
677)
l2
2
and
a(49) = - 6 7 [ f i 73'4(33011+ 12477fl) - 21(9567 + 3616fl)I .
8. Use Proposition 5.3 to calculate
5.2 Calculation of cv
9. Use Proposition 5.4 to show that

i) a(18) = (21 - 6fi)g;,h*(18)
ii) S(9) = 3'I4(d6 - f i ) , as m(1) =
iii) S(81) = 9 f i 3'l4(V'3 + 1)(3 + a)a-l,
+ 11.'
a := [(2 +
where
10. Prove Proposition 5.5.
11. a) Use (5.1.4) to write

T
= *k'
k2
A G ( ~ k')
,
AG(1,
a(p) -flk2
AG(1, k')'
where k := h*(p) and the derivative is with respect to k.

b) Then (as in Section 2.5) show that
T = lim
iern
1
flk'kZpi + [ a ( p) - *k2]
qi
where x, := l / k l , q, := x,, y, := 6 ,
and p, := xt/(x,
+ 1) while
c) Show that convergence is quadratic.

d) When p := 1, this is Algorithm 2.1.
The next exercises develop results in J.M. Borwein [85] and Ramanujan
1141
12. a) Show that with P as in (5.2.8),
Hint:Let q := e-4
Hence show that
in (5.2.10) before differentiating.
Modular Equations and Algebraic Approximations to rr
b) Show that
Thus ~ ( e - + ) is quadratically computable.

c) Evaluate ~ ( e - " 3 ) .
13. Show that
) . J. M. Bonvein [85] these

where Ki := K(ki) and ki := A * ( ~ ~ 'In
identities are studied in detail.
14. Show that when p :=2, (5.2.21) is equivalent to Gauss's identity
(2.5.9).
15. a) Show that
This identity, which follows from Exercise 7c) of Section 3.7, is

entry 2 in Chapter 18 of Ramanujan's second notebook.
rn
b) Prove that
Here ul(n) is the sum of the divisors of n, as in Exercise 12 of

Section 3.7.
5.3 FURTHER FORMULAE FOR a
We commence by establishing a multiplication result for u.

Theorem 5.4
Let p, r>O be given. Let I : = h*(llpr), y := h*(plr), and k : = h*(pr).

Then
and
5.3 Further Fomulae for
Proof. Let s := pr and 4 , =
165
lu
Then y = k ( q , ) and I = k(4:"). We
:! so that with r := K ( y ) ,
For k t = 1 we have L = f l K and (5.3.3) becomes
This is (5.3.1). Then (5.3.2) follows from (5.3.1) and Exercise 6 of Section
5.2. (See Exercise 1.)
For p = r, (5.3.1) reduces to
2 ~( 1,/ f i ,k )
g ( p 2 )=
MP( 1/ f i 7
k)
k := k(e-"')
t p > O and k : = h*(2p). Then

S(2p) :=
kt"
l+k
+R p ( Z 7. k ) .
Proof. Observe that M i l ( y , k ) = 1 + k and R,(y, k ) = k

' = ( 1 - k ) l ( l + k ) . Then
4 2 ~=)( 1 + k ) [ ~ , ( l ,Y ) +
+ y',
while
$(k+ Yl)]
d since R, (1, y ) = R p(y ',k ) ,

.3.6)
1-k
~ ( 2 ~ ) = ( 1 + k ) ~ ~ ( ~ , k ) + $ ( l + k ' ) .
ow (5.3.5) follows from (5.2.15).
166
Modular Equations and Algebraic Approximations to rr
EXAMPLE 5.5. For p : = 3 , k=(2--fl)(fi--*)

yk + y 'k'. Now
which we rewrite as 2 k ( l + fig: + g:).

can write S(6) = [fi(fi + 1)' + 2*(fi
S(6) = g:h*(6)(2fi
while R3(y, k ) = l +
Similarly G k r 2= f i g y k , and we
+ l)]k. Thus
+ 2 f l + fi)
and
~~(6)=~~~*(6)(3-fi)=5fi+6V3-8fi-11.
Corollary 5.2
Let p > 0 and I := A*(1/3p), y := A*(p/3), and k := Ah(3p). Then
and
where
Proof. This follows from Theorem 5.4 and (4.7.9). (See Exercise
3.) 17
5.6. By piecing together various formulae many more values of a
can be obtained. We illustrate this as follows. Given 4 6 ) and h*(6) from
the previous example, we may use Proposition 5.3 to compute a ( $ ) . Then
the functional relation for a , equation (5.1.5), yields a($)= ( 5 a - 6 f l 8 f i + 11) 13 = 0.5138118 . . . . Similarly, given a(3) and h*(3) from Example 5.3, we can use Proposition 5.3 to find that a(:) = 66 + 47fi - 38fl2 7 6 = 0.5138837 . . . .
Indeed numerical computation of the maximum of a [using (5.1.10) and
newton,'^ method] shows that it occurs around 0.709 with a value of
approximately 0.514275. [Note that ( $ + :)I2 is very close to this point.] In
addition, one may observe graphically that a increases up to this value and
then decreases.
EXAMPLE
5.3 Further Formulae for
ff
rom the given formulae and appropriate singular values, many values
an be found in closed form. Some of the cleanest are given below.
Establish (5.3.2) of Theorem 5.4.
Use Corollary 5.1 and the given value of g, to compute a(n) or S(n).
i) g ! o = f i + 2 ,
,
ii) g 8 = ( + )
iii) g:, = (fi
+ I),
iv) gg8=(V%+5)/2,
a(10) = g~,h*(10)(3fi- 4)
a(18) = g;,h*(18)(21 - 6 a )
4 2 2 ) = g$h*(22)(33 - 1 7 f i )
a(58)=g~,h*(58)(99~-444).
(This requires having a tractable form of W,,, which we have not

given but which may be found in Greenhill [1892].)
'"
:
6
vi)
-6
g30
=6
+ 5\/i,
S(30) = g:oh*(30)[(56
+G
+ 38fi)
d o l
Establish Corollary 5.2.

Use the values of G,,, G,,,, and M(p/3) given in Exercises 8 and 9 of
Section 4.7 to prove that:
i)
vi)
and
4 2 1 ) = 3 ~ , ; 6 ( l l +6 f i + 21/;j +
CT(?)=
a)
G y g 3 J T (5m-13V%+49fi-19)
Modular Equations and Algebraic Approximations to s
168
vii)
u(93) = 6 G i ; ( 7 ) 3 ( 1 5 a
+l
3 m + 201fi
+ 217)
Hint:In each case express the right-hand-side of (5.3.7) or (5.3.8) as a

function of (yy'kk')"4 by using the cubic modular equation (4.1.16).
5. The values of a given in Exercise 2 are all expressed in the form
, a quadratic number a,. This is also true for
a ( p ) := g ; ~ * ( ~ ) a for
some other even p.
a) Since g:~*(p)
- 1/(2g;),
deduce that
and estimate the accuracy of the approximation.

b) Show that
="=
14 + 1 0 f i
33 - 172/2
and
"58
140 + 2 6 a
= 99.\/zis - 444
which give four and eight digits of r , respectively.

6. a) Weber gives G:, + G T =~(1 + m ) / 2 and gi4 + :g; = (3 +
m)/2.
Hence evaluate u(17) and 4 3 4 ) as cleanly as possible.
f
= :x solves x3 = 2x + 2 and f,
= :f i x
Similarly,
b)
solves x3 = 2x2 + (2x + l ) ( f i+ 1). Attempt to evaluate a(19) and
a(38).
'
(a)
(m)
7. Use (5.3.4) to evaluate S(25). (Compare Exercise 7 of Section 5.2.)

8. Given that G:, = fi 6, one can show that
9.
Again we have not given a tractable form of W,,.

Show that the perimeter of an ellipse with major axis a and eccentricity
k is given by
where. k := A*(r). In particular, if k := tan ( ~ 1 8 ) ~
5.4
Recursive Approximation to rr
(See Exercise 4 of Section 1.6.) If k := 1/fi

= sin (7~14)
Further such evaluations follow from Table 9.1, which gives K( h*(n))
in I' terms for 11 n I 16.
5.4 RECURSIVE APPROXIMATION TO rr
heorem 5.3 is easily recast as a pth-order iterative method to compute rr.

A General Iteration 5.1
et p # 1 be a positive integer. Let r > 0 and set

a,:=a(r)
and
k,:=h*(r).
For n in N compute kn+, by solving ~ , ( k ; , k: +,) = 0 and let

,
mn := ~ ; ' ( k ~kn+j,)
rn := Rp(kn7kn+l)
-.
Then, for rp2" 2 1,
Proof. This is a straightforward consequence of Theorem 5.3 and

(5.1.11) because a, := a(p2"r).
We may also write this as an identity for rr.
Theorem 5.5
Let p # 1 be a positive integer and let r > 0. Let m, and

Let a, := 1 and a,+, := mila,. Then
E,
be as in (5.4.1).
170
Proof.
?r
Rewrite (5.4.2) as
Then summation yields
and, as a, converges to T-' while a, converges to AG( 1, k;) [since

a, = K(km)lK(ko)],we obtain (5.4.4).
EXAMPLE
5.7.
a) When p := 2, the proof of Proposition 5.3 shows that E, = 2k,+, . In

terms of the AGM: a:+,&, =4an+,cn+,/2=ct12, and (5.4.4) gives a
family of identities extending (2.5.9), or Algorithm 2.2 (the case r := 1).
b) When p := 3, we similarly use Proposition 5.4 to deduce that a;+, E, =
(a, - an+,)(an+ 3an+,)12, where an+, = mi1an can be computed from
the cubic recursion in Proposition 5.4. (See also the cubic iteration
given below.)
We now specialize Iteration 5.1, changing notation as appropriate.
A Quadratic Iteration 5.2
Let r > 0. Let a, := a(r) and ko := h*(r). For n in N let

1- k:,
kn+,:= 1+ k:,
and
(5.4.6ii)
a,,+, := (1 + kn+,)'% - 2 n + 1 f i k n +.l
Then, for r22n2 1,
A Quartic Iteration 5.3
Let r>O. Let a o : = a ( r ) a n d y o : = m . For n in N let
and
5.4 Recursive Approximation to m
O<cw,-rr
-1
r16.4"fie
-drn.\Tr7r
Iteration 5.3 just peforms two steps of Iteration 5.2 as one, with some
mputational saving. The error bound given in each case is very accurate.
A Cubic Iteration 5.4
et r >O. Let cro := a(r) and m o := m(r) = v1+2 f i G ; , / G ; . For n in N let

[(m: - 1)lI3 + 11'
S,
mn
a,+, := s:an - 3 " f i
:= -
mn
s:+2s,-3
2
O<a, - r r - 1 s 1 6 . 3 n f i e - 3 n G " .
One can also provide a cubic iteration using k , instead of s,. This is
mewhat more inelegant and no easier to initialize. [See Exercise 3b).]
Large numbers of initializations for Iterations 5.1, 5.2, and 5.3 are
ailable in the previous sections. We collect some of the cleanest in Tables
information is given in Table 5.3. Recall that
ssary to verify these values lies in Section 4.7 and
e previous sections of this chapter.
When p := 7 we can write the iteration cleanly in terms of u, := k ,114 .
( 1 - unun+l)8= ( 1 - ~ f l ) ( ' - ' : + I )
'unun+l(l- unun+l)[l- unun+~

+(un~n+l)~l
8
un - Un%+l
rn :=3[1+(u,u,+,)~+ ( 1 - u , u ~ + , ) ~ ]
m ; ( l + ufl) + m,r, - 7 ( 1f ufl,,)
3
TABLE 5.2a. Values of G , ' ~ and a(N)for N odd

N
2A*(N)A*'(N) = G , ' ~
a m
TABLE 5.2b. Values of g,12 and a(N)for N even

N
~ A * ( N/ A) * ' ~ ( N=) g,12
f f
(N)
v2-1
$
TABLE 5.3. Selected cubic invariants
174
~r
We present a few illustrations of the behaviour of the algorithms. We

note that a, in the quartic algorithm is just a,, in the quadratic algorithm.
Digits Correct in Quadratic Algorithms
Digits Correct in Cubic Algorithms
Digits Correct in Septic Algorithms
The updates for a(25r) and a(49r) are studied again in Section 9.5, where
solvable versions of the quintic and septic iterations are given.
Additional information on these iterations can be found in various of our

papers. In particular, Bonvein and Bonvein [86]indicates the genesis of the
quadratic iterations and [84b] that of the cubic iterations.
1. Verify the claims in Example 5.7
2.
a) Prove that, in the notation of Theorem 5.5,

fr;
- (k) = 1-C p n a i + l ~ n k =: kO.

K
n=O
b) Observe that this extends Algorithm 1.2.

3. a) Verify Iterations 5.2, 5.3, and 5.4.
b) Use (4.1.24) to produce another form of the cubic iteration.
5.4 Recursive Approximation tb .n
175
a) Determine the initial values for the cubic iteration with r := N / 3

and N := 7, 11, 19, 31.
b) Compute S(45), S(63), 6(243), and S(54).
Derive the following version of the septic (seventh-order) iteration. Let
a, := a(r) and u, := ~ * ( r ) l ' ~
and
, generate un+, decreasingly from
(1 - u:)(l-
u:+~)= (1- unun+l)8.
Let
.- bn
Sn.-
tn :=
an
(1 - uf1+,)(49an- b,) + (1 - u:)(sn - l)b,

8
and
an+,:= S, an+ 7nv7(7- sn - tn)
Then for r7'"
2 1,
Derive the following version of the quintic (fifth-order) iteration (given

in Hughes [84]b. Let a, := a(r) and u, = ~ * ( r ) " ~and
, generate un+,
6
2 2
2
from un+l- un - 5(u,un+l) (un - %+I) -4unun+i[l- (%u,+1)~1=0.
Let
176
Modular Equations and Algebraic Approximations to s
Then
an+':=5cnan 5"+'VF(dn
8
+ u,+'
- c,u,)8
satisfies
O < an- a-' < 1 6 - 5 " - \ / ~ e - ~ " ~ "
7.
for r5'" 2 1.
Show that, in Iteration 5.1, convergence is indeed pth order.
Observe that for a variety of other values of p we have the information to

make Iteration 5.1 entirely explicit. For example, p := 17 is satisfactory since
we have R,,, and MI, is given in (4.6.9). Moreover, (4.6.9) also gives a
':;
= 40 + 1 0 m
form of the modular equation of order 17. Using G:; + G
and
we have an algorithm for a which gives more than
17" digits at step n.
8. Let p r 1, let I := h*(p) and k := h*(49p). Set zp := (GpG4,p)-'/
fi.
a) Show that
b) Ramanujan [14] gives
and
Verify that
where z,
G;:/fi,
and
5.5 Generalized Elliptic Integrals
ii) M;'(A*(~), h*(147)) =
14z3(1- 2,)
d R 3 -' f i t 2
where
c)
Compute that
Verify, from Table 5.2, that

i)
4 2 2 ) = (V?
+ 1)~(33- 17fi)(3fi
- 7fi)(10 - 3 f l )
and
Ramanujan gives the following form of the modular equation of degree

7:
where x := G,, and y := G,. This is Entry 19(ix) in Chapter 19 of the

Second Notebook (Berndt [Pr]). One should compare Exercise 6 of
Section 4.1 and Exercise 12 of Section 4.7.
a) Verify G,, G,,, and GI,,.
b) Show that
where x := g,,,
Verify
that g:,
c)
5.5
and y := g,.
+ g:
=fi
+ 1 and that
GENERALIZED ELLIPTIC INTEGRALS AND RATIONAL

AND ALGEBRAIC SERIES FOR l/s and l / K
We begin with some results on hypergeometric functions [see (1.3.5)].

Changing notation slightly we write
179
Proof. (a) is a special case of Kummer's identity given in Rainville

60, p. 671 or in ErdClyi et al. [53, Section 2.111. It may be verified by
ihowing that both sides satisfy the appropriate hypergeometric differential
:quation (given in Exercise 7 of Section 1.3), are analytic, and agree at
:ero. (b) is a special case of Clausen's product for hypergeometric functions
jven in Slater [66, p. 751 and Exercise 13.
In the sequel we will again use Ramanujan's invariants of (3.2.13)
We also need Klein's absolute invariant J , which was introduced in

rheorem 4.4. This is
Ramanujan [14] talks about "corresponding theories" for K,, s := $ , $ , $ ,

o that for K. For s := $ this is explained by the next result.
4,
Proposition 5.7
2 112
K1,4(h) = (1 + k
K(k)
if2hhf=[(g12+g-12)/2]-1a n d ~ r h r l l f i 0, < k s f i - 1 .
b) K,,,(h) = [I - (kk')2]114~(k)
if 2 h h f = J-11' and o s h r l l f i , 0 l s k s l l f i .
a)
Proof. These may be discovered by piecing together the quadratic and

:ubic transformations given in Erdklyi et al. [53, Section 2.111. They may be
rerified by establishing that both sides satisfy the same differential equation
derived from the appropriate hypergeometric differential equation), and
)oth functions involved have the same finite value at zero.
There is a corresponding relation for K,,,. Since it is a little less concise,
ve consider it at the end of the section. Combining these last two proposiions leads to a variety of alternate hypergeometric expressions for K and
c2.
Theorem 5.6
180
ls
2K
1 1
2k
(ii) - ?r( k ) = k r 2 F ( ; l ; - ( 1 )
0cksfi-1
2K
(iii) - ( k ) = k ' - 1 1 2 2 ~ ,
0sk2s2fi-2
k2
?r
( b ) (iv)
2K ( k )= (1 + k 2 ) - 1 1 2 2 ~ 1
?r
2K
( u ) - ( k ) = ( k r 2?r
(c) (vi)
2K
1 5
1 2 12
- ( k ) = [1 - ( k k r ) 2 ] - 1 1 4 2 ~,1-; 1; J 7r
1
fi'
'
Proof.
(a) We let s := 0 above to deduce ( i ) . Then (ii) follows on replacing q by
- q in the theta function representations of K and (2kk1l2.This is
Jacobi's imaginary transformation of Exercise 7d) in Section 3.2. We
derive (iii) from (ii) by replacing k by k , := (1 - k r ) l ( l+ k ' ) and
using the quadratic transformation K(k,) = [ ( l +k 1 ) / 2 ] K ( k ) of
Theorem 1.2.
( b ) (iv) comes from letting s := $ above. Then ( v ) again follows from
Jacobi7simaginary transformation.
(c) ( v i ) comes from letting s := 4 above. CI
Similarly,
Theorem 5.7
For k restricted as in Theorem 5.6
Proof. We combine Theorem 5.6 and Proposition 5.6. 0

Thus we have provided series for K and K 2 in terms of each of the six
invariants. One can produce other such formulae by further use of transformation identities. For example, Bailey's formula in ErdClyi et al.
[53, ( 2 ) , Section 4.51 with a := 3 and b.:= 1 gives
Note also that we may use (5.5.5) with s := 0 and Theorem 5.6 to
produce similar series for E. We are now ready to build our series. Recall
(5.1.4), which we write as
Thus given a ( N ) and A*(N), we can combine (5.5.10) with Theorem 5.7 to
produce series for 1/71..In like fashion we derive series for 1IK or for the
Gaussian AGM, M ( l , k t ) = n-/2K. In each case we have [(2Kln-)(k)12=
m(k)F(+(k)) for algebraic m and 4, while F(+) has a hypergeometric-type
7 4mF
~ ' + 4 m(/)p(+).Subpower series expansion C:=, an+". Then 4 ~ ~ 1 =
stitution in (5.5.10) leads to
182
Modular Equations and Algebraic Approximations to n
Thus for rational N , the bracketed term is of the form a + nb with a and b
algebraic. We now specialize this for our invariants.
SERIES IN
G,: . For N > 1,
where
a n ( N ):= [ a ( N ) - m k ; ]
SERIES IN
+ n m ( k h 2- k;)
gN: For N r 2,
where
SERIES IN
g4N = 2
114
g N G N : For N
$,
where
vm
c n ( N ):= [ a ( N ) - 1
k',] kh-l
SERIES IN X N :=
where
(gE +
2
g;12)-1
+ n m ( k &+ kh-')
- 4kNkh2
( 1 + k;)2
For N > 2 ,
'
These use the hypergeometric definition of K and (5.5.11). Also using

(5.5.11) and Theorem 5.6(ai) and (aii) leads to, for N > 1,
Modutar Equations and Algebraic Approximations to
184
?r
where
on(N) := [a(N) - m k i ] + n 2 m ( k h 2 - k i )
and for N r 2,
where
Similar formulae exist in the other invariants.

From our formulae for l r - I and the values of a(N) and A*(N) = k,
previously derived we have explicitly computed all but two of the 14 series
which Ramanujan gives without justification in [14, Section 141. Ramanujan
gives series of the form (5.5;13) for N:= 3, 7, 15, of the form (5.5.16) for
ZV:=6,10,18,22,58,andoftheform(5.5.17)forN:=5,9, 13,25,37.He
gives series of the form (5.5.18) for N := 3 and 7 and two in terms of K,,,
which we derive below. In each case manipulation of the formulae produces
the desired result. Indeed a(37) and a(%) were obtained by calculating
e0(37) and do(%) to high precision. In fact, with N := 58, using (5.5.16) and
Exercise 2 of Section 5.3 produces
- "~TN
, it is very
which adds eight digits a term! Since k i behaves like 16e
easy to estimate the convergence rate in each series. For N at all large, the
rate while linear is most impressive. In the exercises we give various other
examples. Bailey [35, p. 961 gives (5.5.14) with N := 2 [equivalently (5.5.15)
with N : = # ] and ascribes this to Ramanujan. The series is
Correspondingly, (5.5.22) with N := 2 yields
while for N := 1, the series in (5.5.20) diverges.
185
We now return to K,,,. From Goursat's exhaustive list of transformations

(Goursat [1881]) we obtain:
Proposition 5.8
For k < 1l f i and h the smaller of the two real solutions of
one has
Proof. Formula (126) in Goursat [I8811 and Propositions 5.6 and 5.7
combine to produce (5.5.26) and (5.5.27). 17
There is a corresponding formula for the larger solution. This implicit
formula for 1 in terms of k can be solved explicitly as follows (Exercise 19).
+ 1)/2*,
For h I(fi
where
and
a,
In each case (s := 4, & ) the transformation from k to h can

