..

U N LEA SHE D

3 14= "3.1415":
(=-* 3 14, 3 14
++ I ++ 3 14159)
for (- 31415 = _,
1; 31415++) write (
3,14), 3141592654[

"0123456789",
"314"
puts*
3141592654=0
;_314=-*"3.141592";}

3,14 1592653589793 2 3846264

~(i5J'$:!.~.l"1

Springer

Arndt Haenel, 1t - Unleashed

Springer
Berlin
Heidelberg
New York
Barcelona
Hong Kong
London
Milan
Paris
Singapore
Tokyo

Jorg Arndt Christoph Haenel

Pi - Unleashed
Translated from the German
by Catriona and David Lischka

With CD-ROM

Springer

Jorg Arndt
Hiihlweg 37
95448 Bayreuth, Germany
Christoph Haenel
Waldfriedenweg 24
82223 Eichenau, Germany

Title of the original German edition:

Pi. Algorithmen, Computer, Arithmetik
Springer-Verlag Berlin Heidelberg 1998,2000
Library of Congress Cataloging-in-Publication Data applied for
Die Deutsche Bibliothek - CIP-Einheitsaufnahme
Arndt. Jorg: Pi-unleashed: with CD-ROM / Jorg Arndt; Christoph Haenel.
Transl. from the German by Catriona and David Lischka. - Berlin; Heidelberg;
New York; Barcelona; Hong Kong; London; Milan; Paris; Singapore; Tokyo:
Springer, 200 I
Dt. Ausg. u.d.T.: [Pi]
ISBN 3-540-66572-2

ISBN 3-540-66572-2 Springer-Verlag Berlin Heidelberg New York

This work consists of a printed book and a CD-ROM packaged with the book, and is
subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data
banks. Duplication of this publication or parts thereof is permitted only under the
provisions of the German copyright law of September 9, 1965, in its current version,
and permission for use must always be obtained from Springer-Verlag. Violations are
liable for prosecution under the German Copyright Law.
Springer-Verlag or the authors make no warranty for representation, either express or
implied with respect to this CD-ROM or book, including their quality, merchantibility,
or fitness for a particular purpose. In no event will Springer or the authors be liable for
direct, indirect, special, incidental, or consequential damages arising out of the use or
inability to use the CD-ROM or book, even if Springer-Verlag or the authors have been
advised of the possibility of such damages.
Springer-Verlag Berlin Heidelberg New York
a member of BertelsmannSpringer Science+Business Media GmbH
Springer-Verlag Berlin Heidelberg 2001
Printed in Germany
The use of general descriptive names, trademarks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the
relevant protective laws and regulations and therefore free for general use.
Cover design: Klinkel + Lopka Werbeagentur, Heidelberg
Typesetting: Camera-ready by the authors
Printed on acid-free paper- SPIN: 10731352
33/3142 GF 543210

Nil desparare
(Do not despair)

Motto from Gauss's Mathematical Diary (1796-1814),

in which the author demonstrates a connection between
the computation of 7r and the arithmetic-geometric mean.

Foreword to the second edition

After the first edition of our book on the fascinating number 7r was
so well received, we are pleased to present the second edition, which
has been completely revised. All the various subject areas have been
expanded, improved and updated. We are grateful to those of our
readers whose suggestions we have adopted.
The update includes a report on the latest world record by the
Japanese Professor Kanada, who has now calculated 7r to 206.1 billion
decimal places, and also the record achieved by Colin Percival, who
in 1999 while still a teenager calculated the 40 trillionth binary digit
of 7r with the assistance of a team of Internet collaborators. In addition there is also a new section on FFT multiplication (multiplication
using fast Fourier transforms), in which this technique that today is
indispensable for modern 7r calculations, is explained in detail with
appropriate examples of code.
The historical section has been completely reworked and considerably enlarged.
We have also expanded the (platform-independent) CD-ROM to
include, in particular, 400 million decimal places of 7r and several freeware high precision libraries. Last but not least, the URLs for many 7r
sites on the Internet have been brought up-to-date.
We would particularly like to thank Professor F .L. Bauer of the
Technical University of Munich for his invaluable ideas and suggestions
for improvements on every topic in the book. We feel honoured to have
enjoyed Professor Bauer's encouragement and advice.
We look forward eagerly to the reactions and verdict of our readers.
We now have a special e-mail address to which you can send your
comments. Please write to pibook@jjj . de.
Have fun!
Jorg Arndt and Christoph Haenel

Foreword to the first edition

The number 7r, about which most people know only that it is expressed
as the ratio of the circumference of a circle to its diameter and begins
with 3.14 ... , was calculated in July 1997 to 51.5 billion decimal places.
In this book you will discover why anyone should be interested in
achieving world records with such bizarre calculations, which are long
enough to fill up thousands of phone directories, and you will also find
many other interesting things about this number. The methods which
are used for these high-precision calculations reflect developments in
mathematics whose significance goes far beyond simply "working out
7r" These new and effective algorithms surprisingly have the pleasant
characteristic that they are easy to understand.
In this book we describe in detail the quest for world records in 7r
and on the enclosed CD-ROM we provide programs which will enable
you yourselves to calculate several million digits of 7r on your computer.
The CD-RO M also contains source code versions of all the programs
and examples, together with more detailed text on topics which had
to be left out of the book due to space constraints. To help you get
started with your own research on the Internet, we have also provided a
collection of URLs. For the statisticians amongst you, we have included
several million digits of 7r on the CD-ROM.
Finally we narrate the main personalities and milestones in the
history of 7r, amongst which you will find not only facts of scientific
interest but also some examples of both the curious and the bizarre.
The following persons provided extremely useful comments and suggestions as we were writing the book: Dieter Beule, Lothar Kruger,
Heinz Pohlmann, Georg Sessler, Mikko Tommila, Ute Zwerschke plus
a particularly co-operative reader who wishes to remain anonymous.
We would like to express our heartfelt thanks to all of them.
We wish you a pleasant journey into the fascinating world of the
number 7r .
August 2000
Jorg Arndt and Christoph Haenel

Contents

1.

The State of Pi Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.

How Random is 7r? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.1 Probabilities......................................
2.2 Is 7r normal? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
2.3 So is 7r not normal? ...............................
2.4 The 163 phenomenon. . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
2.5 Other statistical results. . . . . . . . . . . . . . . . . . . . . . . . . . ..
2.6 The Intuitionists and 7r . . .
2.7 Representation of continued fractions. . . . . . . . . . . . . . ..

21
21
21
24
25
28
30
32

3.

Shortcuts to 7r . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1 Obscurer approaches to 7r . . .
3.2 Small is beautiful .................................
3.3 Squeezing 7r through a sieve. . . . . . . . . . . . . . . . . . . . . . ..
3.4 7r and chance (Monte Carlo methods) . . . . . . . . . . . . . . ..
3.5 Memorabilia......................................
3.6 Bit for bit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
3.7 Refinements......................................
3.8 The 7r room in Paris. . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..

35
35
37
38
39
44
47
49
50

4.

Approximations for 7r and Continued Fractions .....

4.1 Rational approximations. . . . . . . . . . . . . . . . . . . . . . . . . ..
4.2 Other approximations .............................
4.3 Youthful approximations . . . . . . . . . . . . . . . . . . . . . . . . . ..
4.4 On continued fractions. . . . . . . . . . . . . . . . . . . . . . . . . . . ..

51
51
55
63
64

5.

Arcus Tangens . .. ... .... .. ..... .. .... .. ... ... ... ....
5.1 John Machin's arctan formula ......................
5.2 Other arctan formulae. . . . . . . . . . . . . . . . . . . . . . . . . . . ..

69
69
72

Contents

6.

Spigot Algorithms ..................................

6.1 The spigot algorithm in detail ......................
6.2 Sequence of operations. . . . . . . . . . . . . . . . . . . . . . . . . . . ..
6.3 A faster variant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
6.4 Spigot algorithm for e .............................

77
78
80
82
84

7.

Gauss and 7r . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
7.1 The 7f AGM formula. . . .. .. ... .. .. .... .. . .. ... . .. ..
7.2 The Gauss AGM algorithm. . . . . . . . . . . . . . . . . . . . . . . ..
7.3 SchOnhage variant. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
7.4 History of a formula. . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..

87
87
90
92
94

8.

Ramanujan and 7r . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
8.1 Ramanujan's series ................................
8.2 Ramanujan's unusual biography. . . . . . . . . . . . . . . . . . . ..
8.3 Impulses .........................................

103
103
105
110

9.

The Borweins and

113

7r

10. The BBP Algorithm ........ . . . . . . . . . . . . . . . . . . . . . . .. 117

10.1 Binary modulo exponentiation ...................... 120

10.2 A C program on the BBP series. . . . . . . . . . . . . . . . . . . .. 123
10.3 Refinements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 126
11. Arithmetic .......................................... 131

11.1 Multiplication ....................................

11.2 Karatsuba multiplication ...........................
11.3 FFT multiplication ................................
11.4 Division ..........................................
11.5 Square root .......................................
11.6 nth root .........................................
11. 7 Series calculation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..

131
132
135
145
146
149
150

12. Miscellaneous . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
12.1 A 7f quiz .........................................
12.2 Let numbers speak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
12.3 A proof that 7f = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
12.4 The big change ...................................
12.5 Almost but not quite . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
12.6 Why always more'? ................................

153
153
154
155
155
156
158

Contents

XI

12.7 7r and hyperspheres . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..

12.8 Viete x Wallis = Osler ............................
12.9 Squaring the circle with holes .......................
12.10An (in)finite funnel ...............................

158
160
162
164

13. The History of 7r . . . . . . . . . . . . . . . . . . . . ..

13.1 Antiquity ........................................
13.2 Polygons. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
13.3 Infinite expressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
13.4 High-performance algorithms .......................
13.5 The hunt for single 7r digits .........................
Table: History of 7r in the pre-computer era . . . . . . . . . . . . . ..
Table: History of 7r in the computer era . . . . . . . . . . . . . . . . ..
Table: History of digit extraction records . . . . . . . . . . . . . . . ..

165
167
170
185
198
203
205
206
207

14. Historical Notes .....................................

14.1 The earliest squaring the circle in history,? . . . . . . . . . . ..
14.2 A 7r law ..........................................
14.3 The Bieberbach story ..............................

209
209
211
213

15. The Future: 7r Calculations on the Internet ......... 215

15.1 The binsplit algorithm ............................. 215
15.2 The 7r project on the Internet. . . . . . . . . . . . . . . . . . . . . .. 219

16.

7r

Formula Collection ................................ 223

17. Tables ..............................................

17.1 Selected constants to 100 places (base 10) ............
17.2 Digits 0 to 2,500 of 7r (base 10) .....................
17.3 Digits 2,501 to 5,000 of 7r (base 10) ..................
17.4 Digits 0 to 2,500 of 7r (base 16) .....................
17.5 Digits 2,501 to 5,000 of 7r (base 16) ..................
17.6 Continued fraction elements 0 to 1,000 of 7r
17.7 Continued fraction elements 1,001 to 2,000 of 7r

239
239
240
241
242
243
244
245

A. Documentation for the hfloat Library ................

A.1 What hfloat is (good for) ...........................
A.2 Compiling the library. . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
A.3 Functions of the hfloat library. . . . . . . . . . . . . . . . . . . . . ..
A.4 Using hfioats in your own code ......................

247
247
248
248
250

XII

Contents

A.5 Computations with extreme precision. . . . . . . . . . . . . . ..

A.6 Precision and radix. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
A.7 Compiling & running the 7r-example-code ............
A.S Structure of hfloat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..
A.9 Organisation of the files. . . . . . . . . . . . . . . . . . . . . . . . . . ..
A.I0 Distribution policy & no warranty ..................

250
251
253
253
254
255

Bibliography . ........................................... 257

Index ................................................... 265

1. The State of Pi Art

In October 1999 Professor Yasumasa Kanada of the University of

Tokyo announced on the Internet 1 that he had set a new world record
for the calculation of 1f; namely, he had calculated it to 206,158,430,000
decimal places. He had performed two independent calculations using
two different algorithms, each of which had produced identical results
except for the last 45 digits. The programs used to perform the calculations were written by Kanada and his assistant, Daisuke Takahasi.
They had taken 83.5 hours of computer time to execute.
With this latest breakthrough, humankind now possesses 206.1 billion digits of 1f beginning with 3.1415. That is indeed a lot of numbers. Just to have them race past on the screen would take several
weeks. If one were to print out 100 million digits from this 1f (on
around 10,000 sheets of paper), there would still be 206 billion digits
left. They would occupy 30,000 books, each 1000 pages long, although
equally they would also fit on just a few hard disks each weighing less
than a pound.
Yasumasa Kanada's statement came only five months after the
same man announced he had set a new world record of 68.7 billion
decimal places. That announcement in turn came only 18 months after reaching 51.5 billion digits. Since 1981, when the record stood at
2 million digits, it has been beaten 26 times, almost doubling every
year.
For some years Kanada has been the only 1f mathematician intent
on achieving new world records. Up until 1986, when a comparatively
modest 30 million digits was still a world record, a number of competitors were still in the arena. Ten years later this had been whittled down
to two, Kanada in Japan and the Chudnovsky brothers in the USA,
who took it in turns to outdo each other until in 1996 the 8 billionth

ftp://pi.super-computing.org/README.our-1atest~ecord

1. The State of Pi Art

decimal place was reached. Since 1997 and from 51 billion digits, only
Kanada and his team are still in the running.
The computers used to attain these world records have generally been supercomputers which were specially developed for numbercrunching and cost millions of dollars. For example, Kanada achieved
his latest record on a Hitachi SR8000 with 128 processors. Nevertheless, if one has the skill it is possible to make smaller computers perform 7r calculations which are still competitive. In the history of the
world records there have been several examples of this. Thus, for example, a few years ago the above-mentioned Chudnovsky brothers calculated their 8 billion decimal digits of 7r in New York on a computer
which they had built themselves from parts purchased in department
stores [92].
One might expect Kanada to record his 7r calculations in compressed form 2 on around 150 CDs and to offer them for sale to interested parties. However, the professor seriously views his 7r as "noncommercial digits" and will only release them after checking carefully
that they are to be used for scientific purposes. He will only allow the
first few billion digits to be downloaded over the Internet from his
server in Tokyo. This itself is not a trivial task since, with a nprmal
telephone line, downloading the data would take approximately one
day per billion digits (with the associated expense). If the explosion
in the capability of computers continues, it will soon be quicker and
simpler to calculate one's own 7r to a suitable length than to download
it from elsewhere. When one is dealing with a few hundred million
digits this is in fact already the case today. And even a sophisticated
7r program will always be shorter than the number of digits to which
it can calculate.
What is the point of having such a long representation of 7r? One
obvious task is to search through the digits looking for rare or unusual
patterns and possibly for lengthy sequences of the same digits. For the
first few thousand digits, which are shown on page 240ff. of this book,
it is probably still feasible to do this without a computer. Above that
limit, however, one will probably need a search program, for example
our piseek program, which is included on the enclosed CD-ROM along
with the first 400 million digits of 7r and should provide a few hours'
pleasure.
2

Because the decimal digits of 7r are (probably) normally distributed, at best N

digits would compress down to 1/(8IogIQ2)N, around D.4N bytes.

1. The State of Pi Art

As one examines 7r, one is struck by a striking feature which occurs

as early as at position3 768, where 9 occurs six times in succession. This
peculiarity, the probability of which is only 0.08%, even has a name. It
is known as "Feynman point", after the American Nobel prize winner
Richard Feynman (1918-1988). He once claimed that ifhe had to recite
the digits of 7r, he would name them accurately up to this point and
only then say and so on. His words have an elegant double meaning,
since after these six 9s it is of course not at all a case of "and so on" ,
but something quite different follows, in this case an 8. The next block
of six identical numbers occurs only much later, but still 'too early',
at position 193,034. Remarkably, it consists of a second series of 9s.
On the other hand, the first zero in 7r does not occur until position 32, i.e. much later than one would expect. These properties of the
number 7r have facilitated the task of "7r poets" seeking to compose
aids to remembering 7r in which every word has the same number of
letters as the related 7r digit. A zero would terminate such a sequence
by definition. One of these mnemonic verses, which produces a lot of
coffee and 11 pieces of 7r, runs as follows:
May I have a large container of coffee? Cream and sugar?
31415

926

535

Where does the sequence 0123456789 first appear in 7r? It was only
in 1997, at the time of the last but one world record, that this obvious question was finally answered. In fact that sequence does not occur until position 17,387,594,880, i.e. long past the 6.4 billionth digit,
which was previously the furthest that was known. But, if this special
sequence does not occur until that late, another special sequence, consisting of 10 digits all different from each other, occurs relatively early,
commencing in position 60. Even in the 17th century, mathematicians
had calculated 7r this far, as the history of 7r world records shows on
page 205.
Number fanatics might perhaps search in 7r for "self-referential positions", in which the numeric sequence is the same as the position
number. The first such position occurs right at the beginning, namely
in position 1. Other examples are to be found at positions 16,470
and 44,899, but after that there are no further instances until posi3

It is customary to count digit positions from 0, beginning with the 3 in front of

the decimal point. In this way 3 1 4 1 5 correspond to positions 0 to 4.

1. The State of Pi Art

tion 79,873,884. Whether there are any self-referential decimal places

after that we do not know.
It may be more communicative to search through 1[" for particular
telephone numbers or other numbers such as one's own birth date, as
people have always been intrigued by the question, VVhere do I occur
in 1["? (For example, the author's birth dates are at positions 5,407,560
and 14,666,671). If I liked, I could print the position on my business
cards, communicating vital information about what kind of person I
am. On the Internet there is a game like this, although currently it is
only based on 50 million digits 4 .
Somewhere in 1[", everything that is finite is surely to be found,
including, when coded in numeric form, every text in the world, the
shortest, the longest, the cleverest. Naturally that includes this text,
and also of course the Bible, in every language, and every piece of
music. (Moreover, 1[" also contains 1[" itself, albeit only once, whereas
neither 1[" + 1 nor the root of 2 will be found within 1[".)
Human DNA contains about 3.6 billion digits of information, so my
neighbour's DNA could already be contained in the known digits of 1[".
Perhaps, but highly improbable. For example, the number which
Y. Kanada has calculated does not even contain all the 20-digit numbers that are possible, which is 1020 , i.e. far more than the approximately 1011 digits which occur in Kanada's number. Still less does it
contain all the 21-digit numbers, let alone all the numbers of DNA or
Bible length. This makes it unlikely that the Bible, with its estimated
107 digits, occurs in the first "few" digits of 1["; to find the Bible, we
7
would need to possess around 1010 digits, whereas in fact we have
1
little more than 1010
Even worse, it simply is not possible for theoretical reasons to calculate 1[" to so many digits, for the universe consists only of 10 79 elementary particles, and even if one were to turn the entire universe into
a supercomputer, it would only be able to store about 1076 decimal
places, which is far fewer than the 10107 Bible digits mentioned above.
However, it is conceivable that individual positions on this side of that
limit can be calculated.
Anyone who wants to make a name for himself can examine the major issue of whether 1[" is normal, or perhaps more accurately, whether
1[" is not normal. If 1[" were not normal, this would mean that particular
4

http://www.aros.net/-angio/pi_stuff/piquery

1. The State of Pi Art

numbers or blocks of numbers of the same length occur more often or

less often than others, so that, for example, in the 50th billion positions
significantly more sixes occur than in the 60th billion.
Anyone who succeeded in making such a pronouncement would certainly make the headlines. For up to now no one has found any sections
within 7r which are non-normal. Attempts to represent 7r in hexadecimals, under which 7r begins with 3.243F6 A8885, rather than in the customary decimal representation, have likewise been unsuccessful. Even
the so-called simple continued fraction of 7r, i.e. the representation of
7r using an infinite series of fractions in which all the numerators are
equal to 1,
1
7r=3+
1
7+
1
15 +
1
292+under which the fine structure of numbers becomes evident, has not
revealed any regularities, although the simple continued fractions of
some other transcendental numbers, such as the number e, indeed do
exhibit recognisable patterns.
Mathematicians conclude from these observations and from other
indications that 7r is a normally distributed number. But to date they
have not been able to prove this, so it is entirely possible that "somewhere" within 7r a big sensation is waiting to be uncovered. This question is considered further in the next chapter.
It is thus unclear whether 7r is normal. On the other hand, it is definitely known to be irrational, as was demonstrated for the first time in
1766 by the Alsatian mathematician, Johann Heinrich Lambert (17281777). An irrational number is one which cannot be represented as a
ratio of two integers. For example, the fraction f~~ = 3.141 592 ... is a
very good approximation as it specifies 7r to 6 correct decimal places;
nevertheless, it is not equal to 7r, and no other fraction having an integer numerator and denominator no matter how long can produce
the exact value of 7r. The approximation f~~ was discovered already
in the 5th century AD by Tsu Chhung-Chih in China, and for almost
800 years was the best 7r approximation available. This and other approximations are considered in Chapter 4.

l. The State of Pi Art

It is also known that 7r is transcendental. A famous, but extremely

complicated proof of this was provided by the Munich mathematician
Ferdinand Lindemann (1852-1939) in 1882, i.e. more than 100 years
after Lambert's conclusions as to the irrationality of 7r. Lindemann's
proof states that 7r cannot be the root of a polynomial with integral
coefficients, however many terms it may have. It is true, for example, that 97r4 - 2407r 2 + 1492 is approximately 0 (or, more precisely,
-0.02323 ... ), but it is impossible to find such an expression which
gives 0 exactly.
It is astonishing but nevertheless true that today, several hundred
years after the transcendence of 7r was proven, we do not know much
more about 7r.
Admittedly, the constants 7r 2 , e 7r and 7r + log 2 + V2log 3 have
now also been proven to be transcendental. But we still do not know
whether similar quantities such as e + 7r, e 7r, 7r / e, log 7r or 7r e are irrational or not. Many other questions regarding 7r remain unanswered.
Some questions, such as the issue of whether 7r is normal or not, are
definitely unresolved, because no one has an inkling as to what method
one might employ in order to resolve the matter one way or the other.
And when we consider the 206.1 billion digits that are known, all we
know about the decimal sequence of 7r is trivial information such as the
frequency and distribution of (short) blocks of numbers [11, p. 203].
The number 7r is thus one of the oldest subjects of research by
mankind and possibly the one topic within mathematics which has
been researched for the longest. Humans have concerned themselves
with 7r for several thousand years. For example, already in 2000 BC
the Babylonians and Egyptians had discovered approximations for 7r
which were less than 0.02 away from its true value. Over the next,
increasingly enlightened 4000 years, researchers tirelessly never gave
up trying to unravel the secrets of 7r. The fascinating history of this
research is covered in more detail in Chapter 13, page 165ff.
When one considers all the work that has been done on 7r, the meagre amount that is known about it is somewhat surprising. But then
when one considers that scarcely no other field is quite like number
theory in throwing up very simple questions that are so difficult to
answer, perhaps it is not so surprising after all. One example here is
the so-called Last Theorem of Pierre de Fermat (1601-1665) of 1637,
which states that there are no integer solutions for the equation

1. The State of Pi Art

if n is greater than 2. This old and easy-to-state problem remained

unresolved for centuries although countless numbers of (more or less)
educated people endeavoured to resolve it; it was only recently (October 1994) and with a complicated 100-page proof that Fermat's Last
Theorem was finally proven by Andrew Wiles. On the other hand,
problems of number theory which are also easy to formulate always
turn out to defeat all but the very best minds, for example the famous question of whether the number of prime pairs, i.e. sequences of
two prime numbers separated by 2 such as the sequence 3 and 5 or
10,007 and 10,009, is infinite. Simple questions - yet no one has solved
them yet.

With his proof of the transcendence of 7r, Lindemann also settled

another problem which has preoccupied the brains of the best mathematicians and philosophers since the ancient Greeks and which was
already so well-known in 414 Be that Aristophanes made use of it in
his comedy "The Birds" .
This is the problem of the squaring the circle, i.e. the challenge of
finding a square which has exactly the same area as a given circle using
a geometrical construction whose execution requires nothing but an
unmarked straight edge and compass and must be possible in a finite
time. With r as the radius of the circle, one needs to construct side x
of a square whose area x 2 is equal to the area of the circle r 2 7r. The
sides of the square must therefore equal x = rft.
But only certain line lengths can, under the specified conditions, be
determined through geometric constructions, namely ones which are
integers and are produced through normal, rational operations such as
addition or division and/or through the formation of square roots [32,
p. 347]. iv'I7, for example, can be constructed in this way, whereas
3~ cannot. The expression rft does contain a multiplication and a
square root, and hence meets the above requirement for a solution to
the problem. But to allow the circle to be squared, it would have to
be possible to obtain the radicand 7r through the specified operations,
and it is precisely this that does not occur according to Lindemann's
proof. Hence, it is not possible to square the circle with a geometric construction using only compass and ruler. (On the other hand it
is possible if the constraining conditions are varied. For example, in
hyperbolic geometry there may well be a circle, i.e. one whose area

1. The State of Pi Art

is equal to 7r, which can be transformed into a square with the same
area with the aid of (unmarked) compass and straight edge; but in
this - non-Euclidean - geometry a straight line cannot be drawn with
a ruler.)
After Lindemann had resolved the problem of the Squaring the
circle once and for all in 1882 (even if his achievement was only to
demonstrate that the problem cannot be resolved), one would have
expected all further attempts at a solution to cease immediately. It
is all the more amusing to find that despite Lindemann's proof a lot
of people continued to try to solve the problem, and they are still at
it even today. Dudley writes about such "mathematical cranks" in a
monograph [50], in which he mentions amongst others the ship builder
O.Z. (Otto Zimmermann) from Hamburg who published a book entitled "7r is rational" in 1983. In his book Dudley also classifies those who
have attempted to square the circle. According to this, O.Z. is a member of that relatively rare species who keep revising their opinion over
time: before 1975 his construction produced 7r = 3.1415926535576,
then up to January 1976 it was 3.141592653598, while from then on
7r amounted to exactly 3.1428.
is the ratio of the circumference c of a circle to its diameter d;
7r = C : d. This is the classic geometrical definition of the number. The
popular rule that 7r is the circumference of a circle with diameter 1 is
based on this.
A second geometrical definition states that 7r is the ratio of the
area of a circle A to the square of the radius r of the circle i.e. it is
determined by the relationship 7r = A : r2. This results in the obvious
statement that the ratio of the area of the circle to the area of the
surrounding square is given by 7r : 4.
The oldest Indo-Germanic cultures were already aware, thousands
of years Be, that a fixed ratio exists between the circumference and
the diameter of all circles, large or small. The second fixed ratio in
the circle, namely the relationship between the area of the circle and
the square of the radius, was also known unimaginably long ago, even
though this is less obvious than the first. On the other hand, it was
considerably later, perhaps not until ancient Greece, that it was appreciated that both these relationships involve the same ratio, 7r. This fact
is not readily understood. The picture below is perhaps the simplest
way of examining the concept:
7r

1. The State of Pi Art

radius

circumference

In the picture on the left, a circle has been divided into a number of
equal-sized "slices of cake". If one imagines that the slices were taken
out and arranged side by side in a rectangle, they would appear as
shown in the picture on the right. The greater the number of slices,
the more the outline of the rectangle approximates the length of the
circumference of the circle, while the height of the rectangle approaches
the length of the radius of the circle. The rectangle in the picture is
only half occupied by the cake slices, and is therefore twice as big as
the sum of their areas, which is equal to the area of the circle. Thus, if
7r appears in the formula for the area of the circle, then it also figures
in the calculation of the circumference.
In addition to the geometric definition of 7r, there are many other
definitions, including ones drawn from quite different contexts. For
example, 7r appears in interesting probability problems.
What is the probability that a coin tossed an even number of times
will land an equal number of times on its head and on its tail? From
one's schooldays one may recognise the formal treatment of such exercises: if the coin is tossed 2n times and there are two possible outcomes
per toss, then altogether there are 22n possible outcomes, of which
will be the desired outcome. The probability of this is therefore
2n
p = e:)/2 , or

e:)

1 x 3 x 5 x ... x (2n - 1)
(1.1 )
2 x 4 x 6 x ... x
2n
This fraction is difficult to calculate for large values of n. We have
John Wallis (1616-1703) to thank for calculating it before infinitesimal
analysis was invented. Following an interesting train of thought, he
discovered that the infinite product comes to
3x3x5x5x7x7x ...
4
(1.2)
2x4x4x6x6x8x ...
7r
It is not difficult to see [123] that this Wallis product occurs in the
formula for the calculation of p shown above, and in such a manner
that
p=---------~

10

1. The State of Pi Art

p~l/~

(1.3)

which, for example produces a 10% probability with 2n = 62 tosses of

the coin.
We thus have here a definition for 7f which has nothing to do with
circles; although only integers are used, the transcendental number
still appears in the formula.
7f figures in many other probability statements as well. In particular
it occurs in the well-known Gauss error distribution curve, which, for
instance, is used in graduated life tables for life assurance. Some bright
spark [92] remarked it propos of this that 7f even occurs in death.
There are definitions of 7f which are even more abstract, for example, analytical definitions using integrals. Two particularly visually
attractive examples were produced by the Swiss mathematician Leonhard Euler (1707-1783):
7f-4

or

1 vI 0

4
7f

11 VI 2

dx

x dx

x4

= lim - 2 '"'
n-too n ~
k=O

Jn2 -

(1.4)

x4

k2

You can admire a whole range of other

formulae on page 223.

(1.5)
7f

formulae in our collection of

Mathematics professors sometimes use yet another definition in

their lectures.
Under this definition, ~ is declared simply as
the smallest possible zero of the cosine function. For it is the case that the function cos x
is positive at x = 0, but later on, for example when x = 2, it is negative. Since in be- 0
2
x
tween it runs continuously without any sudden changes, it must cross the x-axis at some
point, which is defined as l
It is difficult to believe, but this definition provided grounds for the
important mathematician Edmund Landau, who taught at Gottingen
University in Germany, to lose his chair in 1933, since it was stated
that such a definition was "un-German". You will find further details
of this disgraceful story later on (see page 213).

1. The State of Pi Art

11

The number 7r has fascinated people throughout the ages, but it is

unlikely that there have ever been as many 7r enthusiasts as there are
today. This is suggested by the fact that on the Internet world-wide
web (WWW) there are at least 200 home pages explicitly devoted to 7r
and thousands of online documents referenced by them. It is possible
to obtain everything to do with 7r from the Internet, including some
clever things. Several of the statements made in this book you will also
find somewhere in the Internet. You will find it is quite different with
our CD-ROM, which contains a lot of information you will probably
not be able to read anywhere else. It also contains a large number
of 7r-relevant Internet addresses in the file pilinks. htm and in the
pilinks directory.
There are also websites on which renowned 7r-researchers introduce
themselves, for example, Yasumasa Kanada and the brothers Peter
and Jonathan Borwein, who developed the algorithms which Kanada
used to perform his 7r calculations. The Borweins and a number of
other researchers also make their publications available for reading
and downloading.
There are a number of cheerful7r clubs on the Internet. It is possible
to become a formal member, although in some cases you have to pass
an entry test. This test often consists of reciting by heart the first digits
of 7r. For example the "Freunde der Zahl Pi" (Friends of the number
Pi) in Vienna, which also publishes a journal called "pi vobiscum",
requires successful candidates for membership to be able to recite the
first 100 digits by heart 5 . Some aids for remembering 7r are presented
in Chapter 3.
When 7r is under discussion, often the number e is not far away.
This may be because that number, which is also known as the "base
of natural logarithms" and whose value 2.71 ... is fairly close to 7r, is
also a transcendental number. Another reason could be that the two
numbers are linked by a wonderful formula by Euler (1.10), to which
we shall refer below.
Otherwise, 7r and e do not have a lot in common. e is a relatively
recent discovery, only about 400 years old, whereas 7r is 10 times as
old; e is an offshoot of infinitesimal analysis which was developed in
the 17th century when mathematicians learned how to calculate with
limits. e is just such a limit. A commercial way of explaining e is to ask
to what value an initial sum of money would grow in a year if the rate
5

http://pi314 . at/

12

1. The State of Pi Art

of interest payable on it was 100% per interest period and this interest
period became ever shorter, from a year to half a year, three months
etc. until eventually it became "infinitesimally small". The answer to
this question is that the end value is e times as great as the initial
investment.
The number e has been calculated to far fewer digits than the number 7f. As of February 1999, 200 million digits had been calculated by
Sebastian Wedeniwski, exactly one thousandth of the number of digits
to which 7f has been calculated. (The Chudnovsky brothers are said
to have calculated e to 1 billion digits, but this has not been officially
confirmed.) Due to this disparity it is assumed that e is in principle
more difficult to calculate than 7f, but this has not been proven.
The latest world record makes 7f the mathematical constant which
has been calculated to the greatest number of decimal places. The
inverse value of 7f, i.e. ~ = 0.3183098661 ... , has similarly been calculated to 206.1 billion decimal places. Then comes the square root
of 2 (y'2 = 1.4142135623 ... ), with 137 billion (~ 237 ) decimal places
calculated. All three calculations have been performed by Kanada.
These are followed (as of October 1999) by e = 2.7182818284 ... (as
mentioned above, either 1 billion or 200 million), Riemann's number ((3) = L:~=1 ~ = 1.2020569031 ... (128 million) and In(2) =
0.6931471805 ... (108 million).
A table of such world records can be found on the Internet at
http://www.lacim.uquam.ca/pi/records.html.
Not only are the first 206.1 billion decimal places of 7f known, but
individual digits beyond that are also known. Thus, for example, since
February 1999 we have known the 10 trillionth digit of 7f from its
hexadecimal representation, an A. This and other such astonishing
findings resulted from the use of a new approach to the calculation
of 7f which was discovered quite by surprise in 1995. This subject is
discussed in Chapter 10.
Whereas the decimal representation of the 206.1 billion 7f digits is
still awaiting an artistic representation, the formal representation of 7f
is always an art at the same time. No doubt the fact that there are
grandiose formulae for 7f has contributed to its fascination. Here is a
selection of the available formulae, arranged in historical order:

1. The State of Pi Art

13

1. Fran<;ois Viete (1540-1603) developed the first infinite product for

2/7f in 1593:

v'2. J2+v'2 V2+J2+v'2

(1.6)
7f
2
2
2
The formula shows that 7f can be expressed solely in terms of the
number 2 (Compare the formulae (16.137) on page 234).
2. Lord William Brouncker (c. 1620-1684) made his mark in 1658
with the first ever continued fraction for 7f:
4
12
(1.7)
;: = 1 + -----3"2---

2+-------;;--52
2+--~,.--

72

2+-2+ ...
The unusual feature about this continued fraction is the fact that
it contains a regular pattern (12, 32, 52, ... ), however the price for
this is that it is not a simple continued fraction as its numerators
do not = 1. There is also a simple continued fraction for 7f, but, as
mentioned, it does not exhibit any regularity.
3. From 1650 to at least 1973 virtually all calculations of 7f were
performed using arctan formulae. The arctan formula used most
frequently was first developed by John Machin (1680-1752), who
used it in 1706 to generate what was then a world record of 100
decimal places:
7f
1
1
- = 4 arctan - - arctan (1.8)
4
5
239

4. The next sequence discovered by the Indian mathematician S. Ramanujan (1877-1920) in 1914 has the attraction that it converges
extremely quickly. Every term in this series produces 8 accurate
digits of 7f:

~=
7f

v's

(4n)! (1103

9801 ~ (n!)4

26390n)
396 4n

(1.9)

5. But of all the mathematical formulae, the prize goes to the following theorem by Leonhard Euler dating from 1743. It combines five
base quantities 7f, e, i, 0 and 1) and four basic operators (+, =, .
and exponentiation) into a single expression:

14

1. The State of Pi Art

(1.10)
This formula has always held a certain fascination, possibly because it is attractive to look at, although it is not at all easy to
understand. What it says is that that if the transcendental number
e is first raised to the power of the transcendental number 7r and
then raised again by the imaginary number i, the result is equal to
-1.
After the important American mathematician Benjamin Peirce
(1809-1880) had proved the formula to his students, he cried:
"Gentlemen, that is surely true, it is absolutely paradoxical; we
cannot understand it, and we don't know what it means. But we
have proved it, and therefore we know it is the truth." [39, p. 585].
In the meantime there is at least one relationship between
e which comes out without imaginary numbers [73, p. 98]:

e+1 =2+4I:
1
e -1
r=l (27rr)2

+1

7r

and

(1.11)

This formula establishes a link between the "growth number" e

and ever greater circumferences (27rr).
Again and again the number 7r has attracted poets and philosophers. In the short story Contact by Carl Sagan an extraterrestrial
tells an Earth woman that 7r contains an important message for us in
encrypted form. At some point in 7r, the woman is told, the digits suddenly cease to be randomly distributed numbers and then for a long
time only zeros and ones occur, representing the product of 11 prime
numbers. It is thus an ll-dimensional message. The sender is thus using mathematics to communicate with humans. His message will even
be authenticated, as the zeros and ones only occur in the decimal representation of 7r, which in turn is only meaningful to beings who possess
ten fingers. This important message has been in 7r for trillions of years
waiting until10-fingered mathematicians with fast computers are able
to unravel it - which, of course describes present-day man.
This reflection entirely explains the enthusiasm for 7r all over the
world. It also comes at the right time for 7r researchers, who always
need new reasons to justify their research grants.
Unquestionably, 7r is a "natural constant" which is valid throughout the universe, applying to all living creatures and in the same way.

1. The State of Pi Art

15

It would thus seem well suited to the initiation of communication

with strange beings. It has therefore been suggested that Jr should
be painted on all the space probes which leave our solar system.
Douglas R. Hofstadter, who caused a sensation a few years ago with
three ultra-clever books, presents the following sentence in "Metamagicum" .
If Jr were equal to 3, ...

What Hofstadter means is, firstly, that if Jr were equal to 3, no

one would bother to write the sentence in the first place. Instead, it
would be the case that Jr is equal to 3, so the subjunctive would not be
necessary. Secondly, all circles would be hexagons, as only hexagons
have the perimeter: diameter ratio of 3. And we would probably not
use a round 0, but a hexagon instead.

Algorithms, Computers and Arithmetic

The history of the calculation of Jr may be broken down into three
distinct phases.
The first era began around 250 Be with the Greek mathematician Archimedes of Syracuse. Whereas up to then only experimental
methods had been used, he was the first person to develop a systematic
procedure for arriving at limits for Jr. He calculated the circumferences
of regular polygons which he placed in and around a circle. The outer
polygons have a perimeter that is greater than the circumference of the
circle and therefore provide an upper approximation for Jr, whereas the
inner polygons have a smaller perimeter and therefore produce a lower
approximation. The more sides the polygons used have, the smaller the
difference between these approximations. Archimedes began with regular hexagons and progressed through 12-, 24-and 48-sided polygons
to polygons with 96 sides. Using this method he obtained for Jr a lower
limit of 3i~ (= 3.1408 ... ) and an upper limit of 3~ (= 3.1428 ... ).
Both values are accurate to two decimal places, the first being more
accurate than the second, which is still more accurate than our decimal
3.14.
From then on, anyone who found two statements too much used the
upper limit of 3~ for Jr, which as a result became possibly the longestliving standard in the world. In the Middle Ages, a few scholars still
believed this to be exact. But even more important than the goodness

16

1. The State of Pi Art

of fit of the two approximations was the fact that Archimedes and his
polygon method determined the direction in which virtually all succeeding 7r mathematicians proceeded for nearly 2000 years. Over the
centuries mathematicians increased the number of sides in the polygons used for the calculations until by the end of this era of geometric
approximation in 1630 AD, 7r was known to 39 decimal places.
The second era began in the middle of the 17th century after the
discovery of infinitesimal analysis and infinite expressions. A specific
method, namely the arctan formulae method (see Chapter 5), dominated 7r calculations after that for more than 300 years until around
1980. The first extensive calculation of 7r on this basis produced exactly 100 digits in the year 1706 (using paper and pencil), and the last
one in 1973 (using a computer) somewhat more than 1 million.
We are currently in the third era of 7r calculation. It began around
1980 when mathematicians discovered how to utilise a combination of
three independent developments.
The first development was the speeding up of an apparently simple
arithmetic operation, namely the multiplication of long numbers. As
this is the primary operation involved in calculations of 7r, speeding it
up reduces the total time required by a similar amount. Surprisingly,
the new method of multiplication known as FFT multiplication has
only been around since 1965. Under this method, the time required
to perform multiplication operations no longer rises quadratically as
the number of factors is increased, as had been the case in what was
previously virtually the only multiplication "school", but more or less
linearly. In Chapter 11 we discuss this subject in more detail; a numeric
example in that chapter indicates that the time required to multiply
two factors each of one million digits using the new multiplication
procedure is reduced from one day to three seconds.
The second major breakthrough was the development of highperformance algorithms specifically designed for calculating 7r. They
surpass the arctan formulae by several orders of magnitude.
To the extent that these algorithms still utilise infinite series, they
are much more productive than even the best arctan series. We have
already mentioned such a series produced by Ramanujan with 8 digits
per series term (1.9). An even more effective series was developed by
the Chudnovsky brothers (8.7), who used it in 1989 to calculate 15
digits per term and thus to attain what was then a world record of
1 billion digits.

1. The State of Pi Art

17

Other 7r algorithms which operate in an entirely different fashion

were developed at the beginning of the 1980s primarily by the Canadian brothers, Jonathan (born 1951) and Peter Borwein (born 1953).
One of these algorithms is capable of discovering in each step four
times as many accurate digits as in the previous step. This "quartic"
(Le. gigantic) progression enabled Kanada to calculate his 206.1 billion
7r digits in 1999 in only 20 iterations. The general design of such procedures is discussed in more detail in Chapters 7 and 9. The German
mathematician Carl Friedrich Gauss (1777-1855) had in fact already
developed precisely such an ultra-fast algorithm in the 19th century.
However, for reasons which are not entirely obvious his procedure was
overlooked for over 170 years.
The third development, which is virtually taken for granted these
days, is the performance explosion of the computer. For some time now
we have observed that the "standard computer" doubles its speed at
least every two years. This alone has contributed by a factor of 1000
or more to the explosion in 7r digits since the beginning of the 1980s.
Measured in these terms, the 2 million digits which brought Kanada
his first world record in 1981 are the equivalent of the first 2 billion
digits in his latest world record in 1999.
With all three developments harnessed in parallel, there has been
a huge increase in productivity. Since 1981 the number of known 7r
digits has increased 100,000-fold to the current record of 206.1 billion
digits. This is the equivalent of a growth factor of 190% per year. Few
other technologies have developed so rapidly.

Why?
What motivates people to pursue those elusive 7r digits? For practical
calculations, 10 or even fewer digits are generally sufficient. 39 digits of
7r are sufficient to calculate the volume of the universe to the nearest
atom. Some scientific applications require intermediate results with
significantly more digits than the final result, but we are talking here
of no more than a few hundred digits. It is conceivable that studies of
mathematical problems using computers could require a few thousand
digits, but certainly no more than that [13J. The precision to which
7r can be calculated far surpasses any practical need, so why continue
beyond that?

18

1. The State of Pi Art

One practical use for large 7r programs is to test computer systems.

Calculating billions of 7r digits means executing trillions of arithmetical
and logical single operations, so that if there are any hardware errors in
a computer subjecting it to such an ordeal is bound to uncover them.
In fact there are known instances in which large 7r programs have
brought to light subtle logic errors in computers. Error-free running of
a large 7r program today constitutes an essential ingredient of quality
testing new processors.
The number 7r is the classical and ultimate testbed for numeric
analysis. Because so many decimal places are known with certainty,
new techniques and methods can be verified using 7r as a yardstick.
Every extra digit that is acquired magnifies this advantage.
The main reason for the preoccupation with 7r, however, is the
unresolved questions which this number raises. What we do not know
about 7r is not only enormous, but it is also enormously interesting.
7r research offers extraordinary breadth. It expands far into analysis, number theory, function theory, complexity theory, the study of
algorithms, statistics and other areas beside. Such a wide spectrum
naturally exerts an exceptional fascination.
For 4,000 years 7r has proved an inexhaustible source of new discoveries and surprises. This has been reaffirmed only in the last few years,
which have seen the discovery of novel procedures for the calculation
of 7r, such as the "spigot algorithm" (see Chapter 6) and the "BBP
algorithm" (see Chapter 10).
Empirical data can often be an advantage when approaching theoretical issues, and the digits of 7r provide a wealth of empirical knowledge. Moreover, today we possess aids, such as computer algebra, to
enable us to tackle theoretical questions experimentally, and here too
it is helpful to have a large stock of digits available.
Every world record achieved is crying out to be bettered. Anyone who has the necessary ability and (access to) resources to achieve
something which no one else has yet managed will try. This is true
both of sportsmen and scientists.
7r mathematics does not end with the development of formulae,
theorems and algorithms. The calculation itself is still mathematics,
at least calculations beyond the first million digits. Without in-depth
knowledge of arithmetic, asymptotic behaviour and Fourier transforms
it is not possible to write a 7r program that will keep up with the
competition.

1. The State of Pi Art

19

Someone once said that one can turn a normal person into a 7r fan
but that the reverse is impossible. He was quite right. The enthusiasm
and pleasure which leaps out of texts written by 7r digit-hunters suggests that 7r lives a life of its own in these people. The authors have
certainly been caught by the bug.
And last but by no means least, new discoveries about 7r meet
with the interest even of people who are not mathematicians. Put
another way, 7r matters. 7r is one of those rare objects in mathematics
which act as a showpiece. The hunt for ever more 7r digits provides
the opportunity to gain an insight into the research laboratories of
mathematicians, who otherwise tend to live out of sight of the public.

New goals
For some years, since around 1995, 7r mathematicians have increasingly
been turning to new goals. "Pi: a 2000-year search changes direction,"
so Victor Adamchik and Stanley Wagon describe the new trend [2]. Instead of continually calculating more 7r digits all over again, people are
now devoting ambition and resources to the calculation of individual
digits at the far end of 7r.
The impetus for this new development came from the unexpected
discovery of a method which allows individual hexadecimal digits in 7r
to be calculated without having to calculate all the digits which come
before. The basis for this is the following new formula:

7r

1 (4+

~ 16n
00

8n

1)

1 - 8n + 4 - 8n + 5 - 8n + 6

(1.12)

The inventors of this formula (David Bailey, Peter Borwein and

Simon Plouffe) won world-wide admiration when they presented it in
October 1995 (see Chapter 10). Up to then it was difficult to imagine
that individual digits could be resolved independently. A 7r calculation
is like an enormous "tree" of single operations, and the task of calculating a single "leaf" would appear impossible, rather like removing
a needle from a haystack without moving the hay on top of it to one
side. But with the new formula it is possible, as it were, to fly over 7r
in helicopter and then pick individual locations on which to land. The
critical factor here is the magnitude of 16 n in the denominator of all
the terms in the series.

20

1. The State of Pi Art

When the inventors of the "BBP series" (named after the initial
letters of their names) introduced the new theory, they also delivered
proof of its efficacy with the announcement of the identity of the 10
billionth hexadecimal digit of 7r, which at the time was far off from
discovery.
The new idea seems to have electrified 7r researchers all around the
world. Since then they have found more formulae of the same type as
the BBP series and even an algorithm for generating additional such
formulae. It was with one of these new formulae that digits 100 times
further on in 7r than the last 7r digit calculated by Yasumasa Kanada
were successfully calculated.
The most recent and youngest world record holder in this 7r discipline is Colin Percival, who was only born in 1981 and has only recently
gone up to university. Unlike Kanada, he did not perform his individual digit calculations on a university supercomputer but divided up the
calculations over 1700 computers on the Internet all over the world.
He invited owners of computers to help him with his project, called
the PiHex project, and sent them his calculation program bye-mail,
which the computers then executed in their dead time when they did
not have anything more important to do. Every computer took on a
different subtask and Percival then put the individual pieces together.
In this way after two years elapsed time and almost 700 years of computer time (P90 CPU) he attained a new world record on the 11th
September 2000: he had found the 250 trillionth hexadecimal digit of
7r, an E. Because this E consists of the four binary digits 111 0, Percival
was able to proudly announce: 6 :

The quadrillionth bit of 7r is O.

The one slight imperfection of these 7r calculations is that they only
provide hexadecimal or binary digits of 7r, rather than decimal digits.
To be precise, to the present date we still do not know a way to calculate individual decimal digits in 7r which is quicker than calculating
all the digits which precede them.
But this problem seems to be only a question of time. At any moment someone could devise a "BBP formula" for calculating individual
decimal digits of 7r. We look forward to the day.

http://www.cecm.sfu.ca/projects/pihex/announcelq.html

2. How Random is 7r?

2.1 Probabilities
What is the probability that decimal position s of 7r will contain the
digit z?
At first glance the answer appears quite simple. Because position s
must be occupied by one of the 10 decimal digits 0,1, ... ,9, the probability of anyone of them occurring is exactly 10%.
On the other hand, a so-called probability subjectivist, i.e. a follower of the teachings of Thomas Bayes (1702-1761), would say that
the answer depends on how much he already knows about 7r, since according to this philosophy, "probability" measures the extent of what
one does not know. If he already knows which number occurs in position s, then the probability that it is a z is either 0% or 100%. If on
the other hand he knows nothing about position s, then as far as he
is concerned the probability that the number will be z is 10%. "No!"
mathematicians or followers of classical philosophy will respond, "the
basic question is fundamentally flawed." For the positions of 7r are not
random but precisely determined. For example, in the second decimal
place there is a 4, so there is no point in asking what the probability
is that a 5 will occur in that position. Where certainties reign, there
is no need for probabilities.
But if the basic question is now not a reasonable mathematical
question, is there another question which we can ask about the "randomness" of 7r? There is indeed: "Is 7r normal?"

2.2 Is

'iT

normal?

Mathematicians call a decimal number "normal" when every sequence

of n digits is equally likely to occur. Thus, for example, in a normal

22

2. How Random is 7r?

number the digit 0 occurs with a frequency of 1/10 and the sequence
of digits 357 with a frequency of 1/1000.
If all ten digits occur equally often in a decimal number, it is said
to be "simply normal" .
The question as to whether a number is normally distributed is
only meaningful and interesting if it is of infinite length. The number 7r
has been proven to be irrational, i.e. it is a number that contains an
infinitely long and non-periodic decimal fraction, and therefore it is
possible that it could also be normal. But we do not know this. No one
has succeeded in proving either that 7r is normal or that it is not. Nor
has anyone been able to demonstrate that such a proof is impossible.
If 7r were not normal, then this would manifest itself in individual
digits or sequences of digits occurring at unequal probabilities. For
example, the number 7 might occur more frequently than 3, or perhaps
from some point in the sequence there would be no further instances of
the combination 314159265. Although there are certainly intervals in
which such irregularities (or regularities, depending on how one looks
at it) occur, it is not known today whether this is also the case when
one views the entire sequence of 7r.
In a number which is normal, there should always be a point somewhere where, say, the number 5 occurs a million times in succession.
But we do not know whether such a sequence also occurs in a number such as 7r, of which we do not know whether or not it is normal.
In a random sequence it easy to find an interval which apparently is
non-random.
One thing is certain: the digits of 7r are not random numbers because they can be calculated.
The fact that 7r is not only irrational but also transcendental does
not exclude the possibility of regular patterns occurring within its
number sequence. Conversely, the existence of a pattern in the decimal
sequence of 7r does not mean that 7r is "non-normal". To illustrate
this point, let us consider the artificial number 0.123 ... 10111213 ...
which is made by writing all the natural numbers n in sequence after
the decimal point. Clearly this number follows a pattern, yet it is
normal. This has been proven by Ivan Niven [86].
The preoccupation with the question of whether 7r is normal has
provided the impetus to many statistical studies. The results of these
studies are in part instructive, in part disappointing, in part curious.
We will now report some of those findings.

2.2 Is

'IT

normal?

23

There are several statistical tests available which enable one to

ascertain to what extent the decimal digits in 7r are randomly arranged;
however, the answers are always expressed as statements of probability.
If, for example, such a test is used on a particular roulette table, then
no result can provide 100% certainty that the table has not been rigged,
as even on the "cleanest" table an "impossible" sequence will occur at
some point. However, it is possible to quantify the probability of this.
One statistical test, for example, is the "poker test". In the poker
game, seven different combinations of five cards are considered "poker
hands", e.g. "a pair" or "full house". In the poker test, the actual
number of poker hands is compared with the number that one would
expect. For example, 2 million "poker hands" occur in the first 10
million decimal digits of 7r. Their distribution is as follows [117]:
Poker hand
All different
One pair
Two pairs
Three identical
Full house
Four identical
Five identical

Pattern
abcde
aabcd
aabcd
aaabc
aaabb
aaaab
aaaaa

Actual number
604,976
1,007,151
216,520
144,375
17,891
8,887
200

Expected number
604,800
1,008,000
216,000
144,000
18,000
9,000
200

The comparison does not exhibit any striking departures from the
numbers one would expect. In fact, the X2 test produces the value of
53% for this distribution, which means it is completely normal with
no striking features. Only if percentages above 95% or below 5% were
obtained, would one view a distribution as "suspect".
If one considers smaller intervals, e.g. intervals of 500,000 decimal
places, which are sufficiently large to allow one to expect even the
rarest poker hand, "5 identical", to still occur 10 times, the picture
is different. For example, the interval from 3,000,001 to 3,500,000 has
the following distribution:

24

2. How Random is 7f?

Poker hand
All different
One pair
Two pairs
Three identical
Full house
Four identical
Five identical

Pattern
abcde
aabcd
aabcd
aaabc
aaabb
aaaab
aaaaa

Actual number
30,297
50,263
10,877
7,156
927
459
21

Expected number
30,240
50,400
10,800
7,200
900
450
10

In this case, an opponent would very likely suspect that all was no
longer as it should be, especially if he examined the number of "Five
identical". The probability that this distribution occurs in a random
sequence is only 2.6%, so the alarm bells would definitely start ringing.
On the other hand, in the next equal-sized interval, the probability
reverts once more to 69.5%, and when one considers both intervals
together the overall probability is 32.0%, which would certainly not
arouse suspicion.
There is also the possibility of the opposite occurring, namely that
the distribution of the poker hands matches the expected values so
exactly as once again to look suspicious. Such an interval begins, for
example, in position 4,250,001. In that sequence the distribution of
poker hands is only 0.5% away from the expected figures.

2.3 So is

1('

not normal?

In an early edition of his "Mathematical Games" column Martin Gardner reports a conversation with a "Dr. Matrix" [54]:
"Dr. Matrix borrowed my pencil and rapidly jotted down the value
of 7r to 32 decimals."

(l,

3.141592653589 93238462643383279 0 ...

yY

Yy

2.4 The 163 phenomenon

25

"Mathematicians consider the decimal expansion of 1r a random series, but to a modern numerologist it's rich with remarkable patterns."
He bracketed the two appearances of 26. "Twenty-six, you observe,
is the first two-digit number to repeat. Note how the second 26 marks
the center of a bilaterally symmetric series." Dr. Matrix inserted vertical bars to enclose 18 numerals, then bracketed six other number pairs
shown in the illustration. "The number pairs 79, 32 and 38 on the left
are balanced by the same three pairs in reverse order on the right!" He
also called attention to the sets of five digits on each side of the first 26:
"The first set sums to 20, the number of decimals preceding the second
26. The second set sums to 30, the number of decimals preceding the
second bar. Together they sum to 50, the two-digit number following
the last bar. The sequence between bars starts with the 13th decimal
and 13 is half of 26. The three pairs - 79, 32 and 38 - have six digits
that sum to 32, the pair in the middle as well as the total number of
decimals shown. The 46 and 43 on each side of the second 26 sum to
89, the number pair preceding the first bar. .. "

2.4 The 163 phenomenon

If you were not convinced by Dr. Matrix, then what do you think of
this next case?
The number

(2.1)
looks particularly contrived. It contains nothing but "unusual" components - the transcendental numbers e and 1r and the prime number
163, plus a square root and an exponentiation. Anyone would suppose
that the combination of such numbers must be an "awkward" number.
But in fact expression (2.1) produces the following remarkable result:
e7rv'f63

= 262,537,412,640,768,743.9999999999992...

(2.2)

That is very nearly an integer. The difference between this and an

integer is less than 10- 12 .
This "163 phenomenon" was brought to light by the Scottish mathematician Alexander Aitken (1895-1967) [15, p. 386]. As far as he was
concerned it was no accident. In the mathematical world Aitken has

26

2. How Random is 11"?

become well-known not only for his academic achievements (4 books

and 70 publications) but also for his legendary abilities in mental arithmetic. He was able to break down even large numbers into their factors in his head and to say straight off and with unswerving accuracy
whether a number called out to him was a prime number or not. One
can just hear him saying, "The number 'smells' prime." To elucidate
the 163 curiosity one would no doubt need to be cast of the same
mould.
One would perhaps first try seeing whether there any other values
of n for which e'TrVn is almost integer. In fact there are some. Have a
look at the following table:

n
6
17
18
22
25
37
43
58
59
67
74
163

2,197.990 .. .
422,150.997 .. .
61,4551.992 .. .
2,508,951.998 .. .
6,635,623.9993 .. .
199,148,647.99997.. .
884,736,743.9997 .. .
24,591,257,751.9999998 ...
30,197,683,486.993 ...
147,197,952,743.999998 ...
545,518,122,089.9991. ..
262,537,412,640,768,743.9999999999992 ...

The author of the list, Roy Williams, offered a prize for anyone who
could either convincingly demonstrate that this is a chance occurrence
or else explain to an intelligent university graduate why it is not a
chance occurrence.
One thing is certain: the length of the list alone suggests that not
all of these instances are chance. The list is actually not even complete, as there are many other values of n above 163 for which e'TrVn
produces three or more nines after the decimal point, for example,
n = 232(5 nines), 719(4), 1169(4), 1467(8),4075(5),5773(4).
But knowing this does not make us any the wiser. For the mathematical proofs are so complicated that, however highly one rates university graduates, the prospects of winning the prize are extremely
slim because the phenomenon defies simple explanation. The topic be-

2.4 The 163 phenomenon

27

longs to the theory of modular equations, of which only a few experts

have a full grasp [61J.
For n = 43, 67 and 163 there is the explanation that here the "j
function"
1
j(n) = - + 744 + 196884q + 21493760q2 + 864299970q3 + ... (2.3)
q
with 1.q = _e7rfo
is an integer and furthermore a perfect cube. For n = 43 it has the
value -960 3 , for n = 67 the value -5280 3 and for n = 163 the value
-640320 3 . Why that is so is understood only by specialists. But if
we assume that it does behave like this, then in these cases the phenomenon can be explained from the fact that in the series expansion
(2.3), e 7rfo differs from an integer by the sum of the terms q, q2 etc.,
which are very small and can be neglected.
At any rate, if one knows these facts, then the relationship (2.2)
can be written in even more spectacular fashion as follows:

?je7rv'I63 -

744

= 640319.9999999999999999999999993. . .

