V I S UA L C O M P L E X A N A LY S I S

VISUAL
CO MPL EX
ANALYSIS
T R I S TA N N E E D H A M
UNIVERSIT Y OF SAN FRANCISCO

Foreword by
ROGER PENROSE
Great Clarendon Street, Oxford, OX2 6DP,
United Kingdom
Oxford University Press is a department of the University of Oxford.
It furthers the University’s objective of excellence in research, scholarship,
and education by publishing worldwide. Oxford is a registered trade mark of
Oxford University Press in the UK and in certain other countries
© Tristan Needham 2023
The moral rights of the author have been asserted
First edition published in 1997
25th Anniversary Edition 2023
All rights reserved. No part of this publication may be reproduced, stored in
a retrieval system, or transmitted, in any form or by any means, without the
prior permission in writing of Oxford University Press, or as expressly permitted
by law, by licence or under terms agreed with the appropriate reprographics
rights organization. Enquiries concerning reproduction outside the scope of the
above should be sent to the Rights Department, Oxford University Press, at the
address above
You must not circulate this work in any other form
and you must impose this same condition on any acquirer
Published in the United States of America by Oxford University Press
198 Madison Avenue, New York, NY 10016, United States of America
British Library Cataloguing in Publication Data
Data available
Library of Congress Control Number: 2022948019
ISBN 978–0–19–286891–6 (hbk.)
ISBN 978–0–19–286892–3 (pbk.)
DOI: 10.1093/oso/9780192868916.001.0001
Printed and bound by
CPI Group (UK) Ltd, Croydon, CR0 4YY
Links to third party websites are provided by Oxford in good faith and
for information only. Oxford disclaims any responsibility for the materials
contained in any third party website referenced in this work.
 For
Roger Penrose
and
George Burnett-Stuart
FOREWORD

Roger Penrose

Complex analysis is the theory of functions of complex numbers or, more specif-
ically, holomorphic functions of such numbers. This theory is both profoundly
beautiful and vastly influential, both in pure mathematics and in many areas of
application, particularly in physics, indeed being central to the underlying formal-
ism of quantum mechanics. However, the very concept of a complex number is an
essentially abstract one, depending upon the seemingly absurd notion of a square
root of −1, the square of any ordinary real number being, unlike −1, necessarily
non-negative.
Yet, it should be borne in mind that even the notion of a so-called “real” num-
ber is also an abstraction, and we must move far beyond the immediate notion of
“counting numbers” 0, 1, 2, 3, 4, . . ., and beyond even the fractions 21 , 53 , − 47 , etc., if
we are to express even the square root of 2. But here we are helped by a visual image,
and can perceive a straight line extended indefinitely in both directions to give us
a good intuitive impression of the full array of real numbers. The slightly mislead-
ing term “real” for this imagined array is excusable, as we can indeed imagine a
ruler, or a line of ink drawn on a piece of paper, as providing us with some sort of
conceptual image of this array. This greatly helps our understandings of what the
mathematician’s precise notion of a “real number” is intended to idealize. We are
not concerned with whatever might be the nature of the physics of the particles or
fields that might compose our ruler or ink-line; nor, indeed, do we require any con-
cept of the cosmology that may be relevant to the extension out to infinity of our
imagined ruler or ink-line. Our abstract mathematical notion of a “real number”
remains aloof from any such realities of the actual world. Yet, in a curious reversal
of roles, it is this very mathematical idealization that underlies most of our theories
of the actual world.
So, what can be the driving force behind a need to go beyond these seemingly
ubiquitous “real numbers”? What purpose might there be for the introduction of
a “square root of −1”? Such a number fits nowhere within the span of the real
numbers, and it would appear that we have no reason to demand that the equation
x2 + 1 = 0 have any kind of “solution”. The answer to this desire for such entities
lies in the magic that lies hidden within them, but it is not a magic that immediately
reveals itself. In fact, when the first hint of this magic was actually perceived, in
viii Foreword

the mid 16th century, by Girolamo Cardano, and then more completely by Rafael
Bombelli, these strange numbers were dismissed, even by them, as being as useless
as they were mysterious.
It is of some interest to note that it was not in the equation x2 + 1 = 0 that this
hint of magic was first perceived, but in cubic equations like x3 = 15x + 4, which
has the perfectly sensible solution x = 4. Yet, as Bombelli had noted, Cardano’s
general expression for the solution of such cubic equations necessarily involves a
detour into a mysterious world of numbers of a sort where the equation x2 + 1 = 0
is deemed to have the two solutions, now referred to as the imaginary units x = i
and x = −i, of an algebra—now called complex-number algebra—that had appeared
to be consistent, but not what had been regarded as “real”.
This dismissive attitude did not change much until the mid to late 18th cen-
tury, with Leonhard Euler’s remarkable formula eiθ = cos θ + i sin θ and, even
more importantly, the geometrical representation of the entire family of complex
numbers as points in a Euclidean plane, as initially proposed by Caspar Wessel,
where the algebraic operations on complex numbers are readily understood in
geometrical terms. This provided a kind of 2-dimensional “visual reality” to the
array of complex numbers that could be combined with topological notions, such
as employed initially by Carl Friedrich Gauss, and soon followed by others. The
early to mid 19th century saw many important advances, many of these being due
to Augustin-Louis Cauchy, especially with the beauty and the power of contour
integration, and, perhaps most profoundly, with ideas due to Bernhard Riemann.
The very notion of “complex smoothness” of complex functions was expressed
by use of the Cauchy–Riemann equations, and this provided the powerful con-
cept of a holomorphic function that implies that a power-series expansion always
locally exists, this leading to a vast and powerful theory with numerous magical
properties.
The two revolutions of early 20th century physics both owe a profound debt
to complex-number mathematics. This is most manifest with quantum mechanics,
since the basic formalism of that theory depends fundamentally on complex num-
bers and holomorphic functions. We see a remarkable interplay between quantum
spin and the geometry of complex numbers. The basic equations of Schrödinger
and Dirac are both complex equations. In relativity theory, the transformations
relating the visual field of two observers passing close by each other at differ-
ent relativistic speeds is most easily understood in terms of simple holomorphic
functions. Moreover, many solutions of Einstein’s equations for general relativity
benefit greatly from properties of holomorphic functions, as does the description
of gravitational waves.
In view of the undoubted importance of complex analysis in so much of math-
ematics and physics, it is clearly important that there are basic accounts of these
topics available to those unfamiliar (or only partly familiar) with the basic ideas
 Foreword ix

of complex analysis. In this foreword I have very much stressed how the visual
or geometric viewpoint has been of vital importance, not only to the historical
development of complex analysis, but also to the proper understanding of the
subject. Tristan Needham’s Visual Complex Analysis as originally published in 1997
was, to my knowledge, unique in the extent to which it was able to cover these
fundamental ideas with such thoroughness, visual elegance, and clarity.
With this 25th Anniversary Edition there have been some significant improve-
ments, most particularly in the incorporation of captions to the diagrams. This
makes it easier for the reader to dip into the arguments, as illustrated so elegantly in
such wonderfully expressive pictures, without necessarily having to look through
to find the relevant portion of the text. In any case, I am sure that readers, over a
broad range of relevant knowledge—from those with no prior experience of com-
plex analysis to those already experts—will gain greatly from the charm, distinct
originality, and visual clarity of the arguments presented here.
PREFACE TO THE 25TH ANNIVERSARY EDITION

Introduction
Mathematical reality exists outside the confines of space and time, but books about
mathematics do not. A quarter century after its publication, I am grateful that an
entirely new generation of mathematicians and scientists has continued to embrace
VCA’s unorthodox, intuitive, and, above all, geometrical approach to Complex
Analysis. To mark the occasion, Oxford University Press has graciously permit-
ted me to revisit the work, resulting in the creation of this significantly improved
25th Anniversary Edition.

Principal Changes in the New Edition

Perhaps I should begin by noting what has not changed. As I shall elaborate below,
back in 1997, one of my key mathematical innovations was the application of New-
ton’s geometrical methods from the Principia (now 335 years old) to the Complex
Analysis of Cauchy and Riemann (now more than 200 years old). Therefore, I have
not seen any need to update the main body of the text, which has remained almost
unchanged, save for a few additional corrections.
All of the changes to this new edition of VCA—both in appearance and
substance—came about as a direct result of the knowledge I gained creating my sec-
ond and final book,1 Visual Differential Geometry and Forms: A Mathematical Drama
in Five Acts, Princeton University Press, 2021 (hereafter abbreviated to VDGF). In
particular, I thank Wanda España of PUP for her beautiful work on the design of
VDGF, for this in turn inspired some design elements that I requested of OUP for
this new edition of VCA. Of course I also thank Oxford University Press for accept-
ing some of these suggestions, while also imprinting their own distinctive style on
the work.

• The most obvious change is in the physical dimensions of the book: it has
expanded from 6 ′′ ×9 ′′ to 7 ′′ ×10 ′′ . There are two significant advantages to
this change: (1) the book is more comfortable to hold and read; (2) the 5032

1
Needham (2021).
2
The original edition only contained 501 figures: The figure in this Preface and [6.41] are both new.
xii Preface to the 25th Anniversary Edition

figures—the beating heart of the work!—are now 36% larger, and, one may dare
to hope, 36% clearer!
• Another obvious change is the introduction of the standard numbering sys-
tem for sections, subsections, equations, and figures. In hindsight, my deliberate
avoidance of the standard scheme was perhaps a rather childish, tantrum-like
expression of my disgust with the prevailing, life-sucking reduction of wildly
exciting mathematical ideas to the arid structure enshrined in traditional trea-
tises: “Lemma 12.7.2 implies Theorem 14.3.8”. Let me take this opportunity to
apologize—better 25 years late than never?—to every professor who was ever
brave enough to adopt VCA as the text for their course, only to discover that
they had to struggle to refer their students to any given figure or result!
• The Bibliography has been updated. Not only have previously cited works been
updated to their latest editions, but I have also added a number of new works,
which were not cited in the first edition, for the simple reason that 25 years ago
they did not exist!
• The Index has been improved significantly. Many new entries have been added,
and, perhaps more importantly, wherever a single main entry formerly listed
a long, frustratingly unhelpful string of undifferentiated page numbers, I have
now split it into many individual and helpfully specific subentries. By way of proof,
consider the entry for “Jacobian matrix”, then and now!
• Giving credit to Eugenio Beltrami. In the original edition I pointed out that in 1868
Beltrami discovered3 (and published) the conformal models of the hyperbolic
plane, which Poincaré then rediscovered 14 years later, in 1882. Now, following
the example I set myself in VDGF, I have gone one step further, attempting to put
the record straight by renaming these models as the Beltrami–Poincaré disc and
half-plane. Beltrami also discovered the projective model, and I have renamed
that, too, as the Beltrami–Klein model. Correspondingly, my fancifully named
inhabitants of these hyperbolic worlds have been renamed from “Poincarites”
to Beltrami–Poincarites!
• Twenty five years ago, VCA was on the bleeding edge of what was typograph-
ically possible. Indeed, my editor told me that VCA was the first mathematics
book published by Oxford University Press to be composed in LATEX and yet not
to be typeset using the standard Computer Modern fonts designed by Donald
Knuth, the creator of TEX. I was able to achieve this feat by virtue of Michael
Spivak’s4 then newly created MathTime typeface, and by virtue of the wonderful

3
See Milnor (1982), Stillwell (1996), and Stillwell (2010).
4
Yes, the famous differential geometer!
 Preface to the 25th Anniversary Edition xiii

Y&Y TEX System for Windows.5 I was thereby able to typeset the book in Times
text and (mainly) MathTime mathematics—a vast aesthetic improvement over
Computer Modern, in my humble but strong opinion!
Needless to say, the TEX world has moved on considerably in the past 25 years.
But in order to take advantage of these advances, I was forced to grapple with
the task of updating my original TEX files (including my countless macros) from
the ancient (dead?) language of LATEX2.09 to the modern LATEX 2ε . This effort was
rewarded with the ability to typeset this new edition of VCA using the same type-
faces that I very carefully selected for VDGF, all three of which sprang from the
genius mind of Hermann Zapf (1918–2015): Optima for the headings, Palatino for
the text, and, crucially, the remarkable Euler fonts for the mathematics.
Let me pause for a moment to pay tribute to Zapf’s Euler mathematical fonts.
They would appear to me to be at home onboard the starship of an advanced alien
civilization, yet they also evoke the time-worn stone engravings within an ancient
Greek ruin—their beauty transcends space and time: Picture an inscription on the
Guardian of Forever!6
But my fatal attraction to these fonts brought about conflict between the Euler
mathematics in the text of the new edition and the MathTime mathematics in all
the figures of the original edition. Having been born cursed (and I suppose blessed)
with a compulsion to strive for perfection, I had no choice but to undertake the
self-inflicted, Herculean task of hand-editing (within CorelDRAW) all 501 of my
original, hand-drawn figures! I then output new versions in which each MathTime
symbol is here replaced with its matching Euler counterpart, thereby bringing the
figures into perfect alignment with the new text.
I am the impossibly proud father of remarkable twin daughters, Faith and Hope.
I am also the father of VCA and of VDGF. It is therefore a source of deep, resonating
joy that both sets of twins now look like twins!

• Now let me explain the most fundamental change, the one that took me the
greatest time and effort to accomplish, and the one that I believe transforms this
25th Anniversary Edition into a truly new edition, one that may even be of value
to owners of the original edition.
The 503 figures are the mathematical soul of the work. They crystallize all the
geometrical insights I was able to glean from my many years spent struggling to
understand Complex Analysis. Yet, in the entire original edition, you will not find a
single caption. Why?!
Sadly, the answer is simple: cowardice. As a newly minted DPhil student of Pen-
rose, with no track record or reputation, I feared that the mathematical community

5
Sadly, this system (designed by Professor Berthold Horn of MIT) ceased to exist in 2004.
6
The space–time portal featured in the Star Trek episode, “The City on the Edge of Forever”.
xiv Preface to the 25th Anniversary Edition

would reject or even ridicule7 my Newtonian arguments. I therefore sought to con-

trol the narrative (literally!) by only revealing the arguments within the text proper,
where I could spell out my reasoning in full.
Well, 25 years later, I am certainly older, perhaps a tad wiser, and I have also
gained a modicum of confidence by virtue of the enthusiastic reviews that VCA has
since received in all the major journals. Perhaps more emotionally significant to me
has been the large number of individual notes of appreciation I have received, which
I continue to receive to this day, from readers of all stripes—graduate students,
professors, and working scientists—from around the globe.
So, finally, I have done something that I should have had the courage to do at
the outset:

In this 25th Anniversary Edition, every figure now has a caption

that fully explains its mathematical content. Additionally, many of
these new captions include a title (in bold print) that further distils
the figure’s meaning down to its essence.

This approach8 is directly inspired by the works of my teacher, friend, and mentor,
Sir Roger Penrose—to whom this book is dedicated, and to whom I now offer sin-
cere thanks for his generous Foreword!—most remarkably in his Road to Reality,9
where a single figure’s caption can take up a quarter of a page!
This innovation now makes it possible to read VCA in an entirely new way—as
a highbrow comic book! Much of the geometrical reasoning of the work can now be
grasped simply by studying a figure and its accompanying caption, only turning
to the main text for the complete explanation as needed. Furthermore, instead of
undertaking a systematic, linear reading of the work, you are now invited to skip
and hop about, lighting upon whichever figure happens to catch your eye.

• Newton’s concept of ultimate equality underlies many of the arguments in this

book, but its use was not made explicit in the original edition, and this led some
to suppose that the arguments were less rigorous than they actually were (and
remain). In this new edition I have only occasionally modified the main text, but
I have felt at liberty to explicitly introduce ultimate equalities into some of the
newly minted captions. As in my earlier works,10 I have used the symbol ≍ to
denote this concept.

Let me now expand upon this vitally important point . . .

7
In my defence, such fears were not entirely groundless: When Princeton University Press sent out
draft chapters of VDGF for review, one of the three anonymous reviewers bluntly declared, “This is
not even mathematics!”
8
This is also my approach in VDGF.
9
Penrose (2005).
10
Needham (1993), Needham (2014).
 Preface to the 25th Anniversary Edition xv

More on Newton’s Principia and His Concept of Ultimate Equality

As I explain in the original Preface (immediately following this new Preface), VCA
(and VDGF) could never have come to be had I not undertaken a careful study
of Newton’s Principia in 1982, while I was still a DPhil student of Penrose. As I
struggled to penetrate Newton’s ancient diagrams, I intuitively sensed the power
and beauty of his geometrical reasoning, long before I fully understood it.
But one aspect of Newton’s thinking did make sense to me immediately: why
should Newton settle for studying an equation describing the orbit of a planet
around the Sun when he possessed a geometrical tool of enormous power, capa-
ble of instead analyzing the orbit itself ?! This in turn provoked within me a
life-changing mathematical crisis of conscience: how many other mathematical phe-
nomena had I fooled myself into pretending I had understood, when all I had done
was grasp an equation that described the phenomenon from afar, rather than daring
to stare mathematical reality squarely and geometrically in the face?
I felt excited and strangely fearful, like one of the apes in the opening of Stanley
Kubrick’s, 2001: A Space Odyssey, shrieking and wildly leaping about in the stark
silent presence of the black monolith. I wanted to be that one ape that was brave
enough to touch the monolith, now taking the form of a different ancient reservoir
of deeply alien knowledge: Newton’s mind!
So I did. And once I had done so, I stared down at Newton’s geometrical con-
cept of ultimate equality, as the ape had stared down at the large femur bone on
the ground before him, picked it up, gingerly at first, examined it closely, and then
spent the next 35 years of my life wielding it with all the force and ingenuity I
could muster—a tool to crack open and lay bare to our visual intuition the secrets
of Complex Analysis, and, 25 years later, the secrets of Differential Geometry and
Forms.
The presentation of the geometrical reasoning in VDGF appears to be more rig-
orous than that in VCA because there I explicitly make use of Newton’s concept of
ultimate equality. Let me now steal an excerpt from the Prologue of VDGF in order
to explain what this means.
As I have discussed elsewhere,11 Newtonian scholars have painstakingly
dismantled the pernicious myth12 that the results in the 1687 Principia were first
derived by Newton using his original 1665 version of the calculus, and only later
recast into the geometrical form that we find in the finished work.
Instead, it is now understood that by the mid-1670s, having studied Apollonius,
Pappus, and Huygens, in particular, the mature Newton became disenchanted with

11
See Needham (1993), the original 1997 Preface to VCA, and Needham (2014).
12
Sadly, this myth originated with Newton himself, in the heat of his bitter priority battle with
Leibniz over the discovery of the calculus. See Arnol’d (1990), Bloye and Huggett (2011), de Gandt
(1995), Guicciardini (1999), Newton (1687, p.123), and Westfall (1980).
xvi Preface to the 25th Anniversary Edition

the form in which he had originally discovered the calculus in his youth—which
is different again from the Leibnizian form we all learn in college today—and had
instead embraced purely geometrical methods.
Thus it came to pass that by the 1680s Newton’s algebraic infatuation with power
series gave way to a new form of calculus—what he called the “synthetic method
of fluxions”13 —in which the geometry of the Ancients was transmogrified and
reanimated by its application to shrinking geometrical figures in their moment of
vanishing. This is the potent but non-algorithmic form of calculus that we find in
full flower in his great Principia of 1687.
Let me now immediately spell out Newton’s approach, and in significantly
greater detail than I did in the first edition of VCA, in the vain hope that this
new edition may bolster my efforts in VDGF to inspire more mathematicians and
physicists to adopt Newton’s intuitive (yet rigorous14 ) methods than did the first
edition.
If two quantities A and B depend on a small quantity ϵ, and their ratio
approaches unity as ϵ approaches zero, then I shall avoid the more cumbersome
language of limits by following Newton’s precedent in the Principia, saying simply
that, “A is ultimately equal to B”. Also, as I did in earlier works [(Needham, 1993),
(Needham, 2014)], I shall employ the symbol ≍ to denote this concept of ultimate
equality.15 In short,
A
“A is ultimately equal to B” ⇐⇒ A≍B ⇐⇒ lim
= 1.
B ϵÑ0

It follows [exercise] from the theorems on limits that ultimate equality is an equiv-
alence relation, and that it also inherits additional properties of ordinary equality,
e.g.,

X≍Y&P≍Q =⇒ X · P ≍ Y · Q, and A ≍ B · C ⇐⇒ (A/B) ≍ C.

Before I begin to apply this idea in earnest, I also note that the jurisdiction of
ultimate equality can be extended naturally to things other than numbers, enabling
one to say, for example, that two triangles are “ultimately similar”, meaning that
their angles are ultimately equal.
As I explain in the original Preface, having grasped Newton’s method, I imme-
diately tried my own hand at using it to simplify my teaching of introductory
calculus, only later realizing how I might apply it to Complex Analysis (in VCA),
and later still to Differential Geometry (in VDGF). Though I might choose any num-
ber of simple, illustrative examples [see Needham (1993) for more], I will reuse the
specific one I gave in the original Preface to VCA, and for one simple reason: this

13
See Guicciardini (2009, Ch. 9).
14
Fine print to follow!
15
This notation was subsequently adopted (with attribution) by the Nobel physicist, Subrahmanyan
Chandrasekhar (see Chandrasekhar, 1995, p. 44).
 Preface to the 25th Anniversary Edition xvii

time I will use the “≍”–notation to present the argument rigorously, whereas in the
first edition I did not. Indeed, this example may be viewed as a recipe for trans-
forming many of VCA’s “explanations” into “proofs”,16 merely by sprinkling on
the requisite ≍’s. With the addition of figure captions, I am now able to do some of
this sprinkling myself, but some must still be left to the reader.
Before looking at this example, I suggest you first read the original presentation
of the argument in the original Preface (immediately following this one) and then
return to this point.

Let us show that if T = tan θ, then dT 2

dθ = 1 + T (see figure). If we increase θ
by a small (ultimately vanishing) amount δθ, then T will increase by the length of
the vertical hypotenuse δT of the small triangle, in which the other two sides of
this triangle have been constructed to lie in the directions (θ + δθ) and (θ + π2 ), as
illustrated. To obtain the result, we first observe that in the limit that δθ vanishes,
the small triangle with hypotenuse δT is ultimately similar to the large triangle
with hypotenuse L, because ψ ≍ π2 . Next, as we see in the magnifying glass, the
side δs adjacent to θ in the small triangle is ultimately equal to the illustrated arc
of the circle with radius L, so δs ≍ L δθ; note that we have moved the dot from the
corner of the triangle to the arc, to stress this point. Thus,
dT δT δT L dT
≍ ≍ ≍ =⇒ = L2 = 1 + T 2 .
Ldθ Lδθ δs 1 dθ
So far as I know, Newton never wrote down this specific example, but compare
the illuminating directness of his style of geometrical reasoning with the unillumi-
nating computations we teach our students today, more than three centuries later!
As Newton himself put it,17 the geometric method is to be preferred by virtue of
the “clarity and brevity of the reasoning involved and because of the simplicity of
the conclusions and the illustrations required.” Indeed, Newton went even further,
resolving that only the synthetic method was “worthy of public utterance”.

16
I was already using the ≍ notation (both privately and in print) at the time of writing VCA, and,
in hindsight, it was a mistake that I did not employ it throughout the original edition of VCA.
17
See Guicciardini (2009, p. 231)
xviii Preface to the 25th Anniversary Edition

Newton himself did not employ any symbol to represent his concept of “ultimate
equality”. Instead, his devotion to the geometrical method of the Ancients spilled
over into emulating their mode of expression, causing him to write out the words
“ultimately have the ratio of equality” every single time the concept was invoked
in a proof. As Newton (1687, p. 124) explained, the Principia is “written in words at
length in the manner of the Ancients”. Even when Newton claimed that two ratios
were ultimately equal, he insisted on expressing each ratio in words. As a result, I
myself was quite unable to follow Newton’s reasoning without first transcribing
and summarizing each of his paragraphs into “modern” form (which was in fact
already quite common in 1687). Indeed, back in 1982, this was the catalyst for my
private introduction and use of the symbol ≍.
It is my view that Newton’s choice not to introduce a symbol for “ultimate
equality” was a tragically consequential error for the development of mathemat-
ics. As Leibniz’s symbolic calculus swept the world, Newton’s more penetrating
geometrical method fell by the wayside. In the intervening centuries only a hand-
ful of people ever sought to repair this damage and revive Newton’s approach,
the most notable and distinguished recent champion having been V. I. Arnol’d18
[1937–2010].
Had Newton shed the trappings of this ancient mode of exposition and instead
employed some symbol (any symbol!) in place of the words “ultimately have the
ratio of equality”, his dense, paragraph-length proofs in the Principia might have
been reduced to a few succinct lines, and his mode of thought might still be widely
employed today. Both VCA and VDGF are attempts to demonstrate, very con-
cretely, the continuing relevance and vitality of Newton’s geometrical approach,
in areas of mathematics whose discovery lay a century in the future at the time of
his death in 1727.
Allow me to insert some fine print concerning my use of the words “rigour”
and “proof”. Yes, my occasional explicit use of Newtonian ultimate equalities in
this new edition represents a quantum jump in rigour, as compared to my original
exposition in VCA, but there will be some mathematicians who will object (with
justification!) that even this increase in rigour is insufficient, and that none of the
“proofs” in this work are worthy of that title, including the one just given: I did
not actually prove that the side of the triangle is ultimately equal to the arc of the
circle.
I can offer no logical defence, but will merely repeat the words I wrote in the
original Preface to VCA, 25 years ago: “... suppose one believes, as I do, that our
mathematical theories are attempting to capture aspects of a robust Platonic world
that is not of our making. I would then contend that an initial lack of rigour is a
small price to pay if it allows the reader to see into this world more directly and

18
See, for example, Arnol’d (1990).
 Preface to the 25th Anniversary Edition xix

pleasurably than would otherwise be possible.” So, to preemptively address my

critics, let me therefore concede, from the outset, that when I claim that an assertion
is “proved”, it may be read as, “proved beyond a reasonable doubt”!19

The Continued Relevance of the Unorthodox Contents of VCA

Leaving aside its unorthodox geometrical methods, VCA is also distinguished by
its unorthodox contents. The concept of the complex derivative (or amplitwist, as I
call it) is not even introduced until Chapter 4, and many of the most interesting,
unusual, and important parts of the book make little or no use of analysis.
For many decades before VCA, and now for decades after VCA, almost all20
introductions to Complex Analysis have seemed to follow almost exactly in the
footsteps of the ones that came before—the same topics, explained in the same
order, and in the same manner, like a single-file procession of monks through the
snow, quietly and sombrely intent on stepping into the footprints of the monk who
went before them: Do not disturb the pristine snow lying all about you! But when I
embarked upon VCA, 35 years ago, I was like a happily disobedient child, jumping
about wildly, rolling my whole mind in the snow!
I started from scratch, asking myself, which ideas connected to Complex Analysis
have become the most vital to modern mathematics and physics? A pair of closely related
answers to this question immediately presented themselves, and while I make no
claim to prescience, it is certainly true that the following have only become more
vital to mathematics and physics over the past 25 years:

• Hyperbolic Geometry violates the normal rules of Euclidean geometry, and it

is therefore also called Non-Euclidean Geometry. In the hyperbolic plane there are
infinitely many different lines through a given point that are parallel to a given
line, and the angles in a triangle always add up to less than π, the departure from
π being proportional to the area of the triangle.

Yet, as Poincaré was the first to recognize, this strange geometry arises naturally
across many parts of mathematics and physics. For example, the final figure of this
book reveals how hyperbolic geometry unifies all the methods of solving the two-
dimensional Dirichlet Problem. Furthermore, the visionary insights of Thurston21

19
Upon reading these words, a strongly supportive member of the Editorial Board of Princeton
University Press suggested to my editor that in place of “Q.E.D.”, I conclude each of my proofs in
VDGF with the letters, “P.B.R.D.”!
20
I freely admit that I have not undertaken a systematic study of all the Complex Analysis textbooks
that have been published during the last 25 years, and I am aware that there exist excellent exceptions to
the following generalization, my favourite ones being Shaw (2006) and Stewart and Tall (2018), which
I highly recommend precisely because of their very unusual contents. Incidentally, Shaw just happens
to have been a fellow student of Penrose!
21
For details of Thurston’s Geometrization Conjecture—now Theorem!—see Thurston (1997).
xx Preface to the 25th Anniversary Edition

(1946–2012)—subsequently vindicated by Perelman in 2003, six years after the pub-

lication of VCA—have established that hyperbolic geometry is in some sense more
fundamental than Euclidean geometry.
This fascinating and important geometry is intimately entwined with the complex
numbers, via the Möbius transformations that we shall discuss next. Yet, after a
quarter century, VCA’s long Chapter 6 remains the most complete and the most
geometrical treatment of hyperbolic geometry that I have seen in any introduction
to Complex Analysis.

• Möbius Transformations are mappings of the form

az + b
z ÞÑ M(z) = ,
cz + d
where a, b, c, d are complex constants. Whereas hyperbolic geometry is rarely
even mentioned in introductory texts on Complex Analysis, these transforma-
tions are discussed in all of them, but usually in a perfunctory and superficial
manner. In stark contrast, my long Chapter 3 remains the most complete and
the most geometrical treatment of these transformations that I have seen in any
introductory text.

The reason I lavished such extravagant attention upon these deceptively sim-
ple transformations is that they are possessed of magical powers, manifesting
themselves in multifaceted guises across mathematics and physics. Again, it was
Poincaré who was the first to recognize this. They are the isometries of both the two-
dimensional hyperbolic plane (introduced above) and of three-dimensional hyper-
bolic space; they are the famous Lorentz transformations (isometries) of Minkowski
and Einstein’s four-dimensional spacetime; and when the constants are all integers,
and ad − bc = 1, they form the modular group, describing the symmetries of the
modular functions so important in modern number theory; and the list goes on . . .
The second pair of innovations with respect to the contents of VCA is centred
on the use of vector fields as an alternative means of visualizing complex mappings.
Instead of picturing a point z in C as being mapped to another point w = f(z)
in another copy of C—which is the paradigm in force throughout the first nine
chapters—we instead picture f(z) as a vector emanating from z.

• The Topology of Vector Fields is the subject of Chapter 10. Here we shed
new light on the vital topological concept of the winding number (the subject
of Chapter 7) by instead viewing it through the prism of the index of a singular
point of a vector field. The climax of the topological analysis is a proof of the
glorious Poincaré–Hopf Theorem, (10.4), which relates the indices of a flow on a
closed surface to the surface’s topological genus, which counts how many holes
the surface has.
 Preface to the 25th Anniversary Edition xxi

So far as I can tell, my final22 major innovative topic in VCA has still not had the
impact I believe it deserves, despite my best efforts, and despite the even earlier,
independent efforts of its first principal champion, Professor Bart Braden, who
should also be credited23 with having named the concept in honour of its first
proponent:

• The Pólya Vector Field of a complex mapping z ÞÑ f(z) was introduced by

George Pólya (1887–1985) in his 1974 textbook with Gordon Latta, (Pólya and
Latta, 1974), and it is the subject of VCA’s Chapter 11: At each point z in C
we draw a vector P(z) that is defined to be the complex conjugate of f(z), so
P(z) ≡ f(z). This Pólya vector field then provides a wonderfully vivid means of
visualizing the real and imaginary parts of the integral24 of f(z) along a directed
path (aka a contour) C connecting two points in C, namely, (11.1) on page 549:
Z
f(z) dz = [flow of P along C] + i [flow of P across C] .
C

Here, the imaginary flux component is positive if P flows across C from our left to our
right as we face forward in the direction of travel along C.
Next, it follows immediately from the Cauchy–Riemann equations, (4.7), that
there is a wonderfully vivid physical interpretation of the existence of the complex
derivative (what I call the amplitwist), namely,
If and only if the amplitwist f ′ (z) exists, then the Pólya vector
field P(z) ≡ f(z) is divergence-free and curl-free.
This in turn immediately provides a marvellously physical explanation of one the
central results of Complex Analysis, namely, Cauchy’s Theorem, (8.20), the simplest
version of which states that
If the amplitwist f ′ exists everywhere on and inside a simple
H
closed loop L, then f(z)dz = 0.
L

But if L contains a singularity of f(z), then the integral need not vanish. An example
of central importance is f(z) = (1/z), for which the Pólya vector field is the radial
outward flow from a source at the origin. Choosing L to be an origin-centred circle

22
In fact there are many other innovations scattered across the work, most notably my introduction
of a concept I christened the complex curvature, which I apply to central force fields in Chapter 5, and
to the geometry of harmonic functions in Chapter 12; there’s also the Topological Argument Principle in
Chapter 7, and more besides, but I cannot attempt to catalogue all these ideas and discoveries here.
Some of the observations in Chapter 12 were previously published in Needham (1994), which won the
MAA’s Carl B. Allendoerfer Award in 1995.
23
See Braden (1985), Braden (1987)—which won the MAA’s Carl B. Allendoerfer Award in 1988—and
Braden (1991).
24
Rest assured that the following ideas will all be explained ab initio in the main text; this section
of the Preface is addressed to experienced readers who are already familiar with the fundamentals of
standard Complex Analysis.
xxii Preface to the 25th Anniversary Edition

K of radius r, traced counterclockwise, we obtain an immediate visual and physical

explanation of the iconic fact that
I
1
dz = 2πi.
K z

For clearly there is no flow along K, and since P(z) = (1/z) flows orthogonally
across K (from left to right) with speed (1/r), its flux across K is (2πr)(1/r) = 2π.
Furthermore, this physical interpretation explains why the value of the integral will
remain 2πi if K is continuously deformed into a general loop encircling the source
at the origin, so long as K does not cross that point as it is deformed.
Lastly, the ability of modern computers to quickly and easily draw the Pólya
vector field of any explicit formula for f(z) makes the concept all the more powerful
as a means of visualizing Complex Analysis.
While the evangelical work of VCA (and of Braden’s earlier papers) failed to
have its desired effect in the 20th century, my fervent hope now is that by explic-
itly singling out this concept for praise in this new Preface, more people may take
notice of it, and the gospel of the Pólya vector field may thereby be spread amongst
multitudes of new believers in the 21st century!

The Quiet Revolution

In hindsight, something was clearly in the air. No sooner had VCA been published
than I began to notice the emergence of a wave of kindred works that likewise
challenged the prevailing dominance of an arid, purely formal approach to math-
ematics, and that instead embraced intuitive explanations, the meaning of results,
and, crucially, the revelatory power of geometry.
It appeared that I had unwittingly been enlisted into a resistance movement
of sorts, one in which cell members were not permitted to know each others
identities—perhaps for their own safety?! But, unlike Bourbaki, this quiet revo-
lution had no name and no leaders; it was, to coin a phrase, “By the people, for
the people.” Long may this healthy embrace of meaning, intuition, and geometry
continue!

Cheers!
I raise my glass to the next 25 years!

T. N.
Mill Valley, California
June, 2022
PREFACE

Theories of the known, which are described by different physical ideas, may be equivalent
in all their predictions and hence scientifically indistinguishable. However, they are not
psychologically identical when trying to move from that base into the unknown. For differ-
ent views suggest different kinds of modifications which might be made and hence are not
equivalent in the hypotheses one generates from them in one’s attempt to understand what
is not yet understood.
Feynman (1966)

A Parable
Imagine a society in which the citizens are encouraged, indeed compelled up to
a certain age, to read (and sometimes write) musical scores. All quite admirable.
However, this society also has a very curious—few remember how it all started—
and disturbing law: Music must never be listened to or performed!
Though its importance is universally acknowledged, for some reason music is
not widely appreciated in this society. To be sure, professors still excitedly pore
over the great works of Bach, Wagner, and the rest, and they do their utmost to
communicate to their students the beautiful meaning of what they find there, but
they still become tongue-tied when brashly asked the question, “What’s the point
of all this?!”
In this parable, it was patently unfair and irrational to have a law forbid-
ding would-be music students from experiencing and understanding the subject
directly through “sonic intuition.” But in our society of mathematicians we have
such a law. It is not a written law, and those who flout it may yet prosper, but it
says, Mathematics must not be visualized!
More likely than not, when one opens a random modern mathematics text on a
random subject, one is confronted by abstract symbolic reasoning that is divorced
from one’s sensory experience of the world, despite the fact that the very phenom-
ena one is studying were often discovered by appealing to geometric (and perhaps
physical) intuition.
This reflects the fact that steadily over the last hundred years the honour of
visual reasoning in mathematics has been besmirched. Although the great mathe-
maticians have always been oblivious to such fashions, it is only recently that the
“mathematician in the street” has picked up the gauntlet on behalf of geometry.
xxiv Preface

The present book openly challenges the current dominance of purely symbolic
logical reasoning by using new, visually accessible arguments to explain the truths
of elementary complex analysis.

Computers
In part, the resurgence of interest in geometry can be traced to the mass-availability
of computers to draw mathematical objects, and perhaps also to the related, some-
what breathless, popular interest in chaos theory and in fractals. This book instead
advocates the more sober use of computers as an aid to geometric reasoning.
I have tried to encourage the reader to think of the computer as a physicist would
his laboratory—it may be used to check existing ideas about the construction of
the world, or as a tool for discovering new phenomena which then demand new
ideas for their explanation. Throughout the text I have suggested such uses of the
computer, but I have deliberately avoided giving detailed instructions. The reason
is simple: whereas a mathematical idea is a timeless thing, few things are more
ephemeral than computer hardware and software.
Having said this, the program “f(z)” is currently25 the best tool for visually
exploring the ideas in this book; a free demonstration version can be downloaded
directly from Lascaux Graphics. On occasion it would also be helpful if one had
access to an all-purpose mathematical engine such as Maple or Mathematica. How-
ever, I would like to stress that none of the above software is essential: the entire
book can be fully understood without any use of a computer.
Finally, some readers may be interested in knowing how computers were used
to produce this book. Perhaps five of the 501 diagrams were drawn using output
from Mathematica; the remainder I drew by hand (or rather “by mouse”) using
CorelDRAW, occasionally guided by output from “f(z)”. I typeset the book in LATEX
using the wonderful Y&Y TEX System for Windows26 , the figures being included as
EPS files. The text is Times27 , with Helvetica heads, and the mathematics is princi-
pally MathTime, though nine other mathematical fonts make cameo appearances.
All of these Adobe Type 1 fonts were obtained from Y&Y, Inc., with the exception
of Adobe’s MathematicalPi-Six font, which I used to represent quaternions. Finally,
OUP printed the book directly from my PostScript file.

The Book’s Newtonian Genesis

It is fairly well known that Newton’s original 1665 version of the calculus was dif-
ferent from the one we learn today: its essence was the manipulation of power

25
Sadly, this no longer exists.
26
Sadly, this no longer exists.
27
In this new edition, the headings are Optima, the text is Palatino, and the mathematics is Euler.
 Preface xxv

series, which Newton likened to the manipulation of decimal expansions in arith-

metic. The symbolic calculus—the one in every standard textbook, and the one now
associated with the name of Leibniz—was also perfectly familiar to Newton, but
apparently it was of only incidental interest to him. After all, armed with his power
R 2 R
series, Newton could evaluate an integral like e−x dx just as easily as sinx dx.
Let Leibniz try that!
It is less well known that around 1680 Newton became disenchanted with both
these approaches, whereupon he proceeded to develop a third version of calcu-
lus, based on geometry. This “geometric calculus” is the mathematical engine that
propels the brilliant physics of Newton’s Principia.
Having grasped Newton’s method, I immediately tried my own hand at using it
to simplify my teaching of introductory calculus. An example will help to explain
what I mean by this. Let us show that if T = tan θ, then dT 2
dθ = 1 + T . If we increase
θ by a small amount dθ then T will increase by the amount dT in the figure below.
To obtain the result, we need only observe that in the limit as dθ tends to zero, the
black triangle is ultimately similar [exercise] to the shaded triangle. Thus, in this
limit,

dT L dT
= =⇒ = L2 = 1 + T 2 .
L dθ 1 dθ

Only gradually did I come to realize how naturally this mode of thought could be
applied—almost exactly 300 years later!—to the geometry of the complex plane.

Reading This Book

In the hope of making the book fun to read, I have attempted to write as though
I were explaining the ideas directly to a friend. Correspondingly, I have tried
to make you, the reader, into an active participant in developing the ideas. For
example, as an argument progresses, I have frequently and deliberately placed a
pair of logical stepping stones sufficiently far apart that you may need to pause
xxvi Preface

and stretch slightly to pass from one to the next. Such places are marked “[exer-
cise]”; they often require nothing more than a simple calculation or a moment of
reflection.
This brings me to the exercises proper, which may be found at the end of each
chapter. In the belief that the essential prerequisite for finding the answer to a
question is the desire to find it, I have made every effort to provide exercises that
provoke curiosity. They are considerably more wide-ranging than is common, and
they often establish important facts which are then used freely in the text itself.
While problems whose be all and end all is routine calculation are thereby avoided,
I believe that readers will automatically develop considerable computational skill
in the process of seeking solutions to these problems. On the other hand, my inten-
tion in a large number of the exercises is to illustrate how geometric thinking can
often replace lengthy calculation.
Any part of the book marked with a star (“*”) may be omitted on a first read-
ing. If you do elect to read a starred section, you may in turn choose to omit any
starred subsections. Please note, however, that a part of the book that is starred is
not necessarily any more difficult, nor any less interesting or important, than any
other part of the book.

Teaching from this Book

The entire book can probably be covered in a year, but in a single semester course
one must first decide what kind of course to teach, then choose a corresponding
path through the book. Here I offer just three such possible paths:

• Traditional Course. Chapters 1 to 9, omitting all starred material (e.g., the whole
of Chapter 6).
• Vector Field Course. In order to take advantage of the Pólya vector field
approach to visualizing complex integrals, one could follow the “Tradi-
tional Course” above, omitting Chapter 9, and adding the unstarred parts of
Chapters 10 and 11.
• Non-Euclidean Course. At the expense of teaching any integration, one could
give a course focused on Möbius transformations and non-Euclidean geometry.
These two related parts of complex analysis are probably the most important
ones for contemporary mathematics and physics, and yet they are also the ones
that are almost entirely neglected in undergraduate-level texts. On the other
hand, graduate-level works tend to assume that you have already encountered
the main ideas as an undergraduate: Catch 22!

Such a course might go as follows: All of Chapter 1; the unstarred parts of

Chapter 2; all of Chapter 3, including the starred sections but (possibly) omitting
 Preface xxvii

the starred subsections; all of Chapter 4; all of Chapter 6, including the starred
sections but (possibly) omitting the starred subsections.

Omissions and Apologies

If one believes in the ultimate unity of mathematics and physics, as I do, then a
very strong case for the necessity of complex numbers can be built on their appar-
ently fundamental role in the quantum mechanical laws governing matter. Also, the
work of Sir Roger Penrose has shown (with increasing force) that complex numbers
play an equally central role in the relativistic laws governing the structure of space-
time. Indeed, if the laws of matter and of spacetime are ever to be reconciled, then it
seems very likely that it will be through the auspices of the complex numbers. This
book cannot explore these matters; instead, we refer the interested reader to Feyn-
man (1963, Vol. III), Feynman (1985), Penrose and Rindler (1984), and especially
Penrose (2005).
A more serious omission is the lack of discussion of Riemann surfaces, which I
had originally intended to treat in a final chapter. This plan was aborted once it
became clear that a serious treatment would entail expanding the book beyond
reason. By this time, however, I had already erected much of the necessary scaf-
folding, and this material remains in the finished book. In particular, I hope that
the interested reader will find the last three chapters helpful in understanding Rie-
mann’s original physical insights, as expounded by Klein (1881). See also Springer
(1981, Ch. 1), which essentially reproduces Klein’s monograph, but with additional
helpful commentary.
I consider the history of mathematics to be a vital tool in understanding both the
current state of mathematics, and its trajectory into the future. Sadly, however, I can
do no more than touch on historical matters in the present work; instead I refer you
to the remarkable book, Mathematics and Its History, by John Stillwell (2010). Indeed,
I strongly encourage you to think of his book as a companion to mine: not only does
it trace and explain the development of complex analysis, but it also explores and
illuminates the connections with other areas of mathematics.
To the expert reader I would like to apologize for having invented the word
“amplitwist” [Chapter 4] as a synonym (more or less) for “derivative”, as well the
component terms “amplification” and “twist”. I can only say that the need for some
such terminology was forced on me in the classroom: if you try teaching the ideas
in this book without using such language, I think you will quickly discover what
I mean! Incidentally, a precedence argument in defence of “amplitwist” might be
that a similar term was coined by the older German school of Klein, Bieberbach,
et al. They spoke of “eine Drehstreckung”, from “drehen” (to twist) and “strecken”
(to stretch).
xxviii Preface

A significant proportion of the geometric observations and arguments contained

in this book are, to the best of my knowledge, new. I have not drawn attention to
this in the text itself as this would have served no useful purpose: students don’t
need to know, and experts will know without being told. However, in cases where
an idea is clearly unusual but I am aware of it having been published by someone
else, I have tried to give credit where credit is due.
In attempting to rethink so much classical mathematics, I have no doubt made
mistakes; the blame for these is mine alone. Corrections will be gratefully received
at [email protected].
My book will no doubt be flawed in many ways of which I am not yet aware,
but there is one “sin” that I have intentionally committed, and for which I shall
not repent: many of the arguments are not rigorous, at least as they stand. This is
a serious crime if one believes that our mathematical theories are merely elaborate
mental constructs, precariously hoisted aloft. Then rigour becomes the nerve-
racking balancing act that prevents the entire structure from crashing down around
us. But suppose one believes, as I do, that our mathematical theories are attempt-
ing to capture aspects of a robust Platonic world that is not of our making. I would
then contend that an initial lack of rigour is a small price to pay if it allows the
reader to see into this world more directly and pleasurably than would otherwise
be possible.
T. N.
San Francisco, California
June, 1996
ACKNOWLEDGEMENTS

First and foremost I wish to express my indebtedness to Dr. Stanley Nel. He is my

friend, my colleague, and my Dean, and in all three of these capacities he has helped
me to complete this book. As a friend he offered support when progress was slow
and my spirits were low; as a mathematical colleague he read much of the book
and offered helpful criticisms; as Dean he granted me a succession of increasingly
powerful computers, and when the US Immigration Service sought to have my
position filled by an “equally qualified” American, he successfully fought them on
my behalf. For all this, and much else besides, I offer him my deep gratitude.
Next I would like to thank Prof. John Stillwell of Monash University. The great
value I place on his writings should be clear from the frequency with which I refer
to them in the pages that follow. Also, though I lack his gift for conciseness, I have
sought to emulate elements of his approach in an attempt to give back meaning
to mathematical concepts. Finally, my greatest and most concrete debt arises from
the fact that he read each draft chapter as it was written, and this despite the fact
that we had never even met! The book owes a great deal to his numerous helpful
suggestions and corrections.
I consider myself very fortunate that the mathematics department here at the
University of San Francisco is completely free of political intrigue, rivalry, and other
assorted academic blights. I am grateful to all my colleagues for creating such a
friendly and supportive atmosphere in which to work. In particular, however, I
should like to single out the following people for thanks:

• Nancy Campagna for her diligent proof-reading of half the book;

• Allan Cruse and Millianne Lehmann, not only for granting all my software
requests during their respective tenures as Department Chair, but also for all
their kind and sage advice since my arrival in the United States;
• James Finch for his patience and expertise in helping me overcome various
problems associated with my typesetting of the book in LATEX;
• Robert Wolf for having built up a superb mathematics collection in our library;
• Paul Zeitz for his great faith in me and in the value of what I was trying to
accomplish, for his concrete suggestions and corrections, and for his courage
in being the first person (other than myself) to teach complex analysis using
chapters of the book.
xxx Acknowledgements

Prof. Gerald Alexanderson of Santa Clara University has my sincere thanks for the
encouragement he offered me upon reading some of the earliest chapters, as well
as for his many subsequent acts of kindness.
I will always be grateful for the education I received at Merton College, Oxford.
It is therefore especially pleasing and fitting to have this book published by OUP,
and I would particularly like to thank Dr. Martin Gilchrist, the former Senior Math-
ematics Editor, for his enthusiastic encouragement when I first approached him
with the idea of the book.
When I first arrived at USF from England in 1989 I had barely seen a com-
puter. The fact that OUP printed this book directly from my Internet-transmitted
PostScript files is an indication of how far I have come since then. I owe all this
to James Kabage. A mere graduate student at the time we met, Jim quickly rose
through the ranks to become Director of Network Services. Despite this fact, he
never hesitated to spend hours with me in my office resolving my latest hardware or
software crisis. He always took the extra time to clearly explain to me the reasoning
leading to his solution, and in this way I became his student.
I also thank Dr. Benjamin Baab, the Executive Director of Information Technol-
ogy Services at USF. Despite his lofty position, he too was always willing to roll up
his sleeves in order to help me resolve my latest Microsoft conundrum.
Eric Scheide (our multitalented Webmaster) has my sincere gratitude for writing
an extremely nifty Perl program that greatly speeded my creation of the index.
I thank Prof. Berthold Horn of MIT for creating the magnificent Y&Y TEX System
for Windows28 , for his generous help with assorted TEXnical problems, and for his
willingness to adopt my few suggestions for improving what I consider to be the
Mercedes-Benz of the TEX world.
Similarly, I thank Martin Lapidus of Lascaux Graphics for incorporating many
of my suggestions into his “f(z)”29 program, thereby making it into an even better
tool for doing “visual complex analysis”.
This new printing of the book incorporates a great many corrections. Most of
these were reported by readers, and I very much appreciate their efforts. While I
cannot thank each one of these readers by name, I must acknowledge Dr. R. von
Randow for single-handedly having reported more than 30 errors.
As a student of Roger Penrose I had the privilege of watching him think out loud
by means of his beautiful blackboard drawings. In the process, I became convinced
that if only one tried hard enough—or were clever enough!—every mathemati-
cal mystery could be resolved through geometric reasoning. George Burnett-Stuart
and I became firm friends while students of Penrose. In the course of our endless
discussions of music, physics, and mathematics, George helped me to refine both

28
While Y & Y no longer exists, my gratitude lives on.
29
While “f(z)” no longer exists, my gratitude lives on.
 Acknowledgements xxxi

my conception of the nature of mathematics, and of what constitutes an acceptable

explanation within that subject. My dedication of this book to these two friends
scarcely repays the great debt I owe them.
The care of several friends helped me to cope with depression following the
death of my beloved mother Claudia. In addition to my brother Guy and my father
Rodney, I wish to express my appreciation to Peter and Ginny Pacheco, and to Amy
Miller. I don’t know what I would have done without their healing affection.
Lastly, I thank my dearest wife Mary. During the writing of this book she allowed
me to pretend that science was the most important thing in life; now that the book
is over, she is my daily proof that there is something even more important.
CONTENTS

1 Geometry and Complex Arithmetic 1

1.1 Introduction 1
1.1.1 Historical Sketch 1
1.1.2 Bombelli’s “Wild Thought” 4
1.1.3 Some Terminology and Notation 6
1.1.4 Practice 8
1.1.5 Equivalence of Symbolic and Geometric
Arithmetic 8
1.2 Euler’s Formula 11
1.2.1 Introduction 11
1.2.2 Moving Particle Argument 12
1.2.3 Power Series Argument 13
1.2.4 Sine and Cosine in Terms of Euler’s Formula 15
1.3 Some Applications 16
1.3.1 Introduction 16
1.3.2 Trigonometry 16
1.3.3 Geometry 18
1.3.4 Calculus 22
1.3.5 Algebra 25
1.3.6 Vectorial Operations 30
1.4 Transformations and Euclidean Geometry* 33
1.4.1 Geometry Through the Eyes of Felix Klein 33
1.4.2 Classifying Motions 38
1.4.3 Three Reflections Theorem 41
1.4.4 Similarities and Complex Arithmetic 44
1.4.5 Spatial Complex Numbers? 48
1.5 Exercises 51

2 Complex Functions as Transformations 61

2.1 Introduction 61
2.2 Polynomials 64
2.2.1 Positive Integer Powers 64
xxxiv Contents

2.2.2 Cubics Revisited* 66

2.2.3 Cassinian Curves* 67
2.3 Power Series 71
2.3.1 The Mystery of Real Power Series 71
2.3.2 The Disc of Convergence 75
2.3.3 Approximating a Power Series with a
Polynomial 78
2.3.4 Uniqueness 79
2.3.5 Manipulating Power Series 81
2.3.6 Finding the Radius of Convergence 83
2.3.7 Fourier Series* 86
2.4 The Exponential Function 88
2.4.1 Power Series Approach 88
2.4.2 The Geometry of the Mapping 89
2.4.3 Another Approach 90
2.5 Cosine and Sine 94
2.5.1 Definitions and Identities 94
2.5.2 Relation to Hyperbolic Functions 95
2.5.3 The Geometry of the Mapping 97
2.6 Multifunctions 100
2.6.1 Example: Fractional Powers 100
2.6.2 Single-Valued Branches of a Multifunction 103
2.6.3 Relevance to Power Series 106
2.6.4 An Example with Two Branch Points 108
2.7 The Logarithm Function 110
2.7.1 Inverse of the Exponential Function 110
2.7.2 The Logarithmic Power Series 112
2.7.3 General Powers 113
2.8 Averaging over Circles* 115
2.8.1 The Centroid 115
2.8.2 Averaging over Regular Polygons 118
2.8.3 Averaging over Circles 121
2.9 Exercises 125

3 Möbius Transformations and Inversion 137

3.1 Introduction 137
3.1.1 Definition and Significance of Möbius
Transformations 137
 Contents xxxv

3.1.2 The Connection with Einstein’s Theory of

Relativity* 138
3.1.3 Decomposition into Simple Transformations 139
3.2 Inversion 139
3.2.1 Preliminary Definitions and Facts 139
3.2.2 Preservation of Circles 142
3.2.3 Constructing Inverse Points Using Orthogonal
Circles 144
3.2.4 Preservation of Angles 147
3.2.5 Preservation of Symmetry 149
3.2.6 Inversion in a Sphere 150
3.3 Three Illustrative Applications of Inversion 153
3.3.1 A Problem on Touching Circles 153
3.3.2 A Curious Property of Quadrilaterals with
Orthogonal Diagonals 154
3.3.3 Ptolemy’s Theorem 156
3.4 The Riemann Sphere 157
3.4.1 The Point at Infinity 157
3.4.2 Stereographic Projection 158
3.4.3 Transferring Complex Functions to the Sphere 162
3.4.4 Behaviour of Functions at Infinity 163
3.4.5 Stereographic Formulae* 165
3.5 Möbius Transformations: Basic Results 168
3.5.1 Preservation of Circles, Angles, and Symmetry 168
3.5.2 Non-Uniqueness of the Coefficients 169
3.5.3 The Group Property 170
3.5.4 Fixed Points 171
3.5.5 Fixed Points at Infinity 172
3.5.6 The Cross-Ratio 174
3.6 Möbius Transformations as Matrices* 177
3.6.1 Empirical Evidence of a Link with Linear
Algebra 177
3.6.2 The Explanation: Homogeneous Coordinates 178
3.6.3 Eigenvectors and Eigenvalues* 180
3.6.4 Rotations of the Sphere as Möbius
Transformations* 182
3.7 Visualization and Classification* 184
3.7.1 The Main Idea 184
xxxvi Contents

3.7.2 Elliptic, Hyperbolic, and Loxodromic

Transformations 186
3.7.3 Local Geometric Interpretation of the Multiplier 189
3.7.4 Parabolic Transformations 190
3.7.5 Computing the Multiplier* 191
3.7.6 Eigenvalue Interpretation of the Multiplier* 192
3.8 Decomposition into 2 or 4 Reflections* 194
3.8.1 Introduction 194
3.8.2 Elliptic Case 194
3.8.3 Hyperbolic Case 196
3.8.4 Parabolic Case 197
3.8.5 Summary 198
3.9 Automorphisms of the Unit Disc* 199
3.9.1 Counting Degrees of Freedom 199
3.9.2 Finding the Formula via the Symmetry Principle 200
3.9.3 Interpreting the Simplest Formula
Geometrically* 201
3.9.4 Introduction to Riemann’s Mapping Theorem 204
3.10 Exercises 205

4 Differentiation: The Amplitwist Concept 213

4.1 Introduction 213
4.2 A Puzzling Phenomenon 213
4.3 Local Description of Mappings in the Plane 216
4.3.1 Introduction 216
4.3.2 The Jacobian Matrix 217
4.3.3 The Amplitwist Concept 218
4.4 The Complex Derivative as Amplitwist 220
4.4.1 The Real Derivative Re-examined 220
4.4.2 The Complex Derivative 221
4.4.3 Analytic Functions 223
4.4.4 A Brief Summary 224
4.5 Some Simple Examples 225
4.6 Conformal = Analytic 227
4.6.1 Introduction 227
4.6.2 Conformality Throughout a Region 228
4.6.3 Conformality and the Riemann Sphere 230
4.7 Critical Points 231
4.7.1 Degrees of Crushing 231
 Contents xxxvii

4.7.2 Breakdown of Conformality 232

4.7.3 Branch Points 233
4.8 The Cauchy–Riemann Equations 234
4.8.1 Introduction 234
4.8.2 The Geometry of Linear Transformations 235
4.8.3 The Cauchy–Riemann Equations 237
4.9 Exercises 239

5 Further Geometry of Differentiation 245

5.1 Cauchy–Riemann Revealed 245
5.1.1 Introduction 245
5.1.2 The Cartesian Form 245
5.1.3 The Polar Form 247
5.2 An Intimation of Rigidity 249
5.3 Visual Differentiation of log(z) 252
5.4 Rules of Differentiation 254
5.4.1 Composition 254
5.4.2 Inverse Functions 255
5.4.3 Addition and Multiplication 256
5.5 Polynomials, Power Series, and Rational Functions 257
5.5.1 Polynomials 257
5.5.2 Power Series 257
5.5.3 Rational Functions 259
5.6 Visual Differentiation of the Power Function 260
5.7 Visual Differentiation of exp(z) 262
′
5.8 Geometric Solution of E = E 264
5.9 An Application of Higher Derivatives: Curvature* 266
5.9.1 Introduction 266
5.9.2 Analytic Transformation of Curvature 267
5.9.3 Complex Curvature 270
5.10 Celestial Mechanics* 274
5.10.1 Central Force Fields 274
5.10.2 Two Kinds of Elliptical Orbit 274
5.10.3 Changing the First into the Second 277
5.10.4 The Geometry of Force 278
5.10.5 An Explanation 279
5.10.6 The Kasner–Arnol’d Theorem 280
xxxviii Contents

5.11 Analytic Continuation* 281

5.11.1 Introduction 281
5.11.2 Rigidity 283
5.11.3 Uniqueness 284
5.11.4 Preservation of Identities 286
5.11.5 Analytic Continuation via Reflections 286
5.12 Exercises 293

6 Non-Euclidean Geometry* 303

6.1 Introduction 303
6.1.1 The Parallel Axiom 303
6.1.2 Some Facts from Non-Euclidean Geometry 305
6.1.3 Geometry on a Curved Surface 307
6.1.4 Intrinsic versus Extrinsic Geometry 309
6.1.5 Gaussian Curvature 311
6.1.6 Surfaces of Constant Curvature 313
6.1.7 The Connection with Möbius Transformations 315
6.2 Spherical Geometry 316
6.2.1 The Angular Excess of a Spherical Triangle 316
6.2.2 Motions of the Sphere: Spatial Rotations and
Reflections 317
6.2.3 A Conformal Map of the Sphere 321
6.2.4 Spatial Rotations as Möbius Transformations 325
6.2.5 Spatial Rotations and Quaternions 329
6.3 Hyperbolic Geometry 333
6.3.1 The Tractrix and the Pseudosphere 333
6.3.2 The Constant Negative Curvature of the
Pseudosphere* 335
6.3.3 A Conformal Map of the Pseudosphere 336
6.3.4 Beltrami’s Hyperbolic Plane 339
6.3.5 Hyperbolic Lines and Reflections 342
6.3.6 The Bolyai–Lobachevsky Formula* 347
6.3.7 The Three Types of Direct Motion 348
6.3.8 Decomposing an Arbitrary Direct Motion into
Two Reflections 353
6.3.9 The Angular Excess of a Hyperbolic Triangle 357
6.3.10 The Beltrami–Poincaré Disc 359
6.3.11 Motions of the Beltrami–Poincaré Disc 363
6.3.12 The Hemisphere Model and Hyperbolic Space 367
6.4 Exercises 374
 Contents xxxix

7 Winding Numbers and Topology 385

7.1 Winding Number 385
7.1.1 The Definition 385
7.1.2 What Does “Inside” Mean? 386
7.1.3 Finding Winding Numbers Quickly 387
7.2 Hopf’s Degree Theorem 388
7.2.1 The Result 388
7.2.2 Loops as Mappings of the Circle* 390
7.2.3 The Explanation* 391
7.3 Polynomials and the Argument Principle 393
7.4 A Topological Argument Principle* 394
7.4.1 Counting Preimages Algebraically 394
7.4.2 Counting Preimages Geometrically 396
7.4.3 What’s Topologically Special About Analytic
Functions? 398
7.4.4 A Topological Argument Principle 399
7.4.5 Two Examples 401
7.5 Rouché’s Theorem 403
7.5.1 The Result 403
7.5.2 The Fundamental Theorem of Algebra 404
7.5.3 Brouwer’s Fixed Point Theorem* 404
7.6 Maxima and Minima 405
7.6.1 Maximum-Modulus Theorem 405
7.6.2 Related Results 407
7.7 The Schwarz–Pick Lemma* 407
7.7.1 Schwarz’s Lemma 407
7.7.2 Liouville’s Theorem 409
7.7.3 Pick’s Result 411
7.8 The Generalized Argument Principle 414
7.8.1 Rational Functions 414
7.8.2 Poles and Essential Singularities 416
7.8.3 The Explanation* 419
7.9 Exercises 420

8 Complex Integration: Cauchy’s Theorem 429

8.1 Introduction 429
8.2 The Real Integral 430
8.2.1 The Riemann Sum 430
xl Contents

8.2.2 The Trapezoidal Rule 432

8.2.3 Geometric Estimation of Errors 434
8.3 The Complex Integral 436
8.3.1 Complex Riemann Sums 436
8.3.2 A Visual Technique 439
8.3.3 A Useful Inequality 440
8.3.4 Rules of Integration 441
8.4 Complex Inversion 442
8.4.1 A Circular Arc 442
8.4.2 General Loops 444
8.4.3 Winding Number 446
8.5 Conjugation 447
8.5.1 Introduction 447
8.5.2 Area Interpretation 448
8.5.3 General Loops 450
8.6 Power Functions 450
8.6.1 Integration along a Circular Arc 450
8.6.2 Complex Inversion as a Limiting Case* 452
8.6.3 General Contours and the Deformation
Theorem 453
8.6.4 A Further Extension of the Theorem 454
8.6.5 Residues 455
8.7 The Exponential Mapping 457
8.8 The Fundamental Theorem 458
8.8.1 Introduction 458
8.8.2 An Example 459
8.8.3 The Fundamental Theorem 460
8.8.4 The Integral as Antiderivative 462
8.8.5 Logarithm as Integral 465
8.9 Parametric Evaluation 466
8.10 Cauchy’s Theorem 467
8.10.1 Some Preliminaries 467
8.10.2 The Explanation 469
8.11 The General Cauchy Theorem 472
8.11.1 The Result 472
8.11.2 The Explanation 473
8.11.3 A Simpler Explanation 474
 Contents xli

8.12 The General Formula of Contour Integration 475

8.13 Exercises 478

9 Cauchy’s Formula and Its Applications 485

9.1 Cauchy’s Formula 485
9.1.1 Introduction 485
9.1.2 First Explanation 486
9.1.3 Gauss’s Mean Value Theorem 487
9.1.4 A Second Explanation and the General
Cauchy Formula 488
9.2 Infinite Differentiability and Taylor Series 489
9.2.1 Infinite Differentiability 489
9.2.2 Taylor Series 491
9.3 Calculus of Residues 493
9.3.1 Laurent Series Centred at a Pole 493
9.3.2 A Formula for Calculating Residues 494
9.3.3 Application to Real Integrals 495
9.3.4 Calculating Residues using Taylor Series 497
9.3.5 Application to Summation of Series 498
9.4 Annular Laurent Series 501
9.4.1 An Example 501
9.4.2 Laurent’s Theorem 502
9.5 Exercises 506

10 Vector Fields: Physics and Topology 511

10.1 Vector Fields 511
10.1.1 Complex Functions as Vector Fields 511
10.1.2 Physical Vector Fields 513
10.1.3 Flows and Force Fields 515
10.1.4 Sources and Sinks 516
10.2 Winding Numbers and Vector Fields* 518
10.2.1 The Index of a Singular Point 518
10.2.2 The Index According to Poincaré 522
10.2.3 The Index Theorem 523
10.3 Flows on Closed Surfaces* 525
10.3.1 Formulation of the Poincaré–Hopf Theorem 525
10.3.2 Defining the Index on a Surface 527
10.3.3 An Explanation of the Poincaré–Hopf Theorem 529
10.4 Exercises 532
xlii Contents

11 Vector Fields and Complex Integration 537

11.1 Flux and Work 537
11.1.1 Flux 537
11.1.2 Work 539
11.1.3 Local Flux and Local Work 542
11.1.4 Divergence and Curl in Geometric Form* 543
11.1.5 Divergence-Free and Curl-Free Vector Fields 545
11.2 Complex Integration in Terms of Vector Fields 547
11.2.1 The Pólya Vector Field 547
11.2.2 Cauchy’s Theorem 549
11.2.3 Example: Area as Flux 550
11.2.4 Example: Winding Number as Flux 551
11.2.5 Local Behaviour of Vector Fields* 553
11.2.6 Cauchy’s Formula 555
11.2.7 Positive Powers 556
11.2.8 Negative Powers and Multipoles 557
11.2.9 Multipoles at Infinity 560
11.2.10 Laurent’s Series as a Multipole Expansion 561
11.3 The Complex Potential 562
11.3.1 Introduction 562
11.3.2 The Stream Function 563
11.3.3 The Gradient Field 565
11.3.4 The Potential Function 567
11.3.5 The Complex Potential 569
11.3.6 Examples 572
11.4 Exercises 574

12 Flows and Harmonic Functions 577

12.1 Harmonic Duals 577
12.1.1 Dual Flows 577
12.1.2 Harmonic Duals 580
12.2 Conformal Invariance 583
12.2.1 Conformal Invariance of Harmonicity 583
12.2.2 Conformal Invariance of the Laplacian 584
12.2.3 The Meaning of the Laplacian 586
12.3 A Powerful Computational Tool 587
12.4 The Complex Curvature Revisited* 590
12.4.1 Some Geometry of Harmonic Equipotentials 590
12.4.2 The Curvature of Harmonic Equipotentials 590
 Contents xliii

12.4.3 Further Complex Curvature Calculations 594

12.4.4 Further Geometry of the Complex Curvature 595
12.5 Flow Around an Obstacle 598
12.5.1 Introduction 598
12.5.2 An Example 599
12.5.3 The Method of Images 604
12.5.4 Mapping One Flow Onto Another 611
12.6 The Physics of Riemann’s Mapping Theorem 613
12.6.1 Introduction 613
12.6.2 Exterior Mappings and Flows Round
Obstacles 615
12.6.3 Interior Mappings and Dipoles 618
12.6.4 Interior Mappings, Vortices, and Sources 620
12.6.5 An Example: Automorphisms of the Disc 624
12.6.6 Green’s Function 626
12.7 Dirichlet’s Problem 630
12.7.1 Introduction 630
12.7.2 Schwarz’s Interpretation 631
12.7.3 Dirichlet’s Problem for the Disc 634
12.7.4 The Interpretations of Neumann and Bôcher 637
12.7.5 Green’s General Formula 643
12.8 Exercises 649

Bibliography 653
Index 661
CHAPTER 1

Geometry and Complex Arithmetic

1.1 Introduction

1.1.1 Historical Sketch

Half a millennium has elapsed since complex numbers were first discovered. Here,
as the reader is probably already aware, the term complex number refers to an entity
of the form a + ib, where a and b are ordinary real numbers and, unlike any ordi-
nary number, i has the property that i2 = −1. This discovery would ultimately
have a profound impact on the whole of mathematics, unifying much that had
previously seemed disparate, and explaining much that had previously seemed
inexplicable. Despite this happy ending—in reality the story continues to unfold
to this day—progress following the initial discovery of complex numbers was
painfully slow. Indeed, relative to the advances made in the nineteenth century,
little was achieved during the first 250 years of the life of the complex numbers.
How is it possible that complex numbers lay dormant through ages that saw
the coming and the passing of such great minds as Descartes, Fermat, Leibniz, and
even the visionary genius of Newton? The answer appears to lie in the fact that,
far from being embraced, complex numbers were initially greeted with suspicion,
confusion, and even hostility.
Girolamo Cardano’s Ars Magna, which appeared in 1545, is conventionally
taken to be the birth certificate of the complex numbers. Yet in that work Car-
dano introduced such numbers only to immediately dismiss them as “subtle as
they are useless”. As we shall discuss, the first substantial calculations with com-
plex numbers were carried out by Rafael Bombelli, appearing in his L’Algebra of
1572. Yet here too we find the innovator seemingly disowning his discoveries (at
least initially), saying that “the whole matter seems to rest on sophistry rather
than truth”. As late as 1702, Leibniz described i, the square root of −1, as “that
amphibian between existence and nonexistence”. Such sentiments were echoed in
the terminology of the period. To the extent that they were discussed at all, complex

Visual Complex Analysis. 25th Anniversary Edition. Tristan Needham, Oxford University Press.
© Tristan Needham (2023). DOI: 10.1093/oso/9780192868916.003.0001
2 Geometry and Complex Arithmetic

[1.1] The complex plane enables us to visualize the abstract complex number a + ib as a
simple point in the plane, with Cartesian coordinates (a, b).

numbers were called “impossible” or “imaginary”, the latter term having (unfortu-
nately) lingered to the present day1 . Even in 1770 the situation was still sufficiently
confused that it was possible for so great a mathematician as Euler to mistakenly
√ √ √
argue that −2 −3 = 6.
The root cause of all this trouble seems to have been a psychological or philo-
sophical block. How could one investigate these matters with enthusiasm or
confidence when nobody felt they knew the answer to the question, “What is a
complex number?”
A satisfactory answer to this question was only found at the end of the eigh-
teenth century2 . Independently, and in rapid succession, Wessel, Argand, and
Gauss all recognized that complex numbers could be given a simple, concrete, geo-
metric interpretation as points (or vectors) in the plane: The mystical quantity a + ib
should be viewed simply as the point in the xy-plane having Cartesian coordi-
nates (a, b), or equivalently as the vector connecting the origin to that point. See
[1.1]. When thought of in this way, the plane is denoted C and is called the complex
plane3 .
The operations of adding or multiplying two complex numbers could
now be given equally definite meanings as geometric operations on the two

1
However, an “imaginary number” now refers to a real multiple of i, rather than to a general com-
plex number. Incidentally, the term “real number” was introduced precisely to distinguish such a
number from an “imaginary number”.
2
Wallis almost hit on the answer in 1673; see Stillwell (2010, §14.4) for a detailed account of this
interesting near miss.
3
Also known as the “Gauss plane” or the “Argand plane”.
 Introduction 3

[1.2] [a] Geometrically, the sum of the complex numbers A and B is the diagonal of the
parallelogram with sides A and B. [b] Geometrically, the length of the product of A and
B is the product of their separate lengths, and its angle is the sum of their separate angles.

corresponding points (or vectors) in the plane. The rule for addition is illustrated in
[1.2a]:
The sum A + B of two complex numbers is given by the parallelogram
rule of ordinary vector addition. (1.1)
Note that this is consistent with [1.1], in the sense that 4+3i (for example) is indeed
the sum of 4 and 3i.
Figure [1.2b] illustrates the much less obvious rule for multiplication:
The length of AB is the product of the lengths of A and B, and the angle
of AB is the sum of the angles of A and B. (1.2)
This rule is not forced on us in any obvious way by [1.1], but note that it is at least
consistent with it, in the sense that 3i (for example) is indeed the product of 3 and i.
Check this for yourself. As a more exciting example, consider the product of i with
itself. Since i has unit length and angle (π/2), i2 has unit length and angle π. Thus
i2 = −1.
The publication of the geometric interpretation by Wessel and by Argand went
all but unnoticed, but the reputation of Gauss (as great then as it is now) ensured
wide dissemination and acceptance of complex numbers as points in the plane. Per-
haps less important than the details of this new interpretation (at least initially) was
the mere fact that there now existed some way of making sense of these numbers—
that they were now legitimate objects of investigation. In any event, the floodgates
of invention were about to open.
It had taken more than two and a half centuries to come to terms with complex
numbers, but the development of a beautiful new theory of how to do calculus with
such numbers (what we now call complex analysis) was astonishingly rapid. Most of
the fundamental results were obtained (by Cauchy, Riemann, and others) between
1814 and 1851—a span of less than forty years!
4 Geometry and Complex Arithmetic

Other views of the history of the subject are certainly possible. For example,
Stewart and Tall (2018) suggest that the geometric interpretation4 was somewhat
incidental to the explosive development of complex analysis. However, it should
be noted that Riemann’s ideas, in particular, would simply not have been possible
without prior knowledge of the geometry of the complex plane.

1.1.2 Bombelli’s “Wild Thought”

The power and beauty of complex analysis ultimately springs from the multipli-
cation rule (1.2) in conjunction with the addition rule (1.1). These rules were first
discovered by Bombelli in symbolic form; more than two centuries passed before
the complex plane revealed figure [1.2]. Since we merely plucked the rules out of
thin air, let us return to the sixteenth century in order to understand their algebraic
origins.
Many texts seek to introduce complex numbers with a convenient historical
fiction based on solving quadratic equations,

x2 = mx + c. (1.3)

Two thousand years BCE, it was already known that such equations could be solved
using a method that is equivalent to the modern formula,
h p i
x = 21 m ± m2 + 4c .

But what if m2 + 4c is negative? This was the very problem that led Cardano to
consider square roots of negative numbers. Thus far the textbook is being histor-
ically accurate, but next we read that the need for (1.3) to always have a solution
forces us to take complex numbers seriously. This argument carries almost as little
weight now as it did in the sixteenth century. Indeed, we have already pointed out
that Cardano did not hesitate to discard such “solutions” as useless.
It was not that Cardano lacked the imagination to pursue the matter further,
rather he had a fairly compelling reason not to. For the ancient Greeks mathe-
matics was synonymous with geometry, and this conception still held sway in
the sixteenth century. Thus an algebraic relation such as (1.3) was not so much
thought of as a problem in its own right, but rather as a mere vehicle for solving a
genuine problem of geometry. For example, (1.3) may be considered to represent
the problem of finding the intersection points of the parabola y = x2 and the line
y = mx + c. See [1.3a].
In the case of L1 the problem has a solution; algebraically, (m2 + 4c) > 0 and the
two intersection points are given by the formula above. In the case of L2 the problem
clearly does not have a solution; algebraically, (m2 + 4c) < 0 and the absence of

4
We must protest one piece of their evidence: Although Wallis did attempt a geometric interpreta-
tion of complex numbers in 1673, he did not hit upon the “correct” picture; see footnote 2.
 Introduction 5

[1.3] [a] While some lines intersect a parabola, some do not, so the quadratic equation
x2 = mx + c may or may not have solutions. [b] But x3 = mx + c must always have a
solution.

solutions is correctly manifested by the occurrence of “impossible” numbers in the

formula.
It was not the quadratic that forced complex numbers to be taken seriously, it
was the cubic,
x3 = 3px + 2q.

[Ex. 1 shows that a general cubic can always be reduced to this form.] This equation
represents the problem of finding the intersection points of the cubic curve y = x3
and the line y = 3px+2q. See [1.3b]. Building on the work of del Ferro and Tartaglia,
Cardano’s Ars Magna showed that this equation could be solved by means of a
remarkable formula [see Ex. 2]:
q p q p
3 3
x = q + q2 − p3 + q − q2 − p3 . (1.4)

Try it yourself on x3 = 6x + 6.
Some thirty years after this formula appeared, Bombelli recognized that there
was something strange and paradoxical about it. First note that if the line
y = 3px + 2q is such that p3 > q2 then the formula involves complex numbers.
For example, Bombelli considered x3 = 15x + 4, which yields
√
3
√3
x = 2 + 11i + 2 − 11i.

In the previous case of [1.3a] this merely signalled that the geometric problem had
no solution, but in [1.3b] it is clear that the line will always hit the curve! In fact
inspection of Bombelli’s example yields the solution x = 4.
As he struggled to resolve this paradox, Bombelli had what he called a “wild
thought”: perhaps the solution x = 4 could be recovered from the above expression
√ √
if 3 2 + 11i = 2 + ni and 3 2 − 11i = 2 − ni. Of course for this to work he would
6 Geometry and Complex Arithmetic

have to assume that the addition of two complex numbers A = a+ie e

a and B = b+i b
obeyed the plausible rule,

A + B = (a + i a e = (a + b) + i (e
e) + (b + i b) e
a + b). (1.5)
√
Next, to see if there was indeed a value of n for which 3 2 + 11i = 2+in, he needed
to calculate (2 + in)3 . To do so he assumed that he could multiply out brackets as
in ordinary algebra, so that
e = ab + i (a b
e) (b + i b)
(a + i a e+a
e b) + i2 a e
e b.

Using i2 = −1, he concluded that the product of two complex numbers would be
given by

AB = (a + i a e = (ab − a
e) (b + i b) e + i (a b
e b) e+a
e b). (1.6)

This rule vindicated his “wild thought”, for he was now able to show that (2±i)3 =
2 ± 11i. Check this for yourself.
While complex numbers themselves remained mysterious, Bombelli’s work on
cubic equations thus established that perfectly real problems required complex
arithmetic for their solution.
Just as with its birth, the subsequent development of the theory of complex num-
bers was inextricably bound up with progress in other areas of mathematics (and
also physics). Sadly, we can only touch on these matters in this book; for a full and
fascinating account of these interconnections, the reader is instead referred to Still-
well (2010). Repeating what was said in the Preface, we cannot overstate the value
of reading Stillwell’s book alongside this one.

1.1.3 Some Terminology and Notation

Leaving history behind us, we now introduce the modern terminology and nota-
tion used to describe complex numbers. The information is summarized in the table
below, and is illustrated in [1.4].

Name Meaning Notation

modulus of z length r of z |z|
argument of z angle θ of z arg (z)
real part of z x coordinate of z Re(z)
imaginary part of z y coordinate of z Im(z)
imaginary number real multiple of i
real axis set of real numbers
imaginary axis set of imaginary numbers
complex conjugate of z reflection of z in the real axis z
 Introduction 7

[1.4] Visual summary of the terminology and notation used to describe complex numbers.

It is valuable to grasp from the outset that (according to the geometric view) a
complex number is a single, indivisible entity—a point in the plane. Only when
we choose to describe such a point with numerical coordinates does a complex
number appear to be compound or “complex”. More precisely, C is said to be two
dimensional, meaning that two real numbers (coordinates) are needed to label a point
within it, but exactly how the labelling is done is entirely up to us.
One way to label the points is with Cartesian coordinates (the real part x and
the imaginary part y), the complex number being written as z = x + iy. This is the
natural labelling when we are dealing with the addition of two complex numbers,
because (1.5) says that the real and imaginary parts of A+B are obtained by adding
the real and imaginary parts of A and B.
In the case of multiplication, the Cartesian labelling no longer appears natural,
for it leads to the messy and unenlightening rule (1.6). The much simpler geometric
rule (1.2) makes it clear that we should instead label a typical point z with its polar
coordinates, r = |z| and θ = arg z. In place of z = x+iy we may now write z = r∠θ,
where the symbol ∠ serves to remind us that θ is the angle of z. [Although this
notation is still used by some, we shall only employ it briefly; later in this chapter
we will discover a much better notation (the standard one) which will then be used
throughout the remainder of the book.] The geometric multiplication rule (1.2) now
takes the simple form,

(R∠ϕ) (r∠θ) = (Rr)∠(ϕ + θ). (1.7)

In common with the Cartesian label x + iy, a given polar label r∠θ specifies a
unique point, but (unlike the Cartesian case) a given point does not have a unique
8 Geometry and Complex Arithmetic

polar label. Since any two angles that differ by a multiple of 2π correspond to the
same direction, a given point has infinitely many different labels:

. . . = r∠(θ − 4π) = r∠(θ − 2π) = r∠θ = r∠(θ + 2π) = r∠(θ + 4π) = . . .

This simple fact about angles will become increasingly important as our subject
unfolds.
The Cartesian and polar coordinates are the most common ways of labelling
complex numbers, but they are not the only ways. In Chapter 3 we will meet
another particularly useful method, called “stereographic” coordinates.

1.1.4 Practice
Before continuing, we strongly suggest that you make yourself comfortable with
the concepts, terminology, and notation introduced thus far. To do so, try to
convince yourself geometrically (and/or algebraically) of each of the following facts:
p
Re(z) = 12 [z + z] Im(z) = 2i1
[z − z] |z| = x2 + y2
Im(z)
tan[arg z] = Re(z) z z = |z|2 r∠θ = r (cos θ + i sin θ)

Defining 1
z by (1/z) z = 1, it follows that 1
z = 1
r∠θ = r1 ∠(−θ).

R∠ϕ R 1 x y
= ∠(ϕ − θ) = 2 −i 2
r∠θ r (x + iy) x + y2 x + y2
√
(1 + i)4 = −4 (1 + i)13 = −26 (1 + i) (1 + i 3)6 = 26
√
(1 + i 3)3 (1 + i)5 √
= −4i √ = − 2 ∠ − (π/12) r∠θ = r∠(−θ)
(1 − i)2 ( 3 + i) 2

z1 + z 2 = z1 + z 2 z1 z 2 = z1 z2 z1 /z2 = z1 /z2 .
Lastly, establish the so-called generalized triangle inequality:

|z1 + z2 + · · · + zn | ⩽ |z1 | + |z2 | + · · · + |zn |. (1.8)

When does equality hold?

1.1.5 Equivalence of Symbolic and Geometric Arithmetic

We have been using the symbolic rules (1.5) and (1.6) interchangeably with the
geometric rules (1.1) and (1.2), and we now justify this by showing that they are
indeed equivalent. The equivalence of the addition rules (1.1) and (1.5) will be
familiar to those who have studied vectors; in any event, the verification is suf-
ficiently straightforward that we may safely leave it to the reader. We therefore
only address the equivalence of the multiplication rules (1.2) and (1.6).
 Introduction 9

[1.5] A complex number can be interpreted as a rotation and expansion of the plane.
Geometrically, multiplication by a complex number A = R∠ϕ rotates the plane through
angle ϕ, and expands it by R, so, in particular, parallelograms are transformed into similar
parallelograms. It follows that A(B + C) = AB + AC.

First we will show how the symbolic rule may be derived from the geometric
rule. To do so we shall rephrase the geometric rule (1.7) in a particularly useful
and important way. Let z denote a general point in C, and consider what hap-
pens to it—where it moves to—when it is multiplied by a fixed complex number
A = R∠ϕ. According to (1.7), the length of z is magnified by R, while the angle of z is
increased by ϕ. Now imagine that this is done simultaneously to every point of the
plane:

Geometrically, multiplication by a complex number A = R∠ϕ is a rotation

(1.9)
of the plane through angle ϕ, and an expansion of the plane by factor R.

A few comments are in order:

• Both the rotation and the expansion are centred at the origin.
• It makes no difference whether we do the rotation followed by the expansion,
or the expansion followed by the rotation.
• If R < 1 then the “expansion” is in reality a contraction.

Figure [1.5] illustrates the effect of such a transformation, the lightly shaded
shapes being transformed into the darkly shaded shapes. Check for yourself that
√
in this example A = 1 + i 3 = 2 ∠ π3 .
It is now a simple matter to deduce the symbolic rule from the geometric
rule. Recall the essential steps taken by Bombelli in deriving (1.6): (i) i2 = −1;
(ii) brackets can be multiplied out, i.e., if A, B, C, are complex numbers then
10 Geometry and Complex Arithmetic

[1.6] Geometric
explanation of the geometry of multiplication. [a] iz =

z rotated by π2 . [b] A = 4 + 3i = 5∠ϕ. [c] Since this shaded triangle is similar
to the one in [b], we see that multiplication by 5∠ϕ does indeed rotate the plane by ϕ,
and expand it by 5.

A(B + C) = AB + AC. We have already seen that the geometric rule gives us (i),
and figure [1.5] now reveals that (ii) is also true, for the simple reason that rotations
and expansions preserve parallelograms. By the geometric definition of addition, B + C
is the fourth vertex of the parallelogram with vertices 0, B, C. To establish (ii), we
merely observe that multiplication by A rotates and expands this parallelogram
into another parallelogram with vertices 0, AB, AC and A(B + C). This completes
the derivation of (1.6).
Conversely, we now show how the geometric rule may be derived from the sym-
bolic rule5 . We begin by considering the transformation z ÞÑ iz. According to the
symbolic rule, this means that (x + iy) ÞÑ (−y + ix), and [1.6a] reveals that iz is
z rotated through a right angle. We now use this fact to interpret the transformation
z ÞÑ A z, where A is a general complex number. How this is done may be grasped
sufficiently well using the example A = 4 + 3i = 5∠ϕ, where ϕ = tan−1 (3/4).

5
In every text we have examined this is done using trigonometric identities. We believe that the
present argument supports the view that such identities are merely complicated manifestations of the
simple rule for complex multiplication.
 Euler’s Formula 11

See [1.6b]. The symbolic rule says that brackets can be multiplied out, so our
transformation may be rewritten as follows:

z ÞÑ A z = (4 + 3i)z
= 4z + 3(iz)
π


= 4z + 3 z rotated by 2 .

This is visualized in [1.6c]. We can now see that the shaded triangles in [1.6c] and
[1.6b] are similar, so multiplication by 5∠ϕ does indeed rotate the plane by ϕ, and
expand it by 5. Done.

1.2 Euler’s Formula

1.2.1 Introduction
It is time to replace the r∠θ notation with a much better one that depends on the
following miraculous fact:

eiθ = cos θ + i sin θ ! (1.10)

This result was discovered by Leonhard Euler around 1740, and it is called Euler’s
formula in his honour.
Before attempting to explain this result, let us say something of its meaning and
utility. As illustrated in [1.7a], the formula says that eiθ is the point on the unit
circle at angle θ. Instead of writing a general complex number as z = r∠θ, we can
now write z = r eiθ . Concretely, this says that to reach z we must take the unit

[1.7] [a] Euler’s formula states that eiθ has unit length and points at angle θ; expanding
this by r, we may reach any complex number. [b] The complex velocity V(t) (tangent to
the complex orbit Z(t)) is ultimately equal to M/δ, as δ Ñ 0.
12 Geometry and Complex Arithmetic

vector eiθ that points at z, then stretch it by the length of z. Part of the beauty of
this representation is that the geometric rule (1.7) for multiplying complex numbers
now looks almost obvious:



R eiϕ r eiθ = Rr ei(ϕ+θ) .

Put differently, algebraically manipulating eiθ in the same way as the real function
ex yields true facts about complex numbers.
In order to explain Euler’s formula we must first address the more basic ques-
tion, “What does eiθ mean?” Surprisingly, many authors answer this by defining
eiθ , out of the blue, to be (cos θ + i sin θ)! This gambit is logically unimpeachable,
but it is also a low blow to Euler, reducing one of his greatest achievements to a
mere tautology. We will therefore give two heuristic arguments in support of (1.10);
deeper arguments will emerge in later chapters.

1.2.2 Moving Particle Argument

d x
Recall the basic fact that ex is its own derivative: dx e = ex . This is actually a
d
defining property, that is, if dx f(x) = f(x), and f(0) = 1, then f(x) = ex . Similarly,
kx d
if k is a real constant, then e may be defined by the property dx f(x) = k f(x). To
x
extend the action of the ordinary exponential function e from real values of x to
imaginary ones, let us cling to this property by insisting that it remain true if k = i,
so that
d it
e = ieit . (1.11)
dt
We have used the letter t instead of x because we will now think of the variable
as being time. We are used to thinking of the derivative of a real function as the
slope of the tangent to the graph of the function, but how are we to understand the
derivative in the above equation?
To make sense of this, imagine a particle moving along a curve in C. See [1.7b].
The motion of the particle can be described parametrically by saying that at time t its
position is the complex number Z(t). Next, recall from physics that the velocity V(t)
is the vector—now thought of as a complex number—whose length and direction
are given by the instantaneous speed, and the instantaneous direction of motion
(tangent to the trajectory), of the moving particle. The figure shows the movement
M of the particle between time t and t + δ, and this should make it clear that

d Z(t + δ) − Z(t) M
Z(t) = lim = lim = V(t).
dt δÑ0 δ δÑ0 δ

Thus, given a complex function Z(t) of a real variable t, we can always visualize Z
as the position of a moving particle, and dZ
dt as its velocity.
 Euler’s Formula 13

[1.8] First explanation of Euler’s formula. The orbit Z(t) = eit is characterized by the
fact that its velocity V = iZ = (its position rotated by π2 ). Thus, the particle must always
move at right angles to its current position, and at unit speed. Therefore, after time t = θ
it will have travelled distance θ round the unit circle, subtending angle θ at the origin, so
eiθ = cos θ + i sin θ.

We can now use this idea to find the trajectory in the case Z(t) = eit . See [1.8].
According to (1.11),
velocity = V = iZ = position, rotated through a right angle.
Since the initial position of the particle is Z(0) = e0 = 1, its initial velocity is i,
and so it is moving vertically upwards. A split second later the particle will have
moved very slightly in this direction, and its new velocity will be at right angles to
its new position vector. Continuing to construct the motion in this way, it is clear
that the particle will travel round the unit circle.
Since we now know that |Z(t)| remains equal to 1 throughout the motion, it fol-
lows that the particle’s speed |V(t)| also remains equal to 1. Thus after time t = θ
the particle will have travelled a distance θ round the unit circle, and so the angle
of Z(θ) = eiθ will be θ. This is the geometric statement of Euler’s formula.

1.2.3 Power Series Argument

For our second argument, we begin by re-expressing the defining property
d
dx f(x) = f(x) in terms of power series. Assuming that f(x) can be expressed in
the form a0 + a1 x + a2 x2 + · · · , a simple calculation shows that
x2 x3
ex = f(x) = 1 + x + + + ··· ,
2! 3!
and further investigation shows that this series converges for all (real) values of x.
14 Geometry and Complex Arithmetic

[1.9] Second explanation of Euler’s formula. Imagine the terms of the horizontal power
series for eθ to be rods connected by hinges. Now wrap the series into the illustrated
spiral, with each rod making a right angle with the one before it. The argument in the text
confirms Euler’s miraculous conclusion: the spiral ends at the point on the unit circle at
angle θ. So, imagining the rigid, vertical iθ-rod transformed into flexible string, we may
wrap it onto the unit circle, and its end will coincide with the end of the spiral!

Putting x equal to a real value θ, this infinite sum of horizontal real numbers is
visualized in [1.9]. To make sense of eiθ , we now cling to the power series and put
x = iθ:
(iθ)2 (iθ)3
eiθ = 1 + iθ + + + ··· .
2! 3!
As illustrated in [1.9], this series is just as meaningful as the series for eθ , but
instead of the terms all having the same direction, here each term makes a right
angle with the previous one, producing a kind of spiral.
This picture makes it clear that the known convergence of the series for eθ guar-
antees that the spiral series for eiθ converges to a definite point in C. However, it
is certainly not clear that it will converge to the point on the unit circle at angle θ.
To see this, we split the spiral into its real and imaginary parts:

eiθ = C(θ) + iS(θ),

where
θ2 θ4 θ3 θ5
C(θ) = 1 − + − ··· , and S(θ) = θ − + − ··· .
2! 4! 3! 5!
At this point we could obtain Euler’s formula by appealing to Taylor’s Theorem,
which shows that C(θ) and S(θ) are the power series for cos θ and sin θ. However,
we can also get the result by means of the following elementary argument that does
not require Taylor’s Theorem.
 Euler’s Formula 15

We wish to show two things about eiθ = C(θ) + iS(θ): (i) it has unit length, and
(ii) it has angle θ. To do this, first note that differentiation of the power series C and
S yields
C ′ = −S and S ′ = C,
where a prime denotes differentiation with respect to θ.
To establish (i), observe that
d iθ 2
|e | = (C2 + S2 ) ′ = 2(CC ′ + SS ′ ) = 0,
dθ
which means that the length of eiθ is independent of θ. Since ei0 = 1, we deduce
that |eiθ | = 1 for all θ.
To establish (ii) we must show that Θ(θ) = θ, where Θ(θ) denotes the angle of
eiθ , so that
S(θ)
tan Θ(θ) = .
C(θ)
Since we already know that C2 + S2 = 1, we find that the derivative of the LHS of
the above equation is
 
′ S2 Θ′
[tan Θ(θ)] = (1 + tan Θ) Θ = 1 + 2 Θ ′ = 2 ,
2 ′
C C
and that the derivative of the RHS is
 ′
S S ′C − C ′S 1
= 2
= 2.
C C C
Thus
dΘ
= Θ ′ = 1,
dθ
which implies that Θ(θ) = θ + const. Taking the angle of ei0 = 1 to be 0 [would it
make any geometric difference if we took it to be 2π?], we find that Θ = θ.
Although it is incidental to our purpose, note that we can now conclude (with-
out Taylor’s Theorem) that C(θ) and S(θ) are the power series of cos θ and
sin θ.

1.2.4 Sine and Cosine in Terms of Euler’s Formula

A simple but important consequence of Euler’s formula is that sine and cosine can
be constructed from the exponential function. More precisely, inspection of [1.10]
yields
eiθ + e−iθ = 2 cos θ and eiθ − e−iθ = 2i sin θ,
or equivalently,
eiθ + e−iθ eiθ − e−iθ
cos θ = and sin θ = . (1.12)
2 2i
16 Geometry and Complex Arithmetic

[1.10] eiθ + e−iθ = 2 cos θ and eiθ − e−iθ = 2i sin θ.

1.3 Some Applications

1.3.1 Introduction
Often problems that do not appear to involve complex numbers are nevertheless
solved most elegantly by viewing them through complex spectacles. In this section
we will illustrate this point with a variety of examples taken from diverse areas
of mathematics. Further examples may be found in the exercises at the end of the
chapter.
The first example [trigonometry] merely illustrates the power of the concepts
already developed, but the remaining examples develop important new ideas.

1.3.2 Trigonometry
All trigonometric identities may be viewed as arising from the rule for complex
multiplication. In the following examples we will reduce clutter by using the
following shorthand: C ≡ cos θ, S ≡ sin θ, and similarly, c ≡ cos ϕ, s ≡ sin ϕ.
To find an identity for cos(θ + ϕ), view it as a component of ei(θ+ϕ) . See [1.11a].
Since
cos(θ + ϕ) + i sin(θ + ϕ) = ei(θ+ϕ)
= eiθ eiϕ
= (C + iS)(c + is)
= [Cc − Ss] + i[Sc + Cs],
we obtain not only an identity for cos(θ + ϕ), but also one for sin(θ + ϕ):
cos(θ + ϕ) = Cc − Ss and sin(θ + ϕ) = Sc + Cs.
This illustrates another powerful feature of using complex numbers: every complex
equation says two things at once.
 Some Applications 17

[1.11] [a] Trigonometric identities for cos(θ + ϕ) and sin(θ + ϕ) are easily obtained
by taking the real and imaginary parts of ei(θ+ϕ) = eiθ eiϕ . [b] An identity for tan 3θ in
terms of T = tan θ follows easily from the illustrated fact that (1 + iT )3 has angle 3θ.

To simultaneously find identities for cos 3θ and sin 3θ, consider ei3θ :
   
cos 3θ + i sin 3θ = ei3θ = (eiθ )3 = (C + iS)3 = C3 − 3CS2 + i 3C2 S − S3 .

Using C2 + S2 = 1, these identities may be rewritten in the more familiar forms,

cos 3θ = 4C3 − 3C and sin 3θ = −4S3 + 3S.

We have just seen how to express trig functions of multiples of θ in terms of

powers of trig functions of θ, but we can also go in the opposite direction. For
example, suppose we want an identity for cos4 θ in terms of multiples of θ. Since
2 cos θ = eiθ + e−iθ ,

4
24 cos4 θ = eiθ + e−iθ



= ei4θ + e−i4θ + 4 ei2θ + e−i2θ + 6
= 2 cos 4θ + 8 cos 2θ + 6
=⇒ cos4 θ = 1
8
[cos 4θ + 4 cos 2θ + 3] .

Although Euler’s formula is extremely convenient for doing such calculations,

it is not essential: all we are really using is the equivalence of the geometric and
symbolic forms of complex multiplication. To stress this point, let us do an example
without Euler’s formula.
To find an identity for tan 3θ in terms of T = tan θ, consider z = 1 + iT . See
[1.11b]. Since z is at angle θ, z3 will be at angle 3θ, so tan 3θ = Im(z3 )/Re(z3 ). Thus,
3T − T 3
z3 = (1 + iT )3 = (1 − 3T 2 ) + i(3T − T 3 ) =⇒ tan 3θ = .
1 − 3T 2
18 Geometry and Complex Arithmetic

[1.12] A surprising property of quadrilaterals. [a] If squares are constructed on the sides
of a quadrilateral, the line-segments joining the centres of opposite squares are perpen-
dicular and of equal length! A short algebraic proof is made possible by taking the edges
of the quadrilateral to be complex numbers that sum to zero. [b] An alternative geometric
proof is based on the illustrated fact: If squares are constructed on two sides of an arbitrary
triangle, then the line-segments from their centres to the midpoint m of the remaining side
are perpendicular and of equal length.

1.3.3 Geometry
We shall base our discussion of geometric applications on a single example. In
[1.12a] we have constructed squares on the sides of an arbitrary quadrilateral. Let
us prove what this picture strongly suggests: the line-segments joining the centres of
opposite squares are perpendicular and of equal length. It would require a great deal
of ingenuity to find a purely geometric proof of this surprising result, so instead
of relying on our own intelligence, let us invoke the intelligence of the complex
numbers!
Introducing a factor of 2 for convenience, let 2a, 2b, 2c, and 2d represent complex
numbers running along the edges of the quadrilateral. The only condition is that
the quadrilateral close up, i.e.,

a + b + c + d = 0.

As illustrated, choose the origin of C to be at the vertex where 2a begins. To reach

the centre p of the square constructed on that side, we go along a, then an equal
distance at right angles to a. Thus, since ia is a rotated through a right angle, p =
a + ia = (1 + i)a. Likewise,

q = 2a + (1 + i)b, r = 2a + 2b + (1 + i)c, s = 2a + 2b + 2c + (1 + i)d.

 Some Applications 19

The complex numbers A = s − q (from q to s) and B = r − p (from p to r) are

therefore given by

A = (b + 2c + d) + i(d − b) and B = (a + 2b + c) + i(c − a).

We wish to show that A and B are perpendicular and of equal length. These two
statements can be combined into the single complex statement B = iA, which says
that B is A rotated by (π/2). To finish the proof, note that this is the same thing as
A + iB = 0, the verification of which is a routine calculation:

A + iB = (a + b + c + d) + i (a + b + c + d) = 0.

As a first step towards a purely geometric explanation of the result in [1.12a],

consider [1.12b]. Here squares have been constructed on two sides of an arbitrary
triangle, and, as the picture suggests, the line-segments from their centres to the mid-
point m of the remaining side are perpendicular and of equal length. As is shown in
Ex. 21, [1.12a] can be quickly deduced6 from [1.12b]. The latter result can, of course,
be proved in the same manner as above, but let us instead try to find a purely
geometric argument.
To do so we will take an interesting detour, investigating translations and rota-
tions of the plane in terms of complex functions. In reality, this “detour” is much
more important than the geometric puzzle to which our results will be applied.
Let Tv denote a translation of the plane by v, so that a general point z is mapped
to Tv (z) = z + v. See [1.13a], which also illustrates the effect of the translation on a
triangle. The inverse of Tv , written Tv−1 , is the transformation that undoes it; more
formally, Tv−1 is defined by Tv−1 ◦ Tv = E = Tv ◦ Tv−1 , where E is the “do nothing”
transformation (called the identity) that maps each point to itself: E(z) = z. Clearly,
Tv−1 = T−v .
If we perform Tv , followed by another translation Tw , then the composite
mapping Tw ◦ Tv of the plane is another translation:

Tw ◦ Tv (z) = Tw (z + v) = z + (w + v) = Tw+v (z).

This gives us an interesting way of motivating addition itself. If we had intro-

duced a complex number v as being the translation Tv , then we could have defined
the “sum” of two complex numbers Tv and Tw to be the net effect of performing
these translations in succession (in either order). Of course this would have been
equivalent to the definition of addition that we actually gave.
Let Rθa denote a rotation of the plane through angle θ about the point a. For

−1
example, Rϕ a ◦ Ra = Ra
θ θ+ϕ
, and Rθa = R−θ
a . As a first step towards expressing
rotations as complex functions, note that (1.9) says that a rotation about the origin
can be written as Rθ0 (z) = eiθ z.

6
This approach is based on a paper of Finney (1970).
20 Geometry and Complex Arithmetic

[1.13] Translations and rotations as complex functions. [a] Composition of translations

is equivalent to addition: Tw ◦ Tv (z) = Tw+v (z). [b] The general rotation Rθa can be
performed by translating a to 0, rotating by θ about 0, then translating 0 back to a.

As illustrated in [1.13b], the general rotation Rθa can be performed by translating

a to 0, rotating θ about 0, then translating 0 back to a:


Rθa (z) = Ta ◦ Rθ0 ◦ Ta−1 (z) = eiθ (z − a) + a = eiθ z + k,
where k = a(1 − eiθ ). Thus we find that a rotation about any point can instead be
expressed as an equal rotation about the origin, followed by a translation: Rθa =


Tk ◦ Rθ0 . Conversely, a rotation of α about the origin followed by a translation of
v can always be reduced to a single rotation:
Tv ◦ Rα
0 = Rc ,
α
where c = v/(1 − eiα ).
In the same way, you can easily check that if we perform the translation before the
rotation, the net transformation can again be accomplished with a single rotation:
Rθ0 ◦ Tv = Rθp . What is p?
The results just obtained are certainly not obvious geometrically [try them], and
they serve to illustrate the power of thinking of translations and rotations as com-
plex functions. As a further illustration, consider the net effect of performing two
rotations about different points. Representing the rotations as complex functions,
an easy calculation [exercise] yields
 
Rϕb ◦ R θ
a (z) = ei(θ+ϕ) z + v, where v = aeiϕ (1 − eiθ ) + b(1 − eiϕ ).

Unless (θ + ϕ) is a multiple of 2π, the previous paragraph therefore tells us that

v aeiϕ (1 − eiθ ) + b(1 − eiϕ )
Rϕ
b ◦ Ra = Rc
θ (θ+ϕ)
, where c= = .
1 − ei(θ+ϕ) 1 − ei(θ+ϕ)
[What should c equal if b = a or ϕ = 0? Check the formula.] This result is illustrated
in [1.14a]. Later we shall find a purely geometric explanation of this result, and,
in the process, a very simple geometric construction of the point c given by the
complicated formula above.
 Some Applications 21

[1.14] Composition of rotations. [a] In general, two successive rotations (of θ and ϕ)
about different points (a and b) are equivalent to a single rotation of (θ + ϕ) about a
special point c. [b] But if (θ + ϕ) = 2π then the net result is instead a translation, as can
easily be seen here in the special case θ = ϕ = π.

If, on the other hand, (θ + ϕ) is a multiple of 2π, then ei(θ+ϕ) = 1, and

Rϕ
b ◦ Ra = Tv ,
θ
where v = (1 − eiϕ )(b − a).

For example, putting θ = ϕ = π, this predicts that Rπ

b ◦Ra = T2(b−a) is a translation
π

by twice the complex number connecting the first centre of rotation to the second.
That this is indeed true can be deduced directly from [1.14b].
The above result on the composition of two rotations implies [exercise] the
following:

Let M = Rθann ◦ · · · ◦ Rθa22 ◦ Rθa11 be the composition of n rotations, and let

Θ = θ1 + θ2 + · · · + θn be the total amount of rotation. In general,
M = RΘc (for some c), but if Θ is a multiple of 2π then M = Tv , for some v.

Returning to our original problem, we can now give an elegant geometric expla-
(π/2) (π/2)
nation of the result in [1.12b]. Referring to [1.15a], let M = Rπ m ◦ Rp ◦ Rs .
According to the result just obtained, M is a translation. To find out what transla-
tion, we need only discover the effect of M on a single point. Clearly, M(k) = k, so
M is the zero translation, i.e., the identity transformation E. Thus
−1
R(π/2)
p ◦ Rs(π/2) = (Rπ
m) ◦ M = Rπ
m.

If we define s ′ = Rπ ′
m (s) then m is the midpoint of ss . But, on the other hand,
 
s ′ = Rp(π/2)
◦ Rs(π/2) (s) = R(π/2)
p (s).

Thus the triangle sps ′ is isosceles and has a right angle at p, so sm and pm are
perpendicular and of equal length. Done.
22 Geometry and Complex Arithmetic

(π/2) (π/2)
[1.15] [a] The total angle of rotation of M = Rπ m ◦Rp ◦Rs is 2π, so it is a translation.
But we see that M(k) = k, so the translation vanishes—M is the identity. It follows [see
text] that sm and pm are perpendicular and of equal length. [b] The orbit Z(t) = eat eibt
is a spiral: as eibt rotates round the unit circle with angular speed b, it is stretched by the
rapidly increasing factor eat .

1.3.4 Calculus
For our calculus example, consider the problem of finding the 100th derivative of
ex sin x. More generally, we will show how complex numbers may be used to find
the nth derivative of eax sin bx.
In discussing Euler’s formula we saw that eit may be thought of as the location
at time t of a particle travelling around the unit circle at unit speed. In the same
way, eibt may be thought of as a unit complex number rotating about the origin
with (angular) speed b. If we stretch this unit complex number by eat as it turns,
then its tip describes the motion of a particle that is spiralling away from the origin.
See [1.15b].
The relevance of this to the opening problem is that the location of the particle
at time t is
Z(t) = eat eibt = eat cos bt + i eat sin bt.
Thus the derivative of eat sin bt is simply the vertical (imaginary) component of
the velocity V of Z.
We could find V simply by differentiating the components of Z in the above
expression, but we shall instead use this example to introduce the geometric
approach that will be used throughout this book. In [1.16], consider the movement
M = Z(t + δ) − Z(t) of the particle between time t and (t + δ).
Recall that V is defined to be the limit of (M/δ) as δ tends to zero. Thus V and
(M/δ) are very nearly equal if δ is very small. This suggests two intuitive ways of
 Some Applications 23

[1.16] Geometric evaluation of the velocity: Vδ ≍ M ≍ A + B ≍ eibt |Z|aδ + ieibt |Z|bδ,

so V = (a + ib)Z.

speaking, both of which will be used in this book: (i) we shall say that “V = (M/δ)
when δ is infinitesimal” or (ii) that “V and (M/δ) are ultimately equal” (as δ tends to
zero).
As explained in the new Preface to this 25th Anniversary Edition, both the concept
and the language of ultimate equality were introduced by Newton in his great Prin-
cipia of 1687, which is filled with inspired geometrical constructions, not tedious
calculus computations. Sadly, although it was the mathematical engine that pro-
pelled his extraordinary discoveries, Newton chose to emulate the expository style
of the ancient Greek geometers, writing out his proofs in words: he did not intro-
duce any symbol to represent his critical concept of ultimate equality. We, however,
shall employ the symbol ” ≍ ” to denote Newton’s concept, enabling us to write,
for example, V ≍ (M/δ).
We stress that here the words “ultimately equal” and “infinitesimal” are being
used in definite, technical senses; in particular, “infinitesimal” does not refer to
some mystical, infinitely small quantity.7 More precisely, following Newton’s lead

7
For more on this distinction, see the discussion in Chandrasekhar (1995).
24 Geometry and Complex Arithmetic

in the Principia, if two quantities X and Y depend on a third quantity δ, then we

shall employ the following language and notation:
X
lim = 1 ⇐⇒ “X = Y for infinitesimal δ”.
δÑ0 Y
⇐⇒ “X and Y are ultimately equal (as δ tends to zero)”.
⇐⇒ X ≍ Y.
It follows from the basic theorems on limits that Newton’s concept of “ultimate
equality” is [exercise] an equivalence relation, and that it also inherits other proper-
ties of ordinary equality. For example, since V and (M/δ) are ultimately equal, so
are Vδ and M:
M
V ≍ ⇐⇒ Vδ ≍ M.
δ
We now return to the problem of finding the velocity of the spiralling particle.
As illustrated in [1.16], draw rays from 0 through Z(t) and Z(t + δ), together with
circular arcs (centred at 0) through those points. Now let A and B be the complex
numbers connecting Z(t) to the illustrated intersection points of these rays and
arcs. If δ is infinitesimal, then B is at right angles to A and Z, and M = A + B.
Let us find the ultimate lengths of A and B. During the time interval δ, the angle
of Z increases by bδ, so the two rays cut off an arc of length bδ on the unit circle,
and an arc of length |Z|bδ on the circle through Z. Thus |B| is ultimately equal to
|Z|bδ. Next, note that |A| is the increase in |Z(t)| occurring in the time interval δ.
Thus, since
d d at
|Z(t)| = e = a|Z|,
dt dt
|A| is ultimately equal to |Z|aδ.
The shaded triangle at Z is therefore ultimately similar to the shaded right tri-
angle with hypotenuse a + ib. Rotating the latter triangle by the angle of Z, you
should now be able to see that if δ is infinitesimal then
M = (a + ib) rotated by the angle of Z, and expanded by |Z| δ
= (a + ib)Zδ
d
=⇒ V = Z = (a + ib)Z. (1.13)
dt
Thus all rays from the origin cut the spiral at the same angle [the angle of (a + ib)],
and the speed of the particle is proportional to its distance from the origin.
Note that although we have not yet given meaning to ez (where z is a general
complex number), it is certainly tempting to write Z(t) = eat eibt = e(a+ib)t . This
makes the result (1.13) look very natural. Conversely, this suggests that we should
define ez = e(x+iy) to be ex eiy ; another justification for this step will emerge in the
next chapter.
 Some Applications 25

Using (1.13), it is now easy to take further derivatives. For example, the acceler-
ation of the particle is
d2 d
2
Z= V = (a + ib)2 Z = (a + ib) V.
dt dt
Continuing in this way, each new derivative is obtained by multiplying the previ-
ous one by (a + ib). [Try sketching these successive derivatives in [1.16].] Writing
√
(a + ib) = R eiϕ , where R = a2 + b2 and ϕ is the appropriate value of tan−1 (b/a),
we therefore find that
dn
Z = (a + ib)n Z = Rn einϕ eat eibt = Rn eat ei(bt+nϕ) .
dtn
Thus
dn  at  n  
n
e sin bt = (a2 + b2 ) 2 eat sin bt + n tan−1 (b/a) . (1.14)
dt

1.3.5 Algebra
In the final year of his life (1716) Roger Cotes made a remarkable discovery that
enabled him (in principle) to evaluate the family of integrals,
Z
dx
n
,
x −1
where n = 1, 2, 3, . . .. To see the connection with algebra, consider the case n = 2.
The key observations are that the denominator (x2 − 1) can be factorized into
(x − 1)(x + 1), and that the integrand can then be split into partial fractions:
Z Z   
dx 1 1 1 1 x−1
=2 − dx = 2 ln .
x2 − 1 x−1 x+1 x+1
As we shall see, for higher values of n one cannot completely factorize (xn − 1)
into linear factors without employing complex numbers—a scarce and dubious
commodity in 1716! However, Cotes was aware that if he could break down (xn −1)
into real linear and quadratic factors, then he would be able to evaluate the inte-
gral. Here, a “real quadratic” refers to a quadratic whose coefficients are all real
numbers.
For example, (x4 − 1) can be broken down into (x − 1)(x + 1)(x2 + 1), yielding a
partial fraction expression of the form
1 A B Cx D
= + + + ,
x4 − 1 x − 1 x + 1 x2 + 1 x2 + 1
and hence an integral that can be evaluated in terms of ln and tan−1 . More generally,
even if the factorization involves more complicated quadratics than (x2 + 1), it is
easy to show that only ln and tan−1 are needed to evaluate the resulting integrals.
26 Geometry and Complex Arithmetic

In order to set Cotes’ work on (xn − 1) in a wider context, we shall investigate

the general connection between the roots of a polynomial and its factorization. This
connection can be explained by considering the geometric series,

Gm−1 = cm−1 + cm−2 z + cm−3 z2 + · · · + czm−2 + zm−1 ,

in which c and z are complex. Just as in real algebra, this series may be summed by
noting that zGm−1 and cGm−1 contain almost the same terms—try an example, say
m = 4, if you have trouble seeing this. Subtracting these two expressions yields

(z − c)Gm−1 = zm − cm , (1.15)

and thus
zm − cm
Gm−1 = .
z−c
If we think of c as fixed and z as variable, then (zm − cm ) is an mth -degree
polynomial in z, and z = c is a root. The result (1.15) says that this mth -degree
polynomial can be factored into the product of the linear term (z − c) and the
(m − 1)th -degree polynomial Gm−1 .
In 1637 Descartes published an important generalization of this result. Let Pn (z)
denote a general polynomial of degree n:

Pn (z) = zn + Azn−1 + · · · + Dz + E,

where the coefficients A, . . . , E may be complex. Since (1.15) implies

Pn (z) − Pn (c) = (z − c) [Gn−1 + AGn−2 + · · · + D] ,

we obtain Descartes’ Factor Theorem linking the existence of roots to factorizability:

If c is a solution of Pn (z) = 0 then Pn (z) = (z − c) Pn−1 , where Pn−1 is

of degree (n − 1).
If we could in turn find a root c ′ of Pn−1 , then the same reasoning would yield
Pn = (z − c)(z − c ′ ) Pn−2 . Continuing in this way, Descartes’ theorem therefore
holds out the promise of factoring Pn into precisely n linear factors:

Pn (z) = (z − c1 )(z − c2 ) · · · (z − cn ). (1.16)

If we do not acknowledge the existence of complex roots (as in the early 18th
century) then this factorization will be possible in some cases (e.g., z2 − 1), and
impossible in others (e.g., z2 + 1). But, in splendid contrast to this, if one admits
complex numbers then it can be shown that Pn always has n roots in C, and the fac-
torization (1.16) is always possible. This is called the Fundamental Theorem of Algebra,
and we shall explain its truth in Chapter 7.
Each factor (z−ck ) in (1.16) represents a complex number connecting the root ck
to the variable point z. Figure [1.17a] illustrates this for a general cubic polynomial.
 Some Applications 27

[1.17] Geometric evaluation of polynomials. [a] If the roots of polynomial P are known,
then its value at z can easily be visualized: Connect the roots to z, then add the angles
and multiply the lengths to obtain the direction and length of P(z). [b] First example of
Cotes’s geometrical factorization of xn − 1: Here, x2 − 1 = PC1 · PC2 .

Writing each of these complex numbers in the form Rk eiϕk , (1.16) takes the more
vivid form
Pn (z) = R1 R2 · · · Rn ei(ϕ1 +ϕ2 +···+ϕn ) .
Although the Fundamental Theorem of Algebra was not available to Cotes, let
us see how it guarantees that he would succeed in his quest to decompose xn − 1
into real linear and quadratic factors. Cotes’ polynomial has real coefficients, and,
quite generally, we can show that
If a polynomial has real coefficients then its complex roots occur in complex
conjugate pairs, and it can be factorized into real linear and quadratic factors.
For if the coefficients A, . . . , E of Pn (z) are all real then Pn (c) = 0 implies [exercise]
Pn ( c ) = 0, and the factorization (1.16) contains
(z − c)(z − c) = z2 − (c + c) z + cc = z2 − 2 Re(c) z + |c|2 ,
which is a real quadratic.
Let us now discuss how Cotes was able to factorize xn − 1 into real linear and
quadratic factors by appealing to the geometry of the regular n-gon. [An “n-gon” is an
n-sided polygon.] To appreciate the following, place yourself in his 18th century
shoes and forget all you have just learnt concerning the Fundamental Theorem of
Algebra; even forget about complex numbers and the complex plane!
For the first few values of n, the desired factorizations of Un (x) = xn − 1 are not
too hard to find:
U2 (x) = (x − 1)(x + 1), (1.17)
U3 (x) = (x − 1)(x2 + x + 1), (1.18)
28 Geometry and Complex Arithmetic

U4 (x) = (x − 1)(x + 1)(x2 + 1), (1.19)

 h √ i  h √ i 
U5 (x) = (x − 1) x2 + 1+2 5 x + 1 x2 + 1−2 5 x + 1 ,

but the general pattern seems elusive.

To find such a pattern, let us try to visualize the simplest case, (1.17). See [1.17b].
Let O be a fixed point, and P a variable point, on a line in the plane (which we are
not thinking of as C), and let x denote the distance OP. If we now draw a circle of
unit radius centred at O, and let C1 and C2 be its intersection points with the line,
then clearly8 U2 (x) = PC1 · PC2 .
To understand quadratic factors in this spirit, let us skip over (1.18) to the sim-
pler quadratic in (1.19). This factorization of U4 (x) is the best we could do without
complex numbers, but ideally we would have liked to have decomposed U4 (x) into
four linear factors. This suggests that we rewrite (1.19) as
p p
U4 (x) = (x − 1)(x + 1) x2 + 1 x2 + 1,

the last two “factors” being analogous to genuine linear factors. If we are to inter-
pret this expression (by analogy with the previous case) as the product of the
distances of P from four fixed points, then the points corresponding to the last
two “factors” must be off the line. More precisely, Pythagoras’ Theorem tells us that
√
a point whose distance from P is x2 + 12 must lie at unit distance from O in a
direction at right angles to the line OP.
Referring to [1.18a], we can now see that U4 (x) = PC1 · PC2 · PC3 · PC4 , where
C1 C2 C3 C4 is the illustrated square inscribed in the circle.
Since we have factorized U4 (x) with the regular 4-gon (the square), perhaps we
can factorize U3 (x) with the regular 3-gon (the equilateral triangle). See [1.18b].

[1.18] Cotes’s geometrical factorization of (xn − 1) using regular n-gons: If OP = x,

then [a] x4 − 1 = PC1 · PC2 · PC3 · PC4 ; [b] x3 − 1 = PC1 · PC2 · PC3 .

8
Here, and in what follows, we shall suppose for convenience that x > 1, so that Un (x) is positive.
 Some Applications 29

[1.19] The nth roots of unity: Each vertex of the regular n-gon is a solution of zn = 1.

Applying Pythagoras’ Theorem to this figure,

 h √ i2 
 
1 2
PC1 · PC2 · PC3 = PC1 · (PC2 ) = (x − 1)
2
x+ 2 + 2
3

= (x − 1)(x2 + x + 1),
which is indeed the desired factorization (1.18) of U3 (x)!
A plausible generalization for Un now presents itself:
If C1 C2 C3 · · · Cn is a regular n-gon inscribed in a circle of unit radius centred at
O, and P is the point on OC1 at distance x from O, then Un (x) = PC1 · PC2 · · ·
PCn .
This is Cotes’ result. Unfortunately, he stated it without proof, and he left no clue as
to how he discovered it. Thus we can only speculate that he may have been guided
by an argument like the one we have just supplied9 .
Since the vertices of the regular n-gon will always come in symmetric pairs that
are equidistant from P, the examples in [1.18] make it clear that Cotes’ result is
indeed equivalent to factorizing Un (x) into real linear and quadratic factors.
Recovering from our feigned bout of amnesia concerning complex numbers and
their geometric interpretation, Cotes’ result becomes simple to understand and to
prove. Taking O to be the origin of the complex plane, and C1 to be 1, the vertices
of Cotes’ n-gon are given by Ck+1 = eik(2π/n) . See [1.19], which illustrates the case
n = 12. Since (Ck+1 )n = eik2π = 1, all is suddenly clear: The vertices of the regular
n-gon are the n complex roots of Un (z) = zn − 1. Because the solutions of zn − 1 = 0
√
may be written formally as z = n 1, the vertices of the n-gon are called the nth roots
of unity.

9
Stillwell (2010) has instead speculated that Cotes used complex numbers (as we are about to), but
then deliberately stated his findings in a form that did not require them.
30 Geometry and Complex Arithmetic

By Descartes’ Factor Theorem, the complete factorization of (zn − 1) is therefore

zn − 1 = Un (z) = (z − C1 )(z − C2 ) · · · (z − Cn ),

with each conjugate pair of roots yielding a real quadratic factor,

    
z − eik(2π/n) z − e−ik(2π/n) = z2 − 2z cos 2kπ n + 1.

Each factor (z − Ck ) = Rk eiϕk may be viewed (cf. [1.17a]) as a complex number

connecting a vertex of the n-gon to z. Thus, if P is an arbitrary point in the plane
(not merely a point on the real axis), then we obtain the following generalized form
of Cotes’ result:

Un (P) = [PC1 · PC2 · · · PCn ] eiΦ ,

where Φ = (ϕ1 + ϕ2 + · · · + ϕn ). If P happens to be a real number (again supposed

greater than 1) then Φ = 0 [make sure you see this], and we recover Cotes’ result.
We did not immediately state and prove Cotes’ result in terms of complex num-
bers because we felt there was something rather fascinating about our first, direct
approach. Viewed in hindsight, it shows that even if we attempt to avoid complex
numbers, we cannot avoid the geometry of the complex plane!

1.3.6 Vectorial Operations

Not only is complex addition the same as vector addition, but we will now show
that the familiar vectorial operations of dot and cross products (also called scalar
and vector products) are both subsumed by complex multiplication. Since these
vectorial operations are extremely important in physics—they were discovered
by physicists!—their connection with complex multiplication will prove valuable
both in applying complex analysis to the physical world, and in using physics to
understand complex analysis.
When a complex number z = x + iy is being thought of merely as a vector, we
shall write it in bold type, with its components in a column:
!
x
z = x + iy ⇐⇒ z = .
y

Although the dot and cross product are meaningful for arbitrary vectors in space,
we shall assume in the following that our vectors all lie in a single plane—the
complex plane.
Given two vectors a and b, figure [1.20a] recalls the definition of the dot product
as the length of one vector, times the projection onto that vector of the other vector:

· ·
a b = |a| |b| cos θ = b a,

where θ is the angle between a and b.

 Some Applications 31

[1.20] [a] The dot product is defined to be the orthogonal projection of one vector onto
the other, times the length of the other vector. [b] The cross product is defined to be the
vector orthogonal to both (oriented according to the “right hand rule”), with length equal
to the area of the parallelogram they span.

Figure [1.20b] recalls the definition of the cross product: a × b is the vector per-
pendicular to the plane of a and b whose length is equal to the area A of the
parallelogram spanned by a and b. But wait, there are two (opposite) directions
perpendicular to C; which should we choose?
Writing A = |a| |b| sin θ, the area A has a sign attached to it. An easy way to see
this sign is to think of the angle θ from a to b as lying in range −π to π; the sign
of A is then the same as θ. If A > 0, as in [1.20b], then we define a × b to point
upwards from the plane, and if A < 0 we define it to point downwards. It follows
that a × b = −(b × a).
This conventional definition of a × b is intrinsically three-dimensional, and it
therefore presents a problem: if a and b are thought of as complex numbers, a × b
cannot be, for it does not lie in the (complex) plane of a and b. No such problem
·
exists with the dot product because a b is simply a real number, and this suggests
a way out.
Since all our vectors will be lying in the same plane, their cross products will
all have equal (or opposite) directions, so the only distinction between one cross
product and another will be the value of A. For the purposes of this book we will
therefore redefine the cross product to be the (signed) area A of the parallelogram spanned
by a and b:
a × b = |a| |b| sin θ = −(b × a).
Figure [1.21] shows two complex numbers a = |a| eiα and b = |b| eiβ , the angle
from a to b being θ = (β − α). To see how their dot and cross products are related
to complex multiplication, consider the effect of multiplying each point in C by
a. This is a rotation of −α and an expansion of |a|, and if we look at the image
32 Geometry and Complex Arithmetic

[1.21] The dot and cross products in terms of complex multiplication. Multiplication of C
by a is a rotation of −α and an expansion of |a|. It follows that the dot and cross products
can be expressed as the real and imaginary parts of a b.

under this transformation of the shaded right triangle with hypotenuse b, then we
immediately see that

·
a b = a b + i (a × b). (1.20)

Of course we could also have obtained this by simple calculation:

a b = (|a| e−iα )(|b| eiβ ) = |a| |b| ei(β−α) = |a| |b| eiθ = |a| |b|(cos θ + i sin θ).

When we refer to the dot and cross products as “vectorial operations” we

mean that they are defined geometrically, independently of any particular choice of
coordinate axes. However, once such a choice has been made, (1.20) makes it easy
to express these operations in terms of Cartesian coordinates. Writing a = x + iy
and b = x ′ + iy ′ ,

a b = (x − iy)(x ′ + iy ′ ) = (xx ′ + yy ′ ) + i (xy ′ − yx ′ ),

so
! ! ! !
x′ x′
·
x ′ ′ x
= xx + yy and × = xy ′ − yx ′ .
y y′ y y′

We end with an example that illustrates the importance of the sign of the area
(a×b). Consider the problem of finding the area A of the quadrilateral in [1.22a],
whose vertices are, in counterclockwise order, a, b, c, and d. Clearly this is just the
sum of the ordinary, unsigned areas of the four triangles formed by joining the
 Transformations and Euclidean Geometry* 33

[1.22] Area of a quadrilateral in terms of signed areas of triangles. [a] The area of the
quadrilateral is the sum of the areas of the four triangles, and is therefore given by (1.21).
[b] The area of the quadrilateral is still the sum of the areas of the four triangles, because
the striped area is automatically subtracted from the areas of the other three triangles.

vertices of the quadrilateral to the origin. Thus, since the area of each triangle is
simply half the area of the corresponding parallelogram,
A = 1
2 [(a×b) + (b×c) + (c×d) + (d×a)]
1
 
= 2 Im a b + b c + c d + d a . (1.21)
Obviously this formula could easily be generalized to polygons with more than
four sides.
But what if 0 is outside the quadrilateral? In [1.22b], A is clearly the sum of the
ordinary areas of three of the triangles, minus the ordinary area of the striped tri-
angle. Since the angle from b to c is negative, 12 (b×c) is automatically the negative
of the striped area, and A is therefore given by exactly the same formula as before!
Can you find a location for 0 that makes two of the signed areas negative? Check
that the formula still works. Exercise 35 shows that (1.21) always works.

1.4 Transformations and Euclidean Geometry*

1.4.1 Geometry Through the Eyes of Felix Klein

Even with the benefit of enormous hindsight, it is hard to introduce complex num-
bers in a compelling manner. Historically, we have seen how cubic equations
forced them upon us algebraically, and in discussing Cotes’ work we saw some-
thing of the inevitability of their geometric interpretation. In this section we will
attempt to show how complex numbers arise very naturally, almost inevitably,
from a careful re-examination of plane Euclidean geometry10 .

10
The excellent book by Nikulin and Shafarevich (1987) is the only other work we know of in which
a similar attempt is made.
34 Geometry and Complex Arithmetic

As the * following the title of this section indicates, the material it contains
may be omitted. However, in addition to “explaining” complex numbers, these
ideas are very interesting in their own right, and they will also be needed for an
understanding of other optional sections of the book.
Although the ancient Greeks made many beautiful and remarkable discover-
ies in geometry, it was two thousand years later that Felix Klein first asked and
answered the question, “What is geometry?”
Let us restrict ourselves from the outset to plane geometry. One might begin by
saying that this is the study of geometric properties of geometric figures in the
plane, but what are (i) “geometric properties”, and (ii) “geometric figures”? We
will concentrate on (i), swiftly passing over (ii) by interpreting “geometric figure”
as anything we might choose to draw on an infinitely large piece of flat paper with
an infinitely fine pen.
As for (i), we begin by noting that if two figures (e.g., two triangles) have
the same geometric properties, then (from the point of view of geometry) they
must be the “same”, “equal”, or, as one usually says, congruent. Thus if we had
a clear definition of congruence (“geometric equality”) then we could reverse
this observation and define geometric properties as those properties that are common
to all congruent figures. How, then, can we tell if two figures are geometrically
equal?
Consider the triangles in [1.23], and imagine that they are pieces of paper that
you could pick up in your hand. To see if T is congruent to T ′ , you could pick up
T and check whether it could be placed on top of T ′ . Note that it is essential that
we be allowed to move T in space: in order to place T on top of Te we must first
flip it over; we can’t just slide T around within the plane. Tentatively generalizing,
this suggests that a figure F is congruent to another figure F ′ if there exists a motion of
F through space that makes it coincide with F ′ . Note that the discussion suggests that
there are two fundamentally different types of motion: those that involve flipping
the figure over, and those that do not. Later, we shall return to this important
point.
It is clearly somewhat unsatisfactory that in attempting to define geometry in
the plane we have appealed to the idea of motion through space. We now rectify
this. Returning to [1.23], imagine that T and T ′ are drawn on separate, transparent
sheets of plastic. Instead of picking up just the triangle T , we now pick up the entire
sheet on which it is drawn, then try to place it on the second sheet so as to make
T coincide with T ′ . At the end of this motion, each point A on T ’s sheet lies over
a point A ′ of T ′ ’s sheet, and we can now define the motion M to be this mapping
A ÞÑ A ′ = M(A) of the plane to itself.
However, not any old mapping qualifies as a motion, for we must also capture
the (previously implicit) idea of the sheet remaining rigid while it moves, so that
 Transformations and Euclidean Geometry* 35

[1.23] The triangles T ′ and Te are both “geometrically equal” (“congruent”) to T because
there exists a “motion” that carries T to each of these. But the motion that carries T to Te
requires a flip.

distances between points remain constant during the motion. Here, then, is our
definition:
A “motion” M is a mapping of the plane to itself such that the distance
between any two points A and B is equal to the distance between their (1.22)
images A ′ = M(A) and B ′ = M(B).
Note that what we have called a motion is often termed a “rigid motion”, or an
“isometry”.
Armed with this precise concept of a motion, our final definition of geometric
equality becomes
∼ F ′ , if there exists a motion M such that
F is congruent to F ′ , written F =
(1.23)
F ′ = M(F).
Next, as a consequence of our earlier discussion, a geometric property of a figure
is one that is unaltered by all possible motions of the figure. Finally, in answer to the
opening question of “What is geometry?”, Klein would answer that it is the study
of these so-called invariants of the set of motions.
One of the most remarkable discoveries of the 19th century was that Euclidean
geometry is not the only possible geometry. Two of these so-called non-Euclidean
geometries will be studied in Chapter 6, but for the moment we wish only to
explain how Klein was able to generalize the above ideas so as to embrace such
new geometries.
The aim in (1.23) was to use a family of transformations to introduce a concept
∼
of geometric equality. But will this =-type of equality behave in the way we would
like and expect? To answer this we must first make these expectations explicit. So
as not to confuse this general discussion with the particular concept of congruence
in (1.23), let us denote geometric equality by ∼.

(i) A figure should equal itself: F ∼ F, for all F.

36 Geometry and Complex Arithmetic

[1.24] Klein’s vision of a geometry in terms of a group of transformations. In projective

geometry “geometric equality” is defined using more general transformations than rigid
Euclidean motions. But geometric equality only makes sense if it is an equivalence relation,
and this in turn implies that the generalized transformations satisfy the three requirements
of a group.

(ii) If F equals F ′ , then F ′ should equal F: F ∼ F ′ ⇒ F ′ ∼ F.

(iii) If F and F ′ are equal, and F ′ and F ′′ are equal, then F and F ′′ should also be
equal: F ∼ F ′ & F ′ ∼ F ′′ ⇒ F ∼ F ′′ .

Any relation satisfying these expectations is called an equivalence relation.

Now suppose that we retain the definition (1.23) of geometric equality, but that
we generalize the definition of “motion” given in (1.22) by replacing the family of
distance-preserving transformations with some other family G of transformations.
It should be clear that not any old G will be compatible with our aim of defining
geometric equality. Indeed, (i), (ii), and (iii) imply that G must have the following
very special structure, which is illustrated11 in [1.24].

(i) The family G must contain a transformation E (called the identity) that maps
each point to itself.
(ii) If G contains a transformation M, then it must also contain a transformation
M−1 (called the inverse) that undoes M. [Check for yourself that for M−1 to
exist (let alone be a member of G) M must have the special properties of being

11
Here G is the group of projections. If we do a perspective drawing of figures in the plane, then the
mapping from that plane to the “canvas” plane is called a perspectivity. A projection is then defined to
be any sequence of perspectivities. Can you see why the set of projections should form a group?
 Transformations and Euclidean Geometry* 37

(a) onto and (b) one-to-one, i.e., (a) every point must be the image of some
point, and (b) distinct points must have distinct images.]
(iii) If M and N are members of G then so is the composite transformation N◦M =
(M followed by N). This property of G is called closure.

We have thus arrived, very naturally, at a concept of fundamental importance in

the whole of mathematics: a family G of transformations that satisfies these three12
requirements is called a group.
Let us check that the motions defined in (1.22) do indeed form a group: (i) Since
the identity transformation preserves distances, it is a motion. (ii) Provided it exists,
the inverse of a motion will preserve distances and hence will be a motion itself.
As for existence, (a) it is certainly plausible that when we apply a motion to the
entire plane then the image is the entire plane—we will prove this later—and (b)
the non-zero distance between distinct points is preserved by a motion, so their
images are again distinct. (iii) If two transformations do not alter distances, then
applying them in succession will not alter distances either, so the composition of
two motions is another motion.
Klein’s idea was that we could first select a group G at will, then define a corre-
sponding “geometry” as the study of the invariants of that G. [Klein first announced
this idea in 1872—when he was 23 years old!—at the University of Erlangen, and
it has thus come to be known as his Erlangen Program.] For example, if we choose
G to be the group of rigid motions, we recover the familiar Euclidean geometry of
the plane. But this is far from being the only geometry of the plane, as the so-called
projective geometry of [1.24] illustrates.
Klein’s vision of geometry was broader still. We have been concerned with what
geometries are possible when figures are drawn anywhere in the plane, but suppose
for example that we are only allowed to draw within some disc D. It should be
clear that we can construct “geometries of D” in exactly the same way that we
constructed geometries of the plane: given a group H of transformations of D to
itself, the corresponding geometry is the study of the invariants of H. If you doubt
that any such groups exist, consider the set of all rotations around the centre of D.
The reader may well feel that the above discussion is a chronic case of mathemat-
ical generalization running amuck—that the resulting conception of geometry is (to
coin a phrase) “as subtle as it is useless”. Nothing could be further from the truth!
In Chapter 3 we shall be led, very naturally, to consider a particularly interesting
group of transformations of a disc to itself. The resulting non-Euclidean geometry
is called hyperbolic or Lobachevskian geometry, and it is the subject of Chapter 6. Far
from being useless, this geometry has proved to be an immensely powerful tool in

12
In more abstract settings it is necessary to add a fourth requirement of associativity, namely,
A ◦ (B ◦ C) = (A ◦ B) ◦ C. Of course for transformations this is automatically true.
38 Geometry and Complex Arithmetic

diverse areas of mathematics and physics, and the insights it continues to provide
lie on the cutting edge of contemporary research.

1.4.2 Classifying Motions

To understand the foundations of Euclidean geometry, it seems we must study its
group of motions. At the moment, this group is defined rather abstractly as the set
of distance-preserving mappings of the plane to itself. However, it is easy enough
to think of concrete examples of motions: a rotation of the plane about an arbitrary
point, a translation of the plane, or a reflection of the plane in some line. Our aim
is to understand the most general possible motions in equally vivid terms.
We begin by stating a key fact:
A motion is uniquely determined by its effect on any triangle (i.e., on
(1.24)
any three non-collinear points).
By this we mean that knowing what happens to the three points tells us what must
happen to every point in the plane. To see this, first look at [1.25]. This shows that
each point P is uniquely determined by its distances from the vertices A, B, C of
such a triangle13 . The distances from A and B yield two circles which (in general)
intersect in two points, P and Q. The third distance (from C) then picks out P.
To obtain the result (1.24), now look at [1.26]. This illustrates a motion M map-
ping A, B, C to A ′ , B ′ , C ′ . By the very definition of a motion, M must map an
arbitrary point P to a point P ′ whose distances from A ′ , B ′ , C ′ are equal to the
original distances of P from A, B, C. Thus, as shown, P ′ is uniquely determined.
Done.

[1.25] Each point P is uniquely determined by its distances from the vertices A, B, C of
a triangle.

13
This is how earthquakes are located. Two types of wave are emitted by the quake as it begins:
fast-moving “P-waves” of compression, and slower-moving “S-waves” of destructive shear. Thus the
P-waves will arrive at a seismic station before the S-waves, and the time-lag between these events may
be used to calculate the distance of the quake from that station. Repeating this calculation at two more
seismic stations, the quake may be located.
 Transformations and Euclidean Geometry* 39

[1.26] A rigid motion is uniquely determined by its effect on any triangle: If M maps ABC
to A ′ B ′ C ′ , then an arbitrary point P must be mapped to the illustrated point P ′ = M(P).

A big step towards classification is the realization that there are two funda-
mentally different kinds of motions. In terms of our earlier conception of motion
through space, the distinction is whether or not a figure must be flipped over before
it can be placed on top of a congruent figure. To see how this dichotomy arises in
terms of the new definition (1.22), suppose that a motion sends two points A and
B to A ′ and B ′ . See [1.27]. According to (1.24), the motion is not yet determined:
we need to know the image of any (non-collinear) third point C, such as the one
shown in [1.27]. Since motions preserve the distances of C from A and B, there are
e in the line L
just two possibilities for the image of C, namely, C ′ and its reflection C
′ ′ e
through A and B . Thus there are precisely two motions (M and M, say) that map
e sends C to C.
A, B to A ′ , B ′ : M sends C to C ′ , and M e
e
A distinction can be made between M and M by looking at how they affect
angles. All motions preserve the magnitude of angles, but we see that M also pre-
serves the sense of the angle θ, while Me reverses it. The fundamental nature of this
distinction can be seen from the fact that M must in fact preserve all angles, while
Me must reverse all angles.
To see this, consider the fate of the angle ϕ in the triangle T . If C goes to C ′
(i.e., if the motion is M) then, carrying out the construction indicated in [1.26],
the image of T is T ′ , and the angle is preserved. If, on the other hand, C goes to
e then the image of T is the reflection Te of T ′ in L, and
e (i.e., if the motion is M)
C
the angle is reversed. Motions that preserve angles are called direct, and those that
reverse angles are called opposite. Thus rotations and translations are direct, while
reflections are opposite. Summarizing what we have found,

There is exactly one direct motion M (and exactly one opposite

motion M)e that maps a given line-segment AB to another
e = (M (1.25)
line-segment A ′ B ′ of equal length. Furthermore, M
′ ′
followed by reflection in the line A B ).
40 Geometry and Complex Arithmetic

[1.27] If a motion M maps a given line-segment AB to another line-segment A ′ B ′ of equal

e in A ′ B ′ . Thus,
length, then a third point C must be mapped to either C ′ or its reflection C
there is exactly one direct motion M (and exactly one opposite motion M) f that maps
′ ′
a given line-segment AB to another line-segment A B of equal length. Furthermore,
Me = (M followed by reflection in the line A ′ B ′ ).

To understand motions we may thus consider two randomly drawn segments

AB and A ′ B ′ of equal length, then find the direct motion (and the opposite motion)
that maps one to the other. It is now easy to show that
Every direct motion is a rotation, or else (exceptionally) a translation. (1.26)
Note that this result gives us greater insight into our earlier calculations on the
composition of rotations and translations: since the composition of any two direct
motions is another direct motion [why?], it can only be a rotation or a translation.
Conversely, those calculations allow us to restate (1.26) in a very neat way:
Every direct motion can be expressed as a complex function of the form
(1.27)
M(z) = eiθ z + v.
We now establish (1.26). If the line-segment A ′ B ′ is parallel to AB then the vec-
tors AB and A ′ B ′ are either equal or opposite. If they are equal, as in [1.28a], the
motion is a translation; if they are opposite, as in [1.28b], the motion is a rotation
of π about the intersection point of the lines AA ′ and BB ′ .
If the segments are not parallel, produce them (if necessary) till they meet at M,
and let θ be the angle between the directions of AB and A ′ B ′ . See [1.28c]. First
recall an elementary property of circles: the chord AA ′ subtends the same angle θ
at every point of the circular arc AMA ′ . Next, let O denote the intersection point of
this arc with the perpendicular bisector of AA ′ . We now see that the direct motion
carrying AB to A ′ B ′ is a rotation of θ about O, for clearly A is rotated to A ′ , and
the direction of AB is rotated into the direction of A ′ B ′ . Done.
 Transformations and Euclidean Geometry* 41

[1.28] Every direct motion is a rotation, or else (exceptionally) a translation. In the excep-
tional case that the motion M maps a line segment AB to a parallel segment A ′ B ′ , M is
either [a] a translation, or [b] a rotation of π about the intersection point of the lines AA ′
and BB ′ . However, in the general case [c], M is a rotation of θ about O, where θ and O
are constructed as shown.

The sense in which translations are “exceptional” is that if the two segments are
drawn at random then it is very unlikely that they will be parallel. Indeed, given
AB, a translation is only needed for one possible direction of A ′ B ′ out of infinitely
many, so the mathematical probability that a random direct motion is a translation
is actually zero!
Direct transformations will be more important to us than opposite ones, so we
relegate the investigation of opposite motions to Exs. 39, 40, 41. The reason for the
greater emphasis on direct motions stems from the fact that they form a group
(a subgroup of the full group of motions), while the opposite motions do not. Can
you see why?

1.4.3 Three Reflections Theorem

In chemistry one is concerned with the interactions of atoms, but to gain deeper
insights one must study the electrons, protons, and neutrons from which atoms
are built. Likewise, though our concern is with direct motions, we will gain deeper
insights by studying the opposite motions from which direct motions are built.
More precisely,

Every direct motion is the composition of two reflections. (1.28)

Note that the second sentence of (1.25) then implies that every opposite motion is the
composition of three reflections. See Ex. 39. In brief, every motion is the composition
42 Geometry and Complex Arithmetic

of either two or three reflections, a result that is called the Three Reflections
Theorem14 .
Earlier we tried to show that the set of motions forms a group, but it was not clear
that every motion had an inverse. The Three Reflections Theorem settles this neatly
and explicitly, for the inverse of a sequence of reflections is obtained by reversing
the order in which the reflections are performed.
In what follows, let ℜL denote reflection in a line L. Thus reflection in L1 followed
by reflection in L2 is written ℜL2 ◦ ℜL1 . According to (1.26), proving (1.28) amounts
to showing that every rotation (and every translation) is of the form ℜL2 ◦ ℜL1 . This
is an immediate consequence of the following:

If L1 and L2 intersect at O, and the angle from L1 to L2 is ϕ, then

ℜL2 ◦ ℜL1 is a rotation of 2ϕ about O,
and
If L1 and L2 are parallel, and V is the perpendicular connecting vector from
L1 to L2 , then ℜL2 ◦ ℜL1 is a translation of 2V.
Both these results are easy enough to prove directly [try it!], but the following is
perhaps more elegant.
First, since ℜL2 ◦ ℜL1 is a direct motion (because it reverses angles twice), it is
either a rotation or a translation. Second, note that rotations and translations may
be distinguished by their invariant curves, that is, curves that are mapped into them-
selves. For a rotation about a point O, the invariant curves are circles centred at O,
while for a translation they are lines parallel to the translation.
Now look at [1.29a]. Clearly ℜL2 ◦ ℜL1 leaves invariant any circle centred at O,
so it is a rotation about O. To see that the angle of the rotation is 2ϕ, consider the
image P ′ of any point P on L1 . Done.
Now look at [1.29b]. Clearly ℜL2 ◦ ℜL1 leaves invariant any line perpendicular
to L1 and L2 , so it is a translation parallel to such lines. To see that the translation is
2V, consider the image P ′ of any point P on L1 . Done.
Note that a rotation of θ can be represented as ℜL2 ◦ ℜL1 , where L1 , L2 is any pair
of lines that pass through the centre of the rotation and that contain an angle (θ/2).
Likewise, a translation of T corresponds to any pair of parallel lines separated by
T/2. This circumstance yields a very elegant method for composing rotations and
translations.
For example, see [1.30a]. Here a rotation Rθa about a through θ is being repre-
sented as ℜL2 ◦ ℜL1 , and a rotation Rϕ b about b through ϕ is being represented as
ℜL2′ ◦ℜL1′ . To find the net effect of rotating about a and then about b, choose L2 = L1′
to be the line through a and b. If θ + ϕ ̸= 2π, then L1 and L2′ will intersect at some

14
Results such as (1.26) may instead be viewed as consequences of this theorem; see Stillwell (1992)
for an elegant and elementary exposition of this approach.
 Transformations and Euclidean Geometry* 43

[1.29] The composition of two reflections is a rotation or a translation. [a] If L1 and

L2 intersect at O, and the angle from L1 to L2 is ϕ, then ℜL2 ◦ ℜL1 is a rotation of 2ϕ
about O. [b] If L1 and L2 are parallel, and V is the perpendicular connecting vector
from L1 to L2 , then ℜL2 ◦ ℜL1 is a translation of 2V.