described very simply analytically. Consider the generalized singular value
function A: defined by
Then it transpires that in all three cases,
Modular Equations and Algebraic Approximations to .n
186
(This can be verified from formulae in Goursat [1881].) In other words, the
Nth singular value of K is sent to the Nth, (Nl2)th and (Nl3)th singular
values of K,,,, K,,,, and K,,,, respectively. Thus
and
while various singular values for K,,, are given in Exercise 19b). If we now
define a, by
(5.5.32)
a,(N)=-
Tr
COS ( r s )
4 ~ : 1+2s (
k:=A:(N)
we may use (5.5.5) and (5.5.6) to write
where
b, (N) :=
~~~
1
cos (7rs)
while
The details are left as Exercise 20. Now, with some perseverance, we can
derive series including Ramanujan's missing formulae, which come with
s := 2 and N := 4 and 5 in (5.5.33) [See Exercise 20b) .]
Finally, we observe that the result of Exercise 22 in combination with the
discussion of I' values in Section 9.2 shows that the formulae for K,,,, K,,,,
and K,,, are in essence the only such formulae.
In Section 13 of Ramanujan [14] one finds an explanation of series of the

form (5.5.13) without many details. Then in Section 14, with essentially no
5.5
Generalized Elliptic Integrals
187
explanation, he gives his other 14 series. Hardy quoting Mordell (in

Ramanujan [62]) observes that "it is unfortunate that Ramanujan has not
developed in detail the corresponding theories." The explanation as provided by this section is a bit disappointing, since for all these theories, all we
have are well-concealed versions of the original theory for K. Nonetheless
we can explain all of the beautiful and mysterious series.
1.
2.
3.
4.
Prove the generalized Legendre identity of (5.5.6).

Prove Proposition 5.7.
Verify formulae (5.5.13) to (5.5.23).
Show that
5. Show that
6. Show that
7.
Show that in (5.5.18)
while
and
8. Show that
9. Show that
Modular Equations and Algebraic Approximations to n.
and
Show that
m
~ f i M ( 2 "2-I")
~
2 (-1)"(12n + I)[,] ( $>, 8-"
= ~r
n=O
a) Show, using (5.5.13) with n := 7, that
This series of Ramanujan's has the property that, as J. Holloway

has observed, it can be used to compute the millionth (binary)
digit of l / a without computing the first half million digits. Note
that the terms are exact binary fractions whose numerators grow
roughly like 26nwhile the denominators are 16 .212".
b) Observe that formula (5.5.23) can be recast as
Use Exercise 6 of Section 1.3 to show that when re(c - a - b) > 0,
[This is easy for re(c) > re(b) > 0.1

Ciausen's product formula is
Prove this by showing that both sides satisfy the same generalized
hypergeometric equation and are analytic at zero, with value 1 there.
Verify that
5 .
189
Use Exercise 12 and Theorem 5.6(ai) to compute K ( 1 , f i ) . Similarly

use Theorem 5.6(biv) to compute K(* - 1). (Compare Section 1.6.)
a) Verify that
i) sin (tx) = (t sin x),F1
l+t l-t 3
- - . sin2x
2 .'2'
iii) log ( x + v 1 + x2)= x-,F,($, $; $; -x2).

b) Use Clausen's product to deduce that
1
sin2(t sin-' x) = - 2
c)
t)n
2 (t)n((2x)2n
(2n)!
Similarly deduce Euler's formula (Bromwich [26])
[This is also the limiting case of b). See also Exercise 16 of Section
11.3.1
d) Find similar formulae for sinh [t log (x + l/l+;;i)] and for
log (x +
e) Establish that
m).
and that
f) Prove that
sin ( r t )
gt
-=
a)
Prove that
.=o
(t)n(-t)n
(n!)'
'
Modular Equations and Algebraic Approximations to w
190
b) Show that the previous integral is Y(3) 110. (See Exercises 12d)
and 12e) of Section 11.3.)
18. Let G,(x) := ,F,(a, (2n
+ 1)/2; 2 a + 1; x).
a) From Clausen's product, establish that
b) Show that G-,,,(x) = (1 + G)

12. Thus
(Note that G-,,, must be evaluated as a limit.)

19. a) Establish the solution of (5.5.28). Hint: x : = h"/h2 satisfies a
simpler cubic than h2.
b) Now verify that the following solutions obtain:
i)
~ ? = 2
ii)
g:
iii)
G?
iv)
vi)
gives
= (fi1)'
+f l ) '
H? = 1
gives H? = 2
gives H
& = f i ( V 3 + 1)'
gives H?
g = ( + ) 4
gives H?=
= (2
2(2 + v3)2
3fl
27
( l4fi3&13fi)
Hint: For iii) and vi) use the increasing form of (5.5.28). which
gives H,, in terms of G,. This entails changing th; central sign in
(5.5.28ii).
20. a) Establish the general formula for .rr-' given by (5.5.33) and
(5.3.34).
b) Verify the following values of a,,,(N) + nb,,,(N):
i) N : = 2 gives (6n+1)13fl
ii) N := 4 gives (60n + 8) 127
iii) N := 5 gives (66n + 8) / l 5 f i .
c)
Compute the values of a,,,(N) and b,,,(N) for N := 3 and N := 6.
5.6 Other Approximations
21. By comparing (5.5.33) with (5.5.16) or (5.5.18), verify the assertions

of (5.5.31) for various N.
22. a) Use Exercise 12 to establish that
b) As in Exercise 15, compute K(h*(3)).
I
5.6
OTHER APPROXZMATZONS
We begin with equation (5.2.20), which we rewrite as
where k := h * ( p ) . Then this shows that
Thus on approximating 2 K I r by an algebraic quantity we produce various

approximations for T . The simplest [which also follows from (5.2.15i)l is
v y
'
-\"
vy"
We do better, however, by using Theorem 5.7(a) to write
so that
Ramanujan [14] uses different estimates of ILKIT. Motivated perhaps by

]~
symmetry considerations, he uses (3.2.16) to expand [ 2 K ( k ) l ~ as
( 1 - 2k2)-I + 0 ( k 2 ) and then (3.2.17) and (3.2.18) to obtain
192
Modular Equations and Algebraic Approximations to a~
Then his approximations to rr are

2
with error 0 ( f l e - + ) ,
(5.6'6)
and
3
rr2(p):= Vy - [ a ( p ) ( 4 G Z 4- 1 ) ] / [ ( 1- 2k2)(8GZ4+ 1)]
with error ~ ( f l e - ' ~ (See

) . Exercise 1.) Moreover, (5.6.5) and (5.6.6)
produce very simple approximations. Thus
3(3 + v3)
(5.6.7) ~ ~ ( 2=5 )
5
and
~ ' ( 2 5=)
63(17 + 1 5 a )
25(7 + 1 5 G )
The latter gives 11 digits of rr. Similarly,
84
~ ~ ( 3=7 )
2 1 f i - 101
(5.6.8)
and
~ ~ ( 3=7147
)
1 4 5 m + 1134
2 2 3 9 9 m - 41916
(The values of a and a are not quadratic surds.) For even p it is better to
use
which is again an O(flee-'+)
approximation. Thus we derive
An even more classical approximation to rr is obtained through taking

logarithms of G, or another invariant. Thus in the notation of the previous
section we may write that rr is approximately equal to
(5.6.11i)
fllog ( 8G
E)
log ( 5 )
xN
1
-log (1728J,)
2 f l
We leave it to the reader to estimate the error in each expression. (See

Exercise 4.) For example, when N := 58, (5.6.llii) produces
m,+
-4
a. fi )
5
12
which gives 10 digits of T.Not surprisingly, Ramanujan [14] gives a host of

xamples of this kind. When the invariant is large these give very good
lgebraic approximations for em. Following Shanks [82] one can take this
onsiderably further. In (3.2.9) we gave q-product formulae for
various invariants. Thus (3.1.4) and (3.2.9ii) yield
( 4kt2
-k '
x := 1 ( ~ ) 2= 4
n
m
=q
n=l
(1 + qn)24
and there is a similar expression for (kkt/4)'. We may expand this product as
a power series and compute as many terms as we wish of its reversion. This
will produce a series of the form
= x - 24x2
$i
+ 852x3- 35744x4+ . . .
We may also take logarithms in (5.6.14). Then we can write

m
log x + f l=~
24
log (1 + qn).=24
---n=l
k=l
k
(5.6.16)
which may be expanded as a power series in q. When we substitute (5.6.15)
into (5.6.16) we will recursively compute
for a fixed sequence {a,). Indeed a, := 1, a, := 47, a, := 2488, a, := 138799,

a, := 7976456, and a, := 467232200. In fact one can show (Newman and
2k-1 24n
Shanks [84]) that 24an is the coefficient of qn in Hi,, (1 + q
) .
Moreover, 24an < 64". A now standard trick of replacing q by - q shows that
(5.6.17) still holds for x := -(kk1/4)'. In Shanks [82] several large invariants
are computed (including g,,,, and G,,,,) to which (5.6.17) may be applied.
Modular Equations and Algebraic Approximations to n
194
Each series adds roughly T<N logl,e digits a term. Hence when we use g,,,,
given in (4.7.14), we gain more than 90 digits a term! But of course we have
to compute the logarithm as well. Thus in our context the formula should be
viewed as a rapid series for log (1x1) + m r r , not as a computation of rr.
Approximations (5.6.5) and (5.6.6) were the reason why Ramanujan computed R,. He did not give (5.6.9). Many very large invariants, including
G141ss,G,,,,,, and G2,1,s are derived in Shanks [82].
1. a)
Establish that (5.6.5) and (5i6.6) have the claimed errors without
using the rather deep formulae (3.2.16) to (3.2.18).
b) Show, using (3.2.16) to (3.2.18), that
and
2.
a) Use (5.6.5) and (5.6.6.) to deduce that

.rr,(13) =
3 ( 3 a + 7)
17
and
n;(13) =
1 0 3 a + 125
158
b) Verify (5.6.7) and (5.6.8).

c) Verify that (5.6.9) has the claimed error.
d) Show that
3. a) For even p , the estimate
will often be cleaner than rrl(p). It also gives o ( - - ~ )

in estimation of rr.
b) Obtain
rr4(58) =
c)
66a
3 3 m - 148
Compute q ( 5 8 ) and n;(22).
and
q(22) =
6.\/zz
33 - 1 7 a
error
'
195
a) Estimate the error in each of (5.6.11), (5.6.12), and (5.6.13).

[Compare (2.5.15) .]
b) Use (5.6.12) to estimate rr by
4
log (396)
and
4
log (842/2) .
rn
a) Show that
and
64gZ4= q - l - 24 + 2764 - 2048q2 + - .
with similar expressions for GZ4.
b) Thus
and
6 4 ( ~ ; + ( 3 ~ ' ~-) 24 = elTG + 4372e-lTm + . - . .
c) When g, or G, is a quadratic surd, this gives an expression for the
integer part of era (and the proximate 0's or 9's). Thus the
integer part of elTm is 2,508,951. That of elTm is 199,148,647, and
that of elTm is 24,591,257,751.
a) Show as discovered by Beukers that, with {a,) as in (5.6.17),
b) Combine ai) and Theorem 5.7(ai) to produce a recursion for {a,).

c) Show that a, is an integer.
A variety of other approximations to rr and to p, the perimeter of an

ellipse, can be found in Chapter 18 of Ramanujan7ssecond notebook
and in Ramanujan [14]. For example, given an ellipse of major axis a,
minor axis b, and eccentricity k := (bla)', he gives
i) p = 2 r r a 2F 1( I2 7 - a2 -7 1 .,k 2 ) = r r ( a + b ) 2 ~ l ( - $ , - $ ; 1 ; t )
where t := [(a - b) /(a + b)I2. Then, as t--t 0,
Modular Equations and Algebraic Approximations to .rr
196
ii) p-rr[3(a+ b ) - d ( 3 a + b)(3b+ a)]

and
iii) p
- rr(a + b)[l + 3t(10 + m t ) - 1 ' 2 ]
where the error in ii) is about 2-'t3 and the error in iii) is about
3 .2-l7tS.This is pleasantly developed in Almqvist and Berndt [Pr]
and in Berndt [Pr]. Truncating the approximation in iii) leads to
iv) p - r ( a + b)[l+ QtI2
with an error about 2-'t3. This is due to Nyvoll [78].
a) Prove i).
b) Justify the error estimates in ii), iii), and iv).
One may avoid the Landen transform in a) by following Ivory
[1796]. We write
8. a) Combine (5.1.2) and (5.2.14) to derive that
p = 4 a ~ ( k ) 2rra[~,'(k, f ) - M,(k, f)e,(k, f)] as k+ 0

where wn(k2,f ') = 0.
b) Show that the error is roughly of order ak2".
c) Deduce that to order ak4,
and to order ak8,
d) Use the cubic identities to establish that
when k = A*(?-).Thus, with order ak6,

k := fi- 1 gives
and 2kkf := G - 2
gives
2?ra(5fi
- 6 f i -7fi
+ 9)
l~a[3(2+ f i ) 6 ( 2 f i - 2 f i - 1)
In the notation of the AGM iteration let k : = cola, and let x: :=

cn+l"n+1.
a) Establish that
b) Let x := ( ~ , / 2 SO
) ~ that x = [ $(1 -
e)
/ ( l + *)I4
and
c) Thus with k := A*(r) one has another estimate for M with x of

order e-4"Ti. When r := 1, x = &[@I4 - 1 ) ~ / ( 2 ' /+~l ) I 4 and the
given terms yield 19 digits of T.
From our point of view possibly the most remarkable result in Chapter 18
Ramanujan's second notebook is the following continued fraction identity
en in entry 12. Let n > 0 be fixed. Define
a p2 (2a)' ( 3 ~ ) '-.
(4a)'
An(a, p ) : = n+ n+ n+
n+
n+
..
An(% P ) + L ( P 7 a )
2
henever /3 > a > 0 and
AG(B,
v m )
=1
198
[or, equivalently, whenever K(alP) = (nI2)PJ. Even more surprisingly, a

slight adjustment of the proof given in Berndt [Pr] shows that (5.6.18) holds
for all a, p > 0.
10. (The moments of K and E ) Let
and
Use the differential equations for K and E to establish that
and
Show that
Use contour integration of O/sin 8 (on the infinite rectangle above

[O, .n/2]) to deduce that
which is twice Catalan's constant, p(2) or G.

Establish that Kl = 1, El = 3 , and Eo = $ + P(2).
Observe that all the odd moments are rational while each even
moment is of the form a + bp(2) with a and b rational. Thus all
moments lie in Q(P(2)).
Analogously define KA and EA. Determine recursions for these
conjugate moments. Show that Kh = r 2 / 4 and Eh = r 2 / 8 . Compute Ki and E;.
g) Use the quadratic transformations to show that
h) Show that
Chapter Six
The Complexity of
Algebraic Functions
Abstract. The aim of this chapter is to analyze the complexity of algebraic
functions in general; and of multiplication, division, and root extraction in
particular. There are two primary tools, Newton's method and the fast
Fourier transform.
6.1
COMPLEXITY CONCERNS
It is obviously inappropriate to consider the multiplication of two manythousand-digit numbers to be of equal difficulty to the multiplication of two
single-digit numbers. A reasonable measure of complexity that takes this
into account is the bit complexity. The bit complexity of an algorithm is the
number of single-digit operations required to terminate the algorithm.
singfee-digitoperations include addition, multiplication, logical comparison,
and storage and retrieval of single-digit numbers. We are exclusively interested in how the complexity increases with the size of the problem. For
example, addition of two n-digit integers by the usual algorithm has bit
complexity O,(n)-the
subscript B on the order symbol is for emphasis.
This is a serial notion of complexity in the sense that on a serial machine it is
an appropriate asymptotic measure of the time required for the calculation.
We use the slightly nonstandard notation
a, = O(bn)
and
b, = O(a,) .
6.1 Complexity Concerns
--
201
= fl(b,), we say that {a,) and {b,) are equivalent. Since accessing an
it number requires fl(n) bit operations, it is apparent that "usual"
ion of two n-digit integers is in fact fl(n). These trivial lower bounds
e a consequence of uniqueness considerations-if we change any digit of
of the numbers being added, we change the answer. Thus any algorithm
additioli must at least "inspect" every digit. So, up to a constant, usual
diton is asymptotically optimal. As we shall see later, one of the interestg consequences of this body of theory is that "usual" multiplication is far
om asymptotically optimal.
A detailed approach to complexity requires a model of computation and
perhaps most readily made rigorous in an analysis of Turing machines.
ee, for example, Aho, Hopcroft, and Ullman [74].) This much detail is
necessary for our purposes.
We will usually content ourselves with merely counting single-digit addins and multiplications. In all the algorithms we present the comparisons
ote that the comparison of two n-digit numbers is LR,(n)] and the storage
ncerns will be bounded by the arithmetic operations-provided the alrithms are sensibly implemented. This is almost always transparent and
11 rarely even elicit comment.
Operational complexity counts the number of operations (addition, multiication, division, and extraction of kth roots performed to a precision
ded by the precision of the output). When all the operations in an
ithm are performed to roughly the same precision, this is a useful
easure. The reasons for the particular choice of operations will be made
parent in Section 6.4. Thus the algorithms of Chapter 5 compute n digits
T with operational complexity O,,(log n); once again the subscript on the
der svmb01 is for em~hasis.
of the primary tools for the analysis of algorithms is the use of recursive
tions. The idea is to divide a problem into smaller subproblems that can
e solved by essentially the same technique and then recurse, a strategy
often called "divide and conquer." The reader unfamiliar with this approach
might like to examine the exercises. One of the lessons of complexity theory
is that many of the usual algorithms of mathematics are far from optimal.
Multiplication, taking Fourier transforms, and matrix multiplication are
but three examples. (See Exercise 3 and the next section.) A second lesson
is that good lower bounds are very difficult to obtain. For the analytic
algorithms we are considering, the only lower bounds we can establish are
the trivial ones. Thus unless the algorithm is of the same order, as is the case
for addition, we cannot achieve exact results,
We shall not discuss combinatorial complexity except to mention that an
introduction to this well-developed and important field may be found in
Aho, Hopcroft, and Ullman [74].
202
1. Prove the following. Let a, b > 0 and c > 1. Suppose that f is monotone
on (0, 00).
a) If f(n) < af(nlc)
+ bn and f(1) = d, then

f b ) = o(n>
f(n) = O(n log n)
f(n) = 0(nlogcn)
b) If f(n)
Iaf(nla)
+ bn(1og n)"-'
ifa<c
if a = c
ifa>c
and f(1) = d, then
Hint: Analyze a) with equality. Then establish the general principle

that the equality solution is the maximal solution.
2. Show that the usual algorithms for multiplying and adding two n x n
matrices have complexity f10,(n3) and flo,(n2), respectively. (We are
counting the number of multiplications and additions of the entries of
the matrices.)
3. (Fast matrix multiplication (Strassen 1969))
a)
Show that
can be computed from
and
b) Observe that the above method reduces the multiplication of

2n x 2n matrices to 7 multiplications and 18 additions of n x n
6.1 Complexity Concerns
203
matrices. Thus if W(n) is the operational complexity of multiplying

two n x n matrices, then by iterating the procedure in a)
The final term comes from using usual matrix addition for the 18
additions. The constant a can be chosen independent of n and can
be used to include the "overhead" of actually breaking the problem
up. Use the above inequality to show that
4.
Note that log,7 5 2.81, so the above method is asymptotically faster

than the usual method. [Extensions of this method can reduce the
bound for multiplication down at least to ~ ( n ' . ~ - ) much
,
as in
Knuth [81]. The best lower bound known is the trivial one, en2.]
Suppose A is a nonsingular 2n X 2n triangular matrix. Write
where B, C, and D are n x n matrices.

a) Show that B-' and
D-I
exist, and that
b) Show that a) iterates to produce an algorithm for inverting 2" X 2"

triangular matrices. Let I(n) be the operational complexity of this
algorithm. Show with W(n) as in Exercise 3 that
and hence that
(One can show that in general matrix inversion and matrix multiplication are asymptotically equivalent. See Aho, Hopcroft, and U1lman [74].)
5.
Show that the bit complexity of calculating n! by multiplying 1X 2 X 3 X

.- using usual multiplication is a((n log n)').
Hint: Analyze the complexity of multiplying an n-digit number by an
m-digit number. Use Stirling's formula to estimate the number of digits
in k!. Exercise 10 of Section 6.4 explores this further.
204
6.2
THE FAST FOURIER TRANSFORM (FFT)
Let w be a primitive (n + 1)th root of unity either in C or in a finite field F,,

that is, wn+' = 1 and w k# 1for k < n + 1. In the complex case we may take
:= e2+l(n+l) . Consider the following two problems.
Given n + 1 numbers cu,, . . . , a,, find the coefficients of the unique polynomial p,(z) := a, + a,z + - - + a,zn of degree n
that satisfies
INTERPOLATION PROBLEM.
EVALUATION PROBLEM. Given the coefficients of a polynomial p, of degree

n, calculate the n + 1 values
These are the two directions of the finite or discrete Fourier transform.
The classical approaches to either part of the Fourier transform problem
have operational complexity at least cn2. This is the operational complexity,
for example, of evaluating p, at n + 1 points using Horner's rule [writing
p,(x) = (((anx+ a,-,)x + a,-,)x + . - -)I. We wish to prove that, in fact,
both directions can be solved with complexity Oo,(n log n). Actually, we
only treat the case n + 1:= c2", which is sufficient for our purposes and
somewhat simpler.
Theorem 6.1 (Fast Fourier Transform)
If n + 1= c2" with c an integral constant, then both the interpolation and

the evaluation problem have operational complexity Oo,(n log n).
Proof. We assume c = 1, that is, n + 1= 2", the case for general c is
entirely analogous. We treat evaluation first. Suppose
Let
and
xr(x2) := x(al + a,x2 + . - . + a,xn-l)
Then with y := x2,
6.2 The Fast Fourier Transform (FFT)

= xr(y)
205
+dy)
and q are both polynomials of degree 2"-l - 1. The observation

kes the proof work is that for w an (n + 1)th root of unity,
i 2-
( w ) -(w
(n+1)12+i 2
e, evaluating p(x) at the n + 1 roots of unity in (6.2.1) reduces to

ating r and q each at the (n + 1) I2 points (w2)l, (w2)', . . . , (w2)("+')I2
lgamating the results. Observe that w2 is a primitive (2"-')th root
unity and we can iterate this process. Let F(2") be the number of
ns and multiplications required to evaluate a polynomial of degree
at the 2" points wk, k = 1, . . . ,2", where w is a primitive (2")th root
f unity. As above,
F(2") = 2~(2"-')
+ 2 - 2"
F(1) = 0 .
a1 term comes from the single addition and multiplication required to

lculate each p(wi) from r(w2') and q ( ~ 2 i )The
. recursion (6.2.2) solves as
- rn
F(2") = 2"+'
and the bound for the evaluation problem follows.

The interpolation problem is equivalent to evaluation. This can be seen
as follows. Let w be a primitive (n + 1)th root of unity and let
(; :5 ;:)
(
:-:)
1
w :=
1::
wn
(6.2.5)
...
W-' = n i l
...
wZn
wn2
1
1
1
1
-1
w
w-2
w -4
...
W-,,
w-~"
...
-2
* * .
w-2
and w-I is also a primitive (n + 1)th root of unity. (See Exercise 4.) The
interPolatioh problem can be written as: Find (a,, . . . , a,) so that
W(ao, . . . ,a,,) = (a,,
. . . ,a,,).
206
However, this can be solved by

~ - ' ( a , , , . . . , a,)
= (a,,
. . . , a,)
which is exactly the evaluation problem.

While restricting the FFT to powers of 2 poses no problem for us over @,
it is a nuisance over finite fields. The problem is that F, has primitive kth
roots if and only if k divides m - 1. Hence, our approach is restricted in the
finite case to considering primes of the form m = ~2~+ 1, which are not
particularly abundant. There are many ways around this difficulty. This is
discussed in Winograd [SO].
The FFT is an enormously useful and widely used algorithm. Depending on
exact form and implementation, it can outperform traditional methods for
values of n well below 100. For a history of the FFT consult Cooley, Lewis,
and Welch [67]. While antecedents for FFT methods are plentiful, Cooley
and Tukey [65] are primarily responsible for introducing the FFT in its
modern form as a complexity-reduced method. More extended discussion of
the FFT and related matters may be found in Aho, Hopcroft, and Ullman
[74] and Winograd [SO]. As a theoretical tool the FFT and related methods
are central. They form the basis for the next section's discussion of fast
multiplication. In the exercises we show how these ideas can be used to
construct asymptotically fast polynomial multiplication, division, and interpolation algorithms. It is even possible to accelerate integer factoring
algorithms by FFT methods. In Chapter 10 an application to the estimation
of certain transcendental functions is provided. Our proof of Theorem 6.1
and some of the exercises follow Borodin and Munro [75].
1. a) (Fast polynomial multiplication) Consider the following algorithm
for multiplying polynomials. Given the coefficients of p and q (both
of degree s n - I), compute the coefficients of pq as follows:
Step 1: Evaluate p and q at 2n points wl, ,. . , w2", where w is a
primitive (2n)th root of unity.*
Step 2: Form the 2n products
i = 1 , . . . ,2n.
p(wi)q(wi)
Step 3: Solve the interpolation problem for pq to find the
coefficients.
Show, using an FFT, that the above algorithm has operational
complexity
* Strictly speaking, we should be using (2")th roots of unity where we have established an
EFT. This can be arranged by padding with leading zero terms, if necessary, without changing
the order of complexity.
6.2 The Fast Fourier Transform (Fm)
207
[The usual convolution product algorithm has operational complexitY W 2 ) . 1

b) Given
show that the coefficients of p can be calculated in O,,(n(log n)').

Hint: Treat the problem recursively and recombine the pieces using
a fast polynomial multiplication.
(Fast polynomial division) Given p of degree n and q of degree m s n ,
both with integer coefficients, it is possible to find u and r with
deg r < deg q so that
in O,,(n log n).

Outline: Simplify by observing that it suffices to calculate u since r may
then be computed by Exercise 1. Set x := 1lx. Then
and so
where h r 1. To calculate z7. (and hence u) it suffices to calculate the first

n - m (= deg u) Taylor coefficients of l l q . This can be done by
Newton's method as follows. Suppose that deg si = j - 1 and that
Establish that
[Note that we may assume q(0) # 0.1 Now the computation of si+ := 2si s:q can be performed using an FIT-based polynomial multiplication and
need only be performed using the first 2j - 1 coefficients of q and si. By
starting with an appropriate first estimate of s; [say, so(x) := 1Iq(O)] and
proceeding inductively as above (doubling the number of coefficients
utilized at each stage), show that the required number of terms of the
expansion can be calculated in O,,(n log n). (See Section 6.4.)
208
(Fast polynomial evaluation) Given p of degree n and n + 1 distinct

points x,, . . . ,x,, show that p(xo), . . . ,p(xn) can all be evaluated in
Oop(n(lognIZ).
(x - xi) and let r, be the remainder on dividing
Hint: Let q, ( x ) = II:.$-'
p by q,. Note that rl(xi) =p(xi) for i < n12. Similarly use
q2(x) := 11:,2 (X- xi). Thus two divisions reduces the problem to two
problems of half the size. Use Exercise 2 and evaluate the recursion.
[The other direction of this problem, namely, constructing Lagrange
interpolating polynomials, is also O,,(n(log n)'). See, for example,
Borodin and Munro [75]. In fact, both directions are Oo,(n log n).]
4. Show that if w is a primitive (n + 1)th root of unity, then
3.
.ij={","
i=o
5.
j=Omod(n+l)
otherwise .
Show that (6.2.4) and (6.2.5) are inverse to each other.

(Reversion of power series) Let f(x) := C;=, a,x k be a formal power
series, with known coefficients.
a) Show, as in the proof of Exercise 2, that the first n coefficients of
the formal series expansion of llf(x) can be computed in
Oop(n1% n).
b) Discuss the complexity of computing the coefficients of the formal
inverse off by Newton's method.
6.
(On calculating xn) The S-and-X binary method for calculating xn is

the following rule. Suppose n has binary representation S0S,S2.. - 6,
with So= 1. Given symbols S and X, define
si:=
and con6ruct the sequence
Now let S be the operation of squaring and let X be the operation of

multiplying by x. Let the sequence of operations S1S2.-.S, operate
from left to right beginning with x. For example, for n = 27,
and
6.3 Fast Multiplication
The sequence of calculations for xZ7is then
x+x2+x3+x6+x12+~13+x26+x27.
a) Prove that the above method computes xn and observe that it only
requires storing x, n, and one partial product.
b) Show that the number of multiplications is less than 2 [log, n] .
c) Show that the above method is optimal for the computation of xZm
(considering only multiplications).
d) Show that the S-and-X method is not optimal for computing xl?
An extended discussion of this interesting and old problem is presented
6.3 FAST MULTIPLICATION
We wish to present a strategy for multiplying very large numbers that is

considerably faster than the usual 0,(n2) method. The idea is to exploit the
FFT. Let a and /3 be two n-digit integers and write
a(x) := anxn-l + a , - , ~ " - ~+ . . - + a1
P(x) := bnxn-l
+ bn-lxn-2 + - .. + bl .
Then if ai and bi are the decimal digits of a and 6, respectively, we have

(6.3.3)
a = a(10)
and
/3 = /3(10) .
We can now calculate a - /3 by using the FFT, as in Exercise 1 of the last

section, to compute the coefficients of
(6.3.4)
Y(X):= a(x)P(x) .
Finally we evaluate y(10).

Let T(n) denote the bit complexity of multiplying two n-digit integers (or
equivalently, floating point numbers) by the above method.
The analysis of the complexity of the algorithm that follows assumes that
we are working with complex roots of unity (and that we are working with a
problem of size 2", as can always be arranged by padding with leading
zeros). This introduces rounding error problems into an intrinsically integer
algorithm; however, this is the setting which we have established an
abundance of values for which the FFT works. Since in practice (and in
210
theory) there are plenty of FFT analogues in a finite setting, one may, if one
prefers, consider the entire algorithm performed mod (O(n)).
Step 1: . Evaluate a(x) and p(x) to precision O(1og n) at the 2n points
wl, w2, . . . ,wZnwith w a primitive (2n)th root of unity. Using the FFT, this
has operational complexity
and bit complexity
[It suffices to use w to precision O(1og n) and thus every multiplication in

Theorem 6.1 is of complexity OB(T(logn)). Observe that the coefficients of
a and p are single-digit numbers and that the coefficients of y are hence
O(1og n)-digit numbers. See Exercise 3.1
Step 2 : Form the 2n products
,
computed to O(1og n) digit precision. This is of bit complexity

(6.3.6)
O~(nT(logn)) -
Step 3: Interpolate the coefficients of y to precision O(1og n) using the

FFT. This has operational complexity
O,,(n 1% n)
and bit complexity, as before,
(6.3.7)
OB(n(log n) T(l% n)) -
Step 4 : Evaluate y(10). This has bit complexity

(6.3.8)
OB@>.
This step essentially requires only addition. The coefficients of y are closely
related to the digits of ap except that they may be too large and a "carry"
must be performed.
The total complexity thus satisfies
6.3 Fast Multiplication
T(n) = 0, (n(1og n) T(log n))
T(n) = OB(n(logn)2(log(logn))2 . . .)
with the product terminating when the iterated log is less than 1.
This is not optimal. (See Exercise 2.) The same analysis as above using a
base n representation of a and 3./ reduces the time for multiplication to
OB(n(log n)(log log n) . . .) .
This is still not quite the asymptotically fastest known algorithm. The best
bound is due to Schonhage and Strassen [71] and is
O,(n(log n)(log 1% n)) .
We shall call any multiplication that performs with this speed a SchonhageStrassen multiplication. It should be emphasized that in all the reduced
complexity multiplications we present, the order estimates include the
overhead additions (and storage concerns).
The observation that multiplication is not intrinsically 0,(n2) was made by

Karatsuba in 1962. He proposed an 0,(n1Ogz3) algorithm. (See Exercise 1.)
Subsequent refinements and improvements are due to Toom, Cook,
Schonhage, Strassen, and others, culminating in the Schonhage-Strassen
multiplication of 1971. This algorithm is of the same flavour as the one
presented above using size 2* representations and performing FFT opera-.
tions m 0 d ( 2 ' ~+ 1). A presentation may be found in Aho, Hopcroft, and
Ullman [74]. Knuth [81] has an extended discussion of multiplication
strategies which includes a discussion of the precision concerns of performing a fast multiplication over C. Cook and Aanderaa [69] conjecture that
multiplication is not O,(n). Here one must be careful about the model of
computation allowed. Under some more powerful than usual models O,(n)
multiplication is possible (see Knuth [81]), while under other more restrictive than usual models it can be shown to be not possible.
Once again it is possible to implement a fast multiplication that will
outperform traditional methods for n in the several hundred-digit range. For
the many million-digit calculations of rr discussed in Chapter 11, use of a fast
muftiplication is imperative. For a discussion of the multiplication used in
this setting see Tamura and Kanada [Pr] and Bailey [88].
(An ~ , ( n ' " ~multiplication)

~~)
Observe that
(a + blOn)(c + dlOn) = ac t [(a - b)(d - c) + ac + bd]lOn + bd1oZn
Use this to reduce multiplication of 2n-digit numbers to three multiplications of n-digit numbers and some additions. Show that this can be
used to produce a multiplication of
[This method can be refined to produce an O,(nl+"

are in Knuth [81].]
Construct a multiplication of complexity
algorithm; details
OB(n(logn)(log log n) . ..) .

Hint: Instead of using base 10 representations, use base n representations so that the polynomials (6.3.1) and (6.3.2) are polynomials of
degree n with coefficients of length log n. Now proceed as in the
algorithm of this section.
Discuss the bit complexity of the FFT. Show that if the input is given to
precision O(m) and the output is required to precision O(m), the bit
complexity is
where M(m) is the complexity of whatever multiplication is employed.

6.4 NEWTON'S METHOD AND THE COMPLEXITY OF
ALGEBRAIC FUNCTIONS
We wish to show the equivalence, from a complexity point of view, of

multipiication, division, and root extraction. The primary tool is Newton's
method.
Theorem 6.2
Suppose that f is analytic in a complex neighbourhood of z . Supposef(z) = 0

and f(z) # 0. Then the iterati~n
converges uniformly quadratically to z for initial values x, in some neighbourhood of z.
6.4 Newton's Method
213
The reader unfamiliar with the proof is directed to Exercise 1. Newton's

method is useful for calculating inverse functions. Observe that g-'(y) is a
zero of f(x) = g(x) - y. This is the content of the corollary. The uniform
nature of the convergence is discussed in the exercises.
Corollary 6.1
Suppose that f is analytic and one to one in a neighbourhood of z,. Then

there is a neighbourhood of f(z,) where f-' can be computed uniformly
quadratically by Newton's method.
The quadratic nature of the convergence is only half the story. The other
half is the "self-correcting" nature of Newton's method. Suppose that
f(z) = 0 and that we are computing z by Newton's method. If the nth iterate
x, is perturbed by an amount O(Ix, - zl), then provided we stay in the
domain of uniform quadratic convergence, computing x,,, from the perturbed value of x, will preserve quadratic convergence. In other words, if
x, - z agree through M digits, then the calculation of x,,, need only be
performed to precision 2M.
Let M(n) denote the bit complexity of multiplication of two n-digit
numbers by some method. We make the following regularity assumptions:
2M(n) IM(2n) s 4M(n)
and
M(n) is nondecreasing .
(6.4.2)
Since two multiplications of length n can be viewed as subproducts of a
single multiplication of length 2n, while four multiplications of length n
comprise one of length 2n, the first part of the assumption is reasonable.
Since multiplications of length n can be padded with leading zeros to
multiplications of length n + k, the second part is also reasonable. Of course
it is easy to imagine a perversely designed multiplication for which (6.4.2)
does not hold.
Let D(n) and R(n), respectively, denote the bit complexity of division
and extraction of square roots, where the input and output are to precision
n. We say that two operations are equivalent if the complexity of one is
bounded by the complexity of the other and conversely. For example, we
say multiplication and division are equivalent if, given a multiplication with
bit complexity M(n), we can construct a division with bit complexity
D(n) = O(M(n)); and conversely, given a division we can so construct a
multiplication. The following remarkable theorem, the first part of which is
due to Cook, may be found in Brent [76c].
Theorem 6.3
Multiplication, division, and root extraction are all equivalent.
214
Proof.
(a)
We first construct a division. Applying Newton's method to the

function f(x) := 1lx - y leads to the iteration
which employs only multiplication and addition. Note that
and the quadratic nature of the convergence is manifest. We assume

Iyl r 1 and that y lies in a neighbourhood V bounded away from zero
(which if we are working in floating point, is no restriction). We may
also assume that, by using a usual 0,(n2) division performed to a
fixed low precision, we have already computed x, := x,(y), so that
Ixo - l l y l < 1/10. Assume n is a power of 2. Then log, n iterations of (6.4.3) will by (6.4.4) produce an error bounded by
and hence provide n digits of l l y . Furthermore, by the self-correcting

nature of Newton's method, the kth step of (6.4.3) requires two
multiplications and two additions of precision only 2k. Thus the total
complexity of the iteration is given by
since 2 ~ ( 2 5~~)( 2 ~ " ' )We

. have shown that l l y can be calculated
with complexity O,(M(n)). Hence since a1b = a (1 Ib),
(b) The equivalence of division and multiplication is now obvious since

a
ab = llb '
(c)
Square roots can be extracted by Newton's method applied to x2 - y,

which yields the classical iteration
6.4
Newton's Method
This satisfies
and the quadratic convergence is again apparent. We can proceed

exactly as in (a) to show that
R(n) = O,(D(n)) = O,(M(n))
is Exercise 7.
It is apparent that any time a function may be quadratically computed by
Newton's method from an iteration involving only addition, multiplication,
and division, that function will be of complexity O,(M(n)). This applies to
any algebraic function over Q(x), that is, any function f satisfying an
equation
(6.4.8)
@(x, f(x)) = 0
where @ is a polynomial in two variables with rational coefficients. More

precisely:
Theorem 6.4
Iff is algebraic over Q(x), then the complexity of calculating n digits of f(x)
is O,(M(n)).
The preceding results, of course, assume that we are avoiding the branch
points of the function in question.
Newton's method has a host of refinements and variants. See, for example,
Householder [70] and Exercise 3. The iteration for square roots can in some
form be traced back to the Babylonians, who used one or two steps of the
method.
It should be observed that not only is D(n) = O,(M(n)), but the constant
concealed by the order sign is fairly small. [From (6.4.5) we see that a
constant 8 works. Indeed, for all known multiplications, the additions term
is negligible and a constant of 4 is appropriate.] This is also the case for root
216
extraction. See Brent [76c], where a number of constants for these and
various other equivalences, as in Exercise 7, are established.
Further discussion of the calculation of algebraic functions may be found
in Kung and Traub [78]. Related matters may also be pursued in Lipson
Wl1. a) Prove Theorem 6.1 by observing that
Substitute this into (6.4.1) to get
Use explicit estimates to prove uniformity.

b) Show that for real f and real x,, the tangent to f at x, intersects the
x axis at x,+,.
c) Suppose that f is convex and strictly increasing on [a, b] and that
f(2) = 0 for some 2 in (a, b). Show that if b r x, > 2, then {x,)
decreases to 2 and convergence is guaranteed.
2. Construct the Newton iteration for yllP by inverting xP. Write x,,, llp 2
yllp in terms of (x, - y
) , thus exhibiting explicitly the quadratic
convergence. For which real starting values does the method converge?
3. a) Consider the iteration
where the notation indicates that the derivatives are evaluated at

y,. For sufficiently well-behaved (for example, analytic) f, this
method will find a zero of f with (n + 2)th-order convergence
provided various derivatives are nonvanishing. Prove these assertions. For n : = 0 this is just Newton's method, for n := 1 it is
Halley's method. (See Householder [70].)
b) Let
Xk+,
:= x k ( l + (1 - YX,)
+ (1 - y ~ k ) 2 )
Show that
and that for x, sufficiently close to f i y , x, converges cubically to

1l y .
6.4
Newton's Method
c) Let
x,+, := Qxk(15- 10xky+ 3y x,) .