(2.4)

There are a few other values of n for which approximating series expansions can be used by way of explanation, even if they cannot explain the j function. The Indian mathematician S. Ramanujan
(1887-1920), of whom we shall hear more in Chapter 8, discovered
some such series. However, he was not given to lengthy explanations,
and his essay [96J is no exception. As far as one can tell, he followed
a semi-heuristic path which involved combing through the first few
hundred degrees of modular equations. In the course of so doing he
came across the degrees 22, 37 and 58, which in a certain combination actually produce integers. For example, he found the relationship
e 7rV37 = 64((6 + v'37)6 + (6 - v'37)6) - 24 - 4372e- 7rV37 - ... , wh~se
approximation then "explains" the case of n = 37 in the list if one
leaves out the terms after the - 24.
With these six values of n, those n which are bigger by a quadratic
factor are then also explained, e.g. n = 232 = 58 . 22 or n = 1467 =
163.3 2. But we are still no closer to explaining the (many) other cases,
e.g. the spectacular case of e7rV7T9 which is only 0.000013 away from
an integer.
The assumption that combination of many complex variables inevitably produces a complex result is deceptive, as we have seen. This

28

2. How Random is 7r?

was also the experience of Donald E. Knuth of Stanford University,

California. Knuth, whom we shall encounter several times more in this
book, is the author of the multi-volume work The Art of Computer
Programming, which is one of the most cited (and possibly even most
read) books in the field of computer science. In the introduction to the
second volume, Knuth tells the story of his super-random number generator [77, p. 4], with which he planned to generate the most random
of all random numbers.
To this end Knuth wrote a program which was designed to be especially random. For example, on every pass it branched off to a random
place in the program and for each random number it performed a random number of loop passes. Knuth also coded the program in such
a complicated way that no one could understand it. Then he started
with an initial value, which naturally was chosen at random. He expected that since his program contained so much randomness it would
produce incredibly random numbers.
But what happened? After starting up, the program converged almost immediately on the 10-digit number 6065038420, which thereafter transformed itself back into itself in only 27 cycles. When he
tried a different initial value it took 3178 cycles, but even this value
may be regarded as lamentable.
The conclusion he drew was that one should never generate random
numbers with a random method.
This applies to investigations into 7r as well. Thus, we should add
that not only is the 163 phenomenon no chance occurrence but it even
constitutes the background for one of the fastest methods of calculating 7r. It was on the basis of the underlying theory that the Chudnovsky
brothers developed their brilliant "Chudnovsky series" (8.7) which enabled them in 1989 to break through the one billion decimal places
of 7r mark for the first time

2.5 Other statistical results

Naturally the digits of 7r have been counted by computers in many
different ways. For example, when Yasumasa Kanada issued his statement about his 206.1 billion decimal place world record in October
1999, he published along with it the distribution of the first 200 billion
digits of 7r. It goes like this:

2.5 Other statistical results

Number

1
2
3
4
5
6
7
8
9
Total

29

Occurrences
20,000,030,841
19,999,914,711
20,000,136,978
20,000,069,393
19,999,921,691
19,999,917,053
19,999,881,515
19,999,967,594
20,000,291,044
19,999,869,180
200,000,000,000

As you see, the numbers are very evenly distributed. The number
which deviates most is the number 8, but even this number deviates by
only 0.00001% from the mean. The X2 test produces a probability of
around 55% for the randomness of this distribution, i.e. "statistically
insignificant" .
Kanada also searched through his record-breaking 7r for interesting patterns. He reports, for example, that the sequence 01234567891
occurs five times among the 206.1 billion decimal places of 7r, which
is somewhat striking as one would only expect this sequence to occur
twice. Another peculiarity is the fact that the sequence 543210987654
appears for the first time relatively late in 7r, at the end of the 206
billion, in position 197,954,994,289, although there is a 50% probability that this sequence would appear in the first 69 billion decimal
places. This probability figure is obtained from inserting w = 0.5 and
k = 11 in the formula below (2.5), which states that the probability of
a k-digit number not appearing in a number containing n digits is
logw

n
w

= log(l _ lO-k)
= (1 - lO-kt

(2.5)
(2.6)

According to the inverse formula (2.6), the probability of the fact ascertained by Kanada (despite the inherent difficulty of making such a
statement) is only 13%. Even so, such numbers would not arouse the
suspicion of any statistician.
If one divides up the digits of 7r into blocks of 10 decimals, what
is then the probability that any given block will consist of all dif-

30

2. How Random is 7r?

ferent numbers? One such block could, for example, be the sequence
0123456789.
There are three points to make here.
Firstly, the probability of this case is greater than one might think.
We are looking for 10! hits out of a total 1010 possibilities. This fraction
can be calculated easily with a pocket calculator, and the result is
the astonishingly large value of around 0.036% = 1/2755 .... In other
words, a block of 10 different figures will occur on average as frequently
as every 2755 blocks.
Secondly, in reality this case actually occurs much earlier than is
suggested by this probability, which is already quite large. In fact a
hit occurs as early as the seventh block, as you can see from this list
of the first seven blocks of 10 digits:

3.1415926535
8979323846
2643383279
5028841971
6939937510
5820974944
5923078164

Thirdly, this result is independent of whether the blocks begin before or after the decimal point of Jr, because the same number 4 appears
immediately before as well as at the end of the block.

2.6 The Intuitionists and

7f

Kanada's calculation feat has an interesting historical association, as

was pointed out by Jonathan Borwein [29].
At the beginning of the 20th century the Dutch mathematician
L.E.J. Brouwer developed a theory which bears the name "Intuitionism" . In short, Brouwer questioned "the Law of the Excluded Middle" ,
which had been sacrosanct to mathematicians from the time of Aristotle: A statement is either true or false, "tertium non datur" ~ there
is no third possibility. Brouwer argued that this law cannot be applied
generally. By way of proof he asked, for example, whether it is true
or false that the sequence 123456789 occurs in Jr. Given the state of
Jr research in his day, he could comfortably assume that this question
would never be answered, since to do this, one would "probably" have

2.6 The Intuitionists and

7r

31

to know 6.9 billion decimal places of 71", whereas at the time (1908) only
707 decimal places were actually known. So if one cannot say whether
this statement is true or false then, according to Brouwer, no one can
invoke the Law of the Excluded Middle for this statement.
Today, that question has been answered (thanks to Kanada), and
with it a number of examples which Brouwer and his disciples used
to prove their theory have been rendered of no value. One of these
examples was the following.
One of the implications of tertium non datur is that proof of the
impossibility of an impossible characteristic is also a proof of the characteristic itself. Brouwer's counterarguments went as follows. I write
down the sequence of decimal digits for 71" and underneath it the decimal fraction p = 0.33333 ... , which I terminate as soon as the sequence
0123456789 occurs in 71". If the 9 in the sequence is the kth decimal
place, then p = (10 k - 1)/(3 . 10 k ).
Now let us assume that pis irrational. If that is so, then p cannot be
equal to (10 k -1) / (31 Ok), as that would be a fraction and consequently
a rational number. Hence the stated number sequence should also not
occur in 71". But if it does not occur in 71", then p = 1/3, which again is
a fraction, contrary to the hypothesis.
In this way, the assumption that p is not a rational number leads
to a contradiction. Despite this we cannot assert that p is rational, as
that would mean that we would have to be able to state two integer
numbers p and q for the numerator and denominator of p, and to do
that we would either have to locate a sequence 0123456789 in 71" or else
to prove that there is no such sequence.
This and other examples were thus used by the Intuitionists to support their thesis that the Law of the Excluded Middle could not be
assumed to be always correct. Not only the example cited above but
many other examples were based on the assumption that we would
never know for certain whether the number sequence 0123456789 occurs in 71" or not.
Unfortunately for the theory, we do now know, and so the question
arises of whether this refutes the theory. In fact it does not. The proponents of the theory would only have had to choose a much longer
sequence in their examples, and they can still do that even today, for
example they could suggest a sequence consisting of 100 billion zeros
in 71", and everything would be back to the way it was before. Not only
is Intuitionism not refuted, but in a certain way it is even confirmed,

32

2. How Random is 7l'?

and, indeed, precisely through experimental mathematics, which comprises a lot of 7r research.

2.7 Representation of continued fractions

The decimal sequence of 1l' reveals little that would enable us to draw
conclusions as to the non-randomness of the 1l' digits. So are there
perhaps other representations of 1l' which are more informative'? This
question is actually quite sensible, for every alternative representation
of 7r, such as the hexadecimal representation, is based on a quite different calculation tree, and there is every likelihood that the new variant
of 1l' could look quite different.
A promising approach here is the so-called simple continued fraction
of 7r 1 . As already mentioned on page 1, the "related" transcendental
number e appears quite differently when represented as a continued
fraction compared with its decimal representation. Whereas in its decimal form e begins with 2.7182818284590 ... , expressed as a continued
fraction the number begins like this:
111 1 1 1 1 1 1 1 1 1
(2.7)
e = 2 + l' + 2 + l' + l' + 4 + l' + l' + '6 + l' + l' + '8 + l' +...
The decimal representation of e thus possesses no regularity, whereas
its representation as a continued fraction exhibits a lot of regularity:
the formula is made up of sequences of three fractions whose denominators are of type 1 n 1, in which every n is always 2 higher than the
n in the previous sequence of three terms.
The number 1l' does not exhibit any such characteristics. In the case
of 7r, the continued fraction (apparently) does not follow any construction rule:
111111111111
1l'=3+- - - - - - - - - - 7 + 15 + 1 + 292 + 1 + 1 + 1 + 2 + 1 + 3 + 1 + 14 + ...
One can perhaps conclude from this that 7r is structured even more
randomly than e.
In 1985 William R. Gosper calculated 17 million continued fraction
positions of 1l', so it is possible to perform statistical analyses of these
as well.
The results for the first 8192 elements (3, 7, 15 etc.) include the
following:
1

We go into continued fractions in more detail in Section 4.4 on pages 64.

2.7 Representation of continued fractions

Mean
Mean
Mean
Mean

value
value
value
value

of 1st 2048 elements

of 2nd 2048 elements
of 3rd 2048 elements
of last 2048 elements

33

27.5
15.3
8.2
10.4

Is there anything in this? The mean value of the first 2048 elements
appears to be quite out of line. It is almost twice as big as the mean
value of all the elements and it is also significantly greater than the
mean of the two next equally large sections. The picture does not
change if one splits the number sequence up into smaller sections or
divides it in different ways: every time the first or at least the foremost
mean values are greater than those which come afterwards.
When one examines the data more closely, however, even this discovery turns out not to amount to much. The fact is that quite early
on, at position 431 of the continued fraction expansion of 7r, an unusually large element occurs, 20776, which has the effect of distorting
the statistics. In the first section of 2048 elements shown above, that
single element by itself accounts for 10.14 of the mean value. If one
ignores the influence of that element, then this first section becomes
inconspicuous.
It is unfortunately the case that the simple continued fraction of 7r,
like the decimal (and every other) representation of 7r, contains no
other message than that there is little regularity in the irregularity of
the number.
However, there are other continued fraction representations than
the simple continued fraction and in some of them there is a certain
amount of regularity. One example here is the elegant new continued
fraction for 7r which was developed by 1.J. Lange in 1999:
12
32 52 72
7r = 3 + (2 8)
6+6+6+6+
.
Lange, 1999 [80]

Here, all the elements follow the simple construction rule (2n~1)2
The known 7r digits pass all random number tests with flying
colours. For this reason one can use 7r as a random number generator
- though one must not reveal to "the opponent" where the random
numbers came from. For a lottery company it would not be a good
idea to simply select the next 7 digits of 7r as the hits each week.
So is there at least a ray of hope that a sensation might yet be uncovered in the digits of 7r? Perhaps. The above-mentioned Chudnovsky

34

2. How Random is 7r?

brothers, who have made a significant contribution to 1f research and

are above the suspicion of starting rumours simply in order to create
a sensation, hinted a few years ago (1992) that they might have found
something: "It's unfortunately not statistically significant yet (but) it's
close to the edge of significance." They told the interviewer, Richard
Preston, that they would need many billions of digits in order to be
able to say anything more precise [92J. It seems that what the brothers
were hinting at was a property which eludes common statistical tests,
namely possible "waves" in the decimal sequence. If, for example, the
first, third and fifth billions regularly have fewer than the expected
number of a particular digit and the second, fourth and sixth billions
regularly have more than the expected number of that digit, the X2
test would probably not detect this, but it would still provide major
cause to reflect.
Since April 1999 the Chudnovsky brothers now have their "many
billions of digits", courtesy of their arch-rival in the world record
stakes. They now have the evidence they need to either confirm or
reject their hypothesis. We suspect that it will be a case of the latter,
but we are eagerly awaiting the results.

3. Shortcuts to

7r

Unquestionably the easiest way to arrive at 7r is to follow the cover

of this book, which shows the first few dozen decimal places. 5000
decimal places are listed on page 240f and 400 million decimal places
are contained on our CD-ROM. Another easy way to access 7r is with a
computer algebra system such as Mathematica or a special 7r program,
of which there are several on the CD-ROM and dozens more on the
Internet.
However, this chapter concentrates on approaches to 7r which involve the exercise of a certain amount of initiative.

3.1 Obscurer approaches to

1r

If area A and diameter d of a circle are known, then 7r is equal to 4Aj d2 .

This is precisely the approach which underlies the following curious C
program created by Brian Westley. It calculates 7r to 4 decimal places.
#define _ OO>OO?O:--OO,--F;
int F,OO;
mainO{F _000 ;printf (10%1.3f\n", -4. *F/OO/OO) ; }LOOO
{

- - -- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -- - - - - - - - - - - - -

36

3. Shortcuts to

1f

The "circle" here has an area of 201 (measured using the characters _ and - _) and a diameter of 16 (measured in lines made out of
these characters), so that a value of 3.141 is obtained for "7r". To obtain more decimal places, you need only to make the circle larger, each
additional decimal place requiring the size of the circle to be increased
by a factor of around 10.
The program does not look at all like an ordinary C program, but
it even complies with the ANSI C programming standard. With this
program, Westley took a prize at the International Obfuscated C Code
Contest (IOCCC) in 1988. This competition has taken place every year
since 1984 on the Internet. The winner is the person who creates the
most incomprehensible and most creative C program, which, however,
must be executable. The basic tools used are the concise C syntax
and above all the C pre-processor (as here). The results are usually
artificial yet also artistic. But they are invariably examples of how one
should not write a program.
Here is another example of such programming:
char
_3141592654[3141
]. __ 3141[3141];_314159[31415]._3141[31415] ;main() {register char*
_3_141.*_3_1415. *_3 __ 1415; register int _314._31415. __ 31415.*_31.
_3_14159. __ 3_1415;*_3141592654= __ 31415=2._3141592654[0] [_3141592654
-1] =1 [ __ 3141] =5; __ 3_1415=1;do{_3_14159=_314=0. __ 31415+ +;for( _31415
=0;_314153.14-4)* __ 31415;_31415++)_31415[_3141]=_314159[_31415]= 1;_3141[*_314159=_3_14159]=_314;_3_141=_3141592654+__ 3_1415;_3_1415=
__ 3_1415
+ __ 3141 ;for
C31415 = 3141__ 3_1415 ;
_31415;_31415-._3_141 ++.
_3_1415++){_314
+=_3142 ;
_314=1;_314+=
* _3_1415; _31
=_314159+ _314;
if(! (*_31+1)
)* _31 =_314 /
__ 31415._314
C3141]=_314 '/.
__ 31415 ;* (
_3 __ 1415=_3_141
)+= *_3_1415
= *_31;while(*
_3 __ 1415 >=
31415/3141 ) *
_3 __ 1415+= 10,(*--_3 __ 1415
C3141]; if ( !
) ++; _314= _314
_3_14159 && *
_3_1415)_3_14159
=1. __ 3_1415 =
3141-_31415;}if(
_314+( __ 31415
1= __ 31415 )
while ( ++ *
_3_141==3141/314
)*_3_141--=0
;}while(_3_14159
) ; { char *
__ 3_14= "3.1415";
write3,1),
(--* __ 3_14, __ 3_14
),(_3_14159
++,++_3_14159))+
3.1415926; }
for ( _31415 = 1;
_31415<31411;_31415++)write(
31415% 314-(
3,14),_3141592654[
_31415
] +
"0123456789" ,"314"
[ 3]+1)-_314;
puts *_3141592654=0
,_3141592654))
;_314= *"3.141592";}

3.2 Small is beautiful

37

This work of C-art created by Roemer B. Lievaart looks as if it

is all about the number 7f, especially since it produces 3141 decimal
digits. Ironically enough, it actually calculates not 7f but its cousin,
the number e. Its output therefore begins with 2.7128 ....
We will explain elsewhere which algorithm was used in the program.

3.2 Small is beautiful

Next we present a mini program only 133 characters long, once again
programmed in ANSI C, which we believe is the shortest C program
in the world for the calculation of 7f - naturally only until someone
outdoes us.
a[52514),b,c=52514,d,e,f=le4,g,h;
main 0 {for ( ;b=c-=14;h=printf ("%04d" ,e+d/f))
for(e=d%=f;g=--b*2;d/=g)d=d*b+f*(h?a[b):f/5),a[b)=d%--g;}

Despite its short length, this little program calculates 15,000 decimal places of 7f.
The underlying calculation method is the so-called spigot algorithm, under which the program produces a steady trickle of digits
one at a time without using or requiring previously computed digits.
You can see the method in action for yourself if you run our J AVA
applet spigot/pispigot. htm which is on our CD-ROM. The spigot
algorithm and the little program are discussed in more detail in Chapter 6 from page 77.
This algorithm was discovered by Stanley Rabinowitz, who published it in 1991 in the form of a program written in FORTRAN [94].
The author provided virtually no explanation or proof, instead promising his readers all would be explained later on in a major article. This
appeared a mere four years later in 1995 [95] (with Stanley Wagon
as co-author). One can therefore say that programming of the spigot
algorithm preceded its rationale.
Our version shown above evolved in a number of stages, in which
several authors (Dik T. Winter, Achim Flammenkamp and others)
were involved.

38

3. Shortcuts to

3.3 Squeezing

1T'

7r

through a sieve

can be approximated using a method which recalls the "sieve of

Eratosthenes" devised by the Greek mathematician of that name (284202 Be) as a means of collecting prime numbers.
Beginning with the sequence of natural numbers 1,2,3,4, ... , one
first of all removes every second element, starting from the third element:

11'

1,2,4,6,8,10,12,14,16,18,20,24,26,28,30,32,34, ...
Starting from the fifth element, every third element is now removed:
1,2,4,6,10,12,16,18,22,24,28,30,34, ...
then, from the seventh element, every fourth element:
1,2,4,6,10,12,18,22,24,30,34, ...
and so on. As a general rule, on the kth pass, every (k + l)th element
is removed from the (2k + l)th element.
The procedure produces the sequence of values f(n) = {I, 2, 4, 6,
10, 12, 18, 22, 30, 34, ... }. When these values are sequentially entered
in the formula

f(n)

(3.1)

each result is a closer approximation to 11'. For n = 1,2,4, 8, 16, 32, ...
4096 = 212, the following approximations are obtained:
1,2,2.7,2.9,3.12,3.10,3.08,3.12,3.1405,3.13, 3.1423, 3.1414, 3.1412
Unfortunately the increase in accuracy is only modest. The series converges only of the order of O(n4 / 3 ), which is a little better than linear
convergence. Furthermore, the memory requirements are considerable:
to use f( 4096), as in the present case, for the approximation, one needs
a sieve with 5.34 million holes in it.
K.S. Brown discovered that the sequence of f(n) can also be obtained by intelligent rounding up . His algorithm goes as follows.
Take an integer n and round it to the next multiple of n - 1, then
to the next multiple of n - 2, and so on until the next multiple of 1.
The result is then exactly the same f(n) as one obtains from the sieve
procedure. For example, with n = 10, one obtains the intermediate
values 18, 24, 28, 30, 30, 32, 33, and 34, with the end result f(1O) = 34.

3.4

3.4

7r

7r

and chance (Monte Carlo methods)

39

and chance (Monte Carlo methods)

The needle problem of the Comte de Buffon

During the American Civil War, Captain C.O. Fox was recovering from
a wound in a military hospital. To pass the time, he threw a number
of identical needles in random fashion onto a board on which he had
previously drawn a series of parallel lines each a needle's length apart.
He counted the number of throws and the number of hits, i.e. instances
in which a needle touched or intersected a line.
After 1100 throws, the Captain had determined 7r to two decimal
places. How come?
First of all it seems to have been the Comte de Buffon (1707-1788),
who examined this kind of experiment and in whose honour it is now
known as the Buffon needle problem. In 1777, Buffon showed that the
ratio of hits to throws was 2:7r, or, stated otherwise, that the probability
of a needle thrown at random onto the area coming to rest across one
of the lines was, ~ ~ 63.7%. With this knowledge, Fox was able to
calculate 7r by doubling the number of throws and dividing by the
number of hits.
The interesting thing about the needle problem is that it forges a
link between the "geometric" 7r and the quite different area of probabilities. There are other similar relationships between 7r and chance,
from which other methods of calculating 7r are derived. They are informatively called Monte Carlo methods.

The dartboard algorithm

Imagine a circle with radius r = 1 with a circumscribed square of side
length 2r = 2. "Darts" are thrown at this square, following a random
distribution, whereby a "hit" is deemed to occur every time the dart
lands in the circle. The procedure is roughly similar to the standard
game of darts, except that in this case the "black" is much larger than
in a conventional dartboard.

40

3. Shortcuts to

1f

1
1

The procedure is simplified by only considering the first quadrant.

As the number of throws (n) increases, the number of hits (t) approaches the ratio of the area of the quadrant of the circle to the
quadrant of the board.
lim!
n-+oo n

7r~

A(circle quadrant)
A(board quadrant)
4t
n

= 7rr 2 /4

7r
4

(3.2)
(3.3)

The result is a neat algorithm for approximating 7r. You only need
to simulate a number of throws (as many as possible) and test on each
throw whether the dart lands inside the circle. This test is particularly
simple: if x and yare the co-ordinates of the landing point(O ~ x, y ~
1), then the point will be inside the circle precisely when its distance
from the origin is x 2 + y2 < 1, i.e. x 2 + y2 < l.
In the following C++ program the random co-ordinates x and y
are generated using the standard C++ function rand (); as this function delivers (pseudo-) random numbers between 0 and (the systemdependent variable) RAND..11AX, they must be divided again by RAND..11AX
to obtain co-ordinates in the interval from 0 to 1. The random number
generator is initialised through the standard function srand () with a
time-dependent, and hence varying, value:

vi

II Dartboard-Algorithm for approximating pi

#include
#include
#include
#include

<iostream.h>
<stdlib.h>
<time.h>
<math.h>

3.4

7r

and chance (Monte Carlo methods)

41

int main(void)
{

long
k, n, hits;
const double factor = 1.0

RAND_MAX;

while (1)
{

cout "Enter the no of tosses (or 0 to exit): ";

cin n;
if ( n <= 0) II input <= 0 means end-of-job
break;
II Initialize the random generator
srand((int)clock(;

II Throw n tosses
for (k=hits=O; k < n; ++k)
{

II Find two random numbers within 0 .. 1

double x = rand() * factor;
double y = rand() * factor;
if (x*x + Y*Y < 1.0) II Within circle?
++hits;
II yes: hits += 1
}

double pi_approx = 4.0 * hits I n;

cout "Approximation of pi after "
n " tosses: " pi_approx
" (error="
fabs(M_PI - pi_approx)*100/M_PI
"%)\n";
}

return 0;
}

1r

and coprimality

The probability of two integer numbers chosen at random being relatively prime to each other is 6/1f2. This characteristic can be used for
another 1f approximation procedure, whereby a sufficiently large number of pairs of random numbers is calculated and these are examined
for coprimality. From the ratio of the number of coprime pairs t to the
total number of trials n the following approximation for 1f is obtained:
lim

n--+oo

!n

1f~

(3.4)

fP

(3.5)

1f2

Ball [14] reports a field experiment in which 50 students were asked

to each write down five pairs of random numbers. It turned out that
154 of them were coprime. This led to 6/1f 2 = 154/250, and hence to
1f = 3.12.

42

3. Shortcuts to

11"

Instead of playing this game with students, one should try it with
a computer. The only problem with such a program is how to investigate the results for coprimality. Fortunately, for 2300 years there
has been an algorithm which does this and could be described as the
"grandfather" of all algorithms. It was first noted down by the Greek
mathematician Euclid (ca. 325-265 BC) and appears in Book VII of
his 13-volume Elements. This work is the most successful mathematics
book ever written and has been published more than 1000 times since
antiquity.
The Euclidean algorithm can be written down in only three lines
of source code (e.g. see the program listing below), although it takes
longer to explain in words. Starting from two integer numbers A > 0
and B > 0, it calculates the largest integer which divides into both A
and B without a remainder. If this "highest common factor" (HCF)
= 1, then A and B are coprime to each other. The algorithm utilises
the property that A = l~JB + C, with 0 ::; C < B. For example, let
us consider the integers A = 78 and B = 21. As 78 = 321 + 15, the
highest number which divides into both 78 and 21 is also the largest
number which divides into 21 and 15. Through iteration one obtains:
78 = 321 + 15
21 = 1 . 15 + 6
15

= 26 + 3

6=23+0
Hence 3 is the highest common factor of 78 and 21.
In the following little program we simulate the student experiment
and so arrive at an approximation for 7[,
II
II
II
II
II

Approximation for pi by a Monte-Carlo-Method:

The program finds pairs of random integers. It determines
whether they are relatively prime. The theoretical probability
of these events is 6 I pi-2.

#include <iostream.h>
#include <stdlib.h>
#include <math.h>
int euclid(int u, int v)
{

int rj
while ( (r = u % v) != 0
{

3.4
u

7r

and chance (Monte Carlo methods)

43

= v;
= r;

v
}

return v;
}

int main(void)
{

while (1)
{

int

n, nTries, nHits;

cout "Enter no of tries (or 0 to exit): ";

cin n;
if (n <= 0) break;
srandint)clock());
for (nTries=nHits=O; nTries < n; ++nTries)
{

int A = rand() + 1;
int C = rand() + 1;
if (euclid(A, C) == 1) II A and C are relative prim
++nHits;
}

double f = nHits * 1.0 I nTries;

double pi = sqrt(6.0 I f);
cout "After " nTries
" tries is pi " pi endl;
}

return 0;
}

The following approximations have been obtained through running the

programs a number of times:
Number of trials
10
100
1,000
10,000
100,000
1,000,000

Approximation for
3.464102
3.273268
3.194383
3.142438
3.143913
3.141554

1f

Error
+0.322509 ...
+0.131675 ...
+0.052790 ...
+0.000845 ...
+0.002320 ...
-0.000038 ...

Although Monte Carlo methods are attractive and interesting and

they are easy to implement, they are not very well suited to calculations of 1f, as they converge poorly. Even after a million attempts it is
scarcely feasible to obtain more than four correct decimal places of the
1f. And it is perfectly possible to obtain a poorer approximation of 1f
on a larger number of trials than on a smaller, as can be seen from the
table above. Methods which are based on random numbers are probably the least efficient of all systematic procedures for the calculation
of 1f.

44

3. Shortcuts to

7r

On top of this there are two further fundamental problems. Firstly,

computer programs are not capable of generating truly random numbers as there is inevitably a point at which they reproduce numbers they have already generated, in the same sequence. Secondly,
the "pseudo-" random numbers generated by a computer inevitably
lie between certain limits, so that on this basis alone they are not
truly random. It follows from this that Monte Carlo methods do not
strictly speaking converge on Jr but, rather, oscillate around Jr. However, given the poor convergence obtained with these methods, these
are really only esoteric thoughts.

3.5 Memorabilia
Another way of arriving at decimal places of Jr is to learn them by
heart.
A number of mnemonic verses are available to assist with this. We
mentioned already such a verse in the first chapter. Many people gifted
at playing with words have tried to conjure up texts in which the number of letters in each successive word corresponds to the corresponding
digit of Jr.
An English poem intended to teach Jr to 31 decimal places was
submitted by a person "F. S. R." to the scientific journal Nature in
1905:
Sir,
3

send

In

sacred

truth

rhyme

excelling

and

rigid

spelling

Numerical

sprites

elucidate

For

me

the

lecture's

dull

weight

If Nature gain not you complain

2
6
4
3
3
8
Tho Dr Johnson fulminate.
3
2
7
9

Probably easier to memorize is the following text, written by the

English astrophysicist Sir James Jeans (1877-1946), which produces
24 digits:

3.5 Memorabilia

45

How I want a drink, alcoholic of course, after the heavy

chapters involving quantum mechanics. All of thy geometry,
Herr Planck, is fairly hard ...
Such texts are available in many languages. The Internet address
http://users . hol. gr rxpolakis/PiPhilology . htm mentions more
than 60 poems in 24 languages, among them "14 + X" in English and
"4+X + V" in German. (Some of them seem harder to remember than
the native digits of 7r.)
Soon after the Frenchman Fautet de Lagny (1660-1734) succeeded
in 1717 in calculating 7r to 127 decimal places, P. Decerf tried his
hand at turning this into a work of art and composed a monumental
7r poem 127 words long. Some years later it was discovered that de
Lagny's "7r" contained a transcription error, so that the 112th decimal
digit should have been an 8 rather than a 7. It was then necessary
to change Decerf's poem and publish an "update", something which
rarely happens in the world of poetry [41, p. 153].
Perhaps the longest mnemonic poem in the world was composed by
Michael Keith. He has written a "Cadaeic cadenza" which produces
an awe-inspiring 3865 decimal places of 7r 1 . Cadae is a made-up word
consisting of the letters which correspond to the first five 7r digits
31415.
The first section of the cadenza is a modified version of Edgar Allen
Poe's famous poem of 1844 The Raven. The new work was given the
ambiguous title of Near A Raven. It consists of 18 verses and begins
as follows:

http://users.aol.com/s6sj7gt/cadenza.htm

46

3. Shortcuts to

7r

Poe, E.
Near A Raven
Midnights so dreary, tired and weary.
Silently pondering volumes extolling all by-now obsolete lore.
During my rather long nap - the weirdest tap!
An ominous vibrating sound disturbing my chamber's antedoor.
"This", I whispered quietly, "I ignore" .

3.1
415
926535
897932384
62643383
27950288
419716

Perfectly, the intellect remembers: the ghostly fires, a glittering ember 9399375105
Inflamed by lightning's outbursts,
8209
windows cast penumbras upon this floor. 749445
Sorrowful, as one mistreated, unhappy thoughts I heeded:
92307816
That inimitable lesson in elegance - Lenore 406286
Is delighting, exciting ... nevermore.
2089
The title and the first two verses alone produce 80 digits of 11". These
are followed by a further 16 such verses plus the author's signature for
the other 660 places. Keith worked hard to remain as close to the
original as possible as regards the metre, action, melody and rhythm.
Thus for example he succeeded in retaining the melodious dark refrain
at the end of every verse.
Of the 14 sections of the Cadaeic cadenza, section 3 contains another poem by the mathematician and 11" enthusiast Charles Lutwidge
Dodgson, better known as Lewis Carroll, author of "Alice in Wonderland", which has likewise been adapted to 11", while section 11 even
contains a work by William Shakespeare (who we all know wrote the
famous words, "to 11" or not to 11"").
11" texts reach their first critical point at the 32nd decimal place,
when the first zero appears in the decimal sequence of 11". As there are
no words with zero letters, the authors of mnemonic verses have to
think up something special. Keith's solution to the problem is to use
words of 10 letters, such as, for example, "disturbing". Other authors
require their readers to interpret a punctuation mark as a O. A particularly difficult problem here is the 601st decimal place of 11", where a
sequence of three zeros occurs in succession.
Short after the 740 decimal places of "Near The Raven", in section 2
of his work, Keith had to resolve another tricky decimal sequence in 11",
the so-called Feynman point, for at position 762 there is a devilish
sequence of 7 numbers 9999998! This sequence may be beautiful to

3.6 Bit for bit

47

7r aesthetes, but for poor Mike, who had to find a text containing a
sequence of six 9-letter words and one 8-letter word in succession, it
was bad news. Nevertheless, he succeeded in cracking it.
There are a number of methods of memory training which have
evolved from the example of 7r. These include: the use of recurring
patterns, creation of blocks of fixed or variable length, association of
graphic, rhythmic, musical or colour representations, learning by repetition at varying intervals and situations. There are articles available
on this subject on the Internet.
Anyone who knows a lot of digits of 7r will go far. For example
such a person can take part in a 7r recitation competition or become
a member of a 7r club, as these normally require one to be able to
recite a number of digits of 7r by way of entry test. We have already
mentioned the Austrian club "Freunde der Zahl Pi" (Friends of the
number Pi) which requires of candidates seeking admission that they
should be able to recite 100 decimal places in a public place in front
of a 7r notary. The Swedish "1000 Club,,2 even requires 1000 digits of
7r.
The members of these clubs are not just nut-cases who fill their
heads with digits of 7r, but serious scientists seem to take pleasure in
being able to perform this feat. One of these is Simon Plouffe, who
was one of the three mathematicians who recently developed a novel
7r algorithm, the BBP algorithm (see page 117). Plouffe obtained an
entry in the (French) Guinness Book of Records in 1977 by memorising
4096 digits of 7r. In fact, he said, he actually knew 4400 digits by heart,
but 4096, i.e. 212, was a nicer number.
Since 1995 the world record holder in the recitation of 7r digits has
been Hiryuku Goto, then aged 21, who recited 42,000 digits in nine
hours. Since this new record was established it has been claimed that
Japanese is better suited for memorising sequences of numbers than
other languages. Or perhaps Goto simply had nothing more sensible
to do after programmers had eliminated him from their code ...

3.6 Bit for bit

The inverse value of 7r, 1/7r, may be represented either as = 0.3183 ...
or = 0.010100010111110 .... This is no contradiction, as the second
2

http://www.ts.umu.se/-olleg/pi/club_l000.html

48

3. Shortcuts to

'IT

value is just the binary representation, in which only zeros and ones
occur.
To obtain the places after the decimal point in the binary representation of I/K from the decimal version, keep multiplying the decimal fraction by 2 and writing each calculated position in front of
the point as the next binary digit while keeping the fraction for the
next multiplication. Thus 0.3183 x 2 = 0.6366, hence 0 is the first
binary digit. 0.6366 x 2 = 1.2732, so the second binary digit is 1. Continuing with the post-decimal point result of the last operation, i.e.
with 0.2732 x 2 = 0.5464, the third binary digit is another O. And
so the procedure continues. Of course the precision of binary representation is not any higher than that of the decimal representation,
so that from n accurate decimal places one can obtain at the most
1/ 10glO 2 ~ 3.3 times as many accurate binary digits.
A quite different method for calculating the binary representation
of I/K was discovered by Simon Plouffe of Bordeaux University. His
procedure was proved and generalised in 1995 by Jonathan Borwein
and Roland Girgensohn [36].
In Plouffe's bit recursion method the starting point is ao = tan(l) =
1.5574 .... From this, succeeding values of at, a2, a3, .. . are calculated
using the following rule:
2ak
ak+l = -1--ak2

(3.6)

Each ak is only checked to see whether it is negative or not. If it is

negative, then the next binary digit is a 1, otherwise it is a O.
The first 10 ak and the resulting binary decimal places run as follows:
k

0
1
2
3
4
5
6
7
8
9

Binary position of I/K

+1.5574
0
-2.1850
1
+1.1578
0
-6.7997
1
+0.3006
0
+0.6610
0
+2.3478
0
-1.0406
1
+25.111
0
-0.0797
1
ak

3.7 Refinements

49

Interesting and unexpected as this algorithm is, it conceals a trap.

Namely, before you can calculate 1/7f in this way, you first need the
initial value tan( 1), but this takes longer to calculate than the direct
calculation of 1/7f.

3.7 Refinements
You might already have a few 7f digits and be wondering whether you
could build on them. You can. An elegant method for doing this was
developed by Daniel Shanks 3 [108].
If Po is an approximation for 7f which has n accurate digits, then
the following calculation step produces an improved approximation PI,
which is accurate to three times as many digits:
(3.7)

PI = Po + sin Po

Example: with Po = 3.14, i.e. n = 3, PI = Po + sinpo produces the

new approximation of PI = 3.14159265 .... The new value is accurate
to 9 decimal places.
The derivation is simple if one uses the series formula for sin E and
remembers that sin(7f + x) = - sinx. In the following, let E be the
error of the previous value Po:
PO=7f+E

PI = Po + sin Po
= 7f + E + sin( 7f +

= 7f +
= 7f

E-

+E -

E)

sinE
(:, -

~~ + ~~ - .. -)

E3

=7f+

6 -'"

With n accurate decimal places in the initial value, Po comes to

E 3 /6 <
0.2 . 1O-3n, i.e. at least 3n decimal places.
Shanks shows that it is possible to achieve still more accurate approximations in one step. Thus,

< lO-n. Hence, the accuracy of the new approximation is

In 1961 Shanks together with Wrench passed the 100,000 mark in a

tion [106].

7r

calcula-

50

3. Shortcuts to

P2 =Po+
=

2 sinpo - tanpo
3

7r

IT

(3.8)

+ 20 + ...

(3.9)

improves the approximation Po of 7r by more than five times as many

accurate places. Shanks may have learned this trick from Wildebrod
Snell (1581-1626) (see page 183).

3.8 The

1f'

room in Paris

The only museum in the world to possess a 7r room is the Palais de

la Decouverte in Paris, Av. Franklin Roosevelt. You will find that the
museum does not have a house number as it is very big and forms part
of the Grand Palais building complex.
At the entrance we follow the sign salle pi and a really good article from the Mathematical Intelligencer [67]. They will both take us
to room 31 [sic!], which is round. There we can learn a few (mainly
French) facts about 7r and marvel at the first 707 digits of 7r which are
displayed in three spiral revolutions around our heads.
These 707 decimal places were calculated in 1874 by William
Shanks (see page 195). Unfortunately he made a mistake at the 528th
position which remained undetected for some time. Thus, at the international exhibition "Art and technology in modern life" in 1937, it
was in the 7r room of all places that an incorrect 7r was displayed [48,
p.105].
When the error was discovered by Ferguson in 1945, the museum
corrected its sequence of numbers immediately.
But without success. The rumour that there is an error in the 7r
on show there persists even to this day. Thus the 1997 edition of the
French encyclopaedia Quid contains a photograph of the 7r room as it
appears today, i.e. with the correct 7r, but with the caption,
"In the dome of the Palais de la Decouverte in Paris the first 627
[sic] decimal places of the transcendental number 7r are displayed, but
apparently some of them are incorrect."
Some people just cannot accept the truth!

4. Approximations for
Fractions

7r

and Continued

4.1 Rational approximations

The shortest approximation for 1f is simply 3. This is 4.5% away from
the true value of 1f and in fact this approximation occurs twice in the
Bible (see page 169). The longest approximation is 206.1 billion digits
long and is still not entirely accurate. Only when we write "1f" for 1fl,
are we entirely accurate - everything else is longer and either more or
less of an approximation.
For the purpose of simplification in this chapter we shall use 1f( s) to
designate a 1f approximation to s correct decimal places. The approximation 3 used in the Bible we will thus designate as 1f(0). In Babylon
the approximation 1f(1) = 3k = 3.125 was used more than 1000 years
earlier.
The statement that 1f( s) is a 1f approximation to s correct decimal
places means that it is less than 10- 8 out from the true value, j1f1f( s) I < 10- 8 On the other hand, the converse conclusion does not
hold true: an absolute error of < 10- 8 does not always also guarantee
s correct decimal places. For example, the good approximation 3.1416
differs from 1f = 3.14159265 ... by only 0.000007 ... , which is less than
10- 5 , but it is still only accurate to three decimal places. However, this
particular problem can only occur when the approximation is greater
than 1f.
In the old days before infinite decimal fractions were known and
understood, it was suggested that one might approximate a "ratio"
such as the ratio of the circumference of a circle to its diameter using
a ratio, i.e. using a fraction consisting of a numerator and a denominator. The classic 1f approximation of this kind is that developed by
Archimedes around 250 Be:
1

or one of the many mathematical identities (see the collection which starts on
page 223).

52

4. Approximations for

7r(2)

223
22
71 < 7r < 7

7r

and Continued Fractions

= 7r(2)

(4.1)

The fact that this approximation states an interval makes it particularly attractive, although the manner in which it was discovered has
its drawbacks. (For more on this, see page 171.) The notation used to
specify the interval makes clear that 7r does not exactly equal either
the left-hand or the right-hand limit, whereas representations of 7r have
not always been so honest. Thus, for example, Tsu Chhung-Chih, who
discovered the excellent approximation
7r(6)

355
113

(4.2)

in China c. 480 AD, was convinced that this fraction was exactly equal
to 7r.
An approximation is called "rational" if it can be represented by
a fraction consisting of integer numerators and denominators. Such
representations are extremely common, and all decimal, hexadecimal,
binary etc. digit sequences of 7r are rational approximations since, for
example, 3.14 is the fraction created by dividing 314 by 100. Because
there is a convention that with certain bases such as 10, 16 or 2 we only
"think" the denominator and do not write it, these approximations
are concise and effective. In actual fact, however, the approximation
7r(2) = 3.14 means not just 4 keystrokes consisting of a decimal point
and 3 numerals, but 7 keystrokes consisting of the fraction line and
two sets of 3 numerals in each of the numerator and denominator. But
if one is going to use the 7-keystroke version, there are other fractions,
e.g. the above-mentioned fraction 355/113 = 7r(6), which represent 7r
more accurately, in this case 4 decimal places more accurately.
A good rational approximation for 7r is one in which the ratio of
the number of correct decimal places to the number of digits in the numerator and denominator is particularly high. As we shall see shortly,
the best rational approximations in this respect attain the value 1.
The best source for good approximations of a (transcendental or
irrational) number is its so-called continued fraction. Under this remarkable, but very interesting representation, the denominators are
not, as in the decimal representation, powers of a fixed base number
but a combination consisting of an integer and a fraction.
In the case of 7r, the continued fraction begins as follows:

4.1 Rational approximations

53

7r

= 3 + ---------:;-1----

(4.3)

7 + ---------,,1--15 + -----=-1-1+----.-292 +
1
1 + ...

Alternatively, a tabular representation of its elements looks like this 2 :

bi

.0

.1

.2

.3

.4

.5

.6

.7

.8

.9

O.

3
3
1
2

7
1
84

15
14
2
6

292
1
1
1

1
1
15
2

1
2
3
2

1
2
13

1
2
1
99

2
2
1
3

1
2
4
5

1.

2.
3.

In order to arrive at a rational approximation for 7r all one has to

do is terminate one's simple continued fraction at any point and work
it out. (Incidentally, one never needs to do any cancelling of fractions
to their lowest terms.) For example, after terminating the continued
fraction after the 3, 7, 15 and 1, one obtains the approximations 3,
2:(" f~~ and f~~, which are accurate to 0, 2, 4 and 6 decimal places
respectively.
Of the approximations obtained in this way, those obtained by terminating the continued fraction directly before one of the large elements are especially accurate. When one examines the relative error
in relation to the terminal element, one obtains a curve like this:

The first 2000 elements will be found on page 244f.

54

4. Approximations for

7r

and Continued Fractions

10- 5
10- 10
10- 15
10- 20
ReI. error
10- 25
10- 30
10- 35
10-

40

Termination before continued fraction element:

715292

84

14

15 13

99

It can be seen from the graph that if termination occurs before a large
element such as 15, 292, 84 or 99, the relative error drops off especially
sharply, so that excellent approximations are to be found at these
points:
Termination point
Before

b2=15

Before
Before

b4=292

Before
Before

b12=14
b21 =84

Before

b25=15
b27 =13

Before

b33=99

Approximation
7f(2) =

2:;-

7f(6) = fr~
7f(12) - 5419351
- 1725033
7f(21) - 21053343141
6701487259
(25)
8958937768937
7f
= 2851718461558
(29)
7f
= 428224593349304
136308121570117
7f(37) - 2646693125139304345
-

842468587426513207

In favourable cases, the number of accurate decimal places is almost equal to the sum of the number of digits in the numerator and
denominator. If one terminates the continued fraction before b4 = 292,
b21 = 84 or b33 = 99 (the biggest elements), the number of accurate
decimal places is exactly equal to the number of digits. There are
more such cases further on, e.g. before b77 = 16, b79 = 161 or b80 = 45.

4.2 Other approximations

55

On the other hand, termination before b5 = 1 produces the value

3
13033ti2 = 71'(9), i.e. 2 decimal places fewer than the number of digits in
the numerator and denominator.
If An and Bn are, respectively, the numerator and denominator of
a continued fraction for a number /'1, which is terminated before the
element bn +1, then the relative error in the approximation is

An
1
I( -B - /'1,)1//'1, < A . B . b
n

n+l

(4.4)

This simple relationship was brought to our attention by F.L. Bauer,

and it verifies most effectively the statement that it "pays" especially
well to terminate the continued fraction before a large element, since
this large element then goes in the denominator of the error estimate,
where it has the effect of significantly reducing the value.
No better, i.e. shorter, approximations will be found for the stated
numbers of accurate decimal places. This can be proved from the theory of continued fractions. For example, this remarkable rational approximation discovered by Johann Heinrich Lambert (1728-1777)
71'(25)

1019514486099146
324521540032945

(4.5)

requires an extra 5 keystrokes compared with the equally accurate

continued fraction approximation stated in the 5th row of the above
table.
The table includes several approximations for 71' which have made
history. The origins of 22/7 (Archimedes) and 355/113 (Tsu ChhungChih) have already been mentioned. The two approximations formed
on termination before the elements 14 and 13 were discovered in Japan
in 1766. What is of especial note here is that at that time continued
fractions were still unknown, so that these discoveries must have been
made by a different route. How this happened is apparently not known,
and hence our wonderment is all the greater.

4.2 Other approximations

However accurate rational approximations may be, they neither look
pretty nor are easily remembered. Those qualities are achieved in other,
visually pleasing, "artistic" approximations with impressive symbols or
fine symmetries.

56

4. Approximations for

7r

and Continued Fractions

The Greek philosopher Plato (427-348 BC) is said to have known

this approximation [51, p. 126]:
7r(2) =

J2 + v'3

(4.6)

and his fellow philosopher in India, Zhang Heng (78-139 AD), was the
first to work with the formula
7r(1) =

v10

(4.7)

The following approximation is attributed to the author of the "Divine Comedy", Dante Alighieri (1265-1321), who was highly educated
in mathematical matters:
7r(3)

v'2

= 3 + 10

(4.8)

The authors have been unable to discover who invented the following impressive approximation formula, but it may have been the
Indian astronomer Aryabhata, who was born in India in 476 AD.

7r(4) = 512

2-

2+

2+ ~ 2+ ~2+ V2+ V2+ J2+v'2

(4.9)

One might be tempted to think that the first minus sign was an error, but that is not the case. As one can easily check with elementary
geometry, the expression
v'2 is the perimeter of an octagon
which is inscribed in a circle of radius 1, i.e. with the circumference
27r. Every time the number of sides is doubled, the innermost v'2 is
replaced by the expression
+ v'2 and the factor in front of the outermost root is doubled. In this way, the above approximation specifies
half the perimeter of an inscribed regular 1024-sided polygon.
Due to the large number of 2s under the roots, the formula (4.9) resembles Franc;ois Viete's product of 1593 (1.6). In fact the two formulae
are related and converge in similar fashion. They both approach the
number 7r through inscribed 2n-sided polygons, in the case of Viete's
product through its area and in the case of the above formula through
its perimeter.

8J2 J2

Kochansky (1631-1700) discovered

4.2 Other approximations

J~O - v12

71'(4) =

57

(4.10)

while Carl Friedrich Gauss (1777-1855) was only 14 when he calculated [82, p. 8]:
71'(13)

22 . 2484 . 12983009
7 2485 12983008

(4.11)

The young Friedrich started from the known approximation 2:; and
then divided it into 3.14159265 .... He subtracted the quotient from 1
and converted the remainder to a fraction with the numerator 1. The
denominator is around 2485. He then proceeded in similar fashion with
22

'7 .

2484
2485'

Adrien-Marie Legendre (1752-1833) referred to this approximation

[75]:
71'(9)

= In -1
x

- 2x4

where

1
2

= -(24 - 1)/(24

+ 1)

(4.12)

Srinivasa Ramanujan (1887-1920) was a past master at discovering

approximations. You will find further information about this important
Indian mathematician who was greatly enamoured of 71' in Chapter 8.
In his article entitled Modular Equations and Approximations to 71' [96]
of 1914, he presented many approximation formulae. To introduce the
idea, the essay starts with the simpler approximations, beginning with
an elegant, symmetrical expression [96, p. 34-35]:

9ti-

(4.13)

19
16 J7

(4.14)

71'(3)

=-+

71'(3)

71'(3)

7 ( 1+5
=3

(4.15)

71'(6) _ 99 (
7
)
- 80 7 - 3\1'2

(4.16)

71'(9)

V3)

63 (17 + 15)5)
25
7 + 15)5

(4.17)

These formulae Ramanujan derived from modular equations. They

are followed in the same article by a "curious" approximation which
he discovered "empirically" [96, p. 35]

58

4. Approximations for

11"(8) =

7r

and Continued Fractions

92 + 19
(4.18)
22
Attempts have been made to find out what may have been concealed behind the word "empirically". The most likely explanation is
the following [25, p. 655]. Ramanujan was an expert in continued fractions, and so he probably noticed that the simple continued fraction
of 11"4 possesses a particularly striking term:
4
111111
11" = 97 + 2 + 2 +"3 + "1 + 16539 + "1 +...
(4.19)
This observation would have suggested to him that he should terminate
the continued fraction before the unusually large element 16539 and
use it to find 11"4 ~ 97 +
= 9 +
Ramanujan obtained another empirical approximation by improving the approximation 355/113. From this he found, "simply from creating the inverse value of 1 - (11311"/355)",
4

i2

11"

2 Ii; .

(14) = 355 (1 _ 0.0003)

113
3533

(4.20)

In 1913 Ramanujan published a one-page article [97] containing

a geometric construction for the approximation ~~~. The essay itself
contains nothing remarkable, but what is remarkable, however, is the
fact that this paper bore the title Squaring the Circle and neither in
the title nor in the text did the expression appear with any quotation
marks. Since Lindemann had proved in 1882 that it was impossible to
square the circle, a fact of which Ramanujan was undoubtedly aware,
no mathematician could really afford such a title any more. It is a
testimony of Ramanujan's mathematical self-assurance that he did so.
As far as Ramanujan's real field of 11" approximation was concerned,
the previous approximations were only finger exercises. In fact he came
up with some considerably more accurate formulae [96, p. 31]:

11"(15) =

10
~
In (V 10 + llv2 + V + 7v2)
V142
4
4

(4.21)

11"(18) =

V~~O In ((2V2 + v'iO)(3 + v'iO))

(4.22)

12 In [14(3 + V5)(2 + V2) ( (5 + 2v'iO)+

11"(22) = V310

4.2 Other approximations

J61 + 20v'lo) ]

59

(4.23)

And to crown it all,

4 In [(5+J29)3
7r(31) = \"'522
y'2
(5v29 + 11 V6) x

(/9+: V6 +j5+:V6)'j

(4.24)

These approximations too stem from modular equations and are a

proof of Ramanujan's supreme mastery of this kind of equation and
function.
More recently, the Borwein brothers have derived further approximations of the type Ramanujan developed [32, p. 194]:

7r(2)

= 3(3~ +

7)

(4.25)

_ 103v'I3 + 125
7r (5) 158

(4.26)

7r 7 _
66y'2
( ) - 33J29 - 148

(4.27)

7r(8) =

4
RO

v58

In(396)

7r 9 _
180 + 52y'3
( ) - 45v'93 + 39v'3I - 201 y'3 - 217
7r(9) =

~
In (J29 + 5)
v'58
y'2

(4.28)
(4.29)
(4.30)

Dario Castellanos is another 7r enthusiast, who in 1988 in an academic paper on "The Ubiquitous 7r" [41, p. 79] presented all kinds of
folklore and attractive approximations he had developed himself.
He started by transforming the Ramanujan approximation shown
at (4.18) into an even more attractive form:
7r(8) =

102 _ 2222

222

(4.31)

60

4. Approximations for

7r

and Continued Fractions

He then tried something new. While continuing Ramanujan's procedure to the fifth power, he discovered
7r (8)

77729
254

(4.32)

and was surprised that Ramanujan had not come across this himself. In
addition, Castellanos also discovered these further approximations [41,
p. 79~80, 83]:
7r(6)

473 + 203
303
- 1

(4.33)

7r(6) = 1.09999901 1.19999911 . 1.39999931 . 1.69999961

7r(7) = 2 +

1+

(~~~) 2

(4.35)

934 + 344 + 174 + 88) 1/4

7r(10) = ( 95 +
75 4
17003 + 82 3 _ 103 _ 93 - 63 - 33
7r(11) =
69 5
7r(13) = ( 100 -

21253

(4.34)

+ 2143 + 303 + 372) 1/4

82 5

(4.36)
(4.37)
(4.38)

Simon Plouffe, whom we have already encountered several times in

this book, was indefatigable in his search for approximations. He has
every reason to be pleased with this one:
7r(6)

(689) / In (689)
396
396

or with
7r(8) = In (5280) /

vV

(4.39)

(4.40)

This next approximation creates a link between 7r and the "Golden

ratio" <p = ~+ 1 :
7r(3) =

~<p2

(4.41 )

while e and 7r are linked as follows:

7r(2) = 9 - e
2

(4.42)
(4.43)

4.2 Other approximations

e(3) = V'20 + 1f
e(7) = \l1f 4

61

(4.44)

+ 1f5

(4.45)

These approximations can be transformed into even more impressive forms. Thus, the last approximation (4.45) becomes the attractive
and easy-to-remember expression
(4.46)

in which a remarkable 7 digits (403.4287) are identical on both the

left-and right-hand side.
Another link between 1f and e is made in the approximation formula
developed by James Stirling (1692-1770) for the calculation of n!, Le.
the product 1 . 2 . 3 ..... n:
n!

~ (~) n

)21fn

(4.47)

Stirling, 1730

Quite apart from the good quality of the approximation - the

relative error is less than 1% from n 2: 9 and less than 0.1% from
n 2: 84 - this formula once again gives us cause to wonder at the
apparent ubiquity of 1f. Here it occurs in a problem comprising only
integers.
In our view, Stirling's formula, due to its beauty, its usefulness and
its age, belongs in the "eternal" list of the greatest formulae which
features formulae such as Euler's theorem (1.10), mentioned previously.
Incidentally, the formula can be improved significantly through a
small correction resulting from the asymptotic expansion:
(4.48)

The relative error is now less than 1% from n 2: 3 and less than 0.1 %
from n 2: 9.
For the similar case of the product of all the odd numbers < 2n,
F .L. Bauer [17, p. 49] used the following approximation:
(2n -I)! = (2n -1) (2n - 3) ... 31

J(2r0T

~1f.(n+!)

(4.49)

This approximation follows from Bauer's good 1f formula (16.62),

which converges on 1f significantly better than the Wallis formula or
its derivative (16.61) even though the only difference between the two
is a tiny! in the denominator.

62

4. Approximations for

1t'

and Continued Fractions

The 1[' approximations which Daniel Shanks derived from investigating complex quartic number fields are an order of magnitude better.
His piece de resistance is the following astonishingly simple formula
which produces 1[' to 80 decimal places [107, p. 398]:
With
D :=

~(1071 + 184V34)

(4.50)

E :=

~(1533 + 266V34)

(4.51)

F:= 429 + 304V2

(4.52)

1
C := '2(627 + 442V2)

(4.53)

followed by

d = D+ VD2 -1

(4.54)

e = E+ VE2 -1

(4.55)

F+ VF2 -1

(4.56)

9 = G+ VC2-1

(4.57)

we obtain
6
1['(80) = J35Q2ln(2' d . e . f . g)

(4.58)

In his article, Shanks proves the existence of an even better approximation of this type, which produces as many as 109 correct decimal
places for 1['. But due to the time that this would take he did not
calculate them. But, as he puts it, "It could be done."
Of course, a large number of good approximations for 1[' can be
obtained from terminating infinite 1[' series at an appropriate point. For
example, the Ramanujan series (1.9) shown on page 13, which yields
8 correct decimal places per term, produces an 80-digit approximation
of 1[' if it is terminated after the 10th term.
This method might appear somewhat trivial. On the other hand, an
infinite series which almost produces 1[' and thus converges close to 1['
would not be at all trivial. However many terms one were to calculate
in such a series, their sum would never come to 1['.
In fact there are some such series and they are truly amazing. Of the
two series shown below, the first one produces 1[' to over 18,000 decimal
places and the second one to as many as 42 billion.

4.3 Youthful approximations

In 10
71'(18,000) = 1002

+00

n~oo

1
10(n/100)2 )

63

(4.59)

J. and P. Borwein, 1992 [34J

(4.60)
J. and P. Borwein, 1992 [34J

These series thus converge to numbers which - even theoretically do not equal 71' but which "happen to" agree with 71' in the first 18,000
or 42 billion decimal places, as the case may be.
The Borwein brothers derived these formulae from so-called modular identities, and in fact the path from there to the formulae is surprisingly short [34]. When one considers that they evolved by chance
rather than being derived from the underlying theory they are quite
remarkable and could only too easily deceive one into thinking that
here indeed was the answer to the question of life, the universe and
71'. The Borweins in fact regard the series as an example of "caveat
computat" .
Incidentally, these formulae can be turned into any number of better
approximations for 71' by replacing the values 100 or 10 10 in them with
larger powers of 10.
Unfortunately, the series are not suitable for real calculations of
71'. It is true that they can be easily transformed from the range
-00 ... +00 to the range O... +00, and the terms in the first series
are derived from simple decimal shifts. However, one needs the value
In 10 in the first formula, and this is even more time-consuming to
calculate than 71' itself, and in the second series one would need to
calculate around 30 billion places of eX.

4.3 Youthful approximations

In the Jugend forscht ("youth researches") competition held every year
in Germany, Sven Kabus took a first prize in 1998 in Schleswig-Holstein
with a 71' program that was based on the "Aryabhata" approximation
(4.9). Kabus [70] first of all ascertained the formula analogous to (4.9)
for the perimeter of circumscribed 8-, 16-, 32-, ... -sided polygons and
observed that (4.9) turned up in it, e.g. in the 16-sided polygon

64

4. Approximations for

7r

and Continued Fractions

2UE3

UU3

= -VT2=+=v'~=2=+=v'2=;;2

(4.61)

where UUn and UEn refer to the perimeters of circumscribed and inscribed polygons with 2n+1 sides. Kabus then made the numerically
interesting observation that the perimeters of the circumscribed polygons were always twice as far from the circumference of the circle as
those of the inscribed 2n-sided polygons. He proved

~UUn 11m
1

7r

2UEn

n-+oo 7r -

(4.62)

From the two observations, Kabus then derived the following iterative
algorithm for 7r:

Algorithm (4.63) Kabus

(alg 4.63)

Initialise:

ao

:=

J~ + ~v'2

Iterate: (n = 0,1, 2, ... K - 1)

an+l :=

J~ + ~an

Then:
_

PK

2K+2 .J2-2<iK+y'2-2aK+l
-3y'2+2aK+l

----+

7r

Each iteration improves the 7r approximation by 1.2 places, whereas,

for example, Archimedes's method only produced 0.6 places each time
the number of sides of the polygon was doubled. Adding various cunning improvements in the arithmetic and using a PASCAL program,
Kabus arrived in this way at 5000 decimal places of 7r.

4.4 On continued fractions

Continued fractions are mentioned a number of times in this book and
we would therefore like to say a few words about them at this point.
Continued fractions are the "lost sons" of mathematics classroom
teaching. They are regarded as too advanced for secondary schools
and too elementary for universities, as a result of which they generally
fall between the cracks of both syllabuses [18, p. 129]. Their historical roots go back to the 17th century mathematicians Cataldi, Wallis
and Huygens, and their theory traces its origins to Leonhard Euler,

4.4 On continued fractions

65

who presented them in 1748 in his Introductio in analysin infinitorum [52, pages 303 fl1. The classic textbook on this interesting topic
was written by Oskar Perron [90], and a more recent textbook is that
by C. D. Olds [87].
A continued fraction is a fraction whose numerator is an integer
and whose denominator is the sum of an integer and a fraction which
in turn possesses the same form.
(4.64)

(4.65)

This representation, which rapidly grows to occupy a large physical

area, is often replaced by
bo

all

a21

a31

+ ~ + ~ + jb; + ... +

ani
Ibn

(4.66)

A. Pringsheim, 1898

(4.67)
Regular (Perron) or simple (Olds) continued fractions [90][87] are
ones in which all the numerators ai
1. They can be written even
more concisely like this:
(4.68)

The elements ai and bi of a continued fraction are also known, respectively, as "partial numerator" and "partial denominator" .
Every real number can be represented uniquely by a simple continued fraction. A number is rational if its continued fraction is finite
and irrational if its continued fraction is infinite.
The second theorem immediately proves, for example, that all the
numbers contained in the following list are irrational because their
continued fractions are infinite. The proof of the irrationality of 7r
which Johann Heinrich Lambert wrote in 1766 is also based on this
theorem. Lambert showed that the continued fraction of arctan 1 =

66

7r

4. Approximations for

7r

and Continued Fractions

I 4 is infinite and therefore 7rI 4 and

hence

7r

must be irrational (see

page 192).
The transformation of a number from its decimal representation
into its simple continued fraction form, and vice versa, is quite simple.
It can be performed with any pocket calculator which has an inverse
operation l/x. The two algorithms look like this:
II cf[] : array with the elements of the continued fraction
procedure NumberToCf (number , n, cf[O .. n-l])
{

for k:=O to n-l

{

x
:= floor(number)
cf [k] := x
number:= 1/(number-x)
}

and the reverse procedure for calculating a number from the elements
of its simple continued fraction goes like this:
function CfToNumber(n, cf[O .. n-1])
{

number := cf[n-1]
for k:=n-2 to 0 step -1
{

number := 1/number + cf[k]

}

return number
}

The simple continued fractions of some prominent constants are

shown below3 :

v's+1
2

(Golden ratIO)

(4.69)

= 1.61803398874989484820 ...

= [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1, ...J

V2 = 1.41421356237309504880 ...
= [1,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2, ..

(4.70)

v'3 = 1.732050807568877 29352 ...

(4.71)

= [1,1,2,1,2,1,2, 1,2,1,2,1,2,1,2,1,2,1,2,1,2,1, ... J

---3

As this book is concerned almost exclusively with numbers that are transcendental and, as a very minimum, are irrational, we are naturally only interested
in infinite continued fractions.

4.4 On continued fractions

e = 2.71828182845904523536 . . .

67

(4.72)

= [2,1,2,1,1,4,1,1,6,1,1,8,1,1,10,1,1,12,1,1,14,1, ... J
Euler, 1737

e2 = 7.38905609893065022723 ...

(4.73)

= [7,2,1,1,3,18,5,1,1,6,30,8,1,1,9,42,11,1,1,12,54, ..