[1.30] Geometric reduction of two rotations to a single rotation. [a] The composi-
tion of two rotations Rϕ b ◦ Ra can be expressed as the composition of four reflections:
θ

(ℜL2′ ◦ ℜL1′ ) ◦ (ℜL2 ◦ ℜL1 ). [b] But if we choose L2 = L1′ then the middle two reflections
cancel, and we discover that Rϕ b ◦ Ra = Rc
θ θ+ϕ
.

point c, as in [1.30b]. Thus the middle two reflections cancel, and the composition
of the two rotations is given by

(ℜL2′ ◦ ℜL1′ ) ◦ (ℜL2 ◦ ℜL1 ) = ℜL2′ ◦ ℜL1 ,

which is a rotation Rθ+ϕ

c about c through (θ + ϕ)! That this construction agrees
with our calculation on p. 20 is demonstrated in Ex. 36.
Further examples of this method may be found in Ex. 42 and Ex. 43.
44 Geometry and Complex Arithmetic

1.4.4 Similarities and Complex Arithmetic

Let us take a closer look at the role of distance in Euclidean geometry. Suppose we
have two right triangles T and Te drawn in the same plane, and suppose that Jack
measures T while Jill measures Te. If Jack and Jill both report that their triangles
have sides 3, 4, and 5, then it is tempting to say that the two triangles are the same,
in the sense that there exists a motion M such that Te = M(T ). But wait! Suppose
that Jack’s ruler is marked in centimetres, while Jill’s is marked in inches. The two
triangles are similar, but they are not congruent. Which is the “true” 3, 4, 5 triangle?
Of course they both are.
The point is that whenever we talk about distances numerically, we are presupposing a
unit of measurement. This may be pictured as a certain line-segment U, and when we
say that some other segment has a length of 5, for example, we mean that precisely
5 copies of U can be fitted into it. But on our flat15 plane any choice of U is as good as
any other—there is no absolute unit of measurement, and our geometric theorems
should reflect that fact.
Meditating on this, we recognize that Euclidean theorems do not in fact depend
on this (arbitrary) choice of U, for they only deal with ratios of lengths, which
are independent of U. For example, Jack can verify that his triangle T satisfies
Pythagoras’ Theorem in the form
(3cm)2 + (4cm)2 = (5cm)2 ,
but, dividing both sides by (5cm)2 , this can be rewritten in terms of the ratios of
the sides, which are pure numbers:
(3/5)2 + (4/5)2 = 1.
Try thinking of another theorem, and check that it too deals only with ratios of
lengths.
Since the theorems of Euclidean geometry do not concern themselves with the
actual sizes of figures, our earlier definition of geometric equality in terms of
motions is clearly too restrictive: two figures should be considered the same if they
are similar. More precisely, we now consider two figures to be the same if there
exists a similarity mapping one to the other, where
A similarity S is a mapping of the plane to itself that preserves ratios of distances.
It is easy to see [exercise] that a given similarity S expands every distance by the
same (non-zero) factor r, which we will call the expansion of S. We can therefore
refine our notation by including the expansion as a superscript, so that a general
similarity of expansion r is written Sr . Clearly, the identity transformation is a sim-
ilarity, Sk ◦ Sr = Skr , and (Sr )−1 = S(1/r) , so it is fairly clear that the set of all

15
In the non-Euclidean geometries of Chapter 6 we will be drawing on curved surfaces, and the
amount of curvature in the surface will dictate an absolute unit of length.
 Transformations and Euclidean Geometry* 45

[1.31] Dilative rotations. [a] A central dilation Dro leaves o fixed and radially
stretches each segment oA by r. [b] If this central dilation is followed by (or pre-
ceded by) a rotation Rθo with the same centre, then we obtain the dilative rotation
Dr,θ
o ≡ Ro ◦ Do = Do ◦ Ro .
θ r r θ

similarities forms a group. We thus arrive at the definition of Euclidean geometry

that Klein gave in his Erlangen address:
Euclidean geometry is the study of those properties of geometric figures
(1.29)
that are invariant under the group of similarities.
Since the motions are just the similarities S1 of unit expansion, the group of
motions is a subgroup of the group of similarities; our previous attempt at defining
Euclidean geometry therefore yields a “subgeometry” of (1.29).
A simple example of an Sr is a central dilation Dro . As illustrated in [1.31a], this
leaves o fixed and radially stretches each segment oA by r. Note that the inverse of
(1/r)
a central dilation is another central dilation with the same centre: (Dro )−1 = Do .
If this central dilation is followed by (or preceded by) a rotation Rθo with the same
centre, then we obtain the dilative rotation

o ≡ Ro ◦ Do = Do ◦ Ro ,
Dr,θ θ r r θ

shown in [1.31b]. Note that a central dilation may be viewed as a special case of a
dilative rotation: Dro = Dr,0
o .
This figure should be ringing loud bells. Taking o to be the origin of C, (1.9) says
that Dr,θ
o corresponds to multiplication by r e :
iθ

Dr,θ
o (z) = r e
iθ
z.
Conversely, and this is the key point, the rule for complex multiplication may be viewed
as a consequence of the behaviour of dilative rotations.
Concentrate on the set of dilative rotations with a common, fixed centre o, which
will be thought of as the origin of the complex plane. Each Dr,θ o is uniquely deter-
mined by its expansion r and rotation θ, and so it can be represented by a vector
46 Geometry and Complex Arithmetic

of length r at angle θ. Likewise, DR,ϕ

o can be represented by a vector of length R
at angle ϕ. What vector will represent the composition of these dilative rotations?
Geometrically it is clear that

DR,ϕ
o ◦ Dr,θ
o = Do ◦ Do
r,θ R,ϕ
= DRr,(θ+ϕ)
o ,

so the new vector is obtained from the original vectors by multiplying their lengths
and adding their angles—complex multiplication!
On page 19 we saw that if complex numbers are viewed as translations then
composition yields complex addition. We now see that if they are instead viewed
as dilative rotations then composition yields complex multiplication. To complete
our “explanation” of complex numbers in terms of geometry, we will show that
these translations and dilative rotations are fundamental to Euclidean geometry as
defined in (1.29).
To understand the general similarity Sr involved in (1.29), note that if p is an
(1/r)
arbitrary point, M ≡ Sr ◦ Dp is a motion. Thus any similarity is the composition of
a dilation and a motion:

Sr = M ◦ Drp . (1.30)

Our classification of motions therefore implies that similarities come in two kinds:
if M preserves angles then so will Sr [a direct similarity]; if M reverses angles then
so will Sr [an opposite similarity].
Just as we concentrated on the group of direct motions, so we will now concen-
trate on the group of direct similarities. The fundamental role of translations and
dilative rotations in Euclidean geometry finally emerges in the following surprising
theorem:

Every direct similarity is a dilative rotation or (exceptionally) a translation. (1.31)

For us, at least, this fact constitutes one satisfying “explanation” of complex num-
bers; as mentioned in the Preface, other equally compelling explanations may be
found in the laws of physics.
To begin to understand (1.31), observe that (1.25) and (1.30) imply that a direct
similarity is determined by the image A ′ B ′ of any line-segment AB. First consider
the exceptional case in which A ′ B ′ are of equal length AB. We then have the three
cases in [1.28], all of which are consistent with (1.31). If A ′ B ′ and AB are parallel
but not of equal length, then we have the two cases shown in [1.32a] and [1.32b], in
both of which we have drawn the lines AA ′ and BB ′ intersecting in p. By appealing
to the similar triangles in these figures, we see that in [1.32a] the similarity is Dr,0
p ,
while in [1.32b] it is Dpr,π
, where in both cases r = (pA ′ /pA) = (pB ′ /pB).
Now consider the much more interesting general case where A ′ B ′ and AB are
neither the same length, nor parallel. Take a peek at [1.32d], which illustrates this.
Here n is the intersection point of the two segments (produced if necessary), and
 Transformations and Euclidean Geometry* 47

[1.32] Every direct similarity is a dilative rotation (with the sole exception of an isomet-
ric translation). If a direct similarity S carries a line segment AB to a parallel segment
A ′ B ′ of different length, then either [a] S = Dr,0 r,π
p , or [b] S = Dp , where in both cases
r = (pA /pA) = (pB /pB). [c] For a given θ, there exists a Dq mapping A to A ′ if and
′ ′ r,θ

only if q lies on the circular arc AnA ′ . [d] If Dr,θ ′

q is to simultaneously map A to A and
′ ′ ′
B to B then θ must be the angle between AB and A B , so q must lie on the circular
arc AnA ′ and on the circular arc BnB ′ , so q = n or q = m. But clearly q = n fails, so
q = m is our only hope. Miraculously, thanks to the similarity of the two shaded triangles,
′ ′
Dr,θ
m succeeds (!) in simultaneously mapping A to A and B to B , thereby completing the
proof.

θ is the angle between them. To establish (1.31), we must show that we can carry
AB to A ′ B ′ with a single dilative rotation. For the time being, simply note that if
AB is to end up having the same direction as A ′ B ′ then it must be rotated by θ, so
the claim is really this: There exists a point q, and an expansion factor r, such that Dr,θ
q
carries A to A ′ and B to B ′ .
Consider the part of [1.32d] that is reproduced in [1.32c]. Clearly, by choosing
r = (nA ′ /nA), Dn r,θ
will map A to A ′ . More generally, you see that we can map A
to A ′ with Dr,θ ′
q if and only if AA subtends angle θ at q. Thus, with the appropriate
value of r, Dq maps A to A ′ if and only if q lies on the circular arc AnA ′ . The figure
r,θ

illustrates one such position, q = m. Before returning to [1.32d], we need to notice

one more thing: mA subtends the same angle (marked •) at n and A ′ .
′ ′
Let us return to [1.32d]. We want Dr,θ q to map A to A and B to B . According
′
to the argument above, q must lie on the circular arc AnA and on the circular arc
48 Geometry and Complex Arithmetic

BnB ′ . Thus there are just two possibilities: q = n or q = m (the other intersection
point of the two arcs). If you think about it, this is a moment of high drama. We
have narrowed down the possibilities for q to just two points by consideration of
angles alone; for either of these two points we can choose the value of the expansion
r so as to make A go to A ′ , but, once this choice has been made, either B will map
to B ′ or it won’t! Furthermore, it is clear from the figure that if q = n then B does
not map to B ′ , so q = m is the only possibility remaining.
In order for Dm r,θ
to simultaneously map A to A ′ and B to B ′ , we need to have
r = (mA ′ /mA) = (mB ′ /mB); in other words, the two shaded triangles need to be
similar. That they are indeed similar is surely something of a miracle. Looking at the
angles formed at n, we see that θ + ⊙ + • = π, and the result follows immediately
by thinking of the RHS as the angle-sum of each of the two shaded triangles. This
completes our proof16 of (1.31).
The reader may feel that it is unsatisfactory that (1.31) calls for dilative rotations
about arbitrary points, while complex numbers represent dilative rotations about
a fixed point o (the origin). This may be answered by noting that the images of
AB under Dr,θ r,θ
q and Do will be parallel and of equal length, so there will exist a
translation [see Ex. 44 for details] Tv mapping one onto the other. In other words, a
general dilative rotation differs from an origin-centred dilative rotation by a mere
translation: Dr,θ r,θ
q = Tv ◦ Do . To sum up,

Every direct similarity Sr can be expressed as a complex function of the form

Sr (z) = reiθ z + v.

1.4.5 Spatial Complex Numbers?

Let us briefly attempt to generalize the above ideas to three-dimensional space.
Firstly, a central dilation of space (centred at O) is defined exactly as before, and
a dilative rotation with the same centre is then the composition of such a dilation
with a rotation of space about an axis passing through O. Once again taking (1.29)
as the definition of Euclidean geometry, we get off to a flying start, because the key
result (1.31) generalizes: Every direct similarity of space is a dilative rotation, a transla-
tion, or the composition of a dilative rotation and a translation along its rotation axis. See
Coxeter (1969, p. 103) for details.
It is therefore natural to ask if there might exist “spatial complex numbers” for
which addition would be composition of translations, and for which multiplication
would be composition of dilative rotations. With addition all goes well: the position
vector of each point in space may be viewed as a translation, and composition of

16
The present argument has the advantage of proceeding in steps, rather than having to be discov-
ered all at once. For other proofs, see Coxeter and Greitzer (1967, p. 97), Coxeter (1969, p. 73), and Eves
(1992, p. 71). Also, see Ex. 45 for a simple proof using complex functions.
 Transformations and Euclidean Geometry* 49

these translations yields ordinary vector addition in space. Note that this vector
addition makes equally good sense in four-dimensional space, or n-dimensional
space for that matter.
Now consider the set Q of dilative rotations with a common, fixed centre O. Ini-
tially, the definition of multiplication goes smoothly, for the “product” Q1 ◦ Q2
of two such dilative rotations is easily seen to be another dilative rotation (Q3 ,
say) of the same kind. This follows from the above classification of direct sim-
ilarities by noting that Q1 ◦ Q2 leaves O fixed. If the expansions of Q1 and Q2
are r1 and r2 then the expansion of Q3 is clearly r3 = r1 r2 , and in Chapter 6 we
shall give a simple geometric construction for the rotation of Q3 from the rota-
tions of Q1 and Q2 . However, unlike rotations in the plane, it makes a difference
in what order we perform two rotations in space, so our multiplication rule is not
commutative:

Q1 ◦ Q2 ̸= Q2 ◦ Q1 . (1.32)

We are certainly accustomed to multiplication being commutative, but there is noth-

ing inconsistent about (1.32), so this cannot be considered a decisive obstacle to an
algebra of “spatial complex numbers”.
However, a fundamental problem does arise when we try to represent these
dilative rotations as points (or vectors) in space. By analogy with complex
multiplication, we wish to interpret the equation Q1 ◦ Q2 = Q3 as saying that the
dilative rotation Q1 maps the point Q2 to the point Q3 . But this interpretation is
impossible! The specification of a point in space requires three numbers, but the
specification of a dilative rotation requires four: one for the expansion, one for the
angle of rotation, and two17 for the direction of the axis of the rotation.
Although we have failed to find a three-dimensional analogue of complex num-
bers, we have discovered the four-dimensional space Q of dilative rotations (centred
at O) of three-dimensional space. Members of Q are called quaternions, and they
may be pictured as points or vectors in four dimensions, but the details of how
to do this will have to wait till Chapter 6. Quaternions can be added by ordinary
vector addition, and they can be multiplied using the non-commutative rule above
(composition of the corresponding dilative rotations).
The discoveries of the rules for multiplying complex numbers and for multiply-
ing quaternions have some interesting parallels. As is well known, the quaternion
rule was discovered in algebraic form by Sir William Rowan Hamilton in 1843. It
is less well known that three years earlier Olinde Rodrigues had published an ele-
gant geometric investigation of the composition of rotations in space that contained

17
To see this, imagine a sphere centred at O. The direction of the axis can be specified by its inter-
section with the sphere, and this point can be specified with two coordinates, e.g., longitude and
latitude.
50 Geometry and Complex Arithmetic

essentially the same result; only much later18 was it recognized that Rodrigues’
geometry was equivalent to Hamilton’s algebra.
Hamilton and Rodrigues are just two examples of hapless mathematicians who
would have been dismayed to examine the unpublished notebooks of the great
Carl Friedrich Gauss. There, like just another log entry in the chronicle of his private
mathematical voyages, Gauss recorded his discovery of the quaternion rule in 1819.
In Chapter 6 we shall investigate quaternion multiplication in detail and find
that it has elegant applications. However, the immediate benefit of this discus-
sion is that we can now see what a remarkable property it is of two-dimensional
space that it is possible to interpret points within it as the fundamental Euclidean
transformations acting on it.

18
See Altmann (1989) for the intriguing details of how this was unravelled.
 Exercises 51

1.5 Exercises

1 The roots of a general cubic equation in X may be viewed (in the XY-plane) as
the intersections of the X-axis with the graph of a cubic of the form,

Y = X3 + AX2 + BX + C.

(i) Show that the point of inflection of the graph occurs at X = − A

3.
A


(ii) Deduce (geometrically) that the substitution X = x − 3 will reduce the
above equation to the form Y = x3 + bx + c.
(iii) Verify this by calculation.
2 In order to solve the cubic equation x3 = 3px + 2q, do the following:
(i) Make the inspired substitution x = s + t, and deduce that x solves the
cubic if st = p and s3 + t3 = 2q.
(ii) Eliminate t between these two equations, thereby obtaining a quadratic
equation in s3 .
(iii) Solve this quadratic to obtain the two possible values of s3 . By symmetry,
what are the possible values of t3 ?
(iv) Given that we know that s3 + t3 = 2q, deduce the formula (1.4).
3 In 1591, more than forty years after the appearance of (1.4), François Viète pub-
lished another method of solving cubics. The method is based on the identity
(see p. 17) cos 3θ = 4C3 − 3C, where C = cos θ.
√
(i) Substitute x = 2 p C into the (reduced) general cubic x3 = 3px + 2q to
q
obtain 4C3 − 3C = p√ p.
(ii) Provided that q2 ⩽ p3 , deduce that the solutions of the original equation
are

√  
x = 2 p cos 13 (ϕ + 2mπ) ,

√
where m is an integer and ϕ = cos−1 (q/p p).
(iii) Check that this formula gives the correct solutions of x3 = 3x, namely,
√
x = 0, ± 3.
4 Here is a basic fact about integers that has many uses in number theory: If two
integers can be expressed as the sum of two squares, then so can their product. With
the understanding that each symbol denotes an integer, this says that if M =
a2 + b2 and N = c2 + d2 , then MN = p2 + q2 . Prove this result by considering
|(a + ib)(c + id)|2 .
52 Geometry and Complex Arithmetic

5 The figure below shows how two similar triangles may be used to construct
the product of two complex numbers. Explain this.

6 (i) If c is a fixed complex number, and R is a fixed real number, explain with a
picture why |z − c| = R is the equation of a circle.
(ii) Given that z satisfies the equation |z + 3 − 4i| = 2, find the minimum and
maximum values of |z|, and the corresponding positions of z.
7 Use a picture to show that if a and b are fixed complex numbers then |z − a| =
|z − b| is the equation of a line.
8 Let L be a straight line in C making an angle ϕ with the real axis, and let d be
its distance from the origin. Show geometrically that if z is any point on L then

d = Im[e−iϕ z] .

[Hint: Interpret e−iϕ using (1.9).]

9 Let A, B, C, D be four points on the unit circle. If A + B + C + D = 0, show that
the points must form a rectangle.
10 Show geometrically that if |z| = 1 then
 
z
Im = 0.
(z + 1)2
Apart from the unit circle, what other points satisfy this equation?
11 Explain geometrically why the locus of z such that
 
z−a
arg = const.
z−b
is an arc of a certain circle passing through the fixed points a and b.
12 By using pictures, find the locus of z for each of the following equations:
   
z−1−i z−1−i
Re = 0, and Im = 0.
z+1+i z+1+i
[Hints: What does Re(W) = 0 imply about the angle of W? Now use the
previous exercise.]
 Exercises 53

13 Find the geometric configuration of the points a, b, and c if

   
b−a a−c
= .
c−a b−c
[Hint: Separately equate the lengths and angles of the two sides.]
14 By considering the product (2 + i) (3 + i), show that
π 1 1
= tan−1 + tan−1 .
4 2 3
15 Draw eiπ/4 , eiπ/2 , and their sum. By expressing each of these numbers in the
form (x + iy), deduce that
3π √
tan = 1 + 2.
8
16 Starting from the origin, go one unit east, then the same length north, then (1/2)
of the previous length west, then (1/3) of the previous length south, then (1/4)
of the previous length east, and so on. What point does this “spiral” converge
to?
17 If z = eiθ ̸= −1, then (z − 1) = (i tan θ2 ) (z + 1). Prove this (i) by calculation,
(ii) with a picture.
18 Prove that
h i i(θ+ϕ) h i i(θ+ϕ)
θ−ϕ θ−ϕ
eiθ + eiϕ = 2 cos 2 e 2 and eiθ − eiϕ = 2i sin 2 e 2

(i) by calculation, and (ii) with a picture.

19 The “centroid” G of a triangle T is the intersection of its medians. If the
vertices are the complex numbers a, b, and c, then you may assume that

On the sides of T we have constructed three similar triangles [dotted] of arbi-

trary shape, so producing a new triangle [dashed] with vertices p, q, r. Using
complex algebra, show that the centroid of the new triangle is in exactly the
same place as the centroid of the old triangle!
54 Geometry and Complex Arithmetic

20 Gaussian integers are complex numbers of the form m + in, where m and n are
integers—they are the grid points in [1.1]. Show that it is impossible to draw an
equilateral triangle such that all three vertices are Gaussian integers. [Hints: You
may assume that one of the vertices is at the origin; try a proof by contradiction;
if a triangle is equilateral, you can rotate one side into another; remember that
√
3 is irrational.]
21 Make a copy of [1.12a], draw in the diagonal of the quadrilateral shown in
[1.12b], and mark its midpoint m. As in [1.12b], draw the line-segments con-
necting m to p, q, r, and s. According to the result in [1.12b], what happens
to p and to r under a rotation of (π/2) about m? So what happens to the
line-segment pr? Deduce the result shown in [1.12a].
22 Will the result in [1.12a] survive if the squares are instead constructed on the
inside of the quadrilateral?
23 Draw an arbitrary triangle, and on each side draw an equilateral triangle lying
outside the given triangle. What do you suspect is special about the new trian-
gle formed by joining the centroids (cf. Ex. 19) of the equilateral triangles? Use
complex algebra to prove that you are right. What happens if the equilateral
triangles are instead drawn on the inside of the given triangle?
24 From (1.15), we know that
zn − 1
1 + z + z2 + · · · + zn−1 = .
z−1
(i) In what region of C must z lie in order that the infinite series 1 + z + z2 + · · ·
converges?
(ii) If z lies in this region, to which point in the plane does the infinite series
converge?
(iii) In the spirit of figure [1.9], draw a large, accurate picture of the infinite
series in the case z = 12 (1 + i), and check that it does indeed converge to
the point predicted by part (ii).
25 Let S = cos θ + cos 3θ + cos 5θ + · · · + cos(2n − 1)θ. Show that
sin 2nθ sin nθ cos nθ
S= or equivalently S= .
2 sin θ sin θ
[Hint: Use Ex. 24, then Ex. 18 to simplify the result.]
26 (i) By considering (a + ib)(cos θ + i sin θ), show that
p  
b cos θ + a sin θ = a2 + b2 sin θ + tan−1 (b/a) .

(ii) Use this result to prove (1.14) by the method of induction.

 Exercises 55

27 Show that the polar equation of the spiral Z(t) = eat eibt in [1.15b] is r =
e(a/b)θ .
28 Reconsider the spiral Z(t) = eat eibt in [1.15b], where a and b are fixed real
numbers. Let τ be a variable real number. According to (1.9), z ÞÑ Fτ (z) =


eaτ eibτ z is an expansion of the plane by factor eaτ , combined with a rotation
of the plane through angle bτ.
(i) Show that Fτ [Z(t)] = Z(t + τ), and deduce that the spiral is an invariant
curve (cf. p. 42) of the transformations Fτ .
(ii) Use this to give a calculus-free demonstration that all rays from the origin
cut the spiral at the same angle.
(iii) Show that if the spiral is rotated about the origin through an arbitrary
angle, the new spiral is again an invariant curve of each Fτ .
(iv) Argue that the spirals in the previous part are the only invariant curves of
Fτ .
29 (i) If V(t) is the complex velocity of a particle whose orbit is Z(t), and dt is
an infinitesimal moment of time, then V(t) dt is a complex number along
the orbit. Thinking of the integral as the (vector) sum of these movements,
Rt
what is the geometric interpretation of t12 V(t) dt?
1
(ii) Referring to [1.15b], sketch the curve Z(t) = a+ib eat eibt .
(iii) Given the result (1.13), what is the velocity of the particle in the previous
part.
R1
(iv) Combine the previous parts to deduce that 0 eat eibt dt =
 1 
at ibt 1
a+ib e e 0
, and draw in this complex number in your sketch
for part (ii).
(v) Use this to deduce that
Z1
a (ea cos b − 1) + b ea sin b
eat cos bt dt = ,
0 a2 + b2
and
Z1
b (1 − ea cos b) + a ea sin b
eat sin bt dt = .
0 a2 + b 2
30 Given two starting numbers S1 , S2 , let us build up an infinite sequence S1 , S2 ,
S3 , S4 , . . . with this rule: each new number is twice the difference of the previous two.
For example, if S1 = 1 and S2 = 4, we obtain 1, 4, 6, 4, −4, −16, −24, . . .. Our
aim is to find a formula for the nth number Sn .
(i) Our generating rule can be written succinctly as Sn+2 = 2(Sn+1 − Sn ).
Show that Sn = zn will solve this recurrence relation if z2 − 2z + 2 = 0.
56 Geometry and Complex Arithmetic

(ii) Use the quadratic formula to obtain z = 1 ± i, and show that if A and B
are arbitrary complex numbers, Sn = A(1 + i)n + B(1 − i)n is a solution
of the recurrence relation.
(iii) If we want only real solutions of the recurrence relation, show that B = A,
and deduce that Sn = 2 Re[A(1 + i)n ].
(iv) Show that for the above example A = −(1/2)
h − i, and by iwriting this in
√ (n+4)π
polar form deduce that Sn = 2 n/2
5 cos 4 + tan−1 2 .
(v) Check that this formula predicts S34 = 262144, and use a computer to
verify this.
[Note that this method can be applied to any recurrence relation of the form
Sn+2 = pSn+1 + qSn .]
31 With the same recurrence relation as in the previous exercise, use a computer
to generate the first 30 members of the sequence given by S1 = 2 and S2 = 4.
Note the repeating pattern of zeros.
(i) With the same notation as before, show that this sequence corresponds to
A = −i, so that Sn = 2 Re[−i(1 + i)n ].
(ii) Draw a sketch showing the locations of −i(1 + i)n for n = 1 to n = 8, and
hence explain the pattern of zeros.
(iii) Writing A = a + ib, our example corresponds to a = 0. More generally,
explain geometrically why such a repeating pattern of zeros will occur if
and only if (a/b) = 0, ±1 or b = 0.
 
(iv) Show that SS12 = 12 1 − a b , and deduce that a repeating pattern of zeros
will occur if and only if S2 = 2S1 (as in our example), S1 = S2 , S1 = 0, or
S2 = 0.
(v) Use a computer to verify these predictions.
32 The Binomial Theorem says that if n is a positive integer,
n  
X  
n n n!
(a + b)n = an−r br , where =
r r (n − r)! r!
r=0
are the binomial coefficients [not vectors!]. The algebraic reasoning leading to
this result is equally valid if a and b are complex numbers. Use this fact to show
that if n = 2m is even then
       
2m 2m 2m 2m

− + − · · · + (−1) m+1
= 2m sin mπ2 .
1 3 5 2m − 1
33 Consider the equation (z − 1)10 = z10 .
(i) Without attempting to solve the equation, show geometrically that all 9
solutions [why not 10?] must lie on the vertical line, Re (z) = 21 . [Hint:
Ex. 7.]
 Exercises 57

(ii) Dividing both sides by z10 , the equation takes the form w10 = 1, where
w = (z − 1)/z. Hence solve the original equation.
(iii) Express these solutions in the form z = x+iy, and thereby verify the result
in (i). [Hint: To do this neatly, use Ex. 18.]
34 Let S denote the set of 12th roots of unity shown in [1.19], one of which is ξ =
ei(π/6) . Note that ξ is a primitive 12th root of unity, meaning that its powers yield
all the 12th roots of unity: S = {ξ, ξ2 , ξ3 , . . . , ξ12 }.
(i) Find all the primitive 12th roots of unity, and mark them on a copy of [1.19].
(ii) Write down, in the form of (1.16), the factorization of the polynomial
Φ12 (z) whose roots are the primitive 12th roots of unity. [In general, Φn (z)
is the polynomial (with the coefficient of the highest power of z equal
to 1) whose roots are the primitive nth roots of unity; it is called the nth
cyclotomic polynomial.]
(iii) By first multiplying out pairs of factors corresponding to conjugate roots,
show that Φ12 (z) = z4 − z2 + 1.
(iv) By repeating the above steps, show that Φ8 (z) = z4 + 1.
(v) For a general value of n, explain the fact that if ζ is a primitive nth root
of unity, then so is ζ. Deduce that if n > 2 then Φn (z) always has even
degree and real coefficients.
(vi) Show that if p is a prime number then Φp (z) = 1 + z + z2 + · · · + zp−1 .
[Hint: Ex. 24.]
[In these examples it is striking that Φn (z) has integer coefficients. In fact it
can be shown that this is true for every Φn (z)! For more on these fascinating
polynomials, see Stillwell (1994).]
35 Show algebraically that the formula (1.21) is invariant under a translation by k,
i.e., its value does not change if a becomes a + k, b becomes b + k, etc. Deduce
from [1.22a] that the formula always gives the area of the quadrilateral. [Hint:
Remember, (z + z) is always real.]
(θ+ϕ)
36 According to the calculation on p. 20, Rϕ
b ◦ Ra = Rc
θ
, where
aeiϕ (1 − eiθ ) + b(1 − eiϕ )
c= .
1 − ei(θ+ϕ)
Let us check that this c is the same as the one given by the geometric construc-
tion in [1.30b].
(i) Explain why the geometric construction is equivalent to saying that c
satisfies the two conditions
   
c−b 1 c−a
arg = 2 ϕ and arg = − 12 θ.
a−b b−a
58 Geometry and Complex Arithmetic

(ii) Verify that the calculated value of c (given above) satisfies the first of these
conditions by showing that
" #
c−b sin θ2
= (θ+ϕ)
eiϕ/2 . (1.33)
a−b sin 2

iα


[Hint: Use 1 − e = −2i sin(α/2) eiα/2 .]
(iii) In the same way, verify that the second condition is also satisfied.
37 Deduce (1.33) directly from [1.30b]. [Hint: Draw in the altitude through b of
the triangle abc, and express its length first in terms of sin θ2 , then in terms of
sin (θ+ϕ)
2 .]
38 On page 20 we calculated that for any non-zero α, Tv ◦ Rα
0 is a rotation:

Tv ◦ Rα
0 = Rc , where c = v/(1 − e ).
α iα

However, if α = 0 then Tv ◦ Rα
0 = Tv is a translation. Try to reconcile these facts
by considering the behaviour of Rαc in the limit that α tends to zero.

39 A glide reflection is the composition Tv ◦ℜL = ℜL ◦Tv of reflection in a line L and

a translation v in the direction of L. For example, if you walk at a steady pace in
the snow, your tracks can be obtained by repeatedly applying the same glide
reflection to a single footprint. Clearly, a glide reflection is an opposite motion.
(i) Draw a line L, a line-segment AB, the image A eB
e of the segment under ℜL ,
′ ′
and the image A B of AB under the glide reflection Tv ◦ ℜL .
(ii) Suppose you erased L from your picture; by considering the line-segments
AA ′ and BB ′ , show that you can reconstruct L.
(iii) Given any two segments AB and A ′ B ′ of equal length, use the previous
part to construct the glide reflection that maps the former to the latter.
(iv) Deduce that every opposite motion is a glide reflection.
(v) Express a glide reflection as the composition of three reflections.
40 Let L be a line making angle ϕ (or ϕ+π) with the real axis, and let p be the point
on L that is closest to the origin, so that |p| is the distance to the line. Consider
the glide reflection [cf. previous exercise] G = Tv ◦ ℜL , where the translation is
through distance r parallel to L. Let us fix the value of ϕ by writing v = +r eiϕ .
(i) Use a picture to show that p = ±i|p| eiϕ , and explain the geometric signif-
icance of the ±.
(ii) What transformation is represented by the complex function H(z) = z+r?
−ϕ
(iii) Use pictures to explain why G = Tp ◦ Rϕ
0 ◦ H ◦ R0 ◦ T−p .
(iv) Deduce that G(z) = ei2ϕ z + eiϕ (r ± 2i|p|).
 Exercises 59

(v) Hence describe (in geometric terms) the glide reflection represented by
G(z) = i z + 4i. Check your answer by looking at the images of −2, 2i,
and 0.
e
41 Let M(z) be the representation of a general opposite motion as a complex
function.
e
(i) Explain why M(z) e
is a direct motion, and deduce from (1.27) that M(z) =
iα
e z + w, for some α and w.
(ii) Using the previous exercise, deduce that every opposite motion is a glide
reflection.
42 On p. 20 we calculated that if (θ + ϕ) = 2π then
Rϕ
b ◦ Ra = Tv ,
θ
where v = (1 − eiϕ )(b − a).
(i) Let Q = (b − a) be the complex number from the first centre of rotation to
the second. Show algebraically that v has length 2 sin(θ/2) |Q|, and that its


direction makes an angle of π−θ
2 with Q.
(ii) Give direct geometric proofs of these results by redrawing figure [1.30b] in
the case (θ + ϕ) = 2π.
43 On p. 20 we calculated that
Tv ◦ Rα
0 = Rc ,
α
where c = v/(1 − eiα ).
(i) Show algebraically that the complex number from the old centre of rotation
|v|
(the origin) to the new centre of rotation (c) has length 2 sin(α/2) , and that
π−α


its direction makes an angle of 2 with v.
(ii) Representing both Rα 0 and Tv as the composition of two reflections, use the
idea in [1.30b] to give direct, geometric proofs of these results.
44 Just as in [1.13b], a dilative rotation Dr,θ
p centred at an arbitrary point p may be
performed by translating p to the origin, doing Dr,θo , then translating o back to
p. Representing these transformations as complex functions, show that
Dr,θ iθ
p (z) = r e z + v, where v = p(1 − r eiθ ).
Conversely, if v is given, deduce that
Tv ◦ Dr,θ
o = Dp ,
r,θ
where p = v/(1 − r eiθ ).
45 In the previous exercise you showed that an arbitrary dilative rotation or trans-
lation can be written as a complex function of the form f(z) = az + b, and,
conversely, that every such function represents a unique dilative rotation or
translation.
(i) Given two pairs of distinct points {A, B} and {A ′ , B ′ }, show [by finding
them explicitly] that a and b exist such that f(A) = A ′ and f(B) = B ′ .
(ii) Deduce the result (1.31).
CHAPTER 2

Complex Functions as Transformations

2.1 Introduction

A complex function f is a rule that assigns to a complex number z an image

complex number w = f(z). In order to investigate such functions it is essential
that we be able to visualize them. Several methods exist for doing this, but (until
Chapter 10) we shall focus almost exclusively on the method introduced in the
previous chapter. That is, we shall view z and its image w as points in the complex
plane, so that f becomes a transformation of the plane.
Conventionally, the image points w are drawn on a fresh copy of C, called the
image plane or the w-plane. This convention is illustrated in [2.1], which depicts the
√
transformation z ÞÑ w = f(z) = (1 + i 3)z = 2 eiπ/3 z (cf. figure [1.5], p. 9).
Usually, the real and imaginary parts of z are denoted x and y, and those of the
image point w are denoted u and v, so that w = f(z) = u(z) + iv(z), where u(z)
and v(z) are real functions of z. The precise forms of these functions will depend on
whether we describe z with Cartesian or polar coordinates. For instance, writing
z = x + iy in the above example yields
√ √
u(x + iy) = x − 3 y and v(x + iy) = 3 x + y,
√
while writing z = r eiθ and (1 + i 3) = 2eiπ/3 yields
   
u(r eiθ ) = 2r cos θ + π3 and v(r eiθ ) = 2r sin θ + π3 .

Of course we may also describe the w-plane with polar coordinates so that
w = f(z) = R eiϕ , where R(z) and ϕ(z) are real functions of z. With the same
example as before, the transformation becomes

R(r eiθ ) = 2r and ϕ(r eiθ ) = θ + π3 .

Visual Complex Analysis. 25th Anniversary Edition. Tristan Needham, Oxford University Press.
© Tristan Needham (2023). DOI: 10.1093/oso/9780192868916.003.0002
62 Complex Functions as Transformations

[2.1] A complex function f(z) is most commonly viewed as a mapping of points in one
copy of C (the “z-plane”)
√ to points in another copy of C (the “w-plane”): z ÞÑ w = f(z).
Here, f(z) = (1 + i 3)z = 2 eiπ/3 z is the combined effect of an expansion by factor 2
and a rotation by angle (π/3). Most commonly, we write z = x + iy and w = u + iv.

We shall find that we can gain considerable insight into a given f by drawing
pictures showing its effect on points, curves, and shapes. However, it would be
nice if we could simultaneously grasp the behaviour of f for all values of z. One
such method is to instead represent f as a vector field, whereby f(z) is depicted as a
vector emanating from the point z; for more detail, the reader is invited to read the
beginning of Chapter 10.
Yet other methods are based on the idea of a graph. In the case of a real function
f(x) of a real variable x we are accustomed to the convenience of visualizing the
overall behaviour of f by means of its graph, i.e., the curve in the two-dimensional
xy-plane made up of the points (x, f(x)). In the case of a complex function this
approach does not seem viable because to depict the pair of complex numbers
(z, f(z)) we would need four dimensions: two for z = x+iy and two for f(z) = u+iv.
Actually, the situation is not quite as hopeless as it seems. First, note that
although two-dimensional space is needed to draw the graph of a real function f,
the graph itself [the set of points (x, f(x))] is only a one-dimensional curve, mean-
ing that only one real number (namely x) is needed to identify each point within it.
Likewise, although four-dimensional space is needed to draw the set of points with
coordinates (x, y, u, v) = (z, f(z)), the graph itself is two-dimensional, meaning that
only two real numbers (namely x and y) are needed to identify each point within
it. Thus, intrinsically, the graph of a complex function is merely a two-dimensional
surface (a so-called Riemann surface), and it is thus susceptible to visualization in
ordinary three-dimensional space. This approach will not be explored in this book,
though the last three chapters in particular should prove helpful in understand-
ing Riemann’s original physical insights, as expounded by Klein (1881). See also
 Introduction 63

Springer (1981, Ch. 1), which essentially reproduces Klein’s monograph, but with
additional helpful commentary.
There is another type of graph of a complex function that is sometimes useful.
The image f(z) of a point z may be described by its distance |f(z)| from the origin,
and the angle arg[f(z)] it makes with the real axis. Let us discard half of this infor-
mation (the angle) and try to depict how the modulus |f(z)| varies with z. To do so,
imagine the complex z-plane lying horizontally in space, and construct a point at
height |f(z)| vertically above each point z in the plane, thereby producing a surface
called the modular surface of f . Figure [2.2] illustrates the conical modular surface of
f(z) = z, while [2.3] illustrates the paraboloid modular surface of f(z) = z2 .
A note on computers. Beginning in this chapter, we will often suggest that you
use a computer to expand your understanding of the mathematical phenomenon

[2.2] The height above C of the modular surface of a complex function f(z) is its modulus,
|f(z)|. Here, f(z) = z, producing a cone.

[2.3] The paraboloid modular surface of f(z) = z2 .

64 Complex Functions as Transformations

under discussion. However, we wish to stress that the specific uses of the computer
that we have suggested in the text are only a beginning. Think of the computer as a
physicist would his laboratory—you may use it to check your existing ideas about
the construction of the world, or as a tool for discovering new phenomena which
then demand new ideas for their explanation. In the Preface we make concrete
suggestions (probably of only fleeting relevance) as to how your laboratory should
be equipped.

2.2 Polynomials

2.2.1 Positive Integer Powers

Consider the mapping z ÞÑ w = zn , where n is a positive integer. Writing z =
r eiθ this becomes w = rn einθ , i.e., the distance is raised to the nth power and the
angle is multiplied by n. Figure [2.4] is intended to make this a little more vivid by
showing the effect of the mapping on some rays and arcs of origin-centred circles.
As you can see, here n = 3.
On page 29 we saw that the n solutions of zn = 1 are the vertices of the regular
n-gon inscribed in the unit circle, with one vertex at 1. This can be understood
more vividly from our new transformational point of view. If w = f(z) = zn then
the solutions of zn = 1 are the points in the z-plane that are mapped by f to the
point w = 1 in the w-plane. Now imagine a particle in orbit round the unit circle in
the z-plane. Since 1n = 1, the image particle w = f(z) will also orbit round the unit
circle (in the w-plane), but with n times the angular speed of the original particle. Thus
each time z executes (1/n) of a revolution, w will execute a complete revolution
and return to the same image point. The preimages of any given w on the unit

[2.4] The geometric effect of the power mapping, z ÞÑ zn .

 Polynomials 65

[2.5] As z rotates around the unit circle, starting at 1, its image under z ÞÑ z3 rotates
three times as fast, executing a complete revolution every time z executes one third of
a revolution. Thus the cube roots of 1 form an equilateral triangle. More generally, the
same reasoning explains why the nth roots of unity form a regular n-gon.

[2.6] Extending the reasoning of the previous figure, the cube roots of an arbitrary complex
number c will form an equilateral triangle, and the nth roots will form a regular n-gon.

circle will therefore be successive positions of z as it repeatedly executes (1/n) of a

revolution, i.e., they will be the vertices of a regular n-gon. With w = 1, figure [2.5]
illustrates this idea for the mapping w = f(z) = z3 .
More generally, [2.6] shows how to solve z3 = c = R eiϕ by inscribing an equi-
√
lateral triangle in the circle |z| = 3 R. By the same reasoning, it is clear that the
solutions of zn = c are the vertices of the regular n-gon inscribed in the circle
√
|z| = n R, with one vertex at angle (ϕ/n).
To arrive at the same result symbolically, first note that if ϕ is one value of arg c,
then the complete set of possible angles is (ϕ + 2mπ), where m is an arbitrary
integer. Setting z = r eiθ ,

rn einθ = zn = c = R ei(ϕ+2mπ) =⇒ rn = R and nθ = ϕ + 2mπ,

√
so the solutions are zm = n R ei(ϕ+2mπ)/n . Each time we increase m by 1, zm is
2πi
rotated by (1/n) of a revolution (because zm+1 = e n zm ), producing the vertices
66 Complex Functions as Transformations

of a regular n-gon. Thus the complete set of solutions will be obtained if we let m
take any n consecutive values, say m = 0, 1, 2, . . . , (n − 1).

2.2.2 Cubics Revisited*

As an instructive application of these ideas, let us reconsider the problem of solv-
ing a cubic equation in x. For simplicity, we shall assume in the following that the
coefficients of the cubic are all real.
In the previous chapter we saw [Ex. 1] that the general cubic could always be
reduced to the form x3 = 3px + 2q. We then found [Ex. 2] that this could be solved
using Cardano’s formula,
p p
x = s + t, where s3 = q + q2 − p3 , t3 = q − q2 − p3 , and st = p.
Once again, observe that if q2 < p3 then this formula involves complex numbers.
On the other hand, we also saw [Ex. 3] that the cubic could be solved using
Viète’s formula:
√  
if q2 ⩽ p3 , then x = 2 p cos 13 (ϕ + 2mπ) ,
√
where m is an integer and ϕ = cos−1 (q/p p). At the time of its discovery, Viète’s
“angle trisection” method was a breakthrough, because it solved the cubic (using
only real numbers) precisely when Cardano’s formula involved “impossible”,
complex numbers. For a long time thereafter, Viète’s method was thought to be
entirely different from Cardano’s, and it is sometimes presented in this way even
today. We shall now take a closer look at these two methods and see that they are
really the same.
If q2 ⩽ p3 , then in Cardano’s formula s3 and t3 are complex conjugates:
p p
s3 = q + i p 3 − q 2 and t 3 = s3 = q − i p 3 − q 2 .
These complex numbers are illustrated on the RHS of [2.7]. By Pythagoras’
√
Theorem, they both have length |s3 | = p p, and so the angle ϕ occurring in Viète’s
formula is simply the angle of s3 .
√
Since s3 and t3 lie on the circle of radius ( p)3 , their preimages under the map-
√
ping z ÞÑ z3 will lie on the circle of radius p. The LHS of [2.7] shows these
preimages; note that the three values of t are the complex conjugates of the three
values of s.
According to the Fundamental Theorem of Algebra, the original cubic should
have three solutions. However, by combining each of the three values of s with
each of the three values of t, it would seem that Cardano’s formula x = s + t yields
nine solutions.
The resolution lies in the fact that we also require st = p. Since p is real, this
means s and t must have equal and opposite angles. In the formula x = s + t, each
 Polynomials 67

[2.7] The solution of the cubic x3 = 3px + 2q was long thought to have two com-
p solutions. If q ⩽ p , Cardano’s 1545 formula says that x = s + s, where
2 3
pletely different
3 3 2
s = q+i p
√  1 − q . On the
 other hand, Viète’s
−1
1591 “angle trisection method” yields
√
x = 2 p cos 3 (ϕ + 2mπ) , where ϕ = cos (q/p p). The figure reveals that the two
methods are in fact the same.

of the three values of s must therefore be paired with the conjugate value of t. We
can now see how Cardano’s formula becomes Viète’s formula:
√  
xm = sm + tm = sm + sm = 2 p cos 13 (ϕ + 2mπ) .
In Ex. 4 the reader is invited to consider the case q2 > p3 .

2.2.3 Cassinian Curves*

Consider [2.8a]. The ends of a piece of string of length l are attached to two fixed
points a1 and a2 in C, and, with its tip at z, a pencil holds the string taut. The figure
illustrates the well known fact that if we move the pencil (continuing to keep the
string taut) it traces out an ellipse, with foci a1 and a2 . Writing r1,2 = |z − a1,2 |, the
equation of the ellipse is thus
r1 + r2 = l.
By choosing different values of l we obtain the illustrated family of confocal
ellipses.
In 1687 Newton published his great Principia, in which he demonstrated that the
planets orbit in such ellipses, with the sun at one of the foci. Seven years earlier,
however, Giovanni Cassini had instead proposed that the orbits were curves for
which the product of the distances is constant:
r1 · r2 = const. = k2 . (2.1)
68 Complex Functions as Transformations

[2.8] [a] Ellipses with foci a1 and a2 have equation r1 + r2 = const. In 1687 Newton
proved that the planets orbit the sun in such ellipses, with the sun at one focus. [b] Seven
years earlier, however, Giovanni Cassini had instead proposed that they orbit the sun in
what are now called Cassinian curves, with equation r1 r2 = const. The figure eight curve
is called the lemniscate, and it played a pivotal role in understanding elliptic integrals and
elliptic functions.

These curves are illustrated in [2.8b]; they are called Cassinian curves, and the
points a1 and a2 are again called foci.
The following facts will become clearer in a moment, but you might like to think
about them for yourself. If k is small then the curve consists of two separate pieces,
resembling small circles centred at a1 and a2 . As k increases, these two compo-
nents of the curve become more egg shaped. When k reaches a value equal to half
the distance between the foci then the pointed ends of the egg shapes meet at the
midpoint of the foci, producing a figure eight [shown solid]. Increasing the value
of k still further, the curve first resembles an hourglass, then an ellipse, and finally
a circle.
Although Cassinian curves turned out to be useless as a description of planetary
motion, the figure eight curve proved extremely valuable in quite another context.
In 1694 it was rediscovered by James Bernoulli and christened the lemniscate—it
then became the catalyst in unravelling the behaviour of the so-called elliptic inte-
grals and elliptic functions. See Stillwell (2010) and Siegel (1969) for more on this
fascinating story.
Cassinian curves arise naturally in the context of complex polynomials. A gen-
eral quadratic Q(z) = z2 + pz + q will have two roots (say, a1 and a2 ) and so can
be factorized as Q(z) = (z − a1 )(z − a2 ). In terms of [2.8b], this becomes

Q(z) = r1 r2 ei(θ1 +θ2 ) .

Therefore, by virtue of (2.1), z ÞÑ w = Q(z) will map each curve in [2.8b] to an

origin-centred circle, |w| = k2 , and it will map the foci to the origin.
 Polynomials 69

If we follow this transformation by a translation of c, i.e., if we change z ÞÑ

Q(z) to z ÞÑ Q(z) + c, then the images will instead be concentric circles centred at
c = (image of foci). Conversely, given any quadratic mapping z ÞÑ w = Q(z), the
preimages of a family of concentric circles in the w-plane centred at c will be the
Cassinian curves whose foci are the preimages of c.
In particular, consider the case c = 1 and w = Q(z) = z2 . The preimages of w = 1
are z = ±1, so these are the foci, and the Cassinian curves are thus centred at the
origin. See [2.9]. Since Q leaves the origin fixed, the lemniscate must be mapped
(as illustrated) to the circle of radius 1 passing through the origin. Writing z =
r eiθ , w = r2 ei2θ , and so we see from the figure that the polar equation of the
lemniscate is
r2 = 2 cos 2θ. (2.2)
Returning to [2.8b], the form of the Cassinian curves may be grasped more intu-
itively by sketching the modular surface of Q(z) = (z−a1 )(z−a2 ). First observe that
as z moves further and further away from the origin, Q(z) behaves more and more
like z2 . Indeed, since the ratio [Q(z)/z2 ] is easily seen [exercise] to tend to unity
as |z| tends to infinity, we may say that Q(z) is ultimately equal to z2 in this limit.
Thus, for large values of |z|, the modular surface of Q will look like the paraboloid
in [2.3].
Next, consider the behaviour of the surface near a1 . Writing D = |a1 − a2 | for the
distance between the foci, we see [exercise] that |Q(z)| is ultimately equal to Dr1 as
z tends to a1 . Thus the surface meets the plane at a1 in a cone like that shown in
[2.2]. Of course the same thing happens at a2 .
Combining these facts, we obtain the surface shown in [2.10]. Since a Cassinian
curve satisfies |Q(z)| = r1 r2 = k2 , it is the intersection of this surface with a plane

[2.9] On the left, the Cassinian curves with foci ±1 have equation |(z−1)(z+1)| = const.
It follows that their images under z ÞÑ z2 are the circles illustrated on the right, centred at
1. It then follows immediately that the equation of the lemniscate is r2 = 2 cos 2θ.
70 Complex Functions as Transformations

[2.10] The form of the Cassinian curves in [2.8b] may be grasped intuitively as horizontal
sections of the modular surface of Q(z) = (z − a1 )(z − a2 ).

parallel to C, and at height k2 above it. As k increases from 0 to a large value, it is

now easy to follow the evolution of the curves in [2.8b] by looking at how this inter-
section varies as the plane moves upward in [2.10]. Thus the Cassinian curves may
be viewed as a geographical contour map of the modular surface of the quadratic.
Interestingly, Cassinian curves were already known to the ancient Greeks.
Around 150 BCE, Perseus considered the intersection curves of a torus [obtained
by rotating a circle C about an exterior line l in its plane] with planes parallel to
l. It turns out that if the distance of the plane from l equals the radius of C then
the resulting spiric section of Perseus is a Cassinian curve. See [2.11]; in particular,
note how the lemniscate [dashed] makes its surprise appearance when the plane
touches the inner rim of the torus. We have adapted this figure from Brieskorn and
Knörrer (2012), to which the reader is referred for more details.
Returning to the complex plane, there is a natural way to define Cassinian curves
with more than two foci: A Cassinian curve with n foci, a1 , a2 , . . . , an , is the locus
of a point for which the product of the distances to the foci remains constant. A
straightforward extension of the above ideas shows that these curves are the preim-
ages of origin-centred circles |w| = const. under the mapping given by the nth
degree polynomial whose roots are the foci:

z ÞÑ w = Pn (z) = (z − a1 )(z − a2 ) · · · (z − an ).

Equivalently, the Cassinian curves are the cross-sections of the modular surface of
Pn (z). This surface has n cone-like legs resting on C at a1 , a2 , . . . , an , and for large
values of |z| it resembles the axially symmetric modular surface of zn .
 Power Series 71

[2.11] The Spiric Sections of Perseus (150 BCE). The intersection of a torus with a plane
whose distance from the axis of symmetry equals the radius of the generating circle turns
out to be a Cassinian curve. If the plane touches the inner rim, the lemniscate makes a
surprise appearance!

2.3 Power Series

2.3.1 The Mystery of Real Power Series

Many real functions F(x) can be expressed (e.g., via Taylor’s Theorem) as power
series:
∞
X
F(x) = cj xj = c0 + c1 x + c2 x2 + c3 x3 + · · · ,
j=0

where the cj ’s are real constants. Of course, this infinite series will normally only
converge to F(x) in some origin-centred interval of convergence −R < x < R. But how
is R (the radius of convergence) determined by F(x)?
It turns out that this question has a beautifully simple answer, but only if
we investigate it in the complex plane. If we instead restrict ourselves to the real
line—as mathematicians were forced to in the era in which such series were first
employed—then the relationship between R and F(x) is utterly mysterious. Histori-
cally, it was precisely this mystery1 that led Cauchy to several of his breakthroughs
in complex analysis.

1
Cauchy was investigating the convergence of series solutions to Kepler’s equation, which describes
where a planet is in its orbit at any given time.
72 Complex Functions as Transformations

To see that there is a mystery, consider the power series representations of the
functions
1 1
G(x) = and H(x) = .
1 − x2 1 + x2
The familiar infinite geometric series,
X ∞
1
= xj = 1 + x + x2 + x3 + · · · if and only if − 1 < x < 1, (2.3)
1−x
j=0

immediately yields
∞
X ∞
X
G(x) = x2j and H(x) = (−1)j x2j ,
j=0 j=0

where both series have the same interval of convergence, −1 < x < 1.
It is easy to understand the interval of convergence of the series for G(x) if
we look at the graph [2.12a]. The series becomes divergent at x = ±1 because
these points are singularities of the function itself, i.e., they are places where |G(x)|
becomes infinite. But if we look at y = |H(x)| in [2.12b], there seems to be no reason
for the series to break down at x = ±1. Yet break down it does.
To begin to understand this, let us expand these functions into power series cen-
P
tred at x = k (instead of x = 0), i.e., into series of the form ∞ j
j=0 cj X , where
X = (x − k) measures the displacement of x from the centre k. To expand G we first
generalize (2.3) by expanding 1/(a − x) about k:
1 1 1 1
= =  X

 ,
a−x a − (X + k) (a − k) 1 − a−k

1
[2.12] [a] The convergence of the power series for G(x) = 1−x 2 is readily understood:

clearly it must end when we arrive at the singularities x = ±1. [b] In contrast to this,
2 also stops converging at ±1 is utterly
1
the fact that the power series for H(x) = 1+x
mysterious!
 Power Series 73

[2.13] The radius of convergence is the distance to the nearest singularity. [a] It is intu-
1
itively clear that the radius of convergence R of the power series for G(x) = 1−x 2 centred

at k is√the distance from k to the nearest singularity. [b] The previously mysterious formula
1
R = 12 + k2 for the convergence of the real power series for H(x) = 1+x 2 is explained

by its imaginary singularities at ±i.

and so
X ∞
1 Xj
= , if and only if |X| < |a − k|. (2.4)
a−x (a − k)j+1
j=0

To apply this result to G, we factorize (1−x2 ) = (1−x) (1+x) and then decompose
G into partial fractions:
  X∞  
1 1 1 1 1 1 1
=2 − =2 − Xj ,
1 − x2 1 − x −1 − x (1 − k)j+1 (−1 − k)j+1
j=0

where |X| < |1 − k| and |X| < |1 + k|. Thus the interval of convergence |X| < R is
given by

R = min {|1 − k|, |1 + k|} = (distance from k to the nearest singularity of G).

This readily comprehensible result is illustrated in [2.13a]; ignore the shaded disc
for the time being.
In the case of H(x), I cannot think of an elegant method of finding the expansion
using only real numbers, but see Ex. 9 for an attempt. Be that as it may, it can be
shown that the radius of convergence of the series in X is given by the strange
√
formula R = 1 + k2 . As with Cotes’ work in the previous chapter, we have here a
result about real functions that is trying to tell us about the existence of the complex
plane.
If we picture the real line as embedded in a plane then Pythagoras’ Theorem
√
tells us that R = 12 + k2 should be interpreted as the distance from the centre k
74 Complex Functions as Transformations

of the expansion to either of the fixed points that lie off the line, one unit from 0 in a
direction at right angles to the line. See [2.13b]. If the plane is thought of as C, then
these points are ±i, and
R = (distance from k to ±i).
The mystery begins to unravel when we turn to the complex function
1
h(z) = ,
1 + z2
which is identical to H(x) when z is restricted to the real axis of the complex plane.
In fact there is a sense—we cannot be explicit yet—in which h(z) is the only complex
function that agrees with H on this line.
While [2.12b] shows that h(z) is well-behaved for real values of z, it is clear that
h(z) has two singularities in the complex plane, one at z = i and the other at z = −i;
these are shown as little explosions in [2.13b]. Figure [2.14] tries to make this more
vivid by showing the modular surface of h(z), the singularities at ±i appearing
as “volcanoes” erupting above these points. We will sort through the details in a
moment, but the mystery has all but disappeared: in both [2.13a] and [2.13b], the
radius of convergence is the distance to the nearest singularity.
If we intersect the surface in [2.14] with a vertical plane through the real axis
then we recover the deceptively tranquil graph in [2.12b], but if we instead slice

1
[2.14] Twin volcanoes erupt in the modular surface of h(z) = 1+z 2 , directly above its

singularities at z = ±i. The graphs in [2.12] can now be understood in a unified way, as
the vertical sections of this surface in the imaginary and real directions. The mysterious
radius of convergence of the real power series for h(x) is thereby explained!
 Power Series 75

the surface along the imaginary axis then we obtain the graph in [2.12a]. That this
is no accident may be seen by first noting that G(x) is just the restriction to the real
axis of the complex function g(z) = 1/(1 − z2 ). Since g(z) = h(iz), h and g are
essentially the same: if we rotate the plane by (π/2) and then do h, we obtain g. In
particular the modular surface of g is simply [2.14] rotated by (π/2), the volcanoes
at ±i being rotated to ±1.

2.3.2 The Disc of Convergence

Let us consider the convergence of complex power series, leaving aside for the
moment the question of whether a given complex function can be expressed as
such a series.
A complex power series P(z) (centred at the origin) is an expression of the form
∞
X
P(z) = cj zj = c0 + c1 z + c2 z2 + c3 z3 + · · · , (2.5)
j=0

where the cj ’s are complex constants, and z is a complex variable. The partial sums
of this infinite series are just the ordinary polynomials,
X
n
Pn (z) = cj zj = c0 + c1 z + c2 z2 + c3 z3 + · · · + cn zn .
j=0

For a given value of z = a, the sequence of points P1 (a), P2 (a), P3 (a), . . . is said to
converge to the point A if for any given positive number ϵ, no matter how small, there
exists a positive integer N such that |A − Pn (a)| < ϵ for every value of n greater
than N. Figure [2.15a] illustrates that this is much simpler than it sounds: all it says
is that once we reach a certain point PN (a) in the sequence P1 (a), P2 (a), P3 (a), . . .,
all of the subsequent points lie within an arbitrarily small disc of radius ϵ centred
at A.
In this case we say that the power series P(z) converges to A at z = a, and we write
P(a) = A. If the sequence P1 (a), P2 (a), P3 (a), . . . does not converge to a particular
point, then the power series P(z) is said to diverge at z = a. Thus for each point z,
P(z) will either converge or diverge.
Figure [2.15b] shows a magnified view of the disc in [2.15a]. If n > m > N then
Pm (a) and Pn (a) both lie within this disc, and consequently the distance between
them must be less than the diameter of the disc:

|cm+1 am+1 + cm+2 am+2 + · · · + cn an | = |Pn (a) − Pm (a)| < 2ϵ. (2.6)

Conversely, it can be shown [exercise] that if this condition is met then P(a) con-
verges. Thus we have a new way of phrasing the definition of convergence: P(a)
converges if and only if there exists an N such that inequality (2.6) holds (for arbitrarily
small ϵ) whenever m and n are both greater than N.
76 Complex Functions as Transformations

[2.15] [a] The sequence Pn (a) is said to converge to A if it inevitably enters every disc
centred at A, no matter how small we make its radius ϵ, and the disc then acts like a
black hole: having entered the disc, the sequence can never escape it! [b] The direct
route from Pm (a) to Pn (a) has length |Pn (a) − Pm (a)| and is shorter than the indirect
en (a) − P
route that goes via Pm+1 (a), Pm+2 (a), etc., which has length P em (a). But if P(a)
is absolutely convergent then the indirect route must have a length that tends to zero for
sufficiently large m and n, and therefore the direct route must, too. Therefore, absolute
convergence implies convergence.

The complex power series P(z) is said to be absolutely convergent at z = a if the

real series
∞
X
e
P(z) ≡ |cj zj | = |c0 | + |c1 z| + |c2 z2 | + |c3 z3 | + · · · ,
j=0

converges there. Absolute convergence is certainly different from ordinary conver-

P j
gence. For example, [exercise] P(z) = z /j is convergent at z = −1, but it is not
absolutely convergent there. On the other hand,
If P(z) is absolutely convergent at some point, then it will also be con-
(2.7)
vergent at that point.
Thus absolute convergence is a stronger requirement than convergence.
To establish (2.7), suppose that P(z) is absolutely convergent at z = a, so that (by
e
definition) P(a) is convergent. In terms of the partial sums P en (z) = Pn |cj zj | of
j=0
e
the real series P(z), this says that for sufficiently large values of m and n we can
make [Pen (a) − Pem (a)] as small as we please. But, referring to [2.15b], we see that

en (a) − P
P em (a) = |cm+1 am+1 | + |cm+2 am+2 | + · · · + |cn an |

is the total length of the roundabout journey from Pm (a) to Pn (a) that goes via
Pm+1 (a), Pm+2 (a), etc. Since |Pn (a) − Pm (a)| is the length of the shortest journey
from Pm (a) to Pn (a),
 Power Series 77

[2.16] [a] If P(z) converges at a then it also converges everywhere inside the shaded disc
|z| < |a|. And if it diverges at d then it also diverges in the striped region |z| > |d|. Its
behaviour is therefore unknown only in “the ring of doubt”, |a| ⩽ |z| ⩽ |d|. Testing a point
q half way across this ring then halves the width of the ring. [b] Iterating this q-algorithm,
the ring of doubt shrinks and converges to a definite circle, the circle of convergence, the
interior of which is the disc of convergence, and the radius of which, R, is the radius of
convergence.

en (a) − P
|Pn (a) − Pm (a)| ⩽ P em (a).

Thus |Pn (a) − Pm (a)| must also become arbitrarily small for sufficiently large m
and n. Done.
We can now establish the following fundamental fact:
If P(z) converges at z = a, then it will also converge everywhere inside
(2.8)
the disc |z| < |a|.
See [2.16a]. In fact we will show that P(z) is absolutely convergent in this disc; the
result then follows directly from (2.7).
If P(a) converges then the length |cn an | of each term must die away to zero
as n goes to infinity [why?]. In particular, there must be a number M such that
|cn an | < M for all n. If |z| < |a| then ρ = |z|/|a| < 1 and so |cn zn | < Mρn . Thus,

P em (z) ⩽ M(ρm+1 + ρm+2 + · · · + ρn ) = M [ρm+1 − ρn+1 ],

en (z) − P (2.9)
1−ρ
where the RHS is as small as we please for sufficiently large m and n. Done.
If P(z) does not converge everywhere in the plane then there must be at least
one point d where it diverges. Now suppose that P(z) were to converge at some
point p further away from the origin than d. See [2.16a]. By (2.8) it would then
converge everywhere inside the disc |z| < |p|, and in particular it would converge
at d, contradicting our initial hypothesis. Thus,
If P(z) diverges at z = d, then it will also diverge everywhere outside
(2.10)
the circle |z| = |d|.
78 Complex Functions as Transformations

At this stage we have settled the question of convergence everywhere except

in the “ring of doubt”, |a| ⩽ |z| ⩽ |d|, shown in [2.16a]. Suppose we take a point
q half way across the ring of doubt (i.e., on the circle |z| = |a|+|d| 2 ), then check
whether P(q) converges or not. Regardless of the outcome, (2.8) and (2.10) enable
us to obtain a new ring of doubt that is half as wide as before. For example, if P(q)
is convergent then P(z) is convergent for |z| < |q|, and the new ring of doubt is
|q| ⩽ |z| ⩽ |d|. Repeating this test procedure in the new ring will again halve its
width. Continuing in this manner, the ring of doubt will narrow to a definite circle
|z| = R (called the circle of convergence) such that P(z) converges everywhere inside
the circle, and diverges everywhere outside the circle. See [2.16b]. The radius R is
called the radius of convergence—at last we see where this name comes from!—and
the interior of the circle is called the disc of convergence.
Note that this argument tells us nothing about the convergence of P(z) on the
circle of convergence. In principle, we can imagine convergence at all, some, or
none of the points on this circle, and one can actually find examples of power series
that realize each of these three possibilities. See Ex. 11.
All of the above results immediately generalize to a power series centred at an
P
arbitrary point k, that is to a series of the form P(z) = cj Zj , where Z = (z − k)
is the complex number from the centre k to the point z. Thus, restating our main
conclusion (due to Niels Abel) in general form,
Given a complex power series P(z) centred at k, there exists a circle
|z − k| = R centred at k such that P(z) converges everywhere inside (2.11)
the circle, and P(z) diverges everywhere outside the circle.
Of course one can also have a series that converges everywhere, but this may be
thought of as the limiting case in which the circle of convergence is infinitely large.
Returning to figures [2.13a] and [2.13b], we now recognize the illustrated discs
as the discs of convergence of the series for 1/(1 ∓ z2 ).

2.3.3 Approximating a Power Series with a Polynomial

Implicit in the definition of convergence is a simple but very important fact: if P(a)
converges, then its value can be approximated by the partial sum Pm (a), and by
choosing a sufficiently large value of m we can make the approximation as accurate
as we wish. Combining this observation with (2.11),
At each point z in the disc of convergence, P(z) can be approximated with
arbitrarily high precision by a polynomial Pm (z) of sufficiently high degree.
For simplicity’s sake, let us investigate this further in the case that P(z) is cen-
tred at the origin. The error Em (z) at z associated with the approximation Pm (z) can
be defined as the distance Em (z) = |P(z) − Pm (z)| between the exact answer and
the approximation. For a fixed value of m, the error Em (z) will vary as z moves
 Power Series 79

around in the disc of convergence. Clearly, since Em (0) = 0, the error will be
extremely small if z is close to the origin, but what if z approaches the circle of
convergence? The answer depends on the particular power series, but it can hap-
pen that the error becomes enormous! [See Ex. 12.] This does not contradict the above
result: for any fixed z, no matter how close to the circle of convergence, the error
Em (z) will become arbitrarily small as m tends to infinity.
This problem is avoided if we restrict z to the disc |z| ⩽ r, where r < R, because
this prevents z from getting arbitrarily close to the circle of convergence, |z| = R.
In attempting to approximate P(z) within this disc, it turns out that we can do the
following. We first decide on the maximum error (say ϵ) that we are willing to
put up with, then choose (once and for all) an approximating polynomial Pm (z) of
sufficiently high degree that the error is smaller than ϵ throughout the disc. That
is, throughout the disc, the approximating point Pm (z) lies less than ϵ away from
the true point, P(z). One describes this by saying that P(z) is uniformly convergent
on this disc:
If P(z) has disc of convergence |z| < R, then P(z) is uniformly convergent
(2.12)
on the closed disc |z| ⩽ r, where r < R.
Although we may not have uniform convergence on the whole disc of conver-
gence, the above result shows that this is really a technicality: we do have uniform
convergence on a disc that almost fills the complete disc of convergence, say
r = (0.999999999) R.
To verify (2.12), first do Ex. 12, then have a good look at (2.9).

2.3.4 Uniqueness
If a complex function can be expressed as a power series, then it can only be done
so in one way—the power series is unique. This is an immediate consequence of the
Identity Theorem:
If
c0 + c1 z + c2 z2 + c3 z3 + · · · = d0 + d1 z + d2 z2 + d3 z3 + · · ·
for all z in a neighbourhood (no matter how small) of 0, then the power
series are identical: cj = dj .
Putting z = 0 yields c0 = d0 , so they may be cancelled from both sides. Dividing by
z and again putting z = 0 then yields c1 = d1 , and so on. [Although this was easy,
Ex. 13 shows that it is actually rather remarkable.] The result can be strengthened
considerably: If the power series merely agree along a segment of curve (no matter how
small) through 0, or if they agree at every point of an infinite sequence of points that con-
verges to 0, then the series are identical. The verification is essentially the same, only
instead of putting z = 0, we now take the limit as z approaches 0, either along the
segment of curve or through the sequence of points.
80 Complex Functions as Transformations

We can perhaps make greater intuitive sense of these results if we first recall that
a power series can be approximated with arbitrarily high precision by a polynomial
of sufficiently high degree. Given two points in the plane (no matter how close
together) there is a unique line passing through them. Thinking in terms of a graph
y = f(x), this says that a polynomial of degree 1, say f(x) = c0 + c1 x, is uniquely
determined by the images of any two points, no matter how close together. Like-
wise, in the case of degree 2, if we are given three points (no matter how close
together), there is only one parabolic graph y = f(x) = c0 + c1 x + c2 x2 that can be
threaded through them. This idea easily extends to complex functions: there is one,
and only one, complex polynomial of degree n that maps a given set of (n + 1) points to
a given set of (n + 1) image points. The above result may therefore be thought of as
the limiting case in which the number of known points (together with their known
image points) tends to infinity.
Earlier we alluded to a sense in which h(z) = 1/(1 + z2 ) is the only complex
function that agrees with the real function H(x) = 1/(1 + x2 ) on the real line. Yet
clearly we can easily write down infinitely many complex functions that agree with
H(x) in this way. For example,

cos[x2 y] + i sin[y2 ]
g(z) = g(x + iy) = .
ey + x2 ln(e + y4 )

Then in what sense can h(z) be considered the unique generalization of H(x)?
P
We already know that h(z) can be expressed as the power series ∞ j 2j
j=0 (−1) z ,
and this fact yields [exercise] a provisional answer: h(z) is the only complex func-
tion that (i) agrees with H(x) on the real axis, and (ii) can be expressed as a power
series in z. This still does not completely capture the sense in which h(z) is unique,
but it’s a start.
More generally, suppose we are given a real function F(x) that can be expressed
as a power series in x on a (necessarily origin-centred) segment of the real line:
P P∞
F(x) = ∞ j
j=0 cj x . Then the complex power series f(z) =
j
j=0 cj z with the same
coefficients can be used to define the unique complex function f(z) that
(i) agrees with F on the given segment of the real axis, and (ii) can be expressed as
a power series in z.
For example, consider the complex exponential function, written ez , the geometry
P
of which we will discuss in the next section. Since ex = ∞ j
j=0 x /j!,

ez = 1 + z + 1 2
2! z + 1 3
3! z + 1 4
4! z + ··· .

Note that our heuristic, power-series approach to Euler’s formula [Chapter 1] is

starting to look more respectable!
 Power Series 81

2.3.5 Manipulating Power Series

The fact that power series can be approximated with arbitrarily high precision by
polynomials implies [see Ex. 14] that

Two power series with the same centre can be added, multiplied, and divided
(2.13)
in the same way as polynomials.
If the two series P(z) and Q(z) have discs of convergence D1 and D2 , then the result-
ing series for [P +Q] and PQ will both converge in the smaller of D1 and D2 , though
they may in fact converge within a still larger disc. No such general statement is
possible in the case of (P/Q) = P(1/Q), because the convergence of the series for
(1/Q) is limited not only by the boundary circle of D2 , but also by any points inside
D2 where Q(z) = 0.
Let us illustrate (2.13) with a few examples. Earlier we actually assumed this
result in order to find the series for 1/(1 − z2 ) centred at k. Using the partial fraction
decomposition
1 (1/2) (1/2)
2
= + ,
1−z 1−z 1+z
we obtained two power series for the functions on the RHS, and then assumed that
these power series could be added like two polynomials, by adding the coefficients.
In the special case k = 0 we can check that this procedure works, because we
already know the correct answer for the series centred at the origin:
1
= 1 + z2 + z 4 + z6 + · · · .
1 − z2
Since
1
= 1 + z + z 2 + z3 + z4 + z 5 + · · ·
1−z
1
and = 1 − z + z2 − z3 + z 4 − z5 + · · · ,
1+z
we see that adding the coefficients of these series does indeed yield the correct
series for 1/(1 − z2 ).
Since
  
1 1 1
= ,
1 − z2 1−z 1+z
we can recycle this example to illustrate the correctness of multiplying power series
as if they were polynomials:

[1 + z + z2 + z3 + z4 + z5 + · · · ][1 − z + z2 − z3 + z4 − z5 + · · · ]
= 1 + (1–1) z + (1–1+1) z2 + (1–1+1–1) z3 + (1–1+1–1+1) z4 + · · · ,

which is again the correct series for 1/(1 − z2 ).

82 Complex Functions as Transformations

Next, let’s use (2.13) to find the series for 1/(1 − z)2 :

[1 + z + z2 + z3 + z4 + z5 + · · · ][1 + z + z2 + z3 + z4 + z5 + · · · ]
= 1 + (1+1) z + (1+1+1) z2 + (1+1+1+1) z3 + (1+1+1+1+1) z4 + · · · ,
P
and so (1 − z)−2 = ∞ j
j=0 (j + 1) z .
You may check for yourself that the above series for (1 − z)−1 and (1 − z)−2 are
both special cases of the general Binomial Theorem, which states that if n is any real
number (not just a positive integer), then within the unit disc,
n(n−1) 2 n(n−1)(n−2) 3 n(n−1)(n−2)(n−3) 4
(1 + z)n = 1 + nz + 2! z + 3! z + 4! z + · · · . (2.14)

Historically, this result was one of Newton’s key weapons in developing calculus,
and later it played an equally central role in the work of Euler.
In Exs. 16, 17, 18, we show how manipulation of power series may be used to
demonstrate the Binomial Theorem, first for all negative integers, then for all ratio-
nal powers. Although we shall not discuss it further, the case of an irrational power
ρ may be treated by taking an infinite sequence of rational numbers that converges
to ρ. Later we shall use other methods to establish a still more general version of
(2.14) in which the power n is allowed to be a complex number!
Next we describe how to divide two power series P(z) and Q(z). In order to
P
find the series P(z)/Q(z) = cj zj , one multiplies both sides by Q(z) to obtain
P
P(z) = Q(z) cj zj , and then multiplies the two power series on the right. By the
uniqueness result, the coefficients of this series must equal the known coefficients
of P(z), and this enables one to calculate the cj ’s. An example will make this process
much clearer.
P
In order to find the coefficients cj in the series 1/ez = cj zj , we multiply both
sides by ez to obtain
z 2 z3 z4
1 = [1 + z + + + + · · · ][c0 + c1 z + c2 z2 + c3 z3 + c4 z4 + · · · ]
2! 3! 4!
c c1 c2  2  c0 c1 c2 c3  3
0
= c0 + (c0 + c1 ) z + + + z + + + + z ··· .
2! 1! 0! 3! 2! 1! 0!
By the uniqueness result, we may equate coefficients on both sides to obtain an
infinite set of linear equations:

1 = c0 ,
0 = c0 + c1 ,
0 = c0 /2! + c1 /1! + c2 /0!,
0 = c0 /3! + c1 /2! + c2 /1! + c3 /0!, etc.

Successively solving the first few of these equations [exercise] quickly leads to
the guess cn = (−1)n /n!, which is then easily verified [exercise] by considering
 Power Series 83

the binomial expansion of (1 − 1)m , where m is a positive integer. Thus we find

that

1/ez = 1 − z + 1 2
2! z − 1 3
3! z + 1 4
4! z − 1 5
5! z + · · · = e−z ,

just as with the real function ex .

2.3.6 Finding the Radius of Convergence

P
Given a complex power series P(z) = cj zj , there are several ways of determin-
ing its radius of convergence directly from its coefficients. Since they are formally
identical to the methods used on real series, we merely state them, leaving it to you
to generalize the standard real proofs.
The ratio test says that
cn
R = lim ,
nÑ∞ cn+1
provided this limit exists. For example, if
z2 z3 z4
P(z) = 1 + z + 2
+ 2 + 2 + ··· ,
2 3 4
then
1/n2

1 2
R = lim = lim 1 + n = 1.
nÑ∞ 1/(n + 1)2 nÑ∞

If |cn /cn+1 | tends to infinity then (formally) R = ∞, corresponding to convergence

P∞ j
everywhere in the plane. For example, ez = j=0 z /j! converges everywhere,
because
1/n!
R = lim = lim (n + 1) = ∞.
nÑ∞ 1/(n + 1)! nÑ∞
When the ratio test fails, or becomes difficult to apply, we can often use the root
test, which says that
1
R = lim p ,
nÑ∞ n |c |
n

provided this limit exists. For example, if we first recall [we will discuss this later]
that the real function ex may be written as
 x n
ex = lim 1 + ,
nÑ∞ n
then applying the root test to the series
X∞  j2
j−3
P(z) = zj ,
j
j=1

yields [exercise] R = e3 .
84 Complex Functions as Transformations

On occasion both the ratio and root tests will fail, but there exists a slightly
refined version of the latter which can be shown to work in all cases. It is called
the Cauchy-Hadamard Theorem, and it says that
1
R= p .
lim sup n |cn |
We will not discuss this further since it is not needed in this book.
The above examples of power series were plucked out of thin air, but often our
starting point is a known complex function f(z) which is then expressed as a power
series. The problem of determining R then has a conceptually much more satisfying
answer. Roughly2 ,

If f(z) can be expressed as a power series centred at k, then the radius of

(2.15)
convergence is the distance from k to the nearest singularity of f(z).

Figure [2.17a] illustrates this, the singularities of f(z) being represented as explo-
sions. To understand which functions can be expanded into power series we need
deep results from later in the book, but we are already in a position to verify that
a rational function [the ratio of two polynomials] can be, and that the radius of
convergence for its expansion is given by (2.15).

[2.17] [a] Here, explosions mark the locations of singularities of a function f(z) that we
wish to expand into a power series centred at k, i.e., a series in powers of Z ≡ (z − k).
Then the radius of convergence R = (the distance from k to the nearest singularity.)
[b] Here we see the three discs of convergence that arise from three different choices of
the centre k, assuming that a is the nearest singularity for all three centres.

2
Later [p. 107] we shall have to modify the statement in the case that f(z) is a “multifunction”,
having more than one value for a given value of z.
 Power Series 85

To begin with, reconsider [2.13a] and [2.13b], both of which are examples of
√
(2.15). Recall that in [2.13b] we merely claimed that R = 1 + k2 for the series expan-
sion of h(z) = 1/(1+z2 ) centred at the real point k. We now verify this and explicitly
find the series.
To do so, first note that (2.4) easily generalizes to
X ∞
1 Zj
= , if and only if |Z| < |a − k|, (2.16)
a−z (a − k)j+1
j=0

where a and k are now arbitrary complex numbers, and Z = (z − k) is the complex
number connecting the centre of the expansion to z. The condition |z − k| < |a − k|
for convergence is that z lie in the interior of the circle centred at k and passing
through a. See [2.17b], which also shows the discs of convergence when we instead
choose to expand 1/(a − z) about k1 or k2 . Since the function 1/(a − z) has just one
singularity at z = a, we have verified (2.15) for this particular function.
Earlier we found the expansion of 1/(1 − x2 ) by factorizing the denominator
and using partial fractions. We are now in a position to use exactly the same
approach to find the expansion of h(z) = 1/(1 + z2 ) centred at an arbitrary complex
number k:
 
1 1 1 1 1
= = − .
1 + z2 (z − i)(z + i) 2i −i − z i − z
Applying (2.16) to both terms then yields

∞
X  
1 1 1 1
= − Zj . (2.17)
1 + z2 2i (−i − k)j+1 (i − k)j+1
j=0

The series for 1/(±i − z) converge inside the concentric circles |z − k| = | ± i −

k| centred at k and passing through the points ∓i, which are the singularities of
h(z). But (2.17) will only converge when both these series converge, i.e., in the disc
|z − k| < R where R is the distance from the centre k to the nearest singularity of
h(z). Thus we have confirmed (2.15) for h(z).
In particular, if k is real then (2.17) converges in the disc shown in [2.13b]. If z is
restricted to the real axis then h(z) reduces to the real function 1/(1 + x2 ), and the
expansion of this function into powers of X = (x − k) can be deduced easily from
√ √
(2.17). Since k is now real, |i−k| = 1 + k2 , and we may write (i−k) = 1 + k2 eiϕ ,
where ϕ = arg(i − k) is the appropriate value of tan−1 (−1/k). Thus [exercise]
X∞  
1 sin(j + 1)ϕ
= √ Xj . (2.18)
1 + x2 ( 1 + k 2 )j+1
j=0

Again, we have here a result concerning real functions that would be very difficult
to obtain using only real numbers.
86 Complex Functions as Transformations

The above analysis of 1/(1 + z2 ) can easily be generalized [exercise] to show

that any rational function can be expressed as a power series, with radius of
convergence given by (2.15).

2.3.7 Fourier Series*

On the 21st of December 1807, Joseph Fourier announced to the French Academy
a discovery so remarkable that his distinguished audience found it literally incred-
ible. His claim was that any3 real periodic function F(θ), no matter how capricious
its graph, may be decomposed into a sum of sinusoidal waves of higher and higher
frequency. For simplicity’s sake, let the period be 2π; then the Fourier series is

∞
X
F(θ) = 12 a0 + [an cos nθ + bn sin nθ] ,
n=1

where [see Ex. 20]

Z 2π Z 2π
1 1
an = F(θ) cos nθ dθ and bn = F(θ) sin nθ dθ. (2.19)
π 0 π 0

This optional section is addressed primarily to readers who have already

encountered such series. For those who have not, we hope that this brief discussion
(together with the exercises at the end of the chapter) may whet your appetite for
more on this fascinating subject4 .
In the world of the real numbers there appears to be no possible connection
between the concepts of Fourier series and Taylor series, but when we pass into
the complex realm a beautiful and remarkable fact emerges:

Taylor series and Fourier series of real functions are merely two different ways of
viewing complex power series.

We will explain this cryptic pronouncement by means of an example.

Consider the complex function f(z) = 1/(1 − z). Writing z = r eiθ , one finds
[exercise] that the real and imaginary parts of f(r eiθ ) are given by
   
1 − r cos θ r sin θ
f(r eiθ ) = u(r eiθ ) + i v(r eiθ ) = + i .
1 + r2 − 2r cos θ 1 + r2 − 2r cos θ

Let’s concentrate on just one of these real functions, say v.

3
Later it was found that some restrictions must be placed on F, but they are astonishingly weak.
4
In many areas of mathematics it is hard to find even one really enlightening book, but Fourier
analysis has been blessed with at least two: Lanczos (1966), and Körner (1988).
 Power Series 87

If z moves outward from the origin along a ray θ = const. then v(r eiθ ) becomes
a function of r alone, say Vθ (r). For example,
r
V π4 (r) = √ .
2(1 + r2 ) − 2r
If z instead travels round and round a circle r = const. then v becomes a function
of θ alone, say Ver (θ). For example,

e 1 (θ) = 2 sin θ
V .
2 5 − 4 cos θ
Note that this is a periodic function of θ, with period 2π. The reason is simple
and applies to any V er (θ) arising from a (single-valued) function f(z): each time z
makes a complete revolution and returns to its original position, f(z) travels along
a closed loop and returns to its original position.
Now, to see the unity of Taylor and Fourier series, recall that (within the unit
disc) f(z) = 1/(1 − z) can be expressed as a convergent complex power series:

f(r eiθ ) = 1 + (r eiθ ) + (r eiθ )2 + (r eiθ )3 + (r eiθ )4 + · · ·

= 1 + r(cos θ + i sin θ) + r2 (cos 2θ + i sin 2θ) + r3 (cos 3θ + i sin 3θ) + · · · .

In particular,

v(r eiθ ) = r sin θ + r2 sin 2θ + r3 sin 3θ + r4 sin 4θ + r5 sin 5θ + · · · .

If we put θ = (π/4), we immediately obtain the Taylor series for V π4 (r):

r
√ = V π4 (r) = √1 r + r2 + √1 r3 − √1 r5 − r6 − √1 r7 + √1 r9 + · · · .
2(1 + r2 ) − 2r 2 2 2 2 2

Once again, consider how difficult this would be to obtain using only real numbers.
From this we find, for example, that
 
d98 r
√ = 98!
dr98 2(1 + r2 ) − 2r r=0

e 1 (θ):
If we instead put r = (1/2), we immediately obtain the Fourier series for V
2

2 sin θ e 1 (θ) =
=V 1
2 sin θ + 1
22
sin 2θ + 1
23
sin 3θ + 1
24
sin 4θ + · · · .
5 − 4 cos θ 2

The absence of cosine waves in this series correctly reflects the fact that Ve 1 (θ) is an
2
odd function of θ.
This connection between complex power series and Fourier series is not merely
aesthetically satisfying, it can also be very practical. The conventional derivation
e 1 (θ) requires that we evaluate the tricky integrals in (2.19),
of the Fourier series of V
2
88 Complex Functions as Transformations

whereas we have obtained the result using only simple algebra! Indeed, we can
now use our Fourier series to do integration:
Z 2π  
2 sin θ sin nθ π
dθ = n .
0 5 − 4 cos θ 2
Further examples may be found in Exs. 21, 37, 38.
We end with a premonition of things to come. The coefficients in a Taylor series
may be calculated by differentiation, while those in a Fourier series may be cal-
culated by integration. Since these two types of series are really the same in the
complex plane, this suggests that there exists some hidden connection between dif-
ferentiation and integration that only complex numbers can reveal. Later we shall
see how Cauchy confirmed this idea in spectacular fashion.

2.4 The Exponential Function

2.4.1 Power Series Approach

We have seen that the only complex function expressible as a power series that
generalizes the real function ex to complex values is

ez = 1 + z + 1 2
2! z + 1 3
3! z + 1 4
4! z + ··· ,

which converges everywhere in C. We now investigate the geometric nature of this

function.
Figure [2.18] visualizes the above series as a spiral journey, the angle between
successive legs of the journey being fixed and equal to • = arg(z) = arg(x + iy). In
the special case where this angle is a right angle, we saw in Chapter 1 that the spiral
converges to a point on the unit circle given by Euler’s formula, eiy = cos y+i sin y.
In fact this special spiral enables us to figure out what happens in the case of the
general spiral in [2.18]: for an arbitrary value of z = x + iy, the spiral converges to the
illustrated point at distance ex and at angle y. In other words,

ex+iy = ex eiy .

This is a consequence of the fact that if a and b are arbitrary complex numbers,
then ea eb = ea+b . To verify this we simply multiply the two series:
  
ea eb = 1 + a + 2!1 a2 + 3!1 a3 + · · · 1 + b + 2!1 b2 + 3!1 b3 + · · ·
 2   3 
a + 2ab + b2 a + 3a2 b + 3ab2 + b3
= 1 + (a + b) + + + ···
2! 3!
= 1 + (a + b) + 1
2! (a + b)2 + 1
3! (a + b)3 + · · ·
= ea+b .
 The Exponential Function 89

[2.18] The power series for ez can be visualized as a spiral journey, the angle between
successive legs of the journey being fixed and equal to • = arg(z) = arg(x+iy). In the text
we show that the power series has the property that ea eb = ea+b , so that ex+iy = ex eiy .
Thus the spiral ends at the point that has length ex and angle y, as illustrated.

Here we have left it to you to show that the general term in the penultimate line is
indeed (a + b)n /n!.

2.4.2 The Geometry of the Mapping

Figure [2.19] illustrates the essential features of the mapping z ÞÑ w = ez . Study it
carefully, noting the following facts:

• If z travels upward at a steady speed s, then w rotates about the origin at angular
speed s. After z has travelled a distance of 2π, w returns to its starting position.
Thus the mapping is periodic, with period 2πi.
• If z travels westward at a steady speed, w travels towards the origin, with ever
decreasing speed. Conversely, if z travels eastward at a steady speed, w travels
away from the origin with ever increasing speed.
• Combining the previous two facts, the entire w-plane (with the exception of
w = 0) will be filled by the image of any horizontal strip in the z-plane of
height 2π.
• A line in general position is mapped to a spiral of the type discussed in the
previous chapter.
• Euler’s formula eiy = cos y + i sin y can be interpreted as saying that ez wraps
the imaginary axis round and round the unit circle like a piece of string.
90 Complex Functions as Transformations

[2.19] The geometry of the complex exponential follows from the fact that ex+iy has
length ex and angle y. Thus horizontal lines y = const. map to rays, and vertical lines
x = const. map to circles. Mysteriously, for now, a computer can be used to empirically
confirm that small squares are ultimately mapped to small squares, in the limit that they
shrink and vanish. This is a central mystery of complex analysis, and it will begin to unravel
in Chapter 4.

• The half-plane to the left of the imaginary axis is mapped to the interior of the
unit circle, and the half-plane to the right of the imaginary axis is mapped to the
exterior of the unit circle.
• The images of the small squares closely resemble squares, and (related to this)
any two intersecting lines map to curves that intersect at the same angle as the
lines themselves.

The last of these observations is not intended to be self-evident—in Chapter 4 we

will begin to explore this fundamental property and to see that it is shared by many
other important complex mappings.

2.4.3 Another Approach

The advantage of the power series approach to ez is that it suggests that there is
something unique about this generalization of ex to complex values. The disad-
vantage is the amount of unilluminating algebra needed to decipher the geometric
 The Exponential Function 91

[2.20] [a] The fact that (ex ) ′ = ex implies that the shaded triangle has unit base. It follows
that ynew ≍ (1 + δ) yold as δ vanishes. [b] Now start with x = 0 and yold = 1. Taking
 = (x/n)
δ  and repeatedly moving
 δ to the right, the height is ultimately multiplied by
x n
1+ n x
each time, so ex ≍ 1 + n .

meaning of the series. We now describe a different approach in which the geometry
lies much closer to the surface. The idea is to generalize the real result,
 x n
ex = lim 1 + . (2.20)
nÑ∞ n
Here is one way of understanding (2.20). As we discussed in Chapter 1, f(x) = ex
may be defined by the property f ′ (x) = f(x). Figure [2.20a] interprets this in terms
of the graph of y = f(x). Drawing a tangent at an arbitrary point, the base of the
shaded triangle is always equal to 1. As you see from the figure, it follows that if
the height is yold at some point x, then moving x an infinitesimal distance δ to the
right yields a new height given by

ynew = (1 + δ) yold .

To find the height ex at x, we divide the interval [0, x] into a large number n
of very short intervals of length (x/n). Since the height at x = 0 is 1, the height
at (x/n) will be approximately [1 + (x/n)] · 1, and so the height at 2(x/n) will be
approximately [1 + (x/n)] · [1 + (x/n)] · 1, and so......., and so the height at x =
n(x/n) will be approximately [1 + (x/n)]n . [For clarity’s sake, [2.20b] illustrates
this geometric progression with the small (hence inaccurate) value n = 3.] It is now
plausible that the approximation [1+(x/n)]n becomes more and more accurate as n
tends to infinity, thereby yielding (2.20). Try using a computer to verify empirically
that the accuracy does indeed increase with n.
Generalizing (2.20) to complex values, we may define ez as
 z n
ez = lim 1 + . (2.21)
nÑ∞ n
92 Complex Functions as Transformations

First we should check that this is the same generalization of ex that we obtained
using power series. Using the Binomial Theorem to write down the first few terms
of the nth degree polynomial [1 + (z/n)]n , we get
 z n h z i n(n − 1) h z i2 n(n − 1)(n − 2) h z i3
1+ = 1+n + + + ···
n n 2! n 3! n




1 − n1 2 1 − n1 1 − n2 3
= 1+z+ z + z + ··· ,
2! 3!
which makes it clear that we do recover the original power series as n tends to
infinity.
Next we turn to the geometry of (2.21). In deciphering the power series for ez we
felt free to assume Euler’s formula, because in Chapter 1 we used the power series
to derive that result. However, it would smack of circular reasoning if we were to
assume Euler’s formula while following our new approach to ez , based on (2.21).
Temporarily, we shall therefore revert to our earlier notation and write r∠θ instead
of r eiθ ; the fact we wish to understand is therefore written ex+iy = ex ∠ y.
With n = 6, figure [2.21] uses Ex. 1.5, p. 52, to geometrically construct the suc-
cessive powers of a ≡ [1 + (z/n)] for a specific value of z. [All six shaded triangles
are similar; the two kinds of shading merely help to distinguish one triangle from
the next.] Even with this small value of n, we see empirically that in this particular
case [1 + (z/n)]n is close to ex ∠ y. To understand this mathematically, we will try
to approximate a = [1 + (z/n)] when n is large.

[2.21] Begin with the lightly shaded triangle with vertices 0, 1, and a ≡ [1 + (z/n)], here
illustrated with n = 6. Now successively construct triangles that are all similar to the orig-
inal, as illustrated. We have thereby constructed the successive powers of a, culminating
in an = [1 + (z/n)]n . Our earlier computations demonstrate that this must ultimately
be equal to ex ∠ y, and this concrete example seems to confirm this. But what is the
geometrical explanation?!
 The Exponential Function 93

Let ϵ be a small, ultimately vanishing, complex number. Consider the length

r and angle θ of the number (1 + ϵ) = r∠θ shown in [2.22]. The origin-centred
circular arc [not shown] connecting (1 + ϵ) to the point r on the real axis almost
coincides with the illustrated perpendicular from (1 + ϵ) to the real axis. Thus r is
approximately equal to [1 + Re(ϵ)], and is ultimately equal5 to it as ϵ tends to zero.
Similarly, we see that the angle θ (the illustrated arc of the unit circle) is ultimately
equal to Im(ϵ). Thus

(1 + ϵ) ≍ [1 + Re(ϵ)] ∠ Im(ϵ), as ϵ vanishes.

Now set ϵ = (z/n) = (x + iy)/n. With the same values of z and n as in [2.21],



figure [2.23] shows the approximation b ≡ 1 + n
x
∠ n
y
to a, together with its
successive powers.

[2.22] As ϵ vanishes, we see that (1 + ϵ) ≍ [1 + Re(ϵ)] ∠ Im(ϵ).

[2.23] Geometrical explanation of [2.21]. Taking ϵ = (z/n) = (x
in the previous

+ iy)/n
figure, we may now approximate a ≡ [1 + (z/n)] as b ≡ 1 + n ∠ n , because
x y
 a ≍ b
z n
as
 n goes
to infinity.

 Here we

once again illustrate the case n = 6. Thus, 1 + n ≍
y n x n
1+ n x
∠ n = 1+ n ∠ y ≍ ex ∠ y, and the mystery is solved!

5
Once again, “ ≍ ” denotes Newton’s concept of ultimate equality; see the new Preface.
94 Complex Functions as Transformations

Returning to the general case, the geometry of (2.21) should now be clear. As n
goes to infinity,
h z in h x   y in  x n
1+ ≍ 1+ ∠ = 1+ ∠ y.
n n n n
Thus, in this limit as n goes to infinity, and using (2.20), we deduce that
ex+iy = ex ∠ y,
as was to be shown. In particular, if we put x = 0 then we recover Euler’s formula,
eiy = 1∠ y, and so we are entitled to write ex+iy = ex eiy .
For a slightly different way of looking at (2.21), see Ex. 22.

2.5 Cosine and Sine

2.5.1 Definitions and Identities

In the previous chapter Euler’s formula enabled us to express cosine and sine in
terms of the exponential function evaluated along the imaginary axis:
eix + e−ix eix − e−ix
cos x = and sin x = .
2 2i
Now that we understand the effect of ez on arbitrary points (not merely points on
the imaginary axis), it is natural to extend the definitions of cosine and sine to the
complex functions
eiz + e−iz eiz − e−iz
cos z ≡ and sin z ≡ . (2.22)
2 2i
Of course another way of generalizing cos x and sin x would be via their power
series, discussed in the previous chapter. This leads to the alternative definitions,
z 2 z4 z6 z3 z5 z7
cos z = 1 − + − + ··· , and sin z = z − + − + ··· .
2! 4! 6! 3! 5! 7!
However, by writing down the series for e±iz you can easily check that these two
approaches both yield the same complex functions.
From the definitions (2.22) we see that cos z and sin z have much in common
with their real ancestors. For example, cos(−z) = cos z, and sin(−z) = − sin z.
Also, since ez is periodic with period 2πi, it follows that cos z and sin z are also
periodic, but with period 2π. The meaning of this periodicity will become clearer
when we examine the geometry of the mappings.
Other immediate consequences of (2.22) are the following important generaliza-
tions of Euler’s formula:
eiz = cos z + i sin z and e−iz = cos z − i sin z.
WARNING: cos z and sin z are now complex numbers—they are not the real and
imaginary parts of eiz .
 Cosine and Sine 95

It is not hard to show that all the familiar identities for cos x and sin x continue
to hold for our new complex functions. For example, we still have
cos2 z + sin2 z = (cos z + i sin z)(cos z − i sin z) = eiz e−iz = e0 = 1,
despite the fact that this identity no longer expresses Pythagoras’ Theorem. Simi-
larly, we will show that if a and b are arbitrary complex numbers then
cos(a + b) = cos a cos b − sin a sin b (2.23)
sin(a + b) = sin a cos b + cos a sin b, (2.24)
despite the fact that these identities no longer express the geometric rule for
multiplying points on the unit circle. First,
cos(a + b) + i sin(a + b) = ei(a+b)
= eia eib
= (cos a + i sin a)(cos b + i sin b)
= (cos a cos b − sin a sin b) + i(sin a cos b + cos a sin b),
exactly as in the previous chapter. However, in view of the warning above, we do
not obtain (2.23) and (2.24) simply by equating real and imaginary parts. Instead
[exercise] one first finds the analogous identity for cos(a + b) − i sin(a + b), then
adds it to (or subtracts it from) the one above.

2.5.2 Relation to Hyperbolic Functions

Recall the definitions of the hyperbolic cosine and sine functions:
ex + e−x ex − e−x
cosh x ≡ and sinh x ≡ .
2 2
By interpreting each of these as the average (i.e., midpoint) of ex and ±e−x , it is
easy to obtain the graphs y = cosh x and y = sinh x shown in [2.24a] and [2.24b].
As you probably know, cosh x and sinh x satisfy identities that are remarkably
similar to those satisfied by cos x and sin x, respectively. For example, if r1 and r2
are arbitrary real numbers, then [exercise]
cosh(r1 + r2 ) = cosh r1 cosh r2 + sinh r1 sinh r2 (2.25)
sinh(r1 + r2 ) = sinh r1 cosh r2 + cosh r1 sinh r2 . (2.26)
Nevertheless, [2.24] shows that the actual behaviour of the hyperbolic functions
is quite unlike the circular functions: they are not periodic, and they become arbi-
trarily large as x tends to infinity. It is therefore surprising and pleasing that the
introduction of complex numbers brings about a unification of these two types of
functions.
We begin to see this if we restrict z = iy to the imaginary axis, for then
cos(iy) = cosh y and sin(iy) = i sinh y.
This connection becomes particularly vivid if we consider the modular surface of
sin z. Since | sin z| is ultimately equal to |z| as z approaches the origin, it follows that
96 Complex Functions as Transformations

[2.24] [a] The graph of cosh x may be visualized as the average (i.e., midpoint) of the
graphs of ex and e−x . [b] The graph of sinh x may be visualized in the same way, as the
average of ex and −e−x .

the surface rises above the origin in the form of a cone. Also, | sin(z + π)| = | sin z|,
so there is an identical cone at each multiple of π along the real axis. These are the
only points [exercise] at which the surface hits the plane. Figure [2.25]—which we
have adapted from Markushevich (2005)—shows a portion of the surface. Notice
that this surface also yields the cosh graph, for if we restrict z = (3π/2) + iy to the
line x = (3π/2), for example, then | sin z| = cosh y.
A practical benefit of this unification is that if you can remember (or quickly
derive using Euler’s formula) a trig identity involving cosine and sine, then you can
immediately write down the corresponding identity for the hyperbolic functions.
For example, if we substitute a = ir1 and b = ir2 into (2.23) and (2.24), then we
obtain (2.25) and (2.26).
 Cosine and Sine 97

[2.25] Cutting open the modular surface of the complex function sin z reveals the
previously hidden unity of the circular and hyperbolic functions.

The connection between the circular and hyperbolic functions becomes stronger
still if we generalize the latter to complex functions in the obvious way:
ez + e−z ez − e−z
cosh z ≡ and sinh z ≡ .
2 2
Since we now have
cosh z = cos(iz) and sinh z = −i sin(iz),
the distinction between the two kinds of function has all but evaporated: cosh is the
composition of a rotation through (π/2), followed by cos; also, sinh is the compo-
sition of a rotation through (π/2), followed by sin, followed by a rotation through
−(π/2).

2.5.3 The Geometry of the Mapping

Just as in the real case, sin z = cos(z− π2 ), which means that we may obtain sin from
cos by first translating the plane by −(π/2). It follows from the preceding remarks
that it is sufficient to study just cos z in order to understand all four functions, cos z,
sin z, cosh z, and sinh z. We now consider the geometric nature of the mapping
z ÞÑ w = cos z.
We begin by finding the image of a horizontal line y = −c lying below the real
axis. It is psychologically helpful to picture the line as the orbit of a particle moving
98 Complex Functions as Transformations

[2.26] Understanding the geometry of ez allows us to understand the geometry of cos z.

Consider the illustrated horizontal line z(t) = t − ic (where t is time) traced from left to
right. Then iz travels upward along the illustrated vertical line, and 12 eiz therefore travels
counterclockwise along an origin-centred circle. By the same reasoning, 21 e−iz travels
clockwise along the smaller circle. Finally, cos z is the sum of these two counter-rotating
circular motions, and it traces an ellipse. It turns out that all horizontal lines are mapped
to confocal ellipses, with foci ±1!

eastward at unit speed, whose position at time t is z = t − ic. See [2.26], in which
the line is shown heavy and unbroken. As z traces this line, −z traces the line y = c,
but in the opposite direction. Applying the mapping z ÞÑ iz (which is a rotation of
2 ), the image particles trace the vertical lines x = ±c, again with unit speed and in
π

opposite directions. Finally applying z ÞÑ 21 ez , the image particles orbit with equal
and opposite angular speeds in origin-centred circles of radii 12 e±c .
The image orbit under z ÞÑ w = cos z of the original particle travelling on the
line y = −c is just the sum of these counter-rotating circular motions. This is clearly
some kind of symmetrical oval hitting the real and imaginary axes at a = cosh c
and ib = i sinh c. It is also clear that cos z executes a complete orbit of this oval
with each movement of 2π by z; this is the geometric meaning of the periodicity of
cos z.
I haven’t found a simple geometric explanation, but it’s easy to show symboli-
cally that the oval traced by cos z is a perfect ellipse. Writing w = u + iv, we find
from the figure [exercise] that u = a cos t and v = b sin t, which is the familiar
parametric representation of the ellipse (u/a)2 + (v/b)2 = 1. Furthermore,
p q
a2 − b2 = cosh2 c − sinh2 c = 1,
so the foci are at ±1, independent of which particular horizontal line z travels along.
 Cosine and Sine 99

Try mulling this over. How does the shape of the ellipse change as we vary c?
How do we recover the real cosine function as c tends to zero? What is the orbit
of cos z as z travels eastward along the line y = c, above the real axis? What is the
image of the vertical line x = c under z ÞÑ cosh z? What is the orbit of sin z as z
travels eastward along the line y = c; how does it differ from the orbit of cos z; and
is the resulting variation of | sin z| consistent with the modular surface shown in
[2.25]?
Before reading on, try using the idea in [2.26] to sketch for yourself the image
under z ÞÑ cos z of a vertical line.
As illustrated in [2.27], the answer is a hyperbola. We can show this using the
addition rule (2.23), which yields

u + iv = cos(x + iy) = cos x cosh y − i sin x sinh y.

On a horizontal line, y is constant, so (u/ cosh y)2 + (v/ sinh y)2 = 1, as before. On
a vertical line, x is constant, so (u/ cos x)2 − (v/ sin x)2 = 1, which is the equation
of a hyperbola. Furthermore, since cos2 x + sin2 x = 1, it follows that the foci of the
hyperbola are always ±1, independent of which vertical line is being mapped.
Figure [2.27] tries to make these results more vivid by showing the image of a
grid of horizontal and vertical lines. Note the empirical fact that each small square in

[2.27] Using the same reasoning as the previous figure, vertical lines are mapped by
cos z to hyperbolas, and calculation reveals that these are also confocal, with the same
foci as the ellipses, ±1! Next, recall the reflection property of the conic sections: a ray of
light emitted from a focus is reflected directly towards the other focus by the ellipse, and
directly away from the other focus by the hyperbola, as illustrated. It follows that these
two sets of conic sections are the orthogonal trajectories of each other. Furthermore—and
this is an additional, surprising, and presently mysterious fact—as the size of the squares
in the grid on the left tends to zero, their images under cos z are ultimately squares, too,
as illustrated!
100 Complex Functions as Transformations

the grid is mapped by cos z to an image shape that is again approximately square. This is
the same surprising (and visually pleasing) phenomenon that we observed in the
case of z ÞÑ ez .
We hope your curiosity is piqued—later chapters are devoted to probing this
phenomenon in depth. In the present case of z ÞÑ cos z we can at least give a
mathematical explanation of part of the result, namely, that the sides of the image
“squares” do indeed meet at right angles; in other words, each ellipse cuts each
hyperbola at right angles.
This hinges on the fact that these ellipses and hyperbolas are confocal. To prove
the desired result [exercise], think of each curve as a mirror, then appeal to the
familiar reflection property of the conic sections: a ray of light emitted from a focus
is reflected directly towards the other focus by the ellipse, and directly away from
the other focus by the hyperbola, as illustrated in [2.27].

2.6 Multifunctions

2.6.1 Example: Fractional Powers

Thus far we have considered a complex function f to be a rule that assigns to each
point z (perhaps restricted to lie in some region) a single complex number f(z).
This familiar conception of a function is unduly restrictive. Using examples, we
now discuss how we may broaden the definition of a function to allow f(z) to have
many different values for a single value of z. In this case f is called a “many-valued
function”, or, as we shall prefer, a multifunction.
We have, in effect, already encountered such multifunctions. For example, we
√
know that 3 z has three different values (if z is not zero), so it is a three-valued
multifunction. In greater detail, [2.28] recalls how we can find the three values of
√3 p using the mapping z ÞÑ z3 . Having found one solution a, we can find the other

two (b and c) using the fact that as z = r eiθ orbits round an origin-centred circle,
z3 = r3 ei3θ orbits with three times the angular speed, executing a complete revo-
lution each time z executes one third of a revolution. Put differently, reversing the
direction of the mapping divides the angular speed by three. This is an essential
√
ingredient in understanding the mapping z ÞÑ 3 z, which we will now study in
detail.
√ √ √
Writing z = r eiθ , we have 3 z = 3 r ei(θ/3) . Here 3 r is uniquely defined as the
real cube root of the length of z; the sole source of the three-fold ambiguity in the
formula is the fact that there are infinitely many different choices for the angle θ of
a given point z.
Think of z as a moving point that is initially at z = p. If we arbitrarily choose θ
√
to be the angle ϕ shown in [2.28], then 3 p = a. As z gradually moves away from p,
√ √
θ gradually changes from its initial value ϕ, and 3 z = 3 r ei(θ/3) gradually moves
 Multifunctions 101

[2.28] The cube root is a three-valued multifunction. Since z ÞÑ z3 multiplies angles by

√
3, z ÞÑ 3 z divides angles by 3. Now suppose we start at the illustrated point p at angle
ϕ on the circle of radius R. Then it would seem clear √ that its cube root is the illustrated
√
point p = a at angle (ϕ/3) on the circle of radius 3 R. But now suppose p executes
3
√
a complete revolution and returns to its starting point. Then 3 z rotates at one third the
√
angular speed and arrives at b: a second equally good value of 3 p. If p keeps rotating,
√
executing a second revolution, we arrive at 3 p = c, and a third revolution returns us to
√ √
the original value 3 p = a. Thus the three values of 3 p are a, b, and c, and they form an
equilateral triangle.

away from its initial position a, but in a completely determined way—its distance from
the origin is the cube root of the distance of z, and its angular speed is one third
that of z.
Figure [2.29] illustrates this. Usually we draw mappings going from left to right,
but here we have reversed this convention to facilitate comparison with [2.28].
√
As z travels along the closed loop A (finally returning to p), 3 z travels along
the illustrated closed loop and returns to its original value a. However, if z instead
√
travels along the closed loop B, which goes round the origin once, then 3 z does not
return to its original value but instead ends up at a different cube root of p, namely
b. Note that the detailed shape of B is irrelevant, all that matters is that it encircles
√
the origin once. Similarly, if z travels along C, encircling the origin twice, then 3 z
ends up at c, the third and final cube root of p. Clearly, if z were to travel along a
√
loop [not shown] that encircled the origin three times, then 3 z would return to its
original value a.
√ √
The premise for this picture of z ÞÑ 3 z was the arbitrary choice of 3 p = a, rather
√
than b or c. If we instead chose 3 p = b, then the orbits on the left of [2.29] would
√
simply be rotated by (2π/3). Similarly, if we chose 3 p = c, then the orbits would
be rotated by (4π/3).
√
The point z = 0 is called a branch point of 3 z. More generally, let f(z) be a mul-
tifunction and let a = f(p) be one of its values at some point z = p. Arbitrarily
choosing the initial position of f(z) to be a, we may follow the movement of f(z) as
z travels along a closed loop beginning and ending at p. When z returns to p, f(z)
102 Complex Functions as Transformations

√ √
[2.29] Writing z = r eiθ , we have 3 z = 3 r ei(θ/3) . Once again, let us arbitrarily start at
√3 p = a, and then let p move along the three illustrated paths, A, B, and C. As you see,

which cube root we end up at does not depend on the detailed shape of the path, it only
depends on the net change in θ—how many complete revolutions have been executed
when p returns home. The cube root rotates one third of a revolution each time p loops
around the branch point at the origin.

will either return to a or it will not. A branch point z = q of f is a point such that
f(z) fails to return to a as z travels along any loop that encircles q once.
√
Returning to the specific example f(z) = 3 z, we have seen that if z executes three
revolutions round the branch point at z = 0 then f(z) returns to its original value.
If f(z) were an ordinary, single-valued function then it would return to its original
value after only one revolution. Thus, relative to an ordinary function, two extra
revolutions are needed to restore the original value of f(z). We summarize this by
√
saying that 0 is a branch point of 3 z of order two.
More generally, if q is a branch point of some multifunction f(z), and f(z) first
returns to its original value after N revolutions round q, then q is called an alge-
braic branch point of order (N − 1); an algebraic branch point of order 1 is called
a simple branch point. We should stress that it is perfectly possible that f(z) never
returns to its original value, no matter how many times z travels round q. In this
case q is called a logarithmic branch point—the name will be explained in the next
section.
√
By extending the above discussion of 3 z, check for yourself that if n is an integer
then z(1/n) is an n-valued multifunction whose only (finite) branch point is at z = 0,
the order of this branch point being (n − 1). More generally, the same is true for
any fractional power z(m/n) , where (m/n) is a fraction reduced to lowest terms.
 Multifunctions 103

2.6.2 Single-Valued Branches of a Multifunction

Next we will show how we may extract three ordinary, single-valued functions
√
from the three-valued multifunction 3 z. First, [2.30] introduces some terminology
which we need for describing sets of points in C.
A set S is said to be connected (see [2.30a]) if any two points in S can be con-
nected by an unbroken curve lying entirely within S. Conversely, if there exist
pairs of points that cannot be connected in this way (see [2.30b]), then the set is
disconnected. Amongst connected sets we may single out the simply connected sets
(see [2.30c]) as those that do not have holes in them. More precisely, if we picture
the path connecting two points in the set as an elastic string, then this string may
be continuously deformed into any other path connecting the points, without any
part of the string ever leaving the set. Conversely, if the set does have holes in it
then it is multiply connected (see [2.30d]) and there exist two paths connecting two
points such that one path cannot be deformed into the other.
√
Now let us return to [2.29]. By arbitrarily picking one of the three values of 3 p
√
at z = p, and then allowing z to move, we see that we obtain a unique value of 3 Z
associated with any particular path from p to Z. However, we are still dealing with
a multifunction: by going round the branch point at 0 we can end up at any one of
√
the three possible values of 3 Z.
√
On the other hand, the value of 3 Z does not depend on the detailed shape of the
path: if we continuously deform the path without crossing the branch point then we obtain
√
the same value of 3 Z. This shows us how we may obtain a single-valued function.
If we restrict z to any simply connected set S that contains p but does not contain
√
the branch point, then every path in S from p to Z will yield the same value of 3 Z,

[2.30] Here is the terminology that is used to describe different types of regions in the
plane.
104 Complex Functions as Transformations

[2.31] If we chose a simply connected region S that does not contain the branch point at
0, then no path within S can encircle the branch point at 0, and therefore it is possible to
√ √
define a single-valued branch of 3 z within S. For example, if we arbitrarily take 3 p = a,
√
then let z = p move
√ about within S and arrive at z = Z, say, then 3 z arrives at a unique
3
value f1 (Z) = Z that is independent of the path. But since there are three values of
√ √
3 p, there are three distinct branches of 3 z. Incidentally, note that these three branches

display the ubiquitous (yet mysterious) preservation of small squares.

which we will call f1 (Z). Since the path is irrelevant, f1 is an ordinary, single-valued
√
function of position on S; it is called a branch of the original multifunction 3 z.
Figure [2.31] illustrates such a set S, together with its image under the branch
√
f1 of 3 z. Here we have reverted to our normal practice of depicting the mapping
√
going from left to right. If we instead choose 3 p = b then we obtain a second
√ √
branch f2 of 3 z, while 3 p = c yields the third and final branch f3 . Notice, inci-
dentally, that all three branches display the by now ubiquitous (yet mysterious)
preservation of small squares.
We now describe how we may enlarge the domain S of the branches so as to
obtain the cube roots of any point in the plane. First of all, as illustrated in [2.32], we
draw an arbitrary (but not self-intersecting) curve C from the branch point 0 out to
infinity; this is called a branch cut. Provisionally, we now take S to be the plane with
the points of C removed—this prevents any closed path in S from encircling the
branch point. We thereby obtain on S the three branches f1 , f2 , and f3 . For example,
the figure shows the cube root f1 (d) of d.
What about a point such as e on C? Imagine that z is travelling round an origin-
centred circle through e. The figure illustrates the fact that f1 (z) approaches two
different values according as z arrives at e with positive or negative angular speed.
 Multifunctions 105

√
[2.32] In order to define a single-valued branch of f(z) = 3 z throughout the entire com-
plex plane, we begin by drawing an arbitrary curve C that starts at the branch point
0 and extends out to infinity, as illustrated, and define S to be C with the points of √C
removed. The curve C is called a branch cut. Out of the three possible choices for 3 d,
let us arbitrarily choose the illustrated point f1 (d). Then f1 (Z) is single-valued throughout
S, because it is impossible for z to encircle the branch point at 0 as it travels from d to
Z. But we still need to define f1 (e) at a point such as e that lies on C. By convention, it
is defined to be the value of f1 (z) when z arrives at e travelling counterclockwise, i.e.,
arriving from the right of C. The two other branches are defined in exactly the same way.

If we (arbitrarily) define f1 (e) to be the value of f(z) when z travels counterclockwise

round the circle, then f1 (z) is well defined on the whole plane. Similarly for the
other two branches.
√
Of course the branch cut C is the work of man—the multifunction 3 z is obliv-
ious to our desire to dissect it into three single-valued functions. As we have just
seen, this shows up in the fact that the resulting branches are discontinuous on C,
√
despite the fact that the three values of 3 z always move continuously as z moves
continuously. As z crosses C travelling counterclockwise then we must switch from
√
one branch to the next in order to maintain continuous motion of 3 z: for example,
f1 switches to f2 . If z executes three counterclockwise revolutions round the branch
point, then the branches permute cyclically, each finally returning to itself: using
an arrow to denote a crossing of C,
       

 f 1 
 
 f 2 
 
 f3 
 
 f 1 

f2 Ñ f3 Ñ f1 Ñ f2 .

 f   
 f   
 f   
 f  
3 1 2 3

A common choice for C is the negative real axis. If we do not allow z to cross the
cut then we may restrict the angle θ = arg(z) to lie in the range −π < θ ⩽ π. This is
called the principal value of the argument, written Arg (z); note the capital first letter.
√
With this choice of θ, the single-valued function 3 r ei(θ/3) is called the principal
√
branch of the cube root; let us write it as [ 3 z ]. Note that the principal branch agrees
106 Complex Functions as Transformations

with the real cube root function on the positive real axis, but not on the negative
√
real axis; for example, [ 3 −8 ] = 2 ei(π/3) . Also note that the other two branches
associated with this choice of C can be expressed in terms of the principal branch
√ √
as ei(2π/3) [ 3 z ] and ei(4π/3) [ 3 z ].
It should be clear how the above discussion extends to a general fractional
power.

2.6.3 Relevance to Power Series

Earlier we explained the otherwise mysterious interval of convergence for a real
function such as 1/(1 + x2 ) by extending the function off the real line and into
the complex plane: the obstruction to convergence was the existence of points at
which the complex function became infinite (singularities). We now discuss the
more subtle fact that branch points also act as obstacles to the convergence of power
series.
The real Binomial Theorem says that if n is any real number (not just a positive
integer), then
n(n−1) 2 n(n−1)(n−2) 3 n(n−1)(n−2)(n−3) 4
(1 + x)n = 1 + nx + 2! x + 3! x + 4! x + ··· .
If n is a positive integer then the series terminates at xn and the issue of conver-
gence does not arise. If n is not a positive integer then the ratio test tells us that
the interval of convergence of the power series is −1 < x < 1. This interval is eas-
ily understood when n is negative, because the function then has a singularity at
x = −1. But how, for example, are we to explain this interval of convergence in the
case n = (1/3)?
1 1
Figure [2.33a] shows the graph y = (1 + x) 3 of the real function f(x) = (1 + x) 3 ,
which is well defined for all x since every real number has a unique real cube root.
Looking at this graph, there seems to be no good reason for the series to break
down at ±1, yet break down it does. This is illustrated rather vividly by the dashed
curve, which is the graph of the 30th degree polynomial obtained by truncating the
binomial series at x30 . As you can see, this curve follows y = f(x) very closely
(actually more closely than illustrated) between ±1, but just beyond this interval it
suddenly starts to deviate wildly.
Unlike the case of 1/(1 + x2 ), observe that the mystery does not disappear when
1
we extend the real function f(x) to the complex function f(z) = (1 + z) 3 , because
f(z) does not have any singularities.
We have already discussed the fact [see (2.14) and Exs. 16, 17, 18] that the
Binomial Theorem extends to the complex plane. In the present case it says that
1
f(z) = (1 + z) 3 = 1 + 13 z − 19 z2 + 5 3
81 z − 10 4
243 z + 22 5
729 z − ··· ,
with convergence inside the unit disc shown in [2.33b]. In common with all power
series, the RHS of the above equation is a single-valued function. For example, at
 Multifunctions 107

1
[2.33] [a] The real function f(x) = (1+x) 3 can be expressed as a power series, but only for
−1 < x < 1; the dashed curve shows the 30th degree polynomial obtained by truncating
the binomial series at x30 . But why does convergence break down here?! There are no
1
singularities in sight, and, unlike the resolution of the convergence mystery of 1+x 2 , the
1
complex generalization f(z) = (1 + z) also does not have any singularities! [b] But f(z)
3

does have a branch point at −1. Suppose the series were to converge at the illustrated
point z outside the unit disc. Then the two illustrated paths could encircle the branch
point, yielding two different values of f(z). But the power series is single-valued, so this
is impossible!

z = 0 the series equals 1. But while f(x) was an ordinary single-valued function
of x, the LHS of the above equation is a three-valued multifunction of z, with a
2π
second order branch point at z = −1. For example, f(0) takes three values: 1, ei 3 ,
2π
and e−i 3 . We now recognize that the power series represents just one branch of
f(z), namely the one for which f(0) = 1.
This solves the mystery. For suppose that the series were to converge inside the
larger circle in [2.33b], and in particular at the illustrated point z. Starting at z = 0
with the value f(0) = 1, then travelling along the two illustrated paths to z, we
clearly end up with two different values of f(z), because together the two paths
enclose the branch point at −1. But the power series cannot mimic this behaviour
since it is necessarily single-valued—its only way out is to cease converging outside
the unit disc. We have demanded the impossible of the power series, and it has
responded by committing suicide!
This example shows that a branch point is just as real an obstacle to convergence
as a singularity. Quite generally, this argument shows that if a branch of a multi-
function can be expressed as a power series, the disc of convergence cannot be large
enough to contain any branch points of the multifunction. This strongly suggests
a further generalization of the (unproven) statement (2.15):
If a complex function or a branch of a multifunction can be expressed as
a power series, the radius of convergence is the distance to the nearest (2.27)
singularity or branch point.
108 Complex Functions as Transformations

Much later in the book we will develop the tools necessary to confirm this
conjecture.

2.6.4 An Example with Two Branch Points

Choosing the positive value of the square root, [2.34a] illustrates the graph y =
√
f(x) = 1 + x2 , which is a hyperbola. Again, the Binomial Theorem yields a power
series that mysteriously only converges between ±1, namely,
1
f(x) = (1 + x2 ) 2 = 1 + 12 x2 − 18 x4 + 1 6
16 x − 5 8
128 x + ··· .

The divergence of the series beyond this interval is vividly conveyed by the dashed
curve, which is the graph of the 20th degree polynomial obtained by truncating the
binomial series at x20 .
As before, the explanation lies in C, where f(x) becomes the two-valued mul-
√ p
tifunction f(z) = z2 + 1. This can be rewritten as f(z) = (z − i)(z + i), which
makes it clear that f(z) has two simple branch points, one at i and the other at −i.
These branch points obstruct the convergence of the corresponding complex series,
limiting it to the unit disc shown in [2.34b].
In greater detail, the notation of [2.34b] enables us to write
√
f(z) = r1 r2 ei(θ1 +θ2 )/2 . (2.28)

Here we must bear in mind that the figure illustrates only one possibility (out of
infinitely many) for each of the angles θ1 and θ2 . To see that i is indeed a branch

1
[2.34] [a] The real function f(x) = (1 + x2 ) 2 can be expressed as a power series, but only
for −1 < x < 1; the dashed curve shows the 20th degree polynomialp obtained by trun-
cating the binomial series at x20 . [b] In the complex plane, f(z) = (z − i)(z + i) =
√
r1 r2 ei(θ1 +θ2 )/2 , which makes it clear that f(z) has branch points at ±i, and these
obstruct the convergence of the complex power series beyond the unit disc. If z loops
around i along L, as illustrated, then the net changes in θ1 and θ2 are 2π and 0,
respectively. So f(z) ⇝ eiπ f(z) = −f(z).
 Multifunctions 109

point, suppose we start with the value of f(z) given by the illustrated values of θ1
and θ2 . Now let z travel round the illustrated loop L. As it does so, (z + i) rocks
back and forth, so θ2 merely oscillates, finally returning to its original value. But
(z − i) undergoes a complete revolution, and so θ1 increases by 2π. Thus when z
returns to its original position, (2.28) shows that f(z) does not return to its original
value, but rather to
√ √
f new (z) = r1 r2 ei(θ1 +2π+θ2 )/2 = eiπ r1 r2 ei(θ1 +θ2 )/2 = −fold (z).

Of course the same thing happens if z travels along a loop that goes once round −i,
instead of round +i.
In order to dissect f(z) into two single-valued branches, we appear to need two
branch cuts: one cut C1 from i to infinity (to prevent us encircling the branch point
at i), and another cut C2 from −i to infinity, for the same reason. Figure [2.35a]
illustrates a particularly common and important choice of these cuts, namely, rays
going due west. If we do not allow z to cross the cuts then we may restrict the angle
θ1 = arg(z − i) to its principal value, in the range −π < θ1 ⩽ π. For example, the
angle in [2.34b] is not the principal value, while the one in [2.35a] is. If θ2 is like-
wise restricted to its unique principal value then (2.28) becomes the single-valued
principal branch of f(z), say F(z). The other branch of f(z) is simply −F(z).
Let us return to the previous situation in which we allowed θ1 and θ2 to take
general values rather than their principal values. Figure [2.35b] illustrates the fact
that it is possible to define two branches of f(z) using only a single branch cut
C that connects the two branch points. If z is restricted to the shaded, multiply-
connected region S, then it cannot loop around either branch point singly. It can,

√ √
[2.35] [a] We can define two single-valued branches of f(z) = z2 + 1 = r1 r2 ei(θ1 + θ2 )/2
by making branch cuts going due west from both branch points, and then taking θ1 and
θ2 to be their principal values, as illustrated. [b] In fact we only need a single branch cut
connecting the two branch points, such as C. For then a loop L cannot encircle either
branch point singly, and if it encircles both, as illustrated, then f(z) does not change its
value.
110 Complex Functions as Transformations

however, travel along a loop such as L that encircles both branch points together.
But in this case both θ1 and θ2 increase by 2π, so (2.28) shows that f(z) returns to
its original value. Thus we can define two single-valued branches on S. Finally, we
may expand S until it borders on C.

2.7 The Logarithm Function

2.7.1 Inverse of the Exponential Function

The complex logarithm function log(z) may be introduced as the “inverse” of ez .
More precisely, we define log(z) to be any complex number such that elog(z) = z. It
follows [exercise] that

log(z) = ln |z| + i arg(z).

Since arg(z) takes infinitely many values, differing from each other by multiples
of 2π, we see that log(z) is a multifunction taking infinitely many values, differing
from each other by multiples of 2πi. For example,
√
log(2 + 2i) = ln 2 2 + i(π/4) + 2nπi,

where n is an arbitrary integer.

The reason we get infinitely many values is clear if we go back to the exponen-
tial mapping shown in [2.19], p. 90: each time z travels straight upward by 2πi,
ez executes a complete revolution and returns to its original value. Figure [2.36]
rephrases this using the above example of log(2 + 2i). If we arbitrarily choose the
√
initial value w = ln 2 2 + i (π/4) for log(2 + 2i), then as z travels along a loop that
encircles the origin ν times in the counterclockwise direction, log(z) moves along
a path from w to w + 2νπi. Check that you understand (roughly) the shapes of the
illustrated image paths.
Clearly log(z) has a branch point at z = 0. However, this branch point is quite
unlike that of z(1/n) , for no matter how many times we loop around the origin
(say counterclockwise), log(z) never returns to its original value, rather it continues
moving upwards forever. You can now understand the previously introduced term,
“logarithmic branch point”.
Here is another difference between the branch points of z(1/n) and log(z). As z
approaches the origin, say along a ray, |z(1/n) | tends to zero, but | log(z)| tends to
infinity, and in this sense the origin is a singularity as a well as a branch point.
On the other hand, algebraic branch points can also be singularities; consider
√
(1/ z).
To define single-valued branches of log(z) we make a branch cut from 0 out
to infinity. The most common choice for this cut is the negative real axis. In this
cut plane we may restrict arg(z) to its principal value Arg (z); remember, this is
 The Logarithm Function 111

[2.36] Here we see that 0 is a branch point of log(z). However, this branch point is
quite unlike that of z(1/n) , for no matter how many times we loop around the origin (say
counterclockwise), log(z) never returns to its original value, rather it continues moving
upwards forever. If we make a branch cut along the negative real axis, we can restrict
the angle to be its principal value, in which case we obtain the principal value of log(z),
written Log (z) ≡ ln |z| + i Arg (z).

defined by −π < Arg (z) ⩽ π. This yields the principal branch or principal value of
the logarithm, written Log (z), and defined by
Log (z) ≡ ln |z| + i Arg (z).
√
For example, Log (− 3 − i) = ln 2 − i (5π/6), Log (i) = i(π/2), and Log (−1) = iπ.
Note that if z = x is on the positive real axis, Log (x) = ln(x).
Figure [2.37] illustrates how the mapping z ÞÑ w = Log (z) sends rays to hor-
izontal lines, and circles to vertical line-segments connecting the horizontal lines
at heights ±π; the entire z-plane is mapped to the horizontal strip of the w-plane
bounded by these lines. Study this figure until you are completely at peace with
it. You can see the price we pay for forcing the logarithm to be single-valued: it
becomes discontinuous at the cut. As z crosses the cut travelling counterclock-
wise, the height of w suddenly jumps from π to −π. If we wish w to instead move
continuously, then we must switch to the branch Log (z) + 2πi of the logarithm.
Another problem with restricting ourselves to the principal branch is that the
familiar rules for the logarithm break down. For example, Log (ab) is not always
equal to Log (a) + Log (b); try a = −1 and b = i, for example. However, if we keep
all values of the logarithm in play then it is true [exercise] that
log(ab) = log(a) + log(b) and log(a/b) = log(a) − log(b),
112 Complex Functions as Transformations

[2.37] The geometry of Log (z). Note that the price we pay for having a single-valued
logarithm is that it is discontinuous: as we cross the branch cut, its value jumps.

in the sense that every value of the LHS is contained amongst the values of the RHS,
and vice versa.

2.7.2 The Logarithmic Power Series

If we wish to find a power series for the complex logarithm, two problems imme-
diately arise. First, since a power series is single-valued, the best we can hope for
is to represent a single branch of log(z); let’s choose the principal branch, Log (z).
Second, the origin is both a singularity and a branch point of Log (z) so we cannot
have a power series centred there (i.e., in powers of z); let us therefore try an expan-
sion centred at z = 1, i.e., in powers of (z − 1). [Of course any other non-zero point
would be equally suitable.] Writing Z = (z − 1), our problem, then, is to expand
Log (1 + Z) in powers of Z.
Let us use the abbreviation L(z) = Log (1 + z). Since the branch point of L(z) is
z = −1, the largest disc of convergence we can have is the unit disc. To find the
series we will use the fact that eL(z) = (1 + z). Recall from (2.21) [on p. 91] that
by taking n to be a sufficiently large positive integer, we can approximate eL as
precisely as we wish using [1 + (L/n)]n . Thus, as (1/n) vanishes,6
 n
L L 1
1+ ≍ eL = (1 + z) =⇒ 1 + ≍ (1 + z) n .
n n

6
Once again, “ ≍ ” denotes Newton’s concept of ultimate equality; see the new Preface.
 The Logarithm Function 113

1
There are n branches of (1 + z) n within the unit disc, but since L(0) = 0 we need
1
the branch of (1 + z) n that equals 1 when z = 0. Appealing to the Binomial series
for this principal branch, we obtain
1 1 1 1
L 1 n(n − 1) 2 n(n − 1)( n1 − 2) 3
1+ ≍ 1+ z+ z + z + ··· ,
n n 2! 3!
and hence
( n1 − 1)z2 ( n1 − 1)( 2n
1
− 1)z3 ( n1 − 1)( 2n
1 1
− 1)( 3n − 1)z4
L(z) ≍ z + + + + ··· .
2 3 4
Finally, since this ultimate equality becomes equality in the limit that (1/n) tends
to zero, we obtain the following logarithmic power series:
z2 z3 z4 z5 z6
Log (1 + z) = z − + − + − + ··· . (2.29)
2 3 4 5 6
For other approaches to this series, see Exs. 31, 32.
Using the ratio test, you can check for yourself that this series does indeed con-
verge inside the unit circle. In fact it can be shown [see Ex. 11] that the series also
converges everywhere on the unit circle, except obviously at z = −1. This yields
some very interesting special cases. For example, putting z = i and then equating
real and imaginary parts, we get
√ 1 1 1 1 1 1
ln 2 = − + − + − + ···
2 4 6 8 10 12

 
1 1 1 1 1
and π = 4 1− + − + − + ··· .
3 5 7 9 11
√
Try checking the first series by noting that if z = 1, then ln 2 = 12 ln(1 + z). The
remarkable second series for π was first published by Leibniz in 1676, though it was
already known to Newton and Gregory. It is usually obtained from the power series
for tan−1 x, evaluated at x = 1. Astonishingly, that power series was discovered
centuries before Newton, Gregory, and Leibniz roamed the Earth, by the Indian
mathematician, Mˉadhava of Sangamagrˉama (c.1340 – c.1425); see (Stillwell, 2010,
p. 166).
For other interesting applications of the logarithmic series, see Exs. 36, 37, 38.

2.7.3 General Powers

If x is a real variable then we are accustomed to being able to express x3 , for
example, as e3 ln x . Let’s see whether we can do the same thing using the complex
exponential and logarithm. That is, let us investigate the possibility of writing

zk = ek log(z) . (2.30)
114 Complex Functions as Transformations

Let z = r eiθ , where θ is chosen to be the principal value, Arg (z). Then

e3 Log (z) = e3(ln r+iθ) = e3 ln r ei3θ = r3 ei3θ = z3 .

But the most general branch of log(z) is simply Log (z) + 2nπi, where n is an
integer, so

e3 log(z) = e6nπi e3 Log (z) = e6nπi z3 = z3

is true irrespective of which branch of the logarithm is chosen. Clearly, by the same
argument, (2.30) is true for all integer values of k.
1 1
Next, consider the three branches of z 3 . Recalling that the principal branch [z 3 ]
√
of this function is 3 r ei(θ/3) , where θ again represents the principal angle, you can
1 1
easily check that e 3 Log (z) = [z 3 ]. Thus the general branch of the logarithm yields
1 2nπ 1
e 3 log(z) = ei 3 [z 3 ].

Thus we have again confirmed (2.30), in the sense that the infinitely many branches
1 1
of log(z) yield precisely the three branches of the cube root: [z 3 ], ei(2π/3) [z 3 ], and
1
ei(4π/3) [z 3 ]. By the same reasoning, if (p/q) is a fraction reduced to lowest terms
p p
then e q log(z) yields precisely the q branches of z q .
Finally, note that the RHS of (2.30) is still meaningful if k = (a + ib) is a complex
number. Emboldened by the above successes, we now take (2.30) as the definition of
a complex power. If we use Log (z) in (2.30) then we find that the principal branch
of z(a+ib) is given by [exercise]

[z(a+ib) ] ≡ e(a+ib) Log (z) = ra e−bθ ei(aθ+b ln r) .

If z now travels along a closed loop encircling the origin n times, then log(z) moves
along a path from Log (z) to Log (z) + 2nπi, and z(a+ib) moves along a path from
[z(a+ib) ] to

z(a+ib) = ei2πna e−2πnb [z(a+ib) ].

If b ̸= 0 then the factor e−2πnb makes it obvious that z(a+ib) never returns to
its original value, no matter how many times we go round the origin. Thus z = 0
is a logarithmic branch point in this case. This is still true even if b = 0, provided
[exercise] that the real power a is irrational. Only when a is a rational number does
za return to its original value after a finite number of revolutions, and only when
a is an integer does za become single-valued.
We end with an important observation on the use of “ez ” to denote the single-
valued exponential mapping. Reversing the roles of the constant and variable in
(2.30), we are forced to define f(z) = kz to be the “multifunction” [see Ex. 29] f(z) =
ez log(k) . But if we now put k = e = 2.718 . . . then we are suddenly in hot water: the
exponential mapping “ez ” is merely one branch [what are the others?] of the newly
defined multifunction (2.718 . . .)z . To avoid this confusion, some authors always
 Averaging over Circles* 115

write the exponential mapping as exp(z). However, we shall retain the notation
“ez ”, which is both convenient and rooted in history, with the understanding that
ez always refers to the single-valued exponential mapping, and never to the multifunction
(2.718 . . .)z .

2.8 Averaging over Circles*

2.8.1 The Centroid

This entire section is optional because the chief result to which we shall be led
(Gauss’s Mean Value Theorem) will be derived again later, in fact more than once. It
is nevertheless fun and instructive to attempt to understand the result using only
the most elementary of methods.
Consider a set of n point particles in C, located at z1 , z2 , . . ., zn . If the mass of the
particle at zj is mj then the centroid Z of the set of particles (also called the “centre
of mass”) is defined to be
Pn
j=1 mj zj
Z ≡ Pn .
j=1 mj

If we imagine the plane to be massless, Z is the point at which we could rest the
plane on a pin so as to make it balance.
Throughout this section we shall take the masses of the particles to be equal, in which
case the centroid becomes the average position of the particles:
1 X
n
Z= zj .
n
j=1

This is the case depicted in [2.38a]. An immediate consequence of this definition

P
is that (zj − Z) = 0. In other words, the complex numbers from Z to the particles
cancel. This vanishing sum is illustrated in [2.38b]. Conversely, if some point Z has
the property that the complex numbers connecting it to the particles cancel, then Z
must be the centroid.
Another immediate result is that if we translate the set of points by b, then the
centroid will translate with them, i.e., the new centroid will be Z + b. The same
thing happens if we rotate the set of points about the origin—the centroid rotates
with them. In general,
If Z is the centroid of {zj }, then the centroid of {azj + b} is aZ + b. (2.31)
Given a second set of n points {e e we may add pairs from the
zj } (with centroid Z),
two sets to obtain the set {zj + e
zj }, and it is easy to see that the centroid of the latter
e
is Z + Z. In particular, the centroid Z of {zj = xj + iyj } is the sum X + iY of the
centroid X of the points {xj } on the real axis and the centroid iY of the points {iyj }
on the imaginary axis.
116 Complex Functions as Transformations

∑n
j=1 mj zj
[2.38] [a] The centroid Z of masses mj located at zj is Z ≡ ∑n . In the case that
j=1 mj
the masses are all equal, join Z to the locations of the masses, and sum these connecting
complex numbers (zj − Z) . . . [b] . . . and discover that their sum vanishes! Furthermore,
this vanishing sum characterizes the location of Z.

Our next result will play a minor role at the end of this section, but later we
shall see that it has other interesting consequences. The convex hull H of the set of
particles {zj } is defined to be the smallest convex polygon such that each particle
lies on H or inside it. More intuitively, first imagine pegs sticking out of the plane
at each point zj , then stretch an imaginary rubber band so as to enclose all the pegs.
When released, the rubber band will contract into the desired polygon H shown in
[2.39a]. We can now state the result:

The centroid Z must lie in the interior of the convex hull H. (2.32)

For if p is outside this set, we see that the complex numbers from p to the particles
cannot possibly cancel, as they must do for Z. More formally, we take it as visually
evident that through any exterior point p we may draw a line L such that H and its
shaded interior lie entirely on one side of L. The impossibility of the complex num-
bers cancelling now follows from their lying entirely on this side of L, for they all
must have positive components in the direction of the illustrated complex number
N normal to L. Except when the particles are collinear (in which case H collapses
to a line-segment), the same reasoning forbids Z from lying on H.
As illustrated in [2.39b], an immediate consequence of (2.32) is that

If all the particles lie within some circle then their centroid must also lie
(2.33)
within that circle.

The main result we wish to derive in this section is based on the following fact.
Defining the centre of a regular n-gon to be the centre of the circumscribing circle,

The centre of a regular n-gon is the centroid of its vertices. (2.34)

 Averaging over Circles* 117

[2.39] The centroid must lie within the convex hull. [a] If p lies outside the convex hull
H of the masses, then the connecting complex numbers from p to the masses cannot
vanish, therefore the centroid Z must lie within H. [b] Therefore, if all the particles lie
within some circle then their centroid must also lie within that circle.

By virtue of (2.31), we may as well choose the n-gon to be centred at the origin,
in which case the claim is that the sum of the vertices vanishes. As illustrated in
[2.40a], this is obvious if n is even since the vertices then occur in opposite pairs.
The explanation is not quite so obvious when n is odd; see [2.40b], which illus-
P
trates the case n = 5. However, if we draw zj systematically, taking the vertices
zj in counterclockwise order, then we obtain [2.40c], and the answer is suddenly
clear: the sum of the vertices of the regular 5-gon forms another regular 5-gon. The figure
explains why this happens. Since the angle between successive vertices in [2.40b] is

[2.40] The sum of the vertices of a regular n-gon vanishes. [a] If n is even, it is obvious
that the sum of the vertices of any regular, origin-centred n-gon must vanish, for the
vertices occur in diametrically opposite pairs. [b] In fact the sum of the vertices must also
vanish when n is odd, but this is no longer immediately obvious. Begin by noting that
the angle between successive vertices is (2π/n). [c] Drawing the sum of the vertices in
counterclockwise order, all is suddenly clear: the sum itself forms a regular n-gon, and
therefore it vanishes.
118 Complex Functions as Transformations

(2π/5), this is also the angle between successive terms of the sum in [2.40c]. Clearly
this argument generalizes to arbitrary n (both odd and even), thereby establishing
(2.34). For a different approach, see Ex. 40.

2.8.2 Averaging over Regular Polygons

If a complex mapping z ÞÑ w = f(z) maps the set of points {zj } to the set {wj =
f(zj )}, then the centroid W of the image points may be described as the average of
f(z) over the set {zj } of n points. Writing this average as ⟨f(z)⟩n ,
Xn
⟨f(z)⟩n = n1
f(zj ) .
j=1

Note that if f(z) = c is constant, then its average over any set of points is equal
to c.
Henceforth, we shall restrict ourselves to the case where {zj } are the vertices of
a regular n-gon; correspondingly, ⟨f(z)⟩n will be understood as the average of f(z)
over the vertices of such a regular n-gon. Note that if we write f(z) = u(z) + iv(z),
then
⟨f(z)⟩n = ⟨u(z)⟩n + i ⟨v(z)⟩n . (2.35)
Initially, we consider only origin-centred polygons.
Consider, then, the average of f(z) = zm over the vertices of such a regular n-
gon. Figure [2.41] illustrates the case n = 6. In the centre of the figure is a shaded
regular hexagon, and on the periphery are the images of its vertices under the map-
pings z, z2 , . . ., z6 . Study this figure carefully, and see if you can understand what’s
going on. If we take still higher values of m, then this pattern repeats cyclically: z7
is like z1 , z8 is like z2 , and so on.
For us the essential feature of this figure is that unless m is a multiple of 6, the
image under zm of the regular 6-gon is another regular polygon. [Note that we
count two equal and opposite points as a regular 2-gon, but we do not count a
single point as a regular polygon.] More precisely, and in general,
Unless m is a multiple of n, the image under zm of an origin-centred reg-
ular n-gon is an origin-centred regular N-gon, where N = (n divided by
(2.36)
the highest common factor of m and n). If m is a multiple of n, then the
image is a single point.
Check that this agrees with [2.41]. Try to establish the result on your own, but see
Ex. 41 if you get stuck.
Combining this result with (2.34), we obtain the following key fact: If n >
m then ⟨zm ⟩n = 0. This is easy to generalize. If
Pm (z) = c0 + c1 z + c2 z2 + c3 z3 + · · · + cm zm
 Averaging over Circles* 119

[2.41] Unless m is a multiple of n, the image under zm of an origin-centred regular n-gon

is an origin-centred regular N-gon, where N = (n divided by the highest common factor
of m and n). If m is a multiple of n, then the image is a single point.

is a general polynomial of degree m, then its average over the vertices of the
n-gon is

⟨Pm (z)⟩n = ⟨c0 ⟩n + c1 ⟨z⟩n + c2 ⟨z2 ⟩n + c3 ⟨z3 ⟩n + · · · + cm ⟨zm ⟩n .

If the number n of vertices is greater than the degree m of the polynomial, we

therefore obtain

⟨Pm (z)⟩n = ⟨c0 ⟩n = c0 = Pm (0).

In other words, the centroid of the image points is the image of the centroid. Expressing
this result in the language of averages,

If n > m then the average of an mth degree polynomial Pm (z) over

the vertices of an origin-centred regular n-gon is its value Pm (0) at the (2.37)
centre of the n-gon.