2
2 4
Show that
and that for x, sufficiently close to l l f i , x, converges cubically to

1I*.
d) Show that b) and c) require fewer multiplications than (6.4.3) or
(6.4.6) for computation of reciprocals and square roots. These
are, in practice, very good high precision algorithms.
Let x, := 1 and let
be the iteration (6.4.3) for computing llx. Show that x,,, is the
(2,+' - 1)th Taylor polynomial of f(x) := 1lx expanded around the
point 1. Thus for division we may think of Newton's method as a
means of accelerating the computation of the Taylor series.
Let x, := 1 and let
be the iteration (6.4.6) for computing fi.Show thqt x,+, is,a rational
function r(x) with numerator of degree 2k and denchinator of degree
2, - 1 that satisfies
[This implies that r(x) is the (2,, 2, - 1) Pad6 approximant to fi at 1.

See Section 10.1.1
Invert 1/x2- y to calculate fi without using any divisions.
(Other equivalences; Brent [76c])
a) Show that squaring is equivalent to multiplication by considering
(a + b)' - (a - b)'.
b) Complete the proof of Theorem 6.3 by showing that square root
extraction is equivalent to multiplication.
Hint:
Use S := lo-" for appropriate rn to reduce the computation of x2

to root extraction and O,(n) operations.
218
c) Show that inversion is equivalent to multiplication.

d) Show that pth-root extraction is equivalent to multiplication.
8.
a) Suppose that Cf(n) is the bit complexity of calculating n digits off.

Assume that Cf(n) is increasing and that Cf(2n) r 2 Cf(n) . Suppose that f satisfies the conditions of Corollary 6.1. Show that f has
at worst the same bit complexity as f, and that
b) Write down Newton's method for computing exp from log and log
from exp. These particularly simple iterations combine with a) to
show that the problems of calculating exp and log are effectively
equivalent, from a bit complexity viewpoint.
9.
(Fast base conversion; Schonhage) Let k and j be fixed integers. Show

that an n-"digit" base k number can be converted into base j with bit
complexity
Hint: Break the number to be converted in half (base k), convert each
half, and recombine. (See Knuth [81] for a lengthy discussion of this
problem.)
10. a) Show that n! can be calculated with bit complexity
O,(log n M(n log n)) .

Hint: First calculate 1 x 2, 3 x 4, . . . . Then calculate 1 x 2 x 3 x 4,
5 x 6 x 7 x 8 , . . . etc.
b) Compare this to calculating n! as in Exercise 3 of Section 6.1.
Show that no multiplication can reduce this method below 0(n2).
[The best known bound for n! is OB((loglog n)M(n log n)); see
P. B. Borwein [85].]
Chapter Seven
Algorithms for the

Abstract. We analyze algorithms for the transcendental elementary functions
based on the transformation theory for elliptic integrals and in particular on
the AGM.
7.1
~r
AND
LOG
All the elementary transcendental functions can be calculated with bit

complexity OB(log n M(n)). This is a consequence of the fact that log has
operational complexity O,,(log n), and hence has bit complexity
OB(logn M(n)). The approach to log rests most easily on the logarithmic
asymptotic of K at 1. Before proceeding with this analysis it is convenient to
record the complexity of the algorithms for T,based either on iterating the
modular equation, W,, or more specially on the AGM.
Theorem 7.1
The initial n digits of n- can be calculated with operational complexity
and with bit complexity
Proof. Both Algorithms 2.1 and 2.2 as well as most of those of Chapter
5 perform with the above complexity.
Algorithms for the Elementary Functions
220
In all of the above algorithms some of the calculation may be done to

reduced precision. For example, in Algorithm 2.1 the computation of x, and
y,, which both tend to 1 quadratically, can be performed at successively
lower precision, The saving, however, is only in the constant term, reducing
it by a factor of less than 2.
We record the following estimates:
Theorem 7.2
(:)I=
(71.1)
I K ' ( ~ )-log
(7.1.2)
IEr(k) - 11= 0(lk2log kl)
(7.1.3)
(7.1.4)
I E ' ( ~ )- 111101k2log k (
Kr(k) -log
():I
0(lk2 log kl)
5 101k2log
re(k) > 0 , k-+O

re(k) > 0, k +o
kl
k E (0,
k E (0,
Proof. The relationships for K are in Exercise 4 of Section 1.3. (See

also Exercise 1 of Section 2.3.) For (7.1.2) observe that by (1.3.2), for
O<k<l
The constant in (7.1.4) requires a little additional scrutiny.

The following approach to calculating log is essentially due to Salamin (in
Beeler et al. [72]).
Algorithm 7.1
For x ( i , l ) and n r 3 ,
where
and
7.1
lr
and Log
77
ZAG(1,
This algorithm has operational complexity Oop(logn) and bit complexity
Proof. The estimate (7.1.5) is immediate from Theorem 7.2. The

n of the two elliptic integrals requires precomputing T, which has
Oo,(logn). The final detail is that AG(1,lO-"x) and
be calculated to precision 2n using O(log n) iterations of
n. (See Exercise 1.)
A related algorithm that avoids precomputing T can be established from
gorithm 1.2 (which provides a direct calculation of KrIEr).
et Rr(k) := Kr(k)lE'(k). For x E ( $ ,1) and n r 3,

llog x - Rr(lO-") + Rr(lO-"x)I
ed with the AGM process commencing with a, := 1 and

This algorithm has operational complexity Oop(logn) and bit complexity
The proof is straightforward and is left as an exercise. Both of the above
algorithms are based on an underlying quadratic method for calculating K r
and Er. Algorithms based on pth-order methods can be constructed as in
Chapter 5. A quartic version, for example, can be derived from Exercise 3
of Section 1.4. Instead of calculating Rr from the AGM, we use
1- Zb, 4"[a: - [(a:
+ p:)/2]2]
222
In this, as in other quartic versions of quadratic algorithms, there is a

substantial computational saving (as much as 35%).
The algorithms of this section for log all suffer from the drawback that they
are not truly iterative. Increasing the precision requires choosing new
starting values. This is more of an aesthetic than a computational~problem;
even with iterative methods one only computes to a fixed precision, and
increasing the accuracy usually entails starting at least one of the calculations
all over again. While two different AGM or related processes must be
calculated for the initial value of log, subsequent values require computing
only a single AGM since one of the terms can be reused. These methods are
quite stable, requiring only O(1og log n)-guard digits. They will outcompete
traditional methods--depending enormously on implementation-in the several-hundred-digit range.
The algorithms of this section are the asymptotically fastest known
algorithms for log (see Section 10.2) and are faster than any known
algorithms based on other methods, although OB((log n ) 2 ~ ( n ) )can be
achieved by techniques of Chapter 10. These types of algorithms were first
examined by Salamin (Beeler et al. [72]) and independently by Brent
[76a, b, and c]. Newman [82] gives a self-contained account, as do Bonvein
and Bonvein [84a]. The second algorithm is in Bonvein and Bonvein [84d].
Finally, while we have only presented the algorithms for real k, they
extend naturally into the complex plane; only the error estimates become
slightly more complicated. (See Exercise 1.) Matrix versions due to Stickel
[85] are discussed in Exercise 6.
1. Show that Algorithm 7.1 can be used to calculate log uniformly for
{z E CI lz - 11< f } with operational complexity O,,(log n) and bit
complexity 0, (log n M(n)) .
2. Examine the convergence of the AGM for a, := 1 and b, := lo-".
Specifically, estimate the number of iterations required to produce an
answer within lo-" of the limit. For Algorithms 7.1 and 7.2, find a
reasonable bound on the number of iterations of the AGM required to
produce a 1000-digit precision algorithm for log x , x E ( f , 1).
3. (An asymptotic algorithm for T ) Show that, for n 2 3,
and hence
7.1
l~
and Log
This provides an Oop(logn) algorithm for n. (See Newman [82] or

Bonvein and Borwein [84a].)
4. Given the pth-order modular equation 0, (as in Section 4.5), show how
to construct asymptotic algorithms for log analogous to Algorithms 7.1
and 7.2, but with an underlying pth-order iteration.
5. Show how the series expansion for K' and E' of Section 1.3 can be
combined with Algorithms 7.1 and 7.2 to provide Oop(logn) algorithms
for log that provide n digits of log using starting values 10-"Ikx and
10-"Ik.
6. (The matrix AGM) Let PNdenote the N X N self-adjoint positive
definite matrices and let I denote the N X N identity matrix. Let
A, := A E P N , Bo : = I ,
(7.1.9i)
A,+, := ;(A,
+ B,)
a) Show that if A E P N , then there exists a unique C E PNso that

c2= A. Hint: The iteration C,+, := C, f 4 (A - c:), Co := 0 converges to C.
b) Suppose that X, E PNcommutes with A E PN.Show that, for Xo
sufficiently close to a , Newton's method X, := (X, + AX, l )
converges to
E PNquadratically.
c) Show that the matrix AGM (7.1.9) converges to a matrix
AG(A, I ) E PNand show that A, - B, converges quadratically to
zero.
d) Show that
+, 4
Hint: Imitate the second proof of Theorem 1.1and use the fact that
A, and B, commute.
e) Let
and let Kr(A) = K(-).

Suppose that A E PNand also that
0 < a r llAllm5 b. Show for large n that (7.1.5) holds, namely,
there exists c so that
224
f) Show that this provides an O,,(log n) algorithm for the matrix

logarithm of A E PN and, by inversion, an O,,((log n)') iteration
for the matrix exponential.
For further details, and extensions beyond the positive definite case see
Stickel [85], where computational experience is also indicated.
7.2 THETA FUNCTION ALGORITHMS FOR LOG

We start with the fundamental identity of Chapter 2
and the series expansion
We recall that
The algorithm for log ( l / q ) is now:

Algorithm 7.3
Fix q E ( a , b ) w i t h O < a < b < l .

Step 1 : Calculate K(k) 1%-from (7.2.2).
Step 2 : Calculate k from (7.2.3).
Step 3 : Calculate K(kl) using the AGM commencing with 1 and k.
Step 4 : Calculate log ( l l q ) = %-[K(kl)lK(k)].
This algorithm has operational complexity O,,(fi)
O A f i Wn))
and bit complexity
The algorithm's complexity is determined by the series expansions employed in steps 1 and 2. These "sparse" series yield n-digit accuracy after
.\/iinonzero terms. While the asymptotic complexity is far from optimal, the
algorithm has the advantage of not requiring very small starting values for
the AGM iteration (Step 3). Also, only a single AGM iteration is required.
Sasaki and Kanada [82], who proposed and analyzed the above algorithm,
show that it out performs the methods of Section 7.1 for numbers in the
7.2 Theta Function Algorithms for Log
225
3000-digit range. Note that rr must be precomputed. As Sasaki and Kanada

observe, the algorithm may be accelerated by using
m log
(i)
= log
($)
for various m. This speeds up the series calculation at the expense of the
AGM. For n-digit precision, using m = n effectively reduces this algorithm
to Algorithm 7.1. (See Exercise 1.)
For certain choices of q we get reduction in complexity. If q is any small
integer, then the series expansions are particularly easy to evaluate. (See
Exercise 2.) This leads to a very fast algorithm for rrllog 10 (using base 10
arithmetic) as follows. From (7.2.1), (7.2.2), and (7.2.3) we have
Now for q := 1/ lo4 both of the above series are just sequences of 0's and 1's
and the starting values for the mean iteration can be calculated very quickly
[O,(M(n))]. The remainder of the work involves calculating a single AGM.
Similarly rrllog p is amenable to very fast computation in base p.
Properly interpreted, (7.2.6) remains valid for matrices and provides a
theta-based computation of the matrix logarithm of a positive definite
matrix. (See Exercise 3.)

Further discussion of material in this section may be found in Sasaki and
Kanada [82] and in Bonvein and Bonvein [84d]. Sasaki and Kanada
compare Algorithm 7.3 to algorithms for log based on Taylor series for
log (1 + x) and log [(I + x) l(1- x)] and conclude that for more than (roughly) 100 decimal digits the Taylor series methods are slower.
1. Show that (7.2.4) for various m can be combined with Algorithm 7.3 to
2.
provide algorithms for log of any complexity between O,,(fi) and

O,,(log n).
DISCUSS
the bit complexity of calculating 02(q), 8,( q), and 0,( q), where
q is the reciprocal of a fixed integer. Show, for p integral, that (7.2.6)
can be used to calculate rr/log p with bit complexity O,(log n M(n)).
(See Exercise 9 of Section 6.4.)
226
3.
Consider the matrix AGM of Exercise 6 of the previous section.

Establish a matrix version of (7.2.1) to (7.2.6). Then construct an
algorithm for the matrix log based on (7.2.6).
7.3 THE COMPLEXITY OF ELEMENTARY AND ELLIPTIC

FUNCTIONS
The algorithms of Section 7.1 can be inverted by Newton's method to

provide algorithms for exp with bit complexity O,(log n M(n)) and operational complexity O,,((log n)'). From a bit complexity point of view, this is
the best known bound. Since we can invert log in { l z - 11I$}, we can
produce an algorithm for exp in a complex neighbourhood of zero and
hence have OB(logn M(n)) algorithms for all the trigonometric functions.
Exercises 1 and 2 give some variations. In fact, for any elementary function
we have the following:
Theorem 7.3
Any elementary function f over Q(x) can be calculated uniformly (in

bounded regions where f is single valued and analytic) with bit complexity
and with operational complexity
where s is a constant depending only on f,
Proof. For our purposes the elementary functions are the rational
functions, log and exp, and any function that can be formed from these
functions by a finite number of compositions, multiplications, additions, and
solutions of algebraic equations. (See Davenport [81] or Ritt [48].) The
point of the proof is that such an f is constructed from Q(x) by taking a finite
number of exponentials, logarithms, and solutions of algebraic equations in
these quantities. The number of algebraic equations to be solved determines
the constant s. As in Chapter 6, solution of the algebraic equation in
question can be effected in a time proportional to the complexity of
evaluating the equation.
A number of comments are in order. First, we can formulate the above
theorem for f algebraic over R(x) if we assume that the requisite real
numbers are given. Second, while multiple solutions of an algebraic equation pose no theoretical problem, in practice, determining the "correct root"
7.3 Elementary and Elliptic Functions
can be a major nuisance. For a multiple-valued function the theorem must

be interpreted as guaranteeing some value of the function. This is inevitable.
It is not even clear what it should mean to compute an infinite-valued
function.
That the operational complexity behaves like (log n)Vnstead of (log n)
reflects in part that operational complexity is an inappropriate measure
when Newton's method is involved. While the preceding algorithms for log
require most of the operations to be done to full precision, this is no longer
the case for this approach to other transcendental elementary functions.
Since we can calculate elliptic integrals with bit complexity
OB(logn M(n)) (Exercise 5 of Section 1.4 and Exercise 2 of this section), we
can calculate the Jacobian elliptic functions with similar dispatch.
It is not clear which other nonelementary transcendental functions have
bit complexity OB(logn M(n)). Does, for example, the gamma function?
Nor is it clear whether the bit complexity OB(log n M(n)) is best possible for
the nonalgebraic elementary functions. The best known lower bound for log
and exp is the virtually trivial bound of OB(M(n)). (See Exercise 3.)
The algorithms for exp (see also Exercises 1and 2) require inversion and are
much less satisfactory than the algorithms for log. A direct OB((logn ) 2 ~ ( n ) )
algorithm is presented in Chapter 10. There are a number of issues that
remain unresolved in this discussion. The most obvious is a discussion of
lower bounds. Observe that, by Theorem 6.4, showing that exp does not
have complexity OB(M(n)) would show that exp is a transcendental function. Likewise, showing that T does not have bit complexity OB(M(n))
would imply the transcendence of T.The gap between the known operational complexities for log [O,,(log n)] and exp [O,,((log n)')] is probably specious, but this also is not known. We will show in Section 8.8 that direct
algorithms for exp and log of the type that we derived for K in Chapter 1
cannot exist. There are no quadratically convergent fixed iterations for these
functions.
1. From (2.3.7) and Exercise 3 of Section 2.5,
2.
where a, and c, are generated from the AGM commencing with a, := 1

and b, := k', while a,* is generated from the AGM commencing with
a: := 1 and b: := k. Show that this leads to an OB(logn M(n)) algorithm for e", which begins by solving for a, la,* = ~ ~ 1 2 .
( A more direct approach to tan) As in Exercise 5 of Section 1.4 and
Theorem 2.6 of Section 2.6, we have the Landen transform for
228
1- k:,
k,+l := 1+ k:,
and
tan(c#+,+,-&)==k,',tan&
a) Show that
F(+o, k) = c$o
+ 0(k2)
k -+ 0 .
as
b) Show that
+ O(1-
F(+o, k) = log tan
k)
as
c) Show that if w, := tan (6,, then
[E(?)I
log tan
1+ k:,
[ n (T)]
:(
t)+
n-1
tan-' w, =
m=O
[E(T)]+
1 k:,
log tan
~ ( k :+ (1 - ko))
($ + tan;
1% 6 +
w ~ )
w',+ (1 - ko))
7.3 Elementary and Elliptic Functions
229
d m+
where 6 :=
w,. Show, by inverting c), that tan-' can be
calculated from log with complexity O,(log n M(n)). Show that, by
inverting the above, tan can be calculated from log with complexity
O,(log n M(nN
This approach, which avoids using complex arithmetic to access
the trigonometric functions, is essentially due to Brent [76a].
Let L(n) denote the bit complexity for evaluating n digits of log. Show
that
D(n>= O,(L(n))
and hence, log and exp are at least as complex as multiplication.
Hint: Consider computing the derivative of log.
Chapter Eight

Abstract. In Section 8.1 we dejine abstract means and discuss their behavior.
In Section 8.2 we discuss equivalent means. In the next three sections we
consider general mean iterations and examine their convergence properties.
Later sections concern Taylor expansions of means, multidimensional means,
and related questions. The jinal section considers algebraic mean iterations
and the possibility of extracting elementary limits from such iterations.
8.1 ABSTRACT MEANS
There is a large literature on means but little agreement as to what

exactly constitutes a mean. For our purposes we have
Definition 8.1
(a) A mean is a continuous real-valued function M of two strictly positive

real variables a and b such that
for all a > 0 and b > 0. We denote a A b := inin (a, b ) and

a v b := max (a, b). (Continuity is not always essential. On occasion
we will refer to possibly discontinuous functions satisfying the above
definition as discontinuous means.)
( b ) A mean is strict if, in addition, it is diagonal:
if and only if a = b.
8.1 Abstract Means
A mean is homogeneous if
(8.1.3)
M(Aa, Ab) = AM(a, b)
for a, b, A>0.
) A mean is symmetric if
(8.1.4)
M(a, b) = M(b, a)
for a, b > 0.
A mean is (strictly) isotone if, for a, b > 0,
(8.1.5)
M(a, .)
and
M(., b)
are (strictly) increasing .
We will find it convenient to consider the trace tMof a mean M given by

(x) := M(x, 1). We gather up some useful properties of means whose
proofs are left as Exercise 1.
Proposition 8.1
(a) Every diagonal continuous (strictly), isotone mapping is a (strict) mean.
(b) Suppose that M is symmetric and homogeneous. Then M is isotone if
and only if its trace t, is isotone.
The
isotone, (symmetric), (strict), (homogeneous) means form a con(c)
vex set.
( d ) The symmetric, (homogeneous) means form a uniformly closed convex
set.
Let
M and N be (strict) means. Then any continuous mapping P such
(e)
that
is a (strict) mean.
Corresponding to each mean M we associate another mean M,, defined
by
Then M = (M-,)-, and M-, is strict, symmetric, homogeneous, or

isotone whenever M is, and tM-Jx) = t,'(llx). This is a special way of
building an equivalent mean as we now discuss.
For any strictly monotone (increasing or decreasing) continuous function
f : [w+ -,[w+ we define the mean
232
When f(x) := xp (we write f := L ' ) we denote Mf by M, [consistently with

(8.1.6)]. It is easy to check that Mf is a strict, symmetric, or isotone mean
whenever M is. (See Exercise 2.) To give our discussion some flesh, we
introduce four of the most useful classes of means. In this chapter we
reserve the letters M and N for general means.
The Holder Means
or
E R X let
Then H, is just the arithmetic mean A , and thus H, is a strict, homogeneous, symmetric, isotone mean. Moreover, limp,, H,(a, b) = flis the
geometric mean G and may reasonably be denoted by Ho. (See Exercise
14.) Since H-, IHo IH,, Proposition 8.l(e) shows H, to be a strict mean.
For all p one can unambiguously define H, for a, b 2 0. Then the trace of H,
satisfies
t H p + )=
Note also that (H,)
-,= H-,
[2- , I p , 00)
[O, 2-lIP)
p r o
P<O
and that H, = A,.
Another useful way of building means is based on the next proposition.

Proposition 8.2
Let M be an isotone, homogeneous mean. Then, for p E R,
defines another homogeneous mean, ,M, which is strict of symmetric

whenever M is strictly isotone or symmetric.
Proof.
This is left as Exercise 4a).
The Lehmer Means
For p E R we let
Since L, =,(HI), each L, is a symmetric, homogeneous, strict mean. Since

tLp(x)= (xP + 1) /(xP-I + I), Lp is isotone only for 0 c p 5 1. Moreover,
L, = H,, L,,, = H,, Lo = H-, , and (by Exercise 5 of Section 8.6) these are
e. 1
r n U Y I . LICI
I..=L...U
"dd
the onlv means common to the Lehmer and Holder classes. Again
- there is
difficulty extending Lp to a, b r 0 and
{I
and
tLp(a)= 2
here, here and below, we write f ( ~ for

) lim,,,
P>O
;::
f(x). To see this, use
= (l-,)(Ll)(a,
b) = L(,-,)@, b).
The Gini Means
. Consider f = L~
where q := s - r. Then
[=I
as
8.1.14)
f -l~,,,-,(f(a),
f(b)) =
bs
I/("-r)
=: G,,,(a, b)
n G,,,(a, b). (See Gini [38].) Moreover (8.1.14) shows

hat G,,, is indeed a strict, homogeneous mean.
ous strictly monotone function of a nonnegative variable.
r(a, b) := f-I[
S.bf(.)d~]
b -a
extends to a symmetric, (generally nonhomogeneous), strict, continuous

Proof. The integral mean value theorem gives the conclusion, bar the
continuity. This follows from the continuity of the definite integral.
We immediately apply this to Stolarsky's power means (Stolarsky

[75,801).
Stolarsky ' s Means
MI,p-l by Sp. Then Sp is a homogeneous, symmetric,
234
(8.1.16i)
with
(8.1.16ii)
&(a, b) = lim Sp(a, b) =

P+O
b-a
=: 2 ( a , b).
log b - log a
and
(8.1.16iii)
The mean 2
' is the logarithmic mean and 4 is the identric mean. These
means have
{ i-I
-l/(p-1)
t s p >=
p>0, p f 1
p=l
p r o
and
tsp(a) = a .
Also
so that Sp(a, b) is increasing in (a, b) for all p in R. This, in part, follows

from the inequality (1- p)xP + pxP-l r 1 for 0 s p I1.
All of the means in these classes are piecewise monotone in the sense that
iM(x) has only finitely many sign changes. Thus there is no difficulty in
defining M(0, b) = bM(0,l) = b limxl0 tM(x) in all these homogeneous
cases, and we can freely consider M defined on &:=
!i{(x,
f' Y ) I2
~ 0, y r 0)
and at a. We will do so from now on. Note also that when M is continuous,
tM([O,a ] ) = [0, a ] is equivalent to tM(0)= 0 and tM(m)= a .
An elementary but very useful proposition is next. The proof is left for
Exercise 11.
Proposition 8.4
(Composition)
If M is defined by
where Mi, i = 0,1,2, are means, then M is a mean. If two of M,, MI, M, are
strict, so is M. If all three are homogeneous, symmetric, or isotone, then so
is M.
8.1 Abstract Means
i-
235
For example, M v N and M A N are strict means whenever M and N are.

So is any mean between them, in the sense of Proposition 8.1 (e).
There are many highly pathological means, as Exercise 12 shows. This is
rticularly so in the absence of continuity. For future reference we will say
at a homogeneous mean is ultimately monotone if M(x, 1) and M(l, x ) are
notone in some neighbourhood of zero and a. We write M 5 N if
a , b) 5 N(a, b) for all a, b > 0.
There is a great literature on particular means and very little on means in

general. Much classical information can be found in Hardy, Littlewood, and
Polya [59] and in the other references scattered throughout the chapter.
Prove Proposition 8.1.
Establish that for any mean M and any strictly monotone f : R + - + R+,
Mf(a, b) := f - ' ~ ( f ( a ) , f(b)) defines a mean which is strict, symmetric, or isotone whenever M is. Moreover, iff is bP, p E Rx, then Mf is
homogeneous whenever M is.
Show that Hp(a, b) is a continuous increasing function of p with
limp,, Hp = v and limp,-, Hp = A .
b) Establish the assertions about the Holder means.
c) If M is homogeneous and symmetric, then M, = G, whenever M,
exists as a limit.
a)
a) Prove Proposition 8.2.

b) Establish the assertions about the Lehmer means.
(Zsotonicity of M,) Let @(a):= a log a for a > 0, and let M be a
differentiable mean.
a) Show that Mp(a, b) is isotone in p > 0 if and only if
( 8 . )
@(a) dM(a' b,
da
+ @(b) dM2pd b,
2 @(M(a, b))
b) Suppose that M is also homogeneous. Then
Use a), b), and the convexity of @ to show that Hp is increasing for
p in [W.
d) If M is homogeneous and symmetric and if Mp is increasing for
p > 0, then Mp is increasing for p in R, whenever M,, exists.
c)
236
6. Suppose Mp is isotone in p. Then

i) , M r M P
ii) , M 5 M p
iii) pM 5 Mp-l
p r l
psl
p50.
7. a) Suppose M is symmetric and homogeneous with tM(0)> 0 or with

tM(a)< a , then
lim M , ( a , b ) = a ~ b
p-+ -m
and
limMp(a,b)=avb.
P-=
b) In particular, this holds for Hp.

c) If, in addition, Mp is isotone in p , then
lim ,M(a, b) = a
p+-m
and
lim M(a, b) = a v b
p-tm P
d) In particular, this holds for Lp.

8. a) Show that ,M is isotone in p if g(p) := log M(aP, bP) is always
convex, since then g ( p + 1) - g ( p ) increases with p. (See Beckenbach [SO].)
b) Use ,Cauchy's inequality to show that this holds for
9. a) Establish the assertions about Stolarsky's means.

b) Show that
c) Show that
and deduce that
10. a) Show that SP(a, b) is isotone in p by applying Exercise 5a) and

observing that (8.1.17) becomes $(a, b) r 2 ( a , b). [This in turn
follows by calculus from ( t - 1)2r t log2 t, t > 1.1
8.1 Abstract Means
237
Show that $,(a, b) is isotone in p This can be done by showing

that (8.1.17) becomes
which again reduces to ( t - 1)22 t log2 t.

The generalized mean E , , is defined by
and extended appropriately on .the boundary (as in Leach and

Scholander [78]). Show that each E , , = (S,), for some p and q.
Let h(r) := log /(ar- 1) lrl so that i ( r ) = log $,(a, 1). By b) h(r)
is convex. Thus for s > r,
log E,,,(a, 1) =
S-r
2 log
$,(a, 1)
and E,,, increases in r and s because E,,, = 9,.

The mean
is very classical and is called Heronian mean.

Show that if - l s p 5 $ o r p r 2 , then
with the inequality reversed when p 5 -1 or 5 p 5 2. Moreover
(8.1.19) fails if ( p + 1 ) / 3 is replaced by any smaller number
(larger in the reversed case). (See Stolarsky [80] for details.)
Show that 2, and 9,tend to v as p tends to infinity.
Use the condition of Exercise 8a) to show that ,2 is isotone in p.
Show that E,,,-, = ,2 r 2, for p r 1 and so E,,, 2 2, for r > s >
0.
Show that
11. Prove Proposition 8.4 on the composition of means.
12. Let q be a nonnegative function satisfying

i) q ( l ) = 1
ii) l ~ x < q ( x ) < x v1
xZ1.
238
a) Then
and
are (possibly discontinuous), homogeneous, symmetric, strict

means.
b) Suppose in addition that
iii) q(x) = xq
(2.
-
Then
M(1, x) = q(x) = t,&)
Hence if we take an arbitrary (even analytic) function satisfying iii)

with 1< q(x) < x for x > 1, there is a strict, homogeneous mean
with t, = q.
'
c) Let q(x):=
x > 1, x rational
otherwise
Then M is a densely discontinuous mean which is not ultimately

monotone. Moreover M lies between two continuous means.
13. Let Pn(a, b):=
(icka
k=O
c,=-O,
&,=I.
k=O
Then P,, is a homogeneous strict mean which is symmetric if and only if

ck = C n - k 14. If M is a continuously differentiable mean then Mo := limp,o Mp exists
and is given by
where q := (dMlda)(l, 1).

Hint: Use L'HBpital's rule.
8.2 Equivalence of Means
239
8.2 EQUNALENCE OF MEANS
Let @: R+ + R+ be continuous. We say that a mean M dominates a mean N
M ( @ ( a ) , @ ( b ) ) = @ ( N ( a , b ) ) a,b>O
and we write M >, N or M > N. If M and N dominate each other, we call
M and N equivalent. Since domination is transitive, this is an equivalence
relation which we write -. Unfortunately most of our results demand that
we consider more restrictive notions of equivalence, so we require that @ be
one to one from now on. ~ e n c @
e is monotone.
Theorem 8.1
Suppose that M and N are two means with M >, N .
(a) Suppose that N is homogeneous. For each t > 0 , consider

g,(x) := @(t@-'(x)). Then g, is isotone and
(8.2.1)
M a , b)=g ; l ~ ( g t ( a ) ,
for each a, b in rng (@).

( b ) Suppose that M : = A . Then
@=mP+p
(c)
t
t
:
J
cu>O,p#O
and N = H,, p # O .
Suppose that M := G. Then either
and N = H,, P E R .
It
follows that the only homogeneous means equivalent to A are the
(d)
Holder p means ( H p ,p # 0 ) and the only homogeneous mean equivalent to G is G.
Proof
( a ) By homogeneity of N we have, for t > 0 ,
240
and (8.2.1) follows. Moreover g, is isotone as a composition of two

co-monotone functions.
( b ) Now (8.2.1), with M := A , becomes
(8.2.2)
= g,(a)
2g(?)
+ g,(b)
a, b E rng (@)
This is solved [Exercise la)] by
(8.2.3)
g,(x) = a(t)x + b(t)
x E rng (a)
for some a(t) and b(t) in R. Thus
Now Exercise l b ) shows that this is solved by
(c)
Since Q, is positive, we must have Q, = alp.+p , a > 0, and N is as

claimed.
In this case we have
Let h , := log ( g ,o exp). Then
As above h,(x) = a(t)x
+ b(t) and
for B ( t ) positive. This is solved by
[as in Exercise lc)]. If Q, is of the first type, N = G; while if Q, is of the

second type, N = H,, ,p # 0.
(d) Thus all H,, p # 0, are equivalent and G dominates them all.
A more general program may be undertaken, based on Theorem 8.l(a).

Unfortunately, without extra hypotheses, solutions of (8.2.1) are very hard
to characterize. The following simple result is accessible.
8.2 Equivalence of Means
241
Proposition 8.5
Suppose M is a homogeneous, strict mean with M >, N for some isotone h,

and suppose t,(O) = 0. Then tM(0)= 0.
Proof. Since h(N(a, 1)) = M(h(a), h(l)), we have h(0) = h(N(0,l)) =
M(h(O), h(1)). This is only possible if h(0) = 0, h(0) = h(l), or h(0) = m.
Since h is strictly increasing, only h(0) = 0 can occur. Thus h(0) = 0 =
M(0, h(1)). Since M is homogeneous, tM(0)= 0 as claimed. Cl
As an example we see that Hp>, AG, p > 0 (AG is the Gaussian AGM)
is impossible for h isotone, as is Lp >, AG with p r 1. These considerations
suggest that we can say more if h is required to be onto. We will write
M =: N if M >, N for some one-to-one mapping h of IWt onto IW+ [so that h
must be (strictly) isotone with h(0) = 0, h(m) = or (strictly) antitone with
h(0) = m, h(m) = 01. This is an equivalence relation stronger than -, and we
call it strong equivalence. We will only consider strong equivalence of
homogeneous means. If we let g, := h(th-l), g, is surjective and we observe
that (8.2.1) becomes
Since M is homogeneous, we may replace g, by h, := g,/g,(l). Then h,(l) =

1 and
In the next lemma we give several conditions for (8.2.8) to have only the
trivial solutions (normalized). Under these conditions we have g, = c(t) L and
Then Exercise lb) shows h = cup, a > 0, p # 0, and it follows that the only
homogeneous means strongly equivalent to M are Mp, p # 0. Thus we have
determined the strong equivalence class of M in the cases covered by the
next result.
Lemma 8.1
Suppose M is a homogeneous strict mean.
(a)
The following two conditions imply that (8.2.8) only has the trivial
solution when h,(l) = 1.
(i) M is ultimately monotone and
(ii) h,(ii) =. ii for some 2 # 1, ii > 0.
242
:
(b)
(i) (ai) holds if M is isotone (or piecewise monotone), while
(ii) (aii) holds if M is differentiable and (1) ,i has a unique positive

zero, a # 1 and h is differentiable; or (2) t,(R+) J R+.
Proof.
Given that i and 1 are distinct positive fixed points of h,, it follows
from (8.2.8) and continuity that {c > Olh,(c-) = c) is a closed interval
C. We show that s := sup C is a. If not we argue as follows.
For any c < s in C, M(c, s) < s. Hence M(c, s + E)< s for some
E > 0. But M(s, s + E) > S. Thus M(c, s + E) = s for some % < s, c in C,
E > O . Then for n = 1 , 2 , 3 , . . . ,
s=h:(s) = M(h:(c), h:(s
+ s)) = M(c, h:(s+
s)).
(Here hn :=hn-' 0 h denotes iterated composition.)

Let b, := c-'h:(s + E). Then, as b, # b, (because s + E ~ C )b,, is
a strictly monotone sequence, since h, is strictly isotone. If {b,) were
bounded above, the limit point would be a member of C larger than s.
Thus b, increases without bound and also t,(b,) = slc. This violates
monotonicity of M at infinity. The proof that inf C = 0 is left as
Exercise 2b).
We consider (ii). If h and M are differentiable, we have from (8.2.8)
Since h, is strictly increasing, we have tM(h,(6))= 0 and h,(i) = 5. If

on the other hand t, is not surjective, we argue as follows. Suppose
M(0, 1) > 0. Then
as h,(O) = 0, and M(0,l) < 1 is a second fixed point of h,. Finally if

M(co, 1) < m, we argue with M - , . 0
Condition (ai) holds for H,, L,, and S,. Condition (2) of (bii) holds for
H,, p .iC- 0; for L, , p 5 0 or p r 1; and for S, , p > 0; while condition (1)
holds for L,, p r 1. (See Exercise 3.) It is reasonable, at least for computational purposes, to require that equivalence be defined by differentiable
maps. If this is done, the results are more complete.
Theorem 8.1 can be found in Wimp [84] in slightly different form. This is
partly explained by the fact that exp[(logx + log y)/2] = G(x, y). For
Hardy, Littlewood, ,and Polya [59] or Wimp [84] this means that A >,,, G.
8.3 Compound Means
243
a more formal development this is problematical since means are only

)' and log is not always positive. Thus we rule out the second
Show, using continuity of g,, that (8.2.2) is solved as in (8.2.3).
(Proofs without continuity assumptions can be found in Wimp [84]
and elsewhere. More general results depend on the measurability of
gt-1
Let $(x) := @(x)- @(I).Then @ satisfies
*(tx> =
44x1 + ~ ( t )
for some c(t) in R. Then $(t) = c(t) and thus [a(x) - 11l$(x) =
[$(tx) - $(t) - $(x)]l$(t)$(x) is independent of x if $(x) # 0. Thus
$(xy) = c$(x)$(y) + $(x) + $( y). If c = 0, then $(x) = C log x,
while if c # 0, h(x) := c$(x) + 1 solves h(xy) = h(x) A( y) whose
.
(8.2.4) holds.
solution is L ~ Thus
Show that (8.2.6) has solutions only of form (8.2.7). Hint:Consider
h(x) := log [@(x)]- log [@(I)]. Then h(xt) = a(t)A(x) + h(t). Now
proceed much as in b).
By considering M-, show that if M >, N, M strict and homogeneous, h isotone with t,(m) < m , then t,(m) < w.
Complete the proof of Lemma 8.l(a).
Complete the proof of Lemma 8.l(b).
Verify the final claims of this section.
Calculate the strong equivalence classes within G,,, and E,,s,
If M is a symmetric, homogeneous, strict mean, then either tM(0)=
0 or tM(m)= 03.
Similarly, suppose then that M >, N for some homogeneous N with
h monotone. Show that 0 or OJ lies in rng(h).
8.3 COMPOUND MEANS
We now formalize the notion of a mean iteration. The Gaussian AGM and
Archimedes' method provide the central examples. Let M and N be any two
continuous means. Let a > 0 and b > 0 be given and consider the iteration
ao:=a
bo:=b
a,+, := M a n , b,)
b,+* := N a n , bn)
Under mild hypotheses, Theorem 8.2 shows that the iterates converge to
a common limit, which we call the compound of M and N and denote by
"
244
M @ N(a, b). We will also denote the entire iterative process by [M, N]. We
say that M is comparable to N if one of the following holds:
(a) M ( a , b ) z N ( a , b )
(b) M(a, b) 5 N(a, b)
for
for
a,b>O
a, b > 0
(c)
for
O<a<b
for
0 <b <a
M(a,b)sN(a,b)
and
N(a, b) 5 M(a, b)
When M and N are symmetric, c) cannot occur, and we say M and N are
comparable. In the nonsymmetric case c) can occur, and then M is comparable to N but not conversely. In the next result comparability is only needed
to establish monotonicity of the iterates. (See Exercise 1 and Theorem.8.8
of Section 8.7.)
Theorem 8.2
Let M and N be means with M comparable to N.
(a) Suppose that M or N is strict. Then .EM, N] converges and M @ N is a
mean which is strict if both M and N are.
(b) M @ N is a homogeneous, symmetric, or isotone if each of M and N is.
(c) M @ N is continuous, and the convergence is monotone and uniform
on compact subsets of {(a, b)la, b > 0).
Proof
(a) Suppose that a > b and that M r N. Then a, = M(a, b) r N(a, b) = b,.
Inductively suppose that a, r b,. Then
(8.3.2)
a, r M(a,, b,) = a,,, r b,,, = N(a,, b,) r b,
and {a,) decreases while {b,) increases. Since each bounds the other,
both sequences converge, say, to x and y, respectively. By continuity
we have x = M(x, y) and y = N(x, y). Since M or N is strict, x = y. If
a s b, we may have to exchange the roles of a, and b,. Thus M @ N
exists and satisfies
(8.3.3)
This finishes (a).
MANsM@N~MvN.
8.3 Compound Means
245
The sequences {a,) and {b,) are in fact built by repeated composition
of means. Thus a, = Mn(a, b) and bn = Nn(a, b) for means Mn and N, .
These means are symmetric, homogeneous, or isotone when M and N
are, and so is the limit mean M @ N.
.
M,
Let a, b > O and E > 0 be given. Pick n so that la, - b,l < ~ 1 2Now
and N, are continuous. So we may find S > O with IM,(a, b) M,(ar, br)l < & / 2 and INn(a, b) - N,(ar, br)l < ~ / if2 lar - a [ < S and
Ibr - bl < 6, (ar, br >O). Again assume a > b and M 2 N. Then
M @ N(al, b')
5 Mn(ar,b r ) 5 M,(a,
b)
+ -2E --= N,(a,
b) + E
and
Thus M @ N is continuous. Finally, Dini7s theorem shows that M,

and N, must actually converge uniformly on compact subsets, since
convergence is monotone. 17
The key observation about M @ N is the following 'invariance principle'
which we have already used repeatedly in Chapters 1 and 2 to show that
AG(1, k r ) = 7r/2K.
Theorem 8.3 (Invariance Principle)
@ N exists. Then M @ N is the unique, (continuous) mean
@(M(a, b), N(a, b)) = @(a, b)
Proof. Iteration of (8.3.4) shows that

lim @(an,bn) = @(a, b)
n-m
Thus
@(a, b) = @(M@ N(a, b), M @ N(a, b))
and since @(c, c) = c, Q, = M @ N.
246
General Means and iterations
Observe that we need not verify that @ is a mean, but only that
@(x, x) = x for x > 0 and that @ is a continuous solution of (8.3.4).
(a) Let M : = Hl and N : = H-,. Then H I @ H-, = G. Observe that

G(a, b) := G satisfies
Since fi= x, we must have Hl @ H-,
= G.
(b) Let M(a, b) := 9ab2/(a + 2b)2 and N(a, b) := (a + 2b) 13. Then
and again the invariance principle shows that M @ N(a, b) = a 113b213.

(c)
LetH,@H,=AG:[email protected],
( %)
A @ G ( a , b) = a A G 1, - =
an2K1(b/a)
@(a, b) :=
an2K1(bla)
as Theorem 1.1 (second proof) shows by the invariance principle.

We now distinguish two better structured classes of mean iterations.
Definition 8.2
Let M and N be symmetric means.
(a) Suppose M and N are comparable. Then we write
M@,N:=M@N
and
[M,N],:=[M,N].
We call these Gaussian mean iterations and call @, the Gaussian

product.
(b) Consider the iteration: a, := a > 0, b, := b > 0, and
a,,, := M a , , b,)
b,+l := N(a,+,, b,)
We denote the iteration by [M, N], and the limit by M @, N. We call

these Archimedean mean iterations and call @, the Archimedean
product.
The existence of M @, N is guaranteed by Theorem 8.2. For M @, N
existence comes from the next result.
8.3 Compound Means
Proposition 8.6
Let M and N be symmetric means. Suppose that M is strict. Then M @, N