Stieltjes, c. 1890 [90J

vie = 1.64872127070012814684 ...

(4.74)

Sundman, 1895

=
7r

[1,1,1,1,5,1,1,9,1,1,13,1,1,17,1,1,21,1,1,25,1,1, ... J

= 3.14159265358979323846

...

(4.75)

= [3,7,15,1,292,1,1,1,2,1,3,1,14,2,1,1,2,2,2,2,1,84, ... J

Vii = 1.77245385090551602729 ...

=
7r

(4.76)

[1,1,3,2,1,1,6,1,28,13,1,1,2,18,1,1,1,83,1,4, ... J

= 22.45915771836104547343

...

(4.77)

= [22,2,5,1,1,1,1,1,3,2,1,1,3,9,15,25,1,1,5,4,1, ... J
~ = 1.25992104989487316476 ...

(4.78)

[1,3,1,5,1,1,4,1,1,8,1,14,1,10,2,1,4,12,2,3,2,1, ... J
n

, = n--+oo~
lim" -k

-In(n + 1)

(Euler's constant)

(4.79)

k=l

= 0.57721566490153286060 ...
= [0,1,1,2,1,2,1,4,3,13,5,1,1,8,1,2,4,1,1,40,1,11,3, ... J
Euler, 1734

On closer examination we find there is a recurring pattern in the

continued fractions of the first 6 constants, e.g. in vie the pattern
1, 1, 4n + 1, whereas this does not occur in the second five cases beginning with 7r.
We cannot conclude from this that the simple continued fractions
of many or even of the majority of common constants possess a regular
pattern. In reality, this applies only to a minority, namely to the socalled quadratic irrationalities, one example of which is the "Golden

68

4. Approximations for

7r

and Continued Fractions

ratio" ( = 1 + /5/2), also to a few well-known transcendental numbers, such as the constant e itself and individual algebraic expressions
containing e, e 2 or e 1 / q , where q is an integer.
Interestingly, in almost all irrational or transcendental numbers,
including 7f, the geometric mean of their first n continued fraction
elements converges towards a fixed limit and this limit is the same
in all such numbers. This astonishing discovery was made in 1935 by
Alexander Khintchine (1894-1959) [74]:
lim \/b 1 b2

n--+oo

..... bn

= 2.6854520010 ... = Ko

(4.80)

nr=l

The limit Ko is defined as

[1 + k(k~2) FOg2 k and became known
as the Khintchine constant. The number is difficult to calculate, and
only a modest 110,000 decimal places are available.
For the number 7f, the geometric mean of the first 17,001,303 elements of its simple continued fraction come to 2.686393 [12], which
is fairly close to the Khintchine constant. On the other hand, the geometric mean for the number e, the big exception, moves away as n increases, while in the case of another exception, the constant (Golden
ratio), whose continued fraction consists of pure ones, the geometric
mean does converge, but towards 1 rather than Ko.
If one looks at the table with the approximations for 7f on page 54
in this light, one could form the opinion that continued fractions produce a more compact representation of 7f than the decimal representation. For the table shows, for example, that 7f can be just as well
approximated with 3 continued fraction elements as with 6 decimal
digits. However, this is only the case at particular places, and these
are precisely the ones at which it is especially beneficial to terminate a
continued fraction. Further on in both representations, the advantage
of the continued fraction always disappears, and generally the simple
continued fraction of an irrational number requires just as many elements as its equally accurate decimal fraction possesses decimal places.
The quotient from the number of accurate decimal places of the approximation and the number of continued fraction elements converges
towards a fixed limit which is close to 1, as Alexander Khintchine once
again proved in 1935. The exact value 1.03064 ... = 7f2/(6 In 2 In 10)
was stated somewhat later, in 1937, by Paul Levy.

5. Arcus Tangens

5.1 John Machin's arctan formula

By 2000 Be the Babylonians and Egyptians had arrived at one decimal place approximations for 7f based on the measurement of circles.
Archimedes was the first mathematician to develop a geometry-based
approximation in 250 Be and his polygon method served as the basis
for 7f approximations for the next 2000 years until (by 1630) 39 decimal
places had been calculated.
The second half of the 17th century saw the development of infinitesimal analysis, which made it possible to formulate infinite expressions for 7f. This was an analytical method of calculation by means
of which 7f could be investigated to a much greater depth. Among the
infinite expressions for 7f, a sub-species of series which are based on
the arctan function emerged.
The arcus functions are the inverse of the trigonometric functions
(sin, cos, tan etc). As their name suggests, they constitute (circular)
arcs or sections of them. If x = tan y, then arctan x = y, if Y lies
between -7f /2 and +7f /2. For our purposes the special case of x = 1
is particularly interesting as the arctan function for this case results
in the value 7f / 4 (or 45 degrees of arc).
hence,

7f

4"

= arctan 1

(5.1 )

The arctan function possesses a series which is relatively easy to calculate and was originally discovered by James Gregory (1638?~1675).
For the area under the curve y =
in the interval [0, xl he found:

1;x2

arctan x = {X

io

1 +t
From this he derived the Gregory Series in 1671:

(5.2)

70

5. Arcus Tangens

x3
arctan x = x - -3

+ -x5 - -x7 + ...

(5.3)

Gregory, 1671

To obtain a formula for 7r with this series (5.3), one needs only to set
x = 1 because, as mentioned, arctan 1 = i. The resulting series is
called the Leibniz series:

111

7r

-=1--+---+--
4
3
5
7
9

(5.4)

Leibniz, 1674

No doubt due to its simplicity, "every schoolboy" knows the Leibniz series. However, it is not suitable for numerical calculations of 7r
because its terms get smaller only very slowly. If one terminates the
series after the nth term, the absolute error, i.e. the difference between
the series sum and the true value of 7r, is only ~ l/n. Therefore, for
example, even if one were to calculate 2 billion terms, one would only
obtain 9 accurate decimal places of 7r.
The poor convergence of the Leibniz series (5.4) is thus due to the
fact that it attempts to do too much in one go, so to speak. It determines 7r / 4 by calculating a single arctan, i.e. a single arc. If instead
one assembles 7r / 4 in a suitable fashion from relatively short circular
arcs, i.e. out of several arctan values, one arrives at formulae which
enable 7r to be calculated significantly more quickly.
The simplest of such compound arctan formulae dates back to Leonhard Euler (1707-1783), and goes like this:
7r

4' =

1
1
arctan 2 + arctan 3

(5.5)

Euler, 1738

The formula can be proven using the known trigonometric identity

tan (a

+ fJf.I)

tana + tan (3
1 - tan a tan (3

(5 6)
.

if one makes the substitutions tan a = 1/2, i.e. a = arctan 1/2 and
tan (3 = 1/3 i.e. (3 = arctan 1/3.
The following diagram provides a geometric interpretation of this
formula:

5.1 John Machin's arctan formula

71

B
arctan!
arctan

tan .!3

The sum of the arcs AC = arctan and C E = arctan! produces the

arc AE = arctan 1 = i. If one replaces each of the arctan expressions
in Euler's formula (5.5) with its Gregory series (5.3), one sees that its
terms becomes smaller much more rapidly than those in the Leibniz
series. For example the 100th term of the Leibniz series (16.59) (=
arctan 1) yields just two leading zeros after the decimal point, but the
100th term of arctan! already produces 62, while the 100th term of
arctan produces as many as 98 zeros after the decimal point. This
improvement in convergence offsets by far the disadvantage that two
series have to be calculated now instead of one.
An even better formula can be obtained from two circular arcs as
follows [76, p. 246]:
The number
11111
0: = arctan "5 = "5 - 3.53 + 5.55 - 7.5 7 + ...
(5.7)

is obtained from the Gregory series (5.3) if we set x = ~. So, tan 0: = ~.

Therefore,
2tan 0:
5
tan 20: = -------c~
and
(5.8)
2
1 - tan 0:
12
120
(5.9)
tan 40: = 119
It will be seen that 40: is only a little greater than
a second angle (3 with (3 = 40: - i, then

i. If we introduce

72

5. Arcus Tangens

tan4a
i_ _-_tan
_2
tanf3 = _
1 + tan4a tan i

1
239

(5.10)

f3 is very easy to calculate:

1

(5.11)

f3 = arctan 239
239 - 3 . 2393 + 5 . 2395 - ...
a and f3 together produce
7r

- = 4a -

f3

(5.12)

= 4 arctan 5" -

1
arctan 239

(5.13)

= 4 [ 5"1 -

+ 5 .155

(5.14)

1
3 . 53

]
[1
1
]
- . .. - 239 - 3 . 239 3 + ...

Machin, 1706

This is the so-called Machin formula. It bears the name of the person
who discovered it, John Machin (1680-1752). He used it in 1706 to
calculate 7r to 100 decimal places.
The successive terms of the series of the first summand in the
Machin formula decline by around 1/25. The series therefore converges
with log1025 = 1.39 ... decimal places per series term. Due to the 5 in
the denominator, it can also be calculated relatively easily using paper
and pencil. The series of the second summand is more troublesome due
to the "awkward" 239, but on the other hand it converges better, with
around 4.76 places (= loglO(239 2 )) per term.

5.2 Other arctan formulae

Several systematic methods for deriving arctan relations exist. One of
them is to use the formulas [41, p. 91]:
1 = arctan -1
arctan --b
aa

+ arctan

(a b
b )
-a +1
2

and

(5.15)

Dodgson (Lewis Carroll) ?

1 1 1
arctan - = 2 arctan -2 - arctan 4 3
(5.16)
a
a
a + 3a
with appropriate choices for a and b. Substitution of a = 2 and b = 1
produces the above-mentioned arctan formula of Euler (5.5), while
this formula plus substitution a = 3 and b = 1 results in the arctan
formula (5.17) of Charles Hutton (1737-1823) shown below. Using this
and other methods, over the years various further arctan formulae have

5.2 Other arctan formulae

73

been discovered for the result 7r / 4, especially ones in which more than
two arctan expressions occur. Here is a selection:
7r
1
1
"4 = 2 arctan 2 + arctan "7
(5.17)
Hutton, 1776, Performance Index=3.28

= 3 arctan

4: + arctan 20 + arctan 1985

(5.18)

Loney, 1893 [122], PI =2.73

= 6 arctan

8' + 2 arctan 57 + arctan 239

(5.19)

St0rmer, 1896, PI=2.1O

= 4 arctan

5' -

1
arctan 239

(5.20)

Machin, 1706, PI=1.85

1
1
1
- arctan - 4 arctan 10
239
515

(5.21)

= 8 arctan -

Klingenstierna c. 1730, PI=1.79 [114, p. 296]

= 12 arctan 18 + 8 arctan 57

(5.22)

- 5 arctan 239

Gauss [56, II, p. 524], PI=1.79

1 1 1
+ 2 arctan 443 - 5 arctan 1393 -

= 22 arctan 28

1
- 10 arctan 11018

(5.23)

Escott, PI=1.63

= 44 arctan 57

+ 7 arctan 239
1

+ 24 arctan 12943

1
- 12 arctan 682

+
(5.24)

St0rmer, 1896, PI=1.59

This list contains some of the arctan formulae whose originators are
known. There are many other such formulae, some of whose originators
are known and others not. On our CD-ROM you will find a substantial
collection of them in the ari th directory, including some "monsters"
with 11, 12 or even 13 terms, also an algorithm which can be used to
automatically discover still more.
Along with the names of the originators, the list provides a "performance index" (PI). The smaller this is, the better, as this means that
calculation of 7r requires less effort. The performance index is based on
calculation of the expression 1/ log a 1 + 1/ log a2 + ... + 1/ log an.
From the list it is clear that the Machin formula is one of the best.
It is also the best of the four possible [32, p. 345] arctan formulae which

74

5. Arcus Tangens

comprise only two terms. This explains why for over 250 years it was
the favourite formula of 7r digit hunters.
Nevertheless, other arctan formulae were also widely used for a long
time. For example, Leonhard Euler in 1755 used the following relation:
7r

"4 = 5 arctan "7 + 2 arctan 79

(5.25)

Euler, PI=1.89

Its performance index of 1.89 is worse than the formulae cited above.
But Euler discovered an elegant way to reduce the the amount of work.
He derived the identity

y(

)
2
2.4
2.4.6 3
arctan x = - 1 + -y + --y 2 +
y + ...
x
3
35
357
where y = x 2 /(1 + x 2 ).

(5.26)

Euler, 1755

and inserted the arguments of (5.25) into it. The first one, x = ~,
leads to y = l~O' while the second one, x =
produces y =
The powers of 10 which are now in these y's reduce their decimal
computation to a matter of simple decimal shifts. By this means Euler
was able to calculate 20 decimal places of 7r in less than an hour [32,
p.340].

t9

IM60'

In the previously mentioned arctan formula of Gauss (5.22), the

arguments 1/18 and 1/57 have a similar useful property:
1
( 1
arctan 18 = 18 325

24

+ 3 . 325 2 + 3 . 5 . 3253 + ...

and
1
(1
arctan 57 = 57 3250

24

(5.27)

(5.28)

+ 3.32502 + 3.32503 + ...

Once again the work is simplified through decimal shifts, as a result of

which the terms of the second series can be obtained from those of the
first series. For this reason, the Gaussian arctan formula (5.22) was
promoted to the "best formula for 7r calculations up to 1,000 decimal
places" in the pre-computer era [16].
A 7r calculation using arctan formulae is very easy to programme.
On our CD-ROM there is a quite elementary C program for doing this.
One of the authors (JA) devoted a considerable effort a few years
ago to the search for arctan formulae, in the course of which he discov-

5.2 Other arctan formulae

75

ered some beauties (see formulae (16.124) to (16.132) in our collection

of formulae beginning on page 233). He was looking for expressions
in which the first term had the biggest possible denominator. He was
especially pleased when his computer churned out the following arctan
formula consisting of 11 terms:
~
1
1
"4 = 36462 arctan 390112 + 135908 arctan 485298 +
1
1
+ 274509 arctan 683982 - 39581 arctan 1984933 +
1
1
+ 178477 arctan 2478328 - 114569 arctan 3449051 1
1
-146571 arctan 18975991 + 61914 arctan 22709274 1
1
- 6044 arctan 24208144 - 89431 arctan 201229582 1
- 43938 arctan 2189376182
(5.29)
Arndt [7], 1993

The following two arctan formulae deserve a special mention for a

quite different reason:
~
1
1
1
1
- = arctan - + arctan - + arctan - + arctan - + ...
4
2
5
13
34
00
1
(5.30)
= Larctan~
n=l

2n+l

and

- =

3V5-5 - ~
~F2narctan
2
n=l

(2)
3F 2n+2

+ F 2n+2

(5.31)

Arndt, 1994

These series have an infinite number of arctan summands and are

therefore no good for an effective ~ calculation. But their distinguishing feature is that they contain the famous Fibonacci numbers Fn. Fibonacci numbers are named after their discoverer, Fibonacci, whose real name was Leonardo of Pis a (1180-1240), and go
Fn = 1,1,2,3,5,8,13,21,34, .... Each succeeding number is the sum
of its two predecessors Fn+2 = Fn+1 + F n , n 2: 1.
Fibonacci numbers have various beautiful characteristics and occur
often in nature. In particular, they forge a bridge between mathematics

76

5. Arcus Tangens

and art because the ratio of successive Fibonacci numbers converges

towards the "Golden ratio" <p = ~ (J5 + 1) = 1.61803 .... This Golden
ratio <p has been regarded since antiquity as especially aesthetic in
sculptures, paintings and buildings and logically occurs also in Arndt's
formula (5.31). In this way Fibonacci numbers endow our 7r with an
artistic touch.

6. Spigot Algorithms

The spigot algorithm, developed by Stanley Rabinowitz and Stanley

Wagon [95], is a recent and elegant method of calculating 7r. The algorithm is ideal for running on a personal computer.
1. The spigot algorithm starts producing digits of 11" right from startup and thereafter churns them out at regular intervals. Under all
the other methods, 11" has to be calculated completely in a buffer
and is only output at the end, all at once. By contrast, with the
spigot algorithm the 11" digits trickle out one at a time. It is possible
to watch it at work, so it is well suited for online demonstrations
on the Internet, for example. The CD-ROM which goes with this
book contains the Java applet spigot/pispigot. htm, which shows
the algorithm in action. You only need to load it into a standard
J ava-capable browser.
2. The algorithm works with nice small integers; even for 15,000 digits
of 11" the values of its variables do not climb to more than 32 bits
(including plus and minus signs), so that the long C data type is
sufficient on standard 16-and 32-bit compilers. This means there
are no problems with rounding, deletion or truncation, which can
make life miserable with other algorithms.
3. Implementation of the spigot algorithm does not require any extraneous software, such as, for example, a high precision library.
Everything you need will be found in any standard C compiler.
4. The spigot algorithm is surprisingly fast. Although the time it
requires is of quadratic order, so that it cannot compare with
high-performance algorithms such as the Gauss AGM algorithm
(see page 91), nevertheless it regularly outdoes algorithms that are
based on arctan series (see page 69).
5. The mathematics behind the spigot algorithm is simple.

78

6. Spigot Algorithms

6. The spigot algorithm can be written in only a few lines of source

code. The shortest 7r programs are based on it. A proof of this has
already been provided above (see page 37).

6.1 The spigot algorithm in detail

The starting point is the following simple structured series for 7r:
1
2
3
7r = 2 + -(2 + -(2 + -(2 + ... )))
(6.1)
3
5
7
It can be derived from the Leibniz series (16.59) without great effort
if one just uses the Euler transformation [76, p. 255]. However, we will
not perform that step now.
The series (6.1) can be conceived of as a number in a number system
with a variable base. Normally we encounter only numbers with a fixed
base, which is usually base 10. Every place of such a number must be
multiplied with a value which is higher than the value of the next digit
to the right, by a constant factor known as the "base".
But occasionally we come across numbers in which this factor is
not constant, e.g. the numbers which correspond to the expression "2
weeks, 3 days, 4 hours and 5 minutes". Because the ratio of weeks
to days = 1 : 7, days to hours = 1 : 24 and hours to minutes = 1 :
60, when converting this number to a decimal number (with the unit
"weeks"), three different factors have to be considered, namely 1/7,
1/24 and 1/60. Hence, to answer the question how many weeks are in
the example, we have to calculate the following:

(6.2)
Now compare this expression with the right-hand side of the 7r series
(6.1). You will see the same arrangement of brackets and also the
different bases. On the other hand, in the 7r series all the numerals
are equal, i.e. = 2, whereas in the week example they are different,
i.e. = 2, 3, 4 and 5; the Jr series is also infinite, whereas the example
terminates after the fourth digit.
Back to Jr. The task which the spigot algorithm performs consists
simply of converting the Jr series (6.1) to our base 10 number system,
i.e. to this form:
111
1
Jr = 3.1415 ... = 3 + 10 (1 + 10 (4 + 10 (1 + 10 (5 + ... ))))
(6.3)

6.1 The spigot algorithm in detail

79

Such a task is known in arithmetic as radix conversion and functions - in the present case - as follows:
On every step one 7r decimal place is calculated. To do this, first of
all all the digits in the number to be converted are multiplied by 10
(the new base). Then, starting from the right, every decimal place is
divided by the previous base (2i + 1)/i which applies to this decimal
place. On every division, the remainder is retained and the integer
quotient is carried over to the next decimal place. The most recently
calculated carry-over is the new decimal place of 7r.
The question arises as to how many terms in the 7r series (6.1)
really have to be carried over in order to obtain n decimal places
of 7r, including the digit 3 before the decimal point. In their article,
Rabinowitz and Wagon specify the value LI0n/3j, where the notation
lxj in the customary manner refers to the largest integer ~ x, i.e. it
means l10/3j = 3 and L3j = 3. They even "prove" that this value is
"correct". Unfortunately, it is not correct, as they could have seen from
the values n = 1 and n = 32. We are taking the liberty of correcting
the mistake and take one more decimal place, i.e. l10n/3 + Ij decimal
places, in the (tested) supposition that we are then correct in every
case.
Before we can begin to programme, we must explain the single real
complication in the algorithm.
When the base of the 7r series (6.1) is changed, it is possible for
a digit pair = 10 to occur. It can thus happen that at some point
a 10 occurs at position p, i.e. 3.1415 ... (p)(10). The 1 in the 10 is
an unresolved carry-over and must be added to the previous digit i.e.
3.1415 ... (p + 1)0. It can even happen that in front of such a 10 one
or more 9s have been calculated, so that those places too must be
corrected. The 3.1415 ... (p)99 . .. 9(10) must then be transformed into
3.1415 ... (p + 1)00 ... 00.
This complication means that the 7r program cannot release the
digits it has calculated immediately, but must store them in a buffer
until the next digit or the next few digits has/have been calculated.
When a new digit arrives, one or more digits will still be held in the
buffer. The first of these will definitely be < 9 and the others, if there
are any, exactly = 9. In this way we have the following situation in the
buffer at the point when the new digit q arrives:
p99 ... 9 +--- q
"-v-"

80

6. Spigot Algorithms

There are now three possibilities:

1. q < 9: it turns out not to have been necessary to retain the value
in the buffer. p and any succeeding 99 ... 9 can be output as they
are. q becomes the next decimal digit of p.
2. q = 9: no decision has been made so the number of 9s held in the
buffer is increased by 1 as q is added to it.
3. q = 10: it turns out to have been necessary to store the intermediate values in the buffer because a 1 must now be added to the
sequence of digits held in the buffer. As a result, p is increased by
1, and all the 9s held temporarily become Os. All these decimal
places are now ready and can be output, but the 1 in the 10 is
stored temporarily, Le. p is set to O.
The complication thus arises because the 7r series (6.1) is not unique
in relation to the second positions. For example, can be represented
in two ways, namely by 0+HO+%(2+(3+' .))) and by 0+i(2). There
are series for 7r which are unique, but they are much more long-winded
to calculate than (6.1).

6.2 Sequence of operations

We have now assembled all the elements required to describe the sequence of operations in the spigot algorithm.
The spigot algorithm calculates the first n decimal places of 7r. It
works with a field a[O], a[I], ... , a[N] consisting of N+l = l(10n)/3J+l
integers, where lxJ signifies the largest integer ~ x. Moreover, two
variables p and q are used to record the first and the current provisional
decimal place, also a numerator nines for the number of temporary
nines.
Initialisation: set p = 0 and nines = O.
For i = 0, 1,2, ... , N, set ali] = 2.
Iteration: repeat until n decimal places have been output .
Multiply with the new base. Multiply each ali] by 10 .
Normalise. Beginning on the right, from i = N to i = 1, divide
ali] by (2i+ 1) to obtain a quotient q and a remainder r. Replace
ali] with r. Multiply q by i and add the result (the carry-over)
to the element ali - 1].

6.2 Sequence of operations

81

Calculation of the next provisional digit of Jr. The left-hand digit

a[O] is post-processed. It is divided by 10. The division remainder
replaces a[O], while the quotient q produces the next provisional
1igit of Jr .
Correct the old provisional digits. If q is neither 9 nor 10, then
the first provisional digit up to now, p, and the nines which
succeed it are confirmed and output. The new first provisional
digit p now becomes = q, and nines is set to = O.
If q = 9, then out of the provisional nines only the number nines
is incremented by 1; no digits are output.
If q = 10, the first provisional digit up to now, p, is increased
by 1 and output. The provisional nines become zeros; they are
likewise output. The new first provisional digit, p, becomes = 0
(that is the 1st digit of q). nines is reset to O.
A small improvement can be achieved if one initialises the first
provisional digit of p with a negative value and then intercepts this
value during output. This means that the output begins immediately
with 314 ... instead of initially with a zero.
The following C function spigot 0 is described in the sequence of
operations below.
1*

* function
*
void spigot(digits)
* Spigot program for pi
* 1 digit per loop

*1
#include <stdio.h>
#include <stdlib.h>
void spigot(int digits)
{

int
int

i, nines = 0;
q,
p = -1;
len = 10*digits+3+1;
*a;

int
int

a = malloc(len*sizeof(*a));
for (i=O; i < len; ++i)
a[i) = 2;
while (digits >= 0)

q = 0;
for(i=len; --i >= 1;
{

q += 10L * a[i);
a[i) = q % (i+i+1);

1*
1*
1*
1*

next prelim. digit

previous prelim. digit
len: One more than R+W
array pointer

1* Init a[] with 2's

*1

1* Compensate for the very first digit *1

1* q = carry + 10*a[i)
1* a[i) := q % (2i+1)

82

6. Spigot Algorithms
q /= (i+i+1);
q *= i;

/* carry := iJoor(q,2i+1)*i

q += 10L * a[O];
a[O] = q % 10;
q /= 10;

/*
/*
/*
/*

first digit
q := carry + 10 * a[O]
a[O] = q mod 10
q : next prelim digit

i f (q == 9)

++nines;
else
{

/* q == 9: increment no of 9's

1* q

!= 9: print prelim. digits

i f (p >= 0)

printf("%Olld", p + q/10);
/* p : prevo prel. digit
if (digits < nines)
/* adjust digits to print
nines = digits;
digits -= (nines+1);
while (--nines >= 0)
/* print 9's or O's
printf(q == 10? "0"
"9");
nines = 0;
p = (q == 10 ? 0 : q);
/* set previous prelim. digit

*/
*/
*/
*/

}
}

free(a);
return;
}

At the end of the much-cited article [95] by Rabinowitz and Wagon

there is a PASCAL program which evidently was not written by the authors themselves but by a student, to whom they express their thanks.
This program has also been typed out and tested by some readers of
this book. They had various problems with it, for example, on 16-bit
PASCAL compilers the program stops working from n > 262 because
an integer overflow occurs at that point, or else at n = 1 and n = 32
it prints an incorrect last digit because the series length is too short,
or else in many cases of n it outputs fewer than n digits because variable nines at the end of the program is not yet o. Our above program
attempts to avoid these weaknesses.

6.3 A faster variant

Two improvements can be made to the spigot algorithm which make
it considerably shorter and quicker to run.
First of all, instead of 7r, 10007r is calculated, and hence the series
123
10007r = 2000 + 3(2000 + 5(2000 + 7(2000 + ... )))
(6.4)
is used. The conversion is thus not performed in base 10 but in base
10,000, so that on every pass the program produces four decimal places
instead of only one.

6.3 A faster variant

83

This trick not only has the effect that the entire program becomes
4 times faster, but an even more important effect is that the "complication" mentioned above becomes a lot simpler. The faster variant
always waits only exactly 1 position (consisting of 4 digits) instead
of having a variable number of places. This is sufficient for the first
approx. 50,000 digits of 7f, as up to that point there is no case in which
more than one such 4-digit chain has to be placed in a buffer. This is
only necessary if 4 zeros occur one after the other in a position that
is divisible by 4, and the first time that this occurs in 7f is at digit
position 54,936.
The second improvement is really a textbook tip which, however,
was evidently forgotten in the original formulation of the spigot algorithm. Namely, with a radix conversion, after each result place, one can
shorten the remainder that still has to be converted, by the number of
bits of the result place. Hence after each calculation of one place consisting of 4 decimal digits, the length of the f [] field can be reduced
by L104/3 + 1J = 14 places. This improvement speeds up the program
by an additional factor of 2.
Here is a C program for the faster variant. It is an expanded version
of the mini-program shown on page 37.
/*
* Spigot program for pi to NDIGITS decimals
* 4 digits per loop
* Expanded version
* Thanks to Dik T. Winter and Achim Flammenkamp.
*/

#include <stdio.h>
#include <stdlib.h>
#define NDIGITS 15000
#define LEN
(NDIGITS/4+1) *14

/* max. digits to compute

/* nec. array length

long
long
long
long
long
long
long
long

1* array of 4 digit-decimals*1
1* nominator prevo base
*/

a[LEN];
b;
c = LEN;
d;
e = 0;
f = 10000;
g;
h = 0;

/* index
/* accumulator and carry
/* save prevo 4 digits
1* new base, 4 dec. digits
1* denom prevo base
1* init switch

*/
*/

*/
*/
*/

*1
*1

*/

int main(void)
{

for ( ; (b=c-=14) > 0;

{

for (; --b > 0; )

{

1* outer loop:4 digits/loop*/

/* inner loop: radix conv

*/

84

6. Spigot Algorithms

*=

if (h

b;

==

0)
d += 2000
else
d += a[b]
g=b+b-1;

a[b]
d

= d % g;

1= g;

1* acc *= nom. prev base

*1

f;

1* first outer loop

*1

f;

1* non-first outer loop

1* denom prevo base

*1
*1

1* save carry

*1

h = printf(I%04ld", e+d/f);
1* print prev 4 digits *1
d = e = d % f;
1* save current 4 digits *1
1* assure a small enough d *1
}

return 0;
}

On the basis of the NDIGITS definition, this program calculates exactly 15,000 decimal places of 1r. There is no reason why it has to
stop at this limit if one is not concerned about the portability of the
program and compliance with ANSI C. As most C compilers around
ignore an overflow in the evaluation of integer expressions which depend on plus and minus signs and there is sufficient "breathing space"
in the array length, this program can normally be expanded to calculate twice as many 7r positions (using #define NDIGITS 32500). This
number was still a world record 40 years ago. Even more decimal places
can be calculated with floating point variables and arithmetic.
But however large a data range you select, even with this variant of
the spigot algorithm, i.e. where each place consists of only 4 7r digits
and only one such place is held in the buffer, it is still not possible to
get past 7r digit position 54932. This can only be achieved by removing
at least one of these limiting conditions; the easiest one to remove is
the first.

6.4 Spigot algorithm for e

The spigot algorithm is clearly not limited to the calculation of 7r.
Let us consider the transcendental number e = 2.7182 .... Its series
expansion which is the equivalent of the above 7r series (6.1) goes as
follows:

111
e = 1 + -(1 + -(1 + -(1 + ... )))
(6.5)
1
2
3
Here, as in the 7r series, different bases occur, but this time they
are all of the type that their numerator = 1. This means that when

6.4 Spigot algorithm for e

85

calculating the carry-over, there is no need for a multiplication operation.

More important, however, is the fact that the e series (6.5) is unique,
so that the complication which occurs with 7r does not occur here. A
spigot program for the decimal places of e is therefore simpler.
Here is a 138-character long program for the calculation of e to
15,000 places in the style of the mini-7r program shown on page 37:
/* note: N=15000, LEN=87700 >= 1.4*N*log10(N), 84700=LEN-N/5 */
a[87700),b,c=87700,d,e=le4,f=le5,h;
main 0 {f or ( ; b=c-- , b>84700; h=printf (" %05d" ,e+d/f) ,e=d%=f)
for(;--b;d+=f*(h?a[b) :e),a[b)=d%b,d/=b);}

Back on page 36 we promised to tell you how Lievaart's "obfuscated" program works. Well, it too works with the spigot algorithm
for e.

7. Gauss and

7r

One of the fastest method of calculating 7r, if not the fastest of all, is
almost 200 years old. It was invented by the German mathematician
Carl Friedrich Gauss (1777~ 1855) around 1800. It was subsequently
forgotten and only unearthed 170 years later, when two researchers,
Eugene Salamin [100] and Richard Brent [37] independently rediscovered it at the same time and turned it into the basis for superfast 7r
calculations.
Gauss's calculation method has since formed the basis of a number
of other formulations, as one can tell from the different names by which
the method is known in the literature. Thus we find the "Brent-Salamin
iteration" or the "Gauss-Legendre method" and other variations. In
this book, we refer to the method as the Gauss AGM algorithm as it
bears Gauss's unmistakable hallmark and its distinctive feature is the
arithmetic geometric mean (AGM).

7.1 The

7r

AG M formula

Gauss's formula that was so important in the study of 7r is shown in its

original form on page 101. Using modern notation, it looks like this:
2AGM 2 (1, ~)
7r

= 1
",",00
2k 2
"2 - uk=l
Ck

(7.1)

Gauss, 1809, Salamin, Brent, 1976

The essential element in this formula is the function AGM(a, b),

which produces the so-called arithmetic-geometric mean (AGM) of
two numbers. The AGM is a combination of the arithmetic and the
geometric means. The second variable, the Ck in the sum in the denominator, is directly linked with the AGM.
In everyday life we come across several 'means'. The arithmetic
mean (a + b)/2 of two numbers a and b is a common tool and can

88

7. Gauss and

7r

be used, for example, to arrive at the average of two marks in a certificate of academic achievement. The geometric mean ..;a:b is also
well-known and is used to calculate the average of two or more variables which are linked by a multiplication operation, for example to
calculate the average of two interest rates. In both cases, the word
"mean" refers to a variable which lies "in the middle" between two
starting values.
The AGM is also the average of two variables a and b. However,
this average cannot be calculated in one step, but requires an infinite
number of steps, each of which approximates it more closely. The AGM
calculation rule is "iterative", in that each step builds on and improves
on the results of the previous step.

Algorithm (7.2) AGM rule

(alg 7.2)

Initialise:
ao := a
bo := b
Iterate: (k = 0,1,2, ... )
.- ak+bk

~
~

AGM(a,b)
AGM(a,b)
bk+ 1 := va:;;:1ik
Then converge ak and bk to the same limit AGM(a, b).
a k+l

.-

It can be seen that the ak+ 1 is calculated by taking the arithmetic mean
of its predecessors ak and bk, and the bk+l by taking their geometric
mean.
When Gauss first got interested in the AGM (at the age of 14) he
ascertained its numeric behaviour in a number of cases "manually".
One of four surviving numeric examples begins with the starting values
a = J2 and b = 1 and, applying the rule (alg 7.2), it proceeds as
follows: [56, III, p. 364]1.
Arithmetic mean

Geometric mean

a = 1.414213562373095048802 b = 1.000000000000000000000
al = 1.207106781186547524401 bl = 1.18920711500272106671 7
a2 = 1.198156948094634295559 b2 = 1.198123521493120122607
a3 = 1.198140234793877209083 b3 = 1.198140234677307205798
a4 = 1.198140234735592207441 b4 = 1.198140234735592207439
1

Accurate
places
0
0
4
9
19

Gauss demonstrated from such columns of numbers that, mathematical genius

through he was, he too did not shy away from laborious numeric calculations.
Mathematicians today still stand in awe of him for this reason.

7.1 The

7r

AGM formula

89

In the consecutive iteration steps, ak and bk move towards each

other, ak getting smaller and bk getting bigger. (They would still move
towards each other after crossing even if a had been smaller than b
at the beginning). The value to which they converge, i.e. their limit
values, is called the "AGM of a and b" and is written as AGM(a, b).
In the case of a = V2 and b = 1 the limit value is clearly 1.19814 ....

The AGM always lies between the geometric mean and the arithmetic mean of the starting values.

arithmetic mean It x
arithmetic-geometric mean AGM(l, x)
geometrc mean ..;x
1

x
The main point to emerge from this numeric example of Gauss's is
that ak and bk rapidly converge towards each other. The last column
shows the number of decimal places which are identical for both variables. After only four iterations AGM( V2, 1) is already accurate to 19
decimal places.
In each iteration, the number of correct decimal places approximately doubles. This quadratic convergence is the outstanding quality
of the AGM and is what makes it so interesting to mathematicians. In
our 7r formula (7.1) the speed at which 7r is calculated is determined by
the AGM function (in the numerator) effectively "bequeathing" this
quadratic convergence to it.
The denominator offormula (7.1) contains the variable Ck. This is a
useful auxiliary variable which crops up again and again in the context
of the AGM. The variables ak and bk are generated during the AGM
iterations. The auxiliary variable Ck is defined as half the difference
between them.

90

7. Gauss and

71'

(7.3)

Ck+1 := 2(ak - bk)

identical with

c~+1

a~+1 - b~+l

(ak+1 - ak)2

(7.4)

As a and b move towards each other, c obviously converges towards O.

Some of the basic features of the AGM are the following: With real arguments

ao and bo, 0 < bo ::; ao and the auxiliary variable Ck+l := (ak - bk)/2, the following
apply [32, p. 1-4]:

ak
bk
Ck

ak+l + Ck+l
ak+l - Ck+l
(ak - bk)/2 = cLd(4ak)

=
=
c~ = a~ - b~ = (ak - ak-d

ak+l - bk+l = (ak - bk)2 /(2(..;ak + ~)2)

M(a, b) := lim ak = lim bk
k...-+CXJ
k--+oo

o ::;

AGM(A . a, A . b)

= A' AGM(a, b)
a+b

r-;-

AGM(a,b)

= AGM(-2-,vab)

AGM(l b)
,
Also:

1 + b AGM(l 2Vb)
2
' 1+b

bk ::; bk+l ::; ak+l ::; ak

and:

AGM(ak, bk) = AGM(ak+l, bk+l) = AGM(a, b)

The AGM is a well-defined quadratically convergent iteration for starting values

ao := 1, bo := z, where re(z) > O.

= 2kak, Lk = 2kc;:' and

k
C-k = 2 bie, where aie, bk and cie are generated from the AGM algorithm (alg 7.2)
commencing with ao := ao, bo := Co and Co := bo; note the "exchange" of band c
The AGM is also well defined for negative indices k: a-k

in the last two expressions.

7.2 The Gauss AGM algorithm

From his research into the AGM and various other mathematical deliberations which we will go into below, Gauss arrived at his formula
for 7r which is so important today (7.1). It can be converted directly
into an algorithm for 7r.

7.2 The Gauss AGM algorithm

Algorithm (7.5) Gauss AGM

91

(alg 7.5)

Initialise:

ao

:= 1

bo := 1/V2
80 := 1/2
Iterate (k = 0,1,2, ... , K - 1)
t := ak
ak+1 := (ak + bk ) /2
bk+l := v'tbk
Ck+l := (ak+l - t)2
8k+1 := 8k - 2k+lck+l
Compute an approximation of 1f:
1fK

(aK+b K )2
2s K

The algorithm begins with initialisation of three variables a, band

8. In succeeding iterations the variables a and b are expanded further
according to the AGM rule (alg 7.2) into an approximation for the
term AGM(1, 1/V2) which occurs in the denominator of (7.1). At the
same time the sum 8 in the denominator is calculated using the abovementioned auxiliary variable ck+l = (ak+1 -ak)2. The temporary variable t = ak is necessary because the value ak is used twice more after
it has already been replaced by ak+l. The iteration runs either until a
predefined level of precision has been achieved in the results or else, as
here, until a specified number K of iterations has been performed. At
the end of the procedure, the approximation for 1f is calculated in a
single step outside the loop. Here, instead of the (unachievable) value
AGM(1, 1/V2), the next value aK+1 = (aK+bK)/2 is used. This is the
most accurate approximation which can be achieved in K iterations.

1f.

These few lines thus define one of the best algorithms for calculating
The first three iterations alone produce 19 accurate decimal places:

7. Gauss and

92

Iteration
1
2
3
4
5
6
7
8
9

7r

PK

3.14 ...
3.14159 26 ...
3.14159 26535 89793 238 ...
ditto
ditto
ditto
ditto
ditto
ditto

J\ccurate places
3
8
19
41
84
171
345
694
1392

Every extra iteration produces more than twice as many places as

its predecessor. With 10 iterations the number of accurate decimal
places increases lOOO-fold and only 36 iterations are required to calculate 7r to 206.1 billion decimal places. The absolute error is [32, p. 48]:
7r22K+4e-7r2K+l
PK

<

PK+1

7r -

7r -

- J\GM2(1,

1/V2)

(7r-PK?
2K+17r 2

or

(7.6)

The quadratic decrease in the absolute error (7r - PK) will be evident
from the second expression, and from this one can determine that PK
is always smaller than 7r.

7.3 Schonhage variant

The calculation operations with long variables within the iteration loop
are a decisive factor in determining the performance of the Gauss J\GM
algorithm, especially the long multiplication akbk, the calculation of
the square root of this, and also the squaring operation (ak+1 - ak)2.
(On the other hand the long additions are of less consequence, likewise
the multiplications by powers of 2, as these only involve decimal shifts
of the variables which in any case are represented as binary numbers.)
J\ny improvement in performance is of course highly desirable. The
following version, devised by J\rnold SchOnhage [105, p. 266], significantly speeds up the Gauss J\G M algorithm. This variation avoids the
long multiplication akbk required on every iteration.

7.3 SchOnhage variant

Algorithm (7.7) Gauss AGM, Schonhage variant

93

(alg 7.7)

Initialise:

ao

:=

Ao := 1
Bo := 0.5
So := 0.5
Iterate: (k = 0,1,2, ... , K - 1)
t := (Ak + B k )/4
bk := VIJk
ak+l := (ak + bk)/2
A k+1 := a~+l
Bk+l := 2(Ak+l - t)
Sk+l := Sk + 2k+1 (Bk+l - Ak+I)
Compute the approximation of 7r:
_ AK+BK
PK SK

Schonhage dispenses with the long multiplication of akbk by replacing that product with less time-consuming operations involving other
variables which occur: akbk = 2(a~+1 - i(a~ + b~)). (The squares of a
and b are stored in the variables A and B.) The resulting performance
is around 25% better. This is achieved without any implementation
tricks simply through rearrangement of the operations.
In this way Schonhage avoids another long square root calculation
during initialisation and, during the calculation of the approximation
for 7r, he avoids a long squaring operation.
If one takes advantage of the fact that the bk and Bk can share the
same memory store, then the Schonhage arrangement, like the original
form, requires only 5 long variables, namely a, A, b = B, t and s, thus
avoiding any storage penalty for the speed gained.
It is very easy to convert the Gauss AGM algorithm to a program,
using the Schonhage variation:
a =A=1
B = s = 0.5
II B and b can share storage
for k=1 to N
t = (A+B)/4
b = sqrt(B)
a = (a+b)/2
A = a-2
B = (A-t)*2
s = s + (B-A) * 2-k
end for
pi = (a+b)-2/(2*s)

94

7. Gauss and

7r

For the first few digits of 1f this program can be implemented without any problems in a standard programming language with normal
floating point arithmetic or, better still, with a computer algebra system. Calculation of world record numbers of decimal digits is not,
however, so easy and major implementation problems arise.

Mass storage problem. The long variables a, A, b=B, sand t have to

be sized from the beginning to match the length of the result; but when
one is dealing with hundreds of billions of 1f digits, the main memory
of even the biggest computers is not sufficient, so that the variables
have to be stored outside of main memory and introduced into it in
stages. To do this quickly, considerable processor-specific work must
be done on memory management.
High precision arithmetic. Computers possess only a (very) limited machine precision in their instructions, and many can only handle variables a maximum of 16 (decimal) places long. Hence "long"
operations such as addition or modification of long variables have to
be assembled from "short" operations. The procedure is theoretically
complex and its practical implementation is performance-critical. For
this reason the "high precision arithmetic" that is required is generally the most difficult and the most challenging part of a 1f program.
It usually constitutes the very reason for developing a program in the
first place. However, it is also possible to "purchase" this basic work
(including the memory management functions mentioned above) and
draw it from one of the many high precision libraries (such as our
hfloatlibrary), by integrating such a library into one's program in an
appropriate fashion.

7.4 History of a formula

The AGM
The Gaussian 1f formula (7.1) looks relatively harmless but there is
actually more to it than meets the eye. It is remarkable not just because
it is so effective but also for its mathematical background and the
history of how it came about.
The arithmetic-geometric mean (AGM) and its calculation rule are
less old than one might suppose from its simplicity. Both were discovered some 200 years ago. Surprisingly, the calculation rule for the

7.4 History of a formula

95

AGM is actually older than the AGM itself [46]. The French mathematician Joseph-Louis Lagrange (1736~ 1813) was the first to use the
rule (alg 7.2) in 1785 for the calculation of approximations for elliptic
integrals, although he did not discover the AGM or its relationship
with elliptic integrals himself. It was Gauss who succeeded in making
these discoveries. As David Cox writes, "... we have the amusing situation of Gauss, who anticipated so much in Abel, Jacobi and others,
himself anticipated by Lagrange." [45, p. 315].
It was in 1791, at the tender age of 14, that Gauss discovered the
AGM. His discovery exercised such a fascination on him that during
the next 10 years he worked almost without interruption on developing
an AGM theory and took this work to heights which have not been
surpassed since [57, p. 186].
Then when he was 22 or 23, Gauss wrote an essay (in Latin) on his
discoveries relating to the AGM: de origine proprietatibusque generalibus numerorum arithmetico-geometricorum (On the origin and general characteristics of arithmetic-geometric means) [56, III,pp. 361 ~
374]. However, he never had the essay published and it was only printed
in 1866 as part of his estate. Gauss's lifetime publications included
only one piece on the AGM, and that was in 1818 in his treatise Determinatio attractionis (On the attraction of the elliptic ring) [56, III,
p. 352~353], which contains the third proof of the fundamental identity
(7.13). Additional information is provided in his formal writings.
Very little is known about Gauss's teenage research on the AGM.
However, it seems likely that he possessed the basic knowledge quoted
above very early, and evidently he knew from his research on the lemniscate that special significance is attached to the argument pair V2
and 1. There are also signs that already in 1794 he was aware of the
connection between the AGM and "numerical integrals", as he called
them, which today are referred to as theta functions 2 .
There were two phases to Gauss's research on the AGM, 13 May
1799 marking the start of the second phase. It was on this that Gauss,
now 22 years old, succeeded in forging a link between the AGM and
a quite different area which up to then he had been following quite
separately.

Theta functions are power series whose exponents are square roots.

96

7. Gauss and

'IT

The lemniscate
The second constituent in Gauss's 7r formula (7.1) besides the AGM
is the so-called lemniscate function. This function take its name from
the lemniscate 3 , a curve which Gauss had already studied in detail as
a teenager. The lemniscate takes its name from the Greek word lemniskos, meaning a little ribbon or bow. It looks rather like a sideways

8:

~A
Expressed in polar co-ordinates, its equation is as follows:
r2

cos 2()

(7.8)

When Gauss turned to the lemniscate, this curve was almost exactly 100 years old. Two brothers had discovered it independently in
the same year, 1694, Jakob Bernoulli (1654-1705) and his younger
brother Johann (1667-1748). Jakob's publication about it appeared in
September 1694 and Johann's only a month later. Following publication, the brothers had a violent dispute about which of them could
claim to have invented it [45, p. 311]. The quarrel went so far that
Johann swore he would never return to Basle as long as his brother
was still living4.
Jakob Bernoulli had discovered the lemniscate via the so-called elastic curve, and still expressed it in Cartesian co-ordinates (xx + yy =
aJ xx - yy). This elastic curve is formed if one forces a rod to bend
so far that its ends are perpendicular to an imaginary line connecting
their endpoints.

The lemniscate is the set of all the points for which~roduct of the distances
of two fixed points Hand F2 possesses the value (HH/2)2.
When Jakob died in 1705, Johann succeeded him to his professorship in Basle,
which post he held for another 43 years [83, p. 112].

7.4 History of a formula

~-n

97

o~-----------+

-_.LJB

The equation for the elastic curve is

y-

lo

z2 dz

(7.9)

Vl- z4

More important historically than the equation for the curve, however, was the equation for its arc length AGB. This goes like this:

lo

dz

(7.10)
2.6220575542 ...
o Vf=Z4
Jakob Bernoulli succeeded in writing a formula for this arc length
in 1691. He needed it again three years later when he was looking for
a curve that was analytically more manageable but had to have the
same arc length. He found this curve in 1694 in the lemniscate, whose
half perimeter is in fact given by the integral (7.10). Thus one can say
that the arc length of the lemniscate was known before the lemniscate
itself [45, p. 311].
In the eighteenth century, even before Gauss, the elastic curve and
the lemniscate figured in many mathematical treatises. For example, in
1730 James Stirling (1692-1770) specified approximate values for the
integrals (7.9) on the interval [0,1] and for (7.10), which were accurate
to 17 places. Giovanni Fagnano (1715-1797) discovered some methods
for dividing up the lemniscate arc into n equal parts, where n can be
equal to 2m , 3 2m or 5 2m [45, p. 313].
From 1748 Leonhard Euler (1707-1783) developed a first stage of
the theory of elliptic integrals, starting from the lemniscate. In particular, Euler discovered the remarkable relationship between the two
integrals already cited:
1
1
dz
z 2dz

w=2

Jo

VI - z4 . Jo VI - z4

(7.11)

98

7. Gauss and

7r

Gauss began work on the lemniscate in January 1797 (at the age
of 19), as is borne out by entry no. 51 in his mathematical diary [55,
p. 67]: "Curvam lemniscatam a

dx

Jl-x4

pendent em perscrutari coepi." (I have begun to research the lemniscate

curve, which is a function of ... ). Originally the word "elasticam" had
occurred at this point, but Gauss later crossed it out and replaced it
with "lemniscatam". Soon afterwards he began using the symbol w,
as used in (7.10).
Gauss began his research with the problem of how to divide a
lemniscate into equal parts which was mentioned above in connection with Fagnano. This took him to the "lemniscate functions"
sinlemn(J; (l-z4)-1/2 dz) = x and coslemn(w /2- J;(l-z4)-1/2 dz) =
x, which he defined for complex numbers and among which he found
several relationships. He also found (one year after discovering the regular 17-sided polygon) a construction using only compass and straight
edge, by means of which the lemniscate can be divided into 5 (not 17)
equal parts5. The first three months of 1797 saw a large number of
discoveries in this area, as indicated by four diary entries (nos. 54, 59,
60, 62).
Gauss was particularly encouraged when he noticed the analogy
of the lemniscate functions to the circle functions, for example the
following:
w /2 = 1\1 - z4)-1/2 dz and

7r /2

= 1\1 - z2)-1/2 dz

(7.12)

The analogy is visually even greater when one reflects that in Gauss's
day the symbol n was often still used in lieu of the symbol w.
In 1798 Gauss began to study the expression ~, which turned out
to be the key to further discoveries. Amongst other things he found a
series for the inverse of this expression, and with it he calculated w / 7r
to 15 decimal places.
Suffice to say that by July 1798 Gauss already knew "everything"
about the quotient 7r/w [45, p. 319]. Entry no. 92 of his diary states,
5

Henrik Abel (1802-1829) subsequently continued Gauss's work and found that
exactly the same ratios apply to the lemniscate as to the circle and, in particular,
that division of the lemniscate into 17 parts is possible with compass and straight
edge too.

7.4 History of a formula

99

"We have gained some very elegant details about the lemniscate, which
have exceeded all expectations, and indeed using methods which open
up an entirely new field." He sensed that he had made an important
discovery and under no. 95 (in October 1798) he even wrote, "A new
field of analysis stands before us, namely the investigation of functions
etc." Gauss was so excited that he left his sentence unfinished.

Merger of the two fields

Then came 30 May 1799. Gauss noted: "That the AGM is equal to -[;
between 1 and V2 we have confirmed up to the 11th decimal digit; if
this is proven, then a truly new field of analysis stands before us [56,
X.2, p. 43]."
Expressed as a formula, the discovery, which even today ranks as
sensational, looks like this:

~ = AGM( V2, 1)

(= 1.19814 ... )
(7.13)
w
In this formula Gauss recognised the connection between two apparently non-overlapping areas, namely the AGM and the lemniscate
functions, which in turn are closely related to the elliptic functions.
He was also aware at the time of the significance of his discovery as
his conclusion ("new field of analysis") suggests. This conclusion was
fully confirmed, moreover, in the 19th century by the work of Riemann,
Jacobi et al.
It is striking in the diary entry that Gauss had arrived at his findings not through mathematical deduction but through simple comparison of two numbers. Normally mathematicians are cautious about
inferring a law simply from numeric coincidence. But this was quite
different; Gauss seems to have suspected that this coincidence was no
chance occurrence. This was by no means the only instance in which
numeric results led Gauss to infer mathematical relationships, but it
was without a doubt the most prominent.
The wording in his diary also reveals that Gauss did not yet have
a proof, and this he developed only later. When exactly is not known,
but it is certain that by 23 December 1799 he actually had two proofs,
and in 1818 he even published a third.
The first and second proofs run over the comparison of the series
expansions of the left and right sides of (7.13). The third proof is the
shortest and the most elegant [56, III, pp. 352, 353]. Gauss examined
the complete elliptic integral of the first kind,

100

7. Gauss and

I (a, b) =

7r

r /2 -r===d=e===:::==
J a cos + b sin e

io

2 ()

of which the lemniscate perimeter is a special case: r:v = 2I( V2, 1). He
then showed that this integral remains constant if a and b are iterated in accordance with the AGM rule (alg 7.2): I(a, b) = I(al' bl ) =
I(a2' b2) = .... As ak and bk move towards the same limit value
AGM(a, b), I(a, b) = 7['/(2 AGM(a, b)) applies, which proves the relation (7.13).
It is only a few steps from this proof to the 7[' formula (7.1). In
modern accounts, for example as presented by Nick Lord [81], the path
from there takes one over the following steps which are not difficult and
were not new ground in Gauss's time. The missing pieces were, firstly,
the auxiliary integral L( a, b) = o7r /2 cos 2 ede / J a 2 cos 2 e + b2 sin2 e,
for which L(a, b) + L(b, a) = I(a, b) holds true; secondly, the relationship mentioned above that was discovered by Leonhard Euler(7.11) of L(V2, 1). I(V2, 1) = 7['/4; and, thirdly, the relationship
c6L(a, b) = (c6 - S)I(a, b) with S = L~o 2k-Ic~, which is obtained
from c6L(b, a)-2cI L(b l , al) = c6I(a, b)/2. If the argument pair (V2, 1)
is converted to the pair (1, 1/ V2), a more elegant version of the result
is obtained. And there you have the formula (7.1) and the Gauss AGM
algorithm.

The original
It took us a long while to uncover the original formula of Gauss which
forms the basis of the Gauss AGM algorithm. It appears on page 6 of
handbook 6, Short essays from various fields of mathematics, begun in
May 1809.
This little work is in the possession of the Niedersiichsische Staatsund Universitiitsbibliothek Gottingen (Lower Saxon State and University Library Gottingen) (shelf mark: Cod. Ms. C. Fr. Gauss, Handbuch
6). When we enquired after it, it turned out to be in a pitiful condition,
so that the manuscripts and rare books department first had to have
the book restored before we were able to print the formula (with their
blessing) for the first time as Gauss originally wrote it:

7.4 History of a formula

.'!tit ii..c1o!)
,I,

.. It "::.

1(t .. t'd

" =
I ..

(I

...

101

\f1i.6

y'"

~j ,t

""1)(..,,./...:., "-' '1

tit. ~
+4( I#'l'.~. tt." )

r'c'. 1.("("-+ 4 CIttL'It

+ c'e' ... '1.C'("

\"'41.-

a.~-::.It
,....

The formula in the last line

elll + etc.
} _!
+C'C'+2C"C"+4CIII C III +etc. - 2 aa
e' e'

+ 2 e" e" + 4 e

lll

_ 2mM
7r

(7.14)

is what we were looking for! It is even more general than (7.1). It is

only necessary to set a = A = 1 and b = B = 1/V2, and one obtains
m = M and e2 = C 2 = a 2 - b2 = 1/2. If instead of M and m one sets
AGM(l/V2), and if each stroke is replaced by an index, for example
if we write C3 instead of e"', the Gauss formula is transformed into:
2 ~ 2k-le~ =
~

k=l

! _ 2(AGM(1, 1/V2))2
2

7r

and from this the (7.1) is immediately derived.

(7.15)

102

7. Gauss and

7r

How could it happen ...

Several historians have observed Gauss's hand in the formula (7.1). In

particular, Salamin did so himself when in his paper Computation of 7r
Using Arithmetic-Geometric Mean [100] of 1976 he referred to the root
and added: "It is quite surprising that such an easily derived formula
for 7r has apparently been overlooked for 155 years. The author made
his discovery in December of 1973."
How could it happen that the Gaussian formula and along with
it the Gauss AGM algorithm were lost without trace so that it was
necessary for many years to calculate 7r with only the weaker arctan
formulae instead of with Gauss's formula?
Eugene Salamin commented on it as follows [101]:
"I quite agree that the 7r-AGM algorithm could have been discovered 150 or so years ago. The two key items are the Gauss AGM
formula for computing the elliptic integrals, and the Legendre formula
relating these different elliptic integrals. If you can determine when
these formulas were first published, then you know how long the 7r
method was waiting to be discovered.
It should be pointed out, however, that the AGM method, involving
full precision multiply, divide, and square root, was not very practical
in the days of Gauss and Legendre. It was only when multiplication
techniques substantially faster than the N 2 ones (see Knuth, vol. 2 [77])
were discovered, that the 7r AGM algorithm became more than just a
theoretical formula."
That is all water under the bridge now. In 1976 the number of
known digits of 7r stood at one million, this record having been established in 1973 using arctan series. After Brent und Salamin, this
record was left in the dust. In only three years, computers were already
chomping their way up to 30 million decimal places (see Chapter 13),
all using programs that were based on the Gauss AGM algorithm.

8. Ramanujan and

7r

The story of Srinivasa Ramanujan, the greatest Indian mathematician

of the modern era, could come from a sociology text book on the making
of geniuses. Impoverished youth, inadequate education, misunderstood
by those around him, forced to adopt an independent line. All of a
sudden someone spots the youth's brilliance, appreciates his visionary
thinking, takes him under his wing, encourages him and elicits from
him in the space of a few years an incredible wealth of discoveries. This
creative period is followed by illness, homesickness and untimely death.
In the ensuing period his work is largely lost without trace, known
only to a few privileged persons, forgotten by the world at large. Then
60 years later he enjoys a kind of revival. A few people begin to get
interested in him, enquire into his life and milieu, and suddenly this
interest is kindled into a raging fire as a storm of excitement spreads
through India and across the mathematically minded world. Jubilation
culminates in celebrations, conferences and posthumous honours on
the occasion of the centenary of Ramanujan's date of birth in 1987.
Ramanujan's extraordinary achievement was to tread new ground
in mathematics despite having had little in the way of a formal mathematical education. As such, he stirred the creativity and imagination
of the best brains and inspired them on to extraordinary discoveries
of their own.

8.1 Ramanujan's series

A number of Ramanujan's discoveries are linked with the number 7f and
the issue of how to calculate it. In 1914, at the age of 27, Ramanujan
published an article entitled Modular Equations and Approximations
to 7f [96]. This essay contains around 30 formulae for 7f, including the
following [96, formula (44)]:

104

8. Ramanujan and

1
27rV2

1103
99 2

71'

27493

1 1 3

+ 996 . 2 . 42 +
53883 1 3 1 3 . 5 . 7
99lO . 2 . 4' 42 . 82

+ ...

(8.1)

Ramanujan himself commented that this series converges extremely

rapidlyl. In fact, every summand produces about 8 correct decimal
places, as can be seen just from the first three terms:
0.1125 .... 10

0.273 .... 10- 8

0.229 .... 10- 16

In (8.1), Ramanujan had produced a 7r series which was significantly

better than previous formulae, such as, for example, the arctan formula
of John Machin, which moves towards 7r at the rate of only 1.4 decimal
places per summand. Ramanujan's new series was thus the first one
suitable for taking the calculation of 7r beyond the million mark, and
in fact Gosper calculated 17 million places of 7r with this formula in
1985.
One might ask whether the great speed of convergence of the series
(8.1) is not dissipated by the fact that what the formula calculates is
not actually 7r but 1 divided by 7r, while it also includes another square
root. But in fact this is not the case. Firstly, the calculation of the
square root and of the inverse are operations which are only performed
once and, secondly, procedures for calculating them which converge
very rapidly, in fact quadratically, have been available since the time of
Isaac Newton, i.e. for 400 years. The two terminal operations can thus
be performed in the time it takes to perform a few long multiplication
operations and add only insignificantly to the total time.
It is quite striking how it was in 1914 or even earlier that Ramanujan discovered the series (8.1), yet it was not used until 1985. The 70
years which passed between these two events were not a gestation period required for the series to mature, but the significant event which
occurred in the interim was the invention of the computer. Without
a computer, i.e. using paper and pencil only, it is not sensible to calculate an expression like the Ramanujan series over many terms. This
and other products of Ramanujan's inventive genius therefore had to
be put on ice until the tools needed to exploit them had been invented.

In the literature this formula is generally presented in a slightly more compact

form (see 13.49).

8.2 Ramanujan's unusual biography

105

Ramanujan provided only sparse information on his series (8.1) and

failed to furnish any proof. However, as is evident simply from the title
of the article, the formula is related to modular equations.
It is rare to discover an account of modular equations that is readily comprehensible. Fortunately, however, the Borwein brothers have
written such a description, which we may therefore assume to be authoritative [33, p. 112B]:
The Ramanujan 'IT series (8.1) is based on a 58th order modular
equation [11, p. 202]. Historians can only speculate as to how Ramanujan arrived at it. He himself says only that it was developed
from theories "corresponding" to the standard theory of the theta and
modular functions. We can be certain that it took Ramanujan a large
number of attempts to arrive at manageable modular equations.
In his paper on 'IT, as well as (8.1), Ramanujan lists another 16
series for 1/'IT which do not converge as well but some of which are
quite spectacular for other reasons. One of these series is as follows [96,
Formel (29)]:

~
'IT

k=O

(2n)
42n + 5
n
3

212n+4

(8.2)

This series does not converge particularly well, producing only 1.8
decimal places per term. But it is possible with this series to calculate
the second half of n binary places of 1/'IT without having to calculate
the first half, because the denominators grow twice as quickly as the
numerators. In this respect, Ramanujan had discovered here what was
a kind of precursor to the BBP series discovered only quite recently
(see Chapter 10), which permits the calculation of binary places in the
"middle" of 'IT.