Finally, let us generalize to regular n-gons that are centred at an arbitrary point k,
instead of the origin. Of course when we apply zm to the vertices of such a regular
polygon, the image points do not form a regular polygon. See [2.42], which shows
the effect of z4 on the vertices of a regular hexagon H centred at k, together with
the image of the entire circle on which these vertices lie. Nevertheless, the figure
also illustrates the surprising and beautiful fact that, once again, the centroid of the
image points is the image of the centroid of H. Figure [2.43a] confirms this empirically
120 Complex Functions as Transformations

[2.42] It is a surprising and beautiful fact (here illustrated with n = 6 and m = 4) that the
centroid of the image points of the vertices of a regular n-gon H under z ÞÑ zm is the
image of the centroid of H.

[2.43] [a] Empirical verification of the claim made in the previous figure: the sum of
the connecting complex numbers from k4 to images of the vertices of H vanishes, so k4
is indeed the centroid of these image points. [b] In order to mathematically explain this
e see main text
phenomenon, we view H to be the translation by k of the origin-centred H;
for the argument.

by showing that the sum of complex numbers connecting k4 to the image points is
indeed zero.
Extending our notation slightly, we may write the average of zm over the vertices
of H as ⟨zm ⟩H , so what we must show is that ⟨z4 ⟩H = k4 . It is no harder to treat the
general case of zm acting on the vertices of a regular n-gon H centred at k. First
note that H can be obtained by translating an origin-centred n-gon H e by k. See the
e translates to a vertex zj + k of H, it follows
example in [2.43b]. Since a vertex zj of H
that

⟨zm ⟩H = ⟨(z + k)m ⟩H

e.
 Averaging over Circles* 121

P

But (z + k)m = m m j m−j
j=0 j z k is just an mth degree polynomial which maps
0 to k . Using (2.37), we conclude that if n > m then ⟨zm ⟩H = km , as was to be
m

shown.
Generalizing the argument that led to (2.37), we see that (2.37) is a special case
of the following result:
If n > m then the average of an mth degree polynomial Pm (z) over the
vertices of a regular n-gon centred at k is its value Pm (k) at the centre (2.38)
of the n-gon.

2.8.3 Averaging over Circles

Since at least the time of Archimedes, mathematicians have found it fruitful to think
of a circle as the limit of a regular n-gon as n tends to infinity. We will now use this
idea to investigate the average of a complex function over a circle.
Inscribing a regular n-gon in a given circle, and taking the limit as n tends to
infinity, (2.38) shows that
The average over a circle C of a polynomial of arbitrarily high degree is
(2.39)
equal to the value of the polynomial at the centre of C.
By (2.35), the average ⟨f(z)⟩C of a complex function f(z) = u(z)+iv(z) over a cir-
cle C may be expressed as ⟨f(z)⟩C = ⟨u(z)⟩C + i ⟨v(z)⟩C . Using a familiar idea from
ordinary calculus, the averages of the two real functions u and v may be expressed
as integrals. If C has centre k and radius R, then as θ varies between 0 and 2π,
z = k + R eiθ traces out C. Thus,
Z Z
1 2π 1 2π
⟨u(z)⟩C = u(k + R eiθ ) dθ and ⟨v(z)⟩C = v(k + R eiθ ) dθ.
2π 0 2π 0
More compactly, we may write
Z 2π
1
⟨f(z)⟩C = f(k + R eiθ ) dθ,
2π 0

in which it is understood that the complex integral may be evaluated in terms of

the real integrals above.
Once again denoting a general mth degree polynomial by Pm (z), (2.39) can
therefore be expressed as an integral formula:
Z
1 2π
Pm (k + R eiθ ) dθ = ⟨Pm (z)⟩C = Pm (k). (2.40)
2π 0
For example, if C is centred at the origin and Pm (z) = zm , then
Z Z
1 2π m imθ Rm 2π
⟨z ⟩C =
m
R e dθ = [cos mθ + i sin mθ] dθ = 0,
2π 0 2π 0
in agreement with (2.40).
122 Complex Functions as Transformations

The fact that (2.40) holds for polynomials of arbitrarily high degree immediately
suggests that it might also hold for power series. We shall show that it does.
As usual we will only give the details for origin-centred power series, the gen-
P
eralization to arbitrary centres being straightforward. Let P(z) = ∞ j
j=0 cj z be the
Pm
power series, so that Pm (z) = j=0 cj zj are its approximating polynomials. If the
circle C lies inside the disc of convergence of P(z), then (2.12) implies the following.
No matter how small we choose a real number ϵ, we can find a sufficiently large m
such that Pm (z) approximates P(z) with accuracy ϵ throughout C and its interior. If
we write E(z) for the complex number from the approximation Pm (z) to the exact
answer P(z), then

P(z) = Pm (z) + E(z), where |E(z)| < ϵ

for all z on and inside C, and in particular at the centre k of C.

At this point we could immediately study ⟨P(z)⟩C in terms of its integral rep-
resentation, but it is more instructive to first consider the average ⟨P(z)⟩n of P(z)
over a regular n-gon inscribed in C. Once this is done, we may let n tend to infinity
to obtain ⟨P(z)⟩C .
First note that E(z) maps the vertices of the n-gon to points lying inside an
origin-centred disc of radius ϵ. By (2.33), or directly from the generalized trian-
gle inequality (1.8) on p. 8, the centroid ⟨E(z)⟩n of these points must also lie in this
disc. Choosing n greater than m, say n = (m + 1), (2.38) yields

⟨P(z)⟩m+1 = ⟨Pm (z)⟩m+1 + ⟨E(z)⟩m+1

= Pm (k) + ⟨E(z)⟩m+1
= P(k) + [⟨E(z)⟩m+1 − E(k)] .
The term in square brackets is the connecting complex number from E(k) to
⟨E(z)⟩m+1 , and since both these points lie within a disc of radius ϵ, their connecting
complex number must be shorter than 2ϵ. Finally, since the term in square brackets
may also be interpreted as the connecting complex number from P(k) to ⟨P(z)⟩m+1 ,
we have the following result:

Let m be chosen so that Pm (z) approximates the power series P(z) with
accuracy ϵ on and within a circle C centred at k. If a regular (m+1)-gon
(2.41)
is inscribed in C, then the average ⟨P(z)⟩m+1 of P(z) over its vertices
will approximate P(k) with accuracy 2ϵ.
We have thus transformed an exact result concerning the approximation Pm (z) into
an approximation result concerning the exact mapping P(z).
For example, let C be the unit circle, and let P(z) = ez . If we desire an accuracy
of ϵ = 0.004 everywhere on the unit disc then it turns out that m = 5 is sufficient,
i.e., the approximating polynomial of lowest degree that has this accuracy is
1 2 1 3 1 4 1 5
P5 (z) = 1 + z + 2! z + 3! z + 4! z + 5! z .
 Averaging over Circles* 123

Figure [2.44] shows the image under z ÞÑ ez of C, and in particular it shows

the images of the vertices of a regular hexagon inscribed in C. According to the
result, the centroid of these image points should differ from e0 = 1 by no more than
0.008—an indiscernible discrepancy in a drawing done to this scale. This prediction
is convincingly borne out in [2.45], which shows the sum of the complex numbers
connecting 1 to the images of the vertices of the hexagon. To within the accuracy
of the drawing, the sum is indeed zero!

[2.44] Applying z ÞÑ ez to the unit circle and to the vertices of the inscribed regular
hexagon on the left yields the figure on the right. The centroid of the six image points is
almost the image of the centre of the original hexagon, namely, e0 = 1. As the number
of vertices goes to infinity, the result is ultimately exact.

[2.45] Empirical confirmation of the claim made in the previous figure: the connecting
complex numbers from 1 to the images of the vertices of the hexagon do indeed sum to
zero, at least to within the accuracy of this drawing.
124 Complex Functions as Transformations

In the limit that ϵ tends to zero and m tends to infinity, (2.41) yields a form of
Gauss’s Mean Value Theorem:
If a complex function f(z) can be expressed as a power series, and a circle
C (radius R and centre k) lies within the disc of convergence of that
power series, then
R2π
⟨f(z)⟩C = 2π1 iθ
0 f(k + R e ) dθ = f(k).

In addition to its theoretical importance, this formula can sometimes be used to

evaluate difficult real integrals. For example, the exact version of [2.44] is ⟨ez ⟩C =
R2π
e0 = 1, and this implies [exercise] that 0 ecos θ cos[sin θ] dθ = 2π. See Ex. 43 for
another example of this idea.
 Exercises 125

2.9 Exercises

1 Sketch the circle |z−1| = 1. Find (geometrically) the polar equation of the image
of this circle under the mapping z ÞÑ z2 . Sketch this image curve, which is called
a cardioid.
2 Consider the complex mapping z ÞÑ w = (z − a)/(z − b). Show geometrically
that if we apply this mapping to the perpendicular bisector of the line-segment
joining a and b, then the image is the unit circle. In greater detail, describe the
motion of w round this circle as z travels along the line at constant speed.
3 Consider the family of complex mappings
z−a
z ÞÑ Ma (z) = (a constant).
az − 1
[These mappings will turn out to be fundamental to non-Euclidean geometry.]
Do the following problems algebraically; in the next chapter we will provide
geometric explanations.
(i) Show that Ma [Ma (z)] = z. In other words, Ma is self-inverse.
(ii) Show that Ma (z) maps the unit circle to itself.
(iii) Show that if a lies inside the unit disc then Ma (z) maps the unit disc to
itself.
Hint: Use |q|2 = q q to verify that

|a z − 1|2 − |z − a|2 = (1 − |a|2 ) (1 − |z|2 ).

4 In figure [2.7] we saw that if q2 ⩽ p3 then the solutions of x3 = 3px + 2q are all
real. Draw the corresponding picture in the case q2 > p3 , and deduce that one
solution is real, while the other two form a complex conjugate pair.
5 Show that the mapping z ÞÑ z2 doubles the angle between two rays coming
out of the origin. Use this to deduce that the lemniscate (see [2.9] on p. 69)
must self-intersect at right angles.
6 This question refers to the Cassinian curves in [2.9] on p. 69.
(i) On a copy of this figure, sketch the curves that intersect each Cassinian
curve at right angles; these are called the orthogonal trajectories of the
original family of curves.
(ii) Give an argument to show that each orthogonal trajectory hits one of the
foci at ±1.
(iii) If the Cassinian curves are thought of as a geographical contour map of
the modular surface (cf. [2.10]) of (z2 − 1), then what is the interpretation
of the orthogonal trajectories in terms of the surface?
126 Complex Functions as Transformations

(iv) In Chapter 4 we will show that if two curves intersect at some point p ̸=
0, and if the angle between them at p is ϕ, then the image curves under
z ÞÑ w = z2 will also intersect at angle ϕ, at the point w = p2 . Use this
to deduce that as z travels out from one of the foci along an orthogonal
trajectory, w = z2 travels along a ray out of w = 1.
(v) Check the result of the previous part by using a computer to draw the
√
images under w ÞÑ w of (A) circles centred at w = 1; (B) the radii of
such circles.
(vi) Writing z = x + iy and w = u + iv, find u and v as functions of x and
y. By writing down the equation of a line in the w-plane through w = 1,
show that the orthogonal trajectories of the Cassinian curves are actually
segments of hyperbolas.
7 Sketch the modular surface of C(z) = (z + 1)(z − 1)(z + 1 + i). Hence sketch the
Cassinian curves |C(z)| = const., then check your answer using a computer. To
answer the following questions, recall that if R(z) is a real function of position
in the plane, then R(p) is a local minimum of R if R(p) < R(z) for all z in the
immediate neighbourhood of p. A local maximum is defined similarly.
(i) Referring to the previous exercise, what is the significance of the orthog-
onal trajectories of the Cassinian curves you have just drawn?
(ii) Does |C(z)| have any local maxima?
(iii) Does |C(z)| have any non-zero local minima?
(iv) If D is a disc (or indeed a more arbitrary shape), can the maximum of |C(z)|
on D occur at a point inside D, or must the maximum occur at a boundary
point of D? What about the minimum of |C(z)| on D?
(v) Do you get the same answers to these questions if C(z) is replaced by
an arbitrary polynomial? What about a complex function that is merely
known to be expressible as a power series?
8 On page 69 we saw that the polar equation of the lemniscate with foci at ±1 is
r2 = 2 cos 2θ. In fact James Bernoulli and his successors worked with a slightly
different lemniscate having equation r2 = cos 2θ. Let us call this the standard
lemniscate.
(i) Where are the foci of the standard lemniscate?
(ii) What is the value of the product of the distances from the foci to a point
on the standard lemniscate?
(iii) Show that the Cartesian equation of the standard lemniscate is

(x2 + y2 )2 = x2 − y2 .
 Exercises 127

9 Here is an attempt [ultimately doomed] at using real methods to expand H(x) =

1/(1 + x2 ) into a power series centred at x = k, i.e., into a series of the form
P
H(x) = ∞ j
j=0 cj X , where X = (x − k). According to Taylor’s Theorem, cj =
H(j) (k)/j!, where H(j) (k) is the jth derivative of H.
(i) Show that c0 = 1/(1 + k2 ) and c1 = −2k/(1 + k2 )2 , and find c2 . Note how
it becomes increasingly difficult to calculate the successive derivatives.
(ii) Recall (or prove) that the nth derivative of a product AB of two functions
A(x) and B(x) is given by Leibniz’s rule:
Xn  
n
(AB)(n) = A(j) B(n−j) .
j
j=0

By applying this result to the product (1 + x2 )H(x), deduce that

(1 + k2 )H(n) (k) + 2nkH(n−1) (k) + n(n − 1)H(n−2) (k) = 0.
Because the coefficients in this recurrence relation depend on n, we cannot
solve it using the technique of Ex. 30 on p. 55.
(iii) Deduce from the previous part that the recurrence relation for the cj ’s is
(1 + k2 )cn + 2kcn−1 + cn−2 = 0,
which does have constant coefficients.
(iv) Solve this recurrence relation, and hence recover the result (2.17) on p. 85.
10 Reconsider the series (2.18) on p. 85.
(i) Show that we recover the correct series (missing the odd powers of x) when
the centre k of the series is at the origin.
(ii) Find a value of k such that the series is missing all the powers Xn , where
n = 2, 5, 8, 11, 14, . . .. Check your answer using a computer.
11 Show that each of the following series has the unit circle as its circle of conver-
gence, then investigate the convergence on the unit circle. You can guess the
correct answers by “drawing the series” in the manner of [2.18] on p. 89.
∞
X ∞
X ∞
X
zn zn
(i) zn , (ii) , (iii) .
n n2
n=0 n=1 n=1

[By virtue of (2.29), note that the second series is −Log (1 − z).]
P∞ j
12 Consider the geometric series P(z) = j=0 z , which converges to 1/(1 − z)
inside the unit disc. The approximating polynomials in this case are Pm (z) =
Pm j
j=0 z .
(i) Show that the error Em (z) ≡ |P(z) − Pm (z)| is given by
|z|m+1
Em (z) = .
|1 − z|
128 Complex Functions as Transformations

(ii) If z is any fixed point in the disc of convergence, what happens to the error
as m tends to infinity?
(iii) If we fix m, what happens to the error as z approaches the boundary point
z = 1?
(iv) Suppose we want to approximate this series in the disc |z| ⩽ 0.9, and fur-
ther suppose that the maximum error we will tolerate is ϵ = 0.01. Find the
lowest degree polynomial Pm (z) that approximates P(z) with the desired
accuracy throughout the disc.
13 We have seen that if we set Pn (z) = zn , then the representation of a com-
P
plex function f(z) as an infinite series ∞ n=0 cn Pn (z) (i.e., a power series) is
unique. This is not true, however, if Pn (z) is just any old set of polynomials.
The following example is taken from Boas and Boas (2010). Defining

zn−1 zn
P0 (z) = −1, and Pn (z) = − (n = 1, 2, 3, . . .),
(n − 1)! n!
show that

−2P0 − P1 + P3 + 2P4 + 3P5 + · · · = ez = P1 + 2P2 + 3P3 + 4P4 + · · · .

P P∞
14 Consider two power series, P(z) = ∞ j
j=0 pj z and Q(z) =
j
j=0 qj z , which
Pn P
have approximating polynomials Pn (z) = j=0 pj zj and Qm (z) = m j
j=0 qj z .
If the radii of convergence of P(z) and Q(z) are R1 and R2 then both series are
uniformly convergent in the disc |z| ⩽ r, where r < min{R1 , R2 }. Thus if ϵ is the
maximum error we will tolerate in this disc, we can find a sufficiently large n
such that

Pn (z) = P(z) + E1 (z) and Qn (z) = Q(z) + E2 (z),

where the (complex) errors E1,2 (z) both have lengths less than ϵ. Use this to
show that by taking a sufficiently high value of n we can approximate [P(z) +
Q(z)] and P(z) Q(z) with arbitrarily high precision using [Pn (z) + Qn (z)] and
Pn (z) Qn (z), respectively.
15 Give an example of a pair of origin-centred power series, say P(z) and Q(z),
such that the disc of convergence for the product P(z)Q(z) is larger than either of
the two discs of convergence for P(z) and Q(z). [Hint: think in terms of rational
functions, such as [z2 /(5 − z)3 ], which are known to be expressible as power
series.]
16 Our aim is to give a combinatorial explanation of the Binomial Theorem (2.14)
for all negative integer values of n. The simple yet crucial first step is to write
n = −m and to change z to −z. Check that the desired result (2.14) now takes
 Exercises 129

P∞
the form (1 − z)−m = r=0 cr zr , where cr is the binomial coefficient
 
m+r−1
cr = . (2.42)
r
[Note that this says that the coefficients cr are obtained by reading Pascal’s tri-
angle diagonally, instead of horizontally.] To begin to understand this, consider
the special case m = 3. Using the geometric series for (1−z)−1 , we may express
(1 − z)−3 as

[1 + z + z2 + z3 + · · · ] • [1 + z + z2 + z3 + · · · ] • [1 + z + z2 + z3 + · · · ],

where • simply denotes multiplication. Suppose we want the coefficient c9 of

z9 . One way to get z9 is to take z3 from the first bracket, z4 from the second, and
z2 from the third.
(i) Write this way of obtaining z9 as the sequence zzz • zzzz • zz of 9 z’s and
2 •’s, where the latter keep track of which power of z came from which
bracket. [I got this nice idea from my friend Paul Zeitz.] Explain why c9
is the number of distinguishable rearrangements of this sequence of 11
symbols. Be sure to address the meaning of sequences in which a • comes
first, last, or is adjacent to the other •.


(ii) Deduce that c9 = 11 9 , in agreement with (2.42).
(iii) Generalize this argument and thereby deduce (2.42).
17 Here is an inductive approach to the result of the previous exercise.
(i) Write down the first few rows of Pascal’s triangle and circle the numbers
5

4
3
2
6


3 , 2 , 1 , 0 . Check that the sum of these numbers is 3 . Explain this.
(ii) Generalize your argument to show that
       
n n−1 n−2 n−3
= + + + · · · + 1.
r r r−1 r−2
P
r
(iii) Assume that (1−z)−M = ∞ r=0
M+r−1
r z holds for some positive integer
M. Now multiply this series by the geometric series for (1 − z)−1 to find
(1 − z)−(M+1) . Deduce that the binomial series is valid for all negative
integer powers.
18 The basic idea of the following argument is due to Euler. Initially, let n be any
real (possibly irrational) number, and define
X∞    
n r n n(n − 1)(n − 2) . . . (n − r + 1)
B(z, n) ≡ z where ≡ ,
r r r!
r=0


and n
0 ≡ 1. We know from elementary algebra that if n is a positive integer
then B(z, n) = (1 + z)n . To establish the Binomial Theorem (2.14) for rational
130 Complex Functions as Transformations

p
powers, we must show that if p and q are integers then B(z, q ) is the principal
p
branch of (1 + z) q .
(i) With a fixed value of n, use the ratio test to show that B(z, n) converges
in the unit disc, |z| < 1.
(ii) By multiplying the two power series, deduce that
∞
X r  
X 
r n m
B(z, n) B(z, m) = Cr (n, m) z where Cr (n, m) = .
j r−j
r=0 j=0

(iii) If m and n are positive integers, then show that

B(z, n) B(z, m) = B(z, n + m), (2.43)

and deduce that Cr (n, m) = n+m r . But Cr (n, m) and n+m r are simply
polynomials in n and m, and so the fact that they agree at infinitely many
values of m and n [positive integers] implies that they must be equal for all
real values of m and n. Thus the key formula (2.43) is valid for all real values
of m and n.
(iv) By substituting n = −m in (2.43), deduce the Binomial Theorem for
negative integer values of n.
(v) Use (2.43) to show that if q is an integer then [B(z, q1 )]q = (1 + z). Deduce
1
that B(z, q1 ) is the principal branch of (1 + z) q .
p
(vi) Finally, show that if p and q are integers, then B(z, q ) is indeed the
p
principal branch of (1 + z) q .
19 Show that the ratio test cannot be used to find the radius of convergence of the
√
power series (2.18) on p. 85. Use the root test to confirm that R = 1 + k2 .
R2π
20 Show that if m and n are integers, then 0 cos mθ cos nθ dθ vanishes unless
m = n, in which case it equals π. Likewise, establish a similar result for
R2π
0 sin mθ sin nθ dθ. Use these facts to verify (2.19), at least formally.

21 Do the following problems by first substituting z = r eiθ into the power series
for ez , then equating real and imaginary parts.
P
(i) Show that the Fourier series for [cos(sin θ)] ecos θ is ∞n=0
cos nθ , and
n!
write down the Fourier series for [sin(sin θ)] ecos θ .
R2π
(ii) Deduce that 0 ecos θ [cos(sin θ)] cos mθ dθ = (π/m!), where m is a
positive integer.
√
(iii) By writing x = (r/ 2), find the power series for f(x) = ex sin x.
(iv) Check the first few terms of the series for f(x) by multiplying the series
for ex and sin x.
 Exercises 131

(v) Calculate the nth derivative f(n) (0) using (1.14) on p. 25 of Chapter 1. By
using these derivatives in Taylor’s Theorem, verify your answer to part
(iii).
22 Reconsider the formula,
 z n
ez = lim Pn (z), where Pn (z) = 1 + .
nÑ∞ n

(i) Check that Pn (z) is the composition of a translation by n, followed by a

contraction by (1/n), followed by the power mapping z ÞÑ zn .
(ii) Referring to figure [2.4] on p. 64, use the previous part to sketch the images
under Pn (z) of circular arcs centred at −n, and of rays emanating from −n.
(iii) Let S be an origin-centred square (say of unit side) in the z-plane. With a
large value of n, sketch just those portions of the arcs and rays (considered
in the previous part) that lie within S.
(iv) Use the previous two parts to quantitatively explain figure [2.19] on p. 90.
23 If you did not do so earlier, sketch the image of a vertical line x = k under z ÞÑ
w = cos z by drawing the analogue of [2.26]. Deduce that the asymptotes of
this hyperbola are arg w = ±k. Check this using the equation of the hyperbola.
√ √
24 Consider the multifunction f(z) = z − 1 3 z − i.
(i) Where are the branch points and what are their orders?
(ii) Why is it not possible to define branches using a single branch cut of the
type shown in [2.35b]?
(iii) How many values does f(z) have at a typical point z? Find and then plot
all the values of f(0).
(iv) Choose one of the values of f(0) which you have just plotted, and label
it p. Sketch a loop L that starts and ends at the origin such that if f(0) is
initially chosen to be −1, then as z travels along L and returns to the origin,
f(z) travels along a path from −1 to p. Do the same for each of the other
possible values of f(0).
√
25 Describe the branch points of the function f(z) = 1/ 1 − z4 . What is the small-
est number of branch cuts that may be used to obtain single-valued branches
of f(z)? Sketch an example of such cuts. [Remark: This function is historically
R
important, owing to the fact (Ex. 20, p. 242) that f(x) dx represents the arc
length of the lemniscate. This integral (the lemniscatic integral) cannot be eval-
uated in terms of elementary functions—it is an example of a new kind of
function called an elliptic integral. See Stillwell (2010), for more background and
detail.]
132 Complex Functions as Transformations

26 For each function f(z) below, find and then plot all the branch points and sin-
gularities. Assuming that these functions may be expressed as power series
centred at k [in fact they can be], use the result (2.27) on p. 107 to verify the
stated value of the radius of convergence R.
(i) If f(z) = 1/(eπz − 1) and k = (1 + 2i), then R = 1.
√
(ii) If f(z) is a branch of 5 z4 − 1 and k = 3i, then R = 2.
√ √
(iii) If f(z) is a branch of z − i/(z − 1) and k = −1, then R = 2.
27 Until Euler cleared up the whole mess, the complex logarithm was a source
of tremendous confusion. For example, show that log(z) and log(−z) have no
common values, then consider the following argument of John Bernoulli:

log[(−z)2 ] = log[z2 ] =⇒ log(−z) + log(−z) = log(z) + log(z)

=⇒ 2 log(−z) = 2 log(z)
=⇒ log(−z) = log(z).

What is wrong with this argument?!

28 What value does zi take at z = −1 if we start with the principal value at z = 1
(i.e., 1i = 1), and then let z travel one and a half revolutions clockwise round
the origin?
29 In this exercise you will see that the “multifunction” kz is quite different in char-
acter from all the other multifunctions we have discussed. For integer values
of n, define ln ≡ [Log (k) + 2nπi].
(i) Show that the “branches” of kz are eln z .
(ii) Suppose that z travels along an arbitrary loop, beginning and ending at
z = p. If we initially choose the value el2 p for kz , then what value of kz do
we arrive at when z returns to p? Deduce that kz has no branch points.
Since we cannot change one value of kz into another by travelling round a loop,

we should view its “branches” . . . , el−1 z , el0 z , el1 z , . . . as an infinite set of
completely unrelated single-valued functions.
30 Show that all the values of ii are real! Are there any other points z such that zi
is real?
31 In the case of a real variable, the logarithmic power series was originally dis-
covered [see next exercise] as follows. First check that ln(1+X) can be written as
RX
0 [1/(1+x)] dx, and then expand [1/(1+x)] as a power series in x. Finally, inte-
grate your series term by term. [Later in the book we will be able to generalize
this argument to the complex plane.]
 Exercises 133

32 Here is another approach to the logarithmic power series. As before, let L(z) =
Log (1 + z). Since L(0) = 0, the power series for L(z) must be of the form L(z) =
az + bz2 + cz3 + dz4 + · · · . Substitute this into the equation

1 + z = eL = 1 + L + 1 2
2! L + 1 3
3! L + 1 4
4! L + ··· ,

then find a, b, c, and d by equating powers of z. [Historically the logarithmic

series came first—both Mercator and Newton discovered it using the method
in the previous exercise—then Newton reversed the reasoning of the present
exercise to obtain the series for ex . See Stillwell (2010).]
33 (i) Use [2.26] to discuss the branch points of the multifunction cos−1 (z).
(ii) Rewrite the equation w = cos z as a quadratic in eiz . By solving this
√
equation, deduce that cos−1 (z) = −i log[z + z2 − 1 ]. [Why do we not
need to bother to write ± in front of the square root?]
(iii) Show that as z travels along a loop that goes once round either 1 or −1 (but
√ √
not both), the value of [z + z2 − 1 ] changes to 1/[z + z2 − 1 ].
(iv) Use the previous part to show that the formula in part (ii) is in accord with
the discussion in part (i).
34 Write down the origin-centred power series for (1 − cos z). Use the Binomial
Theorem to write down the power series (centred at Z = 0) for the principal
√
branch of 1 − Z, then substitute Z = (1−cos z). Hence show that if we choose
√
the branch of cos z that maps 0 to 1, then
√ z2 z4 19z6
cos z = 1 − − − − ··· .
4 96 5760
Verify this using a computer. Where does this series converge?
35 What value does (z/ sin z) approach as z approaches the origin? Use the
series for sin z to find the first few terms of the origin-centred power series
for (z/ sin z). Check your answer using a computer. Where does this series
converge?
36 By considering Log (1+ix), where x is a real number lying between ±1, deduce
that
x3 x5 x7 x9 x11
tan−1 (x) = x − + − + − + ··· .
3 5 7 9 11
In what range does this value of tan−1 (x) lie? Give another derivation of the
series using the idea in Ex. 31.
37 (i) Show geometrically that as z = eiθ goes round and round the unit circle
(with ever increasing θ), Im [Log (1 + z)] = (Θ/2), where Θ is the principal
value of θ, i.e., −π < Θ ⩽ π.
134 Complex Functions as Transformations

(ii) Consider the periodic “saw tooth” function F(θ) whose graph is shown
below. By substituting z = eiθ in the logarithmic series (2.29), use the
previous part to deduce the following Fourier series:
sin 2θ sin 3θ sin 4θ
F(θ) = sin θ − + − + ··· .
2 3 4
(iii) Check this Fourier series by directly evaluating the integrals (2.19).
(iv) Use a computer to draw graphs of the partial sums of the Fourier series.
As you increase the number of terms, observe the magical convergence of
this sum of smooth waves to the jagged graph below. If only Fourier could
have seen this on the screen, not just in his mind’s eye!

38 As in the previous exercise, let Θ = Arg (z).

(i) Use (2.29) to show that
 
1 1+z z3 z5 z7
Log =z+ + + + ··· .
2 1−z 3 5 7
(ii) Show geometrically that as z = eiθ goes round and round the unit circle,
  hπi
1 1+z
Im Log = (sign of Θ) .
2 1−z 4
(iii) Consider the periodic “square wave” function G(θ) whose graph is shown
below. Use the previous two parts to deduce that its Fourier series is
sin 3θ sin 5θ sin 7θ
G(θ) = sin θ + + + + ··· .
3 5 7
Finally, repeat parts (iii) and (iv) of the previous exercise.

39 Show that (2.32) is still true even if the (positive) masses of the particles are not
all equal.
 Exercises 135

40 Here is another simple way of deriving (2.34). If the vertices of the origin-
centred regular n-gon are rotated by ϕ, then their centroid Z rotates with them
to eiϕ Z. By choosing ϕ = (2π/n), deduce that Z = 0.
41 To establish (2.36), let z0 , z1 , z2 , . . . , zn−1 be the vertices (labelled counterclock-
wise) of the regular n-gon, and let C be the circumscribing circle. Also, let
wj = zm j be the image of vertex zj under the mapping z ÞÑ w = z . Think
m

of z as a particle that starts at z0 and orbits counterclockwise round C, so that

the image particle w = zm starts at w0 and orbits round another circle with m
times the angular speed of z.
(i) Show that each time z travels from one vertex to the next, w executes
(m/n) of a revolution. Thus as z travels from z0 to zk , w executes k(m/n)
revolutions as it travels from w0 to wk .
(ii) Let wk be the first point in the sequence w1 , w2 , etc., such that wk = w0 .
Deduce that if (M/N) is (m/n) reduced to lowest terms, then k = N. Note
that N = (n divided by the highest common factor of m and n).
(iii) Explain why wN+1 = w1 , wN+2 = w2 , etc.
(iv) Show that w0 , w1 , . . . , wN−1 are distinct.
(v) Show that w0 , w1 , . . . , wN−1 are the vertices of a regular N-gon.
42 Consider the mapping z ÞÑ w = Pn (z), where Pn (z) is a general polynomial
of degree n ⩾ 2. Let Sq be the set of points in the z-plane that are mapped to
a particular point q in the w-plane. Show that the centroid of Sq is independent
of the choice of q, and is therefore a property of the polynomial itself. [Hint:
This is another way of looking at a familiar fact about the sum of the roots of a
polynomial.]
43 Use Gauss’s Mean Value Theorem [p. 78] to find the average of cos z over the
circle |z| = r. Deduce (and check with a computer) that for all real values of r,
Z 2π
cos[r cos θ] cosh[r sin θ] dθ = 2π.
0
CHAPTER 3

Möbius Transformations and Inversion

3.1 Introduction

3.1.1 Definition and Significance of Möbius Transformations

A Möbius transformation1 is a mapping of the form
az + b
M(z) = , (3.1)
cz + d
where a, b, c, d are complex constants. These mappings have many beautiful prop-
erties, and they find very varied application throughout complex analysis. Despite
their apparent simplicity, Möbius transformations lie at the heart of several exciting
areas of modern mathematical research. This is due in large part to their intimate
and somewhat miraculous connection with the non-Euclidean geometries alluded
to in Chapter 1. [This connection is the subject of Chapter 6.] Moreover, these trans-
formations are also intimately connected2 with Einstein’s Theory of Relativity! This
connection has been exploited with remarkable success by Sir Roger Penrose; see
Penrose and Rindler (1984).
Two centuries have passed since August Ferdinand Möbius first studied the
transformations that now bear his name, but it is fair to say that the rich vein of
knowledge which he thereby exposed is still far from being exhausted. For this
reason, we shall investigate Möbius transformations in considerably greater depth
than is customary.

1
Also known as a “linear”, “bilinear”, “linear-fractional”, or “homographic” transformation.
2
According to Coxeter (1967), this connection was first recognized by H. Liebmann in 1905, the very
year that Einstein discovered Special Relativity!

Visual Complex Analysis. 25th Anniversary Edition. Tristan Needham, Oxford University Press.
© Tristan Needham (2023). DOI: 10.1093/oso/9780192868916.003.0003
138 Möbius Transformations and Inversion

3.1.2 The Connection with Einstein’s Theory of Relativity*

Clearly it would be neither appropriate nor feasible for us to explore this connection
in detail, but let us at least briefly indicate how Möbius transformations are related
to Einstein’s Theory of Relativity.
In that theory, the time T and the 3-dimensional Cartesian coordinates (X, Y, Z)
of an event are combined into a single 4-vector (T , X, Y, Z) in 4-dimensional space-
time. Of course the spatial components of this vector have no absolute significance:
rotating the coordinate axes yields different coordinates (X, e Y,
e Z)
e for one and the
same point in space. But if two people choose different axes, they will nevertheless
agree on the value of X e2 + Ye2 + Z
e2 = X2 + Y 2 + Z2 , for this represents the square of
the distance to the point.
In contrast to this, we are accustomed to thinking that the time component T does
have an absolute significance. However, Einstein’s theory—confirmed by innu-
merable experiments—tells us that this is wrong. If two (momentarily coincident)
observers are in relative motion, they will disagree about the times at which events occur.
Furthermore, they will no longer agree about the value of (X2 + Y 2 + Z2 )—this is
the famous Lorentz contraction. Is there any aspect of spacetime that has absolute
significance and on which two observers in relative motion must agree? Yes: mak-
ing a convenient choice of units in which the speed of light is equal to 1, Einstein
discovered that both observers will agree on the value of

Te2 − (X
e2 + Ye2 + Z
e2 ) = T 2 − (X2 + Y 2 + Z2 ).

A Lorentz transformation L is a linear transformation of spacetime (a 4 × 4 matrix)

that maps one observer’s description (T , X, Y, Z) of an event to another observer’s
description (Te, X,
e Y,
e Z)
e of the same event. Put differently, L is a linear transforma-
tion that preserves the quantity T 2 − (X2 + Y 2 + Z2 ), upon which both observers
must agree.
Now imagine that the spacetime coordinate origin emits a flash of light—an
origin-centred sphere whose radius increases at the speed of light. It turns out that
any given L is completely determined by its effect on the coordinates of the light
rays that make up this flash. Here is the next crucial idea: in Ex. 8 we explain how
we may set up a one-to-one correspondence between these light rays and complex numbers.
Thus each Lorentz transformation of spacetime induces a definite mapping of the
complex plane. What kinds of complex mappings do we obtain in this way? The
miraculous answer turns out to be this:
The complex mappings that correspond to the Lorentz transformations
are the Möbius transformations! Conversely, every Möbius transformation (3.2)
of C yields a unique Lorentz transformation of spacetime.
Even among professional physicists, this “miracle” is not as well known as it
should be.
 Inversion 139

The connection exhibited in (3.2) is deep and powerful. Just for starters, it means
that any result we establish concerning Möbius transformations will immediately
yield a corresponding result in Einstein’s Theory of Relativity. Furthermore, these
Möbius transformation proofs turn out to be considerably more elegant than direct
spacetime proofs.
To really understand the above claims, we strongly recommend that after read-
ing this chapter you consult Penrose and Rindler (1984, Ch. 1), as well as Shaw
(2006, Ch. 23) and Needham (2021, §6.4).

3.1.3 Decomposition into Simple Transformations

As a first step towards making sense of (3.1), let us decompose M(z) [exercise] into
the following sequence of transformations:

(i) z ÞÑ z + dc , which is a translation; 



(ii) z ÞÑ (1/z); 
(3.3)
(iii) z ÞÑ − (ad−bc) z, which is an expansion and a rotation; 


c2


(iv) z ÞÑ z + ac , which is another translation.
Note that if (ad−bc) = 0 then M(z) is an uninteresting constant mapping, sending
every point z to the same image point (a/c); in this exceptional case M(z) is called
singular. In discussing Möbius transformations we shall therefore always assume
that M(z) is non-singular, meaning that (ad − bc) ̸= 0.
Of the four transformations above, only the second one has not yet been
investigated. This mapping z ÞÑ (1/z) holds the key to understanding Möbius
transformations; we shall call it complex inversion. The next section examines its
many remarkable and powerful properties.

3.2 Inversion

3.2.1 Preliminary Definitions and Facts

The image of z = r eiθ under complex inversion is 1/(r eiθ ) = (1/r) e−iθ : the new
length is the reciprocal of the original, and the new angle is the negative of the
original. See [3.1a]. Note how a point outside the unit circle C is mapped to a point
inside C, and vice versa.
Figure [3.1a] also illustrates a particularly fruitful way of decomposing complex
inversion into a two-stage process:
(i) Send z = r eiθ to the point that is in the same direction as z but that has
reciprocal length, namely the point (1/r) eiθ = (1/z).
(ii) Apply complex conjugation (i.e., reflection in the real axis), which sends (1/z)
to (1/z) = (1/z).
Check for yourself that the order in which we apply these mappings is immaterial.
140 Möbius Transformations and Inversion

[3.1] [a] Geometric inversion in the unit circle C: z ÞÑ IC (z) ≡ (1/z) leaves the points
of C fixed, and swaps the interior and exterior. When composed with conjugation, it
yields complex inversion, z ÞÑ (1/z). [b] Definition of geometric inversion z ÞÑ IK (z) in
a general circle K.

While stage (ii) is geometrically trivial, we shall see that the mapping in stage (i)
is filled with surprises; it is called3 geometric inversion, or simply inversion. Clearly,
the unit circle C plays a special role for this mapping: the inversion interchanges
the interior and exterior of C, while each point on C remains fixed (i.e., is mapped
to itself). For this reason we write the mapping as z ÞÑ IC (z) = (1/z), and we call
IC (a little more precisely than before) “inversion in C”.
This added precision in terminology is important because, as illustrated in [3.1b],
there is a natural way of generalizing IC to inversion in an arbitrary circle K (say
with centre q and radius R). Clearly, this “inversion in K”, written z ÞÑ e z = IK (z),
should be such that the interior and exterior of K are interchanged, while each point
on K remains fixed. If ρ is the distance from q to z, then we define e z = IK (z) to be
the point in the same direction from q as z, and at distance (R2 /ρ) from q. [Check for
yourself that this definition does indeed perform as advertised.]
As usual, we invite you to use a computer to verify empirically the many results
we shall derive concerning inversion. However, in the case of this particular map-
ping, you can also construct (fairly easily) a mechanical instrument that will carry
out the mapping for you; see Ex. 2.
Although we shall not need it for a while, it is easy enough to obtain a formula for
IK (z). Because the connecting complex numbers from q to z and to e z both have the

3
In older works it is often called “transformation by reciprocal radii”.
 Inversion 141

[3.2] [a] Ordinary reflection in a line. [b] If the radius of K is large, so that it starts to look
straight, then geometric inversion z ÞÑ ez = IK (z) looks very much like ordinary reflection.
More precisely, as z approaches K, then IK (z) is ultimately its reflection in the tangent
to K. [c] Under inversion in K, the triangle aqb is similar to the triangle bqee a.

same direction, and their lengths are ρ and (R2 /ρ), it follows that (e
z−q)(z − q) = R2 .
Solving for e
z,
R2 q z + (R2 − |q|2 )
IK (z) = +q= . (3.4)
z−q z−q
For example, if we put q = 0 and R = 1, then we recover IC (z) = (1/z).
There is a very interesting similarity (which will deepen as we go on) between
inversion IK (z) in a circle K and reflection ℜL (z) in a line L. See [3.2a] and [3.2b].
First, L divides the plane into two pieces, or “components”, which are interchanged
by ℜL (z); second, each point on the boundary between the components remains
fixed; third, ℜL (z) is involutory or self-inverse, meaning that ℜL ◦ ℜL is the identity
mapping, leaving every point fixed. To put this last property differently, consider
a point z and its reflection e z = ℜL (z) in L. Such a pair are said to be “mirror
images”, or to be “symmetric with respect to L”. The involutory property says that
the reflection causes such a pair of points to swap places.
Check for yourself that IK (z) shares all three of these properties. Furthermore,
the black triangle in [3.2b] illustrates the fact that if K is large then the effect of IK on
a small shape close to K looks very much like ordinary reflection. [We will explain
this later, but you might like to check this empirically using a computer.] For these
reasons, and others still to come, IK (z) is often also called reflection in a circle, and
the pair of points z and e z = IK (z) are said to be symmetric with respect to K.
We end this subsection with two simple properties of inversion, the first of which
will serve as the springboard for the investigations that follow. Let us use the sym-
bol [cd] to stand for the distance |c − d| between two points c and d. We hope that
142 Möbius Transformations and Inversion

no confusion will arise from this, the square brackets serving as a reminder that
[cd] is not the product of the complex numbers c and d.
e = IK (b) are their
e = IK (a) and b
In [3.2c], a and b are two arbitrary points, and a
images under inversion in K. By definition, [qa][qe 2
a ] = R = [qb][qbe ], and so
e
[qa]/[qb] = [qb]/[qe
a].
Noting the common angle ∠aqb = ∠e e we deduce that
aqb,
e
e and b,
If inversion in a circle centred at q maps two points a and b to a
e a are similar. (3.5)
then the triangles aqb and bqe
Lastly, let us find the relationship between the separation [ab] of two points, and
the separation [eabe ] of their images under inversion. Using (3.5),

[e
abe ]/[ab] = [qb
e ]/[qa] = R2 /[qa][qb],

and so the separation of the image points is given by

 2

e] = R
[e
ab [ab]. (3.6)
[qa][qb]

3.2.2 Preservation of Circles

Let us examine the effect of IK on lines and then on circles. If a line L passes through
the centre q of K, then clearly IK maps L to itself, which we may write as IK (L) = L.
Of course we don’t mean that each point of L remains fixed, for IK interchanges the
portions of L interior and exterior to K; the only points of L that remain fixed are the
two places where it intersects K.
Matters become much more interesting when we consider a general line L that
does not pass through q. Figure [3.3] provides the surprising answer:
If a line L does not pass through the centre q of K, then inversion in K
(3.7)
maps L to a circle that passes through q.
Here b is an arbitrary point on L, while a is the intersection of L with the perpen-
dicular line through q. By virtue of (3.5), ∠qbeea = ∠qab = (π/2), so b e lies on the
circle having the line-segment qe a as diameter. Done. Notice, incidentally, that the
tangent at q of the image circle is parallel to L.
Note that (3.7) makes no mention of the radius R of K. You may therefore be con-
cerned that in [3.3] we have chosen R so that K does not intersect L; what happens
if K does intersect L? Check for yourself that, while the picture looks somewhat dif-
ferent in this case, the geometric argument above continues to apply without any
modification.
We now give a less direct, but more instructive way of understanding why (3.7)
does not depend on the size of K. We will show that if the result holds for one
 Inversion 143

[3.3] The geometric inverse of a line is a circle through the centre of inversion. By virtue
of [3.2c], the illustrated right triangles are similar, so we see that the geometric inverse of
a line L is a circle that passes through the centre of the inversion, and that its tangent
there is parallel to L.

circle K1 (radius R1 ) centred at q, then it will hold for any other circle K2 (radius R2 )
centred at q.
Let z be an arbitrary point, and let e
z1 = IK1 (z) and e z2 = IK2 (z). Obviously e
z1 and
e
z2 are both in the same direction from q as z, and you can easily check that the ratio
of their distances from q is independent of the location of z:

[q e
z2 ]/[q e
z1 ] = (R2 /R1 )2 ≡ k, say.

Thus

IK2 = Dkq ◦ IK1 , (3.8)

where the “central dilation” Dkq [see p. 45] is an expansion (centred at q) of the
plane by a factor of k. It follows [exercise] that if (3.7) holds for K1 then it also holds
for K2 .
Look again at [3.3]. Since IK is involutory, it simply swaps the line and circle, and
so the image of any circle through q is a line not passing through q. But what happens to
a general circle C that does not pass through q? Initially, suppose that C does not
contain q in its interior. Figure [3.4] provides the beautiful answer:

If a circle C does not pass through the centre q of K, then inversion in K

(3.9)
maps C to another circle not passing through q.

This fundamental result is often described by saying that inversion “preserves

circles”.
It follows from (3.8) that if (3.9) is true for one choice of K, then it will be true
for any choice of K. We may therefore conveniently choose K so that C lies inside
144 Möbius Transformations and Inversion

[3.4] Geometric inversion maps circles to circles! This beautiful and important fact is
readily confirmed by first observing that [3.2c] implies that the two grey angles are equal
to each other, and, likewise, the two black angles are equal to each other. But this means
that the right angle at c, being the difference of the grey and black angles, implies the
illustrated right angle at e
c. Done!

it, as illustrated. Here a and b are the ends of a diameter of C, and they therefore
subtend a right angle at a general point c on C. To understand (3.9), first use (3.5)
to check that both the shaded angles are equal, and that both the black angles are
equal. Next look at the triangle abc, and observe that the external shaded angle at
a is the sum of the two illustrated internal angles: the right angle at c and the black
angle at b. It follows that ∠e aecbe = (π/2), and hence that a e are the ends of a
e and b
diameter of a circle through e c. Thus we have demonstrated (3.9) in the case where
C does not contain q. We leave it to you to check that the same line of reasoning
establishes the result in the case where C does contain q.
The result (3.7) is in fact a special limiting case of (3.9). Figure [3.5] shows a line
L, the point p on L closest to the centre q of the inversion, and a circle C tangent to
L at p. As its radius tends to infinity, C tends to L, and the image circle C e = IK (C)
tends to a circle through q.
Later we will be able to give a much cleaner way of seeing that (3.7) and (3.9) are
two aspects of a single result.

3.2.3 Constructing Inverse Points Using Orthogonal Circles

Consider [3.6a]. The circle C cuts the circle of inversion K at right angles at a and
b. In other words, the tangent T to C at a (for example) passes through q. Under
inversion in K, a and b remain fixed, and T is mapped into itself. Thus the image of
C must be a circle that again passes through a and b and that is again orthogonal
to K. But clearly there is only one circle with these properties, namely C itself. Thus,
Under inversion in K, every circle orthogonal to K is mapped to itself. (3.10)
 Inversion 145

[3.5] Inversion of a line is the limit of inversion of a circle. Here we see that [3.3] is in
fact merely a special limiting case of the general result shown in [3.4]. As the radius of
C tends to infinity, so that C becomes the line L, the image circle C e = IK (C) becomes a
circle through q.

[3.6] Constructing the geometric inverse using orthogonal circles. [a] If C is orthogonal
to K, then it is mapped to itself, because the image must be another circle that again
passes through the fixed points a and b, and the tangent T at a is mapped into itself.
Thus it can only be C itself. [b] It follows that the geometric inverse e
z of a point z may
be constructed as the second intersection point of any two circles through z that are
orthogonal to K.

Figure [3.6a] illustrates two immediate consequences of this result. First, the disc
bounded by C is also mapped to itself, the shaded and hatched regions into which
K divides it being swapped by the inversion. Second, a line from q through a point
z on C intersects C for the second time at the inverse point e z.
Another consequence (the key result of this subsection) is the geometric con-
struction shown in [3.6b], the verification of which is left to you.
The inverse ez of z in K is the second intersection point of any two circles
that pass through z and are orthogonal to K.
146 Möbius Transformations and Inversion

[3.7] Inversion (reflection) in a line is the limit of inversion in a circle. [a] The reflection
e
z of z in the line L can be constructed as the second intersection point of any two circles
through z that are orthogonal to L. [b] The same construction for geometric inversion in
the circle K. As the radius of K tends to infinity, this becomes reflection in the tangent L.

Note that the construction of e z in [3.6a] is the special limiting case in which the
radius of one of the circles tends to infinity, and so becomes a line through q. For
other, less important, geometric constructions of inversion, see Ex. 1.
The previously mentioned analogy between inversion in K and reflection in a
line L now deepens, for the reflection e z = ℜL (z) of z in L can be obtained using
precisely the same construction; see [3.7a]. Note that the line-segment joining z and
e
z is orthogonal to L, and that its intersection p with L is equidistant from z and e z:
[pz]/[p e
z ] = 1.
As illustrated in [3.7b], the segment of L in the vicinity of p can be approximated
by an arc of a large circle K tangent to L at p. Here e z = IK (z) is the image under
inversion in K of the same point z as before. As you can see, there is virtually no
difference between the two figures. More precisely, as the radius of K tends to infin-
ity, inversion in K becomes reflection in L. In particular, [pz]/[p e
z ] tends to unity, or
equivalently, [p e
z ] is “ultimately equal” to [pz]. We can now understand what was
happening in figure [3.2b].
We can also check this result algebraically. First, though, observe that from the
geometric point of view it is sufficient to demonstrate the result for a single choice
of the line L and a single point p on it. Let us therefore choose L to be the real axis,
and let p be the origin. The circle K of radius R centred at q = iR is therefore tangent
to L at p. Using (3.4), we obtain [exercise]

z
IK (z) = .
1 − (iz/R)

Thus as R tends to infinity we find that IK (z) is ultimately equal to ℜL (z) = z, as

was to be shown.
Here is another way of looking at the result. Instead of making K larger and
larger, let z move closer and closer to an arbitrary point p on a circle K of fixed size.
As z approaches p from any direction, IK (z) is ultimately equal to ℜT (z), where T
is the tangent to K at p.
 Inversion 147

[3.8] Conformal and anticonformal mappings preserve and reverse angles, respectively.

Again, we can also get this algebraically using the above equation. If R is fixed
and |z| < R, then [exercise]
iz2 z3
IK (z) = z + − 2 + ··· .
R R
Thus as z approaches p = 0, IK (z) is ultimately equal to ℜL (z) = z, which is
reflection in the tangent to K at p.

3.2.4 Preservation of Angles

Let us begin by discussing what is meant by “preservation of angles”. In the centre
of [3.8] are two curves S1 and S2 intersecting at a point p. Provided these curves are
sufficiently smooth at p, then, as illustrated, we may draw their tangent lines T1 and
T2 at p. We now define the “angle between S1 and S2 ” at p to be the acute angle θ
from T1 to T2 . Thus this angle θ has a sign attached to it: the angle between S2 and S1 is
minus the illustrated angle between S1 and S2 . If we now apply a sufficiently smooth
transformation to the curves, then the image curves will again possess tangents
at the image of p, and so there will be a well-defined angle between these image
curves.
If the angle between the image curves is the same as the angle between the
original curves through p, then we say that the transformation has “preserved”
the angle at p. It is perfectly possible that the transformation preserves the angle
between one pair of curves through p, but not every pair through p. However, if
the transformation does preserve the angle between every pair of curves through p,
then we say that it is conformal at p. We stress that this means that both the magni-
tude and the sign of the angles are preserved; see the right of [3.8]. If every angle at
p is instead mapped to an angle of equal magnitude but opposite sign, then we say
that the mapping is anticonformal at p; see the left of [3.8]. If the mapping is con-
formal at every point in the region where it is defined, then we call it a conformal
mapping; if it is instead anticonformal at every point, then we call it an anticonformal
mapping. Finally, if a mapping is known to preserve the magnitude of angles, but we
148 Möbius Transformations and Inversion

[3.9] Through any given point z not on K, there is precisely one circle orthogonal to K
that passes through z in any given direction.

are unable to say whether or not it preserves their sense, then we call it an isogonal
mapping.
It is easy enough to think of concrete mappings that are either conformal or anti-
conformal. For example, a translation z ÞÑ (z + c) is conformal, as is a rotation and
expansion of the plane given by z ÞÑ az. On the other hand, z ÞÑ z is anticonformal,
as is any reflection in a line. The analogy between such a reflection and inversion
in a circle now gets even deeper, for

Inversion in a circle is an anticonformal mapping.

To see this, first look at [3.9]. This illustrates the fact that given any point z not on
K, there is precisely one circle orthogonal to K that passes through z in any given
direction. [Given the point and the direction, can you think how to construct this
circle?]
As in [3.8], suppose that two curves S1 and S2 intersect at p, and that their tan-
gents there are T1 and T2 , the angle between them being θ. To find out what happens
to this angle under inversion in K, let us replace S1 and S2 with the unique circles
orthogonal to K that pass through p in the same directions as directions S1 and S2 ,
i.e., circles whose tangents at p are T1 and T2 . See [3.10a]. Since inversion in K maps
each of these circles to themselves, the new angle at p e = IK (p) is −θ. Done.
Figure [3.10b] illustrates the effect of z ÞÑ (1/z) on angles. Since this mapping
is equivalent to reflection (i.e., inversion) in the unit circle followed by reflection
in the real axis (both of which are anticonformal), we see that their composition
reverses the angle twice, restoring it to its original value:

Complex inversion, z ÞÑ (1/z), is conformal.

 Inversion 149

[3.10] Geometric inversion is anticonformal, so complex inversion is conformal.

[a] When two circles intersect, the magnitudes of both angles of intersection are the same,
by symmetry. But, taking account of the sense of the angles, it follows that geometric
inversion is anticonformal. [b] Since complex conjugation is also anticonformal, its com-
position with geometric inversion reverses angles twice. Therefore, complex inversion,
z ÞÑ (1/z), is conformal.

By the same reasoning, it follows more generally that

The composition of an even number of reflections (in lines or circles) is

a conformal mapping, while the composition of an odd number of such
reflections is an anticonformal mapping.

3.2.5 Preservation of Symmetry

Consider [3.11a], which shows two points a and b that are symmetric with respect
e and L to L,
e, b to b,
to a line L. If reflection in a line M maps a to a e then clearly the
image points a e and be are again symmetric with respect to the image line L.
e In brief,
reflection in lines “preserves symmetry” with respect to lines.
We now show that reflection in circles also preserves symmetry with respect to
circles:
e
If a and b are symmetric with respect to a circle K, then their images a
e
and b under inversion in any circle J are again symmetric with respect
e of K.
to the image K
To understand this, first note that, since inversion is anticonformal, (3.10) is just a
special case of the following more general result:

Inversion maps any pair of orthogonal circles to another pair of orthogonal circles.

Of course if one of the circles passes through the centre of inversion then its image
will be a line. However, if we think of lines as merely being circles of infinite radius
then the result is true without qualification.
150 Möbius Transformations and Inversion

[3.11] Geometric inversion preserves symmetry. [a] If a and b are symmetric with respect
e in a line M are still symmetric with respect to
e and b
to a line L, then their reflections a
e
the reflection L of L: we say that symmetry is preserved. [b] Since geometric inversion
preserves circles and orthogonality, we see that if a and b are symmetric with respect to
e in a circle J are still symmetric with respect to
e and b
a circle K, then their “reflections” a
the reflection Ke of K: here, too, symmetry is preserved!

The preservation of symmetry result is now easily understood. See [3.11b]. Since
the two dashed circles through a and b are orthogonal to K, their images under
e and they therefore intersect in a pair of
inversion in J are likewise orthogonal to K,
e
points that are symmetric with respect to K.

3.2.6 Inversion in a Sphere

Inversion IS of three-dimensional space in a sphere S (radius R and centre q) is
defined in the obvious way: if p is a point in space at distance ρ from q, then IS (p)
is the point in the same direction from q as p, and at distance (R2 /ρ) from q. We
should explain that this is not generalization for its own sake; soon we will see how
this three-dimensional inversion sheds new light on two-dimensional inversion
in C.
Without any additional work, we may immediately generalize most of the above
results on inversion in circles to results on inversion in spheres. For example, recon-
sider [3.3]. If we rotate this figure (in space) about the line through q and a, then we
obtain [3.12], in which the circle of inversion K has swept out a sphere of inversion
S, and the line has swept out a plane Π. Thus we have the following result:

Under inversion in a sphere centred at q, a plane Π that does not contain

q is mapped to a sphere that contains q and whose tangent plane there is
(3.11)
parallel to Π. Conversely, a sphere containing q is mapped to a plane that
is parallel to the tangent plane of that sphere at q.
 Inversion 151

[3.12] Inversion in a sphere sends a plane to a sphere through the centre of inversion.
Rotating [3.3] (in space) about the line through q and a, we see that inversion in the
sphere S maps the plane Π to a sphere that contains q and whose tangent plane there is
parallel to Π.

By the same token, if we rotate figure [3.4] about the line through q and a, then
we find that

Under inversion in a sphere, the image of a sphere that does not contain
the centre of inversion is another sphere that does not contain the centre of
inversion.

This result immediately tells us what will happen to a circle in space under inver-
sion in a sphere, for such a circle may be thought of as the intersection of two
spheres. Thus we easily deduce [exercise] the following result:

Under inversion in a sphere, the image of a circle C that does not pass
through the centre q of inversion is another circle that does not pass through
(3.12)
q. If C does pass through q then the image is a line parallel to the tangent
of C at q.

The close connection between inversion in a circle and reflection in a line

also persists: reflection in a plane is a limiting case of inversion in a sphere.
For this reason, inversion in a sphere is also called “reflection in a sphere”. Of
152 Möbius Transformations and Inversion

[3.13] Let S1 and S2 be intersecting spheres, and let C1 and C2 be the great circles in
which these spheres intersect any plane Π passing through their centres. Then S1 and S2
are orthogonal if and only if C1 and C2 are orthogonal. Also illustrated is the fact that if we
restrict attention to Π then geometric inversion of space in S1 induces two-dimensional
inversion in the circle C1 .

particular importance is the fact that such three-dimensional reflections again

preserve symmetry:

Let K be a plane or sphere, and let a and b be symmetric points with

respect to K. Under a three-dimensional reflection in any plane or sphere, (3.13)
the images of a and b are again symmetric with respect to the image of K.

We now describe the steps leading to this result; they are closely analogous to the
steps leading to the two-dimensional preservation of symmetry result.
If we rotate figure [3.6a] about the line joining the centres of K and C, we deduce
that
Under inversion in a sphere K, every sphere orthogonal to K is mapped
(3.14)
to itself.

When we say that spheres are “orthogonal” we mean that their tangent planes
are orthogonal at each point of their circle of intersection. However, in order to
be able to easily draw on previous results, let us rephrase this three-dimensional
description in two-dimensional terms:

Let S1 and S2 be intersecting spheres, and let C1 and C2 be the great circles
in which these spheres intersect any plane Π passing through their centres.
Then S1 and S2 are orthogonal if and only if C1 and C2 are orthogonal.

See [3.13]. This figure is also intended to help you see that if we restrict attention to
Π then the three-dimensional inversion in S1 is identical to the two-dimensional
inversion in C1 . This way of viewing inversion in spheres allows us to quickly
generalize earlier results.
 Three Illustrative Applications of Inversion 153

For example, referring back to [3.6b], we find—make sure you see this—that if p
e = IS1 (p) may be constructed as the second intersection point of any
lies in Π then p
two circles like C2 that (i) lie in Π, (ii) are orthogonal to C1 , and (iii) pass through p.
Next, suppose that S1 and S2 in [3.13] are subjected to inversion in a third sphere
K. Choose Π to be the unique plane passing through the centres of S1 , S2 , K, and let
C be the great circle in which K intersects Π. Since IC maps C1 and C2 to orthogonal
circles, we deduce [exercise] that (3.14) is a special case of the following result:

Orthogonal spheres invert to orthogonal spheres. (3.15)

Here we are considering a plane to be a limiting case of a sphere.

Putting these facts together, you should now be able to see the truth of (3.13).

3.3 Three Illustrative Applications of Inversion

3.3.1 A Problem on Touching Circles

For our first problem, consider [3.14], in which we imagine that we are given two
circles A and B that touch at q. As illustrated, we now construct the circle C0 that
touches A and B and whose centre lies on the horizontal line L through the centres
of A and B. Finally, we construct the chain of circles C1 , C2 , etc., such that Cn+1
touches Cn , A, and B.

[3.14] Given any two circles A and B that touch at q, as shown, construct touching
circles, Ci , as shown. Then, very surprisingly, (1) the points of contact of the chain Ci all
lie on a circle that also touches A and B at q, and (2) the height of the centre of Cn above
L is 2n times the radius of Cn , as illustrated for n = 3.
154 Möbius Transformations and Inversion

[3.15] Referring to the previous figure, consider the illustrated (unique) circle K centred
at q that cuts C3 at right angles. Geometric inversion in K maps C3 to itself, and it maps A
and B to parallel vertical lines. The truth of both of the mysterious, bulleted facts below
suddenly becomes obvious!

The figure illustrates two remarkable facts about this chain of circles:

• The points of contact of the chain C0 , C1 , C2 , etc., all lie on a circle [dashed]
touching A and B at q.
• If the radius of Cn is rn , then the height above L of the centre of Cn is 2nrn . The
figure illustrates this for C3 .

Before reading further, see if you can prove either of these results using conven-
tional geometric methods.
Inversion allows us to demonstrate both these results in a single elegant swoop.
In [3.14], we have drawn the unique circle K centred at q that cuts C3 at right angles.
Thus inversion in K will map C3 to itself, and it will map A and B to parallel
vertical lines; see [3.15]. Check for yourself that the stated results are immediate
consequences of this figure.

3.3.2 A Curious Property of Quadrilaterals with Orthogonal

Diagonals
Figure [3.16] shows a shaded quadrilateral whose diagonals intersect orthogonally
at q. If we now reflect q in each of the edges of the quadrilateral, then we obtain four
new points. Very surprisingly, these four points lie on a circle4 . As with the previous
problem, see if you can prove this by ordinary means.

4
I am grateful to my friend Paul Zeitz for challenging me with this problem, which appeared in the
USA Mathematical Olympiad.
 Three Illustrative Applications of Inversion 155

[3.16] A curious property of quadrilaterals with orthogonal diagonals. Consider a quadri-

lateral whose diagonals intersect orthogonally at q. If we reflect q in each of the edges of
the quadrilateral, then we obtain four new points. Very surprisingly, these four points lie
on a circle!

[3.17] Geometric explanation of the previous figure. First, on the left, we use the con-
struction in [3.7a] to represent the reflection of q in an edge as the second intersection
point of any two circles through q whose centres lie on that edge. Here we have chosen
the centres of these circles to be the vertices of the quadrilateral. Shown on the right is
the result of performing an inversion in any circle centred at q. The orthogonal circles
become orthogonal lines, intersecting in a rectangle, so these four points lie on a circle.
But this means that the original four points were also on a circle. Done!

To demonstrate the result using inversion, we first use the construction in [3.7a]
to represent the reflection of q in an edge as the second intersection point of any two
circles through q whose centres lie on that edge. More precisely, let us choose the
centres of these circles to be the vertices of the quadrilateral; see the LHS of [3.17].
156 Möbius Transformations and Inversion

Note that, because the diagonals are orthogonal, a pair of these circles centred
at the ends of an edge will intersect orthogonally both at q and at the reflection of
q in that edge.
It follows that if we now apply an inversion in any circle centred at q, then a
pair of such orthogonal circles through q will be mapped to a pair of orthogonal
lines (parallel to the diagonals of the original quadrilateral); see the RHS of [3.17].
Thus the images of the four reflections of q are the vertices of a rectangle, and they
therefore lie on a circle. The desired result follows immediately. Why?

3.3.3 Ptolemy’s Theorem

Figure [3.18a] shows a quadrilateral abcd inscribed in a circle. Ptolemy (c. CE 125)
discovered the beautiful fact that the sum of the product of the opposite sides is the
product of the diagonals. In symbols,
[ad] [bc] + [ab] [cd] = [ac] [bd].
We note that for Ptolemy this was not merely interesting, it was a crucial tool for
doing astronomy! See Ex. 9. His original proof (which is reproduced in most geom-
etry texts) is elegant and simple, but it is very difficult to discover on one’s own.
On the other hand, once one has become comfortable with inversion, the following
proof is almost mechanical.

[3.18] [a] A cyclic quadrilateral is subject to Ptolemy’s Theorem: [ad] [bc] + [ab] [cd] =
[ac] [bd]. [b] Here is the result of inverting figure [a] in any circle K centred at a. Clearly,
ee
[b c ] + [e
cde ] = [b
ede ]. But (3.6) dictates how the separations of these inverted points
are related to the separations of the original points, and Ptolemy’s Theorem then follows
immediately!
 The Riemann Sphere 157

Inverting figure [3.18a] in a circle K centred at one of the vertices (say a), we
obtain [3.18b], in which
ee
[b c ] + [e
cde ] = [b
ede ].

Recalling that (3.6) tells us how the separation of two inverted points is related to
the separation of the original points, we deduce that
[bc] [cd] [bd]
+ = .
[ab][ac] [ac][ad] [ab][ad]
Multiplying both sides by ([ab] [ac] [ad]), we deduce Ptolemy’s Theorem.

3.4 The Riemann Sphere

3.4.1 The Point at Infinity

In discussing inversion we saw that results about lines could always be understood
as special limiting cases of results about circles, simply by letting the radius tend
to infinity. This limiting process is nevertheless tiresome and clumsy; how much
better it would be if lines could literally be described as circles of infinite radius.
Here is another, related inconvenience. Inversion in the unit circle is a one-to-
one mapping of the plane to itself that swaps pairs of points. The same is true of
the mapping z ÞÑ (1/z). However, there are exceptions: no image point is presently
associated with z = 0, nor is 0 to be found among the image points.
To resolve both these difficulties, note that as z moves further and further away
from the origin, (1/z) moves closer and closer to 0. Thus as z travels to ever greater
distances (in any direction), it is as though it were approaching a single point at
infinity, written ∞, whose image is 0. Thus, by definition, this point ∞ satisfies the
following equations:
1 1
= 0, = ∞.
∞ 0
The addition of this single point at infinity turns the complex plane into the so-
called extended complex plane. Thus we may now say, without qualification, that
z ÞÑ (1/z) is a one-to-one mapping of the extended plane to itself.
If a curve passes through z = 0 then (by definition) the image curve under
z ÞÑ (1/z) will be a curve through the point at infinity. Conversely, if the image
curve passes through 0 then the original curve passed through the point ∞. Since
z ÞÑ (1/z) swaps a circle through 0 with a line, we may now say that a line is just
a circle that happens to pass through the point at infinity, and (without further
qualification) inversion in a “circle” sends “circles” to “circles”.
This is all very tidy, but it leaves one feeling none the wiser. We are accustomed
to using the symbol ∞ only in conjunction with a limiting process, not as a thing
in its own right; how are we to grasp its new meaning as a definite point that is
infinitely far away?
158 Möbius Transformations and Inversion

3.4.2 Stereographic Projection

Riemann’s profoundly beautiful answer to this question was to interpret complex
numbers as points on a sphere Σ, instead of as points in a plane. Throughout the
following discussion, imagine the complex plane positioned horizontally in space.
In order to be definite about which way up the plane is, suppose that when we look
down on C from above, a positive (i.e., counterclockwise) rotation of (π/2) carries 1
to i. Now let Σ be the sphere centred at the origin of C, and let it have unit radius
so its “equator” coincides with the unit circle5 .
We now seek to set up a correspondence between points on Σ and points in C. If
we think of Σ as the surface of the Earth, then this is the ancient problem of how to
draw a geographical map. In an atlas you will find many different ways of drawing
maps, the reason for the variety being that no single map can faithfully represent
every aspect of a curved6 surface on a flat piece of paper. Although distortions of
some kind are inevitably introduced, different maps can “preserve” or “faithfully
represent” some (but not all) features of the curved surface. For example, a map
can preserve angles at the expense of distorting areas.
Ptolemy (c. CE 125) was the first to construct such a map, which he used to plot
the positions of heavenly bodies on the “celestial sphere”. His method is called
stereographic projection, and we will soon see how perfectly it is adapted to our
needs. Figure [3.19] illustrates its definition. From the north pole N of the sphere
Σ, draw the line through the point p in C; the stereographic image of p on Σ is the
point p b where this line intersects Σ. Since this gives us a one-to-one correspondence
between points in C and points on Σ, let us also say that p is the stereographic image
of pb . No confusion should arise from this, the context making it clear whether we
are mapping C to Σ, or vice versa.
Note the following immediate facts: (i) the interior of the unit circle is mapped
to the southern hemisphere of Σ, and in particular 0 is mapped to the south pole, S;
(ii) each point on the unit circle is mapped to itself, now viewed as lying on
the equator of Σ; (iii) the exterior of the unit circle is mapped to the northern
hemisphere of Σ, except that N is not the image of any finite point in the plane.
However, it is clear that as p moves further and further away from the origin (in
any direction), pb moves closer and closer to N. This strongly suggests that N is the
stereographic image of the point at infinity. Thus stereographic projection establishes
a one-to-one correspondence between every point of the extended complex plane
and every point of Σ. Instead of merely speaking of a “correspondence” between

5
Some works instead define Σ to be tangent to the complex plane at its south pole.
6
This concept of “curvature” will be defined more precisely in Chapter 6.
 The Riemann Sphere 159

[3.19] Stereographic Projection maps the surface of the (clear glass) sphere to its equa-
b on the surface of
torial plane. Standing at the north pole, N, fire a laser beam at a point p
the sphere: it goes on to hit the plane at p, called the stereographic projection (or image)
of pb . Conversely, p
b is symmetrically called the stereographic projection (or image) of p.
If p moves along a line in the plane, the laser beam sweeps out a plane, so the beam cuts
the sphere in a circle that passes through N, and its tangent there is parallel to the line
in the plane. This is because the equatorial plane and the tangent plane to the sphere at
N are parallel—both are horizontal—so the plane swept out by the laser beam intersects
these parallel planes in parallel lines.

complex numbers and points of Σ, we can imagine that the points of Σ are the com-
plex numbers. For example, S = 0 and N = ∞. Once stereographic projection has
been used to label each point of Σ with a complex number, Σ is called the Riemann
sphere.
We have already discussed the fact that a line in C may be viewed as a circle pass-
ing through the point at infinity. The Riemann sphere now transforms this abstract
idea into a literal fact:

The stereographic image of a line in the plane is a circle on Σ passing

(3.16)
through N = ∞.

To see this, observe that as p moves along the line shown in [3.19], the line con-
necting N to p sweeps out a plane through N. Thus p b moves along the intersection
of this plane with Σ, which is a circle passing through N. Done. In addition, note
that the tangent to this circle at N is parallel to the original line. This is explained in the
caption.
160 Möbius Transformations and Inversion

[3.20] Stereographic Projection is conformal. If two lines intersect in the plane, their
stereographic projections on the sphere are two circles that intersect at the same angle
at N, for the tangents to these circles at N are parallel to the lines in the plane. But, by
symmetry, the angle between these circles at p b is the same as their angle at N, which
we have just seen is the same as the angle between the lines in the plane. Done!

From this last fact it follows that stereographic projection preserves angles. Con-
sider [3.20], which shows two lines intersecting at p, together with their circular,
stereographic projections. By symmetry, the magnitude of the angle of intersec-
tion between the circles is the same at their two intersection points, p b and N. Since
their tangents at N are parallel to the original lines in the plane, it follows that the
illustrated angles at p and p b are of equal magnitude. But before we can say that
stereographic projection is “conformal”, we must assign a sense to the angle on the
sphere.
According to our convention, the illustrated angle at p (from the black curve to
the white one) is positive, i.e., it is counterclockwise when viewed from above the
plane. From the perspective from which we have drawn [3.20], the angle at p b is
negative, i.e., clockwise. However, if we were looking at this angle from inside the
sphere then it would be positive. Thus

If we define the sense of an angle on Σ by its appearance to an observer

(3.17)
inside Σ, then stereographic projection is conformal.

[HISTORICAL NOTE: It is quite remarkable that although stereographic pro-

jection had been well known since Ptolemy first put it to practical use around
125 CE, its beautiful and fundamentally important conformality was not discovered
 The Riemann Sphere 161

[3.21] [a] Two-dimensional stereographic projection of the Riemann sphere Σ to the

plane is merely the restriction to√Σ of the action of three-dimensional geometric
inversion in the sphere K of radius 2 (here shown in cross section) centred at N that
cuts through Σ along its unit-circle equator in C. [b] Inversion of C in the unit circle
induces a reflection of the Riemann sphere in its equatorial plane, C. If we compose this
with conjugation, which reflects Σ in the vertical plane through the real axis, then we
obtain complex inversion z ÞÑ (1/z). But reflection across two perpendicular planes is
equivalent to a rotation of space about their line of intersection through angle π. Thus,
the mapping z ÞÑ (1/z) in induces a rotation of the Riemann sphere about the real
axis through an angle of π.

for another 1500 years! This was first done around 1590 by Thomas Harriot.7 In
Chapter 6 we shall meet Harriot again, for in 1603 he made another profoundly
important discovery about the geometry of the sphere, this time relating the angles
in a spherical triangle to the triangle’s area—(6.9) on page 317.]
Clearly, any origin-centred circle in the plane is mapped to a horizontal circle
on Σ, but what happens to a general circle? The startling answer is that it too is
mapped to a circle on the Riemann sphere! This is quite difficult to see if we stick to our
original definition of stereographic projection, but it suddenly becomes obvious
if we change our point of view. Look again at [3.12], and observe how closely it
resembles the definition of stereographic projection.
To make the connection precise, let K be the sphere centred at the north pole N
of Σ that intersects Σ along its equator (the unit circle of C). Figure [3.21a] shows a
vertical cross section (through N and the real axis), of K, Σ, and C. The full three-
dimensional picture is obtained by rotating this figure about the line through N
and S. We now see that
√
If K is the sphere of radius 2 centred at N, then stereographic projection
is the restriction to C or Σ of inversion in K.

7
See Stillwell (2010, §16.2). For a short sketch of Harriot’s life, see Stillwell (2010, §17.7). The first
full biography of this very remarkable, unsung hero of mathematics and science—born within a mile
of Stephen Hawking’s place of birth (and mine!)—has at last been published: Arianrhod (2019).
162 Möbius Transformations and Inversion

b is its stereographic projection on Σ, then

In other words, if a is a point of C and a
b = IK (a) and a = IK ( a
a b ).
Appealing to our earlier work on inversion in spheres, (3.12) confirms our claim
that

Stereographic projection preserves circles.

Note that (3.16) could also have been derived from (3.12) in this way.

3.4.3 Transferring Complex Functions to the Sphere

Stereographic projection enables us to transfer the action of any complex function
to the Riemann sphere. Given a complex mapping z ÞÑ w = f(z) of C to itself,
we obtain a corresponding mapping b z ÞÑ wb of Σ to itself, where b b are the
z and w
stereographic images of z and w. We shall say that z ÞÑ w induces the mapping
b
z ÞÑ w b of Σ.
For example, consider what happens if we transfer f(z) = z to Σ. Clearly [exer-
cise],

Complex conjugation in C induces a reflection of the Riemann sphere in

the vertical plane passing through the real axis.

For our next example, consider z ÞÑ e z = (1/z), which is inversion in the unit
circle C. Figure [3.21b] shows a vertical cross section of Σ taken through N and the
point z in C. This figure also illustrates the very surprising result of transferring
this inversion to Σ:

Inversion of C in the unit circle induces a reflection of the Riemann

(3.18)
sphere in its equatorial plane, C.

Here is an elegant way of seeing this. First note that not only are the pair of points
z and e z symmetric (in the two-dimensional sense) with respect to C, but they
are also symmetric (in the three-dimensional sense) with respect to the sphere Σ.
Now apply the three-dimensional preservation of symmetry result (3.13). Since z
and ez are symmetric with respect to Σ, their stereographic images b z = IK (z) and
b
e
z = IK ( e
z ) will be symmetric with respect to IK (Σ). But IK (Σ) = C. Done! A more
elementary (but less illuminating) derivation may be found in Ex. 6.
By combining the above results, we can now find the effect of complex inversion
on the Riemann sphere. In C, we know that z ÞÑ (1/z) is equivalent to inversion
in the unit circle, followed by complex conjugation. The induced mapping on Σ
is therefore the composition of two reflections in perpendicular planes through the real
axis—one horizontal, the other vertical. However, it is not hard to see (perhaps
with the aid of an orange) that the net effect of successively reflecting Σ in any
 The Riemann Sphere 163

two perpendicular planes through the real axis is a rotation of Σ about the real axis
through angle π. Thus we have shown that
The mapping z ÞÑ (1/z) in C induces a rotation of the Riemann sphere
(3.19)
about the real axis through an angle of π.
Recall that the point ∞ was originally defined by the property that it be swapped
with 0 under complex inversion, z ÞÑ (1/z). The result (3.19) vividly illustrates the
correctness of identifying N with the point at infinity, for the point 0 in C corre-
sponds to the south pole S of Σ, and the rotation of π about the real axis does indeed
swap S with N.

3.4.4 Behaviour of Functions at Infinity

Suppose two curves in C extend to arbitrarily large distances from the origin.
Abstractly, one would say that they meet at the point at infinity. On Σ this becomes
a literal intersection at N, and if each of the curves arrives at N in a well defined
direction, then one can even assign an “intersection angle at ∞”. For example, [3.20]
illustrates that if two lines in C intersect at a finite point and contain an angle α
there, then they intersect for a second time at ∞ and they contain an angle −α at
that point.
Transferring a complex function to the Riemann sphere enables one to examine
its behaviour “at infinity” exactly as one would at any other point. In particular,
one can look to see if the function preserves the angle between any two curves
passing through ∞. For example, the result (3.19) shows that complex inversion
does preserve such angles at N, and it is therefore said to be “conformal at infin-
ity”. By the same token, this rotation of Σ will also preserve the angle between two
curves that pass through the singularity z = 0 of z ÞÑ (1/z), so complex inversion
is conformal there too. In brief, complex inversion is conformal throughout the extended
complex plane.
In this chapter we have found it convenient to depict z ÞÑ w as a mapping of C to
itself , and in the above example we have likewise interpreted the induced mapping
b
z ÞÑ w b as sending points on the sphere to other points on the same sphere. How-
ever, it is often better to revert to the convention of the previous chapter, whereby
the mapping sends points in the z-plane to image points residing in a second copy
of C, the w-plane. In the same spirit, the induced mapping b z ÞÑ wb may be viewed
as mapping points in one sphere (the z-sphere) to points in a second sphere (the
w-sphere). We illustrate this with an example.
Consider z ÞÑ w = zn , where n is a positive integer. The top half of [3.22] illus-
trates the effect of the mapping (in the case n = 2) on a grid of small “squares”
abutting the unit circle and two rays containing an angle θ. Very mysteriously,
the images of these “squares” in the w-plane are again almost square. In the next
chapter we will show that this is just one consequence of a more basic mystery,
164 Möbius Transformations and Inversion

namely, that z ÞÑ w = zn is conformal. Indeed, we will show that if a map-

ping is conformal, then any infinitesimal shape is mapped to a similar infinitesimal
shape.
Since stereographic projection is known to be conformal, we would therefore
anticipate that when we transfer the grid from the z-plane to the z-sphere, the result
will again be a grid of “squares”. That this does indeed happen can be seen at the
bottom left of [3.22]; the bottom right of [3.22] illustrates the same phenomenon
as we pass from the image grid in the w-plane to the image grid on the w-sphere.
Quite generally, any conformal mapping of C will induce a conformal mapping of
Σ that will (as one consequence) map a grid of infinitesimal squares to another grid
of infinitesimal squares.
Figure [3.22] not only manifests the conformality of z ÞÑ w = z2 , but it also
illustrates that there exist points at which this conformality breaks down. Clearly,
the angle θ at the origin is doubled; more generally, z ÞÑ w = zn multiplies angles
at 0 by n. Quite generally, if the conformality of an otherwise conformal mapping
breaks down at a particular point p, then p is called a critical point of the mapping.
Thus we may say that 0 is a critical point of z ÞÑ w = zn .
If we restrict ourselves to C then this is the only critical point of this mapping.
However, if we look at the induced mapping of Σ, then the figure makes it clear
that in the extended complex plane there is a second critical point at infinity: angles
there are multiplied by n, just as they were at 0. Thus, more precisely than before,
the claim is that z ÞÑ w = zn is a conformal mapping whose only critical points are
0 and ∞.
Next, we discuss how the behaviour of a complex mapping at infinity may be
investigated algebraically. Complex inversion rotates Σ so that a neighbourhood of
N = ∞ becomes a neighbourhood of S = 0. Thus to examine behaviour near infin-
ity we may first apply complex inversion and then examine the neighbourhood of
the origin. Algebraically, this means that to study f(z) at infinity we should study
F(z) ≡ f(1/z) at the origin. For example, f(z) is conformal at infinity if and only if
F(z) is conformal at the origin.
For example, if f(z) = (z + 1)3 /(z5 − z), then F(z) = z2 (1 + z)3 /(1 − z4 ), which
has a double root at 0. Thus instead of merely saying that f(z) “dies away to zero
like (1/z2 ) as z tends to infinity”, we can now say (more precisely) that f(z) has a
double root at z = ∞.
This process can also be used to extend the concept of a branch point of a multi-
function to the point at infinity. For example, if f(z) = log(z) then F(z) = − log(z).
Thus f(z) not only has a logarithmic branch point at z = 0, it also has one at
z = ∞.
 The Riemann Sphere 165

[3.22] The conformality of z ÞÑ zn . In the next chapter we will begin to address the
remarkable and mysterious fact that z ÞÑ zn is conformal, here illustrated in the case
n = 2 [TOP]. We will also see that conformality suffices to explain the illustrated fact
that a grid of small squares is ultimately mapped to a grid of squares, in the limit that
they vanish. Note, however, that it is not conformal at 0: the angles between the rays are
doubled. We see this too at the south pole of the Riemann sphere, and at the north pole,
corresponding to ∞. Since stereographic projection has been proven to be conformal,
it follows that the conformality of z ÞÑ z2 is transferred to the induced mapping of the
Riemann sphere [BOTTOM], and therefore it, too, ultimately maps squares to squares.

3.4.5 Stereographic Formulae*

In this subsection we derive explicit formulae connecting the coordinates of a point
z in C and its stereographic projection bz on Σ. These formulae will prove useful in
investigating non-Euclidean geometry, but if you don’t plan to study Chapter 6
then you should feel free to skip this subsection.
166 Möbius Transformations and Inversion

To begin with, let us describe z with Cartesian coordinates: z = x + iy. Similarly,

let (X, Y, Z) be the Cartesian coordinates of b z on Σ; here the X- and Y-axes are chosen
to coincide with the x- and y-axes of C, so that the positive Z-axis passes through
N. To make yourself comfortable with these coordinates, check the following facts:
the equation of Σ is X2 + Y 2 + Z2 = 1, the coordinates of N are (0, 0, 1), and similarly
S = (0, 0, −1), 1 = (1, 0, 0), i = (0, 1, 0), etc.
Now let us find the formula for the stereographic projection z = x + iy of the
point b z on Σ in terms of the coordinates (X, Y, Z) of b z . Let z ′ = X + iY be the foot of
the perpendicular from b z to C. Clearly, the desired point z is in the same direction
as z ′ , so
|z|
z = ′ z ′.
|z |
Now look at [3.23a], which shows the vertical cross section of Σ and C taken
through N and bz ; note that this vertical plane necessarily also contains z ′ and z.
From the similarity of the illustrated right triangles with hypotenuses N b
z and Nz,
we immediately deduce [exercise] that
|z| 1
= ,
|z ′ | 1−Z
and so we obtain our first stereographic formula:
X + iY
x + iy = . (3.20)
1−Z
Let us now invert this formula to find the coordinates of b
z in terms of those of
z. Since [exercise]
1+Z
|z|2 = ,
1−Z
we obtain [exercise]
2z 2x + i2y |z|2 − 1
X + iY = = , and Z= . (3.21)
1 + |z|2 1 + x2 + y2 |z|2 + 1
Although it is often useful to describe the points of Σ with the three coordinates
(X, Y, Z), this is certainly unnatural, for the sphere is intrinsically two dimensional.
If we instead describe b z with the more natural (two-dimensional) spherical polar
coordinates (ϕ, θ) then we obtain a particularly neat stereographic formula.
First recall8 that θ measures angle around the Z-axis, with θ = 0 being assigned
to the vertical half-plane through the positive X-axis: thus for a point z in C, the
angle θ is simply the usual angle from the positive real axis to z. The definition
of ϕ is illustrated in [3.23b]—it is the angle subtended at the centre of Σ by the
points N and b z : for example, the equator corresponds to ϕ = (π/2). By convention,
0 ⩽ ϕ ⩽ π.

8
This is the American convention; in my native England the roles of θ and ϕ are the reverse of
those stated here.
 The Riemann Sphere 167

[3.23] Two stereographic formulae. [a] From the similarity of the illustrated triangles, we
immediately obtain a Cartesian formula for z = x + iy = X+iY 1−Z in terms of the Cartesian
coordinates (X, Y, Z) of the point b z on the sphere. This can be inverted to find b
z in terms
of z; see (3.21). [b] If we instead use geographical coordinates (θ, ϕ) to describe b
z on the
sphere, the figure shows that we obtain a particularly elegant and useful polar formula:
z = cot(ϕ/2) eiθ .

If z is the stereographic projection of the point b

z having coordinates (ϕ, θ), then
iθ
clearly z = r e , and so it only remains to find r as a function of ϕ. From [3.23b] it is
clear [exercise] that the triangles N b
z S and N0 z are similar, and because the angle
∠NS b z = (ϕ/2), it follows [exercise] that r = cot(ϕ/2). Thus our new stereographic
formula is
z = cot(ϕ/2) eiθ . (3.22)
We will now illustrate this formula with two applications. In Ex. 8 we also show
how this formula may be used to establish a beautiful alternative interpretation of
stereographic projection, due to Sir Roger Penrose.
As our first application, let us rederive the result (3.19). As above, let bz be a
general point of Σ having coordinates (ϕ, θ), and let b ze be the point to which it is
carried when we rotate Σ by π about the real axis. Check for yourself (perhaps
ze are (π − ϕ, −θ). Thus if e
with the aid of an orange) that the coordinates of b z is the
stereographic image of b ze , then
 
π ϕ −iθ 1 1
e
z = cot − e = e−iθ = ,
2 2 cot(ϕ/2) z
as was to be shown.
For our second application, recall that if two points on a sphere are diametrically
opposite each other (such as the north and south poles) then they are said to be
antipodal. Let us show that
b and q
If p b are antipodal points of Σ, then their stereographic projections
p and q are related by the following formula: (3.23)
q = −(1/p).
168 Möbius Transformations and Inversion

Put differently, q = −IC (p), where C is the unit circle. Note that the relationship
between p and q is actually symmetrical (as clearly it should be): p = −(1/q). To
b has coordinates (ϕ, θ) then q
verify (3.23), first check for yourself that if p b has
coordinates (π − ϕ, π + θ). The remainder of the proof is almost identical to the
previous calculation. For an elementary geometric proof, see Ex. 6.

3.5 Möbius Transformations: Basic Results

3.5.1 Preservation of Circles, Angles, and Symmetry

From (3.3) we know that a general Möbius transformation M(z) = az+b cz+d can be
decomposed into the following sequence of more elementary transformations: a
translation, complex inversion, a rotation, an expansion, and a second translation.
Since each of these transformations preserves circles, angles, and symmetry, we
immediately deduce the following fundamental results:

• Möbius transformations map circles to circles.9

• Möbius transformations are conformal.
• If two points are symmetric with respect to a circle, then their images under a Möbius
transformation are symmetric with respect to the image circle. This is called the
“Symmetry Principle”.

We know that a circle C will map to a circle—of course lines are now included as
“circles”—but what will happen to the disc bounded by C? First we give a useful
way of thinking about this disc. Imagine yourself walking round C moving coun-
terclockwise; your motion gives C what is a called a positive sense or orientation. Of
the two regions into which this positively oriented circle divides the plane, the disc
may now be identified as the one lying to your left.
Now consider the effect of the four transformations in (3.3) on the disc and on
the positively oriented circle bounding it. Translations, rotations, and expansions
all preserve the orientation of C and map the interior of C to the interior of the
image C e of C. However, the effect of complex inversion on C depends on whether
or not C contains the origin. If C does not contain the origin, then Ce has the same
e This is easily
orientation as C, and the interior of C is mapped to the interior of C.
understood by looking at [3.24].
If C does contain the origin then Ce has the opposite orientation and the interior
e
of C is mapped to the exterior of C. If C passes through the origin then its interior
e See [3.25].
is mapped to the half-plane lying to the left of the oriented line C.

9
Remarkably, Carathéodory (1937) proved that this property actually characterizes the Möbius
transformations: they alone have this property!
 Möbius Transformations: Basic Results 169

[3.24] A Möbius transformation maps an oriented circle C to an oriented circle C e in such

e
a way that the region to the left of C is mapped to the region to the left of C.

[3.25] A limiting form of the previous figure has the image disc expand to become a half-
plane, but it is still true that the region to the left of C is mapped to the region to the left
e In the next chapter we will see that this a more universal phenomenon, having little
of C.
to do with the specific geometry of Möbius transformations.

To summarize,
A Möbius transformation maps an oriented circle C to an oriented circle
e in such a way that the region to the left of C is mapped to the region to
C (3.24)
e
the left of C.

3.5.2 Non-Uniqueness of the Coefficients

To specify a particular Möbius transformation M(z) = az+b
cz+d it seems that we need
to specify the four complex numbers a, b, c, and d, which we call the coefficients of
170 Möbius Transformations and Inversion

the Möbius transformation. In geometric terms, this would mean that to specify a
particular Möbius transformation we would need to know the images of any four
distinct points. This is wrong.
If k is an arbitrary (non-zero) complex number then
az + b kaz + kb
= M(z) = .
cz + d kcz + kd
In other words, multiplying the coefficients by k yields one and the same mapping,
and so only the ratios of the coefficients matter. Since three complex numbers are suf-
ficient to pin down the mapping—(a/b), (b/c), (c/d), for example—we conjecture
(and later prove) that
There exists a unique Möbius transformation sending any three points
(3.25)
to any other three points.
In the course of gradually establishing this one result we shall be led to further
important properties of Möbius transformations.
If you read the last section of Chapter 1, then (3.25) may be ringing a bell: the
similarity transformations needed to do Euclidean geometry are also determined
by their effect on three points. Indeed, we saw in that chapter that such similarities
can be expressed as complex functions of the form f(z) = az + b, and so they actu-
ally are Möbius transformations, albeit of a particularly simple kind. However, for
such a similarity to exist, the image points must form a triangle that is similar to the
triangle formed by the original points. But in the case of Möbius transformations
there is no such restriction, and this opens the way to more flexible, non-Euclidean
geometries in which Möbius transformations play the role of the “motions”. This
is the subject of Chapter 6.
Let us make a further remark on the non-uniqueness of the coefficients of a
Möbius transformation. Recall from the beginning of this chapter that the inter-
esting Möbius transformations are the non-singular ones, for which (ad − bc) ̸= 0.
For if (ad − bc) = 0 then M(z) = az+b cz+d crushes the entire plane down to the
single point (a/c). If M is non-singular, then we may multiply its coefficients by
√
k = ±1/ ad − bc, in which case the new coefficients satisfy
(ad − bc) = 1;
the Möbius transformation is then said to be normalized. When investigating the
properties of a general Möbius transformation, it turns out to be very convenient to
work with this normalized form. However, when doing calculations with specific
Möbius transformations, it is usually best not to normalize them.

3.5.3 The Group Property

In addition to preserving circles, angles, and symmetry, the mapping
az + b
z ÞÑ w = M(z) = (ad − bc) ̸= 0
cz + d
 Möbius Transformations: Basic Results 171

is also one-to-one and onto. This means that if we are given any point w in the w-
plane, there is one (and only one) point z in the z-plane that is mapped to w. We
can show this by explicitly finding the inverse transformation w ÞÑ z = M−1 (w).
Solving the above equation for z in terms of w, we find [exercise] that M−1 is also
a Möbius transformation:
dz − b
M−1 (z) = . (3.26)
−cz + a
Note that if M is normalized, then this formula for M−1 is automatically normalized
as well.
If we look at the induced mapping on the Riemann sphere, then we find that a
Möbius transformation actually establishes a one-to-one correspondence between
points of the complete z-sphere and points of the complete w-sphere, including
their points at infinity. Indeed you may easily convince yourself that

M(∞) = (a/c) and M(−d/c) = ∞.

Using (3.26), you may check for yourself that M−1 (a/c) = ∞ and M−1 (∞) =
−(d/c).
Next, consider the composition M ≡ (M2 ◦ M1 ) of two Möbius transformations,
a2 z + b 2 a1 z + b1
M2 (z) = and M1 (z) = .
c2 z + d2 c1 z + d1
A simple calculation [exercise] shows that M is also a Möbius transformation:
(a2 a1 + b2 c1 )z + (a2 b1 + b2 d1 )
M(z) = (M2 ◦ M1 )(z) = . (3.27)
(c2 a1 + d2 c1 )z + (c2 b1 + d2 d1 )
It is clear geometrically that if M1 and M2 are non-singular, then so is M. This is
certainly not obvious algebraically, but later in this section we shall introduce a
new algebraic approach that does make it obvious.
If you have studied “groups”, or if you read the final section of Chapter 1, then
you will realize that we have now established the following: The set of non-singular
Möbius transformations forms a group under composition. For, (i) the identity mapping
E(z) = z belongs to the set; (ii) the composition of two members of the set yields a
third member of the set; (iii) every member of the set possesses an inverse that also
lies in the set.

3.5.4 Fixed Points

As another step towards establishing (3.25), let us show that if a Möbius trans-
formation exists mapping three given points to three other given points, then it is
unique. To this end, we now introduce the extremely important concept of the fixed
points of a Möbius transformation. Quite generally, p is called a fixed point of a
mapping f if f(p) = p, in which case one may also say that p is “mapped to itself”,
172 Möbius Transformations and Inversion

or that it “remains fixed”. Note that under the identity mapping, z ÞÑ E(z) = z,
every point is a fixed point.
By definition, then, the fixed points of a general Möbius transformation M(z)
are the solutions of
az + b
z = M(z) = .
cz + d
Since this is merely a quadratic in disguise, we deduce that

With the exception of the identity mapping, a Möbius transformation has

at most two fixed points.

From the above result it follows that if a Möbius transformation is known to have
more than two fixed points, then it must be the identity. This enables us to establish
the uniqueness part of (3.25). Suppose that M and N are two Möbius transforma-
tions that both map the three given points (say q, r, s) to the three given image
points. Since (N−1 ◦ M) is a Möbius transformation that has q, r, and s as fixed
points, we deduce that it must be the identity mapping, and so N = M. Done.
We now describe the fixed points explicitly. If M(z) is normalized, then the two
fixed points ξ+ , ξ− are given by [exercise]
p
(a − d) ± (a + d)2 − 4
ξ± = . (3.28)
2c
In the exceptional case where (a + d) = ±2, the two fixed points ξ± coalesce into
the single fixed point ξ = (a − d)/2c. In this case the Möbius transformation is
called parabolic.

3.5.5 Fixed Points at Infinity

Let us now briefly outline how the fixed point can be used to classify the Möbius
transformations into just four achetypes. The full mathematical details are worked
out later, in Section 3.7. Recall our earlier observation (3.2), that, miraculously, the
Möbius transformations exactly correspond to the symmetries (called Lorentz trans-
formations) of Minkowski and Einstein’s spacetime. Thus this classification of the
Möbius transformations is important for relativity theory, too! We cannot explore
this in detail here, but see Needham (2021, §6.4) for the technical details of the
spacetime interpretation via Lorentz transformations
Provided c ̸= 0 then the fixed points both lie in the finite plane; we now dis-
cuss the fact that if c = 0 then at least one fixed point is at infinity. If c = 0 then the
Möbius transformation takes the form M(z) = Az+B, which represents, as we have
mentioned, the most general “direct” (i.e., conformal) similarity transformation of
the plane. If we write A = ρ eiα then this may be viewed as the composition of an
 Möbius Transformations: Basic Results 173

[3.26] Classification of Möbius and Lorentz transformations. Each of the four types
of transformation has two names, depending on whether it is viewed as acting on C
[name on the left] or on spacetime [name on the right]. The mathematical classification
depends on analysing the fixed points, and this is done in detail in Section 3.7. For
the technical details of the spacetime interpretation via Lorentz transformations, see
Needham (2021, §6.4).

origin-centred rotation of α, an origin-centred expansion by ρ, and finally a trans-

lation of B. Let us visualize each of these three transformations on the Riemann
sphere.
With α > 0, figure [3.26a] illustrates that the rotation z ÞÑ eiα z in C induces
an equal rotation of Σ about the vertical axis through its centre. Horizontal cir-
cles on Σ rotate (in the direction of the arrows) into themselves and are therefore
called invariant curves of the transformation. This figure makes it vividly clear that
the fixed points of such a rotation are 0 and ∞. Note also that the (great) circles
174 Möbius Transformations and Inversion

through these fixed points (which are orthogonal to the invariant circles) are per-
muted among themselves. This pure rotation is the simplest, archetypal example
of a so-called elliptic Möbius transformation.
With ρ > 1, figure [3.26b] illustrates the induced transformation on Σ corre-
sponding to the origin-centred expansion of C, z ÞÑ ρz. If ρ < 1 then we have a
contraction of C, and points on Σ move due South instead of due North. Again it
is clear that the fixed points are 0 and ∞, but the roles of the two families of curves
in [3.26a] are now reversed: the invariant curves are the great circles through the
fixed points at the poles, and the orthogonal horizontal circles are permuted among
themselves. This pure expansion is the simplest, archetypal example of a so-called
hyperbolic Möbius transformation.
Figure [3.26c] shows the combined effect of the rotation and expansion in [3.26a]
and [3.26b]. Here the invariant curves are the illustrated “spirals”; however, the two
families of circles in [3.26a] (or [3.26b]) are both invariant as a whole, in the sense
that the members of each family are permuted among themselves. This rotation
and expansion is the archetypal loxodromic Möbius transformation, of which the
elliptic and hyperbolic transformations are particularly important special cases.
Finally, [3.26d] illustrates a translation. Since the invariant curves in C are the
family of parallel lines in the direction of the translation, the invariant curves on
Σ are the family of circles whose common tangent at ∞ is parallel to the invariant
lines in C. Since ∞ is the only fixed point, a pure translation is an example of a
parabolic Möbius transformation.
Note the following consequence of the above discussion:

A Möbius transformation has a fixed point at ∞ if and only if it is a

similarity, M(z) = (az + b). Furthermore, ∞ is the sole fixed point if (3.29)
and only if M(z) is a translation, M(z) = (z + b).

Later we will use this to show that each Möbius transformation is equivalent, in a
certain sense, to one (and only one) of the four types shown in [3.26].

3.5.6 The Cross-Ratio

Returning to (3.25), we have already established that if we can find a Möbius trans-
formation M that maps three given points q, r, s to three other given points q e, er, e
s,
then M is unique. It thus remains to show that such an M always exists.
To see this, first let us arbitrarily choose three points q ′ , r ′ , s ′ , once and for all.
Next, suppose we can write down a Möbius transformation mapping three arbi-
trary points q, r, s to these particular three points, q ′ , r ′ , s ′ ; let Mqrs (z) denote
this Möbius transformation. In exactly the same way we could also write down
Mqe er se (z). By virtue of the group property, it is now easy to see that

M = M−1
ee
q e ◦ Mqrs
rs
 Möbius Transformations: Basic Results 175

is a Möbius transformation mapping q, r, s to q ′ , r ′ , s ′ and thence to q e, er, e

s, as was
desired.
Now the real trick is to choose q ′ , r ′ , s ′ in such a way as to make it easy to write
down Mqrs (z). We don’t like to pull rabbits out of hats, but try q ′ = 0, r ′ = 1,
and s ′ = ∞. Along with this special choice comes a special, standard notation: the
unique Möbius transformation mapping three given points q, r, s to 0, 1, ∞ (respectively)
is written [z, q, r, s].
In order to map q to q ′ = 0 and s to s ′ = ∞, the numerator and denominator of
[z, q, r, s] must be proportional to (z − q) and (z − s), respectively. Thus [z, q, r, s] =



k z−qz−s , where k is a constant. Finally, since k r−s = [r, q, r, s] ≡ 1, we deduce
r−q

that
(z − q)(r − s)
[z, q, r, s] = .
(z − s)(r − q)
This is not quite so rabbit-like as it appears. Two hundred years prior to Möbius’
investigations, Girard Desargues had discovered the importance of the expression
[z, q, r, s] within the subject of projective geometry, where it was christened the
cross-ratio of z, q, r, s (in this order10 ). Its significance in that context is briefly
explained in Ex. 14, but the reader is urged to consult Stillwell (2010) for greater
detail and background. We can now restate (3.25) in a more explicit form:
The unique Möbius transformation z ÞÑ w = M(z) sending three points
e, er, e
q, r, s to any other three points q s is given by
e)(er − e
(w − q s) (z − q)(r − s)
e, er, e
= [w, q s] = [z, q, r, s] = . (3.30)
(w − e e)
s)(er − q (z − s)(r − q)
Although we have not done so, in any concrete case one could easily go on to solve
this equation for w, thereby obtaining an explicit formula for w = M(z).
The result (3.30) may be rephrased in various helpful ways. For example, if a
Möbius transformation maps four points p, q, r, s to p e, q
e, er, e
s (respectively) then the
cross-ratio is invariant: [e e, er, e
p, q s] = [p, q, r, s]. Conversely, p, q, r, s can be mapped
e, q
to p e, er, e
s by a Möbius transformation if their cross-ratios are equal.
Recalling (3.24), we also obtain the following:
Let C be the unique circle through the points q, r, s in the z-plane,
oriented so that these points succeed one another in the stated order.
e be the unique oriented circle through q
Likewise, let C e, er, e
s in the w-
e (3.31)
plane. Then the Möbius transformation given by (3.30) maps C to C,
and it maps the region lying to the left of C to the region lying to the
e
left of C.

10
Different orders yield different values; see Ex. 16. Unfortunately, there is no firm convention as
to which of these values is “the” cross-ratio. For example, our definition agrees with Carathéodory
(1964), Penrose and Rindler (1984), and Jones and Singerman (1987), but it is different from the equally
common definition of Ahlfors (1979).
176 Möbius Transformations and Inversion

[3.27] There exists a unique Möbius transformation that maps three ordered points on
e and the region to the left of C
the circle C to three other ordered points on the circle C,
e
is mapped to the region to the left of C.

[3.28] Geometric interpretation of the cross-ratio. The cross-ratio w = [z, q, r, s] can be

vividly pictured as the image of z under the unique Möbius transformation that maps the
oriented circle C through q, r, s to the real axis in such a way that these three points map
to 0, 1, ∞. Also illustrated is the fact that a point p lies on the circle C through q, r, s if
and only if Im [p, q, r, s] = 0.

This is illustrated in [3.27].

This in turn gives us a more vivid picture of the cross-ratio: w = [z, q, r, s] is the
image of z under the unique Möbius transformation that maps the oriented circle
C through q, r, s to the real axis in such a way that these three points map to 0, 1,
∞. If q, r, s induce a positive orientation on C then the interior of C is mapped to
the upper half-plane; if they induce a negative orientation, then the image is the
lower half-plane. This is illustrated in [3.28], from which we immediately deduce
a neat equation for the circle C:

A point p lies on the circle C through q, r, s if and only if

Im [p, q, r, s] = 0. (3.32)
 Möbius Transformations as Matrices* 177

Furthermore, if q, r, s induce a positive orientation on C (as in [3.28]), then

p lies inside C if and only if Im [p, q, r, s] > 0. If the orientation of C
is negative, then the inequality is reversed.
For a more elementary proof of (3.32), see Ex. 15.

3.6 Möbius Transformations as Matrices*

3.6.1 Empirical Evidence of a Link with Linear Algebra

As you were reading about the group property of Möbius transformations, you
may well have experienced déjà vu, for the results we obtained were remarkably
reminiscent of the behaviour of matrices in linear algebra. Before explaining the rea-
son for this connection between Möbius transformations and linear algebra, let us
be more explicit about the empirical evidence for believing that such a connection
exists.
We begin by associating with every Möbius transformation M(z) a correspond-
ing 2 × 2 matrix [M]:
" #
az + b a b
M(z) = ÐÑ [M] = .
cz + d c d

Since the coefficients of the Möbius transformation are not unique, neither is the
corresponding matrix: if k is any non-zero constant, then the matrix k[M] corre-
sponds to the same Möbius transformation as [M]. However, if [M] is normalized
by imposing (ad − bc) = 1, then there are just two possible matrices associated
with a given Möbius transformation: if one is called [M], the other is −[M]; in other
words, the matrix is determined “uniquely up to sign”. This apparently trivial fact
turns out to have deep significance in both mathematics and physics; see Penrose
and Rindler (1984, Ch. 1) and Penrose (2005).
At this point there exists a strong possibility of confusion, so we issue the follow-
ing WARNING: In linear algebra we are—or should be!—accustomed to thinking
of a real 2 × 2 matrix as representing a linear transformation of R2 . For example,
!
0 −1
represents a rotation of the plane through (π/2). That is, when we apply it
1 0


to a vector yx in R2 , we obtain

! ! !  ! 
0 −1 x −y x
= = rotated by (π/2) .
1 0 y x y
 
a b
In stark contrast, the matrix c d
corresponding to a Möbius transformation
generally has complex numbers as its entries, and so it cannot be interpreted as a
linear transformation of R2 . Even if the entries are real, it must not be thought of in
178 Möbius Transformations and Inversion

!
this way. For example, the matrix 01 −10 corresponds to the Möbius transforma-
tion M(z) = −(1/z), which is certainly not a linear transformation of C. To avoid
confusion, we will adopt the following notational convention: We use (ROUND)
brackets for a real matrix corresponding to a linear transformation of R2 or of C, and
we use [SQUARE] brackets for a (generally) complex matrix corresponding to a Möbius
transformation of C.
Despite this warning, we have the following striking parallels between the
behaviour of Möbius transformations and the matrices that represent them:

• The identity Möbius transformation


E(z) = z corresponds to the familiar
1 0
identity matrix, [E] = 0 1 .
 
a b
• The Möbius transformation M(z) with matrix [M] = c d
possesses an
inverse if and only if the matrix possesses an inverse. For recall that [M] is
non-singular if and only if its determinant det[M] = (ad − bc) is non-zero.
• If we look at (3.26), we see that the matrix of the inverse Möbius transformation
M−1 (z) is the same as the inverse matrix [M]−1 . To put this succinctly,

[M−1 ] = [M]−1 .

• In linear algebra we compose two linear transformations by multiplying their

matrices; indeed, this is the origin of the multiplication rule. If we multiply the
matrices [M2 ] and [M1 ] corresponding to the two Möbius transformations M2 (z)
and M1 (z), then we obtain
" # " # " #
a2 b2 a1 b1 a2 a1 + b 2 c1 a2 b1 + b2 d1
= .
c2 d2 c1 d1 c2 a1 + d2 c1 c2 b1 + d2 d1

But look at (3.27)! This is simply the matrix of the composite Möbius transforma-
tion (M2 ◦ M1 )(z). Thus multiplication of Möbius matrices corresponds to composition
of Möbius transformations:

[M2 ] [M1 ] = [M2 ◦ M1 ].

3.6.2 The Explanation: Homogeneous Coordinates

Clearly this cannot all be coincidence, but what is really going on here?! The answer
is simple, yet subtle. To see it we must first describe the complex plane with a
 Möbius Transformations as Matrices* 179

completely new kind of coordinate system. Instead of expressing z = x + iy in

terms of two real numbers, we write it as the ratio of two complex numbers, z1 and z2 :
z1
z= .
z2
The ordered pair of complex numbers [z1 , z2 ] are called homogeneous coordinates
of z. In order that this ratio be well defined we demand that [z1 , z2 ] ̸= [0, 0]. To each
ordered pair [z1 arbitrary, z2 ̸= 0] there corresponds precisely one point z = (z1 /z2 ),
but to each point z there corresponds an infinite set of homogeneous coordinates,
[kz1 , kz2 ] = k[z1 , z2 ], where k is an arbitrary non-zero complex number.
What about a pair of the form [z1 , 0]? By holding z1 fixed as z2 tends to 0, it is
clear that [z1 , 0] must be identified with the point at infinity. Thus the totality of
pairs [z1 , z2 ] provide coordinates for the extended complex plane. The introduction
of homogeneous coordinates thereby accomplishes for algebra what the Riemann
sphere accomplishes for geometry—it does away with the exceptional role of ∞.
Just as we use the symbol R2 to denote the set of pairs (x, y) of real numbers,
so we use the symbol C2 to denote the set of pairs [z1 , z2 ] of complex numbers. To
highlight the distinction between R2 and C2 , we use conventional round brackets
when writing down an element (x, y) of R2 , but we use square brackets for an
element [z1 , z2 ] of C2 .
Just as a linear transformation of R2 is represented by a real 2 × 2 matrix, so a
linear transformation of C2 is represented by a complex 2 × 2 matrix:
" # " # " #" # " #
z1 w1 a b z1 a z1 + b z2
Þ−Ñ = = .
z2 w2 c d z2 c z1 + d z2

But if [z1 , z2 ] and [w1 , w2 ] are thought of as the homogeneous coordinates in C2 of

the point z = (z1 /z2 ) in C and its image point w = (w1 /w2 ), then the above linear
transformation of C2 induces the following (non-linear) transformation of C:
z1 w1 a z1 + b z2 a (z1 /z2 ) + b az + b
z= Þ−Ñ w= = = = .
z2 w2 c z1 + d z2 c (z1 /z2 ) + d cz + d
This is none other than the most general Möbius transformation!
We have thus explained why Möbius transformations in C behave so much
like linear transformations—they are linear transformations, only they act on the
homogeneous coordinates in C2 , rather than directly on the points of C itself.
As with the cross-ratio, homogeneous coordinates first arose in projective geom-
etry, and for this reason they are often also called projective coordinates. See Stillwell
(2010) for greater detail on the history of the idea. We cannot move on without men-
tioning that in recent times these homogeneous coordinates have provided the key
to great conceptual advances (and powerful new computational techniques) in Ein-
stein’s Theory of Relativity. This pioneering body of work on 2-spinors is due to
180 Möbius Transformations and Inversion

Sir Roger Penrose. See Penrose and Rindler (1984, Ch. 1), Penrose (2005), and Shaw
(2006, Ch. 23).

3.6.3 Eigenvectors and Eigenvalues*

The above representation of Möbius transformations as matrices provides an
elegant and practical method of doing concrete calculations. More significantly,
however, it also means that in developing the theory of Möbius transformations
we suddenly have access to a whole range of new ideas and techniques taken from
linear algebra.
We begin with something very simple. We previously remarked that while it is
geometrically obvious that the composition of two non-singular Möbius transfor-
mations is again non-singular, it is far from obvious algebraically. Our new point
of view rectifies this, for recall the following elementary property of determinants:

det{[M2 ] [M1 ]} = det[M2 ] det[M1 ].

Thus if det[M2 ] ̸= 0 and det[M1 ] ̸= 0, then det{[M2 ] [M1 ]} ̸= 0, as was to be shown.

This also sheds further light on the virtue of working with normalized Möbius
transformations. For if det[M2 ] = 1 and det[M1 ] = 1, then det{[M2 ] [M1 ]} = 1. Thus
the set of normalized 2 × 2 matrices form a group—a “subgroup” of the full group
of non-singular matrices.
For our second example, consider the eigenvectors of a linear transformation
   
a b z1
[M] = of C2 . By definition, an eigenvector is a vector z = whose
c d z2
“direction” is unaltered by the transformation, in the sense that its image is simply a
multiple λz of the original; this multiple λ is called the eigenvalue of the eigenvector.
In other words, an eigenvector satisfies the equation
" #" # " #
a b z1 z1
=λ .
c d z2 z2

In terms of the corresponding Möbius transformation in C, this means that z =

(z1 /z2 ) is mapped to M(z) = (λz1 /λz2 ) = z, and so
 
z1
z = (z1 /z2 ) is a fixed point of M(z) if and only if z = is an
z2 (3.33)
eigenvector of [M].

Note that one immediate benefit of this approach is that there is no longer any
real distinction between a finite fixed point and a fixed point at ∞, for the latter
 
z1
merely corresponds to an eigenvector of the form . For example, consider how
0
 Möbius Transformations as Matrices* 181

elegantly we may rederive the fact that ∞ is a fixed point if and only if M(z) is a
similarity transformation. If ∞ is a fixed point then
" # " #" # " #
z1 a b z1 a z1
λ = = .
0 c d 0 c z1

Thus c = 0, λ = a, and M(z) = (a/d)z + (b/d).

Recall that if the matrix [M] represents the Möbius transformation M(z), then
so does the matrix k[M] obtained by multiplying the entries by k. The fact that
eigenvectors carry geometric information about M(z) shows up in the fact that
they are independent of the choice of k. Indeed, if z is an eigenvector of [M] (with
eigenvalue λ) then it is also an eigenvector of k[M], but with eigenvalue kλ:

{k[M]} z = kλ z.

Since the eigenvalue does depend on the arbitrary choice of k, it appears that its
value can have no bearing on the geometric nature of the mapping M(z). Very
surprisingly, however, if [M] is normalized then the exact opposite is true! In the
next section we will show that the eigenvalues of the normalized matrix [M] completely
determine the geometric nature of the corresponding Möbius transformation M(z). In
anticipation of this result, let us investigate the eigenvalues further.
Recall the fact that the eigenvalues of [M] are the solutions of the so-called
 
char-
1 0
acteristic equation, det{[M] − λ[E]} = 0, where [E] is the identity matrix 0 1 . Using
the fact that [M] is normalized, we find [exercise] that the characteristic equation is

λ2 − (a + d)λ + 1 = 0,

which (for later use) may be written as

1
λ+ = a + d. (3.34)
λ
The first thing we notice about this equation is that there are typically two eigen-
values, λ1 and λ2 , and they are determined solely by the value of (a + d). By
inspecting the coefficients of the quadratic we immediately deduce that

λ1 λ 2 = 1 and λ1 + λ2 = (a + d). (3.35)

Thus if we know λ1 , then λ2 = (1/λ1 ). We emphasize this point because it is not

obvious when we simply write down the formula for the eigenvalues:
q
λ 1 , λ2 = 1
2 (a + d) ± (a + d)2 − 4 .
182 Möbius Transformations and Inversion

Aficionados of linear algebra will recognize (3.35) as a special case of the fol-
lowing general result on the eigenvalues λ1 , λ2 , . . . , λn of any n × n matrix
N:
λ1 λ2 . . . λn = det N and λ1 + λ2 + · · · + λn = tr N,
where tr N ≡ (the sum of the diagonal elements of N) is called the trace of N. For
future use, recall the following nice property of the trace function: If N and P are
both n × n matrices, then
tr {NP} = tr {PN}. (3.36)
In the case of 2 × 2 matrices (which is all that we shall ever need) this is easily
verified by a direct calculation [exercise].

3.6.4 Rotations of the Sphere as Möbius Transformations*

This subsection is optional because its main result is only needed in Chapter 6. Fur-
thermore, in that chapter we shall treat the same result in a much better and simpler
way; the only purpose of this subsection is to further illustrate the connections that
exist between Möbius transformations and linear algebra.
Let us investigate what it might mean to say that two vectors p and q in C2 are
“orthogonal”. Two vectors p and q in R2 are orthogonal if and only if their dot
product vanishes:
! !
· ·
p1 q1
p q= = p1 q1 + p2 q2 = 0.
p2 q2
·
Thus it would seem natural to say that p and q are “orthogonal” if p q = 0. This
will not do. In particular, whereas we would like the dot product of any nonzero
   
·
1 1
vector with itself to be a positive real number, we find that = 0, for
i i
example. As it stands, the dot product is not suitable for use in C2 .
·
The standard solution to this difficulty is to generalize the dot product p q to
·
the so-called inner product, ⟨p , q⟩ ≡ p q:
*" # " #+
p1 q1
⟨p , q⟩ = , = p1 q1 + p2 q2 .
p2 q2
We cannot go into all the reasons why this is the “right” generalization, but observe
that it shares the following desirable properties of the dot product:
⟨p , p⟩ ⩾ 0 and ⟨p , p⟩ = 0 if and only if p1 = 0 = p2 ;
⟨p + q , r⟩ = ⟨p , r⟩ + ⟨q , r⟩ and ⟨r , p + q⟩ = ⟨r , p⟩ + ⟨r , q⟩ .
Note, however, that it is not commutative: ⟨q , p⟩ = ⟨p , q⟩.
 Möbius Transformations as Matrices* 183

We now agree that p and q are “orthogonal” if and only if

⟨p , q⟩ = p1 q1 + p2 q2 = 0.

What does this “orthogonality” mean in terms of the points p = (p1 /p2 ) and
q = (q1 /q2 ) whose homogeneous coordinate vectors are p and q? The answer is
surprising. As you may easily check, the above equation says that q = −(1/p), and
so from (3.23) we deduce that
Two vectors in C2 are orthogonal if and only if they are the homogeneous
coordinates of antipodal points on the Riemann sphere.

Suppose we could find a linear transformation [R] of C2 that were analogous to

a rotation—what transformation of the Riemann sphere Σ would be induced by
the corresponding Möbius transformation R(z)? By “analogous to a rotation”, we
mean that [R] preserves the inner product:

⟨[R]p , [R]q⟩ = ⟨p , q⟩ . (3.37)

In particular, [R] maps every pair of orthogonal vectors to another such pair, and
R(z) therefore maps every pair of antipodal points on Σ to another such pair. We
shall not attempt a real proof, but since the transformation of Σ is also known to be
continuous and conformal11 , it can only be a rotation of Σ.
The desired invariance of the inner product (3.37) may be neatly rephrased using
an operation called the conjugate transpose, denoted by a superscript ∗. This opera-
tion takes the complex conjugate of each element in a matrix and then interchanges
the rows and columns:
" #∗ " #∗ " #
∗ p1 ∗ a b a c
p = = [p1 , p2 ] and [R] = = .
p2 c d b d

Since the inner product can now be expressed in terms of ordinary matrix multi-
plication as ⟨p , q⟩ = p∗ q, and since [exercise] {[R]p}∗ = p∗ [R]∗ , we find that (3.37)
takes the form

p∗ {[R]∗ [R]} q = p∗ q.

Clearly this is satisfied if

[R]∗ [R] = [E], (3.38)

and in linear algebra it is shown that this is also a necessary condition.

Matrices satisfying equation (3.38) are extremely important in both mathematics
and physics—they are called unitary matrices. In the present case of normalized 2×2

11
If it were not continuous then it could, for example, exchange points on two antipodal patches of
Σ while leaving the remainder fixed. If it were continuous but anticonformal, then it could map each
point to its antipodal point, or to its reflection in a plane through the centre of Σ.
184 Möbius Transformations and Inversion

matrices, we can easily find the most general unitary matrix [R] by re-expressing
(3.38) as [R]∗ = [R]−1 :
" # " # " #
a c d −b a b
= =⇒ [R] = .
b d −c a −b a

Although we have left some unsatisfactory gaps in the above reasoning, we have
nevertheless arrived at an important truth: The most general rotation of the Riemann
sphere can be expressed as a Möbius transformation of the form

az + b
R(z) = . (3.39)
−bz + a

This was first discovered by Gauss, around 1819.

3.7 Visualization and Classification*

3.7.1 The Main Idea

Although the decomposition (3.3) of a general Möbius transformation M(z) has
proved valuable in obtaining results, it makes M(z) appear much more compli-
cated than it is. In this section we will reveal this hidden simplicity by examining
the fixed points in greater detail; this will enable us to visualize Möbius trans-
formations in a particularly vivid way. In the process we will clarify our earlier
remark that Möbius transformations can be classified into four types, each M(z)
being “equivalent” to one (and only one) of the four types of transformation illus-
trated in [3.26]. The lovely idea behind this classification scheme is due to Felix
Klein.
To begin with, suppose that M(z) has two distinct fixed points, ξ+ and ξ− . Now
look at the LHS of [3.29], and in particular at the family C1 of circles [shown dashed]
passing through the fixed points. If we think of M(z) as a mapping z ÞÑ w =
M(z) of this figure to itself, then each member of C1 is mapped to another member of C1 .
Why?
Still with reference to the LHS of [3.29], suppose that p [not shown] is an arbitrary
point on the line through ξ+ and ξ− , but lying outside the segment connecting the
p
fixed points. If K is the circle of radius [pξ+ ][pξ− ] centred at p, then ξ+ and ξ−
are symmetric with respect to K. Thus K cuts each member of C1 at right angles
(cf. [3.9]). By varying the position of p we thus obtain a family C2 of circles [shown
solid] such that ξ+ and ξ− are symmetric with respect to each member of C2 , and each
member of C2 is orthogonal to each member of C1 .
Now we come to the main idea: to the LHS of [3.29] we apply a Möbius transforma-
tion F(z) that sends one fixed point (say ξ+ ) to 0, and the other fixed point (ξ− ) to ∞. The
 Visualization and Classification* 185

[3.29] The geometric idea behind the classification of the Möbius transformations. Given
a Möbius transformation M(z) with two fixed points, ξ+ and ξ− , two families of circles
naturally arise: the dashed family C1 passing through the fixed points, and the orthogonal
family C2 , such that ξ+ and ξ− are symmetric with respect to each member of C2 . Note
that M(z) must map members of C1 amongst themselves, and likewise members of C2
are mapped amongst themselves. The key idea is to now apply a Möbius transformation
z−ξ+
F(z) = z−ξ −
that sends one fixed point to the origin (south pole of the Riemann sphere)
and the other to ∞ (the north pole). Thus, on the right, the circles C1 become rays, and
the circles C2 become origin-centred circles. Then the induced mapping M f = F◦M◦
F clearly has fixed points 0 and ∞, and must therefore take the form M(
−1 f ez) = me z,
iα
where m = ρ e is called the multiplier; in the illustrated case, ρ = 1 and α = (π/3).
This multiplier m completely characterizes the original Möbius transformation, M, and
therefore serves to classify it. In this manner, we obtain the first three archetypes shown
in [3.26].

RHS of [3.29] shows the image of the LHS under such a Möbius transformation,
the simplest example of which is
z − ξ+
F(z) = .
z − ξ−
[Note that we have not bothered to write this in normalized form.] Since F is
a Möbius transformation, it must map the members of C1 to the circles passing
through 0 and ∞, i.e., to lines through the origin [shown dashed]. Furthermore,
since F is conformal, two such lines must contain the same angle at 0 as the corre-
sponding C1 circles do at ξ+ . We have tried to make this easy to see in our picture
by drawing C1 circles passing through ξ+ in evenly spaced directions, each one
making an angle of (π/6) with the next.
As an aside, observe that we now have a second, simpler explanation of the
existence of the family C2 of circles orthogonal to C1 . Since the illustrated set of
origin-centred circles are orthogonal to lines through 0, their images under F−1
must be circles orthogonal to each member of C1 .
Next, let e e = F(w) be the images under F of z and w = M(z). We
z = F(z) and w
may now think of F as carrying the original Möbius transformation z ÞÑ w = M(z)
186 Möbius Transformations and Inversion

on the left over to a transformation Mf on the right, namely e

z ÞÑ w f e
e = M( z ). More
explicitly,
 

we = F(w) = F (M[z]) = F M F−1 ( ez) ,

and so
f = F ◦ M ◦ F−1 .
M (3.40)

Since Mf is the composition of three Möbius transformations, it is itself a Möbius

transformation. Furthermore, it follows immediately from the construction that
f are 0 and ∞. But we have already seen that if a Möbius
the fixed points of M
transformation leaves these points fixed, it can only be of the form
f e
M( z) = me
z,

where m = ρ eiα is simply a complex number. Geometrically, M f is just a rotation

by α combined with an expansion by ρ.
This complex number m not only constitutes a complete description of the
mapping M f but, as we will see shortly, it also completely characterizes the geo-
metric nature of the original Möbius transformation M. The number m is called
the multiplier of M(z).

3.7.2 Elliptic, Hyperbolic, and Loxodromic Transformations

Before reading on, refresh your memory of the classification (shown in [3.26a,b,c])
of Möbius transformations of the form M( f e z) = m ez.
We call M(z) an elliptic Möbius transformation if M f is elliptic, meaning that the
iα
latter is a pure rotation corresponding to m = e . Since M f is a rotation if and only
if it maps each origin-centred circle to itself, M(z) is elliptic if and only if it maps
each C2 circle to itself. With α = (π/3), the RHS of [3.29] illustrates the effect of M f
on the point e z. On the LHS you can see the corresponding, unambiguous effect of
M: it moves z along its C2 circle till it lies on the C1 circle making angle (π/3) with
the original C1 through z.
Figure [3.30]12 is intended to give a more vivid impression of this same elliptic
transformation. Each shaded “rectangle” is mapped by M(z) to the next one in
the direction of the arrows—some of these regions have been filled with black to
emphasize this. This figure may be viewed as typical, with one exception. Because
we have chosen α = (π/3), six successive applications of M yield the identity, and
one therefore says that M has period 6. More generally, if α = (m/n)2π, where
(m/n) is a fraction reduced to lowest terms, then M has period n. Of course this
is not typical. In general (α/2π) will be irrational, and no matter how many times
we apply M we will never obtain the identity.

12
Shading inspired by Ford (1929).
 Visualization and Classification* 187

[3.30] An Elliptic Möbius transformation has multiplier m = eiα , so that M f is a pure

rotation. See [3.26a]. Each shaded “rectangle” is mapped by M(z) to the next one in the
direction of the arrows—some of these regions have been filled with black to emphasize
this. This figure may be viewed as typical, with one exception. Because we have chosen
α = (π/3), six successive applications of M yield the identity, and one therefore says that
M has period 6.

We call M(z) a hyperbolic Möbius transformation if M f is hyperbolic, meaning that

the latter is a pure expansion corresponding to m = ρ ̸= 1. Since M f is an expansion
if and only if it maps each line through the origin to itself, M(z) is hyperbolic if and
only if it maps each C1 circle to itself. Figure [3.31] illustrates such a transformation
with ρ > 1. Note that if we repeatedly apply this mapping then any shape (such as
the small black square near ξ+ ) is repelled away from ξ+ , eventually being sucked
into ξ− . In this case ξ+ is called the repulsive fixed point and ξ− is called the attractive
fixed point; if m = ρ < 1 then the roles of ξ+ and ξ− are reversed.
Finally, if m = ρ eiα has a general value, and M f is the composition of both a
rotation and an expansion, then M is called a loxodromic Möbius transformation. In
this case neither the C1 circles, nor the C2 circles are invariant. The curves that are
invariant are illustrated in [3.32], which also shows the effect of successive applica-
tions of M to a small square near ξ+ . In studying this figure, you may find it helpful
to note that

The loxodromic Möbius transformation with fixed points ξ± and mul-

tiplier m = ρ eiα is the composition (in either order) of (i) the elliptic
Möbius transformation with multiplier m = eiα and fixed points ξ± ; (3.41)
(ii) the hyperbolic Möbius transformation with multiplier m = ρ and
fixed points ξ± .
188 Möbius Transformations and Inversion

[3.31] A Hyperbolic Möbius transformation has multiplier m = ρ / f is a pure

= 1, so that M
expansion. See [3.26b]. Iterating the mapping pushes everything away from the repulsive
fixed point and sucks everything towards the attractive fixed point.

[3.32] A Loxodromoic Möbius transformation has general multiplier m = ρeiα , with

f expands by ρ and rotates by α. See [3.26c]. Iterating the
ρ ̸= 1 and α ̸= 0, so that M
mapping pushes everything away from the repulsive fixed point and sucks everything
towards the attractive fixed point, moving along the illustrated invariant curves.

Just as in the case of a hyperbolic transformation, note that one fixed point is
repulsive while the other is attractive. In this figure we have taken α > 0 and ρ > 1;
how would it look if α were negative, or if ρ were less than one?
 Visualization and Classification* 189

3.7.3 Local Geometric Interpretation of the Multiplier

In [3.29] we arbitrarily elected to send ξ+ to 0, rather than ξ− . In this sense our
definition of m is clearly ambiguous. How would the new value of m be related to
the old one if we were to instead send ξ− to 0?
Note that (3.40) may be expressed as (F ◦ M) = (M f ◦ F). Writing w = M(z), and
recalling the definition of F, we therefore have
 
w − ξ+ z − ξ+
=m . (3.42)
w − ξ− z − ξ−
[This formula is often called the normal form of the Möbius transformation.] Inter-
changing ξ+ and ξ− in this formula is equivalent to sending ξ− to 0 and ξ+ to ∞,
in which case we obtain
 
w − ξ+ 1 z − ξ+
= .
w − ξ− m z − ξ−
Thus the multiplier has changed from m to (1/m), and both of these values can lay
equal claim to being called “the” multiplier. Let us therefore refine our language
and call the number m occurring in (3.42) the multiplier associated with ξ+ ; we will
sometimes write it as m + to emphasize this. In these terms, we have just shown
that the multipliers associated with the two fixed points are the reciprocals of one another.
Let us try to understand this more geometrically.
Reconsider [3.29], in which the multiplier associated with ξ+ is m = ei(π/3) .
We now seek to interpret m directly in terms of [3.30], without the assistance of
the RHS of [3.29]. The closer we are to ξ+ , the more closely do the members of C2
resemble tiny concentric circles centred at ξ+ . This is easy to understand: (A) as we
examine smaller and smaller neighbourhoods of ξ+ , the C1 circles look more and
more like their tangent lines at ξ+ ; (B) by definition, each C2 cuts every C1 circle
orthogonally.
From these remarks, it is now clear that the local effect of M (in an infinitesimal
neighbourhood of ξ+ ) is a rotation centred at ξ+ through angle (π/3)—this is the
meaning of the multiplier m + = ei(π/3) associated with ξ+ . Of course exactly the
same reasoning applies to the infinitesimal neighbourhood of ξ− , but we see from
[3.30] that the positive rotation at ξ+ forces an equal and opposite rotation at ξ− .
Thus the local effect of M in the neighbourhood of ξ− is a rotation of −(π/3), and
the associated multiplier m − is e−i(π/3) = (1/m + ), as was to be explained.
If we look at [3.31], then we can see the same phenomenon at work in the case
of a hyperbolic transformation. In this figure the multiplier associated with ξ+ is
m = ρ > 1, and this can now be interpreted as saying that the local effect of M
in an infinitesimal neighbourhood of ξ+ is an expansion centred at that point—we
will verify in a moment that the “local expansion factor” is precisely ρ. It is also
clear from the figure that the local effect of M in an infinitesimal neighbourhood
of ξ− is a contraction, so that the multiplier associated with that point is real and
190 Möbius Transformations and Inversion

less than one. However, it is not so clear that this number is precisely (1/ρ), as we
know it must be. This too can be demonstrated geometrically, but let us instead
content ourselves with showing how our original algebraic argument may be re-
interpreted geometrically in terms of the “local effect” of M in the vicinity of each
of the fixed points.
Let us write Z = (z−ξ+ ) and W = (w−ξ+ ) for the complex numbers emanating
from ξ+ connecting that point to z and to its image w = M(z). We have claimed
(and partially verified) that if Z is infinitesimal then the effect of M is to rotate Z by
α and to expand it by ρ: in other words, W = m Z. To verify this, note that (3.42)
can be rewritten as
 
W w − ξ−
=m .
Z z − ξ−
As Z tends to zero, both z and w tend to ξ+ , and so the fraction on the right is
ultimately equal to m . Thus W is ultimately equal to m Z, as was to be shown.
After you have read the next chapter, you will be able to look back at what we
have just done and recognize it as an example of differentiating a complex function.

3.7.4 Parabolic Transformations

We now possess an excellent understanding of Möbius transformations with two
fixed points, so all that remains is to treat the case where M has only one fixed point
ξ, in which case M is called a parabolic Möbius transformation.
Consider the LHS of [3.33], but ignore the arrows for the time being. Here we
have drawn two families of circles: the solid ones all pass through the fixed point
ξ in one direction, and the dashed ones all pass through ξ in the perpendicular
direction. Note that since the two types of circles are orthogonal at ξ, they are also

[3.33] A Parabolic Möbius transformation has a single fixed point, and M( f z

e) = z
e+T
is a translation, with its single fixed point at ∞, as illustrated in [3.26d].
 Visualization and Classification* 191

(by symmetry) orthogonal at their second intersection point. The RHS illustrates
what happens when we send ξ to ∞ by means of the Möbius transformation
1
G(z) = .
z−ξ
Clearly [exercise], the two orthogonal families of circles become two orthogonal
families of parallel lines. Conversely, if we apply G−1 to any two orthogonal fam-
ilies of lines on the right, then on the left we get two orthogonal families of circles
through ξ.
As before, let e
z = G(z) and w e = G(w) be the images on the RHS of z and w =
M(z). Thus the Möbius transformation z ÞÑ w = M(z) on the LHS induces another
Möbius transformation e z ÞÑ w f e
e = M( z ) on the RHS, where
f = G ◦ M ◦ G−1 .
M
f we deduce that M
Since ∞ is the sole fixed point of M, f can only be a translation:

f e
M( z) = e
z + T.

Now suppose that the arrows on the RHS of [3.33] represent the direction of the
translation T . As illustrated, we now draw a grid aligned with T , each shaded
square being carried into the next by M. f On the LHS of [3.33] we thus obtain a
vivid picture of the action of the original parabolic Möbius transformation M: each
solid circle is carried into itself; each dashed circle is carried into another dashed
circle; and each shaded region is carried into the next in the direction of the arrows.
If M(z) = az+b
cz+d is normalized, then we know from (3.28) that it is parabolic if
and only if (a + d) = ±2, in which case ξ = (a − d)/2c. Now let us determine the
corresponding translation T in terms of the coefficients. Since (G ◦ M) = (M f ◦ G),
the so-called normal form of M is given by
1 1
= + T.
w−ξ z−ξ
Since M maps z = ∞ to w = (a/c), we deduce that
1
T= = ±c,
(a/c) − ξ
where the “±” is the arbitrarily chosen sign of (a + d).

3.7.5 Computing the Multiplier*

We have seen how the multiplier m determines the character of a Möbius transfor-
mation, and we now show how we can determine the character of m directly from
the coefficients of M(z) = az+b
cz+d .
Suppose we have already calculated the fixed points ξ± using (3.28), for
example. Since M maps z = ∞ to w = (a/c), we deduce from the normal form
192 Möbius Transformations and Inversion

(3.42) that the multiplier associated with ξ+ is

a − c ξ+
m = . (3.43)
a − c ξ−
For example, consider complex inversion, z ÞÑ (1/z). The fixed points are the
solutions of z = (1/z), namely, ξ± = ±1. Thus the multiplier associated with
ξ+ = 1 is m = −1 = eiπ , which happens to be the same as the multiplier (1/m )
associated with ξ− = −1. Thus complex inversion is elliptic, and an infinitesimal
neighbourhood of either fixed point is simply rotated about that point through
angle π. Try using a computer to check this prediction.
If desired, we can obtain a completely explicit formula for m by substitut-
ing (3.28) into (3.43). If we only want to know the character of the Möbius
transformation, then we can proceed as follows.
It turns out—we will prove it in a moment—that m is related to the coefficients
of the normalized Möbius transformation by the equation,
√ 1
m + √ = a + d. (3.44)
m
Note that the symmetry of this equation implies that if m is a solution, then so
is (1/m ); this is just as it should be. Without bothering to solve (3.44) for m , we
now obtain [exercise] the following algebraic classification: The normalized Möbius
transformation M(z) = az+b cz+d is

elliptic, iff (a + d) is real and |a + d| < 2; 



parabolic, iff (a + d) = ±2;
(3.45)
hyperbolic, iff (a + d) is real and |a + d| > 2;  


loxodromic, iff (a + d) is complex.
Hint: you can get a better feel for this by sketching the graph of y = x + (1/x).
In order to derive (3.44) elegantly, let us use matrices. Rewriting (3.40),

f = [F] [M] [F]−1 =⇒ det[M]
[M] f = det [F][F]−1 det[M] = det[M].

Thus, regardless of whether or not [F] is normalized, [M] is normalized if and only
f is normalized. Since M(z)
if [M] f f =
= m z, its normalized matrix is [exercise] [M]
√ 
m 0
√ . Recalling (3.36), we deduce that
0 1/ m
√ 1  
m + √ = tr [F] [M] [F]−1 = tr [F]−1 [F][M] = tr [M] = a + d,
m
as was to be shown.

3.7.6 Eigenvalue Interpretation of the Multiplier*

If [M] is a linear transformation of C2 , then we saw in (3.33) that its eigenvec-
tors are the homogeneous coordinates of the fixed points of the corresponding
 Visualization and Classification* 193

Möbius transformation M(z). We also claimed that if [M] is normalized then the
eigenvalues completely determine the character of M(z). We can now be more
precise:

If a fixed point of M(z) is represented as an eigenvector (with eigenvalue λ)

of the normalized matrix [M], then the multiplier m associated with the fixed (3.46)
point is given by m = 1/λ2 .

Before proving this result, we illustrate it with the example of complex inver-
sion, z ÞÑ (1/z). We already know that the fixed points are ±1, that the associated
multipliers are both

given
by m = −1, and we easily find [exercise] that the nor-
0 i
malized matrix is i
. If we choose the homogeneous coordinate vector of a
0
" #
z
finite point z to be , then the eigenvectors corresponding to the fixed points
1
" #
±1
z = ±1 are . Since
1
" #" # " # " #" # " #
0 i 1 1 0 i −1 −1
=i and = −i ,
i 0 1 1 i 0 1 1

we see that the eigenvalues are given by λ = ±i, in agreement with (3.46).
√
Returning to the general case, comparison of (3.34) and (3.44) reveals that m
and λ satisfy the same quadratic, so we immediately deduce most of (3.46): the two
reciprocal values of m are equal to the two reciprocal values of λ2 . However, this
does not tell us which value of λ2 yields which value of m , nor is this line of attack
very illuminating. Here, then, is a more transparent approach.
We begin by recalling a standard result of linear algebra, which is valid for n × n
matrices:

If e is an eigenvector of [A] with eigenvalue λ, then e

e ≡ [B]e is an
e
eigenvector of [A] ≡ [B][A][B] , and its eigenvalue is also λ.
−1

This is verified easily:


e e
[A] e = [B][A][B]−1 [B]e = [B][A]e = [B]λe = λ e
e.

Let us return to [3.29], in which the fixed point ξ+ of M (with associated mul-
tiplier m + ) was mapped to the fixed point 0 of M f = (F ◦ M ◦ F−1 ) by means of
z ÞÑ e z = F(z) = z−ξz−ξ− . In terms of linear transformations of C , the eigenvector
+ 2
   
ξ+ 0
of [M] is being mapped by [F] to the eigenvector of
1 1

f = [F] [M] [F]−1 .

[M]
194 Möbius Transformations and Inversion

" #
ξ+
The linear algebra result now tells us that if λ+ denotes the eigenvalue of ,
1
then
" # " #
f 0 0
[M] = λ+ .
1 1
This is true irrespective of whether or not any of the matrices in the above equation
are normalized.
Now suppose that [M] is normalized, as demanded in (3.46). Irrespective of
whether or not [F] is normalized, we have already noted that [M] is normalized
f is normalized. Since the normalized matrix of M(
if and only if [M] f e z) = m+e z is
√ 
f = m+ 0
given by [M] √ , we deduce that
0 1/ m +
" # "√ #" # " #
0 m+ 0 0 1 0
λ+ = √ =√ .
1 0 1/ m + 1 m+ 1
Thus m + = 1/λ2+ , as was to be shown.

3.8 Decomposition into 2 or 4 Reflections*

3.8.1 Introduction
Recall from (3.4) that the formula for inversion or “reflection” in a circle K has the
form
Az + B
IK (z) = .
Cz + D
It follows easily that the composition of any two reflections (in circles or lines) is
a Möbius transformation. Since the composition of two Möbius transformations is
another Möbius transformation, it follows more generally that the composition of an
even number of reflections is a Möbius transformation.
Conversely, in this section we will use the Symmetry Principle [see p. 168] to
show that
Every non-loxodromic Möbius transformation can be expressed as the
composition of two reflections, and every loxodromic Möbius transfor-
mation can be expressed as the composition of four reflections.
In the following, it would be helpful (but not essential) for you to have read the
final section of Chapter 1.

3.8.2 Elliptic Case

Consider [3.34], which depicts the same elliptic transformation shown in [3.29]
and [3.30]. Recall that the LHS shows a Möbius transformation M such that after
 Decomposition into 2 or 4 Reflections* 195

[3.34] Decomposition of an Elliptic Möbius transformation into two reflections. If M

is an elliptic Möbius transformation, and the multiplier associated with one of the fixed
points ξ+ is m = eiα , then M = IB ◦ IA , where A and B are any two circles through the
fixed points such that the angle from A to B at ξ+ is (α/2).

sending ξ+ and ξ− to 0 and ∞ by means of F(z) = (z − ξ+ )/(z − ξ− ), the new trans-

formation on the RHS is a pure rotation M(f ez ) = eiα e
z. In the illustrated example,
α = (π/3) and the dark “rectangle” abutting the line A e is carried into the dark
e
“rectangle” abutting the line B.
As we discussed in Chapter 1 [see p. 42], this origin-centred rotation of α is
equivalent to successively reflecting in any two lines containing angle (α/2) at 0,
e and B.
such as the illustrated lines A e In symbols,

f = ℜe ◦ ℜ e .
M B A

In particular, ℜA e
e maps the dark “rectangle” abutting the line A to the light “rect-
angle”, then ℜBe maps this to the dark “rectangle” abutting the line B. e The figure
tries to make this clear by also showing the successive images of both a point and
a diagonal circular arc of the original dark “rectangle”.
Now think what this means on the LHS of [3.34]. The Symmetry Principle tells us
that if two points are symmetric with respect to the line Ae then their images under
the Möbius transformation F−1 are symmetric with respect to the circle A = F−1 (A) e
through the fixed points. [Recall that in [3.29] the family of such circles was called
e on the RHS becomes reflection (i.e., inversion) in A on the
C1 .] Thus reflection in A
LHS. Of course the same goes for the second reflection in B. e Thus we have shown
the following:

If M is an elliptic Möbius transformation, and the multiplier associated

with one of the fixed points ξ+ is m = eiα , then M = IB ◦ IA where A
(3.47)
and B are any two circles through the fixed points such that the angle
from A to B at ξ+ is (α/2).
196 Möbius Transformations and Inversion

3.8.3 Hyperbolic Case

Figure [3.35] (cf. [3.31]) illustrates a similar result in the case of a hyperbolic Möbius
transformation. Here the multiplier associated with ξ+ is a real number m = ρ,
and the transformation on the RHS is a pure expansion, M( f e z) = ρez. As with a
e
rotation, an expansion can also be achieved using two reflections: if A and B e are any
two origin-centred circles such that
rB e
(radius of B) √
= = ρ, (3.48)
rA e
(radius of A)
e followed by reflection in B
then reflection in A e yields an origin-centred expansion by ρ. In
symbols, this result—which is really the same as (3.8)—says that
f = Ie ◦ I e .
M B A

As in [3.34], the RHS of [3.35] illustrates the successive effect of these two reflec-
tions on a dark rectangle abutting A. e Just as before, the Symmetry Principle
−1
applied to F tells us that the original Möbius transformation on the LHS can be
expressed as
M = IB ◦ IA .
Recall from [3.29] that A and B belong to the family C2 of circles orthogonal
to the family C1 of circles through the fixed points. At the time, we pointed out
an equivalent property of C2 , namely, that the fixed points ξ± are symmetric with
respect to each member of C2 ; this enables us to explain how it is that (IB ◦ IA )
leaves ξ+ and ξ− fixed. In the case of [3.34], this was obvious because each reflection
separately left those points fixed; in the present case, however, IA swaps the points,
then IB swaps them back again, the net effect being to leave them fixed.