exists and
where N*(a, b) := N(M(a, b), b). So by Theorem 8.3 M @, N is the unique

continuous mapping satisfying
with $(x, x) = x for x > 0.

In consequence M @, N is a continuous mean which is strict, homogeneous, or isotone whenever both M and N are.
Proof. This follows from Theorem 8.3 since M is comparable to N*

(See Exercise 4.) 0
The theorem makes no use of symmetry of M and N. However, this is
critical to our further analysis of convergence rates. Finally, suppose that we
are given a function M which satisfies (8.1.1) and (8.1.2) or (8.1.3), but only
for 0 < a < b. We may extend M to a symmetric mean M via
(8.3.5)
~ ( ab):=
, M(a
b,a v b).
Similarly we can extend a function defined only on 0 < b < a by using

M(a v b, a A b). The new mean is homogeneous or strict when M is. Thus
in some of our future iterations we will consider means only defined on the
45" sectors (Exercise 3 among others). Moreover, if we have two comparable
means on 0 < b < a, the extensions are comparable so that our convergence
results apply. Note also that in the definition of comparability we excluded
the possibility that M 5 N for 0 < b < a while N 5 M for 0 < a < b. In this
case the iterates oscillate, and combining two steps results in a comparable
iteration. We call such iterates partially comparable.
Our treatment is a synthesis and extension of that in Schoenberg [77,82],

Lehmer [71], and Foster and Phillips [84b] among others. The term
compound is due to Lehmer [71]. In general it is very hard to determine
M @ N, but easy to verify a limit once one has found it. Schoenberg [77]
gives a geometric proof of the limit of the AGM due to Jacobi.
1. a) Let M and N be strict means. Then [M, N] converges and M @ N is
a strict (continuous) mean.
248
2.
Hint: {a, v b,) and {a, A b,) are monotone sequences, and so
a, v b, decreases to x and a, A b, increases to y. One may suppose
x = (M v N)(x, y). Thus x = y by Proposition 8.4 and M @ N exists. Since M A N 5 M @ N r M v N, the limit is a strict mean.
b) M @ N is symmetric or homogeneous when M and N are.
a) Let M be a strict mean. Then
b) Show that
c)
where as before Mf(a, b) := f -'M( f(a), f(b)).

In particular a) and b) show that
d) '1f M and N are symmetric, then

M@N=N@M.
[More generally, see Exercise 6.1 Thus
M@,N=N@,M.
3. (Carlson's log) Define means M(a, b) := v a - (a + b) 12
N(a, b) := y'[(a + b)/2]. b = M(b, a). Show that
M@N(a, b ) =
\jr2 log (alb)
and
for a + b .
Hence [using Exercise 2b)],

44112 @ N112 = 9
Explicitly, if
._ a,
an+' '-
+a
and
2
b,,+ := b,
+ 2l k K
then the limit is

b-a
logb-loga
'
Note that this is neither a Gaussian nor -an Archimedean iteration.
8.4 Convergence Rates and Some Examples
249
4.
Prove Proposition 8.6 from Theorem 8.3. Note that in this setting
M I N * for O < a < b and M r N * for O < b < a .
5.
a) Show that given strict symmetric means M and N,

M@, N(b, N(a, b)) = N O , M(a, b ) .
b) Show that in the special case in which M = N the limit
M @, M =: @ is characterized by
@(a, b) = @(b, M(a, b))
c)
Show [using b) or otherwise] that
ii)
G@,G(a, b ) = a 113b213
iii) H P @ , H P ( a , b ) = (
a P -t2bP
'IP
pfO.
6 . For any function Q let &(a, b) := Q(b, a). Prove that for any means M
and N,
and
M @, N(b, N(b, a)) = N @, d ( a , b)
8.4
CONVERGENCE RATES AND SOME EXAMPLES
In this section we show that Gaussian iterations typically converge quadratically and Archimedean iterations sublinearly. We also give some more
examples of Gaussian and Archimedean iterations for which we can calculate
the limit.
(a) Let a, denote the area of a regular m -2"-gon inscribed in a unit circle.
Let b, denote the area of the circumscribed regular m .2"-gon. It is
and b,,, = 2a,+,b,l(a,+, + b,) while
easily verified that a,,, =
.
we geometrically vera, = $m sin (2nlm) and b, = m tan ( ~ l m )Thus
ify that a, and b, tend to T. This gives G O , H-,(a,, b,) = n or, on
using homogeneity and replacing 2 r l m by 0,
GO. H-,(sin 9,2 tan
(8.4.1)
(!))
=0
There is no need for m to be integral and (8.4.1) holds for all

0<0<rr.
(b) Consider now the same process but using circumscribed and inscribed
perimeters a, and b,. Now we have a,,, = 2anb,l(an + b,) and
b,,, = q awhile a, = 2 m tan(rr1m) and b, =2m sin(.n/m). Then
a, and b, tend to 2~ and
H-, @, G(tan 9, sin 0) = 6 .
(8.4.2)
This is pursued further in Exercises 1 and 2.
A complete analysis of the iteration of Schwab, Borchardt, Pfaff, and

Gauss is in Miel [83], The central observation is:
Theorem 8.4 (Schwab-Borchardt)
Proof. Schwab established the first case by geometric arguments which

Schoenberg [82] reproduces. Given the formulae, it is simpler to use the
invariance principle. Since a;+, - b f = (af - b i ) /4, this reduces to showing that
+,
(w)= arccos):(
a, < bn
(w)
arccosh):(
bn +
a, > bn
2 arccos
bn+l
(8.4.4)
2 arccosh
Since an+llb,+l= d(a,lb, t- 1) 12, this is just the half-angle formula for cos
or cosh. El
We now turn to study rates of convergence for Gaussian and Archimidean means.
Theorem 8.5
Let M and N both be continuously differentiable symmetric means and

suppose that at least one is strict.
251
(a) Consider the Archimedean iteration [M, N],. Then, if a , # b,,
(b) Consider the Gaussian iteration [M, N], for comparable M and N.
(i) Then, if a, # b,,
(8.4.6)
lim
n-+-
an,,
a,
- bn+l = 0 .
- bn
(ii) Suppose, in addition, that M and N are twice continuously

differentiable. Then, if a, # b,,
(8.4.7)
lim
Ian+l-
la,
',+I1
- b,I2
- IMll(s, s) -Nll(s, ')I

2
where s := M @, N(a, b).
Proof. Since M(c, c) = c for all c, we must have M,,(c, c) + M,,(c, c) =

1 for all c . (Here M i denotes the partial derivative with respect to the ith
variable.) Since M 'and N are symmetric, it follows that MXi(c,c) =
N,i(c, c) = $, i = 1,2.
(a) Let s := (M @, N)(a, b). The mean value theorem gives
and since bn+, = N(a,+,, b,),
Thus
and (a) follows.

(b)
(i) Similarly, (8.4.8) still holds, and also
Thus
ancl - b,+, = o(a, - s) + o(b, - s) = o(a, - b,)
(since s lies between a, and b,). This gives (b).
Given
one more derivative, we have
(ii)
and a similar formula for b,. Subtraction gives (ii). Here we have
[Exercise
used V'M(S, s)(a, - s, b, - s)' = M,,(s, s)(a, - b,)'.
7d).]
Considerably more can be said if M and N are two or three times
continuously differentiable. (See Foster and Phillips [84b].) In addition the
convergence in each case is uniform. For our purposes Theorem 8.5 suffices.
In particular, Archimedean iterations characteristically converge linearly
and Gaussian iterations super linearly (quadratically if twice differentiable).
In light of Jacobi-type methods in the theory of equations, this may seem
counterintuitive. Exercise 3c) shows that (8.4.6) may fail if M and N are not
differentiable. The theorem justifies our separating Gaussian and Archimedean iterations. While symmetry is central to our convergence arguments, it
is not always essential, all that is really needed in (a) and (b)(i) Theorem
8.5 is that the two means have the same gradients on the diagonal. (See
Exercise 11.) Also, (8.4.7) shows that better than quadratic convergence is
possible only if M,, = N,,. (See Exercise 7.) Additional information is given
in Exercise 11.
Finally, let us say that two iterations are equivalent ([M, N]
[M', N'])
M' and N
N'. Strong equivalence is similarly defined. Clearly
if M
equivalence of iterations implies equivalence of the limits, but not conversely. Indeed, if the mapping h is continuously bi-differentiable, the rates of
convergence must be the same. (See Exercise 8.)
-,
-,
-,

Some of the Archimedean considerations are discuss.ed again in Chapter 11.
If we take m := 6 in Example 8.2(b), we have a recursive version of
Archimedes' original method. (See also Edwards [79] and Phillips [81].)
1.
a) Show that (8.4.1) and (8.4.2) are consistent with the general
formula of Exercise 5a) of Section 8.3.
b) Show directly that in both (a) and (b) of Example 8.2 we have
$ (a, - b,). This illustrates why computation of any
a,,, - b,,
large number of digits of rr by this method is impractical.
,-
c)
d,,, -limd,
d, - lim d,
2.
a)
+ 2bn)/3 =: d,
Show that in both cases (a,
satisfies
Show that changes of variables cause no problems for Gaussian or

Archimedean iterations, that is,
i) (M@,N), = M f O g N f
ii) (M @, N)f = Mf O, Nf.
b) Show that (8.4.2) implies that
v1- x2
A @, G(x, 1)= arccos x
c)
Osx<l.
Hint: Let x := cos 8 in (8.4.2) and use aii).

Rederive Theorem 8.4 from b).
The next exercise shows a class of means which is closed under the
Gaussian product. Contrast this with the Gaussian product of Holder
means.
3. Let Q,(a, b) := t(a v b)
+ (1 - t)(a A b) for 0 5 t 5 1.
a) Show that Q, is a homogeneous, symmetric, isotone mean which is

strict for 0 < t < 1.
b) Show that for t r s,
This may easily be done by the invariance principle. Alternatively,

suppose a > b > 0 and observe that the limit must be linear.
C) Show that (a,-b,l=(t-s)"la-bl.
d) Show that Q,
Q, = Qt
Q,,.
o,
4.
o,
Let M and N be symmetric, homogeneous means.

a) Show that M @, N = G if and only if N = M - , .
b) Show that M 8,N = A if and only if N = 2A - M. Note that M is
a mean if and only if 2A - M is.
c) Characterize M @, N = H, .
d) Show that if M(a, b) :=
.\I-,
then
Explicitly,
a n + ,:= d a :
+ b i - anbn
and
bn+,
:=a
converges to d ( a t + bi) 12.

If M : = ( H l + H - , ) I 2 and N : = ( H l + L2)/2 then M @ , N = A .
Explicitly,
.- a: + bi + 6anbn
an+l * -
4(an + bn)
cbnverges to (a,
and
bn:=
3a:
+ 30; + 2anb,
4(an + bn) W,
+ b,) 12.
variant of the AGM) Show that
Explicitly, the common limit of

a b
bn+l:=L
an + bn
and
+-a
2
is 2AG(ao, bo)21(ao+ b,). (See P. B. Borwein [Pr].)

Show that M(a, b) := (a + b + 2ab) /(a + b + 2) is a strict symmetric nonhomogeneous mean with M = M-, .
Show that M @
M(a,
I,b) = (a + 2b + 3ab) l(b + 2a + 3).
Hint: If h(t) := tl(1 + t): then M = A,.
Use Theorem 8.5(b) to show that [H,, H,],, [H,, L,],, and
[L,, L,], converge quadratically, not better, for p # q.
(The quartic AGM) The iteration [A, N],, where N(a, b) :=
[(a3b + b3a)/2]114, converges with order 4. [See Exercise 3 of
Section 1.41
Reformulate the cubic modular equation as a third-order Gaussian
mean iteration to compute AG.
Show that for any symmetric twice differentiable mean,
Mll(C, c) = -M12(~, c).
Verify the assertions about equivalent iterations.
[H,. , H,,], if and only if p m q= qmp.(SupShow that [H,, H,],
pose that p and q are not both zero.)
Hint: One may suppose p = 0 or 1 and use Theorem 8.1.
Thus [HI, H,], and [A, GI, are not equivalent.
9. For O < k < l let M ( a , b ) : = ( l - l l k ) a + ( l l k ) b

ak-'bl[(l - 1lk)a + (llk)bIk-'.
1-llk
Ilk
and N(a,b):=
a) Show that M @ N(a, b) = a

b .
Show
directly
that
convergence
is quadratic. (Theorem 8.5 does
b)
not apply, but gradients do coincide on the diagonal.)
255
If M and N are homogeneous and L := M @ N, show that for x > 0,
Thus although we do not know an explicit formula for L := HI @, L,,

we have L(0, x) = x L ( 4 , l ) .
Suppose that M and N are comparable continuously differentiable,
homogen~ousmeans such that M @ N exists.
Show that if a, # b,,
' 4
8
lim lan+l- bn+ll = IM,1(s, $1 - N,&,

la, - b,l
n-
41
where s := M @ N(a, b).

b) What conditions, other than symmetry, in conjunction with part a)
now guarantee linear or superlinear convergence?
12. Suppose that M and N are continuously differentiable, homogeneous,

symmetric means. Let
a) Show that [M*, N*] still converges superlinearly.

b) For M : = A and N : = G [so that M * ( a , b ) = ( 3 a + b ) / 4 and
121 show directly that convergence is quadN*(a, b) = (a +
ratic.
a)
13. The ten neo-Pythagorean means are fl := A , f2 := G, f, := H, f, := L2

(also called the contraharmonic mean), and
where m : = a ~ b E ,: = a v b, and d : = E - m .
256
Show that f, to f,, are symmetric, homogeneous means which are

not differentiable on the diagonal. Thus Theorem 8.5(a) does not
apply.
b) Show that f7 @, f, = A , and that [f7, f,], converges sublinearly for
a < b but quartically for a > b.
c) Show that f, 8,f,($, 1) = 4, and that we have one-step termination-which cannot happen for strictly comparable means.
These means were originally defined by the Greeks in terms of
proportions. For j := 1 to 10, respectively, x := &(a, b) solves
a)
x-m
m
(1) =--====
(2)
m-x
m
x-m
m
(3) =-== =
(4)
m-x
m
x-rn
x
(5) =r-==
=(6)
m-x
112
E-m
m
(7) -z-==
=m-x
E
(8)
E-m
(9) - = m-x
m
x
x-rn
m
==
m-x
x
x-m
m
=-== = z
m-x
m
x-m
m
=--=
=m-x
x
m
m
m
--x-m
m
=--
(10)
m-m
x
m
-
d) Verify that x = &(a, b) in each case.

e) Verify that these are the only such means.
8.5
CARLSON'S INTEGRALS AND MORE EXAMPLES
This section is largely given to a description of a unified approach to Gauss's

and Borchardt's algorithms due to Carlson [71]. It shows both the possibilities and the limitations of looking for iterative methods based on a
more general hypergeometric transformation.
Let us consider the integral
where a + a ' = 6 + 6'; re(a), re(al) > 0, and p is the beta function. (The
prime is not complementation in this context.) Then obviously
(8.5.2)
~ ( a6,; 6';x2, y 2 ) = ~ ( a6 ;1 , s ; y2,x2)
and R is homogeneous of degree -a in x2 and y2. (See Exercise 1.) We are

interested in
8.5 Carlson's Integrals and More Examples
(8.5.3)
Cij:=Fi@q
i,j=l,2,3,4
where the means Fi, i = 1,2,3,4, are given by
Thus C12= AG, C3, produces Carlson's log (Exercise 3 of Section 8.3), and
C14= A @, G leads to Borchardt's algorithm (Theorem 8.4).
Theorem 8.6 (Carlson)
Let i, j = 1 , 2 , 3 , 4 with i .f j.
(a)
[Fi, q ] converges.
(b)
Cij(a,b)=[R(a;S,S';a2, b2)]- 1/ 2 n
(c)
where ( a ; S, 8 ' ) is given by the (i, j)th entry Table 8.1.

Convergence is linear except for C12= C2,, which is AG, the limit of
the Gaussian AGM.
TABLE 8.1
Since Fl and F2 are symmetric while F3(b, a) = F4(a, b), up to exchange

of 6 and 6' the table is symmetric around the main diagonal because of
(8.5.2). Boxes marked with * correspond to trivial iterations.
Proof.
(a) Since F4cc F2 5 Fl 5 F3 for 0 < b < a, each pair of means is partially
comparable. By Proposition 8.4 on composition, each mean is strict.
Thus Theorem 8.2 establishes (a).
(b) Let us denote Fi(a, b) by A. Make, in (8.5.1), the substitution
t : = s ( s + fi)l(s-I- f;). Then
and
Thus
If we fix i and j, we can determine values of the parameters for which

f,, k # i, j, vanishes in (8.5.4). For example, with i := 3 and j := 4, we
set a := a t : = 1 and have
where S + 6 ' =2. There is a unique value S := 1 so that (8.5.5)

becomes invariant for f3 and f,. By Exercise lb), [R(1; 1,1; a2, b2)]-1'2
is an invariant for [F,, F4] to which Theorem 8.3 applies. Similarly, for
i : = 1 and j : = 2 , we set 6 : = S 1 : = 5, which gives a + a l = land
The only possible invariant has a := a ' = f , and by Exercise lb),

[R($; 4 , f ;a2, b2)]-I is an invariant for [F,, F,]. The rest of the table
is similarly verified. [See Exercise lc).]
(c) The convergence assertions are straightforward. [See Exercise
Id).]
8.3. AS observed before, C,,, C,,, and C,, have been previously
identified. In Exercise 2 we indicate how to recover the previous forms of
the limit. In particular Theorem 8.6 gives integral representations for these
means. Now consider i := 1 and j := 3 (or similarly, j := 1 and i := 4).
Suppose that a > b. Let
EXAMPLE
Then s4= (a2 - b2)l(t + a2) and

4(1- s4)-I'2 ds = -(a2
Since a : = $, 6 : =
3,
and
a':=
+ b2)-112dt .
- b2)lI4(t
$, we have a r = l , P(a, a f ) = 4 , and
Recall that the arclemniscate sine is
and gives the arc length of the lemniscate (r2 = cos 28) from the origin to
the point with radial position x. (See also Theorem 1.7.) Thus
Similarly,
where the hyperbolic arclemniscate is defined by

arcslh x :=
IOx+
(1
s4)-112ds
[See Exercise 2d).] Observe that fiC;;l2(1,

1.7.
0) = K ( 1 / ~ ) by
, Theorem
8.4 (ARCHIMEDEAN
MEANS)Let us consider H, @, H, for p , q =
rtl, 0. Then the previous exercise and Exercise 5 of Section 8.3 show that it
suffices to consider Hl @,HI, Hl @, H,, and Hl @, H-,. Exercise 5c) of
Section 8.3 gives H, @, H,(a, b) = (a + 2b)/3 while H, @, Ho is Borchardt's
algorithm. It remains to study HI @, H-, . With a := /3 := $ , this is a special
case of the mean iteration a, := a > 0 and b, := b > 0,
EXAMPLE
a,,, := aa,
+ (1 - a)b,
bn+l:=
an+lbn
Pa,+, + (1 - P)bn
260
where 0 < a < 1 and 0 < p < 1. In the notation of Exercise 3 of Section 8.4
this computes G : = Q, @ , ( Q p ) - , for a > b. If we let k n := bnlan, we
derive
k-1
-1=(
nil
a p ) ( k i l - 1) = (ap)"+(: - I)
and
Thus
an+1
- -a
+ (1- a)kn
an
and the limit is

G(a, b ) = a f i
alb)
[ I1--a((aapp))"n((ll--alb)
which is the ratio of thetalike products occurring in the q-binomial theorem.

(See Exercise 7 in Section 9.4.) Wimp [84]continues a discussion of similar
calculations, all of which give linear convergence.
If we let a p := $ and a := 4, we have a product expansion for
Hl @, H-I . Foster and Phillips [84a]show how this function can be closely
approximated by elementary functions and can be given an elegant asymptotic expansion.
Many other related algorithms can be found in Wimp [84]. This includes an
extensive discussion of quadratically computable trigonometric integrals (in
Chapter 14).
1. a) Show
that
R ( a ; 6 , s ' ; x2, y2) = y - 2 a ~ ( aS,6
;
6 ' ; 1 - x21y2),
where F is the Gaussian hypergeometric series. (See Exercise 6 in
Section 1.3.)
2
b) Show that R ( a ; 6, 6 ' ; x2, y 2 ) is homogeneous of degree - a in x
and y2. Also R ( a ; S , S 1 ; 1,l)= 1 and R ( a ; 6, S t ; ., -) is continuous.
c) Verify the entries in Table 8.1.
261
d) Show that [Pi, 51, i < j, is linearly convergent for all cases except
i := 1, j := 2. Observe that [F,, F,], which is not Archimedean, has
a convergence rate of 2-" and [F2, F,], which is partially comparable, has oscillatory iterates.
2.
a) Observe that C12(a,b) = ( ~ / 2 ) l I ( ab)

, with
as before.
b) Show that C,,(a, b) :=
a' - bL
2 log (alb)
( a # b)
Hint: Use partial fractions.

Show that C14(a,b) is given by L8.4.3).
Hint: Let ( t + a2)l(t + b2) = : cos 6 or cosh2 6.
d) Show that C13(a,b) is given by (8.5.8) for a < b.
e) Use the invariance principle to verify that
c)
f) This completes the analysis of all of Carlson's means in explicit

form. Attempt to verify part e) directly from Theorem 8.6.
3. Use the invariance of C,, to show that
i) arcsl x = fiarcslh y
where (1 + y4)(l +
ii) arcsl x = 2 arcsl z
where x =: 2zI&7/(1+
m)2, and
=
z4) and z2 < fi- 1.
This is Jacobi's duplication formula for arcsl (Watson [33]).

4.
Show that with M(a, b) = G(a, Hp12(a,b)) =: N(b, a),
5.
( C d s o n [75]) Let a, b > 0 and set
262
a) Show that
ab] = 2F(a, b) .
F[(?)',
b) Using a) (or directly) show that
6. (Carlson [78]) Let a, b, c, A real be given. Assume a + A, b + A,

c + A r 0, and at most one of these is zero. Consider
T(A) := T(a, b, c; A) :=
IAw+
[(a
t)(b + t)(c + r)]-112dt .
Let
k : = h + \l(h +.a)(A + b) + \l(h
+ b)(A + c) + \l(h + a)(A + c) .
a) Show that T( A) = 2T(k).

b) Show that T( A,) = limn,, 2n + l kn-112 , where iteratively A := k,,
kn+l:= k, and A =: k,.
Show
that for a, b, c, d > 0,
c)
[(t + a2)(t + b2)(t + c2)(t + d2)]-'I2 dt = T(A, B, C; 0)
where A := (ab + cd)', B := (ac + bd)', and C := (ad
7. (Tricomi [65]) Let 0 5 k 5 1 be given. Set

R,(a, b) := \lk2a2 + kf2b2
k1:= \l=
and
a) Show that M is a strict mean on 0 < a < b.

b) Show that
where cn(v,) := a,lb, and 0 < v, < 2K.

Hint: Use the half-angle formula for cn
+ bc)'.
8.6 Series Expansions of Certain Means
and deduce that cn(vn) := anlb, satisfies v,,, = v,/2 and b,,,lbn =
va,+l/b,= an+,lbn+,= ~ n ( v ~ / 2 ~ + ' ) .
Recover
Borchardt's algorithm by considering k := 0.
c)
8. Compute a formula for the MacLaurin series for G(1- x, 1) in (8.5.9).
8.6 SERIES EXPANSIONS OF CERTAIN MEANS
For homogeneous means it is particularly easy to compute Taylor series. It
suffices to expand the trace around 1. In the exercises we list various series
taken from Gould and Mays [84], extending results in Lehmer [71].
Similarly,
From these it follows that the only means which are both Holder and
Lehmer means are H-,, A, and G. (See Exercise 5.) Similarly, but slightly
more elaborate, analysis shows that the following theorem holds.
Theorem 8.7
(Lehmer)
HpOgHq= Hs
(a)
if and only if p + q = s = 0.
(b)
Lp O , L q
if and only if p
(4
= Ls
+ q = 2 s = 0,1,2.
Lp O&,= Hs
if and o n l y i f p + q - 1 = s = - 1 , 0 , 1 .
(4
Hp 0, Lq = Ls
if and only if p = -q and s = $.
264
Proof. Sufficiency is easy in each case. Necessity is more elaborate and

in some cases somewhat tedious. (See Exercise 6.) In essence it follows from
the invariance principle and the computation of the first few terms of the
Taylor series of the limit.
Thus only A, G, and H can arise in such compounding. Lehmer continues
by studying in detail the Taylor series of A @ , L,, various properties of
which had been previously examined by Stieltjes in a letter to Hermite
(Hermite and Stieltjes [05, letter 3231). (See Exercise 7.)
Theorem 8.7, while very pretty, is of limited ambit, since it cannot diagnose
whether the compounded'mean is equivalent to a mean in the class.
1. Show that the Holder means satisfy
where
2.
for n L 1.
(Euler) If g(x) := C : = , anxnand gP(x) =: C : = , bnxn, then
3. Show that the Stolarsky means satisfy
where
1
D ( p , n):=
4.
a)
nP
N-l
C (-1)
k=o
P
n-k
Show that the Gini means satisfy
where
) kp+k-n
n-k+l
D(p7 k)
8.6 Series Expansions of Certain Means
b) Show that
m
C~
~ ~ , ~ ( 1x), :=
1 -
( sr,,n)xn
n=O
where
c) Thus G,,, can be computed by convolution. A similar result exists

for Er,*.
5. Show that H, = L, if and only if p = 0 and q = -1; p

p = 4 and q = 0 .
6. Prove Theorem 8.7.
7. (Lehmer) Let
=1
and q = 1; or
a) Show that
where the Pi are defined recursively by P o ( x ) : =1 and

Pn+,(x):= [pn(x)lZ+ 2x2".
Hint: If
an :=
pn
P I P 2 . .. Pn-,
ao:=l
bn :=
Pn+l+ 2x2"
P l P 2 . - Pn
b0:=1+4x
C,
..-- 2x2"
P I P 2 ..Pn
co := 2x
then
an + bn
an+, = 2
266
Now observe that
and use the invariance principle.

b) Show independently' that
where gl := 2 and g,, m r 2, satisfies the recursion
Observe that the gi are integers.

Hint: Show that the series for R satisfies
c) Use b) to compute g, for m 2 20. (g2, = 23335660.)

8.7 MULTIDIMENSIONAL MEANS AND ITERATIONS
Most of the results of this chapter have direct analogues for functions of
more than two variables. Often these follow by similar. arguments. We
concentrate on mean iterations. Let Z := (a,, . . . , a,) be any strictly positive vector.
An N-dimensional mean is any continuous function M such that
for all a,> 0. The mean is strict if
Also M is symmetric if M(Z) = ~
( 6 for
) any permutation b of Z.
Homogeneity and isotonicity are defined analogously. The Lehmer and
Holder means are defined by
8.7 Multidimensional Means and Iterations
and
(8.7.4)
respectively.
Given N N-dimensional means M', . . . , MN, we consider the iteration
[M',...,MN]defined b y i & : = i > O a n d
We write this vectorially as a,+,:= ~ ( i and

~ )denote the common limit
when it exists by @ E l Mi. Again, when each Mi is symmetric, we consider
N
this to be a Gaussian iteration with limit @, Mi. Similarly, if
(8.7.6)
i=l
i-1
:= M'(an+l,an+17 . . , a,+l, a,,
..
, a n)
we have an Archimedean iteration (there are other possible generalizations)

N.
with limit @, Mi. Under mild hypotheses the convergence results of Section
8.4 remaiiZ:alid.
Theorem 8.8
Let M1, M2, . . . , MN be strict N-dimensional means.

(a) Then @ E l Mi exists and is a strict, continuous mean.
(b) Suppose that the means are symmetric and continuously differentiable.
af,. Then, if a , # b n ,
Let a n : = VEl af, and b n : =
/\El
(8.7.7)
lim Ian+lF bn+lI = 0

la,-b,l
n j m
and convergence is superlinear in the Gaussian iteration.

If, in fact, the means are twice continuously differentiable, convergence in the Gaussian iteration is quadratic (uniformly on compact
subsets).
(d) Convergence is linear in the Archimedean iteration for continuously
differentiable symmetric means.
(c)
Proof.
(a) Much as before, a, r an+,r bn+,r b,, and we may suppose a, converges to a and bn converges to b. Let F be any cluster point of {%),
which is bounded. Then b 5 /\
ci 5 V Ni = l ci 5 a, and
El
268
Since all the means are strict, we must have ci = a and ci = b for each i.
Thus a = b, and the iteration converges, say, to &. As before, the limit
is a continuous strict mean.
(b) By symmetry we have MIk(ce)= 1 / N for any i, k and any multiple of
the unit vector e . Thus
i
a,,,
-a=
2 [(a* - a ) + o(a* - a ) ] . ,
k=l
Hence
- ai,+,l= o(a:
-a)
+ ~ ( a i-, a ) = ~ ( a -, b,)
and (8.7.7) follows.

This
is argued as in the two-variable case and relies on the fact that the
(c)
Hessian v2Mi(a, a , . . . , a ) sums to zero and has all diagonal entries
equal to -(N - l)y', where each off-diagonal entry equals yi. Thus for
each i and j,
ii
a,,,
- a'+, =
2 (ah - a:)2 + o((an - bn12)
h<k
(and unless all yi coincide, convergence will be at best quadratic). Now
O c a,,,
- b,+, 5
la, - b,I2 + o(la,
- b,I2)
where B is an easily computable constant depending on the means,

initial values, and dimension. This shows (c).
(d) We leave (d) as Exercise 2b).
EXAMPLE
8.5 (SCHLOMILCH)
Consider
C,+l:= (a,b,c,)
113
=: M3(a,, b,, c,)
This is a quadratically convergent Gaussian iteration, as follows from

Theorem 8.8(c). (Since M' r M2 2 M3, M2 is a strict mean.) Let S(a, b, c)
denote the limit. While we cannot identify S, we can, following Landau,
observe that if ac = b2, then a,c, = b: and thus
269
,,,,,,
where He(a, c) := (a VE
+ c) 13 = E ,(a, c) is the Heronian mean.
While this does not appear to have a closed form, it has an attractive
product expansion,
where xo := x and
While it is difficult to identify two-dimensional compound means, it

appears even harder to identify higher dimensional ones.
There is a satisfactory analogue of Proposition 8.3. Let f be continuous
and strictly monotone as before. Let a,, a,, . . . , a, be distinct positive real
numbers. Let F(,, be any Nth antiderivative for f. We define
Then M J N ,is a strict, symmetric N-dimensional mean. (See Exercise 5.)

We should emphasize that the invariance principle continues to hold:
@ := @ E l Mi is the unique continuous diagonal mapping satisfying
A wealth of information on N-dimensional means can be found in Hardy,

Littlewood, and Polya [59]. The Lehmer means are studied in Beckenbach
[50]. The Schlomilch mean and Landau's contribution are discussed in
Schoenberg [77]. Various N-dimensional quadratic iterations are exhibited in
Wimp [84] and Arazy et al. [85].
1. a) Show that Hp and Lp are strict, homogeneous, symmetric means.
b) Show that (IIL, ai)'lN = limp,, Hp(G) =: Ho(G).
c) Show that Hp is convex for p r 1 and that Lp is convex for
lsp52.
2. a)
b)
c)
d)
Show that @ E l Mi is a continuous strict mean when each Mi is.

Show that convergence is linear in the Archimedean iteration.
Estimate the rate in the three-variable case.
Show that
and
Hint:Use the invariance principle when p
:= 1.
3.
Show that S(1, x ) satisfies (8.7.9). Compare this to the product expansion for 2 ( l , x).
4.
a) Let
and
Show that
and that convergence is quadratic.

b) Let Sk be the kth elementary symmetric mean function of N variables. That is, Sk := f k l ( y ) with fk as in Exercise 6 of Section 11.2.
Thus, So := 1, S, := H I , etc. Let M k := S k / S k - , for 1 5 k 5 n. Show
%,
that M* 2 M*"
and that
&, M i = Ho.
i= 1
c)
In a) replace M 2 by H,,. Thus
Show that the limit mean M satisfies
This is not the general limit.
d) More generally, whenever M is homogeneous and symmetric,

(P = M @ G @ M-, satisfies @(a, f i , c) = fi. Similar results
hold in N dimensions.
5. a) Show that MI,, of (8.7.10) is uniquely defined and is a strict mean.

Hint: Let
Then PN is the Lagrange interpolating polynomial for F(,, at

a,, a,, . . . , a,. (See Exercise 1 of Section 10.1.) Thus
and f
must agree at some point strictly between A
a, and V
a,.
Now show that MINf extends continuously to all strictly positive
variables. The harder part is to show it is strict.
b) Show that the three-dimensional Stolarsky means are appropriately
defined (and are homogeneous) if
PT,)
El
El
for p f 0, 1, 2. Also
(a - b)(b - c)(c - a)
b' c, := 2(b - c) log a 2(c - a) log b + 2(a - b) log c
1"'
while
S1(a' b'
( a - b)(b - c)(a - c)
:=I2(b - c)a log a + 2(c - a) b log b + 2(a - b)c log c
and, using f(x) := log, that the identric mean may be given as
9 ( a , b, c):= S2(a, b, c) := exp[
a2 log a
(a-b)(a-c)
c2 log c
log b
+ ( b -b2
a)(b-c)
c)
Show that the N-dimensional logarithmic mean is given by

N
2,(a,, a,,
. . . , a,)
:= [(-I)N(N - 1)
c,=,
log a,
(a, - aj>
d) Investigate the isotonicity of Sp(a, b, c) as a function of p.
272
6. There are various multidimensional generalizations of the AGM due to

Borchardt [I8881 and others. In four variables one may take
a) Observe that, while the means are not all symmetric, the derivatives
on the diagonal (a = b = c = d ) all coincide, and hence establish
quadratic convergence.
b) Show that when a, := b, and c, := do, the iteration reduces to the
AGM. This iteration shares many of AGM's attributes (Arazy et al.
[85l)
7. (An extended convergence result)
a) Observe that the convergence result Theorem 8.8 continues to hold
if the condition that the means are strict is relaxed to
N
a(q= V ci
i=l
i=l
and @(Z) = A ci implies
i=l
ci =
ci .
i=l
Here
This clearly holds if N - 1 of the means are strict.

b) More interestingly, use a) to establish the following result. Let A be
an entry-positive N by N matrix. Set A, : = A and
for n 2 0. Here *F
represents entrywise square root. Then A,
converges to a constant matrix with entry e(A).
c)
If A :=
( t) then e(A)
( 1 1)then e(A)
= AG(a,
b).
= a(a, b ) .
d) If A :=
e) Similarly consider A, and B, entry positive and
8.8 Algebraic Iterations and Functional Relations
273
Show that this iteration converges to a constant matrix. (See Cohen

and Nussbaum [87].)
8.8 ALGEBRAIC ITERATIONS AND FUNCTIONAL RELATIONS
Among the more familiar and fundamental properties of the exponential
function is that it satisfies the algebraic functional relation
(8.8.1)
f(z>= [f(zln)ln .
Are there any other algebraic functional relations of the form
where /3 and p are algebraic? More precisely, p is an algebraic function if

there exists a polynomial P in two variables with coefficients in some base
field F (for most of this discussion F := Q or F := C ) such that
and R is an algebraic functional relation for f if R is a polynomial in three

variables over F and there exists an algebraic function p so that
The function p is termed an algebraic transformation (for f ) . We assume

throughout that z is a complex variable and unless otherwise indicated that f
is defined on a region U which contains an open set V so that p(V) C U.
[This ensures that (8.8.4) can be continued to any open connected component of the domain of f that contains U.] The collection of all such
transformations satisfying (8.8.4) for some R we denote by T G ( f : F), the
algebraic transformation group o f f over F. If
where M and N arf homogeneous means, then
where M(x) := M(l, x) and p(x) := N ( l , x)lM(l, x). Thus if both M and N
are algebraic, then L satisfies an algebraic functional relation with algebraic
274
transform p. It transpires that the transformation groups of the most

elementary transcendental functions are very simple, too simple in fact, to
support quadratically converging mean iterations. (See Exercise 1.) This, in
part, explains why the algorithms of Chapter 7 for exp and log require the
intermediate.use of nonelementary transcendental functions.
We wish now to compute the algebraic transformation groups of some
familiar functions. However, we need first to establish the transcendence
[over @(z)] of the functions exp, log, and exp(exp). This is considerably
easier than the arguments for the transcendence of numbers such as e. The
arguments are roughly the same for all the above functions. For exp it
proceeds as follows. Suppose exp satisfies an equation of the form
2a, exp (iz)
i=O
where the a i are rational functions, k is assumed minimal, and a, = 1. We

differentiate (8.8.7) to get
k
(8.8.8)
(ai + +ai)exp (iz) = 0 .
i=l
If k 2 2 and we divide (8.8.8) by exp (z), we obtain a lower order expression

than (8.8.7) and violate the minimality of k. To dispose of the k = 1 case we
must show that exp is not a rational function. Since exp is entire, if it were
rational it would in fact have to be polynomial, and if it were polynomial it
would have to have a finite Taylor expansion.
Theorem 8.9
(a) TG(exp: Q) = {az: a rational)
(b) TG(1og: Q) = {zb: b rational)
(c) TG(exp (exp): Q) = {z).
Proof. We will prove, by elementary methods, that

(a')
(b')
(c')
TG(exp: Q ) = {az + b: a rational, exp (b) and b algebraic)

TG(1og: Q) = {azb:log (a) and a algebraic, b rational)
TG(exp (exp): Q) = { z + b: exp (b) and b algebraic).
,
To see that (a), (b), (c) and (a'), (b'), (c') are the same is equivalent to
Lindemann's theorem which guarantees that exp (a) is transcendental for
any algebraic a # 0. (See Exercise 7 of Section 11.2.)
We first prove part (a'). Suppose that p is an algebraic transformation for
exp. Then there exists an expression of the form
(8.8.9)
C ai exp (miz) exp (nip(z)) = 0
i=O
275
where the ai are nonzero algebraic functions in z and the m , and n , are
integers. Equivalently,
k
2 ai exp ( m i z + ap ( z ) ) = 0
i=O
where we assume k is minimal, that is, we assume (8.8.10) contains a

minimum number of distinct nonzero exp ( m i z + n i p ( z ) ) terms. If we divide
(8.8.10) by its last term and differentiate, we get
This expression has one less term than (8.8.10) and contradicts the
minimality of (8.8.10) unless (8.8.11) contains no nonzero terms. This
implies that each term of (8.8.10) must be constant. Thus there exist m and
n integral and a nonzero algebraic function a ( z ) so that
(8.8.12)
a ( z ) exp ( m z + n p ( z ) ) = constant.
Since exp is transcendental, m z + n p ( z ) must be constant. Call this constant

b ; since p is algebraic over Q, b must be algebraic. Specializing (8.8.12) at
z = 0 shows that exp ( b ) must also be algebraic. Part ( a ' ) is completed by
observing that p ( z ) = ( m l n ) z + b is an algebraic transform of exp.
The proof of part ( b ' ) is similar. If p is an algebraic transform for log,
then there exists an expression of the form
.
k
(8.8.13)
C ai[log ( z ) l m i .[log ( p ( z ) ) l n i= 0
i=O
where the ai are algebraic functions and the m i and ni are nonnegative
integers. Let (8.8.13) be minimal in the sense that it has the smallest
maximal degree and contains the fewest distinct terms of maximal degree.
(The degree of a term is m i + ni.) Suppose the term corresonding to i = 0 is
such a maximal term. If we now divide (8.8.13) by a, and differentiate, we
are left with an expression containing fewer maximal terms, which contradicts minimality unless all the transcendental terms vanish under differentiation. In particular (8.8.13) must actually be of the form
(8.8.14)
b log ( 2 ) + c log ( ~ ( 2 )=) d
where b, c, and d are algebraic numbers. Thus
276
(8.8.15)
p(z) = exp ( d / c ) ~ - ~ ' "
and we see that p is algebraic exactly when exp (dlc) is algebraic and -blc
is rational. (Otherwise, by Exercise 2, z - ~ "is transcendental.) This finishes
(b').
The proof of (c') is again similar to (a'). There are, however, a few
additional wrinkles. Suppose now that p(z) is an algebraic transform for
exp (exp (2)). Consider minimal length sums of the form
(8.8.16)
2 ai exp [mi exp (z) - ni exp (&))I

i=O
=0
where the ai are algebraic functions of z, exp (z), and exp (p(z)). Dividing
and differentiating (8.8.16) leads, as in (a'), to a contradiciton unless
(8.8.17)
mi exp (z) - ni exp (p(z)) = constant
and the result follows. One of the wrinkles is that we must establish the
transcendence of exp (exp (2)) over the algebraic functions in z, exp (z), and
exp (p(z)). This is necessary for our minimality argument and can be done
analogously to proving that exp is transcendental. (See Exercise 2.) Cl
The transformations we are considering are algebraic over 0. This is a
natural choice for computationally related questions. The analogous results
for algebraic transformations over C are similar but easier because the
number theoretic details vanish. We have the following:
Theorem 8.10
(a) TG(exp: @) = {az + b: a E Q and b E @).
(b) TG(1og: C)= {azb: b E Q and a E C).
(c) TG(exp (exp) : C) = {z + b : exp (b) E Q) .
If f(z) = g(a(z)), where a(z) is an algebraic function, then
EXAMPLE 8.6 Let p be an algebraic function and suppose p(") is the

identity. [For example, p(z) := rz, where r is an nth root of unity.] Then
is invariant under p , where p(") := p(p("-')). Note that TG(f) is not in

general conjugate to TG(exp) since TG(exp) contains no finite elements of
order greater than 2.
8.8. Algebraic Iterations and Functional Relations
277
' and log, (z), where q is an algebraic number distinct from 0 and 1,
TG(qz :Q) = {az
+ b: a and b rational)
TG(log, :Q) = {azb:a and b rational)
1
F
To derive these results we first observe that (a') and (b') of Theorem 8.9
hold for the above functions (with exp replaced by qz and log replaced by
log,). The only difficult part is to show that qa and a are simultaneously
algebraic exactly when a is rational. This is the celebrated GelfondSchneider theorem. (See Section 11.2.)
z) and
In Exercises 1 and 5 we show that functions like /3(eY("),
/.?(log y(z), z ) , where P(., .) and y(.) are algebraic, cannot support functional equations that possess quadratic fixed points in their domain of analyticity. In particular such functions cannot be the limit of quadratically convergent homogeneous mean iterations. This shows why the elementary
functions are always associated with linearly convergent iterations.
The results of the section underscore the difficulties of analyzing in closed

form the limits of mean iterations. The familiar elementary transcendental
functions can only be limits of fairly trivial iterations. The arguments are
extended in the exercises to cover sin, tan, arccos, and so on. Exercise 6
treats the special case of compounding rational means and shows that the
limit of such a mean iteration is either the kth root of a rational function or
is transcendental.
1. a)
Suppose that
as in (8.8.5), where M ( l , x ) and N ( l , x) are algebraic functions

analytic in a neighbourhood of 1. Show that if L can be quadratically computed by iterating (8.8.6), then p has a fixed point at 1of the
form p(x) := 1+ O(1- x)' as X + 1.
b) Call an algebraic transformation p quadratically attractive at c if
p(c) = c + O(c - 2)' as z-+ c. Suppose that A is algebraically conjugate to p. That is,
278
c) Show that none of exp, log, or exp (exp) nor any function algebraically conjugate to the above functions is the limit of a quadratically
converging algebraic homogeneous mean iteration.
2.
a) Prove that za ( a irrational), log, and sin are transcendental functions over C(z).
b) Show that exp (exp) is transcendental over C(z, exp (z), exp (p(z)))
when p is algebraic.
3. Prove (8.8.20) and (8.8.21) assuming the Gelfond-Schneider theorem.

4.
(On the algebraic transformations of sin, cos, and exponential sums)

Let a(z) be an algebraic function over F := C(z, ez), the field of
rational functions in z and ez. Suppose that a(z) is not algebraic over
C(z). Show that
and calculate TG(cos :C) and TG(sin :C). Outline:

Show that if p E TG(a :C), then there exists a nontrivial relation
d
C bj(z) exp [kjp(z) + hjz]
j=O
with bj E C(z) and k,, hj E H.