8.2 Ramanujan's unusual biography

Srinivasa Ramanujan's short life has been portrayed in detail, the
main impetus to this being the run-up to the centenary of his date
of birth in 1987. The most comprehensive biography is that by Robert
Kanigel [71]. The publisher of Ramanujan's notebooks, Bruce Berndt,
has also collected together many biographical details [21], [22], [23].
In particular, he was able to interview Ramanujan's widow, S. Janaki
Ramanujan, who survived her husband by more than 60 years.

106

8. Ramanujan and

7r

Srinivasa Aiyangar Ramanujan was born in the southern Indian

city of Erode on 22 December 1887 in his mother's family's home. The
family was very poor. After his birth, his mother returned to Kumbakonam, where her husband worked in a clothing business. Kumbakonam is about 155 miles to the south-west of Madras.
Ramanujan showed an early interest in and aptitude for mathematics. But, unlike many famous mathematicians, he had no mentor
but acquired his mathematical knowledge entirely by himself. Up to
the age of 25 it is likely that he only had access to five mathematical
books. The first of these was S.L. Loney's Trigonometry, published in
Cambridge in 1893, and this he worked his way through at the age
of 12. The book discusses logarithms of complex variables, the calculation of Jr, the Gregory series and series sums, amongst other topics.
He himself explored similar avenues later on.
At the age of 15, Ramanujan borrowed a work by George S. Carr,
A Synopsis of Elementary Results in Pure Mathematics from the local
public library. Carr's book was republished in 1970 under a different
title [40]. This "synopsis" was in fact intended as an overview aimed at
preparing university students for their finals in mathematics. The book
lists 6,165 theorems, but without or with only very brief proofs. In any
case (or notwithstanding), this book seems to have aroused Ramanujan's genius. When Ramanujan started documenting his mathematical
discoveries in notebook form in 1904, Carr was evidently the model he
followed.
Already at this time, aged 16, he was completely taken with mathematics and only wanted to study that subject. As a result he failed
his examinations in the other subjects and was sent down. Further attempts to obtain a college education met with no success. "The Collage
of Kumbakonam rejected the one great man they had ever possessed,"
Ramanujan's patron, Hardy, wrote bitterly [63, p. 7]. Consequently,
from the age of 17 to 23, Ramanujan worked in complete isolation. He
entrusted the results of his research to his notebooks but otherwise
kept them to himself. It is very sad that he obtained no mathematical
"feedback" at such an important period of his life.
In 1909 Ramanujan got married. His wife, who was only nine at the
time, was able to provide some interesting details in 1984 (by which
time she had been widowed for 64 years) [22]. In particular, she said
Ramanujan was convinced that he would not have fallen ill in England
if she had gone there with him.

8.2 Ramanujan's unusual biography

107

Ramanujan's first mathematical publication appeared in 1911, and

only from 1912 did his extraordinary mathematical genius begin
to gain some recognition. Two English mathematicians, Sir Francis
Spring and Sir Gilbert Walker, arranged a monthly stipend for him.
After a long search, Ramanujan obtained employment at a harbour
company in Madras in 1912. The proprietor of the company and a
manager who was a mathematician himself took great interest in Ramanujan's work and encouraged him to inform some English mathematicians of his discoveries. His first attempts did not produce any
encouraging results, but then in January 1913 he wrote to Godfrey
H. Hardy, a very well-known English number theoretician at Trinity
College, Cambridge University. Hardy was immediately convinced and
invited Ramanujan to Cambridge. After having to face the objections
of his caste and family (for orthodox Brahmins an ocean crossing was
grounds for being made an outcast), Ramanujan sailed from India
in 1914.
He began one of the most important collaborations in the history
of mathematics. During his time in Cambridge Ramanujan published
several articles, some of which are classified as fundamental. One of
these is the above-mentioned article from 1914 on modular equations
and approximations for 'Jr, which is where the series (8.1) comes from.
This paper is the only one Ramanujan wrote about 'Jr, although additional findings relating to 'Jr are documented in his notebooks.
He wrote several articles jointly with Hardy. "The blend of Hardy's
technical expertise and Ramanujan's raw brilliance produced an unequated collaboration" [33, p. 112AJ, commented the Borwein brothers.
After three years in Cambridge, Ramanujan succumbed to an
unidentified illness in 1917. It was diagnosed as tuberculosis but it
may simply have been vitamin deficiency due to poor nutrition during
the First World War. After two years in sanatoriums and convalescence
homes, he returned to his native country in 1919.
Soon after his return he was offered a chair at the Hindu University
of Benares. Ramanujan was forced to decline for the moment due to his
ill-health but planned to take up the offer when his health improved.
But unfortunately he never recovered from his illness and died on 26
April 1920 at the age of 32 in Kumbakonam.
Ramanujan left 37 publications and a wealth of problems which
he had published in the Journal of the Indian Mathematical Society.

108

8. Ramanujan and

7l"

There were also five notebooks and various unpublished essays and
manuscripts plus 120 theorems in his letters to Hardy.
A simple list of the topics on which Ramanujan worked reads as follows [25, p. 644J: partitions, theta functions, statistical mechanics, Lie
algebra, probability theory, modular forms, elliptic functions, complex
multiplication, hypergeometric series, q series, asymptotic behaviour,
beta integrals.
He also made some comparatively elementary discoveries. There
is an anecdote about a taxi with the registration number 1729. It
was in such a taxi that in around 1918 Hardy drove to a hospital to visit Ramanujan. On entering, Hardy mentioned this number,
upon which Ramanujan commented, "1729 is a very interesting number, it is the smallest number expressible as a sum of two cubes in
two different ways." Ramanujan was referring to the representations
13 + 123 = 93 + 103 , both of which produce 1729. In fact it is true
that no smaller number possesses this characteristic. Ramanujan also
discovered formulae for the construction of higher sums of powers and
used them, for example, to ascertain the identity
44

+ 64 + 84 + 94 + 144 =

15

(8.3)

Ramanujan's skills at handling continued fractions are unsurpassed.

Among his results, for example, we find - without proof, as was so often
the case - the following continued fraction formulae which helped to
convince Hardy of Ramanujan's genius. They are among the most aesthetically pleasing he left, but they are also some of the most difficult
to prove [63, p. 8J:

J +.J5 _ .J5 +
5

_ _ _e_-_;-.,,----_ _

(8.4)

e- 2rr
1 + -------;--e- 4rr
1+
-6rr

1+_e_ _
1 + ...

V5+1
1+

~ (V5-1)~
2

54

2
1

271"VS
e--5-

- - - - = - - - - (8.5)

e- 2rrv5

1+-----=-e- 4rrv5
1+
e- 6rrv5
1+--1 + ...
----;:=-

8.2 Ramanujan's unusual biography

109

Hardy said of these and other formulae containing continued fractions [63, p. 9]: "I had never seen anything in the least like them before.
A single look at them is enough to show that they could only be written down by a mathematician of the highest class. They [the formulae]
must be true because, if they were not true, no one would have had
the imagination to invent them."
Some of Ramanujan's findings have been lost, even though G.E. Andrews at least recovered Ramanujan's "missing notebook" in 1976.
Some of it the University of Madras may have succeeded in losing. At
any rate Mrs. Ramanujan said that shortly after her husband's death
his former teacher had come and claimed all his documents for the
university, but no one has succeeded in finding them there.
Ramanujan's Collected Papers appeared seven years after his death.
In 1936, Hardy gave "twelve lectures on subjects suggested by [Ramanujan's] life and work", which were published in 1940. In the foreword, Hardy writes as follows about his protege [63, p. 6]: "The real
tragedy about Ramanujan was not his early death. It is of course a
disaster that any great man should die young, but a mathematician
is often old at thirty, and his death may be less of a catastrophe than
it seems. Henrik Abel (1802-1829) died at twenty-six and, although
he would no doubt have added a great deal more to mathematics, he
could hardly have become a greater man. The tragedy of Ramanujan
was not that he died young, but that, during his five unfortunate years,
(between eighteen and twenty-three) - his genius was misdirected, sidetracked and to a certain extent distorted."
Several authors, including the Borwein brothers, have succeeded
in recent years in rendering Ramanujan's work much easier to understand and more transparent. A significant contribution here was made
(and is still being made) by Bruce Berndt, who is currently publishing Ramanujan's notebooks with a detailed commentary. Above all,
Berndt is adding in the missing proofs (and in some cases dis proofs )
which Ramanujan omitted to write down or else wrote only incompletely. Four of the five notebooks have already been published [24].
For friends of Jr, notebooks 3 and 4 are the most interesting.

110

8. Ramanujan and

7r

8.3 Impulses
The path which led Ramanujan to his discoveries about 7r has proved
to be of great practical importance and very forward-looking. Several
other researchers have continued down the same route.
The Borwein brothers have succeeded in deriving general formulae
which enable any number of convergent series of the Ramanujan type
to be generated (through fixing of one parameter). Further information
will be found on our CD-ROM in the ari th directory. Following this
model, one of the authors (JA) generated the following series which
produces a staggering 50 correct decimal places per term:
1

(6n )!

A +n B

12n (3n)! (n!)3

In

00

;: = yCI2] ~

(8.6)

where

A := 5280419026080999965452185 +
+ 2361475178400070170568800 v'5 + 32 v'5 .
. (10891728551171178200467436212395209160385656017+
+4870929086578810225077338534541688721351255040 v'5) 1/2
B := 654159204458052267524145750 +

+ 292548889855077669080467200 v'5 +
+ 209664 )3110
. (6260208323789001636993322654444020882161+
+ 2799650273060444296577206890718825190235 v'5)

1/2

J := - [17897749588626020 + 8004116944887336 v'5 +

+ 108 v'5(10985234579463550323713318473 +
+ 4912746253692362754607395912 v'5) 1/2] 3
Arndt [7J, 1994

Another spectacular but also practical "Ramanujan" -type series

was developed by the Chudnovsky brothers. It goes like this:

8.3 Impulses

~=
1f

12
~(_l)k (6k)! 13591409 + 545140134k
J640320 3
(k!)3(3k)!
(640320 3)k

f='o

111

(8.7)

D. and G. ChudnoYsky, 1987

It was with this series, which produces 15 decimal places per series
term, that the Chudnovskys calculated 1f to over 8 billion decimal
places in 1996. The series is used in several computer algebra systems,
e.g. in Mathematica, to calculate 1f, and is thus a very useful series; it
would also be the ideal candidate for anyone seeking to divide up the
task of calculating 1f on the Internet because it can be adapted very
well to the binsplit algorithm (see Chapter 15 on page 215ff.).
In the more recent high-performance calculations of 1f, Ramanujantype series have been replaced by other methods, notably by iterative
algorithms developed by the Borweins.
One of these is this "quintic" algorithm:

Algorithm (8.8) Borwein, quintic

(alg 8.8)

Initialise:
So := 5(y5 - 2)

ao := 1/2
Iterate:

Sk+l

:= (25)/(Sk(Z
where

+ x/z + 1)2)

1
IT

J. and P. Borwein, 1987 [33, p. 115]

produces 5th order convergence on 1/1f, i.e. on every iteration

the number of decimal places found undergoes a fivefold increase.
ak

Such algorithms may not be found in Ramanujan's writings, but

they are closely related to Ramanujan's analysis. For example, this
special algorithm is based on a 5th order modular equation (inside the
relation Sk) whose origin can be traced back to Ramanujan. Thus one
can say with some justification that Ramanujan's hand is still to be
found in 1f calculations today.

9. The Borweins and

7r

If anyone or two figures stand out from the 7r research of recent years,
then it must be Peter and Jonathan Borwein. They have developed
virtually all the high-performance algorithms on which 7r computations
are based today.
The brothers were born in St. Andrews, Scotland, Jonathan in 1951
and Peter in 1953. Their parents emigrated to Canada, where David
Borwein was for many years head of the mathematics institute at the
University of Western Ontario and is now a professor emeritus. It was
"natural" that the brothers should study mathematics, and Jonathan
obtained his PhD in 1974, Peter in 1979.
After completing their studies, they worked - together - at Dalhousie University in Halifax, the capital city of the east Canadian
province of Nova Scotia. There they both became professors as well.
Then in 1992 Jonathan moved from the east to the west of Canada, to
be followed a year later by Peter. Today they both teach and research
at the Simon Fraser University in Burnaby, British Columbia. There
they head the Center for Experimental & Constructive Mathematics
(CECM), which they have had a free hand at establishing.
The Borweins' academic work is already very considerable. Peter
Borwein's list of publications includes five books and almost 100 articles just over a period of 14 years. Jonathan Borwein has published a
similar volume of papers.
7r is not the only research field in which the Borweins are active.
In fact, less than 20% of their publications include 7r in the title. Nevertheless 7r is the main topic with which their names are currently
linked. The main reason for this is that since 1984 the brothers have
developed a large number of high-performance algorithms for the calculation of 7r which have been immediately implemented and executed
by various mathematicians, such as Bailey and Kanada. The Borweins
made their debut in the 7r field in 1984 with the following quadratically

114

9. The Borweins and

7r

convergent algorithm [30J, which they derived from the

Gauss, Salamin and Brent (see Chapter 7):

Algorithm (9.1)

7r

AGM derived, quadratic

7r algorithm of
(alg 9.1)

Initialise:

ao
bo

:=
:=

Po :=

J2
0

2+ J2

Iterate (k = 0,1,2, ... ):

ak+1 :=

~ (..fiik + k)

bk +1 .=
.

v~"

Iak (J!rll)
bk+ak

k +1 )
._ Pk bk+1 (1+a
2
Pk+1 .1+bk+ 1
--t
Then converge the Pk'S quadratically to
1
Pk - 7r < 10 2k+
(k ~ 2)

7r+
7r+, where:

J. and P. Borwein, 1987 [30, p. 360]

But this algorithm was only the starting point for many other algorithms of the iterative type, including some which converge even better.
Among these, the following quartic algorithm made their names especially well known, as since its publication in 1987 it has been used in
most subsequent world record calculations, notably in those achieved
by Yasumasa Kanada [32, p. 170J:

Algorithm (9.2) Borwein, quartic

(alg 9.2)

Initialise:
Yo :=

ao

J2-1
4J2

:= 6 -

Iterate (k = 0,1,2, ... ):

1- {/l-y4
Yk+1 := l+~

+ Yk+d 4 - 22k +3Yk+1(1 + Yk+1 + Y~+l) ~ ~

Then converge the ak quartically to 1/ 7r
ak+1 := ak(l

The error is:

o ak -

~ :::; 16 . 4k . 2e- 4k27r

J. and P. Borwein, 1987 [32, p. 170]

This algorithm is expressed in a more general form together with

many other algorithms in the 7r textbook Jonathan and Peter Borwein
published in 1987 under the title [32J

9. The Borweins and

7r

115

Pi and the AGM

A Study in Analytic Number Theory and Computational Complexity
This important book (a must for all 7r enthusiasts who dare to
try their hand at substantial mathematical fare) covers all aspects
of 7r research, not just the rapid computation of 7r to many decimal
places. For example, the authors go back to the 19th century, to the
analysis of the time, elliptic integrals and functions, theta and modular
functions, Lagrange (1736-1813), Legendre (1752-1833), Gauss (17771855), Jacobi (1804-1851) and Ramanujan (1887-1920).
The whole book is permeated with the respect and admiration of
the two young writers for the achievements of those on whose shoulders
they stand, who researched the subject before them and provided the
groundwork for their own discoveries. The voluminous list of references
are a testament to this, as is the fact that they dedicate the book to
their father, David Borwein.
In addition to the high-performance algorithms themselves, for
which they have become famous, the Borweins explain their general
structure. They prove, for example that the Ramanujan series (8.1) is
the special case n = 58 of a much more general approach. And on the
few pages of their Chapter 5, algorithms pour out one after the other
- quadratic, cubic, quartic, quintic, septic and even nonic algorithms.
The reader is invited to develop additional ones. Incredible. (You will
find more on this subject on our CD-ROM in the directory arith).
Even after publication of their 7r book, the Borweins have remained
true to the number 7r. In 1995 Peter Borwein and two other researchers
came up with the spectacular discovery known as the Bailey-BorweinPlouffe series (see Chapter 10). This series makes possible something
one can scarcely imagine, namely that it might be possible to calculate individual (hexadecimal) digits right in the middle of 7r without
having to calculate all the previous digits first. Just try and imagine
the problem as a simple multiplication operation. From a product of
30-digit numbers, find the 29th digit under the condition that you may
not calculate the digits either to the right or left of it. Sounds difficult
indeed. Now suppose, for example, that you want to find the 123 billionth digit of 7r even though this digit is inextricably linked with its
predecessor digits in a gigantic tree. Yet it is possible.
Just recently a new book by Jonathan and Peter Borwein (together
with Lennart Berggren) has been published [20],

116

9. The Borweins and

7r

Pi: A Source Book

This book contains 70 important original articles on the subject of
in facsimile form, for example the historically important treatises
by Lambert on the irrationality of 7r (1761) and by Lindemann on the
transcendence of 7r (1882). However, the book also includes some important articles by the Borweins themselves. This excellent and useful
book confirms the historic dimension which they see as providing a
setting to their own work.
The Borwein brothers' ground-breaking work has not only ensured
that the number of known 7r digits has increased enormously, but above
all we owe it to the Borweins that 7r research is today held in high
esteem - even among mathematicians, of whom not a few believed
that the whole topic had been laid to rest since Lindemann's proof of
the transcendence of 7r.
On the basis of the results achieved by the Borweins, it has even
been suggested that separate branches of mathematics should be established for 7r and e, which might then be known respectively as 7r mathematics and e mathematics. e mathematics would be linear, explicit,
open to generalisation, easily approachable, highly algebraic and would
lead via the exponential function to topics like one-parameter subgroups, Lie algebra and group representations. By contrast, 7r mathematics would be non-linear, chthonic, barely generalis able, highly analytical and, via modular functions and Ramanujan identities, would
have a significant impact on function theory, number theory and combinational logic [121].
Who knows, perhaps one day there will be an academic degree of
Dr. math. 7r. Anyone holding such a degree would have Jonathan and
Peter Borwein to thank.
7r

10. The BBP Algorithm

The title of the invitation sounded merely interesting:

On the nth digit of a transcendental number OR The 10 billionth hexadecimal digit of 7r is '9'.
However, a sensation was concealed in the text which followed:
"Up to now, it was generally believed that to compute the
nth digit of a transcendental number like 7r was as difficult as
calculating 7r itself. We will show that this is not true . ..
We will give algorithms for this ... These algorithms can be
easily implemented. do not need multiple precision arithmetic,
require virtually no memory, and feature run times that scale
nearly linearly with the order of the digit desired. They make
it feasible to compute, for example, the billionth binary digit
of log 2 or 7r on a modest work station in a few days of run
time ... "
Such was the wording of the title and abstract with which invitations were issued on 30 September 1995 to attend a colloquium at
the Simon Fraser University in Burnaby, Canada. The invitations had
been issued by three well-known 7r mathematicians, David Bailey, Peter Borwein and Simon Plouffe.
Two weeks later at the colloquium the three speakers presented the
algorithms they had announced, gave the necessary proofs for them
and with this overturned the old adage, he who seeks to climb to a
mountain peak (or to calculate the furthermost digits of 7r) must climb
the mountain by foot (or start from the initial "3").
It was a new formula which destroyed an old "paradigm":

7r

1 (4

~ 16n
00

1)

8n + 1 - 8n + 4 - 8n + 5 - 8n + 6

(10.1)

118

10. The BBP Algorithm

With this series it thus becomes possible to calculate any hexadecimal place within 1f. Even better, the series enables the next digits from
any given hexadecimal place of 1f, for instance from the last known
digit, to be added piece by piece without having to calculate a single
prior position. The distinguishing term in the formula is the 16n in the
denominator of every series term.
We will shortly examine the mathematical procedure of the "BBP
series" (as it has since become known). But first let us consider the
question of how the formula came about.
Certainly not by chance, even if luck played some part in the discovery. All three parties are established mathematicians who have been
working with the number 1f for a considerable time. The mathematical
background against which the BBP series was developed has been thoroughly researched. Yet the series was not discovered through mathematical deduction or inference. Instead, the researchers used a tool
called Computer Algebra System and a particular procedure called the
"PSQL algorithm" to generate their series. They themselves write [10]
that they found their formula (10.1) "through a combination of inspired testing and extensive searching" .
Computer Algebra is a scientific area which is concerned with methods for the solution of mathematically formulated problems using symbolic algorithms. Its origin is sometimes attributed to Ada Augusta,
Countess of Lovelace l (1815-1852), who may have been the first person to remark that computers are also suited for handling symbolic
data (such as formulae). Computer Algebra Systems, of which the
best known are MuPAD, Maple or Mathematica, normally work with
mathematical symbols rather than with numbers. One of their use is
for checking the derivations of equations, but they are also used to
identify new relationships, as in the present context or to perform
high-precision calculations.
The PSQL algorithm is used to find integer relationships between
real numbers. A vector of real numbers (Xl, X2, . .. ,xn ) is entered. The
algorithm then looks for a vector of integers (aI, a2, ... , an), which do
not all = 0, such that the relation alxl + a2X2 + ... + anxn = holds
true. It is easy to see why this algorithm was a material factor in discovering the BBP series. The researchers had established that the classic
series for In 2 = L:~=l n.~n can be used to calculate any binary digits of In 2. They therefore wondered whether there might be a similar

The programming language ADA is also named after the Countess.

10. The BBP Algorithm

119

series for 7r. After many failed attempts, they finally fed the PSQL al. h
'th
,\",00
1
,\",00
1
gont m WI Xl = 7r, X2 = un=l (8n+1)16n"'" X8 = un=l (8n+7)16 n '
When the algorithm then churned out the vector (1, -4,0,0,2,1,1,0),
the BBP series was discovered [5SJ.
In this way the series (10.1) was the brainchild not of a human but
a computer. Nevertheless, it still needs a normal proof, which is not
very difficult [13J.
First it is necessary to establish that for every k < S the following
is true:

1/../2

1
o

xk-1

- - -8 dx
1- x

11/../2

00

x k -1+8n dx

n=O

1
= 2k/2

00

n=O

(10.2)

+ k)

16 n (Sn

Then one can write:

1 (4
2
1
1)
Sn + 1 - Sn + 4 - Sn + 5 - Sn + 6

~ 16n
00

11../2 4yi2 - sx 3
1o

4yi2x 4

I-x

Sx 5

dx

(10.3)

If one inserts y = xyi2, one obtains

}0

16y -16
d
y4 - 2y3 + 4y - 4 Y

1
1

4y

11

4y - S

(10.4)
--dydY=7r
o y2 - 2
0 y2 - 2y + 2
where the fraction in the integral on the left has been partly reduced.
The BBP series could certainly also be used to calculate 7r from
scratch, but it is not particularly well suited for such a computation,
so it will never replace algorithms such as, for example, the Gauss
AGM algorithm, which have been known for longer. The strength of
the formula (10.1) clearly lies in its ability to calculate individual digits
"anywhere" in 7r. The computing plan for this is not at all difficult and
moreover it contains a special treat.
During evaluation of the BBP series, one calculates each of the four
summands 8 1 = I:~=o 16n(Jn+1) ' 8 2 , 83, 8 4 separately. Let us consider,
for the sake of example, 8 1 .

120

10. The BBP Algorithm

To obtain the sequence of hexadecimal digits of 81 which begin at

a given position p = d + 1 (p = 1,2, ... ), 8 1 is shifted towards the left
by d hexadecimals places. The part shifted beyond the hexadecimal
point is discarded. The digits are then obtained from the part of the
product 16 d . 8 1 which appears after the hexadecimal point. When it
comes to the divisions by Sn + 1, once again only the places after the
hexadecimal point are of interest, so that only the remainder of these
divisions ever have to be calculated, as a result of which the numbers
carried along are enormously reduced. In this way, the calculation of
8 1 is simplified in the following way:
Decimal portion of 16 d 8 1
00
16 d - n
=
Sn+l modI

n=O

=~
~

n=O

16

d n
-

mod(Sn + 1) mod 1 + ~ _1_ _1_

Sn+ 1
~ 16 n - d Sn+ 1

(10.5)

n=d

Here, reduction through the modulo function is used wherever appropriate, so that only the remainder of each division is considered. The
summand 81 is now represented by two part-sums. In the first of these,
the d - n's are positive so that the powers 16 d - n are in the nominator
and the series terms are > 1. In the second part-sum, on the other
hand, the d - n's are negative and all the series terms are < 1. The
reduction to remainders is only necessary and useful for the first partsum.

10.1 Binary modulo exponentiation

The treat we referred to above occurs during calculation of the numerators of the first part-sum 16 d - n mod(Sn + 1). If the modulo operation
in the first part-sum is initially ignored, then we have the most effective calculation possible for a power of the form xk. There is an old
procedure for this which is described by Knuth [77, p. 461] as follows:
Suppose, for example, that we need to compute x 16 ; we could simply
start with x and multiply by x fifteen times. But it is possible to obtain
the same answer with only four multiplications, if we repeatedly take
the square of each partial result, successively forming x 2 , x4, x 8 , x 16 .
The same idea applies, in general, to any value of k, in the following
way. Write k in the binary number system (suppressing zeros at the

10.1 Binary modulo exponentiation

121

left). Then replace each "I" by the pair of letters SX, replace each
"0" by S, and cross off the "SX" that now appears at the left. The
result is a rule for computing x n , if S is interpreted as the operation of
squaring, and if X is interpreted as the operation of multiplying by x.
For example, if k = 23, its binary representation is 10111; so we
form the sequence SX S SX SX SX and remove the leading SX to
obtain the rule SSXSXSX. This rule states that we should "square,
square, multiply by x, square, multiply by x, square, and multiply by
x; in other words, we should successively compute x 2, x4, x 5, xlO, xu,
x 22 , x 23
According to Knuth this "binary exponentiation" was already
known in India over 2000 years ago. It is recorded in 200 BC in India
in a work called "Chandah-sutra" [77, p. 461]. Knuth notes that this
information was evidently confined to India and outside of India there
was still no mention of the method 1000 years later.
In the problem we are about to consider, a small modification
should be made. As we are only interested in the places after the
hexadecimal point and integer elements are negligible in the calculation, we only need to calculate the remainders of each division and can
forget the quotients. This makes all the numbers a lot smaller. Thus
on every iteration, a modulo operation is performed with the (next)
denominator c = 8n+ 1. The result is "binary modulo exponentiation" .
This can be easily translated into an algorithm:

122

10. The BBP Algorithm

Algorithm (10.6) Binary modulo exponentiation

(alg 10.6)

To calculate r = bk mode, let r := bmode and t be the value of

the greatest power of 2 which is not bigger than k. Set k := k - t.
Then:
while ( t > 1 )

:=

t/2;

r := r2 mode
if(k?::t)
{
k:= k - t
r := b rmode
}

}
At the end of this calculation, r is given the desired value bk mod e.
The algorithm is executed entirely with positive integers which do
not become greater than (e - 1)2. A numeric example from [13] goes
as follows. During the calculation of 349 mod 400, r successively assumes the values 3, 9, 27, 329, 241, 81, 161 and 83, and in fact
349 = 239,299,329,230,617,529,590,083, so that 83 is the correct result.
Calculation of the rest of the formula (10.1) is very simple. The
summands calculated through binary modulo exponentiation have to
be divided by 8n+ 1 and added up. Once again we are only interested in
the portions which come after the hexadecimal point. The procedure is
then expanded to the summands 82 to 84. The sum 481 - 282 - 83 - 84
is then calculated, and its first place after the hexadecimal point is the
desired hexadecimal digit of 1r.
As you can see, the BBP series is surprisingly simple. Neither the
formula (10.1) nor its proof or implementation are particularly subtle or complicated. It is precisely this which makes the algorithm so
interesting.
Even if 7f is the main point of interest, as is so often the case,
especially in this book, it requires only a minor foreshortening of the
results of Bailey, Borwein and Plouffe to narrow it to 7f. In fact it goes
a lot further. The three researchers have not only found additional
formulae of the above type for 7f, 7f2 etc., but they have also discovered
such formulae for other "polylogarithmic" constants, such as log 2.

10.2 A C program on the BBP series

123

Their paper on the subject [10] presents these results in detail as well.
We recommend most strongly that any reader so inclined should look
at the article, which can be easily downloaded from the Internet.

10.2 A C program on the BBP series

Below we present an ANSI C program for calculating the pth hexadecimal place after the hexadecimal point of 7f. p can be up to around 536
million. The program uses the BBP series (10.1) of Bailey, Borwein
and Plouffe.
It is almost the same as the program which David Bailey has placed
on the Internet 2 .
In the program, the central operation of the binary modulo exponentiation is performed in the function expm (). After the description,
the source code should be easy to understand.
The function series () is used to compute each of the 4 summands
in the BBP series. As explained above, each of these summands has
two part-sums. The first consists of the first d = p - 1 terms with
numerators > 1, and the second consists of the succeeding terms with
numerators = 1. The portion of the terms in the first summand which
comes after the hexadecimal point is calculated using binary modulo
exponentiation; the terms in the second summand are all < 1 and are
determined in the normal way using floating point division. From the
second summand, only so many terms are used until, taking roundings
into account, the result is accurate to 8 hexadecimal places, for the
program outputs not just the pth place but also the next 7 places.
To achieve a level of accuracy of 8 hexadecimal places after the hexadecimal point one needs around 42 binary places, corresponding to a
decimal accuracy of around 10- 13 , which is in fact the value of eps in
the program.
In main () the user parameter p is read in, calculation of the four
summands is triggered and finally the post-hexadecimal point portion
from position p is printed as a hexadecimal number.
The time it takes to run the program is a little more than of a
linear order. To compute the millionth hexadecimal place of 7f takes
48 seconds on our Pentium 400 platform, while the 10 millionth place
takes 580 seconds.
2

http://www.cecm.sfu.ca/personal/pborwein

124

lO. The BBP Algorithm

It should be possible to compile the program with any ANSI C

compiler.
1***************************************************** *****
** This ANSI C-Program computes the p-th hexadecimal digit
* of pi (plus the following 7 digits).

** p comes from the command line.

* 0 <= P <= 2-29 (approx 536 million).
* p=l means the first digit after the radix point.
**
*
*
*
*

The program uses the BBP-algorithm of

D. Bailey, P. Borwein und S. Plouffe.
Its central function is expm() which performs the
"binary modulo exponent at ion" .
After D. Bailey 960429

****************************************************** ****1
#include
#include
#include
#include
#include

<stdio.h>
<stdlib.h>
<math.h>
<time.h>
<float.h>

1* Limits:

d_max = p_max - 1
ak_max = 8 * d_max + 6
r_max = ak_max - 1
r_max-2 < 2-(LDBL_MANT_DIG).
With mantissa length = 2m (usually 2m
r_max <= 2-m-1
ak_max <= 2-m
d_max < 2-(m-3)
p_max < 2-(m-3)

*1

#define p_max

64):

pow(2, LDBL_MANT_DIG*0.5 - 3)

1***************************************************** *****
**

expm(long n, double ak)

** Computes 16-n mod ak using binary modulo exponentation.

****************************************************** ****1

double expm(long n, double ak)

{

long double r = 16.0;

long nt;
if (ak == 1) return 0.;
if (n
0) return 1;
if (n == 1) return fmod(16.0, ak);

1* nt: largest power of 2 not greater han n/2. *1

for (nt=1; nt <= n; nt =1)

10.2 A C program on the BBP series

nt

= 2;

1* Binary exponentiation modulo ak. *1

do
{

r = fmodl(r*r, ak);
if n & nt) != 0)
r = fmodl(16.0*r, ak);
nt = 1;
} while (nt != 0);
return r;
}

1***************************************************** *****
**
**

series(double m, long d)
computes sum_k {16-(d-k)mod(8k+m)}/(8k+m),
(k=O, 1, 2, ... , d-l)

****************************************************** ***1
double series(double m, long d)
{

long
k;
double ak = m;
double s = 0., t = 1., x;

1* Sum of terms O.. d-l *1

for (k = 0; k < d; k++)
{

x = expm (d-k, ak);

s += x I ak;
s = fmod(s, 1.0);
ak += 8.0;
}

1* Some additional terms for 8 hex digit accurracy *1

#define eps 0.25e-12 1* = 16-{-10} I 4 *1
while ( (x=t/ak) > eps)
{

s += x;
1= 16.0;
ak += 8.0;

return s;
}

1***************************************************** *****

*
main 0
*
******************************************************* ****1
main(int argc, char *argv[])
{

time_t t_beg, t_end;

double s;
double p = 1000.0;

1* default fuer p *1

125

126

10. The BBP Algorithm

1* Get p from the command line *1

if (argc> 1) p = strtod(argv[l] , NULL);
if (p < 1)
p = 1;
if (p > p_max) p = p_max;
printf("Hex digits %.Of to
Lbeg=clockO;
s = 4*series (1, p-1)
- 2*series (4, p-1)
series (5, p-1)
series (6, p-1);
s += 4;
/* ensure s

%.Of of pi: ", p, p+7);

/*

4, 8i+l */

1* -2, 8i+4 *1
1* -1, 8i+5 *1
1* -1, 8i+6 *1
>= 0 */

t_end=clockO;
printf("%08lX\n", (unsigned long) (s*pow(2, 32);
printf(HElapsed time was: %.1f sec.\nH,
1.0*(t_end-t_beg)/CLOCKS_PER_SEC);
return 0;
}

10.3 Refinements
Bailey, Borwein and Plouffe had hardly announced their "shocking
formula" (10.1) when other researchers sat down at their computers
and attempted to also find beautiful "BBP-like" series. In particular,
BBP formulae were found for other constants, for example, for 71"2 and
log 7, as well as simpler or more effective formulae for 71" itself (to view
some examples, see our collection of formulae on page 227).
The research duo Viktor Adamchik and Stan Wagon win the prize
for the simplest of all the BBP formulae for 71":
00

71"

(_l)n ( 2

=~~

4n

1)

+ 1 + 4n + 2 + 4n + 3

(10.7)

Adamchik, Wagon, 1997 [2]

This formula consists of only 3 terms, i.e. one term fewer than the
original BBP series. But it is not only shorter.
Wagon and Adamchik have determined with this series [lJ that
there is a route to its derivation which simultaneously includes the
proof of the series. With (10.7) one no longer needs to convert the
summands found by the computer into integrals and then resolve them,
as shown above, but instead one can to a certain extent lean back
and leave the computer to manipulate identities. The authors needed
"only" one idea.
As such an idea, they used the expression:

10.3 Refinements

~ (-l)n ( a l
~-n

n=O

4n

a2

a3

a4

+ 1 + 4n + 2 +4n-+-3 +4n-+-4

127

They entered this into their computer algebra system Mathematica 3.0
which then produced the following transformed expression:
1
S(2( 4(a2 arccot 2 - a4log 4 + a4log 5 + a3( 7r /4 + arccot 3 - (log 25) /4) ))

+ a I (7r + 4 arccot 3 + log 25) )

At this point the authors told the computer to perform a few simple
substitutions, e.g. the substitution 7r = arctan 1 + arctan 2 + arctan 3,
following which it produced this still further simplified output:

2a2 + (a2l 7r

a2

+ a3 )

arctan 2 - 2a4log 2 + (al

4""

a3 + a4 )
- 2

log 5

This took care of the most irksome work. Wagon and Adamchik now
needed only to look for such values for aI, a2, ... a4, with which the
coefficients of the second to fourth terms become = 0, while the first
coefficient becomes = 1. This solution they would of course have also
been able to find in 10 seconds by hand, but they got the computer to
do this for them too, and for this purpose they entered

They could hear their computer deliberating, and then the system
delivered quite simply:
{{ a2 -7 2, al -7 2, a3 -7 1, a4 -70}}

And with that the above simple 7r formula (10.7) was born and proven
at the same time! Child's play, really...
Short and instructive though the formula (10.7) is, it is not particularly effective. Fabrice Bellard overcame this shortcoming with the
following series:
00

7r

(_1)n(

1
= 64?; 1024n

32
1
4n + 1 - 4n + 3

64
4
4
- IOn + 3 - lOn + 5 - lOn + 7

256
IOn + 1

1)

+ lOn + 9

(10.8)

Obviously (10.8) is quicker than the two previously cited BBP series
for 7r. Our own measurements confirm the 43% saving which Bellard

128

10. The BBP Algorithm

claims for this series compared with (10.1). This improvement is based
on the fact that Bellard's formula calculates to the higher base 1024,
compared with the base 16 of the BBP series, so that he needs fewer
terms for the same search item. This formula enables one to get further
into 7r with the same amount of effort.
Bellard's series is currently the favourite amongst 7r digit hunters.
With it, Bellard and most recently the student Colin Percival have
broken several records, and as of September 2000 the latest record was
the 10 15 th binary place, a 02 (see page 203).
The BBP series and the many variants which have subsequently
been developed unfortunately can only produce hexadecimal (or binary, octal etc.) places, and this is at least somewhat exotic. It would
be nicer if we had a formula which could produce any decimal place
of 7r.
Now, for some time we have known of a constant whose hindividual
decimal places can be calculated with a BBP-like series. This constant
is In(0.9) and its related series goes like this:
In

(9)
10

L
00

= -

n=l

11
IOn

(10.9)

This series raises hopes that for 7r too a series will one day be
discovered whose denominators contain a power of 10 rather than a
power of 2, so that we can then explore decimal, and hence familiar
terrain.
Unfortunately, this is not very likely. One of the distinguishing features of the BBP formulae for 7r is the fact that they contain arctan expressions which are rational multiples of 7r to base 16. (Unfortunately)
it has meanwhile been shown that there are no arctan expressions with
the base 10, so that the search for BBP formulae for decimal places is
basically a waste of time. But perhaps someone will have a stroke of
luck in some other place.
Simon Plouffe recently showed such a different path for the decimal
place problem. He has specified an algorithm with whose aid individual
digits of 7r can be found for any base. However, the procedure has the
drawback that the time it takes to calculate is of cubic order and hence
slower than the algorithms for calculating a decimal 7r from scratch.
But even this has already been improved on. The previously mentioned Fabrice Bellard has found out how to arrive at any individual
digit of 7r in any base B in quadratic time. Here is his algorithm [19J:

10.3 Refinements

Algorithm (10.10) Calculation of the nth digit

of 1f (base B)

129

(alg 10.10)

Set N:= l(n+E)log2BJ, where E is a small integer (for example,

20) used to ensure accuracy.

sum:=
0
for (every prime number a from a = 3 to a < 2N)

{
vmax:=
m
v
s

llog(2N)j log(a)J
a vmax

0
0
1
1

b
for (k = l,2,oo.,N)

{
b:= V(~,k)' b mod m
a 2k-l
A := av (a,2k-l) A mod m

v:=v-v(a,k)-v(a,2k-1)
if (v > 0)
{
s := s + k b A-I. avmax-v mod m
}

}
s
sum

s . B n - 1 mod m
sum + -!i'i mod 1

}
At the end of the algorithm, the desired decimal place stands at the
beginning of the portion of sum which is after the base B point.

1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1

11. Arithmetic

Computation of 7r requires three things: a high-performance computer,

an efficient algorithm and fast arithmetic. "Why fast arithmetic'?" you
ask. "Isn't that just the same thing as a high-performance computer'?"
Yes and no. Naturally all computers can add up and multiply. This
is their real domain, and high-performance computers do this too,
albeit more quickly. But they can only work with short numbers, generally no longer than 32 or 80 bits, which is a long way short of what
is required for the 7r problem. This means in effect that every 7r program has to assemble its long arithmetic operations out of the short
operations which are all the computer offers.

11.1 Multiplication
The arithmetic operation which has by far the biggest impact on the
length of time required is the multiplication of long numbers. A program for the calculation of 7r spends almost its entire run time performing either multiplication operations or else "higher" operations
such as division or the calculation of square roots, which are actually
implemented as a sequence of multiplications. By contrast, addition
and subtraction playa relatively insignificant role.
What is the most effective way to calculate the product of two
multiplicands? The answer to this question has been the same for
hundreds of years, namely the same way as ever and as is still taught
in school.
Anyone who learned his arithmetic before the era of the pocket
calculator will know how school multiplication works: multiply the
first multiplicand with each digit of the second multiplicand, write the
product on a line underneath, then do the same for the other digits,
displacing the results by one digit as appropriate, and then add all the
lines of products together, taking into account any carryovers.

132

11. Arithmetic

Because under the school method each digit of the first multiplicand is multiplied by each digit of the other multiplicand, if both
multiplicands are N digits long, then a total of N . N = N2 individual multiplication operations must be performed. The time this takes
rises as a function of N 2 , so that it takes four times as long to multiply numbers which are only twice as long. This characteristic pattern
means that if N is large, multiplication of the two numbers takes a
huge amount of time. Even on a computer which can perform 10 million single multiplications per second, multiplication of two numbers
each a million digits long will take more than a day to perform. In a
high-precision computation with multiplicands that are considerably
longer the time required for the necessary calculations is prohibitive.
But in fact there are faster ways of multiplying numbers, and it is
not even difficult to understand how they work.

11.2 Karatsuba multiplication

The first way to accelerate the process is to use the K aratsuba multiplication method. Although this technique has only been known since
the beginning of the 1960s, it is not clear who invented it. Evidently
it was not invented by the Russian scientist Anatolii Karatsuba, after
whom it is named, as in 1962 he only discovered a similar (and more
complicated) method [77, p. 295]. Never mind.
Suppose we want to multiply together (in base 10) two N = 2k -digit
numbers u and v. (For the sake of simplicity we will assume here and in
what follows that the two multiplicands in a multiplication are always
of the same length and that their number of digits is an integral power
of two. More complicated cases we will leave to the programmers.)
Each of these numbers is divided into an upper and lower half, each of
N /2 digits, and designated as UI and uo, and VI and vo, respectively,
such that u = 10N/2 uI + Uo and V = lO N / 2 VI + Vo.
Under the school method, the product of u and v is calculated as
follows:
(11.1)
This entails 4 partial multiplications: UI VI, UOVI, UI Vo and UOVo. The
factors lO N and lO N /2 do not require any multiplication but merely
indicate simple leftwards shifts by Nand N /2 digits respectively.

11.2 Karatsuba multiplication

133

By comparison, the Karatsuba multiplication constructs the product of U and v as follows:

uv = (ION

+ 1ON / 2 )UIVI + 1O N / 2 (UI

+(10 N/ 2 + l)uovo

uo)(vo - VI)

+
(11.2)

Hey presto! Suddenly only 3 partial multiplications are necessary,

namely for the products UIVI, (UI - uo)(vo - vI) and UoVo. This is a
considerable advance, which is only slightly offset by three additional
additions.
A numeric example will make this clearer. Let N = 4, U = 9876,
UI = 98, Uo = 76, v = 5432, VI = 54, Vo = 32 ..
School Multiplication
9854
7654
9832
7632

5292
4104
3136
24 32
536464 32

Karatsuba Multiplication
9854
Ditto
(98 - 76)(32 - 54)
7632
Ditto

5292
5292
-0484
2432
2432
53646432

In the course of this type of multiplication, three multiplications

are executed with half-length operands. Each of these contains three
further multiplications with quarter-length operands etc. Karatsuba
multiplication thus lends itself well to a recursive call. This will be
seen from the following verbal piece of code.

134

11. Arithmetic

Algorithm (11.3) Karatsuba multiplication

(alg 11.3)

The product of the numbers in U and v is calculated and stored in

r. Variables U and v are of length n, r is of length 2n. n must be an
even number. The upper (left) half of U is called UI and the lower
(right) half is called uo. The corresponding two halves of v are VI
and VI. A temporary store t of length 2n is required.
karatsuba(r, u, v, n)

< limit or n is not an even number)

multiply in the normal way,
else

i f (n

{
call karatsuba(r, u, v, n/2) to calculate UI . VI
and store result in the left-hand half of r.
call karatsuba(r+n, u+n/2, v+n/2, n/2) to
calculate Uo . Vo in the right-hand half of r.
copy r to t
add UI . VI (or t) and Uo . Vo (or t + n)
to r + n/2, i.e. to the middle two quarters of r.
calculate lUI - uol in t and Iva - VII in t + n/2
note down the signs.
call karatsuba(t+n, t, t+n/2, n/2) to calculate
lUI - uollvo - VII in t + n.
if UI - Uo and Vo - VI have the same sign,
add lUI - uol'lvo - VII to r + n/2, otherwise
subtract it from r + n/2.

}
return

}
You will notice the three occurrences of the recursive call of the
karatsubaO function.
The variable limit is at least = 1, but in practice it is often greater
than this. In fact, if the operands become too small, Karatsuba multiplication ceases to be worthwhile and is then replaced by a normal
multiplication operation.
The same applies if the length n is an odd number (either from the
start or after the halving steps). "Awkward" operand length are pure
poison to the performance of the Karatsuba multiplication method,

11.3 FFT multiplication

135

which only works effectively when number lengths are powers of two.
This characteristic it shares with FFT multiplication, which is described in the next section.
Our demo implementation of the method, which you will find in
mult/karamult on the CD-ROM, works like this as well.
As we have seen, the beneficial effect of reducing the number of
partial multiplications from 4 to 3 is propagated in succeeding steps of
the procedure. In this way, Karatsuba multiplication requires a total
of only 3 k short multiplications for N = 2k -digit operands, instead of
the N 2 = 4k under the school method. As a result, the total time is
reduced from a quadratic order N 2 to the order N io g2 3 = N1.58496 ....
Applied to the above example, the multiplication of two numbers both
one million digits long no longer takes a day but a mere 5 minutes!
In view of its simplicity, it is surprising that the Karatsuba method
of multiplication is only 40 years old. Neither the ancient astronomers
nor the excellent mathematicians who lived in the centuries preceding
the invention of the computer, nor even the famous "human computers" of the 18th and 19th centuries seem to have had any method more
sophisticated than school multiplication at their disposal, despite their
staggering mathematical feats. At least, none of these eminent authorities reported otherwise [77, p. 295].

11.3 FFT multiplication

The Karatsuba method may be good, but actually there are faster
methods.
The key to greater speed is to perform a transformation. Transformations often work wonders, not just in mathematics. If a problem
cannot be resolved satisfactorily by a direct route, then, with luck, it
will be easier to solve if it is transformed into a different problem. In
the present case, this proved to be so.
Prior to the multiplication, the multiplicands are first changed, or
rather transformed, in a particular way. An operation which is equivalent to normal multiplication but runs more quickly is then executed
on the transformed multiplicands. The results are subsequently transformed back, producing the desired product.
Naturally the detour should not take longer than the direct route.

136

n.

Arithmetic

Perhaps the easiest example of a helpful transformation is one which

could have been reported hundreds of years ago. Imagine how laborious
it would be for our method to perform even simple multiplications
using the Roman numeral system, e.g. to multiply MMMDCCLXXXIX
by XLIX. Perhaps one day an Arab looked over the shoulder of a
Roman and, surprised to see so much unnecessary work, decided to
help him: he transformed the two starting Roman numbers into their
Arabic equivalents, then multiplied them using the Arabic decimal
system, just as we do today, and then transformed the Arabic result
back into Roman numerals. Anyone can see that this method would
have achieved the desired outcome far more quickly.
No doubt you will remember from your schooldays another multiplication procedure which also works with transformations. This entails using logarithms. Before the pocket calculator won the day, this
method was the preferred aid to multiplication, especially as it was
possible to make things even easier with the famous slide rule.
Under this method, first of all the two multiplicands were turned
into logarithms using logarithmic tables, then these logarithms were
added, and finally the sum was converted back. Anyone who had to do
this will be able to remember his astonishment when the product he
was after actually emerged at the end.
Expressed graphically, the logarithmic transformation goes as follows:
a, b
-+ Take logarithms -+ log a, log b

Multiplication

Addition

t
a

x b

(11.4)

Convert back

log a

+ log b

Instead of proceeding along the direct path from a, b to the product

a x b, this type of multiplication makes a detour via the logarithms of
a and b. Here it is making use of the logarithm theorem, according to

which log( a x b) = log a + log b. The logarithms of a and b are therefore

added, following which they are converted back to ordinary numbers
to give the result.
Instead of proceeding along the direct path from a, b to the product
a x b, this type of multiplication makes a detour via the logarithms of
a and b. Here it is making use of the logarithm theorem, according to
which log( a x b) = log a + log b. The logarithms of a and b are therefore

11.3 FFT multiplication

137

added, following which they are converted back to ordinary numbers

to give the result.
Multiplication using logarithms may be elegant for applications in
which speed is all-important and storage is limited, however it is not
considered, simply because there is a quicker method. This is multiplication using fast Fourier transforms, known as FFT multiplication. It
is the fastest known method of multiplying long numbers and is therefore used in all the big 7r computations. The procedure is described in
detail below.
FFT multiplication also takes a detour involving transformations,
as a means of achieving the desired result in less time. The transformation involved is the Fourier transformation or Fourier transform,
as it is generally called, (FT) and its rapid variant is called the fast
Fourier transform (FFT).
Fourier transforms date back to the French mathematician Jean
Baptiste de Fourier (1788-1830) who in his price-winning essay Theorie
analytique de la chaleur (Analytical Theory of Heat) (1822) proposed
them as a means of analysing series. They have since become a common
and indispensable tool in the natural sciences.
Fast Fourier transformation is generally dated to the year 1965, in
which J. Cooley and J. Tukey [43] presented an FFT algorithm to the
astonished expert world for the first time (see page 198). It was subsequently established that such algorithms had been used before and that
the roots are even to be found in the work of C.F. Gauss (1805) [65].
The commutative diagram of FFT multiplication looks like this:
a, b

--+ forward FFT --+

F(a), F(b)

"School"
multiplication

elementwise
multiplication

axb

(11.5)

+--- backward FFT +---

F (a) . F (b)

In the representation shown above, "forward FFT" corresponds to

the conversion into logarithms and "backward FFT" cooresponds to
the conversion out of logarithms in the previous example (11.4). Both
methods also entail three transformations. These are the two forward

138

11. Arithmetic

FFT's for the two multiplicands and one backward FFT for their product.
The essential ingredients of FFT multiplication are set out below.

Polynomials. A polynomial (of degree N in an Unknown x) can generally be written as P(x) = 'L-r:=oPkxk. The Pk are called the coefficients of the polynomial and are always integers here. Examples of
polynomials are A(x) = 8x 3 + 7x 2 + 8x + 9 or B(x) = 4x 2 - X + 2.
Multiplication of polynomials. We use these examples A(x) and
B(x), and multiply them into a product polynomial C(x):

C(x) =

A(x)
x
B(x)
2
2
= (8x + 7x + 8x + 9) x
(4x - X + 2)
= 32x 5 +28x 4 +32x 3 +36x 2
-8x4
-7x 3 -8x 2 -9x
+16x3 +14x2 +16x +18
3

The general formula for the coefficients Ck of the polynomial C (x) =

'L-~~ocixj, which is the product of A(x) = 'L-!Oaixi and B(x) =
'L-f=o bjx j , is quite simple: Ck = 'L-i+j=k aibj. To convince oneself that
this is correct, one needs only to write out the product for the first few
powers of x and compare the coefficients. The elegant name for the
sequence of Ck'S is the linear convolution of the sequences of ai and bj.
This convolution corresponds to the multiplication in the space of the
real numbers.

N umbers are (almost) polynomials. A number like 6789 can also

be expressed as 6 103 + 7 . 10 2 + 8 101 + 9 100 Generally the decimal number dNd N- 1 ... d2d 1 do is equal to 'L-r:=o dk 10k. The digits of a
decimal number are thus the coefficients of a polynomial P(x) which
for x = 10 has exactly the value of that decimal number. It is because
these coefficients are restricted here to the values 0,1, ... 9, that the
word "almost" is included in the section title.
N umber multiplication = polynomial multiplication + carryovers. Numbers can be multiplied through multiplication of the
corresponding polynomials. However the Ck values cannot be 2: 10.
But since the allowed values of dk are not restricted to O... 9, after
the multiplication carryovers have to be formed from left to right and
resolved.

11.3 FFT multiplication

139

As an example, consider the polynomial multiplication of 67 times

89, during which such carryovers occur in three of the four Ck.

ai, bj
ai x b1:
ai x bo:

67
48

Ck

24

do
d1
d2
d3

11

x 89
56
54 63
110 63
6
3
6

:3

Fourier transformation, forward and backwards. The forward

Fourier transform F(a) of a digit string a = {ao, a1, a2, ... , aN-d is a
digit string of equal length with the components ao, a1, a2, ... , aN-i'
which are defined as follows:

aw :=

N-1

N-1

k=O

k=O

I: ak(cos(27fwkjN) + i sin(27fwkjN)) = I: ake27riwk/N

The backward Fourier transform F- 1 ( a) of a sequence of numbers a

is likewise a number sequence of equal length, but with the components
aw := 'L.r:=r} ake-27riwk/N. The difference here is the initial factor
and the minus sign in the exponent of the coefficients.
The direct implementation of these FT definitions leads to the slow
Fourier transform. It consists of two simple loops. The calculation is
performed with complex numbers.

II
II
II
II

-------------------------sft() (slow) fourier transform

n
length of array f, not nec. a power of 2
isign determines forward (> 0) or backward 0) transform

void sft(Complex *f, long n, int isign)

{

Complex *res = new Complex[n];

const double phO = isign > 0 ? +2.0*M_PI/n

-2.0*M_Pl/n;

for (long w=O; w<n; ++w)

{

Complex t = 0.0;
for (long k=O; k<n; ++k)
{

double phi = phO*k*w;

t += f[k] * Complex (cos (phi) , sin(phi));
}

140

11. Arithmetic

res [w] = t;
}

for (long k=O; k<n; ++k)

f[k] = res[k];

II back transform requires normalization by n

i f (isign < 0)
for (long k=O; k<n; ++k) f[k] 1= n;
delete [] res;
}

There are two different reasons why the above abstract definitions
of the FT, apparently a matter of luck, are the key to rapid multiplication: firstly the convolution theorem and secondly the existence of
fast techniques for the calculation of Fourier transforms.

Convolution theorem. The convolution theorem F(a x b) = F(a) .

F(b) applies to Fourier transformed numbers. It states that the customary multiplication of polynomials (i.e. the convolution of the sequences of coefficients with computation time proportional to N 2 ) is
equivalent to an elementwise multiplication of the Fourier transformed
coefficient sequences of multiplicands, for which the computation time
is only proportional to N. Ck = Li+j=k aibj ~ cy = fLyby or, expressed
otherwise, C = F-I(F(a) F(b)). Here, c refers to the convolution product (or, to be more accurate, the coefficient sequence of the convolution
product) and, by analogy, a denotes the sequence ao, al, . .. , aN-I' The
multiplication implied by the dot must be carried out elementwise. The
last equation is a different way of writing the same thing which is expressed in the right-hand column of the commutative diagram of the
FFT multiplication (11.5). The proof for the convolution theorem is
not difficult, and you can find it in the text fxtrem/fxtrem. * on our
CD-ROM.
Fast Fourier transform (FFT). The time it takes to perform the
described (slow) Fourier transform of a digit string of length N is
evidently proportional to N 2 .
If there were no more efficient method than this, FT multiplication would not get one anywhere. But fortunately the fast Fourier
transform (FFT) is available, along with many variants of it. They all
have in common that the computational time is reduced from being
proportional to N2 to only N 10g2 (N), i.e. only a little worse than proportional to N. (Strictly speaking, the computational effort declines to
the order of magnitude of Nlog 2(N)log2(log2(N)); this was the result

11.3 FFT multiplication

141

which Schonhage and Strassen revealed in their classic paper on the

Fast multiplication of large numbers [104] of 1971.) The basic idea is
the same as the idea behind the Karatsuba multiplication method: one
FT of length N is split into less than four FTs of length N/2 (with
exactly four FTs there would be no gain). Under the Karatsuba multiplication method, the N-digit multiplication has to be replaced by
three N /2-digit multiplications, under FFT by only two FTs of length
N /2. Here too the splitting step is applied recursively until the length
reaches one. However an FT of length one is only a "do-nothing":
0,0 = Lg=o ake2rriwkjl = ao In this way every FT can be calculated
through incessant division into shorter sequences and correct assembly
of the partial results.
Cyclic and linear convolution. A final ingredient is the conversion
of cyclic to linear convolutions. In the formula for the convolution
theorem we omitted to mention that the indices k must be understood
as cyclic modulo N, so it should really be Ck = Li+j=k mod N aibj {:}
Cy = O,yb y . To achieve the required linear convolutions from this cyclic
convolution, one lengthens the sequences ao, aI, ... , aN-I and bo, bl,
... , bN-I by N zeros to ao, aI, ... , aN-I, 0, 0, ... , and bo, bl ,
... , bN-I, 0,0, ... 0. It makes no difference here whether the zeros are
added on the left or right. The cyclic convolution of these sequences
produces the linear convolution Co, CI, ... , C2N-2, C2N-I of the original
sequences.

FFT multiplication in action. The demo program shown below

constitutes a complete multiplication with the aid of fast Fourier transforms.

II
II
II

Fast Fourier Transform (FFT) multiplication in action

#include <iostream.h>
#include <Complex.h>
typedef complex<double> Complex;

II
II

function prototypes:

fft(): fast fourier transform

void fft(Complex f[], long n, int isign);
II getdigs(): read digits from stdin into array a[]
long getdigs(char *s, long base, long a[], long d_max);
void print(const char *bla, long a[]
,long n);
void print(const char *bla, Complex f[], long n);

II

see below

II
II
II

getdigs.cc
print.cc
print.cc

142

11. Arithmetic

int mainO
{

cout

"\nDEMO OF THE FAST FOURIER TRANSFORM MULTIPLICATION\n";

typedef unsigned long Digit;

const long nab_max = 1000;
Digit
a [nab_max] , b[nab_max];
Digit
base
10;
long
na, nb;
while ( 1)

II

II
II

max Digit's of each multplier

Digit's of the multipliers

II

length of the multipliers

main loop

II

1) Read the two multipliers, set-up length values

na = getdigs("Multiplicand ", base, a, nab_max);
if (na == 0) break;
nb = getdigs(IMultiplicator", base, b, nab_max);
if (nb == 0) break;
long nc = na + nb; II length of result array c
long nf;
II length of complex arrays fa, fb

II
II

2) Find nf = smallest power-of-2 >= nc

In case of sft(): letting nf = nc; would suffice.
for (nf
2; nf < nc; nf += nf)
cout

II
II

II

II
II
II
II
II

"Base=" base
", na=" na ", nb=" nb
", nc=" nc ", nf=" nf endl;

3) Copy a[] to complex array far], also be] to fb[].

Justify left, pad with zeros to the right:

Complex *fa = new Complex [nf] ;
for (long k=O; k < na; ++k) fa[k] = ark];
for (long k=na; k < nf; ++k) fa[k] = 0;
print("Complex Multiplicand (zero padded right): II fa, nf);
Complex *fb = new Complex[nf];
for (long k=O ; k < nb; ++k) fb[k] = b[k];
for (long k=nb; k < nf; ++k) fb[k] = 0;
print("Complex Multiplicator (zero padded right): ", fb, nf);
4) Perform fast fourier transforms forward on far] and fb[]
fft(fa, nf, +1);
print("Complex Multiplicand (after forward FFT): ", fa, nf);
fft(fb, nf, +1);
print (IIComplex Multiplicator (after backward FFT): ", fb, nf);
5) Multiply elementwise: fa[k] = fa[k] * fb[k]
for (long k=O; k < nf; ++k) fa[k] *= fb[k];
print("After elementwise complex multiplication: ", fa, nf);
6) Perform fast fourier transform back incl. normalization:
fft(fa, nf, -1);
print("Complex Product (after backwards FFT)
: ", fa, nf);
7) Copy result in far] to a real array c[]:
Digit *c = new Digit [nc] ;
for (long k=O; k < nc; ++k) c[k] = (Digit)(fa[k].real()+0.5);
print("Real Product
: II c, nc);
8) Shift right by one; the least significant digit is c[nc-2]
for (long k=nc-2; k>=O; --k) c[k+1] = c[k];
c[O] = 0;
print("After right shift by one
c, nc);
9) carry operation:
for (long k = nc-1; k >= 1; --k)

11.3 FFT multiplication

143

c [k] I base;
Digit carry
c [k]
carry * base;
c [k-l]
+= carry;

print("Final Product (after carry operation)

cout endl;
delete [] fa, fb;
delete [] c;

11,

c, nc);

return 0;
}

II
II
II
II
II

fast fourier transform

(radix2, decimation in frequency)
length of array f, must be a power of 2
determines forward (> 0) or backward 0) transform

fftO
n

isign

void fft(Complex f[], long n, int isign)

{

double pi = isign >= 0 ? +M_PI : -M_PI;

for (long m=n; m > 1; m 1= 2)
{

long mh = m I 2;
double phi = pi I mh;
for (long j=O; j < mh; ++j)
{

double p = phi * j;
Complex cs = Complex(cos(p), sin(p;
for (long tl=j; tl < j+n; tl += m) {
Complex u = f[tl];
Complex v = f[tl+mh];
f [tl] = u+v;
f[tl+mh] = (u-v) * cs;
}

II data reordering:
for (long m=l, j=O; m < n-l; ++m)
{

for (long k=nl; (!j-=k)&k; k=l)

if ( j > m ) II SWAP(f[m], f[j]);

{;}

Complex t = f[m];
f[m] = f[j];
f[j] = t;
}

if ( isign < 0) II normalise if backwards transform

for (long k=O; k < n; ++k) f[k] 1= n;
}

II --------------------------

144

11. Arithmetic

In the program, the function fit (), which performs the fast Fourier
transform, is used instead of the function sft () (slow Fourier transform) shown on page 139. Its source code is shown at the end of the
program listing. Compared with the slow variants, in the fast variant the length of the complex arrays must be a power of two (the
smallest ~ na + nb), because - as mentioned previously - the speed is
achieved through recursive halving of the multiplicands. After forward
and backward transforms, the results are shifted by one digit towards
the right because the transformation leaves the right-most place empty
(the convolution product of digits i and j moves to position i + j, and
simultaneously the convolution product of the right-most places na-1
and nb - 1 move to position na + nb - 2). Finally the carryovers are
resolved. The other operations in the program should be clear from
the foregoing and from the comment statements within the program
itself.
Comparison. The fast Fourier transform passes its speed on to the
FFT multiplication, in which it is called three times. For this reason,
the time taken to perform an FFT multiplication rises only proportionally N log2 N to the length N of the multiplicands. The effect is
considerable when N is large, as the following graph demonstrates.