[3.35] Decomposition of a Hyperbolic Möbius transformation into two reflections. If M

is a hyperbolic Möbius transformation, and the multiplier associated with one of the fixed
points ξ+ is m = ρ, then M = IB ◦ IA , where A and B are any two circles of Apollonius
√
with limit points ξ± such that (rB /rA ) = ρ.
 Decomposition into 2 or 4 Reflections* 197

In the case of an elliptic transformation, (3.47) describes how to pick out a pair
of C1 circles corresponding to any given angle α. In the present case of a hyperbolic
transformation, how are we to pick out a pair of C2 circles corresponding to any
given value of ρ? The answer depends on a third characterizing property of the C2
circles: they are the circles of Apollonius with limit points ξ± .
This terminology reflects Apollonius’ remarkable discovery (c. 250 BCE) that if
a point z moves in such a way that the ratio of the distances of z from two fixed
points ξ± remains constant, then z moves on a circle. Figure [3.35] makes this easy
to understand. As z travels round A, e z = F(z) travels round the origin-centred circle
e of radius rA . But this constant rA is none other than the ratio of the distances of
A
z from two fixed points ξ± :

|z − ξ+ |
rA = | e
z | = |F(z)| = .
|z − ξ− |

Note that this also explains the “limit point” terminology: as the ratio rA tends to
0, the corresponding Apollonian circle A shrinks down towards the limit point ξ+ ;
as rA tends to infinity, A shrinks down towards the other limit point ξ− . Another
bonus of our discussion is a result that is frequently not mentioned in geometry
texts: the limit points defining a family of Apollonian circles are symmetric with respect to
each of these circles.
Since the quantities rA and rB occurring in (3.48) are now expressible purely in
terms of the geometry of the LHS of [3.35], we have solved the problem of picking
an appropriate pair of C2 circles:

If M is a hyperbolic Möbius transformation, and the multiplier asso-

ciated with one of the fixed points ξ+ is m = ρ, then M = IB ◦ IA ,
where A and B are any two circles of Apollonius with limit points ξ±
√
such that (rB /rA ) = ρ.

3.8.4 Parabolic Case

Figure [3.36] is a modified copy of [3.33], and it illustrates how the same idea
applies to a parabolic transformation. Recall that after we have sent the solitary
fixed point ξ to ∞ by means of the Möbius transformation z ÞÑ ez = G(z) = 1/(z−ξ),
f
the new transformation on the RHS is a translation, M( ez) = ez + T.
As we discussed on p. 42, this translation can be expressed as Mf = ℜe ◦ ℜ e ,
B A
e e
where A and B are any two parallel lines such that the perpendicular connecting
complex number from A e to B
e is (T/2). Applying the Symmetry Principle to the
−1
Möbius transformation G , we deduce that (on the LHS)
198 Möbius Transformations and Inversion

[3.36] Decomposition of a Parabolic Möbius transformation into two reflections. A

parabolic Möbius transformation M with fixed point ξ can be expressed as M = IB ◦ IA ,
where A and B are circles that touch each other at ξ.

A parabolic Möbius transformation M with fixed point ξ can be expressed as

M = IB ◦ IA , where A and B are circles that touch each other at ξ.

3.8.5 Summary
Lest the details obscure the simplicity of what we have discovered, we summarize
our results as follows:
A non-loxodromic Möbius transformation M can always be decomposed
into two reflections in circles A and B that are orthogonal to the invari-
(3.49)
ant circles of M. Furthermore, M is elliptic, parabolic, or hyperbolic
according as A and B intersect, touch, or do not intersect.
Recalling (3.41), we also deduce that a loxodromic Möbius transformation M can
always be decomposed into four reflections in circles:
M = {IB ′ ◦ IA ′ } ◦ {IB ◦ IA } = {IB ◦ IA } ◦ {IB ′ ◦ IA ′ } ,
where A and B both pass through the fixed points, and where A ′ and B ′ are both orthogonal
to A and B.
We should stress that these results concern the least number of reflections into
which a Möbius transformation can be decomposed. Thus if a particular Möbius
transformation is expressible as the composition of four reflections, this does not
necessarily imply that it is loxodromic—one might be able to reduce the num-
ber of reflections from four to two. For example, if A and B are lines containing
angle (π/12) at 0, and A ′ and B ′ are lines containing angle (π/6) at 0, then the
 Automorphisms of the Unit Disc* 199

Möbius transformation (ℜB ′ ◦ ℜA ′ ◦ ℜB ◦ ℜA ) represents a rotation of (π/2),

which can be reduced to two reflections in lines containing angle (π/4). As a more
extreme example of this idea of redundant reflections, check for yourself that (3.3)
represents a decomposition of a general Möbius transformation into ten reflections!

3.9 Automorphisms of the Unit Disc*

3.9.1 Counting Degrees of Freedom

An automorphism of a region R of the complex plane is a one-to-one, conformal
mapping of R to itself. If R is a disc (or a half-plane) then clearly we can map it to
itself with a Möbius transformation M, and since M is one-to-one and conformal,
it is (by definition) an automorphism. In this subsection we will find all possible
Möbius automorphisms of the unit disc. These Möbius transformations are impor-
tant for at least two reasons: (i) in Chapter 6 we will see that they play a central
role in non-Euclidean geometry; (ii) in Chapter 7 we will see that they are the only
automorphisms of the disc!
In the following, let C denote the unit circle, let D denote the unit disc (including
C), and let M(z) denote a Möbius transformation of D to itself. Before we try to find
a formula for the most general M, let us see “how many” such Möbius transforma-
tions there are. In other words, how many real numbers (parameters) are required
to specify a particular M?
To illustrate how such counting may be done, let us first show that the set of
all Möbius transformations forms a “six parameter family”. Once we have chosen
three points in C, there is a unique Möbius transformation that maps them to three
arbitrary image points, and each of these 3 image points w = u + iv requires 2
real numbers (u and v) for its specification. If we think of the three original points
as having fixed locations, and the three image points as freely movable, then the
total number of parameters needed to specify a particular Möbius transformation
is thus 3 × 2 = 6. Another suggestive way of describing this fact is to say that the
most general Möbius transformation has six degrees of freedom.
Returning to the original problem, it is clear that we will lose some of these six
degrees of freedom when we impose the condition that M(z) map D to itself. In
fact we lose half of them:

Möbius automorphisms of D have three degrees of freedom. (3.50)

Figure [3.37a] gives one way of seeing this. Here q, r, s may be viewed as having
fixed locations on C, while q e, er, e
s are thought of as freely movable. Provided (as
e, er, e
illustrated) that q s induce the same orientation of C as q, r, s, we know from
200 Möbius Transformations and Inversion

[3.37] [a] Möbius automorphisms of D have three real degrees of freedom. Think of q,
r, s as fixed on C, but q e, er, e
s as freely movable, their locations specified by three angles.
Then we know from (3.31) that the unique Möbius automorphism of D mapping q, r, s
e, er, e
to q s, is given by z ÞÑ e z = M(z), where [e e, er, e
z, q s] = [z, q, r, s].
[b] Finding the most general Möbius automorphism via the Symmetry Principle. If a
Möbius automorphism of D sends a to 0, then the Symmetry Principle tells us that it sends
(1/a) to ∞. It then follows
 easily [see text] that the most general Möbius automorphism
of D is Mϕ a (z) = e
iϕ z−a
a z−1 , which does indeed have three real degrees of freedom.

e, er, e
(3.31) that the unique Möbius automorphism of D mapping q, r, s to q s, is given
by z ÞÑ ez = M(z), where

[e e, er, e
z, q s] = [z, q, r, s].

e, er, e
Since three real numbers are needed to specify q s—their angles, for example—
this establishes (3.50).

3.9.2 Finding the Formula via the Symmetry Principle

According to (3.50), the specification of a particular M requires three bits of infor-
mation. However, we are not obliged to give this information in the form of three
points on C—any data that are equivalent to three real numbers will do equally
well. A particularly useful alternative of this kind is shown in [3.37b]. We specify
which point a inside D is to be mapped to the origin, and we also specify which
point p on C is to be the image of the point 1 (or of some other definite point on C).
Choosing a uses up two degrees of freedom; choosing p uses up the third and last
degree of freedom.
Before pursuing this, we note another consequence of (3.50): we cannot gener-
ally find a Möbius automorphism that simultaneously sends the interior point a to
0 and sends another interior point to some other interior point. These requirements
amount to four conditions on M, while (3.50) tells us that only three such condi-
tions can be accommodated. It is very much as if we were seeking to draw a circle
through four arbitrary points—it can’t be done! However, suppose in this analogy
that we are lucky, and that the four points just happen to be concyclic, then the
 Automorphisms of the Unit Disc* 201

circle that passes through them is unique. By the same token,

If two Möbius automorphisms M and N map two interior points to the same
(3.51)
image points, then M = N.

Returning to [3.37b], note that since C is mapped to itself by M, the Symmetry

Principle tells us that if a pair of points are symmetric with respect to C, then so
are their images. Now we apply this to the symmetric pair of points, a and (1/a)
shown in [3.37b]. Since a is mapped to 0, (1/a) must be mapped to the reflection
of 0 in C, namely, ∞. Thus M must have the form
 
z−a
M(z) = k ,
az − 1
where k is a constant. Finally, we require that p = M(1) be a point on C, so
|1 − a|
1 = |p| = |k| = |k| =⇒ k = eiϕ .
|a − 1|
Thus the choice of p is equivalent to the choice of ϕ. Using the angle ϕ and the point
a to label the transformation, we have discovered that the most general Möbius
automorphism of D is
 
z−a
Mϕa (z) = e iϕ
. (3.52)
az − 1

Note that Mϕ iϕ
0 (z) = −e z = e
i(π+ϕ)
z simply rotates D about its centre 0 through
angle (π + ϕ). The general Möbius automorphism Mϕ a may be interpreted as Ma
0

followed by a rotation of ϕ, and from this point of view the really interesting part
of the transformation is M0a , which we will now abbreviate to Ma . This is the same
Ma whose properties you were asked to investigate algebraically in Chapter 2,
Ex. 3.

3.9.3 Interpreting the Simplest Formula Geometrically*

To find the geometric meaning of
z−a
Ma (z) = , (3.53)
az − 1
we could simply apply our whole arsenal of classification techniques. We ask that
you try this yourself in Ex. 26.
Here we will instead attempt to make sense of Ma “with our bare hands”, as it
were. This is probably more illuminating, and it certainly provides better geometric
sport! Begin by noting that Ma has the property that it swaps a and 0: not only
is Ma (a) = 0, but also Ma (0) = a. According to (3.51), this is the only Möbius
automorphism with this property, so if we can geometrically construct a Möbius
automorphism that swaps a and 0, then it must be Ma .
202 Möbius Transformations and Inversion

[3.38] Geometric meaning of the most general Möbius automorphism. The most general
Möbius automorphism Mϕ a (z) (above) is the composition of a rotation with the funda-
mental mapping Ma (z) = az−a
z−1 . To discover the geometric meaning of Ma (z), we first
observe that Ma swaps a and 0. The Symmetry Principle then guides us [see text] to the
answer: Ma = ℜL ◦ IJ , the order of the reflections being immaterial. The fact that Ma
swaps a and 0 can now be recognized as a special case of the fact that Ma is involutory:
(Ma ◦ Ma ) = E, and every pair of points {z, Ma (z)} is swapped by Ma .

As was explained earlier in [3.6] on page 145, the reflection IJ in any circle J
orthogonal to C will map D to itself, the two regions into which D is divided by J
being swapped. See [3.38]. At this point the obvious thing to do is to find the circle
J such that IJ swaps a and 0. Clearly the centre q of J must lie on the line L through
a and 0, but where?
We can answer this question with the same symmetry argument that we used
earlier. Since a and (1/a) are symmetric with respect to C, their images under IJ
are symmetric with respect to IJ (C) = C. Because we want IJ (a) = 0, we deduce
that IJ (1/a) = ∞. But the point that is mapped to infinity by IJ is the centre of J, so
q = (1/a).
Of course IJ is an anticonformal mapping; to obtain a conformal Möbius auto-
morphism we must compose it with another reflection. However, we have already
successfully swapped a and 0, so this second reflection must leave these points
fixed. The obvious (and only) choice is thus reflection in L. Here, then, is our
geometric interpretation of Ma :

Ma = ℜL ◦ IJ .

Incidentally, observe [exercise] that the order of these reflections doesn’t matter:
we may also write Ma = IJ ◦ ℜL .
Clearly the fixed points ξ± are the intersection points of J and L, and so they
are symmetric with respect to C. Since the reflections occur in orthogonal circles
 Automorphisms of the Unit Disc* 203

[3.39] The true meaning of Ma awaits in Chapter 6: it is a half-turn of the hyperbolic

plane about ξ+ . We can also express Ma as (IL ′ ◦ IJ ′ ), where J ′ and L ′ are any two
orthogonal circles through ξ+ that are orthogonal to C. Here we illustrate some of the
invariant circles, together with the effect of Ma on a “square”. The deeper meaning of
all this will become clear once we have encountered hyperbolic geometry in Chapter 6.
We will then recognize that for inhabitants of the hyperbolic plane, Ma is nothing more
than an ordinary rotation of the plane about ξ+ through angle π! See Section 6.3.11.

through these points, Ma is elliptic, and the multipliers associated with ξ± are both
given by m = eiπ = −1. Thus the effect of Ma on an infinitesimal neighbourhood
of the interior fixed point ξ+ is a rotation of π. The fact that Ma swaps a and 0 can
now be recognized as a special case of the fact that Ma is involutory: (Ma ◦Ma ) = E,
and every pair of points z, Ma (z) is swapped by Ma . Finally, note that we can also
express Ma as (IL ′ ◦ IJ ′ ), where J ′ and L ′ are any two orthogonal circles through
ξ+ that are orthogonal to C. All this is illustrated in [3.39], which also shows some
of the invariant circles, together with the effect of Ma on a “square”.
The deeper meaning of all this will become clear once we have encountered
Hyperbolic Geometry in Chapter 6. We will then recognize that for inhabitants of
the hyperbolic plane, Ma is nothing more than an ordinary rotation of the plane about
ξ+ through angle π! See Section 6.3.11.
While we will return to the geometry of the general Möbius automorphisms Mϕ a
in Chapter 6, we remark here that they can only be elliptic, parabolic, or hyperbolic.
This is because (by construction) they leave C invariant, while a loxodromic Möbius
transformation has no invariant circles. To be more precise, in Chapter 6 we will
use the above interpretation of Ma to show geometrically that
204 Möbius Transformations and Inversion

If we define Φ ≡ 2 cos−1 |a|, then Mϕ

a is

(i) elliptic if |ϕ| < Φ,

(ii) parabolic if |ϕ| = Φ, (3.54)
(iii) hyperbolic if |ϕ| > Φ.
For an algebraic proof, see Ex. 27.

3.9.4 Introduction to Riemann’s Mapping Theorem

Riemann’s doctoral thesis of 1851 contained many profound new results, one of the
most famous being the following, which is now called Riemann’s Mapping Theorem:

Any simply connected region R (other than the entire plane) may be
(3.55)
mapped one-to-one and conformally to any other such region S.
In Chapter 12 we shall discuss this in detail, but for the time being we merely wish
to point out some connections between Riemann’s result and what we have learnt
concerning automorphisms of the disc.
First note that to establish (3.55) in general, it is sufficient to establish it in the
special case that S is the unit disc D. For if FR is a one-to-one conformal mapping
from R to D, and FS is likewise a one-to-one conformal mapping of S to D, then
S ◦ FR is a one-to-one conformal mapping of R to S, as required.
F−1
If M is an arbitrary automorphism of D, then M◦FR is clearly another one-to-one
conformal mapping from R to D. In fact every such mapping must be of this form.
For if e
FR were any other such mapping, then e FR ◦ F−1
R would be some automorphism
e
M of D, in which case FR = M ◦ FR .
Thus the number of one-to-one conformal mappings from R to S is equal to the
number from R to D, which in turn is equal to the number of automorphisms of D.
As we have already said, in Chapter 7 we will show that these automorphisms are
the Möbius transformations Mϕ a , which form a 3-parameter family. Thus (3.55) in
fact implies that the one-to-one conformal mappings from R to S form a three-parameter
family.
 Exercises 205

3.10 Exercises

1 In each of the figures below, show that p and p e are symmetric with respect to
the circle. The dashed lines are not strictly part of the constructions, rather they
are intended to be helpful or suggestive.

2 In 1864 a French officer named Peaucellier caused a sensation by discovering a

simple mechanism (Peaucellier’s linkage) for transforming linear motion (say of
a piston) into circular motion (say of a wheel). The figure below shows six rods
hinged at the white dots, and anchored at o. Two of the rods have length l, and
the other four have length r. With the assistance of the dashed circle, show that
√
e = IK (p), where K is the circle of radius l2 − r2 centred at o. Construct this
p
mechanism—perhaps using strips of fairly stiff cardboard for rods, and draw-
ing pins for hinges—and use it to verify properties of inversion. In particular,
try moving p along a line.

3 Let S be a sphere, and let p be a point not on S. Explain why IS (p) may be con-
structed as the second intersection point of any three spheres that pass through
p and are orthogonal to S. Explain the preservation of three-dimensional
symmetry in terms of this construction.
4 Deduce (3.23), p. 167 directly from (3.18), p. 162.
5 Consider the following two-stage mapping: first stereographically project C
onto the Riemann sphere Σ in the usual way; now stereographically project Σ
206 Möbius Transformations and Inversion

back to C, but from the south pole instead of the north pole. The net effect of this
is some complex mapping z ÞÑ f(z) of C to itself. What is f?
6 Both figures below show vertical cross sections of the Riemann sphere.
(i) In figure [a], show that the triangles p0N and N0q are similar. Deduce
(3.23).
(ii) Figure [b] is a modified copy of [3.21b]. Show that the triangles z0N and
N0 e
z are similar. Deduce (3.18).

7 (i) Use a computer to draw the images in C of several origin-centred circles

under the exponential mapping, z ÞÑ ez . Explain the obvious symmetry of
these image curves with respect to the real axis.
(ii) Now use the computer to draw these same image curves on the Riemann
sphere, instead of in C. Note the surprising new symmetry!
(iii) Use (3.19) to explain this extra symmetry.
8 This exercise continues the discussion of (3.2), p. 138. If a point p in space emits
a flash of light, we claimed that each of the light rays could be represented
by a complex number. Here is one, indirect method of establishing this corre-
spondence. Once again, we choose units of space and time so that the speed
of light is 1. After one unit of time, the expanding sphere of light emitted by
p—made up of particles of light called photons—forms a unit sphere. Thus each
photon may be identified with a point on the Riemann sphere, and hence, via
stereographic projection, with a complex number. Indeed, if the photon has
spherical polar coordinates (ϕ, θ), then (3.22) tells us that the corresponding
complex number is z = cot(ϕ/2) eiθ .
Sir Roger Penrose (see Penrose and Rindler (1984, p. 13)) discovered the
following remarkable method of passing from a light ray to the associated com-
plex number directly, without the assistance of the Riemann sphere. Imagine
that p is one unit vertically above the origin of the (horizontal) complex plane.
At the instant that p emits its flash, let C begin to travel straight up (in the direc-
tion ϕ = 0) at the speed of light (= 1) towards p. Decompose the velocity of the
 Exercises 207

photon F emitted by p in the direction (ϕ, θ) into components perpendicular

and parallel to C. Hence find the time at which F hits C. Deduce that F hits C
at the point z = cot(ϕ/2) eiθ . Amazingly, we see that Penrose’s construction is
equivalent to stereographic projection!
9 In order to analyse astronomical data, Ptolemy required accurate trigonometric
tables, which he constructed using the addition formulae for sine and cosine.
The figures below explain how he discovered these key addition formulae. Both
the circles have unit radius.
(i) In figure [a], show that A = 2 sin θ and B = 2 cos θ.
(ii) In figure [b], apply Ptolemy’s Theorem to the illustrated quadrilateral, and
deduce that sin(θ + ϕ) = sin θ cos ϕ + sin ϕ cos θ.

10 The aim of this question is to understand the following result:

Any two non-intersecting, non-concentric circles can be mapped to concentric

circles by means of a suitable Möbius transformation.

(i) If A and B are the two circles in question, show that there exists a pair of
points ξ± that are symmetric with respect to both A and B.
(ii) Deduce that if F(z) = (z − ξ+ )/(z − ξ− ), then F(A) and F(B) are concentric
circles, as was desired.
11 This exercise yields a more intuitive proof of the result of the previous exercise.
Using different colours for each, draw two non-intersecting, non-concentric cir-
cles, A and B, then draw the line L through their centres. Label as p and q the
intersection points of B with L.
(i) Using corresponding colours, draw a fresh picture showing the images A, e
e e e of A, B, L, q under inversion in any circle centred at p. To get you started,
B, L, q
note that Le = L.
(ii) Now add to your figure by drawing the circle K, centred at q e
e, that cuts A
at right angles, and let g and h be the intersection points of K and L.
208 Möbius Transformations and Inversion

(iii) Now draw a new picture showing the images K ′ , L ′ , h ′ of K, L, h under

inversion in any circle centred at g.
(iv) By appealing to the anticonformal nature of inversion, deduce that A e ′, B
e′
are concentric circles centred at h ′ .
Since the composition of two inversions is a Möbius transformation, you have
proved the result of the previous exercise.
12 Figure (i) below shows two non-intersecting, non-concentric circles A and B,
together with a chain of circles C1 , C2 , . . . that touch one another successively,
and that all touch A and B. As you would expect, the chain fails to “close up”:
C8 overlaps C1 instead of touching it. Figure (ii) shows that this failure to close is
not inevitable. Given a different pair A, B, it is possible to obtain a closed chain
where Cn touches C1 . Here n = 5, but by considering the case where A, B are
concentric, you can easily see that any value of n is possible, given the right A
and B.

Steiner discovered, very surprisingly, that if the chain closes for one choice of
C1 , then it closes for every choice of C1 , and the resulting chain always contains
the same number of touching circles. Explain this using the result of Ex. 10.
13 (i) Let P be a sphere resting on the flat surface Q of a table. Let S1 , S2 , . . . be a
string of spheres touching one another successively and all the same size
as P. If each S-sphere touches both P and Q, show that S6 touches S1 , so
that we have a closed “necklace” of six spheres around P.
(ii) Let A, B, C be three spheres (not necessarily of equal size) all touching one
another. As in the previous part, let S1 , S2 , . . . be a string of spheres (now
of unequal size) touching one another successively, and all touching A, B,
C. Astonishingly (cf. previous exercise), S6 will always touch S1 , forming
a closed “necklace” of six spheres interlocked with A, B, C. Prove this by
first applying an inversion centred at the point of contact of A and B, then
appealing to part (i).
 Exercises 209

The chain of six spheres in part (ii) is called Soddy’s Hexlet, after the ama-
teur mathematician Frederick Soddy—fellow graduate of Merton College,
Oxford!—who discovered it (without inversion!). For further information on
Soddy’s Hexlet, see Ogilvy (1990). Soddy’s full time job was chemistry—in 1921
he won the Nobel Prize for his discovery of isotopes!
14 The figure below shows four collinear points a, b, c, d, together with the (nec-
essarily coplanar) light rays from those points to an observer. Imagine that
the collinear points lie in the complex plane, and that the observer is above
the plane looking down. Show that the cross-ratio [a, b, c, d] can be expressed
purely in terms of the directions of these light rays; more precisely, show that

sin α sin γ
[a, b, c, d] = − .
sin β sin δ
Suppose the observer now does a perspective drawing on a glass “canvas plane”
C (arbitrarily positioned between himself and C). That is, for each point p in C
he draws a point p ′ where the light ray from p to his eye hits C. Use the above
result to show that although angles and distances are both distorted in his
drawing, cross-ratios of collinear points are preserved: [a ′ , b ′ , c ′ , d ′ ] = [a, b, c, d].
15 Show that in both of the figures below, Arg [z, q, r, s] = θ + ϕ. Hence deduce
(3.32), p. 156.
210 Möbius Transformations and Inversion

16 As in figure [3.28], think of the cross-ratio [z, q, r, s] as a Möbius transformation.

(i) Explain geometrically why permuting q, r, s, in [z, q, r, s] yields six differ-
ent Möbius transformations.
(ii) If I(z) is the Möbius transformation that leaves 1 fixed and that swaps 0
with ∞, explain geometrically why I ◦ [z, q, r, s] = [z, s, r, q].
(iii) If J(z) is the Möbius transformation that sends 0, 1, ∞ to 1, ∞, 0, respec-
tively, explain geometrically why J ◦ [z, q, r, s] = [z, s, q, r].
(iv) Employing the abbreviation χ ≡ [z, q, r, s], explain why the six Möbius
transformations in part (i) can be expressed as
χ, I ◦ χ, J ◦ χ, I ◦ J ◦ χ, J ◦ I ◦ χ, I ◦ J ◦ I ◦ χ.
(v) Show that I(z) = (1/z) and J(z) = 1/(1 − z).
(vi) Deduce that the six possible values of the cross-ratio are
1 1 χ χ−1
χ, χ, 1−χ 1 − χ, χ−1 , χ .

17 Show geometrically that if a and c lie on a circle K, and b and d are symmetric
with respect to K, then the point [a, b, c, d] lies on the unit circle. [Hints: Draw
the two circles through a, b, d and through b, c, d. Now think of [z, b, c, d] as a
Möbius transformation.]
18 The curvature κ of a circle is defined to be the reciprocal of its radius. Let M(z) =
a z+b
c z+d be normalized. Use (3.3) to show geometrically that M maps the real line
to a circle of curvature
 
2 d
κ = 2c Im .
c
az+b
19 Let M(z) = cz+d be normalized.
(i) Using (3.3), draw diagrams to illustrate the successive effects of these trans-
formations on a family of concentric circles. Note that the image circles are
generally not concentric.
(ii) Deduce that the image circles are concentric if and only if the original family
of circles are centred at q = −(d/c). Write down the centre of the image cir-
cles in this case. [Note that this is not the image of the centre of the original
circles: M(q) is the point at infinity!]
(iii) Hence show geometrically that the circle IM with equation |c z + d| = 1 is
mapped by M to a circle of equal size. Furthermore, show that each arc of
IM is mapped to an image arc of equal size. For this reason, IM is called
the isometric circle of M.
For applications of the isometric circle, see Ford (1929) and Katok (1992).
 Exercises 211

20 (i) Show that every Möbius transformation of the form

pz + q
M(z) = where |p| > |q|
qz + p
can be rewritten in the form
 
z−a
M(z) = e iθ
, where |a| < 1.
az − 1
[Notice that the converse is also true. In other words, the two sets of
functions are the same.]
(ii) Use the matrix representation of the first equation to show that this set of
Möbius transformations forms a group under the operation of composi-
tion.
(iii) Use the disc-automorphism interpretation of these transformations to give
a geometric explanation of the fact that they form a group.
21 (i) Use the matrix representation to show algebraically that the set of Möbius
transformations
az + b
R(z) = with |a|2 + |b|2 = 1
−bz + a
forms a group under the operation of composition.
(ii) Using the interpretation of these functions given on page 184, explain part
(i) geometrically.
22 Let H be the rectangular hyperbola with Cartesian equation x2 − y2 = 1. Show
that z ÞÑ w = z2 maps H to the line Re(w) = 1. What is the image of this
line under complex inversion, w ÞÑ (1/w)? Referring back to figure [2.9], p. 69,
deduce that complex inversion maps H to a lemniscate!
p
[Hint: Think of complex inversion as z ÞÑ (1/z2 ).]
23 From the simple fact that z ÞÑ (1/z) is involutory, deduce that it is elliptic, with
multiplier −1.
24 (i) Use the Symmetry Principle to show that the most general Möbius trans-
formation of the upper half-plane to the unit disc has the form
 
iθ z−a
M(z) = e ,
z−a
where Im a > 0.
(ii) The most general Möbius transformation back from the unit disc to the
upper half-plane will therefore be the inverse of M(z). Let’s call this inverse
N(z). Use the matrix form of M to show that
a z − a eiθ
N(z) = M−1 (z) = .
z − eiθ
212 Möbius Transformations and Inversion

(iii) Explain why the Symmetry Principle implies that N(1/z) = N(z).
(iv) Show by direct calculation that the formula for N in part (ii) does indeed
satisfy the equation in part (iii).
25 Let M(z) be the general Möbius automorphism of the upper half-plane.
(i) Observing that M maps the real axis into itself, use (3.30) to show that the
coefficients of M are real.
(ii) By considering Im[M(i)], deduce that the only restriction on these real
coefficients is that they have positive determinant: (ad − bc) > 0.
(iii) Explain (both algebraically and geometrically) why these Möbius transfor-
mations form a group under composition.
(iv) How many degrees of freedom does M have? Why does this make sense?
26 Reconsider (3.53), p. 201.
(i) Use (3.45), p. 192 to show that Ma is elliptic.
(ii) Use (3.44), p. 192 to show that both multipliers are given by m = −1.
(iii) Calculate the matrix product [Ma ] [Ma ], and thereby verify that Ma is
involutory.
(iv) Use (3.28), p. 172 to calculate the fixed points of Ma .
(v) Show that the result of the previous part is in accord with figure [3.38].
27 Use (3.45), p. 192 to verify (3.54), p. 204.
CHAPTER 4

Differentiation: The Amplitwist Concept

4.1 Introduction

Having studied functions of complex numbers, we now turn to the calculus of such
functions.
To know the graph of an ordinary real function is to know the function com-
pletely, and so to understand curves is to understand real functions. The key insight
of differential calculus is that if we take a common or garden curve, place it under a
microscope and examine it using lenses of greater and greater magnifying power,
each little piece looks like a straight line. When produced, these infinitesimal pieces
of straight line are the tangents to the curve, and their directions describe the local
behaviour of the curve. Thinking of the curve as the graph of f(x), these directions
are in turn described by the derivative, f ′ (x).
Despite the fact that we cannot draw the graph of a complex function, in this
chapter we shall see how it is still possible to describe the local behaviour of a
complex mapping by means of a complex analogue of the ordinary derivative—the
“amplitwist”.

4.2 A Puzzling Phenomenon

Throughout Chapter 2 we witnessed a very strange phenomenon. Whenever we

generalized a familiar real function to a corresponding complex function, the map-
ping sent infinitesimal squares to infinitesimal squares. At present this is a purely
empirical observation based on using a computer to draw pictures of the map-
pings. In this chapter we begin to explore the theoretical underpinnings of the
phenomenon.
Let’s go back and take a closer look at a simple mapping like z ÞÑ w = z2 . As we
already know, this maps the origin-centred circle |z| = r into the circle |w| = r2 , and
it maps the ray arg(z) = θ into the ray arg(w) = 2θ. An obvious consequence of this

Visual Complex Analysis. 25th Anniversary Edition. Tristan Needham, Oxford University Press.
© Tristan Needham (2023). DOI: 10.1093/oso/9780192868916.003.0004
214 Differentiation: The Amplitwist Concept

[4.1] The tip of the iceberg: small squares map to small squares—but why? Clearly the
mapping z ÞÑ z2 preserves the orthogonality of origin-centred rays and circles. So the
small, ultimately vanishing “squares” on the left must be mapped to small, ultimately van-
ishing rectangles on the right. Very surprisingly, though, squares are ultimately mapped to
squares! As we shall see, this is a consequence of the fact that the mapping is conformal.

is that the right angle of intersection between such circles and rays in the z-plane
is preserved by the mapping, which is to say that their images in the w-plane also
meet at right angles. As illustrated in [4.1], a grid of infinitesimal squares formed
from such circles and rays must therefore be mapped to an image grid composed of
infinitesimal rectangles. However, this does not explain why these image rectangles
must again be squares.
As we will explain shortly, the fact that infinitesimal squares are preserved is
just one consequence of the fact that z ÞÑ w = z2 is conformal everywhere except
at the two critical points z = 0 and z = ∞, where angles are doubled. In par-
ticular, any pair of orthogonal curves is mapped to another pair of orthogonal
curves. In order to give another example of this, we first dismember our map-
ping into its real and imaginary parts. Writing z = x + iy and w = u + iv, we
obtain
u + iv = w = z2 = (x + iy)2 = (x2 − y2 ) + i 2xy.
Thus the new coordinates are given in terms of the old ones by
u = x2 − y2 ,
(4.1)
v = 2xy.
We now forget (temporarily!) that we are in C, and think of (4.1) as simply rep-
resenting a mapping of R2 to R2 . If we let our point (x, y) slide along any of the
rectangular hyperbolas with equation 2xy = const., then we see from (4.1) that its
 A Puzzling Phenomenon 215

[4.2] Again, small squares map to small squares—but why? The pre-images of the hori-
zontal and vertical lines on the right are the two families of hyperbolas on the left, which a
computation confirms are orthogonal to each other, as illustrated. Thus the squares on the
right must have pre-images that are rectangles on the left. But, once again, surprisingly,
the pre-images of the squares are actually squares!

image (u, v) will move on a horizontal line v = const. Likewise, the preimages of
the vertical lines u = const. will be another family of rectangular hyperbolas with
equations (x2 − y2 ) = const. Since their images are orthogonal, the claimed con-
formality of z ÞÑ z2 implies that these two kinds of hyperbolas should themselves
be orthogonal.
Figure [4.2] makes it clear that they are indeed orthogonal. We may verify
this mathematically by recalling that two curves are orthogonal at a point of
intersection if the product of their slopes at that point is equal to −1. Implicitly
differentiating the equations of the hyperbolas, we find that
x2 − y2 = const. ⇒ x − yy ′ = 0 ⇒ y ′ = +(x/y),
2xy = const. ⇒ y + xy ′ = 0 ⇒ y ′ = −(y/x).
Thus the product of the slopes of the two kinds of hyperbola at a point of
intersection is −1, as was to be shown.
Clearly we could carry on in this way, analysing the effect of the mapping on
one pair of curves after another, but what is really needed is a general argument
showing that if two curves meet at some arbitrary angle ϕ, then their images under
(4.1) will also meet at angle ϕ. To obtain such an argument, we shall continue to
pretend that we are living in the less rich structure of R2 (rather than our own
home C) and investigate the local properties of a general mapping of the plane to
itself.
216 Differentiation: The Amplitwist Concept

4.3 Local Description of Mappings in the Plane

4.3.1 Introduction
Referring to [4.3], it’s clear that to find out whether any given mapping is confor-
mal or not will require only a local investigation of what is happening very near
to the intersection point q. To make this clearer still, recognize that if we wish to
measure ϕ, or indeed even define it, we need to draw the tangents [dotted] to both
curves and then measure the angle between them. We could draw a very good
approximation to one of these tangents simply by joining q to any nearby point p
on the curve. Of course the nearer p is to q, the better will the chord qp approx-
imate the actual tangent. Since we are only concerned here with directions and
angles (rather than positions) we may dispense with the tangent itself, and instead
use the infinitesimal vector qp that points along it. Likewise, after we have per-
formed the mapping, we are not interested in the positions of the image points
Q and P themselves; rather, we want the infinitesimal connecting vector QP that
describes the direction of the new tangent at Q. We will call this infinitesimal vector
QP the image of the vector qp. However natural this terminology may seem, note
that this really is a new sense of the word “image”.
Let us now summarize our strategy. Given formulae such as (4.1), which describe
the mapping of the points to their image points, we wish to discover the induced
mapping of infinitesimal vectors emanating from a point q to their image vectors
emanating from the image point Q. In principle, we could then apply the latter
mapping to qp and to qs, yielding their images QP and QS, and hence the angle
of intersection of the image curves through Q.

[4.3] Conformal mappings. The angle ϕ between the curves passing through q on the
left is defined to be the angle between their tangents there. But as p and s merge with q,
this angle ϕ ≍ ∠pqs. Let the images of p, q, s be P, Q, S, respectively. Then to prove
conformality, we need to show that ∠PQS ≍ ϕ ≍ ∠pqs. To do so, we need to show that
the transformation that carries the initial connecting vectors on the left to their “images” on
the right (ultimately) does not alter the angle between them. This (linear) transformation
is called the Jacobian, and its coordinate description in the case of a general mapping is
derived via the next figure.
 Local Description of Mappings in the Plane 217

4.3.2 The Jacobian Matrix

Consider [4.4]. As discussed, the direction ofthe illustrated curve through q is
dx
being described with an infinitesimal vector dy
; the infinitesimal image vec-
 
du
tor dv
gives the direction of the image curve through Q. We can determine the
 
du
component du of dv
as follows:
 
dx
du = total change in u due to moving along
dy

= (change in u produced by moving dx in the x-direction)

+
(change in u produced by moving dy in the y-direction)

= (rate of change of u with x) · (change dx in x)

+
(rate of change of u with y) · (change dy in y)

= (Bx u) dx + (By u) dy,

where Bx = B/Bx etc. Likewise, we find that the vertical component is given by the
formula
dv = (Bx v) dx + (By v) dy.

Since these expressions are linear in dx and dy, it follows (assuming that not all the
partial derivatives vanish) that the infinitesimal vectors are carried to their images
by a linear transformation. The general significance of this will be discussed later,
but for the moment it means that the local effect of our mapping is completely
described by a matrix J called the Jacobian. Thus,

[4.4] The local linear (Jacobian) transformation induced by a general mapping. An

infinitesimal vector along a curve is carried to its image along the image curve by a linear
transformation called the Jacobian, J. As the text explains, du = (Bx u) dx + (By u) dy,
and similarly for dv, leading to the matrix representation of J as (4.2).
218 Differentiation: The Amplitwist Concept

! ! !
dx du dx
Þ−Ñ =J ,
dy dv dy

where the Jacobian matrix is

!
Bx u By u
J= . (4.2)
Bx v By v

We are now in a position to return to the specific mapping z ÞÑ z2 , or more

precisely to the mapping of R2 that we extracted from it. If we evaluate (4.2) for the
mapping (4.1), we find that
!
2x −2y
J= .
2y 2x

The geometric effect of this matrix is perhaps more clearly seen if we switch to
polar coordinates. At the point z = r eiθ —or rather (r cos θ, r sin θ), since for the
moment we are still in R2 —we have
!
cos θ − sin θ
J = 2r .
sin θ cos θ

The effect of the 2r is merely to expand all the vectors by this factor. This clearly
does not affect the angle between any two of them. The remaining matrix is
probably familiar to you as producing a rotation of θ, and hence it too does
not alter the angle between vectors. Since both stages of the transformation pre-
serve angles, we have in fact verified the previous claim: the net transformation is
conformal.

4.3.3 The Amplitwist Concept

We have just seen that the local effect of z ÞÑ z2 on infinitesimal vectors is to expand
them and to rotate them. Transformations of this type (i.e., whose local effect is
produced in these two steps) will play a dominating role from now on, and it will
be very much to our advantage to have vivid new words specifically to describe
them.
If all the infinitesimal vectors (qp etc.) emanating from q merely undergo an
equal enlargement to produce their images at Q, then we shall say that the local
effect of the mapping is to amplify the vectors, and that the magnification factor
involved is the amplification of the mapping at the point q. If, on the other hand,
they all undergo an equal rotation, then we shall say that the local effect of the
mapping is to twist the vectors, and that the angle of rotation involved is the twist
of the mapping at the point q. More generally, the kind of mapping that will con-
cern us will locally both amplify and twist infinitesimal vectors—we say that such
a transformation is locally an amplitwist. Thus “an amplitwist” is synonymous
 Local Description of Mappings in the Plane 219

with “a (direct) similarity”, except that the former refers to the transformation of
infinitesimal vectors, whereas “a similarity” has no such connotation.
[We remind the reader of the discussion in both Prefaces: “infinitesimal” is
being used here in a definite, technical sense—small and ultimately vanishing,
the relationship to other infinitesimals being expressed via Newtonian “ultimate
equalities”.]
We can illustrate the new terminology with reference to the concrete case we
have just analysed: The mapping z ÞÑ z2 is locally an amplitwist with amplification
2r and twist θ. See [4.5]. Quite generally, this figure makes it clear that if a mapping
is locally an amplitwist then it is automatically conformal—the angle ϕ between the
infinitesimal complex numbers is preserved.
Returning to [4.1] and [4.2], we now understand why infinitesimal squares were
mapped to infinitesimal squares. Indeed, an infinitesimal region of arbitrary shape
located at z will be “amplitwisted” (amplified and twisted) to a similar shape at
z2 . Note that here we are extending our terminology still further: henceforth we
will freely employ the verb “to amplitwist”, meaning to amplify and to twist an
infinitesimal geometric object.
All we really have at the moment is one simple mapping that turned out to be
locally an amplitwist. In order to appreciate how truly fundamental this amplitwist
concept is, we must return to C and begin from scratch to develop the idea of
complex differentiation.

[4.5] Amplitwist terminology. The local effect of z ÞÑ z2 on infinitesimal complex num-

bers emanating from z = r eiθ is to amplitwist them. Specifically, here we must amplify
them by 2r—this magnification factor is called the amplification—and we must twist
them by θ—this rotation angle is called the twist. The net transformation is called the
amplitwist: it is equivalent to multiplication by (amplification) ei(twist) = 2r eiθ , in this
example. If a complex mapping is locally an amplitwist, as this one is, then it is also con-
formal, as illustrated. We reserve this terminology for infinitesimal complex numbers, i.e.,
for small complex numbers in their moment of vanishing.
220 Differentiation: The Amplitwist Concept

4.4 The Complex Derivative as Amplitwist

4.4.1 The Real Derivative Re-examined

In the ordinary real calculus we have a potent means of visualizing the derivative
f ′ of a function f from R to R, namely, as the slope of the graph y = f(x). See [4.6a].
Unfortunately, due to our lack of four-dimensional imagination, we can’t draw
the graph of a complex function, and hence we cannot generalize this particular
conception of the derivative in any obvious way.
As a first step towards a successful generalization, we simply split the axes apart,
so that [4.6a] becomes [4.6b]. Note that we have drawn both copies of R in a hori-
zontal position, in anticipation of their being viewed as merely the real axes of two
complex planes.
Next, continuing in the spirit of the previous section, we observe that |f ′ (x)|
describes how much the initial infinitesimal vector at x must be expanded to obtain
its image at f(x). More algebraically, f ′ (x) is that real number by which we must
multiply the initial vector to obtain its image:

f ′ (x) · Ñ =
−Ñ . (4.3)

If f ′ (x) > 0 (as in [4.6b]) then the image of positive dx is a positive df, but if
′
f (x) < 0 then the infinitesimal image vector df is negative and points to the left,
as illustrated in [4.7]. In this case, df can be obtained by first expanding dx by
|f ′ (x)|, then rotating it by π. If we think of f ′ (x) as a point on the real axis of C,
then arg[f ′ (x)] = 0 when f ′ (x) > 0, and arg[f ′ (x)] = π when f ′ (x) < 0. Thus,

[4.6] The real derivative re-examined. [a] The traditional picture of the derivative f ′ as the
slope of the tangent to the graph y = f(x). [b] Let us instead break apart the two copies of
the real axis, as we are forced to break apart the two copies of C in the case of a complex
function, z ÞÑ f(z). Then an infinitesimal movement dx is mapped to df = f ′ dx: in other
words, the local transformation is an amplification by f ′ .
 The Complex Derivative as Amplitwist 221

[4.7] If we think of f ′ (x) as a point on the real axis of C, then arg[f ′ (x)] = 0 when
f ′ (x) > 0, and arg[f ′ (x)] = π when f ′ (x) < 0. Thus, regardless of whether f ′ (x) is
positive or negative, we see that the local effect of f on an infinitesimal vector dx at x is
to expand it by |f ′ (x)| and to rotate it by arg[f ′ (x)].

regardless of whether f ′ (x) is positive or negative, we see that the local effect of
f on an infinitesimal vector dx at x is to expand it by |f ′ (x)| and to rotate it by
arg[f ′ (x)].
With all this fresh in our minds, we now attempt to generalize the notion of
“derivative” to mappings of C.

4.4.2 The Complex Derivative

Consider the effect of a complex mapping f(z) on an infinitesimal complex number
emanating from z. Its image (i.e., the connecting complex number between the two
image points) will be an infinitesimal complex number emanating from f(z). The
generalization of [4.6b] or [4.7] is now [4.8]. On the right, we have drawn this image
complex number in black, and we have also drawn a copy [white] at f(z) of the
original arrow at z. To transform the white arrow into the black image arrow now
requires not only an expansion, but also a rotation. In figure [4.8] it looks as though
we must expand the white arrow by 2, and rotate it by (3π/4). Contrast this with
the case of a real function, where the required rotation angle could only be 0 or π;
in the case of a complex function we need rotations through arbitrary angles.

[4.8] The local effect of a general complex mapping. The effect of a complex function on
an infinitesimal arrow is to rotate it and expand it, which is equivalent to multiplication
by a complex number = (expansion factor)ei(rotation angle) = 2 ei(3π/4) , in the illustrated
example.
222 Differentiation: The Amplitwist Concept

Nevertheless, we can still write down an algebraic equation completely analo-

gous to (4.3), because “expand and rotate” is precisely what multiplication by a complex
number means. Thus the complex derivative f ′ (z) can now be introduced as that
complex number by which we must multiply the infinitesimal number at z to obtain
its image at f(z):

f ′ (z) · Õ

Ñ
= . (4.4)

In order to produce the correct effect, the length of f ′ (z) must be the magnification
factor, and the argument of f ′ (z) must be the angle of rotation. For example, at the par-
ticular point shown in [4.8] we would have f ′ (z) = 2 ei(3π/4) . In fact, in the spirit
of Chapter 1, we need not even distinguish between the local transformation and
the complex number that represents it.
To find f ′ (z) we have looked at the image of a specific arrow at z, but (unlike the
case of R) there are now infinitely many possible directions for such arrows. What
if we had looked at an arrow in a different direction from the illustrated one?
We are immediately in trouble, because a typical mapping1 will do what you
see in [4.9]. Clearly the magnification factor differs for the various arrows, and
likewise each arrow needs to be rotated a different amount to obtain its image.
While we could still use a complex number in (4.4) to describe the transformation
of the arrows, it would have to be a different number for each arrow. There would
therefore be no single complex number we could assign to this point as being the
derivative of f at z. We have arrived at an apparently gloomy impasse: a typical
mapping of C simply cannot be differentiated.

[4.9] Infinitesimal complex numbers in different directions are generally rotated and
expanded by different amounts. In general, a complex function will map an infinitesi-
mal circle to an infinitesimal ellipse, as shown. But this means that infinitesimal radii in
different directions each undergo different rotations and expansions. The mapping is not
locally an amplitwist, and is therefore not “analytic”: f ′ (z) does not exist.

1
Shortly we will justify certain details of [4.9], such as the fact that an infinitesimal circle is mapped
to an infinitesimal ellipse.
 The Complex Derivative as Amplitwist 223

4.4.3 Analytic Functions

We get around the above obstacle in Zen-like fashion—by ignoring it! That is, from
now on we concentrate almost exclusively on those very special mappings that
can be differentiated. Such functions are called analytic (or holomorphic). From the
previous discussion it follows that

Analytic mappings are precisely those whose local effect is an amplitwist: all the
infinitesimal complex numbers emanating from a single point are amplified and
twisted the same amount.

In contrast to [4.9], the effect of an analytic mapping can be seen in [4.10]. For such
a mapping the derivative exists, and simply is the amplitwist, or, if you prefer, the
complex number representing the amplitwist.
At this point you might quite reasonably fear that however interesting such map-
pings might be, they would be too exotic to include any familiar or useful functions.
However, a ray of hope is held out by the humble-looking mapping z ÞÑ z2 , for we
have already established that it is locally an amplitwist, as illustrated in [4.5], and
so it now gains admittance into the select set of analytic functions. In fact, quite
amazingly, we will discover in the next chapter that virtually every function we
have met in this book is analytic! Of course we have already seen plenty of empir-
ical evidence of this in our many pictures showing small “squares” being mapped
to small “squares”.
It should perhaps be stressed that all our recent pictures have been concerned
with local properties, and hence with infinitesimal arrows and figures. For example,
it’s clear from [4.10] that any analytic mapping will send infinitesimal circles
to other infinitesimal circles; however, this does not mean that such mappings
typically send circles to circles. Figure [4.11] (which contains [4.10] at its centre)
illustrates the fact that if we start with an infinitesimal circle and then expand it,
its image will generally distort out of all semblance of circularity. Of course, an
important exception to this is provided by the Möbius transformations, for these

[4.10] Analytic = locally an amplitwist, expanding and rotating all infinitesimal complex
numbers equally. By definition, an analytic mapping is locally an amplitwist, and the
derivative f ′ (z) simply is the amplitwist.
224 Differentiation: The Amplitwist Concept

[4.11] Analytic mappings only send infinitesimal circles to circles. While an analytic
mapping must send infinitesimal circles to circles, it will generally deform an expanding
circle into a shape that becomes less and less circular. The only analytic mappings that
preserve circles of all sizes are the Möbius transformations.

precisely do preserve circles of all sizes. In fact it can be shown that the Möbius
transformations are the only2 ones with this property.

4.4.4 A Brief Summary

The principal kinds of mapping we wish to study in this book are the analytic
(complex-differentiable) ones. Although these will turn out to include almost all
the useful functions, they are nevertheless very special. Their effect on an infinites-
imal disc centred at z is, after translation to f(z), simply to amplify and twist it.
The amplification is the expansion factor, and the twist is the angle of rotation. The
local effect of f is then completely encoded in the single complex number f ′ (z), the
derivative of f, or (as we will often prefer to call it) the amplitwist of f:
f ′ (z) = the derivative of f at z
= the amplitwist of f at z
= (amplification) ei(twist)
′
= |f ′ (z)| ei arg [f (z)]
.
To obtain the image at f(z) of an infinitesimal complex number at z, you just
multiply it by f ′ (z).
Two last points. We have introduced the word “amplitwist” (in addition to
“derivative”) because it is suggestive, and because it will make later reasoning

2
This was proved by Carathéodory (1937).
 Some Simple Examples 225

much easier to explain. However, the student meeting this subject for the first time
should be made aware of the fact that in all other books only the word deriva-
tive is used. Also, note that the two words are synonymous only to the extent
(cf. Chapter 1) that a complex number can be identified with the similarity trans-
formation it produces when each point is multiplied by it. Thus, for example,
“to differentiate” will not mean the same as “to amplitwist”: the former refers to
the act of finding the derivative of a function, while the latter refers to the act of
“amplifying and twisting” an infinitesimal geometric figure.

4.5 Some Simple Examples

In the following examples we have superimposed the image copy of C on the

original one.
z ÞÑ z + c.

This represents a translation of the points by c. As we see in [4.12a], the length

of complex numbers emanating from z is preserved, and hence the amplification is
unity. Equally clearly, since no rotation is induced, the twist is zero. Hence

(z + c) ′ = amplitwist of (z + c) = 1 ei0 = 1.

Notice how this is in complete accord with the familiar rule of real calculus, namely,
d
that dx (x + c) = 1.

z ÞÑ Az.

If A = a eiα , then this represents the combination of an origin-centred expansion

by a, and a rotation by α. It is clear in [4.12b] that any arrow at z (in particular an
infinitesimal one) will suffer precisely the same amplification and twist as do the
points of the plane themselves. Hence

(Az) ′ = amplitwist of (Az) = A.

While the meaning is richer, this is once again formally identical to the familiar
d
result dx (Ax) = A.
z ÞÑ z2 .

Our earlier investigation revealed that at the point z = r eiθ this mapping is
locally an amplitwist with amplification 2r and twist θ, as illustrated in [4.5]. Hence,

(z2 ) ′ = amplitwist of (z2 ) = (amplification) ei(twist) = 2r eiθ = 2z .

Once again, note that this result is formally identical to the formula (x2 ) ′ = 2x
of ordinary calculus. In the next chapter we shall obtain a directly complex and
geometrical demonstration of this fact.
226 Differentiation: The Amplitwist Concept

[4.12] The complex amplitwist formula agrees with the real derivative. [a] The transla-
tion z ÞÑ z + c has amplification 1 and twist 0, so its amplitwist is 1, in accordance with
the real formula, (x + c) ′ = 1. [b] The mapping z ÞÑ Az = a eiα z has amplification a and
twist α, so its amplitwist is a eiα = A, in accordance with the real formula, (Ax) ′ = A.

[4.13] Conjugation is anticonformal and therefore not analytic. The mapping z ÞÑ z is

anticonformal, so it is not locally an amplitwist, and the derivative does not exist.

z ÞÑ z.

Since this mapping is anticonformal, it clearly cannot be analytic, for we have

already observed that if a mapping is locally an amplitwist, then it is automati-
cally conformal. Figure [4.13a] pinpoints the trouble. From the picture we see that
the image at z of any complex number emanating from z has the same length as
the original, and hence the amplification is unity. The problem lies in the fact that
an arrow at angle ϕ must be rotated by −2ϕ to obtain its image arrow at angle −ϕ.
Thus different arrows must be rotated different amounts (which is not a twist) and
hence there is no amplitwist.
 Conformal = Analytic 227

4.6 Conformal = Analytic

4.6.1 Introduction
In [4.5] we saw clearly that any mapping that is locally an amplitwist is also auto-
matically conformal. In terms of complex differentiation, we can now rephrase
this by saying that all analytic functions are conformal. The question then nat-
urally arises as to whether the converse might also be true. Is every confor-
mal mapping analytic, or, in other words, is the local effect of every conformal
mapping nothing more complicated than an amplitwist? If this were the case
then the two concepts would be equivalent and we would have a new way
of recognizing, and perhaps reasoning about, analytic functions. A tempting
prospect!
To dismiss this as a possibility would only require the discovery of a single
function that is conformal and yet whose local effect is not an amplitwist. The
example of complex conjugation, illustrated in [4.13b], shows how important it is
that we take into account the fact that the mapping preserves not only the mag-
nitude of angles, but also their sense. For z ÞÑ z is not analytic, but it is also not a
counterexample to the conjecture, because it is anticonformal.
We have seen that although conjugation does possess an amplification, it fails to
be analytic because it doesn’t have a twist. Let us now consider instead a function
that does possess a twist, but which again fails to be analytic, this time by virtue of
not having an amplification. The effect of such a mapping at a particular point is
illustrated in [4.14]. The three curves on the LHS intersect at equal angles of (π/3),
and on the RHS their images do too. But the picture clearly shows that we are not
dealing with an amplitwist. Imagine that the infinitesimal tangent complex num-
bers to the curves are first twisted, but then rather than being amplified, as they
would be by an analytic function, they are expanded by different factors. Despite

[4.14] A twist without an amplitwist. At a single point it is possible for a mapping to

possess a twist, and therefore to be conformal at that point, and yet not possess an
amplification, and therefore not be an amplitwist.
228 Differentiation: The Amplitwist Concept

this, however, the initial twist ensures that the angle between two curves is pre-
served both in magnitude and sense: the mapping is genuinely conformal at this
point.

4.6.2 Conformality Throughout a Region

If we only insist on conformality at isolated points then such counterexamples do
indeed exist (we’ve drawn one!), but if we require the mapping to be conformal
throughout a region then this nonanalytic behaviour cannot occur.
Imagine that we have a region throughout which the mapping is (i) conformal,
and (ii) sufficiently non-pathological that an infinitesimal line-segment is mapped
to another infinitesimal line-segment. In fact, re-examination of [4.3] reveals that
(ii) must be presupposed in order for (i) even to make sense. For if the infinitesimal
straight piece of curve from q to p did not map to another of the same kind at
Q, then we could not even speak of an angle of intersection at Q, let alone of its
possible equality with ϕ.
Now look at [4.15]. In our conformal region we have drawn a large (i.e., not
infinitesimal) triangle abc, along with its image ABC. Notice that while the straight
edges of abc are completely distorted to produce the curvilinear edges of ABC,
the angles of this “triangular” image are identical with those of the original. Now
imagine shrinking abc down towards an arbitrary point in the region. As we do so,
the sides of its shrinking image will increasingly resemble straight lines [by virtue

[4.15] A mapping that is conformal throughout a region must locally be an amplitwist. If

the mapping is conformal through the region in which the triangle abc shrinks to a point,
then its curvilinear triangular image ABC must ultimately become a similar infinitesi-
mal triangle, which can therefore be obtained by amplitwisting the original infinitesimal
triangle.
 Conformal = Analytic 229

of (ii)], and all the while the angles will remain the same as the original’s. Thus,
any infinitesimal triangle in this region is mapped to another infinitesimal similar
triangle. Since the image triangle merely has a different size and alignment on the
page, it is indeed obtained by amplitwisting the original.
We have thus established the sought-after equivalence of conformal and analytic
mappings:
A mapping is locally an amplitwist at a point p if it is conformal throughout
an infinitesimal neighbourhood of p.
For this reason, the conventional definition of f being “analytic” at p is that f ′ exist
at p and at all points in an infinitesimal neighbourhood of p.
From this result we can immediately deduce, for example, that complex inver-
sion z ÞÑ (1/z) is analytic, for we have already demonstrated geometrically that
it is conformal. By the same token, it follows more generally that all Möbius
transformations are analytic.
For no extra charge, we can obtain a further equivalence simply by concentrating
on distances rather than angles. What we have just seen is that a mapping can-
not possess a twist throughout a region without also having an amplification. In
order to investigate the converse, suppose that a mapping is only known to possess
an amplification throughout a region. Re-examine [4.15] from this point of view.
Unlike the previous case, there is no longer any a priori reason for the image ABC
to betray any features common to the original. However, as we carry out the same
shrinking process as before, the local existence of amplifications begins to reveal
itself.
As the triangle becomes very small, we may consider two of its sides, for
example ab and ac, to be infinitesimal arrows emanating from a vertex. While
we may not yet know anything of angles, we do know that these arrows both
undergo the same amplification to produce their images AB and AC. But if we
now apply this reasoning at one of the other vertices, we immediately find that
in order to be consistent, all three sides must undergo the same3 amplification.
Once again we have been able to deduce that the image triangle is similar to the
original.
However, this time all we know is that the magnitude of the angles in the infinites-
imal image triangle are the same as those in the original. If the sense of the angles
also agree, then the image is obtained by amplitwisting the original, just as before.
But if the angles are reversed, then we must flip the original triangle over as well

3
We only mean “same” in the sense that the variations in amplification are of the same infinitesimal
order as the dimensions of abc. If the amplifications were precisely the same, then extending our
argument to a whole network of closely spaced vertices, we would conclude that the amplification
was constant throughout the region.
230 Differentiation: The Amplitwist Concept

as amplitwisting it. This “flip” may be accomplished by reflecting in any line; in

particular, we may employ reflection in the real axis, z ÞÑ z. Thus if f(z) is a
mapping that is known to possess an amplification throughout an infinitesimal
neighbourhood of a point p, then either f(z) is analytic at p, or else f(z) is analytic
at p.
It is interesting to note that the use of triangles in the above arguments was not
incidental, but instead crucial. Rectangles, for example, would simply not have
sufficed. Take the first argument. Certainly conformality still guarantees us that
an infinitesimal rectangle maps to another infinitesimal rectangle. However, this
image rectangle could in principle have very different proportions from the orig-
inal, and hence not be obtainable via an amplitwist. Try the second argument for
yourself and see how it too fails.
For a computational approach to the above results, see Ahlfors (1979, p. 73).

4.6.3 Conformality and the Riemann Sphere

In the previous chapter we addressed a twin question: “How are we to visualize
the effect of a mapping on infinitely remote parts of the complex plane, or the effect
of a mapping that hurls finite points into the infinite distance?” Our answer was
to replace both complex planes (original and image) with Riemann spheres. We
could then visualize the mapping as taking place between the two spheres, rather
than between the two planes. To a large extent the success of this merely depended
upon the fact that we had gathered up the infinite reaches of the plane to a single
point on the sphere. It did not depend on the precise manner in which we chose
to do this. Why then the insistence on accomplishing this with stereographic pro-
jection, rather than in some other way? Several reasons emerged in the previous
chapter, but the present discussion shows that another compelling reason is that
stereographic projection is conformal.
Only now can we fully appreciate this point, for we have seen that analytic
functions are the conformal mappings of the plane. As illustrated in [4.16], the con-
formality of stereographic projection now enables us to translate this directly into
a statement about Riemann spheres:
A mapping between spheres represents an analytic function if and only if
it is conformal.
We have drawn the spheres separate from the planes to reinforce the idea that
we are entitled to let the plane fade from our minds, and to adopt instead the
sphere as a logically independent base of operations. Indeed, at this stage we could
consider complex analysis to be nothing more than the study of conformal maps
between spheres. But in works on Riemann surfaces it is shown that in order to
embrace the global aspects of many-one functions and their inverses, one must
extend this conception to conformal mappings between more general surfaces,
such as doughnuts.
 Critical Points 231

[4.16] Analytic mappings of induce conformal mappings of the Riemann sphere, for
the simple reason that stereographic projection is conformal.

4.7 Critical Points

4.7.1 Degrees of Crushing

We return to the mapping z2 and note that at z = 0, (z2 ) ′ = 2z = 0. A place such as
this, where the derivative vanishes, is called a critical point. Recall that in the previ-
ous chapter we defined the term “critical point” differently, as a point at which the
conformality of an otherwise conformal mapping breaks down. These two defini-
tions are not at odds with one another. If the derivative f ′ (z) of an analytic mapping
f is not zero at z = p, then we know that f is conformal at p, so conformality can
only break down at points where f ′ (z) = 0. Although it is not obvious, later we
will be able to prove the converse fact that if f ′ (p) = 0 then f cannot be conformal
at p. Thus the two definitions are equivalent.
In terms of the amplitwist concept a critical point could equally well be defined
as a point of zero amplification. This suggests that the effect of an analytic mapping
on an infinitesimal disc centred at a critical point is to “crush it down to a single
image point”. The statement in quotes is not to be taken literally, rather it is to be
understood in the following sense.
Imagine that the disc (radius ϵ) is so tiny that it must be placed under a micro-
scope in order to be seen. Suppose that we have available a whole family of lenses
of increasing power with which to view it: L0 , L1 , L2 , L3 , . . .. For example, L0 has
magnification 1/ϵ0 = 1, so that it’s really no better than using the naked eye. On
the other hand L1 has magnification 1/ϵ1 , and it is thus so powerful that we can
232 Differentiation: The Amplitwist Concept

actually see the disc with it. The lens L2 is even more remarkable in that it magni-
fies by 1/ϵ2 , so that even a small part of our microscopic disc now completely fills
the viewing screen4 .
Let’s switch back to L1 so that we can see the whole disc again, and watch what
happens to it when we apply the transformation z ÞÑ z2 . It disappears! At best we
might see a single dot sitting at the image of the critical point. It is in this sense that
the mapping is crushing. However, if we now attach L2 instead, we can see our
mistake: the dot isn’t a dot, in fact it’s another disc of radius ϵ2 .
For this particular mapping, L2 was sufficient to see that the disc had not been
completely crushed. However, at a critical point of another mapping, even this
might not provide sufficient magnification, and we would require a stronger lens,
say Lm , to reveal that the image of the disc isn’t just a point. The integer m measures
the degree of crushing at the critical point.

4.7.2 Breakdown of Conformality

In addition to being locally crushing, we have stated (but not yet proved) that the
conformality of an analytic function breaks down at its critical points. We can see
this in our example. When the z2 mapping acts on a pair of rays through the crit-
ical point z = 0, it fails to preserve the angle between them; in fact it doubles it.
Thus, just at the critical point, the conformality of z2 breaks down. This is a gen-
eral property. In fact we will show later that the behaviour of a mapping very near
to a critical point is essentially given by zm , m ⩾ 2. Rather than being conserved,
angles at the critical point z = 0 are consequently multiplied by m. We quantify
the degree5 of this strange behaviour by saying that z = 0 is a critical point of order
(m − 1). Notice that this m is the same one as in the previous paragraph: in order
to see the image we have to use the Lm lens.
Despite the fact that conformality breaks down at critical points, we shall con-
tinue to make such bald statements as, “z2 is conformal”. The tacit assumption
is that critical points are being excluded. Indeed we were making this assump-
tion throughout the previous section, for we only concerned ourselves there with
typical points. Later we will see that critical points are, in a mathematically pre-
cise sense, “few and far between”, and this is our excuse for the scant attention
we are presently paying them. Nevertheless, we may safely skirt around this issue

4
In terms of this analogy we could say that most of the diagrams in this chapter, indeed in the rest of
the book, show views of the image complex plane taken through L1 . For example [4.10] depicts a tiny
circle being amplitwisted to produce another circle. However, if we viewed part of this image ‘circle’
with L2 instead of L1 , then deviations from circularity would become visible. Of course the smaller
we make the preimage circle, the smaller these deviations will be.
5
The reason we define it to be (m − 1), rather than m, is that this properly reflects the multiplicity
of the root of the derivative.
 Critical Points 233

only so long as we focus on the effect of the function on separate chunks of its
domain. When one studies Riemann surfaces, one tries to fit all this partial infor-
mation into a global picture of the mapping, and in achieving this the critical points
will play a crucial role. They do so by virtue of yet another aspect of the peculiar
behaviour of a mapping in the vicinity of such points, and it is to this feature that we
now turn.
In the previous chapter we discussed the possibility of critical points being
located at infinity. In particular, we considered z ÞÑ zm . On the Riemann sphere
we drew two straight lines passing through the origin, and we thereby saw that
angles at both z = 0 and z = ∞ were multiplied by m. We therefore conclude that
∞ is a critical point of zm of order (m − 1), just like the origin. Actually, except
for m = 2, we don’t yet know if zm is conformal anywhere! However, in the next
chapter we will see that it is conformal everywhere except at the two critical points
we have just discussed.

4.7.3 Branch Points

First consider the case of a real function R(x) from R to R. In solving problems
on maxima and minima, we learn from an early age the importance of finding the
places where R ′ (x) = 0. Figure [4.17] shows an ordinary graph of y = R(x), empha-
sizing a different aspect of the behaviour of R near to a “critical point” c where
R ′ (c) = 0. Above a typical point t, for which R ′ (t) ̸= 0, the graph is either going up
or going down, so the function is locally one-to-one. However, near c it is clearly
two-to-one.
An analogous significance holds for complex mappings. Typically f ′ (z) ̸= 0, and
so an infinitesimal neighbourhood of z is amplitwisted to an infinitesimal image
neighbourhood of w = f(z), and the two neighbourhoods are clearly in one-to-one
correspondence. However, if f ′ (z0 ) = 0 then (according to our earlier claim) near
to z0 the function behaves like zm , with m ⩾ 2. Thus, if a point is in a close orbit
around z0 , its image will orbit w0 at m-times the angular speed, and corresponding
to each point near w0 there will be m preimages near z0 . Thus w0 is a branch point
of order (m − 1). We conclude that a critical point of a given order maps to a branch
point of the same order.
We began this idea by using an analogy with real functions, but we should also
note an important difference. A real function R(x) is necessarily one-to-one when
R ′ (x) ̸= 0, but (unlike the complex case) it need not be many-to-one when R ′ (x) = 0.
The graph of x3 , for example, is flat at the origin and yet it is still one-to-one in an
infinitesimal neighbourhood of that point. In contrast to this, the complex mapping
z ÞÑ z3 is three-to-one near the origin, due to the existence of complex cube roots.
234 Differentiation: The Amplitwist Concept

[4.17] Breakdown of one-to-oneness near critical points. When R ′ (x) ̸= 0 (as at x = t)

the graph is either going up or down, so the function is one-to-one in the immediate
vicinity of the point. But near a “critical point” x = c, defined by R ′ (c) = 0, we see that
R need not be one-to-one. The same turns out to be true for complex analytic mappings,
only more so: if f ′ (c) = 0, then f is necessarily not one-to-one in the immediate vicinity
of z = c.

4.8 The Cauchy–Riemann Equations

4.8.1 Introduction
To end this chapter we will try to gain a better perspective on where the analytic
functions lie within the hierarchy of mappings of the plane. A benefit of this will be
the discovery of another way (the third!) of characterizing analytic functions, this
time in terms of their real and imaginary parts.
The first thing to do is realize that the “general” mappings (x, y) ÞÑ (u, v) that
we considered earlier were not really as general as they could have been. Picture
part of the plane as being a rolled out piece of pastry on a table. A general mapping
corresponds to “doing something” to the pastry, thereby moving its points to new
locations (the images) on the table. For example, we might cut the pastry in half
and move the two pieces away from each other. This is much more general than
anything we contemplated earlier, for it does not even possess the rudimentary
quality of continuity. That is, if two points are on either side of the cut, then no
matter how close we move them together, their images will remain far apart.
Even if we do insist on continuity, the resulting mappings are still more general
than those we have considered. For example, imagine pressing down the rolling-
pin somewhere in the middle of the pastry, and, in a single roll, stretching the far
side to twice its former size. This certainly is continuous, for bringing two points
together always brings their images together. The problem now lies in the fact that
if two infinitesimal, diametrically opposed arrows emanate from a point beneath
 The Cauchy–Riemann Equations 235

the starting position of the pin, then they each undergo a quite different transfor-
mation. Thus, in an obvious sense (not a subtle complex-differentiation sense) the
mapping isn’t differentiable at this point. Nevertheless, provided we stay away
from this line, the mapping is differentiable in the real sense, and hence subject to
our earlier analysis using the Jacobian matrix.
Another interesting kind of mapping arises from the commonplace operation of
folding the pastry. Suppose we fold it like a letter being placed in an envelope [two
creases]. Three different points will end up above a single point of the table, and
the mapping is thus three-to-one. However, at the creases themselves the mapping
is only one-to-one, and furthermore, differentiability also breaks down there. Nev-
ertheless, provided that we only look at the fold-free portions of the pastry, such
many-one functions are still subject to our previous analysis.
Suppose we play with the pastry in an ordinary way, rolling it (not necessarily
evenly) now in this direction, now in another, then turning it, folding it, rolling
it again, and so forth; then, provided we suitably restrict the domain, we can
still apply our old analysis. While such a mapping is indeed very general, we
hope that this discussion has revealed that (being continuous and differentiable
in the real sense) it is, in fact, already quite high up the evolutionary ladder. It
will therefore not come as such a surprise to learn that the local geometric effect
of such a mapping is remarkably simple, though naturally not as simple as an
amplitwist.

4.8.2 The Geometry of Linear Transformations

We pick up our earlier investigation where we left it. The local effect of the mapping
is to perform the linear transformation encoded in the Jacobian matrix (4.2). If we
can first understand the effect of a uniform linear transformation—corresponding
to a constant matrix—then we shall be finished. For we need only then remember
that our analysis is only applicable locally, the actual linear transformation varying
as it does from one place to the next.
Consider the effect of a uniform linear transformation on a circle C. Since the
Cartesian equation of C is quadratic, the linear change of coordinates induced by
the transformation will lead to another quadratic equation for the image curve.
The image curve E is thus a conic section, and since the finite points of C are not
sent to infinity, this conic must be an ellipse. See [4.18], and compare this also with
[4.9], where the local consequence of this result was illustrated for a non-uniform
transformation acting on an infinitesimal circle.
We have just used an algebraic statement of linearity. The fundamental geometric
fact is that it makes no difference if we add two vectors and then map the result,
or if we map the vectors first and then add them. Convince yourself of these two
simple consequences:
236 Differentiation: The Amplitwist Concept

[4.18] Singular Value Decomposition. A linear transformation maps the circle C to the
ellipse E. Let d be the diameter of C that is mapped to the major axis D. Consider the
chords of C [dashed] that are perpendicular to d. Since these are all bisected by d, their
images must be a family of parallel chords of E such that D is their common bisector.
They must therefore be the family perpendicular to D. We thereby deduce the Singular
Value Decomposition: A linear transformation is a stretch in the direction of d, another
stretch perpendicular to it, and finally a twist.

• Parallel lines map to parallel lines.

• The midpoint of a line-segment maps to the midpoint of the image line-segment.

We now apply these facts to E.

Since all the diameters of C are bisected by the centre of C, it follows that the
image chords of E must all pass through a common point of bisection. Thus the
centre of C is mapped to the centre of E. Drawn in the same heavy line as its image
is the particular diameter d of the circle that is mapped to the major axis D of the
ellipse. Now consider the chords of C [dashed] that are perpendicular to d. Since
these are all bisected by d, their images must be a family of parallel chords of E such
that D is their common bisector. They must therefore be the family perpendicular
to D. All this is summarized in [4.18].
Thus we have obtained the following result:
Local Singular Value Decomposition. The local linear transforma-
tion induced by a (non-linear) complex mapping is a stretch in the (4.5)
direction of d, another stretch perpendicular to it, and finally a twist.

This result also makes sense at the level of counting degrees of freedom. Just as
the matrix has four independent entries, so the specification of our transformation
also requires four bits of information: the direction of d, the stretch factor in this
direction, the perpendicular stretch factor, and the twist.
 The Cauchy–Riemann Equations 237

[REMARKS: This Singular Value Decomposition (SVD for short) is of enormous

importance in Linear Algebra and its real-world applications. It was discovered
by Eugenio Beltrami,6 whom we shall meet again in Chapter 6, because he also
made pivotal discoveries in Hyperbolic Geometry. In the context of the standard
SVD theorem, the two orthogonal expansion factors are called the singular values
of the transformation. In VDGF7 we elaborate on these ideas, and use them to pro-
vide a simple geometrical interpretation of the transpose of a matrix, which we have not
found in any standard Linear Algebra textbook. This in turn can be used to provide
geometrical explanations of other standard results in Linear Algebra.]
The ultimate specialization to analytic functions now simply requires that the
two stretch factors be put equal. This apparently reduces the number of degrees
of freedom from four to three. However, since we are now producing an equal
expansion in all directions, the direction chosen for d becomes irrelevant, and we
are left with only two genuine degrees of freedom: the amplification and the twist.
Note that we now have the following:

An orientation preserving mapping is conformal if and only if it sends

infinitesimal circles to infinitesimal circles.

If a mapping preserves circles in general, then, in particular, it must send infinites-

imal circles to infinitesimal circles, and hence it must be conformal8 . Bypassing
the detailed investigation of the previous chapter, we now see that the confor-
mality/analyticity of Möbius transformations follows from the mere fact that they
preserve circles.

4.8.3 The Cauchy–Riemann Equations

We obtain another characterization of analytic functions if we now ask how we may
recognize a Jacobian matrix for which both expansion factors are equal. This is most
easily answered by considering what kind of matrix corresponds to multiplication
by a complex number, for we already know that this produces the desired type of
linear transformation. Multiplying z = (x + iy) by (a + ib), we get

(x + iy) ÞÑ (a + ib) (x + iy) = (ax − by) + i (bx + ay).

This corresponds to the multiplication of a vector in R2 by the matrix

!
a −b
. (4.6)
b a

6
See Stewart (1993).
7
See Needham (2021, §15.4).
8
For a different proof of this fact, see Sommerville (1958, p. 237).
238 Differentiation: The Amplitwist Concept

Compare this with the Jacobian matrix (4.2),

!
Bx u By u
J= .
Bx v By v
In order for the effect of J to reduce to an amplitwist, it must have the same form
as (4.6), and thus
Bx u = +By v,
(4.7)
Bx v = −By u.
These are the celebrated Cauchy–Riemann equations. They provide us with a
third way of recognizing an analytic function. However, as with the underlying
amplitwist concept, these equations must be satisfied throughout an infinitesimal
neighbourhood of a point in order that the mapping be analytic there [see Ex. 12].
Since (a + ib) is playing the role of the amplitwist, comparison of (4.6) and (4.2)
now yields two formulae for the derivative:

f ′ = Bx u + i Bx v = Bx f, (4.8)
and
f ′ = By v − i By u = −i By f. (4.9)
3
By way of example, consider z ÞÑ z . Multiplying this out we obtain a rather
haphazard looking mess:
u + iv = (x3 − 3xy2 ) + i (3x2 y − y3 ).
However, differentiating the real and imaginary parts, we obtain
Bx u = 3x2 − 3y2 = +By v,
Bx v = 6xy = −By u,
and so the Cauchy–Riemann equations are satisfied. Thus, far from being haphaz-
ard, the special forms of u and v have ensured that the mapping is analytic. Using
(4.8) we can calculate the amplitwist:
(z3 ) ′ = 3(x2 − y2 ) + i 6xy = 3z2 ,
just as in ordinary calculus. Check that (4.9) gives the same answer.
In the next chapter we will sever our umbilical cord to R2 and discover how the
above results can be better understood by directly appealing to the geometry of the
complex plane.
 Exercises 239

4.9 Exercises

1 Use the Cauchy–Riemann equations to verify that z ÞÑ z is not analytic.

2 The mapping z ÞÑ z3 acts on an infinitesimal shape and the image is exam-
ined. It is found that the shape has been rotated by π, and its linear dimensions
expanded by 12. Where was the shape originally located? [There are two
possibilities.]
3 Consider z ÞÑ Ω(z) = z2 /z. By writing z in polar form, find out the geometric
effect of Ω. Using two colours, draw two very small arrows of equal length
emanating from a typical point z: one parallel to z; the other perpendicular to
z. Draw their images emanating from Ω(z). Deduce that Ω fails to produce an
amplitwist. [Your picture should show this in two ways.]
4 The picture shows the shaded interior of a curve being mapped by an analytic
function to the exterior of the image curve. If z travels round the curve counter-
clockwise, then which way does its image w travel round the image curve?
[Hint: Draw some infinitesimal arrows emanating from z, including one in the
direction of motion.]

5 Consider f(x + iy) = (x2 + y2 ) + i (y/x). Find and sketch the curves that are
mapped by f into (a) horizontal lines, and (b) vertical lines. Notice from your
answers that f appears to be conformal. Show that it is not in two ways: (i) by
explicitly finding some curves whose angle of intersection isn’t preserved; and
(ii) by using the Cauchy–Riemann equations.
6 Continuing from the previous exercise, show that no choice of v can make
f(x + iy) = (x2 + y2 ) + iv analytic.
7 (i) If g(z) = 3 + 2i then explain geometrically why g ′ (z) ≡ 0.
(ii) Show that if the amplification of an analytic function is identically zero (i.e.,
f ′ (z) ≡ 0) on some connected region, then the function is constant there.
240 Differentiation: The Amplitwist Concept

(iii) Give a simple counterexample to show that this conclusion does not follow
if the region is instead made up of disconnected components.
8 Use pictures to explain why if f(z) is analytic on some connected region, each of
the following conditions forces it to reduce to a constant.
(i) Ref(z) = 0
(ii) |f(z)| = const.
(iii) Not only is f(z) analytic, but f(z) is too.
9 Use the Cauchy–Riemann equations to give rigorous computational proofs of
the results of the previous two exercises.
10 Instead of writing a mapping in terms of its real and imaginary parts (i.e. f =
u+iv), it is sometimes more convenient to write it in terms of length and angle:

f(z) = R eiΨ ,

where R and Ψ are functions of z. Show that the equations that characterize an
analytic f are now
Bx R = R By Ψ and By R = −R Bx Ψ.

11 Let’s agree to say that “f = u + iv satisfies the Cauchy–Riemann equations”

if u and v do. Show that if f(z) and g(z) both satisfy the Cauchy–Riemann
equations, then their sum and their product do also.
12 For nonzero z, let f(z) = f(x + iy) = xy/z.
(i) Show that f(z) approaches 0 as z approaches any point on the real or
imaginary axis, including the origin.
(ii) Having established that f = 0 on both axes, deduce that the Cauchy–
Riemann equations are satisfied at the origin.
(iii) Despite this, show that f is not even differentiable at 0, let alone analytic
there! To do so, find the image of an infinitesimal arrow emanating from
0 and pointing in the direction eiϕ . Deduce that while f does have a twist
at 0, it fails to have an amplification there.
13 Verify that z ÞÑ ez satisfies the Cauchy–Riemann equations, and find (ez ) ′ .
14 By sketching the image of an infinitesimal rectangle under an analytic map-
ping, deduce that the local magnification factor for area is the square of the
amplification. Rederive this fact by looking at the determinant of the Jacobian
matrix.
15 Let us define S to be the square region given by

a − b ⩽ Re(z) ⩽ a + b and − b ⩽ Im(z) ⩽ b.

 Exercises 241

(i) Sketch a typical S for which b < a. Now sketch its image S e under the
z
mapping z ÞÑ e .
e from your sketch, and write down the ratio
(ii) Deduce the area of S
!
e
area of S
Λ≡ .
area of S

(iii) Using the results of the previous two exercises, what limit should Λ
approach as b shrinks to nothing?
(iv) Find limbÑ0 Λ from your expression in part (ii), and check that it agrees
with your geometric answer in part (iii).
16 Consider the complex inversion mapping I(z) = (1/z). Since I is conformal,
its local effect must be an amplitwist. By considering the image of an arc of an
origin-centred circle, deduce that |(1/z) ′ | = 1/|z|2 .
17 Consider the complex inversion mapping I(z) = (1/z).
(i) If z = x + iy and I = u + iv, express u and v in terms of x and y.
(ii) Show that the Cauchy–Riemann equations are satisfied everywhere
except the origin, so that I is analytic except at this point.
(iii) Find the Jacobian matrix, and by expressing it in terms of polar coordi-
nates, find the local geometric effect of I.
(iv) Use (4.8) to show that the amplitwist is −(1/z2 ), just as in ordinary calcu-
lus, and in accord with the previous exercise. Use this to confirm the result
of part (iii).
18 Recall Ex. 19, p. 210, where you showed that a general Möbius transformation
az + b
M(z) = ,
cz + d
maps concentric circles to concentric circles if and only if the original family
(call it F) is centred at q = −(d/c). Let ρ = |z − q| be the distance from q to z,
so that the members of F are ρ = const.
(i) By considering orthogonal connecting vectors from one member of F to
an infinitesimally larger member of F, deduce that the amplification of M
is constant on each circle of F. Deduce that |M ′ | must be a function of ρ
alone.
(ii) By considering the image of an infinitesimal shape that starts far from q
and then travels to a point very close to q, deduce that at some point in
the journey the image and preimage are congruent.
(iii) Combine the above results to deduce that there is a special member IM
of F such that infinitesimal shapes on IM are mapped to congruent image
242 Differentiation: The Amplitwist Concept

shapes on the image circle M(IM ). Recall that IM is called the isometric
circle of M.
(iv) Use the previous part to explain why M(IM ) has the same radius as IM .
(v) Explain why IM−1 = M(IM ).
(vi) Suppose that M is normalized. Using the idea in Ex. 16, show that the
amplification of M is
1
|M ′ (z)| = 2 2 .
|c| ρ
19 Consider the mapping f(z) = z4 , illustrated below. On the left is a particle p
travelling upwards along a segment of the line x = 1, while on the right is the
image path traced by f(p).
(i) Copy this diagram, and by considering the length and angle of p as it
continues its upward journey, sketch the continuation of the image path.
(ii) Show that A = i sec4 (π/8).
(iii) Find and mark on your picture the two positions (call them b1 and b2 ) of
p that map to the self-intersection point B of the image path.
(iv) Assuming the result f ′ (z) = 4z3 , find the twist at b1 and also at b2 .
(v) Using the previous part, show that (as indicated at B) the image path cuts
itself at right angles.

20 The figure below is a copy of [2.9], p. 69.

(i) Show geometrically that if z moves distance ds along the lemniscate,
increasing θ by dθ, then w = z2 moves a distance 4 dθ along the circle.
(ii) Using the fact that (z2 ) ′ = 2z, deduce geometrically that ds = 2 dθ/r.
(iii) Using the fact that r2 = 2 cos 2θ, show by calculation that
q
r dr = 2 1 − (r4 /4) dθ.
 Exercises 243

(iv) Let s represent the length of the segment of the lemniscate connecting the
origin to the point z. Deduce from the previous two parts that
Zr
dr
s= p ,
0 1 − (r4 /4)
hence the name, lemniscatic integral.

21 (i) Three-Dimensional SVD: By extending the argument given in the text,

show that in three-dimensional space the effect of a linear transformation
is to stretch space in three mutually perpendicular directions (generally
by three different factors), then to rotate it.
(ii) Deduce that a mapping of three-dimensional space to itself is locally a
three-dimensional amplitwist if and only if it maps infinitesimal spheres
to infinitesimal spheres.
(iii) Deduce that inversion in a sphere preserves the magnitude of the angle
contained by two intersecting curves in space.
(iv) Deduce that stereographic projection is conformal.

Remark: In stark contrast to the bountiful conformal mappings of the

plane, Liouville and Maxwell independently discovered that the only angle-
preserving transformation of space is an inversion, or perhaps the composition
of several inversions.
CHAPTER 5

Further Geometry of Differentiation

5.1 Cauchy–Riemann Revealed

5.1.1 Introduction
In the previous chapter we began to investigate the remarkable nature of analytic
functions in C by studying mappings in the less structured realm of R2 . In particu-
lar, the Jacobian provided us with a painless way of deriving the Cauchy–Riemann
characterization of analytic functions, and also of computing their amplitwists.
However, this approach was rather indirect.
In this chapter we will instead study differentiation directly in the complex
plane, primarily through the use of ultimate equality—the powerful geometrical
tool that Newton unleashed in his great Principia of 1687; see both the original and
the new Preface. However, in the original edition these arguments were expressed
mainly in the language of infinitesimals, and we have left this mode of expression
largely intact. But in the captions, all of which are new to this edition, we have felt
free to employ the ≍ –notation.
Our first application of this approach will be the rederivation of the Cauchy–
Riemann (henceforth “CR”) equations, and the discovery of new forms that they
can take on.

5.1.2 The Cartesian Form

Consider a very fine mesh of squares aligned with the real and imaginary axes.
See the top left of [5.1]. Under an analytic mapping each infinitesimal square will
be amplitwisted to produce an image that is also square. We will show that the
CR equations are nothing more than a symbolic restatement of this geometric
fact.

Visual Complex Analysis. 25th Anniversary Edition. Tristan Needham, Oxford University Press.
© Tristan Needham (2023). DOI: 10.1093/oso/9780192868916.003.0005
246 Further Geometry of Differentiation

[5.1] Geometric derivation of the Cartesian Cauchy–Riemann equations. The image

of the real edge ϵ of the small (ultimately vanishing) square on the left is ultimately
equal to ϵ Bx f on the right. Likewise, the image of the imaginary edge iϵ of the square
on the left is ultimately equal to ϵ By f on the right. But if f is analytic, then the image
of the square on the left must ultimately be another square on the right. So one edge
of the image is ultimately equal to the other edge rotated by (π/2), that is, multiplied
by i. Thus, ϵ By f ≍ iϵ Bx f =⇒ By f = i Bx f, which is the compact Cartesian form of
the celebrated Cauchy–Riemann equations.

Zoom in on an individual square and its image, as depicted in the bottom half of
[5.1]. Suppose, as drawn, that the initial square has side ϵ. If we start at z and then
move a distance ϵ in the x-direction, the image will move along a complex number
given by
(change in x)·(rate of change with x of the image f) = ϵ Bx f.
Similarly, if the point moves along the vertical edge by going ϵ in the y-direction,
then its image will move along ϵ By f. Now since these two image vectors span a
square they must be related by a simple rotation of π/2, that is by multiplication
with i. After cancelling ϵ, we thus obtain
iBx f = By f, (5.1)
et voilà! That this is indeed a compact form of the CR equations may be seen by
inserting f = u + iv:
iBx (u + iv) = By (u + iv),
and then equating real and imaginary parts to yield
Bx u = By v and Bx v = −By u, (5.2)
just as before.
 Cauchy–Riemann Revealed 247

To obtain the amplitwist itself, we recall that each infinitesimal arrow is taken to
its image by multiplication with f ′ . Now, since we know what the images are for
the two sides of the square, we easily deduce
ϵ Þ−Ñ ϵ f ′ = ϵ Bx f
⇒ f′ = Bx f
and
iϵ Þ−Ñ iϵ f ′ = ϵ By f
⇒ f′ = −i By f.
These are the same formulae (4.8) and (4.9) that we previously obtained, but now
we understand their geometrical meaning.

5.1.3 The Polar Form

Equation (5.2) is the most common way of writing CR, but it isn’t the only way. It
took this form because we chose to describe both complex planes in terms of their
real and imaginary parts, that is with Cartesian coordinates. Thus we could briefly
describe (5.2) as being the Cart.–Cart. form. In Ex. 10, p. 240 we retained Cartesian
coordinates for the first plane but employed polar coordinates in the image plane;
this led to another form (Cart.–Polar) of CR. As the next example of our geometric
method we will derive the Polar-Cart. form of the equations.
In order to do this, we begin with an infinitesimal square adapted to polar coor-
dinates. See [5.2]. If we start at z and increase r by dr, then we obtain eiθ dr as the

[5.2] Geometric derivation of the polar Cauchy–Riemann equations. If z moves radi-

ally by dr then its image ultimately moves dr·Br f. If the angle of z increases by dθ, ulti-
mately causing z to move r dθ orthogonally, then its image ultimately moves dθ · Bθ f,
as illustrated. If the initial shape is ultimately a square (so that dr ≍ r dθ) then an
analytic f will ultimately map it to another square, so dθ·Bθ f ≍ idr·Br f ≍ irdθ·Br f.
Thus the Cauchy–Riemann equations take the form B θ f = ir B r f.
248 Further Geometry of Differentiation

radial edge. If, on the other hand, we increase θ by dθ, then the point will move
in the perpendicular direction given by ieiθ . As dθ tends to zero, this edge is ulti-
mately equal to an infinitesimal arc of circle of length r dθ; the complex number
describing it will therefore be i eiθ r dθ = i z dθ. It’s also clear from our picture that
initially square ⇐⇒ dr = r dθ. (5.3)
Now look at the image. Just as before, if we increase r by dr then the image will
move along dr · Br f; likewise, changing θ by dθ will move the image along dθ · Bθ f.
If the mapping is analytic then these again span a square, and so the latter must be
i times the former:
dθ · Bθ f = idr · Br f.
Substituting (5.3) into this, and cancelling dθ, we obtain
Bθ f = ir Br f (5.4)
as the new compact form of CR. By inserting f = u + iv, the reader may verify that
(5.4) is equivalent to the following pair of Polar–Cart. equations:
Bθ v = +r Br u (5.5)

Bθ u = −r Br v. (5.6)
By examining the amplitwist that carries each arrow to its image we can also obtain
two expressions for the derivative:
eiθ dr Þ−Ñ eiθ dr · f ′ = dr · Br f
′
=⇒ f = e−iθ Br f (5.7)
and
i z dθ Þ−Ñ i z dθ · f ′ = dθ · Bθ f
=⇒ f ′ = −(i/z) Bθ f. (5.8)
As a simple example let’s take z3 = r3 e3iθ . From (5.7) we obtain
(z3 ) ′ = e−iθ 3r2 e3iθ = 3r2 e2iθ = 3z2 ,
while from (5.8) we obtain
(z3 ) ′ = −(i/z) r3 3ie3iθ = −(i/z) 3i z3 = 3z2 .
In obtaining the same answer from both these expressions we have also verified
that z3 actually was analytic in the first place.
Of the four possible ways of writing CR, only one now remains to be found,
namely the Polar-Polar form. We leave it to the reader to verify that if we write
f = R eiΨ (cf. Ex. 10, p. 240) then CR takes the form
Bθ R = −rR Br Ψ and R Bθ Ψ = r Br R.
 An Intimation of Rigidity 249

5.2 An Intimation of Rigidity

A recurring theme in complex analysis is the “rigidity” of analytic functions. By

this we mean that their highly structured nature (everywhere locally an amplitwist)
enables us to pin down their precise behaviour from very limited information. For
example, even if we are only told the effect of an analytic function on a small region,
then its definition can be extended beyond these confines in a unique way—like a
crystal grown from a seed. In fact, given even the meagre knowledge of how an
analytic mapping affects a closed curve (just the points on the curve, mind you),
we can predict precisely what happens to each point inside! See [5.3].
Later we will justify these wild claims, and in Chapter 9 we will even find an
explicit formula (called Cauchy’s Formula) for w in terms of A, B, C, etc. For the
moment, though, we will obtain our first glimpse of this rigidity by considering a
different kind of partial information.
Consider [5.4]. Origin-centred circles are being mapped to vertical lines, and the
larger the circle, the further to the right is the image, but with no restriction on how
the lines are spaced. How much information do you think we can gather about an
analytic mapping possessing this property? Try meditating on this before reading
further.
Well, we know that f is conformal and that its local effect is just an amplitwist.
Consider the rays emanating from the origin. Since these cut through all the circles
at right angles, their images must cut through the vertical lines at right angles, and

[5.3] Cauchy’s Formula. The structure of an analytic function is so strong that given even
the meagre knowledge of how it affects a closed curve (just the points on the curve, mind
you), we can predict precisely what happens to each point inside! The precise recipe is
given by Cauchy’s Formula, (9.1), page 485.
250 Further Geometry of Differentiation

they are thus horizontal lines. In fact, if we swing the ray around counter-clockwise,
we can even tell whether its image line will move up or down. Look at [5.5], which
depicts the fate of a infinitesimal square bounded by two circles and two rays. We
know that the infinitesimal radial arrow connecting the two circles must map to
a connecting arrow between the lines going from left to right. But since the square
is to be amplitwisted, its image must be positioned as shown. Thus we find that a
positive rotation of the ray will translate the image line upwards.
We have made some good progress, but that we cannot yet have fully cap-
tured the consequences of analyticity can be seen from Ex. 5, p. 239. Despite not
being analytic, the mapping (x + iy) ÞÑ (x2 + y2 ) + i(y/x) was there shown to
possess all the above desiderata. Indeed it would be easy to write down an infini-
tude of nonanalytic functions that would be consistent with the known facts. Very
remarkably, and in stark contrast to this, if we restrict ourselves to analytic map-
pings, then we will prove that there is only one! To show this we must turn to the
CR equations.
In [5.4] we are mapping natural polar objects to natural Cartesian objects, so it’s
clear that we should employ the Polar–Cart. form, namely (5.5) and (5.6). In order
to put them to use, we must first translate [5.4] into ‘Equationspeak’. We could
describe the figure by saying that rotating the point only moves the image up and
down, not side to side; in other words, varying θ produces no change in u: Bθ u = 0.
It follows from (5.6) that Br v = 0. This says that moving the point radially outwards
does not affect the height of the image, and thus that rays are mapped to horizontal

[5.4] There is only one analytic mapping that does this. Suppose that a mapping sends
origin-centred circles to vertical lines, and the larger the circle, the further to the right
is the image line, but with no restriction on how the lines are spaced. It is easy to write
down an infinitude of different non-analytic mappings that do this, but, very remarkably,
if we restrict ourselves to analytic mappings, then we can prove that there is only one!
 An Intimation of Rigidity 251

[5.5] Geometric proof of the claim made in the previous figure. Let us now restrict our-
selves to analytic mappings possessing the property in the previous figure. Since rays are
orthogonal to origin-centred circles, their images under a conformal mapping must be
orthogonal to vertical lines—so the images of rays are horizontal lines. Furthermore, we
know that the infinitesimal outward radial arrow connecting the two illustrated circles
must map to a connecting arrow between vertical lines going from left to right. Since the
square is amplitwisted to its image, we deduce that as the ray rotates counterclockwise,
its horizontal image line moves upwards.

lines. This is old news to us seasoned geometers, but fortunately we have another
equation left:

Bθ V(θ) = r Br U(r). (5.9)

Here we have written v = V(θ) to stress that it is known to depend only on θ;

similarly for u = U(r).
Now (5.9) looks like an impossible equation, for the LHS quite explicitly depends
only on θ, while the RHS is equally emphatic about only depending on r. The
only way out of this is for both of these real quantities to equal a constant, say
A. Dispensing with the superfluous partial derivatives, we thereby obtain
dU dV
r =A and = A.
dr dθ
Integrating these equations we find that

U = A ln r + const. and V = Aθ + const.,

and hence

U + iV = A (ln r + iθ) + B,
252 Further Geometry of Differentiation

where B = const. But we recognize this special combination as none other than
the complex logarithm! Thus
f(z) = A log z + B. (5.10)
Suppose, more generally, that an analytic function g(z) is known to send circles
with centre c to parallel lines making a fixed angle ϕ with the imaginary axis. That
there is no fundamental difference between this and the previous case may be seen
by considering
z ÞÑ e−iϕ g(z + c);
for you may convince yourself that this possesses property [5.4], and hence it too
must equal (5.10). The rigidity of analytic functions has thus led to the rather
remarkable conclusion that the complex logarithm is uniquely defined (up to
constants) as the conformal mapping sending concentric circles to parallel lines.

5.3 Visual Differentiation of log(z)

A fringe benefit of the previous section was the discovery that log(z) actually is
analytic. Since this multifunction finds its simplest representation in Polar–Cart.
form, namely
log z = ln r + i(θ + 2mπ),
we can easily find its derivative using (5.7) or (5.8). For purposes of illustration, we
will now use them both:
(log z) ′ = e−iθ Br log z = e−iθ (1/r) = 1/z,
and
(log z) ′ = −(i/z) Bθ log z = −(i/z) i = 1/z. (5.11)
You notice, of course, how this is formally identical to the case of the ordinary, real
logarithm.
You may be wondering how our previous discussion of the branches of this mul-
tifunction affects all this. For example, it’s interesting how m (which labels the
different branches) does not appear in the result (5.11). The basic philosophy of
this book is that while it often takes more imagination and effort to find a picture
than to do a calculation, the picture will always reward you by bringing you nearer
to the Truth. In this spirit, we now find a visual explanation of (5.11) that will also
make it clear that the answer does not depend on m.
Equations (5.7) and (5.8) were derived by examining the infinitesimal geometry
of a general analytic mapping. Why not then apply this idea to the geometry of a
specific mapping, and thereby evaluate its amplitwist directly?
 Visual Differentiation of log(z) 253

[5.6] Geometric demonstration that [log(z)] ′ = (1/z). Rotating z by the small, ultimately
vanishing angle δ moves it along the white complex number, which ultimately has length
rδ. All of the infinitely many images under log(z) move upwards together, along iδ. Thus,
for all branches, the amplification = (1/r), the twist = −θ, and therefore the amplitwist =
(1/r)e−iθ = (1/z).

Consider [5.6], which shows a typical point z and a few of its infinitely many
images under log. In order to find the amplitwist we need only find the image of
a single arrow emanating from z. The easiest one to find is shown in [5.6], namely
an arrow perpendicular to z. Notice how if z makes an angle θ with the horizontal,
then the perpendicular vector will make an angle θ with the vertical. Also, if it
subtends an infinitesimal angle δ at the origin, then—because it is like a small arc
of a circle—its length will be rδ. Now look at the images of z. Since we have purely
rotated z, its images will all move vertically up through a distance equal to the
angle of rotation δ. To make it easier to see what amplitwist carries the arrow at z
into its image, we have drawn copies of the original arrow at each image point. It
is now evident from the picture that

amplification = 1/r
twist = −θ
=⇒ amplitwist = (1/r) e−iθ = 1/z.

Although all the image vectors emanate from different points in the different
branches, they are all identical as vectors, and so it is clear that the amplitwist does
not depend on which branch we look at.
254 Further Geometry of Differentiation

5.4 Rules of Differentiation

We already know how to differentiate z2 and also log z, so how would you use this
knowledge to find, for example, the derivative of log(z2 log z)? Your immediate
reaction (chain and product rules) is quite correct, and in this section we merely
verify that all the familiar rules of real differentiation carry over into the complex
realm without any changes, at least in appearance.

5.4.1 Composition
The composite function (g ◦ f)(z) = g[f(z)] of course just means “do f, then do
g”. If both f and g are analytic then each of these two steps conserves angles, and
therefore the composite mapping does too. We deduce that g[f(z)] is analytic, and
we now show that the net amplitwist it produces is correctly given by the chain
rule.
Let f ′ (z) = A eiα and g ′ (w) = B eiβ , where w = f(z). Consider [5.7]. An infinites-
imal arrow at z is amplitwisted by f to produce an image at w; then this, in its turn,
is amplitwisted by g to produce the final image at g(w). It is clear from the picture
that

net amplification = AB
net twist = α + β
=⇒ net amplitwist = AB ei(α+β) ,

and thus we obtain the familiar chain rule:

{g[f(z)]} ′ = g ′ (w) · f ′ (z). (5.12)

[5.7] Geometry of the Chain Rule. The local effect of the composition of two analytic
mappings is an amplitwist that is the composition of their separate amplitwists, so we
obtain the familiar chain rule.
 Rules of Differentiation 255

As an example of this we may put g(z) = kz. In the last chapter we showed that
g ′ (z) = k, and so we now conclude from (5.12) that
[k f(z)] ′ = k f ′ (z).

5.4.2 Inverse Functions

Provided we are not at a critical point (where the derivative vanishes), an infinites-
imal disc at z will be amplitwisted to produce an image disc at w = f(z), and these
two discs will be in one-to-one correspondence. See [5.8]. An analytic function thus
always possesses a local inverse in this sense, and we wish to know its derivative.
Clearly, the amplitwist that returns the image disc to its original state has
reciprocal amplification, and opposite twist:
amplification of f−1 at w = 1/(amplification of f at z) = 1/|f ′ (z)|
twist of f−1 at w = − (twist of f at z) = arg[1/f ′ (z)]
=⇒ [f−1 (w)] ′ = 1/f ′ (z). (5.13)
By way of example, consider w = f(z) = log z, for which z = f−1 (w) = ew . From
(5.13) we find that
(ew ) ′ = 1/(log z) ′ = z = ew , (5.14)
in agreement with your calculation in Ex. 13, p. 240. Later we will give a visual
derivation of (5.14).
Both (5.12) and (5.13) could have been derived even more quickly if we had
directly employed the algebraic idea of the image arrow being f ′ times the original
one. We chose instead to keep the geometry to the fore, and reserved the algebra
of multiplication for the final encoding of the results as (5.12) and (5.13). However,
to derive the next two rules by pure geometry would be cumbersome, so we will
use a little algebra.

[5.8] The amplitwist of the inverse of an analytic mapping has the opposite twist and
the reciprocal amplification, and so we obtain the familiar rule [f−1 (w)] ′ = 1/f ′ (z).
256 Further Geometry of Differentiation

5.4.3 Addition and Multiplication

On the far left of [5.9] we see an infinitesimal arrow ξ connecting z to a neighbour-
ing point. The images of these two points under f and (separately) g, are shown in
the middle of the figure. Lastly, we either add or multiply these points to obtain
the two points on the far right. By examining the image vector connecting these
final points we can deduce the amplitwists of (f + g) and fg, respectively. In [5.9]
we see that
A − a ≍ ξf ′ and B − b ≍ ξg ′ ,
so that
(A + B) − (a + b) ≍ ξ (f ′ + g ′ ),
| {z }
ultimate image of ξ

and hence we obtain the addition rule:

(f + g) ′ = f ′ + g ′ . (5.15)
Likewise, we find that
AB − ab ≍ ξ (f ′ b + a g ′ ),
| {z }
ultimate image of ξ

and thereby obtain the product rule:

(fg) ′ = f ′ g + f g ′ . (5.16)

[5.9] If f and g are analytic, then by considering the image of the small, ultimately van-
ishing complex number ξ under first their sum and then their product, we deduce that
(f + g) ′ = f ′ + g ′ , and (fg) ′ = f ′ g + f g ′ , just as in elementary real calculus.
 Polynomials, Power Series, and Rational Functions 257

5.5 Polynomials, Power Series, and Rational

Functions

5.5.1 Polynomials
We can look at the rules of the previous section from a slightly different point of
view. Take rule (5.16), for example. In a way, what is on the RHS is less important
than the fact that there is a RHS. By this we mean that we have here a recipe for
creating new analytic functions: ‘given two such functions, form their product’.
Likewise, each of our other rules can be thought of as a means of producing new
analytic functions from old. The analytic functions are indeed the aristocrats of the
complex plane, but provided they only mate with their own kind, and only in ways
sanctioned by the rules (which allow many forms of incest!), their offspring will
also be aristocrats. For example, suppose we start with only the mapping z ÞÑ z,
which is known to be analytic. Our rules now quickly generate z2 , z3 , . . ., and thence
any polynomial.
Consider a typical polynomial of degree n:
Sn (z) = a0 + a1 z + a2 z2 + · · · + an zn .
We have just seen that this is analytic, and thus it maps an infinitesimal disc at p to
another at Sn (p). Furthermore, the amplitwist that transforms the former into the
latter is, according to (5.15),
Sn′ (z) = (a0 ) ′ + (a1 z) ′ + (a2 z2 ) ′ + · · · + (an zn ) ′ .
We already know how to differentiate the first four terms, and in the next section
we will confirm that in general (zm ) ′ = m zm−1 , as you no doubt anticipated. Thus
Sn′ (z) = a1 + 2a2 z + 3a3 z2 + · · · + nan zn−1 . (5.17)

5.5.2 Power Series

This discussion of polynomials naturally leads to the investigation of power series.
In Chapter 2 we discussed how a convergent power series1
S(z) = a0 + a1 z + a2 z2 + a3 z3 + · · · (5.18)
could be approximated by a polynomial Sn . We explained how the effect of S within
its circle of convergence could be mimicked by Sn , with arbitrarily high accuracy,
simply by taking a sufficiently high value of n.
Of course the question we now face is whether power series are analytic, and if
they are, how are we to calculate their derivatives? We will see that the answers to
these questions are “yes” and “(5.17)”.

1
For simplicity’s sake we shall use a power series centred at the origin. However, as we pointed out
in Chapter 2, this does not involve any loss of generality.
258 Further Geometry of Differentiation

[5.10] Power series are analytic. On the left, we see the images of a small, ultimately
vanishing disc D centred at p under the higher and higher degree polynomial approxi-
mations Sn (z) of a power series, S(z). On the right, we see these same image discs, now
translated to have their common centre at S(p). All these polynomial approximations are
analytic, so three equally spaced points on the rim of D will be amplitwisted to their
images on the rims of the image discs, and so it becomes clear that S(z) is indeed ana-
lytic, and that its amplitwist is the limit of the amplitwists of Sn as n goes to infinity. [For
ease of visualization, here the amplification and twist are both depicted as monotonically
increasing with n, but this will not be the case in general.]

Consider an infinitesimal disc D with centre p. If p is inside the circle of conver-

gence of S, then so is a sufficiently small D. The series (5.18) therefore converges
at all points of D, and thus S maps the disc to some infinitesimal unknown shape
S(D) covering S(p). Now look at the left of [5.10]. This shows a magnified view
of the successive images of D [itself not shown in the figure] under S10 , S100 , S1000 ,
etc. Since each of these polynomials is known to be analytic, each image is a disc.
However, it is also known that these images will coincide, ever more perfectly, with
S(D). Thus S sends infinitesimal discs to other discs, and it is therefore analytic.
We have tried to make this plainer still on the right of [5.10]. Since we are now
only interested in amplitwists, the actual image points are unimportant—we only
care about the connecting arrows between them. To make it easier to watch what is
happening to these arrows we have translated the discs—which doesn’t affect the
vectors—so that their centres all coincide at S(p). By way of illustration, we now
consider the fate, as n increases, of three equally spaced vectors from p to three
equally spaced points (a, b, c) on the rim of D. Each of the analytic mappings Sn
 Polynomials, Power Series, and Rational Functions 259

amplitwists these vectors to three equally spaced image vectors. The figure shows
the gradual evolution2 of these images towards their final state (given by S) as we
successively apply S10 , S100 , etc. The amplitwist that carries the arrows of D into
these images therefore undergoes a corresponding evolution towards a final value.
The amplitwist S ′ that carries the original vectors of D to their ultimate images is
thus mimicked with arbitrarily high accuracy by Sn′ , as n increases. Therefore
S ′ (z) = a1 + 2a2 z + 3a3 z2 + 4a4 z3 + · · · . (5.19)
We have reached a very important conclusion. Any power series is analytic within
its disc of convergence, and its derivative is obtained simply by differentiating
the series term by term. Since the result of this process (5.19) is yet another con-
vergent power series with the same radius of convergence, there is nothing to
stop us differentiating again. Continuing in this manner, we discover that a power
series is infinitely differentiable within its disc of convergence. The reason this is
so important is that we will be able to show later that every analytic function can
be represented locally as a power series, and thus analytic functions are infinitely
differentiable.
This result is in sharp contrast to the case of real functions. For example, the
mileage displayed on the dash of your car is a differentiable function of the time
displayed on the clock. In fact the derivative is itself displayed on the speedometer.
However, in the instant that you hit the brakes, the second derivative (acceleration)
does not exist. More generally, consider the real function that vanishes for nega-
tive x, and that equals xm for non-negative x. This is differentiable (m − 1) times
everywhere, but not m times at the origin. Our complex aristocrats will be shown
to be quite incapable of stooping to this sort of behaviour.

5.5.3 Rational Functions

Earlier we established that the product rule applies to complex analytic functions,
but we neglected to check the quotient rule. We invite you to verify this now, using
the same kind of reasoning that led to (5.16). If you get stuck, there is a hint in
Ex. 9. In any event, the important point is that the quotient of two analytic functions
is also analytic except at the points where it has singularities. In particular, if we
apply this result to polynomials then we can conclude that the rational functions
are analytic.
The fact that the quotient of two analytic functions is again analytic can be looked
at in a rather more geometric way. Let I(z) = (1/z) be the complex inversion
mapping. As we discussed at such length in Chapter 3, I(z) is conformal, and hence
it is analytic. It follows that if g(z) is analytic, then so is [1/g(z)], because this is the

2
For ease of visualization, we have taken both the amplification and the twist to be steadily increas-
ing with n. In general they could exhibit damped oscillations as they settled down to their final
values.
260 Further Geometry of Differentiation

composition (I ◦ g) of two analytic functions. Finally, if f(z) is analytic, the product

rule tells us that f(z) · [1/g(z)] = [f(z)/g(z)] is too.

5.6 Visual Differentiation of the Power Function

We saw in the last section that z2 , z3 , z4 , . . . were all analytic. Composing with com-
plex inversion, it follows that z−2 , z−3 , z−4 , . . . are too. Since the inverse functions
(in the sense of [5.8]) are branches of the multifunctions z±1/2 , z±1/3 , . . . discussed
in Chapter 2, it follows that these too are analytic. Composing zp with z1/q (p, q
integers), it follows that any rational power is analytic. Furthermore, since the
geometric effect of any real power can be reproduced with arbitrary accuracy by
rational powers, it follows that these real powers are also analytic.
The calculation of the derivative of a real power za is similar to the example z3
given on p. 248. We find that
(za ) ′ = a za−1 , (5.20)
just as in ordinary calculus. In fact the real formula (xa ) ′ = a xa−1 can be thought
of as the specialization of (5.20) that results when both z and the infinitesimal arrow
emanating from it are taken to be on the real axis (cf. [4.7], p. 221).
Just as in the case of the complex logarithm, we do not rest at the result (5.20) of a
calculation, but rather we stalk the thing to its geometric lair. Since the amplitwist
is the same for all arrows, we need only find the image of a single arrow in the
direction of our choice. As a first (ill-fated) attempt, consider [5.11], in which we
have chosen an arrow parallel to z. To facilitate comparison, we have drawn a copy
of the initial arrow at the image point. You can see from the picture that
twist = (a − 1)θ BUT amplification = ???????
We are thus half-thwarted, for we cannot see how long the image arrow is. In fact
to figure this out would require precisely the same calculation (general Binomial
Theorem) as is needed in the real case. Oh well, “If at first you don’t succeed, …”
“Try, try an arrow perpendicular to z !” From [5.12], we see that this arrow orig-
inally makes an angle θ with the vertical, and so after magnifying the angle of
z by a, it will make an angle aθ with the vertical. Once again we see that the
twist = (a − 1)θ. However, this time we can see the amplification, simply by rec-
ognizing that each arrow is an infinitesimal arc of a circle. The angle subtended by
the arc has been magnified by a, while the radius of the circle has been magnified
by ra−1 . The net amplification of the arc is therefore a ra−1 . Thus
amplification = a ra−1
twist = (a − 1)θ
=⇒ amplitwist = a ra−1 ei(a−1)θ = a za−1 .
 Visual Differentiation of the Power Function 261

[5.11] First (ill-fated) attempt to geometrically evaluate the amplitwist of z ÞÑ za . An

infinitesimal radial complex number emanating from z = r eiθ is amplitwisted to a radial
complex number emanating from z = ra eiaθ , and therefore the twist = (a − 1)θ.
However, the amplification cannot be geometrically determined from this figure.

[5.12] Second (successful!) attempt to geometrically evaluate the amplitwist of z ÞÑ za .

If we instead consider an infinitesimal complex number orthogonal to z = r eiθ , then
it is amplitwisted to an infinitesimal complex number orthogonal to z = ra eiaθ . There-
fore, once again, we see that the twist = (a − 1)θ. However, this time the amplification
can be geometrically determined: the initial complex number has length ≍ rϵ, while its
image has length ≍ ra (aϵ), so the amplification = a ra−1 . Therefore, the amplitwist =
(za ) ′ = a ra−1 ei(a−1)θ = a za−1 .
262 Further Geometry of Differentiation

In the above figures a = 3, and so there is no ambiguity in the meaning of za

or za−1 . But if a is a fraction, for example, then both za and za−1 are multifunc-
tions possessing many different branches. We urge you to redraw [5.12] in such a
case. For example, if a = (1/3) then the infinitesimal arrow on the left will have
three images on the right, one for each branch of the cube root function. Unlike the
case of the multifunction log(z) (illustrated in [5.6]) these images are obtained by
amplitwisting the original arrow by three different amounts: each branch of za has
a different amplitwist. However, your figure will show you [exercise] that

The amplitwist of each branch of za is given by (za ) ′ = a za /z, provided

(5.21)
that the same branch of za is used on both sides of the equation.

To the best of our knowledge there is no3 direct, intuitive way of understanding the
real result (xa ) ′ = a xa−1 . It is therefore particularly pleasing that with the greater
generality of the complex result (5.21) comes the richer geometry of [5.12] needed
to see its truth.

5.7 Visual Differentiation of exp(z)

We have already seen that (ez ) ′ = ez by calculation, and we will now explain it
geometrically. In [5.13] we have written a typical point z = x + iθ to make it easier
to remember that w = ez = ex eiθ has angle θ. Moving z vertically up through a
distance δ will rotate the image through an angle δ. Being an infinitesimal arc of
circle of radius ex , the image vector has length ex δ; its direction is θ to the vertical.
As usual, we have copied the original arrow at the image so that we may more
clearly see the amplitwist:

amplification = ex
twist = θ
=⇒ amplitwist = ex eiθ = ez .

Actually, we have been a little hasty. We haven’t really shown yet (at least not
geometrically) that ez is analytic: we don’t know if all arrows undergo an equal
amplitwist. Figure [5.13] tells us that if it’s analytic, then (ez ) ′ = ez . To establish
analyticity we need only see that one other arrow is affected in the same way.
In [5.14] we move z an infinitesimal distance δ in the x-direction, thereby mov-
ing the image radially outwards. Now, from ordinary calculus, the amplification
produced by ex along the real axis is ex (cf. [4.6], p. 220), so the length of this image

3
In special cases there are ways. For example, consider a cube of side x. It is easy to visualize that
if we increase the separation of one of the three pairs of faces by δ, we add a layer of volume x2 δ. The
result (x3 ) ′ = 3x2 follows.
 Visual Differentiation of exp(z) 263

[5.13] Geometrical evaluation the amplitwist of z ÞÑ ez . An infinitesimal vertical com-

plex number iδ emanating from z = x + iθ is amplitwisted to a complex number
orthogonal to w = ez = ex eiθ whose length is ≍ ex δ. Therefore the amplification = ex
and the twist = θ, so the amplitwist = (ez ) ′ = ex eiθ = ez .

[5.14] Verification that z ÞÑ ez is locally an amplitwist. If we instead look at the image

of an infinitesimal horizontal complex number emanating from z, we see that it, too,
undergoes the same amplitwist as did the vertical one in the previous figure, confirming
that all infinitesimal complex numbers undergo the same amplitwist.
264 Further Geometry of Differentiation

vector is ex δ. It is now clear that this new arrow in [5.14] has indeed undergone
precisely the same amplification and twist as that in [5.13], thus establishing the
analyticity of ez .

5.8 Geometric Solution of E ′ = E

Up to now we have motivated the definition of the exponential mapping in rather

ad hoc ways. We are now in a position to do so in a logically more satisfying manner,
although the most compelling explanation will have to wait till later.
Consider first the ordinary real function that we write as ex . As we discussed
in Chapter 2, one way of characterizing this function is to say that the slope of its
graph is always equal to its height. An equivalent dynamic interpretation would be
that if the distance of a particle at time t is et , then its speed equals its distance from
us. In either event, this amounts to saying that the function satisfies the differential
equation

E ′ = E. (5.22)

Of course this doesn’t quite pin it down since k ex also obeys (5.22); however, if
we insist that the real solution of (5.22) also satisfy E(0) = 1, then no ambiguity
remains.
The object of this section is to show that the complex exponential function can
be characterized in exactly the same way. If a complex-analytic function E(z) is to
generalize ex then it must satisfy (5.22) on the real axis. We will now show geomet-
rically that (5.22) uniquely propagates ex off the real axis into the plane to produce
the familiar complex exponential mapping. The plan will be essentially to reverse
the flow of logic associated with [5.13] and [5.14].
A typical point z is being mapped to an unknown image w, where w = E(z)
is subject to (5.22). Decoding this equation, we find that it says that vectors ema-
nating from z undergo an amplitwist equal to the image point w. From this alone
we will figure out where w must be! In what follows, try to free your mind from
assumptions based on your previous knowledge of ez .
Consider what happens to the little (ultimately vanishing) square of side ϵ
shown in [5.15]. Because it’s twisted by the angle of w, its horizontal edge becomes
parallel to w, while its vertical edge becomes orthogonal to w. Thus horizontal
movement of z results in radial movement of the image, while vertical movement
results in rotation of the image. The question that now remains is exactly how swift
these radial and rotational motions are. Having used the twist, we now turn to the
amplification.
If z moves at unit speed in the x-direction, then since the amplification is r, the
image moves radially with speed equal to its distance from the origin. But this
 Geometric Solution of E′ = E 265

[5.15] Geometric solution of E ′ = E. The small, ultimately vanishing square is

amplitwisted by E ′ (z) = E(z) = w, so the image of the horizontal edge is radial, and
the sides are amplified by r = |w|. Thus dθ ≍ ϵ: vertical movement of z produces a
numerically equal rotation of w. If z moves at unit speed in the x-direction, then since
the amplification is r, the image moves radially with speed equal to its distance from the
origin. But this is just the familiar property of the ordinary exponential function. Thus E
maps horizontal lines exponentially onto rays.

is just the familiar property of the ordinary exponential function. Thus E maps
horizontal lines exponentially onto rays. If we now insist that E(0) = 1, then the
real axis maps to the real axis, and we thereby recover the ordinary exponential
function. We also know that translating a horizontal line upwards will rotate its
image ray counter-clockwise, but we don’t yet know how fast. In [5.15] dθ is the
infinitesimal rotation produced by moving z through a distance ϵ along the vertical
edge of the square. But since the amplification is r, we know that the image of this
edge has length rϵ, and consequently dθ = ϵ. In other words,
An infinitesimal vertical translation produces a numerically equal rotation. (5.23)
We can now completely describe the mapping produced by E(z). Imagine watch-
ing the image as we move from the origin to a typical point z = x + iθ in a
two-legged journey: first along the real axis to x, then straight up to z. See [5.16]. As
we move to x, the image moves along the real axis from 1 to ex . Repeated applica-
tion of (5.23) then tells us that moving up a distance θ will rotate the image through
an angle θ. For example, we find that E(z) wraps the imaginary axis around the unit
circle in such a way that
E(iθ) = cos θ + i sin θ.
This is our old friend, the celebrated Euler Formula. It also follows directly from
this geometry that the mapping has the property
E(a + b) = E(a) · E(b).
It is now entirely logical to define “ez ” to be E(z), and our work is done.
266 Further Geometry of Differentiation

[5.16] Geometry of the exponential. If we insist that E(0) = 1, then combining the two
facts we deduced in the previous figure, we see that we have arrived at the familiar
geometry of z ÞÑ ez .

As we indicated at the start of this section, there is in fact an even more com-
pelling explanation than the above. We have just used a very natural differential
equation to propagate ex off the real axis; however, it will turn out that even
this equation is superfluous. The rigidity of analytic functions is so great that
merely knowing the values of ex on the real axis uniquely determines its “analytic
continuation” into the complex realm.

5.9 An Application of Higher Derivatives: Curvature*

5.9.1 Introduction
Earlier we alluded to the remarkable fact that analytic functions are infinitely dif-
ferentiable. In other words, if f is analytic then f ′′ exists. In this section we seek to
shed geometric light on the meaning and existence of this second derivative f ′′ . We
shall do so by answering the following question:
If an analytic mapping f acts on a curve K of known curvature κ at p,
then what is the curvature e e at f(p)?
κ of the image curve K
In the next section we shall see that the solution to this problem provides a novel
insight into (of all things!) the elliptical orbits of the planets round the sun at one
focus.
At the risk of ruining the suspense, here is the answer to our question:
" # !
1 f ′′ (p) b
ξ
e
κ= ′ Im +κ , (5.24)
|f (p)| f ′ (p)
 An Application of Higher Derivatives: Curvature* 267

[5.17] Transformation of curvature under z ÞÑ ez . Applying the general formula (5.24)

for the transformation of curvature, the image under z ÞÑ ez of the straight line at angle
ϕ is given by eκ = e−x sin ϕ. Check for yourself that this is in accord with the empirical
evidence in this figure.

where bξ denotes the unit complex number tangent to the original curve at p. Before
explaining this result, let us simply test it on an example.
On the left of [5.17] we have drawn three line-segments, and on the right their
images under f(z) = ez . The segments are distinguished by the value of the angle
ϕ that each makes with the horizontal: ϕ = 0 for the dotted one; ϕ = (π/2) for
the dashed one; and the solid one represents a general value of ϕ. Now look at the
curvature of their images: e κ = 0 for the dotted one; e κ = e−a for the dashed one;
and on the solid image, e κ starts out large and then dies away as we spiral out from
the origin.
In order to compare these empirical observations with our formula, write the
unit tangent as b
ξ = eiϕ and note that if f(z) = ez then f ′′ = f ′ = ez . With z = x+iy,
formula (5.24) therefore reduces to

e
κ = e−x (sin ϕ + κ).

Using the fact that κ = 0 for our line-segments, and that ϕ is constant on each, you
may now easily check the accord between this formula and figure [5.17].

5.9.2 Analytic Transformation of Curvature

We now turn to the explanation of (5.24). The presence of an imaginary part in
this rather daunting formula would seem to bode ill for a purely geometric attack.
Surprisingly, this isn’t the case. Consider [5.18].
On the left is the curve K, with curvature κ at p. Note that we have arbitrarily
assigned a sense to K so as to give κ a definite sign. At the top of the figure is the
268 Further Geometry of Differentiation

[5.18] Geometric derivation of the transformation law for curvature. The analytic map-
ping f transforms the curve K into the curve K. e Curvature is the rate of rotation of the
tangent, so κ ≍ (ϵ/|ξ|) at p is transformed into eκ ≍ (eϵ/|e
ξ|) ≍ (eϵ/|f ′ (p)ξ|) at p
e = f(p).
If the twists at p and at q were the same, then the turning angle ϵ would not change,
but in fact (as we see on the right) the amplitwists at p and q differ by χ ≍ f ′′ (p)ξ, so
there is an extra twist σ ≍ (angle subtended by χ). To geometrically determine this extra
′
twist σ, divide the figure onhthe right
i by f h(p) to obtain
i the figure at the bottom. We now
f ′′ (p)ξ
see that σ ≍ Im(ν) = Im f ′ χ(p) ≍ Im f ′ (p) , thereby completing the proof of the
curvature transformation law, (5.24).
 An Application of Higher Derivatives: Curvature* 269

image curve K e under the mapping f; note that its sense is determined by that of K.
It is the curvature of K e at p
e = f(p) that (5.24) purports to describe.
As illustrated, ξ is a small (ultimately infinitesimal) complex number tangent to
K at p. With centre at p we have drawn a circle through the tip of ξ cutting K in
q. At q we have drawn another small (ultimately infinitesimal) tangent complex
number ζ, and we have marked the angle ϵ of rotation from ξ to ζ. Recall that the
curvature κ at p is, by definition, the rate of rotation of the tangent with respect to
distance along K. Since for infinitesimal ξ the arc pq equals |ξ|, the curvature at p
is therefore
ϵ
κ= . (5.25)
|ξ|
Likewise, at the image points p e and qe on the image curve K e we have drawn the
image complex numbers e e the rotation from e
ξ and ζ, ξ to ζe being e
ϵ. Thus the image
curvature is
e
ϵ
e
κ= . (5.26)
|e
ξ|
Our problem therefore reduces to finding e ϵ and |e
ξ|.
e
Since |ξ| is the length of the amplitwisted image of ξ,
|e
ξ| = (amplification) · |ξ| = |f ′ (p)| · |ξ|. (5.27)
The more interesting and difficult part of the problem is to find e
ϵ.
If ξ and ζ both underwent precisely the same twist, then the turning angle e
ϵ for
the images would equal the original turning angle ϵ. However, the twist at q will
differ very slightly, say by σ, from that at p. Thus
e
ϵ = ϵ + (extra twist) = ϵ + σ. (5.28)
′′
This is how f enters the picture, for it describes how the amplitwist varies.
The function f ′ is a perfectly respectable mapping in its own right, and it may be
drawn like any other. The right-hand side of [5.18] is precisely such a picture. Each
point z is mapped to the complex number that amplitwists infinitesimal complex
numbers emanating from z. In particular, we have drawn the images f ′ (p) and
f ′ (q) of p and q. The statement about infinite differentiability can now be recast in
a more blatantly astonishing form: if f is locally an amplitwist, then f ′ automatically is
too. We have indicated this in the picture by showing the disc at p being mapped
by f ′ to another disc at f ′ (p). This startling fact will now yield to us the value of σ.
The amplitwist that carries the disc at p to the disc at f ′ (p) is f ′′ (p). In particular,
ξ is amplitwisted to
χ = f ′′ (p) ξ.
But looking at the triangle on the right, the sides of which are the known quantities
f ′ (p) and χ, we see that the angle at the origin is precisely the extra twist σ that we
seek.
270 Further Geometry of Differentiation

It is easier to obtain an expression for this angle if we first rotate the triangle
to the real axis. This rotation is achieved quite naturally (see the bottom figure) by
dividing by f ′ (p); the sides of the triangle now become 1 and ν = [χ/f ′ (p)]. Because
σ equals the (ultimately) vertical arc through 1, the figure tells us that
   ′′ 
χ f (p) ξ
σ = arc = Im(ν) = Im ′ = Im .
f (p) f ′ (p)
Thus, from (5.26), (5.27), and (5.28), and taking evaluation at p as understood, we
obtain
 h ′′ i 
Im f ′ξ + ϵ
e
κ= f .
|f ′ | |ξ|
Finally, using (5.25) and noting that b ξ = (ξ/|ξ|) is the unit tangent at p, we do
indeed obtain formula (5.24).
Finally, we note that, alternatively, a very short (but unilluminating) computa-
tional proof can be obtained [exercise] by differentiation of the equation,
Twist = Im log f ′ .

5.9.3 Complex Curvature

Let us take a closer look at formula (5.24), which may be written
" #
f ′′ b
ξ κ
e
κ = Im ′ ′ + ′ .
f |f | |f |
The presence of the second term can be understood as follows. If the plane were
to undergo a uniform expansion by factor R then a circle of radius (1/κ) would
become a circle of radius (R/κ), that is of curvature (κ/R). But a small piece of
a general curve resembles an arc of its circle of curvature4 , and the principal local
effect of f (apart from a curvature-preserving twist) is an expansion by factor |f ′ |.
In addition to this phenomenon, the first term says that the mapping will intro-
duce curvature even when none is originally present: the curvature k of the image
of a straight line (as a function of its direction) is
" #
′′ b
f ξ
k( b
ξ ) = Im ′ ′ .
f |f |

Now consider the fate of all the curves that pass through p in the direction b ξ . The
general formula says that f will not only scale their curvatures by (1/|f ′ |) (as previ-
ously explained), but it will also increase their curvatures by the fixed amount k(b ξ).
In this sense, the first term corresponds to an intrinsic property of the mapping f.

4
The circle that touches the curve at the point in question, and whose curvature κ = (1/radius)
agrees with that of the curve at that point.
 An Application of Higher Derivatives: Curvature* 271

However, k( b ξ ) is not really intrinsic to f since it retains a vestige of the original

curves, namely, their direction b ξ . It would appear that the most natural intrinsic
quantity that can be abstracted from k( b ξ ) is

i f ′′
K≡ . (5.29)
f ′ |f ′ |

We propose to call this complex function K (which does not appear to have been
investigated previously) the complex curvature5 of f.
To see that the complex curvature is indeed a natural quantity, picture K(p) as a
vector emanating from p. We will show that

The projection of K(p) onto a line through p is the curvature of the image
(5.30)
of that line at f(p).

See [5.19], in which K has also been drawn at two additional points. Note how
the increasing length of the projection of K onto the line corresponds to increasing
curvature along the image.

[5.19] Geometric interpretation of the complex curvature. The projection of the complex
curvature K(p) onto a line through p is the curvature of the image of that line at f(p).
Here, the projection grows as we move along the line, so the image curve bends more
tightly.

5
NOTE added in this 25th Anniversary Edition: I am delighted that my discovery of K and its math-
ematical properties has since found several applications (particularly in physics) that go beyond the
two applications that I originally discovered and published here in VCA, namely, to dual central force
fields [this chapter] and to the geometry of harmonic functions [Chapter 12]. For example, see Vitelli
and Nelson (2006) and Mughal and Weaire (2009).
272 Further Geometry of Differentiation

To prove (5.30), recall how the scalar product in R2 can be expressed in terms of
complex multiplication:
" #
h i h i ′′ b
K ξ·
b = Re K b ξ = Im i K b
f ξ
ξ = Im ′ ′ = k( b
f |f |
ξ ),

as was to be shown. This result yields a neater and more intelligible form of (5.24):

e
κ=K ξ ·
b + κ .
|f ′ |
(5.31)

To see how K(p) may be determined geometrically, imagine a short, directed

line-segment S rotating about p. The image f(S) rotates with equal speed about
f(p), and its curvature oscillates sinusoidally: it reaches its maximum value |K(p)|
when S points in the direction of K(p), while it vanishes when S is perpendicular
to K(p).
In fact, to reconstruct K(p) it is sufficient to know the image curvatures κ1 and κ2
for just two positions S1 and S2 of the line-segment. Figure [5.20] illustrates this in
the particularly simple case that S1 and S2 are horizontal and vertical, respectively.
We then have

K = κ1 + iκ2 .

We conclude this section with a different way of looking at K. On the left of

[5.21] is an infinitesimal black shape Q, together with copies obtained by translating
Q a fixed amount |ξ| in various directions ξ. Under an analytic mapping f, Q is
amplitwisted to the similar black shape Q e on the right. As Q translates by ξ, Q
e not
′
only translates by f ξ, but it also rotates and expands. More precisely, the rotation

[5.20] Geometric interpretation of the real and imaginary parts of the complex curva-
ture. The real and imaginary parts of the complex curvature K(p) are the curvatures of
the images of the real and imaginary line segments through p.
 An Application of Higher Derivatives: Curvature* 273

e is just the angle σ on the RHS of [5.18]. This rotation is clearly greatest when
of Q
χ is perpendicular to f ′ (p), pointing counterclockwise along the circle |f ′ | = const.
This occurs when ξ is in the direction of K, for then

χ ∝ f ′′ K ∝ i/f ′ ∝ if ′ .

If we turn the direction of motion of Q by −(π/2), then χ also turns by −(π/2)

to point radially outwards along the ray arg f ′ = const., thereby producing the
greatest increase in |f ′ |.
We now understand [5.21] in greater detail:
e be its image under an analytic
Let Q be an infinitesimal shape, and let Q
mapping f. Then Q e rotates most rapidly, and its size remains constant,
when Q moves in the direction of K. On the other hand, Q e expands most (5.32)
rapidly, and does not rotate, when Q moves in the orthogonal direction
−iK.

In still greater detail, as Q begins to translate in an arbitrary direction b

ξ , let R denote
the rate of rotation of Q e with respect to the distance it moves. Then

b.
R=K ξ ·
This achieves its maximum value Rmax = |K| when Q moves in the direction of K.
e Let E denote the rate of increase of the size6
Similarly, consider the expansion of Q.

[5.21] Another geometric interpretation of the complex curvature. Let Q be an infinites-

imal shape, and let Q e be its image under an analytic mapping f. Then Q e rotates most
rapidly, and its size remains constant, when Q moves in the direction of the complex
curvature K. On the other hand, Q e expands most rapidly, and does not rotate, when Q
moves in the orthogonal direction −iK.

6 e For example, if Q
Here we mean the linear dimensions of Q. e were a disc then we could take its
“size” to be its radius.
274 Further Geometry of Differentiation

e (again with respect to the distance it moves) as a fraction of Q’s

of Q e initial size.
Then [exercise]
b × K.
E=ξ
This achieves its maximum value E max = |K| when Q moves in the direction −iK.
These two results may be viewed as two facets of the single complex equation
b
ξ K = R + iE.
In Chapter 12, having developed the physical concepts of “flux” and “circula-
tion”, we shall return to the complex curvature and see that it has other elegant
properties and applications.

5.10 Celestial Mechanics*

5.10.1 Central Force Fields

If a particle p, moving through space, is constantly being pulled towards (or pushed
away from) a fixed point o with a force that depends only on its distance r from o
then we say that it is in a central force field and that o is the centre of force. No matter
how the force varies with r, it is not hard to show [exercise] that the orbit of p will
always lie in a plane through o.
Another feature of motion in any central force field is that the radius op sweeps
out area at a constant rate A, called the areal speed. A proof of this is given in Ex. 24.
If the mass of p is m then [exercise] the angular momentum h of p is 2mA. The fact
that A is constant is thus a manifestation of the conservation of angular momentum.
In addition to the angular momentum, the total energy E of the particle remains
constant as it orbits. Henceforth, we shall always use a particle of unit mass. Thus if
the particle’s speed is v then the kinetic energy contribution has the definite value
1 2
2 v , while the potential energy contribution is only defined up to a constant. We
shall restrict ourselves to force fields that vary as a power of r, and we may then fix
the constant by arbitrarily assigning zero potential energy to the point where the
field vanishes: if the force grows as a positive power of r, at the origin; if the force
dies away as a negative power of r, at infinity.

5.10.2 Two Kinds of Elliptical Orbit

Consider the attractive linear force field in which, by definition, the force towards
o is proportional to r. This linear force law is extremely important in physics, for if
almost any physical system is slightly disturbed from equilibrium then the restor-
ing force is precisely of this kind. Here is a simple example of what we mean; it
will enable you to experimentally investigate motion in a linear force field. You are
encouraged to do the following, not merely to imagine it.
 Celestial Mechanics* 275

Take a small weight W and suspend it just above a point o of a horizontal table
using several feet of thread, perhaps attached to the ceiling. If you pull W to the
side by just an inch or two then, because the thread is long, W barely rises above the
table’s surface and we may idealize this to a movement on the table. Furthermore,
although the forces acting on W in this displaced position are actually gravity and
the tension in the thread, the net effect [exercise] is as though o were magically
pulling W towards it with a force proportional to r, as was required. To avoid the
possibility of confusion later, we stress that gravity is playing absolutely no essen-
tial role here; it is merely providing one particularly convenient way of simulating
a linear force field.
Now pull W a little bit away from o and give it a gentle flick in a random
direction. You see that the orbit of W is a closed curve traversed again and
again—a beautifully symmetrical oval shape centred at o. But exactly what is
this oval?
It is an ellipse! To demonstrate this, take the tabletop to be C with o as its origin.
Once again take W to have unit mass, and let its location at time t be z(t). For
simplicity’s sake, let the force directed towards the origin equal the distance |z|.
The differential equation governing the motion of W will therefore be z̈ = −z,
the two basic solutions of which are z = e±it . These represent counter-rotating
motions of unit speed around the unit circle. [Try launching W so as to produce
these solutions.] The general solution is then obtained as a linear combination of
these motions:

z = p eit + q e−it , (5.33)

where p and q may, without any real loss of generality, be taken as real and
satisfying p > q.
As is illustrated in [5.22], the addition of such counter-rotating circular motions
results in elliptical motion with the attracting point at the centre. This becomes clear
if we rewrite (5.33) as

z = a cos t + ib sin t ,

where a = p + q and b = p − q. Each of these numbers has a double significance:

a is both the semimajor axis and the point of launch; b is both the semiminor axis
√ √
and the speed of launch. Note that the foci are at ± a2 − b2 = ±2 pq.
Finally, for future use, let us calculate the constant energy E of a particle orbiting
in this field. The potential energy is the work needed to pull the particle away from
the origin out to a distance of r, namely, [exercise] (r2 /2). Thus

E = 12 (v2 + r2 ).

As the particle orbits round the ellipse in [5.22], we see that this expression always
equals 12 (a2 + b2 ).
276 Further Geometry of Differentiation

[5.22] Addition of counter-rotating circular motions results in elliptical motion.

We now turn to a second, more famous example of elliptical motion in a central

force field: the orbits of the planets around the sun. There are two fundamental
differences between this phenomenon and the one above. First, instead of the force
of attraction increasing linearly with distance, here the force of gravity dies away as
the square of the distance from the sun. Second, instead of the centre of attraction
being at the centre of the elliptical orbit, here the sun is at one of the foci.
The ancient Greeks discovered that the ellipse has beautiful mathematical prop-
erties; two thousand years later Newton revealed that it has equally beautiful
physical significance. He discovered that if, and only7 if, the force field is linear or
inverse-square, then elliptical orbits result. In the Principia Newton explicitly drew
attention to this coincidence, calling it “very remarkable”. As the Nobel physi-
cist Subrahmanyan Chandrasekhar (1995, p. 287) observed, “nowhere else in the
Principia has Newton allowed himself a similar expression of surprise.”
We are left with something of a mystery. There appears to be some special
connection between the linear and inverse-square force fields, but what could it
possibly be? Newton himself was able to find a connection, and we shall use com-
plex analysis to find another. For more on both these connections, see Arnol’d
(1990), Needham (1993), and Chandrasekhar (1995).

7
Newton assumed that the force varies as a power of the distance, but it has since been discovered
that the result is still true if we drop this requirement.
 Celestial Mechanics* 277

5.10.3 Changing the First into the Second

The geometry of complex numbers was not yet understood in the time of New-
ton; had it been, he would surely have discovered the following surprising fact.
If we apply the mapping z ÞÑ z2 to an origin-centred ellipse, then the image is
not some strange ugly shape, as one might expect, but rather another perfect ellipse;
furthermore, this ellipse automatically has one focus at the origin! See [5.23]. Before
exploring the implications, let us verify this fact: squaring (5.33),

z ÞÑ z2 = (p eit + q e−it )2 = p2 ei2t + q2 e−i2t + 2pq.

The first two terms correspond to an origin-centred ellipse with foci at ±2pq; the
last term therefore translates the left-hand focus to the origin.
Expressed in dynamical terms, this geometric result states that while leaving the
attracting point fixed at the origin, z ÞÑ z2 transforms an orbit of the linear field into
an orbit of the inverse square field. However, we are only in a position to state the
result in this way because we already know what the orbits in the two fields look
like. Is there instead some a priori reason why z ÞÑ z2 should map orbits of the linear
field to orbits of the gravitational field? If there were such a reason then [5.23] could
be viewed as a novel derivation, or explanation, of the elliptical motion of planets
about the sun as focus.
That there is indeed such a reason was discovered in the first decade of the 20th
century. Several people deserve credit for this beautiful result which, at the time
of this writing, is still not widely known. Apparently Bohlin (1911) was the first to
publish it, not knowing that Kasner (1913) had already discovered a more general
result in 1909. Finally, knowing only of Bohlin’s work, Arnol’d (1990) rediscovered
Kasner’s general theorem.

[5.23] Transformation of the linear force field into the inverse-square force field. Very
remarkably, squaring an origin-centred ellipse results in another ellipse. Furthermore, one
of the foci of the image ellipse is located at the origin! As the text explains, this provides
a novel explanation of the elliptical orbits of the planets around the Sun at one focus.
278 Further Geometry of Differentiation

Before embarking on the details of the explanation, here (following Needham

(1993)) is our plan of attack. In the absence of force a particle will move in a straight
line; bending is therefore the manifestation of force, and this can be quantified in
terms of the curvature of the orbit. Since the mapping z ÞÑ z2 is analytic, we may
use the results of the previous section to find the relationship between the curvature
of an orbit and the curvature of the image orbit produced by the mapping. This
will enable us to find the relationship between the forces that hold the preimage
and image in their respective orbits.

5.10.4 The Geometry of Force

Given an orbit and a centre of force, our aim is to find a purely geometric formula
for the magnitude F of the force F that holds the particle in that orbit. Consider
figure [5.24]. As illustrated, it is conceptually very helpful to decompose F into
components F T and F N that are tangential and normal to the orbit, respectively.
The effect of the component F T is to change the speed v of p without altering its
course. The effect of the component F N is to bend the orbit of p without altering its
speed.
From elementary mechanics we know that if a particle of unit mass moves at
constant speed v round a circle of radius ρ then the force directed towards the centre
is (v2 /ρ). Thus if the curvature of the orbit is κ (as illustrated) then FN = κv2 . If we
call the acute angle between the radius and the normal γ, then it follows that the

[5.24] Geometric determination of the central force law from the orbit. If h is the (con-
stant) angular momentum of the orbiting particle, the magnitude ofthe central
 force field
κ sec3 γ
that holds the particle in its orbit of curvature κ is given by F = h2 r2
.
 Celestial Mechanics* 279

total force acting on p is

F = FN sec γ = κ v2 sec γ.
In order to fully reduce this formula to geometric terms, we need to express v in
geometric terms. This is made possible by the constancy of the angular momentum
h = 2A. If we decompose the velocity v into radial and transverse components vr
and vt , then clearly only the latter generates area. More precisely, h = 2A = r vt =
r v cos γ, so
 sec γ 
v=h . (5.34)
r
Substituting for v in the previous result, we obtain the desired geometric formula
for the force:
 
2 κ sec3 γ
F=h . (5.35)
r2
This result is essentially due to Newton (1687, Prop. VII). Observe that the concept
of time has almost disappeared in this formula, the only vestige being the constant
h that specifies how fast the orbit is traversed.

5.10.5 An Explanation
As z describes an arbitrary orbit, (5.35) tells us the force F needed to hold it in that
orbit. Now apply the mapping z ÞÑ z2 , and let a tilde denote a quantity associated
with the image, e.g., er = r2 . The force e
F needed to hold the image in its orbit is
 
e e 2 e
κ sec3 γe
F=h ,
er 2
and we now seek to relate this to the original force F.
First, to find e
κ, simply put f(z) = z2 into (5.24) and thereby obtain [exercise]
1 h cos γ κ i
e
κ= + .
2 r2 r
Next, observe that since the ray from 0 to z maps to the ray from 0 to z2 , the
conformality of the mapping implies that γ e = γ.
e
Putting these facts into the formula for F, and substituting for the original speed
and force from (5.34) and (5.35), we get
!2  
e
h 1 2
v + 1
rF
e
F= 2 2
. (5.36)
h er 2

Even if F is a simple power law, generally this e F will not be. However, if and only
8
if the original force field is linear , the numerator in the above expression magically

8
However, as we shall see in a moment, it could be the repulsive linear field F = −r instead of the
attractive one F = +r.
280 Further Geometry of Differentiation

becomes the constant total energy E of the particle in the original field:
!2
e
h E
e
F= ! (5.37)
h er 2

The image therefore moves in a field that is inverse-square, as was to be shown.

Here is a fact which may have been bothering you already. The only gravita-
tional orbits we have managed to explain in this way are the ellipses; where are the
hyperbolic orbits which we know are also possible in a gravitational field? In fact
the geometry of z ÞÑ z2 does explain these, the resolution being that gravitational
orbits arise not only as the images of orbits in a linear field that is attractive, but
also of orbits in a linear field that is repulsive, F = −r. The orbits in this field are
hyperbolae with centre (i.e., intersection of asymptotes) at the origin, and z ÞÑ z2
maps these to hyperbolae with one focus at the origin.
The dynamical explanation is almost unchanged: the constant total energy of
the particle in the original repulsive linear field is now given by E = 12 (v2 − r2 ), so
inserting F = −r in (5.36) once again yields (5.37). See Needham (1993) for more
on this, as well as the general result we are about to state, which may be proved in
exactly the same way as the special case above.