Consider minimal (in d ) expressions of the above type and argue,
as in the proof of the transcendence of exp, that for some j,
exp [kjp(z) + hjz] E C(z).
Suppose that a(z) is algebraic over C(z, eg(')) where g(z) is algebraic over C(z), and suppose that a(z) is not algebraic over C(z).
Extend the arguments of Exercises 4 and 1 to show that TG(a, C)
contains no elements that are quadratic at any point where g is
analytic. In particular, such a function cannot be the limit of a
quadratically converging algebraic homogeneous mean iteration.
Suppose that a(z) is algebraic over C(z, log (g(z))) where g(z) is
algebraic over C(z) and suppose that a(z) is not algebraic over
C(z). Show that TG(a : C) contains no elements that are quadratic
at any point where g is analytic and nonzero.
6 . (On the iteration of rational means)
a) Let M and N be homogeneous rational means. (A rational mean is a

mean that is a rational function.) Show that M @ N(l,
. x)
. is convergent in some neighbourhood of 1. Show that either
1) M @ N(1, x) is transcendental
or
2) [M @ N(l, x)]' is a rational function for some integer i.
279
Outline: Suppose f(x, y) = M @ N(x, y) and suppose that f(1, x) is

algebraic. Then one has a finite sum
1C s,(x)[f(l, x)ll = 0,
si
Eq
x)
and
(8.8.22)
2 'i(x7
Y)[~(x,y)I1= 0
where ri(x, y) = x-'s,(ylx). Thus, on substituting,

(8.8.23)
C ri(M7N)[f(M, N)]' = 0
while by inyariance,
Consider a minimal expression of type (8.8.22) and deduce from

(8.8.23) and (8.8.24) that, for some i ,
Observe that
and deduce by passing to the limit that
b)' Let
and
As in Exercise 9 of Section 8.4,
c)
7.
and the convergence is quadratic. In particular case 2) of a) can

occur.
Show that A @, L,(l, x) is transcendental. Thus case 1) of a) can
also occur. (See Exercise 7 of Section 8.6.)
Calculate
sin-' :C)
and
~ ~ ( c o s:-C)' .
Hint:
sin-' z = - i log (iz +
m)
.
Chapter Nine

Abstract. In Section 9.1 we derive the classical formula for r2(n) and its
theta function equivalent. In Sections 9.2 and 9.3 we consider the summation
of various multidimensional series. Results include an alternating series test in
several dimensions, evaluation of various lattice sums, and related invariants.
Section 9.4 gives Watson's quintuple-product identity, and Ramanujan's ,TI
product. Section 9.5 considers quintic and septic multipliers and solvable
9.1
SUMS OF TWO SQUARES
9 We need the following simple lemma on Lambert series whose proof

(Exercise 1) proceeds by expanding both sides of each equation.
1r
5
t
,'
Lemma 9.1
If Iql<1 and u n : = q n l ( l - q " ) , then
Our development hinges on yet another remarkable identity due to

Ramanujan [62]. (See also Hardy and Wright [60].)
282
Proposition 9.1
Let 8 be real with 0 < 0 < T . Let
(9.1.3)
(9.1.5)
T := T ( q , 6 ) :=
cot
(:)
+C
un sin (no)
n=l
1 "
n u n [ l - cos (no)].
T2:= T 2 ( q ,0 ) = 2 n=l
Then
Tl + T , = T ~ .
Proof.
[f cot (;)12+s1+ S2
T'=
where
1 "
S , := un cot
2 .=I
(:)
sin (no)
and
S2 :=
umunsin ( m 0 )sin (no) .
m,n=l
Now
1
- cot
2
(:)
1
1
sin (no) = - + C cos ( k o ) + - cos (no)
2 k=l
2
while
2 sin ( m 0 ) sin (no) = cos [ ( m- n)8]- cos [(m+ n)6] .
Thus
T'
[:
cot
(:)I
+C
k=O
at cos (kB)
9.1 Sums of Two Squares
for constants a k ( q ) which we proceed to evaluate. Now
where the $u: term comes from m = n in S., Thus

1
a, = - 2 u n ( l+ u,,)= - C nu,
2 .=I
2 n=l
by (9.11). For k 2 1, S, contributes to a,,
and S, donates
m-n=k
... ..
...
n-m=k
..
'
m+n=k
for m,n r 1. Thus
Luckily,
and
Hence
Thh shows that

1
) cos ( k 6 ) + 5
which is the desired result.
C kuk[l - cos ( k d ) ]
k=l
284
For 8 := v / 2 this result becomes
on applying (9.1.2). Thus
Now (3.2.23) can be used to show that the right-hand side is 8:(q). Thus
and this last expression is nonnegative for real q. (Hardy and Wright [60] use
Proposition 9.1 to deduce the formula for 0: from that for 8;).
For r = 1 or 3 let dr(k) denote the number of divisors of k congruent to r
modulo 4. Then
and a comparison of the coefficients in (9.1.9) shows that
In other words, the number of representations of a positive integer k as a

sum of two squares, counting order and sign, is 4 times the surplus of
divisors of k congruent to 1 modulo 4 over those congruent to 3 modulo 4.
This recovers Jacobi's classical result, a result also known to Gauss. (See
Dickson [71].) Since
we also have
This is the formula exploited in Section 3.7. Also (9.1.9) shows that
(9.1.12)
8;(q) = 1 4
c
m
(-1)'n-'''2qnm
n,m=l
n odd
and
An alternative recent derivation of (9.1.9), due to Hirschhorn [85],relies

only on the triple-product identity. Begin with (3.1.13) and let a2 := w. This
gives
Now apply the triple product in form (3.1.1) to each of these sums. This
leads to
n (1
m
(9.1.14)
(a - a w l )
- a2qn)(l- a-2qn)(l-
qn)
n=l
Next differentiate each side with respect to a, at 1. (Use logarithmic

differentiation on each product separately.) This, after rearrangement,
produces
286
The right-hand side of (9.1.15) is
where the last equality follows from (3.1.4) and (3.1.7). Thus
which is equivalent to (9.1.9).

In the course of his study of Ramanujan's mock theta functions, Andrews
1861 discovered the following remarkable cubic counterpart of (9.1. l l ) ,
which we write as
It is an easy consequence of (9.1.16) that every number is a sum of three

triangular numbers (a fact originally observed by Fermat and proved by
Gauss). This also implies that a number is a sum of three odd squares
exactly when the number is congruent to 3 modulo 8 (Exercise 71) as
observed by Euler.

A comprehensive account of the development of formulae for sums of 2n
squares, from elliptic considerations and Lambert series, can be found in
Rademacher [73]. Many wonderful related identities are to be found in
Ramanujan [62]. Odd sums are much harder to evaluate. They involve
generating functions of'class numbers as shown in Mordell [16] and Watson
[35l -
1. Prove Lemma 9.1.

2. Verify (9.1.6) and (9.1.7). Show that these are equivalent to (9.1.8).
3. a) Prove formula (9.1.11) for 0;(q).
b) Establish (9.1.12) and (9.1.13).
4.
a) Let 0 tend to .rr in (9.1.8) and evaluate the limit.

.
b) Evaluate (9.1.8) when O : = ~ / 4 7,r / 3 , 2 ~ / 3 Inparticularshowthat
5. Two identities due to Lorenz (Dickson [71, vol. 3, p. 291) are
n odd
where ~ , ( n )is 1 if n = 8 k
+ 1, 8 k + 3 and -1 if n = 8 k + 5 , 8 k + 7 ,
Let d,,+,(k) be the number of divisors of k of the form ai + b.

Show that the number R2(k) of integer solutions of n2 + 2m2 = k is
given by
Similarly, the number R3(k) of integer solutions of n2 + 3m2 = k is

given by
Note that a) follows from Exercise 12 of Section 3.7

Observe that ii) can be used to show that
m
1- q4"
%(~)0~(q~2
) = 4( - - z i ) q n = 4
n=1
1- q
n odd
n=l
n odd
qn + q-"
1 qZn+ q-2n
'
Hence
-6
1) I2 and Fn and Ln are the Fibonacci and Lucas
where p := (1
numbers, as in Section 3.7.
6. ,Use Ramanujan's modular identity of order 3 (Section 5.2) to deduce
that
288
Hence deduce that
where p := (3 - V3)/2. Note that $ CfZ2 ~ ~ ( q ) ~ can

~ (beq written
~ )
( q ~a )closed
.
form for
more compactly as 0i(q)0i(q3) + 1 3 ~ ( ~ ~ ) 8 ~Find
the Lambert series above when q := e -dv7 .
7. Use formula (9.1.16) to establish that every positive integer is the sum
of three triangular numbers and that every number of the form 8k + 3 is
the sum of three odd squares.
9.2 (CHEMICAL) LATTICE SUMS
Sums of the form
arise naturally in chemistry. (Here the prime indicates that we avoid

summing i = j = k = 0.) Indeed, b3(l) can be considered as the potential or
Coulomb sum at the origin of a cubic lattice with alternating unit charges at
all nonzero lattice points. This may be considered as an idealization of a
rocksalt crystal. The quantity b3(l) is called Madelung's constant for NaC1.
Different crystals give rise to different lattice sums. We will also consider its
two-dimensional (laminar) analogue
and its four-dimensional form
There are some nontrivial considerations about the sense in which these
sums converge. (See Exercises 1 and 2.) We assume all sums denote limits of
rectangular summations. These will converge for re(s) > 0. The general form
of these rectangular sums is
9.2 (Chemical) Lattice Sums
289
which, for real-valued 2, can be shown to converge by an alternating series

test. (See Exercise 2.) The convergence in bN(l) is ~ ( n - ' I 2 )so that 10"
terms are needed for O(n) digits. Obviously direct computation is virtually
impossible.
Fortunately, some beautiful analytic reductions are possible. In this
section we illustrate this for b, and b,. The idea is to observe that, for
re(s) > 0.
where Msis the Mellin transform of Section 3.6. Thus
Now (9.1.13) shows that
Thus
and we have factored b2(2s) into a product of Dirichlet L functions (the

alternating J function a and P := L-,) defined by
In particular, b2(2) = -T log2. Equation (9.2.4) is originally due to

Lorenz. Now any reasonable method of computing a ( $ ) and P ( $ ) will
compute b,(l.). Such is possible by various integral or summation techniques. Correspondingly, from Theorem 3.2,
290
and expansion gives
Now we compute these transforms and have
on adding and subtracting m(mk)-Yn each summation. Thus
But .Z,"=, (2n + I)-" = (1 - 2-"((s)
and so
In particular,,
Use of the functional equation for (, equation (3.6.6), allows one to

establish other formulae. In general any identity for any even-dimensional
power of theta functions will convert into a factorization for a matching
lattice sum. (See Exercise 4.) In Exercises 6 through 10 we show how
two-dimensional lattice sums may be used to evaluate elliptic invariants.
An excellent and extensive recent survey of lattice sums can be found in

Glasser and Zucker [go]. This also discusses their chemical origin at some
length. Madelung's constant is analyzed from a mathematical perspective in
Borwein, Borwein, and Taylor [85]. There is ambiguity in the literature as
to whether the constants are positive or negative. We have chosen the sign
as convenient.
1. (Analyticity of lattice sums)
a) Considered as a limit of rectangular sums, show that b2(s) and

b,(s) exist and are analytic for re(s) > 0.
Hint: Use the Mellin transform.
b) Show that
291
Hint: g2(2s) := C;=, (-l)"r,(n)n"\onverges

and is analytic for
re(s) > f . For re(s) > 1, g2(2s) and b2(2s) coincide.
for
c) The analogous sum g3(2s) := C:=, (- l)"r,(n)n-"iverges
s : = 4.
2. (Alternating series test) A mapping a: N+
,
ml, m2, . . . , mN2 0,
R is (N-)monotone if, for
Thus 1-monotonicity is just i(m) r i(m + I), and 2-monotonicity is

C(m, n) + i(m + 1, n + 1) r Z(m, n + 1) + ii(m + 1,n), while 3-monotonicity demands that the alternating sum over any unit cube be
positive if the bottom corner is. We say that is fully monotone if a
and all its restrictions are monotone. Consider
2C
j=l
(-l)Lmii(ml,m,, . . . , m,)
m.=O
a) Prove the following alternating series test given in D. Borwein and

J. Borwein [86]. If a is fully monotone and lim,,,
a(&) = 0, then
the rectangular sums converge alternatingly. Hint: (In two dimensions) show that s, := Crj=o ( - l ) i f i ~ ( i ,j) satisfies
Draw a picture. (The case of b, and b, is spelt out in Borwein,

Borwein, and Taylor [85]. An analogous bounded convergence
test due to Hardy can be found in Bromwich [26].)
b) Suppose that ii is N times continuously differentiable on (R+),.
Show that a is totally monotone if the partial derivatives satisfy
C)
for all partial derivatives with i, < i2 < i, < - . < .,i
Verify that C; (-l)'+jf k(i2+ j2 + k2)-p and E; (-1)'+'(2i
converge for p > 0.
3. Prove that r(s)b,(s)= ~ ~ ( 8 : - 1 ) .
4.
a)
Show that
m
and
+ j)-P
292
b)
Show that
b8(2s) = -16C(s)a(s
- 3)
re (s) > 1
is equivalent to the formula (3.2.25)
c)
Use 020304= 0:,
equation (3.2.4), to show that
5. The hexagonal sum is

m
h2(2s) :=
4(n, m)
re(s) > 0
-- [(n + m12)' + 3(rn12)~]~
where q(n, m) := ${sin [(n + 1)0] sin [(m + l)0] - sin (no) sin [(m 1)0]) and 0 := 27~13.This corresponds to calculating the Coulomb
potential on a regular hexagonal lattice. (See Borwein, Borwein, and
Taylor [85].)
a) For re(s) > 1, show using the cubic modular equation that
b) A formula of Cauchy (Dickson [71, vol. 3, p. 201) gives
Use this and a) to deduce that

h2(2s) = 3(1- 3'-")l(s) L-,(s)
where
c)
L-,(s) := 1 - 2-"
4-"
- 5-
+ 7-"
8-
+ .. . .
This provides an analytic continuation of h2(2s). Thus
is Madelung's constant for the hexagonal lattice. Show that

h2(2) = V3-rlog 3 .
293
(Evaluating invariants) The factorization of two-dimensional zeta

sums into sums of products of L series can be carried a great deal
further. This is described in Glasser and Zucker [80] and in Zucker and
Robertson [76a,b]. By combining number theoretic and transform
techniques, one can explicitly factor all sums whose discriminants are
disjoint (have one form per genus) and a few others. This leads to
formulae such as
and
In these two formulae P is an odd square-free number [congruent to

1 modulo 4 in (9.2.8)] with t distinct factors. The right hand sums over
all divisors of P and one has
which. are primitive L series modulo d. (These can only exist for d = P,
4P, or 8P, and the sign configuration is in fact uniquely specified.) Here
(kin) is the Kronecker (generalized Legendre) symbol. L,, is taken for
p = f1 (mod 4). For example, with P := 29, (9.2.8) becomes
.
Dickson [29] gives an extensive list of disjoint discriminants. In

particular, there are 18 numbers less than 10,000 to which (9.2.8)
applies and 15 numbers less than 10,000 to which (9.2.9) applies.
Indeed, (9.2.8) holds for type one P := 5, 13, 21, 33, 37, 57, 85, 93,
105, 133, 165, 177, 253, 273, 345, 357, 385, 1365 and (9.2.9) holds for
type two P:= 1, 3, 5, 11, 15, 21, 29, 35, 39, 51, 65, 95, 105, 165, 231.
There are only finitely many disjoint discriminants. We shall call such
P solvable.
Implicit in (9.2.8) and (9.2.9) are corresponding theta series identities, and formulae for representations as weighted sums of squares.
a) Show that if q is replaced by - q in (9.2.8) and (9.2.9), we
produce formulae for
!
i)
ii)
2'(-l)m+n(mz+ pn2)"
xt (-l)"'(mz
+ 2pn2)-"
Show that the effect of replacing q by -q in (9.2.8), (9.2.9), or

similar formulae is to replace L+, by -[I- (21d)2'-7 L,,, unless
the L function was multiplied by some factor involving 2-", or d is
even, in which cases it is unchanged.
Recall that (3.2.12) and Exercise 4d) of Section 3.2 gave
4?i2' (-l)"+"(m2 + rn2)-' = - 3
log f
( e )
and
Use these formulae with a) and b) to establish that for the

appropriate P,
i) 2'-'
?i-
-log f 4
fi
( m ) = L14,(1) log 2
?iii) 2'-1 ----log f ;l(GP)

= L-,,(I) log 2
m.
The classical Dirichlet class number formulae (Landau [58]) allow

us to write L,,(l) algebraically. One has, for d > O restricted so
that L,, is primitive,
h(d)
i) L+,(l) = 2 v2 log ~ ( d )
2Tr h(- d )
ii) L-,(I) = - v2 w(d)
Here h(d) is the number of (broadly) e q u i v a p t primitive classes
of reduced forms with discriminant d = b - 4ac or ideals in
~(fi)
where d is D or 4 0 , depending on whether D = 1 (mod 4)
or D = 2,3 (mod 4); ~ ( d is) the fundamental unit in Q ( f i ) , which
may be computed from the fundamental solution of the appropriate Pell's equation (see LeVeque [77] and Hua [82]); and w(d) is
a factor which counts the number of automorphs of the form, and
is 2 except that w(3) = 6 and w(4) = 4. From formula (4.12) in
Zucker and Robertson [76a] we have Dirichlet's formulae, for
d >4,
Also for d > 0
1 d-l
h(d) log ~ ( d =) - 5 2 (d / n) log (sin
n=l
(7)).
Moreover, for disjoint forms, one can observe that the class
number h(-d) must coincide with the number of genera g. The
number of genera is as follows. If d is odd, then g = 2"-', where d
has m distinct prime factors. If d is even and dl4 has m distinct
prime factors (including 2), then g = 2" when dl4 = 0 , 1 , 5 (mod 8)
and g = 2"-l otherwise.
An excellent brief survey of history and of recent advances
regarding the class number can be found in Goldfeld [85]. Now
observe that, for the appropriate P, both $ f 4 and $ f;' will be
products of powers of fundamental units from some of the divisors
of 4 P or 8P. [For our solvable P, h(-4P) = 2' when P = 1 (mod 4)
and h(-8P) = 2'.] In particular, verify that for P := 5, 13, or 37,
since t = 1 in each case
Since h(P) = 1 in each case, we see that
P := 21,33,57,93,
e)
Similarly, we may now verify the values of G:,

given in Exercise 9 and 10 of Section 4.7.
Establish that for P := 5 or 29,
f)
(Compare Table 5.2.)

For all the type one numbers P listed above, we can now
explicitly give Gp. Similarly for the type two numbers we can give
g,,. This accounts for most of the square-free and nonprime
invariants given by Weber or Ramanujan.
Show that
296
7.
(Evaluating singular values) We can proceed further to evaluate

k : = h"(2P) for P of type two. We know that k / 4 =
f;'(m)fl"m).
Hence
-- T log -
V D
(1;)
C' (-1)"(m2 + 2pn2)-'
We already know [Exercise 6c)l the first sum on the right. From some
elementary, hut skillful, theta transformations we may deduce that
On setting s := 1 and substituting above, we derive
a beautiful simplification.
a) Use (9.2.10) to compute h"(2P) for P := 3, 5, 11, 29.
b) Use (9.2.10) to compute h"(210) given in (4.6.12).
c) Observe, as Zucker did, that there are computable in this form
two larger singular values: those for 2 P := 330 and 462. Verify that
and that
8.
(Evaluation of K in terms of I') Selberg and Chowla [67] showed for

all rational numbers r that K(h*(r)) is expressible in closed form using
a finite number of I' values. This relied on Kronecker's remarkable
'Grenz-Formel,' which has
297
as a special case. Here v:= q ( G ) is the eta function of (3.2.9) and

(3.2.11), and y is Euler's constant. Using (9.2.13), Zucker [77] applies
the factorization results described above to explicitly compute K
corresponding to solvable sums in terms of T, surds, and r values. This
leads to the following table of evaluations (Table 9.1). Elsewhere
Zucker has actually given K(A*(210)) [which involves r(n1840) for
(n, 840) = 11. The general formula valid for either r := P (P of type
one) or r := 2 P ( P of type two) is given by
and
4 log ?,I(-)
-.=,c
1
h(-4r)
4r-1
(-4r1~(2)
There is a corresponding formula when in Zucker's terms S(1,1,

(1 + r)/4) is solvable.
a) Verify the contents of Table 9.1, for r = 1 , 2 , . . . ,6.
b) Show that r(m124) is O,,(log n) computable for integral m.
Hint: Express such r values in terms of K at singular values. It
would be interesting to know if this is possible more generally.
/42/2). (Compare Exercise 9 of Section 5.2.)
c) Compute E((3 - fi)
d) Show in general that E(A*(r)) is computable in terms of r values
and algebraic quantities for rational r.
(Conjugate divisors and evaluations of k) Consider P as above and
suppose P := dld2 and Q := d,ld2 for divisors dl and d,. Using
Kronecker's genus character sum formulae one can show that G, and
G, = G,-I must have the same general form for P of type one.
-1
Similarly, g2, and g,, = g,,-I
are paired, as are A"(2P) and A*(2Q).
This is best illustrated with examples.
a)
Show that
The first value may be computed as above. The second is easily

verified from Schlafli's form of the quintic modular equation.
TABLE 9.1. Evaluation of K at the First Sixteen Singular Values
b) Correspondingly
One may verify G,,,, and G,,,, from the corresponding cubic and
quintic equations, and so on.
c) Use the techniques of Section 5.3 to compute ~ ( 1 0 5 ) .
d) Observe that
i)
A*(6)=(2-fi)(fi-fi)
A*($)= A*'(%)=(2 - v 3 ) ( f i + fl)
ii)
A*(10) = ( m - 3)(fi - 1)'

A*(:)= A*'(;) = ( m - 3 ) ( f i +
iii)
A*(58) = (l32/58 - 9 9 ) ( f i - 1)6

A * ( & ) =A*'(?) = (l32/58 - 9 9 ) ( f i +
q2
.
e) Indeed, for all P of type two, A*(2Q) and A*(2P) will be "conjugate", as we illustrate for A*(210). Let
Then A'(210) = IT:=,

urn, and each AY(2Q) is a corresponding
where each E, is +I. In compact form one
product IIi=, u~m,m(Q)
has E,(Q) := 1- 2bm(Q), where bm(Q) is the mth binary digit of
b(2Q) defined by
These correspond to all the reduced forms with discriminant 840.
300
f)
Numerically, find the "conjugate" values for P:= 165 and

P:= 231. [This involves computing 28 products and comparing
them with the theta expansion of k(2Q).]
With these conjugate values and the formulae of Section 5.3 one may
observe that a(2P) is now available in closed form for all P of type two (and
for conjugate divisors Q). Again we illustrate with an exercise.
10. a) Generate a recursive version of formula (5.3.3) and specialize this
expression to produce a formula for cu(2dld2) in terms of R d l , Rd2,
Md2, M2 and the appropriate singular values.
b) Apply this formula with dl := 35 and d 2 := 3 to compute 4210).
Observe that this uses only values of A" given in Exercise 9e).
( /213
f i is incorCompute
a(42). Note: g:2 = (21/2 + ~ ) ' [fi)
c)
rectly, given in Weber [08].
11. Let {a,) and {b,) be given sequences.
a) Show that
if and only if
b) Show that
if and only if
c) Show that, with the notation of Exercise 12 of Section 3.7,
and
m
ii)
C e(n)n-"=-. l(s)
n=l
9.3 Odd-Dimensional Sums and Benson's Formula
Also
where (P is Euler's totient function, which counts the numbers less

than n and relatively prime to n. Thus 4 ( l ) := 1, (P(5)= 4, and
(P(6) =2.
9.3 ODD-DIMENSIONAL SUMS AND BENSON'S FORMULA
While even-dimensional sums usually factor, only a few odd-dimensional

ones factor. The reader can, however, produce identities like (9.2.7) from
Jacobi's identity (3.1.15) or from (3.2.8). (See Exercise 1.) There are
nonetheless many theta-based techniques of which we establish one based
on the theta transform.
Theorem 9.1 (Benson (1956))
m,n=t
(9.3.1)
Proof. By symmetry,
where q := e-'. Thus
The theta transform (2.3.2) leads to
odd
and since
302
r($)= -12,
The internal integral I(n, j, k) was evaluated in Exercise 4 of Section 2.2.

We have
and
Finally, for a > 0,
and Benson's formula follows. 0

The convergence acceleration is astounding. Summing for 0 5 j, k 5 3
produces b,(l) := -1.74756459. . . , which is correct to eight places. Even
the first approximation, by - 1277 sech2(n-lfi),gives - 1.73 . . . .
For s # $ this manipulation leads to an integral involving Bessel functions
of the second kind, and the closed form is lost.' There is, however, an
extension to N dimensions (among others). One can show that
odd
Thus
(9.3.3)
- b,(2)
= 16a2
5 sech2(:.\/m)
= 4 log 2
i , .i .. k = l
odd
and
m
- b2(0) = 277
2 sech2[z (2n + 1)] = 1

n=O
which coincides with Exercise 7eii) of Section 3.7. (See Exercise 2.) For the
final evaluation it helps to know that for primitive L series
as follows from the functional equations for L,,:
(3
L-,(s) = C(s) cos - L - , ( I - s )

L+,(s) = C(S)sin
(9
L+,(1 - s )
where C(s) := 2%s-1k-s+1121'(l- s).

Our derivation of Benson's formula can be found in Glasser and Zucker
[80]. In that paper, and references therein, one finds much further discussion of odd-dimensional sums. They also illuminate the relationship between
the multidimensional zeta functions of Epstein and lattice sums.
1. a)
Show that
where L-,(s) := 1 + 3-" - 5-"

b) Show that
7-"
+ ..
and g(1) = m p ( 0 ) = fi.

c) Derive similar identities from Exercise 5 of Section 4.7.
2. Establish the generalization of Benson's formula (9.3.2) and its special
cases (9.3.3) and (9.3.4). Exercise 4b) of Section 9.2 shows that
b8(6)= -8[(3).
The remaining exercises examine the n-dimensional Hurwitz zeta function. Let d > 0 and define
where si := sign (a,) and i := (a,, a,,
. . . , a,).
Thus
and
3.
a) Show that
Hint: Take a Mellin transform. Then make a logarithmic variable

change and sum the resultant series.
b) Let L,(s, d) =: AN(s, d) and L&, d ) =: A-,(s, d), where .? is
the vector (1,1,. . . , 1 ) in UXN. Show that for d > 1
Hint: Use integration by parts.

c) Combine integration by partial fractions and the recursion (9.3.6)
to show that every sum of the form (9.3.5) factors into a linear
combination of one-dimensional Hurwitz zeta functions (with coefficients depending on s).
4. Let AN(s):=A-,(s, N) and PN(s):=AN(s,N).
a) Show that
and
b)
Deduce that, for appropriate s,
iii)
rn,n=~
(m + n)S
= l ( s - 1) - l(s).
Thus
m+n+l
iv)
V)
c)
(-1)
= 4 4 s - 1)
-,
(Iml + 14)"
-gt
--
Q1
(Iml + lnl)"
= 4 3 - 1).
Show that
N
( N - I)!P,(s)
C a r l ( s - 1+ n)
n=l
where a; are Stirling numbers of the first kind. Thus

P ~ ( s ) =2 [(s
- 3) - S(S - 2)
+ a {(s - 1) - J(S)
and
P&) = & [(s
- 4) -
& ((s - 3) + g 5(s - 2) - g [(s
- 1)
+ l(s) .
There is a similar formula for A,(s).

5. a) Show that
b)
Show that
c)
Show that
2
n,m=l
(-l)"+"
1
= - [(2 + 3-"a(s)
(3n + m)" 3
- a(s
- 1) - L-,(s, I)]
[-
1 7
log 2 - -and so L-,,-,(I, 4) = 3 3
6.
Show that
Hint: (1 + x)(l + x2).
. (1 + x
~ ~=-(1~- xZN)
) /(I
- x).
306
9.4
THE QUINTUPLE-PRODUCT IDENTITY
Jacobi's triple-product identity has an elegant fivefold analogue due to

Watson [29] and Gordon [61]. This is
valid for all complex z and q with I ql< 1 and z f 0. The proof is left as an
exercise. (See Exercise 1.)
If we divide both sides by 1- z-' and let z tend to 1, we derive
Now this yields
on using Euler's pentagonal formula (3.1.10) and (3.1.7). This can be

used to establish a recurrence formula for r2(n), as in Ewell [82]. (See
Exercise 4.)
The identity, implicit in Ramanujan7swork, was discovered by Watson [29]
and rediscovered by Gordon [61]. Further extensions are discussed by
Gordon. Other proofs abound in the literature.
1. Prove (9.4.1). As with the triple-product identity, it is easy to establish
the formula (9.4.1) up to a constant relying on q alone. To evaluate the
constant observe that when z := -1, (9.4.1) reduces to Euler's penta-
gonal identity.
2.
Show that (9.4.1) is equivalent to
9.4 The Quintuple-Product Identity
Let q := qi and z := q-k in (9.4.1) and deduce that

m
where N(j, k) consists of all integers congruent to 0, + k , j, j + k,

j + 2k (mod 2j) (repeated as appropriate).
Deduce that Euler's identity follows for j := 4 and k := 1. For j := 3
and k := 1 one gets
(n2+n)/2
. With f(q) := IIr=, (1 - qn)-l this bewhere g(q) := CF=, q
comes a formula due to Ramanujan
Use g(q) = f(q)lf2(q2) to obtain a functional equation for f(q)

(the partition function).
Use (9.4.3) with j := 6 and k := 1 to obtain
Hence obtain a functional equation for 0,. Thus note that

3e3(q9) = O3 ( q) never has a solution. (Compare Section 4.7.)
Establish equation (9.4.2) and
Use (9.4.2) to derive a recursion for r2(n).

Prove that
Hint: Use Exercise 2.

Show that
308
so that
Observe that this is slightly faster to compute than the original theta
series ratio.
e) Show that
so that
5. a) Show that
b) Replace q by -q in a) to express L-,,(2s - 1) as a lattice sum.

c) Combine (9.4.4) and (3.2.7) to prove that
Observe that at s = 4 this equals 4.

Gordon [61] also gives various congruences, like those given in Chapter 3
for the partition function. For example,
has c3,+, divisible by 3.

6 . There is yet another remarkable identity due to Ramanujan, which
includes both the triple-product and the q-binomial theorems. This is
the ,?Irl sum, whose derivation and uses are accessibly described in
Askey [80]. In standard notation one writes
9.5 Quintic and Septic Multipliers and Iterations
309
where (a; q), :=
k=O
(1 - aqk) and (a; q), := (a; 4)( 4 ; q),
'
This converges at least for I q ( < 1 and I blql< 1x1 < 1.

a) Verify that the triple product-1is contained in (9.4.5).
Hint: Begin by setting a:= c , x : = cx and b:=O. Now let c:=O.
b) For b := q (9.4.5) becomes the q-binominal theorem:
c)
9.5.
valid for 1x1 < 1 and Iql < 1. Verify that Cauchy's binomial theorem
is a special case. Hint: Begin by setting a := q-2n.
Use the q-binomial theorem to express the limit mean of Example
8.4 as a Taylor series.
QUINTIC AND SEPTIC MULTIPLIERS AND ITEZUTZONS
In this section we discuss additional modular and multiplier identities and

give a number of applications. Notations are as in Sections 4.6 and 4.7.
References to entries are all to Chapter 19 in Ramanujan's Second
Notebook (Berndt [Pr]).
Proposition 9.2
(4
(9.5.1)
5M, = 1+ ~G,,,/G;
(9.5.2)
1/M5 = 1+ ~ G , / G &
(9.5.3)
5M5 + 1/M5=2{2+ k l + k'l') .
(b) Let 1/(2t + 1) := M,. Then

(9.5.4)
1 - 212 = M;(I
-llt -t
2 ) d m
Proof.
(a) These are given in Entry 13. Berndt [Pr] provides proofs which may
also be deduced from (4.1.20).
(b) These may similarly be found in Entry 14.
Some Additional. Applications
310
We will write, as in Section 4.7,
Ramanujan also gives the following beautiful counterpart to Theorem 4.11

(Entry 12(iii)).
Theorem 9.2
where for i = 1 or 2
and
This provides a solvable update for M5 despite the nonsolvable nature of W5

) have
(See Exercise 8 of Section 4.5). Indeed with m, := 5 ~ , ( r 5 ' ~we
with x := m, - 1, y and ri as above.

If we combine (95.9) with the following formula for
remarkably simple solvable 5th-order iteration for T.
E,
we obtain a
Proposition 9.3
If r > 0 and s(r) := ~ ; ' ( r )then
Proof.
We begin with E,, as given by (5.2.13), and the explicit formula for R,.
We use Proposition 9.2(a) to rewrite s(r)R5 and 9.2(b) for the terms
involving 1' and k2. This leads reasonably directly to
Now (5.2.14) becomes (9.5.10). (Exercise 2.)
311
The identical manipulations to those in Example 5.3 immediately yield

4 5 ) = ${v3We next list several similar identities for M,.
qiT&Fij>.
Proposition 9.4
4 9 M 2 = 1- + - -I'- - 8 11'
k
k t kk'
1/
k
= 2-
(3213
.
k' - kk' - 8 +-
1'
11'
(b) Let t := (k1)lI4.Then

(9.5.13)
7M7 - 1/M7 = 6 - 16t + 12t2 - 8t3 .
Proof. These are to be found in Entry 19. 0

From (9.5.13) and (4.6.7) we may establish that for r>O and
~ ( r:=
) Mql(r)
where t4 = kl. (See Exercise 3.)

Ramanujan does not give a septic analogue to Theorem 9.2. He does,
however, give quintic and septic updates for the eta-multiplier. Let 77 be
given by (3.2.9) and (3.2.11). Let
so that Np corresponds to Mp. This is the eta-multiplier of order p. Now

(3.2.15) can be written as
where wP(l2,k 2 )= 0.
Theorem 9.3
(4
(9.5.17)
312
Proof.
( a ) In each case one combines (9.5.16) with appropriate multiplier equations. For p := 5 use (9.5.1) and (9.5.2). For p := 3 use the identities
preceding equation (4.7.9). (The details are left as Exercise 6a).)
( b ) We use (9.5.11) and (9.5.12) to write
l l ' ~ , ~ ( l 8+ ~ ; =) kl'
+ Ik' - kk'
and
k k ' ~ ; ( 4 9+ 8 1 ~ ;=) k l r + Ik'
- 11'
We now subtract one from the other and divide by 11'.

We finish the section by listing Ramanujan's updates for N, and N,. Entry
12(i) can be recast as
5 N 5 ( q 5 )= ( p 1 I 5+ v1I5- 1 ) ' / ( 5 ~ , ( ~ ) )
(9.5.21)
where p, and u are the solutions to

pv = - 1 ,
+ Y = 11 + ( 5 ~ , ( q ) ) ~ .
Entry 18(ii) becomes

7 N 7 ( q 7 ):= ( p 1 I 7+ vl',
(9.5.22)
where p ,
+ w1I7- 1 ) ~ / ( 7 N , ( q ) )
and o are the roots of

x3 - ax2 - bx
+ 1 =0
and a and b are given by

a := 57 + 14[7N7(q)12+ [ 7 ~ , ( ~ ) ] ~
b := 289 + 126[7N7(q)]'+ 19[7N7(q)14+ [7N7(q)16.
313
Theorem 9.2 gives 83(q25)solvably in terms of 8,(q5) and 03(q). Likewise

(9.5.22) gives 8,(q4*) solvably in terms of e3($') and B3(q). Thus 04(q) =
8,(-q) is similarly solvable and since k = 8il8, and j is solvable in k we see
that for f any of 8,, k, g , G or J , f ( q P ) is solvable over Q(f(q), f(ql'P)); for
p := 5 and for p := 7. In view of the nonsolvability of the quintic or septic
modular equations for A (and hence k and j), this is at first surprising. What
is happening is that the Galois group for F,, p r 5 and prime, is a
nonsolvable group of order ( p - l ) p ( p + 1). ow ever j(q'lP) is a root of F,
and, since F, is irreducible, it is of order p + 1. Thus the splitting field for F,
over Q,( j(q), j(ql'P)) has order dividing p ( p - 1). For p := 3, 5, 7, and 11,
( p - 1 ) / 2 is prime and the corresponding group is obviously solvable. For
p := 7, for example, we expect seventh roots of cube roots to comprise the
solution ( ( p - 1) I2 = 3). This is consistent with equations (9.5.22) and
Knowing that a solution exists and exhibiting it, particularly in simple form,
can be very different matters. The components of the quintic and septic
algorithms for T are far less complicated than one might initially expect.
Both can be packaged very elegantly. (See Exercises 2 and 7, and Borwein
and Borwein [Pr].)
Combine (9.5.1) and (9.5.2) to obtain Schlafli's form of the quintic
modular equation.
Compute that
iii) ~ , ' ( 315) =
a)
(5 2
iv) ~ ; ' ( 2 / 5 ) = I 6 ( f i + 2 ) ( f i - 1)2 .

Find closed forms for Myl(n) for n := l , 5 , 9, and 25.
Use G,, = [ ( f i + 1) /2][(9 + 1/85)/2]"' and the conjugate nature
of G,,,, to compute 5M,(y). Use Theorem 5.4 to compute ~ ( 8 5 ) .
Verify the formula (9.5.10) for a(25r).
Obtain closed forms for a(n) n =25, 125, 625, 225, and 1225
(Ramanujan [14] givesG,,,,).
Observe that Exercise Ib) (9.5.9) and (9.5.10) combine to give
several explicit iterations for T.For example, we may begin with
)
r : = 1, a(1) = 4, ~ ( 1 =5(1/5-2).
Use Theorem 9.2 to compute G,,,, in terms of G, and G,,,.
314
3.
a) Show that (9.5.14) holds and that
b) Compute the corresponding updates for 6(25r) and 6(49r). In

particular
c) Verify 4 7 ) .
d) Establish that R, and R, are as given in Table 5.1.
4.
Ramanujan also gives (Entry 19)
a) Thus show
b) Compute M , ( $ ) and M,(Y).

5.
Ramanujan in his letters (Hardy [40] p. 353) gives the following

beautiful hybrid identity. Let
Then
a) Verify that G:, =
8 l I 4 ( f i+ 1)
8 - l I 4 ( f i + 1)
and G:,, =
2
2
6. a) Establish the eta-multiplier formulae of Theorem 9.3a) and b).
b) Show that N,(k;,
*
1
k,)= - .
c)
Show that
where x := (kk'lll')113, y := ( ~ l ' l k k ' ) ~and

'~
d) Prove that if x : = a and y : = m
Hint: Use (4.6.8).

Combine (9.5.20) and (9.5.22) to produce a solvable update for
m, := 7~,(r7'"); and so a solvable 7th-order iteration for T.
Ramanujan's letters also contain the modular equations of degree 5 for
K,,, and K,,,. (See Section 5.5.) For K,,, one has
and for K,,, one has

lk + l'k'
+ 8(11'kk')"3{(lk)1'3 + (l'k')1'3) = 1 .
a) Verify that, in the notation of (5.5.34),
b) Similarly
Hint: The appropriate pth-order modular transformation for K,

sends I =: h: (n) to k := (p2n).
Chapter Ten
Other Approaches to the

We examine some of the standard polynomial and rational approximations to elementary functions, particularly to exp and log. We discuss
methods for reducing the complexity of calculating these functions based on
accelerating the evaluation of the approximants. While these methods are
usually less than optimal,,they are of more general application than those of
Chapters 6 and 7.
Abstract.
10.1 CLASSICAL APPROXIMATIONS
We commence with an analysis of the standard approximations to exp on a

disk D, := { l z l l 8 } . The notations we will require are as follows. Let P,
denote the algebraic polynomials of degree at most n with real coefficients.
Let 11 f llA denote the supremum norm of a continuous function f on the set
A , that is,
For a continuous f on an infinite compact set A C C, let
and let
These quantities are, respectively, the error in best uniform polynomial and
10.1 Classical Approximations
317
best uniform rational approximations. The existence of the best approximants is fairly straightforward and is left as Exercise 2.
The most commonly used polynomial approximations to exp are undoubtedly the partial sums of the Taylor series. This is reasonable since in any
neighbourhood of zero the partial sums are asymptotically optimal. (See
also Exercise 3.)
1
Theorem 10.1
(b)
IfpEPn,then
Part (a) follows from the estimate
Proof.
"
i=n+l
1
1
~<-[l+n+2(~+;;;?
( a + l)!
+ ( n + 3 )1( n + 4 ) + .
.)I .
For part (b) we observe that if
where s n is the nth partial sum to exp at zero, then by RouchC's theorem,
p(z) - sn(z)
and
ez - s,(z)
have the same number of zeros (counting multiplicity) in Dl. As ez sn(z) has a zero of order n + 1 at zero, we deduce the contradiction that
p = s,. To finish the proof we need only observe that
1
min lez - sn(z)l> lzl=l
(n + I)!
We now turn to the Pad6 approximants to exp. These are rational

approximations that are a natural extension of the Taylor approximants. We
define
and observe that r,,, is a rational function of z with numerator of degree m

and denominator of degree n. In closed form,
318
Theorem 10.2
( b ) Ifp, q E P,, then
Proof. Let
n
(101.6)
qn(z) := (2n)!
(:)(-Zlv
"=O
Then
Now from (10.1.6), for l z l 5 1,

(10.1.8)
Since
(10.1.9)
= JOw (t - z)"tne-' dt
(($)!
we have from (10.1.7) and (10.1.8),
Part (b) requires showing that for 1 zl = 1,
which follows from the estimates (Exercise 5)
and
The rest of the argument is analogous to part (b) of Theorem 10.1. If

there were a rational function p l q satisfying (b) with the inequality strictly
reversed, then by (10.1.11) and RouchC's theorem,
PI^ - rn,,
and
ez - r,,,
would both have the same number of zeros, and from (10.1.7) we would
deduce that p l q - r,,, has at least 2n + 1 zeros and hence is identically
zero. 0
The fact that r,,, is the Pad6 approximant is a consequence of (10.1.7),
which shows that
In general the (m, n) Pade' approximant to an f (analytic at zero) is the

unique rational function r = plq, where p E P, and q E P,, which satisfies
f(z>- p(z>/q(z) = O(zh
\
where h (in nondegenerate cases h = m + n + 1) is as large as possible. For

n = 0 this defines the nth Taylor polynomial.
The following is a partial list of the standard series and continued fraction
expansions for exp and log.
320
log
z+1
(=)
=2 ( t
1
+ S;i
+9
l + . - .)
z"-.l
lim -= log z
SJO
Pad6 approximants derive their name from H. PadC, a student of Hermite,

who was one of the first to systematically study such approximations at the
end of the last century. The theory of Pad6 approximation may be pursued
in Baker and Graves-Morris [81]. The convergence theory for Pad6 approximants is far more complicated than the analogous well-known theory for
Taylor series. (See Exercise 9.) Except in special cases, such as exp or
functions given by Stieltjes transforms [S da(t) l ( x + t)], analysis of region
or rates of convergence is only partially understood. Theorem 10.2 can be
sharpened to show that
using PadC approximants centered at z := 1/(2n + 1). This is due to Trefethen [84]. (See Exercise 10.) The discussion of the approximations of exp
follows Newman [79], as do Exercises 7 and 8. These exercises illustrate the
different rates of convergence on disks and intervals. Exercise 8 is the n = m
case of a conjecture of Meinardus, namely, that
as n + m+ co. This conjecture has been resolved recently by Braess [84].

(See also Nemeth [77].)
For further discussion of the material of this section the reader is referred
to Cheney [66] or Newman [79]. The various expansions may be found in
Abramowitz and Stegun [64].
In Section 11.3 we will use the PadC approximant to exp to derive a
precise irrationality measure for e.
1. (Lagrange interpolation formula) Given n + 1 points in the plane,
(zi, w,), i = 0, . . . , n, so that z, # zi for i # j, show that there exists a
unique p E P, so that
p(zi) = W ,
i = 0,.
. . ,n .
Show that
where
2.
(Existence of best approximants) Prove that En and R, are well

defined, that is, show that the min is achieved in (10.1.2) and (10.1.3).
Hint: A uniformly bounded sequence of polynomials, all of degree n
has a uniformly convergent subsequence whose limit is a polynomial of
degree at most n.
3.
a)
Suppose that f is entire and that s, is the nth partial sum o f f at

zero. Show that
=1
lim sup
n-'m
Ilf-snllD,
Hint: Show that if f(z) =:
aizi, then for infinitely many m,
Now use the arguments of Theorem 10.1.

b) Suppose f is analytic in a neighbourhood of zero. For fixed n show
that
lim E,(fY
4) - 1
/If-snIID,