Time required

School multiplication

c.
FFT multi lication

N1.583

c' N log N
Operand length

The diagram shows clearly how superior FFT multiplication is to

the Karatsuba and school multiplication methods.

11.4 Division

145

In our example of the multiplication of two one--million-digit (= N)

numbers, the number of individual multiplication operations is reduced
from 1 billion to approx. 20 million (= N log2 N), reducing the time
taken to perform the calculation from more than a day to only 3 seconds. It beats the Karatsuba multiplication method, which itself is
pretty fast, by a factor of around 100. (You can observe these calculation times for yourself if you run our demo program mul t/ speedcmp).

11.4 Division
The conventional method of performing long division as taught in
schools should not be used to divide long numbers u and v. Instead
it is replaced by two operations, namely one multiplication and one
reciprocation, in accordance with the following schema:
u
1
- =u.(11.6)
v
v
We have already dealt in detail with multiplication operations. For
the reciprocation ~ of long numbers, the method still used today is
that devised more than 300 years ago by Isaac Newton (1642-1727)
for determining a zero of a function. As long as certain conditions for
continuity are satisfied, then Newton's method determines a zero of a
function f(x) iteratively with
Xi+l = Xi -

f(Xi)
!'(Xi)

(11.7)

Starting from a starting approximation xo with f(xo) ~ 0, this rule

produces successive values Xl, X2, ... , which tend ever closer to zero.
What is unusual about Newton's method is that it converges
quadratically and on top of that it is self-correcting. At each iteration the number of accurate places of the successor value is doubled,
and it is sufficient here for the predecessor value to be accurate only
to half the number of digits.
In our case, we have to use the function f(x) = v - l/x. Its zero is
the desired inverse value l/v. With the first derivation f'(x) = -1/x2,
the iteration rule runs simply as follows:
(11.8)
If the Xi are accurate to m places, then the successor values Xi+l will
be accurate to at least 2m - 1 places. Normally one continues the

146

11. Arithmetic

iterations until the expression xi(1 - v . Xi) = 0 within the bounds

of the calculation accuracy. Here, the expression (11.8) is calculated
economically just as it stands there, and not as Xi+l = xi(2 - VXi): the
term 1 - v . Xi is so close to 0 that its left half consists of nothing but
zeros, and one therefore only has to multiply with half the digits. An
example will illustrate this more clearly.
Suppose we want to calculate 1/v = 1/7 = 0.14285714 ... to 8
decimal places. Let the initial value be Xo = 0.1. We then obtain:

k
0
1
2
3

Xi

xi(1 - v . Xi)

Xi+l

0.1
0.13
0.1417
0.14284777

0.03
0.0117
0.0011477
0.00009372

0.13
0.1417
0.14284777
0.14285714

Accurate Places
1
2
4
8

8 accurate decimal places are obtained after only three iterations

each involving two multiplications of 1, 2 and 4 digits.
To determine the amount of time required, for the sake of simplicity
we will assume that the desired accuracy of places n of the last approximation is exactly a power of two 2k times the accuracy of places no of
the starting approximation. In that case, k iterations are needed, and
their initial accuracy does not need to be greater than no, 2no, 4no,
... , 2k-1nO' With the known summation properties of powers of two
L::~J 2i = 2k - 1 and the fact that only multiplications with almost
linear time characteristics are used in the iterations steps, in the ideal
case (asymptotic) we obtain for the calculation of the inverse value 1/v
a time requirement of two multiplications of length n times the end
accuracy. At the end there is another such long multiplication, so that
for the division u/v (asymptotic) 3 long multiplications are necessary.
For various reasons it is not possible to perform the long division
in this ideal time when actually implemented. Realistically one would
have to estimate the time at four rather than three long multiplications.

11.5 Square root

As one can see from the modern iterative formulae for the 1f computation, for example the Gauss AGM algorithm, a square root v'd with a

11.5 Square root

147

long radicand d occurs there at least once per iteration, and this has
to be calculated for the full length of the results.
If you can still remember how square roots were calculated at
school, then forget it as there are some far more effective methods
of performing this calculation.

Newton iteration without divisions. This procedure divides the

problem into two steps, rather like in the division, the first being the
calculation of the reciprocal of the square root you are looking for,
I/Vd, and the second being its multiplication by d, since dlVd = Vd.
For 1I Vd, it is then necessary to determine the zero of the function
f(x) = l1x2 - d, and for this Newton's method (11.7) provides the
following iteration rule:
Xi+l

= Xi

+ Xi

1 - d x2
2 z

(11.9)

The iterations normally end when the expression xi(1 - d . xD/2

= 0, within the bounds of the calculation accuracy.
Xi converges quadratically, as in the long division. Once again it is
possible to reduce the amount of calculation effort by only carrying
along the number of digits which are really necessary. Adopting this
approach, it takes about the same time to calculate I/Vd as it does
to perform 3 long multiplications. The calculation of Vd itself entails
another long multiplication.

Coupled Newton iteration. Arnold Schonhage succeeded in arriving at a significantly more effective solution on the basis of an apparently inferior approach [104, p. 257]. His method calculates the square
root Vd directly and ascertains the zero of the function f (x) = d - x2 ,
for which from (11.7) the following iteration rule is derived:
d - x2

Xi+ 1

= Xi + -2-_z
Xi

(11.10)

Each step of this contains an expensive division, which "really" disqualifies the procedure. However, there is a wonderful trick which reduces this drawback so effectively that the total time required is less
than under the first method we mentioned. This is known as the coupled Newton iteration and works with the following double iteration:

148

11. Arithmetic

Algorithm (11.11) Square root by coupled

Newton iteration

(alg 11.11)

Initialise:
Xo :::::; Vd
Vo := 1/(2xo)
Iterate:
Vi+l
Xi+l

+ (1 := Xi + (d := Vi

2XiVi) . 'I7-t-t

1/(2Vd)

X;) . VH!~~

Vd

Instead of dividing, the procedure entails multiplying by an approximate reciprocal Vi+l = 1/(2xi) which is improved before the start of
the current iteration in a coupled iteration from Vi and Xi.
The time required to perform this calculation is asymptotically 7/3
long multiplications, which in reality expand to 3 long multiplications.
For more about this, see the implementation contained in the hfloat
library under hfloat/src/hf/itsqrt. cc
A recursive MuPAD procedure for calculating the square root looks
like this (after [4]):
Sqrt_Coupled_Newton := proc(d. n)
II assume d >= O. n >= 1
local m. g. h. x. e. f. t; II v is static! begin
t := DIGITS;
II save global precision
if n <= 2
then
DIGITS := n;
x := sqrt(d); II Initial value of x
DIGITS := t;
return(x);
end_if;
m := floor(n/2)+1;
x := Sqrt_Coupled_Newton(d. m);
DIGITS := m;
if m <= 2
then
v := 1/(2*x);
else
g := 1 - 2*x*v;
h := g * v; II v from the previous iteration
v := v + h; II now v is of precision m
end_if;
DIGITS := n;
e := d - x-2;
f := e * v;

:=

+ f;

DIGITS := t;
II restore global preCision
return(x); end_proc:

11.6 nth root

149

The variable DIGITS: = is used to specify the degree of precision

required (in decimal places).

11.6 nth root

Some 11" algorithms also use roots with larger exponents, i.e. expressions
of the form \Id with n > 2. This n is usually a constant and moreover
very small, so that one can afford to call a separate algorithm for every
n.
There is an interesting procedure for the construction of such algorithms for any n, whose convergence moreover is user-specifiable.
In every case, the procedure entails calculating the reciprocal of
the desired nth root at the start - as for the square root. The desired
result is then determined via the detour
\Id = 1 . d_1_ . d
(11.12)

\fi5

V'd n - 1

and once the root has been calculated, another multiplication by d is

necessary at the end.
The starting point for this class of algorithm is the identity
1
x

\fi5
with y := 1 - xn D

(11.13)

Converting to a Taylor series produces

_1_ _

vD

nlT'\ -

(1 + It + (1 +
n

(1

n)y2
2 2
n

(1

+ n)(l + 2n)y3
6 3
n

+ n)(l + 2n) ... (1 + (k kInk

l)n)yk

+ ... +

+ ... )

(11.14)

The desired iteration rule for a convergence order k is now obtained

by terminating this series after the (k - 1)th term. Let the resulting
sum be designated <h (n, x). The iteration then goes as follows:

(11.15)

In this way, for example, we obtain an algorithm that converges

quadratically for the reciprocal of a square root (n = 2, D = d):
tJh(2, Xi)
(1 - d x 2 )
= Xi + Xi
2 z

Xi+1 =

(11.16)

150

11. Arithmetic

in fact, precisely the iteration rule (11.9) which we saw in the last
section.
The following table contains the four iteration rules for the cases
n = 2,3 and k = 2,3:

{j~
Order of
convergencE

f/~

D= d'l.
Yi = (1 - xtD)

D=d
Yi = (1 - x;D)
Xi+l = xi(l
Xi+l = xi(l

+ yi/2)
+ yi/2 + 3y; /8)

Xi+l = xi(l
Xi+l = xi(l

+ yi/2)
+ yi/3 + 3y; /8)

In this way we have a number of possibilities for calculating the nth

roots in an effective manner.

11.7 Series calculation

Calculation of poorly converging infinite series is one of the difficult
tasks encountered everyday in working with numbers. The Leibniz
series for 1r / 4 is a particularly glaring case:

~=1-l+~-~+- .. =f)-1)k2k~1

(11.17)

k=O

This converges so poorly that no less than lO n terms have to be calculated for n accurate decimal places.
It is all the more surprising that there is an algorithm which speeds
up calculation of this series enormously. With this algorithm it is possible, for example, to obtain 1000 accurate decimal places from only
1307 summands, instead of a massive 101000. Even better, the algorithm can not only be used for the Leibniz series but also for many
other series. The only firm condition is that the terms of the series
must alternate, i.e. each positive term must be followed by a negative
term, which in turn is followed by another positive term. In addition,
the series terms have to be simple to calculate and the desired precision
of the series sum must not exceed a certain limit, e.g. 1000 decimal
places.
This astonishing procedure was developed by Henri Cohen, F. Rodriguez Villegas and Don Zagier [42]. In their paper, the authors describe a whole class of algorithms for other applications. We will confine
ourselves here to the most elegant of these. It is aptly named sumalt.

11.7 Series calculation

151

The critical discovery lies in the determination of universal coefficients which are independent of the series to be calculated and by
means of which the convergence of the series can be significantly accelerated.
The algorithm approximates the sum s = L:~o ak with alternating
ak through a weighted sum of ao, aI, ... , an-l with universal coefficients Cn,k/ dn . Both Cn,k and dn are integers. For the first n we obtain
the following values:

n
1
2
3
4

dn
3
17
99
577

cn,o
2ao
16ao
9Sao
576ao

Cn,l

Cn ,2

Sal
SOal

32a2

544al

3S4a 2

Cn ,3

12Sa3

The coefficients do not depend on ak, but only on the number of

terms n, and hence on the degree of precision to which the sum is
required to be calculated. For a large class of series ak the relative
error is around (3 + Vs)-n ~ 5.S2S-n, so that for a relative precision
of D decimal places the value n = 1.31 . D is sufficient.
The coefficients Cn,k are obtained through an iteration rule using
an auxiliary variable bn,k.

bn,n-l := 22n -

(ILlS)

Cn,n-l := bn,n-l

(11.19)

(2k

+ l)(k + 1)
+ k)

bn,k-l := bn,k (n _ k)(n

2
Cn,k-l := Cn,k

(k = n - 1, n - 2, ... ,0)

+ bn,k-l

(11.20)
(11.21)

We depart from the original article by calculating the coefficients

Cn,k from k = n - 1 downwards to k = -1, as this makes them easier

to calculate and eliminates one square root calculation. Moreover the

term d n , by which the end result has to be divided, is equal to the last
calculated Cn-l.
In the syntax of the MuPAD computer algebra system, the surprisingly simple algorithm runs as follows:
sumalt := proc(n)
local b,c,k,s;
begin
b := 2-(2*n-1);
c := b;
S :=

0;

II temorary var.
II Cn,k
II

bn,k

152

11. Arithmetic

for k from n-1 downto 0 do

t := (-1)-k I (2*k+1); II here: Leibniz term, general:
s := s + c

ak

t;

b := b * 2*k+1)*(k+1))/(2*(n-k)*(n+k));
c := c + b;
end_for;
s := sic;
I I last Cn,k = dn
return( s );
end_proc:

Just how closely this algorithm approximates the Leibniz series can
be seen from the following table:
n
1
2
5
10
20
30

sumalt(n)
0.666666666666666666666666666666
0.7843137254 90196 0784313725 49019
0.785396636600918492087091551855
0.785398163452432892329702343038
0.785398163397448309920667687680
0.785398163397448309615660848818

Accurate places(D)
0
2
5
9
18
26

4-

0.785398163397448309615660845819

30

7r

In this way D = 26 accurate decimal places of 7r / 4 were determined

with only 30 series terms in the Leibniz series, whereas otherwise this
would have required 1030 series terms. You can tell just how fast the algorithm is from the fact that it took our PC only 2 seconds to calculate
1000 accurate decimal places.

12. Miscellaneous

12.1 A

7r

quiz

One of the few women who is actively engaged in the "7r scene" is Eve
A. Andersson. She has put on the Internet a quiz called "The Pi Trivia
Game" which gives us the "chance to pay tribute to the magnificent
transcendental number that we have all grown to love" .1
The quiz consists of 25 questions from Eve's "7r question database" .
The whole test is too long to print here, but a few of the questions are
provided below:
1. Consider the following series of natural numbers, constructed by

successively larger strings of digits from the beginning of the decimal expansion of the number 7r: 3, 31, 314, 31415, 314159, 3141592,
etc. Out of the first 1000 numbers in this series, how many are
primes'?
e) 58 ?
c) 4
d) 21
b) 34
a) 48
2. What is another name for
a) el numero buono
b) die Ludolphsche Zahl
c) Gesundheit
d) die Eulersche Zahl
e) Drei

7r

in Germany?

3. If one were to find the circumference of a circle the size of the

known universe, requiring that the circumference be accurate to
within the radius of one proton, how many decimal places of 7r
would need to be used?
a) 2 million
b) 39
c) 11
d) 48,000
e) 300

http://eveander.com/trivia/

154

12. Miscellaneous

4. Among the digits of 7r currently known, the concentrations of each

of the digits 0 - 9 are pretty much equal. However, in the first
30 digits of pi's decimal expansion, one number is conspicuously
missing. Which number is it?
a)7
b)2
c)O
d)8
e)6

Solutions:
1.: c), 2.: b), 3.: b), 4.: c).
Most people are surprised at the answer to question 1. Among the
first 1000 initial sequences of 7r there are indeed only four prime
numbers, and they occur right at the beginning: 3, 31, 314159 and
3 1415926535897932384626433832795028841. Whether there are
any other prime numbers greater than 38 digits long at the beginning of 7r is not known; at any rate it is certain that there are no other
prime numbers less than 5000 digits long.

12.2 Let numbers speak

"If you stare at a number long enough, the number will talk back to
you." This conviction of Dario Castellanos is stated in his article The
Ubiquitous 7r [41] dated 1988, a collection of "numerological" findings
which he and others have made over time.
1. The fraction ff~ is known to be a good approximation (to 6 decimal

places) of

v0i ~

7r.

The following approximation for

553
311 + 1

ViF
(12.1)

which is almost equal to the fraction if~ read backwards, is less

well-known.
2. The following approximation of ViF contains a rounded version of
the first 6 digits of 7r (3.141593):

v0i ~ (~) 2 193

14

(12.2)

3. T .E. Lobeck discovered a magical property of 7r. The square on

the left hand side is a standard magic square of the order 5 x 5.
Its lines, columns and diagonals all add up to 65. If one replaces
each value n in the square by the nth digit of 7r, then one obtains
a new number square. In this square, all the line totals also occur
as column totals:

12.3 A proof that

17 24 1 8 15
23 5 7 14 16
4 6 13 20 22
10 12 19 21 3
11 18 25 2 9
Magic square

12.3 A proof that

'IT

1["

2
4
3
6
9
2
7
3
6
5
4
2
1
9
9
4
8
6
3
8
1
3
5
5
3
(17) (29) (25) (24) (23)
Square modified with 7r

=2

155

(24)
(23)
(25)
(29)
(17)

= 2

Look at this sequence of semicircles:

1/2

1
1/2

. 2k
I1m

71"

"2

k-+oo

1["

The semi-circle on the left has diameter 1 and the length of its arc
is
The diameters of the two semicircles immediately to the right
are half as long, but their arcs are i long. Again, the arcs of the
4 semicircles in the middle and the 8 semicircles in the last but one
diagram also total in length. If one continues this an infinite number
of times, the wavy line is eventually transformed into a straight line
of total length i which is equal to the total of all the diameters, i.e. 1.
Hence,

i.

~=1
2

7r

=2

1["

2 k+l =-2

q.e.d.

12.4 The big change

From a FORTRAN manual issued by the Xerox company: "The primary purpose of the DATA statement is to specify names for constants;

156

12. Miscellaneous

instead of writing 3.141592653589793 at every place in the program in

which 1[' occurs, one can assign this value to a variable PI with a DATA
statement and then use that instead of the long form. This also simplifies amending the program in the event that the value of pi should
change."

12.5 Almost but not quite

The most well-known series for
as follows:
1['

1['

~ + ~ - ~ + .. -)

= 4 (1-

is called the Leibniz series and goes

(12.3)

This series is also notorious for its poor convergence which makes it
unsuitable for 1[' calculations. After the nth term the error is still ~. To
obtain 100 accurate decimal places of 1[', for example, ~ would have to
be equal to 10- 100 , i.e. n = 10100 , and that is greater than the number
of particles in the entire universe.
However, the series compensates to some extent for its poor convergence with an unusual property. Sometimes after running past the
first incorrect digit it continues correctly for a good-sized section. Look
at the value which the series produces after 50,000 terms and compare
it with the first correct decimal places of 1[':
50000 (

1)k-1

;k

k=l

whereas

1
1['

=3.14157265358979523846264238327950 ...
11'
11'
11'
(12.4)

= 3.14159265358979323846264338327950 ...

In the results, the 5th decimal place is already incorrect, and this is not
expected to be otherwise in the 50,000th term of the above estimate.
But after this incorrect digit, the next 9 digits are entirely correct!
Then we get another 8 digits correct, and after another incorrect digit
another 8 correct ones. Altogether, only 3 out of the first 33 digits are
incorrect (the ones marked as such above).
R.D. North of Colorado Springs, USA, drew this peculiarity to the
attention of the Borwein brothers. Martin R. Powell had also noticed
this strange feature back in 1982 after 13.5 hours of computer analysis
on his school computer (see [93]). The Borweins and their colleague
K. Dilcher looked into the phenomenon and clarified the mathematical
background in an article [35].

12.5 Almost but not quite

157

The phenomenon is due to the particular form which the residual

sum of the Leibniz series takes, i.e. the sum of all the terms which follow
after the N 12th series term. This asymptotic behaviour amounts to the
following [35]:

_
7f

4L

(_1)k-1 _

2. _ ~

2k _ 1 - N

_ 122

2770 _ 101042

N3 + N5

N7 + N9

++(-1)

m 2 E2m
+ ...
N2m+1

Nll

k=1

(12.5)

J. and P. Borwein, Dilcher,1989

Here the numerators 2, 2, 10, 122, 2770, 101042, ... represent the
Eulerian numbers E2m, multiplied by 2. N must be divisible by 4.
If we set the value 105 for N in (12.5), the first four terms of the
residual sum are then +2 . 10- 5 = 0.00002, -2 . 10- 15 , +1 . 10- 24
and -1.22 .... 10- 33 . However, all this means is that the 5th and
15th decimal places are incorrect by -2 and the 24th decimal place is
incorrect by +1, but the rest of the first 33 places are correct.
In the meantime, the subject has been discussed further in the
literature. In [5] we learn that the asymptotic formula (12.5) could
have been discovered in only one second if the Maple computer algebra
system had been fed this line of code:
asympt(simplifysum(-4*(-1)-j/(2*j-l),
j=1 .. n/2)-Pi) /(-1)-(n/2+1)),n,8);

Before this discovery was made, the Borwein, Borwein and Dilcher
article ranked as an example of the superiority of man over computer,
but this new finding suggests otherwise.
But, to be fair, one must say that the curiosity discovered by North
for decimal places only applies if N is a power of 10. If, for example,
only one single series term is added in (12.4), then the result looks
quite different:
50001 (

1)k-1

4L ;k _1
k=1

= 3.14161265318979923842264278327550 ...

1M'I

11

11

11

1M'I

Now 8 out of the first 33 decimal places are different from

11
1f.

(12.6)

158

12. Miscellaneous

12.6 Why always more?

Why are people always looking for more digits rather than fewer'?
No doubt this was uppermost in the minds of Wei Gong-yi, Yang Ziquiang, Sun Jia-chang and Li Jia-kai of the Academia Sinica Computer
Centre in Peking, who in 1996 published an academic article entitled,
The Computation of 7r to 10,000,000 digits [60]. The article made clear
the trouble which the four authors had had with their calculation and
how they had finally overcome the problems. The article naturally contains a list of references. One of these is another article which sounds
familiar, The Computation of 7r to 29,360,000 Decimal Digits by David
Bailey, 1988.
It would seem that the direction of progress is not always forwards.

12.7

7r

and hyperspheres

A circle has the area 7rr2 and a sphere the volume ~7rr3. In both
formulae 7r appears to the first power.
When considering all multi-dimensional objects, it simplifies matters if one can speak generally in terms of "sphere" and "volume".
Thus, for example, a circle may be understood as a 2-dimensional
sphere and its area as 2-dimensional volume. Thus, the I-dimensional
circle diameter also has a volume which amounts to VI = 2r; 7r does
not occur at all in this expression or only to the Oth power.
How does one calculate the volumes of spheres which have more
than three dimensions ("hyperspheres")? They can be determined
through integral calculus, e.g. using the recursion formula [73, p. 179ff]
(12.7)
This formula takes us from VI to V2, from V2 to V3 and from there
in turn to the volumes of higher spheres which are too complex for us
to imagine:
r

V2

= 21

VI (y)dx

= 21

V3

2ydx

= 41

21 V2(y)dx

= 27r(r 2x

2
21 7ry dx

Jr2 - x 2dx
r

= 21

= 7rr2

(12.8)

7r(r2 - x 2)dx

3
- =-)
3

4 3
= -7rr

(12.9)

12.7

'/r

and hyperspheres

159

Interestingly, it is only in the volume of the 4-dimensional sphere that

our 7l" appears for the first time as a square.
What happens after this'? V5 produces 185 7l"2 r 5, once again with 7l"
to the second power. But in V6 = i7l"3 r 6, 7l" then appears to the third
power.
The general formulae for the volume and the area of a k-dimensional
sphere of radius r go as follows
k/2 k
Vik (r ) -- _1_
(kj2)! 7l" r

(12.11)

n
k/2 k-l
Ok (r ) = (kj2)! 7l" ' r

(12.12)

whereby n! is defined recursively:

n! = n (n - I)! with I! = 1 and (~)! = ~y7r

(12.13)

It will now be clear how the exponents of 7l" are determined, i.e. as 'half
dimension', if necessary rounded to the next lower integer.
This is confirmed in the following list of volumes and areas of selected hyperspheres together with the numeric values in the case r = 1:
Dimensions
k
1
2
3
4
5
6
7
8
9
10
20
100

Volume
Vd1")
21"
'/r1"2
'/r1"3 .

4/3

'/r2 1" 4/2

2
'/r 1"5 .

8/15

'/r3 1" 6/6

'/r3 1"

7 . 16/105
'/r4 r 8/24

'/r4 r 9

32/945

'/r5 r lO /120
lO 20
'/r 1" /1O!
'/r50 r lOO /50!

Surface
Vk(l)

2.0
3.14
4.19
4.93
15.261
5.17
4.72
4.06
3.30
2.55
0.0258
2.10- 40

Dd1")

2
'/r1"/2
2
'/r1" 4
2
'/r 1"3 . 2
2
'/r 1"4 8/3
'/r3 r 5

16/15
7/ 3
'/r4 r 8 32/105
'/r5 r 9/12
lO 19
'/r 1" /10!
'/r50 r 99/50!
'/r3 1" 6 .

'/r4 r

Dk(l)

2.0
6.28
12.57
19.74
26.32
31.01
133.071
32.47
29.69
25.50
0.516
2.10- 38

In the columns with volumes Vk(l) and surfaces Ok(l) for the unit
spheres, it is striking that with 5 and 7 dimensions a maximum occurs

160

12. Miscellaneous

up to which the values continue to get bigger but after which they become smaller and converge towards O. The unit hyperspheres seem to
expand initially as the number of dimensions is increased, and thereafter to contract in on themselves, even though their radiuses remain
unchanged (= 1). "Imagine that if you can" .
Individual sums of these terms are also interesting if one continues
them infinitely. Thus, for example, the volumes of the unit hyperspheres with even dimensions add up to a particularly simple expression,
00

L V k(1) = err

(12.14)

k=l

and the total values of the volumes and areas are very nearly integers.
00

L Vk(1) = err (1 + erf( V7f)) = 45.999326 ...

(12.15)

k=l
00

L Ok(1) = 2(1 + 7ferr (1 + erf( V7f))) = 291.0222 ...

(12.16)

k=l
(;;;)

",,00

where erf ( v 7f = 2 L.Jk=O

(-1) k rr "
k!(2k+l)

Is that an accident?
One might surmise from the formulae for the 3-dimensional sphere
r3 7f) and other 3-dimensional bodies such as the circular cylinder
(r2 H 7f) that 7f generally occurs to the first power in the volume of
smooth 3-dimensional bodies. But in fact this is not the case. For
example the torus ("annulus") has the volume 2Rr2 7f2, and here 7f
occurs to the second power.

(!

12.8 Viete

Wallis

Osler

The history of 7f contains the interesting detail that virtually the two
first infinite expressions for 7f were infinite products, whereas almost all
the other similar expressions are infinite sums. These are the famous
products of Fran<;ois Viete (1540-1603), 1593,

12.8 Viete x Wallis

Osler

161

(12.17)
ViHe, 1593

(1616~ 1703),

and John Wallis

2
7r

1655,

13 35 57 79
22 44 66 88

(12.18)

Wallis, 1655

Although these products are several centuries old (the originals

looked slightly different), the discovery that they are related to each
other and can be united into a single expression is only quite recent.
Thomas Osler demonstrated this in October 1999 with the following
formula [88]:

2
- =
7r

1
2

II

1
2

~+-

n=l

(n radicals)
oo

II

n=l

2p+1n - 1

2P+l n

2p+1 n

+1

.----~-

2P+l n

(12.19)

Osler[88], 1999

Here, each of the two historical products is a special case of a more

general double product. The first part of (12.19) consists of the first
p factors of the Viete product (12.17), whereas the second part is a
"Wallis-like" product. Osler describes it in this way because in the case
of p = 0 it is exactly equal to the Wallis product (12.18) and in other
cases of p it is very similar to the Wallis product, only missing a few
partial products.
The Osler product turns into the Viete product as p tends to infinity
and is equal to the Wallis product when p = O. In the intermediate
cases p = 1, p = 2 etc., we obtain combined Viete- and Wallis-like
products:

162

12. Miscellaneous

_ 0.1 -

p -

'IT

13 . 35 . 57 . 79 . JU.l.
22 44 66 88 1010

11-13 ...
1212

Wallis's original product

= 1.1 =
fi x 35. 79. 11-13. 1517.
'IT

=
P = 2 .1
'IT

V '2

Vfi.
'2

4-4

88

J+
1.
2

1212

1. fi
2 V'2

1616

1921 ...
2020

X 79 . 1517 . 2325 . 3133 ...

88 1616 2424 3232

P=3:~={f.J~+~{f. ~+~J~+~{fx
x 1517 . 3133 . 4749 . 6365 ...
1616

3232

4848

6464

p-+oo:~={f.J~+~{f. ~+~J~+~{f ...

Viete's original product
Every time p is increased by 1, a square root is added to the Viete-like
part and every second factor in the Wallis-like part is removed.
The Wallis-like part in the Osler products can be interpreted as an
error factor of the Viete-like part. One can then see just how accurate
the Viete product is when it is terminated after p factors, as the error
factor is only around 1 + 4- P -1, so that, for example, after 25 factors
we already have 15 accurate decimal places of 1r.
To derive his formula, Osler used the same half-angle functions

for sin x = 2 sin ~ cos ~ and cos ~ =

~ (1 + cos x) which Viete also
used (see page 187) together with the infinite product for sin x =
x n~=1 (1 - n 1: 2 ), which Euler discovered some 150 years after Viete.

12.9 Squaring the circle with holes

Let us stay with the two famous 7r products and now consider a beautiful article by Hansklaus RummIer [99]:
For the individual factors of the Viete product

(12.20)
Viete, 1593

there is a simple geometric explanation. If Un is the perimeter of a

regular 2n-sided polygon which is inscribed in a circle of radius 1, then

12.9 Squaring the circle with holes

163

these factors are equal to the ratio of 2 successive perimeters Un : Un+ 1.

We can now express the right-hand side of (12.20) as follows:

U1 U2 U3 U4
U2 . U3 . U4 . U5 ...

U1

= Uoo

(12.21)

The Viete formula now becomes immediately clear. The inner numerators and denominators cancel out in pairs, leaving the first numerator
U1 and the last denominator U00' U1 is equal to the perimeter of a
digon, i.e. it is equal to twice the diameter 4. Uoo is the perimeter of
a 2O-sided polygon and therefore identical to the circumference 27r,
hence
4
2
U1
(12.22)
Uoo 27r 7r
i.e. equal to the left-hand side of (12.20).
For the Wallis product,
7r
2 . 4 4 6 6 8 8 10
-=--.--.--.--.
4
33 55 77 g.g
=

11 (

- (2n

(12.23)

~ 1)2 )

(12.24)

Wallis, 1655

on the other hand, there is apparently no such geometric proof. But

one can interpret the product via a geometric construction, as follows.
Divide a square of side length 1 into 3 x 3 sub-squares and punch
out the innermost square (left-hand picture):

.............
012345

..

0 .. 7

The area of the figure on the left without the black square clearly
comes to 1 - ~ = 0.888 .... Now take the 8 remaining sub-squares,
divide each one into 5 x 5 squares and once again punch out the middle
of each one. After performing this operation, the total area of the figure

164

12. Miscellaneous

is only (1-~). (1- 2\) = 0.853 ... (centre diagram). In the next step,
each of the remaining 24 squares is divided into 7 x 7 squares and
once again the centre square of each one is punched out (right-hand
diagram).
The process only needs to be continued to infinity to obtain a total
area of 7r / 4 = 0.78539816 .... But significantly earlier than this, after
perhaps 3 or 4 iterations, one obtains an attractive "Wallis pattern".

12.10 An (in)finite funnel

The curve for the equation y = 1/ x is the familiar hyperbola. Consider
the right-hand part, from x = 1 to 00. If this section of the curve is
rotated about the x-axis, the result is a funnel-shaped object whose
volume comes to exactly2 7r.

y=~
x

This may not be particularly remarkable in itself. But how about

this, then? the surface of this funne1 3 is infinite!
In other words, it is possible to fill the funnel but not to paint it in
its entirety.

13. The History of 7r

Mathematical analysis of the circle is one of the oldest challenges to

have faced mathematicians. Statements as to how the circumference
or the area of a circle can be expressed through other variables are
already found in the oldest mathematical documents.
In the beginning, 1[" = 3. This we may assume without reservations
since it corresponds to what one can see by eye. For centuries this value
was sufficient for all practical applications, for example in surveying,
astronomy and architecture. The real history of 1[" only begins when
better approximations than 3 were discovered. This moment in time
varies from culture to culture, but by the start of the second millennium Be such approximations had been developed in several places.
Since then 4000 years of research into 1[" have taken place, and in fact
no single mathematical subject can trace its history so far back into
the past.

The symbol

1r

research may date back to ancient times, but the name "1["" is of
relatively recent provenance. It was only in the 18th century AD that
the concept came to be referred to with the Greek letter 1[". Before
that it was necessary to describe it in a roundabout way, for example
using a description like [114, p. 278] quantitas, in quam cum multiplicetur dyameter, proveniet circumferentia (the quantity which, when the
diameter is multiplied by it, yields the circumference).
The person who invented the symbol 1[" is assumed to have been
the Englishman William Jones (1675-1749), because he used 1[" in the
sense as it is used today in his Synopsis Palmariarum Mathesos, a
compilation of notable mathematical exercises published in 1706. On
page 263 he gives the following statement [68]:
1["

166

13. The History of

7r

... in the Circle, the Diameter is to Circumference as 1 to

16
4
116
4
116
4
"3 - 239 - :3 53 - 2393 + "5 55 - 2395 - ... = 3.14159 ... =

7r

However, neither this nor the other derivations given at this point
in the text are actually the work of Jones, but rather they flowed
from "the accurate and Ready Pen of the Truly Ingenious Mr. John
Machin" (1680-1752), to quote Jones's own Introduction. The symbol
7r must therefore be attributed to the same man who had already
earned his place in history with his arctan formula (5.13) and his record
calculation of 7r to 100 decimal places (see page 192).
Some authors had used the letter 7r even earlier for other variables
relating to the circle. Thus already in 1631 we find it used by William
Oughtred (1574-1660), who was one of John Wallis's teachers. In his
Clavis M athematicae Oughtred used the symbol 7r to denote a length
equal to half the circumference of a circle and the symbol 0 for the half
a diameter when he established the proportion 7 : 22 = 0 : 7r = 113 :
355 [114, p. 302]' so that "our" 7r was the same as "his" 7r : O. 7r was
also used in a similar way in 1669 by Isaac Barrow (1630-1678)[41,
p.91].
For 30 years Jones's or, more likely, Machin's invention was not
repeated by anyone else. Other symbols were tried instead. Johann
Bernoulli (1667-1748), for example, employed the Latin letter c and
Leonhard Euler (1707-1783) used first p in 1734 and then c in 1736.
A revival of the symbol 7r occurred when Euler used it again in
the treatise he wrote in 1736 entitled Mechanica sive motus scientia
analytice exposita. From Euler, "not least due to his extensive correspondence" (Tropfke) with all and sundry, the symbol soon spread
to other mathematicians as well [114, p. 303]. It had finally won the
day when Euler used it in his famous 1748 publication in Latin, Introductio in analysin infinitorum (Introduction to the Analysis of the
Infinite) [52]. In this work, Euler introduced it with the words, "For
the sake of brevity we will write this number as 7r; thus 7r is equal to
half the circumference of a circle of radius 1 or equal to the length of
an arc of 180 degrees." [53, p. 95].
Again, the handy and commonly used term "radius", also appeared
only relatively recently. Contrary to what one would expect, it dates
not from antiquity but first appears in a book published in 1583 [114,
p. 134]. However, from that date on it was virtually the only word used
for the concept. The Greeks too did not call the radius half-diameter,

13.1 Antiquity

167

semi-diameter or another similar term, but instead they referred to

"the straight line from the centre point", and this roundabout description continued to be used long into the Middle Ages by faithful
imitators of the Greek elements.

13.1 Antiquity
Babylon
A clay tablet unearthed in 1936 from the Old Babylonian period, approx. 1900-1600 BC, states that the circumference of an hexagon is
0;57,36 (in base 60) = 96/100 = 24/25 times the circumference of the
circumscribed circle [72, p. 18]. From Uhexagon = 3 d = 24/25 Ucircle =
24/25 7r d we get what is perhaps the oldest approximation to 7r,
7rBabylon

= 3 = 3.125 = 7r - 0.0165...

(13.1)

Egypt

In ancient times it was not absolutely essential to calculate the circumference of a circle, as it could always be measured. On the other
hand it is quite difficult to arrive at an accurate figure for the area of
the circle by measurement, so that it was highly desirable to be able to
calculate it. Problem no. 50 in the Egyptian Rhind Papyrus (named
after A. H. Rhind, a Scot who purchased it in 1858 in Luxor), which
dates back to around 1850 BC, reads as follows [72, p. 18]: "Example
of a round field of diameter 9. What is the area'? Take away 1/9 of the
diameter; the remainder is 8. Multiply 8 times 8; it makes 64. Therefore, the area is 64." The Egyptian scholar was thus using the formula
A = (d - d/9)2 = (8d/9)2. When this is compared with the formula
for the area, A = 7r d2 /4, we obtain the approximation
7rEgypt

(16)2 = 3.16049 ... =

= "9

7r

+ 0.0189 ...

(13.2)

168

13. The History of

7r

How this value came about is suggested by problem no. 48 of the

Rhind Papyrus in which is shown
the figure of an octagon inscribed in
a square of side 9. The author may
perhaps have observed that the area
of the octagon which is 81-4.3 2 /2 =
63 units was somewhat smaller than
the area of the circle, and therefore
added on a small amount to correct
it [72, p. 19].
It is possible that values for 1f either as accurate or even more
accurate were known already several centuries earlier in Egypt. This
is suggested by evidence that as early as 2600 BC aids existed for the
calculation using inscribed irregular polygons. It could be that these
were used to derive good approximations for 1f. For further information
on this point see page 209.

India
The Indian Sulva8utra8 (which means "cord rules", i.e. rules for
building altars of certain shapes with the aid of pieces of cord [51,
p. 102]), which is certainly older than the surviving version dating
from 600 Be [72, p. 4], also describes a calculation of the area of a
circle: "If you wish to turn a circle into a square, divide the diameter
into 8 parts, and again one of these eight parts into 29 parts; of these
29 parts remove 28, and moreover the sixth part (of the one left) less
the eighth part (of the sixth part)." This produces a side length 8 of
the required square of

d~ (7 + 29 (1 - ~ (1 - ~) ) )
1

8=
8

= d 9785
11136

and hence
1flndia =

48

/d

9785
5568 )

= 3.08832 ... =

1f -

0.0532 ...

(13.3)

Another famous approximation to 1f which has been ascribed to the

Chinese may actually also be of Indian origin. This is

13.1 Antiquity

1r

v'1O =

3.16227. . .

169

(13.4)

Tropfke cites a source according to which JIO is already found in the

year 150 BC in the writings of the Indian scholar Umasvati. The author
apparently even believes that he can date this value back to 500 BC
in India [114, p. 277].
Bible

The Bible too mentions a value which is an approximation for 1r, as

indeed one would expect. The architect Hiram of Tyre was commissioned by King Solomon built a circular water reservoir out of ore,
described as a "molten sea". In 1 Kings 7:23 and again using the same
words in 2 Chronicles 4:2, we are told: "And he made a molten sea of
ten cubits from brim to brim ... , and a line of 30 cubits did compass
it round about."
Thirty cubits round about and 10 cubits wide produces a 1r Bible = 3.
This is pretty pathetic, not only when considered in absolute terms,
but also for the time, 550 BC, as more accurate values of 1r had been
known much earlier. Many people have ridiculed the Bible for failing
to come up with anything better, inspiring exegetes to search long
and hard for statements which would present the Bible in a better
light. For example in the 18th century it was declared that the molten
sea must have been a hexagon [114, p. 261]. Whether more recent
endeavours are any better is somewhat unlikely. One example here
is M.D. Stern [111], who concludes from the difference between the
spoken and written versions of the Bible that the actual value of 1r is
obtained through division of both values, giving
333
1rBible = 106 = 3.141509 ... = 1r - 0.000083. . .
(13.5)
This value is accurate to four decimal places, and if it could only
be confirmed, it would certainly silence the mirth at the apparent
inaccuracy of the Bible.
Ancient Greece

All the great Greek mathematicians of the classical era from the 5th
to the 3rd centuries BC worked on problems relating to circles. One
discovery that was especially important for 1r mathematicians was that
of the method of exhaustion. Under this method, which is attributed

170

13. The History of 7r

to Antiphon (c. 430 Be) or Eudoxos (408-355 Be), the area of a twodimensional figure such as a circle can be arrived at through mental
"exhaustion" using ever more elaborate versions of figures whose areas are known, such as polygons. Euclid (c. 330-275 Be) used this
method, for example, to prove the theorem that the areas of circles
are proportional to the squares of their diameters.
Yet none of the eminent Greek mathematicians, from Anaxagoras
(c. 499-428 Be) to Euclid, seems to have improved on the numeric
value of 1f. The Greeks were geometricians; algebra did not interest
them. They were more interested in geometric aspects of circles such
as the problem of squaring the circle. This refers to the problem of
how to construct a square whose area is equal to the area of a given
circle. Several mathematicians had already produced solutions to this
problem when Euclid increased the difficulty of the requirement by
specifying that it must be solved only with an (unmarked) compass
and (unmarked) straight edge. As was only established in 1882, these
additional conditions make the squaring the circle insoluble.
Nevertheless the philosopher Plato (427-348 Be) is supposed to
have obtained what for his day was a very accurate value for 7r, V2 +
03 = 3.146 ... [51, p. 126]. This expression shows that Plato probably
arrived at the formula as the arithmetical mean of the half-perimeters
of the inscribed square 2V2 and the circumscribed hexagon 203. As
these initial values, with 3.464 ... and 2.828 ... , are somewhat weak
approximations for 7r, and yet the end value is quite accurate, one is
attempted to interpret this as a typical case of the non-mathematician
striking it lucky.

13.2 Polygons
After the initial phase largely consisting of empirical approximations,
the Greek mathematician Archimedes of Syracuse (287-212 Be) inaugurated the first theoretical phase in the mathematics of 7r. Using
a new method, he attained a milestone around 250 Be when, for the
first time, he systematically approximated the number 7r and produced
upper and lower limits on its value. In his book The Measurement of
the Circle he stated three theorems about the circle, the third of which
goes as follows:
3. The ratio of the circumference of any circle to its diameter is
less than 3 ~ but greater than 3 i~

13.2 Polygons

171

When expressed as a formula, the theorem states

1
10
(13.6)
3+->7r>3+7
71
and places 7r between the limits 3.14084 ... and 3.14285 ... , which are
accurate to two decimal places l .
"This marks the origin of the value 7r = 3~, which spread victoriously from country to country and from book to book", writes
Tropfke [114, p. 273]. "In Alexandria it quickly replaced the old value of
(16/9?, as it was just as convenient but more accurate. From Alexandria news of Archimedes's value for 7r spread to India, and it has even
been demonstrated to have been known in distant China in the fifth
century AD."
Archimedes's "paper" [6] has only survived in parts; nearly all the
passages in the Doric dialect which Archimedes spoke have been lost.
Nevertheless it has been possible to largely reconstruct Archimedes's
proof.
To prove his theorem, Archimedes calculated the perimeters of two
regular 96-sided polygons such that one of them touched the circle
from the outside and the other from the inside. The perimeter of the
circumscribed polygon is slightly larger than the circumference of the
circle and thus provides an upper limit for 7r, while the perimeter of
the inscribed polygon is slightly smaller, yielding a lower limit for 7r.
Archimedes's 96-sided polygons were obtained by calculating a sequence of polygons in which each successive polygon had twice as
many sides as its predecessor. He began with hexagons, whose geometric properties had been researched in-depth in ancient Greece.
From hexagons, Archimedes then moved on to 12-sided polygons, and
from there to 24-, 48-and finally 96-sided polygons in a total of four
similar steps. The method of calculation entails first of all determining
all the circumscribed polygons and then all the inscribed polygons,
rather than first calculating the two hexagons and then the two 12sided polygons etc. in criss-cross manner. Sometimes it is described
otherwise.
Archimedes obtained the necessary formulae from the figures below:

Today the left- and right-hand limits are generally transposed so that the theorem
is written as 3 +
< 7r < 3 + ~, but in Archimedes's time it was customary
to write it as shown above; moreover it was the upper limit which he calculated
first.

172

13. The History of

iT

Circumscribed polygons

Inscribed polygons

C
D

(CO

+ OA) : CA = OA: AD
OA 2 + AD2 = D02

(13.7)
(13.8)

(AB

+ AC) : BC = AD : DB (13.9)
(13.10)
AD2 + DB2 = AB2

The labels in the pictures are the same as those used by Archimedes.
Lines AG and BG denote the half sides of a circumscribed or inscribed
polygon with n-sides, AD and B D denote the half sides of the "succeeding" 2n-sided polygon; G D and AD thus have the effect of halving
angles AGG and BAG.
The formulae (13.7) and (13.9) can be derived from elementary
geometry, but this is not a trivial task. Archimedes had to calculate
them without trigonometric half-angle formulae, and with little in the
way of formulae to refer to.
The formulae, which enable the transition from polygons of n sides
to polygons of 2n sides, gave Archimedes in the circumscribed as in
the inscribed case two completely symmetrical sequences of numbers
of a6, a12, a24, ... , a96 and b6 , b12, b24 ,., b96 , which are related to
each other as follows:
a2n

= an + bn

b2n

Ja~n +

(13.11)
c2

(13.12)

Writing this using our standard trigonometric notation, which

Archimedes did not yet possess, this implies an = c/tan 0:, bn =
c/sin 0:, 0: = 7r In, where c is an arbitrary scaling constant which helps
to make the calculations easier. Archimedes's two formulae (13.11)
and (13.12) are probably the first recursion formulae in history and
inevitably propelled him from the hexagon to the 96-sided polygon.

13.2 Polygons

173

After deriving these formulae, Archimedes "only" needed to input

an initial approximation and he could then start calculating. The starting ratios OA : AC and AC : CB come in both hexagons to = vis, and
the accuracy of all subsequent calculations depends on the accuracy of
this vis. As Archimedes had no other means at his disposal, he had to
use two approximations, and without a word of explanation he chose
these two:
1351 > V3 265
(13.13)
780
> 153
Scarcely any puzzle has so exercised the brains of mathematics historians as the question of how Archimedes could have arrived at these
preferred values, which are only 2 10- 5 and even 5 10- 7 away from
the true value, which is 1.7320508 .... The simplest explanation [64,
II, p. 51J is that he obtained them from the inequalities

+ 2a >

~
2
V a
b

+ > a + 2a + 1

(13.14)

which were known long before his time, probably already in the time
of the Babylonians. In the above formulation a 2 is the next lower or
next higher perfect square below or above the radicand a 2 + b. This
formula (13.14) produces the following on the first occasion that a = 2
and b = -1 are used:
1

2-->V3>2-7
5
->V3>4

Of the two limits, the left one provides a better approximation but
the right one contains the smaller numbers and therefore Archimedes
could have chosen 5 as an approximation for 3v1s = V27 and used
13.14 once again.
2
2
5 + 10 > 3V3 > 5 + U
26
19
15 > V3 > U
If he now used the 26 from the left-hand side as an approximation
for ff.152 = )675, then after applying (13.14) again he could have
arrived directly at 13.13 via
1
1
26 - - > 15V3 > 26 - 52
51
1351 >
V3 > 1325 = 265
780
765
153

174

13. The History of

7r

In any case, Archimedes took the numerators 1351 and 265 from
(13.13) for the a6, the denominators 780 and 153 for the c, and 2 c
(from "Pythagoras") for the b6 . He then performed four iterations for
each of the 12-, 24-, 48-, and 96-sided polygons. In each iteration he
had to approximate a square root using a fraction, and for this he
evidently used the same effective approximation method (13.14) that
he had used for the initial \1'3. In any case the basic inaccuracy of the
area of the 96-sided polygon instead of the area of the circle remained
greater by factors (8 and 3) than that of the calculation. The mathematical procedure, most of which has survived (only the expressions
in square brackets have been added) can be represented as follows [64,
II, p. 55]:
Inscribed polygons
Circumscribed polygons
b
a
b
c
n
a
c
153 6
1351
1560
265
306
780
2911
571 > [V571 2 + 1532 ] 153 12
< V2911 2 + 780 2 780
< 30131.1.
> 591~
2 4
2
780 1
1162~ [J(1162~)2 + 153 ]153 24 59241.1.
2 4

> 1172~

< V1823 2 + 2402 240

1823

< 1838t1

2334~ (J(1162~)2 +

1532 ]153

48

> 2339~
4673~

153

96

3661 t1
1007

< V10072 + 66 2
< 1009i

2016i < J(2016i)2

+ 66 2

66

< 2017i
IHere the fractions of a and c are reduced to simpler expressions.

At the end Archimedes had, with C96 : a96 and C96 : b96, calculated
approximations for the ratio of polygon side to diameter whose relative
error was less than 4.10- 5 both for the circumscribed and the inscribed
96-sided polygons. Through multiplication of the fractions C96 : b 96 and
C96 : a96 by 96 he arrived at
667~

96153

7r

667~

< 4673.! = 3 + 46731. < 3 + 4672.! = 3 + ;;

2

(13.15)

and
96 . 66
7r

> 20171.

6336

284~

= 20171. = 3 + 2017.!

4 4 4

10

> 3 + 71

(13.16)

13.2 Polygons

175

and thus completed the famous proof of his famous theorem.

Voila! What is the most remarkable thing about Archimedes's Jr
calculation? The accuracy of the result? The fact that it produced two
limits for Jr, which was the first indication that there is no simple value
for Jr7 Or the method of calculation itself, which was quite novel in its
manner and scope 7
Archimedes's book, The Measurement of the Circle - actually a
misnomer as it is in fact concerned with calculations rather than measurements - contains two additional theorems.
The first of these states that every circle has the same area as
a right-angled triangle one of whose sides is equal to the radius
and whose hypotenuse is equal to the circumference. This theorem
Archimedes proved indirectly by showing that the area of the circle
cannot be either greater or smaller than the area of such a triangle,
hence it must be equal to the triangle. His second theorem states that
the ratio of the circle to the square of its diameter is more or less
11 : 14. This is obvious from the third theorem stated above, as is
shown by the relationships 3 + ~ = 2:; = 4

U.

Archimedes's approximation method can in principle be used to

calculate Jr to any degree of accuracy, and in fact up to the 17th century many of his successors did nothing more than calculate polygons
with ever greater numbers of sides. In any case, his method remained
without a serious rival for nearly 2000 years.
Unfortunately, however, Archimedes's polygon method does not
converge particularly well. Even when the lengths of the sides of circumscribed and inscribed 96-side polygons are calculated exactly, the
absolute error of the 7r approximation thus achieved always remains
0.00112 ... or -0.00056 .... Every time the number of sides is doubled,
the error reduces only by around 1/4. When formulated accurately,
regular polygons with n sides produce (at least) l2log lO n - 1.19J correct decimal digits. In this way, for example, 10 I8 -sided polygons are
required to calculate 7r to just 36 places. This approach thus guaranteed a lot of sweat.
Archimedes himself seems to have improved his calculation. Heron
(1st century AD) refers to a work by Archimedes which has since been
lost, in which the limits are stated at 211875 : 67441 = 3.14163 ... and
197888 : 62351 = 3.1738 ... [114, p. 273]. But however accurate the

176

13. The History of 7r

first value is, there must be something wrong overall as both values
are greater than 7r.
400 years after Archimedes's time the Greek astronomer Ptolemy
(150 AD) also came up with the approximation 3.1416. It may be
that he discovered this approximation by simply deriving the approximate value 38'30" = 3
from the sexagesimal representations of
Archimedes's limits, 3~ = 38'34.28" and 3 ~? = 38'27.04" [114,
p. 275J, or else he may have separately summed the numerators and
denominators from their fractions 3 + 1/7 = 154/49 and 3 + 10/71 =
223/71 to arrive at the mean value of 377/120 [51, p. 127J.

No

Rome
As far as we know, the Romans did not make any further progress in
the mathematical analysis of circles. As late as 15 BC no more precise
a value than 7r = 3 appears to have been either required or known. Surveying in Roman times was evidently not particularly accurate [114,
p.277J.

China
When Wang Mang came to the throne at the end of the Western
Han dynasty (206 BC-24 AD) he commissioned the astronomer and
calendar expert Liu Xin to develop a standard measurement for his
empire. Liu Xin's response was to produce a cylindrical vessel made
of bronze. It is assumed that around 100 copies of this were made and
distributed throughout the country. One of these has been preserved
in the Peking Palace Museum. From analysing it, historians have concluded that Liu Xin must have possessed a 7r approximation which
was significantly better than 3 and may even have been the accurate
value of 3.1547. This would then have been around the beginning of
the Common Era.
This and also everything else we know about the history of 7r in
China comes from an paper [78J by Lam Lay-Yong and Ang Tian-Se,
most of which has been confirmed by [51].
Around a century later Zhang Heng (78-139 AD) improved the
value of 7r experimentally in the following manner. In his day the ratio
of the area of a square to an inscribed circle was assumed to be 4 : 3.
By analogy, Zhang Heng calculated that the ratio of the volume of a
cube to that of an inscribed sphere was its square, i.e. 4 2 : 32 = 16 : 9.

13.2 Polygons

177

As he probably realised that this was too large, he corrected it to 8 : 5.

When calculating in the other direction, to be consistent he took the
root J8 : y'5 for the circle : square ratio and from this he calculated
7r at Vi6 = 3.1622 ....
In the third century, Wang Fan (217-257 AD) came up with the
circumference: diameter ratio of ~~ = 3.155 ....
He was followed by two important scholars whose 7r calculations
were systematic and produced better results. The first of these was
Liu Hui c. 263 AD. He started with a circle of radius 10 and, using
Pythagoras's theorem, he calculated the areas of inscribed polygons
starting with the hexagon and proceeding upwards to polygons with
192 sides. His calculation ended with the following inequality:
64
169
= A 192 < A < A96 + 2(A 192 - A 96 ) = 314
(13.17)
314
625
625
Here, A refers to the area of the circle and An to the area of an inscribed
polygon with n sides.
E
D
F
Liu Hui obtained his upper limit from
the inscribed polygon with double the
number of sides, unlike Archimedes
who had obtained his upper bound
from the circumscribed polygon with
o
the same number of sides. The expression 2(A2n - An) means n times rectangle AEFC.
At the end of his calculation Liu Hui then established the following
approximation from his two limits, which differ by ~~~:
36
A 192 + (13.18)
625
From this he then arrived at the following approximation for 7r
314+ ...!.
7r ~
(13.19)
102 25 = 3 . 1416

Liu Hui, 263 AD

which differs from the true value of 7r by 7.346 10- 6 .

Liu Hui did not give any justification for the relationship (13.18),
so we can only guess as to how he arrived at it. No doubt it was clear
to him that the true value he was looking for was closer to the 192sided polygon than to the 96-sided polygon, so he probably used a

178

13. The History of 'iT

weighted mean instead of the arithmetic mean. In this way he may

have divided the difference 169 - 64 = 105 by 3 instead of halving it,
which would have resulted in the approximation 314 + 99/625 (which
possesses almost the same error, but with the opposite sign). We can
well understand how the scholar was unwilling to leave things there:
the fraction ~i5 "is crying out" for a small manipulation which would
enable it to be replaced by the fraction ~g~, which is much easier
to shorten and produces 2~' This was the final crowning touch to a
formula (13.19) which is both accurate and easy to use.
Liu Hui's method is similar in many respects to that of Archimedes
(see pages 171ft'.): both mathematicians used the exhaustion method
applied to polygons and started with a hexagon, proceeding from there
in a number of stages in which the number of sides was doubled. And
both men were the only 7r mathematicians of ancient times to calculate
two values for 7r, namely an upper and a lower limit,
Unlike Archimedes, Liu Hui only used inscribed polygons and calculated their areas rather than their perimeters. But above all, he used
a decimal place system which he had inherited from his ancestors and
which by his day already possessed the decimal zero. Although he subsequently reverted to fractions, he was able to dispense with fractions
during the calculations. This meant that the calculation process was
significantly simpler and serves as a particularly good example of the
capability of the Chinese in ancient times to handle large numbers.
Liu Hui was, for example, aware that he needed to calculate radicands
from square roots to 10 decimal places accurately in order to obtain 5
correct decimal places in the result. Nevertheless, the high accuracy of
(13.19) is not due to his use of decimal numbers but to the fact that
the mean of the upper and lower limit was almost optimally weighted.
(That this was the case was first established in 1621 by Wildebrod
Snell, see page 183.)
The second important ancient Chinese 7r scholar was Tsu ChhungChih (or Zu Chongzhi) (429-500). He improved the precision of 7r from
Liu's value by around 2 powers of ten to the limits

3.1415926 <

7f

< 3.1415927

(13.20)

This interval, with its 7 accurate decimal places, held the world
record for 800 years, although to all intents it remained a secret as
even in China the two limits were not rediscovered until the fourteenth
century.

13.2 Polygons

179

Another interesting feature of the result is that Tsu Chhung-Chih

was one of the first to express these numbers using decimal notation,
when he wrote [116, p. 312]:
3 zhang, 1 chi, 4 cun, 1 fen, 5 li, 9 hao, 2 miao, 7 hu
In this expression, zhang, chi etc. are all units of length which behave
like 1 : 10 : 100 : ....
If Tsu used the same method of calculation as Liu, as seems likely,
he used polygons with 12,288 sides to calculate 7r and had to record
his interim results with up to 13 decimal places.
Tsu also found a second value which was numerically less accurate
but more visually attractive. This is as follows:
7r:::::J

355
113 = 3.1415929 ...

(13.21)

This decimal fraction, which is accurate to 6 decimal places, agrees

with the fourth partial quotient of the simple continued fraction of
7r (see page 54), but it was definitely not obtained in this way. It is
assumed that he obtained the fraction from two less accurate values
which were then combined fortuitously to produce an especially good
mean.

India
In 499 AD the astronomer Aryabhata (born 476 AD) wrote a work
called Aryabhatiya. His theorem 10 (out of 133 theorems) goes as follows [72, p. 203]:
"Add 4 to 100, multiply by 8 and add 62,000. This is the approximate circumference of a circle of which the diameter is
20,000."
This produces a value for 7r of 62832/20000 = 3.1416. However, this
value is suspected of being of Greek origin, possibly even originating
from Archimedes, and of having found its way to India like many
other things. But Aryabhata may have calculated it himself, for he
demonstrated that if a is the length of the side of a polygon with n sides
which is inscribed in a circle of diameter 1, and b is the length of a side
in a polygon with twice as many sides 2n, then b2 = (1 - ~) /2.
Beginning with a hexagon he calculated the side lengths of polygons
with 12, 24, ... , 384 sides. The perimeter of the last polygon he stated

180

13. The History of

7r

as )9.8694, and from this he calculated the previously mentioned value

of 3.1416 by approximation of this root [14, p. 341].
The Hindu Brahmagupta (born 598 AD) was fascinated by the
discovery that the perimeters of regular polygons with 12, 24, 48 and 96
sides and diameter 10 are )965, )981, )986 and )987. He concluded
from this that if the number of sides were doubled again, the perimeter
would tend towards )1000. Using this logic he arrived (c. 650 AD) at
7'1 = )1000/10 = JIO = 3.16227 ... [41, p. 68].

Choresmia (today Uzbekistan)

In c. 830 the Choresmian Alkarism made his mark on the history of 7'1
by coming up with 3 values of 7'1,22/7, JIO and 62832/20000. The first
of these was intended to serve as a mean value, the second was meant
for geometricians and the third for astronomers. We also have Alkarism
or Abu Ja'far Mohammed ibn Musa al-Khowarizmi to thank for our
present-day word algorithm, which is based on his name. He wrote a
famous book entitled Kitab al jabr w'al-muqabala; and it is from the
title of that work that our word algebra [77, p. 197] is derived.