5.10.6 The Kasner–Arnol’d Theorem

The power laws F ∝ r and e F ∝ er −2 are examples of what Arnol’d calls dual force
laws, and both he and Kasner discovered that they constitute just one example of
duality. Here is the general result:

Associated with each power law F ∝ rA there is precisely one power law
e e
F ∝ er A that is dual in the sense that orbits of the former are mapped to
orbits of the latter by z ÞÑ zm , and the relationships between the forces
and the mapping are:

e + 3) = 4 (A + 3)
(A + 3)(A and m = .
2
To their result we add the following point of clarification on the role of energy:

In general, positive energy orbits in either the attractive or repulsive

field F ∝ rA map to attractive orbits in the dual field, while negative
energy orbits map to repulsive ones. However, if −3 < A < −1 (e.g.,
gravity) then these roles are reversed. In all cases, zero energy orbits map
to force-free rectilinear orbits.
 Analytic Continuation* 281

5.11 Analytic Continuation*

5.11.1 Introduction
Throughout this book we have stressed how functions may be viewed as geometric
entities that need not be expressed (nor even be expressible) in terms of formulae.
As an illustration of the limitations of formulae, consider

G(z) = 1 + z + z2 + z3 + · · · .

This power series converges inside the unit circle |z| = 1, and consequently it is
analytic there. Figure [5.25] shows a grid of little squares inside this circle being
amplitwisted by z ÞÑ w = G(z) to another such grid lying to the right of the verti-
cal line Re (w) = 12 , which itself is the image of the circle. Now this circle is certainly
a barrier to the power series formula for G, since it clearly diverges at 1: geomet-
rically, the image of the circle extends to ∞. However, the circle is not a barrier to
the geometric entity that the formula is attempting to describe.
Consider a somewhat different-looking power series centred at −1:
"    2 #
1 z+1 z+1
H(z) = 1+ + + ··· .
2 2 2

This series is analytic inside a larger circle of convergence |z + 1| = 2. Despite the

apparent difference, H(z) maps the previously considered solid grid inside |z| = 1
to precisely the same grid on the right of x = (1/2) as G did: H = G inside |z| = 1.
But now the grid may be extended to the dotted one lying outside |z| = 1, and H
amplitwists it to the dotted grid lying to the left of x = (1/2). We say that H is an
analytic continuation of G to the larger disc. An obvious question is whether H is the
only analytic continuation of G to this region. As we hope [5.25] makes palpable,
the rigidity imposed by being locally an amplitwist does indeed force the mapping
to grow in a unique way.
The objective of this final section will be to make this rigidity clearer, and also to
describe one method (due to Schwarz) of explicitly finding the mapping in regions
beyond its original definition. Before doing this, however, we will complete our
discussion of [5.25].
The figure makes it plain that H is no more the end of the line than G was: it too
can be continued. But if we cling to power series then the scope of our description
of the mapping that underlies both G and H will be strictly limited. This is because
such series only converge inside discs, and if we try to expand any disc then it will
eventually hit the singularity at z = 1 and then be unable to go round it. Thus any
power series will necessarily miss out at least half of the potential domain of the
mapping. On the other hand, as you may have already noticed, the Möbius trans-
formation 1/(1 − z) is analytic everywhere except z = 1, and it agrees with both
282 Further Geometry of Differentiation

[5.25] Analytic Continuation. The origin-centred power series G(z) = 1 + z + z2 + z3 + · · ·

only converges inside the unit disc. The image w = G(z) of the grid of small “squares” on
the left (inside the unit disc) is shown [solid] on the right, in the region Re (w) > 12 . But if
we extend the grid [dotted] on the left beyond the unit circle, then there is a unique analytic
continuation of the image grid [dotted] on the right into the region 14 < Re (w) < 12 ,
expressible as a power series that converges inside the larger circle |z + 1| = 2. If we
abandon power series altogether, then the grid can in fact be extended uniquely to the
entire plane, the complete analytic continuation being given by z ÞÑ 1/(1 − z), with the
singularity z = 1 being mapped to the point at ∞.

G and H within their circles of convergence; it thus constitutes the complete ana-
lytic continuation of the mapping. [We encourage you to use this fact to check the
details of the figure.] The simplicity of this example is perhaps misleading. Usu-
ally one cannot hope to capture the entire geometric mapping within a single closed
expression such as 1/(1 − z).
When one stares at a figure like [5.25] one starts to sense the rigid growth of the
mapping due to the analytic requirement that an expanding mesh of tiny squares
must map to another such mesh. It also becomes clear how the mapping itself is
oblivious to the different formulae with which we try to describe it. Indeed we have
seen that the two circles—such formidable and impenetrable barriers to the power
series—have only a slight significance for the mapping itself: both hit z = 1, so both
images extend to ∞.
 Analytic Continuation* 283

5.11.2 Rigidity
The essential character of analytic rigidity is captured in the following result:

If even an arbitrarily small segment of curve is crushed to a point by an

analytic mapping, then its entire domain will be collapsed down to that
point.
The theory of integration to be developed in the following chapters will provide a
convincing explanation of this fact. For the present, though, we can obtain a good
measure of insight into its truth by extending our previous discussion of critical
points (page 231). This may give the illusion of dispensing with integration theory,
but as we pointed out at the time, that discussion also had to draw on future results.
We now recap the relevant facts concerning critical points.
The amplification vanishes at a critical point p, leading to the impression that an
infinitesimal disc centred there is crushed down to a point. However, this is merely
a ‘trick of the light’ due to low magnification of the image plane. If the order of p is
(m − 1), so that the mapping locally resembles zm , then an infinitesimal disc at p of
radius ϵ will be mapped [m-fold] onto a vastly smaller disc of radius ϵm . In terms
of the microscope analogy this means that we must use the Lm lens to see that the
image isn’t a perfect point. The greater the order, the greater the degree of crushing
at p, and the greater the power of the first lens that will reveal the nonpointlike
image.
Now observe that, calculationally speaking, the role of the increasingly high-
powered lenses that fail to resolve the image is taken over by the increasingly high-
order derivatives that vanish at p:

L1 , . . . , Lm−1 show nothing, but image visible with Lm
f(z) ∼ z ⇐⇒
m
f ′ (p) = 0, f ′′ (p) = 0, . . . , f(m−1) (p) = 0, but f(m) (p) ̸= 0

In short, the higher the derivative that vanishes at p, the greater the degree of
crushing at p.
We now apply this insight to the given situation. Let s be the (possibly) tiny seg-
ment that is crushed by f(z). The amplification of f at a point of s may be read off
by looking in any direction. By choosing to look along s we find that the amplifica-
tion vanishes at each point of s. The entire segment is therefore made up of critical
points for which f ′ = 0. Now think of f ′ as an analytic mapping in its own right, just
as we did in [5.18]. We have just seen that this mapping automatically possesses the
same property as f did: it crushes s to a point. We conclude that its derivative must
also vanish on s. Clearly there is no end to this; all the derivatives of f must vanish,
and, correspondingly, infinitesimal discs centred on s must be totally crushed.
This means that there is at least a sheathlike region surrounding s which is com-
pletely crushed by f. But if we take a new curve lying in this region, the whole line
284 Further Geometry of Differentiation

of thought may be repeated to deduce that f must crush a still larger region. The
collapse of the function therefore proceeds outwards (at the speed of thought!) to
the entire domain.

5.11.3 Uniqueness
Suppose that A(z) and B(z) are both analytic functions defined on a region that
happens to be the same size and shape as California. Suppose, further, that A and
B both happen to have the same effect on a tiny piece of curve, say a fallen eyelash
lying in a San Francisco street. This tiny measure of agreement instantly forces them
into total agreement, even hundreds of miles away in Los Angeles! For (A − B) is
analytic throughout California, and since it crushes the eyelash to 0, it must do the
same to the entire state.
We can express this slightly differently. If we arbitrarily specify the image points
of a small piece of curve s, then in general there will not exist an analytic function
that sends s to this image. However, the previous paragraph assures us that if we
can find such a function on a domain including s, then it is unique.
This is the “compelling reason” we referred to earlier in connection with the
uniqueness of the generalization of ex to complex values. For if an analytic gener-
alization E(z) exists, then we see that it will be uniquely determined by the values
of ex on even a small piece of the real axis. Of course knowing this does not help in
the least to find out what E(z) actually is. The value of our previous derivations of
explicit expressions for E(z) therefore remains undiminished. On the other hand,
the new knowledge is not without practical implications. Consider these three very
different-looking expressions:
 z n
lim 1 + , ex (cos y + i sin y) , 1 + z + z2 /2! + z3 /3! + · · · .
nÑ∞ n
They are all analytic, and they all agree with ex when z is real. Thus, without further
calculation, we know they must all be equal to each other, for they can only be
different ways of expressing the unique analytic continuation of ex .
New and important aspects of uniqueness emerge when we consider domains
that merely overlap, rather than coincide. Let g(z) and h(z) be analytic functions
defined on the sets P and Q shown in [5.26a]. If they agree on even a small segment
s in P ∩ Q then they will agree throughout P ∩ Q. If we imagine that we initially
only know about g on P, then we may think of h as describing the same geometric
mapping as g but with the domain P extended to encompass Q. We are encouraged
in this view by the fact that g uniquely determines this analytic continuation. For
suppose h∗ were another continuation of g into Q. On s we would then have h∗ =
g = h, but this forces h∗ = h throughout their common domain Q.
The functions G(z) and H(z) of the introduction furnish a concrete example of
the above, where P happens to lie wholly within Q. The function 1/(1 − z) then
constitutes the analytic continuation of H to the rest of the plane.
 Analytic Continuation* 285

[5.26] Analytic continuation along alternative routes. [a] If g(z) and h(z) are analytic
functions defined in overlapping regions P and Q, and they agree on even a small segment
s in this overlap, then they must agree throughout P ∩ Q, and we are entitled to think of
h(z) as the unique analytic continuation of g(z) from P into Q. [b] Although the analytic
continuations along PQR · · · S and along PQ eR
e · · · S are both uniquely determined, they
need not agree on S: in this case we have a multifunction.

[5.27] Analytic continuation of log z. In P we can define a single-valued branch of log z

given by f(z) = ln r+i θ, where −π < θ ⩽ π. The figure shows how 1 maps to 0, and how
P maps to the surrounding region f(P). The definition of log z then has unique analytic
e but they disagree at −1, for example.
continuations into Q and into Q,

Just as g was continued from P to Q, so we may continue the process along a

whole chain of overlapping sets P, Q, R, . . . , S, as in [5.26b]. We thereby obtain a
unique analytic continuation of g to S. But what if we chose an alternative route
e R,
such as P, Q, e . . . , S ? Once again the continuation of g to S is unique, but this need
not agree with the first continuation. The idea of analytic continuation has thus led
very naturally to the idea of multifunctions.
Consider [5.27]. In P we can define a single-valued branch of log z given by f(z) =
ln r + i θ, where −π < θ ⩽ π. The figure shows how 1 maps to 0, and how P maps
to the surrounding region f(P). If we define g(z) = ln r + iΘ (with − π2 < Θ ⩽ 3π 2 )
286 Further Geometry of Differentiation

then since g = f on P ∩ Q, it must be the analytic continuation of f to Q. Likewise

e is g̃(z) = ln r + iΘ,
its continuation to Q e where − 3π < Θ e ⩽ π , for example. In
2 2
the region surrounding −1 we now have two unique continuations of one and the
same function f. But despite this common ancestry, they clearly disagree with each
other: g(−1) = iπ, while g̃(−1) = −iπ.

5.11.4 Preservation of Identities

In this subsection we will show that any identities that hold for real functions must
continue to hold for their analytic generalizations to C (assuming such exist). This
is easiest to explain through examples.
First we consider an important example dealing with power series. Suppose that
the real function f(x) can be represented by a convergent power series

f(x) = a + bx + c x2 + d x3 + · · · .

We therefore know that the complex series

F(z) = a + bz + c z2 + d z3 + · · · .

is convergent and hence analytic. But since F(x) = f(x) on the real axis, it follows
that F is the unique analytic continuation of f to complex values. In other words, the
transition from f to its analytic continuation does not change the formula (series).
For our next example we consider a real identity involving two variables: ex ·ey =
ex+y . It will help to appreciate the argument if you can be temporarily stricken with
amnesia, so that the complex function ez and its associated geometry suddenly
mean nothing to you. Suppose that an analytic continuation of ex to complex values
exists, and call it E(z). We can now show that E must be subject to precisely the same
law, and without even knowing what E is!
Let Fζ (z) ≡ E(ζ)·E(z), and let Gζ (z) ≡ E(ζ+z). First note that for fixed ζ both F(z)
and G(z) are analytic functions of z. Now suppose that ζ is real, so that E(ζ) = eζ . If
z now moves on a segment of the real axis then it follows from the real identity that
F(z) = G(z); but from our recent results we know this implies that they are equal
everywhere. If we hold z fixed instead, then analogous reasoning yields Fζ = Gζ ,
and we conclude that

E(ζ) · E(z) = E(ζ + z)

for complex values of both ζ and z. It should be clear that this reasoning extends
to any identity, even one involving more than two variables.

5.11.5 Analytic Continuation via Reflections

Quite distinct from questions of existence and uniqueness is the problem of actu-
ally finding an analytic continuation. The above ideas and results are mute on this
 Analytic Continuation* 287

[5.28] Extending an analytic mapping via reflection. If the analytic function f maps P to
Q, then we can use it to construct an analytic mapping from P to Q, given by f⋆ (z) ≡ f(z).
As we see, the two reflections reverse angles twice, resulting in a conformal mapping.

issue, although it could reasonably be claimed that the persistence of identities

is a practical help. We next explain a Symmetry Principle (due to Schwarz) which
enables one to find a continuation easily and explicitly, albeit under rather special
circumstances.
We first describe how it is possible to use two reflections to construct a new
analytic function from an old one. Suppose an analytic function f is defined on a
region P, the image of which is Q (see [5.28]). Let P and Q be the reflections of these
regions across the real axis. We can now use f to construct an analytic mapping from
P to Q, namely

f⋆ (z) ≡ f(z).

The figure explains why f⋆ is conformal, and hence analytic. All three stages,
a ÞÑ a ÞÑ f(a) ÞÑ f(a), preserve the magnitude of an angle at a; the first reflec-
tion reverses the sense, then f preserves the reversed sense, and finally the second
reflection undoes the damage, restoring the angle to pristine condition at f⋆ (a).
In general this mapping f⋆ will not be a continuation of f in any sense, rather
it is an entirely new mapping. This should become clear if you imagine moving P
downwards until some of it crosses the real axis. P and P now overlap so that P ∩ P
constitutes a common domain for f and f⋆ , but we hope you can see that there is no
reason for them to agree with each other. This is clearer still if we take an example:
exercise
f = (rotation of ϕ) =⇒ f⋆ = (rotation of − ϕ).

Although it is generally not a continuation of f, this new mapping f⋆ (together with

its soon to be introduced generalization to circles) is very useful in its own right. In
288 Further Geometry of Differentiation

[5.29] Analytic continuation via reflection. Suppose that the analytic function f defined
in P is the complex extension of a real function defined on L. Then the function f⋆ (z) ≡
f(z) introduced in the previous figure is the analytic continuation of f into P. Generalizing,
if f maps a line-segment L (not necessarily real) to another line-segment L b , then we
can analytically continue from one side of L to the other by using the fact that points
symmetric in L map to points symmetric in L b.

Chapter 12 we will show that it is intimately connected with the so-called method
of images of electrostatics and fluid dynamics.
We now turn to the special circumstance under which f⋆ is the analytic continu-
ation of f. Suppose that f is itself the complex generalization of a real function, and
let P have a part L of its boundary along the real axis, as in [5.29]. Since f is real on L,
the image set Q will also border on the real axis. Unlike the general situation previ-
ously considered, f and f⋆ will now automatically agree on their common domain
P ∩ P = L, for if z is real then

f⋆ (z) = f(z) = f(z) = f(z).

We can now think of f and f⋆ as being two parts of a single analytic mapping F
on P ∪ P. Indeed, by considering what happens to the two halves of an infinites-
imal disc centred on L, it’s clear that F is analytic there, for the image is another
infinitesimal disc. [What happens if we are at a critical point?] Once again, notice
how different this is from the case of real functions, for we could easily join two
pieces of graph together with a kink at the join; their values would then agree,
while their derivatives would not.
Of course if f is already defined in P (as well as P) then f⋆ must simply repro-
duce the mapping that’s already there. For example, the formula for the complex
generalization sin z is valid everywhere, so it should be subject to the symmetry
f⋆ (z) = f(z). Indeed if we follow the three steps of a ÞÑ f⋆ (a) then we do find that
eia − e−ia e−ia − eia
a ÞÑ a ÞÑ ÞÑ = sin a .
2i −2i
 Analytic Continuation* 289

We can rephrase our result in a more symmetric and slightly generalized [exer-
cise] form. If f maps a line-segment L (not necessarily real) to another line-segment
b , then we can analytically continue from one side of L to the other by using the
L
b.
fact that points symmetric in L map to points symmetric in L
This sounds very reminiscent of the conservation of symmetry by Möbius trans-
formations that we discovered in Chapter 3, and indeed by fusing these two
symmetry principles we can obtain a significant generalization of our result. Sup-
pose that instead of mapping a particular9 line to a line, f sends a part C of a circle
to a part C b of another circle. We can reduce this to the previous case by using
two Möbius transformations to send C ÞÑ L, and C b ÞÑ Lb . We deduce that points
symmetric in C map to points symmetric in C b.
As a mixed example, imagine that f maps part of the unit circle to part of the
real axis. If f is only known inside the circle then the above result tells us [exercise]
that there is an analytic continuation to the exterior given by
 
† 1
f (z) ≡ f .
z
The complete analytic function F is then defined to be f inside the circle, and f†
outside the circle. By construction, this function sends symmetric pairs of points to
conjugate images: F† (z) = F(z).
Using what is now known as Schwarzian reflection10 , in 1870 H. A. Schwarz was
able to generalize his Reflection Principle beyond lines and circles to more general
curves. We end this chapter with a description of Schwarz’s simple, yet fascinating,
idea. The key is to use an analytic function to fake conjugation.
We know that reflecting every point across the real axis (z ÞÑ z) is not an analytic
function. However, given a sufficiently smooth11 curve K, it is possible to find an
analytic function SK (z) that selectively sends just the points of K to their conjugates:

z∈K =⇒ SK (z) = z .

Davis and Pollak (1958) christened SK the Schwarz function of K. We can now define
the Schwarzian reflection of z across K to be e
z = ℜK (z), where

ℜK (z) ≡ SK (z) .

To see why this is a good idea, consider [5.30]. First note that points on K are
unaffected, in accord with the ordinary notion of reflection, e.g.,

e = SK (q) = (q) = q .
q

9
We stress “particular”, because if a general line were sent to a line then the mapping could only be
linear. Similarly, in the new case, if a general circle were sent to a circle then the mapping would have
to be a Möbius transformation.
10
Schwarz (1972a).
11
The curve must in fact be “analytic”. On this point, see Davis (1974), which also contains many
interesting applications of the Schwarz function. A more advanced work is Shapiro (1992).
290 Further Geometry of Differentiation

[5.30] Schwarzian Reflection. The Schwarz function SK of a curve K is defined to be

the analytic mapping that mimics conjugation on K: if z lies on K, then SK (z) = z. We
can then define Schwarzian reflection across K to be z ÞÑ e z = ℜK (z) ≡ SK (z). As we
see, very close to K, Schwarzian reflection becomes ordinary reflection. If K is a line or
circle, then Schwarzian reflection is precisely reflection/inversion in K.

Next, observe that since SK is analytic, an infinitesimal disc centred at q is

amplitwisted (not reflected) to a disc centred at q. Furthermore, by noting how
qp is mapped to qp, it follows that on K

amplification = 1 and twist = −2ϕ =⇒ S K′ = e−i2ϕ ,

where ϕ is the angle that the tangent to K makes with the horizontal. It is now clear
from the symmetry of the figure that if the point a is on the infinitesimal circle, then
e is indeed its reflection across the tangent of K. Thus, at least very close to K, z ÞÑ e
a z
is a reasonable generalization of the reflection concept. Furthermore, reflecting in
K twice yields the identity mapping, as it should. For since ℜK is anticonformal,
ℜK ◦ ℜK is conformal, i.e., analytic. But since this function maps each point of K to
itself, and since an analytic function is determined by its values on a curve, ℜK ◦ℜK
must be the identity mapping.
We leave it to the exercises for you to show that if K is a line or a circle then e z
is just the ordinary reflection, even if z is far from K. For example, the unit circle C
may be written as zz = |z|2 = 1, so that on C we have z = (1/z). Thus its Schwarz
function is SC (z) = (1/z), and so ℜC (z) = (1/z), which is just inversion in C.
 Analytic Continuation* 291

[5.31] Schwarzian reflection in an ellipse.

Let us give a less trivial example, namely, reflection in the ellipse E with equation
(x/a)2 + (y/b)2 = 1. Writing x = 12 (z + z) and y = 2i 1
(z − z), then solving for z in
terms of z, we find [exercise] that
1 h p i
SE (z) = 2 (a 2
+ b 2
) z − 2ab z 2 + b 2 − a2 .
a − b2
With a = 2 and b = 1, for example, Schwarzian reflection is given by
h p i
ℜE (z) = 13 5 z − 4 z2 − 3 ,
which is illustrated in [5.31]. We encourage you to verify this figure with your
computer, as well as to examine the effect of ℜE on other shapes.
With the proper concept of reflection in hand, we may now generalize the above
method of analytic continuation across lines and circles to more general curves.
Let f be an analytic mapping defined in a region P bordering on a curve L that is
smooth enough to possess a Schwarz function, and let L b = f(L) be the image of L
under f. Much as in figures [5.28] and [5.29], we may now analytically continue f
across L by demanding that points that are symmetric in L map to points that are
symmetric in Lb . Thus [draw a picture!] the continuation f‡ of f to P
e = ℜL (P) is

f‡ = ℜLb ◦ f ◦ ℜL .
e for it is the compo-
By the same argument as in [5.28], this is indeed analytic in P,
sition of one conformal mapping with two anticonformal mappings. Also, f‡ = f
e is then subject to
on L. The complete analytic function F given by f in P and f‡ in P
‡
the symmetry F = F. This is Schwarz’s Symmetry Principle.
292 Further Geometry of Differentiation

Our previous results are just special cases of this construction. For example, if L
and Lb are segments of the real line then ℜL (z) = ℜ b (z) = z, so f‡ = f⋆ , as before.
L
b is a segment of
Similarly, if L is an arc of the unit circle (so that ℜL (z) = 1/z) and L
‡ †
the real line (so that ℜLb = z), then f = f , as before.
 Exercises 293

5.12 Exercises

·
1 Show that if f = u + iv is analytic then (∇u) (∇v) = 0, where ∇ is the
“gradient operator” of vector calculus. Explain this geometrically.
2 Show that the real and imaginary parts of an analytic function are harmonic,
i.e., they both automatically satisfy Laplace’s equation:

∆Φ = 0 ,

where ∆ (which is often instead written ∇2 ) is defined by ∆ ≡ Bx2 + By2 and is

called the Laplacian. [In Chapter 12 we will see that this equation represents a
crucial link between analytic functions and physics.]
3 Use the previous exercise (not calculation) to show that each of the following
is “harmonic”.
(i) ex cos y.
2 2
(ii) e(x −y ) cos 2xy.
(iii) ln |f(z)|, where f(z) is analytic.
4 What is the most general function u = a x2 +b xy+c y2 that is the real part of an
analytic function? Construct this analytic function, and express it in terms of z.
5 Which of the following are analytic?
(i) e−y (cos x + i sin x).
(ii) cos x − i sin y.
(iii) r3 + i3θ.
 
(iv) r er cos θ ei(θ+r cos θ) .
6 Solve the Polar CR equations given that Bθ v ≡ 0. Express your answer in terms
of a familiar function, and interpret everything you have done geometrically.
7 Use the Cartesian CR equations to show that the only analytic mapping that
sends parallel lines to parallel lines is the linear mapping. [Hint: Begin with
the case of horizontal lines being mapped to horizontal lines. How does this
translate into ‘Equationspeak’? Now solve CR.]
8 Calculate, then draw on a picture, a possible location for log(1 + i). Draw a
small shape at 1 + i. Use the amplitwist of log(z) to draw its image. Verify this
using your computer.
9 Derive the quotient rule in an analogous way to the product rule (see
page 256). [Hint: Multiply top and bottom of (A/B) by (b − ξg ′ ).]
10 Consider the polynomial P(z) = (z − a1 )(z − a2 ) . . . (z − an ).
294 Further Geometry of Differentiation

(i) Show that the critical points of P(z) are the solutions of
1 1 1
+ + ··· + =0.
z − a1 z − a2 z − an
(ii) Let K be a circle with centre p. By considering the conjugate of the
equation in (i), deduce that p is a critical point if and only if it is the
centre of mass of the inverted points IK (aj ).
(iii) Show that the equation in (i) is equivalent to
z − a1 z − a2 z − an
+ + ··· + = 0,
|z − a1 |2 |z − a2 |2 |z − an |2
and by interpreting the LHS as a (positively) weighted sum of the vectors
from z to the roots of P(z), deduce Lucas’ Theorem: The critical points of a
polynomial in C must all lie within the convex hull of its zeros. This is a com-
plex generalization of Rolle’s Theorem in ordinary calculus. [Hint: Use the
fact that (2.32) on page 116 is still valid even if the masses are not equal.]
11 Use (ez ) ′ = ez to show that the derivatives of all the trig functions are given
by the familiar rules of real analysis.
12 Provided it is properly interpreted, show that (zµ ) ′ = µ zµ−1 is still true even
if µ is complex.
13 (i) If a is an arbitrary constant, show that the series
a(a − 1) 2 a(a − 1)(a − 2) 3
f(z) = 1 + az + z + z + ···
2! 3!
converges inside the unit circle.
(ii) Show that (1 + z)f ′ = af.
(iii) Deduce that [(1 + z)−a f] ′ = 0.
(iv) Conclude that f(z) = (1 + z)a .
14 As we pointed out in Chapter 3, stereographic projection has a very practical
use in drawing a conformal map of the world. Once we have this map we
can go on to generate further conformal maps, simply by applying different
analytic functions to it. One particularly useful one was discovered (using
other means) by Gerhard Mercator in 1569. We can describe it (though he
could not have) as the result of applying log(z) to the stereographic map.
(i) Look up both a stereographic map and a Mercator map in an atlas,
and make sure you can relate the changes in shape you see to your
understanding of the complex logarithm.
(ii) Imagine plotting a straight-line course on a Mercator map and then
actually travelling it on the high seas. Show that as you sail, the reading
of your compass never changes.
 Exercises 295

15 (i) By noting that the unit tangent (in the counterclockwise direction) to an
origin-centred circle can be written as b ξ = i(z/|z|), show that formula
(5.24) for the curvature of the image of such a circle can be written as
h i
1 + Re z f ′′
′
e
κ= f .
|z f ′ |

(ii) What should this formula yield if f(z) = log z? Check that it does.
(iii) What should this formula yield if f(z) = zm ? Check that it does. What
is the significance of the negative value of e
κ when m is negative? [Hint:
Which way does the velocity complex number of the image rotate as z
travels counterclockwise round the original circle?]
16 As illustrated below, a region is called convex if all of it is visible from an arbi-
trary vantage point inside. Let an analytic mapping f act on an origin-centred
circle C to produce a simple image curve f(C), the interior of which is convex.
(i) From the formula of Ex. 15, deduce that if f maps the interior of C to the
interior of f(C), then the following inequality holds at all points z of C:
 
z f ′′
Re ⩾ −1.
f′

(ii) What is the analogous inequality when f maps the interior of C to the
exterior of f(C). [Hint: Ex. 4, p. 239.]

17 Let S be a directed line-segment through a point p = x + iy.

(i) Let f(z) = ez . Without calculation, decide which direction of S yields an
image f(S) having vanishing curvature at f(p).
(ii) The complex curvature K must therefore point in one of the two orthogo-
nal directions. Which? By considering the image of S when it points in this
direction, deduce the value of |K|, and thereby conclude that K(p) = ie−x .
(iii) Use (5.29) to verify this formula.
296 Further Geometry of Differentiation

(iv) Repeat as much as possible of the above analysis in the cases f(z) = log(z)
and f(z) = zm , where m is a positive integer. [In neither of these cases
will you be able to see the exact value of |K(p)|.]
(v) According to the geometric reasoning in Ex. 18, p. 241, the amplification
of a Möbius transformation M(z) = az+b cz+d is constant on each circle
centred at −(d/c). Thus the complex curvature of M should be tangent
to these concentric circles. Verify this by calculating K.
(vi) Use a computer to verify figure [5.21] for all four mappings above.
18 Let two curves C1 and C2 emerge from a point p in the same direction. Two
examples are illustrated below.

Although in both cases the angle at p is zero, there is a great temptation to

say that the curves on the right meet at a smaller “angle” than those on the
left. Any putative definition of such an “angle” Θ should (presumably) be
conformally invariant: if the curves are mapped to C e1 and C e2 by a mapping
f that preserves ordinary angles (i.e., an analytic mapping), then the new
“angle” Θe should equal the old “angle” Θ.
(i) Newton (1670) attempted to define such a Θ as the difference of the
curvatures of C1 and C2 at p: Θ ≡ κ1 − κ2 . Use (5.31) to show that this
e = Θ/|f ′ (p)|.
definition is not quite conformally invariant: Θ
(ii) Consider an infinitesimal disc D (radius ϵ) centred at p. Let c1 and c2 be
the centres of curvature of C1 and C2 , and let D be the difference between
the angular sizes of D as seen from c1 and c2 . Show that D = ϵΘ. If a
conformal mapping f is applied to C1 , C2 , and D, deduce that D e = D. [Of
course this is not what we were after: D is (a) infinitesimal, and (b) not
defined by the curves alone. The discovery of a true conformal invariant
had to await Kasner (1912). See Ex. 10, p. 649.]
 Exercises 297

19 In more advanced work on Möbius transformations (e.g., Nehari (1952) and

Beardon (1984)), an important role is played by the so-called Schwarzian
derivative {f(z), z} of an analytic function f(z) with respect to z:
 ′′  ′  2
f 1 f ′′
{f(z), z} ≡ − .
f′ 2 f′
(i) Show that the Schwarzian derivative may also be written as
 2
f ′′′ 3 f ′′
{f(z), z} = ′ − .
f 2 f′
(ii) Show that {az + b, z} = 0 = {(1/z), z}.
(iii) Let f and g be analytic functions, and write w = f(z). Show that the
Schwarzian derivative of the composite function g[f(z)] = g[w] is given
by the following “chain rule”:

{g(w), z} = [f ′ (z)]2 {g(w), w} + {f(z), z}.

(iv) Use the previous two parts to show that all Möbius transformations
have vanishing Schwarzian derivative. [Hint: Recall that the mappings
in part (ii) generate (via composition) the set of all Möbius trans-
formations.] Remark: Ex. 19, p. 481 shows that the converse is also
true: If {f(z), z} = 0 then f = Möbius. Thus Möbius transformations are
completely characterized by their vanishing Schwarzian derivative.
(v) Use the previous two parts to show that the Schwarzian derivative is
“invariant under Möbius transformations”, in the following sense: if M
is a Möbius transformation, and f is analytic, then

{M[f(z)], z} = {f(z), z}.

20 Think of the real axis as representing time t, and let a moving particle
w = f(t) (where f(z) is analytic) trace an orbit curve C. The velocity is then
v = ẇ = f ′ (t).
(i) Use (5.24) to show that the curvature of C is
Im(v̇/v)
e
κ= .
|v|
(ii) Argue that this result does not in fact depend on C being produced by
an analytic mapping, but is instead true of any motion for which the
velocity v and acceleration v̇ are well-defined.
(iii) Show that the formula may be rewritten as
Im(v v̇)
e
κ= .
|v|3
298 Further Geometry of Differentiation

(iv) Deduce that it may also be written vectorially as

|v×v̇|
e
κ= .
|v|3
By considering C to be the “osculating plane”—see Needham (2021)—of
a curve in 3-dimensional space, we see that this formula holds in that
case also.
21 In 3-dimensional space, let (X, Y, Z) be the coordinates of a moving particle. If
X = a cos ωt, Y = a sin ωt, Z = bt, then the path traced by the particle is a helix.
(i) Give interpretations for the numbers a, ω, and b.
(ii) If a and ω remain fixed, what does the helix look like in the two limiting
cases of b becoming very small or very large? What if a and b remain
fixed while ω becomes very small or very large?
(iii) What limiting values would you anticipate for the curvature of the helix
for each of the limiting cases considered in (ii)?
(iv) Use Ex. 20(iv) to show that the curvature of the helix is
a ω2
e
κ= ,
b2 + a2 ω2
and use this to confirm your hunches in (iii).
22 Continuing from Ex. 20, take f(z) to be a general Möbius transformation:
at + b
w = f(t) = ,
ct + d
where ∆ ≡ (ad − bc) ̸= 0. Show that the curvature of this path is
 2  
2c d
e
κ= Im ,
∆ c
in agreement with Ex. 18, p. 210. The fact that this is constant provides a new
proof that the image is a circle, for only circles have constant curvature.
23 As another continuation of Ex. 20, let us see how the Schwarzian derivative
{f(z), z} of Ex. 19 arises rather naturally in the context of curvature.
(i) Show that
de
κ 1
= ′ Im{f(z), z}.
dt |f |
[This formula was discovered by G. Pick. For elegant applications,
see Beardon (1987). For another connection between curvature and
Schwarzian derivatives, see Ex. 28(iii).]
(ii) Use Ex. 19, part (iv) to deduce that if f(z) is a Möbius transformation then
d
e
dt κ = 0. Why is this result geometrically obvious?
 Exercises 299

24 Let the position at time t of a moving particle in C be z(t) = r(t) eiθ(t) .

(i) Show that the acceleration of the particle is
 
z̈ = r̈ − r θ̇2 eiθ + [2ṙθ̇ + r θ̈] ieiθ .
(ii) What are the radial and transverse components of the acceleration?
(iii) If the particle is moving in a central force field, with the centre of force
at the origin, deduce that the areal speed A = (r2 θ̇/2) is constant. For
a beautiful geometric proof of this fact, see Newton (1687, Prop. 1), a
comic-strip rendition of which may be found in Needham (2021, §11.7.3).
25 Sometimes the circle of convergence of a power series is so densely packed
with singularities that it becomes a genuine barrier for the geometric map-
ping, beyond which it cannot be continued. This is called a natural boundary.
An example of this is furnished by
f(z) = z + z2 + z4 + z8 + z16 + · · · ,
which converges inside the unit circle. Show that every point of |z| = 1 is either
a singularity itself, or else has singularities arbitrarily near to it. [Hint: What is
f(1)? Now note that f(z) = z + f(z2 ), and deduce that f is singular when z2 = 1.
Continuing in this manner, show that the 2n -th roots of unity are all singular.]
26 Unlike inversion in a circle, show that Schwarzian reflection in an ellipse E
(see figure [5.31]) does not interchange the interior and the exterior. Indeed,
how does ℜE (z) behave for large values of |z|?
27 (i) If L is a line passing through the real point X, and making an angle σ
with the horizontal, then show that its Schwarz function is
SL (z) = z e−i2σ + X(1 − e−i2σ ) .
(ii) If C is a circle with centre p and radius r, show that its Schwarz function is
r2
SC (z) = p + .
z−p
(iii) Verify the claim that in both these cases z ÞÑ ℜ(z) is the ordinary
reflection, even if z is far from the curve.
28 Let a be a point on a (directed) curve K having Schwarz function S(z).
(i) Show that the curvature of K at a is
i S ′′ (a)
κ ≡ ϕ̇ = · ′ ,
2 [ S (a) ]3/2
where ϕ is the angle in [5.30], and the dot denotes differentiation with
respect to distance l along K (in the given sense). Deduce that
|κ| = |S ′′ /2|.
300 Further Geometry of Differentiation

[Hints: Since S is analytic, so is S ′ . Thus to calculate S ′′ = dS ′ /dz we need

only find the change dS ′ in S ′ produced by an infinitesimal movement dz
of z, taken in any one direction of our choosing. At a let us choose dz along
K, so that dz = eiϕ dl. The corresponding change in S ′ is then determined
solely by the shape of K, for the values of S ′ on K are given by S ′ = e−2iϕ .]
(ii) Deduce that the centre of curvature of K at a is {a + 2[S ′ (a)/S ′′ (a)]}.
(iii) Show that the rate of change of the curvature of K is given by the
“Schwarzian derivative” [Ex. 19] of the Schwarz function:
i
κ̇ = {S(z), z}.
2S ′
29 Check the result of Ex. 28 (i) by applying it to the results of Ex. 27.
30 Let a be a point on a curve K having Schwarz function S(z). By the still
unproven result on the infinite differentiability of analytic functions, S(z) may
be expanded into a Taylor series in the vicinity of a:
1 ′′ 1
S(z) = S(a) + S ′ (a)(z − a) + S (a)(z − a)2 + S ′′′ (a)(z − a)3 + · · · .
2! 3!
(i) Show that the Schwarz function of the tangent line to K at a is given by
the first two terms of the series above. This reconfirms something we saw
in [5.30]: very close to a, reflection in the tangent is a good approximation
to Schwarzian reflection.
(ii) It is natural to suspect that a better approximation to ℜK (z) would be
inversion in the circle of curvature (call it C) of K at a. Let’s verify this.
Use Ex. 28(ii) and Ex. 27(ii) to find SC , and show that it may be written as
 −1
2S ′ 2(S ′ )2 S ′′
SC (z) = a + ′′ + 1 − (z − a) ,
S S ′′ 2S ′
where it is understood that the derivatives are all evaluated at a. Show
that the first three terms in the binomial expansion of SC agree with
those of S, but that they generally differ thereafter. [Hint: You will need
the fact that (S ′ /S ′′ ) = − (S ′ )2 /S ′′ on K. Prove this.]
(iii) If the curvature κ of K were constant then K would be identical to its
circle of curvature. The fact that S and SC disagree beyond the third term
thus reflects the fact that κ does change. One is thus led to guess that the
faster κ changes, the greater the discrepancy between ℜK and inversion
in C. Continuing from the last part, use Ex. 28(iii) to verify this hunch in
the following precise form:

SC (z) − S(z) ≈ (i/3)[S ′ ]2 κ̇ (z − a)3 .

 Exercises 301

31 Let C and D be intersecting circles. Let us say that “D is symmetric in C” if

reflection (inversion) in C maps D into itself. We know this occurs if and only
if D is orthogonal to C, so

D is symmetric in C ⇐⇒ C is symmetric in D.

Briefly, we may simply say that “C and D are symmetric”. Let’s see what
happens if we generalize C and D to intersecting arcs possessing Schwarz
functions, and generalize inversion to Schwarzian reflection.
(i) Explain why the statement “D is symmetric in C” is the same as “if
the point d lies on D then ℜD [ℜC (d)] = ℜC (d)”. Must the arcs be
orthogonal?
(ii) If D is symmetric in C, deduce that the mappings (ℜD ◦ ℜC ) and
(ℜC ◦ ℜD ) are equal at points of D.
(iii) Using the fact that these two mappings are analytic (why?), deduce that
C must also be symmetric in D. Thus, as with circles, we may simply say
that C and D are symmetric.
CHAPTER 6

Non-Euclidean Geometry*

6.1 Introduction

6.1.1 The Parallel Axiom

We have previously alluded to the remarkable discovery (made in the 19th century)
that there exist geometries other than Euclid’s. In this optional chapter we begin to
explore the beautiful connections that exist between these so-called non-Euclidean
geometries and the complex numbers. Since this Introduction summarizes many
of the key ideas and results, you may wish to read it even if you cannot afford the
time to read the entire chapter.
One way to approach Euclidean geometry is to begin with definitions of such
things as “points” and “lines”, together with a few assumptions (axioms) concern-
ing their properties. From there one goes on, using nothing but logic, to deduce
further properties of these objects that are necessary consequences of the initial
axioms. This is the path followed in Euclid’s famous book, The Elements, which
was published around 300 BCE.
Of course Euclidean geometry did not suddenly spring into existence as a fully
formed logical system of axioms and theorems. It was instead developed gradu-
ally as an idealized description of physical measurements performed on physically
constructed lines, triangles, circles, etc. Though the ancients did not think of it in
this way, Euclidean geometry is thus not simply mathematics, it is a physical theory
of space—a fantastically accurate theory of space.
Euclidean geometry is not, however, a perfect theory: modern experiments
have revealed extremely small discrepancies between the predictions of Euclidean
geometry and the measured geometric properties of figures constructed in physical
space. These departures from Euclidean geometry are now known to be gov-
erned, in a precise mathematical way, by the distribution of matter and energy
in space. This is the essence of a revolutionary theory of gravity (General Relativity)
discovered by Einstein in 1915.

Visual Complex Analysis. 25th Anniversary Edition. Tristan Needham, Oxford University Press.
© Tristan Needham (2023). DOI: 10.1093/oso/9780192868916.003.0006
304 Non-Euclidean Geometry*

It turns out that the larger the figures examined, the larger the deviations from
the predictions of Euclidean geometry. However, it’s important to realize just how
small these deviations typically are for figures of reasonable size. For example, sup-
pose we measure the circumference of a circle having a radius of one meter. Even
if our measuring device were capable of detecting a discrepancy the size of a single
atom of matter, no deviation from Euclidean geometry would be found! Little won-
der, then, that for two thousand years mathematicians were seduced into believing
that Euclidean geometry was the only logically possible geometry.
It is a marvellous tribute to the power of human mathematical thought that non-
Euclidean geometry was discovered a full century before Einstein found that it was
needed to describe gravity. To locate the seeds of this mathematical discovery, let
us return to ancient Greece.
Euclid began with just five axioms, the first four of which never aroused con-
troversy. The first axiom, for example, merely states that there exists a unique line
passing through any two given points. However, the status of the fifth axiom (the
so-called parallel axiom) was less clear, and it became the subject of investigations
that ultimately led to the discovery of non-Euclidean geometry:

Parallel Axiom. Through any point p not on the line L there exists
(6.1)
precisely one line L ′ that does not meet L.
Figure [6.1a] illustrates the parallel axiom, and it also explains why this axiom can-
not be experimentally tested, at least as stated. As the line M rotates towards L ′ ,
the intersection point q moves further and further away along L. Our geometric
intuition is based on figures drawn in a finite portion of the plane, but to verify
that L ′ never meets L, we need an infinite plane. We can certainly try to imagine
what an infinite plane would be like, but we have no first hand experience to back
up our hunches.
These are very modern doubts we are expressing. Historically, mathematicians
fervently believed in (6.1), so much so that they thought it must be a logically nec-
essary property of straight lines. But in that case they ought to be able to prove it
outright, instead of merely assuming it as Euclid had done.
Many attempts were made to deduce (6.1) from the first four axioms, one of the
most penetrating being that of Girolamo Saccheri in 1733.1 His idea was to show
that if (6.1) were not true, then a contradiction would necessarily arise. He divided
the denial of (6.1) into two alternatives:

Spherical Axiom. There is no line through p that does not meet L. (6.2)
or

Hyperbolic Axiom. There are at least two lines through p that do not meet L. (6.3)

1
See Stillwell (2010).
 Introduction 305

[6.1] Two forms of the Parallel Axiom. [a] The Parallel Axiom seems beyond empirical
verification: As the line M rotates towards L ′ , the intersection point q moves further and
further away along L, so to be certain that the supposedly unique parallel line L ′ never
intersects L, we need to go all the way to infinity! [b] However, an alternative formulation
of the Parallel Axiom is subject to empirical verification or refutation: The angles in every
triangle must sum to π, so that E = 0.

Our naming of (6.2) will become clear shortly, but the use of “hyperbolic” in
connection with (6.3) is more obscure, though standard.
In the case of (6.2), Saccheri was indeed able to obtain a contradiction, provided
“lines” are assumed to have infinite length. If we drop this requirement, then we
obtain a non-Euclidean geometry called spherical geometry. This is the subject of the
following section.
In the case of (6.3), Saccheri and later mathematicians were able to derive
very strange conclusions, but they were not able to find a contradiction. As we
now know, this is because (6.3) yields another viable non-Euclidean geometry,
called hyperbolic geometry. Of the two non-Euclidean geometries obtained from (6.2)
and (6.3), hyperbolic geometry is by far the more intriguing and important: it is
an essential tool in many areas of contemporary research. Furthermore, there is
even a sense (to be discussed later) in which hyperbolic geometry subsumes both
Euclidean and spherical geometry.

6.1.2 Some Facts from Non-Euclidean Geometry

Let’s take our first look at how these new geometries differ from Euclid’s. A very
familiar theorem of Euclidean geometry states that in any triangle T ,

(Angle sum of T ) = π.

As indicated in [6.1b], this result is actually equivalent to the parallel axiom. It

follows that in non-Euclidean geometry the angle sum of a triangle differs from π.
To measure this difference, we introduce the so-called angular excess E:

E(T ) ≡ (Angle sum of T ) − π.

306 Non-Euclidean Geometry*

Euclidean geometry is thus characterized by the vanishing of E(T ).

Note that, unlike the original formulation of the parallel axiom, this statement
can be checked against experiment: construct a triangle, measure its angles, and
see if they add up to π. Gauss was the first person to ever conceive of the possibil-
ity that physical space might not be Euclidean, and he even attempted the above
experiment, using three mountain tops as the vertices of his triangle, and using
light rays for its edges. Within the accuracy permitted by his equipment, he found
E = 0. Quite correctly, Gauss did not conclude that physical space is definitely
Euclidean in structure, but rather that if it is not Euclidean then its deviation from
Euclidean geometry is extremely small.
Let us return from physics to mathematics. Using pure logic to work out the
consequences of (6.2) and (6.3), both Gauss and Johann Heinrich Lambert indepen-
dently discovered that the two non-Euclidean geometries departed from Euclid’s
in opposite ways:
• In spherical geometry the angle sum is greater than π: E > 0.
• In hyperbolic geometry the angle sum is less than π: E < 0.
Furthermore, they discovered the striking fact that E(T ) is completely determined
by the size of the triangle. More precisely, E(T ) is simply proportional to the area
A(T ) of the triangle T :
E(T ) = K A(T ), (6.4)
where K is a constant that is positive in spherical geometry, and negative in
hyperbolic geometry.
Several interesting points can be made in connection with this result:
• Although there are no qualitative differences between them, there are never-
theless infinitely many different spherical geometries, depending on the value of
the positive constant K. Likewise, each negative value of K yields a different
hyperbolic geometry.
• Since the angle sum of a triangle cannot be negative, E ⩾ −π. Thus in hyperbolic
geometry (K < 0) no triangle can have an area greater than |(π/K)|.
• In non-Euclidean geometry, similar triangles do not exist! This is because (6.4)
tells us that two triangles of different size cannot have the same angles.
• Closely related to the previous point, in non-Euclidean geometry there exists
an absolute unit of length. For example, in spherical geometry we could define it
to be the side of the equilateral triangle having angle sum 1.01π. Similarly, in
hyperbolic geometry we could define it to be the side of the equilateral triangle
having angle sum 0.99π.
 Introduction 307

• A somewhat more natural way of defining the absolute unit of length is in terms
of the constant K. Since the radian measure of angle is defined as a ratio of
lengths, E is a pure number. On the other hand, the area A has units of (length)2 .
It follows that K has units of 1/(length)2 and so it can be written as follows in
terms of a length R: K = +(1/R2 ) in spherical geometry; K = −(1/R2 ) in hyper-
bolic geometry. Later we will see that this length R can be given a very intuitive
interpretation P.
• The smaller the triangle, the harder it is to distinguish it from a Euclidean tri-
angle: only when the linear dimensions are a significant fraction of R will the
difference become obvious. This is why Gauss chose the biggest triangle he
could in his experiment. Einstein’s theory explains why Gauss’s triangle was
nevertheless much too small: the weak gravitational field in the space surround-
ing the earth corresponds to a microscopic value of K and hence to an enormous
value of R. It would have been a different story if Gauss had been able to perform
his experiment in the vicinity of a small black hole!

6.1.3 Geometry on a Curved Surface

We began this book by discussing how the complex numbers met with enormous
initial resistance, and how they were finally accepted only after they were given a
concrete interpretation, via the complex plane. The story of non-Euclidean geometry
is remarkably similar.
Gauss never published his revolutionary ideas on non-Euclidean geometry, and
the two men who are usually credited for their independent discovery of hyper-
bolic geometry are János Bolyai (1832) and Nikolai Lobachevsky (1829). Indeed,
hyperbolic geometry is frequently also called Lobachevskian geometry, perhaps
because Lobachevsky’s investigations went somewhat deeper than Bolyai’s. How-
ever, in the decades that followed their discoveries, Bolyai’s work was completely
ignored, and Lobachevsky’s met only with vicious attacks.
The decisive figure in the acceptance of non-Euclidean geometry was Eugenio
Beltrami. In 1868 he discovered that hyperbolic geometry could be given a concrete
interpretation, via “differential geometry”. For our purposes, differential geometry
is the study of curved surfaces by means of ideas from calculus. What Beltrami dis-
covered was that there exists a surface (the so-called pseudosphere shown in [6.2])
such that figures drawn on it automatically obey the rules of hyperbolic geome-
try2 . Psychologically, Beltrami’s pseudosphere was to hyperbolic geometry as the
complex plane had been to the theory of complex numbers.

2
This oversimplification does not do justice to Beltrami’s accomplishments. Later in this chapter
we shall see what Beltrami really did!
308 Non-Euclidean Geometry*

[6.2] The Pseudosphere. In 1868 Eugenio Beltrami finally removed the mystery and scep-
ticism that had overshadowed the strange, abstract laws of hyperbolic geometry ever
since their discovery, 40 years earlier. He did so by recognizing that the hyperbolic laws
simply are the intrinsic laws of geometry that hold true within a surface called the pseudo-
sphere (shown here) that has constant negative Gaussian curvature. The pseudosphere
extends upward indefinitely, but the base circle is a boundary beyond which it cannot be
extended.

To explain what we mean by this, let us first discuss how we may “do geometry”
on a more general surface, such as the surface of the strange looking vegetable3
shown in [6.3]. The idea of doing geometry on such a surface is essentially due to
Gauss and (in greater generality) to Riemann.
The first thing we must do is to replace the concept of a straight line with that
of a geodesic. Just as a line-segment in a flat plane may be defined as the short-
est route between two points, so a geodesic segment connecting two points on
a curved surface may be defined (provisionally) as the shortest connecting route
within the surface. For example, if you were an ant living on the surface in [6.3],
and you wanted to travel from a to b as quickly as possible, then you would fol-
low the illustrated geodesic segment. The figure also shows the geodesic segment
connecting another pair of points, c and d.
Here is a simple way you can actually construct such geodesic segments: take a
thread and stretch it tightly over the surface to connect the points a and b. Provided

3
European readers may think this an imaginary vegetable, but Americans can buy it in the
supermarket.
 Introduction 309

that the thread can slide around on the surface easily, the tension in the thread
ensures that the resulting path is as short as possible. Note that in the case of cd,
we must imagine that the thread runs over the inside of the surface. In order to deal
with all possible pairs of points in a uniform way, it is therefore best to imagine the
surface as made up of two thinly separated layers, with the thread trapped between
them.
In fact there exists a wonderfully simple and practical (yet surprisingly little
known) method of constructing geodesics on any part of a surface. Cut a long,
straight, narrow strip of masking tape (aka painter’s tape) and start to lay one end
down on the surface in the direction you want your geodesic to go. Holding the
free end up away from the surface with one hand, use your other hand to gently
run your finger along the tape, allowing it to choose its own path as it sticks itself
to the surface as it goes. Behold! You have created a geodesic, and this construction
works equally well on ab and cd! In the case of ab, you can easily check the validity
of the construction by stretching a string over the surface: it will coincide with path
of your sticky tape! For the explanation of why this works (and many applications)
see Needham (2021).
It is now obvious how we should define distance in this geometry: the distance
between a and b is the length of the geodesic segment connecting them. Figure
[6.3] shows how we can then define, for example, a circle of radius r and centre p
as the locus of points at distance r from p. To construct this circle we may take a
piece of thread of length r, hold one end fixed at p, then (keeping the thread taut)
drag the other end round on the surface.
Given three points on the surface, we may join them with geodesics to form a
triangle; [6.3] shows two such triangles, ∆1 and ∆2 . Now look at the angles in ∆1 .
Clearly E(∆1 ) > 0, like a triangle in spherical geometry, while E(∆2 ) < 0, like a
triangle in hyperbolic geometry.

6.1.4 Intrinsic versus Extrinsic Geometry

Clearly it is the curvature of the surface that causes E(∆1 ) and E(∆2 ) to differ from
their Euclidean value E = 0. However, it cannot be the precise shape of the surface
in space that is involved here. To see this, imagine that from the vegetable in [6.3]
we were to cut out a patch of the skin containing ∆1 . Suppose that this patch is
made of fairly stiff material that does not stretch if we try to bend it a little. [As it
happens, the skin of this vegetable is actually like this!] We can now gently bend the
patch into infinitely many slightly different shapes: its so-called extrinsic geometry
has been changed by our stretch-free bending. For example, the curves in space
making up the edges of ∆1 are no longer the same shape as before.
310 Non-Euclidean Geometry*

[6.3] The intrinsic geometry of a surface is the geometry experienced by small,

2-dimensional creatures living within the surface. For them, a “straight line” is the shortest
route between two points, such as ab or cd. A “circle” of radius r is the set of points that
lie at a fixed distance r from its centre. “Triangles” are no longer subject to the Parallel
Axiom, so now their angles need not add up to π: E(∆1 ) > 0, and E(∆2 ) < 0.

On the other hand, if you were an intelligent ant living on this patch, no geomet-
ric experiment you could perform within the surface would reveal that any change
had taken place whatsoever. We say that the intrinsic geometry has not changed.
For example, the curves into which the edges of ∆1 have been deformed are still
the shortest routes on the surface. Correspondingly, the value of E is unaffected by
stretch-free bending: E is governed by intrinsic (not extrinsic) curvature.
To highlight this fact, consider [6.4]. On the left is a flat piece of paper on which
we have drawn a triangle T with angles (π/2), (π/6), (π/3). Of course E(T ) = 0.
Clearly we can bend such a flat piece of paper into either of the two (extrinsically)
curved surfaces on the right4 . However, intrinsically these surfaces have undergone
no change at all—they are both as flat as a pancake! The illustrated triangles on
these surfaces (into which T is carried by our stretch-free bending of the paper) are

4
Of course the conical example on the far right cannot be obtained by bending a rectangle.
 Introduction 311

[6.4] A bent piece of paper remains intrinsically flat. A flat piece of paper can be bent
into all kinds of extrinsically curved shapes in space, but its internal, intrinsic geometry
remains totally flat, and triangles constructed within the surface continue to obey all the
normal laws of Euclidean geometry, such as Pythagoras’s Theorem, and E = 0.

identical to the ones that intelligent ants would construct using geodesics, and in
both cases E = 0: geometry on these surfaces is Euclidean.

6.1.5 Gaussian Curvature

In 1827 Gauss published a beautiful and revolutionary analysis5 of the intrinsic and
extrinsic geometry of surfaces, in which he revealed that remarkable connections
exist between the two. Here we will simply state some of his most important con-
clusions, in their most general form. For explanations of these general results we
refer you to works on differential geometry; see the recommendations at the end
of this chapter. However, only special cases of the general results are needed to
understand non-Euclidean geometry, and these will be separately verified in the
course of this chapter.
For a surface such as [6.3], it is clear that some parts are more curved than others.
Furthermore, the kind of bending also varies from place to place. To quantify the
amount (and type) of bending of the surface at a point p, Gauss introduced a quan-
tity K(p). This function K(p), whose precise definition will be given in a moment, is
called the Gaussian curvature6 . The greater the magnitude of K(p), the more curved
the surface is at p. The sign of K(p) tells us qualitatively what the surface is like in
the immediate neighbourhood of p. See [6.5]. If K(p) < 0 then the neighbourhood
of p resembles a saddle: it bends upwards in some directions, and downwards in
others. If K(p) > 0 then it bends the same way in all directions, like a piece of a
sphere.

5
Gauss (1827).
6
Other names are intrinsic curvature, total curvature, or just plain curvature.
312 Non-Euclidean Geometry*

[6.5] The sign of the Gaussian curvature K(p) tells us the shape of the surface at p.

As we will now start to explain, it is no accident that we have used the same sym-
bol to represent Gaussian curvature as we earlier used for the constant occurring
in (6.4)—they are the same thing!
Gauss originally defined K(p) as follows. Let Π be a plane containing the nor-
mal vector n to the surface at p, and let κ be the (signed) curvature at p of the
curve in which Π intersects the surface. The sign of κ depends on whether the cen-
tre of curvature is in the direction n or −n. The so-called principal curvatures are
the minimum κmin and the maximum κmax values of κ as Π rotates about n. [Inci-
dentally, Euler had previously made the important discovery that these principal
curvatures occur in two perpendicular directions.] Gauss defined K as the product
of the principal curvatures:
K ≡ κmin κmax .
Note that this definition is in terms of the precise shape of the surface in space
(extrinsic geometry). However, Gauss (1827) went on to make the astonishing dis-
covery that K(p) actually measures the intrinsic curvature of the surface, that is, K is
invariant under bending! Gauss was justifiably proud of this result, calling it the The-
orema Egregium (Latin for “remarkable theorem”). As an example of the result, you
may visually convince yourself that K = 0 everywhere on each of the intrinsically
flat surfaces in [6.4].
The intrinsic significance of K is exhibited in the following fundamental result:
If ∆ is an infinitesimal triangle of area dA located at the point p, then
E(∆) = K(p) dA. (6.5)
Since E and dA are defined by the intrinsic geometry, so is K = (E/dA). Once again,
we refer you to works on differential geometry for a proof of (6.5).
It follows from (6.5) [see Ex. 1] that the angular excess of a non-infinitesimal
triangle T is obtained by adding up (i.e., integrating) the Gaussian curvature over
the interior of T :
 Introduction 313

ZZ
E(T ) = K(p) dA. (6.6)
T

As Beltrami recognized, and as we now explain, this lovely result of differential

geometry brings us very close to a concrete interpretation of the non-Euclidean
geometries.

6.1.6 Surfaces of Constant Curvature

Consider a surface such that K(p) has the same value K at every point p; we
call this a surface of constant curvature. For example, a plane is a surface of con-
stant curvature K = 0, as are the other surfaces in [6.4]; a sphere is an example
(not the only one) of a surface of constant positive curvature; and the pseudo-
sphere in [6.2] is an example (not the only one) of a surface of constant negative
curvature.
In the case of a surface of constant curvature (and only in this case) we find that
(6.6) takes the form,
ZZ
E(T ) = K dA = K A(T ).
T

But this is identical to the fundamental formula (6.4) of non-Euclidean geometry!

Thus, as Beltrami realized,

Euclidean, spherical, and hyperbolic geometry can all be interpreted

concretely as the intrinsic geometry of surfaces of constant vanishing,
positive, or negative curvature.

Figure [6.6] illustrates this using the simplest surfaces of each type. To obtain an
added bonus, recall that we previously associated an absolute unit of length R with
a non-Euclidean geometry by writing K = ±(1/R2 ). The bonus is that this length
R now takes on vivid meaning: in spherical geometry R is simply the radius of
the sphere, while in hyperbolic geometry it is the radius of the circular base of the
pseudosphere (called the radius of the pseudosphere). These two interpretations
will be justified later.
The requirement of constant curvature can be understood more intuitively by
reconsidering the discussion at the end of Chapter 1. There we saw that a central
idea in Euclidean geometry is that of a group of motions of the plane: one-to-one
mappings that preserve the distance between all pairs of points. For example, two
figures are congruent if and only if there exists a motion that carries the first into
coincidence with the second. In order that this basic concept of equality be available
in non-Euclidean geometry, we require that our surface admits an analogous group
of motions. If we take one of the triangles on the surface in [6.3], it’s clear that we
cannot slide it to a new location and still have it fit the surface snugly, because the
way in which the surface is curved at the new location is different: variation in the
curvature is the obstruction to motion.
314 Non-Euclidean Geometry*

[6.6] Euclidean, Spherical, and Hyperbolic geometry are the intrinsic geometries of
surfaces of constant Gaussian curvature.

This intuitive explanation can be clarified by appealing to (6.5). First, though,

we wish to eliminate a possible confusion. The triangle on the flat plane in [6.6]
can clearly be slid about and rotated freely, but what about the triangles on the
(extrinsically) curved surfaces in [6.4]? After all, these surfaces are intrinsically flat,
and so Beltrami would have us believe they are therefore just as good as the plane
for doing Euclidean geometry. If we imagine these triangles as completely rigid then
it’s clear that if we try to move them to another location on the surface, they will
no longer fit snugly against the surface. But if the triangle is instead cut out of a
piece of ordinary (bendable but unstretchable) paper, then it can be slid about and
rotated freely, always fitting perfectly against the surface. This is the kind of motion
we are concerned with.
In order to clarify the connection between constant curvature and the existence
of motions, consider an infinitesimal (bendable but unstretchable) triangle located
at p. If its angular excess is E and its area is dA, then (6.5) tells us that the Gaussian
curvature of the surface at p is given by K(p) = (E/dA). Now suppose that there
exists a motion that carries this triangle to an arbitrary point q on the surface. We
may have to bend the triangle to make it fit against the surface at q, but since we
are not allowed to stretch it, the values of E and dA do not alter. Thus K(q) =
(E/dA) = K(p), and the surface has constant curvature.
Finally, let us return to the specific models of spherical and hyperbolic geom-
etry shown in [6.6]. Clearly the triangle on the sphere can be slid about and
rotated freely. In fact here, as on the plane, no bending is needed at all, because
the sphere not only has constant intrinsic curvature, it also has constant extrinsic
curvature.
What about hyperbolic geometry on the pseudosphere? It is certainly much
less obvious, but the fact [to be proved later] that the pseudosphere has constant
curvature guarantees that a bendable but unstretchable triangle can be slid about
 Introduction 315

and rotated freely, always fitting perfectly snugly against the surface. Exercise 15
shows how you can build your own pseudosphere; once built, you can verify this
surprising claim experimentally.

6.1.7 The Connection with Möbius Transformations

As we established in Chapter 1, if the Euclidean plane is identified with C then
its motions (and similarities) are represented by the particularly simple Möbius
transformations of the form M(z) = az + b. One of the principal miracles we wish
to explain in this chapter is that the motions of spherical and hyperbolic geometry
are also Möbius transformations!
The most general (direct) motion of the sphere is a rotation about its centre.
Stereographic projection onto C yields a conformal map of the sphere, and the
rotations of the sphere thus become complex functions acting on this map. As
we showed algebraically in Ex. 21 on page 211 of Chapter 3, they are the Möbius
transformations of the form
az + b
M(z) = .
−bz + a
This was first discovered by Gauss, around 1819. In the next section we will red-
erive this result in a more illuminating way, and we will also explore the connection
with Hamilton’s “quaternions”, which we first met on page 49.
Following the same pattern, it is also possible to construct conformal maps (in C)
of the pseudosphere, thereby transforming its motions into complex functions. One
of the most convenient of these conformal maps is constructed in the unit disc. The
motions of hyperbolic geometry then turn out to be the Möbius automorphisms of
this circular map, which we first derived in (3.52) and then further investigated in
Ex. 20 on page 210:
az + b
M(z) = .
bz + a
This beautiful discovery was made by Henri Poincaré in 1882; see Poincaré (1985).
It seems magical enough that the motions of all three of the two-dimensional
geometries are represented by special kinds of Möbius transformations, but there’s
more! In Chapter 3 [see (3.2) on page 138] we saw that the general Möbius
transformation
az + b
M(z) =
cz + d
has deep significance for physics: it corresponds to the most general Lorentz trans-
formation of spacetime. Might it also have significance in non-Euclidean geometry?
As we will explain at the end of this chapter, in 1883 Poincaré made the startling
discovery that it represents the most general (direct) motion of three-dimensional
hyperbolic space!
316 Non-Euclidean Geometry*

6.2 Spherical Geometry

6.2.1 The Angular Excess of a Spherical Triangle

The geodesics on the sphere are the great circles, that is the intersections of the
sphere with planes through its centre. Thus if you were an ant living on the sphere,
these great circles are what you would call “lines”.
Figure [6.7a] illustrates a general triangle ∆ on a sphere of radius R obtained
by joining three points using such “lines”. Without appealing to (6.6), which is a
deep result in differential geometry, let us show directly that the angular excess
E(∆) obeys the law (6.4), and that the constant K is indeed the Gaussian curva-
ture, K = (1/R2 ). The elegant argument that follows is usually attributed to Euler,
who rediscovered it, but it was in fact first discovered by Thomas Harriot in 1603.7
Yes, this is the same Thomas Harriot who discovered the fundamentally important
conformality of stereographic projection, (3.17)!
Prolonging the sides of ∆ divides the surface of the sphere into eight trian-
gles, the four triangles labelled ∆, ∆α , ∆β , ∆γ each being paired with a congruent
antipodal triangle. This is clearer in [6.7b]. Since the area of the sphere is 4πR2 , we
deduce that

[6.7] Harriot’s 1603 proof that (∆) = (1/R2 ) (∆). Extending the sides of ∆ that
contain the angle α, we obtain the shaded wedge on the right, which occupies (α/2π)
of the complete sphere, so A(∆) + A(∆α ) = 2αR2 ; likewise for β and γ. Since each of
the four triangles on the left has an antipodal partner of equal area—as can be seen more
clearly on the right—it follows that together they occupy half the sphere: A(∆)+A(∆α )+
A(∆β ) + A(∆γ ) = 2πR2 , and the result follows immediately.

7
See Stillwell (2010, §17.6).
 Spherical Geometry 317

A(∆) + A(∆α ) + A(∆β ) + A(∆γ ) = 2πR2 . (6.7)

On the other hand, it is clear in [6.7b] that ∆ and ∆α together form a wedge whose
area is (α/2π) times the area of the sphere:

A(∆) + A(∆α ) = 2αR2 .

Similarly,

A(∆) + A(∆β ) = 2βR2 ,

A(∆) + A(∆γ ) = 2γR2 .

Adding these last three equations, we find that

3A(∆) + A(∆α ) + A(∆β ) + A(∆γ ) = 2(α + β + γ)R2 . (6.8)

Finally, subtracting (6.7) from (6.8), we get

A(∆) = (α + β + γ − π)R2 .

In other words,

E(∆) = K A(∆), where K = (1/R2 ), (6.9)

as was to be shown.

6.2.2 Motions of the Sphere: Spatial Rotations and Reflections

In order to understand the motions (i.e., one-to-one, distance-preserving map-
pings) of the sphere, we must first clarify the idea of “distance”. If two points a and
b are not antipodal then there exists a unique line (great circle) L passing through
them, and a and b divide L into two arcs of unequal length. The “distance” between
the points can now be defined as the length of the shorter arc. But if the points are
antipodal then every line through a automatically passes through b, and the dis-
tance between the points is defined to be the length πR of any of the semicircular
arcs connecting them.
We can now generalize the Euclidean arguments given in the final section of
Chapter 1. There we saw that a motion of the plane is uniquely determined by the
images a ′ , b ′ , c ′ of any three points a, b, c not on a line: the image of P is the unique
point P ′ whose distances from a ′ , b ′ , c ′ equal the distances of P from a, b, c. We
leave it to you to check that this result (and the reason for it) is still true on the
sphere.
On the sphere, as on the plane, we may consistently attribute a sense to angles—
by convention an angle is positive if it is counterclockwise when viewed from
outside the sphere. As happened in the plane, this leads to a division of spher-
ical motions into two types: direct (i.e., conformal) motions, and opposite (i.e.,
anticonformal) motions.
318 Non-Euclidean Geometry*

[6.8] Reflection of the sphere in a line. [a] If Π is the plane through the line (great circle)
L, then reflection RΠ of space across Π induces the anticonformal reflection RL of the
sphere across L. [b] This reflection RL (a) can instead be expressed in intrinsic terms that
make sense to the inhabitants of the sphere: travel some distance d along the orthogonal
“line” M from a until you arrive at L, then travel an equal distance d again.

As in the plane, the simplest opposite motion of the sphere is reflection RL in

a line L. This may be thought of as the transformation induced on the sphere by
reflection RΠ of space in the plane Π containing L. See [6.8a], which illustrates how
the positive angle θ in the illustrated spherical triangle is reversed by RL .
If you were an intelligent ant living on the sphere, the above construction of RL
as the restriction of RΠ to the sphere would be meaningless to you. However, it
is not hard to re-express RL in intrinsically spherical terms. See [6.8b]. To reflect
a in L, first draw the unique line M through a that cuts L at right angles8 . If d is
the distance we must crawl along M from a to reach L, then RL (a) is the point we
reach after crawling a further distance of d. Of course M actually intersects L in
two antipodal points, but we will arrive at the same RL (a) irrespective of which of
these two points is used in the construction.
We now turn to direct motions. The obvious example of a direct motion is a
rotation of the sphere about an axis V passing through its centre. Less obvious is
the fact (to be proved shortly) that these rotations are the only direct motions. To
avoid ambiguity in the description of such rotations, we introduce the following
standard convention. First note that specifying the axis V is equivalent to specifying
either of its antipodal intersection points (say p and q) with the sphere. Now pick
one of these, say p. Suppose that the effect of the rotation on a small line-segment
issuing from p is a positive rotation of θ—recall that this means counterclockwise as
seen from outside the sphere. In this case the motion can be unambiguously described
as a “positive rotation of θ about p”; see [6.9b]. We will write this rotation as Rθp .
Check for yourself that Rθp = R−θ q .

8
If L is thought of as the equator, then when a is one of the poles there are infinitely many M’s—
pick any one you like.
 Spherical Geometry 319

[6.9] Rotation of the sphere via two reflections. [a] If the planes Π1 and Π2 meet at
angle (θ/2) then (RΠ2 ◦ RΠ1 ) is a rotation of space through angle θ about their line of
intersection. [b] The induced transformation of the sphere is therefore a rotation Rθp by θ
about the intersection point p of the lines on the sphere: RL2 ◦ RL1 = Rθp .

In Chapter 1 we saw that every direct motion of the plane was the composition
of two reflections: a rotation if the lines intersected; a translation if the lines were
parallel. We will now see that a similar phenomenon occurs on the sphere, but
because every pair of lines intersect, the composition of two reflections is always a
rotation—the sphere has no motions analogous to translations.
Figure [6.9a] illustrates the composition (RΠ2 ◦ RΠ1 ) of two reflections of space.
Here the planes Π1 and Π2 intersect in a line with direction vector v, and the angle
from Π1 to Π2 is (θ/2). Restricting attention to any one of the shaded planes orthog-
onal to v, we see that the transformation induced by (RΠ2 ◦ RΠ1 ) is (Rl2 ◦ Rl1 ),
where l1 and l2 are the lines in which Π1 and Π2 intersect the plane. Since (Rl2 ◦ Rl1 )
is a rotation of the plane through θ about the intersection point of l1 and l2 , it
is now clear that (RΠ2 ◦ RΠ1 ) is a rotation of space through angle θ about the
axis v.
Figure [6.9b] translates this idea into spherical terms. If Π1 and Π2 pass through
the centre of the sphere, and the lines (great circles) in which they intersect the
sphere are L1 and L2 , then

RL2 ◦ RL1 = Rθp .

In other words,

A rotation Rθp of the sphere about a point p through angle θ may be

expressed as the composition of reflections in any two spherical lines (6.10)
that pass through p and contain the angle (θ /2).
320 Non-Euclidean Geometry*

Note that there is precisely one line P that is mapped into itself by Rθp . If we orient
P in agreement with the rotation (as illustrated) then we obtain a one-to-one corre-
spondence between oriented lines and points: P is called the polar line of p, and p
is called the pole of P.
In the case of the plane we used the analogue of (6.10) to show that the composi-
tion of two rotations about different points was equivalent (in general) to a single
rotation about a third point; exceptionally, however, two rotations could result in
a translation. As you might guess, in the case of the sphere there are no exceptions:

The composition of any two rotations of the sphere is equivalent to a

(6.11)
single rotation. Thus the set of all rotations of the sphere forms a group.

Figure [6.10a] shows how this may be established using exactly the same argument
that was used in the plane. In order to find the net effect of (Rϕ
q ◦ Rp ), draw the
θ

lines L, M, N in the illustrated manner. Then

Rϕ
q ◦ Rp = (RN ◦ RM ) ◦ (RM ◦ RL ) = RN ◦ RL = Rr .
θ ψ

This beautiful geometric method of composing spatial rotations was discovered by

Olinde Rodrigues in 1840.
Note that in the plane the total amount of rotation produced by rotations of θ
and ϕ is simply the sum (θ + ϕ), but on the sphere we have a more complicated
rule. If A is the area of the white spherical triangle, and K = (1/R2 ) is the Gaussian
curvature of the sphere, then the formula for the angular excess implies that

[6.10] [a] Composition of rotations: Rϕq ◦ Rp = (RN ◦ RM ) ◦ (RM ◦ RL ) = RN ◦ RL =

θ
ψ e that
Rr . [b] There is exactly one direct motion M (and exactly one opposite motion M)
′ ′
maps a given line-segment ab to another line-segment a b of equal length. Furthermore,
Me = (RL ◦ M), where L is the line through a ′ and b ′ .
 Spherical Geometry 321

ψ = θ + ϕ − 2KA.

We may now complete the classification of the motions of the sphere. As we have
remarked, there is precisely one motion of the sphere that carries a given spherical
triangle abc to a given congruent image triangle. Figure [6.10b] helps to refine this
result. Using the same logic as was used in the plane, we see [exercise] that

There is exactly one direct motion M (and exactly one opposite motion
e that maps a given line-segment ab to another line-segment a ′ b ′ of
M)
e = (RL ◦ M ), where L is the line through (6.12)
equal length. Furthermore, M
′ ′
a and b .

Figure [6.10b] also shows how we may construct M. Draw the line P through a
and a ′ , and let p be its pole. With the appropriate value of θ, it’s clear that Rθp
will carry the segment ab along P to a segment of equal length emanating from
a ′ ; finally, an appropriate rotation Rϕ ′ ′ ′
a ′ about a will carry this segment into a b .
ϕ
Thus M = (Ra ′ ◦ Rp ), which is equivalent to a single rotation by virtue of (6.11).
θ

Combining this fact with (6.12), we deduce that

Every direct motion of the sphere is a rotation, and every opposite motion is
(6.13)
the composition of a rotation and a reflection.

As a simple test of this result (and your grasp of it) consider the antipodal mapping
that sends every point on the sphere to its antipodal point. Clearly this is a motion,
but how does it accord with the above result?

6.2.3 A Conformal Map of the Sphere

The sphere merely provides one particularly simple model of what we have called
spherical geometry. As Minding (1839) discovered, any surface of constant9 Gaus-
sian curvature K = (1/R2 ) has exactly the same intrinsic geometry as a sphere of
radius R. To see that such surfaces exist, take a Ping-Pong ball and cut it in half: as
you gently flex one of the hemispheres you obtain infinitely many surfaces whose
intrinsic geometry is identical to the original sphere.
Figure [6.11] illustrates that even if we restrict attention just to surfaces of revo-
lution, the sphere is not the only one of constant positive curvature. Though they
hardly look like spheres, an intelligent ant living on either of these surfaces would
never know that he wasn’t living on a sphere. Well, that’s almost true: eventually
he might discover points at which the surface is not smooth, or else he might run
into an edge. In 1899 H. Liebmann proved that if a surface of constant positive
curvature does not suffer from these defects then it can only be a sphere.

9
If the curvature is not constant, two surfaces can have equal curvature at corresponding points and
yet have different intrinsic geometry.
322 Non-Euclidean Geometry*

[6.11] Non-spherical surfaces of constant positive curvature. There exist surfaces of con-
stant positive curvature that are not a portion of the sphere, but Minding proved that their
intrinsic geometry is identical to that of the sphere, at least so long as we stay away from
the pointed tips on the left, or the edges on the right.

The sphere also has the advantage of making it obvious that its intrinsic geom-
etry admits a group of motions: in [6.11] it’s certainly not clear that figures can be
freely moved about and rotated on the surface without stretching them. Neverthe-
less, the above discussion shows that the actual shape of a surface in space is a
distraction, and it would be better to have a more abstract model that captured the
essence of all possible surfaces having the same intrinsic geometry.
By the “essence” we mean knowledge of the distance between any two points,
for this and this alone determines the intrinsic geometry. In fact—and this is a fun-
damental insight of differential geometry—it is sufficient to have a rule for the
infinitesimal distance between neighbouring points. Given this, we may determine
the length of any curve as an infinite sum (i.e., integral) of the infinitesimal seg-
ments into which it may be divided. Consequently, we may also identify the “lines”
of the geometry as shortest routes from one point to another, and we can likewise
[exercise] determine angles.
This leads to the following strategy for capturing the essence of any curved sur-
face S (not necessarily one of constant curvature). To avoid the distraction of the
shape of the surface in space, we draw a map (in the sense of a geographical atlas)
of S on a flat piece of paper. That is we set up a one-to-one correspondence between
points bz on S and points z on the plane, which we will think of as the complex plane.
Now consider the distance d b s separating two neighbouring points b b on
z and q
S. In the map, these points will be represented by z and q = z + dz, separated
by (Euclidean) distance ds = |dz|. Once we have a rule for calculating the actual
separation d b s on S from the apparent separation ds in the map, then (in principle)
we know everything there is to know about the intrinsic geometry of S.
 Spherical Geometry 323

The rule giving d bs in terms of ds is called the metric. In general d b

s depends on
the direction of dz as well as its length ds: writing dz = eiϕ ds,
db
s = Λ(z, ϕ) ds. (6.14)
According to this formula, Λ(z, ϕ) is the amount by which we must expand the
apparent separation ds in the map—located at z, and in the direction ϕ—to obtain
the true separation d b s on the surface S.
We will now carry out the above strategy for the sphere. It follows from (6.9)
that it is impossible [exercise] to draw a map of the sphere that faithfully represents
every aspect of its intrinsic geometry. How we choose to draw our map therefore
depends on which features we wish to faithfully represent. For example, if we want
lines (great circles) on the sphere to be represented by straight lines in the map, then
we may employ the so-called central projection, in which points are projected from
the centre of the sphere onto one of its tangent planes. This yields the so-called
projective map or projective model of the sphere. See [6.12]. Here, the price that we
pay for preserving the concept of lines is that angles are not faithfully represented:
the angle at which two curves meet on the sphere is not (in general) the angle at
which they meet on the map.
For most purposes it is much better to sacrifice straight lines in favour of pre-
serving angles, thereby obtaining a conformal map of the surface. In terms of (6.14),
a map is conformal if and only if the expansion factor Λ does not depend on the
direction ϕ of the infinitesimal vector dz emanating from z:
db
s = Λ(z) ds. (6.15)
[Recall that we established this fact in Chapter 4.] The great advantage of such a
map is that an infinitesimal shape on the surface is represented in the map by a
similar shape that differs from the original only in size: the one on S is just Λ times
bigger.

[6.12] Central projection maps the lines of the sphere to the lines of the map.
324 Non-Euclidean Geometry*

In the case of the sphere we already know of a simple method of constructing a

conformal map, namely, via stereographic projection. For simplicity’s sake, hence-
forth we shall take the sphere to have unit radius so that it may be identified with the
Riemann sphere Σ of Chapter 3. Unlike [6.12], the “lines” of this conformal map do
not appear as straight lines. In fact it’s not too hard to see [exercise] that great circles
on Σ are mapped to circles in C that intersect the unit circle at opposite points.
Formula (6.15) may be paraphrased as saying that a map is conformal if infinites-
imal circles on S are represented in the map by infinitesimal circles (rather than
ellipses). Of course stereographic projection satisfies this requirement since it pre-
serves circles of all sizes. Figure [6.13a] illustrates this with an infinitesimal circle of
radius d bs on Σ being mapped to an infinitesimal circle of radius ds in C. To com-
plete the stereographic map we must find its associated metric function Λ—that is
the ratio of the two radii in [6.13a].
Consider the vertical cross section of [6.13a] shown in [6.13b], and recall that we
showed in Chapter 3 [see p. 161] that stereographic projection is a special case of
inversion:
√
If K is the sphere of radius 2 centred at N, then stereographic projection
is the restriction to C or Σ of inversion in K.

Next, consider (3.6) on p. 142, which describes the effect of inversion on the sepa-
ration of two points. By taking the limit in which the two points coalesce, we may
apply this result to [6.13b] to obtain [exercise]

2
db
s = ds.
[Nz]2

[6.13] Geometric derivation of the conformal metric under stereographic projection.

[a] A small, ultimately vanishing circle of radius d b
s on Σ is mapped to a circle of radius
ds in C. [b] Recalling that stereographic projection is the restriction to Σ of inversion
in the sphere K, (3.6) yields d bs ≍ [Nz] 2
2 ds ≍
2
1+|z|2
ds. This can also be obtained
more directly, without using (3.6), by [exercise] choosing d b s parallel to C; see Needham
(2021, p. 47) for the details.
 Spherical Geometry 325

This can also be obtained more directly, without using (3.6), by [exercise] choosing
db
s parallel to C. Finally, applying Pythagoras’ Theorem to the triangle Nz 0, we
obtain
2
db
s = ds. (6.16)
1 + |z|2
This flat conformal map with metric (6.16) is the desired abstract depiction of all
possible surfaces of constant Gaussian curvature K = +1.

6.2.4 Spatial Rotations as Möbius Transformations

Quite generally, suppose that S is a surface of constant Gaussian curvature (so that
it possesses a group of motions) and suppose we have drawn a conformal map of
S with metric (6.15). Any motion of S will induce a corresponding transformation
of this map in C. Since direct motions of the curved surface are conformal, the
conformality of the map implies that the induced complex functions must also be
conformal and hence analytic. Purely in terms of C, we may therefore identify a
function f(z) as a motion if it is analytic and it “preserves the metric” (6.15). That is,
suppose that the analytic function z ÞÑ e z = f(z) sends two infinitesimally separated
points z and (z+dz) to e z and ( e
z +d e
z ). Then f(z) is a motion if and only if the image
separation d es = |d e
z| is related to the original separation ds = |dz| by

Λ( e
z)de
s = Λ(z) ds.

[Likewise, opposite motions of S correspond to the anticonformal mappings of C

z = f ′ (z) dz, this is equivalent to demanding that
that satisfy this equation.] Since d e
f satisfy the following differential equation:
Λ(z)
f ′ (z) = .
Λ[f(z)]
Returning to the particular case S = Σ, and to the particular conformal map
obtained by stereographic projection, the direct motions of all possible surfaces of
constant Gaussian curvature K = +1 become the set of analytic complex functions
that satisfy
1 + |f(z)|2
f ′ (z) = . (6.17)
1 + |z|2
In principle, we could find these complex functions without ever leaving C. How-
ever, it is simpler and more illuminating to return to the motions of Σ, described by
(6.10) and (6.13). When we apply stereographic projection to these motions, what
complex functions are induced in C?
The first step is clearly to find the complex function induced by a reflection RLb of
Σ in a line Lb . Consider [6.14a], which shows a new intrinsic method of constructing
the reflection RLb (bz ) of a point b
z on Σ, namely [exercise], as the second intersection
326 Non-Euclidean Geometry*

[6.14] Reflection of Σ in a line induces reflection (inversion) of in the stereographic

image of that line. [a] The reflection RLb (bz ) of a point b
z on Σ across the line Lb is the
b
second intersection point of any two circles centred on L that pass through b z . [b] The
same construction after stereographic projection to C, revealing the standard orthogonal-
circle construction of geometric inversion in L.

point of any two circles centred on L b and passing through bz . Note that these two
b
circles are orthogonal to L . Figure [6.14b] shows what this construction looks like in
the stereographic map. Since stereographic projection preserves circles and angles,
the two circles orthogonal to L b and passing through b z are mapped to two circles
orthogonal to L and passing through z. The second intersection point of these circles
is thus the reflection IL (z) of z in L! To sum up,

Reflection of Σ in a line induces reflection (inversion) of C in the

(6.18)
stereographic image of that line.

[For a different proof of (6.18), one that is perhaps even more natural than the one
above, see Ex. 2.] As an important special case, note that if Lb is the intersection of
Σ with the vertical plane through the real axis, then reflection of Σ in L b induces
complex conjugation, z ÞÑ z.
Now let’s find the complex functions corresponding to rotations of Σ. Figure
[6.15] illustrates a rotation R ψ b be
b . Let b
b of Σ through angle ψ about the point a
a
the antipodal point to a b , so that b = −(1/a) [see (3.23), p. 167]. These points a b
b
and b lie on the axis of the rotation and remain fixed; correspondingly, a and
b will be the fixed points of the induced transformation of C. Furthermore, it is
clear geometrically that the effect of the induced transformation on an infinites-
imal neighbourhood of a is a rotation about a [exercise], and, by virtue of our
conventions, the rotation angle is negative ψ.
 Spherical Geometry 327

[6.15] A rotation of the sphere stereographically projects to an Elliptic Möbius trans-

formation of . According to (6.10), R ψ b 2 ◦ RL
b b
b 1 , where L 1 and L 2 are any two
b = RL
a
lines passing through a b such that the angle between them is (ψ/2). Since stereographic
projection preserves circles and angles, the images in C of these lines will be two circles
passing through the fixed points a and b, and containing angle (ψ/2) there. So (6.18)
shows that the induced transformation of C is effected by reflecting across these two
intersecting circles, yielding an Elliptic Möbius transformation with multiplier m = e−iψ .

According to (6.10),

Rψ b 2 ◦ RL
b = RL
a b 1,

where L b 1 and L
b 2 are any two lines passing through a b (and hence also through b b)
such that the angle between them is (ψ/2). Since stereographic projection preserves
circles and angles, the images in C of these lines will be two circles L1 and L2 passing
through the fixed points a and b, and containing angle (ψ/2) there. It follows from
(6.18) that the transformation R ψa induced by the rotation R a
ψ
b is

a = IL2 ◦ IL1 .
Rψ

Thus R ψa is a Möbius transformation! See [6.16], which illustrates a rotation of

ψ = (π/3). Referring back to (3.47) on p. 195, and recalling that the “multi-
plier” describes the local effect of a Möbius transformation in the immediate
neighbourhood of a fixed point, we have found that

A rotation Rψb of Σ stereographically induces an elliptic Möbius trans-

a
formation Rψ ψ
a of C. The fixed points of Ra are a and −(1/a), and the
multiplier m associated with a is m = e−iψ .
328 Non-Euclidean Geometry*

[6.16] The Elliptic Möbius transformation of induced by rotating the sphere by

ψ = (π/3). The fixed points of the rotation of the sphere are the antipodal points ab and
b
b on the axis of rotation, and these map to the fixed points of the Möbius transformation,
a and b = −(1/a). The Möbius transformation is effected by reflecting in any two circles
through these fixed points making angle −(ψ/2), such as L1 and L2 .

A straightforward matrix calculation [see Ex. 4] based on (3.42), p. 189, yields

the following explicit formula for the matrix of R ψ
a:
 i(ψ/2) 2 
 ψ e |a| + e −i(ψ/2)
2i a sin(ψ/2)
Ra =  . (6.19)
2i a sin(ψ/2) e −i(ψ/2)
|a| + e
2 i(ψ/2)

Note that this is in agreement with (3.39), p. 184: rotations of Σ induce Möbius
transformations of the form
Az + B
Rψa (z) = . (6.20)
−Bz + A
By virtue of (6.13), this formula represents the most general direct motion of Σ.
We have already noted that z ÞÑ z corresponds to a reflection of Σ, and it follows
[exercise] that the most general opposite motion is represented by a function of the
form
 
Az + B
z ÞÑ .
−B z + A
Figure [6.10b] provided a very elegant geometric method of composing rotations
of space. The above analysis now opens the way toan equally elegant method of
ψ
computing the net rotation produced by R ωb ◦ Ra
p b . All we need do is compose
the corresponding Möbius transformations:
 ω   ω  ψ
Rp ◦ Rψ
a = Rp Ra .
An otherwise tricky problem has been reduced to multiplying 2 × 2 matrices!
 Spherical Geometry 329

In practice, rotations are frequently expressed in terms of a unit vector v pointing

along the axis of rotation, with a b at its tip. However, (6.19) is currently expressed
in
h terms
i of the stereographic image a of the point a b . Let us therefore re-express
Rψ
a in terms of the components l, m, n of the unit vector

v = l i + m j + n k, l2 + m2 + n2 = 1.
Referring back to (3.20) on p. 166, we see that a and v are related as follows:
l + im 1+n
a= and |a|2 = .
1−n 1−n
Substituting these expressions into (6.19), and removing the common factor of
2/(1 − n), we obtain [exercise]
 
h i cos(ψ/2) + in sin(ψ/2) (−m + il) sin(ψ/2)
Rψv = . (6.21)
(m + il) sin(ψ/2) cos(ψ/2) − in sin(ψ/2)
h i
You may check for yourself that this matrix is “normalized”: det R ψ v = 1. This
makes life that much easier, for when we multiply two such matrices the resulting
matrix will be of precisely the same form. Thus, by comparing the result with (6.21),
we may read off the net rotation.
For example, suppose we perform a rotation of (π/2) about i, followed by a
rotation of (π/2) about j. The Möbius matrix of the net rotation will therefore be
" # " # " #
1 1 −1 1 1 i 1 1 − i −1 + i
√ √ = . (6.22)
2 1 1 2 i 1 2 1+i 1+i
Comparing this with (6.21), we see [exercise] that this is rotation of ψ = (2π/3)
about the axis v = √13 (i + j − k).

6.2.5 Spatial Rotations and Quaternions

This is all rather elegant, but in fact the above method of composing rotations
can be streamlined still further. To see how, let us resume the story of Hamilton’s
quaternions, which were introduced at the close of Chapter 1.
On the morning of Monday, 16 October 1843, Hamilton went for a walk with
his wife. In the back of his mind was a problem with which he had wrestled fruit-
lessly for more than ten years—the search for a three-dimensional analogue of the
complex numbers, one that would permit vectors in space to be multiplied and
divided. As we indicated in Chapter 1, Hamilton was unable to solve this prob-
lem for the simple reason that no such analogue exists. However, as he passed
Brougham Bridge, he suddenly realized that the prize which had eluded him in
three-dimensional space was indeed attainable in four-dimensional space!
In the two-dimensional complex plane, we may think of 1 and i as unit basis
“vectors” in terms of which a general complex number may be expressed as
330 Non-Euclidean Geometry*

z = a1 +b i. The algebra of C amounts to stipulating that multiplication distributes

over addition, that 1 is the identity (i.e., 1z = z1 = z), and that i2 = −1.
In four-dimensional space, Hamilton introduced four basis vectors 1, I, J, K in
terms of which a general vector V (which Hamilton called a quaternion) could be
expressed as

V = v1 + v1 I + v2 J + v3 K, (6.23)

where the coefficients are all real numbers. To define the product of two such
quaternions, Hamilton took 1 to be the identity, and he took I, J, K to be three
different square roots of −1, each analogous to i:

I2 = J2 = K2 = −1. (6.24)

As in ordinary algebra, Hamilton insisted that multiplication distribute over addi-

tion, but in order to render division possible he was forced to make a leap that
was revolutionary in its time: non-commutative multiplication. More precisely,
Hamilton postulated that

IJ = K = −JI, JK = I = −KJ, KI = J = −IK. (6.25)

These relations probably look familiar: they are formally identical to the vector
products of the basis vectors i, j, k in three-dimensional space. For example, i×j =
k = −j×i. We can use this analogy between i, j, k and I, J, K to express the product
of two quaternions in a particularly simple way.
First, let’s use the analogy to simplify the notation (6.23). As in ordinary algebra,
we suppress the identity 1 in the first term and write v 1 = v, which Hamilton called
the scalar part of V. Next we collect the remaining three terms into V ≡ v1 I + v2 J +
v3 K, which Hamilton called the vector part of V. Thus (6.23) becomes

V = v + V.

In the special case where the scalar part v vanishes, Hamilton called V = V a pure
quaternion. Historically, the concept of a pure quaternion was the forerunner of the
idea of an ordinary vector in space. In fact the very word “vector” was coined by
Hamilton in 1846 as a synonym for a “pure quaternion”.
If we multiply V by another quaternion W = w + W, then (6.24) and (6.25) imply
[exercise] that

·
V W = (v w − V W) + (v W + w V + V×W). (6.26)

In particular, if V and W are pure (i.e., v = 0 = w) then this reduces to

·
V W = −V W + V×W. (6.27)

Historically, this formula constituted the very first appearance in mathematics of

the concepts of the dot and cross products. Thus, initially, these vectorial operations
 Spherical Geometry 331

were viewed as merely two facets (the scalar and vector parts) of quaternion multi-
plication. However, it did not take physicists long to realize that the scalar product
and the vector product were each important in their own right, independently of
the quaternions from which they had both sprung.
Further results on quaternions will be derived in the exercises; here we wish
only to explain the connection between quaternions and rotations of space. This
connection hinges on the idea of a binary rotation, which means a rotation of space
though an angle of π. The appropriateness of the word “binary” stems from the
fact that if the same binary rotation is applied twice then the result is the identity.
According to (6.21), the Möbius transformation corresponding to the binary
rotation about the axis v = l i + m j + n k is
" #
in −m + il
π
[R v ] = .
m + il −in
Now, forgetting about quaternions for a moment, let us redefine 1 to be the identity
matrix, and I, J, K to be the binary rotation matrices about i, j, k, respectively. Thus
" # " # " # " #
1 0 0 i 0 −1 i 0
1= , I= , J= , K= .
0 1 i 0 1 0 0 −i
As a simple check, note that the Möbius transformation corresponding to the
Möbius matrix K is K(z) = −z. Make sure you can see why this is as it should
be.
Now we can state the surprising connection with quaternions: under matrix mul-
tiplication, these binary rotation matrices obey [exercise] exactly the same laws (6.24) and
(6.25) as Hamilton’s I, J, K. It follows that quaternion multiplication is equivalent
to multiplying the corresponding 2 × 2 matrices obtained by replacing Hamilton’s h i
1, I, J, K with the matrices above. Conversely, the general rotation matrix R ψ
v in
(6.21) can be expressed [exercise] as the quaternion

Rψ
v = cos(ψ/2) + V sin(ψ/2), (6.28)

where V = l I + m J + n K. This elegant formula is much easier to remember than

(6.21)!
To compose two rotations of space, we need only multiply the corresponding
quaternions. For example, the calculation (6.22)—in which a rotation of (π/2) about
i was followed by a rotation of (π/2) about j—now becomes
√1 (1 + J) √12 (1 + I) = 12 (1 + I + J − K).
2

Once again, but more easily than before, we deduce that this is rotation of ψ =
(2π/3) about the axis v = √13 (i + j − k).
Quaternions also yield a very compact formula for the effect of R ψ v on the posi-
tion vector P = X i + Y j + Z k of a point in space. Suppose that R ψ rotates e If
P to P.
v
332 Non-Euclidean Geometry*

we represent P by the pure quaternion P = X I + Y J + Z K, and likewise represent

e then
e as P,
P
e = R ψ P R −ψ .
P (6.29)
v v
This result was first published by Arthur Cayley in 1845, though he later conceded
priority to Hamilton. Not only is the result elegant, it is also practical. For example,
Hoggar (1992) discusses how (6.29) can be used to smooth the motion of a rotating
object in a computer animation, while Horn (1991) has used it in research connected
with robotic vision!10
Here we will give the most intuitive explanation of (6.29) that we have been able
to think of; Exs. 7, 8 give two more. Begin by noting that any multiple of P is rotated
to the same multiple of P. e To establish (6.29) in general, it is therefore sufficient to
establish it for the case where P and P e are unit vectors whose tips p b and pb
e lie on
the unit sphere. As before, let ab be the point at the tip of v.
 
−ψ
Consider the following composition of three rotations: R ψ ab ◦ Rpb ◦ Ra
θ
b . Cer-
tainly this is equivalent to a single rotation, and [6.17] helps us to see what it is.
Let C be the invariant circle of R ψ b and pb
e , and let w be an
b passing through p
a
infinitesimal vector emanating from p b
e and tangent to C. Note that any vector ema-
nating from a point on C will be carried by R ±ψ b
a into a vector making the same

[6.17] Geometric derivation of the quaternion rotation formula. A rotation of θ about

b
e , turning w into w ′′′ , can be expressed in terms of an equal rotation about p
p b , because
−ψ ′ ′′ ′′′
the effect of R θpbe = R ψab ◦ R θ
b
p ◦ R b
a is w Ñ
Þ w Ñ
Þ w Ñ
Þ w . Expressed in terms of
−ψ
quaternions, R θep = R ψ
v R p R v . Finally, putting θ = π, the binary rotations R p and
θ π

R π become the pure quaternions P and P, e thereby proving the quaternion rotation
e
p
e = Rψ
formula discovered by Hamilton and rediscovered by Cayley: P −ψ
v P Rv .

10
NOTE for the 25th Anniversay Edition: There have no doubt been countless other important
applications of this idea over the past 25 years, but I have not researched them for this new edition.
 Hyperbolic Geometry 333

angle with C. This justifies the illustrated effect w ÞÑ w ′ ÞÑ w ′′ ÞÑ w ′′′ of the three
b
rotations. Thus the net effect w ÞÑ w ′′′ is a rotation of θ about p
e:
−ψ
R θpbe = R ψ
b ◦ Rp
a b ◦ Ra
θ
b .

This geometric fact may be expressed in terms of Möbius matrices, or equivalently

in terms of quaternions:
−ψ
R θpe = R ψ
v Rp Rv .
θ

Finally, if we put θ = π then the binary rotations R π

p and R p
π
e are simply the pure
quaternions P and P,e so we are done.
Further Reading. For more on the historical significance of (6.29), see Altmann
(1989); for the details of how Hamilton was led to quaternions, see Waerden
(1985); for discussion of the connections with modern mathematics and physics,
see Penrose and Rindler (1984), Yaglom (1988), Stillwell (1992), and Penrose (2005).

6.3 Hyperbolic Geometry

6.3.1 The Tractrix and the Pseudosphere

Having studied the intrinsic geometry of surfaces of constant positive Gaussian
curvature, we now turn to the intrinsic geometry of surfaces of constant nega-
tive curvature. Just as there are infinitely many surfaces with K > 0, so there are
infinitely many with K < 0. Beltrami called such surfaces pseudospherical. Accord-
ing to the previously stated result of Minding, all pseudospherical surfaces having
the same negative value of K possess the same intrinsic geometry. To begin to
understand hyperbolic geometry, it is therefore sufficient to examine any pseudo-
spherical surface. For our purposes, the simplest one is the pseudosphere, so let us
explain how this surface may be constructed.
Try the following experiment. Take a small heavy object, such as a paperweight,
and attach a length of string to it. Now place the object on a table and drag it by
moving the free end of the string along the edge of the table. You will see that
the object moves along a curve like that in [6.18a], where the Y-axis represents the
edge of the table. This curve is called the tractrix, and the Y-axis (which the curve
approaches asymptotically) is called the axis. The tractrix was first investigated by
Newton, in 1676.
If the length of the string is R, then it follows that the tractrix has the following
geometric property: the segment of the tangent from the point of contact to the Y-axis
has constant length R. This was Newton’s definition of the tractrix. As an interesting
aside, it follows [exercise] that the tractrix can be constructed as shown in [6.18b],
namely, as an orthogonal trajectory through the family of circles of radius R centred
334 Non-Euclidean Geometry*

on the axis. This provides a good method of quickly sketching a fairly accurate
tractrix.
Returning to [6.18a], let σ represent arc length along the tractrix, with σ = 0
corresponding to the starting position X = R of the object we are dragging. Just as
the object is about to pass through (X, Y), let dX denote the infinitesimal change
in X that occurs while the object moves a distance dσ along the tractrix. From the
similarity of the illustrated triangles, we deduce that

dX X
=− =⇒ X = R e−σ/R . (6.30)
dσ R
The pseudosphere of radius R may now be simultaneously defined and constructed
as the surface obtained by rotating the tractrix about its axis. Remarkably, this sur-
face was investigated as early as 1693 (by Christiaan Huygens), two centuries prior
to its catalytic role in the acceptance of hyperbolic geometry.

[6.18] Newton’s tractrix. [a] Tie a small paperweight to a piece of string of length
R. On a table top, start with the string running along one edge, at right angles to
the other edge, the Y-axis. If you move the free end of the string along the Y-axis
edge of the table, the paperweight will be dragged along the illustrated curve, called
the tractrix, which Newton first investigated in 1676. We see that dX X
dσ = − R and so
−σ/R
X = Re . [b] The tractrix can also be constructed as an orthogonal trajectory
through the family of circles of radius R centred on the Y-axis.
 Hyperbolic Geometry 335

6.3.2 The Constant Negative Curvature of the Pseudosphere*

In this optional section we offer a purely geometric proof that the pseudosphere
does indeed have constant negative Gaussian curvature. More precisely, we will
use the extrinsic definition of K as the product of the principal curvatures to show
that the pseudosphere of radius R has constant negative curvature K = −(1/R2 ). Later
we will give a purely intrinsic demonstration of this fact, so you won’t miss too
much if you skip the following argument.
Let r and er be the two principal radii of curvature of the pseudosphere of radius
R. As with any surface of revolution, it follows by symmetry11 that

er = radius of curvature of the generating tractrix,

r = the segment of the normal from the surface to the axis,

as illustrated in [6.19a]. The problem of determining the Gaussian curvature

1
K=−
r er
is thereby reduced to a problem in plane geometry, which is solved in [6.19b].

[6.19] Geometric proof that the pseudosphere has constant negative Gaussian curva-
ture. [a] The principal directions of maximum and minimum curvature will always occur
in the orthogonal directions in which a surface has local mirror symmetry. [See Needham
(2021).] Clearly, one direction in which the pseudosphere has mirror symmetry is straight
up the pseudosphere, along a meridian tractrix generator. The other principal direction is
the orthogonal sideways direction along a circle of latitude. Thus, K = −1/(r er). [b] From
this figure we can deduce [see text] that the pseudosphere does indeed have constant
negative Gaussian curvature, K = −(1/R2 ).

11
For a full discussion of all these ideas, see Needham (2021).
336 Non-Euclidean Geometry*

By definition, the tractrix in this figure has tangents of constant length R. At the
neighbouring points P and Q, figure [6.19b] illustrates two such tangents, PA and
QB, containing angle •. The corresponding normals PO and QO therefore contain
the same angle •. Note that AC has been drawn perpendicular to QB.
Now let’s watch what happens as Q coalesces with P, which itself remains fixed.
In this limit, O is the centre of the circle of curvature, PQ is an arc of this circle, and
AC is an arc of a circle of radius R centred at P. Thus,12
PQ AC AC R
er ≍ OP and ≍ • ≍ =⇒ ≍ .
OP R PQ er
Next we appeal to the defining property PA = R = QB of the tractrix to deduce
[exercise] that as Q coalesces with P,

BC ≍ PQ.

Finally, using the fact that as Q coalesces with P the triangle ABC is ultimately
similar to the triangle T AP, we deduce that
r AC AC R
≍ ≍ ≍ .
R BC PQ er
Behold!13
1 1
K=− =− 2.
r er R

6.3.3 A Conformal Map of the Pseudosphere

Our next step is to construct a conformal map of the pseudosphere. Recall the bene-
fits of such a map in the case of a sphere: (1) it simultaneously describes all surfaces
of curvature K = +1; (2) it provides an elegant and practical description of the
motions as Möbius transformations. Both of these benefits persist in the present
case of negatively curved surfaces; in particular, the (direct) motions of hyperbolic
geometry again turn out to be Möbius transformations!
For simplicity’s sake, henceforth we shall take the radius of the pseudosphere to be
R = 1, so our map will represent pseudospherical surfaces of curvature K = −1.
[NOTE: This is the conventional choice in almost all texts on hyperbolic geome-
try.] As a first step towards a conformal map, [6.20a] introduces a fairly natural
coordinate system (x, σ) on the pseudosphere.
The first coordinate x measures angle around the axis of the pseudosphere, say
restricted to 0 ⩽ x < 2π. The second coordinate σ measures arc length along
each tractrix generator (as in [6.18a]). Thus the curves x = const. are the tractrix

Once again, “ ≍ ” denotes Newton’s concept of ultimate equality; see the new Preface.
12
13
One of my most treasured possessions is an email from Bill Thurston saying that he liked my
Newtonian proof of this fundamentally important fact!
 Hyperbolic Geometry 337

[6.20] Constructing a conformal map of the pseudosphere. [a] Having specialized to the
standard case, R = 1, we have X = e−σ . [b] Let us decide to use the angle x around the
axis of symmetry as the x-coordinate in this map. If we insist that the map be conformal
then all small (ultimately vanishing) distances must be scaled equally, regardless of direc-
tion. Since the sideways movement X dx on the left is divided by X in passing to the map
dy
on the right, all distances must be divided by X, so dσ = X1 = eσ and therefore y = eσ +k.
Choosing k = 0, we find that the distance db s on the pseudosphere is related
p to the dis-
tance ds in the map by the conformal metric formula, db s = ds/y = dx2 + dy2 /y.

generators of the pseudosphere [note that these are clearly geodesics], and the
curves σ = const. are circular cross sections of the pseudosphere [note that these
are clearly not geodesics]. Since the radius of such a circle is the same thing as the
X-coordinate in [6.18a], it follows from (6.30) that

The radius X of the circle σ = const. passing through

the point (x, σ) is given by X = e−σ .

In our map, let us choose the angle x as our horizontal axis, so that the tractrix
generators of the pseudosphere are represented by vertical lines. See [6.20b]. Thus
a point on the pseudosphere with coordinates (x, σ) will be represented in the map
by a point with Cartesian coordinates (x, y), which we will soon think of as the
complex number z = x + iy.
If our map were not required to be special in any way, then we could simply
choose y = y(x, σ) to be an arbitrary function of x and σ. In stark contrast to
338 Non-Euclidean Geometry*

this, our requirement that the map be conformal leaves (virtually) no freedom in
the choice of the y-coordinate. Let’s try to understand this.
Firstly, the tractrix generators x = const. are orthogonal to the circular cross
sections σ = const., so the same must be true of their images in our conformal map.
Thus the image of σ = const. must be represented by a horizontal line y = const.,
and from this we deduce that y = y(σ) must be a function solely of σ.
Secondly, on the pseudosphere consider the arc of the circle σ = const. (of radius
X) connecting the points (x, σ) and (x + dx, σ). By the definition of x, these points
subtend angle dx at the centre of the circle, so their separation on the pseudo-
sphere is X dx, as illustrated. In the map, these two points have the same height
and are separated by distance dx. Thus in passing from the pseudosphere to the
map, this particular line-segment is shrunk by factor X. [We say “shrunk” because
we’re dividing by X, but since X ⩽ 1 this is actually an expansion.] However, since
the map is conformal, an infinitesimal line-segment emanating from (x, σ) in any
direction must be multiplied by the same factor (1/X) = eσ . In other words, the
metric is
db
s = X ds.
Thirdly, consider the uppermost black disc on the pseudosphere shown in
[6.20a]. Think of this disc as infinitesimal, say of diameter ϵ. In the map, it will be
represented by another disc, whose diameter (ϵ/X) may be interpreted more vividly
as the angular width of the original disc as seen by an observer on the pseudo-
sphere’s axis. Now suppose we repeatedly translate the original disc towards the
pseudosphere’s rim, moving it a distance ϵ each time. Figure [6.20a] illustrates the
resulting chain of touching, congruent discs. As the disc moves down the pseudo-
sphere, it recedes from the axis, and its angular width as seen from the axis therefore
diminishes. Thus the image disc in the map appears to gradually shrink as it moves
downward, and the equal distances 8ϵ between the successive black discs certainly
do not appear equal in the map.
Having developed a feel for how the map works, let’s actually calculate the y-
coordinate corresponding to the point (x, σ) on the pseudosphere. From the above
observations (or directly from the requirement that the illustrated triangles be
similar) we deduce that
dy 1
= = eσ =⇒ y = eσ + const.
dσ X
The standard choice of this constant is 0, so that
y = eσ = (1/X).
Thus the entire pseudosphere is represented in the map by the shaded region lying
above the line y = 1 (which itself represents the pseudosphere’s rim), and the
metric associated with the map is
 Hyperbolic Geometry 339

p
ds dx2 + dy2
db
s = = . (6.31)
y y
For future use, also note that an infinitesimal rectangle in the map with sides dx
and dy represents a similar infinitesimal rectangle on the pseudosphere with sides
(dx/y) and (dy/y). Thus the apparent area dx dy in the map is related to the true
area dA on the pseudosphere by
dx dy
dA = . (6.32)
y2

6.3.4 Beltrami’s Hyperbolic Plane

In the Introduction we gave the impression that Beltrami had succeeded in inter-
preting abstract hyperbolic geometry as the intrinsic geometry of the pseudo-
sphere. This is really not possible, and it is not what Beltrami claimed.
The abstract hyperbolic geometry discovered by Gauss, Bolyai, and Loba-
chevsky is understood to take place in a hyperbolic plane that is exactly like the
Euclidean plane, except that lines within it obey the hyperbolic axiom (6.3):
Given a line L and a point p not on L, there are at least two lines through
p that do not meet L.
The constant negative curvature of the pseudosphere ensures that it faithfully rep-
resents all consequences of this axiom that deal only with a finite region of the
hyperbolic plane. An example of such a consequence is the theorem that the angu-
lar excess of a triangle is a negative multiple of its area, and this does indeed hold
on the pseudosphere.
However, the pseudosphere will not do as a model of the entire hyperbolic plane,
because it departs from the Euclidean plane in two unacceptable ways:

• The pseudosphere is akin to a cylinder instead of a plane. For example, a closed

loop in the plane can always be shrunk to a point, but a loop on the pseudosphere
that wraps around the axis cannot be.
• In the hyperbolic plane, as in the Euclidean plane, a line-segment can be
extended indefinitely in either direction. We have already remarked that the
tractrix generators of the pseudosphere are clearly geodesic, and we would
therefore like to interpret them as hyperbolic lines. But although such a tractrix
extends indefinitely up the pseudosphere, in the other direction it terminates
when it hits the rim.

Beltrami pointed out that the first of these problems can be resolved as follows.
Imagine the pseudosphere covered by a thin stretchable sheet. To obtain the map in
[6.20b], we cut this sheet along a tractrix generator and unwrap it onto the shaded
340 Non-Euclidean Geometry*

region. Of course to make it lie flat and fit into this rectangular region, the sheet
must be stretched—the metric (6.31) tells us how much stretching must be applied
to each part. But now imagine the sheet as wrapping round and round the pseudo-
sphere infinitely many times14 , like an endless roll of cling film15 . By unwrapping
this infinitely long sheet (stretching as we go) we can now cover the entire region
above y = 1. According to this interpretation, a particle travelling along a horizon-
tal line in the map would correspond to a particle travelling round and round a
circle σ = const. on the pseudosphere, executing one complete revolution for each
movement of 2π along the line.
Now let us explain how the conformal map solves our second problem—the
pseudosphere’s edge. In terms of extrinsic geometry, this edge is an insurmountable
obstacle: we cannot extend the pseudosphere smoothly beyond this edge while
preserving its constant curvature. However, we only care about the pseudosphere’s
intrinsic geometry, and we have seen that if we measure distance using d b s = dsy ,
this is identical to the region y > 1 in [6.21].
Imagining yourself as a tiny two-dimensional being living in [6.21], walking
down a line x = const. is exactly the same thing as walking down a tractrix on
the pseudosphere. Of course on the pseudosphere your walk is rudely interrupted
at some point p b on the rim (σ = 0), corresponding to a point p on the line y = 1.
But in the map this point p is just like any other, and there is absolutely nothing
preventing you from continuing your walk all the way down to the point q on
y = 0.
Why stop at q? The answer is that you will never even get that far, because q
is infinitely far from p! Suppose that you are the illustrated small disc on the line
y = 2, and that I am standing outside your hyperbolic world, watching as you walk
at a steady pace towards y = 0. Of course you remain the same hyperbolic size as
you walk, but to me you appear to shrink. This is made particularly vivid by the
illustrated Euclidean interpretation [exercise] of your hyperbolic size d b s = ds
y :

The hyperbolic diameter of an infinitesimal disc centred at (x + iy) is

(6.33)
the angle it subtends at the point x on the real axis.
Thus your apparent size must shrink so that you subtend a constant angle, and
although all your hyperbolic strides are the same length, to me they look shorter
and shorter, and you appear to be travelling more and more slowly.
For example, suppose you are walking at a steady speed of ln 2. As illustrated,
integration of (dy/y) shows [exercise] that you reach y = 1 after one unit of time,
y = (1/2) after two units of time, y = (1/4) after three units of time, etc. Thus,
viewed from outside your world, each successive unit of time only halves your

14
Stillwell (1996) points out that this is probably the very first appearance in mathematics of what
topologists now call a universal cover.
15
For Americans, read “plastic wrap”.
 Hyperbolic Geometry 341

[6.21] The pseudosphere has an edge, but the hyperbolic plane is uniform and infinite.
Suppose “You” start at the point on the pseudosphere corresponding to y = 2 and walk
down a tractrix generator at a steady rate of ln 2. Then in the conformal map you reach
y = 1 (the pseudosphere’s rim—the edge of your world!) after just one unit of time. But
if we imagine you living in the map, perceiving distances around you via the hyperbolic
metric, then you can continue your downward journey, arriving at y = (1/2) after two
units of time, y = (1/4) after three units of time, etc. Although you actually stay the same
size, in the map you appear to shrink, for your true size is the angle dbs that you subtend
at the horizon. Thus, viewed from outside your world, you shrink with each successive
unit of time, only halving your distance from the horizon, y = 0; therefore, you will never
reach it. Your world is infinite and uniform: no place or direction seems any different from
any other.

distance from y = 0, and therefore you will never reach it. [An appropriate name
for this phenomenon might be “Zeno’s Revenge”!]
We now possess a concrete model of the hyperbolic plane, namely, the entire
shaded half-plane y > 0 with metric d b s = dsy . The points on the real axis are
infinitely far from ordinary points and are not (strictly speaking) considered part of
the hyperbolic plane. They are called ideal points, or points at infinity. The complete
line y = 0 of points at infinity will be called the horizon16 .
Studying hyperbolic geometry by means of this map is like studying spheri-
cal geometry via a stereographic map, without ever having seen an actual sphere.
This is not as bad as it sounds. After all, by constructing geographical maps
through terrestrial measurements, man developed a good understanding of the
surface of the Earth centuries before venturing into space and gazing down on its
roundness!
Still, it would be nice to have the analogue of a globe instead of a mere atlas.
The pseudosphere only models a portion of the hyperbolic plane, but might there

16
For reasons that will be clear shortly, another name is the circle at infinity.
342 Non-Euclidean Geometry*

exist a different surface that is isometric to the entire hyperbolic plane? Sadly, in
1901 Hilbert proved17 that every pseudospherical surface necessarily has an edge
beyond which it cannot be smoothly extended while preserving its constant nega-
tive curvature. Thus the upper half-plane with metric (6.31) is as good a depiction
of the hyperbolic plane as we are going to get.
However, just as an atlas uses different kinds of maps to represent the surface of
the Earth, so we can and will use different types of maps to represent the hyperbolic
plane. The particular map we have obtained is conventionally called the Poincaré
upper half-plane, but there is also one called the Poincaré disc, and another called the
Klein disc. Poincaré obtained the first two models in 1882, while Klein obtained the
third in 1871.
We cannot let the names of these models pass without comment. Anyone with
even a passing interest in the history of mathematics will know that ideas are fre-
quently (usually?) named after the wrong person. In fact18 , the three models above
were all discovered by Beltrami! As we shall see, Beltrami obtained these three mod-
els (in 1868, 14 years before Poincaré) in a beautifully unified way, from a fourth
model consisting of a map drawn on a hemisphere. And in case you’re wondering,
yes, the hemisphere model is Beltrami’s, too!
In this 25th Anniversary Edition we shall dogedly attempt to set history straight,
as we have previously attempted to do in VDGF, by giving both men equal credit,
and renaming the conformal maps as the Beltrami–Poincaré half-plane and disc
models.

6.3.5 Hyperbolic Lines and Reflections

Before we get going, let’s indicate where we are going, focusing just on direct
motions. In Euclidean geometry, every direct motion is the composition of reflec-
tions in two lines. We have seen that the same is true in spherical geometry, and we
will soon show that it is again true in hyperbolic geometry. Since two Euclidean
lines must intersect or be parallel, there are just two kinds of direct Euclidean
motions: rotations and translations. The absence of parallel lines on the sphere
implies that its direct motions can only be rotations. Conversely, the multitude of
parallel lines in the hyperbolic plane yields a geometry that is richer than Euclid’s,
containing rotations, translations, and a third kind of motion that has no Euclidean
counterpart.
To avoid confusion, let us use the prefix “h-” to distinguish hyperbolic concepts
from their Euclidean descriptions in the map. For example, an “h-line” will mean a
“hyperbolic line” (i.e., a geodesic), while a “line” will refer to an ordinary straight

17
Hilbert (1965, Vol. 2, pp. 437–448), with English translation available in Hilbert (1902).
18
See Milnor (1982), Beltrami (1868a), Beltrami (1868b), and Stillwell (2010).
 Hyperbolic Geometry 343

line in the map. Let us also define H{z1 , z2 } to be the h-distance (measured using
dbs = dsy ) between z1 and z2 . For example, if dz is infinitesimal, then

|dz|
H{z + dz, z} = .
Im z
Finally, let us define an h-circle of h-radius ρ and h-centre c to be the locus of
points z such that H{z, c} = ρ.
Since tractrix generators of the pseudosphere are clearly geodesic, vertical lines
in the map should also be geodesic, i.e., they should be examples of h-lines. Figure
[6.22a] confirms this directly by showing that

The (unique) shortest route between two vertically separated points is the
(6.34)
vertical line-segment L connecting them.

To see this, compare L with any other route, such as M. Let ds1 be an infinitesimal
segment of L at height y, and let ds2 be the corresponding element of M cut off by
horizontal lines through the ends of ds1 . Since
ds1 ds2
db
s1 = < = db
s 2,
y y
the total hyperbolic length of L is less than M’s. Done. From this we can deduce
that

H{(x + iy1 ), (x + iy2 )} = |ln(y1 /y2 )| . (6.35)

Through a given point of the pseudosphere we obviously have geodesics in all

directions, not just tractrix generators; what do these more general h-lines look like
in the map? The answer is very beautiful and unexpected:

[6.22] [a] Vertical lines are h-lines (geodesics), because d b s 1 = ds1 /y < ds2 /y = d b
s 2.
[b] Inversion in a semicircle orthogonal to the horizon preserves h-distance. Under
z ÞÑ ez = IK (z), we see that d be
s = de y = ds/y = d b
s/e s . Since distance is preserved in
this particular direction, it must be preserved in all directions.
344 Non-Euclidean Geometry*

Every h-line is either a half-line orthogonal to the horizon, or else a semicircle

(6.36)
orthogonal to the horizon.
Before we prove this, it’s important to realize that if you were an inhabitant of the
hyperbolic plane, there would be no way for you to distinguish between the semi-
circular h-lines and the vertical h-lines: every line is exactly like every other, it’s just
our map that makes them look different. What about the fact that the semicircles
have two ends on the horizon, whereas the vertical h-lines appear to only have one?
The answer is that, in addition to the points on the real axis, there is one more point
at infinity, and all the vertical h-lines meet there. According to (6.31), as we move
upward along two neighbouring, vertical h-lines, the h-distance between them dies
away like (1/y), and they converge to a single point at infinity; this is particularly
vivid on the pseudosphere. Finally, note that even in terms of the map, a vertical
h-line may be viewed as just a special case of a semicircular h-line by allowing the
radius to tend to infinity.
We will prove (6.36) by first establishing another equally beautiful fact, one that
is fundamental to all that follows:
Inversion in a semicircle orthogonal to the horizon is an opposite motion of
(6.37)
the hyperbolic plane.
To see why this is true, consider the inversion z ÞÑ e z = IK (z) illustrated in [6.22b].
We need to show that IK (z) does not alter the h-length d b s of any infinitesimal
line-segment ds emanating from z. However, because our model of the hyperbolic
plane is conformal, we need only show that IK (z) preserves the h-length of any sin-
gle ds, in a direction of our choosing. Choosing ds orthogonal to the radius qz of
K (as illustrated), the anticonformality of inversion implies that the image d e s is
also orthogonal to this radius. Thus, by virtue of the illustrated similar triangles, it
follows [exercise] that
des ds
dbe
s = = = db s,
e
y y
as was to be shown.
To establish (6.36), consider [6.23a]. First, the figure shows that two points a and
b [Re(a) ̸= Re(b)] can always be joined by a unique arc L of a semicircle orthogonal
to the real axis: to construct the centre c, simply draw the perpendicular bisector of
ab. As illustrated, let q be one of the ends of this semicircle. Now we need to show
that L is the shortest (smallest h-length) route from a to b.
We show this by applying an inversion z ÞÑ e z = IK (z), where K is any circle
centred at q. This carries the arc L into a vertical line-segment L, e and (6.37) tells us
e and L have equal h-length. More generally, any route M from a to b has the
that L
same h-length as the route M f = IK (M) from a e Thus if L were not the shortest
e to b.
 Hyperbolic Geometry 345

[6.23] [a] Geometric proof that semicircles orthogonal to the horizon are h-lines
(geodesics). Construct L as shown, then invert it in any circle K centred at q, obtain-
ing IK (L) = L,e as shown. Since L e has been proven to be the shortest route from a e to
e and since IK preserves h-distance, L must be the shortest route from a to b. [b] If K
b,
is an h-line, then IK = RK is (literal) h-reflection in K. For if e z = IK (z) then all circles
through z and e z cut K orthogonally, and so the illustrated h-line P does, too. The h-lengths
of zm and me z are equal, for they are interchanged by the h-distance-preserving IK , so
e
z = RK (z), as claimed.

route from a to b, then Le would not be the shortest route from a e in violation
e to b,
of (6.34). Done.
Incidentally, note that this construction also enables us (in principle) to calculate
the h-distance between any two points in the hyperbolic plane:
 
e e
Im a
H{a, b} = H{e a, b} = ln ,
e
Im b
by virtue of (6.35). Later we shall be able to derive a more explicit formula.
The fact that a semicircle orthogonal to the real axis is an h-line strongly suggests
the following re-interpretation of (6.37): hyperbolic plane in the h-line
Inversion in a semicircle K orthogonal to the horizon is a reflection RK of the
(6.38)
hyperbolic plane in the h-line K.
In symbols, RK (z) = IK (z). Before proving this, let’s be clear what we mean by
reflection. Just as we would in Euclidean and spherical geometry, we begin the
construction of RK (z) by drawing the h-line P that passes through z and cuts K
perpendicularly, say at m. Then RK (z) is defined to be the point on P that is the
same h-distance from m as z.
To prove (6.38), consider [6.23b], in which e
z = IK (z). First recall that every circle
through z and e z is automatically orthogonal to K. In particular, the unique h-line
through z and ez must be orthogonal to K, and hence it is the desired “P” of the previ-
ous paragraph. Finally, recall that IK maps P into itself, swapping the segments zm
and ez m. Thus, since IK is a motion, these two h-line segments have equal h-length,
as was to be shown.
346 Non-Euclidean Geometry*

[6.24] The hyperbolic plane satisfies the Hyperbolic Axiom, for we see that there are
infinitely many h-lines [shown dashed] through p that fail to meet the h-line L: they are said
to be ultra-parallel to L. The two h-lines that meet L at the horizon are called asymptotic.
The h-distance of p from L is defined as in Euclidean geometry, as the length of the unique
line segment pq that meets L at right angles.

Conversely, if we are given any two points z and ez, then we may draw the per-
e
pendicular h-bisector K, and RK swaps z and z. Also note that z and its reflection
e
z = RK (z) are the same h-distance from every point k on K, just as in Euclidean and
spherical geometry. This is easily proved: since IK is a motion, and e k = IK (k) = k,
it follows that H{z, k} = H{e z, e
k} = H{e
z, k}.
It is becoming clear that hyperbolic geometry has much in common with
Euclidean geometry. However, now that we know what h-lines look like, [6.24]
shows that hyperbolic geometry really is non-Euclidean: there are infinitely many
h-lines through p [shown dashed] that do not meet the h-line L. Such h-lines are
said to be ultra-parallel to L.
Separating the ultra-parallels from the h-lines that do intersect L, we see that
there are precisely two h-lines that fail to meet L anywhere within the hyperbolic
plane proper, but that do meet it on the horizon. These two h-lines are called
asymptotic19 .
As in Euclidean geometry, the figure makes it clear that there is precisely one
h-line M passing through p that cuts L at right angles (say at q). In fact [exercise]
M may be constructed as the unique h-line through p and RL (p). The existence of
M makes it possible to define the distance of a point p from a line L in the usual
way, namely, as the h-length of the segment pq of M.
Since M and L are orthogonal, RM = IM maps L into itself, swapping the two
ends on the horizon. It follows [exercise] that RM swaps the two asymptotic lines,

19
Another commonly used name is parallel.
 Hyperbolic Geometry 347

[6.25] The same geometry as the previous figure, but in the case where L is a vertical
half-line.

and that M bisects the angle at p contained by the asymptotic lines. The angle
between M and either asymptotic line is called the angle of parallelism, and is usually
denoted Π. As one rotates the line M about p, its intersection point on L moves off
towards infinity, and Π tells you how far you can rotate M before it starts missing
L entirely.
Finally, [6.25] merely serves to illustrate the same concepts and terminology as
[6.24], but in the case where the h-line L happens to be represented as a vertical
half-line instead of a semicircle.

6.3.6 The Bolyai–Lobachevsky Formula*

This brief, optional subsection nicely illustrates how the preceding ideas may be
used to solve a significant, concrete problem: finding the angle of parallelism, Π.
In Euclidean geometry the analogue of the two asymptotic lines is the unique
parallel line through p, and since this is perpendicular to M, the analogue of Π is a
right angle. On the other hand, in hyperbolic geometry it is clear that Π is always
acute, and that its value decreases as the distance D ≡ H{p, q} of p from L increases.
More precisely, both Bolyai and Lobachevsky obtained this result:

Bolyai–Lobachevsky Formula : tan(Π/2) = e−D ,

and from this they were able to derive many of their other results. We now
give a simple geometric proof of this so-called Bolyai–Lobachevsky Formula. Green-
berg (2008) has called this “one of the most remarkable formulas in the whole of
mathematics”, but for us it will be of only incidental interest.
348 Non-Euclidean Geometry*

[6.26] Geometric proof of the Bolyai–Lobachevsky Formula. To find the h-length D of

the arc pq, apply z ÞÑ e
z = RC (z). This carries
 the
 arc pqinto the
 illustrated vertical
eq. Applying (6.35), D = ln [epm]
line-segment p [qm] [cm]
= ln [epm] = |ln tan(Π/2)| =
− ln tan(Π/2), the last equality following from the fact that Π is acute. Thus, tan(Π/2) =
e−D , as was to be shown.

First note that it is sufficient to establish the formula using [6.25], rather than
[6.24]. This is because we may transform [6.24] into [6.25] by performing an inver-
sion (i.e., a hyperbolic reflection) in any semicircle centred at one of the ends of
L.
Figure [6.26] reproduces the essential elements of [6.25]. In order to find the h-
length D of the arc pq, let us apply the h-reflection z ÞÑ e z = RC (z), where C is
the illustrated semicircle that is centred at the end c of M, and that passes through
q. This carries the arc pq into the illustrated vertical line-segment peq. By virtue of
e, i.e., the ratio
(6.35), it only remains to find the ratio of the y-coordinates of q and p
of the Euclidean distances [qm] and [e pm].
From the fact that the radius pm is orthogonal to the circle M it follows [exercise]
that the angle pmc equals Π. It then follows [exercise] that the angle ce pm equals
(Π/2), as illustrated. Thus
   
[qm] [cm]
D = ln = ln = |ln tan(Π/2)| = − ln tan(Π/2),
[e
pm] [e
pm]
where the last equality follows from the fact that tan(Π/2) < 1, because Π is acute.
Thus tan(Π/2) = e−D , as was to be shown.

6.3.7 The Three Types of Direct Motion

As we have pointed out, the “Poincaré upper half-plane” was first discovered by
Beltrami, and in this 25th Anniversary Edition we are biting the bullet (as we already
 Hyperbolic Geometry 349

did in VDGF) calling it instead the Beltrami–Poincaré upper half-plane. What Poincaré
does deserve sole credit for—enormous credit!—is the realization that hyperbolic
geometry is intimately connected with complex analysis. The cornerstone of this
connection is the fact that the (direct) motions of the hyperbolic plane are Möbius
transformations. Let us outline how this comes about.
If L1 and L2 are two h-lines, then the composition
M ≡ RL2 ◦ RL1
of h-reflection in these lines will be a direct motion of the hyperbolic plane. Since
every h-reflection is represented in the map by inversion in a circle, we immediately
deduce that any direct motion of the form M is represented by a (non-loxodromic)
Möbius transformation M(z). Furthermore, later we will show that every direct
motion is of the form M; indeed, we will even give an explicit geometric construc-
tion for decomposing an arbitrary direct motion into two h-reflections. Supposing
this already done, we see that every direct motion is represented as a (non-loxodromic)
Möbius transformation.
Conversely, suppose that M(z) is an arbitrary Möbius transformation that maps
the upper half-plane to itself. Then it follows that M(z) must map the real axis (the
horizon) into itself. But a loxodromic Möbius transformation cannot possess such
an invariant line: its strangely shaped invariant curves were illustrated in [3.32] on
p. 188. Thus M(z) is non-loxodromic, and from (3.49), p. 198, we deduce that M(z)
is the composition of inversion in two circles orthogonal to the real axis. Thus the
most general Möbius transformation of the upper half-plane to itself represents a direct
hyperbolic motion of the type M above.
One way to discover the algebraic form of these Möbius transformations is to
use the formula (3.4), p. 141: inversion in a circle K centred at the point q on the
real axis, and of radius R, is given by
q z + (R2 − q2 )
IK (z) = .
z−q
Composing two such functions, we find [exercise] that a motion of type M corre-
sponds to a Möbius transformation
az + b
M(z) = , where a, b, c, d are real, and (ad − bc) > 0. (6.39)
cz + d
Recall that in Ex. 25, p. 212, you showed that this is the form of the most general
Möbius transformation of the upper half-plane to itself. Thus we have agreement
with the conclusion of the previous paragraph.
So much for the overview—now let’s look in detail at the direct motions M. We
know from [6.24] or [6.25] that there are just three possible configurations for the h-
lines L1 and L2 , and correspondingly M ≡ RL2 ◦ RL2 is one of three fundamentally
different types:
350 Non-Euclidean Geometry*

(i) If the h-lines intersect, then M is called a hyperbolic rotation.

(ii) If the h-lines are asymptotic, then M is a new kind of motion (peculiar to
hyperbolic geometry) called a limit rotation.
(iii) If the h-lines are ultra-parallel, then M is called a hyperbolic translation.

We can now reap the rewards of all our hard work in Chapter 3, for these three
types of motion are just the three types of non-loxodromic Möbius transformation:
(i) h-rotations are the “elliptic” ones; (ii) limit rotations are the “parabolic” ones;
and (iii) h-translations are the “hyperbolic”20 ones. At this point, you might find
it helpful to reread the discussion of these Möbius transformations at the end of
Chapter 3.
We already understand these Möbius transformations, so it only remains to look
at them afresh, through hyperbolic spectacles. That is, imagine that you belong
to the race of Beltrami–Poincarites—tiny, intelligent, two-dimensional beings who
inhabit the hyperbolic plane. To you and your fellow Beltrami–Poincarites, h-lines
really are straight lines, the real axis really is infinitely far away, etc. What will you
see if the above motions are applied to your world?
Let us begin with h-rotations. Figure [6.27] illustrates the elliptic Möbius trans-
formation—let’s call it Rϕ a —that arises in the case where the h-lines intersect at a,
and the angle from L1 to L2 is (ϕ/2). [We have chosen to illustrate ϕ = (π/3).] Thus
Rϕ iϕ
a has fixed points a and a, and the multiplier associated with a is m = e . As in

[6.27] A Hyperbolic Rotation, results from reflecting across intersecting h-lines, L1 and L2 .

20
Try not to be confused by this unrelated use of the word “hyperbolic”.
 Hyperbolic Geometry 351

Chapter 3, each shaded “rectangle” is mapped by Rϕ a to the next one in the direction
of the arrows—some of these regions have been filled with black to emphasize this.
Consider how all this looks to you and your fellow Beltrami–Poincarites. For
example, you see each black “rectangle” as being exactly the same shape and size
as every other. To understand Rϕ a better, we begin by noting that (in terms of the
map) its effect on an infinitesimal neighbourhood of a is just a Euclidean rotation
of ϕ about a. But since the map is conformal, this implies that a Poincarite standing
at a will also see his immediate neighbourhood undergoing a rotation of ϕ.
More remarkably, however, the Poincarite at a will see the entire hyperbolic
plane undergoing a perfect rotation of ϕ. Every h-line segment ap he constructs
emanating from a is transformed by z ÞÑ e z = Rϕa (z) into another h-line segment
aep of equal length, making angle ϕ with the original. If the Poincarite gradually
increases ϕ from 0 to 2π, then he sees pe tracing out an h-circle centred at a, while in
the map we see p e miraculously tracing out a Euclidean circle! Thus the illustrated
Euclidean circles orthogonal to the h-lines through a are all genuine hyperbolic cir-
cles, and a is their common h-centre. Let us record this remarkable result, adding
a detail that is not too hard to prove [exercise]:

Every h-circle is represented in the map by a Euclidean circle, and its

h-centre is the intersection of any two h-lines orthogonal to it. Alge-
braically, the h-circle with h-centre a = (x + iy) and h-radius ρ is
represented by the Euclidean circle with centre (x + iy cosh ρ) and
radius y sinh ρ.

As a stepping stone to the limit rotations, [6.28] introduces a new type of curve
in the hyperbolic plane. On a line L in Euclidean geometry, let p be a fixed point,
let a be a moveable point, and let C be the circle centred at a that passes through p.
If we let a recede to infinity along L, then the limiting form of C is a line (through p
and perpendicular to L). Figure [6.28a] shows that it’s a different story in the hyper-
bolic plane. As a recedes towards the infinitely remote point A on the real axis, the
limiting form of C is a (Euclidean) circle that touches the real axis at A. This is nei-
ther an ordinary h-circle, nor an h-line: it is a new type of curve called a horocycle.
Figure [6.28b] shows that horizontal (Euclidean) lines are also horocycles. Note that
if K is any circle centred at A then the h-reflection RK = IK transforms [6.28a] into
[6.28b]. Thus the Beltrami–Poincarites cannot distinguish between these two types
of horocycle.
Now consider [6.29], which illustrates the parabolic Möbius transformation that
results from h-reflection in h-lines L1 and L2 that are asymptotic at A. Referring
to [6.27] and [6.28], you can now understand why this is called a limit rotation: it
may be viewed as the limit of the h-rotation Rϕ a as a tends to the point A on the
horizon. Note some of the interesting features of this picture: the invariant curves
are horocycles touching at A; each such horocycle is orthogonal to every h-line that
352 Non-Euclidean Geometry*

[6.28] Horocycles. [a] Remarkably, an h-circle C appears in the map as an ordinary

Euclidean circle, but its centre a is the intersection of any two h-lines that meet C orthog-
onally. Keeping p fixed, let us keep increasing the h-radius of C indefinitely. The centre
a gradually moves toward the point A on the horizon, and the circle ultimately becomes
the illustrated horocycle. [b] If we instead let the centre a move upward indefinitely, the
limiting horocycle takes the form of a horizontal line. If K is any circle centred at A then
the h-reflection RK = IK transforms the first kind of horocycle in [a] into this new one,
so inhabitants of the hyperbolic plane cannot tell them apart!

ends at A; and any two such horocycles cut off the same h-length on every h-line
that ends at A.
In terms of the map, the simplest limit rotation occurs when the asymptotic h-
lines L1 and L2 are represented as vertical Euclidean half-lines, say separated by
Euclidean distance (α/2). In this case, M = (RL2 ◦ RL1 ) is represented in the map
by the composition of two Euclidean reflections in parallel lines. Thus M is just a
Euclidean translation z ÞÑ (z + α) of the upper half-plane, and the invariant curves
are horizontal lines, which are again horocycles, but now of the form shown in
[6.28b]. Note that this Euclidean translation is not an h-translation. This is particu-
larly clear if we visualize the effect of M on the pseudosphere, where it becomes a
rotation through angle α about the pseudosphere’s axis.
Figure [6.30] illustrates the third and final type of motion, the h-translation
(hyperbolic Möbius transformation) resulting from h-reflection in two ultra-
parallel h-lines. First note that there is precisely one h-line L that is orthogonal
to both L1 and L2 . Unlike a Euclidean translation, this h-line L is the only h-line
that is mapped into itself; it is called the axis of the h-translation. Despite this dif-
ference, the name “h-translation” is appropriate, for every point on the h-line L
is moved the same h-distance (say δ) along L. If we assume that the axis L has
a direction assigned to it, then we may unambiguously denote this h-translation
by TLδ .
In Euclidean geometry, the invariant curves of a translation are the parallel lines
in the direction of the translation. However, [6.30] shows that the invariant curves
of TLδ are not h-lines, but rather arcs of Euclidean circles connecting the ends e1 and
 Hyperbolic Geometry 353

[6.29] A Hyperbolic Limit Rotation results from reflecting across asymptotic h-lines, L1
and L2 . This transformation has no analogue in Euclidean geometry, but it can be thought
of as the limit of the h-rotation Rϕ
a as a tends to the point A on the horizon. Note that
the invariant curves are horocycles.

e2 of L. These are called the equidistant curves of L, because every point on such a
curve is the same h-distance from the h-line L. Make sure you can see this.
In terms of the map, the simplest h-translation occurs when the ultra-parallel h-
lines L1 and L2 are represented by concentric Euclidean semicircles, say centred at
the origin for convenience. In this case, the two h-reflections (i.e., inversions) yield
a central dilation z ÞÑ kz, where k is the real expansion factor. The axis of this h-
translation is the vertical line through the origin (the y-axis), and the equidistant
curves are all other (Euclidean) lines through the origin (cf. [6.20] and [6.21]). Note
that this Euclidean expansion is a similarity transformation of the map, but it is not
a similarity transformation of the hyperbolic plane—there are none!
Having completed our survey of these three types of direct motion, it’s impor-
tant to note that they not only look very different in terms of their effect on
the map, but they also have unique fingerprints in terms of the intrinsic hyper-
bolic geometry. To put this another way, Beltrami–Poincarites can tell these motions
apart. For example, of the three, only h-rotations have invariant h-circles, and only
h-translations have an invariant h-line.

6.3.8 Decomposing an Arbitrary Direct Motion into Two Reflections

We will now show that the h-rotations, limit rotations, and h-translations are the
only direct motions of the hyperbolic plane. That is, an arbitrary direct motion M
can always be decomposed into two h-reflections: M ≡ (RL2 ◦ RL1 ).
The first step is a familiar lemma: an arbitrary hyperbolic motion M (not necessar-
ily direct) is uniquely determined by its effect on any three non-collinear points. As in
354 Non-Euclidean Geometry*

[6.30] A Hyperbolic Translation results from reflecting across non-intersecting (ultra-

parallel) h-lines, L1 and L2 . Unlike Euclidean geometry, there is only one h-line L
orthogonal to L1 and L2 , and it is the only invariant h-line. The other invariant curves
are arcs of Euclidean circles connecting the ends e1 and e2 , and they are not h-lines.
They are called equidistant curves, because every point on such a curve is the same
h-distance from the h-line L.

Euclidean geometry, this will be established if we can show that the location of
a point p is uniquely determined by its h-distances from any three non-collinear
points a, b, c. Consider [6.31a], in which we have supposed (for simplicity’s sake
only) that the h-line L through a and b is represented by a vertical line in the map.
Through the point p, draw h-circles centred at a, b, and c. Since c does not lie on L
(by assumption), we see that p is the only point at which the three circles intersect.
Done.
Now suppose that an arbitrary motion carries two points a and b to the points
a and b ′ in [6.31b]. By the above result, the motion will be determined once we
′

know the image of any third point p not on the line L through a and b. Draw-
ing the illustrated h-circles with h-centres a ′ and b ′ and with h-radii H{a, p} and
H{b, p}, we see that the two intersection points p ′ and p e are the only possible
images for p. Furthermore, since the h-line L ′ through a ′ and b ′ is necessar-
ily orthogonal to the h-circles centred at those points, we also see that p ′ and
e are symmetric with respect to L ′ , i.e., p
p e = IL ′ (p ′ ) = RL ′ (p ′ ). Thus we have
shown that

There is exactly one direct motion M (and exactly one opposite motion
e that maps a given h-line segment ab to another h-line segment
M)
e = (RL ′ ◦ M), where L ′ is (6.40)
a ′ b ′ of equal h-length. Furthermore, M
′ ′
the h-line through a and b .
 Hyperbolic Geometry 355

[6.31] An arbitrary hyperbolic motion M is uniquely determined by its effect on any

three non-collinear points. [a] The location of a point p is uniquely determined by its
h-distances from any three non-collinear points a, b, c. [b] Let M be an arbitrary motion,
and let a ′ = M(a) and b ′ = M(b). Then M will be determined once we know the
image of any third point p not on the line L through a and b. Drawing the illustrated
h-circles with h-centres a ′ and b ′ and with h-radii H{a, p} and H{b, p}, we see that the
two intersection points p ′ and pe are the only possible images for p, and that these two
possibilities are simply reflections of each other in L ′ . This completes the proof of (6.40).

We shall now give an explicit geometric construction for decomposing an arbitrary

direct motion M into two h-reflections. First note that (6.40) implies that M is deter-
mined by its effect on any two points, no matter how close together they are. Though
it is not essential, the following construction is particularly clear if we choose the
points to be infinitesimally separated.
Let us therefore take the two given points to be z and (z + dz), and their given
images under M to be w = M(z) and (w+dw) = M(z+dz). Figure [6.32] illustrates
this idea. Our task is to find two h-reflections that will simultaneously carry z to w,
and dz to dw. [Incidentally, since M must be conformal, it can be thought of as an
analytic function, so we may write dw = M ′ (z) dz.]
First, carry z to w using the h-translation TLδ , where δ = H{z, w}, and L is the
unique h-line from z to w. Note that since TLδ is conformal, it carries dz to an
infinitesimal vector d e z (of equal h-length) making the same angle with L as dz.
Next, apply the h-rotation Rθw , where θ is the angle from d e z to dw. This leaves w
where it is, and it rotates d e
z to dw. Since the net transformation carries z to w, and
dz to dw, it must be M:

M = Rθw ◦ TLδ .

Implicitly, this formula decomposes M into four h-reflections, because TLδ and
Rθw can both be decomposed into two h-reflections. However, [6.32] illustrates that
we can always arrange for two of the four h-reflections to cancel. Defining m to
356 Non-Euclidean Geometry*

[6.32] Decomposing an arbitrary direct motion M into two h-reflections. Imagine two
points close together, z and z + dz, being mapped to their images w and w + dw. First
carry z to w with the unique h-translation TLδ (where δ = H{z, w}) along the h-line from
z to w. This carries dz conformally to de
z, still making the same angle with L. Now rotate
dez about w by θ to obtain dw. Thus, M = Rθw ◦ TLδ . This implicitly decomposes M
into four reflections, but we can always arrange for the middle two to cancel. Here,
M = (RC ◦ RB ) ◦ (RB ◦ RA ) = RC ◦ RA = Rϕ a , but clearly this construction may just as
easily yield an A and a C that are asymptotic or ultra-parallel, in which case M is a limit
rotation or an h-translation.

be the h-midpoint of the h-line segment zw, draw h-lines A and B orthogonal to
L and passing through m and w, respectively. Then TLδ = (RB ◦ RA ). If we now
draw an h-line C through w making angle (θ/2) with B, then Rθw = (RC ◦ RB ).
Thus, as we set out to show, every direct motion can be decomposed into two
h-reflections:

M = (RC ◦ RB ) ◦ (RB ◦ RA ) = RC ◦ RA .