Part a) illustrates that the Taylor approximants behave globally

like best polynomial approximants to entire functions on disks.
The story is different on different shaped regions. Part b) shows
that locally the Taylor approximants are always optimal.
4.
Prove that (10.1.4) has the representation (10.1.5).
5. Establish the estimates (10.1.12) and (10.1.13).
6. Establish the expansions (10.1.14) to (10.1.22).
7. (Polynomial approximation to exp on [-I, 11) Let s, be the nth

partial sum of exp at zero.
a) Show that if p(z) is a polynomial of degree n , then p(z)p(llz) is a
polynomial of degree n in the variable z + 11.2.
If x is the real part of z, where 121 = 1, then ex = ez12ei12=
b) er/2e1/2z
. On IzI = 1 approximate e Z l 2by s,(z/2) and approximate
ell2' by s,(1 122) and estimate the errors.
c) Use part a) to construct polynomial approximations to exp that
satisfy
Note the approximation is
d) Modify part c) to show that
(This is in fact asymptotically optimal.) Hint: Consider the method

with approximation centered at 11n.

>
8.
(Rational approximation to exp on [-I, 11) Show that
1.
Hint: Proceed as in 7). First observe that Exercise 7a) holds for
rational functions of degree n. The approximation is given by
where lzl= 1 and z := x + iy . Use estimates like those in the proof of

Theorem 10.2 to prove the result.
9. The (n, 1) Pad6 approximant p, to f := E;=, aizi has denominator
a,+,z - a, and, provided a, # 0, satisfies
a) Show that iff is entire, then there is a subsequence of {p,) that

converges to f uniformly on any given compact subset of C.
b) Show that there exists an entire function so that the full sequence
{p,) does not converge uniformly on any open set in @.
Hint: Show that the poles of the p, can be dense in @.
Exercise 9, due to Beardon and Perron, illuminates some of the problems
inherent in uniform convergence questions for Pad6 approximants. It was
conjectured that subsequential convergence holds for the sequence of (m, k)
Pad6 approximants (k fixed) and for the sequence of (m, m) Pad6 approximants. Much of the conjecture concerning convergence along rows (k fixed)
was recently settled by Buslaev, Gonchar, and Suetin [84]. They show, for
example, that iff is entire, some subsequence of the (m, k) Pad6 approximants (k fixed) converges uniformly to f on compact subsets of C. These
conjectures, due variously to Baker, Gammel, Graves-Morris, Wills, and
others are discussed in Baker and Graves-Morris [81]. A more complete
convergence theory is available if one is prepared to settle for weaker types
of convergence, for example, convergence in measure.
10. (More on the Pad6 approximants to exp) Let
Show, as in the proof of Theorem 10.2, that
Show that, as m + n + 03,
Hint: Observe that (u - l)"um is essentially a "spike" at u :=

ml(m + n) and that
Show that
zo
rn
~ r n , n ( :=
~)
m!(n + m - k)!
(m - k)!k! z
zo
" n!(n + m - k)!

(n - k)!k! ( - l ) k ~ k.
qm.n(z) :=
Recall that
Show that
J," tne-' dt = n!
P m ,n
:= Pm,n
n!
and
Qm,n
..-
qm,n
m!
are polynomials with integer coefficients of degree m and n,

respectively.
Show that, as n, m --+ 03,
The convergence is uniform on compact subsets.

Hint: Examine the coefficients of pm,,(z)l(n + m)!
Show that, as m, n + ~ ,
The convergence is uniform on compact subsets that avoid any
325
zeros of the denominator sequence. Observe that, by e), only

finitely many of these zeros lie in any compact set. Further details
may be found in Trefethen [84] or Braess [84].
g) Show that
Hint: Consider the Pad6 a roximant centered at 2p2/(2n + 1).

/(2n + 1) in parts e) and f). This gives
That is, replace z by z - 2p PP
the upper bound. Use RouchCYstheorem, as in the proof of
Theorem 10.2, to derive the lower estimate.
11. (On the main diagonal Pad6 approximants to log)
I
a) Suppose S, and T, are polynomials of degree n and suppose that
T,(x) log x - S,(x) = O(1- x)~"" .

Show that, if T,(x) := to + t,x
b)
+ . + t,xn, then
Show that
if we normalize so that t, := 1. [This is the denominator of the

(n, n) Pad6 approximant to log at the point 1.1 Observe that T, is
of degree n and has integral coefficients.
c) Show (with the above normalization, t, := 1) that d, -S, has
integer coefficients, where d, := LCM(1, . . . , n).
d) Let
s,( 4
%,(x) := log x - T,(x) '
Show that
%, (x) :=
du
326
Hint: Observe that gn:= O(1- x)'"+'. Now differentiate to get
Thus
Since the middle term is a polynomial of degree 2n with lead

coefficient 1, we must have
Show that
e) Show that
and that for x 2 0,
f)
Show that, for x E ( 1 - 6 , 1 + a ) ,
where c, > 0 and d, > 0 depend only on 6, 0 < 6 < 1.

10.2 REDUCED COMPLEXITY METHODS
We are primarily concerned with methods that accelerate the evaluation of

the elementary function by reducing the complexity of evaluating one of the
standard approximants. Most of the approximants listed in the previous
section evaluated by usual methods (such as, Horner's rule) provide between
O(n) and O(n log n) digits for n arithmetic operations. The slight difference
comes from the more rapid convergence of the Taylor polynomials for exp
10.2 Reduced Complexity Methods
327
than those for log. (See Exercise 1.) We proceed to examine three methods
of complexity reduction. While none of these methods is as fast as those of
Chapter 7, they all have their own particular advantages. The second
method, based on the fast Fourier transform, for example, applies to most
of the special functions.
10.2.1 Acceleration Based on Functional Equations
The exponential satisfies the functional equation
This allows us to reduce the calculation of the exponential to a small region

about the origin and then to approximate in that region using a Taylor or a
Pad6 approximant. From estimates like those of Section 10.1 and (10.2.1)
we have
and
where s, and r,,, are, respectively, the nth Taylor polynomial and the (n, n)
Pad6 approximant to exp at zero. Both of these above estimates provide
O
and
OB(v'EM(n))
methods for the evaluation of exp.

The functional relation for log is
f(z2) = 2 f k )
(10.2.4)
Combined with (lO.l.l9), this leads to
Truncation after n terms leads to an approximation on lz - 11 < 4 that has

error 0(4-"') and yields an algorithm for log with complexity
Oop()
and
OB(v'EM(n)).
These are good intermediate range estimates of exp and log. For
328
n := 100, (10.2.3) provides in excess of 12,000-digit accuracy at the expense

of roughly 300 full precision multiplications. (300 terms of the Taylor series
provides roughly 600 digits.)
Variations of the above methods can be used to construct O,,(fi)
algorithms for the circular functions and their inverses. (See Exercise 2.)
Since any elliptic function (Section 1.7) satisfies an algebraic "half-angle"
formula, we can, as above, construct O,,(nl/'+") algorithms for elliptic
functions provided we have expansions ,available at the origin. (Unfortunately, convenient closed forms of the Taylor series for sn, for example, are
not available.)
10.2.2
Acceleration Based on the FFT
It is possible, based on FFT methods, to evaluate a polynomial or rational

function of degree n at n + 1 distinct points in O,,(n(log n)2). (See Exercise
3 of Section 6.2.) Our aim is to use this observation to accelerate the
evaluation of partial sums or related approximations. We illustrate with log.
Start with
and consider
We can write
n-1
where
Now, for fixed z, evaluate pn(kn) at k = 0,1, . . . , n - 1 in O,,(n(log n)')

(Exercise 4) and evaluate s n 2 in a further O(n) operation. Since for 1215 1,
sn2(z) provides fi(nz] digits of log z, we have constructed an algorithm for
log which is 0,,(n" (log n)"
With care, this idea can be extended to produce
).
(10.2.10)
O,,(nl~'(log n)')
and
O,(nl"(log n ) ' ~ ( n ) )
algorithms for a variety of nonelementary transcendental functions. Almost

any function with regular Taylor coefficients is susceptible to such analysis.
Algorithms for the gamma function and the hypergeometric functions, of

complexity given by (10.2.10), are presented in Exercises 6 and 7.
We can combine the methods of Section 10.2.1 with the above to
construct an
(10.2.11)
~ , , ( n ' / ~ ( l n)')
o~
and
~ , ( n " ~ ( l n)'M(n))
o~
algorithm for log. This merely requires truncating (10.2.5), say, after n2
terms and evaluating the truncation using the FFT method in O,,(n(log n)')
steps. Note that this approximation provides 0(n3) digits of log.
Likewise algorithms can be constructed for exp and the trigonometric
functions with complexity given by (10.2.11).
The relative difficulty of implementing these algorithms renders them
largely of theoretical interest.
10.2.3 Acceleration Based on Binary Splitting
If we wish to evaluate the constant e by summing the Taylor series at 1, then
we can and should take advantage of the reduced length of each individual
operation. With this in mind consider
and
where
By construction,
With x := 1, (10.2.14) approximates e with an error of less than el(2")!, and

hence provides fl(n2") digits of e. We can use (10.2.13) to recursively
evaluate e with bit complexity O,((log n)M(n)).
Brent [76c] shows how to modify this to provide an O,((log n)'M(n))
algorithm for exp. This is outlined in Exercise 8.
Modifications and variations lead to O,((log n)'M(n)) algorithms for all
the elementary functions. Particular values of various of the nonelementary
special functions are also amenable to this analysis. So, for example, is
330
Euler's constant. [See Exercises 10 and 11.1 Exercise 9 outlines an

O,((log n ) ' ~ ( n ) ) algorithm for rr based on recursive evaluation of arctan.
It is perhaps worth underlining the observation that these acceleration
methods apply only to the bit complexity. The operational complexity is not
reduced. Nonetheless, the algorithm for the number e implicit in (10.2.12),
(10.2.13), and (10.2.14) is asymptotically as fast as any known and has the
virtue of being implementable using only integer addition and multiplication, except for a single final division.
Walz [Pr] has studied classes of asymptotic methods using extrapolation
and elimination techniques. These often outperform AGM-based methods
in the "microcomputer range" (less than 20 digits). Interestingly, for the
complete elliptic integral of the first kind he finds the AGM to be always
superior. For incomplete integrals this is not always so.
Much of the material of this section is due to Brent [76c]. In particular,

Exercises 3 and 8 follow Brent closely.
1. a) Show that truncating the series in (10.1.14) and using Horner's
rule leads to an O,,(nllog n) algorithm for exp.
b) Show that (10.1.18), (10.1.19), and (10.1.20) all lead to O,,(n)
algorithms for log using usual methods for evaluating the polynomials or rational functions in question.
c) Show, for n := 2" and S := 1/2", that (10.1.17) and (10.1.22) lead
to O,,(n) algorithms for exp and log.
d) Analyze the complexity of (lO.l.l5), (lO.l.l6), and (10.1.21).
Assume in all parts that the method in question is used on a compact region bounded away from the boundary of the domain of
convergence.
2.
a)
Use the functional relation

2 arctan
(y-z 2 z+ 1 + 1
= arctan
and the expansion
to construct on ~ , , ( n " ~ )algorithm for arctan on

b) Use the functional relation
[COS
(;)I2
1 ~ 1 %q 5 1.
cos z + 1
to construct an ~ , , ( n " ~ )algorithm for cos on Dl.
c)
Show, in general, that if
and f(z/2) is an algebraic function of f(z), then one can construct

an algorithm of complexity
0,,(n"')
and
~,(n"~~(n))
3. Instead of approximating ez'2n by sn(z/2") in (10.2.2), approximate

ez12n
,
- 1 by sn(z/2") - 1. Then repeatedly use the relation
4.
to evaluate e" - 1. Show that this avoids requiring 0 ( n U 2 )guard digits.

This modification and its obvious analogue for the Pad6 approximant allow the calculation of exp without loss of significant digits beyond
the O(1og n) loss inherent in performing nU2operations.
Show that p,(y) of (10.2.9) can be written as
where w, and v, are polynomials of degree n. Use FFT methods to

show that the coefficients of wn and vn can all be calculated in
O,,(n(log n)') and hence, that pn(y) can be evaluated at n points
in O,,(n(log n)'). (See Exercise lb) of Section 6.2.)
~ algorithm for exp by writing
5. a) Construct an ~ , , ( n ' ' ~ ( l on)')
where
tN(rn) := (Nm)
XNm !
X
+
xN- 1
x2
+ (Nm + l)(Nm
+ 2) +..
+ (Nm + 1) . . (N(m + 1) - 1)
and evaluating tN(m) for m = 0, . . . , N - 1 . Care must be taken

to compute (Nk)! in the requisite time.
lo~
b) Show how part a) can be used to construct an ~ , , ( n ~ ' ~ (n)2)
algorithm for exp.
332
c)
Construct an ~ ~ , ( n ~ ~n)')
' ( l algorithm
o~
for the Bessel function of
order zero,
6. (An O,(n"'(log
n)2) algorithm for the gamma function)
a) From the definition
show, by breaking the integral at N and expanding, that

(10.2.18)
T(s) = N
(-l)k N~
+
2- 7
k=O
IN
m
e-'ts-l dt.
b) Show for s E [I, 21 that
c)
Use b) and FFT methods to construct an algorithm for the gamma

function of complexity
~ ~ , ( n ~ ~n)')
' ( l o ~and
0,(n1"(1og n ) ' ~ ( n ) ).
7. ( ~ , ~ ( n ~ ~ n)')
' ( l oalgorithms
~
for hypergeornetric functions) We now
consider a (general) hypergeometric function to be a function
where anla,-, := R(n) for some fixed rational function R. In this

problem R is assumed to have rational (or precomputed) coefficients
and f is assumed to have a nonzero radius of convergence.
a) Show that the Gaussian hypergeometric series F(a, b; c; z) of
(1.3.5) is hypergeometric by the above definition. Show that,
provided a , b, and c, c # 0, -1, -2, . . . , are rational, F(a, b; c; z)
satisfies the additional assumptions.
b) Show that sin (e),
ez, log(1- z), E l n , and K l n are all
hypergeometric functions, up to a rational normalization.
c) For fixed n let
and
Observe that
and show that

Snz-, = Q(0) + znQ(n)T(n) + z 2 " ~ ( 2 n ) ~ ( 2+n ). . .
+ Zn(n-ll
Q(n(n - 1))T(n(n - 1))
d) Show that
can all be evaluated in O,,(n(log n)').

Hint: Consider the rational function of y
n-l
V(y) :=
fl R(i + y)
i=O
and observe that
e)
Now first compute the coefficients of V by recursively breaking the

problem in half and using a fast multiplication. Then calculate V at
the points V(O), V(n), V(2n), . . . , V((n - 1)n). (See Exercises 1
and 3 of Section 6.2.)
Show that
Q(01, Q(n),.
Q((n
can all be evaluated in O,,(n(log n)').
- l>n>
Hint: Note that Q is a rational function of k of degree bounded by

n degree(R). Now show that the coefficients of Q as functions of
k can be calculated in O,,(n(log n)') by proceeding recursively.
To do this consider
f)
and use a fast polynomial multiplication to recombine the pieces.

Finally evaluate Q(O), Q(n), . . . , Q((n - 1)n) as before.
Use the preceding parts to construct
~ , , ( n ~ ~ n)')
~ ( l o ~and
~ , ( n " ~ ( l no)~' ~ ( n ) )
algorithms for any hypergeometric function.

8.
(An O,((log n ) ' ~ ( n ) ) algorithm for exp)

a) Show that (10.2.13) can be used to recursively evaluate e with bit
complexity 0, ((log n) M(n)).
b) Use (10.2.13) for rational plq, where p2 5 q 5 2N, to compute
ePfq.Show that the bit complexity is as in Exercise 8a).
Suppose
x E [0, 1) is a binary 2"-digit number. Show that x can be
c)
written as
whereqk:=2
d) Write
2k
2k-1
andOspk<2
with pk and qk as in c). Show that this gives an O,((log n ) ' ~ ( n ) )

algorithm for exp on [0, 1).
9.
(An O,((log n ) ' ~ ( n ) ) algorithm for

a)
Consider the expansion
T)
and the recursion
where, for a < b,

p(a, b) := (2a + 1). (2a + 3). - . (2b - 1) .
Show that
calculates the (2"" - 1)th partial sum of arctan (llx).

Show that for a fixed integer x > 1 the above recursion computes
arctan ( l l x ) in
OB((h? n ) 2 ~ ( n ).)
This gives an OB((logn ) ' ~ ( n ) ) algorithm for .rr from
or any similar arctan formula.

Let f be a hypergeometric function defined as in Exercise 7 (with
the same additional assumptions). Show, for fixed p l q rational
inside the region of convergence of the expansion (10.2.19), that
f(plq) can be calculated with bit complexity
Show, for fixed rational p l q , that I'(plq) can be calculated with

bit complexity
336
11. (On the complexity of Euler's constant)

Euler-Mascheroni constant y is defined by
1
.. - log
y .- ,
,
,
rn
(10.2.20)
+ -+
Euler's constant or the

= 0.5772156649.
. ..
It is related to the gamma function by the formula
a) Use the recursion
to prove that if I' has an expansion of form (10.2.21), then y is

given by (10.2.20).
b) The exponential integral E l is defined by
Show that
(10.2.23)
-El(x)
=y
x>o.
+ log x + kC= 1 k . k!
c) Use (10.2.23) to construct an
algorithm for y by choosing x roughly of size 6n.

This method, suggested by Sweeney [63], is a reasonably efficient
method for computing y. Brent and McMillan [80] present a
number of algorithms for this computation. They calculate over
29,000 partial quotients of the continued fraction for y, As a
consequence they show that if y is rational the denominator of y
exceeds
Chapter Eleven
Abstract. The first section of the chapter deals with the history of the
calculation of n and related matters, while the second section deals with its
transcendence. The third section looks at irrationality measures and includes a
proof of the irrationality of c(3). This chapter is largely self-contained and
indeed contains considerable related number theory, especially in the exercises.
11.1
ON THE HISTORY OF THE CALCULATION OF n
The history of .rr presumably begins with man's first attempts at estimating
the perimeter or area of a circle of given radius and as such starts at the
dawn of recorded history. The Egyptian Rhind (or Ahmes) Papyrus which
dates from approximately 2000 B.c., gives a value of (1619)' = 3.1604. . . for
T. Various other early Babylonian and Egyptian estimates include 3, 3 $ ,
and 33. Implicit in the Bible (1 Kings 7: 23) is a value 3: "And he made a
molten sea, ten cubits from the one brim to the'other; it was round all
about. . . and a line of thirty cubits did compass it round about."
Mathematical interest in n comes into sharp focus in the classical Greek
period. The Greeks investigated the problem of "squaring the circle.'' This
question and its final resolution over two millenia later will be pursued in
the next section. Currently we wish to review the primary Western developments in the calculation of n.
Archimedes of Syracuse (287-212 B.c.)provided the first major landmark
in the quest for digits of n. By considering inscribed and circumscribed
polygons of 96 sides, Archimedes gave the estimate
A salient feature of Archimedes' method is that it can, in principle, be used

to provide any number of digits of n-.
If an denotes the length of a circumscribed regular 6.2"-gon and 6,
denotes the length of an inscribed regular 6.2"-gon about a circle of radius
112, then
This two-term iteration, starting with a, := 2 f i and b, := 3, can be used to

calculate n-. (See also Section 8.4.) The fourth iteration yields a, = 3.1427. . .
and 6, = 3.1410. . . and corresponds to estimating n- using polygons with 96
sides.
If we observe that
we again see that the error is decreased by a factor of approximately 4 with

each iteration. Variations of this modern formulation of Archimedes'
method provided the basis for virtually all extended precision calculations of
n. for the next 1800 years, culminating with Ludolph van Ceulen (15401610) who correctly computed 34 digits. The limitations of this method stem
from the relatively slow convergence and from the need to extract square
roots. (See Exercise 1.)
Fran~oisVibta (1540-1603) gave the first infinite expansion
which he derived by considering a limit of areas of inscribed 2"-gons. (See

Exercise 2.) John Wallis (1616-1703) through a complicated calculation
demanding prodigious numerical insight derived the infinite product expansion
This appears in his Arithmetica In$nitorum.of 1655. A few years later Lord
Brouncker (1620-1684), the first president of the Royal Society, recast this
as the continued fraction.
11.1 On the History of the Calculation of n
The Scottish mathematician James Gregory (1638-1675) in 1671 provided

the underlying method for the next era in the history of the calculation of T.
He showed that
and hence, on setting x := 1, that
a formula independently discovered in 1674 by Leibniz (1646-1716). By the

beginning of the eighteenth century Abraham Sharp under the direction of
the English astronomer and mathematician E. Halley had obtained 71
correct digits of rr using Gregory's series (11.1.7) with x :=
namely,
m,
It is the techniques of calculus that so expanded the scope for calculating,

and it is perhaps not surprising that Isaac Newton (1642-1727) himself
calculated rr to 15 digits sometime in 1665-66. He used the series
which is essentially an arcsin expansion. (See Exercise 4.) Newton was later
to write: "I am ashamed to tell you to how many figures I carried these
computations, having no other business at the time." John Machin (16801752) derived the formula which bears his name:
rr
- = 4 arctan
4
($)- arctan (&).
Coupled with Gregory's series for arctan this provides a very attractive
method for calculating rr since the first term is well suited to decimal
arithmetic and the second term converges very rapidly. Machin calculated
100 digits this way in 1706. In the same year William Jones published his A
New Introduction to the Mathematics, where he denoted the ratio of the

circumference to the diameter by the Greek letter 'IT, presumably for the
first letter of periphery. It was, however, Leonard Euler (1707-1783) who
popularized the use of the symbol. Euler derived numerous series and
products for 'IT and rr2. Among the best-known are
'IT
and
-90 = x' =an' I

'IT
.=I
The explicit summation of (11.1.12) had eluded Liebniz and also the
Bernoulli brothers, Jacques and Jean. The method by which Euler derived
his evaluations of C;=, 1lnZkis outlined in Exercise 7. This is to be found in
Euler's Introducio in Analysin Infinitorurn of 1748. The Machin-like formula
(11.1.14)
'IT
= 20 arctan
($)+ 8 arctan ($)
coupled with the expansion
where y := x2/(l + x2), allowed Euler to compute 20 digits of 'IT in under an

hour.
The next 200 years saw little change in the methods employed to calculate
'IT. In 1844 Johann. Dase (1824-1861), a calculating prodogy, used the
formula
(11.1.16)
'IT
- = arctan
4
(i)+
arctan
(') + arctan (i)
to produce 205 digits of 'IT. (Dase's arithmetical abilities were awesome-he

could multiply 100-digit numbers together in his head, a feat which took him
roughly 8 hours.)
-The zenith (or nadir depending on your perspective) in premachine
calculations was achieved by William Shanks (1812-1882), who published
607 purported digits of 'IT, of which 527 were correct. Later Shanks published an extension to 707 digits. This was also incorrect after the 527th
digit. These calculations took Shanks years and were performed in an
entirely straightforward fashion using no tricks or shortcuts. (See W. Shanks
[1853].) The mistakes went unnoticed until 1945, when D. F. Ferguson, in
one of the final hand calculations, produced 530 digits. Ferguson produced
808 digits in 1947, using a desk calculator and the formula
11.1 On the History of the Calculation of .rr
Thus dawns the computer age. In June 1949 ENIAC (Electronic Numerical Integrator and Computer) was used to evaluate 2037 digits of rr using
Machin's formula and 70 hours elapsed time. An analysis of the distribution
of the digits was carried out by Metropolis, Reitwiesner, and von Neumann.
By 1958, Genuys had computed 10,000 digits on an IBM 704 in 100 minutes,
once again using Machin's formula. Felton had performed a 10,000-digit
calculation in 1957; however, due to machine error it was only correct to
7480 digits. In 1961 D. Shanks and Wrench [62] used the identity
(11.1.18)
~=24arctan(!j)+8arctan(-&)+4arctan(&)
and under 9 hours on an IBM 7090 to produce 100,000 digits of T.This was
checked using the formula
1
(11.1.19) T = 48 arctan
+ 32 arctan
- 20 arctan (%) .
(&)
($)
The million-digit mark was set by Guilloud and Bouyer in 1973 on a CDC
7600. The calculation, which took just under a day, used (11.1.19) with
(11.1.18) as a check.
. Kanada, Tamura, Yoshino, and Ushiro [Pr] calculated in excess of 16
million digits using an AGM based algorithm, Algorithm 2.2, and checked
10 million digits using (11.1.19). The 16 million-digit calculation took under
30 hours on a HITAC M-280H and used an FFT-based fast multiplication.
At the end of 1985 the record belonged to W. Gosper. He calculated 17
million terms of the continued fraction expansion for T and so in excess of
this number of decimal digits-after a radix conversion from a binary
computation. His method is based on a very careful evaluation of Ramanujan7s series (5.5.23) on a Symbolics 3670. (A remarkable feat considering
the size of the machine.) As is surprisingly often the case with these large
scale calculations, Gosper uncovered subtle design flaws which had not
surfaced in smaller calculations.
In January 1986, D. H. Bailey [88] computed 29,360,000 decimal digits of
T on the CRAY-2 at the NASA Ames Research Center. This calculation
used only 12 steps of the quartic algorithm (5.4.7) with r := 4. This results
in computing a(z5O), which agrees with T-' to more than 45 million places.
The calculation took less than 28 hours and was verified with a 40-hour
computation of 25 steps of Algorithm 2.1. It is amusing to observe that the
quartic calculation requires well under 100 full precision multiplications,
divisions, and root extractions.
In July 1986, Kanada reclaimed the record with a computation of 2''
decimal digits. He again used Algorithm 2.2, verified in September using
(5.4.7) with r := 4, but reduced the elapsed time to 5 hours and 56 minutes
on a S-810120 super computer. This represented a speed-up by a factor of
15. His previous computation now used only 96 minutes of CPU time. Plans
were to compute 227 decimal digits (over 100 million) at the end of 1987.
Nor is the end in sight. It will probably be the case than hundreds or
thousands of millions of digits will be calculated by the end of the century.
(This is now more a matter of will than anything else. Indeed, by the time of
second printing Kanada had similarly computed 201 million digits.) Apart
from observations like "the sequence 314159 appears in the digits of n
commencing at digit 9,973,760," there is little we care to say about the
digits. They have, however, been subjected to considerable scrutiny. It is an
open question as to whether a is normal. That is, do all sequences of
integers appear with the same frequency in the digits (are one-tenth of the
digits 7, one-hundredth of the consecutive digit pairs 23, etc.)? On the basis
of the first 30 million digits the answer appears positive. This, of course, is
no great help in deciding the normality issue. (See Wagon [85].)
In terms of utility, even far-fetched applications such as measuring the
circumference of the universe require no more digits than Ludolph van
Ceulen had available-but then utility has had little to do with this particular story.
This section presents only the highlights of the quest for digits. The matter
may be pursued in detaiI in Beckmann [77], a most useful though rather
individualistic history, and in Le Petit Archimdde [80]. Schepler's chronography [50] and Wrench's history [60] are also of interest. Details of the more
recent calculations may be found in Tamura and Kanada [Pr], where a
compendium of Machin-like identities is provided.
There is also a considerable collection of T-related trivia. For example,
the Indiana House of Representatives attempted to legislate the value of .rr
in Bill 246 of 1897. The bill, which appears to proclaim n to be several
different incredibly inaccurate values, including 4 and 64/25 (see Beckmann
[77], and Singmaster [85]), passed the House and only floundered in the
Senate on the apparently chance intercession of C . A. Waldo, a professor at
Purdue. Keith [86] gives a 402 digit mnemonic for T.
Show that the algorithm of (11.1.1) and (11.1.2) calculates n by
showing that a, and b, are as advertised.
Prove (11.1.3) and estimate how many iterations of (11.1.1) and
(11.1.2) are required to calculate 35 digits of T . This should be
compared to Bailey's [88] calculation which uses the same operations.
Prove, from the product expansions for sin and cos, that
sin 8
(11.1.20) 8 =
($1 < a .
cos ($12) cos ($12') cos ($12~).Alternatively deduce (11.1.20) in an elementary fashion by setting
11.1 On the History of the Calculation of .rr
In:=cos
($1 . - . (t)
(;)
cos
In=
sin 0
2" sin (012")
and showing- that
cos
'
c) Set 8 := 7r12 and use the formula cos (012)

deduce Vikta's formula (11.1.4)
3.
m to
a) Prove Wallis's formula (11.1.5) in the form
Hint: Show that
and
b) Establish the corresponding formula for e:
c) Show that the volume of the 2n-dimensional unit sphere is .rrnln!

while the (2n 1)-dimensional unit sphere has the volume
2'"+'[n! l(2n + l)!]?rn. Find a unified formula for these two cases.
4.
Deduce (11.1.10) roughly as Newton did. Show that
and that
344
Pi
5.
a) Deduce Machin's formula '

77
- = 4 arctan
4
(i)
- arctan
(&)
as follows. Let 0 := arctan $ . Then

tan 20 =
2 tan 0 - -5
1- tan2 0 12
and
Hence
tan 40 - - =
-1+tan40 - -1
1 tan48
239
'
b) Show that arctan satisfies the addition formula

arctan x + arctan y = arctan
c) Show that
(11.1.21)
arctan
(i)
= arctan
(-)p +1q
(-)
X Y < ~ .
+ arctan (
and that if 1+ P2 = qr,

(11.1.22)
arctan
(&)
+ arctan (&)arctan (i)
.
=
Formula (11.1.21) was known to Euler. Bromwich [26] attributes (11.1.22)

.to Charles Dodgson (Lewis Carroll).
6.
(Machin-like formulae)
a) Show, for integral ai and bj, that
k v = arctan
($) + arctan (2.)+ . . . + arctan (2)
where k is an integer if and only if

(a,
+ ib,)(a, + ib,) . .. (a, + ib,)
has zero imaginary part.
11.1 On the History of the Calculation of
Hint: Consider (a,

fact that
+ ib,). . . (a, + ib,)

1
arctan z = 2i log
345
~r
= rei0, 101 < T,
and use the
(-1.
(This gives an algorithmic check for Machin-like formulae.)

b) Show, for positive integral u, v, and k and integral rn and n, that
rn arctan
():
+ n arctan
if and only if (1 - i)k(u + i)"(v

c) Verify
T
- = 4 arctan
-=
5-
-=
-=
(i)
+ i)"
- arctan
kT
is real.
(A)
(i)
(f) + arctan
2 arctan ( f) arctan (f )
2 arctan (i)+ arctan (f )
arctan
(t) 7
(Machin, 1706)
(Euler, 1738)
(Hermann, 1706)
(Hutton, 1776) .
These are, in fact, all the nontrivial solutions of b). This was a
problem of Grave's solved by Stprrmer in 1897. The problem can
be reduced to finding integral solutions of 1+ x2 = 2yn or 1+ x2 =
yn, n r 3, n odd. (See Ribenboim [84].) Much related material on
Machin-like formulae occurs in Lehmer [38] and Todd [49].
7.
Prove Brouncker's continued fraction by showing that
Hint: If
s := a,
then
+ a, + a,a2 + a,a2a3 + - .
This is a nonsimple continued fraction. The convergents satisfy

a similar recursion to that given in Exercise 2 of Section 11.3.
Apply this to
so that
and
X
arctan x =
x2
1+
3-x2+
9x2
25x2
5-3x2+ 7-5X2...
'
Now set x := 1. (According to Beckmann, Brouncker merely announced

his result-the above derivation is essentially due to Euler.)
8. Consider the series
Observe that sin x = 0 exactly when x = + k r . Now observe that on

setting y := x 2,
exactly when y = (krr)', k = 1 , 2 , 3 , . . .. If (11.1.24) were a polynomial,

we would know that the sum of the reciprocals of the roots of (11.1.23)
equals the negative of the coefficient of y and in general the sum of the
reciprocals of the powers would be expressible in terms of the coefficients and Bernoulli numbers. Thus we would deduce that
1
i-(kr12
and
21 - 1
( k ? ~ ) 90
~
k=l
Use the product expansion for sin [Section 2.2, Exercise Id)] to make
the above argument of Euler's rigorous. (See also Exercise 14 of
Section 11.3.)
9. In computing r from (11.1.19) one must evaluate arctan ( & ) and
arctan ( $ ). Use (11.1.15) to observe that
I
L
k
11.2 On the Transcendence of a
347
Thus terms of the second series are just decimal shifts of terms of the
first series. (See Ballantine [39].) How does this affect the'complexity
of calculating the two arctans?
10. Prove that the number
0.12345678910111213~. . n(n + 1)
- ..
is normal. A proof may be found in Niven 1561.

11.2 ON THE TRANSCENDENCE OF
?r
The problem of "squaring the circle" is the problem of constructing a square

of the same area as a given circle of radius 1, or alternatively given a line
.
rules of
segment of unit length of constructing a segment of length ~ 3 7 The
construction allow for the use of an unmarked straightedge and an unmarked compass. A more precise definition of constructible is provided in
Exercise 1. In fact, the constructible numbers are exactly those numbers
which can be obtained from the integers by a finite sequence of rational
operations and extraction of square roots. (See Exercise 1.) Thus constructible numbers are algebraic and the transcendentality of .rr shows the
impossibility of the problem.
The Greek notion of number, based on geometric construction, made
consideration of such problems more natural than they perhaps seem today.
Indeed the problem had arisen by the fifth century B.C. Anaxagoras, who
died in 428 B.c., had, according to Plutarch, considered it while in jail. His
contemporary, Hippocrates of Chios, the author of one of the first geometry
texts, also considered the question. The other classical Greek problems of
"duplicating the cube" and "trisecting the angle" also arose in this period.
The "Delian problem" of duplicating the cube (in volume), so named
because the oracle of Apollo at Delos had prescribed duplicating the cubical
altar as a Teans of halting the plague of 428 B.c., is equivalent to
constructing fi.(The impossibility of solving these problems is also discussed in Exercise 1.)
By 414 B.C. attempts at constructing .rr had become so numerous that
Aristrophanes refers to "circle squarers" in his play "The Birds." The term
came to refer to people who attempt the impossible. However, attempting
the futile is not always a waste of time. As Boyer [68, p. 711 points out:
The better part of Greek mathematics, and of much later mathematical

thought, was suggested by efforts to achieve the impossible-or, failing
this, t o modify the rules. The Heroic Age failed in its immediate
objective, under the rules, but the efforts were crowned with brilliant
success in other respects.
It is hard to know whether more energy has been consumed by attempts
at circle squaring or by calculations of rr. While doomed attempts to square
the circle flourish to this day, ,by the eighteenth century it had become
accepted in the mathematical community that the problem was probably
impossible.- In 1755 the French Academy of Sciences refused to examine
any more quadratures while as early as 1668 Gregory had attempted to
prove their impossibility. The first substantial step in this direction was due
to Lambert (1728-1777), who proved rr irrational in 1761. A subsequent
more rigorous proof was provided by Legendre (1752-1833). Legendre
proved the irrationality of rr2 in his Ele'ments de Gbome'trie of 1794 and
commented:
It is probable that the number rr is not even contained among the
algebraic irrationalities. . . But it seems to be very difficult to prove
this strictly.
(See Exercise 3.) This belief was shared by Euler. Liouville established the
existence of transcendental numbers in 1840, and in 1873 Hermite proved e
transcendental. It had been proved irrational by Euler in 1737. Finally in
1882 F. Lindemann [I8821 extended Hermite's proof to cover the transcendence of rr, thus laying to rest a 2300-year-old problem. This was simplified by
Weierstrass [67] in 1885, Hilbert [I8931 in 1893, and many others. Lindemann in fact established more generally that
for distinct algebraic numbers a,, . . . , a, and nonzero algebraic numbers

p!, . . . , pn. (See Exercise 7.) The transcendentality of rr follows since
- 1= 0. This is the signal achievement of the nineteenth century with
regard to transcendental number theory. Note that Lindemann's theorem
implies the transcendence of cos a , sin a , and tan a for algebraic a # 0 and
also the transcendence of log p for P algebraic, P # 0 or 1.
In 1900 Hilbert, as the seventh of his 23 problems posed at the International Congress in Paris, asked whether a Pis transcendental for a algebraic
(a Z O,1) and p an algebraic irrational. This was solved independently by
Gelfond and Schneider in 1934. (See Niven [56]). It is interesting to note
that Hilbert had speculated that this problem would probably resist solution
longer than the Riemann hypothesis or Fermat's last theorem. In 1966
Baker substantially generalized the Gelfond-Schneider theorem by showing
11.2 On the Transcendence of rr
349
that any nonvanishing linear combination of logarithms of algebraic numbers with algebraic coefficients is transcendental. That is, if ari and Pi are
nonzero algebraic numbers, then Po + Xi"=, a,. log Pj # 0. (See Baker [75].)
Note that ii = e-"I2, and so transcendence of e" follows from the GelfondSchneider theorem.
;
For a discussion of constructibility one might consult Famous Problems in

Elementary Geometry by Klein [1897]. The treatment we give in Exercise 1
follows Clark [71]. The exercises on transcendental numbers follow Baker
[75], Hardy and Wright [60], Hua [82], and LeVeque [77], and for the most
part are simplifications of the original arguments. The particularly simple
proof of the irrationality of r 2(outlined in Exercise 3) is due to Niven [56].
The reader interested in pursuing these matters in depth is directed to Baker
[75], Mahler [67], or Lang [66a].
1. (On constructible numbers, doubling the cube, and trisecting the
angle)
Constructible numbers can be defined by:
(i) The points (0,O) and (0, 1) are constructible.

(ii) Lines joining constructible numbers are constructible.
(iii) Circles with constructible centers and constructible radii (that
is, the radius is the distance between two constructible points)
are constructible.
(iv) The points of intersection of constructible lines and circles are
constructible.
It is the points of (iv) that form the constructible numbers.
a) Prove that the constructible numbers are a subset of the algebraic
numbers.
b) Show that perpendicular bisectors are constructible. See the hint
provided in the illustration. Hint:
c) Show that a circle through three given points is constructible. Sec

the hint provided in the illustration. Hint:
d) Show that the constructible numbers form a field. See the hint
provided in the illustration. Hint:
e) A number field C is constructible if C = Q(c,, . . . , c,), where

c,, . . . , c, are all constructible. Show that if C is constructible,
then C has degree 2" over Q.
f) Show that real extensions of degree 2 are constructible.
g) (The impossibility of doubling the altar) Use e) to show that % is
not construcible and so the Delian problem is not solvable.
h) (The impossibility of trisecting the angle) Show that constructing
an angle 8 is equivalent to constructing cos 0 and show that a 60"
angle is constructible. Show that
11.2 On the Transcendence of n
and hence, that cos 20" is a root of
Show that the above polynomial is irreducible over Q and hence, by

part e), that a 20" angle is not constructible. Show that the
constructibility of 8, given a rational cos38, depends only on
whether 4x3 - 3x - cos 38 factors over Q. Show that a 30" angle is
constructible.
The following series of exercises is on transcendental and algebraic
numbers. Recall that a is algebraic of degree n if a is the root of an