Middle Ages
In Europe in around 1220 Leonardo of Pisa (1180-1240'1), better
known by the name Fibonacci, independently calculated the value
864/275 = 3.14181 ... from a 96-sided polygon, without drawing on
Archimedes's work. His approach was more accurate than Archimedes's,
but not as systematic, so according to Tropfke his 3 correct decimal
places owe more to chance than to his calculation method.
It is said that Dante Alighieri (1265-1321), author of "The Divine Comedy", possessed the approximation 7'1 = 3 + )2/10 =
3.14142 ... [41, p. 68].
At any rate it was Dante who composed that secular verse which
could aptly summarise all the endeavours of 7'1 researchers 2 :
As the geometer his mind applies
To square the circle, nor for all his wit
Finds the right formula, howe'er he tries, ...
2

Paradise, Canto XXXIII, lines 133-135, translated by Dorothy Sayers and Barbara Reynolds.

13.2 Polygons

181

Nicely put indeed. But unfortunately in this stanza he is not talking

about the calculation of 7r but about the "squaring the circle" , which
in Dante's day was still viewed as an unresolved problem rather than
one that is intrinsically insoluble.
Leonardo da Vinci (1452-1519) came up with a "mechanical" approximation. The inventor of many machines once rolled a circular
cylinder whose height was equal to half the radius of the cross-section
for one revolution. The area covered while such a cylinder is rolling
takes the form of a rectangle and can be measured easily. It is equal
to the area of the cross-section of the cylinder, i.e. the area of a circle,
7rr2 [26, p. 27].

All in all, the European Middle Ages saw little progress in the matter of 7r. Thus in 1464 the man regarded as the greatest European
mathematician of the fifteenth century, Regiomontanus (1436-1476),
whose real name was Johannes Muller, used a value of 3.14243 for
7r. And Adrianus Metius (1571-1635), rated equally important, was
using nothing better than Tsu Chhung-Chih's approximant 355/113
which the Chinese mathematician had discovered 1000 years earlier,
and even this he only found through a happy accident when he calculated the mean values of the numerators and denominators of two
limits (377/120 and 333/106) [14, p. 343]. At a time when the mathematicians themselves could not produce anything better than this, we
should praise the approximation of 7r = 3i = 3.125 which the artist
Albrecht Duerer (1471-1528) used, 3500 years after the Babylonians
had first invented it. This value can at least be well constructed with
a compass and ruler.

182

13. The History of

7r

By far the best approximation to 7f for 180 years was calculated by

the Persian astronomer of the observatory in Samarkand, Al-Khashi,
c. 1430. He calculated the following value for 27f using polygons with
3 . 2 28 sides [77, p. 198]:
27f

= 6.2831853071 79586 5

This value is accurate to 16 decimal places.

While Al-Khashi may in principle have used the same method
as Archimedes, he performed his calculations using a different approach [69, pp. 314-319]. Especially impressive is the way he continuously monitored his results for errors, which allowed him to ignore
redundant decimal places. Al-Khashi performed his calculations using
the sexagesimal system, but then transformed his 27f to the decimal
system.

The modern era

As far as 7f research is concerned in Europe, the sun did not rise again
until the end of the sixteenth century, when it did so in most spectacular fashion. In 1579 the lawyer and amateur mathematician Fran<.;ois
Viete (1540-1603) discovered 9 decimal places for 7f. In so doing, he
combined Archimedes's model with the principles of trigonometry,
which had in the meantime become established, to calculate polygons
with 3.2 17 = 393216 sides. His results are as follows [114, p. 284]:
. 180
180
n . sm < 7f < n . tan n
n
and with n = 393216,
3.1415926535 < 7f < 3.1415926537

To find the values of the sine and tangent, Viete applied the formula
2 sin 2 () /2 = 1 - cos () [14, p. 343] iteratively.
Several other Dutchmen then went on to leave their mark on the
history of 7f. The first of these was Adrian van Romanus (1561-1615),
who in 1593 achieved distinction by calculating polygons with one billion sides in order to arrive at 15 accurate decimal places. Three years
later, in 1596, he was surpassed by Ludolph van Ceulen (1539-1610), a
professor of mathematics at the University of Leyden. At first Ceulen
calculated only 20 digits, for which he used the perimeters of the inscribed and circumscribed regular polygons of 60 . 233 ~ 480 billion
sides.

13.2 Polygons

183

He continued his calculations afterwards, but did not publish them

himself. A posthumous publication from the year 1615 specifies 32
decimal places derived from polygons of 262 ::::: 4.10 18 sides and in 1621
as many as 35 places were attributed to him. These 35 places were
very likely carved on his tombstone (which has since been lost). But
Ceulen's record lived on, with the result that up to the time of the
First World War 7r was still known in Germany as the "Ludolphian
number".
Shortly before the polygon method of calculating 7r was replaced
by the series method it underwent yet one last significant improvement. In particular, Wildebrod Snell (1581-1626) succeeded in 1621
in obtaining 34 digits of 7r from 23 -sided polygons, whereas Ludolph
van Ceulen, using the same method as Archimedes, had only managed
16. Snell divided every sector of the polygon one more time by 3 and
constructed from each polygon side two further sides which contained
the arc more accurately. His method was based on the inequality [14,
p.344]
3 sin 0
. 0
0
(13.22)
-:-------=-O < 0 < 2 sm - + tan2 + cos
3
3
The lower limit had already been discovered by the German cardinal
Nicolaus Cusanus (1401-1464) [114, p. 281], while the upper limit was
discovered by Snell himself. The accuracy of the procedure can be
easily inferred today from the series expansions of sine and tangent.
For it appears that the third order terms in the expression 2 sin x+tan x
drop out, resulting in a quadratic error improvement. Moreover, the
upper limit of Snell's inequality (13.22) confirms the correctness of Liu
Hui's (intuitive) approach in his calculation of 7r 1400 years earlier (see
page 177).
It was using Snell's method that Grienberger, the last mathematician of the polygon era, which had lasted almost 2000 years, calculated
7r to 39 decimal places in 1630.
The seventeenth century saw the discovery of other alternatives to
Archimedes's method. Whereas under Archimedes's method, first of
all the perimeters of all the circumscribed polygons are calculated and
only then are the perimeters of all the inscribed polygons calculated,
the new approaches involved working in "criss-cross" fashion: first of all
a pair comprising one inner and one outer polygon is calculated, each
possessing n sides, and then a similar pair of polygons with 2n sides
is calculated, followed by a further pair with 4n sides etc.

184

13. The History of

7r

Such an approach was developed by James Gregory (16387-1675),

whom we will meet again as one of the originators of the Leibniz series
(see page 188). Continuing in the style of Archimedes, he held the
radius of the circle constant over all the iterations, but calculated in criss-cross fashion - the areas In and in of the circumscribed and
inscribed polygons instead of the perimeters. Beginning with a square
and the radius J2, he repeated his procedure, which results in linear
convergence on 7r, as shown below:
_ 2 i2n' In
I 2n - .
Z2n + In

(13.23)

Gregory, c. 1667

The left-hand formula had already appeared in the Middle Ages, in

the writings of Jordanus Nemorarius (-12377), but Gregory is unlikely
to have known of Nemorarius's work.
Another innovation was subsequently made by Rene Descartes
(1596-1650), who came up with an idea which Cusanus had already
tried 200 years earlier, but without success: every time one moves on
to the next iteration, it is no longer the radius of the circle that is held
constant but the perimeter of the polygons. From a square of perimeter P an octagon with an equal perimeter is constructed, and from this
in turn a 16-sided polygon with an equal perimeter etc. [114, p. 289].
As the number of sides increases, the radiuses of the circumscribed
and inscribed circles in each case converge from above and below towards the common value of P/(27r). Beginning with the square and
the perimeter P = 8, Descartes's recursion rule runs as follows:
R4

= J2, r4 =

1; r2n

1
"2(rn

+ R n ),

R2n

Jr2n . Rn

(13.24)

Descartes, c. 1649

Descartes's procedure converges linearly towards the value 4/7r.

At this point we shall take a detour 160 years forward in time
to 1809, when Carl Friedrich Gauss (1777-1855) came up with the
following, almost identical recursion rule which also converges towards
7r, but a lot more quickly3:
R4

= J2, r4 =

1; r2n

1
"2(rn

+ R n ),

R2n

Jrn . Rn

(13.25)

Gauss, 1809

Strictly speaking, the R2n and r2n converge towards the value ..jhn . 7r with
14 = 1/2; hn = In - %(R~n - r~n)' This results from the formula (7.1) on
page 87

13.3 Infinite expressions

185

It is difficult to believe, but the small difference of rn instead of r2n

in the formula on the right-hand side actually results in quadratic
convergence. The recursion no longer produces a constant improvement
on every step but instead it delivers each time twice the improvement
achieved during the previous step, yielding, for example, not just 6
accurate decimal places of 7f after 10 recursion steps but over 1000. A
mathematical miracle!
Whereas Descartes's method is based on elementary geometry,
Gauss's iteration rule is derived from the quite different theory of
elliptic functions. The fact that the notation rand R is used in both
formulae should not be taken to mean that the Gauss formula refers,
like the Descartes formula, to the radiuses of circles.
The details, background and the story of how the Gauss algorithm
came about are discussed in depth in Chapter 7, where the regrettable
historical accident which resulted in this algorithm being overlooked
for one-and-a-half centuries, unti11976, is also discussed.

13.3 Infinite expressions

The second era in the calculation of 7f began with infinite products and
infinite series (sums). After almost 2000 years during which mathematicians had had only geometric methods at their disposal, these
provided them at last with an analytical tool.

Beginnings in India
The first series for 7f were noted down in India in the fifteenth century,
possibly even earlier. They are thus at least 100 years older than the
first European series (1593). This fact might have been known for 100
years, but the essay by Charles M. Whish on this subject dating from
the year 1835 On the Hindu quadrature of the circle and the infinite
series of the proportion of the circumference to the diameter exhibited
in the four Sastras, Tantra Sangraham, Yukti-Bhasa, Karana-Paddhati
and Sadratnamala was overlooked for 100 years. It was not until around
1940 that this essay resurfaced and its essentials were confirmed.
No less than eight 7f series are found in the Sanskrit texts YuktiBhasa and Yukti-Dipika, among them, for example, the two following
series, the first of which begins like the Leibniz series:

186

13. The History of 7r

7r

1
p-1

p/2

-~l--+---+"'=r----

p2+1

(13.26)

Nilakantha. 15th century [89]

444

7r

= 3 + 33 _ 3 - 53 _ 5 + 73 - 7 - 93 - 9 + ...

(13.27)

Nilakantha, 15th century [89]

All eight of the series are included in our collection of formulae, items
(16.6) to (16.13).
We are indebted to S. Parameswaran [89] and Rinjan Roy [98] for
making us aware of them. According to Parameswaran, the origins
of these series are complicated, especially as they were passed down
orally. The Yukti-Dipika was composed by Sankaran (c. 1500-1560),
who claims to have received directions from Jyestha-devan, who for
his part wrote down the Yukti- Bhasa. This contains detailed proofs
for the series as well as the series themselves. These two scholars in
turn received instruction in astronomy and mathematics from Kelallur Nilakantha Somayaji (c. 1444 - c. 1545) [sic!], the author of the
work Tantra Sangraham. If the series were not already composed, then
the latest person who could have written them is Nilakantha. Recent
research suggests that some of the series could actually have been formulated earlier and, in particular, some of them are linked with the
name of Madhavan, who lived from c. 1340 to 1425.
Even more surprising, if that is possible, than the great age of these
Indian series is the fact that they are quite simply better than the first
European series.
The author Nilakantha provided a residual sum along with the series (13.26). This gives an estimate of the order of magnitude of p-5.
(See formula (12.5)). As far as we know, Europe had to wait another
hundred years for an estimate that would be as accurate as this, until
Euler produced one in 1750.
The second series (13.27) is significantly better than the first
(13.26) as regards convergence. For example, the first series required
1010 terms to produce 10 correct decimal places of 7r, whereas the
second required fewer than 103 .
A third series (16.6) is even more convergent, in that 28 terms
produce as many as 16 accurate decimal places.

13.3 Infinite expressions

187

Infinite products

In Europe, Fran<;ois Viete, whom we mentioned above, produced an

infinite expression, or, to be more precise, an infinite product. in 1593
he achieved this feat, remarkable both ;'historically and aesthetically,
in his work Variorum de Rebus Mathematicis [115]:

(13.28)
Viete, 1593

The product is also numerically interesting as it converges very

quickly to 1[', generating 15 accurate decimal places after only 25 terms.
Viete arrived at his formula through a simple but cunning trick.
He applied the standard formula sin(x) = 2sin(x/2) cos(x/2) recursively and thus obtained sin(x) = 2n sin(x/2 n ) cos(x/2) ... cos(x/2 n ).
As n becomes bigger, 2n sin(x/2n) converges to x. To calculate the
cos expressions, Viete used another standard formula cos(x/2) =

J~

(1 + cos x). He then only needed to substitute 1['/2 for x, i.e. to

set sin(x) = 1 and cos (x) = 0, and - hey presto! - there was his clever
formula.
Curiously in the history of 1[', the Viete product was followed by
another product. This is curious because apart from this product and
Viete's product, there are few other infinite products for 1[' but a large
number of infinite series. In any case in 1655 John Wallis (1616-1703)
reported the following infinite product in his Arithmetica Infinitorum,
derived by successive interpolations:

i
1['

3 x 3 x 5 x 5 x 7 x 7 x 9 x 9 x 11 x 11 x 13 x 13 .. (13.29)
2 x 4 x 4 x 6 x 6 x 8 x 8 x 10 x 10 x 12 x 12 x 14 ..
Wallis, 1655

Of course no mathematics student today would be allowed to write

the product in this way because it results in : . To show the unconditional convergence, two numerators and two denominators must each
be combined together: ~ = ~:~ . ~:~ ... [114, p. 292].
One can say that Wallis computed the Integral f01 xmykdx (with
x 2 + y2 = 1) for integral values of m and k [20, p. 68-80]. Wallis's
derivation is historically significant as it inspired Newton to a similar
approach which ultimately led him to the binomial series.

188

13. The History of 7r

Even if the Wallis product can be calculated easily, it cannot be

used for effective 7r calculations because after 100 terms, for example,
it is only accurate to 7.8 . 10-3.
William Viscount Brouncker (c. 1620-1684) was the first President
of the Royal Society and "manipulated" (Beckmann) the Wallis product into a continued fraction for 7r, the first in the history of 7r, in 1658:
4

:;;: = 1 + ----3-;2:---

(13.30)

2+--~-

52

2+--2+ ...
Brouncker, 1658

According to Beckmann [18, p. 131J we can only guess as to how

Brouncker arrived at this continued fraction. Leonhard Euler was the
first person to derive it systematically in 1775, i.e. over 100 years later,
from the series expansion of arctan 1.
At the time when the first series were discovered, infinitesimal analysis did not yet exist. This explains why the mathematicians concerned
found the derivations so difficult.
Infinite sums

The independent discoveries by Isaac Newton (1643-1727) and Gottfried Wilhelm Leibniz (1646-1716) of differential and integral calculus
in the second half of the 17th century heralded a boom in series formulae for 7r.
Some time in 1665-1666, Isaac Newton himself succumbed to the
temptation to calculate 7r with the aid of one of the new series. He used
his Method of Fluxions and Infinite Series to derive a quite effective
series for 7r (see (16.63)) which is essentially an expansion based on
the arcsin function. Using this, he obtained 15 digits of 7r (which, of
course, was no world record at that time). Newton later wrote, "I am
ashamed to tell you to how many figures I carried these computations,
having no other business at the time" [32, p. 339J. His words show only
too well the two minds which characterise many 7r "digit hunters" .
The date 15 February 1671 is of historical importance as the date
on which James Gregory (1638'1-1675) stated the first series for arctan
(without explaining how he derived it), in a letter [20, p. 89J. If r is

13.3 Infinite expressions

189

the radius, a the arc and t the associated tangens, then the Gregory
series reads as follows [114, p. 293]:

t3
a = t - 3r 2

t5

+ 5r4

t7
- 7r6

t9

+ 9r8

- ...

which we write in modern form (r = 1):

x3 x5 x7 x9
arctan x = x - - + - - - + - - ...
3
5
7
9

(13.31)

(13.32)

Gregory himself did not perform the easy step of substituting 1 for
x which leads to the so-called LeibnizSeries:
1["
1 1 1 1
4 = 1 - "3 + "5 - "7 + 9 - . . .
(13.33)
Leibniz derived this series independently of Gregory and cited it in
letters to several friends from 1674 onwards. In the meantime we know
that the same series was already known much earlier in India (see
page 185).
The first person to make use of the Gregory series to obtain an approximation for 1[" was the astronomer Abraham Sharp (1651~1742). In
for x and in this way discov(13.32) he substituted tan (1[" /6) =
ered a series (see (16.67)) which produces about 1 digit per 2 terms.
This he used in 1699 (almost 40 years after Gregory's invention) to
calculate 1[" relatively easily to 71 correct decimal places. This was the
very first world record that was not based on Archimedes's polygon
method.

Jl73

Leonhard Euler

18th century mathematics was dominated by the Swiss mathematician

Leonhard Euler (1707~ 1783). His almost inconceivably extensive life's
work, consisting of over 700 books and essays, which take up 80 (still
not completely published) volumes, attests to his status as the most
versatile and prolific mathematician in the history of mathematics.
His creativity and productivity were not impaired in the least by the
persistent problems he suffered with his eyes, which resulted eventually
in complete loss of his sight.
Even though he is sometimes identified more with the number e,
Euler has the distinction of having decisively increased our understanding of 1[". He came up with an exceptional number of formulae for

190

13. The History of 1':

including the one which is commonly rated as the most fascinating of all time, because it unites in one formulae 5 basic quantities of
mathematics:

1f,

(13.34)

You would not believe how much has already been said and written
about this formula! In the introduction we quoted the enthusiastic
words of Benjamin Peirce, but he was only one of many who have
marvelled effusively at the formula. Only a few years ago (1990) it
was voted the "most beautiful of all" by the readers of Mathematical
Intelligencer [120]4
Euler did not actually express "his" formula in this way but like
this (1743, in Miscellanea Berolinesia [41, p. 89]):

eiz = cos z + i sin z

(13.35)

Euler, 1743

which through z = 1f turns into (13.34) because cos 1f = -1 and sin 1f =

O. This expression was not entirely new. Roger Cotes (1682-1716) had
already published the equivalent equation in 1714

Rx =

log (cos x + RSinx)

(13.36)

Cotes, 1714

albeit using a different notation [39, p. 481]. Finally, mathematicians

like writing Euler's theorem somewhat cryptically, for example
1f

= -i In( -1)

(13.37)

David Wells, commenting on the results of the reader survey quoted

above, wondered whether small and "inessential" changes to the formula (13.34) could also change its aesthetic value and thus the degree
of favour it would find among readers. For example, he wondered how
the following equivalent formula would have scored:
(13.38)

Among Euler's most spectacular results in the 1f domain is his solution for the sum of the reciprocal perfect square roots, i.e. for the
. I1+1+1+
1 + .. '. M any great math
"
expreSSIOn
4'
9"
+ n7
ematIcIans,
4

In second place was another of Euler's formulae, namely that which defines the
relationship between the number of vertices (V), the number of edges (E) and
the number of faces (F) in a polyhedron: V - E + F = 2. in 14th place was an
Indian 1': series dating from the 15th century (16.10).

13.3 Infinite expressions

191

among them Leibniz and the brothers Jakob (1654~ 1705) and Johann
Bernoulli (1667~1748), had tried their hand at this apparently simple
exercise without success. Euler was the first one to succeed, in 1736,
in resolving this sum with the elegant expression,

00

1
k2

7f2

= "6 = 1.64493406. . .

(13.39)

k=l

By way of proof, Euler considered the equation

(13.40)

sinx = 0

For sin x, he took the series that was already familiar to Newton
sinx = x - x 3/3! + x 5 /5! - x 7 /7! +...
(13.41)
6
2
4
= x (1 - x /3! + x /5! - x /7! + ... )
(13.42)
or, after substitution of y = x 2 ,
= x (1 - y/3! + y2/5! - y3/7!

+ ... )

(13.43)

In order for the right-hand side of the equation to = 0, either x or the

expression in brackets must = O. The known solutions of (13.40) x =
7f, 27f, 37f thus hold true for the expression in brackets where y =
7f2, 227f2, 3 27f2, ... correspond to them. Euler then used the theorem
from algebra, which states that the sum of the inverse values of the
n solutions of the polynomial 1 + alX + a2x2 + ... + anx n is equal to
-aI, i.e. to the negative of the first coefficient. (Example: the equation
-~(x - 1)(x - 2)(x - 3) = 0 has the solutions 1.2 and 3. The sum
of their inverse values is equal to
and this is also the negative
coefficient of x, as one can establish by multiplying out the left-hand
side of the equation.) Euler assumed that this must also apply to the
infinite sum contained in the brackets, which has the same form. This
first coefficient in (13.43), set negative, is 1/3! = ~; this value must
thus equal the sum of the inverse values of the solutions

II

127f2

+ 227f2 + 327f2 + ... = 6

(13.44)

If both sides are multiplied by 7f2, one obtains the sought-after solution

to this unusual problem.

In Euler's previously mentioned primary work Introductio in analysin infinitorum of 1748, he delivered another proof for the sum
(13.39), which satisfies more stringent requirements.
This work also contains "truckloads" (Beckmann) of other 7f formulae, e.g. similar summation expressions for 7f3, 7f4, 7f6 etc. as well

192

13. The History of 7r

as many more beautiful identities. It would actually require a separate monograph to do justice to the wealth of Euler's discoveries on
7r. Nevertheless, a small selection of his pronouncements is included in
our collection of formulae on page 224ff.
Another of Euler's achievements was, as mentioned above, to make
the symbol 7r popular. Then, at the age of 72 in 1779, he began calculating 7r and arrived at 20 decimal places using an arctan formula he
had developed himself and a series of his own for the arctan function
(see page 74). This only took him an hour.
But Euler was not the only eighteenth century mathematician to
add to our collective knowledge of 7r. Already in the 15th century it was
suspected that 7r could not be a rational number, i.e. that it could not
be represented as a quotient of two integers. Huygens and Leibniz were
quite convinced of this. But it was not until 1766 that the Alsatian
mathematician Johann Heinrich Lambert (1728-1777) produced the
proof that 7r is irrational in [79], His proof begins with the statement
of an infinite continued fraction for tan v, which he expressed like this:
1
tan v = - - - - - - - - - = - 1 - - - - (13.45)
1: v 1
3: v 1
5:v-----7: v -l&c.
Lambert then demonstrated that the right-hand side of the equation must be irrational if v i= 0 is a rational number. But because
tan( 7r / 4) = 1 is rational, it follows that 7r / 4, and therefore also 7r
itself, must be irrational.
After Lambert, Legendre proved in 1794, that 7r 2 is also irrational.
Arctan formulae

in 1706 John Machin (1680-1752) outdid Sharp's 71 decimal places

with exactly 100 decimal places, all of them were correct. He was the
first mathematician to use a combination of two arctan series in a
single expression. His arctan formula became extremely popular:
1
1
7r = 16 arctan - - 4 arctan (13.46)
5
239
Machin calculated each of the two arctan expressions using Gregory's
senes

13.3 Infinite expressions

193

111
1
arctan - = - - - 3 + - 5 -...
(13.47)
x
x
3x
5x
For the first arctan expression with x = 5 he had to calculate
around 70 terms in the series (13.47), but for the second expression
with x = 239 he only needed 20 terms.
John Machin did not publish his famous achievement himself; his
(1 + 100)-digit 7r and his arctan formula appeared in William Jones's
previously mentioned work, the Synopsis Palmariorum Mathesos[68],
in which the symbol7r was also proposed for the first time. In this way,
this work by a self-made mathematician contains two major innovations regarding 7r.
For more than 200 years afterwards no one came up with a better method of calculating 7r than to use combined arctan formulae.
However, mathematicians soon learned how to construct other such
relations after the model of Machin's formula which might offer scope
for optimisation. C.F. Gauss (1777-1855) developed some especially effective variants. Thus it came about that the arctan method remained
in use for more than 250 years until 1970, although of course this is far
short of the life of Archimedes's method. Even the early computer calculations almost exclusively made use of arctan formulae, and one of
those used most frequently was actually the good old Machin formula
(5.20).
After Machin many other mathematicians took up the task, and
every few years a new 7r world record would be set. But as the number of decimal places grew, so too did the risk of making a mistake.
Thus we hear of several instances of calculation errors which in each
case were revealed inevitably the next time the record was beaten. For
example, although Fautet De Lagny (1660-1734) discovered 27 digits
more than Machin in 1719, it turned out when the world record was
next beaten by Vega (1754-1802), that his 112th decimal place was
incorrect. However, as this was the only incorrect digit, it seems likely
that it was a case of a transcription error rather than that the calculation itself was at fault. Vega himself also made a mistake - of the
143 digits he produced in 1789, only 126 were correct. Five years later
Vega improved on his own record with a 7r that was 140 digits long,
but the last four of these were also incorrect.
Even today, the last digits calculated are always the most suspect.
The current world record holder Yasumasa Kanada cut off 200 digits
from his 206.1 billion 7r digits to be on the safe side, although his

194

13. The History of

7f

test calculation only differed from the actual calculation in the last 45
digits. No doubt he thought there were enough digits anyway.
Euler too was a victim of De Lagny's error, as he reproduced De
Lagny's number in his Intmductio. Since then, publishers thinking of
issuing a new edition of this important and interesting work have been
faced with the dilemma as to whether they should simply correct the
error or transfer it for the sake of historical truth. A recent English
edition dating from 1988 chose the latter approach [52, p. 101], while
an older German edition published in 1885 [53, p. 95] decided on the
first approach.
In the 18th century, important mathematicians were also at work
on 7r in Japan, although they remained behind their European counterparts. in 1722 Takebe Kenko (1664-1739) obtained 41 decimal places
for 7r 2 using a series for 7r 2 (16.86), while in 1739 Matsunaga obtained
50 decimal places. Thereafter, according to Beckmann [18, p. 102]' the
Japanese seem to have had more sense than their European colleagues;
they continued to study series yielding 7r, but wasted no more time on
digit hunting. Today Beckmann would not be able to say this, as in
recent years no one has pursued the hunt for 7r digits more vigorously
than the Japanese Yasumasu Kanada and his team.
The Austrian mathematician Lutz von Strassnitzky (1803-1852)
exploited an unusual opportunity which came his way. in 1840, the
"famous mental computer" Zacharias Dase (182(}-1861) visited him
in Vienna and attended his lectures on elementary mathematics. As
he was whiling away the time with the "most colossal, but pointless
calculations" , Strassnitzky persuaded him to do perform some research
which "at least he would be able to use" , namely the calculation of 7r
to 200 decimal places. Among the formulae it was suggested he should
use, Dase chose the following arctan formula which does not converge
as well as Machin's formula but is well suited for paper and pencil
(and rubber!) [112]:
1
1
1
1
-47r = arctan - + arctan - + arctan (13.48)
258
In under two months Dase had finished. Dase's 200 decimal places were
8 fewer than William Rutherford had achieved twenty years earlier, but
they were still a world record as only 152 of Rutherford's digits were
actually correct.
Dase was known throughout Germany for his calculation skills. He
was able, for example, to multiply two 8-digit numbers in his head

13.3 Infinite expressions

195

in 54 seconds, while 200-digit numbers took him only 8 hours and

45 minutes, a feat which one can scarcely imagine. He offered to calculate some mathematical tables, so Gauss suggested he should expand
the existing prime factorisation tables. Dase took up this suggestion
and, with financial support from the Hamburg Academy of Sciences,
he calculated the prime numbers in all the numbers between 7 and
9 million. Beckmann concludes his report on Zacharias Dase with the
following commentary: "Evidently Carl Friedrich Gauss, an innovator
in so many areas, was also the first to introduce payment for computer
time" [18, p. 107J.

The last records obtained using paper and pencil

The sad tale of William Shanks (1812-1882) took place against the
backdrop of the mid-19th century. Shanks performed no fewer than
four IT calculations by hand. In 1851 he calculated first 315 and then
530 decimal places. Two years later he attained 607, and finally twenty
years later as many as 707. In 1853 he published a complete book,
which contained not only the 607 digits but a detailed description of
how he had performed the 530 digit calculation. In the foreword he
wrote:
"Towards the close of the year 1850, the Author first formed the
design of rectifying the Circle to upwards of 300 places of decimals. He
was fully aware, at that time, that the accomplishment of his purpose
would add little or nothing to his fame as a Mathematician, though it
might as a Computer; nor would it be productive of anything in the
shape of pecuniary recompense at all adequate to the labour of such
lengthy computations."
This was somewhat pessimistic, as Shanks's version of IT was accorded the great honour of being exhibited at the big world exhibition
in Paris in 1937 (see also page 50).
Shanks's calculations concealed a labour of Hercules. In the course
of his calculations he had to determine some logarithms accurately up
to 137 places and to calculate the exact value of 2721. A Victorian
commentator observed, "These huge chains of calculations not only
prove the persistence and accuracy of this mathematician, but they
are also a proof that computational skill and courage are increasing in
society." [119, p. 51J
But fate had dealt Shanks an unlucky hand. To quote J.W. Wrench:
"(His 607 decimal places were) incorrectly calculated beyond 527 dec-

196

13. The History of

7r

imal places. The accuracy was further vitiated by a blunder committed by Shanks in correcting his copy prior to publication, with the
result that similar errors appear in decimal places 460-462 and 513515. These errors persist in Shanks's first paper of 1873 containing
the extension to 707 decimals of his earlier approximation. His second paper of that year, which contained his final approximation to 7f
(of 707 decimals), gives corrections of these errors; however, there appears an inadvertent typographical error in the 326th decimal place of
his final value. In retrospect, we now realize that Shanks's first value
(530 decimal places) published in 1853 was the most accurate he ever
published." [122]
The error in the 528th digit of Shanks's various 7r calculations remained unproven for many years. The person who got the furthest to
checking it was Richter in 1854; however the latter did arrive at the
wrong decimals. It was not until 1946, 90 years after Shanks, that the
error was demonstrated by Daniel Ferguson, incidentally in what may
have been the last 7f calculation to have been undertaken by hand.
However, Shanks's 7r had been suspected of being incorrect much
earlier. The first person who became suspicious about the accuracy of
Shanks's 7r was the mathematician Augustus de Morgan (1806-1871),
a severe critic of the would-be circle squarer and inventor of the morbus cyclometricus, defined as the circle-squaring disease. He counted
the frequency of every digit occurring in the 607 decimal version of
Shanks's number and concluded from the deficiency of 7's that Shanks
must have made a mistake. Unfortunately he was right.
This was all very sad as it only went to show how unfair the world
can be: a person can sweat away at his labours for years only to have
someone else glance at the results and show that they are wrong in a
matter of an hour.
As the 19th century came towards its close, there was a major
new breakthrough on 7r research. What had already been clear to
mathematical scholars such as Michael Stifel (1487-1567) in the 16th
century became certain fact in 1882, when it was established beyond
doubt that 7r is not only irrational but also transcendental. Stated formally this means that 7r cannot be the root of any algebraic equation
containing only rational coefficients. A corollary of this is that it is
not possible to square the circle with a compass and straight edge.
Proof of this was certainly in the air by 1873, when Charles Hermite
(1822-1901) proved the transcendence of e. Decisive proof that 7r too

13.3 Infinite expressions

197

is transcendental was provided a few years later, in 1882, by Ferdinand

Lindemann (1852-1939). His rationale, which is very heavy going, was
later simplified a number of times and today is available in an almost
elementary form, e.g. see [49, pp. 103-109]. Like the proof of 1T'S irrationality, the proof of its transcendence is once again indirect. The
argument goes that if eX is the root of a rational algebraic equation,
then x must be transcendental (and vice versa). In fact, we know from
Euler's famous formula (13.34) that ein is a rational number, i.e. -l.
Therefore i1T must be transcendental and hence 1T also.
The last world record calculation of 1T to have been performed by
hand, or at most with the aid of an electromechanical hand-operated
desk calculator, was performed in the years 1945-1949. Ferguson and
John Wrench Jr. produced the last and best value yet, 1120 decimal
places, before the advent of the computer era.

The computer era

It was at the end of the 1940s that computers entered the 1T arena,
signalling an end to the era of only a few hundred 1T digits.
In 1949 one of the first electronic computers, the ENIAC (Electronic
Numerical Integrator and Computer), was programmed to calculate
2,037 decimal places of 1T, using John Machin's arctan formula (5.20).
The calculation took 70 hours to perform, i.e. more than 2 minutes
per digit. In 1957, F.E. Felton attempted to calculate 10,000 decimal
places, but due to a hardware fault only the first 7,480 places were
correct. The goal of reaching 10,000 digits was achieved the following
year by F. Genuys on an IBM 704 in 100 minutes [13].
On 29 July 1961, two IBM researchers, Daniel Shanks and John
W. Wrench Jr., reached the 100,000 mark. They used an IBM 7090,
which took 8 hours and 43 minutes to perform the calculation. They
wrote a detailed account, still quite readable even today, of their
project [106], in which they bring out the struggle they had with their
algorithms. Like their predecessors they only had arctan series (one
from St0rmer (5.19) for the actual calculation and one from Gauss
(5.22) to check it), but they came up with some interesting optimisation tricks in the course of their work.
Several other calculations increased the number of known 1T digits
by 1973 to 1 million places, using only classical arctan formulae. The
calculation of 1T to 1 million decimal places took just under one day
on a CDC 7600.

198

13. The History of 7r

Even though computers were becoming more powerful, the calculation of still more decimal places would have run into insurmountable
obstacles at some point. Daniel Shanks, whom we mentioned above,
still believed in 1961 that no one would ever succeed in calculating
more than 1 million 7r digits. Castellanos [41] commented on this statement by quoting the Chinese saying that it is dangerous to make predictions especially when they refer to the future.
The Chinese saying did not stop Peter Borwein predicting in a
lecture in 1995 that we would never discover the 1051 th digit of 7r.
Even if 1051 is a very big number and Peter Borwein an extremely
distinguished 7r researcher, this is a bold statement indeed.

13.4 High-performance algorithms

The third era in the hunt for 7r digits began in about 1980 when the
world record still stood at around 1 million decimal places. It was
around this time that the two breakthroughs occurred which together
would enable the calculation of 7r digits to new orders of magnitude.

Fast multiplication
The first of these occurred in 1965 when the process of multiplying
two (long) numbers was speeded up dramatically and to a certain extent out of proportion and the old "school multiplication" approach
was superseded. Under the new method, known as fast Fourier transform (FFT) multiplication, the two multiplicands are first "Fourier"transformed; they are then subjected to an pointwise multiplication
process which can be executed in linear time, and finally the vector
product is transformed back again. The critical factor in this multiplication process is the fact that the three Fourier transforms can
be executed very quickly, i.e. at a rate better than quadratically (see
pages 135ff.).
The discovery of this multiplication algorithm came more or less
by chance:
"During a session of the scientific advisory committee of the American President [Lyndon B. Johnson in 1964] Richard W. Garwin established that John W. Tukey, who was one of the attendees, was trying

13.4 High-performance algorithms

199

to develop some programs for the Fourier transform 5 . Garwin ... asked
... and Tukey provided an outline of what was later to be the famous
Cooley-Tukey algorithm."
It just goes to show how a President has his uses!
"Garwin went to the computer centre of the IBM research centre in
Yorktown Heights to have the procedure programmed. James W. Cooley had not been working at the IBM research centre for long and, in
his own words, was entrusted with the assignment because he was the
only one who did not have anything important to do ... " [38, p. 8].
It was only a short while after publication of the Cooley-Tukey
algorithm that the search for historical sources began. Someone who
read their paper described an earlier program of his own with similar
performance and attributed it to a method by G.C. Danielson and
C. Lanczos dating from 1942. The authors for their part referenced
essays by C. Runge and H. Koenig dating from the years 1903 and
1924 [44]. But these were not the earliest attempts at using FFT. It
seems that once again it was C.F. Gauss who was the first to have
had the all-important idea of speeding up the Fourier transform and
in fact he did so 100 years (1805) before anyone else [65].
The time required for school multiplication is of quadratic order,
whereas FFT multiplication, of which there are today many variants,
only requires the order of N log2 N (see page 140). With FFT multiplication, the time it takes to multiply n-digit multiplicands does not
increase much more quickly than n itself.
1r-specific algorithms

But as well as FFT multiplication, a second branch needed to be added

to the tree of knowledge before today's 7r feats would become feasible, namely the discovery of high-performance algorithms developed
specifically with the aim of calculating 7r more quickly and to additional decimal places. This was achieved from 1976, and as a result
the arctan methods which had been virtually all that was available up
to then were completely replaced.
One of the new methods was based on series which the Indian
mathematician Srinivasa Ramanujan (1887-1920) had developed back
5

Moreover, John Tukey is the inventor of the word "bit", short for binary digit,
in 1946, and he also seems to be responsible for introducing the term "software"
in the sense as it is used today, in 1958 ("The Economist" June 3rd 2000, p. 96)

200

13. The History of

7r

in 1914. One ofthese is the following series, which is significantly faster,

by around a factor of 5, than the best series known before then:

.!. = vIS
7f

~ (4n)! . [1103 + 26390n]

9801 ~ (n!)4
396 4n

(13.49)

A second significant discovery in this area was the Gauss AGM algorithm. It consists of an iteration method such that, when executed,
the number of correct 7f digits is doubled during each iteration stage,
thus converging quadratically on 7f. In this way, for example, the first 9
iterations produce successively 1, 4, 9, 20, 42, 85, 173,347 and 697 correct digits of 7f [13, p. 53].
This algorithm was independently discovered in 1976 by Eugene
Salamin and Richard Brent. But the underlying formula had already
been discovered 170 years earlier by the German mathematician Carl
Friedrich Gauss (see Chapter 7). Salamin at least was aware of the
origin of the algorithm when he published it.
In 1985, the Borwein brothers added to the stock of 7f-specific highperformance algorithms when they published iteration procedures in
the style of the Gaussian AGM algorithm which converge at a rate
better than quadratically. Thus, in the "quartic Borwein iteration" four
times as many correct digits are produced during each iterative step
so that in 1999 the present world record holder, Yasumasa Kanada,
only needed twenty iterations to calculate 206.1 billion digits of 7f.
The Borweins derived their method from the same theory of modular
functions which had also led Ramanujan to his series.
It was thanks to these discoveries that, in an unprecedented boom
spanning the 19 years from 1981 to 1999, the 7f world record was
bettered a total of 26 times, often by more than double the previous
record, from 2 million to currently 206.1 billion digits. Our table on
page 206 lists the individual landmarks.
The new

1T'

digit hunters

It is striking that 18 out of the 39 world records are linked with the
name Yasumasa Kanada (the stress falls on the second syllable). This
man is Research Adviser in the Department of Science at Tokyo University. He heads a laboratory, Kanada laboratory, which specialises
amongst other things in "high performance computing". To perform

13.4 High-performance algorithms

201

this work, Kanada has used special supercomputers, most recently a

Hitachi SR8000 which contains 128 processors. His feats of calculation have been accomplished with these and ever more sophisticated
implementations of the Borwein algorithms.
Unfortunately Kanada is not given to talking about his accomplishments. At the time of writing (April 2000) he has only just brought out
a paper on his third most recent world record (51.5 billion digits) [113]
(in Japanese) and he has issued only brief statements on the Internet
regarding his latest records (68.7 and 206.1 billion digits). Kanada appears not to answer all the e-mails which are sent to him, so we know
relatively little about the details of his work.
Even less has been divulged by Kanada's two "rivals", the Chudnovsky brothers, who have six entries in the list of world records.
However, an excellent biographical report about the two brothers, The
Mountains of Pi, by Richard Preston appeared in the New Yorker in
March 1992 [92]. This essay describes the lonely life of David and
Gregory Chudnovsky in their apartment in Manhattan, where they
achieved their world records on a home computer called m-zero which
they assembled from parts purchased in department stores. The money
used to finance their work came from their wives, for the Chudnovskys
have been unsuccessful at obtaining university posts although they
have established track records in number theory. One reason for this,
Preston remarks bitterly, is that Gregory suffers from a muscular disorder Myasthenia gravis which keeps him bed-ridden.
David Blatner, author of The Joy of 7r [28], an attractive little book
which came out in 1997, apparently tried to track down the Chudnovsky brothers. He found out that they had recently opened a new
office at the Brooklyn Polytechnic University, called the Institute for
Mathematics and Advanced Supercomputing, which has a staff of precisely two (i.e. the Chudnovsky brothers themselves). They no longer
live in Manhattan. Shortly before moving, they ran one more 7r calculation on their m-zero, this time to 8 billion digits, which, including
checking, took up a week of computing time. This was a new record.
They then gave up the chase to achieve further records and destroyed
their home computer.
It seems that the experts do not trust the Chudnovskys entirely.
When Yasumasa Kanada announced one of his world records, the
Chudnovsky brothers said that they had already gone way past that
figure. Whereupon Kanada sent them a few digits from "his" 7r, along

202

13. The History of

7r

with the position from where he had taken them. He asked the brothers
to send him an analogous section from their 1f calculation for verification purposes. But the Chudnovsky brothers have never responded.
The only others to have held a world record since 1981 apart from
Kanada or the Chudnovsky brothers are Gosper and Bailey.
As mentioned above, Gosper calculated 17 million decimal places
in 1985, using the Ramanujan series (13.49). His computer was not a
supercomputer but a comparatively simple Symbolics 3670 workstation which, however, compared with conventional computers, has the
advantage of unlimited arithmetical precision. When announcing his
record, Gosper stressed the fact that he had not calculated simple decimal places but more sophisticated continued fraction places. His sequence thus did not begin with 3,1,4,1,5, but with 3,7,15,1,292. Gosper
hoped to attain a deeper understanding of the nature of the number 1f
by adopting this representation form, since for number theoreticians a
continued fraction is definitely a more informative alternative. Sadly
the continued fraction did not reward the effort of his calculations
but only suggested "normal probability" (see also Chapter 04 on this
point). Nevertheless, the calculation did confirm that the continued
fraction of 1f does not follow any model, which is actually a useful
piece of information, since, for example, the numbers e, v'2 or J3 do
follow a model.
To crown it all, Gosper ended up having to convert his continued
fraction to the decimal form solely in order to be able to check it.
David Bailey is a scientist at the NASA Ames Research Center in
California, USA. He has published extensively on high-performance
calculations and is one of the authors of the BBP algorithm. Between
7 and 9 January 1986 in a total of 28 hours he calculated 1f to 29 million decimal places and wrote a fine essay about it [9J. In the essay he
states amongst other things that the main purpose of his calculation
was to test the hardware (of a Cray-2 supercomputer), the operating
system and the compiler. He wrote his calculation program in FORTRAN because the compiler for that programming language was able
to exploit the vector processor properties of the Cray particularly well.
The actual calculation required 12 iterations of the quartic Borwein
algorithm and checking it required 24 iterations of the quadratic Borwein algorithm. The report also contains statistics on the frequencies
of different numbers, but these do not produce any striking results.

13.5 The hunt for single

13.5 The hunt for single

11"

iT

digits

203

digits

We have already quoted the observation by Adamchik and Wagon that

the "2000-year search [for ever more 7r digits is changing] direction".
Today more and more is heard about a new type of 7r record, which
was proclaimed with the words "The forty trillionth bit of 7r is 0" .
The background to this is the fascinating "BBP algorithm" , which
is named after the initial letters of its authors, David Bailey, Peter Borwein and Simon Plouffe, and was only published recently, in September
1995. With this algorithm it is possible to calculate individual hexadecimal positions in the middle of 7r without having to calculate all the
preceding digits. This method is described in detail in Chapter 10.
This algorithm has opened up a new stage in the quest for 7r digits,
namely the search for digits as far removed from the beginning of
7r as possible. The first landmark was set by the three researchers
themselves with the 10 billionth hexadecimal7r digit, corresponding to
the 12 billionth decimal place. At the time of publication of their data
(September 1995), this digit was almost twice as far into 7r as the last
known digit obtained by the conventional route.
Within only a year later a student called Fabrice Bellard6 had
caught up. He discovered an even better formula and used it to calculate the 100 billionth (October 1996) and the 250 billionth (September
1997) hexadecimal digits of 7r, which were a factor of 10 and 25, respectively, more distant. But even these achievements were overtaken
almost immediately by the disclosure of some hexadecimal digits in
the trillionth area such as the 1.25 trillionth (August 1998), the 10
trillionth (February 1999) and the 250 trillionth (11 September 2000) 7,
which in turn bettered the previous records by factors, respectively, of
5, 20 and 1000. The holder of both records is 19-year-old Colin Percival, a student at the Simon Fraser University in Burnaby, Canada,
the same establishment where the Borwein brothers and Simon Plouffe
work. No doubt we will hear many more great things about this young
man. He began his university studies in mathematics at the age of
13 in parallel to his secondary school education, and has successfully
completed an unusually major project while still in his teens.
Although the formula Percival used was not new (in fact, it was
the Bellard formula, 10.8), for the calculation itself he used an unusual
6

http://www-stud.enst.fr/-bellard/
http://www.cecm.sfu.ca/projects/pihex/announcelq.html

204

13. The History of

7r

approach which may well become a standard approach in years to come

(see also Chapter 15). What was new about it was that he obtained the
co-operation of 1734 Internet computers based in 56 different countries
which all calculated parts of his project in their "dead time". Percival
then put the results together. In this way he obtained the benefit of
almost 700 years of computing time in an elapsed period of only two
years.
No doubt he will produce another, even more spectacular result
soon.
With this we end our excursion through history and through 4000
years of research into 71". The most important milestones are summarised one more time in the tables on the following pages. What
we know now is the beginning and the middle. The end is not yet in
sight.

13.5 The hunt for single 11' digits

Who?
Babylonians
Egyptians
Indians
Bible
Plato
Archimedes
Zhang Heng
Ptolemy
Wang Fan
Liu Hui
Tsu Chhung-Chih

When 7
20007 BC
2000? BC
6007 BC
4407 BC
c. 380 BC
c. 250 BC
c. 130 AD
150
c.250
263
c.480

No. of
decimals
1
1
0
0
2
2
1
3
1
5
7
6
3
1
3
3
3
16
9
9
15
20
35
34
39
15
71
100
111
50
136
152
200
248
261
530
530
620
710
808
1,120

Algorithm 1
3+ 1/8
4 (8/9)2
4 (9785/11136?
11'=3
V2+v'3
pgm(96)

v'lO

205

Comment

3
2;1

< 11' <

377/120
142/45
pgm(192)
pgm(12288)

11' -- 355
113
Aryabhata
499
pg(384)
Brahmagupta
6407
v'lO
Alkarism
830
62832/20000
Fibonacci
1220
864/275
Dante
c. 1320
3+ V2/1O
AI-Khashi
1430
pgm(3.2 28 )
211'
Nilakantha
c. 1501
104348/33215
Viete
1579
pgm(32 17 )
pgm(220)
Romanus
1593
Ludolph Van Ceulen
1596
pgm(15 . 235 )
pgm(262)
Ludolph Van Ceulen
1615
Snell
1621
pgm(23O )
Grienberger
1630
last use of p'gons
pgmO
Newton
1665
N
first use of series
Sharp
1699
G
Machin
1706
M
De Lagny
1719
G
127 total
Matsunaga
1739
H
Vega
1794
E
140 total
Rutherford
1824
E
208 total
Dase
1844
S
Clausen
1847
M
Lehmann
1853
H
W. Shanks
1853
M
Ferguson
1945
L
Ferguson
1946
L
last hand calc'ion
Ferguson
1947.01
L
desk calculator
Ferguson, Wrench
1948.01
M
Smith, Wrench
1949.06
M
1 pgm(x) = Polygons with x sides.
Series: E=Euler (5.25), G=Gregory (5.3), H=Hutton (5.17), L=Loney (5.18),
M=Machin (5.20), N=Newton (16.63), S=Strassnitzky (16.108),
Table 13.1. History of 11' in the pre-computer era

206
Who7

13. The History of 7r

When 7 No. of decimals Alg

Time Computer

Reitwiesner
1949.09
2,037 M
70 h ENIAC
Nicholson, Jeenel
3,092 M
1954.11
0:13 h NORC
Felton
7,480 K
1957
33 h Pegasus
Genuys
1958.01
10,000 M
1:40 h IBM 704
Felton
1958.05
10,020 K
33 h Pegasus
Guilloud
1959
16,167 M
4:18 h IBM 704
8hanks, Wrench
1961.07
100,265 8
8:43 h IBM 7090
Guilloud, Filliatre
1966.02
250,000 G
41:55 h IBM 7030
Guilloud, Dichampt 1967.02
500,000 G
28:10 h CDC 6600
Guilloud, Boyer
1,001,250 G
1973
23:18 h
Miyoshi, Kanada
1981
2,000,036 K
137:18 h FACOM
Guilloud
2,000,050
1981-82
Tamura
1982
2,097,144 G2
7:14 h MELCOLM
Tamura, Kanada
1982
4,194,288 G2
2:21 h HIT M-280H
Tamura, Kanada
1982
8,388,576 G2
6:52 h HIT M-280H
Kanada, Yoshino,
Tamura
1982
16,777,206 G2
< 30 h HIT M-280H
Gosper
1985.10
17,526,200 R
8ymbolics 3670
Bailey
1986.01
29,360,111 B4
28 h CRAY-2
Kanada, Tamura
1986.09
33,554,414 G2
6:36 h HIT 8-810/20
Kanada, Tamura
1986.10
67,108,839 G2
23 h HIT 8-810/20
Kanada et al.
1987.01
134,217,700 G2
35:15 h NEC 8X-2
Kanada, Tamura
1988.01
201,326,551 G2
5:57 h 8-820
Chudnovsky's
1989.05
480,000,000 7 C
:::::: 6 mon CRAY-2
Chudnovsky's
1989.06
525,229,270 C
> 1 mon7 IBM-3090
Kanada, Tamura
1989.07
536,870,898 G2
67:13 h HIT 8-820/80
Chudnovsky's
1989.08 1,011,196,6917 C > 2 mon 7 IBM-3090
Kanada, Tamura
1989.11
1,073,741,799 G2
74:30 h HIT 8-820/80
Chudnovsky's
1991.08 2,260,000,0007 C
250 h 7 m-zero
Chudnovsky's
1994.05 4,044,000,0007 C
unknown unknown
Takahashi, Kanada 1995.06
3,221,225,466 B4
36:52 h HIT 8-3800
Takahashi, Kanada 1995.08
4,294,967,286 B4
113:41 h HIT 8-3800
Takahashi, Kanada 1995.10
6,442,450,938 B4
116:38 h HIT 8-3800
Chudnovsky's
1996.03
8,000,000,000 C
unknown 1 week 7
Takahashi, Kanada 1997.04 17,179,869,142 G2
5:11 h HIT 8R2201
Takahashi, Kanada 1997.05 34,359,738,327 B4
15:19 h HIT 8R2201
Takahashi, Kanada 1997.07 51,539,607,510 B4
29:03 h HIT 8R2201
Takahashi, Kanada 1999.04 68,719,470,000 B4
32:54 h HIT 8R8000
Takahashi, Kanada 1999.09 206,158,430,000 G2
37:21 h HIT 8R8000
IT 8eries: M=Machin (5.20), G=Gauss (5.22), K=Klingenstierna (5.21),
8=8t0rmer (5.19), R=Ramamanujan (13.49), C=Chudnovsky (8.7).
Iterations: G2=Gauss AGM (alg 7.5), B4=Borwein quartic (alg 9.2)
Table 13.2. History of

7r

in the computer era [113]

13.5 The hunt for single

Who?
Bailey,
P. Borwein,
Plouffe
Bellard
Bellard
Percival
Percival
Percival

When?

1995.09
1996.10
1997.09
1998.08
1999.02
2000.09

Hexadecimal
Position
lO lD
1011
2.5.10 11

1.25. 10 12
10.10 12
250.10 12

1r

digits

Formula

Hexadecimal
Sequence

BBP (10.1)
Bellard (10.8)
Bellard (10.8)
Bellard (10.8)
Bellard (10.8)
Bellard (10.8)

921C73C683
9C381872D2
87F72B1DC9
07E45733CC
AOF9FF371D
E6216B069C

Table 13.3. History of digit extraction records

...

...
...
...

...

...

207

14. Historical Notes

14.1 The earliest squaring the circle in history?

On the circumference of the circle of radius 5, there are 12 points
whose co-ordinates are integers. These are the 4 points at which the
circle is intersected by the co-ordinate axes (5,0) and (0, 5) and the
8 angular vertices of the right-angled triangles whose non-hypotenuse
sides are 3 and 4, i.e. points (4, 3) and (3, 4). This is illustrated
in the left-hand diagram, which shows the first quadrant of the circle:
(05)

The 12 points define an (irregular) 12-sided polygon, 8 of whose

sides are of length J2 . vb, and the other 4 are of length J2. Its
diameter is 10 and its perimeter 8 x J2yb + 4 x J2 = 30.9550 ... ;
both together produce an approximation for 7r of 3.09550 ....
If the circle is enlarged by a factor of 5 to radius 25 (centre diagram),
we find the original 12 points (or rather, their equivalents) (25, 0),
(20, 15), (15, 20) and (0, 25) once again, but this time there
are another 8 extra points which have integer co-ordinates (24, 7)
and (7, 24). The side lengths of the irregular 20-sided polygon thus
defined add up to 12 x 5J2 plus 8 x 4V5, and they produce the approximation 7r ~ 3.12814 ....
If the circle is magnified a second time by a factor of 5, we obtain 8 further points which have integer co-ordinates, (117, 44) and
(44, 117) (diagram on the right). The resulting 28-sided polygon

210

14. Historical Notes

with diameter 250 has 20 sides of length 25V2 and 8 sides of length
3V2y'5, and the approximation is now 3.13200 ....
It will be seen that every time the radius is increased by a factor
of 5, 8 new vertices occur whose co-ordinates are integersl; even more
important for calculation purposes is the fact that all the side lengths
of the resulting 36-, 44- etc. sided polygons can be expressed solely
using square roots of 2 and/or 5. This captivatingly simple procedure,
which would doubtless have thrilled Pythagoras, was discovered by
Franz Gnaedinger of Zurich [59J; the complete proof was contributed
by Christoph Poeppe. In arriving at this discovery, Gnadinger identified additional simple geometric constructions for the formation of the
vertex co-ordinates and also devised simple methods to approximate
the square roots from 2 and 5 to any degree of accuracy. For the roots
from 2, for example, he specifies the following number stack, whose
construction rule is easy to guess:
1

1
2

2
4

3
5

10

7
12

17
29

24
41

70

58
99

140

The fractions 10/7 and 7/5 are simple approximations for V2. 24/17
and 17/12 are better approximations, and 140/99 and 99/70 are already very good approximations for V2.
Now Franz Gnadinger is first and foremost an Egyptologist, and so
he looked to see whether the good old Egyptians (who of course were
very young in the pharaonic era, due to their low life expectancy, and
looked at the world with just the same curiosity that young people do
today) had known of this method. His conclusion is that they probably
did.
To start with, the Egyptians had a motive for measuring circles.
The hieroglyph for the sun god Re was a circle. If one could only
understand the circle and fathom its secret number, one would be able
to share in the power of Re.
1

The co-ordinates of the new points are generated iteratively as follows:

Xo := 0, yo:= 1 Thereupon for k = 0, 1,2, ... :
X2k+l
X2k+2

= 3Y2k - 4X2k,
Y2k+l = 4Y2k + 3X2k
= 4Y2k+l - 3X2k+l, Y2k+2 = 3Y2k+l + 4X2k+l

14.2 A

7r

law

211

But the Egyptians were also quite precocious in mathematical matters. They already knew about Pythagorean triangles (right-angled
triangles whose side lengths are integers), the simplest of which have
side lengths which are multiples of 3, 4 and 5 and were referred to by
the Egyptians as "holy triangles" .
According to Gnadinger, all the elements of a method for approximating 7r and even the most important numbers are to be found in
King Zoser's funerary complex at Saqqara, in the Cheops pyramid and
the Chephren pyramid, for example:

Saqqara: 17/12 for root 2

Saqqara: numerous Pythagorean triads
Chephren pyramid: half cross-section = holy triangle
Cheops pyramid: holy triangle 15-20-25 in the royal chamber
Cheops pyramid: 140/99 for root 2 and 161/72 for root 5
Cheops pyramid: basex2:height = 22:7 or approximately 7r

Gnadinger is especially interested in the visionary architectural genius Imhotep, who c. 2600 BC designed the monumental and graceful step pyramid at Saqqara near the former Memphis, to the fame
and eternal life of Pharaoh Zoser. Gniidinger cites other Egyptologists
in support of his view that this Imhotep could well have developed
extremely good approximations for 7r based on the above procedure
and the rational approximations for the roots of 2 and 5 which have
been handed down. After just two iterations, with irregular 20-sided
polygons, he would have arrived in systematic manner at the approximation 7r ~ 3.128 ... , which is significantly better than the value of
(~6? = 3.160 ... which is found in the Rhind Papyrus of 1850 BC
and was probably discovered empirically (see page 167). This more
accurate version of 7r would have been more than 700 years older.

14.2 A

1T'

law

100 years ago a law prescribing the value 3.2 for 7r was very nearly
passed in the state of Indiana, USA. The story is no doubt mainly
amusing to Americans, perhaps because the event took place on American soil and in any case everything turned out well in the end, or
perhaps another reason why they like it could be that a former, less
popular vice-president came from Indiana.

212

14. Historical Notes

At the end of the nineteenth century, a country doctor named

Dr. Edward Johnston Goodwin (1828'1-1902) lived in the town of Solitude, Indiana. According to his biography, this man discovered the
true value of the number 71' by supernatural means in the first weeks
of March 1888.
Strictly speaking, the one revelation must have included several values of 71', since, in a reader's letter to the fledgling "American Mathematical Monthly" 2, he actually gave 5 different values of 71', from 2.56
to 4.0.
He applied for copyright for his discovery in several states, including
Germany and Austria. He then persuaded a representative of his home
town to introduce a Bill on his new mathematical discovery.
The proposed Bill was written by Dr. Goodwin himself. In this
Bill he described 71' in words and - as he had previously done in the
American Mathematical Monthly - in several ways. House Bill No. 246
contains, for example, the passage:
... disclosing the fourth important fact, that the ratio of the
diameter and circumference is as five-fourth to four ...
This "fact" results in 71' = 16/5 = 3.2. Elsewhere in the text three
other values of 71' are given, namely, the argument that the area =
(circumference/4)2 produces 71' = 4, the argument that the area = 6
(circumference/4) 2/5 results in 71' = 10/3 = 3.33333, while 71' = 32/9 =
3.555556 is obtained from diameter 2 = 92 and circumference/4 = 8.
The Bill met with its first hurdle in the Indiana house of representatives on 5 February 1897. After deliberation in the responsible
committee, the Bill was passed to the senate for confirmation, with
the recommendation that it should be become law.
In defence of the representatives it must be said that their primary
consideration was the commercial good of their state and in mathematical affairs they relied entirely to the reputation of Dr. Goodwin.
One representative argued that if this Bill were passed, then the author
would give his home state the right to use his discovery free of charge
and, for example, to publish it gratis in school textbooks whereas everyone else would have to pay a royalty to use it.
Luckily for the reputation of Indiana Professor C.A. Waldo of Purdue University happened to be visiting the house of representatives
that day and listened in on the committee session. He pointed out
2

The magazine had actually existed earlier but under the name of "Ladies Diary" .

14.3 The Bieberbach story

213

the shortcomings of the Bill to the senators, who placed all further
readings of the Bill on hold, and the matter was permanently laid to
rest.
One of the historians recounting this event, David Singmaster, felt
that Goodwin was due particular credit as one of only very few "circle
squarers" to have possessed several values for 7r [110]. The mean value
of all the 7r'S which Goodwin served up in articles, interviews and in
the above draft Bill, was calculated by Singmaster at 3.28907.
This story even appeared in the German newspaper "Die ZEIT"
(1997/28,4 July 1997). There it appeared with the following comment:
"Before we laugh too loudly at the legislature of Indiana or at the
general ignorance which prevailed in 1897, we should pause a moment
and reflect on what fate the draft Bill would have met, had it been the
subject of a referendum today.

14.3 The Bieberbach story

During the Nazi era Hitler's Germany formed the backdrop to a disgraceful episode concerning 7r. Between 1909 and 1933 the Jewish
mathematician Edmund Landau (1877-1938) pursued a distinguished
teaching career at the University of G6ttingen, a renowned centre of
excellence in mathematics whose prominent figures included, for example, David Hilbert, Felix Klein and Emmy Noether. In his lectures
on differential and integral calculus he used the definition which today
is regarded as standard that 7r /2 is the smallest positive zero of cos x.
In the telegram style which was characteristic of his writings, Landau
wrote in a textbook, "The universal constant from theorem 262 will
always be denoted by 7r [51, p. 130]."
In 1933 Landau was dismissed from his chair on grounds of his
race. An important colleague of Landau's subsequently added insult to
injury by gratuitously suggesting that Landau's dismissal was justified
on mathematical/psychological grounds.
This Berlin colleague of Landau's was the function theoretician
Ludwig Bieberbach. In June 1934 he wrote the following lines in a
treatise on Personality structure and mathematical creativity:
"In this way. " the ultimate reason behind the courageous rejection which the students at G6ttingen University meted out to a great
mathematician, Edmund Landau, was that his un-German style in research and teaching had become intolerable to German sensitivities.