In the illustrated example, it so happens that the h-lines A and C intersect, and
so the motion is an h-rotation: M = Rϕ a , where a is the intersection of A and C, and
(ϕ/2) is the angle between them. However, it is clear that this construction may
just as easily yield an A and a C that are asymptotic or ultra-parallel, in which case
M is a limit rotation or an h-translation.
Summarizing what we have shown, and recalling (6.39),

Every direct motion of the hyperbolic plane is the composition of two

h-reflections, and is thus an h-rotation, a limit rotation, or an h-
translation. In the Beltrami–Poincaré upper half-plane, all such motions
are presented by Möbius transformations of the form
 Hyperbolic Geometry 357

az + b
M(z) = , where a, b, c, d are real, and (ad − bc) > 0.
cz + d
Finally, returning to [6.32] and appealing to (6.40), the unique opposite motion
e carrying z to w and dz to dw is given by three h-reflections:
M
e = R L ′ ◦ RC ◦ RA .
M

Here L ′ is the illustrated h-line passing through w and (w + dw), i.e., passing
through w in the direction dw. This decomposition does not, however, yield the
e for that, and for the formula describing the
simplest geometric interpretation of M;
general opposite motion, see Ex. 24.

6.3.9 The Angular Excess of a Hyperbolic Triangle

Joining three points in the hyperbolic plane with h-line segments yields (by
definition) a hyperbolic triangle. Our objective will be to show that the angular excess
E(T ) of such a hyperbolic triangle T is given by

E(T ) = (−1)A(T ). (6.41)

As we pointed out in the Introduction, this says (amongst other things) that the
angles of ∆ always add up to less than π, and that no matter how large we make
∆, its area can never exceed π. Referring to the differential geometry result (6.6),
we also see that in establishing this formula we will have provided an intrinsic21
proof of the fact that the hyperbolic plane is a surface of constant negative curvature
K = −1.
We have already remarked that Christiaan Huygens investigated the pseu-
dosphere as early as 1693, and to get acquainted with hyperbolic area we will
now confirm one of his surprising results: the pseudosphere has finite area. In
the upper half-plane the pseudosphere is represented by the shaded region
{0 ⩽ x < 2π, y ⩾ 1} shown in [6.20], and (6.32) implies that this region of infinite
Euclidean area does indeed have finite hyperbolic area:
ZZ Z 2π Z ∞ Z 2π Z∞
dx dy dy
A(pseudosphere) = dA = 2
= dx 2
= 2π,
x=0 y=1 y x=0 y=1 y

as Huygens discovered.
Figure [6.33a] illustrates a triangle on the pseudosphere. If the uppermost ver-
tex moves up the pseudosphere indefinitely, then the angle at that vertex tends to
zero, and the edges meeting at that vertex tend to asymptotic lines, namely, trac-
trix generators meeting at infinity. Such a limiting triangle, two of whose edges
are asymptotic, is called an asymptotic triangle. In order to establish (6.41) for ordi-
nary triangles, we first establish it for asymptotic triangles. Figure [6.33b] illustrates

21
Recall that earlier we used the pseudosphere to give an extrinsic proof.
358 Non-Euclidean Geometry*

[6.33] [a] An asymptotic triangle results when the top vertex moves up the pseudosphere
indefinitely. [b] The angular excess of an asymptotic triangle. Suppose the finite edge of
the asymptotic triangle T is an arc of the unit circle. A simple calculation [see text] then
shows that A(T ) = π − α − β. Taking the third angle of T to be zero, this indeed accords
with E(T ) = (−1)A(T ).

such a triangle T in the upper half-plane, the asymptotic tractrix generators becom-
ing vertical half-lines. By Huygens’ result, T clearly has a finite area A(T ), and
because the asymptotic edges meet at angle zero, the result we wish to establish
is A(T ) = (π − α − β).
To simplify the derivation of this result, [6.33b] supposes that the finite edge of
T is an arc of the unit circle. This does not involve any loss of generality, because
an arc of a circle of radius r centred at x = X may be transformed into an arc of the
unit circle by applying the limit rotation z ÞÑ (z − X), followed by the h-translation
z ÞÑ (z/r). From [6.33b] we now deduce that
Z cos β Z ∞  Z cos β
dy dx
A(T ) = √ 2
dx = √ ,
x=cos(π−α) y= 1−x2 y x=cos(π−α) 1 − x2

and writing x = cos θ then yields the desired result:

Zβ
− sin θ dθ
A(T ) = = π − α − β.
π−α sin θ

On the left of [6.34] is a general triangle, say of area A. By applying a suitable

h-rotation about one of the vertices, we can bring one of the edges into a vertical
position, as illustrated on the right of [6.34]. This makes it clear that the area A of the
triangle may be viewed as the difference of the areas of two asymptotic triangles:
one with angles α and (β+θ); the other with angles (π−γ) and θ. Finally, applying
the above result for asymptotic triangles, we deduce (6.41):
 Hyperbolic Geometry 359

[6.34] The angular excess of a general hyperbolic triangle. [a] A general hyperbolic trian-
gle of area A can be rotated so that one edge becomes vertical. [b] But now A is clearly
the difference of the areas of two asymptotic triangles: one with angles α and (β + θ); the
other with angles (π − γ) and θ. It follows immediately that E(T ) = (−1)A(T ), thereby
confirming that the hyperbolic plane has constant negative Gaussian curvature K = −1.

A = [π − α − (β + θ)] − [π − (π − γ) − θ]
= π−α−β−γ
= −E.

6.3.10 The Beltrami–Poincaré Disc

In addition to the upper half-plane model, Beltrami (1868b) constructed another
extremely useful conformal map of the hyperbolic plane, this time inside the unit
disc. Fourteen years later Poincaré rediscovered this map, which is now universally
(and unfairly) known as the Poincaré disc. Following the example we set ourselves
in VDGF, let us now attempt to set the record straight by given both men equal
credit, calling it the Beltrami–Poincaré disc.
Figure [6.35a] illustrates the first step of the construction, which is to map the
entire upper half-plane into the unit disc by means of the inversion

z ÞÑ e
z = IK (z),

where K is the illustrated circle centred at −i and passing through ±1. In order for
this disc to represent the hyperbolic plane, its metric must be inherited from the
upper half-plane. That is, we must define the h-separation H{e e of two points in
a, b}
the disc to be the h-separation H{a, b} of their preimages in the upper half-plane.
Note that this implies [exercise] that the h-lines of the disc are precisely the images
of h-lines in the upper half-plane.
360 Non-Euclidean Geometry*

[6.35] Mapping the Beltrami–Poincaré Half-Plane to the Beltrami–Poincaré Disc.

[a] The anticonformal inversion z ÞÑ e z = IK (z) maps the upper half-plane to the interior
of the unit disc. To retain the hyperbolic geometry, we define the h-separation H{e e
a, b}
of two points in the disc to be the h-separation H{a, b} of their preimages in the upper
half-plane. H-lines now become arcs of circles meeting the unit circle at right angles, and
the horizon is now the unit circle itself, also known as the circle at infinity. [b] Composing
this inversion with conjugation, we recover a conformal mapping to the unit disc, which
we call the Beltrami–Poincaré disc model of the hyperbolic plane, in which h-lines now
take the form of arcs of circles orthogonal to the horizon, such as L, A, U.

Before moving on, try staring at [6.35a] until the following details become clear:
(i) ±1 remain fixed and i is mapped to 0; (ii) the entire shaded part of the upper
half-plane is mapped to the shaded bottom half of the unit disc; (iii) the remaining
part of the upper half-plane (i.e., the top half of the unit disc) is mapped into itself;
(iv) h-lines in the disc are the images of h-lines in the upper half-plane, and these are
arcs of circles orthogonal to the unit circle; (v) the entire horizon of the hyperbolic
plane is represented by the unit circle, with the common point at infinity of vertical
h-lines in the upper half-plane being represented by −i.
At this point we have obtained a map of the hyperbolic plane within the unit
disc. However, since IK (z) is anticonformal, so is our map: angles in the upper
half-plane are currently represented by equal but opposite angles in the disc. If we
now apply z ÞÑ z, which reflects the disc across the real axis into itself, then angles
are reversed a second time, and we obtain the conformal Beltrami–Poincaré disc.
The net transformation from the Beltrami–Poincaré upper half-plane to the
Beltrami–Poincaré disc is thus the composition of z ÞÑ IK (z) and z ÞÑ z, and this is
a Möbius transformation, say D(z). Since D(z) maps i to 0 and −i to ∞, it is clear
that D(z) must be proportional to (z − i)/(z + i). Finally, recalling that a Möbius
transformation is uniquely determined by its effect on three points, and noting that
±1 remain fixed, we deduce [exercise] that
 Hyperbolic Geometry 361

iz + 1
D(z) = . (6.42)
z+i
Alternatively, this may be derived by brute force [exercise] using the formula for
inversion, (3.4), p. 141.
Since D(z) preserves angles and circles, it is easy to transfer the basic types of
curve in the hyperbolic plane from the Beltrami–Poincaré upper half-plane to the
Beltrami–Poincaré disc. Figure [6.35b] illustrates that h-lines are represented by
arcs of circles orthogonal to the unit circle (such as L, A, U), including diameters
such as I. Incidentally, since the horizon is now represented by the unit circle, you
can understand why the horizon is also called the circle at infinity.
The terminology for h-lines is the same as before: I intersects L, A is asymptotic
to L, U is ultra-parallel to L, and a Euclidean circular arc E connecting the ends of
L is an equidistant curve of L. It is also easy to see that a Euclidean circle C lying
strictly inside the unit disc represents an h-circle, though its h-centre a does not
generally coincide with its Euclidean centre. Finally, the horocycles in [6.28a] and
[6.28b] are represented in the Beltrami–Poincaré disc by circles such as H that touch
the unit circle.
Now let us find the metric in the Beltrami–Poincaré disc. Ex. 19 shows how this
may be done by brute calculation, but the following geometric approach22 is much
more enlightening and powerful. First, [6.36a] recalls the earlier observation (6.33):
if ds is the infinitesimal Euclidean length of a horizontal line-element emanating
from z, then the angle between L and E is its hyperbolic length d b s = [ds/ Im(z)].
Note that in purely hyperbolic terms, L is an h-line orthogonal to ds, and E is an
equidistant curve of L. If we apply an h-rotation Rϕ z then L is carried into another
h-line L , and E is carried into an equidistant curve E ′ of L ′ , and the angle between
′

L ′ and E ′ is the same as before. Thus we have the following general construction:

Through one end of ds, draw the h-line l orthogonal to ds, and through the
other end of ds draw the equidistant curve e. Then the h-length db
s of ds is (6.43)
the angle of intersection (on the horizon) of l and e.

Now the beauty of interpreting db s as an angle in this way is that the Möbius
transformation D to the Beltrami–Poincaré disc is conformal, and so the above
construction of db s is valid there too!
Figure [6.36b] illustrates an infinitesimal disc of Euclidean radius ds centred at
z = r eiθ in the Beltrami–Poincaré disc. Because the map is conformal, the h-length
dbs of ds is independent of the direction of ds, so we may simplify the construction
(6.43) by choosing ds orthogonal to the diameter l through z. The equidistant curve
e is then the illustrated arc of a Euclidean circle through the ends of l.

22
We merely rediscovered this angular interpretation of hyperbolic distance, which we believe
originates with Thurston (1997). However, our explanation (and our applications) of the idea differ
somewhat from Thurston’s.
362 Non-Euclidean Geometry*

[6.36] Geometric derivation of the disc metric. [a] If ds is the infinitesimal Euclidean
length of a horizontal line-element emanating from z, then the angle between the orthog-
onal h-line L and the equidistant curve E is its hyperbolic length d b s = [ds/ Im(z)]. But
performing a rotation about z we see that d b s may equally be measured as the angle
between the h-line L ′ and its equidistant curve E ′ . [b] Since the mapping from the half-
plane to the disc is conformal, this angular interpretation of hyperbolic distance applies
here, too! Choose ds orthogonal to the diameter l through z. The equidistant curve e
is then the illustrated arc of a Euclidean circle of radius ρ through the ends of l, and
ρdb s = 1. [c] AB = A ′ B ′ . [d] So, 2ρ ds = (1 − r)(1 + r) = 1 − |z|2 , and the hyperbolic
disc metric is therefore d b 2
s = 1−|z| 2 ds.

To turn this picture of d b

s into a formula, begin by noting that if ρ is the radius
of the circle containing the arc e, then [draw a picture!]

ρdb
s = 1.

Next recall (or prove) the familiar property of circles illustrated in [6.36c], namely,
that all chords passing through a fixed interior point are divided into two parts
whose lengths have constant product: AB = A ′ B ′ . Applying this result to the copy
of [6.36b] shown in [6.36d], we obtain [exercise]

2ρ ds = (1 − r)(1 + r) = 1 − |z|2 .
 Hyperbolic Geometry 363

Thus the metric of the Beltrami–Poincaré disc is

2
db
s = ds. (6.44)
1 − |z|2
Note the remarkable similarity to the formula for the metric of the sphere under
stereographic projection, (6.16)!
Since the Euclidean line-segment connecting 0 to z is also an h-line segment,
we can now find the h-separation of these points by simple integration along the
line-segment:
Z |z| Z |z|  
2 dr 1 1
H{0, z} = 2
= + dr,
0 1−r 0 1+r 1−r
and so
 
1 + |z|
H{0, z} = ln . (6.45)
1 − |z|
As a simple check of this formula, note that as z moves toward the unit circle (the
horizon), H{0, z} tends to infinity, as it should.

6.3.11 Motions of the Beltrami–Poincaré Disc

In the upper half-plane we found that every direct motion was the composition
of two h-reflections, and every opposite motion was the composition of three h-
reflections. Since the intrinsic geometry of the Beltrami–Poincaré disc is identical
to the upper half-plane, this result must still be true, so it only remains to find
out what h-reflection means in the disc. In the upper half-plane we saw that h-
reflection in an h-line K meant geometric inversion in K, and the same is true in the
Beltrami–Poincaré disc!
This is easy to understand. In the upper half-plane, q is the h-reflection of p
in K means that p and q are symmetric (in the sense of inversion) in K. In order
to make the Beltrami–Poincaré disc isometric to the upper half-plane, we insisted
that the mapping z ÞÑ e z = D(z) preserve hyperbolic distance. In particular, qe is
e
e in K. But D(z) is a Möbius transformation, and so the Symmetry
the h-reflection of p
e and q
Principle [see p. 168] implies that p e as was to be shown.
e are symmetric in K,
Thus every direct motion M of the Beltrami–Poincaré disc has the form

M = RL2 ◦ RL1 = IL2 ◦ IL1 ,

where L1 and L2 are h-lines, namely, arcs of circles orthogonal to the unit circle. As
in the upper half-plane, every direct motion is therefore a non-loxodromic Möbius
transformation. We already know that there are just three hyperbolically distin-
guishable types of direct motion, and the distinction between them in terms of L1
and L2 is the same as before: we get an h-rotation when they intersect, a limit rota-
tion when they are asymptotic, and an h-translation when they are ultra-parallel.
364 Non-Euclidean Geometry*

[6.37] A Hyperbolic Rotation results from reflecting across intersecting h-lines, L1 and L2 .
[a] The typical case. [b] If the rotation is about the centre, then it reduces to an ordinary
Euclidean rotation.

We will discuss the formula for these Möbius transformations in a moment, but
first let’s draw pictures of them.
Figure [6.37a] shows a typical h-rotation; note the appearance of h-circles with a
common h-centre. Figure [6.37b] illustrates the pleasant fact that if L1 and L2 inter-
sect at the origin (in which case they are Euclidean diameters) then the resulting
h-rotation manifests itself as a Euclidean rotation.
In this connection, we offer a word of warning. As Euclidean beings, we suffer
from an almost overwhelming temptation to regard the centre of the Beltrami–
Poincaré disc as being special in some way. One must therefore constantly remind
oneself that to the Beltrami–Poincarites who inhabit the disc, every point is indis-
tinguishable from every other point. In particular, the Beltrami–Poincarites do not
see any difference between [6.37a] and [6.37b].
Figure [6.38a] illustrates a typical limit rotation generated by an L1 and an L2
that are asymptotic at a point A on the horizon. Once again note that the invariant
curves are horocycles touching at A, and that these are orthogonal to the family of
h-lines that are asymptotic at A.
Finally, [6.38b] illustrates a typical h-translation. Once again, note that there is
precisely one invariant h-line [shown in bold], and that the invariant equidistant
curves are arcs of circles through the ends of this axis.
From our work in the upper half-plane we know that the three types of motion
pictured above are the only direct motions of the Beltrami–Poincaré disc, and we
now turn to the formula that describes them. We know that every direct motion
is a Möbius transformation that maps the unit disc into itself, and at the end of
 Hyperbolic Geometry 365

[6.38] [a] A Limit Rotation results from reflecting across asymptotic h-lines,
L1 and L2 . [b] A Translation results from reflecting across ultra-parallel h-lines,
L1 and L2 .

Chapter 3, with malice aforethought, we investigated these “Möbius automor-

phisms” of the unit disc. We found [see (3.52), p. 201] that the formula representing
the most general one Mϕ a (z) is
z−a
Mϕ iϕ
a (z) = e Ma (z), where Ma (z) = .
az − 1
Thus Mϕ a is the composition of Ma and a rotation of ϕ about the origin.
Recall that Ma swaps a and 0: Ma (a) = 0 and Ma (0) = a. More generally,
Ma swaps every pair of points z, Ma (z): the transformation is involutory. This is
explained by [6.39a], which recalls the result illustrated in [3.39], p. 203:
Ma = IB ◦ IA ,
where B is the diameter through a, and where A is the circle centred at (1/a) that
is orthogonal to the unit circle.
Hyperbolic geometry gives us a fresh perspective on this result: the intersection
point m of A and B is the h-midpoint of 0 and a, and A itself is the perpendicular
h-bisector of 0a. Furthermore, the inversions in A and B are h-reflections. Thus
Ma is the composition of two h-reflections in perpendicular h-lines through m,
and so
The unique Möbius automorphism Ma that swaps a and 0 is the h-rotation
Rπ
m through angle π about the h-midpoint m of the h-line segment 0a.

An immediate benefit of this insight is that we can now easily find the formula
for the h-separation of any two points, a and z. The h-rotation Ma brings a to the
origin, and we already know the formula (6.45) for the h-distance of a point from
there:
366 Non-Euclidean Geometry*

 
1 + |Ma (z)|
H{a, z} = H{Ma (a), Ma (z)} = H{0, Ma (z)} = ln ,
1 − |Ma (z)|
and so
 
|a z − 1| + |z − a|
H{a, z} = ln . (6.46)
|a z − 1| − |z − a|

Now let us resume and complete our discussion of Mϕ a . As illustrated in [6.37b],

the Euclidean rotation z ÞÑ e z represents the h-rotation Rϕ
iϕ
0 . Thus the most gen-
eral Möbius automorphism of the disc may be interpreted as the composition of
two h-rotations:
ϕ
a = R0 ◦ Rm .
Mϕ π

Figure [6.39b] shows how to compose these h-rotations, using the same idea as
was used in both Euclidean and spherical geometry. The h-rotation Rϕ 0 is the com-
position of h-reflections in any two h-lines through 0 (diameters) containing angle
(ϕ/2). Thus, choosing the first h-line to be B, and calling the second h-line C, we
deduce that

a = (RC ◦ RB ) ◦ (RB ◦ RA ) = RC ◦ RA .
Mϕ

Thus Mϕ a is an h-rotation, limit rotation, or h-translation according as A and C are

intersecting, asymptotic, or ultra-parallel.

[6.39] Geometric interpretation and classification of the general Möbius automorphism,

Mϕ a . [a] The unique Möbius automorphism Ma = IB ◦ IA that swaps a and 0 is the
h-rotation Rπm through angle π about the h-midpoint m of the h-line segment 0a. [b] The
most general Möbius automorphism is Mϕ ϕ
a = R0 ◦ Rm = IC ◦ IB ◦ IB ◦ IA = IC ◦ IA . Thus
π
ϕ
Ma is an h-rotation, limit rotation, or h-translation according as A and C are intersecting,
asymptotic, or ultra-parallel.
 Hyperbolic Geometry 367

Thinking of a as fixed and ϕ as variable, the critical value ϕ = Φ separating

the h-rotations from the h-translations occurs when C is in the position C ′ [shown
dashed] asymptotic to A at p. It is not hard to see [exercise] that the triangle pa0 is
right angled, and so it follows that cos(Φ/2) = |a|, or
Φ = 2 cos−1 |a|.
This explains the result (3.54), p. 204, which you proved algebraically in Ex. 27,
p. 212.
To sum up,
The most general Möbius automorphism Mϕ a of the disc is a direct
hyperbolic motion, and it is (i) an h-rotation if ϕ < Φ; (ii) a limit
rotation if ϕ = Φ; and (iii) an h-translation if ϕ > Φ.
Finally, recall from Ex. 20, p. 210, that the set of Möbius transformations of the
form Mϕ a is identical to the set of the form
Az + B
M(z) = , where |A| > |B|.
Bz + A
Comparing this with (6.20), we see that there are striking similarities not only
between the conformal metrics of the sphere and the hyperbolic plane, but also
between the Möbius transformations that represent their direct motions.

6.3.12 The Hemisphere Model and Hyperbolic Space

Figure [6.40a] illustrates how we may obtain two new models of the hyperbolic
plane. Following Beltrami (1868b), let us stereographically project the Beltrami–
Poincaré disc from the south pole S of the Riemann sphere onto the northern
hemisphere. Defining the h-separation of two points to be the h-separation of their
preimages in the disc, we have a new conformal map of the hyperbolic plane, called
the hemisphere model. The h-lines of this model are the images of h-lines in the disc,
and since stereographic projection preserves circles as well as angles, we deduce
[exercise] that h-lines are (semi-circular) vertical sections of the hemisphere. What do
equidistant curves and horocycles look like?
The hemisphere was Beltrami’s primary model of the hyperbolic plane, and
it was by applying the above stereographic projection to this hemisphere that he
discovered the Beltrami–Poincaré disc. In fact by projecting his hemisphere in dif-
ferent ways, Beltrami obtained (in a unified way) almost all the models in current
use.
For example, by projecting the hemisphere vertically down onto the complex
plane (see [6.40a]) he obtained a new model of the hyperbolic plane inside the unit
disc. This is traditionally called the “Klein model”, but, again, we shall rename it the
Beltrami–Klein model. Less controversially, it is also called the projective model. Since
a small circle on the hemisphere is clearly projected to an ellipse in the disc, the
368 Non-Euclidean Geometry*

Beltrami–Klein model is not conformal. This is a serious disadvantage, but it is com-

pensated for by the fact that the vertical sections of the hemisphere are projected
to (Euclidean) straight lines: h-lines in the Beltrami–Klein model are straight Euclidean
chords of the unit circle. Note the analogy with figure [6.12], in which geodesics on the
sphere are represented by straight lines in the map; Ex. 14 reveals that this analogy
is more than superficial.
Other properties of the Beltrami–Klein model will be explored in the exercises,
but right now we have bigger fish to fry! Up to this point we have focused on
developing the geometry of the hyperbolic plane, the negatively curved counterpart
of the Euclidean plane. The geometry of this Euclidean plane may be thought of as
being inherited from the geometry of three-dimensional Euclidean space. That is,
if (X, Y, Z) are Cartesian coordinates in this space, then the Euclidean distance ds
between two infinitesimally separated points is given by
p
ds = dX2 + dY 2 + dZ2 ,
and two-dimensional Euclidean geometry is obtained by restricting this formula
to the points of an ordinary plane.
The question therefore arises whether there might exist a negatively curved
(whatever that might mean) counterpart of three-dimensional Euclidean space,
such that the geometry induced on each “plane” within this space would auto-
matically be the geometry of the hyperbolic plane. We shall now show that this
three-dimensional hyperbolic space does indeed exist.
To do so, let us find the metric of the hemisphere model. Because the stereo-
graphic projection of the Beltrami–Poincaré disc onto the hemisphere is conformal,
it follows that dbs is once again given by the construction (6.43). Since db s is inde-
pendent of the direction of ds on the hemisphere, we may once again simplify the
construction by choosing ds in an auspicious direction. In the Beltrami–Poincaré
disc the best choice of ds was orthogonal to the diameter through the point of inter-
est, and the best choice on the hemisphere is simply the stereographic projection of
this configuration.
Thus in [6.40b] we have chosen the h-line l to be the vertical section of the hemi-
sphere passing through the north pole and the point from which ds emanates. Thus
l and e are both halves of great circles: the plane of l is vertical, the plane of e
is inclined at angle d b s to the vertical, and the intersection of these planes is the
illustrated diameter of the unit circle lying directly beneath l.
Now let the coordinates of the point at which we have drawn ds be (X, Y, Z),
where the X and Y axes coincide with the real and imaginary axes of C, so that Z
measures the height of the point above C. Since ds is orthogonal to l, and since
the vertical plane of l is orthogonal to the hemisphere, we see that ds is horizontal.
Thus the angle that ds subtends at the point (X, Y, 0) directly beneath it is (ds/Z).
But this angle is just the angle between the planes of l and e! Thus the metric of the
hemisphere model is
 Hyperbolic Geometry 369

[6.40] [a] The Beltrami Hemisphere Model is obtained by stereographically projecting

the Beltrami–Poincaré disc model from the south pole, S, onto the northern hemisphere.
Since stereographic projection preserves circles and angles, h-lines are semi-circular
vertical sections of the hemisphere. Projecting vertically downwards yields the non-
conformal Beltrami–Klein model (aka projective model) of the hyperbolic plane, in which
h-lines now appear as straight Euclidean chords of the unit circle. [b] Conformality implies
that the angular interpretation of hyperbolic distance given in [6.36] applies here, too.
Therefore, the hyperbolic metric on the hemisphere is d b s = (ds/Z).

ds
db
s = . (6.47)
Z
This formula only describes the h-separation of points on the hemisphere, but
there is nothing preventing us from using it to define the h-separation of any two
infinitesimally separated points in the three-dimensional region Z > 0. This region
lying above C, with h-distance defined by (6.47), is called the half-space model of
three-dimensional hyperbolic space, denoted H3 . Without going into detail, it is clear
from (6.47) that the points of C are infinitely h-distant from points that lie strictly
above C. Thus C represents the two-dimensional horizon or sphere at infinity of
hyperbolic space.
At the moment it is a mere tautology that the geometry induced on the hemi-
sphere by (6.47) is that of a hyperbolic plane. To begin to see that there is real meat
on this idea, let us consider some simple motions of hyperbolic space. Clearly d b s
is unaltered by a translation parallel to C, so this is a motion. It is also clear that
dbs is unaltered by a dilation (X, Y, Z) ÞÑ (kX, kY, kZ) centred at the origin. More
generally, a dilation centred at any point of C will preserve d b
s , so this too is a motion.
By applying these two types of motion to the origin-centred unit hemisphere
that we have been studying, we see that
In the half-space model, H3 , every hemisphere orthogonal to C is a
(6.48)
hyperbolic plane, H2 .
370 Non-Euclidean Geometry*

[6.41] Hyperbolic 3-space, H3 . Hyperbolic planes appear as vertical half-planes and as

hemispheres orthogonal to C. The intersection of two such H2 planes is a hyperbolic line:
a vertical half-line, or a semicircle orthogonal to C.

In Euclidean geometry the intersection of two planes is a line, and this suggests
that an h-line should be the intersection of two hyperbolic planes. Thus we antici-
pate that every semicircle orthogonal to C is an h-line, for every such semicircle is the
intersection of two hemispheres orthogonal to C. Note that this agrees with what
we already know: the h-lines of the hemisphere model are semicircles orthogonal to
C. Figure [6.41] attempts to make this all much more vivid,23 showing the vertical
H2 walls filled with circles that are all the same hyperbolic size; compare this to
[6.20], page 337.
Let us return to two-dimensional geometry for a moment. Figure [6.42] illus-
trates how Beltrami obtained the upper half-plane model from his hemisphere
model. From a point q on the rim of the hemisphere, we stereographically project
onto the tangent plane at the point antipodal to q—actually, any plane tangent to
this one would do equally well. Since this preserves circles and angles, we see that a
typical h-line of the hemisphere is mapped to a semicircle orthogonal to the bottom
edge of the half-plane, while an h-line passing through q is mapped to a vertical
line.

23
This figure was not in the original edition of VCA. I drew it about 15 years later, for VDGF, where
it appears as figure [6.6].
 Hyperbolic Geometry 371

[6.42] How Beltrami discovered the upper half-plane model. From a point q on the rim
of the hemisphere, stereographically project onto the tangent plane at the point antipodal
to q. Since this preserves circles and angles, we see that a typical h-line of the hemisphere
is mapped to a semicircle orthogonal to the bottom edge of the half-plane, while an h-line
passing through q is mapped to a vertical line. The angular interpretation of hyperbolic
distance given in [6.36] applies here, too. We therefore recover the familiar hyperbolic
metric in the upper half-plane: d b s = (ds/Z).

Well, since these are the h-lines, it certainly looks like we have obtained the
Beltrami–Poincaré upper half-plane, but to make sure, let’s check that its metric
is really given by (6.31). Since stereographic projection is conformal, we may yet
again use the construction (6.43). Choosing l to be the image of an h-line through
q, the figure immediately reveals that the metric is d b s = (ds/Z). Apart from a
change of notation, this is indeed the same as (6.31).
We have thus returned to the half-plane that began our journey, but we have
returned wiser than when we left. Looking at (6.47) we now recognize this half-
plane orthogonal to C as a hyperbolic plane within hyperbolic space. This reveals
the true role of the stereographic projection in [6.42].
We know that stereographic projection from q is just the restriction to the hemi-
sphere of inversion IK in a sphere K centred at q. Using the same argument (figure
[6.22b]) as in the plane, we see that IK preserves the metric (6.47), so it is a motion of
hyperbolic space, carrying h-lines into h-lines and carrying h-planes into h-planes.
Furthermore, (6.48) tells us that K is a hyperbolic plane in this hyperbolic space,
and we therefore suspect that IK is reflection in this h-plane. This can be confirmed
[exercise] by generalizing the argument in [6.23b]. Thus we have the following
generalization of (6.38):
Inversion in a hemisphere K orthogonal to the horizon is reflection RK
of hyperbolic space in the h-plane K.
372 Non-Euclidean Geometry*

It is beyond the scope of this book to explore the motions of hyperbolic space24 .
However, let us at least describe one particularly beautiful result.
Just as an arbitrary direct motion of an h-plane is the composition of two h-
reflections in h-lines within it, so an arbitrary direct motion of hyperbolic space
is the composition of four reflections in h-planes within it. Thus, in the half-space
model with horizon C, such a motion is the composition of four inversions in
spheres centred on C. If we restrict attention to the points of C then inversion in
such a sphere K is equivalent to two-dimensional inversion of C in the equato-
rial circle in which K intersects C. Conversely, inversion of C in a circle k extends
uniquely to an inversion of space: simply construct the sphere with equator k.
Finally, then, every direct motion of hyperbolic space can be uniquely repre-
sented in terms of C (the horizon) as the composition of inversion in four circles,
and this is none other than the most general Möbius transformation,
az + b
z ÞÑ M(z) = ,
cz + d
of the complex plane! Poincaré discovered this wonderful fact in 1883; see Poincaré
(1985).
We have seen that the direct motions of the hyperbolic plane, the Euclidean plane
and the sphere are subgroups of this group of general Möbius transformations. As
we shall now see, this fact has a remarkable geometrical explanation.
Hilbert’s result on surfaces of constant negative curvature shows that three-
dimensional Euclidean geometry cannot accommodate a model of the hyperbolic
plane. Amazingly, however, three-dimensional hyperbolic space does contain sur-
faces whose intrinsic geometry is Euclidean! In fact these surfaces are the horospheres
that generalize the horocycles. Analogously to [6.28], horospheres are Euclidean
spheres that touch C, as well as planes Z = const. that are parallel to C.
Vertical planes orthogonal to C look flat in our model of hyperbolic space, but
in reality they are intrinsically curved hyperbolic planes. However, a horosphere
Z = const. not only looks flat, it really is flat. For its metric, inherited from the
metric (6.47) of the surrounding space, is just

db
s = (constant) ds,

and this is the metric of a Euclidean plane!

The motions of Euclidean plane geometry may now be viewed as the motions
of hyperbolic space that map this intrinsically flat horosphere into itself. Clearly,
these are the composition of reflections in vertical planes, i.e., h-reflections in h-
planes orthogonal to the horosphere. In this manner, the direct motions of the
Euclidean plane manifest themselves on the horizon C as a subgroup of the Möbius
transformations.

24
Thurston (1997) is an excellent source of information on these motions.
 Hyperbolic Geometry 373

As for spherical geometry, we begin by defining an h-sphere as the set of points at

constant h-distance from a given point (the h-centre). It is not hard to see that these
h-spheres are represented in the half-space model by Euclidean spheres, though
their h-centres do not coincide with their Euclidean centres.
Though it is not immediately obvious in this model, it can be shown [see Ex. 27]
that the intrinsic geometry of such an h-sphere is the same as that of an ordi-
nary sphere (of different radius) in Euclidean space. As with the horosphere, the
motions of this h-sphere may be viewed as the motions of hyperbolic space that
map the h-sphere into itself. Again, these are the composition of h-reflections in h-
planes orthogonal to the h-sphere, and again we arrive at a subgroup of the Möbius
transformations.
Clearly the motions of the hyperbolic plane may also be viewed in this way, and
so we have arrived at a fitting high point on which to end this chapter:
Two-dimensional hyperbolic, Euclidean, and spherical geometry are all
subsumed by three-dimensional hyperbolic geometry.
Further Reading. For a short, masterful overview of Differential Geometry, see
Penrose (1978). For a complete, elementary, and geometrical treatment of Differential
Geometry (and Differential Forms), in the same spirit as this book, see Needham
(2021). For more on its fundamental applications to physics, particularly General
Relativity, see Misner, Thorne, and Wheeler (1973), Wald (1984), Ludvigsen (1999),
Penrose (2005), Thorne and Blandford (2017), and Schutz (2022). For more on the
use of Differential Forms in Differential Geometry, see do Carmo (1994) and O’Neill
(2006), and for their applications to physics see Schutz (1980), Arnol’d (1989), and
Dray (2015). For more on hyperbolic geometry itself, see Stillwell (1992), Stillwell
(1996), Stillwell (2010), and Thurston (1997). Finally, for the connections between
topology and differential geometry, see Hopf (1956).
374 Non-Euclidean Geometry*

6.4 Exercises

1 Draw a geodesic triangle ∆ on the surface of a suitable fruit or vegetable. Now

draw a geodesic segment from one of the vertices to an arbitrary point of the
opposite side. This divides ∆ into two geodesic triangles, say ∆1 and ∆2 . Show
that the angular excess function E is additive, i.e., E(∆) = E(∆1 ) + E(∆2 ). By
continuing this process of subdivision, deduce that (6.5) implies (6.6).
2 Explain (6.18) by generalizing the argument that was used to obtain the special
case (3.18), on p. 162. That is, think of reflection of the sphere in terms reflection
of space in a plane Π, as in [6.8], p. 318. Also, think of stereographic projection
as the restriction to the sphere of the three-dimensional inversion IK , where K
√
is the sphere of radius 2 centred at the north pole of Σ (see [6.13b]). Now let
a be a point on Σ, and consider the effect of IK on a, Π, and RΠ (a).
b be it stereographic image on Σ. If C
3 Let C be a circle in C, and let C b is a great
circle, then (6.18) says that IC stereographically induces reflection of Σ in Cb,
but what transformation is induced if C is an arbitrary circle? Generalize the
argument of figure [6.14] to show that IC becomes projection from the vertex v
b . That is, if w = IC (z) then w
of the cone that touches Σ along C b is the second
intersection point of Σ with the line in space that passes through the vertex v
and the point b z . Explain how (6.18) may be viewed as a limiting case of this
more general result.
4 Use (3.42), p. 189 to show that if the Möbius transformation M(z) has fixed
points ξ± , and the multiplier associated with ξ+ is m , then

" #−1 " √ #" #

1 −ξ+ m 0 1 −ξ+
[M] = √ .
1 −ξ− 0 1/ m 1 −ξ−

By putting ξ+ = a, m = e−iψ , and ξ− = −(1/a), deduce (6.19), p. 328. [Hint:

Remember that you are free to multiply a Möbius matrix by a constant.]
5 Show that the Möbius transformations (6.20) do indeed satisfy the differential
equation (6.17).
6 (i) The conjugate V of a quaternion V = v + V is defined to be the conjugate
transpose V∗ of the corresponding matrix. Show that V ≡ V∗ = v − V,
and deduce that V is a pure quaternion (analogous to a purely imaginary
complex number) if and only if V = −V.
(ii) The length |V| of V is defined (by analogy with complex numbers) by |V|2 =
V V. Show that |V|2 = v2 + |V|2 = |V|2 .
 Exercises 375

(iii) If |V| = 1, then V is called a unit hquaternion.

i Verify that R ψ
v [see (6.28)] is a
∗
unit quaternion, and that R ψ ψ
v = Rv = R −ψ
v .
(iv) Show that V W = W V and deduce that |V W| = |V| |W|. Thus, for example,
the product of two unit quaternions is another unit quaternion.
(v) Show that A is a pure, unit quaternion if and only if A2 = −1.
(vi) Show that any quaternion Q can be expressed as Q = |Q | R ψ v for some v
and some ψ.
e = Q P Q , where
(vii) Suppose we generalize the transformation (6.29) to P ÞÑ P
Q is an arbitrary quaternion. When interpreted in this way, deduce that Q
represents a dilative rotation of space, and the product of two quaternions
represents the composition of the corresponding dilative rotations. [This
confirms the claim at the end of Chapter 1.]
7 [Do the previous exercise before this one.] The following proof of (6.29) is based
on a paper of H. S. M. Coxeter (1946).
(i) Use (6.27) to show that the pure quaternions P and A are orthogonal if and
only if P A + A P = 0.
(ii) If A has unit length, so that A2 = −1, deduce that the previous equation
may be expressed as P = A P A.
(iii) Now keep the pure, unit quaternion A fixed, but let P represent an arbi-
trary pure quaternion. Let ΠA denote the plane with normal vector A that
·
passes through the origin, so that its equation is P A = 0. Now consider
the transformation
P ÞÑ P ′ = A P A . (6.49)
Show that (a) P ′ is automatically pure, and |P ′ | = |P|, so that (6.49) rep-
resents a motion of space; (b) every point on ΠA remains fixed; (c) every
vector orthogonal to ΠA is reversed. Deduce that (6.49) represents reflection
RΠA of space in the plane ΠA .
(iv) Deduce that if the angle from ΠA to a second plane ΠB is (ψ/2), and the
unit vector along the intersection of the planes is V, then the rotation Rψ
v =
(RΠB ◦ RΠA ) is given by
e = (B A) P (A B) = (−B A) P (−B A).
P ÞÑ P
(v) Use (6.27) to show that −B A = cos(ψ/2) + V sin(ψ/2), thereby simultane-
ously proving (6.29) and (6.28).
8 Here is another proof of (6.29). As in the text, we shall assume that P is a
b on the unit sphere. If we represent the
unit vector with its tip at the point p
376 Non-Euclidean Geometry*

stereographic images p and p e of p b

b and p
e by their homogeneous coordinate
e
vectors p and p in C , then we know that the rotation is represented as
2

e = Rψ
p ÞÑ p v p,
where R ψ
v is being thought of as a 2 × 2 matrix.
(i) Show that in homogeneous coordinates, (3.21), p. 166, becomes
2 p1 p2 |p1 |2 − |p2 |2
X + iY = and Z = .
|p1 |2 + |p2 |2 |p1 |2 + |p2 |2
(ii) To simplify this, recall that all multiples of p describe the same point p in
√
C. We can therefore choose the “length” of p to be 2:
⟨p , p⟩ ≡ |p1 |2 + |p2 |2 = 2.
With this choice, show that the above equations can be written as
" # " # " #
1 + Z X + iY p1 p1 p1 p2 p1
= = [ p1 p2 ] = p p∗ .
X − iY 1 − Z p2 p1 p2 p2 p2
(iii) Verify that
" #
1+Z X + iY
= 1 − iP.
X − iY 1−Z
(iv) Deduce that
h i∗
e = R ψ (1 − i P) R ψ = 1 − i R ψ P R −ψ ,
1 − iP v v v v
from which (6.29) follows immediately.
9 (i) Figure [6.40a] gave a two-step process for carrying a point z in the
Beltrami–Poincaré disc to the corresponding point z ′ in the Klein model.
Explain why the net mapping z ÞÑ z ′ of the disc to itself is the one shown
in figure [a] below, where C is an arbitrary circle passing through z and
orthogonal to the unit circle U.
(ii) Figure [b] is a vertical cross section of [6.40a] through z and z ′ . Deduce that
|z ′ | 1 a 2
= and = .
a b |z| b
By multiplying these two equations, deduce that z ′ = 1+|z|
2z
2.

[Thus we have a geometric explanation of the result (3.21), p. 166.]

(iii) This formula can be derived directly from figure [a], without the assis-
tance of the hemisphere. Redraw the figure with C chosen orthogonal to 0z.
Explain geometrically why the centre of C may be viewed as either IU (z ′ ),
or as the midpoint of z and IU (z). Conclude that
1/ z ′ = IU (z ′ ) = 1
2
[z + IU (z)] = 1
2
[z + (1/z)] ,
from which the result follows immediately.
 Exercises 377

10 Think of the sphere as the surface of revolution generated by a semicircle. Con-

struct a conformal map of the sphere by strict analogy with the construction of
the map of the pseudosphere in [6.20]. Show that this is the Mercator map that
you obtained in Ex. 14, p. 294.
11 (i) In the hyperbolic plane, show that the h-circumference of an h-circle of
h-radius ρ is 2π sinh ρ. [Hint: Represent the h-circle as an origin-centred
Euclidean circle in the Beltrami–Poincaré disc.]
(ii) Let the inhabitants of the sphere of radius R draw a circle of (intrinsic)
radius ρ. Use elementary geometry to show that the circle’s circumference
is 2πR sin(ρ/R). Show that if we take the radius of the sphere to be R = i,
then this becomes the formula in part (i)! [Compare this with Ex. 14.]
12 Let L and M be two intersecting chords of the unit circle, and let l and m be the
intersection points of the pairs of tangents drawn at the ends of these chords.
See figure [a] below. In the Beltrami–Klein model, show that L and M repre-
sent orthogonal h-lines if and only if l lies on M (produced) and m lies on L
(produced).
378 Non-Euclidean Geometry*

13 Let z = r eiθ denote a point in the Beltrami–Klein model of the hyperbolic

plane.
(i) On the hemisphere model lying above the Beltrami–Klein model, sketch
the images under vertical upward projection of some circles r = const.
and some rays θ = const. Although angles are generally distorted in
the Beltrami–Klein model, deduce that these circles and rays really are
orthogonal, as they appear to be. Also, note that the Euclidean circles
r = const. are also h-circles.
(ii) Figure [b] above shows a vertical cross section of the hemisphere model
and Beltrami–Klein model taken though a ray θ = const. If the point z
moves outward along this ray by dr, let ds denote the movement of its ver-
tical projection on the hemisphere. Explain why the two shaded triangles
are similar, and deduce that ds = (dr/Z).
(iii) Use the metric (6.47) of the hemisphere to conclude that the h-separation
dbs r of the points in the Beltrami–Klein model with polar coordinates (r, θ)
and (r + dr, θ) is given by
dr
db
sr = .
1 − r2
[Remarkably, this means that the formula for H{0, z} differs from the
formula (6.45) in the Beltrami–Poincaré disc by a mere factor of two!]
(iv) Use the same idea (of projecting onto the hemisphere) to show that the
h-separation d b
s θ of the points (r, θ) and (r, θ + dθ) is given by
r dθ
dbsθ = √ .
1 − r2
(v) Deduce that the h-separation d b s of the points (r, θ) and (r + dr, θ + dθ) is
given by the following formula of Beltrami (1868a):
dr2 r2 dθ2
db
s2= + . (6.50)
(1 − r2 )2 1 − r2
 Exercises 379

14 (i) The figure above superimposes the stereographic and projective [see
p. 323] images of a great circle on the unit sphere. Let zs and zp be the
stereographic and projective images of the point whose spherical-polar
coordinates are (ϕ, θ). Referring to (3.22), p. 167, zs = cot(ϕ/2) eiθ . Show
that zp = [− tan ϕ] eiθ , and deduce that
2zs
zp = .
1 − |zs |2
Compare this with Ex. 9!

(ii) Sketch the curves on the hemisphere that are centrally projected to the cir-
cles |zp | = const. and to the rays arg zp = const. Although angles are
generally distorted in the projective model, observe that these circles and
rays really are orthogonal, as they appear to be.
(iii) Now let the sphere have radius R, and write zp = r eiθ for the projective image
of the point (ϕ, θ). Thus r = −R tan ϕ. Show that if zp moves a distance
dr along the ray θ = const., then the corresponding point on the sphere
moves a distance d b s r given by
dr
db
sr = .
1 + (r/R)2
(iv) Likewise, show that the separation d b s θ of the points on the sphere corre-
sponding to the points (r, θ) and (r, θ + dθ) in the map is given by
r dθ
dbsθ = p .
1 + (r/R)2
(v) Deduce that the spherical separation d b s corresponding to the points (r, θ)
and (r + dr, θ + dθ) is given by
dr2 r2 dθ2
db
s2= + .
(1 + (r/R)2 )2 1 + (r/R)2
(vi) Here is a crazy idea, essentially due to Lambert; see Penrose (2005, § 2.6).
Perhaps we can get a surface of constant negative curvature K = −(1/R2 ) =
+1/(iR)2 by allowing the radius R of the sphere to take on the imaginary
value iR. Verify that if we substitute R = i into the above formula, then it
becomes the Beltrami metric (6.50) of the hyperbolic plane! [To make true
sense of this idea, one must turn to Einstein’s relativity theory; see Thurston
(1997) and especially Penrose (2005).]
15 Take a stack of ten sheets of paper and staple them together, placing staples
along three of the edges. Use a pair of compasses to draw the largest circle
that will fit comfortably inside the top sheet. Pierce through all ten sheets in
the centre of the circle. With heavy scissors, cut along the circle to obtain ten
380 Non-Euclidean Geometry*

identical discs, say of radius R. Repeat this whole process to double the number
of discs to 20.
(i) Cut a narrow sector out of the first disc, and tape the edges together to
form a shallow cone. Repeat this process with the remaining discs, steadily
increasing the angle of the sector each time, so that the cones get sharper
and taller. Ensure that by the end of the process you are making very
narrow cones, using only a quarter disc or less.
(ii) Stack these cones in the order you made them. Explain how it is that you
have created a model of a portion of a pseudosphere of radius R. Create weird
new (extrinsically asymmetric) surfaces of constant negative curvature by
holding the tip of your structure and moving it from side to side!
(iii) Use the same idea to create a disc-like piece of “hyperbolic paper”, such as
you would get if you could simply cut out a disc from your pseudosphere.
Press it against the pseudosphere and verify that you can freely move it
about and rotate it on the surface.
16 By holding a fairly short piece of string against the surface of the toy pseudo-
sphere of the previous exercise, draw a segment of a typical geodesic. Extend
this segment in both directions, one string-length at a time. Note the surpris-
ing way the geodesic only spirals a finite distance up the pseudosphere before
spiralling down again.
(i) Use the upper half-plane to verify mathematically that the tractrix genera-
tors are the only geodesics that extend all the way up to the top.
(ii) Let L be a typical geodesic, and let α be the angle between L and the trac-
trix generator at the point where L hits the rim σ = 0. Show that the
maximum distance σmax that L travels up the pseudosphere is given by
σmax = | ln sin α|.
17 Suppose we have a conformal map of a surface in the xy-plane, with the metric
given by (6.15):
p
db
s = Λ ds = Λ(x, y) dx2 + dy2 .

An elegant result from differential geometry [see Needham (2021, §4.4)] states
that the Gaussian curvature at any point on the surface is given by

1
K=− ∆(ln Λ),
Λ2

where ∆ ≡ (Bx2 + By2 ) is the Laplacian. Try this out on the metric (6.16) of the
stereographic map of the sphere, and on the metrics (6.31) and (6.44) of the
half-plane and disc models of the hyperbolic plane.
 Exercises 381

18 Use the Beltrami–Poincaré disc to rederive the formula tan(Π/2) = e−D for the
angle of parallelism. [Hint: Let one of the h-lines be a diameter.]
19 To derive the metric (6.44), consider the mapping (6.42) z ÞÑ w = D(z) from
the upper half-plane to the Beltrami–Poincaré disc. An infinitesimal vector dz
emanating from z is amplitwisted to an infinitesimal vector dw = D ′ (z) dz
emanating from w, and (by definition) the h-length d b
s of dw is the h-length of
dz. Verify (6.44) by showing that
2 |dw| |dz|
= = db
s.
1 − |w|2 Im z
20 Consider the mapping z ÞÑ w = Mϕ a (z) of the Beltrami–Poincaré disc to itself.
Use the calculational approach of the previous exercise to show that z ÞÑ w is
a hyperbolic motion, i.e., it preserves the metric:
2 |dw| 2 |dz|
= db
s = .
1 − |w| 2 1 − |z|2
21 In the upper half-plane, the h-rotation Rϕ
i through angle ϕ about the point i is
given by the following Möbius transformation:
cz + s
Rϕi (z) = , where c = cos(ϕ/2) and s = sin(ϕ/2).
−s z + c
Prove this in three ways:
′
(i) Show that Rϕ ϕ
i (i) = i and Ri (i) = eiϕ . Why does this prove the result?
(ii) Use the formula for inversion [(3.4), p. 141] to calculate the composition
Rϕi = (RB ◦ RA ), where A and B are h-lines through i, and the angle from
A to B is (ϕ/2). [Hint: Take A to be the imaginary axis, and use a diagram
to show that the semicircle B has centre −(c/s) and radius (1/s).]
(iii) Describe and explain the geometrical effect of applying (D ◦ Rϕ
i ◦ D ) to
−1

the Beltrami–Poincaré disc, where D is the mapping (6.42) from the upper
half-plane to the Beltrami–Poincaré disc. Deduce that
(D ◦ Rϕ
i ◦ D )(z) = e
−1 iϕ
z.
Re-express this equation in terms of products of Möbius matrices, and
solve for the matrix [Rϕ
i ].
382 Non-Euclidean Geometry*

22 (i) Referring to figure [a] above, show that the h-separation of two points in
the upper half-plane may be expressed in terms of a cross-ratio as

H{a, b} = ln[a, B, b, A].

[Hint: Apply an h-rotation centred at a to bring b into a position vertically

above a.]
(ii) Show that the same formula applies to the Beltrami–Poincaré disc in figure
[b].
b and b
23 (i) Let a z be two point on Σ, and let a and z be their stereographic images
in C. If S{a, z} is the distance (on the sphere) between ab and b
z , show that
z−a
S{a, z} = 2 tan−1 .
az + 1
[Hint: Use (6.20) to bring a to the origin.]
(ii) Show that the h-distance formula (6.46) in the Beltrami–Poincaré disc can
be re-expressed as
z−a
H{a, z} = 2 tanh−1 .
az − 1
24 (i) Use a simple sketch to show that z ÞÑ f(z) = −z is an opposite motion of
the Beltrami–Poincaré upper half-plane.
e be an arbitrary opposite motion. By considering (f ◦ M),
(ii) Let M e show that

e az + b
M(z) = , where a, b, c, d are real, and (ad − bc) > 0.
cz + d
(iii) Use this formula to show that M e has two fixed points on the horizon (the
real axis). If L is the h-line whose ends are these two points, explain why L
is invariant under M. e
(iv) Deduce that Me is always a glide reflection: h-translation along L, followed
by (or preceded by) h-reflection in L.
25 Given a point p not on the h-line L, draw an h-circle C of h-radius ρ centred at
p. Draw the h-line orthogonal to L through p, cutting L in q. Draw an h-circle
C ′ of h-radius ρ centred at q, and let q ′ be one of the intersection points with
L. Through q ′ draw the h-line orthogonal to L, cutting C at a and b. Show that
the h-lines joining p to a and b are the two asymptotic lines! [Hint: Take L to
be a vertical line in the upper half-plane.] What happens if we perform this
construction in the Euclidean plane?
26 Sketch a hyperbolic triangle ∆ with vertices (in counterclockwise order) a, b,
and c. Let ξ be an infinitesimal vector emanating from a and pointing along
the edge ab. Carry ξ to b by h-translating it along this edge. Now carry it to c
 Exercises 383

along bc, and finally carry it home to a, along ca. In Euclidean geometry these
three translations would simply cancel, and ξ would return home unaltered.
Use your sketch to show that in hyperbolic geometry the composition of these
three h-translations is an h-rotation about vertex a through angle E(∆).
[Suppose that ∆ is instead a geodesic triangle on an arbitrary surface S of
variable curvature. If an inhabitant of S wants to translate ξ along a “straight
line” (a geodesic), all he has to do is keep its length constant, and keep the
angle it makes with the line constant. This is called parallel transport, and it
plays a central role in differential geometry; see Needham (2021, Act IV). The
above argument still applies, and so when ξ is parallel transported round ∆, it
returns home rotated through E(∆). By virtue of (6.6), this angle of rotation is
the total amount of curvature inside ∆.]
27 Generalize the transformation from the upper half-plane to the Beltrami–
Poincaré disc to obtain a model of hyperbolic space in the interior of the
unit sphere. Describe the appearance of h-lines, h-planes, h-spheres, and horo-
spheres in this model, and explain why an h-sphere is intrinsically the same as
a Euclidean sphere of different radius.
CHAPTER 7

Winding Numbers and Topology

In this chapter we shall investigate a simple but immensely powerful concept—the

number of times a loop winds around a point. In Chapter 2 we saw that this concept
was needed to understand multifunctions, and in the next chapter we will see that it
plays an equally crucial role in understanding complex integration. However, only
the first two sections [up to (7.2)] of the present chapter are actually a prerequisite
for that work; the rest may be read at any time. If you are in a rush to learn about
integration, you may wish to skip the rest of the chapter and return to it later.

7.1 Winding Number

7.1.1 The Definition

As the name suggests, the winding number ν(L, 0) of a closed loop L about the origin
0 is simply the net number of revolutions of the direction of z as it traces out L once
in its given sense. A nut on a bolt admirably illustrates the concept of “net rotation”:
spin the nut this way and that way for a while; the final distance of the nut from
its starting point measures the net rotation it has undergone.
Figure [7.1] shows six loops and their corresponding winding numbers. You can
verify these values by starting at a random point on each curve and tracing it out
with your finger: starting with zero, add one after each positive (= counterclock-
wise) revolution of the vector connecting the origin to your finger, and subtract
one after each negative (= clockwise) revolution. When you have returned to your
starting point, the final count is the winding number of the loop.
It is often useful to consider the winding number of a loop about a point p other
than the origin, and this is correspondingly written ν(L, p). Instead of counting the
revolutions of z, we now count those of (z − p). For example, the shaded region in
[7.1] can be defined as all the positions of p for which ν(L, p) ̸= 0. Try shading this
set for the other loops.

Visual Complex Analysis. 25th Anniversary Edition. Tristan Needham, Oxford University Press.
© Tristan Needham (2023). DOI: 10.1093/oso/9780192868916.003.0007
386 Winding Numbers and Topology

[7.1] The winding number ν of a closed loop L counts the number of times L encircles
a point.

7.1.2 What Does “Inside” Mean?

A loop is called simple if it does not intersect itself; for example, circles, ellipses, and
triangles are all simple. Although a simple loop can actually be very complicated
[see Ex. 1] it seems clear, though it is hard to prove, that it will divide the plane
into just two sets, its inside and its outside. However, in the case of a loop that is
not simple, such as [7.2], it is no longer obvious which points are to be considered
inside the loop, and which outside. The winding number concept allows us make
the desired distinction clearly.

[7.2] Using the winding number to define the “inside” and the “outside” of a closed
loop. A non-simple loop L partitions the plane into multiple simply connected regions,
Dj , and the winding number is constant in each one, taking the value νj . The “inside”
of L is then defined to be the union D1 ∪ D3 of the Dj ’s with νj ̸= 0, and the “outside”
is the union D2 ∪ D4 of the Dj ’s with νj = 0.
 Winding Number 387

A typical loop such as L will partition the plane into a number of sets Dj (four
in this case). If the point p wanders around within one of these sets then it seems
plausible that the winding number ν(L, p) remains constant. Let’s check this.
Concentrate on just a short segment of L. As z traverses it, the rotation of
(z − p) will depend continuously on p unless1 p crosses L. In other words, if we
move p a tiny bit then the rotation angle will likewise only change a tiny bit.
Since the winding number of L is just the sum of the rotations due to all its seg-
ments, it follows that it too depends continuously on the location of p: a tiny
movement of p to p e can only produce a tiny change [ν(L, pe) − ν(L, p)] in the
winding number. But since this small difference is an integer, it must be exactly
0. Done.
Since L winds round each point of Dj the same number of times, it follows that
we can attach a winding number νj to the set as a whole. Verify the values of νj
given in the figure.
The “inside” can now be defined to consist of those Dj for which νj ̸= 0, while
the remaining Dj constitute the “outside”. Thus in [7.2] we find that D1 ∪ D3 is the
inside, while D2 ∪ D4 is the outside.
The “correctness” of this definition will become apparent in the next chapter.

7.1.3 Finding Winding Numbers Quickly

In [7.2] we found the winding numbers directly from the definition: we strenuously
followed the curve with our finger (or eye) and counted revolutions. For a really
complicated loop this could literally become a headache. We now derive a much
quicker and more elegant method of visually computing winding numbers.
If a point r moves around without crossing a loop K then ν(K, r) remains con-
stant, but what happens when the point does cross K? Consider [7.3]. On the far left,
close to the loop K, is the point r; the rest of K is off the picture, and the number
of times it winds round r is ν(K, r). The time-lapse pictures in [7.3] show r moving
towards the loop, which itself deforms so as to avoid being crossed, finally ending
up at the point s.
Now since the moving point never crosses the loop, the winding number
remains constant throughout the process. But on the far right, the new loop can
be thought of as the union of the old loop K, together with the new circle L. Thus,

ν(K, r) = ν(K, s) + ν(L, s) = ν(K, s) − 1

⇒ ν(K, s) = ν(K, r) + 1 .

1
Consider the behaviour of the rotation due to a short segment of L as p crosses it.
388 Winding Numbers and Topology

[7.3] Crossing Rule: If the loop is travelling from our left to our right as we cross it,
ν ⇝ ν + 1. For imagine that K deforms and moves out of our way, so that we never cross
it, then the winding number does not change. But on the far right, K has evolved into
K and a clockwise circle, L, so ν(K, r) = ν(K, s) + ν(L, s) = ν(K, s) − 1, and therefore
ν(K, s) = ν(K, r) + 1, as claimed.

Imagining ourselves at r, looking towards K as we approach it, we may express

this result in the form of the following very useful crossing rule:

If K is moving from our left to our right [our right to our left] as we
(7.1)
cross it, its winding number around us increases [decreases] by one.

Using this result, it is incredibly quick and easy to find the νj ’s for even the most
complicated loop. Try it out on [7.2]. Starting your journey well outside L, where
you know that the winding number is zero, move from region to region, using
crossing rule (7.1) to add or subtract one at each crossing of L.
An immediate consequence of this idea is a connection between n = ν(K, p)
and the number of intersection points of K with a ray emanating from p. Suppose
that the ray is in general position in the sense that it doesn’t pass through any self-
intersection points of K, nor is it tangent to K. If a point q on this ray is sufficiently
distant from p then clearly K cannot wind around it; thus as we move along the
ray from p to q the winding number changes by n. But the winding number only
changes when we cross K, and only one unit per crossing. The ray must therefore
intersect K at least |n| times. However, in addition to these |n| necessary crossings
there may be additional cancelling pairs of crossings. In general, then, the number
of intersection points will be |n|, or |n| + 2, or |n| + 4, etc. Figure [7.4] illustrates
these possibilities for a case in which n = 2, each intersection point being marked
with ⊕ or ⊖ according as the winding number increases or decreases as it is crossed.

7.2 Hopf’s Degree Theorem

7.2.1 The Result

We have discussed the fact that for a fixed loop and a continuously moving point,
the winding number only changes when the point crosses the loop. But it is clear
 Hopf’s Degree Theorem 389

[7.4] A ray emanating from p intersects K at least |ν(K, p)| times. For as we move along
the ray towards infinity, the winding number must eventually drop to zero. But the initial
value n = ν(K, p) only changes when we cross K, and only one unit at a time, so there
must be at least |n| such crossings. But, in addition to these |n| necessary crossings, there
may be additional cancelling pairs of crossings.

that the same must be true of a fixed point and a continuously moving loop: the
winding number of the evolving loop can only change if it crosses the point, and it
changes by ±1 according to the same crossing rule as before. Thus if a loop K can
be continuously deformed into another loop L without ever crossing a point p, the
winding numbers of K and L around p will be equal.
It is natural to ask if the converse is also true: if K and L wind round p the same
number of times, is it always possible to deform K into L without ever crossing p?
This is certainly a more subtle question, but by drawing examples you will be led
to suspect that it is true. In this section we will confirm this hunch, so establishing
that

A loop K may be continuously deformed into another loop L, without

ever crossing the point p, if and only if K and L have the same winding (7.2)
number round p.

At the end of the next chapter, this will turn out to be the key to understanding one
of the central results of complex analysis.
The result in (7.2) is the simplest example of a remarkable topological fact,
called Hopf’s Degree Theorem, that is valid in any number of dimensions. In
the 2-dimensional complex plane, a point can be surrounded using a closed 1-
dimensional curve—a loop. In 3-dimensional space, a point can be surrounded
using a closed 2-dimensional surface. Just as a circle in the plane winds once
390 Winding Numbers and Topology

around its centre, so a sphere in space encloses its centre just once. More gener-
ally, self-intersecting loops in the plane may enclose a point several times, and this
is precisely what ν counts. Similarly, it is possible to define a more general con-
cept (degree) that counts the number of times a surface surrounds a point in space.
Hopf’s Theorem now says that one closed surface may be continuously deformed
into another, without ever crossing p, if and only if they enclose p the same number
of times. Indeed, Hopf’s Theorem says the same is true of n-dimensional surfaces
enclosing points in (n + 1)-dimensional space!

7.2.2 Loops as Mappings of the Circle*

As a first step to understanding (7.2), we will look at loops in a new way. Let C be
a rubber band in the shape of the unit circle. We may now deform C into the shape
of any desired loop L. At the end of the deformation process, each point z of C has
been brought to a definite image point w on L, and thus L may be thought of as the
image of C under a continuous mapping w = L(z). See [7.5].
As θ varies from 0 to 2π, z = eiθ moves round C once and w moves round L
once, the length R and angle Φ of w varying continuously with θ. We may write

w = L(eiθ ) = R(θ) eiΦ(θ) ,

where R(θ) and Φ(θ) are continuous functions. By rotating L (if necessary) we can
ensure that L(ei0 ) is a positive real number, so that we may set Φ(0) = 0. The net
rotation of w after it has returned to its starting point is then given by Φ(2π) = 2πν.
Clearly, the varying length of w is something of a red herring when it comes to
understanding winding numbers, and we now remove this distraction by pulling
each point w of L radially onto the point w b = w/|w| on the unit circle, so obtaining

[7.5] A closed loop L may be viewed as the image of the unit circle C under a continuous
mapping L. As θ goes from 0 to 2π, z = eiθ traverses C once, and w = L(eiθ ) =
R(θ) eiΦ(θ) traverses L once. If Φ(0) = 0 then the net rotation of w after it has returned
to its starting point is given by Φ(2π) = 2πν.
 Hopf’s Degree Theorem 391

[7.6] [a] The loop L can be gradually pushed onto the unit circle by moving w along
b − w). As s increases from 0 to 1, Ls (z) = w + s (w
(w b − w) moves radially from w to
b on the unit circle. [b] With a value of s close to 1, the loop has almost evolved into its
w
b , given by w
final form, L b =L b (eiθ ) = eiΦ(θ) .

a standardized representation Lb . We can even give an explicit prescription for grad-

b
ually deforming L into L . See [7.6a]. Since (wb − w) is the complex number from a
b
point on L to its destination on L , the point that is a fraction s of the way there is

b − w) .
Ls (z) = w + s (w (7.3)

As s varies from 0 to 1, Ls (C) gradually (and reversibly) changes from L into L b.

Figure [7.6b] shows Ls (C) for a value of s close to 1. Finally, note the obvious fact
b , the origin is not crossed.
that as we gradually pull L radially onto L
With lengths disposed of in this way, we are now dealing with a mapping L b
b
from the unit circle C to the standardized loop L on the unit circle, where

w b (eiθ ) = eiΦ(θ) .
b =L (7.4)

In this context, it is common to speak of the degree of the mapping L b which pro-
duces Lb , rather than of the “winding number” of L b (or L). The single real function
b
Φ(θ) completely describes the mapping L, and [7.7] shows how we can immedi-
ately read off the degree of L b (i.e. ν) from the graph of Φ(θ). Make sure you are
comfortable with the meaning of such a graph. For example, if z moves at unit
speed round C, what does the slope (including the sign) of the graph represent?

7.2.3 The Explanation*

The archetypal mapping of degree ν is Jc ν
ν (z) = z , for which Φ(θ) = νθ. Its
b
straight-line graph is shown in [7.7]. As z travels once round C at unit speed, w
c
travels once round Jν with speed |ν|, completing |ν| circuits of the unit circle
[counterclockwise if ν > 0; clockwise if ν < 0].
392 Winding Numbers and Topology

[7.7] Gradual evolution of a loop into the archetypal loop of the same winding number.
If we imagine Lb in the previous figure to be an elastic string wrapped around a cylin-
der, the slack will automatically be taken up, and the actual mapping L b of degree ν
b
will evolve into the archetypal mapping of degree ν, namely, J ν (z) = zν , for which
Φ(θ) = νθ. This can be achieved explicitly by following the illustrated evolution of
Φt (θ) = Φ(θ) + t [νθ − Φ(θ)] from t = 0 to t = 1.

To see how a typical standardized loop L b of winding number ν is related to the

archetypal loop c Jν , we return to the example in [7.6b], for which ν = 2. Thinking
of the unit circle as the boundary of a solid cylinder, and recalling that the loop is
an elastic band that wishes to contract, what will happen if we release L b ? The slack
will be taken up, and L b will automatically contract itself into the archetypal loop
cJ2 . This convincing mental image of the rubber band contracting into c Jν can be
formalized to show that
Any Lb of winding number ν can be continuously deformed into the
archetypal loop c
Jν , and vice versa.
The process of “taking up the slack” can be explicitly described in terms of the
b . As t varies from 0 to 1, the graph of
graph of the Φ that describes L
Φt (θ) = Φ(θ) + t [νθ − Φ(θ)]
continuously and reversibly evolves from the graph of the general Φ into the
straight-line graph of the archetype. The dashed curve in [7.7] is the graph of Φt
for a value of t close to one. Defining
b t (eiθ ) = eiΦt (θ) ,
L (7.5)
b t (C) therefore evolves continuously and reversibly from L
L b into c
Jν , as t varies
from 0 to 1.
 Polynomials and the Argument Principle 393

The explicit two-stage deformation given above [(7.3) followed by (7.5)] allows
us to deform any loop of winding number ν into the archetypal loop c Jν , and with-
out the origin ever being crossed. Conversely, by reversing these steps, cJν may be
deformed into any loop of winding number ν. This demonstrates (7.2), for if K and
L both have winding number ν, we may first deform K into c Jν , and then deform
c
Jν into L.

7.3 Polynomials and the Argument Principle

Let A, B, and C be the complex numbers from the fixed points a, b, and c to the
variable point z. Figure [7.8] shows a circle Γ and its image f(Γ ) under the cubic
mapping
f(z) = (z − a) (z − b) (z − c) = ABC .
Notice that Γ encircles two zeros of the mapping, while f(Γ ) has a winding number
of 2 about zero. This is no accident. Since angles add when we multiply complex
numbers, the number of revolutions executed by ABC is just the sum of the revolu-
tions executed separately by each of A, B, and C. But as z goes round Γ once, A and
B both execute a complete revolution, while the direction of C merely oscillates.
Thus ν[f(Γ ), 0] = 2.
If we enlarged Γ so that it encircled c, then C would also execute a complete
revolution, and the winding number would increase to 3. Once again, the number
of points inside Γ that are mapped to 0 is the winding number of the image about
that point.

[7.8] If a simple loop Γ winds once around m roots of a polynomial P(z), then
ν[P(Γ), 0] = m. In the illustrated cubic example, the number of revolutions executed
by f(z) = ABC is just the sum of the revolutions executed separately by each of A, B,
and C. But as z goes round Γ once, A and B both execute a complete revolution, while
the direction of C merely oscillates, so f(z) winds around 0 twice.
394 Winding Numbers and Topology

It is clear that this result is independent of the circularity of Γ , and that it gener-
alizes to the case of a polynomial P(z) of arbitrary degree: If a simple loop Γ winds
once around m roots of P(z), then ν[P(Γ ), 0] = m.
Roots are simply preimages of 0, and from the geometric viewpoint there is noth-
ing special about this particular image point. Consequently, in future we will look
at the preimages of a general point p and we will call these preimages p-points of
the mapping.
The Argument Principle is a tremendous extension of the above result. Not
only does it apply to general analytic mappings but it also contains the converse
statement that the winding number tells us the number of preimages:

If f(z) is analytic inside and on a simple loop Γ , and N is the number of

(7.6)
p-points [counted with their multiplicities] inside Γ , then N = ν[f(Γ ), p].

The meaning of the expression “counted with their multiplicities” will be explained
in the next section.
We wish to stress that this result is only peripherally connected with the confor-
mality that has been so central to all our previous thinking. In fact the Argument
Principle is a consequence of a still more general topological fact concerning map-
pings that are merely continuous. Our main effort will therefore be directed
towards understanding the general result (due to Poincaré), of which the Argument
Principle is merely a special case.

7.4 A Topological Argument Principle*

7.4.1 Counting Preimages Algebraically

Even in the simple case of [7.8] we find that we must be careful how we count
preimages. If we move the root b towards the one at a, zero continues to have two
preimages inside Γ , while the image of Γ continues to wind round zero twice. How-
ever, when b actually arrives at a there is apparently only a single 0-point inside
Γ (namely a) despite the fact that f(Γ ) winds round 0 twice. Thus a must now be
counted twice if (7.6) is to remain true.
Algebraically, the resolution lies in the fact that a is now a “double-root”, the
factored form of f(z) = ABC = A2 C containing the square of A = (z − a). More
generally, if the factorization of a polynomial contains the term An then we say
that the root a is a 0-point of algebraic multiplicity n, and we must count it with this
multiplicity in (7.6).
When n > 1 there is a further significance to the point a—it is a critical point of
the polynomial. In the cubic mapping f(z) = ABC, let a, b, and c be real, so that
f(z) is real-valued on the real axis. The far left of [7.9] shows the ordinary graph
 A Topological Argument Principle* 395

[7.9] When roots coalesce, the derivative must vanish. The same applies in C, for if
f ′ (a) ̸= 0 then an infinitesimal disc centred at a is amplitwisted to an infinitesimal disc
centred at 0, so that points close to a cannot map to 0.

of f in this case. As b moves towards a, the slope at a is forced to decrease, finally

vanishing at the moment of b’s arrival.
In general this vanishing of the derivative (now amplitwist) must occur wher-
ever two or more roots of a polynomial coalesce. Look again at [7.9]. If f ′ (a) ̸= 0
then the graph is not flat at a and so neighbouring points cannot map to zero; but
this is precisely what we insist on when we merge b into a. Essentially the same
thing happens when we return to C, for if f ′ (a) ̸= 0 then an infinitesimal disc cen-
tred at a is amplitwisted to an infinitesimal disc centred at 0, so that points close to
a cannot map to 0.
This conclusion can be refined. If the root a of a polynomial P(z) has multiplic-
ity n then P may be factorized as An Ω(z), where Ω(a) ̸= 0. It follows by simple
calculation [exercise] that the first (n − 1) derivatives of P vanish at a, so that a is
a critical point of order at least (n − 1). We shall see in a moment that the order
actually is (n − 1).
We next seek to extend the idea of counting preimages “with algebraic multi-
plicities” to analytic mappings in general. Suppose that a is a p-point of an analytic
mapping f(z), i.e., it is a 0-point of f(z) − p. What should the algebraic multiplicity
of this root be? It is only possible to answer this question because of the remarkable
fact that an analytic f can always be represented as a convergent Taylor series in the
neighbourhood of a non-singular point. Thus if ∆ = (z − a) is the small complex
number from a to a nearby point z, we may write

f ′ (a) f ′′ (a) 2
f(z) − p = f(a + ∆) − f(a) = ∆+ ∆ + ··· .
1! 2!
The first nonzero term on the right is the one that dominates the local behaviour of
f(z) − p and decides what the multiplicity of a should be. Typically a will not be a
critical point [f ′ (a) ̸= 0] and so this local behaviour is like ∆ to the first power; we
say that a is a simple root with multiplicity +1.
Now consider the rarer case in which a is a critical point. If the order of the
critical point is (n − 1), so that f(n) is the first nonvanishing derivative at a, then
the dominant first term is proportional to ∆n , and we correspondingly define the
396 Winding Numbers and Topology

algebraic multiplicity2 of a to be n. The analogy between this definition and that for
polynomials may be brought to the fore by setting

f(n) (a) f(n+1) (a) f(n+2) (a) 2

Ω(z) = + ∆+ ∆ + ··· ,
n! (n + 1)! (n + 2)!

where f(n) (a) is the first nonvanishing derivative. The previous equation can now
be written in “factorized” form as

f(z) − p = Ω(z) ∆n , (7.7)

where Ω(a) ̸= 0. From this point of view, the only difference between a general
analytic mapping and a polynomial is that the latter has a single, “once and for all”
factorization, while the former generally requires a different factorization of type
(7.7) in the neighbourhood of each p-point.

7.4.2 Counting Preimages Geometrically

Recall that we wish to explain (7.6) as a special case of a more general result dealing
with mappings that are merely continuous. But since the very notion of algebraic
multiplicity is meaningless for such general mappings, how can we even frame a
proposition of type (7.6)?
What is needed is a geometric way of counting preimages that will agree
with the previous algebraic definition if we specialize to analytic mappings. To
discover the appropriate definition we should therefore return to analytic map-
pings and ask, “What is the geometric fingerprint of a p-point of given algebraic
multiplicity?”
Consider the effect of an analytic f on an infinitesimal circle Ca centred at a simple
p-point a. Since f ′ (a) ̸= 0, Ca is amplitwisted to an infinitesimal circle centred at
p. We see that the winding number (+1) of this image round p is the same as the
algebraic multiplicity of a. In fact, quite generally, if the algebraic multiplicity is
n then the winding number of the image will also be n. This is the sought-after
geometric fingerprint.
To verify this statement, remember that the local behaviour of f near to a is given
by (7.7):

f(z) = f(a + ∆) = p + Ω(z) ∆n ,

with Ω(a) ̸= 0. Thus the basic explanation is that as ∆ revolves round Ca once,
∆n rotates n-times as fast, and therefore f(z) completes n revolutions round p. If

2
Also known as order or valence.
 A Topological Argument Principle* 397

[7.10] Topological Multiplicity as generalization of Algebraic Multiplicity. On the left

are p-points of the continuous mapping h, i.e., points that are mapped to p by h, and
Γa is a loop around a (and only a). If we continuously deform Γa into another such loop
e
Γa without crossing a (or any other p-point) then h(Γa ) will continuously deform, but its
winding number around p will not change. Thus ν(a) ≡ ν[h(Γa ), p ] is really a topologi-
cal invariant associated with the point a, and it is called its topological multiplicity. If we
specialize h to be a complex polynomial, then the topological multiplicity is equal to the
algebraic multiplicity of a as a repeated root of h(z) = p.

Ω were a constant then this argument would be beyond reproach, and we can now
do a little calculation to show that a variable Ω does not disturb the conclusion:

ν[f(Ca ), p ] = ν[f(Ca ) − p, 0] = ν[∆n Ω(Ca ), 0]

= n ν[∆, 0] + ν[Ω(Ca ), 0] . (7.8)

As we shrink Ca down towards a, Ω(Ca ) will shrink down towards Ω(a), but
since Ω(a) ̸= 0 this implies that the image of a sufficiently small Ca will not wind
round 0: ν[Ω(Ca ), 0] = 0. Since ν[∆, 0] = 1, we conclude that ν[f(Ca ), p ] = n, as
claimed.
We may now broaden our horizon and use the above idea to define “multiplic-
ity” for a mapping h(z) that is merely continuous. Let Γa be any simple loop round
a that does not contain other p-points. Figure [7.10] shows such a loop as well as
some other p-points b, c, etc. If we continuously deform Γa into another such loop
e
Γa without crossing a (or any other p-point) then h(Γa ) will continuously deform
into h(eΓa ) without ever crossing p, and so

ν[h(e
Γa ), p ] = ν[h(Γa ), p ] .

Thus, without specifying Γa further, we may unambiguously define the topological

multiplicity3 of a to be

ν(a) ≡ ν[h(Γa ), p ] .

3
Also known as the local degree of h at a.
398 Winding Numbers and Topology

In the case of the mapping in [7.10] we see that ν(a) = −2. If h happened to be
analytic then a would also possess an algebraic multiplicity n, but by deforming
Γa into the infinitesimal circle Ca , we find that the two kinds of multiplicity must
agree: ν(a) = n.

7.4.3 What’s Topologically Special About Analytic Functions?

From the geometric point of view, conformal (analytic) mappings are infinitely
richer in structure than mappings that are merely continuous. However, from
the point of view of topological multiplicity there are only a few distinctions, the
following being one of the most striking:
ν(a) is always positive for analytic functions, while it can be negative for
nonanalytic functions.
For example, the mapping in [7.10] cannot possibly be analytic. The positivity of
ν(a) for analytic functions has already been established, so we need only look more
closely at the possibility of negative multiplicities for nonanalytic functions.
Since general continuous mappings can actually behave in rather wild ways, let
us restrict ourselves to nonanalytic mappings that are at least differentiable in the
real sense. For example, consider h(z) = z. The unique preimage of p is a = p,
and any simple loop Γa round this point is reflected by h into a loop that goes once
round p in the opposite direction. Thus ν(a) = −1.
More generally, recall [see p. 236] that the local effect of such a mapping at a p-
point a is (after translation to p) a stretch by some factor ξa in one direction, another
stretch by some factor ηa in the perpendicular direction, and finally a rotation
through some angle ϕa . For example, conjugation has (taking the first expansion
to be horizontal) ξa = +1, ηa = −1, ϕa = 0. Of course these values of ξa , ηa , and
ϕa are only independent of a because h(x + iy) = x − iy depends linearly on x and
y; most mappings have values that do depend on the point a.
An infinitesimal circle Ca centred at a is generally distorted into an infinitesi-
mal ellipse Ep centred at p, and if the two expansion factors have the same sign
then the mapping preserves orientation so that Ep circulates in the same sense as Ca
and ν(a) = +1. However, if ξa and ηa have opposite signs then the mapping is
orientation reversing: it turns Ca inside out, so that Ep goes round p in the opposite
direction and ν(a) = −1. Our previous example of conjugation was of this type. In
summary, we have

ν(a) = the sign of (ξa ηa ).

The local linear transformation at a is encoded by the Jacobian matrix J(a), and
we can use its determinant det[J(a)] to give a more practical formula for the topo-
logical multiplicity. We know from linear algebra that the determinant of a constant
2 × 2 matrix is the factor (including a sign for orientation) by which the area of a
 A Topological Argument Principle* 399

figure is expanded. Likewise, det[J(a)] measures the local expansion factor for area
at a, and this is just (ξa ηa ). Thus

ν(a) = the sign of det[J(a)] . (7.9)

Of course this formula is vacuous if det[J(a)] = 0. Geometrically this means that

the transformation is locally crushing at a; just as for analytic mappings, such a
place is called a critical point. However, while the local crushing at a critical point of
an analytic mapping is perfectly symmetrical in all directions, this is not true of the
more general mappings presently under consideration. For example, if f(x + iy) =
x − iy3 then [exercise] det[J] = −3y2 , so although f leaves the horizontal separation
of points alone, all the points on the real axis are critical points as a result of crushing
in the vertical direction.
This example also serves to illustrate another difference:

The critical points of an analytic mapping can be distinguished purely

on the basis of topological multiplicity; those of a nonanalytic mapping
cannot.
For analytic functions we have seen that ν(a) = +1 if and only if a is not critical. In
the nonanalytic case ν(a) = ±1 if a is not critical, but it is also possible for a critical
point to have one of these multiplicities. Indeed, you can check [exercise] that the
above example yields ν(a) = −1 for noncritical and critical points alike.
One final difference:
ν(a) is never zero for analytic mappings, but it can vanish for nonana-
lytic mappings.
In the next section we will provide a simple example of a nonanalytic mapping
possessing such p-points of vanishing topological multiplicity. Can you think of
an example for yourself?

7.4.4 A Topological Argument Principle

Let Γ be a simple loop, and let h(z) be a continuous mapping such that only a finite
number of its p-points lie inside Γ . We will show that

The total number of p-points inside Γ (counted with their topological

(7.10)
multiplicities) is equal to the winding number of h(Γ ) round p.
If h is analytic this reduces to (7.6). The rest of the chapter will be devoted to mining
and extending this simple yet profound result4 .

4
When interpreted in terms of vector fields (as we shall do in Chapter 10) this is the key to a very
surprising and beautiful fact called the Poincaré–Hopf Theorem.
400 Winding Numbers and Topology

[7.11] The winding number ν[h(Γ), p] equals the sum of the topological multiplicities
of the p-points inside Γ. For this winding number will not change value if we deform Γ
as shown into a sum of loops around the individual p-points within.

Before explaining (7.10), let us describe one of its immediate consequences. As

in [7.2], h(Γ ) will generally partition the plane into a number of sets, and the above
result then says that every point in Dj has the same number of preimages lying
inside Γ , namely, νj . For example, if h(Γ ) is a simple loop then it partitions the
plane into just two sets, namely, its interior and its exterior. If p is in the interior
then the result says that the total number of p-points inside Γ is 1. But if h is ana-
lytic then these p-points must have strictly positive multiplicities, and so there is
exactly one preimage for each point inside h(Γ ). In other words, we have shown
that
If an analytic function h maps Γ onto h(Γ ) in one-to-one fashion, then it
also maps the interior of Γ onto the interior of h(Γ ) in one-to-one fashion.
This is Darboux’s Theorem.
To explain (7.10), consider [7.11]. This shows three p-points a, b, and c lying
inside Γ while others lie scattered outside. The essential idea is that we can grad-
ually deform Γ (as shown) into the doubly-pinched loop αβγδγβα, which we will
call e
Γ . Since no p-points were crossed in the deformation process, h(e Γ ) will wind
round p the same number of times as h(Γ ). The rest is almost obvious: e Γ is made
up of Γa = αβα, Γb = βγβ, Γc = γδγ, and the winding numbers of their images
round p are, by definition, the topological multiplicities of a, b, and c.
We will spell this out in perhaps unnecessary detail. Let K be a path that is
not necessarily closed, and define R(K) to be the net rotation of h(z) round p as
z traverses K. For example, if K is closed then R(K) = 2πν[h(K), p ]. Then,
 A Topological Argument Principle* 401

2π ν[h(Γ ), p ] = 2π ν[h(e
Γ ), p ]
= R(αβγδγβα)
= R(αβ) + R(βγ) + R(γδ) + R(δγ) + R(γβ) + R(βα)
= R(αβα) + R(βγβ) + R(γδγ)
= R(Γa ) + R(Γb ) + R(Γc )
= 2π[ν(a) + ν(b) + ν(c)] .
Clearly this idea extends to any number of p-points a1 , a2 , etc. lying inside Γ :
X
ν[h(Γ ), p ] = ν(aj ) . (7.11)
aj inside Γ

7.4.5 Two Examples

Let us immediately illustrate the result with a concrete mapping:
h(x + iy) = x + i|y| .
In terms of our pastry analogy [p. 234] this corresponds to making a crease along
the real axis and folding the bottom half of the plane onto the top half; in fact, since
no stretching is required in this example, folding a simple piece of paper will do!
If Im(p) > 0 then p has two preimages, a1 = p and a2 = p. Figure [7.12] shows
that ν(a1 ) = +1 and ν(a2 ) = −1, and in accord with (7.11), it also shows that if Γ
contains a1 and a2 then ν[h(Γ ), p ] = 0.

[7.12] Folding a piece of paper across the x-axis is a continuous mapping h, explicitly
given by h(x + iy) = x + i|y|. If Im(p) > 0, as shown, then p has two preimages,
a1 = p and a2 = p, so ν(a1 ) = +1 and ν(a2 ) = −1. Therefore, (7.11) predicts that
ν[h(Γ ), p ] = ν(a1 ) + ν(a2 ) = (+1) + (−1) = 0, which is indeed true!
402 Winding Numbers and Topology

In general, note that ν[h(Γ ), p ] = 0 merely implies that either there are no preim-
ages inside Γ or the preimages have cancelling topological multiplicities, as above.
However, if f is analytic and ν[f(Γ ), p ] = 0 then the conclusion is quite definite:
there are no preimages inside Γ . Later we shall return to this important point.
Returning to the example, observe that if p = X is real then there is only one
preimage, namely X, and ν(X) = 0. We can look at this in a nice way: as we
move p towards X, the two preimages a1 and a2 also move towards X, and when
they finally coalesce at X their opposite multiplicities annihilate. As we previously
pointed out, such points of vanishing multiplicity can only exist for nonanalytic
mappings.
Figure [7.13] shows a second more elaborate example, in which we subject a unit
disc of pastry to a three-stage transformation H that leaves the boundary Γ (the unit
circle) fixed: H(Γ ) = Γ . Here are the three stages: (A) form a “hat” by lifting up the
part of the disc lying inside the dashed circle, some of the pastry dough outside the
dashed circle being stretched to form the side of the cylinder. (B) radially stretch
the disc forming the top of this “hat” till its radius is greater than one. (C) press
down flat, i.e., project each point vertically down onto the plane.
If we pick a point p from the image set [bottom left] then the number of preim-
ages (counted naively) lying in the original disc is the number of layers of pastry
lying over p. In the final picture [bottom left] we have used the degree of shading
to indicate the number of these layers: one over the lightly shaded inner disc; two

[7.13] (A) Form a “hat” by lifting up the part of the disc lying inside the dashed circle,
some of the pastry dough outside the dashed circle being stretched to form the side of the
cylinder. (B) Radially stretch the disc forming the top of this “hat” till its radius is greater
than one. (C) Press down flat, i.e., project each point vertically down onto the plane. The
final number of layers [bottom left] is indicated by the shading: one layer over the lightly
shaded inner disc; two over the darkly shaded outer ring; and three over the black ring.
You can now geometrically verify that (7.11) works in all three regions.
 Rouché’s Theorem 403

over the darkly shaded outer ring; three over the black ring. Make sure you can see
this.
We can now check (7.11). For example, if p lies in the darkly shaded outer ring,
ν[H(Γ ), p ] = ν[Γ , p ] = 0 and so the multiplicities of the preimages of such points
should sum to zero. By following the effect of the transformation on little loops
round each of the two preimages, we confirm this prediction: one preimage has
multiplicity +1 while the other has multiplicity−1.
Check for yourself that (7.11) continues to work if p instead lies in the inner disc
or in the black ring.

7.5 Rouché’s Theorem

7.5.1 The Result

Imagine walking a dog round and round a tree in a park, both you and the dog
finally returning to your starting points. Further imagine that the dog is on one of
those leashes of variable length, similar to a spring-loaded tape measure. On one
such walk you keep the leash short, so that the dog stays at your heel. It is then clear
that the dog is forced to walk round the tree the same number of times that you do.
On another walk, though, you decide to let out the leash somewhat so that the
dog may scamper about, perhaps even running circles around you. Nevertheless,
provided that you keep the leash short enough so that the dog cannot reach the tree, then
again the dog must circle the tree the same number of times as you.
Let the tree be the origin of C, and let your walk be the image path traced by f(z)
as z traverses a simple loop Γ . Also, let the complex number from you to the dog
be g(z), so that the dog’s position is f(z) + g(z). The requirement that the leash not
stretch to the tree is therefore

|g(z)| < |f(z)| on Γ .

Under these circumstances, the previous paragraph states that

ν[(f + g)(Γ ), 0] = ν[f(Γ ), 0] .

But the Argument Principle then informs us that

If |g(z)| < |f(z)| on Γ , then (f + g) must have the same number of zeros
inside Γ as f.

This is Rouché’s Theorem.

Note that while |g(z)| < |f(z)| is a sufficient condition for (f + g) to have the
same number of roots as f, it is not a necessary condition. For example, consider
g(z) = 2f(z).
404 Winding Numbers and Topology

7.5.2 The Fundamental Theorem of Algebra

A classic illustration of Rouché’s Theorem is the Fundamental Theorem of Algebra,
which states that a polynomial

P(z) = zn + A zn−1 + B zn−2 + · · · + E

of degree n always has n roots. The basic explanation is simple: if |z| is large, the first
term dominates the behaviour of P(z) and the image of a sufficiently large origin-
centred circle C will therefore wind n times round 0; the Argument Principle then
says that P(z) must have n roots inside C.
Rouché’s Theorem merely allows us to make the above idea more precise. Let
f(z) = zn be the first term of P(z) and let g(z) be the sum of all the rest, so
that f + g = P. Now let C be the circle |z| = 1 + |A| + |B| + · · · + |E|. Using the
fact that |z| > 1 on C, it is not hard to show [exercise] that |g(z)| < |f(z)| on C, and
since f has n roots inside C (all at the origin), Rouché says that P must too.
Notice that we have not only confirmed the existence of the n roots, but have also
narrowed down their location: they must all lie inside C. In the exercises you will
see how Rouché’s Theorem can often be used to obtain more precise information
on the location of the roots of an equation.

7.5.3 Brouwer’s Fixed Point Theorem*

Sprinkle talcum powder on a cup of coffee and give it a stir. The little white specks
will swirl around and eventually come to rest, the speck that was originally at z
finally ending up at g(z). If we stir it in a nice symmetrical way then the speck in
the centre will remain motionless and its final position will be identical with its
starting position. Such a place, for which g(z) = z, is called a fixed point of g.
Now stir the coffee in a really complicated way and let it again come to rest.
Incredible as it may seem, at least one speck will have ended up exactly where it
started! This is an example of Brouwer’s Fixed Point Theorem, which asserts that any
continuous mapping of the disc to itself will have a fixed point. Exercise 15 shows
this to be true, but for the moment we wish to demonstrate a slightly different
result: there must be a fixed point if the disc is mapped into its interior and there
are at most a finite number of fixed points.
Let the disc D be |z| ⩽ 1; the condition that g map D into the interior of D is then
|g(z)| < 1 for all z in D. Let m(z) be the movement of z under the mapping, i.e., the
connecting complex number from z to its destination g(z):

m(z) = g(z) − z .

A fixed point then corresponds to no movement: m(z) = 0. Now let f(z) = −z. On
the boundary of D (the unit circle) we have

|g(z)| < 1 = |f(z)|

 Maxima and Minima 405

and so Rouché’s Theorem says that m(z) = g(z) + f(z) has the same number of
roots inside D as f has, namely, one.
If g is merely continuous then there can actually be several fixed points, some of
which will necessarily have negative multiplicities, while if g is analytic then there
can literally only be one.

7.6 Maxima and Minima

7.6.1 Maximum-Modulus Theorem

Take another look at the nonanalytic mapping H of [7.13], and note how the image
of the disc “spills over” the image of its boundary: points inside Γ end up in the
darkly shaded ring outside H(Γ ). The central observation of this section is that such
spilling over is quite impossible in the case of an analytic mapping:

If f is analytic inside and on a simple loop Γ then no point outside f(Γ )

(7.12)
can have a preimage inside Γ .

Let’s see why. The Argument Principle tells us that the sum of the multiplicities
ν(aj ) of the p-points inside Γ is ν[f(Γ ), p ], but if p is outside f(Γ ) then (by definition)
this is zero. Since ν(aj ) is strictly positive for an analytic function, we conclude that
points outside f(Γ ) have no preimages inside Γ . On the other hand, if p lies inside
f(Γ ) then ν[f(Γ ), p ] ̸= 0, and so there must be at least one preimage inside Γ . [Unlike
(7.12), this is also true of nonanalytic mappings.]
Figure [7.14] illustrates an analytic f sending the shaded interior of Γ strictly
to the shaded interior of f(Γ ). Since f(Γ ) winds round the darker region twice, its
points have two preimages in Γ ; we can think of this as arising from the overlap of
two lightly shaded regions, one preimage per lightly shaded point.
One aspect of the “overspill” produced by H in [7.13] is that the points z which
end up furthest from the origin (i.e., for which the modulus |H(z)| is maximum) lie
inside Γ . Conversely, the absence of overspill for an analytic f means that

The maximum of |f(z)| on a region where f is analytic is always achieved

by points on the boundary, never ones inside.

This is called the Maximum-Modulus Theorem and it is illustrated in [7.14]: the

maximum of |f(z)| is |T | = |f(t)|, where t lies on Γ .
The only exception to this result is the trivial analytic mapping z ÞÑ const.,
which sends every point to a single image point. To put this more positively, if we
know that |f| achieves its maximum at an interior point then f(z) = const.
As an elementary example, consider this problem. Let F(z) be the product of the
distances from z to the vertices a, b, c, d of a square. If z lies inside or on the edge
of the square, where does the maximum value of F occur? It is certainly tempting
406 Winding Numbers and Topology

[7.14] Maximum-Modulus Theorem: The maximum of |f(z)| on a region where f is

analytic is always achieved by points on the boundary, never ones inside. For the Argu-
ment Principle tells us that the sum of the multiplicities ν(aj ) of the p-points inside Γ is
ν[f(Γ ), p ], but if p is outside f(Γ ) then (by definition) this is zero. Since ν(aj ) is strictly
positive for an analytic function, we conclude that points outside f(Γ ) have no preimages
inside Γ . On the other hand, if p lies inside f(Γ ) then ν[f(Γ ), p ] ̸= 0, and so there must
be at least one preimage inside Γ .

to guess that the maximum will occur at the centre of the square, but this is wrong.
Since F(z) = |(z − a) (z − b) (z − c) (z − d)| is the modulus of an analytic mapping,
the maximum must in fact occur somewhere on the edge of the square. The exact
location can now be found [exercise] using nothing more than ordinary calculus.
Returning to matters of theory, recall from Chapter 2 that the “modular surface”
of f is the surface obtained by lifting each point z vertically to a height |f(z)| above
the complex plane. If we look at the portion of this surface lying above Γ and its
interior, the result says that the highest point always lies on the edge, never inside.
Although the absolute maximum of the height always occurs on the edge, could
there perhaps be a local maximum of |f| at an interior point a, so that the surface
would have a peak above a? No! For if we cut out the piece of the surface lying
above the interior of any small loop γ round a, the highest point will fail to lie on
the edge. Thus a modular surface has no peaks. Further aspects of the modular
surface are investigated in the exercises.
This absence of local maxima is re-explained in [7.14]. The Argument Principle
says that since γ contains a, f(γ) must wind round A = f(a) at least once. This
makes it clear that there are always points on γ which have images lying further
from the origin than A. More formally (cf. [7.4]), any ray emanating from A must
intersect f(γ), and by choosing the ray to point directly away from the origin, the
intersection point is guaranteed to lie further from the origin than A. Thus |f| cannot
have a local maximum.
 The Schwarz–Pick Lemma* 407

7.6.2 Related Results

As [7.14] illustrates, the Maximum-Modulus result is only one of several that follow
from (7.12). For example, unless there is a 0-point inside Γ , at which |f(z)| = 0,
the point Q closest to the origin (for which |f(z)| is minimum) must also be the
image of a point q lying on the boundary Γ . Naturally enough, this is called the
Minimum-Modulus Theorem.
Thus if we cut out the piece of the modular surface lying above Γ and its interior,
the lowest point will always lie on the edge, unless, that is, the surface actually hits
the complex plane at an interior 0-point of f. By the same token, there can be no
pits in the surface [local minima of |f|] except at 0-points.
As before, the only exception to all this is the mapping z ÞÑ const., for which
every point yields the smallest (and only) value of |f|. Thus if we know that |f|
achieves a positive minimum at an interior point then f(z) = const.
If f = u + iv is analytic then [cf. Ex. 2, p. 293] u and v are automatically “har-
monic”. As we shall see in Chapter 11, this means that these functions are intimately
connected with numerous physical phenomena: heat flow, electrostatics, hydrody-
namics, to name but a few. It is therefore of significance that [7.14] shows that u and
v are also subject to the principle that their maxima and (nonzero) minima can only occur
on Γ , never inside Γ . As before, if a maximum or minimum occurs at an interior point,
the harmonic function must be constant.

7.7 The Schwarz–Pick Lemma*

7.7.1 Schwarz’s Lemma

Thinking of the unit disc as the Beltrami–Poincaré model for non-Euclidean
geometry, we saw in Chapter 6 that a special role was played by the Möbius
transformations of the form
 
ϕ iϕ z−a
Ma (z) = e = eiϕ Ma (z), (7.13)
az − 1

where a lies inside the disc. These one-to-one mappings of the disc to itself act as
rigid motions, for they preserve non-Euclidean distance.
Apart from a digression on Liouville’s Theorem, this section continues the work
(begun in Chapter 6) of exhibiting the beautiful pre-existing harmony that exists
between non-Euclidean geometry and the theory of conformal mappings. Our first
new piece of evidence that these two disciplines somehow “know” about each
other is the following:

Rigid motions of the hyperbolic plane are the only one-to-one analytic
(7.14)
mappings of the disc to itself.
408 Winding Numbers and Topology

There are of course many other kinds of analytic mapping of the disc to itself, but
according to (7.14) they must all fail to be one-to-one. For example, z ÞÑ z3 maps
the disc to itself, but it is three-to-one.
Observe that this result establishes a claim we previously made [see p. 204] in
connection with Riemann’s Mapping Theorem. There we explained that there are
as many mappings of one region to another as there are automorphisms of the
disc. We already knew that these automorphisms included 3-parameter’s worth of
Möbius mappings, and (7.14) now tells us that there are no more.
To verify (7.14) we will first establish a lemma (of great interest in itself) due to
Schwarz:
If an analytic mapping of the disc to itself leaves the centre fixed,
then either every interior point moves nearer to the centre, or else the
transformation is a simple rotation.

The example f(z) = z2 shows that the mapping need not be a rotation in order
for boundary points to keep their distance from the centre. However, at an interior
point we have |z| < 1, and so |f(z)| = |z|2 < |z|, in accord with the result.
Let f be any analytic mapping of the disc to itself leaving the centre fixed, so
that |f(z)| ⩽ 1 on the disc, and f(0) = 0. We wish to show that either |f(z)| < |z|
at interior points, or else f(z) = eiϕ z. To this end, consider the ratio F of image to
preimage:
f(z)
F(z) ≡ .
z

At first sight this may look undefined at 0, but a moment’s thought shows that as
z approaches the origin, F(z) approaches f ′ (0).
From the previous section we know that the maximum modulus of an analytic
function on the disc can only occur at an interior point if the function is constant,
otherwise it’s on the boundary circle |z| = 1. Thus if p is an interior point and z
varies over the unit circle C, then

|F(p)| ⩽ (max |F(z)| on C) = (max |f(z)| on C) ⩽ 1.

Thus it is certainly true that no interior point can end up further from the centre. But
if even a single interior point q remains at the same distance from the centre then
|F(q)| = 1, which means that F has achieved its maximum modulus at an interior
point. In this case F must map the entire disc to a single point of unit modulus, say
eiϕ , so that f(z) = eiϕ z is a rotation. Done.
The result is illustrated in [7.15]. If f is not a rotation then every point z on a circle
such as K is mapped to a point w = f(z) lying strictly inside K, and the shaded
region is compressed as shown. If we shrink K down towards the origin then f
will amplitwist it to another infinitesimal circle centred at the origin, but having a
 The Schwarz–Pick Lemma* 409

[7.15] Schwarz’s Lemma: If an analytic mapping f(z) of the disc to itself leaves the
centre fixed, then either every interior point moves nearer to the centre, or else the trans-
formation is a simple rotation. As explained in the text, this follows from applying the
Maximum-Modulus Theorem to F(z) ≡ f(z)/z.

smaller radius. Thus the amplification of f at the origin must be less than one. We
can only have |f ′ (0)| equal to one in the case of a rotation.
We can now return to (7.14). As in Schwarz’s Lemma, first suppose that the
mapping f leaves the centre fixed, but now take f to be one-to-one, so that it has
a well-defined analytic inverse f−1 which also maps the disc to itself and leaves the
centre fixed. By Schwarz’s Lemma, f sends an interior point p to a point q that is
no further from centre than p. But f−1 is also subject to Schwarz’s Lemma, and so
p = f−1 (q) must be no further from the centre than q. These two statements are
only compatible if |q| = |f(p)| = |p|. Thus f must be a rotation, which is indeed a
rigid motion of type (7.13).
Finally, suppose that the one-to-one mapping f does not leave the centre fixed,
but instead sends it to c. We can now compose f with the rigid motion Mc which
sends c back to 0. We thereby obtain a one-to-one mapping (Mc ◦ f) of the disc
to itself which does leave the centre fixed, and which must therefore be a rotation
Mϕ 0 . But this means [exercise] that

f = Mc ◦ Mϕ
0

is the composition of two rigid motions, and so is itself a rigid motion. Done.

7.7.2 Liouville’s Theorem

The constant mapping f(z) = c crushes the entire plane down to the single image
point c. We now ask whether it is possible for an analytic mapping to compress
410 Winding Numbers and Topology

the entire plane down to a region lying inside a finite circle, without going to the
extreme of completely crushing it to a point.
If we merely demand that the mapping be continuous then this can happen. For
example,
z
h(z) =
1 + |z|
maps the entire plane to the unit disc. Returning to analytic mappings, we notice
that complex inversion, z ÞÑ w = (1/z), manages to conformally compress the
infinite region lying outside the unit circle of the z-plane into the unit disc of
the w-plane. This looks quite hopeful: of the original plane only a puny unit disc
remains to be mapped.
To think like this is to completely forget the rigidity of analytic mappings. Hav-
ing decided to use complex inversion to map the region outside the unit circle, we
cannot change the rules when it comes to mapping the remaining disc: the mapping
z ÞÑ (1/z) acting in the exterior can only be analytically continued to the interior
in one way, namely as z ÞÑ (1/z). The requirement of analyticity thereby forces the
“puny” disc to explode, producing an image of infinite size.
We will now show that
An analytic mapping cannot compress the entire plane into a region
lying inside a disc of finite radius without crushing it all the way down
to a point.
This is Liouville’s Theorem. To understand this we must generalize Schwarz’s
Lemma slightly. Suppose that an analytic function w = f(z) leaves the origin fixed
and compresses the disc |z| ⩽ N to a region lying inside the disc |w| ⩽ M. By the
same reasoning as before, we find that if p lies inside the original disc (boundary
circle K) then
 
|f(z)| M
|F(p)| ⩽ [max |F(z)| on K] = max on K ⩽ .
N N
Hence,
M |p|
|f(p)| ⩽ .
N
But if f compresses the whole plane to a region lying inside the disc of radius M,
then the above result will continue to hold true no matter how large we make N.
Therefore f(p) = 0 for all p, and we are done.
Finally, if f does not leave the origin fixed, but instead sends it to c, we may apply
the previous argument to the function [f(z) − c]. This is the composition of f with
the translation which sends c back to 0. Since the image of the plane under f lies
inside the disc |w| ⩽ M, the translation of −c will produce a region lying inside the
disc |w| ⩽ 2M. The previous inequality then becomes
 The Schwarz–Pick Lemma* 411

2M |p|
|f(p) − c| ⩽ .
N
Once again letting N tend to infinity, we conclude that f(p) = c for all p. Done.

7.7.3 Pick’s Result

We now turn to a second, rather beautiful piece of evidence that non-Euclidean
geometry is intimately connected with the theory of conformal mapping. Recon-
sider [7.15]. Schwarz’s Lemma informs us that (with the exception of rotations) the
distance between interior points and the origin is decreased. This result has two
blemishes, both related to an exaggerated emphasis on the origin: (i) we require that
the mapping leave the origin fixed; (ii) only distances from the origin are shown to
decrease.
Consider (ii) first, and for the moment let us simply put up with (i) by continuing
to assume that our analytic mapping leaves the origin fixed. Although we did not
demonstrate it, perhaps a more symmetrical result holds true—with the exception
of a rotation, will the mapping automatically decrease the distance between any pair of
interior points?
Sadly, no. Consider the effect of f(z) = z2 (which leaves the origin fixed) on the
two interior points a = (3/4) and b = (1/2). The original separation is |a−b| = 0.25,
while the separation of the images is |f(a) − f(b)| = 0.3125. The distance between
the pair of points has increased.
But now consider the effect of exactly the same mapping on exactly the same two
points from the point of view of the Beltrami–Poincarites5 . When they measure the
distance between a and b, it is found [see (6.46), p. 366] to equal H{a, b} = 0.8473,
while the separation of the images is H{f(a), f(b)} = 0.7621. The hyperbolic distance
has decreased! Choose any other pair of points for yourself and examine the effect
of z ÞÑ z2 on their hyperbolic separation.
Pick’s splendid discovery6 was that even if we drop the requirement that the ori-
gin be a fixed point, this decrease in hyperbolic distance is a universal phenomenon:
Unless an analytic mapping of the disc to itself is a hyperbolic rigid motion,
(7.15)
the hyperbolic separation of every pair of interior points decreases.

5
Recall from Chapter 6 that this is the race of beings who inhabit the Beltrami–Poincaré model of
the hyperbolic plane.
6
Georg Alexander Pick [1859–1942] is perhaps best known for his wonderfully simple but com-
pletely unexpected Pick’s Theorem, which determines the area A of a lattice polygon in terms of the
number of lattice points Inside it (I), and on its Boundary (B), namely, A = I + (B/2) − 1. However,
his greatest contribution to science may have been the fact that he fought to appoint Einstein at the
German University of Prague in 1911, became close friends with him there—playing violin together
in a musical quartet—and finally introduced Einstein to the tensor calculus of Ricci and Levi-Civita,
without which Einstein could not have formulated General Relativity in 1915! [This is documented in
Goodstein (2018, p. 90), though, strangely, Einstein himself only seems to have publicly credited his
friend Marcel Grossmann for this critical piece of guidance.] Pick died at the hands of the Nazis at
Theresienstadt concentration camp on the 26th of July, 1942, at the age of 82.
412 Winding Numbers and Topology

Because this result contains Schwarz’s Lemma as a special case [we shall clarify
this shortly] it is often called the Schwarz–Pick Lemma. Despite the startling nature
of the result, we can actually understand its essence very simply; we need only ask
the question, “How do the Beltrami–Poincarites view [7.15]?”
Because their concept of angle is identical to ours, it follows that their concept
of an analytic function is also the same as ours—f appears conformal both to us
and to them. In addition, we both agree that rays emanating from the origin are
straight lines along which we may measure distance. Consequently, the Beltrami–
Poincarites willingly concede that w is closer to 0 than z is, although they violently
disagree with our quantitative determination of exactly how much closer it is. Now
recall that there is a small (psychological) flaw in the Beltrami–Poincaré model:
0 appears special to us because it is the centre of the disc, but to the Beltrami–
Poincarites who inhabit an infinite, homogeneous plane 0 is utterly indistinguishable
from any other point of their world.
The above explanation is formalized in [7.16]. In the top left figure we see that
the Beltrami–Poincarites have marked an arbitrary point a, drawn a few concen-
tric circles centred there, and on the outermost of these they have marked a second
point b. They (and we) now consider the effect of an analytic mapping f of their
world to itself. The point a is sent to some image point A = f(a) [top right] and
likewise b is sent to B = f(b). In order to compare the separation of A and B
with that of a and b, the Beltrami–Poincarites perform a rigid motion (MA ◦ Ma
would do) that moves the circles centred at a to circles (of equal hyperbolic size)
centred at A. Consequently, the hyperbolic separation of a and b will have been
decreased [increased] by f according as B lies inside [outside] the outermost of
these circles. In anticipation of Pick’s result we have drawn it inside, corresponding
to a decrease in hyperbolic separation. However, observe that as in our previous
numerical example, to us Euclideans it looks as though the separation has been
increased.
In order to show us poor blind Euclideans that the circles centred at a and A
really are concentric and of equal sizes (so enabling us to see that B really has got-
ten closer) the Beltrami–Poincarites perform the illustrated rigid motions Ma and
MA . These respectively move a and A to the origin [bottom left and bottom right
figures], yielding circles that are as concentric to us as they always were to them.
Ma moves b to z = Ma (b), while MA moves B to
w = MA (B) = (MA ◦ f)(b) = (MA ◦ f ◦ Ma )(z).
We shall abbreviate (MA ◦ f ◦ Ma ) to F, so that w = F(z).
We can now see that the following are all equivalent:
H{A, B} < H{a, b} ⇐⇒ H{0, w} < H{0, z} ⇐⇒ |w| = |F(z)| < |z|.
But F is an analytic mapping of the disc to itself which leaves the origin fixed, and
so it is subject to Schwarz’s Lemma. Thus unless F is a rotation—in which case
 The Schwarz–Pick Lemma* 413

[7.16] The Schwarz–Pick Lemma is Pick’s brilliant generalization of Schwarz’s Lemma,

stating that the hyperbolic separation of every pair of points must decrease. The
Beltrami–Poincarites (our imaginary inhabitants of the Beltrami–Poincaré disc) have no
way of distinguishing the bottom half of this figure [aka the already-established Schwarz’s
Lemma] from the top half, which is Pick’s result. Q.E.D.!

f = (MA ◦ F ◦ Ma ) is a rigid motion—we must have |w| = |F(z)| < |z|, as depicted.
Done.
Finally, let us express the Schwarz–Pick Lemma in symbolic form. If f is not a
rigid motion then |F(z)| < |z|, which may be written out more explicitly as

|(MA ◦ f ◦ Ma )(z)| < |z|,

which in turn can be written as

|(MA ◦ f)(b)| < |Ma (b)|.

Thus
B−A b−a
< .
AB − 1 ab − 1

If we move b closer and closer to a, then da = (b − a) becomes an infinitesimal

vector emanating from a whose image under f is an infinitesimal vector dA =
(B − A) emanating from A. The above inequality now becomes
414 Winding Numbers and Topology

|dA| |da|
< ,
1 − |A| 2 1 − |a|2

which we may interpret [cf. (6.44), p. 363] as saying that, provided f is not a rigid
motion, the hyperbolic length of dA is less than that of its preimage da. This is the
infinitesimal version of (7.15).

7.8 The Generalized Argument Principle

7.8.1 Rational Functions

We have now seen that there are many powerful and surprising consequences of
the Topological Argument Principle as restricted to analytic functions. Still others
are described in the exercises. However, in all our previous work we have only
examined mappings which are free of singularities in the region under considera-
tion. We now lift this restriction and find that there is a generalization of (7.6) which
applies to this case also.
We began our discussion of the Argument Principle by looking at the pro-
totypical analytic functions without singularities—the polynomials. In order to
understand the generalization to analytic functions with singularities, we should
correspondingly begin with rational functions.
As in [7.8], let A, B, and C be the complex numbers from the fixed points a, b,
and c to the variable point z. The left-hand side of [7.17] shows an expanding circle
Γ at three successive stages of its growth: Γ1 , Γ2 , and Γ3 . The right-hand side shows
the evolution of the image of Γ under the rational mapping

(z − a) (z − b) 1
f(z) = = AB · . (7.16)
(z − c) C

By the time Γ has grown into Γ1 it has enclosed a, and ν [f(Γ1 ), 0] = 1,

in accord with the ordinary Argument Principle. As Γ continues to grow it
crosses the other 0-point at b, and the winding number of f(Γ ) correspondingly
increases to ν [f(Γ2 ), 0] = 2. Now comes the new phenomenon. As Γ crosses
the singularity at c the winding number of its image decreases by one so that
ν [f(Γ3 ), 0] = 1.
The explanation is simple. As z traverses Γ3 , the winding number of f(z) is
the sum of the revolutions executed separately by A, B, and (1/C). The first two
go round once, but as C rotates counterclockwise, (1/C) rotates in the opposite
direction, finally executing one complete negative revolution. By the same token,
if the denominator of f instead contained Cm then (1/Cm ) would execute −m
revolutions, and the winding number would become

ν [f(Γ3 ), 0] = 2 − m.
 The Generalized Argument Principle 415

[7.17] The Generalized Argument Principle. Consider f(z) = (z − a) (z − b) · 1

=
(z − c)
A B · C . Traversing Γ1 , A executes a complete revolution, while B and C merely oscillate,
1

so ν[f(Γ1 ), 0] = 1. Expanding Γ to enclose b, both A and B now execute complete rev-

olutions, while C merely oscillates, so ν[f(Γ2 ), 0] = 2. But when Γ expands still further
to Γ3 , enclosing the pole at c (marked by the explosion!) then (1/C) executes a negative
revolution, decreasing the winding number by one: ν[f(Γ3 ), 0] = 1. In general, if N and
M are the number of interior p-points and poles, both counted with their multiplicities,
then ν[f(Γ ), p ] = N − M.

As with counting zeros, we could say in this case that c was a singularity [or pole,
as we shall now call such places] of multiplicity m.
The previous equation is an example of the Generalized Argument Principle:

Let f be analytic on a simple loop Γ and analytic inside except for a finite
number of poles. If N and M are the number of interior p-points and (7.17)
poles, both counted with their multiplicities, then ν [f(Γ ), p ] = N − M.

Simply by allowing an arbitrary number of factors on the top and bottom of (7.16)
we see that this result is certainly true when f is any rational function.
Before explaining why it works in general, let us develop a more vivid under-
standing of how it works in the case of our example (7.16). We have certainly shown
that as Γ crosses c the winding number drops from 2 to 1, but exactly how does this
unwinding occur?
If we look at the image plane just as Γ crosses c then f(Γ ) undergoes a sudden
and violent change of shape as it leaps to infinity and then returns, but this leaves
us none the wiser. However, if we instead watch its evolution on the Riemann sphere
then we gain a new and delightful insight into the process.
Figure [7.18] (which should be scanned like a comic strip) illustrates this. At
the time of the first picture [top left] Γ has already enclosed the two roots, and its
image is seen to wind round the origin twice. Now follow the evolution of f(Γ )
416 Winding Numbers and Topology

[7.18] Reducing the winding number by passing over the point at infinity. The Riemann
sphere provides a vivid explanation of the previous figure. As Γ expands from Γ2 to Γ3 ,
crossing the pole at c, the loop on the Riemann sphere simply slides over the north pole,
unwinding the loop and reducing its winding number from 2 to 1.

through the remaining pictures. As Γ crosses c [top right] there is no longer any
excitement—f(Γ ) merely slides across the north pole, and this is how the unwinding
is achieved. Try using a computer to animate the evolution of the image f(Γ ) on the
Riemann sphere as Γ expands through the roots and poles of a rational function f
of your choosing.

7.8.2 Poles and Essential Singularities

In generalizing the ordinary Argument Principle we had to ask ourselves how
we should count p-points of a general analytic function. The factorization (7.7)
brought out the analogy with polynomials and gave us a satisfactory definition
of the algebraic (and topological) multiplicity of a p-point.
The method of extending (7.17) from rational functions to analytic functions with
singularities is essentially the same. The only complication is that there are actually
two possible kinds of singularity for an otherwise analytic function.
The first kind of singularity is called a pole. It is by far the most commonly
encountered type in applications of complex analysis, and it is the only type to
which (7.17) applies. Here’s the definition. If f(z) approaches ∞ as z approaches
a from any direction then a is a pole of f. We can understand the terminology by
 The Generalized Argument Principle 417

thinking of the modular surface of f, for there will be an infinitely high spike or
“pole” above the point a. Figure [2.14] on p. 74 is an example of this.
Since f is analytic, it follows that F(z) ≡ [1/f(z)] is also analytic and has a root at
a. If this root has multiplicity m then the factorization (7.7) of F is

F(z) = (z − a)m Ω(z), (7.18)

where Ω is analytic and nonzero at a; in fact we know that Ω(a) = F(m) (a)/m!.
The local behaviour of f near a is therefore given by
e
Ω(z)
f(z) = , (7.19)
(z − a)m
e
where Ω(z) = [1/Ω(z)] is analytic and nonzero at a. This expression brings out the
analogy with rational functions and enables us to identify m as the algebraic mul-
tiplicity or order of the pole at a. We call a pole simple, double, triple, etc., according
as m = 1, 2, 3, etc.
Note that we have also found a way of calculating the order of a pole, namely,
as the order of the first nonvanishing derivative of (1/f). Once you have identified
the locations of the poles, you may use this method [exercise] to find the orders of
the poles of the following functions:
1 cos z 1
P(z) = ; Q(z) = ; R(z) = .
sin z z2 (ez − 1)3
You should have found that P has a simple pole at each multiple of π; Q has a
double pole at 0; and R has a triple pole at each multiple of 2πi.
One more piece of terminology. If the only singularities in some region of an
otherwise analytic function are poles, the function is called meromorphic in that
region.
In addition to poles, it is also possible for an otherwise analytic function to pos-
sess what are called essential singularities. We shall postpone detailed discussion of
such places to a later chapter, but it is clear that the behaviour of a function f in
the vicinity of an essential singularity s must be very strange and wild. If f were
bounded in the vicinity of s then s would not be a singularity at all, but on the
other hand f(z) cannot approach ∞ as z approaches s from all directions, for then
s would only be a pole.
Consider the standard example g(z) = e1/z , which clearly has a singularity of
some type at the origin. If we write z = r eiθ then
cos θ
|g(z)| = e r .

Figure [7.19] depicts the modular surface. If z approaches 0 along the imaginary
axis then |g(z)| = 1. But if the approach is instead made along a path lying to the
left of the imaginary axis (where cos θ < 0) then g(z) tends to 0. Finally, if the
418 Winding Numbers and Topology

[7.19] The modular surface of g(z) = e1/z reveals its essential singularity at the origin.
cos θ
Since |g(z)| = e r , the behaviour at 0 depends drastically upon the direction in which
we approach it. If cos θ < 0 as we approach, then the height |g(z)| drops to zero, and
the surface hits C. If we approach precisely along the tightrope of the imaginary axis (for
which cos θ = 0) then our height holds precariously steady at 1. But if cos θ > 0 as we
approach, then the height explodes to infinity faster than any polynomial!

approach path lies to the right of the imaginary axis then g(z) tends to ∞. In fact,
not only will |g(z)| become infinite in this case, but the rate at which it zooms off to
∞ is quite beyond the ken of any pole.
To see this, reconsider (7.19). The greater the order m, the faster the growth of f
as the pole at a is approached. However, no matter how great the order happens to
be, we know that (z − a)m dies away fast enough to kill this growth, in the sense
that its product with f remains bounded. Indeed, the order of a pole can be defined
as the smallest power of (z − a) which will curb the growth of f in this way.
Compare this with the growth of g(z) as its essential singularity is approached,
say along the positive real x-axis. To confirm that g grows faster than any mero-
morphic function, we need only recall from ordinary calculus that

eλ
lim xm e1/x = lim = ∞,
xÑ0 λÑ∞ λm

no matter how great the value of m.

 The Generalized Argument Principle 419

7.8.3 The Explanation*

In order to explain (7.17) let us return to the interpretation of (7.19). If we think of f
as mapping into the Riemann sphere Σ, then the north pole (∞) is an image point
like any other, and the poles of f are simply its preimages, ∞-points if you will. As
we now explain, this means that the topological multiplicity of an ∞-point can be
defined in exactly the same way as that of any other p-point, namely, as the number
of times that the image of a small loop round a winds round f(a).
Reconsider the mapping F in (7.18). By virtue of (7.8), we know that a sufficiently
small circle Ca centred at a will be mapped to a small loop F(Ca ) winding round
the origin m times. On Σ, the stereographic projection of F(Ca ) therefore winds
round the south pole m times, counterclockwise as seen from inside Σ. Because
complex inversion (which sends F(Ca ) to f(Ca ) = 1/[F(Ca )]) rotates Σ about the
real axis by π, thereby swapping 0 and ∞, this means that f(Ca ) will be a small loop
winding m times round ∞. Since it winds counterclockwise as seen from inside Σ,
its stereographic projection in the plane is therefore a very large loop winding m
times clockwise around the origin, i.e., with winding number −m.
As an aside, observe that it now makes sense to rewrite (7.19) as

e
f(z) = (z − a)−m Ω(z),

and to correspondingly think of a pole of order m as being a root of negative

multiplicity −m.
Shifting our attention away from the origin, we now consider winding numbers
around an arbitrary (finite) point p. By making Ca sufficiently small, we can be
certain that f(Ca ) will be so large that it will wind −m times round p. But if we
expand Ca into any simple loop Γa without crossing any p-points or other poles,
then the winding number of the image round p cannot change. In other words,

If a is a pole of order m and Γa is any simple loop containing a but no

(7.20)
p-points and no other poles, then ν [f(Γa ), p ] = −m.

Finally, reconsider figure [7.11]. You may now easily convince yourself that the
argument leading to (7.11) remains valid if some of the aj ’s are poles instead of
p-points. Let’s call these singular points sj . Thus
X X
ν[f(Γ ), p ] = ν [f(Γaj ), p ] + ν [f(Γsj ), p ] .
p-points poles

Using (7.20), this implies that

ν[f(Γ ), p ] = [number of p-points inside Γ ] − [number of poles inside Γ ],

as was to be shown.
420 Winding Numbers and Topology

7.9 Exercises

1 A “simple” loop can get very complicated (see diagram). However, if we imag-
ine creating this complicated loop by gradually deforming a circle, it is clear
that it will wind round its interior points precisely once. Let N(p) be the num-
ber of intersection points of the simple loop with a ray emanating from p
(cf. [7.4]). What distinguishes the possible values of N(interior point) from
those of N(exterior point)? In place of the crossing rule (7.1), you now possess
(for simple loops) a much more rapid method of determining whether a point
is inside or outside.

You can use this result to play a trick on a friend F: (1) So that foul play cannot be
suspected, get F to draw a very convoluted simple loop for himself; (2) choose a
random point in the thick of things and ask F if it’s inside or not, i.e., starting at
this point, can one escape through the maze to the outside?; (3) after F has been
forced to recognize the time and effort required to answer the question, get him
to choose a point for you; (4) choosing a ray in your mind’s eye, scan along it
and count the intersection points. Amaze F with your virtually instantaneous
answer!
2 Reconsider the mapping L b in (7.4) of the unit circle to itself, and the associated
graph of Φ(θ) in [7.7]. If Φ ′ (a) > 0 then the graph is rising above the point θ =
a, and small movement of z will produce a small movement of the w having the
same sense. We say that Φ is orientation-preserving at a and that the topological
multiplicity ν (a) of z = eia as a preimage of w = eiΦ(a) is +1. Similarly, if
Φ ′ (a) < 0 then the mapping is orientation-reversing and ν (a) = −1. In other
words,
ν (a) = the sign of Φ ′ (a).
Compare this with the 2-dimensional formula (7.9).
(i) In [7.7], explain how the complete set of preimages of w = eiA can be found
by drawing the family of horizontal lines Φ = A, A ± 2π, A ± 4π, etc.
 Exercises 421

(ii) If the set of preimages is typical in the sense that Φ ′ ̸= 0 at any of them,
what do we obtain if we sum their topological multiplicities? Thus to say
that the degree of Lb (the winding number of L) is ν is essentially to say
b is ν-to-one. [Hint: In [7.4], consider a ray as describing the location
that L
of w.]
3 For each of the following functions f(z), find all the p-points lying inside the
specified disc, determine their multiplicities, and by using a computer to draw
the image of the boundary circle, verify the Argument Principle.
(i) f(z) = e3πz and p = i, for the disc |z| ⩽ (4/3).
(ii) f(z) = cos z and p = 1, for the disc |z| ⩽ 5.
(iii) f(z) = sin z4 and p = 0, for the disc |z| ⩽ 2.
4 Reconsider [7.8].
(i) Use a computer to draw the image under a cubic mapping

f(z) = (z − a) (z − b) (z − c)

of an expanding circle Γ , and observe the manner in which the winding

number increases as Γ passes through the roots a, b, and c. In particular,
observe that the shape that marks the birth of a new loop is this: ≺.
(ii) If f ′ (p) ̸= 0 then a little piece of Γ passing through p is merely
amplitwisted to another almost straight piece of curve through f(p).
Deduce that ≺ shapes can only occur when Γ hits a critical point. Explain
why the particular shape ≺ is consistent with a critical point of order 1.
(iii) Observe that there are only two point in the evolution of Γ at which a ≺
shape is produced. Explain this algebraically in terms of the degree of f ′ .
(iv) Let T be the triangle with vertices a, b, and c. There are many ellipses
which can be inscribed in T so as to touch all three sides, but show that
there is only one (call it E) that touches T at the midpoints of the sides.
Hard Show that the two critical points of f are the foci of E !
5 As in the text, let ξa , ηa , ϕa denote the two perpendicular expansion factors
and the rotation angle used to describe the SVD decomposition of the local
linear transformation at a produced by a mapping. By considering the case of
a rotation by (π/4), for which J is constant, show that ξ and η are generally not
the eigenvalues λ1 and λ2 of the Jacobian J. However, confirm for this example
that det(J) = λ1 λ2 = ξη.
6 Even in three or more dimensions the local linear transformation induced by a
mapping f at a point a can still be represented by the Jacobian matrix J(a), and
if a is not a critical point then its topological multiplicity ν (a) as a preimage
422 Winding Numbers and Topology

of f(a) is still given by (7.9). If n is the number of real negative eigenvalues of

J(a), counted with their algebraic multiplicities, show that

ν (a) = (−1)n .

[Hint: Since the characteristic equation det[J(a) − λ I] = 0 has real coefficients,

any complex eigenvalues must occur in conjugate pairs.]
7 Consider the nonanalytic mapping h(z) = |z|2 − i z.
(i) Find the roots of h.
(ii) Calculate the Jacobian J, and hence find det(J).
(iii) Use (7.9) to calculate the multiplicities of the roots in (i).
(iv) Find the image curve traced by h(z) as z = 2 eiθ traverses the circle |z| = 2,
and confirm the prediction of the Topological Argument Principle.
(v) Gain a better understanding of the above facts by observing that h(z) =
z (z − i), and then mimicking the analysis of [7.8].
(vi) Use the insight of the previous part to find ν (i/2), which cannot be done
with (7.9).
8 Let Q(t) be a real function of time t, subject to the differential equation

dn Q dn−1 Q dQ
cn n
+ cn−1 n−1
+ · · · + c1 + c0 Q = 0.
dt dt dt
Recall that one solves this equation by taking a linear superposition of special
solutions of the form Qj (t) = esj t . Substitution into the previous equation
shows that the sj ’s are the roots of the polynomial

F(s) ≡ cn sn + cn−1 sn−1 + · · · + c1 s + c0 .

Note that Qj (t) will decay with time if sj has a negative real part. The issue of
whether or not the general solution of the differential equation decays away
with time therefore reduces to the problem of determining whether or not all n
roots of F(s) lie in the half-plane Re(s) < 0. Let R be the net rotation of F(s) as s
traverses the imaginary axis from bottom to top. Explain the following result:
The general solution of the differential equation will die away if and only if

R = nπ.

This is called the Nyquist Stability Criterion, and an F that satisfies this condition
is called a Hurwitz polynomial. [Hints: Apply the Argument Principle to the loop
consisting of the segment of the imaginary axis from −iR to +iR, followed by
one of the two semicircles having this segment as diameter. Now let R tend to
infinity.]
 Exercises 423

9 Referring to the previous exercise, consider the differential equation

d3 Q
− Q = 0.
dt3
(i) Find R for this equation. Does it satisfy the Nyquist Stability Criterion?
(ii) Confirm your conclusion by explicitly solving the differential equation.
10 If a is real and greater than 1, use Rouché’s Theorem to show that the equation

zn e a = e z

has n solutions inside the unit circle.

11 (i) Applying Rouché’s Theorem to f(z) = 2z5 and g(z) = 8z − 1, show that all
five solutions of the equation 2z5 + 8z − 1 = 0 lie in the disc |z| < 2.
(ii) By reversing the roles of f and g, show that there is only one root in the
unit disc. Deduce that there are four roots in the ring 1 < |z| < 2.
12 We can formalize, and slightly generalize, our explanation of Rouché’s
Theorem as follows:
(i) If p(z) and q(z) are nonzero on a simple curve Γ , and e
Γ is the image curve
under z ÞÑ p(z) q(z), show that

ν [e
Γ , 0] = ν [p(Γ ), 0] + ν [q(Γ ), 0].

(ii) Write
 
g(z)
f(z) + g(z) = f(z) 1 + = f(z) H(z).
f(z)
If |g(z)| < |f(z)| on Γ , sketch a typical H(Γ ). Deduce that

ν [H(Γ ), 0] = 0.

Using the previous part, obtain Rouché’s Theorem.

(iii) If we only stipulate that |g(z)| ⩽ |f(z)| on Γ , then parts of H(Γ ) could actually
coincide with the circle |z − 1| = 1, rather than lying strictly inside it, and
ν [H(Γ ), 0] might not be well-defined. However, show that if we further
stipulate f + g ̸= 0 on Γ , then ν [H(Γ ), 0] = 0 as before. Deduce that ν[(f +
g)(Γ ), 0] = ν[f(Γ ), 0].
13 Let w = f(z) be analytic inside and on a simple loop Γ , and suppose that f(Γ )
is an origin-centred circle.
(i) If ∆ is an infinitesimal movement of z along Γ and ϕ is the correspondingly
infinitesimal rotation of w, show geometrically that
f′ ∆
= iϕ.
f
424 Winding Numbers and Topology

(ii) As z traverses Γ , explain why ν [∆, 0] = 1 and ν [iϕ, 0] = 0.

(iii) Referring to (i) of the previous exercise, show that

ν [f(Γ ), 0] = ν [f ′ (Γ ), 0] + 1.

(iv) Deduce from the Argument Principle that f has one more root inside Γ
than f ′ has. This is sometimes called Macdonald’s Theorem, though I believe
its essence goes back as far as Riemann.
(v) From this we deduce, in particular, that f has at least one root inside Γ .
Derive this fact directly by considering the portion of the modular surface
lying above Γ and its interior.
14 In contrast to analytic mappings, it is perfectly possible for a continuous non-
analytic mapping to completely crush pieces of curve or even areas without
crushing the rest of its domain. Let us give a concrete example to show that the
Topological Argument Principle does not apply to this case. With r = |z|, the
mapping h(z) = ϕ(r) z will be a continuous function of z if ϕ(r) is a continu-
ous function of r. Consider the continuous mapping h of the unit disc to itself
corresponding to
ϕ(r) = 0, 0 ⩽ r < (1/2);
ϕ(r) = 2r − 1, (1/2) ⩽ r ⩽ 1.

(i) Describe this mapping in visually vivid terms.

(ii) Taking Γ to be the circle |z| = (3/4) and letting p = 0, try (and fail) to make
sense of (7.11).
15 The version of Brouwer’s Fixed Point Theorem established in the text fell short
of the full result in two ways: (A) we assumed that |g| < 1 on D rather than
|g| ⩽ 1; (B) we essentially used the Topological Argument Principle, which
the previous exercise shows to be useless in the general case of a continuous
mapping having infinitely many p-points in a finite region. Let’s remove these
blemishes. Once again let m(z) = g(z) − z be the movement of z, and suppose
that Brouwer’s result is false, so that m ̸= 0 throughout the disc |z| ⩽ 1. Obtain
the desired contradiction as follows:
(i) By assumption, m(z) = g(z) − z = g(z) + f(z) does not vanish on the unit
circle C. Use Ex. 12(iii) to show that if |g| ⩽ 1 then ν [m(C), 0] = 1.
(ii) Let Cr be the circle |z| = r, so that C1 = C. By considering the evolution of
ν [m(Cr ), 0] as r increases from 0 to 1, obtain a contradiction with (i).
The key fact ν [m(C), 0] = 1 can be obtained more intuitively. Draw a typical
movement vector m(z) emanating from z and note that it makes an acute angle
with the inward unit normal vector (−z) to C, also drawn emanating from z.
 Exercises 425

But clearly this normal vector undergoes one positive revolution as z traces C.
Deduce that the vector m is also dragged round one revolution.
16 Let f(z) be an odd power of z, and consider its effect on the unit circle C. Note
two facts: (1) if p is on C then f(−p) points in the opposite direction to f(p); (2)
ν [f(C), 0] = odd, in particular it cannot vanish. This is only one example of a
general result. Show that (1) always implies (2), even if f is merely continuous.
[Hints: If f is subject to (1), what can we deduce about the net rotation R of f(z)
as z traverses the semicircle from p to −p? How is the rotation produced by the
remaining semicircle related to R ?]
17 Consider a spherical balloon S resting on a plane. If we gradually deflate S, each
point will end up on the plane so that we have a continuous mapping H of S into
the plane. Observe that the north and south poles, which are antipodal, have
the same image. The Borsuk-Ulam Theorem says that any continuous mapping
H of S into the plane will map some pair of antipodal points to the same image.
Consider the mapping F(p) = H(p) − H(p⋆ ), where p⋆ is antipodal to p. The
theorem then amounts to showing that F has a root somewhere on S. Prove this.
[Hints: It is sufficient to examine the effect of F on just the northern hemisphere.
By taking the boundary of this hemisphere (the equator) to be the circle C of
the previous exercise, deduce that ν [H(C), 0] ̸= 0.]
18 Let f be analytic on a simple loop Γ , and let p be a preimage of a point on f(Γ )
at which |f| is maximum. If ξ is a tangent complex number to Γ at p, and in the
same counterclockwise sense as Γ , show geometrically that ξf ′ (p) points in the
same direction as if(p). What is the analogous result at a positive minimum of
|f| ?
19 (i) If p is not a critical point of an analytic function f, show geometrically that
the modulus of f increases most rapidly in the direction [f(p)/f ′ (p)].
(ii) In terms of the modular surface above p, this direction lies directly beneath
the tangent line to the surface having the greatest “slope” (i.e., tan of the
angle it makes with the complex plane). Show that the slope of this steep-
est tangent plane is |f ′ (p)|, and note the analogy with the slope of the
ordinary graph of a real function.
(iii) What does the modular surface look like at a root of order n ?
(iv) What does the modular surface look like above a critical point of order
m ? Using the case m = 1, explain why such places are called saddle points.
(v) Rephrase Macdonald’s Theorem [Ex. 13] in terms of the number P of pits
and the number S of saddle points in the portion A of the modular surface
lying above Γ and its interior.
426 Winding Numbers and Topology

(vi) It is a beautiful fact that, expressed in this form, Macdonald’s Theorem

can be explained almost purely topologically. The following explanation
is adapted from Pólya (1954), though I believe the basic idea goes back
to Maxwell and Cayley. Since |f| is constant on Γ , the boundary of A is a
horizontal curve K, and since f is analytic, K is higher than the rest of A.
Also, recall that there are no peaks. Suppose for simplicity that f and f ′
have only simple roots (P and S in number) so that the pits are cone-like,
and the saddle points really look like saddles or (more geographically)
like mountain passes.
Now imagine a persistent rain falling on the surface A. The pits grad-
ually fill with water and so become P lakes, the depths of which we shall
imagine are always equal to each other. What happens to the number of
lakes as the water successively rises past each of the S passes? How many
lakes are left by the time that the water has finally risen to the level of K?
As required, deduce that
P = S + 1.
(vii) Generalize the above argument to roots and critical points that are not
simple.
20 Let f(z) and g(z) be analytic inside and on a simple loop Γ . By applying the
Maximum Modulus Theorem to (f − g), show that if f = g on Γ then f = g
throughout the interior.
21 Let R(L) be the net rotation of f(z) round p as z traverses a loop L. For example,
if L does not contain p then
1
ν [L, p] = R(L).
2π
By taking this formula to be the definition of ν, make sense of the statement
that the Generalized Argument Principle (7.17) remains valid even if there are
some poles and p-points on Γ , provided that we count these points with half
their multiplicities.
22 In [7.11] we used the idea of deformation to derive the argument principle. The
figure below shows another method. The interior of Γ , containing various p-
points and poles, has been crudely partitioned into cells Cj in such a way that
each one contains no more than a single p-point or a single pole. Now think of
each cell as being a loop traversed in the conventional direction; this sense is
indicated for two adjacent cells in the figure.
(i) What is the value of ν [f(Cj ), p ] if Cj (1) is empty; (2) contains a p-point of
order m; (3) contains a pole of order n ?
 Exercises 427

(ii) Obtain the Argument Principle by showing that

X
ν [f(Cj ), p ] = ν [f(Γ ), p ].
j

[Hint: If an edge of a cell does not form part of Γ then it is also an edge of an
adjacent cell, but traversed in the opposite direction. What is the net rotation of
f(z) round p as z traverses this edge once in one direction, then in the opposite
direction?]
CHAPTER 8

Complex Integration: Cauchy’s Theorem

8.1 Introduction

In the last few chapters our efforts to extend the idea of differentiation to complex
mappings have been amply rewarded. By innocently attempting to generalize the
real derivative we were quickly led to the amplitwist concept, and the subject then
came to life with a character all its own. While many of the results cast familiar
shadows onto the world of the reals, many did not, and striking indeed was the
flavour of the arguments used to grasp them. The ability of z to freely roam the
plane unleashed in us a degree of visual imagination that had to remain dormant so
long as we could only watch the real number x forlornly pacing its one-dimensional
prison.
This little hymn to the glory of the complex plane can be sung again in the context
of integration, only louder. If differentiation breathed life into the subject, then inte-
gration could be said to give it its soul. Only after we have understood this soul will
we be able to demonstrate such fundamental facts as the infinite differentiability
of analytic mappings1 .
Rb
In ordinary calculus the symbol a has a clear meaning. However, if we wish to
generalize this to C then the need for new ideas is immediately apparent, for how
are we to get from a to b? In R there was only one way, but a and b are now points in
the plane, so we must specify some connecting path (called a contour) “along which
to integrate”. It is then natural to ask whether the value of the integral depends
upon the choice of this contour.
In general the value of the integral will depend on the route chosen. For example,
we will shortly meet an integral of a complex mapping that yields, when evaluated
for a closed contour, the area enclosed by the contour—a flagrant dependence of

1
Since the 1960’s it has actually become possible to do such things without integration, thanks to
pioneering work by G. T. Whyburn, and others. Nevertheless, integration still appears to provide the
simplest approach.

Visual Complex Analysis. 25th Anniversary Edition. Tristan Needham, Oxford University Press.
© Tristan Needham (2023). DOI: 10.1093/oso/9780192868916.003.0008
430 Complex Integration: Cauchy’s Theorem

value on contour. It should be made clear from the outset that while differentiation
only made sense for the strictly limited set of analytic functions, this is not the case
for integration. Indeed, the example just cited involves the integration of a non-
analytic function.
The principal aim of this chapter (beyond the mere construction of an integral
calculus) will be the discovery of conditions under which the value of an integral
does not depend on the choice of contour. One such result is an analogue of the
Fundamental Theorem of real analysis, and in deference to that subject it bears the
same name. However, in the complex realm this is actually a misnomer, for there
exists a still deeper result which has no counterpart in the world of the reals. It is
called Cauchy’s Theorem.
As we have said, it is not only possible, but sometimes useful to integrate non-
analytic functions. However, it should come as no surprise to learn that new
phenomena arise if we concentrate on the integrals of mappings that are analytic.
Cauchy’s Theorem is the essence of these new phenomena. Essentially it says that
any two integrals from a to b will agree, provided that the mapping is analytic
everywhere in the region lying between the two contours. Almost all the fundamental
results of the subject (including some already stated) flow from this single horn of
plenty.

8.2 The Real Integral

8.2.1 The Riemann Sum

As we did with differentiation, we begin by re-examining the more familiar idea of
integrating a real function. The historical origin of this process, and still the princi-
pal means of visualizing it, is the problem of evaluating the area under the graph
of a function.
We first approximate the sought-after area with rectangles. See [8.1]. Dividing
the interval of integration into n line-segments ∆j (the bases of the rectangles), we
randomly select one point xj from each segment, and take the height of the corre-
sponding rectangle to be the height of the curve above the point, namely, f(xj ). The
area of each rectangle is then f(xj ) ∆j , and thus the total rectangular approximation
to the area under f is
X
n
R≡ f(xj ) ∆j . (8.1)
i=1

The quantity R is called a Riemann sum. Finally, by simultaneously letting n tend to

infinity while each ∆j shrinks to nothing, R will tend to the desired area.
In [8.1] we could afford to be indifferent to the precise choice of xj within each
∆j because we had our eye on this final limiting process. As each ∆j shrinks, the
 The Real Integral 431

Pn
[8.1] A Riemann Sum R ≡ j=1 f(xj ) ∆j approximates the area under a curve with
rectangles, the exact answer being obtained in the limit that n goes to infinity and each
∆j goes to zero.

freedom in the choice of xj becomes more and more limited, and the influence of
the choice on the area of the rectangle likewise diminishes. However, if we are
unwilling or unable to actually carry out the limiting process, then, as we shall
now see, we can ill afford to be so blasé in our choice of xj .
You probably dimly remember some professor showing you (8.1) before, and
perhaps you even evaluated a couple of examples by means of it. However, this
was no doubt quickly forgotten once you set eyes upon the Fundamental Theorem
of Calculus. In order to integrate x4 , why bother with taking the limit of some
complicated series when we know that the answer must be that function which
differentiates to x4 , namely, 51 x5 ?
The Fundamental Theorem is a wonderful thing, but one must remem-
ber that many quite ordinary functions simply do not possess an antideriva-
tive that is expressible in terms of elementary functions. To take a simple
example, the Normal Distribution of statistics requires a knowledge of the area
2
under the curve e−x , and this can only be computed numerically, perhaps via a
Riemann sum.
When doing a numerical calculation with (8.1), it would require an infinite
amount of time to find even a single area with perfect precision. It is therefore
important to be able to obtain good approximations to limnÑ∞ R while using only
a finite value of n. Several such methods exist: Simpson’s rule and the Trapezoidal
432 Complex Integration: Cauchy’s Theorem

rule, to name just two that may be familiar. Since the Trapezoidal rule will most
readily lend itself to complex generalization, we will now review it.

8.2.2 The Trapezoidal Rule

As the name suggests, we now use trapezoids instead of rectangles to approximate
the area. Though not strictly necessary, it is convenient to make all the ∆j the same
length. See [8.2].
It is clear from the figure that even a very modest value of n will yield a quite
accurate estimate. Since [8.2] is not of the same form as [8.1], the associated Trape-
zoidal Formula (which we won’t bother to state) is not quite of the type (8.1).
Nevertheless, if we wish to continue to use (8.1), it is not hard to find a Riemann
sum that closely mimics the trapezoidal sum, and hence which retains the latter’s
accuracy.
First note that the shaded trapezoidal estimate shown in [8.2] is identical to the
rectangular one in [8.3], in which we have taken the height of each rectangle to be
the height of the chord at the midpoint of ∆j . Finally, to recover a Riemann sum,
we can replace the height of the chord by the height of the curve at that point.
See [8.4].
In other words, the Riemann sum (8.1) will yield an accurate approximation to
the integral, using only a modest value of n, provided that we choose each xj to
be at the midpoint of its ∆j . We will call this the Midpoint Riemann Sum, and write
it as RM .

[8.2] The Trapezoidal Rule replaces the rectangles with trapezoids, and clearly yields an
accurate approximation to the integral while using only a modest value of n.
 The Real Integral 433

[8.3] The exact same area as the previous figure, but now represented as rectangles whose
heights are given by the midpoints of the illustrated chords.