irreducible polynomial of degree n with integer coefficients. If a satisfies no
such algebraic equation, it is transcendental.
2.
(On Liouville numbers)

a) Prove that if a is algebraic of degree n, then for all E > 0 and for all
c>o,
has only finitely many solutions with p and q integral.

Hint: Suppose a satisfies
Then, provided p l q is not a root of p,
So by the mean value theorem, for p l q E [ a - 1, a

-
+ 11,
for some D, and the result follows.

A much deeper result of Roth (see Baker [75]) shows that a)
holds with q n + E replaced by q2+".
b) Use a) to show that
a : =10 z ~and
1
,,=I
p:=xn=1
lo2"'
are transcendental by showing that the partial sum approximations

would violate a).
Modify the construction of b) to exhibit uncountably many transcendental numbers. (In fact, all infinite subseries of a are transcendental, which exhibits an uncountable set of such numbers.)
Show that the algebraic numbers are countable and hence, that
almost all numbers are transcendental. Numbers that can be proved
transcendental because they can be approximated too rapidly by
rationals are called Liouville numbers. More precisely, a is a
Liouville number if for every m there exists a rationalplq, q >i,so
that
Prove that Liouville numbers are transcendental.

Show that the set of Liouville numbers has measure zero.
Show that the Liouville numbers have Baire category 11. Show, in
fact, that the complement of the Liouville numbers is of Baire
category I. Actually it is the union of Q and a nowhere dense F,.
(Recall that a set is category I if it is the countable union of
nowhere dense sets and that a set is nowhere dense if its closure has
empty interior. A set is category I1 if it is not category I. See
Oxtoby [80].)
3. (On the irrationality of e and n-)
a) Prove that e is irrational.

Hint: Suppose e = qlp and consider
Show that this is an integer strictly between 0 and 1.

b) Show that for any positive integer n,
f( x ) :=
x"(1- x)"
n!
has the property that f and all its derivatives are integer valued at
x = 0 and x = 1 and that for x E (0, I),
c). Prove that eP is irrational for integer p # 0.

Outline: Suppose eP = alb with a, b E N. With f as in b), set
11.2 On the Transcendence of m
Show, by differentiating ePxF(x),that
is an integer but that the left-hand side above lies strictly between 0
and 1 for large n.
d) Deduce from c) that eq is irrational for all rational q #O.
e) Show that rr2 (and hence a ) is irrational.
4"Outline: Suppose .rr2 = alb with a, b E N . Let f be as in b) and
consider
Then
d
dx
- {G(x) sin (TX) - n-G(x) cos (ax)) = r2anf(x)sin (ax)
and
rr
jolanf(x) sin ( m ) dx
= G(O)
+ G(1)
However, G(0) and G ( l ) are integers while as in c) the integral on

the left is strictly between 0 and 1 for large n.
4.
(On the transcendence of e) This exercise outlines Hilbert's proof of

the transcendence of e, a proof which, as LeVeque [77] puts it, is "as
elegant as it is mysterious."
a) Recall that, for k an integer,
and thus if p is polynomial with integer coefficients,
b) Suppose e algebraic. Then there are integers ai so that
let
and let
Observe that S + T = 0.
c) Show that, for infinitely many m, Slm! is a nonzero integer. To do
this, observe that if y := x - k, then
where the pk are particular polynomials with integer coefficients.

Now use a) to see that the first term of S is divisible by m!; that the
rest are divisible by (m + I)!; and that for infinitely many m the first
term is not divisible by (m + I)!.
I TI lm! = 0. To do this let
d) Show that lim,,,
M := Omax
s x s n ((x- l ) ( x - 2 ) . . - ( x -
n)l(x+ 1)
and show that
whence
e) Observe that, since S + T = 0, c) and d) lead to a contradiction, and

hence e cannot be algebraic.
5.
(On the transcendence of r r ) The exercise outlines Baker's [75] synthesis of Hilbert's proof of the transcendence of n.
11.2 On the Transcendence of
355
?r
a) Suppose 0, := in is algebraic. Let 1 be the lead coefficient of the

minimal polynomial for in and let 02,. . . , 8, be the other roots of
the minimal polynomial. Since e'" = - 1,
Show that expanding the above yields a nonzero integer n such that
where q := 2, - n and each ai is a nonzero sum of some subset of

the 0, of the form C SiOi, Si = 0 or 1. Show that any elementary
symmetric function of la,, . . . , la,, with integer coefficients is integer valued. (See Exercise 6.)
b) Let
where
for some prime p. Let
Show, by using a) and repeated integration by parts, that for

m : = ( n +l ) p - 1,
c) Show that J is an integer. To do this, observe that the right-hand

term is a symmetric function of la,, . . . , lan.
d) Show that for j # p - 1,
while for j = p - 1,
f "-"(0)
- - . a,)'.
= ( p - I)!(- l)nP(ala;
Thus for p sufficiently large,

( p - I)! 1 f ('-I' (0)
and
p !'4 f('-')
(0)-
Pi
356
e) Show that, for j s p - 1, and p sufficiently large
Show that
Hence with d)
f) Show, however, that for some M independent of p ,
This contradicts e) for large p and finishes the proof.

6 . (On symmetric polynomials) A polynomial in n variables is symmetric
if it remains unchanged by any permutation of the variables. An
elementary symmetric polynomial fi in the variables x, , . . . ,x, is defined
by
a) Show by induction that any symmetric polynomial in n variables

with integer coefficients can be written as a polynomial with integer
coefficients in the elementary symmetric functions.
b) Let sk := x: + . + xk. Prove Newton's identities:
and
7.
(Lindemann's theorem) This proof of Lindemann's theorem follows

Baker [75] and assumes some general familiarity with algebraic integers.
(An algebraic number whose minimal polynomial has lead coefficient 1:
observe that if a is an algebraic number and the lead coefficient of the
minimal polynomial is 1, then lor is an algebraic integer.)
11.2 On the Transcendence of m
357
Lindemann's Theorem
If a,, . . . , a, are distinct algebraic numbers and P,, . . . , P, are nonzero
algebraic numbers, then
a) Suppose that
Show that one can assume that

(i) The Pi are ordinary integers.
(ii) There exist integers 0 = no < n, < -
- - < n, = n so that, for each T,
%,+I?' ' , an,+l

is a complete set of conjugates and
b) Let I be an integer so that la,, . . . ,la, and @,,

algebraic integers. For p a large prime, let
For 1 5 i 5 n, let
where
Show that
is an algebraic integer and that
( P-
1 f :j)(ak)
. . . ,1&
are
and for p sufficiently large

P!~fi"(ak)
j=p-1,
k=i
while
!
1 f(
a )
otherwise
(An algebraic number a is divisible by h if a l h is an algebraic

integer.)
c) Show that J, . J,,I is a nonzero integer divisible by ( p - I)! and
. . L ( p - I)!.
hence
Show
that
there exists C independent of p so that
d)
IJ,.
I J,I
and that this contradicts c).

8.
(Lengths and measures)

a) Suppose P is a polynomial with integer coefficients of length L and
degree D. Suppose a is an algebraic number with minimal polynomial of degree d and length 1. (The length is the sum of the absolute
value of the coefficients.) Show that either
Hint: Suppose a, := a has minimal polynomial

Q(x) := qdxd+ ' ' ' -!- qO
and let a;,
. . . ,adbe the remaining roots
is a nonzero integer [if P(a,) # 01. Since
of Q. Show that
11.2 On the Transcendence of rr
where the last inequality requires showing
b)
and follows from part b) of this exercise.

Let P and Q be polynomials and let
Show that
i)
A P Q ) = p*.(P)AQ)
ii)
p ( P ) 5 length (P)
iii) p ( x - c ) = p ( x - I c I )
iv) p((x - cn)"") = p(x - c)
v> p(x - c) = m M l , IcI)
( p is called the measure of P.)

The next exercise constructs a form
where the pi are polynomials of degree S m with integer coefficients. Such

forms can be viewed as higher dimensional Pad6 approximants. Setting
x := 1 in the above leads to a polynomial in e that can, in conjunction with
Exercise 8, be used to prove the transcendence of e. Setting x := in, so that
ei" = -1, leads to a proof of the transcendence of T. Such forms were
examined by Hermite. It was, however, Mahler [31] who showed how to
base the transcendence proofs on them.
\
9. (Another proof of the transcendence of e and r r )

a) Suppose that we can find Vn so that
where each pi is of degree m and pn := xm + .
. Then
Pi
360
where each si is of degree m. In particular
and
b) Prove that, for sufficiently smooth G, if
then G("+')(x) = f(x). Use this to show that
e'l+f2+"'+f,,
dt, . . . dt, .
c) Thus
where
and where
d) Use c) to show that V,,,, defined as in a), exists and is uniquely

defined for all m and n.
e) Show that if C, is a circle of radius 6 > n, then
Hint: Use the residue theorem to see that V,,, is of the right form.
11.2 On the Transcendence of
361
?r
Observe by expanding elx and evaluating on circles of large radius

that
v,,,
= O(x
(n+l)(m+l)-1
1.
Show that, with 6 < 1,
Hint: Show by the residue theorem that p k is a polynomial of

degree m and that, with e),
Note that
The n := 1 case of e) and f) provides an alternate derivation of

the Pad6 approximant, equation (10.1.15).
Use f) to show that
has integer coefficients [d, := LCM(1, . . . , n)].

Let n be fixed. Show from c) that for m sufficiently large,
V,(x)#O
for x#O.
Note that this is trivial for real x # 0.

Let D > e and let
From c) show that

Wn,,(x) =
akjxiekx
Osksn
Osjsm
where the akj are integers. Use c) and the fact that d, < en('+") for
large n (Exercise 6 of Section 11.3) to show that for m + n large,
Show from f) with 6 := 1In, that for m + n large,
j) Let x # 0 and n be fixed. Then for m sufficiently large,
and
where c , and c, are constants, independent of n and m.

k) Show that e is transcendental.
Hint: Let Pm(e)= Wn,,(l) for fixed large n. Show that as m+m,
Pm(e) is sufficiently small that, by Exercise 8, e must be transcendental.
1) Show that rr is transcendental.
Hint: Let x := irr in i) so that e'" = -1. Now proceed as in k).
Observe that W,,,(irr) is a polynomial of degree m in rr.
11.3 ZRRATIONALZTY MEASURES

We examine the rate of approximation of e, rr, and log 2 by rationals. For
example, we show, for p and q integers and q sufficiently large, that
Estimates such as (11.3.1) are termed irrationality measures. The expected

rate of rational approximation is as follows: If f(x) is a positive nonincreasing function, then
has infinitely many integer solutions in p and q for almost all a exactly when
11.3 Irrationality Measures
Thus, with probability 1, we expect
to have infinitely many solutions, while
usually has only finitely many solutions. This result is due to Khinchin [64].
(See also Exercise 2 of Section 11.2 and Exercise 1 of this section.)
It is standard to the theory of continued fractions that if
then p l q is a convergent of the simple continued fraction for a , while of any

two consecutive convergents of the continued fraction, at least one of them
satisfies (11.3.6). (See Exercise 2.) Roth's theorem states that for algebraic
a,
has at most finitely many solutions. (See Baker [75].)

For specific transcendentals exact estimates are usually unknown, but e
can be analyzed very precisely.
Theorem 11.1
If
Sn
k=0
(11.3.8)
tn :=
(2n - k)!
(n - k)!k!
" (2n- k)!

k=(,
( n - k)!k!
then
s,
Ie-<l=-
1 log log tn
2 t: log tn [ I + O ( ~ ) I
Proof. From Exercise 10 of Section 10.1, with s, := Pn,,(l) and

we have
tn := QnYn(l)
and
(11.3.11)
The result now follows, as an application of Stirling's formula gives
2n+1=
2 log t,
[ l + o(l)]. 0
log log t,
Theorem 11.2
If p , q E Z and
then, for some n, p / q = sn/tn,where s, and t, are as in Lemma 11.1.

Proof.
(11.3.13)
The continued fraction for (e - 1)/2 is given by

e-1
-=LO, 1,6,10,14,18,. . .]:=[a,, a,,

2
. . .].
If p, := (s, - t,) 12 and q, := t,, then one can verify that
This, with Theorem 11.1, shows that p,/q, are the convergents for the
simple continued fraction for (e - 1)/2 and that expansion (11.3.13) holds.
(See Exercise 2.) In particular, if
and by Exercise 29, ( p - q) /2q is a convergent of (11.3.13).

Corollary 11.1
Let 0 < 6 < 1 and let u be a positive integer. Then
has infinitely many integer solutions, while

1 - 8 loglog q
2~
q2 log q
has at most finitely many integer solutions.
l e l ~ u - ~ l ~ -
The above corollary due to Davis [79] shows that el'" is atypical with
respect to the rate of rational approximation [compare (11.3.4) and
(11.3.5)]. For v = 1, Corollary 11.1 can be deduced from Theorem 11.1
without reference to Theorem 11.2 on continued fractions. (See Exercise 4.)
For general v the proof is left as Exercise 5.
We now turn to irrationality measures for T , l ( 2 ) , and 5(3). The
approach follows Beuker's [79] elegant treatment of ApCry's startling proof
of the irrationality of J(3).
Lemma 11.1
Let r and s be nonnegative integers.
for some n in E
r>s
1
tb)
Id Id
- xrys log xy
- xy
dx dy
r=s>O
for some n in
where d r := LCM(1,2, . . . ,r).
Proof. Consider
If we expand (1 - xy)-' and integrate term by term, we get
Pi
366
which on setting 6 = 0 establishes the r = s case of ( a ) . For r > s ,
which establishes the rest of ( a ) . If we differentiate I with respect to 6 and

set 6 = 0 , we get
and part ( b ) follows from differentiating (11.3.19) and (11.3.20).
Theorem 11.3
5(2) = m2/6 is irrational.

Proof. Let p, be the "shifted" Legendre polynomial
1
n!
p,(x) := -
(11.3.22)
[$J ' x n ( l - x)"
and note that
k=O
()
(n + k)!
n!k! ( - l ) k ~ k
is a polynomial of degree n with integer coefficients. Consider
I" .* --
lo1
16
( 1 - Y )pn("1 dX dY
1 - xy
(11.3.24)
y"(1- y)"x"(l - x)"

dx d y
( 1 - xy)"+'
where the equality follows on integrating n times by parts with respect to x.

Since for 0 I.x , y 5 1,
[with equality for x = y

ll.l(a),
(fi
- 1)'/2] we have, by (11.3.24) and Lemma
367
On the other hand by the same lemma applied to the first form of I,,
where
~=&(n)(;)
(n + k)!
! !
a, is an integer and d, = LCM(1,2, . . . , n). It is a simple consequence of

the prime number theorem (see Exercise 6) that, for any 77 > 1,
Furthermore, Stirling's formula and a little calculus lead to the estimate
where c, > 0, c,, c,, and c, are constants. In fact, van der Poorten [79] gives
From (11.3.26), (11.3.27), (11.3.29), and (11.3.30) we deduce that for

sufficiently large n,
where y, := d i p, and a, are integers, and where

(11.3.32)
6 :=
log [(I + fi)/(fi

- 1)15
-1=0.092159..
log {[(I + f l ) / 2 l s e 2 )
Thus (11.3.31) proves the irrationality of 5(2) = rr2/6.
Theorem 11.4
For p and q integers and q sufficiently large,
and
.> O .
Proof. From (11.3.30), (11.3.31), (11.3.32), and Exercise 3 we deduce

that, given E > 0, if
for sufficiently large q, then plq = anlyn for some n. [Here a,, y,, and
6 := 0.092. . . are as in (11.3.31).] Thus we need only verify (11.3.33) for
p l q = an/yn.This follows from the observation that, for small q > 0 and for
large n,
[See (11.3.24) and (11.3.25).] In conjunction with (11.3.27) this leads to
One now finishes the result by using estimates (11.3.29) and (11.3.30) to
show that
[
(- 1 1 2 - ]
2->3
Yn
1
11.86
Yn
for large n and small q.

The irrationality measure for rr follows from the irrationality measure for
.rr2 since, for large q,
The first irrationality measure for .rr was due to Mahler [53] (see also
Mahler [67]), who showed that
This was later refined by Mignotte [74] to
and by Chudnovsky and Chudnovsky [84] to
1
(which can be marginally sharpened).
q large
Theorem 11.5
[(3) is irrational.
Proof. Consider
where, as in (11.3.22),
We observe that
and rewrite (11.3.35) as
An n-fold integration by parts with respect to x yields
which, on substituting w := (1 - z)/[l - (1 - xy)z], becomes
where the last equality follows from an n-fold integration by parts with
respect to y. For 0 5 x, y, w 5 1,
and from (11.3.36),
370
From Lemma 11.1(6),
where a ; , PA, and d i are integers. Since e 3 ( f i- I ) <~ 1 and f i A grows at

most geometricalIy [see (11.3.42)],we deduce that there exists S t > 0 so that
for large n. This proves the irrationality of 5(3).

Corollary 11.2
For p and q integers and q sufficiently large,
Proof. The proof is similar to the proof of Theorem 11.3 and 11.4. One
first calculates explicitly from (11.3.35) and (ll.3.23),
From Stirling's formula one can show that
c,nc2(1+ 1/2)4n5
(11.3.42)
+ 16)~"
I
:c3nc4(l
for constants c, > 0 , c,, c,, and c,. In fact,
The remainder of the proof follows much as in Theorem 11.4, and the
details are left as Exercise 8.
Similar considerations lead to the irrationality measures for log.
Theorem 11.6
Let p, q , and n be integers. Then
(11.3.44)
and for any
I b P ~>p
I
E
q large
> 0, and fixed n > N,

n+l
q large .
371
The proof of the theorem is sketched in Exercise 9 or in Alladi and

Robinson 1791. Similar results may be found in Baker [75]. The best known
irrationality estimate for log 2 is also due to Chudnovsky and Chudnovsky
[84], who show that the constant can be reduced to 4.13. . . in (11.3.44).
Results on the rational approximation of e may be found in Adams [66],

Bundschuh [71], and Davis [79]. Adams shows that the number of relatively
prime integer solutions of
behaves like 3 log nl(log log n) (rather than the expected c log n that holds
for almost all numbers).
Transcendence estimates of type
where p is an algebraic number of degree d and height h (the height is the

modulus of the maximum coefficient of the minimal polynomial) due to
Feldman are discussed in Baker [75].
Irrationality measures are often difficult. The best known estimate for e"
is
ler-
>
1
qc log log
c a constant.
Apiry's proof of the irrationality of ((3) does not obviously extend to

other values of 5. It is not known whether ((2n -t- 1) is irrational for n > 1.
Equally, whether (log r)l.rr, e + v , Catalan's constant, and Euler's constant
are irrational is unknown.
1. Let f be a positive nonincreasing function so that
Show that the set of a for which
has infinitely many solutions has (Lebesque) measure zero. (The other
half of Khinchin's theorem is more delicate.)
Hint: Fix N. For each q > N consider intervals of radius f(q) /q around
the q points O/q, 1/q, . . . , ( q - l ) / q . The measure of the union of all
these intervals is bounded by
2.
(On simple continued fractions)

let
For integral a,, a, r 0, a, > 0, i # 0,

1
[a,, a,, . . . ,a,] := a,+
a1 +
a2 +
an
and let
[a,, a,, . . .I := f,i,mm
[a,, a,, . . . , a,].
Unless otherwise specified, we assume the continued fraction is

infinite. The nth convergent is defined by
a) Show that the convergents satisfy
Deduce that pn and q, are relatively prime. Note that {p,) and
{ qn) are increasing sequences.
b) Show that continued fractions are well defined, that is, show that
the limit in the definition always exists.
Rational
numbers have two representations since
c)
[a,,
and
. . . , a,]
= [a,,
. . . , an-, + 11
,-,,
a, = 1
[a,,. . . ,a,] = [a,, . . . , a

a, - 1,1] an # I .
Show that a number is rational if and only if it has a finite simple
continued fraction. [See e).]
d) Let
4 := [a,,
373
a,+, , . . .] where a := [a,, a,, . . .]. Show that
Thus {~z,klz,) T a and ~z,+,lq2,+1) ae) One constructs the simple continued fraction to a as follows
(1 1 := integer part). Let
a,:=La]
and
a,:=&
and proceed inductively to let
unless an- [a,] = 0, in which case the algorithm terminates.

Show that
a =[ao, a,,
. . .]
and if the continued fraction is infinite, then the representation is

unique.
f) Show from d) that
g) (Best approximation property) Show that if 0 < q < q, and if

~ 1 +4p,lq,, then
Hint: Prove I P, - qna 1 < - q a 1, Show first that one may asp:=up,+
sume 9,-,<q<q,.
Now write q : = u q , + v q
vpn-1, and ~ n - q n a = u ( ~ n - q n a ) + v ( ~ n - l - q n - i a ) .
Show
that u and v are nonzero integers and that u(p, - q,a) and
v(pn-, - qn-Ia ) have the same sign.
h) Show that if for z > 1,
and
,-,,
then rls and p l q are consecutive convergents of the continued

fraction for a.
Hint: Write p l q as a continued fraction and show that if
then
(Y
= [ a o , . . . , a n , z]
and
z = [ a n + l , a n + , 7- . . I .
[There are two representations for p l q [part c)]. Choose the one
for which n satisfies ps - qr = (-I)"-'.]
i) Show that if
then p l q is a convergent of a.
Hint: Suppose
and
Write
where pilqi is the ith convergent to p l q . Show that z > 1 and

apply h).
j) Of two consecutive convergents at least one satisfies
Hint: By a), if the above fails, then
Hunvitz has shown that of any three consecutive convergents at

least one satisfies
This cannot be sharpened since, for example,
for q sufficiently large. (See, for example, Hua [82].)

k) (Fibonacci numbers) Let
Fl:=F,:=l.
Fn+,:= Fn+ Fn-,
Show that {F,+,IF,) are the convergents to (1 + f i ) / 2 . Show
that
1) Show that
= [a, 2a, 2a, 2a,
. . .].
m) Show that a periodic continued fraction (one where a,+, = a, for

some I and all large k) represents a quadratic irrational (the root
of a quadratic equation with integer coefficients).
n) (Lagrange) Show that a quadratic irrational has a periodic continued fraction. Thus with m) this characterizes quadratic irrationals.
Hint: Suppose rcr2 + s a + t = 0 and a = [a,, a,, a,, . . .I. Then by
d)
7
Substitute this into the quadratic equation for a to obtain integers

A,, B,, and C, with
Express A,, B,, and C, in terms of r, s , and t and show that

A, # 0. Show that
B: - 4A,Cn = s2- 4rt

and that
Pi
Show, using f), that
and
Hence A , 1 , I B,, 1 , and I C, I are bounded independently of n. Thus

Bn3,Cng)for three distinct
( A n 1 ,Bnl, C,,) = ( A n 2 ,Bn2,Cnl)= (Ang,
indices n,, n,, and n,, and by (11.3.52) one of
which implies the periodicity of a.

o) Show, using i), that any integral solution of Pell's equation
n2 - dm2 = 1
d a positive integer
a.
has nlrn a convergent of the continued fraction of

This outline of the basic theory follows Hardy and Wright [60].
3. Suppose there exists a sequence of rationals {p,lq,) and 6 > 0 so that
and
Then either
=
4
or for
4,
for some n
> 0 and for q > c,
Hint: Let l a - p l q l = l l q
$q:+*':q < $q;. Then
1+1/6+&'
E'
>0, and choose n so that
and if plq #pnlqn,
The right-hand side, however, is less than 1 for q large.

Refine Exercise 3 to deduce Corollary 11.1, with v := 1, directly from
Theorem 11.1.
5. (Irrationality measure for el'") Let v be a positive integer. As in
Exercise 10 of section 10.1, let
4.
and
a) Show that
and that
Hence
b) As in Exercises 3 and 4, show that if plq # snltnfor some n then
for large q, where c, is a positive constant depending only on v.

c) Deduce Corollary 11.1.
6. Let d, := LCM(1, . . . ,n).
a) Show that there is a constant C so that
Pi
Hint:
n I
(r2. :,=
pi)
(2n) !
and
(2n)! 14"
n!n!
where the product is taken over the primes pi, n < p i 5 2n. Thus
b) Let ~ ( n denote
)
the number of primes less than or equal to n. The
prime number theorem asserts that
n
~(n)log n
(See, for example, Hardy and Wright [60].) Use this to prove that
.dn= O(esn)
Hint:
d,
for any 6 > 1 .
npfi
5 nT(n)
where the product is taken over the primes 5 n and where ai is the
largest integer so that pfi 5 n.
7.
a) Show that
,=e,
ala2 . ' '

- -1 x x(x + a,)
(X+ a,). . (x + a,)
b) Show that
Hint: Use a) with x := n2 and a, := - k2.

c) Set
- k)!
an,, := -21 (k!)'(n
k3(n+k)! '
Show that
(x
+ a,)
and
d) Let N-+a in c) to deduce that
(This exercise follows van der Poorten [79] and is essentially due
to Apkry.)
8. Finish the proof of Corollary 11.2.
9. (Irrationality measures for log)
a) Let p,(x) := {dldx)'xn(l
- x)".
Then .as in (11.3.23),
Estimate p,(-m). In particular show that for m E N there exist

c, > 0, c,, c,, and c, so that
Hint: By Stirling's formula, up to a power of n
(i )( n
k , rnk
behaves like
where k := a n . It is now a calculus exercise to maximize the above

by differentiating with respect to a. The maximum occurs at
a := q i q z i j .
b) Prove that for large q,
Pi
Outline: Let p, be as in (11.3.22) and let
ffn
=-
dn
+ p,(-1)
log 2
where a, is an integer and d, is as in Exercise 6. (Note that

(1 + x) 1 [p, (x) - p,(- l)].) Furthermore, integrating In by parts n
times with respect to x yields
Estimate / I / from above and below by computing that the maximum of x ( 1 - x ) l ( l + x ) occurs at x : = a - - 1 . For large n,
c)
From (11.3.56), (11.3.58), and part a) prove (11.3.55).

Prove, for E > 0 and n > N,, that
Compare (11.3.4) and (11.3.5),

Hint: Consider
In :=
1+ xlm
and proceed as in b). Note that
Show that B,(q) and B,(q) are irrational for rational q := 1 In,
n = 1 , 2 , 3, . . . .
b) (Euler) Show that m4 n4 = p 2 with m, n, p integral implies
mn = 0.
C) Use b) to, show that B2, B3, and 8, are never all nonzero and
rational.
10. a)
11. a) Show that
Outline: Let
and
Show that
and that
Thus prove (11.3.59) by showing that S,+O.

b) Evaluate 5(2) = v2/6 from a) and Gregory's formula.
12. a) Show that for 0 5 x s 1, one has the following functional equation
for the dilogarithrn C z = , x"ln2:
Hint: Show by taking derivatives that the left-hand side is zero.

Evaluate the constant by letting x-+ 0.
b) (Euler) Show that
c)
Let ~ i j ( x )denote the trilogarithm C;=, x"ln3. Show that, for

0 < x < 1, the following identity due to Landen holds:
Pi
382
d) Deduce that Li3(l) = 5(3) and that

2
7
7r
1
i) ~ i , ( + )= 8~ i , ( l ) - log 2 + - log3(2)
12
6
ii) ~
3
4
i , ( =~-J ~)i , ( l +
) - log -15
7r2
(3-2v3)
e) Show that
j,logrM+l)121
5(3) = Li3(l) = 10
t2 coth (t) dt
and combine this with Exercise 17a) of Section 5.5 to provide

another verification of Exercise 7d). This is discussed in Lewin
181, Sec. 6.31.
13. ' (Euler) Establish
a)
7r[cot (7rx) - cot ( m ) ] =
Hint: Use x := $ and
x--
a-x
(X- n)(a - n)
383
14. (Evaluation of l(2m)) Define {B,), the Bernoulli numbers, by
a) Show that
b) Show, from (11.3.61), that

TZ
cot (TZ) =
2~iz
+ riz
e2rriz- 1
and from the product expansion for sin that
c)
Show that
d) Thus
15.
(Evaluation of P(2m
(11.3.62)
+ 1))
-=
cos z
n=o
a) Show that Eo = 1 and
b) Show that
Define {E,,), the Euler numbers, by
(2n)!
1.1 < 5 .
where, as before,
c)
Show that
d) Thus
There are similar evaluations for.more general L functions.

16. (Series involving
)(:
a) Show that
Hint: Let f := (arcsin x) /d 1- x2. Show that
Show that
also satisfies the above differential equation. (Compare Exercise

16 of Section 5.5.)
b) Show that
Hint:Use a) and the fact that

d
2x arcsin x
x - (arcsin x)' =
lK-7
dx
c) Show, by differentiating in a), that
--
x
-- x2 2 + x arcsin
2 312
1-x
(1-x)
'
d) Specialize the above series or the derivatives of the above series to

show that
2 ( - ~ ) ~ --l 1
--
v3+1
log 25
2
e)
Show that
and that
Thus
f) From e) deduce that
and that
( 2 n + l)l6"
=-
'
Further material is available in Lewin [81], Lehmer [85], and

Zucker [85]. There is an interesting evaluation of Comtet's, namely,
g) We conclude with Ramanujan's
with G denoting Catalan's constant.

Hint: Use part a), and parts b) and h) of Exercise 10 in Section
5.6.
Bibliography
[I8811 N. H. Abel, Oeuvres Complktes (Grondahl and Son, Christiania, 1881).
1641 M. Abramowitz and I. Stegun, Handbook of Mathematical Functions (Dover, New
York, 1964).
W.
W. Adams, "Asymptotic Diophantine Approximation to e," Proc. Natl. Acad. Sci.
[66]
55 (1966), 28-31.
[66] L. V. Ahlfors, Complex Analysis, 2nd ed. (McGraw-Hill, New York, 1966).
[74] A. V. Aho, J. E. Hopcroft, and J. D. Ullman, The Design and Analysis of Computer
Algorithms (Addison-Wesley, Reading, MA, 1974).
[79] K. Alladi and M. L. Robinson, "On Certain Irrational Values of the Logarithm,"
Lecture Notes in Mathematics 751, (Springer, Berlin, 1966).
[Pr]
[prl
G. Almkvist, "A Solution to a Tantalizing Poblem," Fiboracci Quart. (in press).
-and B. C. Berndt, "Gauss, Landen, Ramanujan, the Arithmetic-Geometric Mean,
Ellipses, the Ladies Diary, m, and the American Revolution," MAA Monthly (in press).
G.
E. Andrews, The Theory of Partitions (Addison-Wesley, Reading, MA, 1976).
[76]
,
"An Introduction to Ramanujan's 'Lost' Notebook," MAA Monthly 86 (1979),
[79]
89-108.
"The Fifth and Seventh Order Mock Theta Functions," Trans. AMS 273 (1986),
[86] -,
113-134.
1741 T. M. Apostol, Mathematical Analysis, 2d ed. (Addison-Wesley, Reading, MA, 1974).
[76a] -,
Introduction to Analytic Number Theory (Springer, New York, 1976).
[76b] -,
Modular Functions and Dirichlet Series in Number Theory (Springer, New York,
1976).
[85] J. Arazy, J. Claesson, S. Janson, and J. Peetre, "Means and Their Iterations," in Proc.
19th Nordic Congress of Math., J.R. Stefhson (ed.), Reykjavik (1985).
[80] R. Askey, "Ramanujan's Extensions of the Gamma and Beta Functions," MAA
Monthly 87 (1980), 346-359.
[81] R. Backstrom, "On Reciprocal Series Related to Fibonacci Numbers," Fibonacci Quart.
19 (1981), 14-21.
D.
H. Bailey, "The Computation of rr to 29,360,000 Decimal Digits Using Borweins'
[88]
Quartically Convergent Algorithm", Math. Comput. 50 (1988), 283-296.
[35] W. N. Bailey, Generalized Hypergeometric Series (Cambridge Univ. Press, London,
1935).
A.
Baker, "Approximation to the Logarithms of Certain Rational Numbers," Acta
[64]
Arith. 10 (1964), 315-323.
Transcendental Number Theory (Cambridge Univ. Press, London, 1975).
[75] -,
[81] G. A. Baker and P. Graves-Morris, Pad&Approximants, vols. I and I1 (Addison-Wesley,
Reading, MA, 1981).
388
Bibliography
J. P. Ballantine, "The Best (?)Formula for Computing T to a Thousand Places," M A A

Monthly 46 (1939), 499-501.
[50] E. F. Beckenbach, "A Class of Mean Value Functions," M A A Monthly 52 (1950), 1-6.
[61] - and R. Bellman, An Introduction to Inequalities (Random House, New York,
1961).
P.
Beckmann, A History of Pi, 4th ed. (Golem Press, Boulder, CO, 1977).
[77]
[72] M. Beeler, R. W. Gosper, and R. Schroeppel, "Hakmem," MIT Artificial Intelligence
Lab., MIT, Cambridge, MA, 1972.
[27] E. T. Bell, "Algebraic Arithmetic," AMS Colloq. 7, 1927.
[61] R. Bellman, A Brief Introduction to Theta Functions (Holt, Rinehart and Winston, New
York, 1961).
[Pr] B. C. Berndt, The Jacobian Elliptic Functions, Chapter 18 of Ramanujan's Second
Notebook (Springer, to be published).
[Pr] B. C. Berndt, Modular Equations of Degrees 3, 5, and 7 and Associated Theta Function
Identities, Chapter 19 of Ramanujan's Second Notebook (Springer, to be published).
[79] F. Beukers, "A Note on the Irrationality of l(2) and 6(3)," Bull. London Math. Soc. 11
(l979), 268-272.
S.
Bhargava and Chandrashekar Adiga, "On Some Continued Fraction Identities of
[84]
Srinivasa Ramanujan," Proc. AMS 92 (1984), 13-18.
[Pr] A. J. Biagioli, "A Proof of Two Identities of Ramanujan."
[73] G. Birkhoff, A Source Book in Classical Analysis (Harvard Univ. Press, Cambridge,
MA, 1973).
and G. C. Rota, Ordinary Differential Equations, 2d ed. (Blaisdell, Waltham, MA,
[69]
1969).
[I8881 C. W. Borchardt, Gesammelte Werke (Berlin, 1888).
[66] A. Borel, S. Chowla, C. S. Herz, K. Iwasawa, and J. P. Sene, "Seminar on Complex
Multiplication," Lecture Notes in Mathematics 21 (Springer, Berlin, 1966).
[75] A. Borodin and I. Munro, The Computational Complexity of Algebraic and Numeric
Problems (American Elsevier, New York, 1975).
[86] D. Borwein and J. M. Borwein, "Alternating Series in Several Dimensions," MAA
Monthly (1986) to appear.
,
-,
and K. F. Taylor, "Convergence of Lattice Sums and Madelung's Con[85]
stant," J. Math. Phys. 26 (1985), 2299-3009.
[85] J. M. Bonvein, "Some Modular Identities of Ramanujan Useful in Approximating T,"
Proc. AMS 95 (1985), 365-371.
and P. B. Borwein, "A Very Rapidly Convergent Product Expansion for T," BIT
1831
23 (1983), 538-540.
[84a] -and -,
"The Arithmetic-Geometric Mean and Fast Computation of Elementary
Functions," SZAM Rev. 26 (1984), 351-365.
[84b] - and -,
"Cubic and Higher Order Algorithms for n," Can. Math. Bull. 27
(1984), 436-443.
[84c] -and -,
"Explicit Algebraic Nth Order Approximations to Pi," in Approximation
Theory and Spline Functions, NATO AS1 Ser., S. P. Singh, J. H. W. Burry, and B.
Watson, Eds. (ReideI, Dordrecht, 1984).
"Reduced Complexity Calculation of Log," Dalhousie Technical Report.
[84d] -,
"More Quadratically Converging Algorithms for T," Math Comput. 46
[86] -- and -,
(1986), 247-253.
[Pr] J. M. Borwein and P. B. Borwein, "Approximating T with Ramanujan's Solvable
Modular Equations" (in press).
[39]
Bibliography
389
P. B. Borwein, "On the Complexity of Computing Factorials," J . Algorithms 6 (1985),

376-380.
[prl -, "Problem," SIAM Rev. (in press).
[53] F. Bowman, Introduction to Elliptic Functions (English Universities Press, London,
1953).
C.
B. Boyer, A History of Mathematics (Wiley, New York, 1968).
1681
D.
Braess, "On the Conjecture of Meinardus on Rational Approximation to e: 11," J.
[84]
Approx. Theory 40 (1984), 375-379.
[76a] R. P. Brent, "Fast Multiple-Precision Evaluation of Elementary Functions," J. ACM 23
(l976), 242-251.
"Multiple-Precision Zero-Finding Methods and the Complexity of Elementary
[76b] -,
Function Evaluation," in Analytic Computational Complexity, J . F. Traub, Ed.
(Academic Press, New York, 1976).
[76c] -,
"The Complexity of Multiple-Precision Arithmetic," in Complexity of Computational Problem Solving, R. S. Anderssen and R. P. Brent, Eds. (Univ. of Queensland
Press, Brisbane, 1976).
and E. M. McMillan, "Some New Algorithms for High-Precision Calculation of
[80]
Euler's Constant," Math. Computat. 34 (1980), 305-312.
[83] D. M. Bressoud, "An Easy Proof of the Rogers-Ramanujan Identities," J . Number
Theory 16 (l983), 235-241.
[I8751 C. Briot and J. C. Bouquet, Theorie des Fonctions Elliptiques (Paris, 1875).
[26] T. J. I. A. Bromwich, An Introduction to the Theory of Infinite Series, 2d ed.
(Macmillan, London, 1926).
[71] P. Bundschuh, "Irrationalitatsmasse fiir en, a # 0 rational oder Liouville-Zahl," Math.
Ann. 192 (1971), 229-242.
[84] V. I. Buslaev, A. A. Gonchar, and S. P. Suetin, "On the Convergence of Subsequences
of the nth Row of a Pad6 Table," Math. U.S.S.R. Sbornik 48 (1984), 535-540.
[71] L. Carlitz, "Reduction Formulae for Fibonacci Summations," Fibonacci Quart. 9
(1971), 449-466, 510.
[71] B. C. Carlson, "Algorithms Involving Arithmetic and Geometric Means," M A A
Monthly 78 (1971), 496-505.
"The Logarithmic Mean," M A A Monthly 79 (1972), 615-618.
[72] -,
[75] ---, "Invariance of an Integral Average of a Logarithm," MAA Monthly 82 (1975),
379-382.
---,
Special Functions of Applied Mathematics (Academic Press, New York, 1977).
[77]
,
"Short
Proofs of Three Theorems on Elliptic Integrals," SIAM J . Math. Anal. 9
[78]
(1978), 524-528.
[57] J. W. S. Cassels, An Introduction to Diophantine Approximation (Cambridge Univ.
Press, London, 1957).
[I8741 A. Cayley, "A Memoir on the Transformations of Elliptic Functions," Phil. Trans.
164 (1874), 397-456.
[I8951 -,
An Elementary Treatise on Elliptic Functions (Bell and Sons, 1895; republished
Dover, New York, 1961).
[85] K. Chandrasekharan, Elliptic Functions (Springer, Berlin, 1985).
[66] E. W. Cheney, Introduction to Approximation Theory (McGraw-Hill, New York, 1966).
[84] D. V. Chudnovsky and G. V. Chudnovsky, "Pad6 and Rational Approximation to
Systems of Functions and Their Arithmetic Applications," Lecture Notes in Mathematics
1052 (Springer, Berlin, 1984).
A.
Clark, Elements of Abstract Algebra (Wadsworth, Belmont, CA, 1971).
[71]
[85]
Bibliography
J. E. Cohen and R. D. Nussbaum, "Arithmetic-Geometric Means of Positive Matrices"

Math. Proc. Camb. Phil. Soc. 101 (1987), 209-219.
S. A. Cook and S. 0. Aandeiaa, "On the Minimum Computation of Functions," Trans.
AMS 142 (1969), 291-314.
J. W. Cooley and J. W. Tukey, "An Algorithm for the Machine Calculation of Complex
Fourier Series," Math. Comp. 19 (1965), 297-301.
-, P. A. Lewis, and P. D. Welch, "History of the Fast Fourier Transform," Proc.
ZEEE 55 (1967), 1675-1677.
D. A. Cox, "The Arithmetic-Geometric Mean of Gauss," L'Enseignement Math. 30
(1984), 275-330.
-, "Gauss and the Arithmetic-Geometric Mean," Notices of the AMS 32 (1985),
147-151.
J. H. Davenport, "On the Integration of Algebraic Functions," Lecture Notes in
Computer Science 102 (Springer, Berlin, 1981).
C. S. Davis, "A Note on Rational Approximation," Bull. Austral. Math. Soc. 20 (1979),
407-410.
C. Denninger, "On the Analogue of the Formula of Chowla and Selberg for Real
Quadratic Fields," J. Reine Ang. Math. 351 (1984), 171-191.
L. E. Dickson, Introduction to the Theory of Numbers (Dover, New York, 1929).
-, History of the Theory of Numbers, vols. 1-111 (1919, 1920, 1923), reprinted
Chelsea, New York, 1971).
M. Dutta and L. Debnath, Elements of the Theory of Elliptic and Associated Functions
(World Press Private, Calcutta, 1965).
P. Du Val, Elliptic Functions and Elliptic Curves (Cambridge Univ. Press, London,
1973).
A. Eagle, The Elliptic Functions as They Should Be (Galloway and Porter, Cambridge,
1958).
C. H. Edwards, Jr., The Historical Development of the Calculus (Springer, New York,
1979).
A. ErdClyi et al., Higher Transcendental Functions, vols. 1-3 (McGraw-Hill, New York,
1953-1955).
J. A. Ewell, "An Easy Proof of the Triple-Product Identity," M A A Monthly 88 (1981),
270 -272.
, "Consequences of Watson's Quintuple Product Identity," Fibonacci Quart. 20
(1982), 256-262.
-, "A Simple Proof of Fermat's Two-Square Theorem," M A A Monthly 90 (1983),
635-637.
J. A. Ewell, "On the enumerator for sums of three squares," Fibonacci Quart. 24
(1986), 151-153.
[84a] D. M. E. Foster and G. M. Phillips, "The Arithmetic-Harmonic Mean," Math. Comput.
42 (1984), 183-191.
[84b] -and -,
"A Generalization of the Archimedean Double Sequence," J. Math.
Anal. Appl. 101 (1984), 575-581.
[I8661 C. F. Gauss, Werke (Gottingen, 1866-1933), vol. 3, pp. 361-403.
[38] C. Gini, "Di una Formula Compressiva delle Medie," Metron. 13 (1938), 3-22.
[80] M. L. Glasser and I. J. Zucker, "Lattice Sums in Theoretical Chemistry," Advances and
Perspectives 5 (1980), 67-139.
D.
Goldfeld, "Gauss' Class Number Problem for Imaginary Quadratic Fields," Bull.
[85]
AMS 13 (1985), 23-37.
Bibliography
391
B. Gordon, "Some Identities in Combinatorial Analysis," Quart. J. Math. 12 (1961),

285-290.
[84] H. W. Gould and M. E. Mays, "Series Expansions of Means," J. Math. Anal. Appl. 101
(1984), 611-621.
[I8811 E. M. Goursat, "Sur 1 ' ~ ~ u a t i oDiffCrentielle
n
LinCaire qui Admet pour IntCgrale la
SCrie Hypergiomttrique," Ann. Sci. I'Ecole Normale Sup. (2) 10 (1881), 3-142.
[SO] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series and Products (Academic
Press, New York, 1980).
[I8921 A. G. Greenhill, The Applications of Elliptic Functions (1892; republished Dover, New
York, 1959).
[09] H. Hancock, Theory of Elliptic Functions (1909, republished Dover, New York, 1958).
[28] M. Hanna, "The Modular Equations," Proc. London Math. Soc. 28 (1928), 46-52.
[75] E. R. Hansen, A Table of Series and Products (Prentice-Hall, Englewood Cliffs, NJ,
1975).
G.
H. Hardy, Ramanujan (Cambridge Univ. Press, London, 1940).
[40]
-,
J. E. Littlewood, and G. Polya, Inequalities, 2d ed. (Cambridge Univ. Press,
[59]
London, 1959).
[60] -and E. M. Wright, An Introductibn to the Theory of Numbers, 4th ed. (Oxford
Univ. Press, London, 1960).
[77] P. Henrici, Applied and Computational Complex Analysis, vols. 1 and 2 (Wiley, New
York, 1974, 1977).
[12] C. Hermite, Oeuvres, vol. 3 (Paris, 1912), pp. 150-181.
[17] ---, Oeuvres, vol. 4 (Paris, 1917), pp. 357-377.
[05] -and T. J. Stieltjes, Correspondence d'Hermite et de Stieltjes, vol. 2 (GauthiersVillars, Paris, 1905).
[I8931 D. Hilbert, Math. Ann. 43 (1893), 216-219.
[85] M. D. Hirschhorn, "A Simple Proof of Jacobi's Two Square Theorem," M A A Monthly
92 (1985), 579-580.
[70] A. S. Householder, The Numerical Treatment of a Single Nonlinear Equation (McGrawHill, New York, 1970).
L.
K. Hua, Introduction to Number Theory (Springer, Berlin, 1982).
[82]
D.
Hughes, Master's Thesis, Dalhousie Univ., Halifax, NS, 1984.
[84]
[I7961 J. Ivory, "A New Series for the Rectification of the Ellipsis; Together with Some