214

14. Historical Notes

A people which has seen how alien desires for dominion are gnawing
at its identity, how enemies of the people are working to impose their
alien ways on it, must reject teachers of a type alien to it."
The English mathematician Godfrey H. Hardy, whom we encountered earlier as S. Ramanujan's patron (see page 106), responded to
Bieberbach in these words [109]:
"There are many of us, many Englishmen and many Germans,
who said things during the (First) War which we scarcely meant and
are sorry to remember now. Anxiety for one's own position, dread
of falling behind the rising torrent of folly, determination at all costs
not to be outdone, may be natural if not particularly heroic excuses.
Prof. Bieberbach's reputation excludes such explanations of his utterances; and I find myself driven to the more uncharitable conclusion
that he really believes them true."

15. The FUture:

7r

Calculations on the Internet

In this chapter we introduce you to a procedure which enables 1f to be

computed using a distributed technique and at the same time allows
the results thus computed to be recycled for later computations. This
approach means that large parts of the computational work can be
shared out over many small computers on the Internet. This has the
major advantage that ultra-expensive supercomputers only have to be
used for those parts of the calculation which only they are capable of
performing.
The underlying algorithm is called the binary splitting algorithm,
or binsplit algorithm for short. This algorithm, which has been around
for some time, was recently re-examined in more detail by Bruno
Haible [62J. It is precisely the 1f series which have long since ceased
to be used in the quest for world records that are now returning to
favour. The critical factor here is, however, rapid convergence. With
this in mind it is not surprising that the previously mentioned series
developed by the Chudnovsky brothers (8.7), which converges with 15
decimal places per term, is now the candidate for the binsplit algorithm.

15.1 The binsplit algorithm

Just for once in this book, we invite you to attempt the task of arriving
at a correct derivation yourself. But don't worry, this derivation is very
short, very elementary and once you have done it you will be that much
wiser.
To recap, the time taken to naively total a sum ~f=r} ak is proportional to N2. To improve it, one calculates the ratio rk of successive
summands:
(15.1)

216

15. The Future:

7r

Calculations on the Internet

(To avoid treating k = 0 as a separate case, one sets

a-I

:= 1.) Hence,

N-I

L ak =: ro (1 +

rl

(1

+ r2 (1 + r3 (1 + ... (1 + rN-d .)))) (15.2)

k=O

N ow we define

rm,n := rm (1

+ rm+1 ( ... (1 + rn) ... ))

where

m <n

(15.3)
(15.4 )

Then,

(15.5)
(No need to believe it, just work it out!)
In particular,
n

ro,n

(15.6)

= Lak
k=O

Moreover, the following also holds:

(15.7)

... + rm ..... rx + rm ..... rx . [rx+1 + ... + rx+1 ..... rnl

x

= rm,x +

II rk . rx+l,n

(15.8)

k=m

hence

(15.9)
(where m ::; x < n). All the terms listed here are rational numbers,
and therefore consist of two integers.
In this way we can formulate the binsplit algorithm by simply specifying a recursive binsplit function:
rational function r (rational function a, int m, int n)
{

rational ret;
if m==n then
{

ret := a(m)/a(m-1)
}

else
{

x := floor( (m+n)/2 )

15.1 The binsplit algorithm

ret

;=

217

r(a,m,x) + a(x) / a(m-l) * r(a,x+l,n)

return ret
}

Here, a (k) is a function which returns the kth term of the series to
be summed. As an example let us take arctan(l/lO):
function a(int k)
{

if k<O then
else

return 1
return (-1)-k/((2*k+l)*10-(2*k+l))

The call r(a,O ,N) returns 2:~=o ak.

If one looks at the terms which occur on the call r ( . ,0,7), it will
be apparent that this procedure is better than the naive route. Here
first of all is the recursive structure:

__

rO,7
--------~A_--------~

TO,3
__----A"------,

r_----~A"------,

rO,1
~

TO,O

TI,1

r2,2

r3,3

T4,4

T5,5

T6,6

r7,7

Every node corresponds to a call to the function r, the top one to

the user call r ( . ,0,7,); all the lower nodes are recursive calls of r by
itself.
Specifically, for our arctan(l/lO) example we obtain

__

+ 4~~8:55070~~010~0~3~030~d"0

. . --------------_

------------~A
20930417

4964393497
643500000000

+ 210000000
-417
70000

+299
3000

-7637
990000

__---A....---.

+-lo

3
- 500

7
-900

1
- 300

9
-1100

-16357
1950000

_--A. . .--_
-

11
1300

13
-1500

The numerators and denominators in row k (counted from the top)

are approximately N 10g(N)/2k digits long, where N is the number of
digits in the end result. If we make the simplifying assumption that the
time required to perform a multiplication operation increases linearly
with N I , it can be shown ([62]), that the total time required under
1

Instead of the realistic Nlog(N) log(log(N)), which can be achieved with the best
known multiplication method, namely the SchCinhage-Strassen method [104].

218

15. The Future:

7r

Calculations on the Internet

reasonable assumptions is proportional to M(N) . (log(N))2, where

M(N) indicates the time taken for a multiplication operation oflength
N (In row k of the tree there are proportionally 2k multiplication
operations of length Nlog(N)/2k.) In this way, series with linear but
sufficient convergence can be considered once more as candidates for
achieving 7r world records.
We have actually cheated a little: it is true that the terms r (m, n)
which occur in our tree are of a length proportional to n-m, but the
a (n) which occur during recursion are proportionally n long, so that
on the right-hand side of the tree they are always of maximal length.
This is not actually so bad as the quotients a (k) / a (j) which occur
are again only proportionally k- j long. We therefore specify a function
which calculates this quotient without a detour involving relatively
long terms:
rational function a(int k, int j)
{

if j<O then
else

return (-1)-k/2*k+l)*10-(2*k+l))
return (-1)-(k-j)*(2*j+l)/(2*k+l)/10-(2*(k-j)

The binsplit function which is effectively doing the calculation thus

goes:
function r(function a, int m, int n)
{

rational ret;
if m==n then
{

ret := a(m,m-l)
}

else
{

x := floor( (m+n)/2 )
ret := r(a,m,x) + a(x,m-l) * r(a,x+l,n)
}

return ret
}

A really nice characteristic of the binsplit algorithm is the fact

that each fraction, once calculated, can be reused for later, expanded
calculations. Let us assume that in the latest (N-long) world record
r ( . ,0, N-1) has already been calculated. Let us suppose that technical
progress means that we now have a computer which for the first time
can calculate numbers 2N digits long. We therefore want to work out
r ( . ,0, 2N -1) for a new world record. This is not a problem, since
according to formula (15.9) the following applies (with x = N - 1):

15.2 The

rO,2N-l

rO,N-l

7f

+ aN-I' rN,2N-l

project on the Internet

219

(15.10)

We already have rO,N-l, so what we need to compute is rN,2N-l.

However, this can actually be done on the "old" computer and could
therefore also be calculated in advance. The new supernumbercruncher
would then actually only need to perform the last but one iteration.

15.2 The

7r

project on the Internet

Armed with the binsplit algorithm, we would now like to describe a

project for calculating 1f on the Internet. Instead of numbers of digits,
what we are concerned with here is memory size (there is only one
proportionality constant throughout, typically, for example, 2 [decimal places per byte]). We will assume that our supercomputer has 1
terabyte (TB) of memory, i.e. 240 bytes. This amount of memory can
be halved 40 times until we are down to one byte; thus the binsplit
recursion depth is 40 (in reality, a little smaller if the series converges
rapidly). We now divide the work into 3 layers:
Right at the top is a supercomputer. This will take over as few rows
as possible (in the tree structure), ideally only the ones which really
cannot be handled by any other computer .
For the lower levels, we need as many voluntary Internet participants as possible (for example, you!) who have computer time available on their workstations (which must have a reasonable amount of
memory). They will take over as many rows as possible. But as the
amount of memory available on such PCs is significantly lower than
1 TB, we plan to increase the amount of virtual memory by a factor
of 4 to 16 using "mass storage FFTs,,2. A typical PC has around 64
MB, which mass storage FFTs can then effectively transform into
256 MB (= 228 bytes). This means that the Internet volunteers can
each take over 28 rows, i.e. almost two-thirds of the necessary CPU
work.
There is no doubt that the existing CPU resources are easily sufficient for this: CPU cycles equivalent to 500,000 years or a whole
geological era of CPU time, capable of executing a million instructions per second, were recently applied on a distributed project (an
2

These are fast Fourier transform (FFT) algorithms which can handle data volumes larger than the RAM storage available; they work with files on the hard
disk.

220

15. The Future:

7r

Calculations on the Internet

attempt to break an encryption code). Ideally one would only use

periods during which the computers are not doing anything for this
project. There is one remaining problem: since for reasons of economy we do not want to burden the mainframe with all 12 remaining
rows, there is a yawning gap .
Hence, a company which manufactures really nice computers is making available to us enough computing time on a computer fitted with
16 G B of RAM (and a vast amount of hard disk storage space), on
which we will be able to calculate the rows up to and including 128
gigabytes 3 . In this way the supercomputer will only need to process
the rows with operations requiring 256, 512 and 1024 GB, a total of
3 out of 40.
Meanwhile we must not forget that the Chudnovsky series also
requires a square root to be calculated. But this is not a major problem:
for roots of integers, any number of good rational approximations can
be specified. If ~
is used as a starting approximation for v'd, then the
Va
iteration produces
Un+l

u; + dv;

--:=----=..:.--,;,;:,
Vn+l

2un

(15.11)

Vn

an improved approximation with twice the precision (see section 11.5).

If one really wants to transform the huge fraction into the digit sequence 3.14159 ... , another division operation is necessary, as the binsplit algorithm produces the numerators and denominators separately.
This division entails 4 multiplication operations (see section 11.4).
The question of how to transfer the data thus generated requires
some thought. Transfer of the (spatially distributed) lower rows to the
middle rows will certainly take place over the Internet, in chunks of
several hundred megabytes. It is therefore essential to have a good
Internet connection. The chunks will add up to 2 TB in all, 1 TB
for the numerators and 1 TB for the denominators. The transfer to
the supercomputer will probably be performed by post, since at the
end of the day it is several orders of magnitude cheaper to send a few
kilograms of storage media than to transfer 2 TB over the Internet.
Here the binsplit algorithm will demonstrate its compatibility with
truly traditional methods of data transfer.
3

Here too mass storage FFT will be used.

15.2 The

7r

project on the Internet

221

The project is feasible. We have analysed it and it looks good. It

would be the start of a new era in 7r world records.

16.

Formula Collection

7r

circumference
diameter

1T

(16.1)

area of circle

(16.2)

1T

= radius of circle 2
= 4 arctan 1

1T

= arctan 1 + arctan 2 + arctan 3

1T

1T

(16.3)
(16.4)
(16.5)

-iln(-l)

Nilakantha series, 15th century

1T

= v'i2 (1 _ _1_ +
3.3 1

_1_ _ _
1_ + _1__ ... )
5 . 32
7.33
9.3 4

(16.6)

Nilakantha, 15th century [89]

p/2
1 p2+1
where p is the last odd denominator + 1
1

1T

4'::::d-:3+5-'7+"''f p -

(16.7)

Nilakantha, 15th century [89]

2

1T

4' ~

1-

:3 + 5 - '7 +

...

'f

1
7+1
p - 1 ~ (p2 + 4p + 1)

(16.8)

where p is the last odd denominator +1

Nilakantha, 15th century [89]

1
1
1
1
16 = 15 + 4 . 1 - 3 5 + 4 . 3 + 55 + 4 . 5 - 75 + 4 . 7 + ...
1T

(16.9)

Nilakantha, 15th century [89]

1T

4
4
4
4
3 + 33 _ 3 - 53 _ 5 + 73 _ 7 - 93 - 9 + ...

(16.10)

Nilakantha, 15th century [89]

1T

444
~ 2 + 22 _ 1 - 42 _ 1 + 6 2 _ 1 -

- ... 'f

p2 _ 1 2(p + 1)2 + 4

where p is the last even perfect square in the series

Nilakantha, 15th century [89]

(16.11)

224

16.
7r

"8 =

Formula Collection

7r

1
22 - 1

+ 62 -

+ 102 -

+ 142 -

+ 182 -

(16.12)

+ ...

Nilakantha, 15th century [89)

7r

'2 -

(16.13)

42 - 1 - 82 - 1 - 122 - 1 - 162 - 1 _ ...

Nilakantha, 15th century [89)

Formulae of Leonhard Euler (1707-1783)

e

iTr

= -lor:

7r

-iln(-I) or:

eTr

2i

(16.14)

Euler, 1743

= 1 +x x2

arctan x

1+

3'2

X2

1 + x2

+ 23 .. 45

X2

1 + x2

) 2

+ ...)

(16.15)

Euler, 1755

7r

28
10

[1 + ~ (2.) + ~ (2.)2 + ...]+

100

30336 [

+ 100000

1+

3.5

2 (

3'

100

144)
100000

24 (

+ 35

144 )
100000

+ ...

(16.16)

Euler, 1779 [114, p. 295J

7r=41

~dx

(16.17)

Euler, 1738

26
7r

-5-

dx

io

(16.18)

V1- X4

Euler, 1739
7r

= 4

dx

f01

io VI -

X4

x dx

io VI -

(16.19)

X4

Euler, 1748

7r

= n-+oo
lim - 2
n

~
~

Jn

k2

(16.20)

k=O

Euler, 1738
7r

. [1 + - 1 + (1+

lIm

n-->oo

4n

6n 2

-2 - n
12

1 )]

+- - + ... + -n 2 + 22
n2 + n Z

(16.21)

Euler [41, (3-12))

7r

3\1'3 (1 _ .!:.
2

+ .!:. _ .!:. + .!:. _ .!:. + ... + _1_ _ _1_ + ... )

4
5
7
8
3n + 1
3n + 2

(16.22)

Euler, 1739

1i =

00

7r

l: 2n+1 tan 2n+1

n=1
Euler

(16.23)

16.
1(

1
7r 2 =

1
1
1
1 - 22 - 32 - 52
1

+ (27)2 +
1
- 232

+ (2.3)2

Formula Collection
1

+ (2.5)2

1
1
1
1
(35)2 -172 - 19 2 + (37)2

1
1
(2 13)2 - 29 2

1
- 72

7r

225

1
1
- 112 - 132 + (16.24)

1
(211)2

1
)
(2 3 . 5)2 - ...

(16.25)
(16.26)

Euler, 1737

(16.27)
Euler, 1736
7r 2

32

22

6
2
7r

8 =

22 - 1 . 32 - 1 . 52 - 1 . 72 - 1 ...

(16.28)

1
}2

(16.29)

+ 32 + 52 + 72 ...

Euler, 1748

(16.30)
Euler, 1748

- = -1 - -13+ -1 - -13 ...

7r

32

53

(16.31)

Euler, 1748

4
7r

4
90

7r

=
=

9
680

00

(16.32)

k4 (2k)

k=l
Euler, 1738

1
14

(16.33)

+ 24 + 34 + ...

Euler, 1748

4
96

7r

(16.34)

= 1 + 34 + 54 + 74 + ...
Euler, 1748

(16.35)
Euler, 1748

26
7r

1
27! ( 1
76977927 224 126

+ 226 + 326 + ...

(16.36)

Euler, 1738
7r

= ~ + 4n (
n

+ B2

__
1_
n2

+ 12

B6

+ __
1_ + ... + __
1_)
n 2 + 22
n2 +n2
+ BlO
B14

~+
e 21l'n

3. 22 . n 6
5.24 . n lO
7.26 . n14 + . . .
(16.37)
where the B; are the Bernoulli numbers: B2 = 1/6, B6 = 1/42 ...
n2

Euler [41, (3-13)J

(16.38)
Euler, 1748

226

16.

Formula Collection

1l'

1l'

11

13

17

19

2' = "2 . "6 . "6 . 10 . 14 . 18 . 18 ...

(16.39)

where the numerators are the odd prime numbers

and the denominators are even numbers, not divisible by 4,
which differ from the numerators by 1.
Euler [20, p. 656]

1 2 6 12 20 30 42
-=1+- 1+1+1+1+1+1+1+ ..
2
1l'

(16.40)

Euler. 1739

3
-7r
4

1 3 8 15 24
- 2+2+2+2+2+ ..

= 2+ -

(16.41)

Euler, 1739

Sin 7rZ

7r

IT (1 - ~:)

(16.42)

k=1

For

= 1/2, (16.91) applies.

Euler, 1735

7r

- . - = r(z)r(1- z)

z real, non-integer

sm 7rZ

Io eOO

(16.43)

where rex) :=
t - dt for all real x > 0
Euler's Gamma function
X

Formulae of S. Ramanujan (1887-1920)

1

;: =

(2n) 3 42n + 5

L=

n=O

(16.44)

212nH

Ramanujan [96, (29)], 1914

7r

_1_ ~(_l)k (4n)!) (1123 + 21460n]

3528 f;:a
(n!)44 4k
882 2n

(16.45)

Ramanujan [96, (39)], 1914

~ (4n)! [1103 + 26390n]

= vI8

7r

9801

f;:a (n!)4

(16.46)

396 4n

Ramanujan [96, (44)], 1914

~2 = 1 +

! (!)
2

+ 13
24

(!)
5

~
246

(!)
7

+ ...

(16.47)

Ramanujan [41, (8-1)]

V"2

7re

1 + 1 3 + ~ + 1 35 7 + ... +
1 1 234
+- - - - 1+1+1+1+1+ ..
Ramanujan [41, p. 89]

(16.48)

16.

7r

"2 In 2 =
3
48

7r

1
1 + "2

13

Ramanujan

[41. (8-2)J

7r

. 32 + 2.4

1
. 52

135

1
. 72

+ 2.4.6

1.3

Ramanujan

[41, (8-3)]

7r

Formula Collection

+ .. .

1.3 .5

+ 4" (In 2) = 1 + "2 . 33 + 2 . 4 . 53 + 2 . 4 . 6

227

(16.49)

1
. 73

+ ...

(16.50)

BBP-like series
(_I)n

~~
00

7r

(2+
4n

Adamchik and Wagon,

=~
00

7r

1 (4+

16n

8n

1)

1997

[2]

1)

1
26

(2
- 4n + 1 -

(-1
210n

~
00

1
4n + 3

2
8

= ~ -.!.. (
~ n3

n=1

7r

=~

-.!..

~ n3
n=l

1)

+ IOn + 9

384 + 45/2 _ 1215/2)

n +2
2n + 1
2n + 3
Wagon, 1997 [2]

-12
n +1

Adamchik and

(16.53)

+ IOn + 1 +

26
22
22
- IOn + 3 - IOn + 5 - IOn + 7
Bellard, 1996 [19]
7r 2

(16.52)

1 - 8n + 4 - 8n + 5 - 8n + 6

Borwein, Bailey and Plouffe, 1995

7r

(16.51 )

+ 4n + 2 + 4n + 3

(-238
n+1

+
+

(16.54)

285/2 _ 667/32 _ 5103/16 +

2n + 1
4n + 1
4n + 3

+ 35625/32)

(16.55)

4n+5
Adamchik and Wagon,

7r

1
226

00

1997 [2J

(-1
(227
230n
20n + 1

222
6n+l
2 15

+---

220
60n+15
212

2 25
12n + 1

2 19
IOn+3
212

225
IOn + 1
2 18
20n+7
2 10

224

+ 20n + 3 +
2 15
12n+5
29

------------+

+ 20n + 9 + 30n + 15 + 20n + 11 - - - - -::-::----:-=12n + 7

20n + 13
5
3
27
2
22
2
2
- IOn + 7 - 60n + 45 +--+
+---+
6n + 5
20n + 17
IOn + 9

+ 12n+ 11
Pschill, 1999

1)

+ 12n+ 19

(16.56)

228

16.

7r

7r

Formula Collection

2=1(3
1)
4n + 1 - 4n + 3

= v'3 ~ gn

or, general:
2

7r

(_I)jn

00

(j

(16.57)

(_I)I-1 3j

-l) .>

=- - " " - - ""

3j - 2 v'3 ~ 3jn
~ 2jn + 2l- 1
n=O
1=1

,J

2,3,4 ...

(16.58)

Pschill, 1999

Other series
(16.59)
Leibniz, 1674
7r

7r

1111
= 2V2 ( 1 + 1
3" - 5 - "7 + "9 + 11
=

}~~

+ + ... )

(16.60)

22n

- -

c:) .:; :

(16.61)

Knopp after {41J

(16.62)
Bauer [17], converges a lot better

7r

= 24

v'3 t/4
( 32 + 10

3V3

=4

+ 24

Jx - x 2 dx

( 1
1
12 - 5. 25

(16.63)

28 . 27 - 72 . 29

1)

~ 3 . 5 ... (2k - 3)
- ~ 4 . 6 ... (2k) . (2k + 3) 22k+ 3

(16.64)

Newton, 1665, after {32, p. 339J

7r

'6

arCS1ll

(16.65)

13

= 2 + 2 . 3 . 23 + 2 . 4

1
. 5 25

135

1
2 . 4 . 6 . 7 27

+ ...

(16.66)

Newton, 1676

V3

7r

- = arctan6
3

V; (

1 - 3\

(16.67)

+ 3/' 5 -

31

3 .7

3/ 9 -

.. .)

(16.68)

Sharp, 1699 [114, p. 294J

7r

= 16v'3 (_1_
+ 5 . 72. 33 + 9 . 113 . 3 5 + ... )
1 .3 .3
Used by De Lagny in 1719

(16.69)

16.

L
oo

7r

- - 1+
2 -

8=0

7r

Formula Collection

2 . 4 . 6 ... 28
1
.3 . 5 . 7 ... (28+1) 2 8

229

(16.70)

After Stirling, 1730, [85, p. 81J

(In 2?
~ 1
12 - - 2 - + L 82

7r

(16.71)

28

8=1

Legendre, 1811, [85, p. 81J

7r

(Si:X)

x=-oo

7r

+00 (.Sl:X
= L

(16.72)

)2

(16.73)

x=-oo

(16.74)

(16.75)
(16.76)
Catalan, 1948 [41, p. 86J

1
7r

= 3+ 6

(00 (_I)n+l)
~ L~=1 k2

(16.77)

J. and P. Borwein [32, p. 101J

7r

740025
where

(00
3P(n))
~ G~) 2
n- 1

pen)

(16.78)

20379280

885673181 n 5 + 3125347237n 4

:= -

- 2942969225 n + 1031962795 n
- 196882274n + 10996648
3

Bellard, 1996 [19J

00 n!22 n+ 1
7r = ~ (2n+ I)!

7r

(16.79)

1
234
2+ -(2+ -(2+ -(2+ -(2+ ".))))

Series for spigot algorithm (see page 78)

1
23
35
47
3 + 60 (8 + 7.8.3 (13 + 10.11 . 3 . (18 + 13.14.3 (23 +.

(16.80)

,,)))) (16.81)

Gosper
7r

n=O

(25n - 3)n!(2n)!
2n-l(3n)!

Gosper, 1974

(16.82)

230

16.

Formula Collection

7r

297 =

( 1

25 ~ ben)

7r=

n2

(n

1)

(16.83)

+ 1)2

where ben) refers to the number of odd decimal digits in n

(b(90l) = 2, b(811) = 2, b(406) = 0)
J. and P. Borwein, 1992[34]

7r

= 1
3J3~ ne:)

(16.84)

Comtet, 1974

7r

18

n=l
Euler, 1748
2 _

7r

(16.85)

n2(2n)
n

2n 1
+ (n!)2]

4 1 + ~ (2n
[

(16.86)

+ 2)!

Takehe Kenko (1722) [98, p. 306]

7r 2

= 72 ~ (2 - v'3t
L
n2(2n)
n=l

(16.87)

Crandall, 1994 [47]

4
7r

3240 ~

(16.88)

= 17 L n4(2n)
n=l

Comtet, 1974

! =
7r

12
~(_I)k (6k)! 13591409 + 545140l34k
)640320 3
(k!)3(3k)!
(640320 3)k

t:o

(16.89)

D. and G. Chudnovsky, 1987

Products

(16.90)
Viete, 1593

:;;: =
=

11

= (

33 55 77
24 . 46 . 68 . . .

1)
1 + 4n(n + 1)

13 35 57 79
2 . 2 . 2 . 4. 4 . 6 . 6 . 8 . 8 . . .

00

211

(16.91)

1)
1 - 4n 2

(16.92)

WaIlis, 1655

:;;: =

II

n=l

VI

fl

2 + 2 2 + 2 2 + ... + 2 V2 x
(n radicals)

Osler[88], 1999

II= 2 2P+1n'
+ n - 1 2 + n +1
2P+1n
P

n=l

(16.93)

16. 7r Formula Collection

00

err = 32

II (a~+l)
j=O

231

2- j + 1

(16.94)

where ao = 1, bo = 1/v'2, aj+1 = (aj + bj )/2, bj+1 = .,;a;:Tij

Salamin [100, p. 569)

Continued fractions
For an explanation of the notation, see (4.67) on page 65.
7r

111111111111
3+ - - 7 + 15 + 1 + 292 + 1 + 1 + 1 + 2 + 1 + 3 + 1 + 14 + ...
12

32

52

72

92

-=1+- 7r
2+2+2+2+2+ ...

(16.95)
(16.96)

Brouncker, 1658

4
7r

2 13 35 57
1+7+ 8 + 8 + 8 + ...

(16.97)

Euler, 1783 [114, p. 292]

7r
2 13 35 57
'2=1+'3+4+ 4 + 4 + ...

(16.98)

Euler, 1783 [114, p. 292)

4
12
22 32 42
-=1+7r
3+5+7+9+ ..
7r 2
1 1.1 1.2 2.2 2 .3 3 .3 3.4
-=1+- 6
1+ 1 + 1 + 1 + 1 + 1 + 1 + ...
12
14 24 34 54
- 2 = 1+- 7r
3 + 5 + 7 + 9 + ...
~ = 1-! 23
12 45
2
3- 1 - 3
1 - ...

(16.99)
(16.100)
(16.101)
(16.102)

Stern, 1833

12
32 52 72
7r=3+- 6+6+6+6+ ..
Lange, 1999 [80)

(16.103)

232

16.

7r

Formula Collection

Arctan relations
With the notation [xJ

arctan ~:

2:: = [lJ

(16.104)

i = 2[v'2J + [2v'2J
i = 2[3V3J + [4V3J

(16.105)

Wetherfield, 1996

(16.106)

~ = [2J + [3J

(16.107)

Euler, 1738

~ = [2J + [5J + [8J

(16.108)

Strassnitzky, 1844
7r

-4 = 2[2J - [7J

(16.109)

Hermann, 1706

~=

2[3J

+ [7J

(16.110)

Hutton, 1776

~ = 3[4J + [20J + [1985J

(16.111)

Loney, 1893

~ = 4[5J - [70J + [99J

(16.112)

Euler, 1764

~ = 4[5J - [239J

(16.113)

Machin, 1706

~ = 5[7J + 2[79/3J

(16.114)

Euler, 1755 [114, p. 295]

~ = 6[8J + 2[57J + [239]

(16.115)

St0rmer, 1896

~ = 7[10J + 8[100J + [682J + 4[1000J + 3[1303J - 4[90109J - 2[500150J (16.116)

Wrench Jr. [41, (9-34)]

~ = 8[10] - [239] - 4[515]

(16.117)

Klingenstierna, c. 1730 [114, p. 296]

~ = 8[101/10J - [239J + 4[52525)

(16.118)

Gauss [56, II, p. 524]

~ = 12[18) + 8[57] - 5[239]

(16.119)

Gauss [56, II, p. 524]

22[28J

+ 2[443J -

5[1393J - 1O[11018J

(16.120)

Escott

~ = 12[38J + 20[57) + 7[239J + 24[268J

Gauss [56, II, p. 525]

(16.121)

16.

= 44[57] +

71'

Formula Collection

7[239] - 12[682] + 24[12943]

233

(16.122)

St0rmer, 1896

~ = 88[172] + 51[239] + 32[682] + 44[5357] + 68[12943]

(16.123)

St0rmer, 1896

~ = 88[192] + 39[239] + 100[515] - 32[1068] - 56[173932]

(16.124)

Arndt [7], 1993

~ = 322[577] + 76[682] + 139[1393] + 156[12943] +

+ 132[32807] + 44[1049433]

(16.125)

Arndt [7], 1993

~ = 1587[2852] + 295[4193] + 593[4246] + 359[39307] +

+ 481[55603] + 625[211050] - 708[390112]

(16.126)

Arndt [7], 1993

= 1074[4246] +

1257[5357] + 1731[6107] + 295[12943] +

+ 625[19703]- 481[32807]- 1042[39307] + 398[390112]

(16.127)

Arndt [7], 1993

~ = 7162[12943] + 3796[32807] + 2558[34208] +

+ 2729[44179] - 708[51387] + 2192[114669] - 2805[157318] - 3696[485298] - 2407[24208144]

(16.128)

Arndt [7], 1993

= 50539[51387] +

1555[114669] - 6601[157318] _ 20678[390112] - 5617[485298]- 64126[617427] +

+ 10958[1984933]- 30569[3449051] +
+ 23407[22709274] + 25433[24208144]

(16.129)

Arndt [7], 1993

~ = 36462[390112] + 135908[485298] + 274509[683982]4

_ 39581[1984933] + 178477[2478328] - 114569[3449051] - 146571[18975991] + 61914[22709274] - 6044[24208144]- 89431[201229582] - 43938[2189376182]

(16.130)

Arndt [7], 1993

~ = 446879[683982] + 172370[1635786] - 193720[1984933] +

+369078[2478328] + 18231[3014557] + 21339[3449051]- 154139[6225244]- 110109[18975991] + 80145[22709274]- 223183[24208144]- 107662[201229582]- 216308[2189376182] (16.131)
Arndt [7], 1993

~ = 872408[1984933] + 619249[2298668] + 369078[2478328] +

+
+
+

18231 [3014557] - 1217159[5033696] + 911989[6225244] +

783649[18975991]- 70886[22709274]- 374214[24208144]1044789[168623905] + 339217[201229582] - 446879[2848622638] +
402941[2189376182]
(16.132)

Arndt, [7], 1993

00

~=

[n2

n=l
Knopp

+ n + 1]

(16.133)

234

16.

Formula Collection

'If

00

~=

2: [F2n+1]

(16.134)

n=1

where Fi : Fibonacci numbers, Fl

1, F2

1, Fi+2

Fi+l

+ Fi, i 2::

Arndt, [7], 1994

3\1'5 - 5 ~
2
- L.J Hn [(3Hn+2

'4'If =

3
/
+ F 2n+2)
2]

(16.135)

n=1
Arndt, [7], 1994

Miscellaneous formulae
'If

1-i
2iln -1-'

(16.136)

+t

Fagnano [118, p. 199]

'If

[2(1/2)!]2

(1/2)!(-1/2)!(-1/2)
oo

where r(x) := fo e-

e-

dt for

= r 2 (1/2) = (1/2r(-1/2)? (16.137)

all real x > 0

Euler's Gamma function

2

= ((2)

(16.138)

where ((8)

'If

= lim

~~=1

';8

Riemann's zeta function

(16.139)

Inlcm(F1 , ... ,Fn )

n--+oo

where lem: least common multiple,

F i : Fibonacci numbers, F1 = 1, F2 = 1, Fi+2

Fi+1

+ F;, i 2::

Matiyasevich [84]

'If

-11

22
7

x4(1 - X)4 dx

(16.140)

1 + x2

Backhouse, [8], 1995

e+1=2+4~

1
~(2'1f'T')2+1

e-1

(16.141 )

Kempermann 1995 [73, p. 98]

i 6 = II
00

'If

k=O

2 (k_l)

(k+2)

27(k+~)(k+!)
0

Bellard, 1996 [19]

0]
11

(16.142)

16. 7r Formula Collection

4= ~ (I!

177r
360

!)2 .

k- 2

+ 2 + ... + k

235

(16.143)

k=l
A u- Yeung, 1993, after http://www . nas. nasa .gov/Pubs!NASnelis/97!OS/math. html

37 6
22:S0 = (2 (3)

+L

(1
1)2
+ "2 + ... + k .

00

k-

(16.144)

"=1

Bailey, http://www .nas. nasa.gov/PubsiNASnews/97 /05/math.html

Iterative algorithms
Algorithm (16.145) Archimedes, 250 Be, linear

(alg 16.145)

lni tialise:
ao := v'3
bo :=2

Iterate:
ale+1 := ale
bk+1 :=

+ ble

Jl

+a%+l

--+

6 2k+ 1/ 7r _

6.2"+1/ 7r +

--+

Then:
bk+l

62 k +1

<

<

ak+l

7r
6.2 k +1
Archimedes, c. 250

Be

Algorithm (16.146) Descartes, c. 1649, linear

(alg 16.146)

Initialise:
po :=.)2
qo := 1

Iterate:
PIe+1 := Pk+qk
2
qk+1 := VPk+ 1qle

Then:
_ 4_

qk+l

< 7r <

--+
1
--+

.1
7r

.1

_4_

Pk+l

Descartes, c. 1649 [114, p. 289]

Algorithm (16.147) Bit recursion

Initialise:
ao := tan(l)

Iterate:
:= (2ak)/(I- a%)

._ {I
b(x) .0

If x < 0
Else

Then:
.1
7r

,,",00

6k=0 27'+T
Plouffe from [36]
-

(alg 16.147)

236

16.

7r

Formula Collection

Algorithm (16.148) Gauss AGM

(alg 16.148)

Initialise:

ao := 1
bo := 1/V2
So := 0.5
Iterate:
ak+l
bk+1
C%+1
S"+1

+ bk)/2
v'akbk

:= (ak
:=

:= a%+l - b%+! = (ak+l - ak)2

:=

Sk - 2 k+ 1c%+1

(a n +bn )2

Then:

Pn
7r -

Pn

2s n

7r

7r2n+4e-7f2n+l

AGM2(1,1/V2)

Gauss, c. 1800, Brent, 1976, Salamin, 1976

Algorithm (16.149) Gauss AGM, Schonhage variant

(alg 16.149)

Initialise:

ao := 1
Ao := 1
Bo := 0.5
So := 0.5

Iterate:

+ Bk)/4

T := (Ak
bk
ak+!
Ak+l
Bk+l
Sk+1

Vlfk

:=
:=
:=
:=
:=

a%+l

An+Bn

(ak

+ b,,)/2

2(Ak+1 - T)
Sk + 2k+ 1(Bk+l - Ak+1)

Then:
pn

8n

7r

SchOnhage [105, p. 266]

Algorithm (16.150) Borwein, quadratic

(alg 16.150)

Initialise:

ao := V2
bo := 0
Po := 2 + V2

Iterate:
ak+l := (y'ak + vh)/2
bk+! := (y'ak(l + bk/(ak + bk)
Pk+1 := (Pk bk+1(1+ak+1/(1+bk+1)

Then:
IPn -

7r1

<

_1_

102 n
J. and P. Borwein, 1984 [30, p. 360]

--t

7r

16.

7r

Formula Collection

Algorithm (16.151) Borwein, cubic

237

(alg 16.151)

Initialise:

ao := 1/3
So

:=

(\1'3 -

1)/2

Iterate:
rk+1 := 3/(1 + 2(1 - SI)1/3)
Sk+1 := (rk+1 - 1)/2
ak+1 := r~+1 - 3k(r~+1 - 1)

1
71'

J. and P. Borwein, 1987 [13, p. 53]

(alg 16.152)

Algorithm (16.152) Borwein, quartic

Initialise:

4V2
V2 - 1

ao := 6 Yo :=

Iterate:
Yk+1 := 1- y~)-1/4 - 1)/1 - y~)-1!4 + 1) -4 0+
ak+1 := akl + Yk+1)2)2 - 2 2k +3Yk+1 1 + Yk+1)2 - Yk+1)
Then:

<

an -

7r-

::::

---+

1
71'

16 4k 2e- 4fi2 71'

J. and P. Borwein, 1987 [32, p. 170]

(alg 16.153)

Algorithm (16.153) Borwein, quintic

Initialise:
So

:=

5(j5 - 2)

ao

:=

1/2

Iterate:

+ x/z + I?)

Sk+1 := (25)/(Sk(Z
where

x:=..... -1
Sk

r/

+7
5
+ J~yn2---4'-x03)
5k (8&2- 5 + J~Sk-("-S""'%----=2S-k-+""-;:5"))

y:= (x - I?

z :=
ak+1

(~

= s~ak

(Y
-

Then:
an -

< 16 5 n e-7I'5

J. and P. Borwein, 1987 [33, p. 115]

238

16.

iT

Formula Collection

Algorithm (16.154) Borwein, nonie

(alg 16.154)

Initialise:

Iterate:

ao
ro

:=

1/3

:=

So

:= (1 -

(v'3-1)/2
r8)1/3

:= 1 + 2rk
u := (9rk(1+rk+rml/3

v := t + tu + u 2
m := (27(1 + Sk + s~/v
2

ak+l

S1<+1

+ 3 2k - 1 (1 - m)
:= (1- rd)/((t + 2u)v)
:=

mak

rk+l := (1 -

sD

J. and P. Borwein, 1987 [13, p. 53]

.!.
.".

17. Tables
17.1 Selected constants to 100 places (base 10)
~

3.1415926535 8979323846 2643383279 5028841971 6939937510

5820974944 5923078164 0628620899 8628034825 3421170679

1/~

0.3183098861 8379067153 7767526745 0287240689 1929148091

2897495334 6881177935 9526845307 0180227605 5325061719

6/~2

0.6079271018 5402662866 3276779258 3658334261 5264803347

9293073654 1913650387 2577341264 7147255643 5537310256

~2

9.8696044010 8935861883 4490999876 1511353136 9940724079

0626413349 3762200448 2241920524 3001773403 7185522318

~3

= 31.0062766802

~4

= 97.4090910340

9982017547 6315067101 3952022252 8856588510

7694144538 1038063949 1746570603 7566701032 6028861930
0243723644 0332688705 1112497275 8567268542
1691467859 3899708554 5682719619 0121867234 7529925509

1.7724538509 0551602729 8167483341 1451827975 4945612238

7128213807 7898529112 8459103218 1374950656 7385446654

1.4645918875 6152326302 0142527263 7903917385 9685562793

7174357255 9371383936 4979828626 6145682067 8203538208

1.3313353638 0038971279 7534917950 2808533093 6622381810

4258453707 4828667007 6101723561 4968245891 0567069459

2.7182818284 5904523536 0287471352 6624977572 4709369995

9574966967 6277240766 3035354759 4571382178 5251664274

~e

= 22.4591577183

e~

= 23.1406926327

6104547342 7152204543 7350275893 1513399669

2249203002 5540669260 4039911791 2318519752 7271430315
7926900572 9086367948 5473802661 0624260021
1993445046 4095243423 5069045278 3516971997 0675492196

e~/4

2.1932800507 3801545655 9769659278 7382234616 3764199427

2334858015 9186570268 6418923693 4126522812 5781694047

In~

1.1447298858 4940017414 3427351353 0587116472 9481291531

1571513623 0714721377 6988482607 9783623270 2754897077

log2~

= 1.6514961294 7231879804 3279295108 0073350184 7692676304

log10~

= 0.4971498726 9413385435 1268288290 8988736516 7832438044

1529406788 5154881029 6358454143 8960264792 8098541017

2446134053 4999249471 1208955267 4655547386 4642912223
e~.,(f63 =

26253741264 0768743.
9999999999 9925007259 7198185688 8793538563 3733699086
2707537410 3782106479 1011860731 2951181346 1860645041

240

17. Tables

17.2 Digits 0 to 2,500 of

~

1T

(base 10)

= 3.1415926535 8979323846 2643383279 5028841971 6939937510

(0051)
(0101)
(0151)
(0201)
(0251)
(0301)
(0351)
(0401)
(0451)
(0501)
(0551)
(0601)
(0651)
(0701)
(0751)
(0801)
(0851)
(0901)
(0951)
(1001)
(1051)
(1101)
(1151)
(1201)
(1251)
(1301)
(1351)
(1401)
(1451)
(1501)
(1551)
(1601)
(1651)
(1701)
(1751)
(1801)
(1851)
(1901)
(1951)
(2001)
(2051)
(2101)
(2151)
(2201)
(2251)
(2301)
(2351)
(2401)
(2451)

5820974944
8214808651
4811174502
4428810975
4564856692
7245870066
7892590360
3305727036
0744623799
9833673362
6094370277
0005681271
1468440901
4201995611
5187072113
5024459455
7101000313
5982534904
1857780532
3809525720
0353018529
5574857242
8175463746
8583616035
9448255379
9331367702
2533824300
6782354781
5570674983
3211653449
6369807426
8164706001
1613611573
4547762416
5688767179
8279679766
7392984896
0674427862
4677646575
9465764078
2671947826
4962524517
5709858387
6868386894
8459872736
4390451244
4169486855
0168427394
6456596116

5923078164
3282306647
8410270193
6659334461
3460348610
0631558817
0113305305
5759591953
6274956735
4406566430
0539217176
4526356082
2249534301
2129021960
4999999837
3469083026
7838752886
2875546873
1712268066
1065485863
6899577362
4541506959
4939319255
6370766010
7747268471
8989152104
3558764024
6360093417
8505494588
8720275596
5425278625
6145249192
5255213347
8625189835
0494601653
8145410095
0841284886
2039194945
7396241389
9512694683
8482601476
4939965143
4105978859
2774155991
4469584865
1365497627
5848406353
5226746767
3548862305

0628620899
0938446095
8521105559
2847564823
4543266482
4881520920
4882046652
0921861173
1885752724
8602139494
2931767523
7785771342
4654958537
8640344181
2978049951
4252230825
5875332083
1159562863
1300192787
2788659361
2599413891
5082953311
0604009277
4710181942
0404753464
7521620569
7496473263
2164121992
5869269956
0236480665
5181841757
1732172147
5741849468
6948556209
4668049886
3883786360
2694560424
0471237137
0865832645
9835259570
9909026401
1429809190
5977297549
8559252459
3836736222
8079771569
4220722258
8895252138
7745649803

8628034825
5058223172
6446229489
3786783165
1339360726
9628292540
1384146951
8193261179
8912279381
6395224737
8467481846
7577896091
1050792279
5981362977
0597317328
3344685035
8142061717
8823537875
6611195909
5338182796
2497217752
6861727855
0167113900
9555961989
6208046684
6602405803
9141992726
4586315030
9092721079
4991198818
4672890977
7235014144
4385233239
9219222184
2723279178
9506800642
1965285022
8696095636
9958133904
9825822620
3639443745
6592509372
8930161753
5395943104
6260991246
1435997700
2848864815
5225499546
5593634568

3421170679
5359408128
5493038196
2712019091
0249141273
9171536436
9415116094
3105118548
8301194912
1907021798
7669405132
7363717872
6892589235
4771309960
1609631859
2619311881
7669147303
9375195778
2164201989
8230301952
8347913151
8890750983
9848824012
4676783744
2590694912
8150193511
0426992279
2861829745
7509302955
3479775356
7727938000
1973568548
0739414333
2725502542
6085784383
2512520511
2106611863
4371917287
7802759009
5224894077
5305068203
2169646151
9284681382
9972524680
0805124388
1296160894
8456028506
6672782398
1743241125

17.3 Digits 2,501 to 5,000 of 'IT (base 10)

17.3 Digits 2,501 to 5,000 of

(2501)
(2551)
(2601)
(2651)
(2701)
(2751)
(2801)
(2851)
(2901)
(2951)
(3001)
(3051)
(3101)
(3151)
(3201)
(3251)
(3301)
(3351)
(3401)
(3451)
(3501)
(3551)
(3601)
(3651)
(3701)
(3751)
(3801)
(3851)
(3901)
(3951)
(4001)
(4051)
(4101)
(4151)
(4201)
(4251)
(4301)
(4351)
(4401)
(4451)
(4501)
(4551)
(4601)
(4651)
(4701)
(4751)
(4801)
(4851)
(4901)
(4951)

1507606947
2287489405
9009714909
2265880485
5428584447
2135969536
0374200731
8687519435
8191197939
9839101591
5679452080
6638937787
0306803844
6660021324
1005508106
4011097120
2305587631
7321579198
7229109816
5725121083
6711136990
3515565088
8932261854
6549859461
2332609729
5836041428
1809377344
4492932151
2131449576
2807318915
6655730925
4769312570
3348850346
8455296541
7002378776
1127000407
6343285878
2021149557
0990796547
2996148903
9389713111
0480109412
8530614228
7716692547
9769265672
8888007869
6171196377
4252308177
6222247715
7805419341

9451096596
6010150330
6759852613
7564014270
9526586782
2314429524
0578539062
0643021845
9520614196
9561814675
9514655022
0830390697
7734549202
3408819071
6587969981
6280439039
7635942187
4148488291
9091528017
5791513698
8658516398
4909989859
8963213293
6371802709
9712084433
1388303203
4030707469
6084244485
8572624334
4411010446
4711055785
5863566201
1136576867
2665408530
5913440171
8547332699
5698305235
6158140025
3761255176
0463994713
7904297828
1472213179
8137585043
4873898665
1463853067
2560290228
9213375751
0367515906
8915049530
4473774418

7r

(base 10)

0940252288
8617928680
6554978189
4775551323
1051141354
8493718711
1983874478
3191048481
6342875444
1426912397
5231603881
9207734672
6054146659
0486331734
6357473638
7595156771
3125147120
6447060957
3506712748
8209144421
3150197016
9823873455
3089857064
8199430992
5732654893
8249037589
2112019130
9637669838
4189303968
8232527162
3763466820
8558100729
5324944166
6143444318
2749470420
3908145466
8089330657
0126228594
5675135751
2962107340
5647503203
4764777262
0633217518
4949450114
3609657120
4721040317
1495950156
7350235072
9844489333
4263129860

7971089314
9208747609
3129784821
7964145152
7357395231
0145765403
0847848968
0053706146
0643745123
4894090718
9301420937
2182562599
2520149744
6496514539
4052571459
5770042033
5329281918
5270695722
5832228718
0067510334
5151168517
2833163550
2046752590
4488957571
8239119325
8524374417
2033038019
9522868478
6426243410
0105265227
6531098965
3606598764
8039626579
5867697514
5622305389
4645880797
5740679545
1302164715
7829666454
4375189573
1986915140
2414254854
2979866223
6540628433
9180763832
2118608204
6049631862
8354056704
0963408780
8099888687

5669136867
1782493858
6829989487
3746234364
1342716610
5902799344
3321445713
8067491927
7181921799
6494231961
6213785595
6615014215
2850732518
0579626856
1028970641
7869936007
2618612586
0917567116
3520935396
6711031412
1437657618
7647918535
7091548141
2828905923
9746366730
0291327656
7621101100
3123552658
7732269780
2111660396
2691862056
8611791045
7877185560
5661406800
9456131407
2708266830
7163775254
5097925923
7791745011
5961458901
2870808599
5403321571
7172159160
6639379003
7166416274
1900042296
9472654736
0386743513
7693259939
4132604721

241

242

17. Tables

17.4 Digits 0 to 2,500 of 1T' (base 16)

~

(0051)
(0101)
(0151)
(0201)
(0251)
(0301)
(0351)
(0401)
(0451)
(0501)
(0551)
(0601)
(0651)
(0701)
(0751)
(0801)
(0851)
(0901)
(0951)
(1001)
(1051)
(1101)
(1151)
(1201)
(1251)
(1301)
(1351)
(1401)
(1451)
(1501)
(1551)
(1601)
(1651)
(1701)
(1751)
(1801)
(1851)
(1901)
(1951)
(2001)
(2051)
(2101)
(2151)
(2201)
(2251)
(2301)
(2351)
(2401)
(2451)

3.243F6A8885
2EFA98EC4E
29B7C97C50
A698DFB5AC
F12C7F9924
574E69A458
4AEE7B54A4
FOCA417918
D71577C1BD
48986263E8
86AF7C72E9
169B87931E
6B4BB9AFC4
EC8032EF84
ACC50F6D6F
5E21C66842
960FA728AB
F1651D39AF
A5C33B8B5E
62363F7706
49F1C09B07
794C3B976C
246368FB6F
2CCC814544
COCBA85745
Al00C6402C
16DFFD616B
483EOODF82
287EFFC3AC
FOllA010FA
4565D28E49
E8EF20CADA
EAAD8E716B
F6E2FBF212
F191B3A8Cl
OFB56F74E8
D2ADA8D916
B5FA86C754
7E4900250E
lE5563911D
8326037662
5100529A53
6FB5571BE9
2EC5855664
B5B32944DB
EDB266ECAA
4C29A05913
D6A1D29C07
6382E9C602
7F358452AO

A308D31319
6C89452821
DD3F84D5B5
2FFD72DBDO
A19947B391
FEA3F4933D
lDC25A59B5
B8DB38EF8E
314B2778AF
144055CA39
93B3EE1411
AFD6BA336C
BFE81B6628
5D5DE98575
F383F44239
F6E96C9A67
5133A36EEF
017666CA59
BEE06F75D8
lB.FEDF7242
5372C98099
EOBD04C006
AF3E6C53B5
AF5EBD09BE
C8740FD20B
7279679F25
152F501EC8
9E5C57BBCA
6732C68C4F
3D98FD2183
BC4BFB9790
36774C01DO
93D5AOD08E
2B648888B8
AD2F2F2218
18ACF3D6CE
5FA2668095
42F5FB9D35
2D2071B35E
59DFA6AA78
95CFA911C8
2915D60F57
lFF296EC6B
53B02D5DA9
75092EC419
8C71699A17
40E4183A3E
EFE830F54D
lECC5E0968
E286B79C53

8A2E037073
E638D01377
B547091792
lADFB7B8El
6CF70801F2
7EOD95748F
9C30D5392A
79DCB0603A
2FDA55605C
6A2AAB10B6
636FBC2A2B
24CF5C7A32
219361D809
B1DC262302
2EOB4482A4
OC9C61ABD3
OB6C137A3B
3E82430E88
85C1207340
9B023D37DO
lB7B25D479
BAC1A94FB6
1339B2EB3B
E3D004DE33
5F39B9D3FB
FEFB1FA3CC
AD0552AB32
6F8CA01A87
5573695B27
B84AFCB56C
E1DDF2DAA4
7E9EFE2BFl
D1DOAFC725
12900DF01C
BEOE1777EA
89E299B4A8
770593CC73
CFEBCDAFOC
226800BB57
C14389D95A
19684E734A
3FBC9BC6E4
2AODD915B6
9F8FA108BA
2623AD6EA6
FF5664526C
3F54989A5B
2D38E6F025
6B3F3EBAEF
05AA500737

44A4093822
BE5466CF34
16D5D98979
AFED6A267E
E2858EFC16
728EB65871
F26013C5D1
180E6C9EOE
60E65525F3
B4CC5C3411
A9C55D7418
5381289586
CCFB21A991
EB651B8823
84200469C8
88F06A51AO
E4BA3BF050
8CEE861945
lA449F56Cl
D724DOOA12
D8F6E8DEF7
409F60C45E
52EC6F6DFC
4AFD660F28
DB5579COBD
8EA5E9F8DB
3DB5FAFD23
562EDF1769
BOBBCA58C8
2DD1D35B9A
CB7E3362FB
lFB495DBDA
E08E3C5B2F
4FAD5EA068
752DFE8B02
4FEOFD13EO
14211A1477
7B3E89AOD6
B8EOAF2464
537F207D5B
41B3472DCA
2B60A47681
636521E7B9
47996E8507
B049A7DF7D
C2B19EEl19
429D656B8F
5DC14CDD20
C93C971814
3E07841C7F

299F31D008
E90C6CCOAC
FB1BD1310B
96BA7C9045
636920D871
8BCD588215
B023286085
8BB01E8A3E
AA55AB9457
41E8CEA154
31F6CE5C3E
773B8F4898
487CAC605D
893E81D396
F04A9E1F9B
D2D8542F68
7EFB2A98Al
6F9FB47D84
6AA64ED3AA
48DBOFEAD3
E3FE501AB6
5C9EC2196A
511F9B3095
07192E4BB3
lA60320AD6
3222F83C75
876053317B
DBD542A8F6
E1FFA35DB8
53E479B6F8
1341CEE4C6
4DAE909198
8E7594B78F
8FC31CD1CF
lFA1E5AOCC
B77CC43B81
E6AD206577
411BD3AE1E
369BF009B9
A202E5B9C5
7B14A94A1B
E6740008BA
F9B6FF3405
6A4B7A70E9
9CEE60B88F
3602A57509
E4D699F73F
868470EB26
6B6A70A168
DEAE5C8E7D

17.5 Digits 2,501 to 5,000 of 1r (base 16)

17.5 Digits 2,501 to 5,000 of 7r (base 16)

(2501)
(2551)
(2601)
(2651)
(2701)
(2751)
(2801)
(2851)
(2901)
(2951)
(3001)
(3051)
(3101)
(3151)
(3201)
(3251)
(3301)
(3351)
(3401)
(3451)
(3501)
(3551)
(3601)
(3651)
(3701)
(3751)
(3801)
(3851)
(3901)
(3951)
(4001)
(4051)
(4101)
(4151)
(4201)
(4251)
(4301)
(4351)
(4401)
(4451)
(4501)
(4551)
(4601)
(4651)
(4701)
(4751)
(4801)
(4851)
(4901)
(4951)

44EC5716F2
lA3CB574B2
22F547013A
D0030EECC8
B3314E548B
795679B072
5512721F2E
9F9ABCE94B
18473215D9
02133AE4DD
D7A3C76B3C
153C6E86E3
06FA2965DC
6AC6150EBA
1939260F19
BC4595A67B
218568AB98
2829076170
AA0363CFB5
622CA79CAB
BB5040685A
9995F7997E
OE358829C7
83C3FF1AC2
F2EF34C6FF
14203E13EO
EF6ABBB565
816250CF62
749247848A
0058428D2A
D65FECF16C
07808419F8
ABFC800BCA
D3105588CD
F16DFF2015
8140F7F64C
0068D40824
4614214F74
BFBC09EC03
2547E6ABCA
1462D74869
067AF4D6B6
D9F385B9EE
E397B23A6E
B397454056
4B55A867BC
5EF47E1C90
Cl1548E4C6
79A45D886E
34602A9C60

B8B03ADA37
25837A58DC
E5E58137C2
C73EA4751E
384F6DB908
BCAF89AFDE
6B7124501A
7D8CEC7AEC
08DD433B37
71DFF89El0
11183B5924
4570EAE96F
B999E71DOF
94E2EA78A5
C279605223
C883B17F37
02EECEA50F
ECDD477561
735C904C70
5CABB2F384
323C2AB4B3
623D7DA8F8
E61FD696DE
4696CDB30A
EAFE28ED61
45EEE2B6A3
4F3B1D41CD
A1F25B8D26
OB5692B285
OC55F5EA1D
223BDB7CDE
509EE8EFD8
DC9E447A2E
675FDA79E3
3E21E78FBO
261C946929
713320F46A
BF8B88404D
BD97857FAC
OA9A285078
00680ECOA4
AACE1E7CD3
39D7AB3B12
FA74DD5B43
ACBA489527
Al159A58CC
29317CFDF8
6D2248C113
17325F51EB
DFF8E8A31F

F0500CODFO
0921BDD191
DADCC8B576
41E238CD99
6F420D03F6
9A771FD993
DDE69F84CD
3ADB851DFA
24C2BA1612
314E5581AC
A509F28FE6
B1860E5EOA
803E89D652
FC3C531EOA
A708F71312
Dl018CFF28
DB2F953B2A
9F151013CC
A239D59E9E
6E648B1EAF
319EE9D5CO
37889A97E3
DFA17858BA
EB532E3054
EE7C3C735D
AAABEADB6C
2105D81E79
46FC8883AO
095BBFOOAD
ADF43E233F
3759CBEE74
5561D99735
C3453484FD
674340C5C4
3D4AE6E39F
34411520F7
43B7D4B750
95FC1D96B5
6DD031CB85
25530429F4
27A18DEE4F
375FECCE78
4E8B1DC9FA
326841E7F7
55533A3A20
A9296399El
E80204272F
3FC70F86DC
D59BCOD1F2
636C1BOE12

lC1F040200
13F97CA92F
349AF3DDA7
3BEAOE2F32
OA04BF2CB8
0810B38BAE
877A584718
63094366C4
A14D432A65
77D65Fll19
ED97F1FBFA
5A3E2AB377
66C8252E4C
2DF4F2F74E
B6EBADFE6E
C332DDEFBE
EF7DAD5B6E
A830EB61BD
OBCBAADE14
19BDFOCAAO
21B8F79B54
2D7711ED93
9957F584A5
8FD948E46D
4A14D9E864
4F15FACB4F
9E86854DC7
C1C7B6A37F
19489D1462
70613372FO
604085F2A7
A969A7AAC5
D567050E1E
3465713E38
2BDB83ADF7
7602D4F7BC
0061AF1E39
91AF70F4DD
0496EB27B3
OA2C86DAE9
3FFEA2E887
A399406B2A
F74B6D1856
CA7820FBFB
838D87FE6B
DB33A62A4A
7080BB155C
07F9C9EE41
BCC18F4111
B4C202E132

B3FFAEOCF5
F694324773
A94461460F
80BBAl183E
129024977C
12DCCF3F2E
7408DA17BC
64C3D2EF1C
C451509400
9B043556Fl
9EBABF2C1E
lFE71C4E3D
C9789Cl0B3
A7361D2B3D
EAC31F66E3
6C5AA56558
2F841521B6
960334FE1E
EECC86BC60
2369B9655A
OB19875FAO
5F16681281
lB2272639B
BC312858EB
B7E342105D
DOC742F442
E44B476A3D
1524C369CB
B17423820E
928D937E41
CE77326EA6
OC06C25A04
9EC9DB73DB
D83D28F89E
E93D5A6894
F46B2ED4A2
F62E972445
D366A02F45
55FD3941DA
B66DFB68DC
AD8CB58CEO
4220FE9E35
26A36631EA
OAF54ED8FE
A9B7D09695
563F3125F9
05282CE395
041FOF4047
3564257B78
9EAF664FDl

243

244

17. Tables

17.6 Continued fraction elements 0 to 1,000 of 11'

= 3+ [ 7 15 1
(0021) 84
2 1
(0041)
1 6 8
(0061)
8
1 1
(0081)
1 22 1
(0101)
5
4 1
(0121)
7 3 3
(0141)
2 18 1
(0161)
1 2 13
(0181)
1 1 1
(0201)
1 23 1
(0221)
4 127 14
(0241)
1 29 3
(0261)
2
1 12
(0281)
1 48 16
(0301) 20
1 1
(0321)
3
6 11
(0341) 24 4 4
(0361) 58
5 1
(0381)
1 1 1
(0401)
1 13 1
(0421)
1
1 1
(0441)
7
2 1
(0461)
1 1 3
(0481)
5 2 2
(0501)
2 2 1
(0521)
2
1 1
(0541)
1 2 3
(0561) 11
1 1
(0581) 26
1 1
(0601) 106
2 2
(0621)
2
1 4
(0641)
1 1 3
(0661)
1 29 4
(0681)
7
1 2
(0701) 13
1 5
(0721)
5
1 1
(0741)
1 4 2
(0761)
1 3 1
(0781)
3
137
(0801)
1 12 1
(0821) 15 8 8
(0841)
235
(0861)
1
1 3
(0881)
2
262
(0901)
1
3 2
(0921)
3
2 1
(0941)
2
1 1
(0961)
1 465
(0981)
211

7r

292
1 1
1 15 3
1 7 1
2
1 6
2 2 1
2 2 8
1 1 7
3 7 30
1 2 1
4
1 1
15
1 3
5
1 3
1 1 2
1 4 1
1 4 5
5 4 1
4
3 1
5 2 1
2
1 2
3
1 1
1 1 3
4 15 1
94 55 1
50
2 16
2 28 1
1 19 2
2
1 1
1 3 2
2 1 2
3 2 1
131
23
1 18
1

29
4
4

7 17
3 5
1 1
1 6

16 125 1
1
6
4
2
47
4
20
2
1
3
1
2
2

1 1
1 3
2 4

2 2
1 3
2 1
1 1
2 7

1 2
13 1
2 3
1 1
4 1
1 5
2 4
1 1
4 1
1 5
7 1
13 7
2 1
1 1
2 1
436 8
1 1
4 1
1 1
1 2
24 1
4 44
1 2
5 1
1 13
7 1
1 1
2 2
1 2
1 1
12 10
2 10
1 1
3 1
130
1 1
4 1
5 10
2 2
1 3
13 12
2 12
1

61

159
1 8

6 1

2 1
3 2

4 4
2 1

1 4

2 3

1
4
7
5
2
2
9
1
12
3
16
9
3
1
1
1

3
1 14 2 1 1
2
6 6 99 1 2
1
2 1 1 12 1
2
2 3 1 2 4
24
1 2 1 3 1
2
26 1 4 1 1
7
2 3 1 57 1
3
3 3 1 2 8
1
1 3 3 28 1
2
1 6 1 4 1
1
2 1 21 2 1
1
1 1 1 1 5
1
1 1 3 1 1
7
1 1 2 1 11
4
3 1 2 3 1
2
2 1 1 1 1
2 5
4 6 9 1 5
6 1
1 1 4 3 2
2 2
7 1 15 1 4
1 1
1 1 1 9 1
2 4
10 5 12 3 3
1 4 20776 1 1 1 1
1 1
3 1 1 32 5
2 1
4 6 3 1 3
1 5
43 1 4 3 5
72 3
1 2 3 7 11
1 33
7 19 1 19 3
2 4
1 1 1 4 2
2 1
7 2 27 1 1
1 1
15 1 36 4 2
7 1
2 1 111
3 1
6 6 13 8 6
1 2
5 5 1 1 2
141 1
2 7 7 2 2
1 12
4 18 10 2 8
11 2
4 2 1 1 3
2 1
19 1 4 1 1
4 14
1 1341 1 4
2 1
4 1 14 1 2
425
427 2 7 1
1 3
1 6 1 133
2 12
1 12 1 2 5
1 9
14 1 1 3 2
376 2
107 1 10 3 2
6 1
111 2 8
1 1
516 1 2
2 5
4 7 1 4 2
2 2
512 1 1
3 1
1 5 1 1 1
1 2
1 2 4 2 1

2 2 2
2 6 3
1 1 3
4 16 1
2 1 1
8 2 42
18 1 9
1 1 2
10 3 2
120 2 1
1 2 9
4 1 1
10 3 1
3 1 7
2 2 1
1 5 1
1 5 15
2 1 1
8 1 1
4 3 15
21 1 2
1 1 1
1 14 1
3 1 2
3 1 4
1 2 1
1 4 1
3 1 1
6 2 1
2 1 22
824
2

2
5
1
161
10
2
19
1
20
1
1
3
3
1
2
2
1
2
4
1
1
23
1
2
1
1
1
1
9
2

1
1
1
45
2
1
1
15
1
3
6
1
1
4
5
1
11
1
2
2
34
1
1
2
1
2
1
1
6
1

5
2

3
1
6
1
4

11 1 6
1
711
7
122 2
3 12 1
1 39
214
1 10
181
1
2
818 3
3
126 1
1
1 5 3
1
5
3 1 1
1
1
1 9 3 22 13
231 1 2 10
273 3
5 42
6 1 1
1
1
261
1
2
10 1 6
1 129
112 2
1
227 6
2
1

17.7 Continued fraction elements 1,001 to 2,000 of 11"

17.7 Continued fraction elements 1,001 to 2,000 of

(1001) 193
(1021)
2
(1041) 13
(1061)
6
(1081)
1
(1001)
7
(1121)
1
(1141)
2
(1161)
1
(1181)
1

1
8
2
1
1
2
1
2
6
4

3
1
1
4
1
2
3
1
1
1

9
8
3
12
4
2
4
4
1
11

1
1
1
1
1
1
1
1
2
2

3
2
3
1
3
1
60
7
1
4

(1201) 15 8
(1221) 156 88
(1241)
2 11
(1261)
1 1
(1281)
3 1
(1301) 11 3
(1321)
3 2
(1341)
5 1
(1361)
9 2
(1381)
1 1

8
1
1
1
7
1
2
2
1
2

2
1
5
4
13
7
41
2
5
1
3
2
1
3
1
5
2
1
1 34

13
1
7
1
9
1
2
18
5
1

35 2 1
8 1
3 6 3
1 3
15 2 1
1 2
1 1 13
1 3
4 1 1
4 3
1 1 2 114 12
3 72 2
1 1
3 1 2
1 5
7 1 8
1 1
10 2 1
1 13
3
3
1
1
1
1
1
1
4
5

5 1
1 4
63 37
12 1
2 2
2 1
2 1
4 3
93 1
6 1

(1401) 26 1
1 4 12 1 3 2 1
(1421) 20 1 17 1
6 16 3 3 2
(1441)
2 1
2 1 45 3 1 1 3
(1461)
8 2 47 2
3 1 1 1 15
(1481)
2 1 161 1 26 2 1 2 1
(1501)
5 16
3 1
1 1 1 1 5
(1521)
9 5
2 3
2 5 7 1 32
(1541)
9 1
7 1 102 53 1 1 1
(1561)
4 1
1 3
2 3 1 1 2
(1581)
1 1
1 22
1 2 3 1 3
(1601)
(1621)
(1641)
(1661)
(1681)
(1701)
(1721)
(1741)
(1761)
(1781)

3
1
1
4
1
4
1
1
1
5

1 63 1
3
1 1
3
1 39
7
1 17
2
2 2
1
1 2
1
1 25
1
1 1
1
1 1
2
1 1

1
1
1
1
12
2
1
1
4
1

1
1
12
14
1
1
5
1
1
1

2
2
12
2
7
14
4
1
1
2

1
9
86
1
1
5
1
1
2
1

1 1
1 1
4 1
4 1
1 39
4 1
1 50
1 1
5 4
1 1
6
1
1
1
1
2
14
1
4
3

5 1 2
2 2 18
1 2 1
9 1 8
1 1 1
1 2 1
13 1 5
3 4 2
3 1 1
26 2 24

1
7
1
2
5
2
1
4
1
4

1
1
4
1
2
3
1
1
1
7

9
2
5
1
1
2
1
2
1
15

10 123 3
4
2 1
1
1 2
10
1 1
2
1 3
1
1 1
1 58 7
1
1 2
2 73 1
1 16 28

1
2
2
1
3
1
1
5
1
17

1
1
1
1
1
1
3
21
1
1

1
2
3
1
44
1
94
1
2
67

1
1
1
2
14 18
1
1
2 10
2
5
2
4
1 10
9
2
5
7

8
8
9
3
1
8
3
41
1
1

1
12
3
1
14
3
4
1
2
23

3
1
2
1
1
2
5
5
1
3

3
53
2
10
2
1
1
1
1
3

1
1
2
1
1
2
1
4
1
1

(1801)
1 4
1 5
1 1 8
(1821)
2 1
3 1 11 2 10
(1841)
3 4
1 18
3 9 7
(1861)
4 1 54 2 196 10 3
(1881)
1 3
6 3 16 1 31
(1901)
3 1
2 1
1 1 4
(1921) 16 4
1 2
4 1 4
(1941)
1 30
1 1
3 2 2
(1961)
3 6
1 1
2 4 4
(1981)
1 6
6 2
1 1 10

1
1
5
1
1
1
1
1
24
6

1
4
1
32
5
13
1
1
1
4

5
1
2
40
1
88
2
10
1
1

2 10 1
2 3 1
1 7 1
55 1 5
442 4
1 1 1
4 1 1
1 7 1
1 1 1
3 9 1

1
1
23
1
1
14
9
6
3
2

7i"

6 6
1 1
2 1
7 1
1 4
1 2
1 1
2 6
1 1
2 3

4 59 4
3
2 1
4
2 18
5
1 1
5
2 1
3
1 3
2
1 1
1 148 1
3
1 1
2
5 5

9 3
2 41
1 16
15 1
1 51
4
9
3 1
1 3
5
1
2 1
4 2
4 14
2 2
1 18 528 12
63 1 97 1
4
4
4 1
7 1
3 12
15 2
8 2
2
3
6 1
1 14
1 80
2 2
4 3
1
1

25
2
1
1
3
34
1
1
1
7

2
18
2
1
6
1
17
4
1
2

3 56
13
1
2
2
2
9
1
1
1 19
1
1
6
2
1
1
1
1

245

2
5
2
1
4
10
1
1
11
1

2
2
3
12
1
5
3
2
1
1

10
5
6
2
1
1
1
1
3
2
2
3
43
4
1
13
4
3
1
1
1
4
1 155
6 23
1
6
1 11

2
5
1
1
7
2
1
1
2
1

1
1
4
2
3
1
2
8
3
57

2
1
6
3
2 217
3
3
10
1
3 27
2
1
1 35
1
1
16
1

3 1
4 31
2 1
2 2
1 3
1 1
3 1
1 2
1 2
1 1

1
6
1
1
3
3
1
1
2
5

A. Documentation for the hfloat Library

The following is (almost) the documentation that comes with the hfloat
package1 .