[8.4] The Midpoint Riemann Sum, RM , mimics the previous figure, but now uses the
height of the graph at each midpoint, instead of the height of the chord. As an approxi-
mation to the integral, a typical Riemann sum that uses the height at a random point has
a total error that dies away as ∆, but we will show that the Midpoint Riemann Sum is
much more accurate, with an error that dies away as ∆2 .
434 Complex Integration: Cauchy’s Theorem

8.2.3 Geometric Estimation of Errors

We have said that using midpoints in (8.1) will yield accurate results, but how accu-
rate is “accurate”? First reconsider the case where the xj were chosen randomly,
and suppose that all the ∆j ’s have the same length. Re-examination of [8.1] reveals
that the difference between the actual area lying above each ∆j , and the area of the
approximating rectangle, will be of order ∆2 . Since the total number of rectangles is
of order 1/∆, it follows that the total error will be of order ∆, and thus, as claimed,
it will die away as n increases and ∆ shrinks. We will now show that using RM ,
or the almost equivalent Trapezoidal rule, produces a much smaller total error—in
fact an error that dies away as the square of ∆.
This standard result on the decay of the error can be found in many advanced
calculus books, but rather than repeat the standard calculation, we will supply a
novel geometric2 account. Figure [8.5a] shows a magnified view of the top of one of
the rectangles used in RM . Shown are the chord AB bounding the trapezoid used
in [8.2], and the line-segment DeC e bounding the rectangle used in RM . Notice that
P and Q (the midpoints of these line-segments) will lie directly above the point xj
that is being used in RM .
Visually, it is easy to compare the area under AB [the Trapezoidal rule] with
the actual area under the curve, but the same cannot be said of the area under

[8.5] [a] Start of geometric determination of the accuracy of the Midpoint Riemann Sum.
The area of the rectangle occuring in RM is the area under D e C,
e but this is the same as
the area under the tangent, DC, so the error is the shaded region between DC and the
curve. This error is certainly less than the area of ABCD, which is (PQ) · ∆. [b] As the
chord PQR revolves about the fixed point Q, the product PQ · QR remains constant.

2
While many of the arguments in this book were merely inspired by Newton’s mode of thought in
the Principia, we have here an example that is very close to his actual methods.
 The Real Integral 435

eC
D e [the RM rule]. However, note that if we rotate D eCe about P (keeping the ends
glued to the verticals) until it becomes tangent at DC, the area beneath it will remain
constant [why?]. Thus we are instead free to visualize each term of RM as being the
area lying beneath a tangent such as DC. It is now clear that the actual area lies
between the two values furnished by AB and DC, and that the error induced by
using either rule cannot exceed the area of the small quadrilateral ABCD, namely
[exercise], (PQ) · ∆. In order to find this area we will use the elementary property
of circles that is illustrated in [8.5b]: As the chord PQR revolves about the fixed point
Q, the product PQ · QR remains constant.
Over a sufficiently tiny distance we can consider any segment of curve to be
interchangeable with its tangent. However, over somewhat larger distances (or if
we simply require greater accuracy) we must instead replace it by a segment of its
circle of curvature, that is, the circle whose curvature κ agrees with that of the curve
at the point in question. In [8.6] we have drawn this circle for the segment at P. The
above result now informs us that

[8.6] Conclusion of geometric determination of the accuracy of the Midpoint Riemann

Sum. Let us approximate the curve in the vicinity of P with its circle of curvature, of radius
(1/κ). By virtue of the previous figure, PQ · QR = (AQ)2 . But in the limit of vanishing
∆, we have AQ ≍ DP and QR ≍ PR, so PQ = (AQ)2 /QR ≍ (DP)2 /PR. If θ is
the angle of the tangent at P, then DP
= 12 ∆ sec θ, and PR = (2/κ) cos θ. Therefore,
area (ABCD) = PQ · ∆ ≍ 18 κ sec3 θ ∆3 .
436 Complex Integration: Cauchy’s Theorem

PQ · QR = (AQ)2 . (8.2)
As ∆ shrinks, both (AQ/DP) and (QR/PR) tend to unity, so in this limit we may
substitute DP for AQ and PR for QR. But if the tangent at P makes an angle θ
with the horizontal (in which case OP makes angle θ with the vertical) then DP =
1
2 ∆ sec θ, and PR = (2/κ) cos θ. Substituting these into (8.2), we obtain the result


area (ABCD) = PQ · ∆ ≍ 18 κ sec3 θ ∆3 . (8.3)


If M denotes the maximum of 81 κ sec3 θ over the integration range (which we take
to be of length L) then each such error will be less than M ∆3 . Since the number of
these error terms is (L/∆), we conclude that
total error < (LM) ∆2 ,
and this indeed dies away in the manner originally claimed. [At this point you may
care to look at Ex. 1]
Because the order of the induced error is the same for both RM and the Trape-
zoidal rule, we will tend not to distinguish between them when it comes to their
complex generalizations. This said, there remains one curious pedagogical point
still to be made.
Figure [8.5a] makes it clear that curves deviate less from their tangents than from
their chords, and thus one would anticipate that while the order of the error is the
same for both rules, RM would actually yield the more accurate value of the two.
This is indeed the case, and in fact [see Ex. 2] one can show that it is twice as accurate.
In addition to this accuracy, RM is, if anything, easier to remember and use than the
Trapezoidal formula. It is therefore doubly puzzling that the Trapezoidal formula
is taught in every introductory calculus course, while it appears that the midpoint
Riemann sum RM is seldom even mentioned.

8.3 The Complex Integral

8.3.1 Complex Riemann Sums

In the case of real integration we began with a clear geometric objective (“Find the
area!”) and then invented the integral as a means to this end. In the complex case
we will reverse this process, that is, we will first blindly attempt to generalize real
integrals (via Riemann Sums) and only afterwards will we ask ourselves what we
have created. First, in this chapter, we will find one way of picturing an integral
as a single complex number; then, in Chapter 11, we will use an entirely different
point of view to see that, separately, the real and the imaginary parts of an inte-
gral each possess a vivid geometric (and physical) significance. But to guess the
relevant geometric entities in advance, and then to invent the complex integral as
the appropriate tool with which to find them, would require a prodigious leap of
 The Complex Integral 437

R
[8.7] Complex Riemann Sums. To approximate K f(z) dz, we approximate K with many
small complex chords, ∆j , evaluate f(z) at an arbitrary point zj within each segment
P
of K, then evaluate the sum, f(zj ) ∆j . Given the established superior accuracy of the
Midpoint Riemann Sum (RM ) in the real case, here we have immediately chosen each
zj to be the midpoint of each segment. As we shall explain, the Riemann sum can be
visualized more easily if we choose all the ∆j ’s to have the same length. [The turning
angles ϕj illustrated here are for later use.]

imagination—one that historically never took place. A moment’s thought reveals

that this is similar to the case of differentiation, for there we began with the slope
concept, and through an initially blind process of extrapolation we arrived at the
very different (but no less intuitive) idea of the amplitwist.
Consider [8.7]. In order to integrate a complex mapping f(z) between the points
a and b, we have specified a connecting curve (contour) along which to perform the
integration. This curve (call it K) now plays the role of the interval of integration,
and just as in [8.1], we break it down into small steps ∆j , which we may conve-
niently choose to be of equal length. The difference between this and [8.1] is that
now the steps are not all in the same direction.
In order to construct a Riemann sum, we randomly pick one point zj from each
little segment of K, and then we form the sum of the products f(zj ) ∆j . Finally, as
we increase their number and decrease their lengths, the ∆j will follow K ever more
perfectly, and the Riemann sum will tend to a limiting value (provided only that
the mapping is continuous) that serves as our definition of the complex integral,
written
Z
f(z) dz .
K
438 Complex Integration: Cauchy’s Theorem

[8.8] The values of the function along the contour of integration: wj = f(zj ). [The
angles τj between wj and wj+1 are for later use.]

Just as in the real case, we may obtain an accurate estimate of the integral without
passing to the limit, simply by choosing the zj to be at the midpoints of the segments
of K, rather than at random points. In fact this is the choice that we have illustrated
in [8.7]. Once again, this especially accurate Riemann sum will be denoted RM .
To begin to understand the geometry of RM , consider [8.8]. This shows the image
of K under the mapping z ÞÑ w = f(z), and in particular the image wj of the zj that
was singled out in [8.7]. The corresponding term of RM is then ∆ e j ≡ wj ∆j , and
we will choose to think of this as the arrow that results when wj “acts on” ∆j ,
expanding it by |wj | and rotating it by arg(wj ).
Having obtained each ∆ e j in this manner, we go on to join all these little arrows
together (tail to tip), as in [8.9]. The value of RM , and hence the approximate value
of the integral, is then the connecting complex number between the start and the
finish3 . Notice that since the answer is a connecting arrow, the point at which we
begin drawing RM is irrelevant.
While [8.9] is intended primarily to convey the general idea, it is in fact a faith-
ful evaluation of the specific RM corresponding to [8.7] and [8.8], and you may
now begin to convince yourself of this. This is perhaps most easily achieved by
concentrating on the lengths of the ∆ e j separately from their angles. As w traces
out the image curve in [8.8], its length diminishes, and this produces a corre-
sponding shrinking of the ∆ e j in [8.9]. Likewise, the increasing angle of w results
in progressively greater rotations of the ∆j .

3
The great physicist Richard Feynman used a similar kind of picture to explain his quantum-
mechanical “path integrals”, which are also complex, though they differ from contour integrals. See
Feynman (1985).
 The Complex Integral 439

[8.9] Visualizing the Midpoint Riemann Sum, Rm . The angle of each term ∆ e j ≡ wj ∆j
Pe
in RM = ∆j is the sum of the angles of wj and ∆j , so the illustrated turning angle
e j between successive steps along the Riemann sum is therefore the sum of the turning
ϕ
e j = ϕ j + τj .
angles ϕj and τj in the previous two figures: ϕ

8.3.2 A Visual Technique

While it is not strictly necessary to choose the same length for all the ∆j , the benefit
e j are simply proportional to |wj |,
of this choice is probably clear: the lengths of the ∆
and it is therefore easy to follow the evolution of ∆ e j by eye. On the other hand, it
ej.
is not so easy to visually follow the evolution of the angle of ∆
As we travel along the ∆j in [8.7], we pass through a sequence of sharp bends.
The turning angle at a typical bend is drawn in [8.7], and is denoted ϕj . What will
be the turning angle ϕ e j at the corresponding bend of the Riemann sum? If, for
example, wi+1 pointed in the same direction as wj , then ∆i+1 and ∆j would both
suffer the same rotation, and the turning angle ϕe j of the Riemann sum would equal
the original turning angle ϕj . More generally, if the angle of w increases by τj (see
[8.8]), then the turning angle will also increase by τj . Thus,
e j = ϕ j + τj .
ϕ (8.4)

This simple observation helps to reduce the difficulty of visualizing RM . It is

no longer necessary to look at the angle of each wj (which may be large and hard
to gauge by eye) and to try and imagine the direction of the rotated ∆. In fact we
need now only do this once, to find ∆ e 1 , thereby ensuring that RM heads off in the
correct initial direction. Thereafter, each successive ∆ e is laid down at an angle ϕ
e to
e
its predecessor, and these ϕj may be readily estimated by eye, using (8.4).
440 Complex Integration: Cauchy’s Theorem

Let us spell this out in detail with reference to the concrete example furnished by
[8.7] and [8.8]. In [8.9], we get RM started in the right direction by rotating ∆1 by α,
thereby obtaining ∆ e 1 which points at angle α + β. We can now draw the rest of RM
using only (8.4). To lay down the next ∆ e we need to know ϕ e 1 = ϕ1 + τ1 . The small
positive τ1 clearly kills off just a fraction of the negative ϕ1 , resulting in a slightly
smaller negative bend in RM . Much the same happens when we lay down ∆ e 2 . The
angle ϕ3 at the next bend is positive, and it is therefore increased by τ3 , which itself
is about twice as big as τ1 and τ2 were. You should now be in a position to follow
the rest of RM ’s progress in far greater detail than you could before.
Although the above idea will shortly prove its worth on a theoretical level, it
is clearly not terribly practical. However, in Chapter 11 we will use an entirely
different approach to obtain a second, less strenuous, means of visualizing com-
plex integrals. We will thereby make a double fallacy of an assertion that is to be
found in most texts—assuming they even consider it worthy of note!—namely, that
complex integrals possess no geometric interpretation. Perhaps the mere frequency
with which this myth has been reiterated goes some way to explaining how it has
acquired the status of fact.

8.3.3 A Useful Inequality

In figure [8.9] it is clear that if we were to straighten out all the bends in RM then it
would get longer. Furthermore, the length of the straightened version would just
be the sum of the ∆ e j . Thus
X
|RM | ⩽ |wj | · |∆j | ,

e j = 0. If M denotes the maximum distance from

with equality if4 and only if all ϕ
the origin to the image curve in [8.8], it follows that
X
|RM | ⩽ M |∆j | .

But the sum on the right is just the length of the polygonal approximation to K,
and hence it cannot exceed the actual length of K. Passing to the limit where RM
becomes the integral, we deduce that
Z
f(z) dz ⩽ M · (length of K) . (8.5)
K

For example, if f(z) = (1/z)2 and K is the circle |z| = r, then (8.5) implies that
R R
K f(z) dz ⩽ (2π/r). In particular this implies that limrÑ∞ K f(z) dz = 0. This is a
typical (albeit simplistic) application of (8.5): quite often one wishes to demonstrate

4
The method of visualizing complex integrals in Chapter 11 will enable us to express this condition
for equality in a particularly simple form. See Ex. 6, p. 574.
 The Complex Integral 441

the ultimate vanishing of an integral as K evolves through some family, such as

circles of increasing radius. Without knowing the exact value of any of the integrals,
(8.5) shows that it is sufficient to demonstrate that the maximum size of f(z) on K
dies away faster than the length of K grows.

8.3.4 Rules of Integration

Because the complex integral has been defined in complete analogy with the real
one, it follows that the former will inherit many of the properties of the latter. We
now list some of these shared properties:
Z Z
c f(z) dz = c f(z) dz
K K
Z Z Z
[f(z) + g(z)] dz = f(z) dz + g(z) dz
K K K
Z Z Z
f(z) dz = f(z) dz + f(z) dz
K+L K L
Z Z
f(z) dz = − f(z) dz.
−K K
The meaning of the first two equations is self-evident, but the last two require some
clarification.
If L begins where K left off (see [8.10a]), then to integrate along K + L means
to integrate along K and then to continue integrating along L, and the resulting
integral is then just the sum of the two separate integrals. Notice that the contour
is allowed to have a kink in it. In fact the definition of “contour” merely requires
that the number of such kinks not be infinite.
Lastly, the fourth rule is analogous to swapping the limits on a real integral,
for −K is defined to be the same as K, but traversed in the opposite direction
(see [8.10b]).
However familiar you may be with these rules in real calculus, and however
readily they may lend themselves to complex generalization, we would never-
theless urge you to make a new and separate peace with each of these results,
preferably in terms of pictures such as [8.7], [8.8], and [8.9].
Recall from the Introduction that our main objective is the discovery of condi-
tions under which an integral between two points in the plane does not depend on
the connecting route chosen. The last two rules above may be used to recast this
problem into a neater form. Suppose that the two paths K and K e in [8.10c] both
yield the same value for the integral between a and b. It follows that
Z Z Z Z Z
0= f(z) dz − f(z) dz = f(z) dz + f(z) dz = f(z) dz.
K e
K K e
−K e
K−K
442 Complex Integration: Cauchy’s Theorem

[8.10] [a] If L begins where K left off, then to integrate along K + L means to integrate
along K and then to continue integrating along L, and the resulting integral is then just the
sum of the two separate integrals. [b] Changing K to −K reverses the sign of the integral,
for −K is defined to be the same curve as K, but traversed in the opposite direction. [c] If
e then the
the integral from a to b along K equals the integral along an alternative route K,
e e
integral along the closed loop (K − K) = (K followed by − K) must be zero. Conversely,
if the integral vanishes for all closed loops then all curves between a and b will yield the
same value for the integral. In brief: Path independence is equivalent to vanishing loop
integrals.

Thus equality of the two integrals is equivalent to the vanishing of the integral
taken along the closed loop (K − K) e = (K followed by − K). e Conversely, if the
integral vanishes for all closed loops then all curves between a and b will yield the
same value for the integral. In brief: path independence is equivalent to vanishing loop
integrals. The centrepiece of complex analysis is the link between this phenomenon
and analyticity. Cauchy’s Theorem consists in recognizing that the vanishing of
loop integrals is the nonlocal manifestation of a local property of the mapping,
namely, that it is an amplitwist everywhere inside the loop.

8.4 Complex Inversion

8.4.1 A Circular Arc

Probably the single most important integral in all of complex analysis is that of the
complex inversion mapping, z ÞÑ 1/z. While the truth of this assertion will only
emerge gradually, this is the reason for the great attention we will now lavish on
this particular example.
We begin with the simplest case, namely, where the path of integration K is an
arc of the origin-centred circle of radius A (see [8.11a]). As in [8.7], we divide this
path into small (ultimately vanishing) steps of equal length. The turning angles ϕj
clearly all have the same value, say ϕ. Since the angle that each ∆ subtends at the
 Complex Inversion 443

[8.11] The integral of (1/z) along a circle. Since ∆1 and w1 are ultimately vertical and
horizontal respectively, it follows that RM (in [c]) initially heads off in the vertical direc-
tion. But since the turning angle τ of wj = (1/zj ) in [b] is the opposite of the turning angle
ϕ of ∆ in [a], ϕe = ϕ + τ = 0. In other words, RM has no bends, and so it continues on
in the imaginary direction. We wanted to illustrate the turning-angle approach for later
use, but here is a simpler explanation. Since the ∆j located at angle θ on the circle itself
points at θ to the vertical (as illustrated) and has length Aϕ, we see that multiplying it by
w = (1/A)e−iθ rotates it by −θ, producing a vertical ∆ e = w∆ = iϕ.

origin is also given by ϕ, it follows that |∆| = Aϕ. As z travels round the circle its
image w = 1/z travels round a circle of radius 1/A in the opposite direction (see
[8.11b]), and thus w shrinks each ∆ to produce a ∆ e of length ϕ.
Since ∆1 and w1 are ultimately vertical and horizontal respectively, it follows that
RM (which we choose to begin drawing at the origin in [8.11c]) initially heads off in
a vertical direction. But now we observe that τ = −ϕ, and consequently that ϕ e = 0.
In other words RM has no bends, and so it continues on in the imaginary direction5 .
Thus, irrespective of the radius, the integral equals i times the total angle Ψ through

5
We have used the turning angle idea in order to make the subsequent generalization to other
powers of z straightforward, but there is actually no need for it in the present case. Since the ∆j located
at angle θ on the circle itself points at θ to the vertical and has length Aϕ, we see that multiplying it
by w = (1/A)e−iθ rotates it by −θ, producing a vertical ∆ e = w∆ of length ϕ.
444 Complex Integration: Cauchy’s Theorem

which z turned on its journey along K. Convince yourself that this formulation of
the result remains valid even if K begins at a random point of the circle instead of
on the real axis.
In particular, and of crucial importance, is the case where z continues all the way
round the circle to form a closed loop. The value of the integral is then 2πi. The alert
reader will immediately be perplexed by this result. Why? Because it appears to fly
in the face of Cauchy’s Theorem. We have previously demonstrated geometrically
that complex inversion is analytic, so how can its loop integral fail to vanish?! The
resolution lies in the fact that Cauchy’s Theorem requires that the mapping be ana-
lytic everywhere inside the loop. But our loop encloses the origin, and just at this
one point the analyticity of complex inversion breaks down.

8.4.2 General Loops

The above discussion not only explains why our loop integral failed to vanish, but
it also leads us to anticipate that if the loop does not enclose the origin, then the
integral will vanish. We will now show that this is the case, thereby lending some
credence to Cauchy’s Theorem.
The ease with which we were able to evaluate the integral in [8.11] resulted from
the fact that the ∆j were all orthogonal to the zj . Figure [8.12 a] shows a more typ-
ical ∆ possessing a radial component in addition to a transverse one. As you see,
∆ can be decomposed into a transverse component r dθ making an angle θ with
the vertical, and an orthogonal radial component dr. To obtain the corresponding
piece ∆e of R (see [8.12b]) we multiply by w, thereby rotating these components into

[8.12] The integral of (1/z) along a general contour. [a] The movement ∆ along a general
contour can be broken down into a radial component dr and an orthogonal component
rdθ making angle θ with the vertical. [b] Multiplication by w = (1/z) therefore yields
e = w∆ = (dr/r) + idθ.
∆
 Complex Inversion 445

[8.13] The integral of (1/z) along a closed loop. [a] Instead of choosing ∆j ’s of equal
length, divide the path up, as shown, using closely spaced concentric circles centred at
the origin. Consider the pair ∆j and ∆k cut off by the same two circles. From the previous
figure, we deduce that real components of ∆ e j and ∆e k in the Riemann sum will cancel.
But every ∆ belongs to such a pair (or pairs), because if the closed loop L passes from the
interior of a circle to the exterior, then in order to join up with itself back in the interior,
it must recross the circle in the opposite direction somewhere else. Thus the integral
of (1/z) along a closed loop is always imaginary. The previous figure also implies that
this imaginary answer = i (net angle of rotation as the loop is traversed). [b] The
illustrated loop L does not enclose the origin, so the integral vanishes. [c] If the loop does
wind around 0 once, then the integral is 2πi.

the vertical and horizontal directions, as well as shrinking their lengths to dθ and
(dr/r), respectively.
Let us now see what happens if we stick all these ∆ e j together for a closed loop
such as L (see [8.13a]) that does not encircle the origin. In order to accomplish this
we will forsake our previous choice of equal lengths for all the ∆j , and instead
divide the path up, as shown, using closely spaced concentric circles centred at the
origin6 . Consider the illustrated pair ∆j , ∆k lying between adjacent circles. That
the ∆’s always do occur in such pairs is a consequence of L being a loop. For if

6
This only fails if part of L coincides with such a circle, but in that event we already know that the
contribution to the integral is iΨ.
446 Complex Integration: Cauchy’s Theorem

L passes from the interior of a circle to the exterior, then in order to join up with
itself back in the interior, it must recross the circle in the opposite direction some-
where else. Of course L may weave back and forth across a circle many times (e.g.
at a, a⋆ , b, b⋆ , c, c⋆ ), but the crucial point is that these crossings always occur in
oppositely directed pairs.
In [8.13b] we see the consequence of this for R. From [8.12b] it’s clear that for a
pair such as ∆ e j and ∆ e k , the horizontal components cancel. Since we have seen that
every ∆ belongs to such a pair, it follows that R will have no horizontal component
for any closed loop, whether or not the origin is encircled. It also follows from
[8.12b] that the height of this vertical Riemann sum is obtained by adding up all
the signed angles that the ∆’s subtend. For a loop such as [8.13a], which does not
encircle the origin, this sum is zero: as z traces out L its direction merely oscillates,
rather than executing a complete revolution. Thus, as illustrated in [8.13b], R closes
up on itself. On the other hand, if we translated L to any location where it encircled
the origin then z would execute a complete revolution, and [8.13b] would change
into [8.13c].

8.4.3 Winding Number

Let’s recap. If a closed loop does not encircle the origin then the complex inversion
mapping is analytic everywhere inside it, and in accord with Cauchy’s Theorem the
integral dutifully vanishes. If the origin is encircled, then the integral is no longer
required to vanish by the theorem: the enclosed region now contains a point at
which the mapping is not analytic. Indeed we found that for an origin-centred circle
the answer was not zero, but 2πi. Furthermore, the general investigation revealed
that we would have obtained exactly the same answer if we had instead used an
elliptical loop, or even a square loop. For if we distort the circle into one of these
more general shapes, then all that happens to R is that it meanders about (illustrated
in [8.13c]) on its net vertical journey to 2πi, instead of marching straight there as it
did in [8.11c].
We see that what really matters is not the shape of the loop, but rather its winding
number about the origin. Thus we may summarize our findings tidily as follows:
If L is any closed loop, then
I
1
dz = 2πi ν(L, 0), (8.6)
z
L

where the integral sign with a circle through it (which is a standard symbol) serves
to remind us that we are integrating around a closed contour. Figure [8.14] shows
various loops and the corresponding value of the integral of (1/z) round each of
them. Finally, note that (8.6) can easily be generalized [exercise] to
I
1
dz = 2πi ν(L, p). (8.7)
z−p
L
 Conjugation 447

H
[8.14] The previous figure proves that L (1/z) dz = 2πi ν(L, 0).

8.5 Conjugation

8.5.1 Introduction
In the introduction we stressed that integration makes sense for any continu-
ous complex mapping, regardless of whether or not it is analytic. However, the
relatively lawless non-analytic functions give rise to integrals that behave less pre-
dictably than their analytic counterparts. In particular, Cauchy’s Theorem has no
jurisdiction here, and we therefore have no reason to anticipate path independence
or, equivalently, vanishing loop integrals. As an example of this type of behaviour,
we will show presently that the loop integral of the non-analytic conjugation map-
ping z ÞÑ z yields the area enclosed by the loop. Assuming this result for the
moment, let us use the examples z and (1/z) to spell out more clearly the differences
between the non-analytic and analytic cases.
In the analytic case, provided that the special point z = 0 was not enclosed, the
loop integral vanished. Even when the integral of (1/z) did not vanish, its possible
values were still neatly quantized in units of 2πi; one unit for each time the special
point z = 0 was enclosed by the loop. As we will see later, this behaviour is typical,
although a more general mapping may well possess several special points (at which
analyticity breaks down) dotted about in the plane. Once again, the integral is not
sensitive to the precise shape of the loop. Provided that none of the special points
are enclosed by the loop, then the integral vanishes. However, if some of the points
are enclosed, then each one makes its own distinctive contribution (generally not
2πi) to the integral, one unit for each time it’s encircled. The value of the integral
is just the sum of these discrete contributions.
448 Complex Integration: Cauchy’s Theorem

Contrast all this with our non-analytic example. The area of the loop (and hence
the integral of z) will almost never vanish. Furthermore, instead of being deter-
mined by stable topological properties, the value of the integral is sensitive to the
detailed geometry of the loop. Finally, the value is not neatly quantized, but instead
varies continuously as the loop changes shape.

8.5.2 Area Interpretation

Let us now verify the area interpretation of the integral of z. Recall from Chapter 1
that Im(ab) is just twice the area of the triangle spanned by a and b. As z traces
the loop L in [8.15a], think of the area it sweeps out as being decomposed into
triangular elements, as illustrated. Thus

2 (element of area) = Im[(z + ∆)z] = Im[z∆].

Adding these elements together, we obtain the imaginary part of the Riemann sum
corresponding to the integral of z. Thus we conclude that
I
Im z dz = 2 (area enclosed) .
L

This result can be further simplified by noticing that z and (1/z) both point in the
same direction. It follows that we could draw a picture very similar to [8.12], the
e we would multiply by r instead of dividing by
only difference being that to obtain ∆

H
[8.15] L z dz = 2i (area enclosed by L). [a] Using H(1.20) on page 32, we see that
2 (element of area) = Im[(z + ∆)z] = Im[z∆], so Im L z dz = 2 (area). But z has the
same direction as (1/z), and the argument given in [8.12] easily generalizes, so the inte-
gral is purely imaginary, thereby proving the result. [b] The geometric meaning of the
integral is unaltered if 0 lies outside L, for although the top triangle includes area that lies
outside L, this extra area is cancelled by the negative area of triangles like the one at the
bottom. Compare this to [1.22b] on page 33.
 Conjugation 449

it. The argument that followed from [8.12] therefore remains valid, and we deduce
that the integral of z around a closed loop is purely imaginary. Thus
I
z dz = 2i (area enclosed) . (8.8)
L

Next we ask how this formula would change if the origin were outside the loop.
Figure [8.15b] shows that the pleasing answer is, “Not at all!” The point is that the
integral adds up the signed areas subtended by the ∆’s at the origin. On the far
side, ∆ carries z counterclockwise, yielding a positive element of area; but on the
near side z is moving clockwise, yielding a negative element of area. When these
are added, the unwanted area lying outside the contour simply cancels, leaving
behind just the area enclosed.
As a simple example, consider a circle C of radius r centred at a, the equation of
which is r2 = |z − a|2 = (z − a) (z − a). Solving this for z, and using (8.7), we find
that
I I I
2 1
z dz = a dz + r dz
z−a
C C C
2
= 0 + r 2πi
= 2i (area enclosed).

From what we have done so far you might be inclined to think that the integral
of z could never vanish for a nontrivial loop. That this is false can be seen from the
figure eight loop in [8.16a]. This may be thought of as the union of two separate
loops. The top one is traversed in a positive sense and correspondingly yields its
ordinary area A1 ; but the bottom one is traversed in a negative sense and yields the

H
[8.16] [a] If L is a figure eight, then L z dz = 2i(A
H 1 −A2 ). Note that if L were symmetrical,
P
then the integral would vanish. [b] In general, L z dz = 2i inside νj Aj = 2i [2A1 + A3 ]
in this example.
450 Complex Integration: Cauchy’s Theorem

negative (−A2 ) of its ordinary area. Thus the integral is 2i(A1 − A2 ), and if the loop
were symmetrical then this would vanish.

8.5.3 General Loops

To finish off this example we will explain how the winding number concept can
be used to evaluate the integral for more complicated loops. A typical loop such as
[8.16b] will partition the plane into a number of sets Dj , and in the last chapter we
defined the “inside” to consist of those Dj for which the corresponding winding
number νj ̸= 0; the remaining Dj constitute the “outside”. We can now state the
general result and leave you to ponder its truth:
I X
z dz = 2i νj Aj , (8.9)
L inside

where Aj denotes the area of Dj . For example, in the case of [8.16b] we obtain
H
L z dz = 2i [2A1 + A3 ]. We hope you can already see why this example is correct,
but a full explanation of the general formula (8.9) will be provided later in this
chapter.

8.6 Power Functions

8.6.1 Integration along a Circular Arc

Having understood the integral of (1/z) it is easy to understand the integrals of
other powers. Once again let us begin by integrating along the circular arc K that
was used in [8.11]. The result we will obtain is formally identical to the real result
ZB
1  m+1 
xm dx = B − Am+1 (m ̸= −1),
A m+1
but the difference is that in the complex case we can actually see it!
Figure [8.17a] shows a contour like that in [8.11a], while the transition from
[8.11b] to [8.17b] represents the change from complex inversion to a general integer
power w = zm . Although the primary purpose of [8.17] is to convey the general
argument, you will better understand its details if I tell you that it actually depicts
the special case m = 2.
As z travels along K, w travels round an image circle of radius Am , and with an
angular speed that is m times as great. Thus
e = Am (Aϕ) = Am+1 ϕ ,
|∆|
and
e = τ + ϕ = mϕ + ϕ = (m + 1) ϕ .
ϕ
e have the same length and the same turning angle, it follows that
Since all the ∆’s
RM is a polygonal approximation to an arc of a circle, the centre of which we have
 Power Functions 451

RB  m+1 
[8.17] A zm dz = m+1 1
B − Am+1 , because [a] |∆| = Aϕ, [b] τ = mϕ and
e = Am+1 ϕ, so [c] ϕ
|∆| e = τ + ϕ = (m + 1) ϕ, and therefore ρ = |∆|/
e ϕ e = Am+1 /(m + 1).

chosen to place at the origin in [8.17c]. We will now determine the angle subtended
by this arc, and also its radius.
The angle that each ∆ e subtends at the origin is the same as the turning angle ϕ,
e
namely, (m + 1) times the angle subtended by each ∆. Thus
angle of FINISH = (m + 1)Ψ.
Also, if ρ is the radius, we see from the figure that

e = |∆|
e Am+1
ρϕ =⇒ ρ= .
m+1
We therefore conclude that if m ̸= −1 then
RM = FINISH − START
1 h m+1 i(m+1)Ψ i
= A e − Am+1
m+1
1  m+1 
= B − Am+1 (8.10)
m+1
452 Complex Integration: Cauchy’s Theorem

which, as promised, is formally identical to the real result. We hope you will agree
that it’s rather fascinating how we have been able to visualize this result in a way
that would not have been possible in the real case.
As we have said, [8.17] actually depicts the concrete case m = 2, and before
continuing you may care to sketch another case for yourself; m = −2 might be a
fun one.

8.6.2 Complex Inversion as a Limiting Case*

As in ordinary calculus, we see that the case m = −1 (complex inversion) stands
out from the crowd. Nevertheless, we can still understand the behaviour of this
special power as a limiting case of other powers. With a little care about branches,
the above result can be seen to persist even if we relax the requirement that m be
an integer. As m gradually approaches −1 the radius ρ grows, and so RM looks less
e tend to ϕ. Thus in the limit
and less curved; at the same time the lengths of the ∆’s
that m tends to −1 we see that RM will go straight up the imaginary axis to iΨ. This
is illustrated in [8.18]. The variable n ≡ m + 1 measures the difference between m
and −1, and it is therefore a good label for the Riemann sums shown in the figure.
Returning to the case of integer powers, we next observe that for a complete
circular loop, there is a striking and fundamental difference between complex
inversion and all others powers: if m ̸= −1 then the integral vanishes. This is

[8.18] The integral of (1/z) as a limiting case of the integral of zm . Here, n ≡ m − (−1)
measures the “distance” of zm from the ultimate target of z−1 . But there is also a fun-
damental difference between (1/z) and all other powers: its integral around a complete
circle is 2πi, whereas the integral for all other powers is zero!
 Power Functions 453

because RM will now go round in a complete circle |n| times [clockwise if n < 0;
counterclockwise if n > 0], thereby returning to its beginning.

8.6.3 General Contours and the Deformation Theorem

Thus far we have only established (8.10) for the case where A and B are connected
by a simple arc, but in fact it is true for almost any contour. Take the case n > 0 first.
Since zm is then analytic throughout the plane, it follows directly from Cauchy’s
Theorem that all contours will yield the same value. However, when n < 0 the
situation is a little bit more subtle.
Just as complex inversion suffers a breakdown of analyticity at the origin, so too
do all the other negative powers of z. Therefore Cauchy’s Theorem only guarantees
that two connecting paths yield the same integral provided that together they do
not enclose the origin. For a loop that does enclose the origin, the integral is not
required to vanish, and indeed in the case of z−1 it equals 2πi.
Nevertheless, our direct evaluation reveals that for all other negative powers
the integral round a circular loop does vanish, despite not being required to7 . We
will now derive a new form of Cauchy’s Theorem that enables us to show that the
vanishing of the integral is not a fluke resulting from the special circular shape of
the loop.
Consider [8.19a]. The two loops J and L both encircle a singularity of some map-
ping, and so neither integral is required to vanish by Cauchy’s Theorem. However,
if the mapping is analytic in the shaded region lying between the loops, then we
will now show that the two integrals must be equal. First consider the contribution

[8.19] Deformation Theorem. [a] Deform L by pushing the segment pq outward to form
the illustrated bump. Since the mapping is analytic between the two paths connecting p
and q, it follows from Cauchy’s Theorem that both integrals are equal, and since the rest
of L hasn’t changed, it follows that integral with bump = integral without bump. [b] We
can continue to deform L until it becomes J, and the value of the integral will not change.

7
In Chapter 11 we will give a physical explanation for this difference between complex inversion
and the rest of the negative powers.
454 Complex Integration: Cauchy’s Theorem

to the integral round L that comes from the piece between p and q. Suppose that
we deform L slightly by replacing this segment by the bump in the figure. Since
the mapping is analytic between the two paths connecting p and q, it follows from
Cauchy’s Theorem that both integrals are equal. Also, since the rest of L hasn’t
changed, it follows that integral with bump = integral without bump. All we need do
now, to obtain the stated result, is to let the bump grow and change shape (see
[8.19b]) until L has evolved into J.
The crucial idea is this

Deformation Theorem. If a contour sweeps only through analytic

(8.11)
points as it is deformed, the value of the integral does not change.

Thus, if you imagine the contour to be a rubber band, and the singularity to be a
peg sticking out of the plane (thereby obstructing motion past it), the integral has
the same value for all shapes into which the rubber band can be deformed.
We can immediately apply this Deformation Theorem to our problem. For if the
mapping is a negative power of z other than z−1 then the established fact that the
integral vanishes for a circular loop implies it continues to vanish for any loop into
which the circle can be deformed without crossing the singularity at the origin.
Thus formula (8.10) is path independent even for negative powers.
The Deformation Theorem also provides us with a much simpler derivation
of the result (8.6) governing the general loop integral of the complex inversion
mapping. Imagine taking a length of elastic string and winding it around an origin-
centred circle ν times, finally joining the ends together to form a closed loop. From
our earlier work, it follows that the value of the integral is then 2πiν. But the Defor-
mation Theorem says that this will be the value of the integral for any loop into
which the elastic string may be deformed without being forced over the peg (sin-
gularity) at the origin. Finally, by the Hopf Degree Theorem, the loops into which
it can so be deformed are those with winding number ν.

8.6.4 A Further Extension of the Theorem

Our ‘dynamic’ version of Cauchy’s Theorem can be further extended to embrace
mappings that have several singularities. Consider a loop L (see [8.20a]) encircling
two singularities (pegs) of some mapping; the generalization to more singularities
will be obvious. If we deform L without forcing it over a peg then we know that the
integral will remain constant. The process [8.20a]Ñ[8.20b]Ñ[8.20c] is an example of
such a deformation. The situation in [8.20c] is now rather interesting. The contour
has become pinched together at q, and the value of the integral can be thought
of as the sum of the two separate integrals taken round the touching circles. But
now, by the same reasoning as usual, we may separately distort these circles so
that [8.20c]Ñ[8.20d]. Thus we conclude that
 Power Functions 455

[8.20] Deformation Theorem with multiple singularities. We can deform L into J + K

without changing the value of the integral.
I I I
f(z) dz = f(z) dz + f(z) dz . (8.12)
L J K

To illustrate (8.12), consider f(z) = 2/(z + 1) which has singularities at z = ±i.

2

We can evaluate the integral round any loop C by noting this alternative expression:
i i
f(z) = − .
z+i z−i
Applying (8.7) therefore yields
I
f(z) dz = 2π[ν(C, i) − ν(C, −i)].
C
Assuming (as in [8.20a]) that L encloses both singularities, use this formula to verify
(8.12) for this particular function.

8.6.5 Residues
Since we now possess a fairly complete understanding of the loop integrals of
power functions, it is relatively easy to integrate simple rational functions: we need
only find the decomposition into so-called partial fractions, and then integrate term
by term. Indeed, this is precisely what we did in the example of the last paragraph.
Here is a slightly more complicated example: the integral of f(z) = z5 /(z + 1)2
taken round the contour K in [8.21]. This has a second-order pole at z = −1, and by
writing the numerator as [(z + 1) − 1]5 , we quickly find that
 
1 1
f(z) = − +5 − 10 + 10 (z + 1) − 5 (z + 1)2 + (z + 1)3 .
(z + 1)2 z+1
456 Complex Integration: Cauchy’s Theorem

[8.21] The Residue of a singularity. Let f(z) = z5 /(z + 1)2 , which has a second-order
pole at z = −1. If f(z) is decomposed into partial fractions then the only part that is not
removed by integration is its “residue”—that which remains, defined to be the coefficient
of the complex inversion term, 1/(z+1). This is the only part that does not integrate away
into thin air. The value of the integral therefore has only two ingredients: the residue and
the winding number, and the result is 2πi times their product, which turns out to be
−20πi for this particular function and contour.

But we know that the loop integral of powers other than −1 is zero, and so only the
complex inversion term [in square brackets] can contribute. In detail,
I
f(z) dz = 5 · 2πi ν(K, −1) = −20πi.
K

Thus the value of the integral has been determined by just two factors: the winding
number of the loop, and the amount (i.e., coefficient) of complex inversion con-
tained in the decomposition of the mapping. Because this latter number is the only
part of the function that remains after we integrate, it is called the residue of the
function at the singularity. Quite generally, the residue of f(z) at a singularity s is
denoted Res [f(z), s]. Thus in the above example, Res [z5 /(z + 1)2 , −1] = 5.
In fact the residue concept has a significance that extends far beyond sim-
ple rational functions, as our next example will illustrate. We have previously
[page 259] alluded to the remarkable fact that analytic functions are infinitely dif-
ferentiable, or equivalently, that they can always be represented by a power series
(Taylor’s) in the vicinity of a nonsingular point. For example, the Taylor series
centred at the origin for sin z is
1 3 1 1
sin z = z − z + z 5 − z7 + · · · .
3! 5! 7!
Clearly no such expansion can be possible at a singular point of a mapping.
Nevertheless, we may recover an analogous result near singularities simply by
broadening our notion of a power series to include negative powers. Such a series
is called a Laurent series.
 The Exponential Mapping 457

Consider (sin z)/z6 . This is singular at the origin, but by simple division of the
above Taylor series we obtain the following Laurent series in the vicinity of the
singularity:
 
sin z 1 1 1 1 1 1
= 5− + − z + z3 − · · · .
z6 z 3! z3 5! z 7! 9!
Once again, the residue of the function is defined to be the coefficient of the complex
inversion term: Res [(sin z)/z6 , 0] = 1/5! in this case. If a power series converges at
every point on a contour, then we may accept for the moment that it makes sense
to integrate the series term by term. Once again we see that for a closed loop the
sole contribution to the integral comes from the residue. For example, if K is the
contour in [8.21] then
I
sin z 1 2πi
dz = 2πi ν(K, 0) = − .
z6 5! 5!
K

The above examples of evaluating loop integrals in terms of residues are

instances of Cauchy’s Residue Theorem. We will return to these matters at the end
of this chapter, and, in greater detail, in the following chapter. For the moment we
simply remark that if a mapping possesses several singularities, then a residue can
be attributed to each one.

8.7 The Exponential Mapping

In the case of the exponential mapping the easiest contour along which to integrate
is a vertical line-segment, say L (see [8.22a]). Once again we will find the result to
be formally identical to its real counterpart:
Z
ez dz = eB − eA . (8.13)
L

As z travels from A up to B in [8.22a], its image under w = ez will travel round

the arc shown in [8.22b]. In order to verify (8.13), we will now show that (provided
we choose to begin drawing RM at eA ) this arc is also the precise path taken by the
Riemann sum.
First note that since ∆1 and w1 are effectively vertical and horizontal, respec-
tively, it follows that RM will head off in the required vertical direction. Also
e have the same length, namely, |∆|
observe that all the ∆’s e = eA |∆|. Finally, since L
has no bends (i.e., ϕ = 0), ϕe = τ = |∆|. Because the ∆’se all have the same length
and turning angle, RM will follow an arc of a circle. It only remains to show that if
we begin drawing it at eA , then this is the same arc as in [8.22b].
The next section will reveal the simplest way of seeing this, but the following
direct argument is quite straightforward. We first verify that the two arcs have the
458 Complex Integration: Cauchy’s Theorem

R
[8.22] Geometric evaluation of L ez dz. [a] Let L be the illustrated vertical segment
from A to B = A + iθ. [b] If we choose to begin drawing RM at eA then the illustrated
arc
R traced by w = ez is also the precise path taken by the Riemann sum, and therefore
e e
L e dz = e − e . First, all the ∆’s have the same length, namely, |∆| = e |∆|. Since L
z B A A

has no bends (i.e., ϕ = 0), ϕ e = τ = |∆|. Because the ∆’s

e all have the same length and
turning angle, RM will follow an arc of a circle, and it only remains to verify its radius.
Try this yourself, or see the text for details.

e subtends at the centre of its circle will be the

same radius. The angle that each ∆
e Therefore
same as its turning angle ϕ.

arc e
|∆|
radius = = = eA ,
angle e
ϕ
as required. Lastly, the total angle subtended by the arc at its centre is just the sum
e j = |∆j |, namely θ. The identity of the two arcs is thus established.
of all the ϕ
Since ez is singularity-free, Cauchy’s Theorem assures us that its loop integral
always vanishes. Thus (8.13) must in fact be valid for any path from A to B.

8.8 The Fundamental Theorem

8.8.1 Introduction
Through specific geometric constructions, combined with the use of Cauchy’s
Theorem, we have already learnt a good deal about the integrals of some of the
most important functions. However, there are two immediate problems still to be
resolved: one is pragmatic, while the other is aesthetic.
 The Fundamental Theorem 459

The pragmatic one is that the formulae (8.10) and (8.13) are only known to hold
for certain special configurations of the points A and B: the derivation of (8.10)
assumes that they are equidistant from the origin; while for (8.13) they are assumed
to be vertically separated. To be sure, our various forms of Cauchy’s Theorem guar-
antee us that the integrals in question will continue to be path-independent, no
matter what the locations of A and B. But the problem is that we haven’t yet estab-
lished that these path independent values will continue to be given by the same
formulae as before. In this section we shall see that they are.
The aesthetic concern lies in the manner in which we derived path-independence
for negative powers of z. Recall that we were only able to apply Cauchy’s Theorem
after having explicitly produced an example (a circle) of a loop integral that van-
ishes in spite of enclosing the singularity. Although this was neat enough in itself,
one is left with the feeling that Cauchy’s Theorem cannot be the most direct
way of understanding a loop integral that continues to vanish in the presence of
singularities.
A resolution of both these problems is provided by the so-called Fundamental
Theorem of Contour Integration—a result that is formally identical to its similarly
named counterpart in ordinary calculus. The naming of this theorem is not entirely
appropriate, at least in the context of complex analysis. After all, so far we have
managed quite well without it, suggesting that if this theorem is “Fundamental”,
then Cauchy’s must be “Super-Fundamental”!

8.8.2 An Example
As our first example of this theorem, let us return to the exponential mapping of
the last section in order to discover why (8.13) is valid for any pair of points, not
just ones that are vertically separated. As so often happens in mathematics, all that
is required is a very slight shift in viewpoint.
Figure [8.23] depicts a curve K (connecting a pair of typical points A and B)
being mapped by ez to the curve K e connecting eA and eB . Now let us forget (for a
moment) all about integration and Riemann sums, and instead look at the figure
from the point of view of differentiation.
All the little arrows emanating from a point on K will be mapped to images ema-
nating from a point on K. e In particular, if the arrow ∆ is a little chord of K [tangent,
e will likewise be a little directed chord of
in the limit that it shrinks], then its image ∆
e But for an analytic mapping, such as we are now considering, the original arrows
K.
are sent to their images by a simple amplitwist:
e = (amplitwist of ez ) · ∆ = ez ∆ .
∆ (8.14)

If we now add up all these vector chords of K e then we obtain the connecting vector
V between its start and its finish. But (8.14) tells us that this vector V may also be
460 Complex Integration: Cauchy’s Theorem

[8.23] The Fundamental Theorem in the case of ez . Under z ÞÑ e z = ez , the contour K

e A B
connecting A to B maps to the curve K connecting e to e , the connecting complex
number from start to finish being V = eB − eA . But each ∆ along K is amplitwisted to
e = (amplitwist of ez ) · ∆ = ez ∆ along K,
∆ e so K
e simply is the Riemann sum, and therefore
R z B A
K e dz = V = e − e .

interpreted as the Riemann sum corresponding to the integral of ez along K. We

have thus established the continued validity of (8.13) for all positions of A and B:
Z
ez dz = V = eB − eA .
K

To emphasize the path-independence of the construction, imagine choosing a dif-

ferent contour from A to B. The image curve (i.e., the new Riemann sum) will
then take a different route from eA to eB , but of course the vector V will be quite
unaffected.

8.8.3 The Fundamental Theorem

The Fundamental Theorem amounts to a restatement of the above idea in general
R
terms. Suppose that we wish to evaluate K f(z) dz by the method above. We must
seek an analytic mapping F(z) whose amplitwist F ′ (z) is given by f(z). Assuming
that such an F has been found [whether this animal even exists will be discussed
shortly], we may then draw [8.24], which depicts the image curve K e under the
mapping F. With the same terminology as before, (8.14) now becomes

e = [amplitwist of F(z)] · ∆ = f(z) ∆ .

∆
 The Fundamental Theorem 461

[8.24] The Fundamental Theorem in the general case. If F maps K to K e and F ′ = f then
e e simply is the
· ∆ = f(z) ∆, and therefore K
∆ is amplitwisted to ∆ = [amplitwist of F(z)] R
Riemann sum for the integral of f. Therefore, K f(z) dz = V = F(B) − F(A).

Just as before, we conclude that K e is actually the path taken by the Riemann sum
of f, and that the vector V is once again the path-independent value of the integral:
Z
f(z) dz = V = F(B) − F(A) . (8.15)
K

As in ordinary calculus, the function F cannot be unique, for b F ≡ F + const.

shares the same amplitwist. In terms of a figure like [8.24] this corresponds to the
fact that the effect of bF only differs from that of F by a translation; but this has no
effect on the connecting arrow V.
In ordinary calculus, a real continuous function f(x) always possesses an anti-
derivative F(x) for which F ′ = f. Of course it may not be easy to find, and it may
2
not even be expressible in terms of elementary functions (e.g. f(x) = e−x ), but
at least it exists. In the complex realm, on the other hand, we know that analytic
functions are very special, and so we should not be surprised if the existence of such
a function is no longer assured. Remember that when such an F exists, the integral
of f is path-independent. It follows, for example, that no such function can exist for
the non-analytic mapping f(z) = z. Indeed, we may fall back on the still unproven
result concerning infinite differentiability to see that, quite generally, analyticity of
f is a necessary condition for the existence of F. For if F is analytic, then so too is its
derivative F ′ , namely f.
When presented with the integral of a non-analytic function, it is therefore hope-
less to seek an anti-derivative for use in (8.15)—no such function can exist. For the
462 Complex Integration: Cauchy’s Theorem

special case z ÞÑ z it is possible to extend the area interpretation (hence evaluation)

to contours that are not closed, but for a general non-analytic mapping no such
interpretation will be available. Although such integrals are of much less interest
to us than those of analytic functions, in the next section we shall nevertheless find
a method of evaluating them.
Before returning to more general considerations, let us give a couple more exam-
ples of the theorem in action. Consider f(z) = z2 . If we define F(z) = 13 z3 then F ′ = f,
and thus the path taken by the Riemann sum as we integrate along a contour will
just be its image under z ÞÑ 31 z3 . This allows us to look at the construction [8.17]
in a new light. Recall that while this figure is concerned with the integration of a
general power, it actually depicts the special case z2 . In agreement with our new
general result, we see that z ÞÑ 31 z3 does indeed map the contour in [8.17a] to its
Riemann sum in [8.17c].
As an example of how the theorem also resolves our aesthetic concern over path-
independence for negative powers of z, reconsider f(z) = (1/z2 ), for which we hope
you did actually draw the analogue of [8.17], as suggested. Without appealing to
Cauchy’s Theorem (thus avoiding the attendant anxiety over the singularity at the
origin), we see that since (−1/z) ′ = (1/z2 ), all contours8 between A and B will yield
the same value for the integral:
ZB
1 1 1
dz = − .
A z2 A B

Notice, incidentally, that path-independence has allowed us to reinstate the famil-

RB
iar symbol A without fear of ambiguity.
Instead of having to use Cauchy’s Theorem to extrapolate from the vanishing
of the integral for a circle to its vanishing for more general loops, the conclusion is
now immediate: since B = A for a closed loop, the above expression vanishes.

8.8.4 The Integral as Antiderivative

We have seen that the existence of an antiderivative F (defined by F ′ = f) implies
path-independence for the integral of f. We will now show, conversely, that path-
independence implies the existence of F.
Let us first give another simple example of the Fundamental Theorem. Since
(sin z) ′ = cos z, the integral of cos z will be path-independent, and if we integrate
from the origin, for example, to a variable endpoint Z, then we obtain a well-defined
function of Z:
ZZ
F(Z) = cos z dz = sin Z .
0

8
We exclude contours that actually pass through the singularity.
 The Fundamental Theorem 463

We note, without surprise, that this function is the antiderivative of cos Z. If we

began our integration at an arbitrary point, instead of at the origin, then the
result would only differ by a constant, and so it would still be a perfectly good
antiderivative.
In order to establish the claim of the first paragraph, it is only necessary for us to
show that the above example is typical. If the integral of a mapping f is known to
be path-independent, and A is an arbitrary fixed starting point, then we will show
that
ZZ
F(Z) ≡ f(z) dz (8.16)
A
is the antiderivative whose existence is sought. That is, we will verify that infinites-
imal arrows emanating from a point P are merely amplitwisted to produce their
images under this mapping F, and that the amplitwist at P is just f(P) i.e.,
F ′ (P) = f(P).
First we shall need a simple observation on differences of integrals. See [8.25a].
Two paths L and M are shown connecting the point A to the distinct points P and
Q. We know that for any function f(z),
Z Z Z
f(z) dz − f(z) dz = f(z) dz .
M L −L+M
The path (−L + M) for the right-hand integral is the round-about route from P to
Q shown in [8.25b]. But if the integral is known to be path-independent then we
may replace this path with the straight one S. Returning to the notation of (8.16),
we thus have
Z
F(Q) − F(P) = f(z) dz .
S
In the limit that Q coalesces with P, S becomes (with a minor abuse of terminol-
ogy) an infinitesimal ‘vector’ ∆ emanating from P, and its image ∆e under F will be
given by the left-hand side of the above equation. Thus

R R
[8.25] [a] Given
R R a fixed point R A, suppose we evaluate M f(z) dz and L f(z) dz. [b] Then
M f(z) dz − L f(z) dz = −L+M f(z) dz is an integral along an indirect route from P to
Q. If the integral is path-independent,
RZ we can replace the indirect route
R with the direct
route S. Defining F(Z) ≡ A f(z) dz, we deduce that F(Q) − F(P) = S f(z) dz.
464 Complex Integration: Cauchy’s Theorem

Z
e=
F : ∆ ÞÑ ∆ f(z) dz .
∆

But if ∆ is infinitesimal then the above integral equals f(P) ∆, thereby establishing
the original claim:
e = f(P) ∆ .
F : ∆ ÞÑ ∆

We may now appeal to Cauchy’s Theorem to relate the required path-indepen-

dence to analyticity. If f is analytic throughout some region then its integral is sure
to be path-independent, and therefore F will exist. In other words, every analytic
mapping must itself be the derivative of another analytic mapping.
We conclude this subsection with an interesting application of the analyticity
of F. In the previous subsection we completely ignored our previous geometric
constructions, and instead appealed to the Fundamental Theorem to show, for
example, that
ZB
1
z2 dz = (B3 − A3 ),
A 3

for all positions of A and B. This was apparently a clear improvement on [8.17]
where such formulae were merely established in the special case that A and
B were equidistant from the origin. However, we will now see that analytic-
ity makes it possible, paradoxically, for this special case to contain the general
case.
Consider these two functions:
ZZ
1
F(Z) = z2 dz G(Z) = (Z3 − A3 ) ,
A 3
both of which we now recognize as being analytic. From [8.17] we know that if Z
moves along the origin-centred circle passing through A then

F(Z) = G(Z) .

But by the uniqueness property of analytic functions [page 284] this identity must
continue to hold even if Z wanders off the circle, thereby establishing the general
result.
By applying exactly similar reasoning to the exponential mapping, we may like-
wise extrapolate the validity of (8.13) for vertically separated points (established
by [8.22]) to deduce that
ZZ
ez dz = eZ − eA ,
A

even if Z wanders off the vertical line through A.

 The Fundamental Theorem 465

8.8.5 Logarithm as Integral

In the light of the Fundamental Theorem, we are inclined to jump to the conclusion
that because (log z) ′ = (1/z),
ZZ
1
dz = log Z , (8.17)
1 z

just as in real analysis. In a sense, this is correct, but a little care is required.
The subtlety is, of course, that the singularity at the origin causes the integral
of (1/z) not to be single-valued. Thus we must specify the contour K from 1 to
Z before the integral in (8.17) becomes well defined. On the other hand, until we
choose one of the infinitely many values θ(Z) for the angle of Z, the RHS of (8.17)
is also not well defined. These two difficulties now cancel each other out in the
following way.
√
In [8.26] we have drawn three different contours for the specific case Z = 1+i 3.
If we let θK (Z) stand for the net rotation as we follow K, then
θK0 (Z) = (π/3)
θK1 (Z) = (π/3) + 2π
θK2 (Z) = (π/3) + 4π .
In a sense, including the contour in the definition of angle has rendered it single-
valued. Notice that this definition does not depend on the precise shape of K, but
only on how many times the origin is encircled.

RZ
[8.26] Why 1 (1/z) dz = log Z is a multifunction. From [8.12] we deduce that
RZ
1 (1/z) dz = logK (Z) = ln |Z| + i θK (Z), where θK (Z) denotes the net change in the
angle as z traverses K. So the infinitely many values of log Z arise from the number of
times K loops around 0 before finally arriving at Z, each extra revolution adding 2πi to
the value of log Z .
466 Complex Integration: Cauchy’s Theorem

In this way, we may absorb the means of reaching Z into the definition of log(Z)
in order to obtain a single-valued answer:

logK (Z) = ln |Z| + i θK (Z) .

The unambiguously correct version of (8.17) then reads,

Z
1
dz = logK (Z) .
K z

Of course, the multiple-valued nature of log has merely been disguised, not done
away with. Nevertheless, by pursuing the above idea one is led to consider so-
called Riemann surfaces, whereby multifunctions can be rendered single-valued. But
that is a story for another day.

8.9 Parametric Evaluation

When more elegant means are not available it is nevertheless possible (in principle)
to evaluate a contour integral by expressing it in terms of ordinary real integrals.
We shall now briefly describe and illustrate this method.
The basic idea is to think of the contour L as being traced by a moving particle
whose position at time t is z(t). Next, instead of building the Riemann sum (hence
the integral) from very small vectors that are chords of L, we may equally well use
very small vectors that are tangent to L. This is done using the tangential complex
velocity v = dzdt : the chord representing the movement during the instant of time
δt may be replaced by the tangential vector v δt. Thus if L is traced out during the
time interval a ⩽ t ⩽ b, then
Z Zb
f[z] dz = f[z(t)] v dt .
L a

For example, suppose that L is one counterclockwise circuit of the circle with
radius ρ and centre q, and that f[z] = z. We know from our earlier work that the
answer should be 2πiρ2 . Since z(t) = q + ρ eit (0 ⩽ t ⩽ 2π) and v = i ρ eit , we
obtain
Z Z 2π
z dz = (q + ρ e−it ) iρ eit dt
L 0
Z 2π Z 2π
= iρq (cos t + i sin t) dt + iρ2 dt
0 0

= 2πiρ2 ,

as anticipated.
Naturally, the point of this method is not to confirm previously known results,
but rather to evaluate integrals that we couldn’t do before. For example, with the
 Cauchy’s Theorem 467

same contour, but with f[z] = z2 , the answer can no longer be guessed. However,
you should now find it easy to discover that the answer is 4πiqρ2 .
By way of contrast with the non-analytic examples above, and as further practice
R
with this method, confirm (using the same contour L) that L z2 dz = 0, as predicted
by either Cauchy’s Theorem or the Fundamental Theorem. Likewise, confirm that
R it
E z dz = 0, where E is an origin-centred ellipse. [Hint: recall that z(t) = p e +
−it
qe moves on such an ellipse.]
For our last examples, take the contour to be a section of the parabola y = x2
between 0 and 1 + i; in temporal terms this can be represented as z(t) = t + it2 (0 ⩽
t ⩽ 1). Integrate z along this contour, first using the Fundamental Theorem, then
parametrically. Likewise, use the Fundamental Theorem to evaluate the integral
for ez . By equating the imaginary part of your answer with the imaginary part of
the parametric evaluation, deduce that
Z1
(2t cos t2 + sin t2 ) et dt = e sin 1 .
0

This result can be verified easily [exercise] without using complex numbers. Later,
though, we shall meet real integrals that cannot readily be evaluated by such ordi-
nary means, but which suddenly do become easy when viewed as arising from a
complex integral.

8.10 Cauchy’s Theorem

8.10.1 Some Preliminaries

Having repeatedly witnessed the utility of Cauchy’s Theorem in this chapter, it
is perhaps time that we checked that it is true! We begin with the case where the
contour C is a “simple” closed curve, i.e., without self-intersections. See [8.27].
We have filled the interior of C with a grid of small squares, of side length ϵ,
aligned with the real and imaginary axes. We have then shaded all those squares
that lie wholly within C, and taken the contour K to be the boundary of this
shaded region, traversed counterclockwise. Because we have drawn relatively
large squares (in order to make the picture clear), K is presently only a crude
approximation to C. However, as we let ϵ shrink, the shaded region fills the interior
of C ever more completely, and K follows C ever more precisely. Thus, in order to
see whether or not the integral of a mapping f along C vanishes, it is sufficient to
instead investigate the behaviour of the integral of f along K, as ϵ shrinks to zero.
[This is justified in greater detail in Ex. 20.]
Next we seek to relate this integral along K to the behaviour of the mapping
inside the shaded region that it bounds. Consider the sum of all the integrals of
f taken counterclockwise round each of the infinitesimal shaded squares. This
468 Complex Integration: Cauchy’s Theorem

[8.27] Cauchy’s Theorem (step one): Fill the interior of C with a grid of small (ultimately
vanishing) squares, of side length ϵ, aligned with the real and imaginary axes. The sum
of the integrals around all these squares yields the integral around the boundary, K, for
every interior edge is matched with an oppositely directed abutting edge that cancels it.

counterclockwise sense of integration is illustrated in [8.27] for two adjacent

squares. When we add the integrals from these two squares, their common edge is
traversed twice, once in each direction, and hence the integrals along it cancel. But
this is true of every edge that lies in the shaded region, so that when we sum the
integrals for all the shaded squares, the only edges that do not self-destruct in this
manner are those that make up K:
I X I
f(z) dz = f(z) dz . (8.18)
K shaded squares □

The investigation of the integral of f along C has thus been reduced to the study of
the local effect of f on infinitesimal squares in the interior region.
It should be stressed that the discussion thus far is equally applicable to non-
analytic and analytic mappings. For example, with f(z) = z, (8.18) simply says
[see (8.8)] that the area inside K is the sum of the areas of the shaded squares. In
order to understand Cauchy’s Theorem, we must specialize to the case (illustrated
in [8.27]) where the local effect of f is an amplitwist throughout the interior of C.
First, though, let us try to guess how the magnitude of a typical integral in the
above summation will depend on ϵ (as the squares shrink) for a general mapping.
Experience with real integration, as well as the inequality (8.5), might lead one
to guess that the integral round an infinitesimal square would die away at the same
rate as its perimeter, that is, as ϵ. This is false. The fact that the square is a closed
contour, together with the fact that complex integration is a type of vectorial sum-
mation, implies that the integral must decay much faster than this. In the above
example of conjugation, we know that the exact value of each term is 2iϵ2 , and this
 Cauchy’s Theorem 469

leads us to the correct guess, namely, that the terms die away as the square of ϵ. We
shall verify this in detail shortly, but for the moment the following rough argument
will suffice.
We know that for a general mapping, the integral round K—hence the summa-
tion in (8.18)—will be nonzero and finite. This leads us to believe that each term
must die away with the reciprocal dependence on ϵ as governs the growth of the
number of terms in the series. But the number of terms grows as (fixed area inside C,
divided by the area of each square), that is as (1/ϵ2 ). Thus the magnitude of each term
is expected to die away as ϵ2 . If our original guess had been correct, the order of the
sum in (8.18) would have been ϵ (1/ϵ2 ), yielding an infinite result as the squares
shrunk. Conversely, any contributions to the terms involving powers of ϵ greater
than two, cannot have any influence on the final result.

8.10.2 The Explanation

Let us return to [8.27] and to the explanation of Cauchy’s Theorem. The analytic
mapping f amplitwists the infinitesimal shaded squares on the left to the infinitesi-
mal squares on the right, and [8.28] shows a magnified view of a typical such square
and its image (the black ones in [8.27]). According to our especially accurate mid-
point Riemann sum (RM ), the integral along the bottom edge of this square can
be approximated by the single term A ϵ: the image of the midpoint a, times the
number along this edge. This conforms to our first, wrong guess concerning the
dependence on ϵ of the complete integral round the square. But if we now add this
to the integral along the opposite edge, the answer is

[8.28] Cauchy’s Theorem explained. According to our especially accurate midpoint Rie-
mann sum, RM , the integral along the bottom edge of this square can be approximated by
the single term Aϵ: the image of the midpoint a, times the number along this edge. If we
add this to the integral along the opposite edge, the answer is A ϵ+C (−ϵ) = (A−C) ϵ =
pϵ . Likewise, the contribution from the remaining two edges is also of order ϵ2 , namely,
(B − D) iϵ = iqϵ. But since f is locally an amplitwist,H the image is a square, and so
iq = (q rotated through a right angle) = −p. Therefore, □ f(z) dz = ϵ (p + iq) = 0.
470 Complex Integration: Cauchy’s Theorem

A ϵ + C (−ϵ) = (A − C) ϵ = pϵ .
Even if f is merely differentiable in the real sense, rather than locally an amplitwist,
|p| will still be proportional to ϵ, and the magnitude of pϵ will therefore be pro-
portional to ϵ2 , as anticipated. Likewise, the contribution from the remaining two
edges is also of order ϵ2 , namely, (B − D) iϵ = iq ϵ.
Perhaps you have already seen the light: if f is locally an amplitwist, the image
is a square, and so
iq = q rotated through a right angle = −p
I
=⇒ f(z) dz = ϵ (p + iq) = 0 . (8.19)
□
We conclude from (8.18) that the vanishing of loop integrals for analytic mappings
is indeed the nonlocal manifestation of their local amplitwist property!
Contrast this with non-analytic mappings. See [8.29]. Provided that a mapping is
differentiable in the real sense, we know [see page 236] that its local effect is expan-
sion (by different factors) in two perpendicular directions, followed by a twist.
Thus the image of an infinitesimal square will generally be a parallelogram; p and
q will not have equal length, nor will they be orthogonal. As we see, p and iq no
longer cancel, and ϵ (p + iq) is of order ϵ2 . When we add up the terms of (8.18), of
order (1/ϵ2 ) in number, the answer will therefore be nonzero and finite.
Conjugation provides a particularly striking example of this noncancellation for
non-analytic mappings. See [8.30]. In the terms of the previous paragraph we could
say that its expansion factors are everywhere 1 and −1 (in the horizontal and verti-
cal directions), and that its twist is zero. The image of the square is again a square,
but there is a crucial difference between [8.28] and [8.30]. Because this mapping is
anticonformal, it reverses the orientation of the square, and we see that

[8.29] The non-analytic case. The shrinking square is ultimately mapped to a (non-
square) parallelogram. As we see, p and iq no longer cancel, and ϵ(p + iq) is of order
ϵ2 . When we add up the terms of (8.18), of order (1/ϵ2 ) in number, the answer will
therefore be nonzero and finite.
 Cauchy’s Theorem 471

H
[8.30] □ z dz = 2i(area of □), because ϵ(p + iq) = ϵ(iϵ + iϵ) = 2i ϵ2 .
I
z dz = ϵ (p + iq) = ϵ (iϵ + iϵ) = 2i ϵ2 = 2i (area of □).
□

Returning to questions of analyticity, and comparing (8.19) with [8.29], we obtain

a converse to Cauchy’s Theorem. If all the loop integrals of f are known to vanish,
then, in particular, they will vanish for infinitesimal squares, such as the one on the
left of [8.29]. Thus, p + iq = 0. But it is clear that this can only happen if the image
is another infinitesimal square with the same orientation as the original (cf. [8.30]).
Thus the local effect of f must be an amplitwist. This converse is called Morera’s
Theorem.
As with other new ideas in this book, we have not attempted to present the
arguments in rigorous form; “insight”, not “proof”, is ever our watchword. For
example, consider these objections (ascending in severity) to the geometric argu-
ment of [8.27] and [8.28]: no matter how small the square, the sides of the image
will not be perfectly straight (though they will meet in perfect right angles); the
midpoint a will not be mapped to the exact midpoint of the image; and despite the
undoubted accuracy of RM for a very small contour, it will not yield the exact value
of the integral.
Nevertheless, it seems plausible that [8.28] and its associated reasoning remain
unimpeached when it comes to the evaluation of the dominant ϵ2 contribution.
Indeed, the example of [8.30] lends at least some credence to the irrelevance of
the above objections, for in that case we know the answer is correct to this order of
ϵ. [In fact it comes out exactly right, but that is a fluke.] More generally, recall that
parametric evaluation revealed that the real and imaginary parts of any contour
integral can be expressed as ordinary real integrals. This means that we may carry
over to the complex realm our previous determination (8.3) of the error induced by
RM in real analysis. Thus each of the integrals along the four edges of the square
will differ from their RM -values by an amount that dies away at least as fast as ϵ
cubed [cf. Ex. 21 and Ex. 22]. But as we have previously argued, as ϵ shrinks to zero,
such contributions can have no effect on the sum in (8.18). Although we shall not
dwell on them, other objections can be treated in a similar way.
472 Complex Integration: Cauchy’s Theorem

8.11 The General Cauchy Theorem

8.11.1 The Result

Consider a mapping f that is analytic except at the singularity marked in [8.31].
Must the integral of f round K vanish, or not? You see the problem. For a simple
loop without self-intersections (such as in [8.27]) it is perfectly clear whether a sin-
gularity is lurking inside, and consequently whether Cauchy’s Theorem applies.
But in [8.31], it is not even clear what “inside” means, let alone how this might
relate to Cauchy’s Theorem.
Recall that we encountered such problems before when trying to integrate z
round complicated loops [see (8.9), as well as the discussion on p. 387]. Our solu-
tion was to define the “inside” to be all the points for which the winding number
does not vanish, and conversely, the “outside” to be all the points for which it does
vanish.
With these definitions in place, the completely general version of Cauchy’s
Theorem is stunning in its simplicity:

If an analytic mapping has no singularities “inside” a loop, its

(8.20)
integral round the loop vanishes.

This section is devoted to understanding this beautiful result.

First let us answer our opening question. In [8.31], K does not wind around the
singularity, and therefore (according to the theorem) the integral should vanish. In

[8.31] As in Section 7.1.2, we define the “inside” of a closed loop to be the set of points
for which the winding number is not zero. Then if an analytic mapping has no singularities
inside a loop, its integral round the loop vanishes, as it does in this example.
 The General Cauchy Theorem 473

the process of understanding this particular instance of the theorem we shall be led
to a completely general argument for its validity.

8.11.2 The Explanation

As in [8.16b], the contour K in [8.31] partitions the plane into a number of disjoint
regions; in particular, the inside of K is made up of D1 , D2 , and D3 . See [8.32]. Let
Cj be the boundary of Dj , traversed counterclockwise. So as not to clutter up the
picture, instead of actually drawing these contours in [8.32], we have merely indi-
cated (with ellipses) their common counterclockwise sense. Also shown (in boxes)
are the winding numbers of K around each of the regions Dj . Since there are no sin-
gularities inside the Dj that make up the inside of K, our basic version of Cauchy’s
Theorem applies to each of the simple contours Cj , and we have
I
f(z) dz = 0 . (8.21)
Cj

Now comes the crucial observation. The integral round K can be expressed as a
linear combination of the integrals round the Cj ’s that bound the interior Dj ’s. In
the case of [8.32],
I I I I
f(z) dz = f(z) dz − f(z) dz + 2 f(z) dz . (8.22)
K C1 C2 C3

Consider, for example, the contour C3 , for which the counterclockwise sense hap-
pens to agree with the direction of K. On the other hand, C1 traverses this portion of
K in the opposite direction. Consequently, in (8.22), we end up integrating twice in
the correct direction, and once in the opposite direction; the net result is to integrate

P
[8.32] K = j νj Cj = C1 − C2 + 2 C3 in this example. [The explanation is in the next
figure.]
474 Complex Integration: Cauchy’s Theorem

P
[8.33] Proof that K = j νj Cj . Using the “crossing rule” (7.1) on page 388, we see
that νj = νk + 1. Thus the contour Cj in the direction of K will always occur precisely
one more time than the contour Ck in the opposite direction—the net result is that K is
traversed once in the correct direction.

along this part of K once in the correct direction. You should check for yourself that
all of K is correctly accounted for in this way. Substituting (8.21) into (8.22), we
have confirmed the prediction of the general theorem for this particular contour.
Since (8.22) is clearly true for any function f, we may abstract it away and write
the equation as
K = C1 − C2 + 2 C3 .
Notice that the coefficient of Cj in this sum is none other than the winding number
νj of K about Dj , and that we may therefore rewrite the previous equation as
X
K= νj Cj . (8.23)
j

From the above example, it is clear that to prove the general version of Cauchy’s
Theorem we need only show that (8.23) is true for any K.
Consider [8.33], which shows a portion of an arbitrary contour K sandwiched
between two of the regions (Dj and Dk ) into which it partitions the plane; also
shown is the counterclockwise sense of their boundaries (Cj and Ck ). Using the
“crossing rule” (7.1) on page 388, we deduce that νj = νk + 1. Thus, in (8.23),
we find that the contour Cj in the direction of K will always occur precisely one
more time than the contour Ck in the opposite direction—the net result is that K is
traversed once in the correct direction. Done.
As previously explained, in establishing (8.23) we have also deduced the General
Cauchy Theorem (8.20).

8.11.3 A Simpler Explanation

The Deformation Theorem (8.11) was also deduced from the basic Cauchy theorem
[unproven at that time], and we will now use it to give a simpler and more intuitive
explanation of the General Cauchy Theorem.
 The General Formula of Contour Integration 475

[8.34] If a closed contour can be shrunk to a point without crossing a singularity, the
integral round it vanishes. Hopf’s Degree Theorem tells us that this shrinking process is
possible if and only if the contour does not wind around any singularities.

Suppose that a contour can be deformed and shrunk down to a point without
ever crossing a singularity of an otherwise analytic function. By inequality (8.5),
the value of the integral will be zero at the end of this shrinking process. But by the
Deformation Theorem, the value of the integral remains constant throughout this
process. In other words,
If a closed contour can be shrunk to a point without crossing a singularity,
the integral round it vanishes. (8.24)
To wrap this up, we clearly need a way of recognizing when this shrinking pro-
cess is possible. For example, is it possible for the contour K in [8.31]? Figure [8.34]
shows that it is. Therefore (8.24) implies that the integral along K vanishes, in
agreement with the general theorem.
The two theorems are clearly very closely related. In fact we can now deduce the
General Cauchy Theorem from (8.24) by observing that the winding number of the
final shrunken loop vanishes; Hopf’s Degree Theorem then tells us that
The shrinking process in (8.24) is possible if and only if the contour does not
wind around any singularities.

8.12 The General Formula of Contour Integration

H
Consider the general problem of evaluating K f(z) dz, where K is a general (pos-
sibly self-intersecting) loop, and where f possesses several singularities s1 , s2 , etc.,
inside K. Figure [8.35] illustrates such a situation.
476 Complex Integration: Cauchy’s Theorem

H P
[8.35] The General Residue Theorem: KHf(z) dz = 2πi j ν(K, H sj ) Res [f(z), sj ]. We
P P
know that K = j νj Cj , and therefore K f(z) dz = j νj Cj f(z) dz. The Deforma-
H
tion Theorem then tells us that each Cj f(z) dz can be expressed as a sum of integrals
around the illustrated
H loops σj encircling the singularities within each Cj . Finally, each
such integral σk f(z) dz = 2πi Res [f(z), sk ].

Here the inside of K consists of two simply connected regions, D1 [lightly

shaded] and D2 [darkly shaded], with boundaries C1 and C2 . Since K winds once
round points in the lightly shaded region (ν1 = 1) and twice round points in the
darkly shaded region (ν2 = 2), the general result (8.23) correctly predicts that
X
K= νj Cj = 1 · C1 + 2 · C2 . (8.25)
j

As illustrated, let σj be a simple (counterclockwise) contour containing sj but no

other singularities of f, and let us define
I
Ij ≡ f(z) dz.
σj

By virtue of the Deformation Theorem (8.11), we know that the integral Ij has a
characteristic value that does not depend on the size or shape of σj . Furthermore,
as we saw in [8.20], if a simple loop contains several singularities, then the inte-
gral round that loop is the sum of I-values of the singularities it contains. In our
example,
I I
f(z) dz = I1 + I2 + I5 and f(z) dz = I3 + I4 .
C1 C2

Finally, using (8.25), we deduce that

I
f(z) dz = 1 · [I1 + I2 + I5 ] + 2 · [I3 + I4 ],
K
 The General Formula of Contour Integration 477

in which each I-value has been multiplied by the number of times K winds round the
corresponding singularity. Since (8.23) is valid for arbitrary loops it follows that this
conclusion is too. As the grand finale to this chapter, we have thus obtained the
following completely general formula for the loop integral of an analytic function:
I X
f(z) dz = ν(K, sj ) Ij .
K j

The final icing on the cake is an efficient method of computing the Ij ’s. In the
next chapter we will verify our previous claim that in the neighbourhood of each
sj there exists a unique Laurent series [see page 456], the coefficient of the complex
inversion term being (by definition) the residue Res [f(z), sj ]. Granted this, we see
that Ij = 2πi Res [f(z), sj ]. Thus
I X
f(z) dz = 2πi ν(K, sj ) Res [f(z), sj ]. (8.26)
K j

This is the General Residue Theorem. Note that it contains the General Cauchy
Theorem as the special case in which each ν(K, sj ) = 0.
We will also see in the next chapter that it is possible to find the residues in this
formula directly, without going to the trouble of finding the whole Laurent series.
Thus, even before exemplifying its use, it should be clear that in (8.26) we have a
result of great practical and theoretical power.
478 Complex Integration: Cauchy’s Theorem

8.13 Exercises

1 Thinking of x as representing time, z(x) = x + if(x) is a parametric description

of the ordinary graph, y = f(x).
(i) Show that the complex velocity is v = 1 + i tan θ, where θ is the angle
between the horizontal and the tangent to the graph. Also, show that
complex acceleration is a = i f ′′ .
(ii) Recall from Ex. 20 on page 297 that the curvature of the orbit is κ =
[Im (a v)]/|v|3 . Deduce from (i) that
κ sec3 θ = f ′′ (x) .
(iii) From (ii), deduce that the error equation (8.3) can be written as
1 ′′
area (ABCD) = f (x) ∆3 .
8
2 In [8.6], show that
 
area between the chord AB and the curve
lim = 2.
Ƅ0 area between the tangent CD and the curve

In other words, RM is twice as accurate as the Trapezoidal formula.

3 In the integration of an ordinary real function f(x), let L denote the length of the
integration range, and let M denote the maximum size of f ′′ (x) in this range.
From the previous two exercises, deduce the standard result,
1 2
total Trapezoidal error < 12 LM∆ .
Likewise, deduce the somewhat less familiar result,
total RM error < 241
LM∆2 .
H
4 Write down the values of C (1/z) dz for each of the following choices of C, then
confirm the answers the hard way, using parametric evaluation.
(i) |z| = 1.
(ii) |z − 2| = 1.
(iii) |z − 1| = 2.
5 Evaluate parametrically the integral of (1/z) round the square with vertices
±1 ± i, and confirm that the answer is indeed 2πi.
6 Confirm by parametric evaluation that the integral of zm round an origin-
centred circle vanishes, except when m = −1.
7 Hold a coin (of radius A) down on a flat surface and roll another one (of radius
B) round it. The path traced by a point on the rim of the rolling coin is called
an epicycloid, and it is a closed curve if A = nB, where n is an integer.
 Exercises 479

(i) With the centre of the fixed coin at the origin, show that the epicycloid can
be represented parametrically as
h i
z(t) = B (n + 1) eit − ei(n+1)t .
(ii) By evaluating the integral in (8.8) parametrically, show that
area of epicycloid = πB2 (n + 1) (n + 2) .
8 The figure below shows four simple loops, and in each case we have indi-
cated how much shaded area is enclosed. Use parametric evaluation to verify
equation (8.8) for each of the four loops.

9 What is the generalization of (8.8) to the case where the contour is not closed?
10 Use (8.23) to verify (8.9).
11 The perfect symmetry of figure [8.18] results from integration round the unit
circle. Roughly how would this figure look if we instead used a somewhat
larger circle?
12 Let K be the contour in [8.21].
(i) Evaluate the following integral by factoring the denominator and putting
the integrand into partial fractions:
I 
z
dz .
z2 − iz − 1 − i
K

(ii) Write down the Laurent series (centred at the origin) for (cos z/z11 ). Hence
find
I
cos z 
dz .
z11
K

13 This exercise illustrates how one type of difficult real integral may be evaluated
easily using a complex integral.
Let L be the straight contour along the real axis from −R to +R, and let J be
the semi-circular contour (in the upper half plane) back from +R to −R. The
complete contour L + J is thus a closed loop.
480 Complex Integration: Cauchy’s Theorem

(i) Using the partial fraction idea of the previous exercise, show that the
integral
I
dz
4
(z + 1)
L+J

vanishes if R < 1, and find its value if R > 1.

(ii) Using the fact that z4 +1 is the complex number from −1 to z4 , write down
the minimum value of |z4 + 1| as z travels round J. Now think of R as large,
and use inequality (8.5) to show that the integral round J dies away to zero
as R grows to infinity.
(iii) From the previous parts, deduce the value of
Z +∞
dx
4
.
−∞ (x + 1)

14 (i) The integral

Z +∞
dx
−∞ (x2 + 1)
is easily found by ordinary means, but evaluate it instead by the method
of the previous exercise.
(ii) Likewise, evaluate
Z +∞
dx
−∞ (x2 + 1)2
by ordinary means and then by contour integration. [Hint: The quickest
way to find the partial fraction decomposition for this function is to square
the decomposition of 1/(z2 + 1).]
15 (i) Use the Fundamental Theorem to write down the value of
Z a+ib
ez dz .
0

(ii) Equate the answer with the one obtained by parametric evaluation along
the straight contour from 0 to (a + ib), and deduce that
Z1
a (ea cos b − 1) + b ea sin b
eax cos bx dx = ,
0 a2 + b 2
and
Z1
b (1 − ea cos b) + a ea sin b
eax sin bx dx = .
0 a2 + b 2
(iii) Prove the results in (ii) by ordinary methods.
 Exercises 481

16 (i) Show that when integrating a product of analytic functions, we may use
the ordinary method of “integration by parts”.
(ii) Let L be a contour from the real number −θ to +θ. Show that
Z
z eiz dz = 2i (sin θ − θ cos θ),
L
and verify this by taking L to be a line-segment and integrating parametri-
cally.
17 Let
 n
1 1
f(z) = z+ ,
z z
where n is a positive integer.
(i) Use the Binomial Theorem to find the residue of f at the origin when n is
even and when n is odd.
(ii) If n is odd, what is the value of the integral of f round any loop?
(iii) If n = 2m is even and C is a simple loop winding once round the origin,
deduce from part (i) that
I
(2m)!
f(z) dz = 2πi .
(m!)2
C
(iv) By taking C to be the unit circle, deduce the following result due to Wallis:
Z 2π
(2m)!
cos2m θ dθ = 2m−1 π.
0 2 (m!)2
(v) Similarly, by considering functions of the form zk f(z) where k is an
integer, evaluate
Z 2π Z 2π
cos θ · cos kθ dθ and
n
cosn θ · sin kθ dθ
0 0
18 Let E be the elliptical orbit z(t) = a cos t + ib sin t, where a and b are positive
and t varies from 0 to 2π. By considering the integral of (1/z) round E, show
that
Z 2π
dt 2π
2 cos2 t + b2 sin2 t
= .
0 a ab
19 Let us verify the claim of Ex. 19, p. 297, that if a function has vanishing
Schwarzian derivative, then it must be a Möbius transformation. Following
Beardon (1984), suppose that {f(z), z} = 0, and define F ≡ (f ′′ /f ′ ).
(i) Show that 1/F(z) − 1/F(w) = −(z − w)/2.
(ii) Deduce that d
dz log f ′ (z) = −2/(z − a), for some constant a.
(iii) Perform two further integrations to conclude that f(z) is a Möbius trans-
formation.
482 Complex Integration: Cauchy’s Theorem

20 In [8.27], consider the white fragments of squares sandwiched between

K and C.
(i) Show that the sum of the integrals round these fragments equals the
difference between the integrals round C and K.
(ii) As ϵ shrinks, what is the approximate size of each term in the above series?
(iii) Roughly how many terms are there in the series?
(iv) From the previous parts, what do you conclude about the difference
between the integrals round C and K, as ϵ shrinks to nothing?
21 Let K be the straight contour from a − (ϵ/2) to a + (ϵ/2), where ϵ is a short
complex number in an arbitrary direction.
(i) Use the Fundamental Theorem to integrate z2 along K, and then write
down the value obtained by using a single term in RM . Show that the
1 3
error induced by RM is 12 ϵ.
(ii) As in part (i), find both the exact value and the RM value for the integral
of ez along K. By expanding eϵ/2 as a power series, deduce that the error
1 a 3
in this case is roughly 24 e ϵ.
(iii) Repeat the error analysis of the previous parts for the non-analytic func-
tion z2 . [You will need to use parametric evaluation to find the exact value
of the integral.]
22 Let K be the short contour of the previous exercise. Suppose that f(z) possesses
a Taylor series centred at a that converges at points of K:

f ′ (a) f ′′ (a) 2 f ′′′ (a) 3

f(a + h) = f(a) + h+ h + h + ··· .
1! 2! 3!

[The existence of such a series for any analytic function is derived in the next
chapter.]
(i) By integrating this series along K, show that the difference between the
1 ′′
exact integral and the RM value is roughly 24 f (a) ϵ3 . Verify that the
results of the first two parts of the previous exercise are in accord with
this finding.
(ii) Use the series to show that the complex number from the image of the
midpoint of K to the midpoint of the images of the ends of K is roughly
1 ′′ 2
4 f (a) ϵ . As ϵ shrinks, are these two types of midpoint distinguishable
under the magnifying lens that produces figure [8.28]?
(iii) From the Fundamental Theorem, deduce that the existence of such a series
implies the vanishing of the integral of f round loops within the disc of
convergence.
 Exercises 483

23 Let f(z) be analytic throughout a region which contains a triangle with vertices
a, b, c, and hence with edges A ≡ (c − b), B ≡ (a − c), C ≡ (b − a). Given a
pair of point p and q, let us define wpq as a kind of average of f(z) along the
line-segment pq:
Zq
1
wpq ≡ f(z) dz.
(q − p) p
Show that this complex average mapping sends the sides of the triangle abc
to the vertices wab , wbc , wca of a similar triangle! We merely rediscovered this
result, which is apparently due to Echols (1923).
[Hint: Show that Awbc + Bwca + Cwab = 0, and use A + B + C = 0.]
24 Let K be a closed contour, and let ν be its winding number about the point a.
Show that
I z 
e
dz = 2πi ν ea .
z−a
K
(z−a)
z
[Hint: Write e as e e a
, and expand e(z−a) as a power series.] This is a
special case of Cauchy’s Integral Formula (explained in the next chapter), which
states that if f is analytic inside K, then
I
f(z)
dz = 2πi ν f(a) .
(z − a)
K
25 Consider the image of the disc |z| ⩽ R under the mapping z ÞÑ k zm . As
the radius sweeps round the disc once, its image sweeps m times round the
image disc of radius |k| Rm . Thus we may sensibly define the area of the image
to be m π (|k| Rm )2 . With this understanding, show that if a mapping has a
convergent power series
f(z) = a + b z + c z2 + d z3 + · · · ,
then the area of the image is just the sum of the areas of the images under each
of the separate terms of the series:
area of image = π (|b|2 R2 + 2 |c|2 R4 + 3 |d|2 R6 + · · · ) .
This is Bieberbach’s Area Theorem.
Hint: Recall that the local area expansion factor is |f ′ |2 , so the image area is
ZZ Z R  Z 2π 
′2 ′ ′
|f | dx dy = iθ iθ
f (r e ) f (r e ) dθ r dr.
|z|⩽R 0 0
26 (i) Show that if f is an analytic function without singularities or p -points on
a loop L, then
I
1 f ′ (z)
ν [f(L), p ] = dz.
2πi f(z) − p
L
484 Complex Integration: Cauchy’s Theorem

(ii) Now let

(z − a1 )A1 (z − a2 )A2 · · · (z − an )An
f(z) = ,
(z − b1 )B1 (z − b2 )B2 · · · (z − bm )Bm
and by considering (log f) ′ , find (f ′ /f).
(iii) In part (i) put p = 0 and take L to be a simple loop containing the roots a1
to ar and containing the poles b1 to bs . Thereby obtain a calculation proof
of the Generalized Argument Principle in the case of rational functions:
X
r X
s
ν [f(L), 0 ] = Aj − Bj
j=1 j=1

= (number of interior roots) − (number of interior poles).

CHAPTER 9

Cauchy’s Formula and Its Applications

9.1 Cauchy’s Formula

9.1.1 Introduction
One of the principal objectives of this brief chapter is to tie up various loose ends
from previous chapters. In particular, we have previously claimed (but have not
yet explained) three important properties of an analytic function f(z):

• We can differentiate f(z) as many times as we please—it is “infinitely differen-

tiable”.
• In the vicinity of an ordinary point, f(z) can be expressed as a Taylor series.
• In the vicinity of a singularity, f(z) can be expressed as a Laurent series.

The classical explanation1 of these facts hinges on the following result. If f(z) is
analytic on and inside a simple loop L, and if a is a point inside L, then
I
1 f(z)
dz = f(a). (9.1)
2πi z − a
L

This is called Cauchy’s Formula—it constitutes the precise statement of the “rigidity”
of analytic functions that we depicted in [5.3], p. 249. That is, the formula says that
the values of f on L rigidly determine its values everywhere inside L.
We will give two explanations of (9.1), both of which are firmly rooted in
Cauchy’s Theorem.

1
In the late 1950s a new approach was developed using topological ideas like those in Chapter 7,
and it was our original intention to employ that approach here. However, having lacked both the time
and the imagination to reduce the idea to its visual essentials, we have reluctantly fallen back on an
integral-based approach. For more on the topological approach, see Whyburn (1955), Whyburn (2015)
and Beardon (1979).

Visual Complex Analysis. 25th Anniversary Edition. Tristan Needham, Oxford University Press.
© Tristan Needham (2023). DOI: 10.1093/oso/9780192868916.003.0009
486 Cauchy’s Formula and Its Applications

9.1.2 First Explanation

Since f(z) is assumed analytic inside L, the function [f(z)/(z − a)] is also analytic
there, except that it has a single singularity at z = a. Thus it follows from (8.11),
p. 454, that the value of the integral in (9.1) will not change if L is deformed into its
interior without crossing a.
Let Cr be a circle of radius r, centred at a, and lying strictly inside L. Referring to
[9.1a], we may deform L into such a circle without crossing a, and hence without
altering the value of the integral:
I I
1 f(z) 1 f(z)
dz = dz. (9.2)
2πi z − a 2πi z−a
L Cr

The virtue of this transformation is that the integral round Cr turns out to have a
simple and helpful interpretation.
First recall that the average value ⟨f⟩Cr of f(zθ ) as zθ = a + r eiθ travels round
the circle Cr is defined by
Z
1 2π
⟨f⟩Cr ≡ f(zθ ) dθ.
2π 0
In the previous chapter we saw geometrically that if θ increases by dθ, causing
zθ to move dz along the circle, then dz/(z − a) = i dθ. Substituting this into (9.2),

[9.1] First explanation of Cauchy’s Formula. [a] If f(z) is analytic inside L then so is
f(z)/(z − a), except at z = a, so the Deformation Theorem implies that we can deform L
into the circle Cr of radius r without changing the value of the integral. But on Cr we can
interpret the integral as the average value ⟨f⟩Cr of f(z) on Cr , and, by the Deformation
Theorem again, this average value is independent of the radius. If zj are n equally spaced
points around Cr then ⟨f⟩Cr can be visualized as the limit as n Ñ ∞ of the centroid of
the image points f(zj ). [b] As Cr shrinks towards a, these image points cluster more and
H f(z)
L z−a dz = ⟨f⟩Cr = f(a).
1
more tightly around f(a), so 2πi
 Cauchy’s Formula 487

we find that the original integral round L may be interpreted as the average value
of f on any of the circles Cr :
I
1 f(z)
dz = ⟨f⟩Cr .
2πi z − a
L

Note in particular that ⟨f⟩Cr is independent of the radius r of the circle. To

complete the derivation of (9.1), it therefore only remains to show that this radius-
independent average is the value of f at the centre a.
To better grasp the meaning of the average value ⟨f⟩Cr , imagine n equally spaced
points z1 , z2 , . . . , zn on Cr , and let w1 , w2 , . . . , wn be their images under z ÞÑ w =
P
f(z). The ordinary average Wn ≡ n1 n j=1 wj of these image points is their centroid,
and ⟨f⟩Cr is the limiting position of Wn as n tends to infinity. [For a more detailed
discussion of averages and centroids, consult the final section of Chapter 2.]
Now shrink the circle Cr towards its centre a, as illustrated in [9.1a]. Even if f is
merely continuous (rather than analytic), f(Cr ) will shrink to f(a), as indicated in
[9.1b]. Since the images w1 , w2 , . . ., wn of any n points on Cr will all converge to
f(a), so will their centroid Wn . Thus
lim ⟨f⟩Cr = f(a),
rÑ0

and this completes our first explanation of Cauchy’s Formula.

9.1.3 Gauss’s Mean Value Theorem

In the course of the above investigation we have also picked up an interesting
bonus result:
If f(z) is analytic on and inside a circle C centred at a, then the average value
of f on C is its value at the centre: ⟨f⟩C = f(a).
If we go on to split f into real and imaginary parts as f = u+iv, then we immediately
deduce that ⟨u⟩C + i ⟨v⟩C = u(a) + iv(a), and so
⟨u⟩C = u(a) and ⟨v⟩C = v(a).
Thus if a real function Φ is either the real or the imaginary part of an analytic
complex function, then its average on a circle is its value at the centre.
But if we are given a function Φ, how can we tell whether there exists an analytic
function whose real or imaginary part is equal to Φ? In Ex. 2, p. 293 you showed that
a necessary condition is that Φ be harmonic, i.e., that it satisfy Laplace’s equation,
∆Φ ≡ (Bx2 + By2 )Φ = 0.
In fact in Chapter 12 we will see that this is also a sufficient condition, yielding
Gauss’s Mean Value Theorem:
The average value of a harmonic function on a circle is equal to the value
of the function at the centre of the circle.
488 Cauchy’s Formula and Its Applications

9.1.4 A Second Explanation and the General Cauchy Formula

What will happen to Cauchy’s Formula if the loop L is not required to be simple?
As in the previous chapter, it is now important to carefully define the “inside” of
L as the set of points about which L has non-vanishing winding number:

“inside” = { p | ν[L, p] ̸= 0 }.

Suppose in [9.2] that f has no singularities “inside” L. Then the only interior
singularity of [f(z)/(z − a)] will be the one at z = a. Here, L winds round a twice,
and it is clear that L may be deformed into a small circle centred at a and traversed
ν[L, a] = 2 times. By virtue of Cauchy’s Formula for simple loops, we deduce that
1
H f(z)
2πi L z−a dz = 2 f(a).
More generally, this line of reasoning suggests the following General Cauchy
Formula: If f(z) is analytic on and “inside” a general loop L, then
I
1 f(z)
dz = ν[L, a] f(a). (9.3)
2πi z − a
L

That this is always true is not quite clear from the above line of reasoning. Certainly
Hopf’s Theorem [(7.2), p. 389] guarantees that without crossing the singularity at
a, L may be deformed into a circle centred at a and traversed ν[L, a] times. But the
singularities of f may be scattered in the midst of L, although (by assumption) none
lie “inside” L. So is it clear that this deformation can always be performed without
crossing any of these singularities?

1
H f(z)
[9.2] Second explanation and the General Cauchy Formula: 2πi L z−a dz = ν[L, a] f(a).
Define Fa (z) to be the non-infinitesimal analogue of the amplitwist, expanding and
e so that V
rotating V to its “image” V, e = Fa (z) V. Indeed, as V vanishes, the analogy
becomes exact: Fa (a) ≡ limzÑa Fa (z) = f ′ (a). Since Fa (z) = [f(z) − f(a)]/[z − a]
is analytic inside L, its integral
H f(z) H dz vanishes, by Cauchy’s Theorem, and therefore
H f(z)
1 1 1
0 = 2πi L z−a dz − f(a) 2πi L z−a = 2πi L z−a dz − ν[L, a] f(a), proving the result.
 Infinite Differentiability and Taylor Series 489

We encourage you to pursue this idea, but we shall now present a different
approach which yields (9.3) cleanly and directly. Consider the mapping z ÞÑ e
z=
f(z) in [9.2], and let us define
f(z) − f(a) e e
z−a
Fa (z) ≡ = .
z−a z−a
If V ≡ (z − a) is pictured as a vector emanating from a, and V e ≡ (ez−a e) is pic-
tured as its image emanating from a e, then Fa (z) describes the amount of rotation
e
and expansion that carries V into V = Fa (z) V. Thus Fa (z) is the non-infinitesimal
analogue of the amplitwist f ′ (a) that carries an infinitesimal vector ξ into its image
e
ξ = f ′ (a) ξ, and
Fa (a) ≡ lim Fa (z) = f ′ (a).
zÑa

Since f(z) is assumed to be analytic and to have no singularities “inside” L, it

follows that the same is true of Fa (z). Thus the General Cauchy Theorem [(8.20),
p. 472] implies that
I
1
Fa (z) dz = 0.
2πi
L

In other words,
I I
1 f(z) 1 dz
0 = dz − f(a)
2πi z−a 2πi z−a
L L

I
1 f(z)
= dz − ν[L, a] f(a),
2πi z−a
L

as was to be shown.

9.2 Infinite Differentiability and Taylor Series

9.2.1 Infinite Differentiability

Returning to the case where L is a simple loop, let us show that if f(z) is analytic
inside L then so is f ′ (z). From this it will follow by induction that f(z) is infinitely
differentiable.
What we must show is that if f is conformal, then so is f ′ . In other words, if f ′ is
thought of as a mapping z ÞÑ e z = f ′ (z), then each infinitesimal vector ξ emanating
from a must be rotated and expanded the same amount to obtain the image vector e ξ
e ′′
emanating from a. That is, there is a single complex number f (a) (the amplitwist
of f ′ ) such that e
ξ = f ′′ (a) ξ.
490 Cauchy’s Formula and Its Applications

Our first step is to obtain a neat expression for f ′ (a) in terms of the values of f(z)
on L. Applying Cauchy’s Formula to the analytic function Fa (z), we deduce that
I
′ 1 Fa (z)
f (a) = Fa (a) = dz
2πi z − a
L

I I
1 f(z) f(a) dz
= 2
dz − .
2πi (z − a) 2πi (z − a)2
L L

Since the second integral vanishes,

I
′ 1 f(z)
f (a) = dz. (9.4)
2πi (z − a)2
L

Now let’s use this to find the image e z = f ′ (z) of a short vector ξ
ξ under z ÞÑ e
emanating from a. Ignoring a term proportional to ξ2 , we find [exercise] that
 
I
e 2 f(z)
ξ ≡ f ′ (a + ξ) − f ′ (a) =  dz ξ.
2πi [z − (a + ξ)]2 (z − a)
L

Allowing ξ to become infinitesimal, we deduce the desired result: every infinitesi-

mal ξ emanating from a is amplitwisted to eξ = f ′′ (a) ξ, where
I
2 f(z)
f ′′ (a) = dz. (9.5)
2πi (z − a)3
L

Observe that since

   
d 1 1 d2 1 2
= and = ,
da z − a (z − a)2 da2 z − a (z − a)3

both (9.4) and (9.5) are precisely what we would get if we simply differentiated the
formula
I
1 f(z)
f(a) = dz
2πi z − a
L

with respect to a. Continuing in this way, we are led to conjecture that the nth
derivative f(n) may be represented as
I
(n) n! f(z)
f (a) = dz. (9.6)
2πi (z − a)n+1
L

This is indeed true, as we shall see in a moment.

 Infinite Differentiability and Taylor Series 491

9.2.2 Taylor Series

Now let us show that if f(z) is analytic on and inside an origin-centred circle C of
radius R, then f(z) may be expressed as a power series that converges inside this
disc:

f(z) = c0 + c1 z + c2 z2 + c3 z3 + · · · .

As we saw in Chapter 5, such a power series is infinitely differentiable within its

disc of convergence. Thus the existence of the power series expansion will provide
a second proof of the infinite differentiability of analytic functions. It also follows
that the coefficients cn may be expressed as

f(n) (0)
cn = , (9.7)
n!

so the power series is actually a Taylor series, and the coefficients do not depend
on R:

f ′′ (0) 2 f(3) (0) 3

f(z) = f(0) + f ′ (0) z + z + z + ··· .
2! 3!

To establish the existence of this series, we return to Cauchy’s Formula (9.1).

With a change of notation, this may be rewritten as
I I  
1 f(Z) 1 f(Z) 1
f(z) = dZ = dZ.
2πi Z−z 2πi Z 1 − (z/Z)
C C

See [9.3]. Since z is inside the circle on which Z lies, |z| < |Z| = R, and |(z/Z)| < 1.
1
Thus 1−(z/Z) may be viewed as the sum of an infinite geometric series, and
I
1 f(Z)  
f(z) = 1 + (z/Z) + (z/Z)2 + (z/Z)3 + · · · dZ.
2πi Z
C

Provided it makes sense to integrate this infinite series term by term, we deduce
that f(z) can indeed be expressed as a power series:
∞
X I
n 1 f(Z)
f(z) = cn z where cn = dZ. (9.8)
2πi Zn+1
n=0 C

Furthermore, comparing this formula with (9.7), we also deduce (9.6).

To verify that this term-by-term integration is legitimate, consider the sum
PN−1
fN (z) ≡ n
n=0 cn z of the first N terms of the series (9.8). The result will be
established if we can show that fN (z) tends to f(z) as N tends to infinity.
492 Cauchy’s Formula and Its Applications

[9.3] If f(z) is analytic inside C, then it has a Taylor expansion there:

P∞ n 1
H f(Z)
f(z)= n=0 cn z ,where cn = 2πi
h i C Zn+1 dZ.
h
Cauchy’s Formula can be written as
i
1
H f(Z) 1 1
f(z) = 2πi C Z 1−(z/Z) dZ, and then 1−(z/Z) can be expanded as a geo-
metric series. But the coefficients can also be obtained by differentiation, in the
usual way, and therefore we obtain a formula for the nth derivative as an integral:
H f(Z)
f(n) (a) = n! cn = 2πi
n!
C Zn+1 dZ.

Since
1 (z/Z)N
− [1 + (z/Z) + (z/Z)2 + · · · + (z/Z)N−1 ] = ,
1 − (z/Z) 1 − (z/Z)
it follows that
I
1 (z/Z)N f(Z)
f(z) − fN (z) = dZ.
2πi (Z − z)
C

Finally recall [see (8.5), p. 440] that the modulus of an integral cannot exceed the
product of the length of the path and the maximum modulus of the integrand at
points on the path. If M stands for the maximum value of |f(Z)/(Z − z)| on C, then
it follows that

|f(z) − fN (z)| ⩽ RM|(z/Z)|N .

Thus limNÑ∞ fN (z) = f(z), as was to be shown.

What is the radius of convergence of the series we have obtained? We know that
if f is analytic inside C then the series (9.8) converges to f(z) in that disc. Thus, refer-
ring to [9.3], C may be expanded up to the dashed circle, where it first encounters
a singularity of f. More generally, f(z) may be a single-valued branch of a multi-
function, and as we learnt in Chapter 2, branch points then act as obstacles just as
much as singularities. Thus the radius of convergence is the distance from the centre of
the expansion to the nearest singularity or branch point.
 Calculus of Residues 493

One final point. We chose the origin as the centre of the expansion in order to
avoid algebraic clutter, but this choice really involves no loss of generality. For sup-
pose we instead choose the centre to be at a, meaning that we wish to expand f(z)
in powers of ξ ≡ (z−a). If f(z) is analytic at a then F(ξ) ≡ f(a+ξ) = f(z) is analytic
at the origin of the ξ-plane, and so it possesses an origin-centred Taylor expansion,
∞
X ∞
X
F(n) (0) n f(n) (a)
F(ξ) = ξ =⇒ f(z) = (z − a)n .
n! n!
n=0 n=0

Alternatively [exercise], the existence of this series may be deduced by directly

generalizing the argument leading to the origin-centred series. Either way, we
conclude that
If f(z) is analytic, and a is neither a singularity nor
a branch point, then f(z) may be expressed as the fol-
lowing power series, which converges to f(z) within the
disc whose radius is the distance from a to the nearest
singularity or branch point:
∞
X I
n f(n) (a) 1 f(z)
f(z) = cn (z − a) , where = cn = dz.
n! 2πi (z − a)n+1
n=0 L

9.3 Calculus of Residues

9.3.1 Laurent Series Centred at a Pole

Suppose that a is a pole of an analytic function f(z), i.e., limzÑa f(z) = ∞. In
Chapter 7 we investigated poles by assuming the existence of Taylor series (which
we have just proven), and we found [see (7.19), p. 417] that near a we could express
f(z) as
ϕ(z)
f(z) = ,
(z − a)m
where ϕ(z) is analytic, and ϕ(a) ̸= 0. Recall that the positive integer m is called
the “order” of the pole, and that the greater the order of the pole, the faster f(z)
approaches ∞ as z approaches a.
We know that ϕ(z) can be expressed as a Taylor series centred at a:
∞
X ϕ(n) (a)
ϕ(z) = cn (z − a)n , where cn = .
n!
n=0

Hence we deduce that

494 Cauchy’s Formula and Its Applications

If an analytic function f(z) has a pole of order m at a, then in the vicinity

of this pole, f(z) possesses a Laurent series of the form

c0 c1 cm−1
f(z) = + + ··· + + cm + cm+1 (z − a) + · · · .
(z − a)m (z − a)m−1 (z − a)
Recall that the coefficient of 1/(z − a) is called the “residue” of f(z) at a, denoted
Res [f, a]. Also recall the crucial significance of the residue in evaluating integrals:
if L is a simple loop containing a but no other singularities of f, then
I
f(z) dz = 2πi Res [f, a].
L

More generally, suppose that L is not required to be simple, and that f(z) has several
poles, at a1 , a2 , etc. The existence of the Laurent series was the missing ingredient
in our discussion of this situation in the previous chapter. Having established that
f does indeed possess a Laurent expansion in the vicinity of each of its poles, we
have also verified the General Residue Theorem [(8.26), p. 477]:
I X
f(z) dz = 2πi ν[L, an ] Res [f, an ]. (9.9)
n
L

9.3.2 A Formula for Calculating Residues

It is easy enough to find an explicit formula for the residue at a pole. Looking at
the derivation of the Laurent series above, we see that
ϕ(m−1) (a)
Res [f, a] = cm−1 = .
(m − 1)!
Since ϕ(z) = (z − a)m f(z), we deduce that
If a is an mth order pole of f(z), then
 m−1
1 d
Res [f(z), a] = [(z − a)m f(z)] . (9.10)
(m − 1)! dz
z=a

From this general result one can derive other results that speed up the calculation
of residues in commonly encountered special cases. For example, suppose that f =
(P/Q) has a “simple” (i.e., order 1) pole at a as a result of Q having a simple root
at that point. In that case,
P(z)
Res [f(z), a] = lim (z − a)f(z) = lim h i.
zÑa zÑa Q(z)−Q(a)
z−a

P(z)
Thus, If f(z) = , and a is a simple root of Q, then
Q(z)
P(a)
Res [f(z), a] = ′ . (9.11)
Q (a)
 Calculus of Residues 495

For example, consider f(z) = ez /(z4 − 1), which has simple poles at z = ±1, ±i.
If L is the circle |z − 1| = 1 then z = 1 is the only pole inside L, so (9.11) yields
I
ez ez
dz = 2πi Res [f, 1] = 2πi = 12 πie.
z4 − 1 4z3 z=1
L

We can actually check this using Cauchy’s Formula. Since

(z4 − 1) = (z − 1)(1 + z + z2 + z3 ),

we may write f(z) = F(z)/(z − 1), where F(z) ≡ ez /(1 + z + z2 + z3 ). Since F(z) is
analytic inside L,
I I
ez F(z)
dz = dz = 2πi F(1) = 12 πie,
z4 − 1 z−1
L L

just as before.

9.3.3 Application to Real Integrals

In the exercises of the previous chapter we saw how certain kinds of real integrals
could be expressed in terms of complex contour integrals. According to (9.9), the
evaluation of contour integrals amounts to calculating residues, and we have just
seen that this is straightforward. Thus the Residue Theorem leads to a powerful
method of evaluating real integrals.
Historically, Cauchy’s success in evaluating previously intractable real integrals
was one of the first tangible signs of the power of his discoveries. Many mod-
ern texts (e.g., Marsden et al. (1999)) continue to celebrate this success with very
detailed discussions of how the Residue Theorem may be applied to real integrals.
However, there can be little doubt that this application is less important than it used
to be. Today,2 when faced with a tricky integral, a physicist, engineer, or mathe-
matician is less likely to start calculating residues, and is more likely to reach for
a computer. We will therefore only do a couple of illustrative examples, though
further examples may be found in the exercises.
R+∞
In Ex. 14, p. 480 we evaluated −∞ (x2 + 1)−2 dx using partial fractions. To redo
this problem using residues, we integrate f(z) = 1/(z2 + 1)2 along the simple loop
(L+J) shown in [9.4a]. Here L is the segment of the real axis from −R to +R, and J is
the semi-circular contour (in the upper half plane) back from +R to −R. Rewriting
f(z) as f(z) = 1/(z + i)2 (z − i)2 , we see that the only singularities are the second
order poles at z = ±i. Thus if R > 1 (as illustrated) then (9.10) yields

2
Here, “today” refers to 1997! Today, in 2022, my claim is only more true than it was then.
496 Cauchy’s Formula and Its Applications

I
d 1 −2 π
f(z) dz = 2πi Res [f, i] = 2πi = 2πi = .
dz (z + i)2 z=i (2i) 3 2
L+J

But
I Z +R Z
dx
f(z) dz = + f(z) dz,
−R (x + 1)2
2
J
L+J

and, as you showed in the original exercise, the integral along J tends to zero as R
tends to infinity. Thus
Z +∞
dx π
2 2
= .
−∞ (x + 1) 2
The famous physicist Richard Feynman once bet3 his colleagues, “I can do by
other methods any integral anybody else needs contour integration to do.” It is a
tribute to complex analysis that Feynman lost this bet. Nevertheless, we can check
the above integral using a trick that frequently did enable Feynman to dispense
with residues: differentiation of a simpler integral with respect to a parameter.
Consider the elementary result,
Z +∞    +∞ π
dx 1 −1 x
2 2
= tan = .
−∞ x + a a a −∞ a
Differentiating this with respect to a yields
Z +∞
2a π
2 + a2 )2
dx = 2 ,
−∞ (x a
and substituting a = 1 then confirms our residue calculation.
For our second example, we will evaluate
Z 2π
dθ
I≡ , a > 1,
0 cos θ+a
by rewriting it as a contour integral round the unit circle C. See [9.4b]. As illustrated,
cos θ is the midpoint of z and (1/z), and dz is perpendicular to z and has length dθ:
in symbols, cos θ = 21 [z + (1/z)] and dz = iz dθ. Substituting into I,
I I
(dz/iz) dz
I= 1 = −2i 2 + 2az + 1
.
2 [z + (1/z)] + a z
C C

Since the singularities p and q of the integrand satisfy pq = 1, only one of them
lies inside C—in fact p and q are geometric inverses. Thus [exercise],
 
1 4π 2π
I = 4π Res ,q = =√ .
(z − p)(z − q) (q − p) a2 − 1

3
See Feynman (1997).
 Calculus of Residues 497

R+∞
[9.4] Using residues to evaluate real integrals. [a] To evaluate (x2 + 1)−2 dx,
H −∞
consider the integral of f(z) = (z2 + 1)−2 . Then, L+J f(z) dz = 2πi Res [f, i] = π2 .
H R+∞
But as R Ñ ∞, J f(z) dz Ñ 0, and therefore −∞ (x2 + 1)−2 dx = π2 . [b] To evaluate
R2π
I ≡ 0 cosdθ iθ 1
θ+aH, where a > 1, let z = e , so that cos θ = 2 [z + (1/z)] and dz = izdθ.
dz
Then I = −2i C z2 +2az+1 . Factorizing the denominator into (z −p)(z −q), we obtain
h i
1 4π
I = 4π Res (z−p)(z−q) , q = (q−p) = √a2π
2 −1
.

9.3.4 Calculating Residues using Taylor Series

In order to calculate a residue using (9.10), one must first know the order m of the
pole. If f(z) is built out of simple functions whose Taylor series are known, then
the quickest method of finding m is by manipulating these series. Furthermore,
this approach may be used to calculate the residue itself, often more easily than via
formula (9.10). A few examples should suffice to explain the method.
For our first example, let f(z) = (sin2 z/z5 ), which clearly has a singularity of
some kind at the origin. For small values of z, sin z ≈ z, so f(z) ≈ (1/z3 ), and the
order of the pole is therefore m = 3. By taking more terms of the Taylor series for
sin z we can find more terms in the Laurent expansion of f(z), and hence find the
residue:
 2   3 
1 z3 1 2 z
f(z) = z− + · · · = 5 z − 2z + ···
z5 6 z 6
1 1
= − + ···
z3 3z
1
=⇒ Res [f, 0] = − .
3
In order to appreciate how efficient this is, try checking the result using formula
(9.10) instead.
Our next example will have valuable consequences. Let g(z) = (1/z2 ) cot(πz),
which is clearly singular at the origin. To find the order of this pole, and its residue,
we begin by calculating the Laurent series of cot(πz). When doing such a calcula-
tion, it is important to remember that we are not trying to find the whole Laurent
498 Cauchy’s Formula and Its Applications

series. We just want the (1/z) term of g, which will come from the z term of cot πz,
so that’s as far as we need go:
h i
(πz)2
cos πz 1 − 2! + · · ·
cot πz = =h i
sin πz πz − (πz) 3
+ ··· 3!
  −1
1 (πz)2 (πz)2
= 1− + ··· 1 − + ···
πz 2 6
   
1 (πz)2 (πz)2
= 1− + ··· 1 + + ···
πz 2 6
1 πz
= − + ··· .
πz 3
In particular, note for future use that Res [cot(πz), 0] = (1/π).
Returning to the original function g, we find that
1 π
g(z) = 3
− + ··· ,
πz 3z
and so the origin is a triple pole with Res [g, 0] = −(π/3). Again, try checking this
using formula (9.10) instead.
Continuing with this example, it’s clear that g(z) also has a singularity at each
integer n. To find the residue at n, we could write z = n + ξ and expand g as a
Laurent series in powers of ξ. However, this is unnecessary. Since (1/z2 ) is non-
singular at n, and since cot[π(n + ξ)] = cot πξ,
Res [(1/z2 ) cot(πz), n] = (1/n2 ) Res [cot(πz), n]
= (1/n2 ) Res [cot(πz), 0]
= 1/(πn2 ).
More generally, note that if f(z) is any analytic function that is non-singular at n,
then
1
Res [f(z) cot(πz), n] = f(n). (9.12)
π
This may also be verified [exercise] using (9.11).

9.3.5 Application to Summation of Series

Historically, 1+ 212 + 312 + 412 +· · · was the first simple-looking series that mathemati-
cians were unable to sum using elementary algebraic methods. After the Bernoulli
family had tried and failed, Euler finally cracked the problem in 1734 by means of a
brilliantly unorthodox argument4 . The answer he found was as unexpected as his
methods:
X∞
1 π2
2
= .
n 6
n=1

4
See Ex. 13 and Stillwell (2010).
 Calculus of Residues 499

P∞ 1
[9.5] Evaluation of −∞ h using iRes [f(z)cot (πz), n] = π f(n). If we
f(n)
take f(z) = 1/z2 , then Res cot(πz)z2
, n = π 1n2 . Also, one finds [see text] that
h i H cot(πz) P
Res cot(πz)
z2
1
, 0 = − π3 . Therefore, 2πi S z2
dz = − π3 + π2 N 1
n=1 n2 . But as
N Ñ ∞, one finds [see text] that this integral goes to zero, and we thereby obtain
P
Euler’s extraordinary discovery of 1734: ∞ 1 π2
n=1 n2 = 6 .

Today such results can be derived in a systematic way using residues. Recon-
sider the function g(z) = (1/z2 ) cot(πz) above. With N a positive integer, let S be
the origin-centred square with vertices (N + 21 )(±1 ± i) shown in [9.5]. Adding up
the residues of the illustrated singularities inside S,
I X−1 X
N
1
g(z) dz = Res [g(z), 0] + Res [g(z), n] + Res [g(z), n]
2πi
S n=−N n=1

π 2X 1
N
= − + .
3 π n2
n=1

As we will now see, the integral on the LHS tends to zero as N tends to infinity,
and from this fact we immediately deduce Euler’s result.
To show that the integral of g(z) = (1/z2 ) cot(πz) does indeed tend to zero as
S expands, we must show that the size of the integrand dies away faster than the
perimeter (8N + 4) of S grows. First the easy part: |g(z)| = 1/z2 · | cot(πz)|, and on
S we clearly have |z| > N, so 1/z2 < (1/N2 ).
Next we must examine the size of
eiπz + e−iπz
| cot(πz)| = iπz
e − e−iπz
on the four edges of S. We begin with the horizontal edges, y = ±(N + 12 ). Since
e±iπz = e∓πy , it is not hard to see [exercise] that if N is reasonably large then
| cot(πz)| is very close to 1. Thus for sufficiently large N, | cot(πz)| will certainly be
less than 2, for example.
500 Cauchy’s Formula and Its Applications

Finally, on the vertical edges we have z = ±(N + 21 ) + iy, and it follows [exercise]
that
1 − e−2πy
| cot(πz)| = ⩽ 1.
1 + e−2πy
For sufficiently large N, we have established that | cot(πz)| < 2 everywhere on
S, so by virtue of (8.5), p. 440,
I
2
g(z) dz ⩽ (Max |g| on S)(perimeter of S) < (8N + 4).
N2
S

Since the RHS tends to zero as N tends to infinity, we are done.

More generally, let f(z) be an analytic function such that |f(z)| < (const.)/|z|2 for
sufficiently large |z|. Then it is clear that the above argument applies equally well
to the integral of f(z) cot(πz):
I
1
0 = lim f(z) cot(πz) dz
NÑ∞ 2πi
S
X
= Res [f(z) cot(πz)]
all poles
∞
X X
= Res [f(z) cot(πz), n] + Res [f(z) cot(πz)]
n=−∞ poles of f(z)
∞
X X
1
= f(n) + Res [f(z) cot(πz)],
π n=−∞
poles of f(z)

where the last equality follows from (9.12).

Thus
If f(z) is an analytic function such that |f(z)| < (const.)/|z|2 for suffi-
ciently large |z|, then

∞
X X
f(n) = −π Res [f(z) cot(πz)]. (9.13)
n=−∞ poles of f(z)

Of course if any of the poles of f(z) happen to be integers, then these values of n
are understood to be excluded from the LHS of (9.13).
P
Note that while symmetry enables us to calculate sums like ∞ (1/n2 ) and
P∞ 4
P∞
n=1
3
n=1 (1/n ) using (9.13), we cannot use (9.13) to calculate a sum like n=1 (1/n ).
What, you might ask, is the sum of this last series? The answer is that nobody knows!5

5
This was written in 1997, but, 25 years later, it remains a mystery.
 Annular Laurent Series 501

As a further interesting illustration of (9.13), consider f(z) = 1/(z − w)2 , where

w is an arbitrary (non-integer) complex number. Geometrically, |z − w| is the dis-
tance w to z, and this makes it easy to see that |f(z)| satisfies the requirement of the
theorem. Since the only singularity of f(z) is a double pole at z = w,
∞
X  
1 cot(πz)
= −π Res , w .
n=−∞
(n − w)2 (z − w)2

Using formula (9.10),

 
cot(πz) d π
Res 2
,w = cot(πz) =− 2
.
(z − w) dz z=w sin (πw)
Thus we obtain the remarkable result,
π2 1 1 1 1 1
2
= ··· + 2
+ 2
+ 2+ 2
+ + ··· .
sin (πw) (2 + w) (1 + w) w (1 − w) (2 − w)2
Such series were first discovered by Euler in 1748. What is remarkable about such
a formula is that the periodicity of the function on the LHS is explicitly exhibited
by the series on the RHS. That is, if you change w to (w + 1), the series is clearly
unaltered.

9.4 Annular Laurent Series

9.4.1 An Example
We have seen that the Laurent series is the natural generalization of the Taylor
series when the centre of the expansion is a pole rather than a non-singular point.
However, this is by no means the only situation in which Laurent series are needed.
For example, consider
1
F(z) = ,
(1 − z)(2 − z)

whose simple poles are illustrated in [9.6a]. Since F is analytic within the unit
disc, it possesses a Taylor series in powers of z. This may be found most easily by
splitting F into partial fractions:
1 1
F(z) = −
(1 − z) (2 − z)
for |z|<1 for |z|<2
z }| { z }| {
X∞ ∞
X
1 1
= − = zn − 12 (z/2)n
(1 − z) 2[1 − (z/2)]
n=0 n=0
  n
2 + 4 z + 8 z + · · · + 1 − (1/2) z + ··· , for |z| < 1.
1 3 7 2 n+1
=
502 Cauchy’s Formula and Its Applications

The pole at z = 1 means that outside the unit disc F cannot be expressed as a
power series in z. However, in the shaded annulus 1 < |z| < 2 it can be expressed
as a Laurent series in z:
for |z|>1 for |z|<2
z }| { z }| {
X∞ ∞
X
1 1
F(z) = − − =− (1/z)n+1 − 12 (z/2)n
z[1 − (1/z)] 2[1 − (z/2)]
n=0 n=0
2 3
1 1 1 1 z z z
= ··· − 3
− 2− − − − − − ··· , for 1 < |z| < 2.
z z z 2 4 8 16
Finally, in the region |z| > 2 beyond the annulus we obtain [exercise] a different
Laurent series:
1 3 (2n−1 − 1)
F(z) = 2
+ 3 + ··· + + ··· , for |z| > 2.
z z zn

9.4.2 Laurent’s Theorem

What we have just seen is an illustration of a general phenomenon. See [9.6b].

If f(z) is analytic everywhere within an annulus A centred at a, then

f(z) can be expressed as a Laurent series within A. In fact, if K is any
simple loop lying within A and winding once round a,

1
[9.6] Laurent Series. [a] Let F(z) = (1−z)(2−z) . Then F(z) has a Taylor series within the unit
disc, but the singularity at z = 1 means that no power series can exist beyond. However, in
the shaded annulus, 1 < |z| < 2, it is possible to express F(z) as a so-called Laurent series
2 3
that has positive and negative powers of z: F(z) = · · ·− z13 − z12 − z1 − 12 − z4 − z8 − z16 −· · · .
Moving outside the annulus, a Taylor series is again impossible because of the two poles
closer in, but, here too, a Laurent series exists. However, it is now a different Laurent series:
n−1
F(z) = z12 + z33 + · · · + (2 zn−1) + · · · . [b] Laurent’s Theorem. If f(z) is analytic everywhere
P H
within the annulus, then f(z) = ∞ n 1 f(Z)
n=−∞ cn (z − a) , where cn = 2πi K (Z−a)n+1 dZ.
 Annular Laurent Series 503

∞
X I
1 f(Z)
f(z) = cn (z − a)n , where cn = dZ. (9.14)
n=−∞
2πi (Z − a)n+1
K

Before establishing this result, which is called Laurent’s Theorem, we make the
following observations regarding its significance:

• The surprising thing about the result is the existence of a Laurent series, not the
fact that it converges in an annulus. Since we know that a power series in (z − a)
will converge inside a disc centred at a, it follows [exercise] that a power series
in 1/(z − a) will converge outside a disc centred at a. Since a Laurent series is (by
definition) the sum of a power series in (z − a) and a power series in 1/(z − a),
it follows that it will converge in an annulus.
• Previously we were able to deduce the existence of a Laurent series only in the
vicinity of a pole. The present result is much more powerful: as indicated by
the question marks in [9.6b], we make no assumptions at all concerning the
behaviour of f(z) in the disc D bounded by the inner edge of the annulus. In
practice, the outer edge of the annulus may be expanded until it hits a singu-
larity s of f(z), and the inner edge may likewise be contracted until it hits the
outermost singularity lying in D.
• If there are no singularities in D, then the inner edge of the annulus may be
completely collapsed, thereby transforming the annulus into a disc. In this case,
(9.14) does not contain any negative powers. For if n is negative then f(z)/(z −
a)n+1 is analytic everywhere inside K, and so cn = 0. In this way we recover the
existence of Taylor’s series as a special case of Laurent’s Theorem.
• Suppose that a is a singularity and that for a sufficiently small value of ϵ there
are no other singularities within a distance ϵ of a. In this case one says that a is an
isolated singularity of f(z). Applying Laurent’s Theorem to the annulus 0 < |z −
a| < ϵ, we find that there are just two fundamentally different possibilities: the
principal part of the Laurent series either has finitely many terms, or infinitely
many terms. Recall that in the latter case we have (by definition) an “essential
singularity”. See p. 417, where we considered the example
1 1 1
e1/z = 1 + + 2
+ + ··· .
1! z 2! z 3! z3
To sum up,

An isolated singularity of an analytic function is either a pole or an es-

sential singularity.

Now let us establish (9.14). In order to simplify the calculations, we will only
treat the case a = 0, illustrated in [9.7a]. Here, z is a general point in the annulus, C
and D are counterclockwise circles such that z lies between them, and L is a simple
loop round z, lying within the annulus.
504 Cauchy’s Formula and Its Applications

[9.7] Proof of Laurent’s Theorem. [a] The initial key step mimics theH rewriting of
1 f(V)
Cauchy’s Formula that was used to prove Taylor’s Theorem: f(z) = 2πi L V−z dV =
H h i H h i
1 f(Z) 1 1 f(W) 1
2πi C Z 1−(z/Z) dZ + 2πi D z 1−(W/z) dW. The theorem then follows
quickly by expanding both square brackets into geometric series, in z and (1/z), respec-
tively. [b] This picture justifies the previous equation by showing that L can indeed be
deformed within the annulus into (C) + (−D), without changing the value of the integral.

First, by Cauchy’s Formula,

I I I
1 f(V) 1 f(Z) 1 f(W)
f(z) = dV = dZ − dW,
2πi V − z 2πi Z − z 2πi W − z
L C D

where the second equality follows from the fact that L may be deformed within
the annulus into (C) + (−D), as indicated in [9.7b].
Next, we rewrite the above equation as
I   I  
1 f(Z) 1 1 f(W) 1
f(z) = dZ + dW.
2πi Z 1 − (z/Z) 2πi z 1 − (W/z)
C D

The significance of this is that |(z/Z)| < 1 and |(W/z)| < 1, so both integrands on the
RHS can be expanded into geometric series, very much as we did in the example
of [9.6a].
Referring back to the derivation of the Taylor series (9.8), the integral round C
can be expressed as
 
I   X∞ I
1 f(Z) 1  1 f(Z)
dZ = dZ zn .
2πi Z 1 − (z/Z) 2πi Zn+1
C n=0 C
 Annular Laurent Series 505

Essentially identical reasoning [exercise] also justifies term-by-term integration in

the case of the integral round D:
 
I   X∞ I  n
1 f(W) 1  1 n−1  1
dW = W f(W) dW .
2πi z 1 − (W/z) 2πi z
D n=1 D
Thus the existence of the Laurent series is established:
d3 d2 d1
f(z) = · · · + 3 + 2 + + c0 + c1 z + c2 z2 + · · · ,
z z z
where
I I
1 m−1 1 f(Z)
dm = W f(W) dW and cn = dZ.
2πi 2πi Zn+1
D C
Finally, the following two observations enable us to tidy up the result. First, by
the Deformation Theorem [p. 454], the integrals defining dm and cn do not change
their values if we allow C to contract and D to expand till they coalesce into the same
circle. Indeed, we may replace both C and D with any simple loop K contained in
the annulus and winding round it once. Second, if we write m = −n then the inte-
gral defining the coefficient d−n of zn has integrand W −n−1 f(W) = f(W)/W n+1 ,
which is the same as the integrand for the cn ’s. Thus, as was to be shown, the
Laurent series may be expressed in the compact form of (9.14):
X∞ I
n 1 f(Z)
f(z) = cn z , where cn = dZ.
n=−∞
2πi Zn+1
K
506 Cauchy’s Formula and Its Applications

9.5 Exercises

1 If C is the unit circle, show that

Z 2π I
dt i dz
2
= .
0 1 + a − 2a cos t (z − a) (az − 1)
C
Use Cauchy’s Formula to deduce that if 0 < a < 1, then
Z 2π
dt 2π
2
= .
0 1 + a − 2a cos t 1 − a2
2 Let f(z) be analytic on and inside a circle K defined by |z − a| = ρ, and let M be
the maximum of |f(z)| on K.
(i) Use (9.6) to show that
n!M
f(n) (a) ⩽ n .
ρ
(ii) Suppose that |f(z)| ⩽ M for all z, where M is some constant. By putting
n = 1 in the above inequality, rederive Liouville’s Theorem [p. 410].
(iii) Suppose |f(z)| ⩽ M|z|n for all z, where n is some positive integer. Show that
f(n+1) (z) ≡ 0, and deduce that f(z) must be a polynomial whose degree
does not exceed n.
3 (i) Show that if C is any simple loop round the origin, then
  I
n 1 (1 + z)n
= dz.
r 2πi zr+1
C
(ii) By taking C to be the unit circle, deduce that
 
2n
⩽ 4n .
n
For other interesting applications of complex analysis to problems involving
binomial coefficients, see Bak and Newman (2010).
4 The Legendre polynomials Pn (z) are defined by
1 dn  2 
Pn (z) = n n
(z − 1)n .
2 n! dz
These polynomials are important in many physical problems, including the
quantum mechanical description of the hydrogen atom.
(i) Calculate P1 (z) and P2 (z), and explain why Pn has degree n.
(ii) Use (9.6) to show that
I
1 (Z2 − 1)n
Pn (z) = dZ,
2πi 2n (Z − z)n+1
K
where K is any simple loop round z.
 Exercises 507

p
(iii) By taking K to be a circle of radius |z2 − 1| centred at z, deduce that
Z p
1 π
Pn (z) = (z + z2 − 1 cos θ)n dθ.
π 0
(iv) Check that this last formula yields the same P1 (z) and P2 (z) as you obtained
in part (i).
5 If C denotes the unit circle, show that
Z 2π I
sin2 θ i (z2 − 1)2 π
dθ = − 2
dz = .
0 5 − 4 cos θ 4 z (z − 2)(2z − 1) 4
C

6 Let f(z) be an analytic function with no poles on the real axis, and such that
|f(z)| < (const.)/|z|2 for sufficiently large |z|. By integrating f(z) eiz along the
contour (L + J) shown in [9.4a], deduce that
Z +∞ Z +∞ X
f(x) cos x dx + i f(x) sin x dx = 2πi Res [f(z) eiz ].
−∞ −∞ upper half-plane

[Hint: First show that if y > 0, then |eiz | < 1.]

7 Use the result of the previous exercise to do the following problems, in which
we assume that a > 0.
(i) Show that
Z +∞
cos x π
2 2
dx = e−a .
−∞ x + a a
(ii) Evaluate
Z +∞
x sin x
dx.
−∞ (x2 + a2 )2
n
8 Let Fn (z) = 1/(1 + z ), where n is even.
(i) Use (9.11) to show that if p is a pole of Fn then Res [Fn , p] = −(p/n).
(ii) With the help of part (i), show that the sum of the residues of Fn in the
upper half-plane is a geometric series with sum 1/[in sin(π/n)].
(iii) By applying the Residue Theorem to the contour (L + J) shown in [9.4a],
deduce that
Z∞
dx π
n
= . (9.15)
0 1+x n sin(π/n)
(iv) Although the above derivation breaks down when n is odd [why?], use a
computer to verify that (9.15) is nevertheless still true.
9 Continuing from the previous question, consider the wedge-shaped contour K
shown below.
508 Cauchy’s Formula and Its Applications

(i) Use the Residue Theorem to show that if n = 2, 3, 4, . . ., and R > 1 (as
illustrated), then
I
dz 2πi i(π/n)
n
=− e .
1+z n
K
(ii) Show that
I h i Z∞
dz i(2π/n) dx
lim = 1 − e .
RÑ∞ 1 + zn 0 1 + xn
K
(iii) Deduce that (9.15) is indeed valid for odd n as well as even n.
P
10 Use (9.13) to show that ∞ 4 4
n=1 (1/n ) = (π /90).

11 Show that if f(z) is an analytic function such that |f(z)| < (const.)/|z|2 for
sufficiently large |z|, then
∞
X X
(−1)n f(n) = −π Res [f(z) cosec(πz)].
n=−∞ poles of f(z)

In this formula, it is understood that if any of the poles of f(z) happen to be

integers, then these values of n are excluded from the LHS.
12 Use the result of the previous question to do the following:
(i) Show that
1 1 1 π2
1− + − + ··· = .
4 9 16 12
(ii) Find the sum of the series
1 1 1 1 (−1)n+1
− + − + ··· + + ··· .
2 5 10 17 (n2 + 1)
13 (i) Show that
∞
X 1 π cot πz
= .
n=−∞
z2 −n2 z
 Exercises 509

(ii) Show that the previous equation can be rewritten as

∞
1 X 2z
cot z = + .
z z2 − n2 π2
n=1

(iii) Show that the previous equation can be rewritten as

∞
X
d d
[ln(sin z/z)] = ln(z2 − n2 π2 ).
dz dz
n=1

(iv) By integrating along any path from 0 to z that avoids integers, and then
exponentiating both sides of the resulting equation, deduce that
   
z2 z2 z2
sin z = z 1 − 2 1− 2 2 1 − 2 2 ··· .
π 2π 3π
[Hint: Recall that limzÑ0 (sin z/z) = 1.]
This famous formula was discovered by Euler in 1734, in a much simpler
way, via an extraordinary leap of imagination; see Stillwell (2010) for the
details.
(vi) Multiplying out the brackets on the right, we see that the z3 terms arise
from taking a z2 term from a single bracket, and 1’s from all the others.
But, on the left hand side, the power series for sin z is known. Thus,
 
1 3 1 1 1 1
z − z + · · · = z − 2 + 2 2 + 2 2 + 2 2 + · · · z3 + · · · .
3! π 2π 3π 4π
Euler was thereby able to sum a series that had baffled the greatest math-
ematicians that had come before him, making the wonderful and startling
discovery that
1 1 1 π2
1+ + + + · · · = .
22 32 42 6
CHAPTER 10

Vector Fields: Physics and Topology

10.1 Vector Fields

10.1.1 Complex Functions as Vector Fields

Throughout the course of this book we have relied on a single means of visual-
izing a complex function, namely, as a mapping of points in one complex plane
to points in another. This idea has proved to be extremely powerful, for in terms
of it the complex derivative is nothing more complicated than a local amplitwist.
Despite its many virtues, in this chapter we shall abandon the mapping paradigm
and introduce a completely new one in its place, thereby gaining a host of fresh
insights into the subject and revealing surprising connections with physics.
The new picture of a complex function f(z) involves only a single complex plane.
As before, the variable z is thought of as a point in this plane, but now comes the
new idea: the value of f(z) is pictured as a vector emanating from z. The resulting dia-
gram of points with attached vectors is called the vector field of f. Figures [10.1a]
and [10.1b] illustrate the vector fields of z2 and (1/z), respectively; before reading
further you should study them carefully and convince yourself of their correctness.
Try doing a sketch of the vector fields of some other powers, then compare them
with accurate ones done by your computer. Also use the computer to examine the
vector fields of ez , log z, and sin z.
The vector field concept remedies a significant defect in the mapping point of
view. Although we can learn a lot about a mapping by looking at the images of
specific shapes, we do not get a feel for its overall behaviour. But if we let our eyes
roam over the vector field of a complex function we do get such a view, in much the
same way as we can survey the behaviour of a real function by scanning its graph.
Just as a complex mapping determines a vector field, so a vector field determines
a mapping—the two concepts are equivalent. More explicitly, given the vector V
issuing from the point z, we translate the tail of V to the origin and define the image
of z to be the point at the tip.

Visual Complex Analysis. 25th Anniversary Edition. Tristan Needham, Oxford University Press.
© Tristan Needham (2023). DOI: 10.1093/oso/9780192868916.003.0010
512 Vector Fields: Physics and Topology

[10.1] Vector field representation of f(z): Draw the complex number f(z) emanating
from z. [a] The vector field of f(z) = z2 . [b] The vector field of f(z) = (1/z).

[10.2] A vector field defines a mapping. [a] This vector field corresponds to the mapping
z ÞÑ (1/2)z. [b] This vector field corresponds to the mapping z ÞÑ (1/2)iz.

Consider the examples in [10.2a] and [10.2b]. If z lies on a circle of radius r then
the vector field in [10.2a] is radial and has length (r/2); in [10.2b] the vector field
has the same length but is tangential instead of radial. Check that when viewed as
mappings, [10.2a] corresponds to an expansion of the plane by (1/2), while [10.2b]
corresponds to the same expansion followed by (or preceded by) a rotation through
a right angle.
 Vector Fields 513

If the vectors in [10.2a] were instead directed inwards, what would the corre-
sponding mapping be? If the vectors in [10.2b] were flowing clockwise, what would
the corresponding mapping be?

10.1.2 Physical Vector Fields

Since a vast range of physical phenomena find their most natural description as
vector fields, the potential utility of the new way of looking at complex mappings
should be obvious.
For example, consider the astonishingly complex array of electromagnetic dis-
turbances zipping through the space around you. The visible light carrying these
words to your retina, the totality of television and radio programs simultaneously
being broadcast to your home—all this constitutes only a small part of the fren-
zied activity. But it is a remarkable fact that this great tangle of signals is in fact
completely described by just two vector fields! At each instant of time t there is
an electric vector E(p, t) and a magnetic vector B(p, t) emanating from each point
p in space, and these two vector fields constitute the complete description of the
electromagnetic field.
If we are to describe such physical vector fields with complex mappings, two
problems immediately present themselves. A television set is fixed in space, and
the way in which it produces its picture is by monitoring how the electromagnetic
vector fields at its location vary in time. But a complex mapping is a timeless thing—
it assigns the vector f(z) to the point z once and for all. This is the first of our two
problems. Thus if we are not to radically alter our conception of a complex map-
ping, the only types of physical vector fields we can describe in this manner are
those that do not vary with time. We shall call such vector fields steady.
Fortunately, steady vector fields are both common and important in physics. For
example, the unwavering character of the orbits of the planets reflects the fact that
the gravitational field of the sun does not vary with time. In fact Newton informs
us that this time-independent force on a particle of unit mass located at a point p
in space may be represented by a vector emanating from p, directed to the centre
c of the sun, and with a length equal to M/[cp]2 , where M is the mass of the sun.
Drawing these vectors throughout space we obtain a steady vector field.
The above examples of the electromagnetic and gravitational vector fields illus-
trate our second problem—they exist in three-dimensional space, whereas the
complex plane can only accommodate a two-dimensional vector field. There is no
getting around this problem, but once again it is fortunate that there are certain
important types of physical phenomena which are intrinsically two-dimensional
in nature, and which can therefore be described in the complex plane. Let us begin
with the flow of electricity within a sheet of conducting material.
514 Vector Fields: Physics and Topology

[10.3] The Phase Portrait of V(z) is the set of streamlines along which V flows: mathe-
matically, they are the curves that have V as their tangent vectors. Given such a picture,
one can immediately deduce the direction of V(z) at z, but its magnitude is invisible.
However, in many physical examples it is possible to draw the phase portrait in a spe-
cial way, such that the closer together the streamlines, the stronger the flow. This is the
case here, where electrodes are attached to a metal plate at A and B, and electricity flows
between them—you see that the spacing is smallest, and therefore the flow greatest, along
AB. Note that the streamlines are arcs of circles; this is proved geometrically in the next
figure.

Take two wires and connect them to a battery, then touch the ends to two points
A and B of a thin copper plate. Almost instantly a steady flow of electric current
from one electrode to the other will be set up in the plate. See [10.3]. At each point
z of the plate we now represent this flowing current by a time-independent vector
in the direction of the flow, and with a length equal to the strength of the current
there. Picturing the plate as a portion of C, the flow is thus expressed as a complex
function V(z).
Rather than drawing the actual vector field in [10.3] we have instead shown the
paths along which the electricity flows. Such a picture is called the phase portrait
of the vector field, and the directed curves along which the flow occurs are called
the integral curves or streamlines of the vector field. As illustrated, the streamlines
of this example are in fact arcs of circles connecting the two electrodes. We shall
justify this shortly.
Phase portraits are easy to take in visually and are thus a common way of rep-
resenting vector fields. By definition the vector field is everywhere tangent to the
streamlines, and thus its direction can be recovered from the phase portrait. On the
other hand, it would seem that a phase portrait would necessarily fail to include
the information about the lengths of the vectors. This is true in general, but for
 Vector Fields 515

many vector fields that arise in physics it will be shown that there exists a special
way of drawing the phase portrait so that the strength of the flow is manifested
as the crowding together of the streamlines: the closer together the streamlines, the
stronger the flow1 . Later we shall explain this idea in detail, but for the moment we
remark that [10.3] has actually been drawn in this special way. For example, as we
approach the line-segment connecting the electrodes the streamlines become more
and more crowded together, corresponding to a stronger and stronger current.

10.1.3 Flows and Force Fields

One and the same vector field (or phase portrait) can represent many quite differ-
ent physical phenomena. For example, reconsider the copper plate in [10.3], and
imagine that it is now sandwiched between two layers of material which do not
conduct heat. Remove the electrodes and instead of supplying electricity at a con-
stant rate at A, let us supply heat. Likewise, let us remove heat at the same constant
rate at B. After a short time a steady pattern of heat flow from A to B will be estab-
lished within the copper plate. In this steady state we may assign to each point a
vector in the direction that the heat is flowing there, and having a length equal to
the intensity of the heat flow.
Remarkably, in this steady state the physical laws governing the behaviour of
the heat are identical to those which previously described the electricity, and thus
the phase portrait [10.3] for the electric current is also the phase portrait for the new
heat flow.
Here is yet another interpretation of [10.3]. In attempting to understand the
flow of real liquids, such as water, it is helpful to consider an idealized fluid with
the properties of being frictionless, incompressible, and “irrotational”—the precise
meaning of the last term will be explained later. Imagine a thin layer of such an
ideal fluid sandwiched between two horizontal plates, one of which has two small
holes, A and B. If we now connect the holes with a fine tube that passes through a
pump, a steady flow will be set up in the layer of fluid, and at each point we may
draw its velocity vector. The phase portrait of this steady vector field is once again
given by [10.3]!
Although there are certainly important differences between these three interpre-
tations of [10.3], we may nevertheless lump them together in one class, for they are
all flows of something. Whether it is electricity, heat, or liquid, in each case the vec-
tor field can be thought of as the velocity of flowing “stuff”, and the streamlines
are the paths along which this stuff flows.
A physically quite distinct class is comprised of force fields. For example,
although we previously discussed how the gravitational field of the sun could be
represented as a steady vector field, the vector at a point in space is no longer the

1
Faraday was the first to conceive of vector fields in this way; Maxwell then rendered the idea
mathematically precise and exploited it to the hilt.
516 Vector Fields: Physics and Topology

velocity of some flowing substance, rather, it represents the force experienced by a

unit mass placed there. In the context of force fields, integral curves are called lines
of force rather than streamlines. Here the lines of force are rays coming out of, or
rather entering, the centre of the sun. Although this force field is three-dimensional,
spherical symmetry2 means that it will be the same on any plane drawn through the
centre of the sun. It can therefore be completely described by a complex function.
Although there is nothing actually flowing along the lines of force, we can switch
back to the flow point of view by pretending that there is, thereby interpreting a
force field as the velocity field of a flowing substance. This is not mere sophistry:
it is a remarkable fact that for the most common and important force fields (e.g.,
gravitational and electrostatic) this imaginary flowing substance behaves exactly like
our previously considered ideal fluid.
To illustrate this, we turn to an example in electrostatics: equal and opposite
charges (per unit length) are induced on two long wires which are then held parallel
to each other in empty space. To each point in space we now attach the force vector
that a unit electric charge would experience there; this force field is (by definition)
the electric field E, and its phase portrait is the same on each plane drawn per-
pendicular to the wires. Taking [10.3] to be such a plane, with the wires piercing
through at A and B, the phase portrait of this force field is exactly the one shown
there for the flow of ideal fluid.

10.1.4 Sources and Sinks

In order to make a quantitative analysis of [10.3], we introduce the concepts of
(two-dimensional) sources and sinks. Thinking in terms of our layer of ideal fluid,
a source of strength S is a point at which we pump in S units of fluid per unit of time.
Figure [10.4a] illustrates the symmetric velocity vector field V(z) of an isolated
source at the origin.
Given a curve (open or closed) in a general flow, the amount of fluid flowing
across it in each unit of time is called the flux. Clearly the flux across an element
of the curve is just its length times the component of the velocity perpendicu-
lar to the curve. The total flux across the curve is then the sum (i.e., integral) of
these elementary fluxes. Returning to the specific case of [10.4a], our assumption
of incompressibility says that the flux across any simple loop round 0 must be the
same as the amount of fluid S being pumped in at 0. Since the flow is orthogonal
to the origin-centred circle C of radius r, we deduce that

2πr |V| = S.