Observations on the Evolution of the Formula (a2 + b2 - 2ab(cos +)")," Trans. Roy.
Soc. Edinburgh 4 (1796), 177-190.
[I8291 G. J. Jacobi, "Fundamenta Nova Theoriae Functionum Ellipticarum," 1829; Jacobi's
Gesammelte Werke, vol. 1 (Berlin, 1881-1891), pp. 49-239.
[84] E. Kaltofen and N. Yui, "On the Modular Equation of Order 11," Proc. 1984
MACSYMA User's Conf., pp. 472-485.
[Pr] Y. Kanada, Y. Tamura, S. Yoshino, and Y. Ushiro, "Calculations of rr to 10,013,395
Decimal Places Based on Gauss-Legendre Algorithm and Gauss Arctangent Relation,"
preprint (1983).
[57] I. Kaplansky, An Introduction to Differential Algebra (Hermann, Paris, 1957).
[86] M. Keith, "Circle digits a self-referential story," The Mathematical Intelligencer 8
(1986), 56-57.
[64] A. Khinchin, Continued Fractions (Chicago Univ. Press, Chicago, 1964).
[24] L. V. King, On the Direct Numerical Calculation of Elliptic Functions and Integrals
(Cambridge Univ. Press, London, 1924).
[61]
Bibliography
392
[I8971 F. Klein, Famous Problems of Elementary Geometry, trans]. by W. Beman and D. E.

Smith (Ginn and Co., Boston, 1897).
"Development of Mathematics in the 19th Century," 1928, Trans., Math. Sci.
[79] -,
Press, R. Hermann Ed. (Brookline, MA, 1979).
[I8921
and R. Fricke, Vorlesungen uber die Theorie der Elliptischen Modulfunktionen
(Teubner, Leipzig, 1890, 1892).
[81] D. Knuth, The Art of Computer Programming, vol. 2: Seminumerical Algorithms 2nd
ed. (Addison-Wesley, Reading, MA 1981).
A. Krazer, Lehrbuch der Theta-Funktionen (Chelsea, New York, 1971).

H. T. Kung and J. F. Traub, "All Algebraic Functions Can be Computed Fast," J.
ACM 25 (1978), 245-260.
[I8991 E. Landau, "Sur la Strie des Invers de Nombres de Fibonacci," Bull. Soc. Math. France
27 (1899), 298-300.
[58] E. Landau, Elementary Number Theory (Chelsea, New York, 1958).
[66a] S. Lang, Introduction to Transcendental Numbers (Addison-Wesley, Reading MA,
1966).
[66b] -,
Introduction to Diophantine Approximation (Addison-Wesley, Reading, M A
1966).
[73] -,
Elliptic Functions (Addison-Wesley, Reading, M A , 1973).
[78] E. B. Leach and M. C. Sholander, "Extended Mean Values," M A A Monthly 85 (1978),
84-90.
[1811] A. M. Legendre, Exercises de Calcul Int&graI, vols. 1-3 (Paris, 1811-1819).
[38] D. H. Lehmer, "On Arctangent Relations for n," M A A Monthly 45 (1938),
657-664.
"On the Compounding of Certain Means," J. Math. Anal. Appl. 36 (1971),
[71] -,
183-200.
"Interesting Series Involving the Central Binomial Coefficient," M A A Monthly 92
[85] -,
(1985), 449-457.
[66] J. Lehner, A Short Course in Automorphic Functions (Holt, Rinehart and Winston,
New York, 1966).
[SO] Le Petit Archimzde, no. 64-65, "Numkro Spicial .rr" (1980).
[77] W. J. LeVeque, Fundamentals of Number Theory (Addison-Wesley, Reading, MA,
1977).
[81] L. Lewin, Polylogarithms and Associated Functions (North Holland, New York, 1981).
[74] Tung-Po Lin, "The Power Mean and the Logarithmic Mean,' MAA Monthly 81 (1974),
879-883.
[I8821 F. Lindemann, " ~ b e rdie Zahl .rr," Math. Ann. 20 (1882), 213-225.
[I8841 J. Liouville, "Sur des Classes Trks ten dues de Quantitks dont la Valeur n'est ni
Rationnelle ni m6me Rkductible B des Irrationnelles Algibriques," Compt. Rend. 18
(1844), 883-885, 910-911.
3.
D. Lipson, Elements of Algebra and Algebraic Computing (Addison-Wesley, Reading,
[81]
MA, 1981).
[66] W. Magnus, Formulas and Theorems for the Special Functions of Mathematical Physics
(Springer, Berlin, 1966).
K.
Mahler, "Zur Approximations der Exponentialfunktion und des Logarithmus, I, 11,"
[31]
J. Reine Ang. Math. 166 (1931), 118-136, 137-150.
"On the Approximation of T," Indagiones Math. 15 (1953), 30-42.
[53] -,
[71]
[78]
Bibliography
393
-, "Application
of Some Formulae by Hermite to the Approximation of Exponentials and Logarithms," Math. Ann. 168 (1967), 200-227.
D. Masser, "Elliptic Functions and Transcendence," Lecture Notes in Mathematics 437
(Springer, Berlin, 1975).
G. Miel, "On Calculations Past and Present: The Archimedean Algorithm," M A A
Monthly 90 (1983), 17-35.
M. Mignotte, "Approximations Rationalles de ?r et Quelques Autres Nombres," Bull.
Soc. Math. France, Mem. 37 (1974), 121-132.
D. S. Mitrinovic, Analytic Inequalities (Springer, New York, 1970).
L. J. Mordell, "Note on Class Relation Formulae," Messenger of Math. 45 (1916),
75-80.
G. Nemeth, "Relative Rational Approximation of the Function ex," Math. Notes 21
(1977), 325-328.
R. Nevanlinna and V. Paatero, Introduction to Complex Analysis (Addison-Wesley,
Reading, MA, 1969).
D. J. Newman, "Approximation with Rational Functions," AMS Regional Conf. Ser. in
Mathematics, no. 41 (1979).
-,
"Rational Approximation Versus Fast Computer Methods," Lectures on Approximation and Value Distribution (Presses de l'Universit6 de MontrBal, Montreal, 1982),
pp. 149-174.
-, "A Simplified Version of the Fast Algorithms of Brent and Salamin," Math.
Comput. 44 (1985), 207-210.
M. Newman and D. Shanks, "On a Sequence Arising in Series for T," Math. Comput.
42 (1984), 191-217.
I. Niven, "Irrational Numbers," Carus Math. Monograph 11 (MAA, 1956).
M. Nyvoll, 'Tilnoermelsesformler for Ellipsebuer," BIT 25-26 (1978), 70-72.
J. Oxtoby, Measure and Category, 2d ed. (Springer, New York, 1980).
G. M. Phillips, "Archimedes the Numerical Analyst," M A A Monthly 88 (1981),
165-169.
-, "Archimedes and the Complex Plane," M A A Monthly 91 (1984), 108-114.
H. Rademacher, Topics in Analytic Number Theory (Springer, Berlin, 1973).

E. D. Rainville, Special Functions (Macmillan, New York, 1960).
K. G. Ramanathan, "On Ramanujan's Continued Fraction," Acta Arith. 43 (1984),
209-225.
S. Ramanujan, "Modular Equations and Approximations to ?r," Quart. J. Math. 45
(1914), 350-372.
-, Collected Papers (Chelsea, New York, 1962).
R. Rankin, Modular Forms and Functions (Cambridge Univ. Press, London, 1977).
H. E. Rauch and A. Lebowitz, Elliptic Functions, Theta Functions and Riemann
Surfaces (Williams and Wilkins, Baltimore, MD, 1973).
P. Ribenboim, "Consecutive Powers," Expo. Math. 2 (1984), 193-221.
V'
-, "Representation of Real Numbers by Means of Fibonacci Numbers," L'Enseignement Mathematiques 3 (1985), 249-259.
J. F. Ritt, "Meromorphic Functions with Addition or Multiplication Theorems," Trans.
AMS 29 (1927), 341-360.
-, Integration in Finite Terms (Columbia Univ. Press, New York, 1948).
M. Rosenlicht, "Integration in Finite Terms," M A A Monthly 79 (1972), 963-972.
E. Salamin, "Computation of rr Using Arithmetic-Geometric Mean," Math. Comput. 30

(1976), 565-570.
T.
Sasaki and Y. Kanada, "Practically Fast Multiple-Precision Evaluation of log (x)," J.
[82]
Inf. Process. 5 (1982), 247-250.
[50] H. C. Schepler, "The Chronology of Pi," Math. Mag. (1950), 165-170, 216-228,
279-283.
[77] I. J. Schoenberg, "On the Arithmetic-Geometric Mean," Delta 7 (1977), 49-65.
Mathematical Time Exposures (MAA, Washington, DC, 1982).
[82] -,
B.
Schoenberg,
Elliptic Modular Functions (Springer, Berlin, 1976).
[76]
A.
Schijnhage
and
V. Strassen, "Schnelle Multiplikation Grosser Zahlen," Computing 7
1711
(1971), 281-292.
[67] A. Selberg and S. Chowla, "On Epstein's Zeta-Function," J. Reine Ang. Math. 227
(1967), 86-110.
D.
Shanks, "Dihedral Quartic Approximations and Series for m," J. Number Theory 14
[82]
(1982), 397-423.
J. W. Wrench, Jr., "Calculation of rr to 100,000 Decimals," Math. Comput. 16
[62] -and
(1962), 76-99.
[I8531 W. Shanks, Contributions to Mathematics Comprising Chiefly of the Rectijication of the
Circle to 607 Places of Decimals (G. Bell, London, 1853).
[85] D. Singmaster, "The Legal Values of Pi," Math. Zntelligencer 7 (1985), 67-72.
1661 L. J. Slater, Generalized Hypergeometric Functions (Cambridge Univ. Press, London,
1966).
E.
U. Stickel, "Fast Computation of Matrix Exponential and Logarithms," Analysis 5
[85]
(1985), 163-173.
[74] K. B. Stolarsky, Algebraic Numbers and Diophantine Approximation (Marcel Dekker,
New York, 1974).
"Generalizations of the Logarithmic Mean," Math. Mag. 48 (1975), 87-92.
1751 -,
"The Power and Generalized Logarithmic Means," MAA Monthly 87 (1980),
1801 -,
545-548.
[63] D. Sweeney, "On the Computation of Euler's Constant," Math. Comput. 17 (1963),
170-178.
[Pr] Y. Tamura and Y. Kanada, "Calculation of rr to 4,196,393 Decimals Based on
Gauss-Legendre Algorithm," Math Comput. (in press).
[I8931 J. Tannery and J. Molk, Fonctions Elliptiques, vols. 1 and 2 (1893; republished Chelsea,
New York, 1972).
1491 J. Todd, "A Problem on Arctangent Relations," MAA Monthly 8 (1949), 517-528.
[75] -, "The Leminiscate Constants," Communication of the ACM 18 (1975), 14-19.
Basic Numerical Mathematics, vol. 1 (Academic Press, New York, 1979).
[79] -,
[84] L. Trefethen, "The Asymptotic Accuracy of Rational Best Approximations to ez on a
Disk," J . Approx. Theory 40 (1984), 380-383.
[65] F. G. Tricomi, "Sull 'Algoritmo Iterativo del Borchardt e su di una sua Generalizzazione," Rend. Arc. Mat. Palermo 14 (1965), 85-94.
[79] A. van der Poorten, "A Proof that Euler Missed. . . ,ApBry's Proof of the Irrationality
of 5(3)," Math. Zntelligencer 1 (1979), 195-203.
1851 S. Wagon, "Is rr Normal," The Mathematical Zntelligencer 7 (1985), 65-67.
[74] M. Waldschmidt, "Nombres Transcendants," Lecture Notes in Mathematics 402 (Springer, New York, 1974).
1761
Bibliography
395
J. L. Walsh, "Interpolation and Approximation by Rational Functions in the Complex

Domain," AMS Colloq. Publ. 20, 5th ed. (1969).
[Pr] G. Walz, "Effiziente Berechnung spezieller Funktionen mittels asymptotischer Entwicklungen und Eliminationsprozeduren (Mannheim).
1291 G. N. Watson, "Theorems Stated by Ramanujan (VII): Theorems on Continued
Fractions," J . London Math. Soc. 4 (1929), 39-48.
"Some Singular Moduli (I)," Quart. J. Math. 3 (1932), 81-98.
[32] ,
,
"The
Marquis and the Land Agent," Math. Gaz. 17 (1933), 5-17.
[33]
,
"Generating
Functions of Class Numbers," Compos. Math. 1 (1935), 39-68.
[35]
H.
Weber,
Lehrbuth
der Algebra (1908 republished Chelsea, New York, 1980).
[08]
K.
Weierstrass,
Werke,
vol. 2 (republished Frankfurt, 1967), pp. 341-362.
[67]
A.
Weil,
Elliptic
Functions
According to Eisenstein and Kronecker (Springer, New York,
[76]
1976).
[27] E. T. Whittaker and G.N. Watson, A Course of Modern Analysis, 4th ed. (Cambridge
Univ. Press, London, 1927).
J.
Wimp, Computation with Recurrence Relations (Pitman, London, 1984).
[84]
S.
Winograd, "Arithmetic Complexity of Computations," SIAM Reg. Conf. Ser. in
[80]
Applied Mathematics 33 (1980).
[60] J. W. Wrench, Jr., "The Evolution of Extended Decimal Approximations to m," Math.
Teacher 53 (1960), 644-650.
I.
J. Zucker, "The Evaluation in Terms of r-Functions of the Periods of Elliptic Curves
[77]
Admitting Complex Multiplication," Math. Proc. Cambridge, Phil. Soc. 82 (1977),
111-118.
,
"The Summation of Series of Hyperbolic Functions," SIAM J. Math. Anal. 10
[79]
(1979), 192-206.
"Some Infinite Series of Exponential and Hyperbolic Functions," SIAM J. Math.
[84] -,
Anal. 15 (1984), 406-413.
''On the Series EL, (?)-'k-'' and Related Sums," J. Number Theory 10 (1985),
[85] -,
92-102.
[76a] -and M. M. Robertson, "Some Properties of Dirichlet L-Series," J. Phys. A 9
(1976), 1207-1214.
A Systematic Approach to the Evaluation of Ef,,,,,,,)(am2 + bmn +
[76b] -and -,
cn2)-'," J . Phys. A 9 (1976), 1215-1225.
[69]
Symbol List
Symbol
Description
Arithmetic-Geometric Mean Iteration
Section
Formula
1.1.1 and
1.1.2
Common limit of the AGM
1.1.5
Common limit of the AGM

Variables in the AGM
Legendre form of the AGM
k :
1.1.11
Complementary modulus
the derivative off

1.3.1
complete elliptic integral of the 1st kind
1.3.2
complete elliptic integral of the 2nd kind
K'(k) := K(kl)
1.3.3
Ef(k):= E(kl)
1.3.4
Hypergeometric function
1.3.5
.. :(2i-1)
(2i-1)!!:=1-3.5.
1.4.4
1.4.5
( x ) :=
e t
Gamma function
1.6.4
Symbol
Description
( x , ) :=
1-t
)d ,
Beta function
Weierstrass function
r2(n):=number of representations of n
as a sum of two squares
Oi(s):= Oi(q)
where
q := e-""
k ( q ) := k = '3;(q)1'3:(q)
k 1 ( q ):= k' = 0 : ( q ) 1 ' 3 : ( ~ )
W )=
639)
= e-K'(k)IK(k)
k(s) := k ( q )
where
q := e-"
General theta functions

Theta functions in q , z := 0
Theta functions in z , q suppressed
Theta functions in z , q and j suppressed
A*@):= k ( q )
where
q = e-&
Section
Formula
Symbol
Description
Section
3.2
3.2
q 1 1 1 2 ~ o , Eta function
3.2
q-lIz4~3
3.2
v9q"'2~l
3.2
q-1'24~z
3.2
3.2
3.2
3.2
Gaussian or q-binomial coefficients
3.3
3.3
r4(n) := number of representations of n as a sum

of four squares
3.5
3.5
w(n) := ul(n)
+ ul(odd (n))
3.5
9-1
S(p, q) := C,
e-""Z~'4
3.5
r=o
M( f ) := M*(f ) :=
f(x)*-ld.
Mellin transform
3.6
c,
[(s) :=
C n-'
n=l
L(p):=
Riemann zeta function
pn
x- -
"=I 1- p n ,
3.6
3.6
(PI < 1 ,
Fo:=O,Fl:=l,F,,+,:=F,+F ,-,,
Lambert series
Fibonaccinumbers
3.7
3.7
(Ex. 3)
LO:=2,L,:=l,L,+,:=L,+L
,-,,
Lucasnumbers
3.7
X := { i m ( ~>) 0)
X*:= X U {im)U {Q)
4.3
Inhomogeneous modular group
4.3
4.3
4.3
Fundamental set of r-group
4.3
Fundamental set of A-group
4.3
Formula
Symbol
Description
Section
4.3
(Ex. 1)
4.3
( E x . 1)
4.3
(Ex.1)
4.3
(Ex. 1)
J(t) := -
- A(*)
h2(t)(l - ~(t))' '
+
27
{(:
Klein's absolute invariant
4.3
:):ad-bc=p;a,b,c,dEZ
transformations of order p
pth order modular equation for A

Fp(x,j)=n(x-ji),
ji:=joAi
r=o
pth order modular equation for j

Q adjoin the pth roots of unity
P
n , ( v , U ) :=
I1 (V - u ; )
i=O
Mp(l, k ) : =
O'(q) K(k)
-e ; ( q l l p ) K(1) '
Multiplier of order p
4.6
Formula
Symbol
Description
Section
~ ~ (k )1: =
, -.pkk"
d M p (I, k ) + M;'(I, k)12- pk2
MP(Lk ) dk
a( p) := RP(k1,k ) ,
k := e-+
For n E N, compute k,,,
5.2
by sdving wP(k:, k:+,) = 0
m,, := ~ ; ' ( k , , ,k + , )
:= [rn,r,
+ m : ( l + k:) - p ( l +
5.4
5.4
k:+,)]/3
5.4
a,,,, := m i a , - pnv%,,
T(a + n )
(a)" := -= a(a + 1) . . . (a + n - I ) ,
r(a)
G :=
.=o
b , = fl(a,)
.l.Z!L
(2n + 1)' '
5.4
5.4
r" := RP(k">k"+J
E,
5.2
rising factorial
Catalan's constant
if b, = O(a,)
and
a, = O(b,)
5.5
5.6
(Ex. 10)
6.1
Bit complexity
6.1,
Operational complexity
6.1
Fast Fourier Transform

Bit complexity of multiplication
6.2
6.4
6.3
(Ex. 3)
Bit complexity of division
6.4
Bit complexity of root extraction
6.4
a A b := min(a, b )
8.1
a v b :max(a, b )
8.1
t,(x) := M(x, I ) , M a mean, the trace
8.1
- {O)
Mf(a, b ) := f - ' ~ ( f ( a ) f@))
,
Mp denotes Mf when f(x) = xP
8.1
8.1
8.1
Formula
Symbol
Description
a
b)=
aP +
bP
Section
'/P
a, b > 0.
Holder means
Lehmer means
Stolarsky means
3 ( a , b ) := &(a, b ) ,
Logarithmic mean
$(a, b ) := &(a, b ) ,
Identricmean
dominance
M ( 4 ( a ) , 4 ( b ) ) = (P(N(a,b ) ) ,
if M > 4,N and N > &M,
equivalence
Compound Mean of M and N
M comparable to N
The Gaussian product (Def. 8.2)

The Gaussian mean iterative process (Def. 8.2)
The Archimedean product (Def. 8.2)
The Archimedean mean iterative process (Def. 8.2)
N*(a, b ) = N(M(a, b ) , b )
~ ( ab ), := M(a
Cij:=Fi@q
b, a v b)
(i, j=1,2,3,4)
Formula
Symbol
Description
Section
8.7
L p ( 4 :=
H i ) :=
Lehmer means
(i
a,)/(:
ar-q ,
,=I
,=I
(N 5 a ) ' , HOlder means
8.7
8.7
i=1
The common limit (when it exists) of [MI, . . . , M"]

a',+, := Mi(<), or vectorially
8.7
<+,= M(c,,)
8.7
N-dimensional Gaussian product
8.7
N-dimensional Archimedean product
8.7
He(a, c) := a
8.7
Heronian mean
8.7
8.7
8.8
The algebraic transformation group off over F
9.2
9.2
9.2
9.2
9.2
9.2
(Ex. 6)
9.4
(Ex. 7 )
MJk"
>
9.5
kpd
Np := v2(q)1v2(q1'p) ,
~,:={z@IIzI56}
given by 3.2.9
9.5
10.1
Algebraic polynomials of degree n
10.1
llf llA := ;g lf(4I
10.1
EAf, A):=,"E$Ilf - pllA
10.1
Formula
Description
Symbol
= p ~I g
l f -nP / ~ ~ I A
R n ( f ,4
R , ( f , A ) := R J f )
y := l h [ 1 + 2
dn
d , :=' L C M ( ~2,, . . . ,n)
[ao,a , , . . . ,a,,]
1
[a,, a , , . . . ,a,] := a, + ---1
a, + . a2 +
1
+ -13 + - - + m - log m ] ,
Euler's constant
Section
Formula
10.1
10.1.3
10.2
10.2.20
11.3
11.3
(Ex. 2 )
simple continued fraction

[a,, a , , . . .I
lim [a,, a , ,
n--
. . ., a , ]
11.3
(Ex. 2 )
4
B,=--
and
B,,+,:=O,
2
Bernoulli numbers
1
(-1)"~,,2"
cos z ,=,
(2n)!
Euler numbers
-=
n=l,2,3,
?T
IZI< z'
Author Index
Numbers in parenthesis indicate references in the Bibliography. "Pr" indicates references in press.
Abel, 5, 28
Abramowitz and Stegun (64), 15, 321
Adams (66).37 1
Aho, Hopcroft and Ullman (74), 201, 203, 206,
21 1
Alladi and Robinson (79), 371
Almqvist (Pr), 99
Almqvist and Berndt (Pr), 3, 196, 371
Andrews (76), 67, 86
Andrews (86). 286
Apery, 365, 37 1, 379
Apostol (74), 37, 40
Apostol (76a). 86
Apostol (76b), 116
Arazy et al. (85), 269, 272
Archimedes, 243, 249, 337, 338
Aristophanes, 347
Askey (go), 89, 308
Backstrom @I), 98, 99
Bailey, D. (88), 211, 341, 342
Bailey, W., 181
Bailey, W. (35), 184
Baker, 323, 348
Baker (75), 349, 351, 354, 356, 363, 371
Baker and Graves-Morris @I), 320, 323
Ballantine (39). 347
Baxter, 80
Beckenbach (50). 236, 269
Beckmann, 346
Beckmann (77). 342
Beeler et a/. (72). 220, 222
Bell (27), 36
Bellman (61), 39, 55, 56, 86, 89
Benson, 301,302, 303
Berndt (Pr), 142, 150, 177, 196, 309
Bernoulli, Jacques, 340

Bernoulli, Jean, 340
Beukers, 195
Beukers (79), 365
Bhargava and Chandrashekar Adiga (84). 81
Biagioli (FV), 146
Birkhoff.(73), 10, 18, 116
Birkhoff and Rota (69), 9
Bohr, 89
Borchardt, 250, 256, 257, 263
Borchardt (1888), 272
Borodin and Munro (75). 206, 208
Borwein, Borwein and Taylor (85). 290, 291,
292
Bonvein, D. and Bonvein, J. (86), 291
Borwein, J.M. (85). 163, 164
Bonvein. P.B. (85), 218
Borwein, P.B. (Pr), 254
Borwein and Borwein (84a). 44. 50, 51. 222.
223
Bonvein and Borwein (84b), 108, 174
Borwein and Borwein (84c), 174
Borwein and Borwein (84d), 222,225
Borwein and Borwein (86), 174
Borwein and Borwein (FV), 313
Bouyer, 341
Bowman (531, 29
Boyer (681, 347
Braess (841, 325
Brent, 51
Brent (76a), 48, 222, 229
Brent (76b). 222
Brent (76c), 213, 216, 217, 222, 329, 330
Brent and McMillan (80). 336
Bressoud, 78
Bressoud (83),77
Author Index
Bring, 136
Bromwich (26), 189, 291, 344
Brouncker, 338, 345
Bundschuh (71), 371
Buslaev, Gonchar and Suetin (84), 323
Carlitz (71), 71
Carlson, 248, 257, 261
Carlson (71), 7, 256
Carlson (73, 261
Carlson (78). 262
Catalan, 66
Cauchy, 76, 77, 292
Cayley (1874), 106, 133, 138, 140, 142
Cayley (1895). 22, 102, 138
Chandrasekharan (85). 116
Cheney (66), 321
Chudnovsky and Chudnovsky (84). 368, 371
Clark (7 I), 349
Clausen, 179, 188, 189, 190
Cohen and Nussbaum (87), 273
Comtet, 386
Cook, 69, 213
Cook and Aanderaa (69), 69
Cooley, Lewis and Welch (67), 65
Cooley and Tuckey (65), 65
Cox (85), 15
Dase, 340
Davenport (8 1 ), 226
Davis (79), 365, 37 1
Denninger (84). 89
Dickson (29), 293
Dickson (71), 55, 66, 85, 141, 284, 287, 292
Dirichlet, 86
DuVal (73), 23, 30, 133
Eagle (58), 29
Edwards (79), 252
Epstein, 303
Erdelyi er a/. (53), 10, 30, 55, 178, 179, 181
Euler, 3, 18, 64, 67, 77, 189, 264, 286, 306,
307, 340, 344, 346, 348, 381, 382, 383
Ewell @I), 66
Ewell (82), 306
Ewell (83), 86
Ewell (86), 151
Feldman, 371
Felton, 341
Ferguson, 340
Fermat, 82, 286
Foster and Phillips (84a). 260
Foster and Phillips (84b). 247, 252
Gammel, 323
Gauss, 1, 3, 5, 8, 10, 28, 43, 51, 85, 86, 250,
256, 284, 286
Gauss (1866). 5, 7, 35, 44, 48, 52, 65
Gelfond, 348
Gini (38), 233
Glasser and Zucker (go), 68, 290, 293, 303
Goldfield (85), 295
Gordon (61). 306, 308
Gosper, 341
Gould and Mays (84). 263
Goursat (l881), 185, 186
Gradshteyn and Ryzhik (go), 10
GravC, 345
Graves-Morris, 323
Greenhill (1892), 133, 167
Gregory, 339, 348, 381
Guilloud, 341
Hancock (09), 29
Hanna (28), 133
Hardy, 80, 187
Hardy (40). 141, 314
Hardy, Littlewood and Polya (59). 235, 242,
269
Hardy and Wright, 80
Hardy and Wright (60). 66, 67, 85, 86, 95,
281, 284, 349, 376, 378
Hermite, 135, 264, 348, 359
Hermite and Stieltjes (05). 264
Hilbert, 348. 353, 354
Hiibert (1893). 348
Hirschhorn (85), 285
Holloway, 188
Householder (70), 2 15, 2 16
Hua (82), 294, 349, 375
Hughes (84), 175
Jacobi, 5, 10, 20, 22, 28, 38, 52, 65, 68, 73,
74, 85, 106, 143, 180, 252, 261, 284, 301,
306
Jacobi (18291, 36, 106
Jones, 339
Joubert, 142
Kaltofen and Yui (84), 133
Kanada, 224, 225, 341
Karatsuba, 21 1
Keith (86). 342
King (24), 16, 18, 44, 52, 59, 60
Kiein, 116
KIein (1897), 349
Author Index
Klein (79), 7
Klein and Fricke (1892), 116, 126
Knuth @I), 203, 21 1, 212, 218
Kronecker, 135, 293, 296, 297
Kummer, 179
Kung and Traub (78), 216
Lagrange, 3, 82, 321, 375
Laguerre, 99
Lambert, 99, 348
Landau, 268, 269
Landau (1899), 94
~ a n d a u(58), 86, 294
Landen, 381
Lang (66), 349
Lang (73), 29, 30, 116, 123
Leach and Scholander (78), 237
Legendre, 3, 10, 23, 26, 152, 153, 154, 178,
348
Lehmer, 264, 265
Lehmer (38), 345
Lehmer (71), 247, 263
Lehmer (85), 386
Lehner (66), 1 16
Leibniz, 339, 340
Le Petit Archimede (80), 342
LeVeque (77), 294, 349
Lewin (81), 382, 386
Lindemann, 274, 348
Lindemann (1882), 348
Liouville, 10, 36, 116, 348
Lipson (81), 216
Lorenz, 287, 289
Machin, 339, 341, 344
MacMahon, 67
Mahler (3 I), 359
Mahler (53). 368
Mahler (67), 349, 368
Meil (83), 250
Meinardus, 321
Metropolis, 341
Mignotte (74), 368
Mobius, 100
Mollerup, 89
Mordell, 187
Mordell (l6), 286
Nemeth (77), 321
Newman (79), 32 1
Newman (82), 7, 222, 223
Newman (85), 7
Newman, M., and Shanks (84). 193
Newton, 123, 207, 208, 212, 339, 343, 356
Niven (56), 347, 348, 349

Nyvoll (78), 196
Padt, 320
Pfaff, 250
Phillips (8 I), 252
Picard, 119
PoincarC, 1 16
Poisson, 38
Polya, 86
Rademacher (73), 56, 71, 85, 90, 286
Rainville (60), 179
Ramanathan (84), 8 1
Ramanujan, 69, 80, 81, 84, 100, 138, 140, 144,
146, 150, 154, 157, 158, 161, 164, 179, 184,
186, 187, 188, 194, 195, 197, 287, 295, 306,
307, 308, 309, 310, 311, 312313, 314, 315,
386
Ramanujan (14). 158, 163, 179, 184, 186, 191,
193, 195, 313
Ramanujan (62), 80, 187, 28 1, 286
Rankin (77), 116
Reitwiesner, 341
Ribenboim (84), 345
Ribenboim (85), 95
Riemann, 89
Ritt (48), 226
Rogers, 80
RouchC, 317, 319
Salamin, 51, 220, 222
Salamin (76), 48, 49, 51
Sasaki, 225
Sasaki and Kanada (82), 224, 225
Schepler (50), 342
Schlafli, 147, 297
Schlomilch, 268, 269
Schneider, 348
Schoeneberg (76), 116, 123, 126
Schoeneberg (77), 247, 269
Schoeneberg (82), 247, 250
Schonhage, 21 l
Schonhage and Strassen (7 I), 2 11
Schroter, 11 1, 133
Schur, 80
Schwab, 250
Selberg and Chowla (67). 296
Shanks, D. (82). 193, 194
Shanks, D. and Wrench (62), 341
Shanks, W ., 340
Shanks, W. (1853), 340
Author Index
Singmaster (85), 342
Slater (66), 178, 179
Stickel (85), 222, 224
Stieljes, 264
Stolarsky (75), 233
Stolarsky (go), 237
Stormer, 345
Strassen, 202, 21 1
Sweeney (631, 336
Tamura and Kanada (Pr), 2 11, 342
Tannery and Molk (1893), 138, 141
Todd (49), 345
Todd (75), 394
Todd (79), 7
Toom, 211
Trefethen (84), 321, 325
Tricomi (65), 262
van Ceulen, 342
van der Poorten, 379
Vieta, 338, 343
von Neumann. 341
Wagon (85), 342
Waldo, 342
Wallis, 338, 343

Walz (Pr), 330
Watson (29), 306
Watson (32). 140
Watson (33), 3, 59, 261
Watson (35), 286
Weber, 168, 295
Weber (08), 68, 123, 138, 140, 300
Weierstrass, 29, 30, 141
Weierstrass (67). 348
Whittaker and Watson (27), 10, 26, 27, 29, 55,
59, 72
Wills, 323
Wimp (84), 7, 242, 260, 269
Winograd (go), 206
Wrench (60). 342
Wright, 80
Zucker, 100, 141, 296, 297
Zucker (77), 140, 297
Zucker (79), 100, 154
Zucker (84), 75
Zucker (85), 386
Zucker and Robertson (76a), 293, 294
Zucker and Robertson (76b), 293
Subject Index
Afgebraic addition theorem, 29, 30, 32

Algebraic function, 273, 276
Algebraic functional relation, 273
Algebraic integrals, 5
Algebraic number, 351, 352
Algebraic series for 117~and 1/K, 181-190
Algebraic transformation, 273, 274, 278
Algebraic transformation group, 273
Alpha, singular value of second kind, 152
cubic, 160
monotonicity, 153
quadratic, 158
quartic, 161
quintic, 310
recursions, 156
septic, 3 11
theta function form, 154
Alternating series test, 291
Archimedean mean iterations, 246,249, 251,252
Archimedean means, 259
Archimedean product, 246, 249, 250,
253,262
Archimedes' method, 249, 338
Arclemniscate, 25
Arclemniscate sine, 259
Arithmetic-geometric mean (AGM), I
AGM relation, 1
complex starting values, 15
calculation of 7 ~ .341
computational performance, 330
continued fraction, 188
Gaussian, 181, 241, 257
Jacobi's identity, 35
Legendre form, 3, 45
matrix AGM, 223, 226
multidimensional AGM, 272
quartic AGM, 17, 51, 254
theta forms, 146
variant of, 254
Benson's formula, 301, 303

Bernoulli numbers, 383
Bessel function transform, 39
Best approximants:
existence, 321
uniform polynomial, 3 16
uniform rational, 3 17
Best approximation property, 373
Beta function, 24, 88
Bit complexity, 200
Borchardt's algorithm, 250, 256, 263
Brouncker's continued fraction, 339, 345
Calculating powers of x. 208
Carlson's integrals, 256
Carlson's log, 248, 257
Catalan's constant, 198, 371, 386
Cauchy's binomial theorem, 76
Class number, 141, 295
Clausen's hypergeometric product, 179, 188, 189,
190
Comparable means, 244, 246, 247
Complement, 3, 178
Complementary integral, 8
Complementary modulus, 8
Complete elliptic integrals, 7
complementary integrals, 8
cubic algorithm for K. f07
differential equations, 9
E in terms of theta functions, 42
of first kind (K), 7
higher order transformations, 102ff
homogeneous forms of K and E, 12
K in terms of:
gamma function, 297, 298
theta functions, 35
moments of K and E. 198
quadratic transformations, 12
quadratically convergent algorithms, 14
Subject lndex
Complete elliptic integrals (Continued)

of second kind ( E ) , 7
series expansions of, 8
Complexity:
of algebraic functions, 215
bit, 200
elementary functions, 226
elliptic integral calculations, 227
Jacobian elliptic functions, 227
Lambert series, 98
log, 219
lower bound for log and exp, 227
operational, 201
n , 219
of theta computation, 95
transcendental functions, 219, 226
Complexity of algebraic functions, 215
Compound of means, 243
Conjugate divisors, 297
Constructibility, 347, 349-35 1
Continued fraction, 372-377
Contra-harmonic mean, 255
Convergents, 372
Cubic invariants, 173
Cubic modular indentities, 142
Cubic recursions for Ramanujah's invariants, 145
Diagonal mapping, 230
Dilogarithm, 38 1
Dirichlet class number formulae, 294, 295
Dirichlet L functions, 289
Discrete Fourier transform, 204
Discriminants, 294
Disjoint discriminants; 293
Domination, 239
e:
irrationality of, 352
transcendence of, 353, 359
Elementary symmetric mean function, 270
Elementary symmetric polynomial, 356
Ellipse, arc length of, 8
Elliptic functions, 29, 30, 3 1
cn, 29
degenerate elliptic functions, 29
dn, 29
fundamental parallelogram, 29, 53
Jacobian, 29
lattice, 29
odd elliptic functions, 31
order, 30
period of, 29
sn, 29, 57
Elliptic integrals
(see Complete and Incomplete)
Equivalence of operations, 2 13
Equivalent iterations, 252
Equivalent mean, 23 1, 239
Estimates of e , 363
Eta-multiplier, 3 11, 3 14
of order p. 3 11
Euler-Mascheroni constant, 336
Euler numbers, 383
computation of, 336
Euler's addition theorem, 18
Euler's formula, 189
Euler's identity, 64
Euler's pentagonal number theorem, 64, 66, 80,
306
Euler's totient function, 301
Evaluation problem, 204
Exponential:
algorithms for, 227
approximations to, 317, 318, 320, 321, 322,
323, 327, 330, 33 1, 335
series for. 320
Exp(n), quadratic algorithm for, 50
Factorial calculation, 218
Fast base conversion, 2 18
Fast computation, 50
Fast Fourier transform (FFT), 204,206.21 1, 212
Fast matrix multiplication, 202
Fast multiplication, 209
Fast polynomial division, 207
Fast polynomial evaluation, 208
Fast polynomial multiplication, 206
Fibonacci numbers, 94, 151, 287, 375
Fibonacci sequences and series, 91, 97, 98, 100
Finite Fourier transform, 204
Fully monotone, 29 1
Fundamental limit theorem, 5
first proof, 5
second proof, 6
third proof, 12
fourth proof, 21
Fundamental regions, 113
Fundamental sets, 1 13
Fundamental unit, 294
Galois group:
for F,, 123
for W, and W,, 136
Gamma function, 24, 27, 28, 87, 332, 336
duplication formula, 88, 90
evaluation of K and E , 25, 27, 189, 191, 297,
298
Subject Index
Gauss's multiplication formula, 90
Gaussian binomial coefficients, 76
Gaussian mean iterations, 246, 249, 251, 252
Gaussian product, 246, 248, 253
Gaussian sums, 83, 86
Gelfond-Schneider theorem, 277, 278, 348
Generalized complete elliptic integrals:
of first kind, 178
of second kind, 178
transformations, 179, 180, 185
Generalized Legendre symbol, 293
Generalized singular value function, 185
Genera, 295
Geometric series, 91
Gini's means, 233, 264
Gregory's series, 339
Halley's method, 216
Height, 371
Heronian mean, 237, 269, 270
Hexagonal lattice, 292
Hexagonal sum, 292
Holder's means, 232, 235, 263, 264
Homogeneous mean, 23 1
Homogeneous multidimensional mean, 266
Homer's rule, 204
Huwitz zeta function, 303, 304
Hyberbolic arclemniscate, 259
Hyberbolic function identities, 74, 75, 100
Hypergeometric functions, 332-334
Hypergeometric functions and series:
Gaussian hypergeometric series, 8
generalized, 178
hypergeometric differential equation, I I
Identric mean, 234
Incomplete elliptic integrals, 10
Eulers addition theorem, 18
of first kind, 10, 58, 60
of second kind, 10
of third kind, 10
Inhomogeneous modular group, T, 1 13, 1 17
Interpolation problem, 204
Invariance principle, 245, 246, 260
Invariants, evaluating, 293
Irrationality measures, 362
for exp(l/v), 363, 364, 371, 377
for exp(n), 371
for log, 370, 379
for n , 367, 368, 371
for n Z , 367
for 5(3), 369, 370
Irrationality of e. 352
Irrationality of n , 352
Isotone mean, 23 1
Isotone multimimensional mean, 266
Jacobian elliptic function, 29
cn, 29
dn, 29
half angle formula for cn, 262
half angle formula for sn, 30
sn, 29, 57
Jacobi's differential equation, 20, 22
Jacobi's duplication formula for arcsl, 260
Jacobi's formula for r4(n), 81
Jacobi's identity, 35
Jacobi's imaginary quadratic transformation, 73,
180
Jacobi's triple-product identity, 62, 65. 66, 72,
301, 306
corollaries, 64, 65
finite form, 77
first proof, 62
Khinchin's theorem, 362, 371
Klein's absolute invariant, 115, 179
Kronecker symbol, 293
Kummer's indentity, 179
Lagrange interpolation formula, 321
Lambert series, 91, 99, 281, 286, 288, 300
Landen transform, 17, 57, 59
in terms of elliptic functions, 58
in terms of theta functions, 57
Laplace transform, 39
Lattice sums, chemical, 288, 290
Legendre's relation, 24, 26, 27, 49, 152, 153
generalized, 178
theta function proof, 43
Lehmer's means, 232, 235, 263
Lemniscate sine, 5, 28
Length, 358
Lengths and measures, 358
Lindemann's theorem, 348, 357
Liouville numbers, 35 1 , 352
Liouville's function, 100
Liouville:~principle, 54
Liouville's summation principle, 36
Logarithm:
approximations to, 325, 327, 328, 330
bit complexity, 219, 222
calculation, 220, 221, 222, 223
complex, 222
operational complexity, 219, 222
series for, 320
theta function algorithms for, 224
Subject Index
Logarithmic mean, 234, 248, 261
Lucas numbers, 95, 97, 100, 287
Machin's formula, 339, 340, 341, 344
Madelung's constant, 288, 292, 301
Matrix logarithm, 224, 225, 226
Mean, 230, 231
discontinuous, 230, 238
equivalent mean, 23 1
homogeneous, 23 1, 266
isotone, 23 1, 266
strict, 230, 266
symmetric, 231, 266
trace of, 23 1
Mean iterations, 243
algebraic, 273-280
Archimedean, 246
arithmetic-harmonic, 4
Carison's, 257
Gaussian, 246
harmonic-geometric, 4
Lehmer's 265
multidimensional, 267
pathology, 253
rates of convergence, 251, 267, 272
Schlomilch's, 268
Tricomi's, 265
Means:
classes:
Gini, 233, 264
Heronian, 237, 269, 270
Holder, 232, 235, 263, 264
identric, 234
Lehmer, 232, 235, 263
logarithmic, 234, 248, 261
Neo-Pythagorean, 255
series expansions of, 263-266
Stolarsky, 233, 236, 264
Measure, of polynomial, 359
Meinardus conjecture, 32 1
Mellin transform, 87, 90, 289, 290, 301
Modular equations, 103, 3 15
cubic, 104, 107, 109, 110
degree of, 125
of degree 15, 3 I4
of degree 23, 133
elliptic, 112
endecadic, 106
for j, 123
for K , , and K,,4, 315
for A, of order p, 121, 140
octicity of u-v form, 126, 134
quadratic, 109
quintic, 105, 107, 109, 135, 136, 297, 313
septic, 106, 112, 313

solvable, 310, 311, 313
u-v form, 126, 127-132, 134
Modular functions:
I?-modular functions, 114, 118
A-modular function, 114, 118, 121, 133, 134
A-modular group, 113, 117
Modular transformations, 103
Modulus, 8
Multidimensional Archimedean iteration, 267,
269
Multidimensional Gaussian iteration, 267, 268,
270
Multidimensional invariance principle, 269
Multidimensional mean, 266, 269
Multiplier, 103, 105, 106
Multiplier of order p, 136
cubic, 138, 144, 149
of degree 13, 138
of degree 17, 138
quintic, 138, 309
septic, 138, 311
Neo-Pythagorean means, 255
Newton's identities, 356
Newton's method, 212, 214,215, 216, 217, 218,
223
N-monotone, 29 1
Nome, 41
Non-prime invariants, 295
Nth convergent, 372
Operational complexity, 201
Order of convergence, 2
Pad6 approxiwant, 319, 320, 323, 325,327, 331,
359
Partially comparable iterates, 247
Partition congruences, 84, 308
Partition function, 67, 307
Partitions of natural number, 67
Pell's equation, 376
Pendulum, period of, 8
Pentagonal number, 66, 67
Perimeter of ellipse, 8, 168, 195-197
Periodic continued fraction, 375
Pi:
AGM identities, 48, 52, 169, 197
algorithms, 46, 48, 170, 171, 175; 222, 310,
315, 335
approximations to, 168, 191, 192, 195, 197
complexity of calculating, 219
computation of, 337-342
cubic iteration, 171, 174
Subject Index
general iteration, 169
irrationality, 348, 352
irrationality measure, 367
legislation on, 342
mnemonic for, 342
normality of, 342
quadratic algorithm for e x p ( ~ )50
,
quadratic iteration, 170, 174
quartic iteration, 170, 174
quintic iteration, 175, 3 13
septic iteration, 171, 174, 175
series for I/T, 181-190
transcendence of, 347, 348, 352, 354
Picard's theorem, 1 19
Piecewise monotone, 234
Pochhammer symbol, 178
Poisson summation formula, 36, 89
Prime number theorem, 378
Primitive binary form, 141
Properly equivalent forms, 141
Q-binomial coefficients, 76
Q-binomial theorem. 308
Quadratically attractive transformation, 277
Quadratic computation, 50
Quadratic modular equation, 102
Quadratic reciprocity, 86
Quintic multipliers, 309
Quintuple-product identity, 143, 146, 306
Ramanujan's continued fraction identity, 8 1
Ramanuian's invariants. 70. 179
Ramanujan's modular identity:
of order 3, 287
of order 5, 310, 312
of order 7, 312
sum, 308
Ramanujan's ,P,
Ramanujan's multiplier of second kind, 157, 159
Rational mean iteration, 278
Recursion, 201
Recursion formula for r2(n), r3(n), 15 1, 306, 307
Reduced complexity methods, 327
acceleration based on:
binary splitting, 329
on FFT, 328
on functional equations, 327
Reversion of power series, 208
Riemann hypothesis, 90
Rising factorial, 178
Rogers-Ramanujan identities, 65, 78, 80
Bressoud's "easy proof," 78
Roth's theorem, 363
Schlafli's equation, 147
413
Schonhage-Strassen multiplication, 21 1
Schroter's formula, I I I
Schwarz derivative, 20, 118
Septic multipliers, 309
Simple continued fractions, 372
Singular invariants, 141
Singular moduli, 139
Singular value function, 152
of first kind, 152
generalized, I85
of second kind, 152ff
Singular values, 26.49, 139, 140, 141, 172, 173,
296, 297
Solvability of quintic in modular terms, 135
Solvable modular equations, 310, 31 1, 313
Solvable numbers:
of type one, 293
of type two. 293
Square-free invariants, 295
Stirling numbers, 305
Stirling's formula, 90
Stolarsky's means, 233, 236, 264
identric, 234, 271
logarithmic, 234, 271
multidimensional, 271
Strict mean, 230
Strict multidimensional mean, 266
Strong equivalence, 241, 243
of iterations, 252
Sums of squares:
of two squares, 82, 285, 290
of three squares, 66, 151, 286
of four squares, 81
of others, 71, 287, 292, 293
Supremum norm, 316
Symmetric mean, 231
Symmetric multidimensional mean, 266
Symmetric polynomial, 356
Tan, computation of, 227
Theta functions, 10, 33, 91, 93
basic identities, 64,67, 68,70,71, 73,74, 1 I I
finite transformation, 86, 87
general theta functions, 52
one-dimensional heat equation, 56
theta transformation formulae, 38, 44, 54, 87
Theta series, 91
Trace of mean, 231
Transcendence of e, 348, 353, 359
Transcendence of T, 347, 348, 352, 354
Transcendental functions, 274
Transcendental number, 35 1, 352
Transformation of order p. 119
Transformations of complete elliptic integrals:
414
Subject index
Transformations of complete elliptic integrals

(Continued)
algebraic transformation for K, 21
of generalized integrals, 179, 180, 185
higher order, 102f
quadratic transformations for E and K, 12, 13
quadratic transformations for K. 36
Triangular numbers, 286, 288
Trilogarithm, 38 1
Triple-product identity, 62, 65, 66, 72, 306
Tschimhaus transformation, 136
Turing machines, 20 1
Ultimately monotone, 235, 241
ViBta's formula, 338, 343

Wallis' formula, 338, 343
Weierstrass function, 27, 30, 141
5(2), irrationality of, 366
5(3), irrationality of, 369
5(3), series for, 190, 379
Zeta function, 87, 88
functional equation for, 89
relation to prime distribution, 90

Borwein - Pi and The AGM (Wiley, 1987)

Uploaded by

Copyright:

Available Formats

Borwein - Pi and The AGM (Wiley, 1987)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Borwein - Pi and The AGM (Wiley, 1987)

Uploaded by

Copyright:

Available Formats

What is the book about?

What is the book about?

What mathematical topics does it cover?

What mathematical topics does it cover?

Pi and the AGM

A Study in Analytic Number Theory

Copyright 01987 by John Wiley & Sons, Inc.

(Canadian Mathematical Society series of monographs

To our mathematician father, David Borwein

Five thousand digits of pi.

From the eleventh iteration of

When I was a student, abelian functions were, as an effect of the Jacobian

tions helped produce a better book. The cheerful technical assistance of P.

Halifax, Nova Scotia

brief discussion of a number of topics including elliptic integrals and

Jonathan M. Borwein was born in St. Andrews Scotland in 1951. In 1971 he

Complete Elliptic Integrals and the Arithmetic-Geometric

The Arithmetic-Geometric Mean Iteration 1

Theta Functions and the Arithmetic-Geometric Mean

A Theta Series Solution to the AGM 33

Jacobi's Triple-Product Identity 62

Some Number Theoretic Applications 81

Higher Order Transformations

4.1 A First Approach to Higher Order Transformations 102

Chapter 5 Modular Equations and Algebraic Approximations to

Singular Values of the Second Kind 152

The Complexity of Algebraic Functions

Complexity Concerns 200

Chapter 7 Algorithms for the Elementary Functions

General Means and Iterations

Abstract Means 230

~arlson'sIntegrals and More Examples 256

Some Additional Applications

Sums of Two Squares 281

Other Approaches to the Elementary Functions

10.1 Classical Approximations 316

11.1 On the History of the Calculation of

Complete Elliptic Integrals and

THE ARITHMETIC-GEOMETRIC MEAN ITERATION

One of the jewels of classical analysis is the arithmetic-geometric mean

It is customary and useful to introduce an auxiliary variable,

If we assume that 0 < b, s a,, then from the arithmetic-geometric mean

The Arithmetic-Geometric Mean Iteration

Whence we observe that a, and b, converge to a common limit determined

AM(a, b) = M(Aa, Ab)

1.1 The Arithmetic-Geometric Mean iteration

we have that k, 4 1 quadratically. In fact, the function g defined by

is the unique solution of (1.1.10) analytic in a neighbourhood of 1 that

Comments and Exercises

The Arithmetic-Geometric Mean Iteration

1. Deduce that, for the AGM,

Hence the AGM is well defined for negative n. Show that

Observe that this formula avoids the subtractive cancellation problems

Consider the harmonic-geometric mean iteration

3. Consider the arithmetic-harmonic mean iteration

Show, for a,,

the above iteration converges quadrati-

1.2 Gauss's Derivation of the Fundamental Limit Formula

Show that the AGM is a well-defined quadratically convergent iteration