A.I What hfloat is (good for)

hfloat (for 'huge floats') is a library package for doing calculations

with floating point numbers of extreme precision. It is optimised for

computations with 1000 ... several million digits. The computations
can be done in (almost) arbitrary radix.
The library contains routines for addition, subtraction, multiplication, division, n-th power, square root, n-th root, logarithm, exponentiation and many more.
There are implementations of several superlinear converging algorithms for the computation of pi=3.14159265 ... (in src/pi/). The
computation of 1 million decimal digits of pi takes about 2 minutes
on an AMD Athlon/800.
Code examples for the usage of the library are in examples/.
Code for a binary that collects the pi algorithms can be found in
calcpi/.
Included is the fxt-library, containing many FFT-implementations,
code for convolution, correlation, spectrum and much more.
High precision computations test your systems reliability (hardware
and compiler): every little error results in garbage digits.
Digits of appropriate constants can be used as high quality 'random'
numbers ego for cryptographic stuff.
You may (and are encouraged to) use the fast multiplication in your
own noncommercial bignum software.
hfloat is online at http://www.jjj.de/hfloat/ , the author of hfloat is Jorg
Arndt.

248

A. Documentation for the hfloat Library

A.2 Compiling the library

hfloat is developed under Linux with GNU C It has been compiled
under Linux on intel (386, 486, 586 and compatibles and ia64), power
pc, alpha and 8/390. Compilation under other UNIXes should be possible with few or no changes, especially if GNU C is used. Basically
you need a decent make-utility and a C-compiler 2 , GNU C and GNU
make are strongly recommended. To compile the library
Type 'make dep' (for dependency files)
Type 'make lib' to make the library

A.3 Functions of the hfloat library

Member functions of the hfloat class are declared in src/include/
hfloat.h
Static member junctions, they affect some property of the whole
hfloat class:
hfloat: : default_prec 0 returns the default precision that can be
used (unit is LIMBs)3;
hfloat: : defaul Lprec (unsigned long m) sets the default precision m;
call it once at the beginning of your code
hfloat: : radix 0 returns the radix, which is the same for all
hfloats;
use hfloat: : radix (unsigned long r) to set the radix.
Note that after a radix change all hfloats contain garbage data in
their mantissa so they must be reassigned new values before they
are read.
normally you will use hfloat: : radix (unsigned long r) once at
the beginning of your code

N onstatic member junctions, they affect one particular instance of

an hfloat:
the constructor hfloat 0 creates an hfloat with
default_precision()
2

expect problems with crippled Microsoft compiler-oids

for the notion of a LIMB see page 251

A.3 Functions of the hfloat library

249

the constructor hfloat (unsigned long n) creates an hfloat with

n LIMBs
Create hfloats like hfloat a; (for default precision) or
hfloat a(1024); (for a precision of 1024 LIMBs).
Direct initialization like hfloat a = 1234; is not possible because
the constructor hfloat(unsigned long) has the explicit modifier.
Instead say: hfloat a; a = 1234;
the copy constructor hfloat (const hfloat &h) creates a copy of
the hfloat h
the( assignment) operator
for arguments (i.e. right hand side quantities) int, long, unsigned
long, double, strings (i.e. char *) and hfloat,
sizeO returns the size of the mantissa (in LIMBs)
size (unsigned long s) resizes the mantissa to s LIMBs
precO returns the current working precision (in LIMBs)
prec(unsigned long p) sets the working precision to p LIMBs, resizes if necessary.
exp 0 returns the exponent (with respect to LIMBs);
exp (long) sets the exponent
signO returns +1,0,-1 as usual;
sign(int s) sets the sign
Operators for hfloats are declared in srcl include/hfloatop. h:
the shortcut operators +=,
*=, 1=
for right arguments long and hfloat,
comparison operators
! = , >= , <= , < ,
for comparisons of hfloats with hfloats or longs
comparison operators < , >
for comparisons of hfloats with doubles

>

Functions for hfloats are declared in src/include/hfloatfu.h:

functions of the type func(src,result)
where func E inv, sqr, sqrt, isqrt, cbrt, log4, exp, ...
functions of the type f(src1,src2,result)
where func E add, sub, mul, di v, pow, root, iroot, log, ...
The function names should be self-explanatory.
Print hfloat. h, hfloatfu. hand hfloatop. h now!
4

logarithm to base 2.71828 ...

250

A. Documentation for the hfloat Library

There are also the binary operators +,

* , / and functions
that return a hfloat, like fune (sre) where fune E inv, sqrt, ...
but do not use them if you are after performance: they create temporary hfloats and are there only to allow lazy coding.
E.g. instead of
x = a + b,
use
x += b,
x = a,
or
add(a, b, x) ;
Instead of
x = exp(a);
use
exp(a,x);
Note that the result is always the rightmost argument.

A.4 Using hfloats in your own code

In order to write your own code that uses hfloats
you must #inelude sre/inelude/hfloat.h to get the functions of
the hfloat lib.
use hfloat: : default_pree (n) where n is the precision in LIMBs
(use a power of two)
use hfloat : : radix (rx) where rx is the radix (use 10000 for decimal
numbers or 65536 for hex numbers)
use hfverbosi ty: : telLall 0 if you like to have many operations
echoed, hfverbosi ty: : hush_all 0 for silent operations
when compiling yoursre. ee that uses hfloats link it against the
library libhfl t. a
for extreme precisions increase the maximal workspace size, as described on page 250.
Cf. examples/ ex* . ee for some simple examples of how to use
hfloats. Look into examples/index. txt for what's there.

A.5 Computations with extreme precision

In order to do computations with the maximal possible precision on
your computer you have to manually set the maximal workspace size:

A.6 Precision and radix

251

Set the environment variable NOSWAP .-BYTES according to the size of

physical RAM where no swapping will occur, e.g. with bash say:
export NOSWAP.-BYTES=32M
See src/fxt/auxil/workspace. cc for the default.
Do not give the total amount installed RAM, your machine will
swap to death if you try really use it for hfloat. Currently if you don't
give a power of 2 then the size is increased to the next bigger power
of 2. You may append 'k' for kilobyte or 'M' for megabyte.

A.6 Precision and radix

For the quick readers I begin with the resumee:
for decimal digits and precisions up to 223 LIMBs 5 use radix 10,000
(for even greater precisions choose radix 1,000)
for hexadecimal digits and precisions up to 218 LIMBs6 use radix
65,536 (for even greater precisions choose radix 4,096)
The mantissa of a double consists of (typically) 53 bits 7 , its radix
(base of representation) is 2. The analogous quantity in the mantissa
of a hfloat is a LIMB (typedef'd to an unsigned 16 bit quantity). So the
radix of a hfloat can be in the range 2 ... 65536( = 216 ).
If one wants to get decimal numbers one would not use radix 10
but the greatest power of 10 that is :::; 216 , which is 10000. Due to
the implementation of the convolution there is a second restriction
to the radix: The cumulative sums Ck have to be represented exactly
enough to distinguish every (integer) quantity from the next bigger
(or smaller) value. The highest possible value for a Ck, Cm, must not
jump to Cm 1 due to numerical errors. For radix R and a precision
of N LIMBs
Cm

= N (R -

1)2

This is N times the product of two 'nines' (R - 1), which can appear
in the middle term of the convolution. Cm needs
log2(N (R - 1)2) = 10g2 N
5
6

+ 2 log2(R -

1)

corresponding to more than 32 million decimal digits

corresponding to more than 1 million hexadecimal digits
Of which only the 52 least significant bits are physically present, the most signicant bit is implied to be always set.

252

A. Documentation for the hfloat Library

bits to be represented exactly.

Due to numerical noise there must be a few more bits for safety.
The Ck are computed using doubles. We need to have
M 2: log2 N

+ 2 log2 (R -

1)

+S

where S :=safetybits and M :=mantissabits. With log2(R - 1) <

log2(R) we have equivalently
Nmax(R) = 2M-S-2

[og2(R)

Thus if we want base 2 numbers we would first choose radix R =

216 but this can (if we have M = 53 mantissabits and require S =
3 safetybits) only be used for precisions up to
Nmax(radix = 65,536) = 253-3-2.16 = 218 = 256kiloLIMBs
(corresponding to 4096kilo bits = 1024kilo hex digits).
Nmax(radix = 32,768) = 220 = IMegaLIMBs
(corresponding to 15360kilo bits = 3840kilo hex digits).
Nmax(radix = 16,384) = 222 = 4MegaLIMBs
(corresponding to 57344kilo bits = 14336kilo hex digits).
For decimal numbers
Nmax(radix = 10,000) = 253-3-2.13.29 = 223 .42> 8MegaLIMBs
(N = 23 corresponding to 32M ega decimal digits).

Nmax(radix = 1,000) = 230.07 > IGigaLIMBs

(N = 30 corresponding to 3Giga decimal digits).
If the LIMBs weren't restricted to 16 bits:

Nmax(radix = 100,000) = 216 .78 > 16kiloLIMBs

(N = 16 corresponding to 80kilo decimal digits).

Nmax(radix = 1,000,000) = 210.13 > lkiloLIMBs

(N = 10 corresponding to 6kilo decimal digits).

A.7 Compiling & running the 7r-example-code

253

A.7 Compiling & running the 7r-example-code

To compile the code in the directory ealepi/
Type 'make pi'.
cd to the bin directory, find the example binary pi there.
Type 'pi --help' for usage information (the help text can also be
found in the file doe/pihelp. txt).
Type 'pi 10' to run the program. It will compute 4096 decimal digits
of 7r (in about 1 second on a not very ancient computer).
See the file result. txt for the digits of 7r. All but the last few
digits (10 or so) of the output should be correct. Check correctness
like this:
Type 'pi 10 314' to run the program selftest which uses several
algorithms and outputs how many LIMBs differ in the results.
Type 'pi 8 0 65536' to get 1024 hexadecimal digits of 7r.
Type 'pi 18 16' to get> 1 million decimal digits of 7r.

A.8 Structure of hfloat

The hfloat code is divided into two layers:
The class hftoat (top, user interface) layer. A hftoat consists of a
unique id number (for internal use only), sign and exponent plus and
a pointer to class hfdata. Currently there is a one-to-one correspondence between hftoats and hfdata (it possibly will remain like this).
The arithmetical operators and functions like mul (srel, sre2, dest)
are here. The iterations (for root extraction, inversion, ... ) are also in
this layer.
Declarations are in sre/inelude/hfloat.h (class), sre/inelude/
hfloatop.h (operators) and sre/inelude/hfloatfu.h (functions).
Source files are in sre/hf/.
The class hfdata (mantissa operations, number crunching) layer. A
hfdata consists of a pointer to the digit field, the size and precision of
the mantissa data. Typical function names are dLsomething ( ... ) ,
Declarations are in sre/inelude/hfdata.h (class) and sre/inelude/
hfdatafu.h (functions). Source files are in sre/dt/.

254

A. Documentation for the hfloat Library

A.9 Organisation of the files

directory
bini
doc/
examples/
calcpi/
src/
src/fxt/
src/fxt/mult/
src/include/
src/tz/
src/hf/
src/dt/
src/pi/
testing/
*/bucket/

which files are there

binaries and libs
the main documentation
simple example code, look there!
7f example
(source directory)
source for the fxt-librarry
source for the fft-multiplication
header files for the hfioat-librarry
source for the transcendental functions
source for the hfioat-Iayer
source for the hfdata-layer
source for the 7f-library
cryptic test routines, better ignore
garbage & trash

Selected files in some directories:

src/ incl ude/: (declarations)
- hfloat . h: the hfioat class
- hfloatfu. h: hfioat functions
- hfloatop. h: hfioat operators
- hfverbosi ty . h: adjust how talkative the hfioat operations are
- hfdata. h: the hfdata class
- hfdatafu. h: hfdata functions
src/pi/: implementations of the 7f-algorithms
src/hf/:
- hfloat. cc: hfioat class member functions
- hfloatop. cc: hfioat class operators
- ini t. cc: magic initialiser for the hfioat class
- it*. cc: code for iterations (inverse, sqrt, inverse n-th root)
src/dt/:
- hfdata. cc: hfdata class member functions
- dU. cc: hfdata functions

A.I0 Distribution policy & no warranty

255

A.I0 Distribution policy & no warranty

legal stuff --~
The hfloat package and code is freeware, you may use it at no cost
for noncommercial purposes.
The copyright remains by the author (Jorg Arndt, arndt@jjj .de).
Obvious exceptions are the included pieces of code by other authors.
You may (and are encouraged to) give hfloat away free of charge.
Always give a pointer to the original hfloat if you redistribute pieces
of it.
You are not allowed to make money with hfloat (or parts of it) III
any way.
Before putting hfloat on a CD or similar get my agreement.
WARNING: hfloat is distributed 'as is', you are using hfloat at your
own risk. I will take no responsibility for potential damage that might
be caused.
--~

- - - end of legal stuff - - Please be nice and give me credits if you use hfloat for anything interesting/noticable/scientific. I am interested in hearing about your
project.
If you use hfloat for scientific purposes then cite it in your publications as you would cite an ordinary publication.
Please check for newer versions before passing hfloat on. I tend to
upload new versions even for small changes.
If you notice errors in the code or the documentation please take the
time and drop me an email.

I
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j
j

Bibliography

1. Victor Adamchik, Stan Wagon, A Simple Formula for 'IT, American Mathematical Monthly, Vol 104 (1997) 852-855, also in [20, pp. 557-559J.
(Cited on p. 126)
2. Victor Adamchik, Stan Wagon, Pi: A 2000- Year Search Changes Direction,
Education and Research, Vol 5 (1996), No.1, 11-19,
online at http;llmembers. wri. comlvictor/articles/pi .html
(Cited on p. 19, 126, 227)
3. Association pour Ie Developpement de la Culture Scientifique (ADCS), 61 rue
Saint-Fuscien, 8000 Amiens France, Le nombre 'IT, Sonderheft zur Zeitschrift
Le PETIT ARCHIMEDE.
4. Timm Ahrendt, Schnelle Berechnung der komplexen Quadratwurzel auf hohe
Genauigkeit, Logos Verlag, Berlin, 1996.
Online at http://web.informatik.uni-bonn.de/II/staff/ahrendt/
AHRENDTliteratur.html
(Cited on p. 148)
5. Gert Almquist, Many Correct Digits of 'IT, Revisited, American Mathematical
Monthly, Vol. 104 (1997), No.4, 351-353.
(Cited on p. 157)
6. Archimedes, Measurement of a Circle, in The Works of Archimedes, Ed. by
T.L. Heath, Cambridge University Press, 1897, also in [20, pp. 7-14J.
(Cited on p. 171)
7. Jorg Arndt, Remarks on arithmetical algorithms and the compuation of 'IT,
Online at http://www . j j j . del j oerg. html.
(Cited on p. 75, 110, 233, 234)
8. Nigel Backhouse, Pancake functions and approximations to 'IT. The Mathematical Gazette, Vol. 79 (1995), 371-374.
(Cited on p. 234)
9. David H. Bailey, The Computaion of 'IT to 29,360,999 Decimal Digits Using
Borweins' Quarlically Convergent Algorithm, Mathematics of Computation,
Vol. 50, No. 181 (Jan. 1988), 283-296, also in [20, pp. 562-575J.
(Cited on p. 202)
10. David Bailey, Peter Borwein and Simon Plouffe, On The Rapid
Online
at
Computation of Various Polylogarithmic Constants,
http://www . cecm. sfu. carpborwein, also in [20, pp. 663-676J.
(Cited on p. 118, 123)
11. David H. Bailey, Jonathan M. Borwein and Peter B. Borwein, Ramanujan,
Modular Equations, and Approximations to Pi or How to compute One Billion
Digits oj Pi, American Mathematical Monthly, Vol. 96 (1989), 201-219, also
in [20, pp. 623-641J.
(Cited on p. 6, 105)

258

Bibliography

12. David H. Bailey, Jonathan M. Borwein and Richard E. Crandall, On

the Khintchine Constant, Mathematics of Computation, Vol. 66, No. 217
(Jan. 1997), 417-43l.
(Cited on p. 68)
13. David H. Bailey, Jonathan M. Borwein, Peter B. Borwein and Simon Plouffe,
The Quest for Pi, The Mathematical Intelligencer, Vol. 19 (1997), No.1, 5056.
(Cited on p. 17, 119, 122, 197, 200, 237, 238)
14. Walter William Rouse Ball, Mathematical Recreations and Essays, 11th edition with revisions, Macmillan & Co Ltd, London, 1963.
(Cited on p. 41, 180, 181, 182, 183)
15. Walter William Rouse Ball, Harold Scott Macdonald Coxeter, Mathematical
Recreations and Essays, Toronto, 1974.
(Cited on p. 25)
16. J. P. Ballantine, The best (?) formula for computing IT to a thousand places,
American Mathematical Monthly, Vol. 46 (1939), 499-50l.
(Cited on p. 74)
17. Friedrich L. Bauer, Decrypted Secrets, Methods and Maxims of Cryptology,
Second, Revised and Extended Edition, Springer-Verlag, Heidelberg, 2000.
(Cited on p. 61, 228)
18. Petr Beckmann, A History of IT (PI), Fifth Edition, The Colem Press, Boulder, Colorado, 1982.
(Cited on p. 64, 188, 194, 195)
19. Fabrice Bellard, Computation of the n'th digit of pi on any base in O(n 2 ),
Online at http: / hlww-stud. enst . fr rbellard/pi/.
(Cited on p. 128, 227, 229, 234)
20. Lennart Berggren, Jonathan Borwein, Peter Borwein, Pi: A Source Book,
Springer-Verlag, New York, 1997.
(Cited on p. 115, 187, 188, 226)
21. Bruce C. Berndt, Ramanujan-l00 Years Old (Fashioned) or 100 Years New
(Fangled)?, The Mathematical Intelligencer, Vol. 10 (1988), No.3, 24-29.
(Cited on p. 105)
22. Bruce C. Berndt, A Pilgrimage, The Mathematical Intelligencer, Vol. 8
(1986), NO.1, 25-30.
(Cited on p. 105, 106)
23. Bruce C. Berndt and Robert A. Rankin, Ramanujan, Letters and Commentary, American Mathematical Society, London Mathematical Society, 1995.
(Cited on p. 105)
24. Bruce C. Berndt, Ramanujan's Notebooks, Part 1(1985), II(1989), III(1991),
1V(1994), Veto appear), Springer-Verlag, New York.
(Cited on p. 109)
25. Bruce C. Berndt and S. Bhargava, Ramanujan - For Lowbrows, American
Mathematical Monthly, Vol. 100 (1993), 644-656.
(Cited on p. 58, 108)
26. Eugen Beutel, Die Quadratur des Kreises, Mathematische Bibliothek, No. 12,
Verlag von B.C. Teubner, Leipzig und Berlin, 1913.
(Cited on p. 181)
27. Ludwig Bieberbach, Personlichkeitsstruktur und mathematisches Schaffen,
Forschungen und Fortschritte, 10. Jahrg., Nr. 18 (20. Juni 1934), 235-237.
28. David Blatner, The Joy of IT, Walker and Company, New York, 1997.
(Cited on p. 201)

Bibliography

259

29. Jonathan M. Borwein, Brouwer-Heyting Sequences Converge, Mathematical

Intelligencer, Vol. 20, No.1, 1998, 14-15.
(Cited on p. 30)
30. Jonathan M. Borwein and Peter B. Borwein, The Arithmetic-Geometric Mean
and Fast Computations of Elementary Functions, SIAM Review, Vol. 26, No.3
(July 1984), 351-366, aiso in [20, pp. 537-552J.
(Cited on p. 114, 236)
31. Jonathan M. Borwein, Peter B. Borwein, More Ramanujan-type Series for
1/71', Proceedings of the Centenary Conference, Univ. of Illinois at UrbanaChampaign, June 1-5, 1987, 359-374.
32. Jonathan M. Borwein, Peter B. Borwein, Pi and the AGM - A Study in
Analytic Number Theory and Computational Complexity, Wiley, N.Y., 1987.
(Cited on p. 7, 59, 73, 74, 90, 92, 114, 188, 228, 229, 237)
33. Jonathan M. Borwein, Peter B. Borwein, Ramanujan and Pi, Science and Applications; Supercomputing 88: Volume II, 117-128, 1988. also in [20, p. 588595J
(Cited on p. 105, 107, 111, 237)
34. Jonathan M. Borwein, Peter B. Borwein, Strange Series and High Precision
Fraud, American Mathematical Monthly, Vol. 99 (1992), 622-640.
(Cited on p. 63, 230)
35. Jonathan M. Borwein, Peter B. Borwein and K. Dilcher, Pi, Euler Numbers,
and Asymptotic Expansions, American Mathematical Monthly, Vol. 96 (1989),
681-687, also in [20, pp. 642-648J.
(Cited on p. 156, 157)
36. Jonathan M. Borwein and Roland Girgensohn, Addition theorems and binary
expansions, Canadian Journal of Mathematics, Vol. 47 (1995), 262-273.
(Cited on p. 48, 235)
37. Richard P. Brent, Fast Multiple-Precision Evaluation of Elementary Functions, Journal of the ACM, Vol. 23, No.2, April 1976, 242-251, also in [20,
pp.424-433J.
(Cited on p. 87)
38. E. Oran Brigham, The Fast Fourier Transform, Prentice Hall, 1994.
(Cited on p. 199)
39. David M. Burton, Burton's History of Mathematics: An Introduction,
Wm. C. Brown Publishers, Third Edition, 1995.
(Cited on p. 14, 190)
40. George S. Carr, Formulas and Theorems in Pure Mathematics, Second Edition, Chelsea Publishing Company, 1970.
(Cited on p. 106)
41. Dario Castellanos, The Ubiquitous 7r, Mathematical Magazine 61 (1988), 6798 (Part I) and 148-163 (Part II).
(Cited on p. 45, 59, 60, 72, 154, 166, 180, 190, 198, 224, 225, 226, 227, 228,
229,232)
42. Henri Cohen, F. Rodriguez Villegas and Don Zagier Convergence acceleration of alternating series, Experimental Math., to appear. downloadable from
http://www.math.u-bordeaux.fr/cohen/
(Cited on p. 150)
43. J.W. Cooley, and J.W.Thkey, An algorithm for the machine calculation of
complex Fourier series, Mathematics of Computation Vol. 19 (1965), No. 90,
297-301.
(Cited on p. 137)

260

Bibliography

44. James W. Cooley, Peter A.W. Lewis and Peter D. Welch, Historical Notes on
the Fast Fourier Transform, Proceedings of the IEEE, VoL 55 (October 1967),
No. 10, 1675-1677.
(Cited on p. 199)
45. David A. Cox, The arithmetic-geometric mean of Gauss, L'Enseignment
MatMmatique, t. 30 (1984), 275-330, also in [20, pp. 481-536].
(Cited on p. 95, 96, 97, 98)
46. David A. Cox, Gauss and the Arithmetic-Geometric Mean, Notices of the
American Mathematical Society 32 (1985), 147-151.
(Cited on p. 95)
47. Richard E. Crandall, Projects in Scientific Computation, Springer-Verlag,
New York, Inc., 1994.
(Cited on p. 230)
48. Jean-Paul Delahaye, Pi - die Story, Birkhiiuser-Verlag, Basel, 1999. Franzosische Originalausgabe: Le fascinant nombre 7r, Pour La Science, 1997.
(Cited on p. 50)
49. Drinfel'd, Quadratur des Kreises und Transzendenz von 7r, VEB Deutscher
Verlag der Wissenschaften, Berlin, 1980.
(Cited on p. 197)
50. Underwood Dudley, Mathematical Cranks, The Mathematical Association of
America, 1992.
(Cited on p. 8)
51. Heinz-Dieter Ebbinghaus, Reinhold Remmert, et aL, Numbers, SpringerVerlag, New York, 1990.
(Cited on p. 56, 168, 170, 176, 213)
52. LeQnhard Euler, Introduction to Analysis of the Infinite, Translation of: Introductio in analysin infinitorum, Book I, Translated by John D. Blanton,
Springer-Verlag, New York, 1988, Chapter 10 also in [20, pp. 112-128].
(Cited on p. 65, 166, 194)
53. Leonhard Euler, Einleitung in die AnalysiS des Unendlichen, Erster TeiL Ins
Deutsche iibertragen von H. Maser, Berlin. Verlag von Julius Springer. 1885.
(Cited on p. 166, 194)
54. Martin Gardner, Some comments by Dr. Matrix on symmetrics and reversals,
Scientific American, January 1965, 110-116.
(Cited on p. 24)
55. Carl Friedrich GauB, Mathematisches Tagebuch, 1796-1814, Reihe Ostwalds
Klassiker der exakten Wissenschaften, Akademische Verlagsgesellschaft Geest
& Portig K.-G., Leipzig 1985. Also in [56, X, pp. 483-574].
(Cited on p. 98)
56. Carl Friedrich GauB, Werke, Gottingen, 1866-1933.
(Cited on p. 73, 88, 95, 99, 232)
57. C.F. GauB, Nachlass zur Theorie des Arithmetisch-Geometrischen Mittels und
der Modulfunktion, herausgegeben von Harald Geppert, Reihe Ostwald's Klassiker der exakten Wissenschaften, Nr. 225, Akademische Verlagsgesellschaft,
Leipzig, 1927.
(Cited on p. 95)
58. Roland Girgensohn, personal E-Mail, 1997.
(Cited on p. 119)
59. Franz Gnadinger,
Primary Hill And Rising Sun, Online at
http://www.access.ch/circle/.
(Cited on p. 210)

Bibliography

261

60. Wei Gong-yi, Yang Zi-qiang, Sun Jia-chang, Li Lia-kai, The Computation of
7T to 10,000,000 Decimal Digits J. Num. Method & Compo Appl., 17:1(1996),
78-81.
(Cited on p. 158)
61. GroB, Zagier, On singular moduli, J. Reine Angew. Math., 355(1985), 191220.
(Cited on p. 27)
62. Bruno Haible und Thomas Papanikolaou, Fast multipliprecision evaluation
of series of rational numbers, Online at
http://www.informatik.th-darmstadt.de/TI/Veroeffentlichung/TR/
Welcome. html.
(Cited on p. 215, 217)
63. Godfrey H. Hardy, Ramanujan, Twelve Lectures on Subjects Suggested by his
Life and Work, Chelsea Publishing Company, New York, 1940.
(Cited on p. 106, 108, 109)
64. Sir Thomas Heath, A History of Greek Mathematics, Volume II From
Aristarchus To Diophantus, Dover Publications, 1981, first published 1921.
(Cited on p. 173, 174)
65. M. Heidemann, D. Johnson, C. Burrus, Gauss and the history of the Fast
Fourier transformation, IEEE ASSP Magazine 1, 1984, 14-21.
(Cited on p. 137, 199)
66. Joseph E. Hofmann, Geschichte der Mathematik, Sammlung Goschen Band
226 (Erster Teil) , Band 875 (Zweiter Teil) , Band 882 (Dritter Teil), Walter
de Gruyter & Co, Berlin 1953.
67. Dirk Huylebrouck, The 7T-Room in Paris, The Mathematical Intelligencer,
Vol. 18, No.2, 1996, 51-53.
(Cited on p. 50)
68. William Jones, Synopsis Palmariorum Mathesos, London, 1706, In extracts
in [20, pp. 108-109].
69. A.P. Juschkewitsch, Geschichte der Mathematik im Mittelalter, Teubner Verlag, 1964.
(Cited on p. 182)
70. Sven Kabus, Untersuchung verschiedener 7T-Algorithmen am PC, (PC investigation of various 7T algorithms), entry in the Jugend forscht (Youth researches)
competition, 1998, Junge wissenschaft (Friedrich-Verlag), 58 (1999),34-42.
(Cited on p. 63)
71. Robert Kanigel, The Man Who Know Infinity, Macmillan Publishing Company, 1991.
(Cited on p. 105)
72. Victor J. Katz, A History of Mathematics, HarperCollins College Publishers,
1993.
(Cited on p. 167, 168, 179)
73. Theo Kempermann, Zahlentheoretische Kostproben. Verlag Harri Deutsch,
Thun, Frankfurt am Main, 1995.
(Cited on p. 14, 158, 234)
74. Alexander Khintchine, Continued Fractions, Nordhoff, Groningen 1963
[Moskow 1935].
(Cited on p. 68)
75. Louis V. King, On the Direct Numerical Calculation of Elliptic Functions and
Integrals, Cambridge University Press, Cambridge, 1924.
(Cited on p. 57)

262

Bibliography

76. Konrad Knopp, Infinite sequences and series Dover Publications, New York,
1956
(Cited on p. 71, 78)
77. Donald E. Knuth, The Art of Computer Programming, Volume 2 Seminumerical Algorithms, Third Edition, 1998, 4th Printing August 1999.
(Cited on p. 28, 102, 120, 121, 132, 135, 180, 182)
78. Lam Lay-Yang and Ang Tian-Se, Circle Measurement in Acient China, Historia Mathematica, Vol. 13 (1986), 325~340, also in [20, pp. 20~44].
(Cited on p. 176)
79. Johann Heinrich Lambert, Memoire sur quelques proprietes remarquables
des quantites transcendantes circulaires et logarithmiques, Memoires de
l'Academemie de sciences de Berlin, [17] (1761), 1768, 265~322, also in [20,
pp. 129~140].
(Cited on p. 192)
80. L.J. Lange An elegant new continued fraction for n, American Mathematical
Monthly, Vol 106 (May 1999) 456~458.
(Cited on p. 33, 231)
81. Nick Lord, Recent calculations of n: the Gauss-Salamin algorithm, The Mathematical Gazette, Vol. 76 (1992), 231~242.
(Cited on p. 100)
82. Philipp Maennchen, GaufJ als Zahlenrechner, in [56, X.2,1-73].
(Cited on p. 57)
83. Eli Maar, e: the story of a number, Princeton University Press, 1994.
(Cited on p. 96)
84. Yuri V. Matiyasevich, Richard K. Guy, A New Formula for n, American
Mathematical Monthly, Vol. 93, (October 1986), 631~635.
(Cited on p. 234)
85. Niels Nielsen, Die Gammafunktion, Chelsea Publishing Company, Bronx, New
York, 1965 (first published 1906).
(Cited on p. 229)
86. Ivan Niven, Irrational Numbers, John Wiley and Sons, 1956.
(Cited on p. 22)
87. C.D. Olds, Continued Fractions, Mathematical Association Of America, 1963.
(Cited on p. 65)
88. Thomas J. Osler, The Union of Vieta's and Wallis's Products for Pi, American Mathematical Monthly, Vol. 106, (October 1999), 774~776.
89. S. Parameswaran, Whish's showroom revisited. The Mathematical Gazette,
76(1992), 28~36.
(Cited on p. 186, 223, 224)
90. Oskar Perron, Die Lehre von den Kettenbruchen, Band I und II, B.G. Teubner
Verlagsgesellschaft, Stuttgart, 3. Auflage, 1957.
(Cited on p. 65, 67)
91. G.M. Phillips, Archimedes and the numerical analyst, American Mathematical
Monthly, Vol. 88(1981), 165~169, also in [20, pp. 15~19].
92. Richard Preston, The Mountains of pi, The New Yorker, Mar. 2, 1992, 36~67.
(Cited on p. 2, 10, 34, 201)
93. Martin R. Powell, Significant insignificant digits, The Mathematical Gazette,
66(1982), 220~221.
(Cited on p. 156)
94. Stanley Rabinowitz, Abstract 863-11-482: A Spigot Algorithm for Pi, Abstracts Amer. Math. Society, 12(1991), 30.
(Cited on p. 37)

Bibliography

263

95. Stanley Rabinowitz and Stanley Wagon, A Spigot Algorithm for the Digits of
Pi, American Mathematical Monthly, Vol. 103 (March 1995), 195-203.
(Cited on p. 37, 77, 82)
96. Srinivasa Ramanujan, Modular equations and approximations to 71",
Quart. J. Math. 45 (1914) 350-372, also in [20, pp. 241-257].
(Cited on p. 27, 57, 58, 103, 105, 226)
97. Srinivasa Ramanujan, Squaring the Circle, J. Indian Math. Soc. 5(1913), 132,
also in [20, p. 240].
(Cited on p. 58)
98. Ranjan Roy, The Discovery of the Series Formula for 71" by Leibnitz, Gregory and Nilakantha, Mathematics Magazine 63 (1990), 291-306, also in [20,
pp.92-107].
(Cited on p. 186, 230)
99. Hansklaus RummIer, Squaring the Circle with Holes, American Mathematical
Monthly, Vol. 100 (November 1993), 858-860.
(Cited on p. 162)
100. Eugene Salamin, Computation of 71" Using Arithmetic-Geometric Mean, Mathematics of Computation, Vol. 30, No. 135, July 1976, 565-570, also in [20,
pp. 418-423].
(Cited on p. 87, 102, 231)
101. Eugene Salamin, E-Mail to Jorg Arndt, 1997.
(Cited on p. 102)
102. Norbert Schappacher, Edmund Landau's Goitingen: Prom the Life and Death
of a Great Mathematical Center, The Mathematical Intelligencer, Vol. 13,
No.4, 12-18.
103. L. Schlesinger, Uber Gauss Funktionentheoretische Arbeiten, in [56, X.2, 990].
104. Arnold SchOnhage und Volker Strassen, Schnelle Multiplikation grofJer Zahlen,
Computing 7 (1971), 281-292.
(Cited on p. 141, 147, 217)
105. Arnold SchOnhage, A. F. W. Grotefeld/E. Vetter, Fast Algorithms, A Multitape Turing Machine Implementation, BI Wissenschaftsverlag, Mannheim,
1994.
(Cited on p. 92, 236)
106. D. Shanks and J. W. Wrench, Calculation of Pi to 100,000 Decimals, Mathematics of Computation, vol. 16 (1962), 76-79, also in [20, p. 326-329].
(Cited on p. 49, 197)
107. Daniel Shanks, Dihedral Quartic Approximations and Series for 71", Journal of
Number Theory 14(1982), 397-423.
(Cited on p. 62)
108. Daniel Shanks, Improving an Approximation for Pi, American Mathematical
Monthly, Vol. 99 (1992) No.3, 263.
(Cited on p. 49)
109. Allen Shields, Klein and Bieberbach, Mathematics, Race, and Biology, The
Mathematical Intelligencer, Vol. 10 (1988), No.3, 7-11.
(Cited on p. 214)
110. David Singmaster, The legal values of 71", The Mathematical Intelligencer,
Vol. 7 (1985), No.2, 69-72, also in [20, pp. 236-239].
(Cited on p. 213)
111. M.D. Stern, A Remarkable Approximation to 71", The Mathematical Gazette,
Vol. 69 (1985), 218-219, also in [20, pp. 460-461].
(Cited on p. 169)

264

Bibliography

112. Lutz von Strassnitzky, Der Kreis-Umfang fur den Durchmesser 1 auf 200 Decimalstellen berechnet von Herrn Z. Dahse in Wien. Journal fUr die reine und
angewandte Mathematik, Ed 27, 1844, 198.
(Cited on p. 194)
113. D. Takahasi, Y. Kanada, Calculation of Pi to 51.5 Billion Decimal Digits
on Distributed Memory and Parallel Processors, Transactions of Information
Processing Society of Japan, Vol. 39, No.7, (1998).
(Cited on p. 201, 206)
114. Johannes Tropfke, Geschichte der Elementarmathematik, Vierter Band,
Ebene Geometrie, Dritte Aufiage, Walter De Gruyter & Co, 1940.
(Cited on p. 73, 165, 166, 169, 171, 175, 176, 182, 183, 187, 189, 224, 228,
231, 232, 235)
115. Franciscus Vieta (Francpis Viete), Variorum de Rebus Mathematicis, 1593,
Caput XVIII, 400, auch (tw.) in [20, p. 53-56].
(Cited on p. 187)
116. Alexei Volkov, Zhao Youquin and His Calculation of7r, Historia Mathematica,
Vol. 24, No.3, August 1997, 301-331.
(Cited on p. 179)
117. Stan Wagon, Is 7r Normal? The Mathematical Intelligencer, Vol. 7 (1985),
No.3, 65-67, also in [20, pp. 557-559].
(Cited on p. 23)
118. Sebastian Wedeniwski, Piologie, Eine exakte arithmetische Bibliothek in
C++, Edition 1.3, http://www.hipilib.de/
(Cited on p. 234)
119. David Wells, The Penguin Dictionary of Curious and Interesting Numbers,
Penguin Books, Middlesex, 1986.
(Cited on p. 195)
120. David Wells, Are These the Most Beautiful? The Mathematical Intelligencer,
Vol. 12, No.3, 37-41.
(Cited on p. 190)
121. Jet Wimp, Book Review of Pi and the AGM in SIAM Review, Oct. 1988,
530-533.
(Cited on p. 116)
122. J. W. Wrench, Jr., The evolution of extended decimal approximations to 7r,
The Mathematics Teacher, vol. 53 (Dec. 1960), 544-650. also in [20, p. 319325].
(Cited on p. 73, 196)
123. Robert M. Young, Probability, pi, and the primes, Serendipity and experimentation in elementary calculus, The Mathematical Gazette, Vol. 82 (Nov.
1998), 443-451.
(Cited on p. 9)

Index

163 phenomenon, 25
Abel, 95, 98, 109
ADA,118
Ada of Lovelace, 118
Adamchik, 19, 126, 203, 227
AGM,87-102
- basic features, 90
- rule, 88
Aitken, 25
Al-Khashi, 182, 205
algebra, 180
algorithm, 180
- Archimedes, 235
- BBP, 119
- binsplit, 215, 216, 218
- Borwein
cubic, 237
nonic,238
quadratic, 114, 236
quartic, 114, 237
quintic, 111, 237
- dartboard, 39
- Descartes, 235
- Gauss AGM, 77, 90, 91, 93, 114, 119,
200, 236
- Kabus,64
- spigot, 77
- sumalt, 150
Alkarism, 180, 205
Andersson, 153
Andrews, 109
Antiphon, 170
approximation
- Archimedes, 52, 171
- Borweins, 59, 62
- Castellanos, 59
- Plouffe, 60
- Ramanujan, 57, 58
- Shanks, 62
- Tsu Chhung-Chih, 52, 179
approximations for 7r, 51-64

Archimedes, 15,69,170-177,205,235
arctan relations, 232-234
Arcus Tangens, 69-76
Aristophanes, 7
arithmetic, 131-152
arithmetic-geometric mean see AGM
87
Arndt, 75, 76, 110, 233, 234
Aryabhata, 56, 179, 205
Au-Yeung, 235
Babylon, 69, 167, 205
Backhouse, 234
Bailey, 19, 117, 123,202,206, 207, 227,
235
Ball,41
Barrow, 166
Bauer, 55, 61, 228
Bayes, 21
BBP algorithm, 117-129
BBP series, 118, 119
BBP-like series, 227-228
Beckmann, 188, 191, 194, 195
Bellard, 127, 128, 207, 227, 229, 234
Berggren, 115
Berndt, 105, 109
Bernoulli, Jakob, 96, 191
Bernoulli, Johann, 96, 166, 191
"best (arctan) formula", 74
Bible, 4, 51, 169, 205
Bieberbach, 213
binary modulo exponentiation, 120
binsplit algorithm, 215, 216, 218
bit recursion, 48, 235
Blatner, 201
Borwein, J., 30, 48
Borwein, J. and P., 11, 17,59,63,111,
113, 114, 157, 200, 229, 230, 236-238
- algorithm
- - cubic, 237
nonie, 238
-- quadratic, 236

266

Index

- - quartic, 114, 237

-- quintic, 111, 237
Borwein, P., 19, 117, 198,207,227
Borweins and 71", 113-116
Boyer, 206
Brahmagupta, 180, 205
Brent, 87, 200, 236
Brent-Salamin iteration, 87
Brouncker, 13, 188, 231
Brouwer, 30
Brown, 38
Buffon,39

computer algebra system, 118

computers used in world records, 206
Comtet, 230
continued fraction, 64-68
- ,,67
- 4>, 66
- 71", 32, 53, 67, 188, 231
- 7I"e, 67
- y'2,66
- )3,66
- ij2,67
- ft, 67
- e, 32, 67

C/C++

- e2 , 67

program
- 71" approximation using dartboard
algorithm, 40
- 71" approximation using test for
coprimality, 42
- BBP algorithm, 123, 124
- fast Fourier transform (FFT)
multiplication, 141
- fft () fast Fourier transform, 143
- obfuscated (Roemer), 36
- obfuscated (Westley), 35
- sft 0 slow Fourier transform, 140
- spigot algorithm for 71" (1 digit per
pass), 81
- spigot algorithm for 71" (4 digits per
pass), 83
- spigot algorithm for 71", short form, 37
- spigot algorithm for e, 85
Cadaeic cadenza, 45
Carr, 106
Carroll, 46, 72
Castellanos, 59, 60
Catalan, 229
Cataldi,64
CECM,113
Ceulen see Ludolph van Ceulen 182
CfToNumber 0, 66
Chandah-sutra, 121
Cheops pyramid, 211
Chephren pyramid, 211
2
X test, 23, 29
China, 176
Choresmia, 180
Chronicles, Book of, 169
Chudnovsky, D. and G., 1, 2, 33, 110,
111, 201, 206, 230
circle, squaring the, 7, 170, 196
Clausen, 205
Cohen, 150
compass and straight edge, 7, 98, 170

Cooley, 137, 199

Cooley-Tukey algorithm, 199
coprimality and 71", 41, 42
Cotes, 190
coupled Newton iteration, 147
Cox, 95
Crandall, 230
Cusanus, 183, 184
Danielson, 199
Dante Alighieri, 56, 180, 205
dartboard algorithm, 39, 40
Dase, 194, 195, 205
DATA statement, 155
Decerf,45
definitions of 71", 8
Descartes, 184, 235
Dichampt, 206
Dilcher, 157
division, 145
DNA,4
Dodgson, see Caroll 46
Dudley, 8
Duerer, 181

Ve, 67
e, 11, 14, 25, 32, 37, 60, 67, 84, 85, 116,
189, 190, 202
- series for spigot algorithm, 84
e mathematics, 116
e 2 ,67
educational poems, 44-47
Egypt, 69, 167, 205, 210
elastic curve, 96, 97
Eratosthenes, 38
error distribution curve, 10
Escott, 73, 232
Euclid, 170
euclid (), 42
Euclidean algorithm, 42

Index

Euler, 10, 11, 13, 64, 70, 74, 97, 100,

166, 186, 188-190, 194, 224-226,
230-232
- formulae, 224-226
Euler's famous 7T theorem, 14, 61, 190
Eulerian numbers, 157
exhaustion method, 169
expmO,124
exponentiation, binary, 120, 121
Fagnano, 97, 98, 234
Felton, 197, 206
Ferguson, 50, 205
Fermat, 6
- Last Theorem, 7
Feynman,3
Feynman point, 3, 46
FFT fast Fourier transform, 137
FFT multiplication, 16, 137
Fibonacci, 75, 180, 205
Fibonacci numbers, 75, 234
Filliatre, 206
Flammenkamp,37
formula collection 7T, 13-14, 223-238
FORTRAN, 37,155,202
Fourier, 137
Fox, 39
funnel, an (in)finite, 164

r(n), Euler's Gamma function, 226,

234
'Y, Euler's constant, 67
Gardner, 24
Garwin, 198
Gauss, 57, 73, 87, 137, 184, 193, 195,
197, 199, 200, 232, 236
- AGM algorithm, 90, 91, 102, 114
-- derivative (Borwein), 114
- - Schonhage variant, 93
- and 7T, 87-102
- error distribution curve, 10
Gauss-Legendre method, 87
Gauss~GMO, 93
Genuys, 197, 206
Girgensohn, 48
Gnaedinger, 210
Golden ratio, , 60, 66, 68, 76
Goodwin, 212
Gosper, 32, 104, 202, 206, 229
Goto,47
Greece, 169
Gregory, 69, 70, 184, 188
- Series, 70, 189

267

Grienberger, 183, 205

Guilloud, 206
Haible,215
Hardy, 107, 108, 214
Heath, 257
Hermann, 232
Hermite, 196
hfioat library, 94, 148, 247-254
high-performance algorithms, 16, 198
Hilbert, 213
Hiram of Tyre, 169
historical notes, 209-214
history of digit extraction records, 207
history of 7T, 165-207
history of 7T in the computer era, 206
history of 7T in the pre-computer era,
205
Hofstadter, 15
holy triangle, 211
Hutton, 73, 232
Huygens, 64, 192
hyperspheres, 158
Imhotep, 211
India, 168, 179, 185, 205
individual digits of 7T
- any base, 128
- hexadecimal base, 118
Internet 7T clubs, 11
Intuitionism, 30
IOCCC - International Obfuscated C
Code Contest, 36
irrationality of 7T, 5, 65, 192
j Function, 27

Jacobi, 95
Japan, 194
Java program
- spigot algorithm for
Jeans, 44
Jeenel,206
Johnson, 198
Jones, 165, 193
The Joy of 7T, 201
Jugend forscht, 63
Jyesttha-devan, 186

7T,

77

Kabus, 63, 64
Kanada, 1, 11, 12, 17, 20, 28, 114, 200,
206
Kanigel, 105
Karana-Paddhati, 185
Karatsuba, 132

268

Index

Karatsuba multiplication, 134

Keith, 45
Kempermann, 234
Khintchine, 68
Khintchine constant, 68
Kings, Book of, 169
Klein, 213
Klingenstierna, 73, 232
Knopp, 228, 233
Knuth, 28, 120, 121
Koenig, 199
Kumbakonam,106
Lagny, 45, 193, 205, 228
Lagrange, 95
Lam Lay-Yong, 176
Lambert, 5, 65, 192
Lanczos, 199
Landau, 10, 213
Lange, 33, 231
Legendre, 57, 192
Lehmann, 205
Leibniz, 70, 188, 191, 192, 228
- Series, 70, 157
lemniscate, 95-99
Leonardo da Vinci, 181
Leonardo of Pisa (Fibonacci), 75, 180
Levy, 68
library
- hfioat, 247-254
Lievaart, 37, 85
Lindemann, 6, 197
Liu Hui, 177, 183, 205
Liu Xin, 176
Loney, 73, 106, 232
Lord, 100
Ludolph van Ceulen, 182, 205
Ludolphian number, 153, 183
m-zero,201
Machin, 13, 72, 73, 166, 192, 205, 232
Machin formula, 72
Madhavan, 186
Maple (computer algebra system), 118
Maple program
- Leibniz Series, 157
Mathematica (computer algebra
system), 35, 111, 118, 127
Matiyasevich, 234
Matsunaga, 194, 205
Metius, 181
Miyoshi,206
mnemonic verses, 44-47

modular equations, 57, 105

modulo exponentiation, binary, 120
Monte Carlo methods, 39
Morgan, 196
multiplication, 16, 131-145, 198
- FFT, 137
- Karatsuba, 132
- school method, 131
- using Fourier transforms, 137
- using logarithms, 136
MuPAD (computer algebra system),
118
MuPAD program
- Sqrt Coupled Newton, 148
- sumalt, 151
Nemorarius, 184
new goals, 19
Newton, 104, 145, 188, 205, 228
Nicholson, 206
Nilakantha, 186, 205, 223, 224
- formulae, 223-224
Niven, 22
Noether, 213
normality of 1r, 4, 21-34
North, 156
nth root calculation, 149
NumberToCf (), 66
numerical integrals, 95
numerology, 25, 154
obfuscated C see IOCee 36
Olds,65
Osler, 161, 230
Oughtred, 166
Parameswaran, 186
Peirce, 14
Percival, 20, 128, 203, 207
performance index of arctan formulae,
73
Perron, 65
(Golden ratio), 60, 66
1r, 1-265
- Symbol, 165
1r AGM formula, 87
1r Clubs, 11
1r Continued Fraction, 53, 231
1r Mathematics, 116
1r Room in Paris, 50
1r and , 60
1r and e, 11, 14, 60
1r calculations on the Internet, 215-221
1r definitions, 8

Index

formula collection, 13-14, 223-238

law, 211
Jr quiz, 153
Jr series from India, 15th century, 186
Jr, Digits 0 to 5,000 (base 10), 240-241
Jr, Digits 0 to 5,000 (base 16), 242-243
Jr, predictions, 198
Jr = 2, 155
Pi and the AGM, ll5
Pi: A Source Book, ll6
PI see performance index of arctan
formulae 73
Pi Trivia Game, 153
PiHex project, 20
pispigot.htm,77
Plato, 56, 170, 205
Plouffe, 19, 47, 48, 60, ll7, 128, 207,
227, 235
Poe, 45
Poeppe, 210
polygons, 170, 175
positions, self-referential, 3
Preston, 34, 201
prime pairs, 7
Pringsheim, 65
probabilities, 21
probability subjectivism, 21
Pschill, 227
PSQL algorithm, ll8
Ptolemy, 176, 205

Jr
Jr

Rabinowitz, 37, 77, 82

radius, 166
radix conversion, 79
Ramanujan, 13, 27, 57, 58, 104, 105,
199, 200, 214, 226, 227
- biography, 105-109
- formulae, 226-227
- squaring the circle, 58
Ramanujan and Jr, 103-ll1
random number generator, 27, 40
random numbers, 44
randomness of Jr, 21-34
Regiomontanus, 181
Reitwiesner, 206
residual sum, 186
Rhind Papyrus, 167
Rogers, 65
Romanus, 182, 205
Rome, 176
rounding up, approximation through,
38
Rummler, 162

Runge, 199
Rutherford, 194, 205
Sadratnamala, 185
Sagan, 14
Salamin, 87, 102, 200, 231, 236
Salomon, 169
Sastras, 185
Schonhage, 92, 236
series calculation, 150
Shakespeare, 46
Shanks, D., 49, 62, 197, 206
Shanks, W., 50, 195, 205
Sharp, 189, 205, 228
shortcuts to Jr, 35-50
sieve procedure, 38
Simon Fraser University, ll3
Singmaster, 213
slide rule, 136
Smith,205
Snell, 50, 178, 183, 205
spigot algorithm
- faster variant, 82
- pseudocode, 80
spigot algorithms, 77-85
Spring, 107
square root calculation, 146
squaring the circle, 7, 170, 181, 196
- with holes, 162
state of pi art, 1-20
Stern, 169, 231
Stieltjes, 67
Stifel, 196
Stirling, 61, 97, 229
Strassnitzky, 194, 232
St!15rmer, 73, 197, 232, 233
Sulvasutras, 168
sumalt algorithm, 150
symbol Jr, 165
tables, 239-245
Takahashi, 1, 206
Takebe, 194, 230
Tamura, 206
Tanta Sangraham, 185
tertium non datur, 30
theta functions, 95
time needed by world records, 206
transcendence of Jr, 5, 197
Tropfke, 169, 171, 180
Tsu Chhung-Chih, 5, 178, 205
Tukey, 137, 198
Umasvati, 169

269

270

Index

universe, 4, 17, 153

Vega, 193, 205
Viete, 13, 160-162, 182, 187, 205
Villegas, 150
VVagon, 19, 77, 82, 126, 203, 227
VValdo,212
VValker, 107
VVallis, 9, 64, 161, 163, 187
- product, 9
VVan Mang, 176
VVangFan, 177,205
waves in the 7r sequence, 34
VVedeniwski, 12
VVells, 190
VVestley, 35
VVetherfield, 232
VVhish,185

VViles, 7
VVilliams, 26
VVinter, 37
world records, 12
- list of, 205-207
VVrench Jr., 197, 205, 206, 232
VVVVVV, 11
Yoshino, 206
Yukti-Bhasa, 185
Yukti-Dipika, 185
Zagier, 150
zeta function (8),234
Zhang Heng, 56,176,205
Zimmermann, 8
Zoser pyramid, 211
Zu Chongzhi see Tsu Chhung-Chih 178

Jorg Arndt
Christoph Haenel

UNLEASHED
Never in the 4000-year history of research into n
have results been so prolific as at present. In their
book Jorg Arndt and Christoph Haenel describe
the latest and most fascinating findings of
mathematicians and computer scientists in the
field of n.Attention is focussed on new methods
of computation whose speed outstrips that of
predecessor methods by orders of magnitude.
The book comes with a CD-ROM containing not
only the source code of all programs described,
but also related texts and even complete libraries.

415,
-31415
) += "*_3_ 14

141S

-3-1 1

) :t+-.314= 314

14159 && *
1, 3 1415 =
314+(- 31415
while -(-++ *
)* 3 141--=0
j

) ; { char *
write (<3,1),

3,1415926535897932384626433832795028841971693993751058209749
3.1415926: }
31415<314131415% 314-(
31415
] +
T 3]+1)- 314;
,_3141592654)}
ISBN 3-540-66572-2

http://www.springer.de