Writing z = r eiθ , we find that the vector field of the source is therefore

2
This is an idealization—like the earth, the sun is somewhat flattened at the poles.
 Vector Fields 517

[10.4] [a] A Source. Fluid is pumped in at a constant rate at 0 and travels outward sym-
metrically. Its velocity at z = r eiθ must be proportional to (1/r), for the total amount of
fluid crossing a circle of radius r per unit of time (the flux) will be this velocity multiplied
by 2πr. [b] The streamlines of a doublet are arcs of circles that pass through the source
and the sink. To prove that D = V⊕ + V⊖ is tangent to the illustrated arc of the circle ApB,
we must prove that • = ⊙. The angles ApB and pst are clearly equal, but we also have
(ts/ps) = |V⊕ |/|V⊖ | = Bp/Ap. So, the two shaded triangles are similar, and therefore
• = ⊙. Done.

   
S eiθ S 1
V(z) = |V| e iθ
= = .
2π r 2π z
[We note without proof that this is also the electric field on a plane at right angles
to a very long wire carrying a uniform charge of S per unit length.] The source in
[10.3] is at A instead of at the origin, and so it is described by
 
S 1
V⊕ (z) = .
2π z − A

A sink may be thought of as a source with a negative strength: it is a place where

fluid is pumped out rather than in. In each of the flow experiments which [10.3]
purports to describe, the sink at B has the same strength as the source at A, and so
its vector field is
 
S 1
V⊖ (z) = − .
2π z − B
We now know the vector fields V⊕ (z) and V⊖ (z) which would be produced by
the source or the sink in [10.3] if each were present on its own, but what is the flow
when they are both present together? [Incidentally, this combination of a source
518 Vector Fields: Physics and Topology

and a sink of equal strength is called a doublet.] The answer is perhaps slightly
clearer if we switch to the equivalent electrostatic problem of parallel charged wires
through A and B. A unit charge at z is repelled by A with force V⊕ (z), and attracted
by B with force V⊖ (z). The net force D(z) of the doublet acting on the charge is then
simply the vector sum of the two separate forces:
 
S 1 1 S (A − B)
D(z) = V⊕ (z) + V⊖ (z) = − = . (10.1)
2π z − A z − B 2π (z − A)(z − B)
We will now show geometrically that, as claimed in [10.3], the net force at p is
tangent to the circle through A, p, and B. Consider [10.4b]. It is easy to see [exercise]
that D will be tangent to the circle if and only if the angles marked • and ⊙ are
equal, so this is what we must demonstrate. As illustrated, the angles ApB and pst
are clearly equal. But we also have
ts |V⊕ | Bp
= = .
ps |V⊖ | Ap
Thus the two shaded triangles are similar, and therefore • = ⊙, as was to be shown.

10.2 Winding Numbers and Vector Fields*

10.2.1 The Index of a Singular Point

Let us confine all our discussions to vector fields for which the direction is well-
defined and continuous at all but a finite number of points. These exceptional
places, where the vector field vanishes or becomes infinite, are called singular
points3 . They are easy to spot in a phase portrait, usually as the intersection points of
distinct streamlines. Figure [10.5] shows the phase portraits in the vicinity of some
simple types of singular points, together with their names and their “indices”—a
term which we must now explain.
Figure [10.6] is a magnified view of the simple crosspoint (also called a saddle
point) shown in [10.5]. Round this singular point s we have drawn a simple loop
Γs , and at some of its points we have also drawn the vectors V. Since Γs does not
pass through any singular points, the direction of V is well-defined and continuous
everywhere on it. Thus we can count the net number of revolutions of V(z) as z
traverses Γs . We call this number the index IV [Γs ] of the loop Γs with respect to
the vector field V. When it is clear which vector field is being considered, we may
simplify this notation to I [Γs ]. For example, in [10.6] we see that I [Γs ] = −1. Note
that we have drawn the vectors on Γs only to make it easier to see the value of
the index; actually, since only the directions of the vectors are required, the phase
portrait is sufficient on its own.

3
Otherwise known as critical points or singularities—terms to which we have already attached
different meanings.
 Winding Numbers and Vector Fields* 519

[10.5] The values of the index (I) for various singular points. The geometric definition
of I is explained in the next figure.

If we continuously deform Γs without crossing s (or any other singular point)

then the value of I [Γs ] will also vary continuously, and since it’s an integer, it
will therefore remain constant. Thus we may unambiguously define the index of
a singular point s to be the index of any loop that winds round s once, but does
not wind round any other singular points. It should not cause any confusion if
we abuse our notation slightly and call this index I(s). Applying this definition
to loops of your choosing, you may now verify each of the given values of I
in [10.5].
Before moving on, we observe three properties of the index:

(i) There is nothing to stop us applying the above definition to a non-singular

point, but in this case the index must vanish. Choosing Γs to be a very small
loop, the non-singular nature of s implies that all the vectors on Γs will point
in roughly the same direction, and so I(s) = 0.
520 Vector Fields: Physics and Topology

[10.6] The index of a singular point is the net number of revolutions executed by the
vector field as we traverse a simple (counterclockwise) loop around the singular point.
Here, I = −1.

(ii) If V undergoes a certain rotation as we traverse a piece of curve, then (−V)

undergoes the same rotation. Thus if we reverse the direction of the flow in
each of the above phase portraits, the index will remain the same. For example,
a source must have the same index as a sink, namely, I = 1.
(iii) Just as the index is insensitive to the precise shape of Γs , so it is insensitive to
the precise shape of the streamlines. Imagine that [10.6] is drawn on a rubber
sheet which we gradually stretch, so producing a new distorted phase portrait.
The direction of V at each point of Γs will undergo a continuous change, and
so its net revolutions upon traversing Γs will likewise vary continuously. The
index must therefore remain constant.
Clearly, our new concept of “index” is related to our old concept of “winding num-
ber”, but how? If we instead think of V as a mapping, sending the points of Γs to
those of a new loop V(Γs ), then a moment of thought reveals that the index of Γs is
just a new interpretation of the winding number of its image loop:
IV [Γs ] = ν [V(Γs ), 0]. (10.2)
This makes it clear [see p. 397] that the index I(s) of a point s is the same thing as
its topological multiplicity ν(s) as a preimage of 0. In particular, if V is analytic then
I (root of order n) = n and I (pole of order m) = −m.
Check this for the examples in [10.1] .
If you have not done so yet, we urge you to use a computer to draw the vec-
tor fields of some simple polynomials and rational functions. Notice how roots
and poles show up just as vividly as the corresponding x-intercepts and vertical
 Winding Numbers and Vector Fields* 521

asymptotes occurring in the graph of a real function. Notice how easy it is to zoom
in on the vector field to find their precise locations. But now we can do much better
than the real case, for we can also immediately read off the precise nature of a root
or singularity simply looking at its index!
The following example illustrates the fact that a vector field contains more
information than an ordinary graph. If we sketch the graphs of
(x − 1)2 (x − 1)4
F(x) = and G(x) =
(x + 2)3 (x + 2)7
the results will be qualitatively the same: both look something like a parabola near
x = 1; both have branches going to opposite ends of the vertical asymptote at
x = −2; both look something like (1/x) for large x.
Now use the computer to draw the corresponding vector fields when x is
replaced by z. Striking indeed are the differences! As we traverse a small loop
around the root at z = 1, F makes two positive revolutions while G makes four;
doing the same at the pole z = −2, F makes three negative revolutions while
G makes seven; and on a very large origin-centred circle, F makes one negative
revolution while G makes three.
Returning to the general significance of (10.2), consider the ordinary winding
number ν [L, 0] of a loop L. This can now be viewed as the index of L with respect
to the vector field of the identity mapping:
ν [L, 0] = Iz [L].
Figure [10.7] illustrates this result with Iz [L] = 1. The winding number of L around
a general point a is likewise just its index with respect to the vector field (z − a):
ν [L, a] = I(z−a) [L].

[10.7] Index of the vector field V(z) = z is the winding number of the loop: ν [L, 0] =
Iz [L].
522 Vector Fields: Physics and Topology

10.2.2 The Index According to Poincaré

Figure [10.8a] shows a loop L and a vector field V evaluated on it. Let us use this
simple example (for which it is obvious that IV [L] = 1) to explain a quick method
(due to Poincaré) of finding the index in more complicated cases.
Consider all the places on L (a, b, c in our case) where V points in one arbitrarily
chosen direction. Let P be the number of these places at which V(z) rotates in the
positive sense as z passes through it, and let N be the number at which it rotates in
the negative sense. Even in relatively complicated cases, P and N are usually quick
and easy to find. We now obtain the index as the difference of these two numbers:
IV [L] = P − N. (10.3)
In our case P = 2 because of the positive rotation at a and c, and N = 1 because of
the negative rotation at b. Thus IV [L] = 1, as it should.
To see just how fast and efficient (10.3) is to use, try it out on each of the examples
in [10.5], choosing L to be a counterclockwise circle centred at s, and choosing the
direction to be vertically upward. For extra practice, try it again with a different
choice of direction.

[10.8] Poincaré’s method for quickly finding the index. [a] First, choose an arbitrary
direction, then find all the places zj on L where V(zj ) points in this direction. Count the
number P of zj ’s where V(z) is rotating in the positive, counterclockwise direction as we
pass through zj , and likewise count the number N of places where it is rotating in the
negative direction. Then, IV [L] = P−N. In the illustrated case, P = 2, N = 1, and so the
formula correctly predicts that IV [L] = 2 − 1 = 1. [b] Consider the image V(L) of L under
the mapping corresponding to the vector field V. On the ray in the chosen direction, start
far away at q, where the winding number ν = 0. The intersection points of the ray with
V(L) are the images V(zj ) of the zj ’s, and a positive rotation of V at zj corresponds to a
left-to-right crossing, so the crossing rule (7.1) increases ν by one. Likewise, a negative
rotation of V reduces ν by one. Thus we have proved Poincaré’s formula.
 Winding Numbers and Vector Fields* 523

Although the truth of (10.3) is probably clear at an intuitive level, it is neverthe-

less instructive to deduce it from the “crossing rule” (7.1), p. 388, for computing
winding numbers.
Figure [10.8b] shows the image V(L) of L when V is viewed as a mapping.
[Check that it really is the image!] In these terms, the required index is just
ν [V(L), 0]. Draw the ray from 0 in the previously chosen direction, and let the
point q travel along it (starting far away), ending up at the origin. In its journey,
q will thus cross V(L) at the points V(c), V(b), and V(a). In the vector field pic-
ture [10.8a], the positive rotation of V at c now implies (in [10.8b]) that q sees
V(L) directed from left to right as it approaches the first crossing at V(c). Oppo-
sitely, the negative rotation at b implies that V(L) is directed from right to left as
q approaches V(b). But as we previously argued in Chapter 7, ν [V(L), 0] is the
number of points P at which V(L) is directed from left to right (as seen by q as it
approaches), minus the number of points N at which it is directed from right to left.
Done.

10.2.3 The Index Theorem

With the connection between indices and winding numbers established, the Topo-
logical Argument Principle can be reinterpreted in terms of vector fields:
The index of a simple loop is the sum of the indices of the singular points it contains.
Using a neater argument than the one given in Chapter 7, we can now extend this
theorem to multiply connected regions. As illustrated in [10.9], recall that this means
that the region has holes in it; two in our case.
The shaded region consists of the points which are inside C and outside B1 and
B2 . In general there could be more holes, say g of them, with counterclockwise
boundary curves B1 , B2 , …, Bg . As illustrated, suppose that we have a vector field
on such a region, and let s1 , s2 , . . ., sn be the singular points within the region. In our
case there are only two: s1 is a dipole, and s2 is a saddle point. The generalization
of the Argument Principle is this:
X
g X
n
I [C] − I [Bj ] = I [sj ].
It is called the Index Theorem.
Perhaps using (10.3), for practice, verify that in our example, I [C] = 2, I [B1 ] = 0,
and I [B2 ] = 1, so that the LHS of the Index Theorem equals 1. But the RHS is
I (dipole) + I (saddle point) = 2 + (−1) = 1,
so confirming the prediction of the theorem in this case.
To understand this result, consider [10.10]. Using the dashed curves, break the
region into curvilinear polygons in such a way that each one contains at most one
524 Vector Fields: Physics and Topology

P Pn
[10.9] The Index Theorem: I [C] − g I [Bj ] = I [sj ]. This is certainly true in the
illustrated example, for I [C] − I [B1 ] − I [B2 ] = 2 − 0 − 1 = 1, and I [s1 ] + I [s2 ] =
2 + (−1) = 1.

[10.10] Proof of the Index Theorem. Using the dashed curves, break the region into
curvilinear polygons in such a way that each one contains at most one singular point,
and let their counterclockwise boundaries be Kj . If we sum the indices of all the Kj ’s
then we obtain the RHS of the Index Theorem. On the other hand, the index of a single
Kj is obtained by looking at how much the vector field rotates as one travels along each
edge of Kj , then adding up these net rotation angles. But when we sum the indices of all
the Kj ’s, each interior edge [dashed] is traversed twice, once in each direction, and the
associated angles of rotation therefore cancel. The remaining edges of the Kj ’s together
make up C and −B1 , −B2 , etc. Summing the associated angles of rotation (divided by
2π) therefore yields the LHS of the Index Theorem. Done.
 Flows on Closed Surfaces* 525

singular point, and let their counterclockwise boundaries be Kj . If we sum the

indices of all the Kj ’s then we obtain the RHS of the Index Theorem. For if Kj does
not contain a singular point then its index vanishes, while if it does contain one
then its index is (by definition) the index of that singular point.
On the other hand, the index of a single Kj is obtained by looking at how much
the vector field rotates as one travels along each edge of Kj , then adding up these
net rotation angles. But when we sum the indices of all the Kj ’s, each interior edge
[dashed] is traversed twice, once in each direction, and the associated angles of
rotation therefore cancel. The remaining edges of the Kj ’s together make up C and
−B1 , −B2 , etc. Summing the associated angles of rotation (divided by 2π) therefore
yields the LHS of the Index Theorem. Done.

10.3 Flows on Closed Surfaces*

10.3.1 Formulation of the Poincaré–Hopf Theorem

If a curved surface S in space is “smooth” in the sense that there exists a tangent
plane at each of its points, then it makes sense to speak of a vector field that is every-
where tangent to S. Intuitively, we may picture such a vector field as the velocity
of a fluid that is flowing over S.
Figure [10.11] shows the streamlines of two such flows on the sphere. Notice that
both possess singular points: [10.11a] has two vortices, while [10.11b] has a dipole.
In fact there can be no vector field on the sphere that is free of singular points. This is

[10.11] Poincaré–Hopf Theorem: If a vector field on a smooth surface of genus g has

only a finite number of singular points, then the sum of their indices is χ = (2 − 2g).
Two examples: [a] I (vortex) + I (vortex) = 1 + 1 = 2 = χ(sphere). [b] I (dipole) = 2 =
χ(sphere).
526 Vector Fields: Physics and Topology

one consequence of an extremely beautiful result called the Poincaré–Hopf Theorem,

the formulation of which we will now sketch.
It is not immediately obvious how to give a precise definition of the “index”
of a singular point on a curved surface, but for the moment let us accept that this
integer exists, and that its value is the same as for an analogous singular point in
the plane. Thus if we sum all the indices in [10.11a] we obtain

I (vortex) + I (vortex) = 1 + 1 = 2,

while if we do the same for [10.11b] we obtain

I (dipole) = 2.

Try drawing your own streamlines on an orange, then sum the indices of the
singular points. Is this a coincidence??
There are no coincidences in mathematics! In the case of the sphere, the Poincaré–
Hopf Theorem states that if we sum the indices of any vector field on its sur-
face, we will always get 2 for the answer. Indeed, it says that we will get this
answer for any surface that is topologically a sphere, that is to say, any surface
into which the sphere may be changed by a continuous and invertible trans-
formation. If we imagine the sphere to be made of rubber, examples of such
transformations and surfaces are given by stretching without tearing. The surfaces
of the plum and the wineglass in [10.12a] are two examples of such topological
spheres.
The sphere is the boundary of a solid ball, and other closed surfaces may likewise
be obtained as the boundaries of other solid objects. For example, the surface of a
doughnut is called a torus (top of [10.12b]), and it is clear that this surface is topo-
logically the same as the beach toy at the bottom. But it seems equally clear that no
amount of stretching and bending can turn these surfaces into a sphere—[10.12a]
and [10.12b] are topologically distinct types of surface. Figure [10.12c] shows yet a
third topologically distinct class. Obviously we could continue this list indefinitely
just by adding more holes.
We shall not develop the topological ideas4 necessary to prove it, but once again
it seems clear that these classes of topologically distinct closed surfaces can be
classified purely on the basis of their number of holes. This number g is called
the genus of the surface (see [10.12]). We can now formulate the general result:

If a vector field on a smooth surface of genus g has only a finite number

of singular points, then the sum of their indices is χ ≡ (2 − 2g). (10.4)

4
See “Further Reading”, at the end of this chapter.
 Flows on Closed Surfaces* 527

[10.12] The genus (g) and the Euler characteristic (χ) of a closed surface. The genus
counts the number of holes in a closed surface. The Euler characteristic χ ≡ (2 − 2g)
turns out to be ubiquitous in topology, and it is therefore a more natural way of classifying
surfaces. Furthermore, χ in fact has its own, independent geometrical definition, and it is
then a theorem that χ = (2 − 2g). See Ex. 12.

The number χ ≡ (2 − 2g) occurring in this theorem is called the Euler characteris-
tic of the surface, and it turns out to be ubiquitous in topology. It is therefore more
natural to classify our surfaces using χ rather than g. See [10.12]. Furthermore, χ in
fact has its own, independent geometrical definition, and it is then a theorem that
χ = (2 − 2g). See Ex. 12.
An immediate consequence of (10.4) is that a vector field without any singular
points can exist only on surfaces of vanishing Euler characteristic, i.e., the topo-
logical doughnuts. Even then, the theorem does not actually guarantee that such a
vector field exists, it merely says that if there are singular points then their indices
must cancel. However, you can see for yourself [draw it!] that on a doughnut there
do exist at least two topologically distinct vector fields without any singular points:
one that circulates around its axis of symmetry, and one that circulates through the
hole.

10.3.2 Defining the Index on a Surface

In order to give a precise definition of the “index” of one of the singular points in
[10.11], we should presumably draw a loop round it on the surface, then find the
528 Vector Fields: Physics and Topology

net rotation of the vector field as the loop is traversed. But wait, rotation relative to
what?
To answer this question, we first re-examine the familiar concept of rotation in
the plane. Figure [10.13a] shows that (in the plane) the rotation of V(z) along L
can be thought of as taking place relative to a fiducial vector field having horizontal
streamlines, say U(z) = 1. If we define ∠UV to be the angle between U and V, and
let δL (∠UV) be the net change in this angle along L, then our old definition of the
index is

[10.13] Definition of the index of a singular point on a curved surface. [a] The original
definition of the index in the plane can be thought of as the net rotation of V relative to a
horizontal fiducial vector field U: IV [L] = 2π1
δL (∠UV). [b] The definition and value of
the index does not change if we replace U with any vector field that is nonsingular on and
inside L. [c] Now imagine that [b] is drawn on a rubber sheet. If we continuously stretch it
1
into the form of the curved surface then not only will 2π δL (∠UV) remain well-defined,
but its value will not change.
 Flows on Closed Surfaces* 529

1
IV [L] = δL (∠UV). (10.5)
2π
If we continuously deform the horizontal streamlines of U in [10.13a] to produce
those in [10.13b] then, by the usual reasoning, the RHS of (10.5) will not change.
Thus we conclude that this formula yields the correct value of the index if we
replace U with any vector field that is nonsingular on and inside L.
Now imagine that [10.13b] is drawn on a rubber sheet. If we continuously stretch
it into the form of the curved surface in [10.13c] then not only will the RHS of (10.5)
remain well-defined, but its value will not change. To summarize: if s is a singular
point of a vector field V on a surface S, we define its index as follows. Draw any
nonsingular vector field U on a patch of S that covers s but no other singular points;
on this patch, draw a simple loop L going round s; finally, apply (10.5), that is count
the net revolutions of V relative to U as we traverse L.

10.3.3 An Explanation of the Poincaré–Hopf Theorem

We can now give a very elegant derivation of theorem (10.4), due to Hopf (1956)
himself. The argument proceeds in two steps. First, we show that on a surface of
given genus, all vector fields yield the same value for the sum of their indices;
second, we produce a concrete example of a vector field for which the sum equals
the Euler characteristic. This proves the result.
Suppose that V and W are two different vector fields on a given closed surface
S. See [10.14]. If vj are the singular points of V (marked •) and wj are those of W
(marked ⊙), we must show that
X X
IV [vj ] = IW [wj ].
Much as we did in [10.10], we divide up S into curvilinear polygons (dashed)
such that each one contains at most one vj and one wj . taken counterclockwise
as viewed from outside S. To find the indices of V and W along boundary of one
of the polygons Kj , draw any nonsingular vector field U on the polygon and then
use (10.5). The difference of these indices is then
1  
IW [Kj ] − IV [Kj ] = δKj (∠UW) − δKj (∠UV)
2π
1
= δK (∠VW),
2π j
which is explicitly independent of the local vector field U.
From this we deduce that
X X X
IV [vj ] − IW [wj ] = (IV [Kj ] − IW [Kj ])
all polygons

1 X
= δKj (∠VW)
2π
all polygons

= 0,
530 Vector Fields: Physics and Topology

[10.14] Proof that the sum of the indices is the same for all vector fields. Let V (with
singular points vj marked •) and W (with singular points wj marked ⊙) be two dif-
ferent vector fields on the surface S. Divide up S into curvilinear polygons (dashed)
such that each one contains at most one vj and one wj . Now consider one of these
polygons and its boundary Kj . Draw any nonsingular vector
 field U on the polygon.
Then IW [Kj ] − IV [Kj ] = 2π1
δKj (∠UW) − δKj (∠UV) = 2π 1
δKj (∠VW), which is
explicitly independent of the local vector field U. If we now sum this over all polygons,
P
each edge is traced twice, in opposite directions, so δKj (∠VW) = 0, and therefore
P P
IV [Kj ] = IW [Kj ]. Done.

because every edge of every polygon is traversed once in each direction, producing
equal and opposite changes in ∠VW. We have thus completed the first step: the
sum of the indices is independent of the vector field.
Since the index sum for the example in [10.11a] is 2, we now know that this
is the value for any vector field on a topological sphere. The second step of the
general argument is likewise to produce an example on a surface of arbitrary genus
g, such that the sum is χ = (2 − 2g). Figure [10.15] is such an example for g = 3, the
generalization to higher genus being obvious. Here we imagine that honey is being
poured onto the surface at the top—it then flows over the surface, finally streaming
off at the very bottom. As the figure explains, and as was required, the sum of the
indices is indeed equal to χ.
Further Reading. These topological ideas—in combination with ideas in the
next two chapters—open the door to the important subject of Riemann surfaces. In
particular, we hope you will find it easier to read Klein (1881), which champions
 Flows on Closed Surfaces* 531

[10.15] A specific vector field on a surface of genus g for which the sum of the indices is
χ. We imagine that honey is being poured onto the surface at the top—it then flows over
the surface, finally streaming off at the very bottom. As the figure explains, and as was
required, the sum of the indices is indeed χ. But we know that this sum is the same for
all vector fields on the surface, and we have therefore shown that the sum of their indices
must always equal χ, completing the proof of the Poincaré–Hopf Theorem.

Riemann’s original approach to multifunctions in terms of fluid flowing over a sur-

face in space. See also Springer (1981, Ch. 1), which essentially reproduces Klein’s
1881 monograph, but with additional helpful commentary. For a good introduction
to the more abstract, modern view of Riemann surfaces, see Jones and Singerman
(1987). Finally, for more on topology itself, we recommend Hopf (1956), Prasolov
(1995), Stillwell (2010), Fulton (1995), and Earl (2019) .
532 Vector Fields: Physics and Topology

10.4 Exercises

1 Show both algebraically and geometrically that the streamlines of the vector
field z2 are circles that are tangent to the real axis at the origin. Explain why
the same must be true of the vector field 1/z2 .
2 Use a computer to draw the vector field of 1/(z sin2 z). Use this picture to
determine the location and order of each pole.
3 Use a computer to draw the vector field of

P(z) = z3 + (−1 + 5i) z2 + (−9 − 2i) z + (1 − 7i).

Use this picture to factorize P(z), and check your answer by multiplying out
the brackets.
4 Suppose that one of the streamlines of a vector field V is a simple closed loop
L. Explain why L must contain a singular point of V.
5 Find the index of each of the three singular points shown below.

6 Observe that the neighbourhood of every singular point we have examined in

this chapter is made up of sectors of one of the three types shown below, called
elliptic, parabolic, and hyperbolic. Let e, p, and h denote the number of each type
of sector surrounding a singular point.
 Exercises 533

(i) Verify that the index of each of the three singular points in the previous
question is correctly (and painlessly) predicted by Bendixson’s Formula:

I = 1 + 12 (e − h).

(ii) Explain this formula.

7 Given a vector field V, defined on a circle C, let a vector field W be constructed
on C in the manner illustrated below. If IV [C] = n, find IW [C]. [This problem
is taken from Prasolov (1995, Ch. 6), and the answer may be found there, too.]

8 If f and g are continuous, one-to-one mappings of the sphere S to itself, then

their composition f ◦ g is too. Let us prove that at least one of these three mappings
must possess a fixed point. We proceed by the method of contradiction.
(i) Show if the result is false then, for each point p of S, the points p, f(p), and
[f ◦ g](p) must be distinct.
(ii) In this case, deduce that there is a unique, directed circle Cp on S passing
through these three points in the stated order.
(iii) Imagine a particle orbiting on Cp at unit speed, and let V(p) be its velocity
vector as it passes through p. Since p was arbitrary, V is a vector field
on S.
(iv) By appealing to the Poincaré–Hopf Theorem, obtain the desired contra-
diction.
9 Continuing from the previous exercise, apply the result as follows:
(i) By taking g = f, deduce that (f ◦ f) has a fixed point.
(ii) By taking g = (antipodal mapping), deduce that either f has a fixed point,
or f maps some point to its antipodal point.
10 Arbitrarily choose a collection of points s1 , s2 , . . . , sn on a closed, smooth sur-
face S. By attempting to draw examples on the surface of a suitable fruit or
vegetable, investigate the following claim: There exists a flow on S whose only
singular points are s1 , s2 , . . . , sn , and the type of singular behaviour (dipole,
vortex, etc.) at all but one of these points may be chosen arbitrarily.
534 Vector Fields: Physics and Topology

11 Imagine the surface of the unit sphere divided up into F polygons, the edges
all being “straight lines on the sphere”, i.e., great circles. Let E and V be the
total number of edges and vertices that result from dividing up the sphere in
this way.
(i) Let Pn be an n-gon on the unit sphere. Use (6.9), p. 317, to show that
A(Pn ) = [angle sum of Pn ] − (n − 2)π.
[Hint: Join the vertices of Pn to a point in its interior, thereby dividing it
into n triangles.]
(ii) By summing over all polygons, deduce that
F − E + V = 2.
[This argument is due to Legendre (1794); the result itself is a special case of
the result in the following exercise.]
12 Let S be a smooth closed surface of genus g, so that its Euler characteristic is
χ(S) = 2 − 2g. As in [10.14], let us divide S into F polygons, and let E and V be
the total number of edges and vertices, respectively.
(i) Draw a simple example on the surface of an orange and convince yourself
(by drawing it) that we may obtain a consistent flow over the entire surface
whose only singular points are (1) a source inside each of the F polygons;
(2) a simple saddle point on each of the E edges; (3) a sink at each of the
V vertices. We call this a Stiefel vector field, in honour of Hopf’s student,
Eduard Stiefel [1909–1978]. See Frankel (2012, §16.2b). For an example of
a Stiefel vector field, see the front cover of VDGF, Needham (2021).
(ii) By applying the Poincaré–Hopf Theorem to such a flow on the general
surface S, deduce the following remarkable result, which we call the Euler-
Lhuilier Formula:
F − E + V = χ(S).
NOTES: (1) This is actually the normal definition of χ, and it is then a theorem
that χ(S) = 2−2g. (2) Euler discovered the original version of this formula
in 1750 for topological spheres, with χ = 2; this is called Euler’s Polyhedral
Formula. For a masterful, mathematically accurate, yet riveting account of
this history and the connected mathematical ideas, see Richeson (2008).
(iii) Verify this result for your example in (i), then try it out on a doughnut.
13 The figure below shows all the normals that may drawn from the point p to
the smooth surface S. Let R(q) be the distance from p to a point q of S, and let
us say that q is a critical point of R if the rate of change of R vanishes as q begins
to move within S; we need not specify the direction in which q begins to move
because we are assuming that S has a tangent plane at q.
 Exercises 535

(i) Explain why pq is normal to S if and only if q is a critical point of R.

(ii) The level curves of R on S are the intersections of S with the “onion” of
concentric spheres centred at p. Sketch these level curves in the vicinity
of the illustrated critical points of R. Notice the distinction between points
where R has a local maximum or minimum, versus points where R may
increase or decrease depending on the direction (within S) in which one
moves away from the critical point.
(iii) Imagine that p generates an attractive force field, so that every point par-
ticle in space experiences a force F directed towards p. For example, we
could imagine that p is the centre of the Earth, and that F is the Earth’s
gravitational field. If a particle is constrained to move on S, then the only
part of F to which it can respond is the projection FS of F onto S. Sketch the
streamlines of FS . How are they related to the level curves of R in (ii)?
(iv) You have just seen that the critical points of R are the singular points of FS .
How does the index I(q) of a singular point of FS distinguish between the
types of critical point discussed in (ii)?
(v) Let us define the multiplicity of a normal pq to be this index I(q). Use the
Poincaré–Hopf Theorem to deduce that
The total number of normals (counted with their multiplicities) that may
be drawn to S from any point p is independent of both the location of p
and the precise shape of S, and is equal to χ (S).
536 Vector Fields: Physics and Topology

[This lovely result is essentially due to Reech (1858), though he did not
express it in terms of χ(S), nor did he use an argument like the one above.
With hindsight, Reech’s work is a clear harbinger of Morse Theory, which
it predates by some 70 years.]
(vi) Verify Reech’s theorem for a couple of positions of p in the case where S is
a torus (doughnut).
CHAPTER 11

Vector Fields and Complex Integration

11.1 Flux and Work

We promised long ago that there was a more vivid way of understanding complex
integrals than the geometric Riemann sum of Chapter 8. In this section we lay the
foundations for this elegant new approach. If you are already familiar with vector
calculus then you can skip this section and go directly to Section 11.2.

11.1.1 Flux
In order to define the flux a little more carefully than before, consider [11.1]. At each
point of the directed path K we introduce a unit tangent vector T in the direction
of the path, and a unit normal vector N pointing to our right as we travel along K.
In terms of the corresponding complex numbers, this convention amounts to
T = iN.
The figure also shows how a vector field X (which we will first think of as the
velocity of a fluid flowing over the plane) can be decomposed into tangential and
normal components:
·
X = (X T ) T + (X N) N. ·
Only the second of these components carries fluid across K, and the amount
flowing across an infinitesimal segment ds of the path in unit time (i.e. its flux) is
·
thus (X N) ds. This is a refinement over our previous definition in that the flux
now has a sign: it is positive or negative according as the flow is from left to right
or from right to left. The total flux F [X, K] of X across K is then the integral of the
fluxes across its elements:
Z
F [X, K] = (X N) ds.
K
·
Check for yourself that the flux satisfies
F [−X, K] = F [X, −K] = −F [X, K].

Visual Complex Analysis. 25th Anniversary Edition. Tristan Needham, Oxford University Press.
© Tristan Needham (2023). DOI: 10.1093/oso/9780192868916.003.0011
538 Vector Fields and Complex Integration

[11.1] Decomposition of a vector into its tangential and normal components. Note that
the normal is defined to go to the right of the direction of travel, so that T = iN.

[11.2] If we picture X as the velocity of a fluid, then the flux, F [X, K], of X across K
represents the net flow of fluid out of R per unit time. [a] Mathematically, the flux is the
integral of the outward normal component of X. [b] The flux can be visualized as the limit
of the sum of the illustrated signed areas: positive if the flow is outward, and negative if it
is inward.

The flux concept is further illustrated in [11.2] for the case where K is a simple
closed loop bounding the shaded region R. Figure [11.2a] shows the normal com-
ponents of X, the signed magnitudes of which we must integrate to obtain F [X, K].
Figure [11.2b] shows how we might make an estimate of this flux. We replace K
by a polygonal approximation with directed edges ∆j , and at the midpoint of each
one we draw the normal component of X. The flux is then approximately given by
the algebraic sum of the signed areas of the shaded rectangles. In this case there is
clearly more positive area than negative, so the flux is positive. As the ∆j ’s become
shorter and more numerous, the approximation of course gets better and better.
 Flux and Work 539

In the case of the simple loop K in [11.2a] there is another interesting way of
looking at the flux:

F [X, K] = [fluid leaving R per unit time] − [fluid entering R per unit time].

Henceforth we will always take our fluid to be incompressible. Thus, provided there
are no sources or sinks in R, what flows into R must also flow out of R:

F [X, K] = 0.

Indeed, we may turn this around and define a flow to be sourceless in a region if
all simple loops in that region have vanishing flux. The simplest example of such
a flow without any (finite) sources or sinks is X = const. If the loop does contain
a source, for example, then incompressibility says that the flux equals the strength
of the source.
Although we will only concern ourselves with two-dimensional flows, we
should at least mention the concept of flux in three dimensions. If a fluid is flowing
through ordinary space, it no longer makes sense to speak of the flux across a curve,
but it does make sense to speak of the rate at which the fluid crosses a surface. If
N now stands for the normal to this surface, then the flux across an infinitesimal
·
element of area dA is once again given by (X N) dA. The total flux is then obtained
by integrating this quantity over the whole surface. Just as in two dimensions, the
incompressibility of a three-dimensional flow is equivalent to the statement that all
closed surfaces (that do not contain sources or sinks) have vanishing flux.
Lastly, we should point out that although the word “flux” is Latin for “flow”,
it is standard practice to retain this terminology when applying our mathematical
definition to any vector field X, regardless of whether it actually is the velocity of
a flowing substance. For example, the electric field represents a force, but one of
the four fundamental laws of electromagnetism says that we can think of it as an
incompressible flow in which positive and negative electric charges act as sources
and sinks, so that its flux through a closed surface in space equals the net charge
enclosed.

11.1.2 Work
So far we have only studied the normal component of X; we turn next to its tan-
gential component. To do so, let us now imagine that X is a force field rather than a
flow.
If a particle on which a force acts is displaced infinitesimally then we know from
elementary physics that the work done by the field (i.e. the energy it expends) is
the component of the force in the direction of displacement, times the distance
moved. Thus if the particle moves an amount ds along K then the work done by X is
·
(X T ) ds. As with flux, this definition contains a sign, the physical significance of
540 Vector Fields and Complex Integration

which we will explain shortly. If the particle is moved along the entire length of K,
the total work done by the field is then
Z
W [X, K] = (X T ) ds.
K
·
Figure [11.3a] illustrates the tangential components of X on K, the signed magni-
tudes of which we must add up to obtain W.
Just as for F, check that W satisfies

W [−X, K] = W [X, −K] = −W [X, K].

Note that, unlike F, no modification of W is needed if we wish to extend the idea

to three-dimensional force fields: it still makes perfectly good sense to consider
the work done by the field as a particle is moved along a curve in space, and the
formula is as before.
Figure [11.3b] illustrates a simple thought-experiment for interpreting both the
magnitude and the sign of W. We imagine that the plane in which the force field
acts is made of ice on which a very small ice-puck of mass m can slide without fric-
tion. We now construct a narrow frictionless channel in the shape of K, just wide
enough to accommodate the puck which we fire into it with speed vin . On the ini-
tial leg of the journey we see that the force opposes the motion, and thus if vin is
not sufficiently great, the puck will slow, stop, and return whence it came. Clearly,
though, if we fire the puck with sufficient speed it will overcome all resistance and
emerge at the end of the channel, say with speed vout . Let the initial and final kinetic
energies of the puck be Ein and Eout , so that

[11.3] [a] If we picture X as a force field, then the integral of the tangential component
is the work, W [X, K], done in moving a particle along K. [b] If a puck of mass m is
fired with speed vin into a frictionless channel following K, and it emerges with speed vout ,
then the conservation of energy implies that the work is the change in the kinetic energy:
W = 21 mv2out − 12 mv2in .
 Flux and Work 541

Ein = 12 mv2in and Eout = 12 mv2out .

One of the most sacred principles of physics is the “conservation of energy”,

which states that energy can never be created or destroyed, only transformed from
one kind into another. Thus the energy W expended by the force field on the puck
does not disappear but instead is transformed into the change in the puck’s kinetic
energy:

W [X, K] = Eout − Ein

 
m vout + m vin
= (vout − vin )
2
= [average momentum] (change in speed).

This formula also gives clear meaning to the sign of the work: it is the sign of the
change in speed. Thus if W is positive the field expends energy speeding up the
puck and increasing its kinetic energy, while if W is negative then the puck has to
give up some of its kinetic energy in doing work against the field.
Next, imagine that we bend K round so that the ends almost join to form
a closed loop. When the puck travels along the corresponding channel it will
therefore emerge at essentially the same place that it entered. Suppose it were to
emerge with greater speed than it entered. Joining the ends of the channel together,
the puck would therefore go round the loop faster and faster, gaining energy
with each circuit—energy that could be harnessed to solve the world’s energy
crisis!
Although we may construct mathematical examples for which this happens, if
no energy is supplied from outside the puck/field system then a physical force
field will not behave in this way; it will conserve energy so that the puck returns to
its starting point with exactly the same speed with which it was launched1 . Such a
field is called conservative. Mathematically, X is conservative if and only if

W [X, any closed loop] = 0. (11.1)

Just as we applied the concept of flux to vector fields that were not flows, so we
may apply the concept of work to vector fields that do not represent force. How-
ever, in this general setting it is standard practice to call W [X, K] the circulation of
X along K rather than the work. As with “flux”, this terminology originates from
thinking of X as representing a flow. To see why, take K to be a closed loop and
consider the following thought-experiment of Feynman (1963). Imagine that the
fluid flowing over the plane with velocity X is instantaneously frozen everywhere

1
If energy can be supplied from outside the system, then the work need not vanish for a closed loop.
In fact the operation of all electrical machines depends on the ability of a moving magnet to create an
electric field that can speed up our puck. However, there is still no violation of energy-conservation
since work is being done to move the magnet. See Feynman (1963).
542 Vector Fields and Complex Integration

except within the narrow strip where our channel used to be. The “circulation” is
then [exercise] the speed with which the unfrozen fluid flows (or circulates) round
K, times the length of K.
If this circulation vanishes for every closed loop then the flow is said to be irro-
tational. Just as “circulation” means W [X, K], irrespective of the physical nature of
X, so with equal generality “irrotational” is short for the mathematical statement
(11.1). Thus a conservative force field could also be described as irrotational.

11.1.3 Local Flux and Local Work

At present our definition of a sourceless and irrotational vector field X is that
F [X, any closed loop] = 0 and W [X, any closed loop] = 0. (11.2)
Our next objective is to show that there are two very simple local properties of X
that are equivalent to the non-local ones above.
To do this we must calculate limiting behaviour of the flux and the work for a
small loop that shrinks to nothing, i.e., for an “infinitesimal loop”. Though it is not
entirely obvious, later we will show that this limiting behaviour is independent of
the shape of the infinitesimal loop. We are thus free to simplify the calculations by
choosing the loop to be a small square centred at the point of interest, say z, and
having horizontal and vertical edges of length ϵ. See [11.4].
Accurate estimates of F and W can now be found by evaluating X at the mid-
points (a, b, c, d) of the sides, then summing the appropriate components. In the
limit that ϵ shrinks to nothing, this approximation becomes exact, as does the
following equation, which we will need in a moment:
P(a) − P(c) ≍ ϵ Bx P(z),

[11.4] Local flux and local work. Accurate, ultimately exact estimates of F and W around
a small, ultimately vanishing square can be found by evaluating X at the midpoints (a,
b, c, d) of the sides, then summing the appropriate components. Writing X = P + iQ, we
find that F [X, □ ] ≍ (Bx P + By Q)(area of □ ), and W [X, □ ] ≍ (Bx Q − By P)(area of □ ).
 Flux and Work 543

where Bx P(z) means Bx P evaluated at z.

For the flux we find

F [X, □ ] ≍ ϵ P(a) + ϵ Q(b) − ϵ P(c) − ϵ Q(d)

≍ ϵ [{P(a) − P(c)} + {Q(b) − Q(d)}]
≍ ϵ2 [Bx P(z) + By Q(z)].

This expression can be simplified by considering the formal dot product of the
gradient operator ∇ with the vector field:
! !
∇ X= · Bx
By
P
Q
· = Bx P + By Q.

·
This quantity ∇ X is called the divergence of X, and in terms of it we have

·
F [X, □ ] ≍ [∇ X(z)] (area of □ ). (11.3)

In the next section we will see that (11.3) is true if □ is replaced by an infinitesimal
loop of arbitrary shape. This important result explains the term “divergence”, for
·
it says that ∇ X is the local flux per unit area flowing away from z, i.e., diverging
from z. In future we will abbreviate “local flux per unit area” to “flux density”.
Repeating the above analysis for the work, we find [exercise]

W [X, □ ] ≍ [∇×X(z)] (area of □ ), (11.4)

where the formal cross product is defined by

! !
Bx P
∇×X = × = Bx Q − By P.
By Q

The quantity ∇×X is called the curl of X. Geometrically, it measures the extent
to which X ‘curls around’ the point z. Physically, in terms of force fields, the above
result says that the curl is the local work per unit area, or work density. There
is also a vivid interpretation in terms of flows. If we drop a small disc of paper
onto the surface of the flowing liquid at z, in general it will not only start to move
(translate) along the streamline through z with speed |X(z)|, but it will also rotate
about its centre with some angular speed ω(z). It can be shown that the aspect of
X which determines the rate of rotation ω is none other than the curl:
1
ω(z) = 2 [∇×X(z)].

For this reason “curl” is sometimes denoted “rot”, which is short for “rotation”.

11.1.4 Divergence and Curl in Geometric Form*

The above expression for the divergence was obtained by considering the flux
out of a shape having no connection with the flow. Greater insight is gained by
544 Vector Fields and Complex Integration

[11.5] Divergence and curl in geometric form. Let P be the illustrated orthogonal
trajectory of the streamlines, let p measure distance along it, and let s denote dis-
tance along the streamline S. Also, let κP and κS denote the curvatures of P and S,
·
respectively. Then ∇ X = Bs |X| + κP |X| and ∇×X = −Bp |X| + κS |X|.

considering the flux out of an infinitesimal “rectangle” R, two sides of which are
segments of streamlines of X, while the other two sides are segments of orthogonal
trajectories through the streamlines. See [11.5].
Here z is the point down to which R will ultimately be collapsed in order to find
the divergence there, S and P are the streamline and orthogonal trajectory through
z, and s and p are arc-length along S and P, the direction of increasing p being
chosen to make a positive right angle with X.
The net flux out of R is the difference between the fluxes entering and leaving.
The flux entering is |X| dp, while the flux leaving is the same expression evalu-
ated on the opposite side of R. It is now clear that two factors contribute to more
fluid leaving than entering: (1) greater fluid speed |X| as the fluid exits; (2) greater
separation dp of the streamlines as the fluid exits.
The second factor is clearly governed by how much the direction of X changes
along dp, in other words, by the curvature κP of P at z. More precisely, if δ denotes the
increase in a quantity as we move ds along the streamline, then [exercise] δ(dp) =
κP ds dp. Thus
(Net flux out of R) = δ{|X| dp}
= (δ|X|) dp + |X| δ(dp)
= (Bs |X| + κP |X|) (area of R).
The flux density is therefore
·
∇ X = Bs |X| + κP |X|. (11.5)
 Flux and Work 545

In fact [exercise] this formula is still true for a three-dimensional vector field,
provided that there exists2 a surface P orthogonal to the streamlines, and κP is taken
to be the sum of its principal curvatures.
Turning to the circulation round R, identical reasoning yields [exercise] an
equally neat formula for the curl:
∇×X = −Bp |X| + κS |X|, (11.6)
where κS is the curvature at z of the streamline S.
Although we suspect that (11.5) and (11.6) must have been known to the likes of
Maxwell, Kelvin, or Stokes, we have not found any reference to these formulae in
modern literature.

11.1.5 Divergence-Free and Curl-Free Vector Fields

From the definition (11.2) and the results (11.3) and (11.4) it follows that if X is
sourceless and irrotational throughout some region R, then at each point of R we
have

·
∇ X=0 and ∇×X = 0.
The vector field is then said to be divergence-free and curl-free in R.
For example, consider the vector field of a point source with strength S:
!
S S x/(x2 + y2 )
X(z) = ⇐⇒ X = .
2πz 2π y/(x2 + y2 )
This should have zero flux density (i.e. divergence) everywhere except at the ori-
gin, where it should be undefined. Check that this is so. Recall that we previously
claimed that this was also the electrostatic field of a long, uniformly charged wire.
We can now see that this makes physical sense in that the field is locally conser-
vative. Thus if we fire our puck (which must now carry electric charge in order to
experience the force) round an infinitesimal loop, it will return to its starting point
with its kinetic energy unchanged. To verify this statement you need only check
that the field is curl-free.
We have seen that a sourceless and irrotational field is divergence-free and curl-
free. To end this section we wish to establish the converse result: if the divergence
and curl vanish throughout a region, the flux and work vanish for all simple loops
in that region. We will then have,
A vector field is sourceless and irrotational in a simply connected region if and only
if it is divergence-free and curl-free there.
To understand this converse, consider [11.6] which essentially reproduces part of
[8.27], p. 468. Let us now recycle the line of reasoning associated with that figure.

2
The condition for existence is that the curl either vanish or be orthogonal to the vector field.
546 Vector Fields and Complex Integration

[11.6] Gauss’s Theorem and Stokes’s Theorem. Using the same logic as in the proof of
Cauchy’s Theorem, summing integrals of the normal or the tangential component of X
around the boundaries of the shaded squares yields the integral along the outer boundary
K. But then the flux and work around each square can be expressed in terms of the
divergence and curl of X, via (11.3) and (11.4). This immediately yields Gauss’s Theorem
(11.7) and Stokes’s Theorem (11.8).

We begin by noting that as the grid gets finer and finer, the flux or work for K
becomes the flux or work for C. Next we relate these quantities to the divergence
and curl inside K. Check for yourself that exactly the same mathematical reasoning
which previously yielded
I X I
f(z) dz = f(z) dz ,
K shaded squares □

now yields
X
F [ X, K] = F [ X, □ ]
shaded squares

and
X
W [ X, K] = W [ X, □ ].
shaded squares

However, in the present context these results become accessible to physical intu-
ition. The first says that the total amount of fluid flowing out of K is the sum of
fluxes out of the interior squares. What does the second one say?
Now let the squares of the grid shrink so as to completely fill the interior R of
C. Using (11.3) and (11.4) and replacing the sum over squares by a double integral
over infinitesimal areas dA, we obtain Gauss’s Theorem,
 Complex Integration in Terms of Vector Fields 547

ZZ
F [ X, C] =
R
·
[∇ X] dA, (11.7)

and Stokes’s Theorem,

ZZ
W [ X, C] = [∇×X] dA. (11.8)
R

From these we see that if the divergence and curl vanish everywhere in R then the
flux and work for C also vanish, as was required.
Again following the logic in Chapter 8, consider what happens to the flux and
work as we continuously deform a closed contour, or an open contour with fixed
end points. You should be able to see that (11.7) and (11.8) imply two deformation
theorems:
If the contour sweeps only through points at which the divergence
(11.9)
vanishes, the flux does not change.

If the contour sweeps only through points at which the curl vanishes,
(11.10)
the work does not change.

11.2 Complex Integration in Terms of Vector Fields

11.2.1 The Pólya Vector Field

Consider
Z
H(z) dz
K

from the vector field point of view. See [11.7]. In forming a Riemann sum with
terms H dz we now have the minor advantage that H = |H| eiβ and dz = eiα ds
are not drawn in separate planes, as they were in Chapter 8. However, we still face
the problem that H dz = |H| ei(α+β) ds involves the addition of angles, which is not
easy to visualize. Just as it is more natural to subtract vectors [yielding connecting
vectors] than to add them, so it is also more natural to subtract angles, for this yields
the angle contained between two directions.
The simple and elegant solution to our problem is to consider a new vector field:
instead of drawing H(z) at z we draw its conjugate H(z) = |H| e−iβ . We shall call
this the Pólya vector field of H. Before showing how this solves our problem, let us
offer (i) a caution and (ii) a reassurance:

(i) The Pólya vector field of H is not obtained by reflecting the picture of the ordi-
nary vector field for H in the real axis, for this would attach H(z) to z instead of
z. This will become very clear if you (or your computer) draw the Pólya vec-
tor fields of z and z2 , for example. Comparison with [10.1], p. 512, reveals that
the resulting phase portraits (not the vector fields themselves) are identical to
548 Vector Fields and Complex Integration

those of (1/z) and (1/z2 ). This is because zn points in the same direction as
(1/zn ).
(ii) As we will see in a moment, much is gained by representing H by its Pólya
vector field, but we also wish to stress that nothing is lost: the new field con-
tains exactly the same information as the old one. For example, it is clear that
the index of a loop L merely changes sign when we switch to the Pólya vector
field:
IH [L] = −IH [L].
Thus an nth order root of an analytic H still shows up clearly in its Pólya vector
field as a singular point, but now with index −n instead of n. Likewise, a pole
of order m produces a singular point of index m instead of −m.
Returning to integration, the great advantage of the Pólya vector field is that the
angle θ that it makes with the contour (see [11.7]) is given by θ = α − (−β), and
this is precisely the angle we were trying to visualize—the angle of the term H dz
in the Riemann sum. Better still, we find that
H dz = |H| eiθ ds
 
= |H| cos θ + i |H| sin θ ds
 
= · ·
H T + i H N ds.
Thus the real and imaginary parts of each term in the Riemann sum are the work
and flux of the Pólya vector field for the corresponding element of the contour. We
have thus discovered a vivid interpretation (due to Pólya3 and first championed

R
[11.7] How the Pólya
 vector fieldH makes it possible to visualize K H(z) dz. We can
· ·
see that H dz = H T +i H N ds, so this immediately yields the extremely impor-
tant
R and useful physical and geometrical interpretation of the complex contour integral:
K H(z) dz = W [H, K] + i F [H, K].

3
See Pólya and Latta (1974).
 Complex Integration in Terms of Vector Fields 549

by Bart Braden4 .) of the complex integral of H in terms of the work and flux of its
Pólya vector field along the contour:
Z
H(z) dz = W [H, K] + i F [H, K]. (11.11)
K

This interpretation is rendered particularly useful by the fact that a computer

can instantly draw the Pólya vector field of any function you wish to integrate. You
can then quickly get a feel for the value of the integral by looking at how much the
field flows along and across the contour. For example, the integral of (z2 z) along
the line-segment from 1−i to 1+i is clearly a positive multiple of i. Why? For more
on the nitty-gritty of estimating integrals with (11.11), see Braden (1987).
Our interest in (11.11) will be less in this practical aspect, and more in its theoreti-
cal import: ideas about flows and force fields can shed light on complex integration,
and vice versa. In what follows we shall give examples in both directions.

11.2.2 Cauchy’s Theorem

Given a picture of the vector field of a complex mapping H(z) = u+iv, how can we
tell whether or not H is analytic? To my knowledge there is no satisfactory answer
to this question as posed. However, there is an answer if we instead look at the
Pólya vector field, and it is an answer that exhibits a beautiful connection between
physics and complex analysis:
The Pólya vector field of H is divergence-free and curl-free if and only if H is analytic.
(11.12)
The verification is a simple calculation:
! !
∇ H=· Bx
By
u
−v
· = Bx u − By v,

and
! !
Bx u
∇×H = × = −(Bx v + By u).
By −v

Thus the divergence and curl of H will both vanish if and only if the Cauchy–
Riemann equations are satisfied. Note for future use that these two equations are
really two aspects of a single complex equation,
i Bx H − By H = ∇×H + i ∇ H, · (11.13)
the vanishing of the LHS being the compact form of the CR equations, (5.1) on
page 246.

4
See Braden (1985), Braden (1987)—which won the MAA’s Carl B. Allendoerfer Award in 1988—and
Braden (1991).
550 Vector Fields and Complex Integration

With this connection established, we now have a second, physical explanation

of Cauchy’s Theorem which is scarcely less intuitive than the geometric one in
Chapter 8. For if H is analytic everywhere inside a simple loop K bounding a region
R, its Pólya vector field in R will have (as a flow) zero flux density and (as a force
field) zero work density. This means that there is no net flux of fluid out of R, and
that a puck fired round K returns with its kinetic energy unchanged. From (11.11)
we see that the integral of H round K must vanish.
A more mathematical version of this physical explanation was given at the end
of the last section in terms of the theorems of Gauss and Stokes. Restating that argu-
ment in the present context, for a simple loop K bounding a region R, substitution
of (11.7) and (11.8) into (11.11) yields
I ZZ ZZ
H(z) dz =
R
[∇×H] dA + i
R
·
[∇ H] dA, (11.14)
K

which vanishes if H is curl-free and divergence-free in R.

11.2.3 Example: Area as Flux

As a fun and instructive example, let us reconsider the result
I
z dz = 2iA (11.15)
K

in the light of the physically intuitive theorems of Gauss and Stokes.

Observe that the Pólya vector field of H(z) = z is H(z) = z, which flows radially
outwards from the origin, like a source. However, unlike a source, here the speed of
the flow increases with distance, making it clear that this flow cannot be divergence-
free. Indeed, calculating its flux density, we find that
! !
·
∇ H=
Bx
By
· x
y
= 2.

In other words, in each unit of time, 2 units of fluid are being pumped into each
unit of area. The flux of fluid out of K is therefore 2A. On the other hand the flow
is curl-free:
! !
Bx x
∇×H = × = 0,
By y

so there is no circulation round K. Inserting these facts into (11.14) we obtain (11.15).
Figure [11.8] is a concrete example of this new way of looking at (11.15), the
shape of K having been chosen so as to make the values of the circulation and flux
obvious.
Clearly H(z) = z has no circulation along either of the arcs, and it has equal
and opposite circulations along the line-segments. The total circulation round K
 Complex Integration in Terms of Vector Fields 551

[11.8] Area as one half the flux of H = z. Clearly, H(z) = z has no circulation along
either of the arcs, and it has equal and opposite circulations along the line-segments.
The total circulation round K therefore vanishes. Equally clearly, there is no flux across
the line-segments, but there is across the arcs. The larger arc has length aϕ and the
speed of the fluid crossing it is a, so the flux across it is a2 ϕ; similarly, for the smaller
arc it is b2 ϕ. Thus, F [z, K] = (fluid out) − (fluid in) = 2 (shaded area). Therefore,
H
K z dz = W [z, K] + i F [z, K] = 2iA.

therefore vanishes. Equally clearly, there is no flux across the line-segments, but
there is across the arcs. The larger arc has length aϕ and the speed of the fluid
crossing it is a, so the flux across it is a2 ϕ; similarly, for the smaller arc it is b2 ϕ.
Thus,
 
F [z, K] = (fluid out) − (fluid in) = 2 12 a2 ϕ − 12 b2 ϕ = 2 (shaded area).
Before moving on, let us clear up a paradoxical feature of the vector field z: fluid
is being pumped in uniformly throughout the plane, and yet the flow appears to
radiate from one special place, namely, the origin. The resolution (see [11.9]) lies in
the trivial identity z = z0 + (z − z0 ), which says that the flow from the origin is the
superposition of the sourceless, irrotational field z0 and a copy of the original flow,
but now centred on the arbitrary point z0 instead of the origin.

11.2.4 Example: Winding Number as Flux

Next, let us see how the Pólya vector field also breathes fresh meaning into the
fundamentally important formula
I
1
dz = 2πi ν [L, 0]. (11.16)
z
L

According to (11.11),
I
1
dz = W [(1/z), L] + i F [(1/z), L].
z
L
552 Vector Fields and Complex Integration

[11.9] The identity z = z0 + (z − z0 ) says that the flow from the origin is the superposition
of the sourceless, irrotational field z0 and a copy of the original flow, but now centred on
the arbitrary point z0 instead of the origin.

[11.10] Winding Number interpreted as the flux of a source of strength (1/2π). As we

discussed in the new Preface, the Pólya vector field of (1/z), namely (1/z), represents
the illustrated source of strength 2π at the origin. Choosing K to be an origin-centred
circle K of radius r, traced counterclockwise,
H we obtain an immediate visual and physical
explanation of the iconic fact that K z1 dz = 2πi. For clearly there is no flow along K, and
since P flows orthogonally across K (from left to right) with speed (1/r), its flux across
K is (2πr)(1/r) = 2π. Furthermore, this physical interpretation explains why the value of
the integral will remain 2πi if K is continuously deformed into a general loop encircling
the source at the origin, so long as K does not cross that point as it is deformed. And if
the loop encircles the source ν times, then the intercepted flux is clearly 2πν. Thus if the
strength of the source is reduced to (1/2π), then ν simply is the flux.

But the Pólya vector field (1/z) is an old friend—it is a source of strength 2π located
at the origin.
Figure [11.10] illustrates the intuitive nature of the result from the new point of
view. If a loop does not enclose the source, just as much fluid flows out as in; if a
 Complex Integration in Terms of Vector Fields 553

simple loop does enclose the source, it intercepts the full 2π of fluid being pumped
in at the origin; more generally, a loop will accrue 2π of flux each time it encircles
the source.
To finish the explanation of (11.16) we must show that a source is pure flux, i.e.
every loop has vanishing work or circulation. Since a source is curl-free except at
the origin, Stokes’ Theorem guarantees vanishing work for simple loops that do
not contain 0. If the loop does contain 0 then it’s not so obvious. However, it is
obvious for an origin-centred circle. You can now finish the argument for yourself
by appealing to the Deformation Theorem (11.10).
In connection with another matter, consider the shaded sector in [11.10]. The
same amount of fluid will cross each segment of a contour which passes through it,
but the sign of the flux will depend on the direction of the contour. Try meditating
on the connection between this fact and the crossing rule for winding numbers
[(7.1), p. 388].

11.2.5 Local Behaviour of Vector Fields*

·
We previously showed that ∇ H and ∇×H represent the flux density and work
density of H for infinitesimal squares. However, in order for the formulae (11.7)
and (11.8) to really make sense it is necessary that these interpretations persist for
infinitesimal loops of arbitrary shape. Let us now place (11.7) and (11.8) on firmer
ground by verifying the shape-independent significance of the divergence and curl.
To do so we will first analyse the local behaviour of a general Pólya vector field H in
the neighbourhood of the origin. The generalization to points other than the origin
will be obvious.
As the point z = x + iy moves towards the origin, H(z) will ultimately equal the
following formula, in which the partial derivatives are evaluated at 0:
H(z) − H(0) ≍ x Bx H + y By H = 1
2 (z + z) Bx H − 2i (z − z) By H

1
= 2
[Bx H − i By H] z + 12 [Bx H + i By H] z.
This will become exact in the limit that |z| shrinks to nothing.
Turning to the Pólya vector field itself, and substituting (11.13), we find

· z
H(z) ≍ H(0) + (∇ H) + (∇×H) + C z,
2
iz
2
(11.17)

where C = 12 [Bx H − i By H]. Note that if H is analytic, in which case H is sourceless

and irrotational, then (11.17) correctly reduces to the first two terms of Taylor’s
series: H(z) = H(0) + H ′ (0)z + · · · .
The meaning of the decomposition (11.17) is illustrated in [11.11]. Unless
H(0) = 0, the constant first term dominates: vectors near the origin differ lit-
tle from the vector at the origin. The remaining three terms correct this crude
approximation. The second term describes a vector field (cf. figure [11.8]) that is
554 Vector Fields and Complex Integration

[11.11] Local behaviour of a general vector field. In the limit that z Ñ 0, the general Pólya
·
vector field is given by this ultimate equality: H(z) ≍ H(0)+ 12 (∇ H)z+ 12 (∇×H)iz+C z.
Unless H(0) = 0, the constant first term dominates: vectors near the origin differ little
from the vector at the origin. The remaining three terms correct this crude approximation.
The second term describes a vector field (cf. [11.8]) that is irrotational and has constant
divergence, equal to that of H at the origin. The third term describes a vector field that is
sourceless and has constant curl, equal to that of H at the origin. The final term is both
irrotational and sourceless.

irrotational and has constant divergence, equal to that of H at the origin. The third
term describes a vector field that is sourceless and has constant curl, equal to that
of H at the origin. The final term is both irrotational and sourceless.
Note that this decomposition is geometrically meaningful because the appear-
ance of each of the component vector fields is qualitatively unaffected by the
value of its coefficient5 . We hope these observations make the formula (11.17) both
plausible and meaningful.
Now let us return to the original problem. Let K be a small simple loop of arbi-
trary shape round the origin, and let A be the area it encloses. We wish to show that
the divergence and curl of H at 0 are the limiting values of the flux per unit area
and work per unit area as K shrinks to the origin. Using (11.17) in (11.11) we find

5
This is obvious for the source and vortex terms, but not for the last term; see Ex. 10.
 Complex Integration in Terms of Vector Fields 555

W [H, K] + i F [H, K]
I
= H(z) dz
K
I I I
 
= H(0)
K
1
·
dz + 2 ∇ H − i ∇×H z dz + C z dz.
K K

This becomes exact as K shrinks to nothing. But even if K is not small, we know
that the exact values of these three integrals are
I I I
dz = 0, z dz = 2iA, z dz = 0.
K K K

Thus
 
W [H, K] + i F [H, K] = ∇×H + i ∇ H A. ·
Equating real and imaginary parts, we obtain the desired results.

11.2.6 Cauchy’s Formula

The Pólya vector field also allows us to cast the mathematical explanation of
Cauchy’s Formula into a form that is more accessible to physical intuition.
Consider the function
f(z)
H(z) = ,
(z − p)
where f(z) is analytic. Since H is analytic except at p, its Pólya vector field H will
have vanishing flux and circulation densities except at p. Thus if C is a simple loop
round p, all of its flux and circulation must have originated at p. To find W [H, C]
and F [H, C] we should therefore examine H in the immediate vicinity of p.
If f(p) = A + iB, then very close to p the Pólya vector field H will be indistin-
guishable from
   
A − iB 1 i
=A −B .
z−p z−p z−p

Figure [11.12] illustrates this field for positive A and B, as well as showing the
geometric significance of the algebraic decomposition above.
The first term is familiar as a source at p of strength 2πA, a negative value for
A corresponding to a sink. The second term is a multiple of the less familiar field
i/(z − p) which represents a vortex6 at p. It is easy to see that the circulation round
one of its circular streamlines is 2π, so this will also be its value for any simple
loop round p—we say that the vortex has strength 2π. On the other hand its flux
vanishes for all loops. While a source is pure flux, a vortex is pure circulation.

6
We are now using this term in a narrow sense—previously “vortex” referred to all vector fields
of this topological form.
556 Vector Fields and Complex Integration

[11.12] A more physical view of Cauchy’s Formula. If f(p) = A + iB, then as z Ñ p, the
local behaviour of H(z) ≡ f(z)/(z−p) is H(z) ≍ A[1/(z−p)]−B[i/(z−p)], as illustrated.
The first term is a source (pure flux) of strength 2πA. The second term represents a vortex
(pure circulation) such that the circulation
H f(z) round one of its circular streamlines is −2πB. If
C is a simple loop around p, then C (z−p) dz = W [H, C] + i F [H, C] = −2πB + i 2πA =
2πi f(p).

These observations give us a more physical view of Cauchy’s Formula:

I
f(z)
dz = W [H, C] + i F [H, C]
(z − p)
C
= −2πB + i 2πA
= 2πi f(p).

11.2.7 Positive Powers

If n is a positive integer then zn is analytic everywhere and its Pólya vector field zn
is correspondingly divergence-free and curl-free. Our physical version of Cauchy’s
Theorem therefore gives
I
zn dz = W [z n , C] + i F [z n , C] = 0.
C
At least in the case of an origin-centred circle we can make this much more vivid7 .
Figure [11.13] illustrates the behaviour of z and z2 on such a circle. It now seems
clear that as much fluid flows into each shaded disc as flows out, so that F = 0,
and also (when viewed as force fields) that no net work is done in transporting a
particle round the boundary of each disc, so that W = 0.
We can make this idea precise. First note that for any vector field on the circle, the
work and flux will not change their values if we perform an arbitrary rotation of
the diagram about the centre of the circle. Next, let us exploit the attractive symme-
tries of these particular vector fields. Rotating the picture of z through (π/2) clearly

7
In the particular case of z2 this has also been observed by Braden (1991), though he did not supply
the general argument which follows.
 Complex Integration in Terms of Vector Fields 557

H
[11.13] Pólya vector field explanation of C zn dz = W [z n , C] + i F [z n , C] = 0. Draw
the Pólya vector field z n of zn on an origin-centred circle. Rotating this picture about the
origin through any angle cannot change either the work or the flux. But rotating the picture
by π/(n + 1) yields the negative of the original field, changing the sign of both the work
and the flux. Since the work and the flux must remain unchanged and reverse sign, they
must both vanish.

yields the negative of the original field and, correspondingly, the negative of the
original work and flux. Since W and F are simultaneously required to remain the
same and to reverse sign, they must both vanish.
The same argument applies to z2 under a rotation of (π/3), and to z n under a
rotation of π/(n + 1). Use your computer to check this for n = 3. To understand
this symmetry better, consult Ex. 10.

11.2.8 Negative Powers and Multipoles

Consider the negative power functions (1/zm ), where m is a positive integer. Their
Pólya vector fields (1/zm ) will be divergence-free and curl-free except at the sin-
gularity at the origin. Thus if a simple loop C does not enclose the origin, its
circulation and flux will vanish. However, since we know from the case m = 1
that singularities are capable of generating flux and circulation, it remains some-
thing of a mystery that (except for m = 1) W and F also vanish if C does enclose the
singularity.
In the case of an origin-centred circle we can visualize this result exactly as for
positive powers. Figure [11.14a] illustrates this for the so-called dipole field (1/z2 ).
The argument is also the same as before: this vector field is reversed under a rota-
tion of π, and for (1/zm ) it is reversed under a rotation of π/(m − 1). Knowing that
W and F vanish for the circle tells us [see (11.9), (11.10)] that they will continue
to vanish for any loop into which we may deform the circle without crossing the
origin.
558 Vector Fields and Complex Integration

[11.14] The Dipole Pólya vector field (1/z 2 ). [a] If we draw the field on a circle, then
a rotation of π reverses the field, so by the same reasoning as the previous figure,
H 2 2 2 m
C (1/z ) dz = W [(1/z ),HC] + i F [(1/z ), C] = 0. In general, (1/z ) is reversed under a
m
rotation of π/(m − 1), so C (1/z ) dz = 0. [b] The full phase portrait of the dipole field.
Note that the streamlines are perfect circles.

Let us now go beyond this geometric explanation in search of a compelling phys-

ical explanation. Figure [11.14b] shows the phase portrait of the dipole (1/z2 ), the
streamlines of which are apparently circular; a simple geometric argument [exer-
cise] confirms their perfect circularity. Where have we seen something like this
before? Answer: the doublet field consisting of a source and sink of equal strength
S (see [10.3], p. 514). It therefore looks as though we can obtain the dipole simply by
coalescing the source and sink. This solves our mystery in a surprising and elegant
fashion: neither the source nor the sink generate circulation, and a loop enclosing
both receives equal and opposite fluxes.
This explanation is essentially correct. However, as the sink and source move
closer and closer together, a greater and greater proportion of the fluid from the
source is swallowed up by the sink before it can go anywhere, and at the moment
of coalescence the source and sink annihilate each other, leaving no field at all. Let
us investigate this algebraically using (10.1), p. 518.
Suppose that the source and sink approach the origin along a fixed line L making
an angle ϕ with the real axis. This streamline of symmetry L is called the axis of the
doublet. Putting A = ϵ eiϕ = −B, the doublet field (10.1) becomes
 
2ϵS e−iϕ 1
D(z) = , (11.18)
2π (z 2 − ϵ2 e−i2ϕ )
 Complex Integration in Terms of Vector Fields 559

which dies away as the source/sink separation 2ϵ tends to zero. The solution is
to increase the strength S in inverse proportion to the separation 2ϵ, so that 2ϵS
remains constant. If we call this real constant 2πk, the limiting doublet field (as
ϵ → 0) is

k e−iϕ
D(z) = ,
z2
i.e. the general dipole field obtained by rotating [11.14] by +ϕ and scaling up the
speed of the flow by k, which we may think of as the “strength” of the dipole. Thus
the Pólya vector field of (d/z2 ) is a dipole whose axis points in the direction of d,
and whose strength is |d|. The complex number d is called the dipole moment.
We created the dipole by coalescing equal and opposite sources, increasing their
strength so as to avoid mutual annihilation. Continuing this game, we ask, “What
will happen if we coalesce equal and opposite dipoles, increasing their strength so as
to avoid mutual annihilation?” Figure [11.15] reveals the pleasing answer. Figure
[11.15a] represents a pair of equal and opposite dipoles located at ±ϵ and hav-
ing real dipole moments ±d, while [11.15b] is the Pólya vector field of (1/z3 ). The
resemblance is striking, and we can show algebraically that [11.15b], which is called
a quadrupole, is indeed the appropriate limiting case of [11.15a].
The field for [11.15a] is
 
1 1 z
Q(z) = d − = 4dϵ 2 . (11.19)
(z − ϵ)2 (z + ϵ)2 (z − ϵ2 )2

[11.15] The Quadrupole Pólya vector field (1/z 3 ). [a] Dipoles of strength ±d at ±ϵ are
merged by letting ϵ Ñ 0. But in order to form a quadrupole, the dipole moments d must
grow in inverse proportion to the separation ϵ, so that dϵ remains constant, otherwise
the two opposite dipoles would simply annihilate each other, leaving no field. [b] The full
phase portrait of the resulting quadrupole.
560 Vector Fields and Complex Integration

Once again letting the strength d grow in inverse proportion to the separation, so
that k = 4dϵ remains constant, the coalescence of the dipoles yields the quadrupole:
k
Q(z) = 3 .
z
In general, the Pólya vector field of (q/z3 ) is called a quadrupole with quadrupole
moment q.
We have thus explained the vanishing circulation and flux of (1/z3 ): each of
the dipoles in [11.15a] is known not to generate any circulation or flux, so the
quadrupole in [11.15b] won’t either. You are invited to continue this line of thought
by showing (geometrically and algebraically) that the fusion of two quadrupoles
yields the so-called octupole field, (1/z4 ), and so on.
Dipoles, quadrupoles, octupoles, etc., are collectively known as multipoles. Sim-
ilarly, dipole moments, quadrupole moments, etc., are collectively known as
multipole moments.

11.2.9 Multipoles at Infinity

Although there is no mystery surrounding the vanishing circulation and flux for
positive powers, it would still be nice to find a physical explanation analogous to
the one for negative powers. To see how this can be done, we begin by considering
the constant function f(z) = a, the Pólya vector field of which is a flow of constant
speed |a| in the direction a.
Standing in the midst of this flow, the fluid seems to originate far over the hori-
zon in the direction −a and to disappear over the horizon in the opposite direction,
as though both a source and a sink were present at infinity. To make sense of this
idea, stereographically project the streamlines onto the Riemann sphere. Since the
streamlines are parallel lines in the direction a, their projections are circles which all
pass through the north pole in the same direction. We thus obtain a picture similar
to [10.11b], p. 525: a dipole at infinity!
Let us analyse this further. If we were standing at the midpoint of the source
and sink of the doublet in [10.3], p. 514, the flow in our vicinity would have
approximately constant speed and direction. As the source and sink recede from
us towards infinity, ultimately coalescing there to form a dipole, the approxima-
tion to a constant field gets better and better. The snag is that in this process the
magnitude of the field at any finite point dies away to nothing.
We see this algebraically in (11.18): D(z) → 0 as ϵ → ∞. However, if we let S
grow in proportion to the separation, so that (S/ϵ) = const. = kπ, say, then as
ϵ → ∞ the doublet field tends to the constant field D(z) = −k eiϕ .
Given that z0 yields a dipole at infinity, what might the Pólya vector field of z1
correspond to? Use your computer to see that it is a quadrupole at infinity. Verify
this algebraically using (11.19). Continuing in this fashion, one finds [exercise] that
z2 corresponds to an octupole at infinity, and so on.
 Complex Integration in Terms of Vector Fields 561

11.2.10 Laurent’s Series as a Multipole Expansion

The above ideas shed new light on the Laurent series and the Residue Theorem.
Suppose that an otherwise analytic function f(z) has a triple pole at the origin. We
know from Chapter 9 that f(z) will have a Laurent series of the form

q d ρ
f(z) = 3
+ 2 + +a + b z + c z2 + · · · . (11.20)
z
| z
{z z}
P(z)

In the vicinity of the singularity, the behaviour of f is governed by its principal part
P, the Pólya vector field of which is
q d ρ
P(z) = 3
+ 2+ .
z z z
This we now recognize to be the superposition of a quadrupole, a dipole, and
a source-vortex combination of the type shown in [11.12]. Thus the principal
part of the Laurent series amounts to what a physicist would call a multipole
expansion.
To visually grasp the meaning of such an expansion, consider [11.16] which
illustrates a typical P. Very close to the singularity, the field is completely dom-
inated by the quadrupole with its characteristic four loops, but as we move
slightly further away the quadrupole’s influence wanes relative to the dipole.
Indeed, at this intermediate range we clearly see the characteristic two loops
of a dipole. Finally, at still greater distances, both the quadrupole and the
dipole become insignificant relative to the source-vortex, the precise form of
which is determined solely by the residue ρ. Compare with [11.12], in which
ρ = A + iB.
Continuing our outward journey, now well beyond the unit circle, the entire
principal part becomes negligible relative to the remaining terms of (11.20). First
a becomes important, then bz takes over, and so on. Thus as we approach infinity
the field at first resembles a dipole, then a quadrupole, and so on. However, unlike
the approach to the pole, on the journey to infinity we may experience multipoles
of greater and greater order, without end.
Of course in general f may possess other singularities and (11.20) will cease to
be meaningful when |z| increases to the distance of the nearest one. Nevertheless,
in the region where it is valid, we may still think of the non-negative powers as
representing multipoles at infinity.
To recap, Laurent’s series and the Residue Theorem may be conceived of physi-
cally as follows. The only term capable of generating circulation and flux is (ρ/z),
which may itself be decomposed into a vortex of strength W = −2π Im(ρ) and
a source of strength F = 2π Re(ρ). All the other terms correspond to multipoles
which generate neither circulation nor flux; a finite collection of these reside at the
pole, while the rest are at infinity.
562 Vector Fields and Complex Integration

[11.16] Multipole expansions and the Residue Theorem. Consider the Pólya vector field
f(z) of an analytic function in the vicinity of a triple pole. The dominant behaviour
near the singularity is given by the conjugate of the principal part of the Laurent series:
P(z) = zq3 + zd2 + ρz , which we recognize to be the superposition of a quadrupole, a
dipole, and a source–vortex combination of the type shown in [11.12]. Thus the principal
part of the Laurent series amounts to what a physicist would call a multipole expansion.
Very close to the singularity, the field is completely dominated by the quadrupole with its
characteristic four loops, but as we move slightly further away the quadrupole’s influence
wanes and we see the characteristic two loops of a dipole. Finally, at still greater distances,
both the quadrupole and the dipole become insignificant relative to the source–vortex,
the precise form of which is determined solely by the residue ρ. Neither the quadrupole
nor the dipole generates any circulation or flux,H so if L encircles the pole, the circulation
and flux is entirely due to the source–vortex: L f(z) dz = 2πiρ.

11.3 The Complex Potential

11.3.1 Introduction
Phase portraits are so convenient that it is easy to forget that in general they cannot
represent the lengths of the vectors. In this section we shall see that if a vector field is
either sourceless or irrotational (or both) then there exists a special way of drawing
the phase portrait so that the lengths are represented.
 The Complex Potential 563

Although we shall ultimately be concerned with the Pólya vector fields of

analytic functions, which are both sourceless and irrotational, it is more instruc-
tive to analyse the implications for sourcelessness and irrotationality separately.
Nevertheless, in view of the final objective, we shall continue to write the vector
field as H.

11.3.2 The Stream Function

First let H be a sourceless flow of fluid. The Deformation Theorem (11.9) tells us that
the flux across a curve connecting two given points is independent of the choice of
the curve. Thus if K is any contour from an arbitrary fixed point a to a variable
point z, the flux across it, namely
Ψ (z) ≡ F [H, K],
will be a well-defined function of z, called the stream function. If we choose a dif-
ferent point a then the new stream function will only differ from the old one by an
additive constant.
Suppose that z lies anywhere on the streamline through a. See [11.17]. Choosing
K to be the portion of the streamline from a to z, we see that Ψ(z) = 0. Similarly,
suppose that q lies anywhere on a streamline through another point p. Taking K to
be a path from a to p, followed by the section of the streamline from p to q, we see
that Ψ(q) = Ψ(p). In other words,
The streamlines are the level curves of the stream function Ψ.
Instead of constructing the phase portrait by drawing random streamlines, sup-
pose we do it as follows: choose a number k and draw just those streamlines for which
Ψ = 0, ±k, ±2k, ±3k, . . . See [11.17]. Having drawn the phase portrait in this spe-
cial way, the speed of the flow is represented by the crowding together of the
streamlines. Let’s justify this claim and make it more precise.
Since no fluid crosses the streamlines, we may think of the region lying between
two adjacent ones as a tube down which fluid flows. Any curve connecting the
two sides will have the same flux, namely k. Adapting the language of Faraday
and Maxwell, we may thus describe the tube more quantitatively as a k-flux tube.
The shaded area in [11.17] is part of one such tube, the initial and final cross-
sections (lengths ϵ1 and ϵ2 ) having been drawn perpendicular to the flow. If k is
chosen small, the speed v = |H| of the flow will be approximately constant across
these ends, say v1 and v2 . Thus the fluxes into and out of the shaded region (which
must both equal k) are approximately ϵ1 v1 and ϵ2 v2 . As k is chosen smaller and
smaller, these expressions become more and more accurate:
k k
v1 = and v2 = . (11.21)
ϵ1 ϵ2
In order to maintain a constant flux k, the speed must decrease as the tube widens.
564 Vector Fields and Complex Integration

[11.17] The Stream Function Ψ and the k-flux tubes of a sourceless flow. If H is source-
less, and K is any contour connecting an arbitrary fixed point a to a variable point
z, the flux across it is independent of K, yielding the well-defined stream function,
Ψ (z) ≡ F [H, K]. Now choose a number k and draw just those streamlines for which
Ψ = 0, ±k, ±2k, ±3k, . . . Any cross section of a channel between neighbouring stream-
lines therefore intercepts the same flux k, and so we call it a k-flux tube. This special phase
portrait (pioneered by Faraday and Maxwell) has the great advantage that the speed of
the flow is now visible via |H| ≍ (k/ϵ): the denser the streamlines, the faster the flow.

To summarize:
Let the phase portrait of a sourceless vector field be constructed using k-
flux tubes. If k is chosen small, the speed of the flow at any point will be
(11.22)
approximately given by k divided by the width of the tubes in the vicinity
of the point. For infinitesimal k, the result is exact.
However, since the number of k-flux tubes passing through a given region will
vary inversely with k, our phase portrait will get very cluttered if k is chosen too
small. In practice (cf. [10.3], p. 514) we get a good feel for the speed of the flow with
relatively few streamlines.
Let’s apply these ideas to the simple (non-analytic) example H(z) = i z. The
Pólya vector field is then
!
y
H = −i z ⇐⇒ H = ,
−x

the streamlines of which are clockwise circles round the origin, the speed of the
flow round each one being equal to its radius. See [11.18].
Although this vector field is not irrotational [∇×H = − 2], it is sourceless
·
[∇ H = 0], and thus it possesses a stream function. For convenience’s sake, let’s
choose a = 0. We already know that the streamlines are origin-centred circles, so
to find the value of Ψ on the streamline of radius R we must find the flux for any
 The Complex Potential 565

[11.18] The Pólya vector field H = −i z, drawn using k-flux tubes.

path from the origin to any point on this circle. Choosing the path to be the portion
of the positive real axis from 0 to R, we see that
!
0
ds = dx and N = .
−1
Thus
Z ZR
·
Ψ = (H N) ds =
0
x dx = 12 R2 .

Knowing the stream function we are now in a position to draw the special
phase portrait. Choosing k = (1/2) we find that the radii of the streamlines
√ √ √
are 1, 2, 3, . . . Figure [11.18] illustrates these streamlines, and qualitatively
confirms the prediction of (11.22). As we move outward from the origin the
streamlines become more crowded together, reflecting the increasing speed of
the flow.

11.3.3 The Gradient Field

We have seen in geometrical terms how it is possible to reconstruct a sourceless
vector field H from a knowledge of its stream function Ψ. In order to find a simple
formula for H in terms of Ψ, we need the concept of the gradient field ∇Ψ. This is
defined to be the vector field
! !
Bx Bx Ψ
∇Ψ = Ψ = ⇐⇒ ∇Ψ = Bx Ψ + i By Ψ.
By By Ψ

The gradient field ∇Ψ has a simple geometric interpretation in terms of the

streamlines of [11.17]. To see this, we express the infinitesimal change dΨ resulting
from an infinitesimal movement dz = dx + i dy as a dot product:
566 Vector Fields and Complex Integration

[11.19] The field in terms of the stream function: H = −i ∇Ψ. First, the direction of ∇Ψ
is the one that is orthogonal to the streamlines and along which Ψ increases. Thus −i ∇Ψ
points in the direction of H. Second, |∇Ψ| ≍ (k/ϵ) ≍ |H|. Done.
! !
dΨ = (Bx Ψ) dx + (By Ψ) dy =
Bx Ψ
By Ψ
· dx
dy
·
= ∇Ψ dz.

If dz is tangent to a streamline then dΨ = 0, so ∇Ψ has vanishing dot product with

this direction. Also, Ψ increases when dz makes an acute angle with ∇Ψ. Thus
The direction of ∇Ψ is the one that is orthogonal to the streamlines and
(11.23)
along which Ψ increases. Thus −i ∇Ψ points in the direction of H.
See [11.19].
So much for the direction of ∇Ψ; what about its magnitude? In [11.19] (which is
basically a copy of [11.17]) we imagine that k is infinitesimal. Choosing dz = eiθ ds
in the direction of ∇Ψ, we find dΨ = |∇Ψ| ds. In particular, if we let ds equal ϵ (the
width of the k-flux tube) then dΨ will equal k. Thus
|∇Ψ| ≍ (k/ϵ).
But this is precisely the formula we previously obtained for the speed v = |H| of
the flow! Thus |∇Ψ| = |H|.
Combining this result with (11.23) we obtain the following simple formula for H
in terms of Ψ:
!
By Ψ
H = −i ∇Ψ ⇐⇒ H = . (11.24)
−Bx Ψ
Try this out on our previous example H(z) = i z, the Pólya vector field of which
had stream function Ψ = (x2 + y2 )/2.
Now consider the question, “What additional condition must be satisfied by Ψ
if H is also required to be irrotational?” The answer is that it must satisfy Laplace’s
equation:
 The Complex Potential 567

∆Ψ ≡ Bx2 Ψ + By2 Ψ = 0.
Solutions of this equation are called harmonic, so we may restate the result as
follows:
A sourceless field is irrotational if and only if its stream function is harmonic.
The verification is a simple calculation:
! !
Bx By Ψ
∇×H = × = −∆Ψ.
By −Bx Ψ

11.3.4 The Potential Function

Next suppose that H is a force field which is known to be conservative (irrotational).
In this case it is the work rather than the flux which must be path-independent.
Thus if K is any contour from an arbitrary fixed point a to a variable point
z, the work done by the field in moving the particle along K is a well-defined
function of z,
Φ(z) = W [H, K].
This is called the potential function, though there are several pseudonyms depend-
ing on the context: e.g., in electrostatics it is called the “electrostatic potential”, in
hydrodynamics it is called the “velocity potential”, and in the case of flowing heat
it is already familiar as the temperature. As with the stream function, changing the
choice of a merely changes Φ by an additive constant.
Let’s investigate Φ as we did Ψ. The level curves Φ = const. are called equipo-
tentials; what is their geometric significance? As [11.20] illustrates, the answer is
that
The equipotentials are the orthogonal trajectories through the lines of force. (11.25)
The reason should be clear. A certain amount of work Φ(p) is done in moving the
particle from a to p, but then no additional energy is expended in moving it to q
along the orthogonal trajectory through p. Thus Φ(q) = Φ(p).
Instead of illustrating random equipotentials, [11.20] mimics the special con-
struction used in [11.17]: we draw just those equipotentials for which Φ = 0, ±l, ±2l,
±3l, . . .. In this picture the same amount of work l is required to move the particle
from each equipotential to the next. Let us therefore call the region lying between
two such adjacent equipotentials an l-work tube.
Suppose that l is chosen small, and consider the work done in moving a particle
along the correspondingly short cross-section δ in [11.20]. In the limit of vanishing
l we find that
l
|H| = . (11.26)
δ
568 Vector Fields and Complex Integration

[11.20] The Potential Function Φ of a conservative force field. If K is any contour con-
necting an arbitrary fixed point a to a variable point z, the work done by the conservative
(irrotational) field in moving the particle along K is a well-defined function of z, indepen-
dent of K, called the potential function, Φ(z) = W [H, K]. Just as we did with the stream
function, we draw just those equipotentials for which Φ = 0, ±l, ±2l, ±3l, . . . In this
picture the same amount of work l is required to move the particle from each equipoten-
tial to the next, so we call the region lying between two such adjacent equipotentials an
l-work tube. The equipotentials are the orthogonal trajectories through the lines of force.
Also, |∇Φ| ≍ (l/δ) ≍ |H|, so H = ∇Φ.

Thus the magnitude of the force is represented by the crowding together of the
equipotentials:

Let the equipotentials of a conservative force field be constructed using

l-work tubes. If l is chosen small, the magnitude of the force at any point
(11.27)
will be approximately given by l divided by the width of the tubes in the
vicinity of the point. For infinitesimal l, the result is exact.

Since the gradient field ∇Φ is automatically orthogonal to the equipotentials and

has magnitude (l/δ), we may combine (11.25) and (11.27) into the simple formula
!
Bx Φ
H = ∇Φ ⇐⇒ H = . (11.28)
By Φ

Lastly, suppose that H is required to be sourceless. Since

! !
∇ H=· Bx
By
·
Bx Φ
By Φ
= ∆Φ, (11.29)
 The Complex Potential 569

we see that
A conservative force field is sourceless if and only if its potential function
(11.30)
is harmonic.

11.3.5 The Complex Potential

We now know two things about a vector field H that is irrotational and sourceless:
(i) both Φ and Ψ exist; (ii) it is the Pólya vector field of an analytic function. In this
section we shall attempt to illuminate the connections between these two facts.
Since Φ and Ψ both exist, we may superimpose pictures of types [11.17] and
[11.20], thereby simultaneously dividing the flow into mutually orthogonal k-flux
tubes and l-work tubes. Before drawing this picture let us choose the increment of
work to be numerically equal to the increment of flux: l = k.
Let us call the intersection of a k-flux tube with a k-work tube a k-cell. We already
know that the sides of each k-cell meet at right angles, so for small k they will be
approximately rectangles. The sides of such a rectangle will be the previously con-
sidered widths ϵ and δ of the two kinds of tube. But combining the results (11.21)
and (11.26) we see that
k k
= |H| = =⇒ δ = ϵ.
ϵ δ
Thus

In the limit of vanishing k, the k-cells are squares. (11.31)

The LHS of [11.21] illustrates such a division into approximately square k-cells.
We have labelled Φ = 11k and Ψ = 3k, but we have left it to you [exercise]
to label the remaining streamlines and equipotentials; this can only be done in
one way.
Note that once such a special phase portrait (including the equipotentials) has
R
been drawn with a small value of k, the value of L H dz is easy to find. For if L
crosses m equipotentials and n streamlines, an accurate estimate of the integral
will be k(m + in). If L crosses an equipotential or streamline more than once, how
should m and n be counted?
We mention in passing that there is an interesting physical interpretation of the
k-cells which is due to Maxwell (1881). Suppose that the vector field represents the
flow of a fluid having unit mass per unit area. In the limit of vanishing k, the speed
v will be constant throughout any particular cell, and the kinetic energy of the fluid
in that cell will be
 2
2 1 2 k
1
kinetic energy = 2 (area) v = 2 ϵ = 21 k2 .
ϵ
570 Vector Fields and Complex Integration

[11.21] The Complex Potential, Ω = Φ + iΨ. If H(z) is analytic, then H(z) is both irro-
tational and sourceless, so we may simultaneously define and superimpose l-work tubes
orthogonal to the streamlines with k-flux tubes along the streamlines, thereby dividing the
region of the flow into small “rectangles” (as l and k tend to zero). We now go one step
further and choose l = k, dividing the region into squares bounded by equipotentials
and streamlines. This makes it vividly clear that if we combine the potential and stream
functions into the single complex potential Ω = Φ + iΨ, then the mapping Ω is locally
an amplitwist, i.e., Ω ′ exists. Indeed, we will show that H = Ω ′ .

Thus
Each k-cell contains the same amount of energy, and the total energy in a
region is thus obtained by counting the number of k-cells contained within
it.
If we reinterpret the vector field as an electrostatic field, and correspondingly rein-
terpret “energy” as electrostatic energy, the result is still valid; this was the context
in which Maxwell discovered it.
The result (11.31) is intimately connected with ideas of complex analysis. To
see this, let us combine the potential and stream functions into a single complex
function Ω called the complex potential:
Ω(z) = Φ(z) + i Ψ(z).
Returning to the dominant point of view of this book, think of Ω as a mapping. The
RHS of [11.21] shows the image of the special phase portrait under this mapping:
The complex potential maps streamlines to horizontal lines and equipo-
tentials to vertical lines. Furthermore, each square k-cell is mapped to a
square of side k. Thus Ω is an analytic mapping.
We may check this symbolically. Equating (11.24) and (11.28) we obtain
! !
Bx Φ By Ψ
= ,
By Φ −Bx Ψ
which are the CR equations for Ω.
 The Complex Potential 571

What is the amplitwist of the complex potential? By considering the effect of Ω

on the black k-cell in [11.21] we see that if the streamline through z makes an angle
θ with the horizontal, the twist of Ω at z is −θ, which we recognize as the angle
of H(z). We also see that the amplification of Ω is (k/ϵ), which we recognize as
|H| = |H|. Thus
Ω ′ = H.

Since H is the derivative of an analytic function, it must itself be analytic. We

have thus obtained a second, more geometrical proof that the class of sourceless,
irrotational vector fields is the same as the class of Pólya vector fields of analytic
functions.
The result Ω ′ = H can be checked symbolically. Substituting one of the CR
equations for Ω into (11.28), we obtain

H = ∇Φ = Bx Φ + i By Φ = Bx Φ − i Bx Ψ = Bx Ω = Ω ′ .

When we thought of an analytic function f as a conformal mapping, f ′ repre-

sented its amplitwist. But since any such function may instead be thought of as
the complex potential of a flow, we now have another interpretation of differentia-
tion: f ′ is the conjugate of the velocity of the flow described by f. Correspondingly,
we also have a new interpretation of critical points: they are the places where the
velocity vanishes. Such places are called stagnation points in the flow.
By analysing the implications of sourcelessness and irrotationality separately,
we have been able to understand the Pólya vector fields of non-analytic functions
that may possess a stream function or a potential function, but not both. If we had
instead restricted ourselves from the outset to the Pólya vector fields of analytic
functions, the complex potential could have been obtained more rapidly (but less
revealingly) as follows.
If L is any contour from an arbitrary fixed point a to a variable point z, we may
define
Z
ΩL (z) = H(w) dw = W [H, L] + i F [H, L].
L

But, as we saw in Chapter 8, if H is analytic then this integral is independent of L,

and the well-defined function
Zz
Ω(z) = H(w) dw = Φ(z) + iΨ(z)
a

is in fact the antiderivative of H. More explicitly, the image Ω(L) of a contour L

from p to q is the path taken by the Riemann sum for the integral of H along L.
The value of the integral is then the vector connecting the start of Ω(L) to its finish,
namely, Ω(q) − Ω(p).
572 Vector Fields and Complex Integration

[11.22] Ω(z) = z+ z1 represents the superposition of a uniform eastward flow and a dipole.

11.3.6 Examples
(1) We previously claimed that the streamlines of the dipole H = (1/z 2 ) in [11.14b]
were perfect circles, and we asked you to provide a simple geometric proof. A
second demonstration is obtained by finding the complex potential:
1 1 1
H= 2
=⇒ Ω ′ = 2 =⇒ Ω = − + c.
z z z
The streamlines are the images under Ω−1 (z) = −1/(z − c) of horizontal lines. The
result follows from the fact that inversion sends straight lines to circles through
the origin.

(2) A uniform eastward flow has complex potential Ω = z. If we insert a dipole

of complex potential Ω = (1/z) into this flow then the new flow will be the
superposition of the two individual flows and thus will have complex potential
1
Ω(z) = z + .
z
Using your computer you may verify that the streamlines and equipotentials are
as shown in [11.22]. Note how the streamlines emanating from the dipole are
deformed out of perfect circularity by the uniform flow, but that this distortion
diminishes as the origin is approached.

(3) A source of strength 2π at the origin has vector field H = (1/z). If we choose to
measure work and flux along a path L emanating from z = 1 then [see p. 465] the
complex potential is
 The Complex Potential 573

[11.23] The complex potential of the source (1/z) is a multifunction: although the work
is single-valued, the flux increases by 2π with every extra revolution around the source.
But if we restrict attention to a simply connected region D that does not contain the
source, then Ω becomes single-valued.

Ω(z) = Φ(z) + i Ψ(z) = logL (z) = ln |z| + i θL (z).

While the work Φ is single-valued, the flux Ψ is a multifunction whose values differ
from each other by multiples of 2π. This makes perfect sense since each time L
encircles the source it intercepts the full 2π of fluid being pumped in there. Note
that the single-valued inverse function Ω−1 (z) = ez does indeed map horizontal
and vertical lines to the source’s streamlines and equipotentials.
If we wish to obtain a single-valued complex potential we may do so by con-
fining our attention to any simply connected region not containing the source. The
shaded region D in [11.23] is an example. Any two paths from 1 to z that lie wholly
within D may be deformed into each other without ever leaving D, hence without
crossing the source, hence without altering the flux. For example, we see that for
the particular choice of D in [11.23], the unique values of Ψ at (1 + i) and at (2 + 2i)
are (π/4) and (9π/4). However, a different choice of D might well yield different
values of Ψ at these two points.
More generally, if D is any simply connected region not containing any singular-
ities of an otherwise analytic H, the Pólya vector field H will possess a single-valued
complex potential in D.
574 Vector Fields and Complex Integration

11.4 Exercises

1 For each of the following vector fields X verify that the geometric formulae
(11.5) and (11.6) yield the correct values for the divergence and for the curl:
(i) X = (1/z).
(ii) X = z.
(iii) X = x2 , where z = x + iy.
(iv) X = y2 , where z = x + iy.
(v) X = i(1/r2 )eiθ , where z = reiθ .
2 For each of the following vector fields X, calculate F [X, C] and W [X, C] for
the given loop C, then check your answers by substituting the results of the
previous question into (11.7) and (11.8).
(i) X = x2 , and C is the edge of the rectangle a ⩽ x ⩽ b, −1 ⩽ y ⩽ 1, traversed
counterclockwise.
(ii) X = i(1/r2 )eiθ , and C is the edge of the region a ⩽ r ⩽ b, 0 ⩽ θ ⩽ π,
traversed counterclockwise.
3 Use a computer to draw the Pólya vector field of f(z) = 1/[z sin z] and thereby
identify the locations and orders of the poles of f(z). For each of the following
H
choices of C, numerically estimate C f(z) dz by making on-screen measure-
ments of the vectors, then estimating the flux and circulation round C. In each
case check your estimate by calculating the exact answer using residue theory.
(i) Let C be a small circle centred at −π.
(ii) Let C be a small circle centred at 0.
(iii) Let C be a small circle centred at π.
(iv) Let C be a small circle centred at 2π.
(v) Let C be the boundary of the rectangle 1 ⩽ x ⩽ 7, −1 ⩽ y ⩽ 1.
4 Repeat parts (i) and (ii) of the previous question using f(z) = z cosec 2 z.
5 Let L be a contour from the real number −θ to +θ. By choosing L to be a line-
segment, and then sketching the Pólya vector field at points along L, show that
R iz
L z e dz is purely imaginary. Verify this by calculating the exact value of the
integral.
6 All complex analysis texts recognize the great utility of the inequality

Z Z

L
f(z) dz ⩽
L
·
|f(z)| |dz|, (11.32)
 Exercises 575

but none that we know of have sought to answer the question, “When does
equality hold?” This is probably because no elegant answer is forthcoming (cf.
our attempt in Chapter 8) without the concept of the Pólya vector field. How-
ever, armed with the Pólya vector field, we have what we shall call Braden’s
Theorem8 :
Equality holds in (11.32) if and only if the contour L cuts the streamlines
of the Pólya vector field of f at a constant angle.
Explain Braden’s Theorem.
7 Continuing from the previous question, suppose that f(z) = z.
(i) Show that if L is a segment of the spiral with polar equation r = eθ , then
the condition of Braden’s Theorem is met.
(ii) Verify by explicit calculation that equality does indeed hold in (11.32), as
predicted.
8 Consider the flow created by (2n + 1) sources, each of strength 2π, located at

0, ±π, ±2π, . . . , ±nπ.

(i) If Ωn (z) denotes the complex potential of this flow, show that
     
z2 z2 z2
Ωn (z) = ln z 1 − 2 1 − 2 2 ··· 1 − 2 2 + const.
π 2π nπ
(ii) Ignoring the constant, and referring to Ex. 13 on p. 508, deduce that as the
number of sources increases without limit, Ωn (z) tends to Ω(z) = ln[sin z].
(iii) Check that this answer makes sense by using a computer to draw the
velocity vector field, V = Ω ′ .
9 (i) Explain why the derivative of the complex potential of a source yields the
complex potential of a dipole.
(ii) Referring to the previous question, draw a sketch predicting the appear-
d
ance of the flow whose complex potential is Ω(z) = dz ln[sin z]. Check
your answer by getting the computer to draw this flow.
10 Reconsider the term C z in the local decomposition (11.17) of a general vector
field. See [11.11].
(i) Show that the visual appearance of the vector field C z is essentially inde-
pendent of the value of C. More precisely, show that if C = eiϕ then
increasing ϕ merely causes the entire picture of the vector field C z to
rotate, in fact exactly half as fast as eiϕ rotates.

8
See Braden (1987). We independently recognized this fact, probably at about the same time as
Braden himself.
576 Vector Fields and Complex Integration

(ii) To make the result vivid, create a computer animation of the vector field
eiϕ z as ϕ increases from 0 to π.
(iii) More generally, show that if n is an integer and F(z) stands for either zn
or z−n , then the vector field of eiϕ F is obtained by rotating the vector field
of F through ϕ/(n + 1). [Note that the n = −1 fields (including sources
and vortices) are exceptional.]
11 Consider a flow such that the inverse complex potential is Ω−1 (w) = w + ew .
Use a computer to draw the streamlines, and verify mathematically that the
picture may be interpreted as the flow out of a channel −π ⩽ Im(z) ⩽ π,
Re(z) ⩽ −1.
12 Consider the flow with complex potential
 
1 ez + 1
Ω(z) = .
2 ez − 1
Use a computer to draw the streamlines, and verify mathematically that the pic-
ture may be interpreted as the flow that results when the dipole with complex
potential Ω(z) = (1/z) is confined to the channel −π ⩽ Im(z) ⩽ π.
13 Continuing from the previous question, what would the new complex potential
be if fluid were flowing down the channel with speed v prior to the insertion of
the dipole? Check your answer by using a computer to draw the streamlines.
14 Suppose that the doublet consisting of a source of strength 2π at z = 1 and a
sink of equal strength at z = −1 is inserted into the uniform flow with real, pos-
itive velocity v. Locate the “stagnation points” (singular points of zero velocity)
of the net flow, and describe (perhaps with the aid of a computer animation)
how they move as v varies from 0 to 3.
15 If two sources are located at opposite corners of a square, and two sinks are
located at the other two corners, and all four are of equal strength, then show
that the circle through these four points is a streamline. Check this by getting
the computer to draw the complete flow.
16 Show that the streamlines produced by two vortices of equal strength are Cassi-
nian curves (figure [2.8b], p. 68) whose foci are the locations of the two vortices.
[Note that your reasoning immediately generalizes: Cassinian curves with n foci
are the streamlines of n equal vortices placed at the foci.]
17 Show that the streamlines [11.15b] and the equipotentials of a quadrupole are
lemniscates (see [2.9], p. 69).
CHAPTER 12

Flows and Harmonic Functions

12.1 Harmonic Duals

12.1.1 Dual Flows

As in the previous chapter, let H = Ω ′ be a steady, sourceless, irrotational vector
field with complex potential Ω = Φ+iΨ. If at each point we rotate H through a fixed
angle ϑ then we obtain the Pólya vector field of the analytic function Hϑ ≡ e−iϑ H,
namely, Hϑ = eiϑ H. Thus this rotated vector field is automatically sourceless and
irrotational, and its complex potential is Ωϑ = e−iϑ Ω. Writing Ωϑ ≡ Φϑ + i Ψϑ , the
potential and stream functions are therefore

Φϑ = (cos ϑ) Φ + (sin ϑ) Ψ and Ψϑ = (cos ϑ) Ψ − (sin ϑ) Φ.

Henceforth we shall concentrate on the particularly simple and important case

in which ϑ = +(π/2). After rotating H through this right angle we obtain the Pólya
b . Thus H
vector field of Hπ/2 , for which we shall use the special symbol H b ≡ Hπ/2 =
iH. In complex analysis, the standard terminology is to say that H b is “conjugate”
1
to the original flow H. However, I know of no mathematical connection between
this concept and the familiar one of complex conjugation. Furthermore, our use
of Pólya vector fields (involving genuine complex conjugation) brings these two
senses of “conjugate” into direct conflict, for the complex conjugate of the original
flow is not the “conjugate” flow.
Fortunately, in other areas of mathematics (e.g., topology) there is another term
that is commonly used to describe this idea. We therefore propose to call Hb the dual
of H. Similarly, let us call the potential and stream functions of the dual flow the
dual potential and the dual stream function.

1
Linguistically, the common origin of both terms is the Latin word “conjugatus”, meaning joined
together.

Visual Complex Analysis. 25th Anniversary Edition. Tristan Needham, Oxford University Press.
© Tristan Needham (2023). DOI: 10.1093/oso/9780192868916.003.0012
578 Flows and Harmonic Functions

[12.1] Evolution of a source into its dual vortex: As ϑ goes from 0 to (π/2), the field
b ≡ Hπ/2 =
Hϑ ≡ eiϑ H = eiϑ (1/z) rotates from a pure source into its dual pure vortex, H
iH = (i/z).

Later we shall see that the concept of a dual flow is very useful. For example,
having found the flow of a fluid round an obstacle, the dual flow represents the
electric field which solves an analogous problem in electrostatics.
As interesting examples of dual flows, consider what happens in the vicinity
of a singularity. Figure [12.1] illustrates how, as ϑ varies from 0 to (π/2), a source
gradually evolves into a dual vortex of equal strength. Note (cf. [11.12], p. 556) that
the intermediate flow may also be viewed as a superposition of the original flow
and its dual. Indeed, this is true quite generally:
b.
Hϑ = (cos ϑ)H + (sin ϑ)H

Check for yourself that the type of qualitative change of flow exhibited in [12.1]
does not occur in the case of higher multipoles. For example, the dual of a dipole
is just another dipole. As ϑ varies from 0 to (π/2), are all the intermediate flows
dipoles, as well? See Ex. 10 of the previous chapter.
Observe that in passing from a flow to its dual the roles of the streamlines and
equipotentials are interchanged: the streamlines of the dual flow are the equipoten-
tials of the original, while the equipotentials of the dual flow are the streamlines
of the original. Symbolically, this interchange of roles is manifested in the fact that
the dual potential and stream functions are
b = +Ψ and Ψ
Φ b = −Φ.

The difference of sign in these two equations is easily understood when we

look at [12.2], which depicts a typical flow and its dual. [Streamlines are solid and
equipotentials are dashed.] Recall that if we think of these pictures as force fields,
work is done by the field when a particle moves along a line of force, so the original
and dual potentials increase in the illustrated directions. Similarly, when thought
of as a fluid flow, the flux across a directed segment of curve is positive when the
 Harmonic Duals 579

[12.2] In passing from a flow to its dual, the streamlines and equipotentials are
interchanged.

fluid crosses it from left to right, so the original and dual stream functions increase
in the illustrated directions. We now see clearly that Φ b and Ψ increase in the same
direction, while Ψ b and Φ increase in opposite directions.
Given a complex potential Ω = Φ + iΨ, we may thus think of Ψ as either
the stream function, or as the dual of the potential function. Likewise, Φ may be
thought of as either the potential function, or as minus the dual of the stream func-
tion. Since any analytic function f = u + iv may be thought of as a complex
potential, we may extend this language and say that v is dual to u, and that −u
is dual to v.
Finally, we cannot resist at least mentioning two miraculous connections
between the above ideas and the study of soap films, also known as minimal
surfaces, characterized by vanishing mean curvature.
Imagine yourself standing on the tangent plane to a saddle-shaped surface. In
the case of a soap film, as you turn around, the average (or “mean”) curvature
vanishes: the surface bends away from the tangent plane equally and oppositely as
you turn around.
First miracle: Each complex analytic function H(z) describes the shape of a
minimal surface, and vice versa. Second miracle: Varying ϑ causes the minimal
surface corresponding to Hϑ (z) to undergo stretch-free bending: all these minimal
surfaces have identical intrinsic geometry. For example, if H corresponds to the so-
b corresponds to the so-called catenoid, and [12.3] illustrates
called helicoid, then H
the stretch-free bending of one into the other, each intermediate surface itself being
a minimal surface.
For an elementary introduction to the fascinating subject of minimal surfaces, see
Hildebrandt and Tromba (1985); for the mathematical details, see Nitsche (1989).
580 Flows and Harmonic Functions

[12.3] Miraculously, complex analytic functions correspond to minimal surfaces. In

particular, the helicoid and catenoid correspond to dual complex functions, and the evo-
lution as ϑ goes from 0 to (π/2) is isometric: the intrinsic geometry does not change
throughout the evolution, and all the intermediary surfaces are minimal, too!

NOTE added in this 25th Anniversary Edition: An especially good undergraduate

introduction to this subject has now appeared in Woodward and Bolton (2019). The
discussion includes the 1982 breakthrough discovery by Celso Costa of new mini-
mal surfaces (beyond planes, catenoids, and helicoids) that are closed and without
self-intersections. Furthermore, unusually, the authors provide a thorough discus-
sion of the generalization to surfaces of constant mean curvature (CMC surfaces), in
which the average curvature is the same at every point of the surface, but its value
does not vanish (as it must do for soap films).

12.1.2 Harmonic Duals

We know that both the real and the imaginary parts of an analytic function are
automatically harmonic. It is therefore natural to wonder if, conversely, every har-
monic function is the real (or imaginary) part of some analytic function. As we shall
see, this is indeed the case. That is, given a harmonic function u we can always
find another harmonic function v, the harmonic dual of u, such that f = u + iv
is analytic. [Again, the standard terminology is that v is the “harmonic conjugate”
of u.]
 Harmonic Duals 581

We make two remarks before proceeding. First, if v is an harmonic dual then so

is v + const., and consequently v will only be uniquely determined if we impose
additional conditions, such as v vanishing at a particular point. Second, the har-
monic dual of a single-valued function may itself be a multifunction. Witness the
case u = ln |z|, illustrated in [12.1], for which v = arg(z).
Given an irrotational vector field, we know how to construct a potential func-
tion. But, conversely, if we are given a real function Φ(z) then we may construct an
irrotational vector field H for which Φ is the potential function, namely,
H = ∇Φ.
If Φ is harmonic then we know [(11.30), p. 569] that H will be sourceless, and so it
will possess a stream function Ψ. Since H is irrotational, Ψ is harmonic. The complex
potential Ω = Φ + iΨ will then be an analytic function having as its real part the
given harmonic function Φ. In other words we have shown that
The harmonic dual of a given harmonic function Φ is the stream function
of the vector field ∇Φ.
Alternatively, Ψ is the potential function for the dual of ∇Φ.
This result means that facts about analytic functions can sometimes be recast as
facts about harmonic functions. For example, in Chapter 9 we saw that if f = u + iv
is analytic then ⟨f⟩ = f(p), where ⟨f⟩ denotes the average of f over any circle centred
at p. It follows that the harmonic real part of f obeys the law ⟨u⟩ = u(p). But we now
know that if u is any given harmonic function then we may construct an analytic
function for which it is the real part. We thus obtain Gauss’s Mean Value Theorem:
The average value of a harmonic function on a circle is equal
to the value of the function at the centre of the circle.
We obtain another example by reconsidering [7.14], p. 406, in which we saw that
if f = u + iv is analytic in some region whose boundary is Γ , then the maximum of
u occurs on Γ . The existence of harmonic duals therefore implies that
If a function is harmonic in some region, its maximum occurs on the
boundary of that region.
The same goes for a (nonzero) minimum of a harmonic function.
Next we give explicit formulae for the construction of harmonic duals. To make
Ψ unique, let us demand that it vanish at some point a. Then if K is any path from
a to p, we have the flux formula
Z
Ψ(p) = (∇Φ) N ds.
K
·
Alternatively, in terms of complex integration, we have
Z 
Ψ(p) = Im (∇Φ) dz .
K
582 Flows and Harmonic Functions

As we have seen, if we restrict ourselves to a simply connected region through-

out which Φ is harmonic, these integrals are single-valued. However, if the region
is not simply connected, or if Φ has singularities, then (in general) Ψ will be a
multifunction.
We illustrate these formulae with the example Φ = x3 −3xy2 , which is easily seen
to be harmonic. Choose a = 0, let p = X + iY, and choose K to be the line-segment
between them, which may be represented parametrically as z = x + iy = (X + iY)t,
where 0 ⩽ t ⩽ 1. Since
! ! !
3x2 − 3y2 3X2 − 3Y 2 1 Y
∇Φ = = t , N= √
2
,
−6yx −6YX X2 + Y 2 −X
√
and ds = X2 + Y 2 dt, the first formula yields [exercise]

Ψ = 3X2 Y − Y 3 .

Alternatively, since

∇Φ = (3x2 − 3y2 ) + i 6xy = 3z2 ,

the second formula yields

Z X+iY 
Ψ = Im 3z2 dz = Im (X + iY)3 = 3X2 Y − Y 3 .
0

The simplicity of the second method depended crucially on our ability to express
∇Φ(x, y) as a function Ω ′ (z) of z, but it is not always so obvious how to do this.
However, there does exist a systematic method of doing this in the case where Φ
is defined in a region containing a segment of the real axis.
Let V(x) be the vector field evaluated on the real x-axis, i.e., V(x) = ∇Φ(x, 0). If
Φ(x, y) is an explicit formula in terms of the familiar functions (powers, trigono-
metric, exponential) that possess complex analytic generalizations, then V(x) is
such a formula also. Hence if in the formula for V(x) we now replace the sym-
bol x with the complex variable z then we obtain an analytic function V(z) which
agrees with Ω ′ (z) when z is real. But, as we saw in Chapter 5, this implies that the
two functions must continue to agree when z becomes complex.
Thus our recipe for finding Ω ′ (z) as an explicit formula in z is to calculate
∇Φ(x, y), set y = 0, then substitute z for x:

Ω ′ (z) = ∇Φ(z, 0). (12.1)

For example, if Φ = cos[cos x sinh y] esin x cosh y then [exercise]

∇Φ(x, y) = cos[cos x sinh y] esin x cosh y cos x cosh y + F,

where F stands for three terms which vanish when y = 0. Using (12.1) we get
Ω ′ (z) = esin z cos z, and hence Ψ = Im esin z .
 Conformal Invariance 583

12.2 Conformal Invariance

12.2.1 Conformal Invariance of Harmonicity

Let w = f(z) be a complex analytic function of z, which we will think of as a confor-
mal mapping (rather than as a vector field) from the z-plane to the w-plane. Using
f, any real function Φ(z) in the z-plane may be copied over (or “transplanted”) to
e
a function Φ(w) in the w-plane by defining
e
Φ[f(z)] ≡ Φ(z). (12.2)

In other words, corresponding points in the two planes are assigned equal function
values. We will now show (first symbolically then geometrically) that
e
Harmonicity is conformally invariant: Φ(w) is harmonic if and only if
(12.3)
Φ(z) is harmonic.
e
As before, think of Φ(w) e ≡ ∇Φ.
as the potential of the vector field V e If and only
e e e e
if Φ is harmonic, V possesses an analytic complex potential Ω(w) = Φ(w)+i Ψ(w), e
where the stream function Ψ e is the harmonic dual of Φ.
e Since f is analytic, so is its
composition with Ω: e

e
Ω(z) ≡ Ω[f(z)] = Φ(z) + iΨ(z).

Thus Φ(z) is the real part of an analytic function, and so it is harmonic.

There is a very simple geometric idea behind this important result. Figure [12.4]
illustrates a visual means of checking whether or not a given real function Φ is
harmonic. Once again, think of Φ as the potential of the force field V = ∇Φ. We
know that Φ is harmonic if and only if V admits a complex potential. This we know
occurs if and only if the field may be divided into a grid of (infinitesimal) square
k-cells.
To check this we should therefore construct a “test grid”:

(i) With a small value of k, draw the equipotentials Φ = 0, ±k, ±2k, ±3k, . . ..
(ii) Choose one of the resulting k-work tubes [shaded in the figure] and draw
line-segments across it in such a way that the tube is divided into squares.
(iii) Extend these line-segments into lines of force [dashed] of V, i.e., orthogonal
trajectories through the equipotentials.

Then Φ is harmonic if and only if these lines of force divide each k-work tube into squares.
Figure [12.4a] illustrates this test for a Φ that is harmonic, while [12.4b] illustrates
it for one that is not. The result (12.3) can now be seen as nothing more than a
statement of the conformal invariance of this geometric test. Let us spell this out.
Equation (12.2) defines the potential of each point in the z-plane to be the same
as its image point (under f) in the w-plane. Thus f maps the k-work tubes of Φ to
584 Flows and Harmonic Functions

[12.4] Geometric test for harmonicity of Φ. STEP ONE: With a small value of k, draw
the equipotentials Φ = 0, ±k, ±2k, ±3k, . . . STEP TWO: Divide a k-work tube [shaded]
into squares. STEP THREE: Extend the edges of the squares into lines of force. Then Φ
is harmonic if and only if these lines of force divide each k-work tube into squares.
So [a] is harmonic, and [b] is not.

[12.5] Conformal invariance of the harmonicity test for Φ. If f is conformal, then the
image test grid will pass the harmonicity test if and only if the original test grid passes the
test. Here, the initial grid passes, so its image does, too.
e See [12.5]. Finally, since f is conformal, the constructed test
the k-work tubes of Φ.
grid for Φ will be composed of squares if and only if the image grid is composed
of squares. Figure [12.5] illustrates the case where the potentials are harmonic.

12.2.2 Conformal Invariance of the Laplacian

The result (12.3) is merely a special case of the following more general result on the
conformal invariance of the Laplacian operator ∆:

e 1
∆Φ(w) = ∆Φ(z). (12.4)
|f ′ (z)|2
 Conformal Invariance 585

[12.6] Geometric explanation of ∇ V(w) · e ·

= [1/|f ′ (z)|2 ] ∇ V(z). Let us explain this for
f(z) = cz; in the general case, we need only then replace c with the local amplitwist,
f ′ (z). Compare the flux leaving the shaded disc on the left with the flux leaving the shaded
image disc on the right. Since the separation of adjacent equipotentials is scaled up by
|c|, the strength of the field on the rim of the image disc is scaled down by |c|, while the
circumference of the rim is scaled up by |c|. The net effect is that the flux of V e out of the
image disc is the same as the flux of V out of the original disc. Finally, since the area of
the disc is scaled up by |c|2 , the flux density is scaled down by |c|2 . Done.

We will give two explanations of this result.

For the first explanation, we rephrase the result in terms of flux densities. Just
as we did for Φ e in the w-plane, let us construct the vector field V ≡ ∇Φ in the
z-plane. We wish to understand this:

· ·
e 1
∇ V(w) = ∇ V(z). (12.5)
|f ′ (z)|2
Now consider [12.6], which illustrates a toy model of the phenomenon. The
potential Φ(z) = (S/4) |z|2 generates a vector field V = (S/2) z of uniform diver-
·
gence ∇ V = S. With a small value of k, the LHS of [12.6] shows the special
equipotentials Φ = 0, ±k, ±2k, ±3k, . . ., for which the strength of the field is
inversely proportional to the separation of the curves. Now apply the mapping
w = f(z) = cz, which is a rotation and an expansion by |c|. By definition, these
e
expanded circles are equal-valued equipotentials of Φ(w), so that

e S
Φ(w) = Φ(z) = 41 S |z|2 = 1
|w|2 .
4 |c|2
The field Ve = ∇Φ e therefore has uniform flux density ∇ V ·
e = S/|c|2 , proving (12.5)
for this case.
More intuitively still, in [12.6] compare the flux leaving the shaded disc on the
left with the flux leaving the shaded image disc on the right. Since the separation
of adjacent equipotentials is scaled up by |c|, the strength of the field on the rim of
586 Flows and Harmonic Functions

the image disc is scaled down by |c|, while the circumference of the rim is scaled
up by |c|. The net effect is that the flux of Ve out of the image disc is the same as the
flux of V out of the original disc. Finally, since the area of the disc is scaled up by
|c|2 , the flux density is scaled down by |c|2 .
To employ this idea in the general setting, it is only necessary to recognize
that the local behaviour of a general potential is very similar to our toy potential,
and that the local effect of a general analytic mapping f is very similar to our toy
mapping, with |f ′ | playing the role of |c|.
We will not spell this out completely because we will shortly be able to give a
second explanation which is even simpler. However, according to (11.17), p. 553,
the behaviour of V very near to z0 is expressible as

·
(z − z0 )
V(z) = [∇ V(z0 )] + Y(z),
2
where Y is sourceless and, of course, irrotational. Correspondingly, the local
behaviour of the potential is2

·
Φ(z) = 14 [∇ V(z0 )] r2 + Υ(z), (12.6)

where Υ is harmonic, and r = |z − z0 | is the small distance from z0 . Having made

explicit the connection with the toy model, we leave the remaining details to the
interested reader.

12.2.3 The Meaning of the Laplacian

Given a real function Φ(z) in the z-plane, we have seen that its gradient vector field
∇Φ is a geometric quantity, independent of the coordinates used to describe z. We
have also seen that the divergence of a vector field measures its flux-density, so
it too is geometrically defined. It follows that the Laplacian ∆Φ = ∇ ∇Φ must ·
possess a coordinate-independent interpretation.
In order to state this interpretation, recall that if C is a circle centred at p then
⟨Φ⟩ denotes the average value of Φ on C. We will show that

The Laplacian of Φ at p measures the amount by which the average value

of Φ on an infinitesimal circle centred at p exceeds the value of Φ at p
itself. More precisely, if r is the infinitesimal radius of this circle then (12.7)

⟨Φ⟩ − Φ(p) = 14 r2 ∆Φ.

Note that this result is in accord with Gauss’s Mean Value Theorem, which says
that if Φ is harmonic then ⟨Φ⟩−Φ(p) = 0 for circles of any size, not just infinitesimal
ones. In fact if you have already convinced yourself of (12.6) then [exercise] you

2
This may also be derived directly by taking the Taylor series for Φ and rewriting it in a rather
unobvious way.
 A Powerful Computational Tool 587

may derive (12.7) by using the fact that the harmonic function Υ obeys Gauss’s
Mean Value Theorem.
Before giving a more direct derivation of (12.7), let us return to Gauss’s Mean
Value Theorem itself and rederive it without appealing to complex analysis3 . Let
V = ∇Φ be the vector field of the potential function Φ. The flux of V out of a
(non-infinitesimal) circle C of radius r is then [exercise]

F [V, C] = 2πr Br ⟨Φ⟩.

Thus if Φ is harmonic, so that V is sourceless, then F = 0, by (11.7), p. 547. Since

Br ⟨Φ⟩ = 0, we see that ⟨Φ⟩ is independent of the radius of C. Shrinking C down to
p, we deduce that this radius-independent value must be Φ(p). Done.
·
Now suppose that V is not sourceless, but that its flux-density ∇ V = ∆Φ is con-
stant. Gauss’s Divergence Theorem [(11.7), p. 547] then yields F [V, C] = πr2 ∆Φ.
Inserting this into the previous result, we find that

Br ⟨Φ⟩ = 1
2 r ∆Φ,

which may be integrated to yield the formula in (12.7). To complete the explana-
tion of (12.7) it is only necessary to observe that the Laplacian of an arbitrary Φ is
constant within an infinitesimal circle.
Knowing the meaning of the Laplacian, it is a simple matter to understand its
conformal invariance as expressed in (12.4). The analytic mapping f amplitwists an
e centred at p
infinitesimal circle C centred at p to an infinitesimal circle C e, the new
′ e
radius being er = |f (p)| r. By definition, Φ(e
p) = Φ(p). Likewise, the values of Φ e at
points of Ce are the same as those of Φ at the preimages on C, so ⟨Φ⟩ e on C e equals
⟨Φ⟩ on C. Thus (12.7) implies
e p) = |f ′ (p)|2 r2 ∆Φ(e
r2 ∆Φ(p) = er 2 ∆Φ(e e p),

from which (12.4) follows immediately.

12.3 A Powerful Computational Tool

A zealot might wish for an ideal world in which calculation would always be
relegated to the confirmation of insights provided by geometry. Alas, even this
author must confess to occasional lapses in which calculation has preceded under-
standing! We now describe a powerful computational tool which in many areas of
complex analysis provides a considerable saving of labour. In the next section the
study of the “complex curvature” [cf. Chapter 5] will provide a good showcase for
its simplicity and elegance.

3
Previously we got it from ⟨f⟩ = f(p), which in turn came from Cauchy’s formula.
588 Flows and Harmonic Functions

The gradient operator ∇ of vector calculus acts on a real function R(x, y) to

produce the gradient vector field
! !
Bx Bx R
∇R = R= ,
By By R

and we are free (as we have previously done) to think of this as a complex function

∇R = Bx R + i By R.

From this we may abstract the complex gradient operator ∇, together with the
conjugate operator ∇:

∇ = Bx + i By and ∇ = Bx − i By .

These two operators open the way to an exciting new method of calculation.
Given a vector field
!
u
f= ,
v

we have seen how the real version of ∇ may be formally dotted or crossed with f to
·
yield its divergence ∇ f or its curl ∇×f. The interpretations of these quantities as
flux and work densities shows them to be truly geometric, that is to say, coordinate-
independent. However, there would seem to be no natural way of applying ∇
directly to f to obtain a new vector field ∇f. However, if we replace ∇ by its com-
plex version ∇, and replace the vector field f by the complex function f = u + iv,
then there is a natural definition:

∇f = (Bx + i By ) (u + iv) = (Bx u − By v) + i (Bx v + By u).

The equivalent expression

∇f = ∇u + i ∇v = ∇u + ∇v rotated π
2

helps to see that ∇f is geometrically meaningful (because ∇u and ∇v both are).

The power of the complex gradient derives from the following fundamental
result [exercise]:
A complex function f is analytic if and only if ∇f = 0, in which case we
also have ∇f = 2f ′ .
You may easily verify the following useful properties of ∇:

• ∇(f + g) = ∇f + ∇g.
• ∇(fg) = f∇g + g∇f.
• If f is analytic then ∇f[g(z)] = f ′ [g(z)] ∇g. For example,
∇eg(z) = eg(z) ∇g.
 A Powerful Computational Tool 589

• The concepts of divergence and curl are neatly subsumed by ∇:

·
∇f = ∇ f + i ∇×f.

Similarly,

·
∇f = ∇ f − i ∇×f,

which shows, once again, that a vector field is sourceless and irrotational if and
only if it is the Pólya vector field of an analytic function.
• The Laplacian operator ∆ can be expressed neatly as ∇∇ = ∆ = ∇∇.

In the next section, and in the exercises at the end of the chapter, you will see
the strength of the new technique. You may also find that exercises from previous
chapters are solved more readily by this method. For the moment, here are just two
examples of the use of the complex gradient.
The first is simply to observe how neatly the theorems of Gauss and Stokes
[p. 547] may be combined into a single complex result: If C is the boundary curve
of a simply connected region R then
I ZZ
f dz = i ∇f dA.
R
C

If f is analytic (∇f = 0) we immediately obtain Cauchy’s Theorem. This is not a

new explanation, of course, merely a mathematically streamlined version of our
previous physical one.
Our second example is another derivation of the result (12.4). Let z = x + iy and
w = u + iv, so that the complex gradient operators in the two planes are ∇z =
Bx + i By and ∇w = Bu + i Bv . The result we wish to prove can then be expressed
(less ambiguously than before) as
e
∇z ∇z Φ[f(z)] e
= |f ′ |2 ∇w ∇w Φ(w).

Since w = f(z), a straightforward application of the chain rule yields [exercise]

∇z = f ′ ∇w and ∇z = f ′ ∇w . (12.8)

For example,

∇z f(z) = f ′ ∇w w = f ′ (Bu − i Bv )(u + iv) = 2f ′ ,

as it should. Returning to the problem, we easily obtain

e
∇z ∇z Φ[f(z)] = e
∇z f ′ ∇w Φ(w)
e
= (∇z f ′ ) ∇w Φ(w) e
+ f ′ ∇z ∇w Φ(w)
e
= |f ′ |2 ∇w ∇w Φ(w).
590 Flows and Harmonic Functions

12.4 The Complex Curvature Revisited*

12.4.1 Some Geometry of Harmonic Equipotentials

Given a real function Φ(x, y), figure [12.4] provided a geometric test for harmonic-
ity. However, the first step of the test is not purely geometric in that it uses the
values of Φ, not just the geometry of the curves Φ = const. This leads us to pose
the following more subtle problem. Given a family of curves E filling a region of the
plane, how may we decide whether or not there exists a harmonic function Φ such that E
is its family of equipotentials?
For example, let E be the set of origin-centred circles. If we assign potentials to
these curves according to the rule Φ(z) = |z| then our previous test yields figure
[12.4b], which shows that this potential is not harmonic. But the question we are
now asking is whether these same curves can be assigned potentials according to
a different rule that is harmonic? In fact they can: let Φ(z) = ln |z|.
When the family does admit a harmonic assignment of potential in this way,
we shall simply extend our use of the word “harmonic” to the family of curves
itself. Thus we would say that the family of origin-centred circles is harmonic. The
opening question may then be rephrased succinctly:
What geometric property of E determines whether it is harmonic?
One way to answer this question is to generalize the test in [12.4] to the one in
[12.7]:
(i) Choose two members of E that are very close together.
(ii) Draw line-segments across the region between them [shaded] so as to divide
it into squares.
(iii) Extend these line-segments into streamlines [dashed], i.e., orthogonal trajec-
tories through E.
(iv) Choose one of the resulting flux tubes [darkly shaded] and draw line-
segments across it so as to divide it into squares.
(v) Extend these line-segments into members of E.
Then E is harmonic if and only if the resulting grid is composed of squares.
We already know that a family of concentric circles is harmonic, and [12.7a]
shows how symmetry guarantees that it passes the test. Figure [12.7b] shows that
a family of similar, concentric ellipses is not harmonic.

12.4.2 The Curvature of Harmonic Equipotentials

We now turn to a second, more elegant answer to our question. Let S be the fam-
ily of curves (streamlines) orthogonal to E. When one looks at [12.7] one gets the
 The Complex Curvature Revisited* 591

[12.7] Geometric harmonicity test for a family of curves, E. STEP ONE: Choose two
members of E that are very close together. STEP TWO: Draw line-segments across the
region between them [shaded] so as to divide it into squares. STEP THREE: Extend these
line-segments into streamlines [dashed], i.e., orthogonal trajectories through E. STEP
FOUR: Choose one of the resulting flux tubes [darkly shaded] and draw line-segments
across it so as to divide it into squares. STEP FIVE: Extend these line-segments into mem-
bers of E. Then E is harmonic if and only if the resulting grid is composed of squares.
[a] Concentric circles are harmonic. [b] Concentric ellipses are not harmonic.

feeling that in order for a grid of squares to form, the bending of the curves of E
must be connected in some special way with the bending of the curves of S. By
examining the curvatures of the two types of curve we shall see that this is indeed
the case.
First let us attach directions to the curves of E and S, so that their curvatures will
have well-defined signs. Choose the direction for S arbitrarily, but then define the
direction of E to be that of a tangent vector to S rotated through a positive right angle.
Through any given point w0 there passes one member C1 of S and one member C2
of E. Let the curvatures of C1 and C2 be κ1 and κ2 , and let s1 and s2 be the arc lengths
along C1 and C2 . We then have the following striking result:

E is harmonic if and only if

Bκ1 Bκ2
+ = 0. (12.9)
Bs1 Bs2
In other words, if the rates of change of the curvatures along the two types of
curve are exactly equal and opposite. Note that both these rates of change are well-
defined even in the absence of a choice of direction for the curves, for reversing
such a choice changes the sign of both κ and Bs.
Figure [12.8] illustrates the new test for the concentric circles and ellipses we
considered in [12.7]. Since the circles and lines of [12.8a] have constant curvature,
592 Flows and Harmonic Functions

[12.8] Geometric curvature test for harmonicity: E is harmonic if and only if

Bκ1 Bκ2
Bs1 + Bs2 = 0. [a] Concentric circles are harmonic. [b] Concentric ellipses are not
harmonic.

(12.9) is trivially satisfied. In [12.8b] it is clear that both curvatures are decreasing
at the indicated point, so (12.9) is violated there.
The result (12.9) seems to have been first published by Bivens (1992). We had
also hit upon the same result, but as a natural consequence of investigating our
complex curvature concept, which we introduced in Chapter 5. Here is the pertinent
result:
The complex curvature vector field of an analytic mapping f is
(12.10)
automatically sourceless, and its stream function is Ψ = 1/|f ′ |.

To see the connection between the two results, first observe that a family E in the
w-plane is harmonic if and only if it is the image set of vertical lines in the z-plane
under an analytic mapping w = f(z). For if E is harmonic then its potential function
Φ(w) is the real part of a complex potential z = Ω(w) which maps E to vertical lines,
and so f(z) ≡ Ω−1 (z) has the required property. Conversely, if E is the image set of
vertical lines under an analytic f then it is harmonic, indeed its harmonic potential
is the real part of the complex potential Ω(w) ≡ f−1 (w).
Thus if E is harmonic then the curves C1 and C2 through w0 = f(z0 ) are the images
under f of the horizontal and vertical lines through z0 . But referring to [5.20], p. 272,
we see that curvatures occurring in (12.9) are simply the real and imaginary parts
of the complex curvature of f at z0 :

K(z0 ) = κ1 + iκ2 .

Since infinitesimal horizontal and vertical movements dx and dy are amplified by

|f ′ (z0 )| to movements ds1 and ds2 along C1 and C2 , the flux-density of the complex
 The Complex Curvature Revisited* 593

curvature is
 
·
Bκ1 Bκ2 ′ Bκ1 Bκ2
∇ K= + = |f (z0 )| + .
Bx By Bs1 Bs2
The result (12.10) therefore implies that equation (12.9) is a necessary condition for
E to be harmonic; the question of sufficiency will be addressed shortly.
To prove (12.10) we will use the complex gradient technique of the previous
section; later we will give a proper geometric explanation. We must show that Ψ =
1/|f ′ | is the stream function for the complex curvature

i f ′′
K= ,
f ′ |f ′ |
in other words [cf. (11.24), p. 566], K = −i ∇Ψ.
Since
 
1 i
−i ∇Ψ = −i ∇ = ′ 2 ∇|f ′ |,
|f ′ | |f |
we need to know ∇|f ′ |. Because f is analytic, so is f ′ , and this implies ∇f ′ = 0 and
∇f ′ = 2f ′′ . Hence

2|f ′ | ∇|f ′ | = ∇|f ′ |2 = ∇(f ′ f ′ ) = f ′ ∇f ′ + f ′ ∇f ′ = 2f ′ f ′′

f ′ f ′′
=⇒ ∇|f ′ | = .
|f ′ |
Substituting this into the previous equation, we obtain the desired result:

i f ′ f ′′ i f ′′
−i ∇Ψ = = = K.
|f ′ |3 f ′ |f ′ |
The fact that harmonicity implies (12.9) can also be understood without appeal-
ing to ideas from complex analysis. Let S and E be the streamlines and orthogonal
trajectories of a vector field X. With the present notation, the results (11.5) and (11.6)
from the previous chapter [see p. 544] become

·
B|X| B|X|
∇ X= + κ2 |X| and ∇×X = − + κ1 |X|.
Bs1 Bs2
In order for E (or S) to be harmonic, X must be divergence-free and curl-free, so
B B
κ1 = ln |X| and κ2 = − ln |X|,
Bs2 Bs1
from which (12.9) immediately follows.
We conclude this section by establishing the converse result that equation (12.9)
is a sufficient condition for E to be harmonic. In the w-plane, let Θ(w) be the angle
that the curve C1 through w makes with the horizontal. The angle of the curve C2
through w is therefore Θ + (π/2). Since the curvatures of C1 and C2 are the rates of
594 Flows and Harmonic Functions

change of these two angles with respect to the distances along the curves, we then
have
BΘ BΘ
κ1 = and κ2 = .
Bs1 Bs2
Next we calculate the Laplacian of Θ; the reason will be clear in a moment.
Because the Laplacian is coordinate-independent, we may choose our coordinate
directions tangent to C1 and C2 , obtaining
   
B BΘ B BΘ Bκ1 Bκ2
∆Θ = + = + .
Bs1 Bs1 Bs2 Bs2 Bs1 Bs2
Thus equation (12.9) implies that Θ(w) is harmonic, in which case it is the real
part of an analytic function, say, G(w). We may now define an analytic function
H(w) = e−iG ∝ e−iΘ such that H ∝ eiΘ is everywhere tangent to S. Thus S
and E are the streamlines and equipotentials of the Pólya vector field of an analytic
function. Done.

12.4.3 Further Complex Curvature Calculations

Suppose once again that S and E are the images in the w-plane of horizontal and
vertical lines in the z-plane under an analytic mapping w = f(z). They are then the
streamlines and equipotentials of the vector field Ω ′ in the w-plane, where Ω is the
complex potential z = Ω(w) = f−1 (w) mapping S and E in the w-plane back to
horizontal and vertical lines in the z-plane.
At present, the curvatures κ1 and κ2 at a point w0 = f(z0 ) in the w-plane are
represented as the components of the complex curvature of f at z0 in the z-plane:
Kf (z0 ) = κ1 + iκ2 . [We have added the subscript f because we will shortly be con-
sidering the complex curvature of more than one mapping.] But suppose we think
of the complex potential Ω(w) as fundamental, not merely the inverse of f; how
can we express the curvatures directly in terms of Ω?
We shall answer this question by deriving a remarkably simple relationship
between the complex curvature Kf (z) of f at z, and the complex curvature KΩ (w)
of Ω at the image point w. Since f ′ (z) = 1/Ω ′ (w), equation (12.8) yields

Kf (z) = −i ∇z (1/|f ′ |)
i
= − ′
∇w |Ω ′ |
Ω
|Ω ′ |2
= − ′ [−i ∇w (1/|Ω ′ |)]
Ω
= −Ω ′ (w) KΩ (w). (12.11)

This is interesting. Recall that an infinitesimal complex number ϵ emanating

from z is amplitwisted to yield the image complex number f ′ (z) ϵ emanating from
 The Complex Curvature Revisited* 595

w. Thus, since (12.11) may also be written as

KΩ (w) = −f ′ (z) Kf (z), (12.12)

we see that transforming Kf (z) as if it were an infinitesimal vector yields an image

vector at w which is simply the negative of KΩ (w).
In the next section we shall shed some geometric light on this result, but for the
moment (12.11) yields the desired formula (also known to Bivens (1992)) for the
curvature of the streamlines and equipotentials of a sourceless, irrotational vector
field in terms of its complex potential:

|Ω ′ | Ω ′′
κ1 + iκ2 = −i .
(Ω ′ )2
Next we turn to the curl of K. The complex curvature is the Pólya vector field of
the function (−if ′′ /f ′ |f ′ |), and the presence of |f ′ | in the denominator prevents this
from being analytic. Thus while K has been shown to be divergence-free, it cannot
also be curl-free. What then is its curl?
To find it, recall that

·
∇K = ∇ K + i ∇×K.

Since ∇f ′ = 0 = ∇f ′′ ,
 
i f ′′ f ′′ 1 f ′′  |f ′′ |2
∇K = ∇ = ′ i∇ = K = −i .
f ′ |f ′ | f |f ′ | f′ |f ′ |3

·
Thus ∇ K = 0, which we already knew, and
|f ′′ |2
∇×K = − .
|f ′ |3
Although we have just differentiated K, note that the result does not depend
on any higher derivatives of f than occur in K itself. Indeed, the result may be
re-expressed as

∇×K = −|f ′ | |K|2 . (12.13)

Since K is geometrically defined by f, and since the curl operator is also geomet-
ric, the curl of K must encode some (presumably simple) geometric information
about the mapping f. Unfortunately, we have not yet succeeded in decoding this
information.

12.4.4 Further Geometry of the Complex Curvature

The above results were derived by pure calculation in order to illustrate the
complex gradient technique. We now revert to form and seek more geometrical
explanations, beginning with (12.10).
596 Flows and Harmonic Functions

[12.9] Geometric proof that the complex curvature K has stream function Ψ = 1/|f ′ |.
We know from (5.32) that the streamlines of the complex curvature are the level curves
of the amplification, |f ′ |. It follows readily that −i∇Ψ points in the direction of K. Next,
observe that under the mapping f ′ , the movement dz1 along K is mapped to a movement
tangent to the circle f ′ = const., and therefore dz2 is mapped to a radial movement of
d|f ′ | |f ′′ |
length d|f ′ | = |f ′′ (a) dz2 |. Finally, |∇Ψ| = − |dz
dΨ
2|
= |f1′ |2 |dz 2|
= |f ′ |2 = |K|. Done.

The geometrically derived result (5.32), p. 273, says, in part, that the stream-
lines of the complex curvature are the level curves of the amplification, |f ′ |. Figure [12.9]
illustrates this: sections of adjacent streamlines are mapped by f ′ to arcs of origin-
centred circles. It also shows how the infinitesimal complex number dz1 along the
streamline through the point a is amplitwisted by f ′′ (a) to a complex number at
f ′ (a) that points counterclockwise along the circle |f ′ (z)| = |f ′ (a)| = const. Cor-
respondingly, the orthogonal number dz2 = −i dz1 is amplitwisted to a complex
number at f ′ (a) that points radially outwards.
In order for K to be sourceless it must be of the form K = −i∇Ψ, so that ∇Ψ (the
direction of maximum increase of Ψ) must be directed as shown, and Ψ must be
a function of |f ′ |. Since |f ′ | increases in the direction dz2 , while ∇Ψ is in the oppo-
site direction, Ψ must be a decreasing function of |f ′ |. Thus if Ψ is any decreasing
function of |f ′ | then −i∇Ψ will be a sourceless vector field in the direction of K. It
only remains to show that for the particular function Ψ = 1/|f ′ |, the magnitudes also
agree, i.e., |∇Ψ| = |K|.
Let d|f ′ | and dΨ be the changes in |f ′ | and Ψ that result from the movement dz2 .
From the picture we see that d|f ′ | = |f ′′ (a) dz2 |, so that with Ψ = 1/|f ′ |,

dΨ 1 d|f ′ | |f ′′ |
|∇Ψ| = − = ′2 = ′ 2 = |K|,
|dz2 | |f | |dz2 | |f |

as was to be shown.
 The Complex Curvature Revisited* 597

[12.10] Geometric proof that −KΩ (w) = f ′ (z) Kf (z). The horizontal line-segment at
z is mapped by f to a segment of curvature κ1 , with unit tangent b ξ . The inverse mapping
Ω unbends this piece of curve, so, by the general transformation law of curvature, (5.31),
KΩ ξ·′
b +κ1 /|Ω ′ | = 0. Therefore, the component of −KΩ in the direction of b
′
ξ is κ1 /|Ω ′ | =
κ1 |f |. Likewise, the orthogonal component is κ2 |f |. Thus −KΩ is obtained by expanding
Kf by |f ′ |, and rotating it by arg(bξ ). But arg(b
ξ ) = arg(f ′ ), so −KΩ (w) = f ′ (z) Kf (z).
Done.

Next we give a more geometric derivation of (12.12). Figure [12.10] shows a hor-
izontal line-segment at z being mapped by f to a segment of curve at w whose
curvature is κ1 and whose unit tangent is b
ξ . The inverse mapping Ω unbends this
piece of curve and sends it back to the straight line-segment at z. Hence, by the
general transformation law of curvature, (5.31), page 272,

·
b + κ1 /|Ω ′ | = 0.
KΩ ξ

In other words, as illustrated, the component of −KΩ in the direction of b ξ is

′ ′ ′
κ1 /|Ω | = κ1 |f |. Likewise, the orthogonal component is κ2 |f |.
We now see that −KΩ is obtained by expanding Kf by |f ′ |, and rotating it by
arg(bξ ). But since an infinitesimal real number ϵ at z is amplitwisted to a complex
number ϵf ′ (z) at w which points in the direction of b
ξ , we see that arg(b
ξ ) = arg(f ′ ).
Hence

−KΩ (w) = f ′ (z) Kf (z),

as was to be shown.
It is also possible to give a more geometric derivation of the result (12.13). How-
ever, we shall not bother to do this since we have not yet been able to establish the
significance of that result.
598 Flows and Harmonic Functions

[12.11] Finding the flow around a given obstacle, B. Given a frictionless flow, imagine
that we instantly freeze the shaded region between streamlines, turning it into an obstacle,
B, around which the flow continues, undisturbed. The image of the boundary under the
conformal complex potential is a horizontal line, and, conversely, the problem of finding
the flow around a given B therefore amounts to finding a conformal mapping Ω with this
property.

12.5 Flow Around an Obstacle

12.5.1 Introduction
Consider a typical fluid flow such as [12.11]. From our assumption that the fluid
has no viscosity [this is most certainly an idealization] it follows that if we were to
suddenly freeze the fluid within a flux tube, such as the shaded region in [12.11],
then the unfrozen fluid would continue to flow in exactly the same way as before.
The same idea applies if [12.11] instead represents flowing heat: if the shaded region
of the metal plate were suddenly replaced with material which did not conduct
heat then there would be no disturbance to the flow of heat in the remainder of the
plate. This is much less of an idealization than in the case of fluid flow.
Conversely, if we insert an obstacle into a flow then the new disturbed flow must
be such that the boundary B of this obstacle is a streamline, or is made up of seg-
ments of streamlines. If we think of the complex potential Ω of this disturbed flow
as a mapping, this means that Ω maps B to a horizontal line, or to segments of a
horizontal line.
The problem of finding flows around a given B therefore amounts to finding
conformal mappings Ω with this property. In fact, since the complex potential of
a given flow is only defined up to a constant, we may further demand that the
horizontal image line be the real axis. Alternatively, this characterization may be
 Flow Around an Obstacle 599

[12.12] The flow around a disc can be realized as the previously considered super-
position of a uniform flow with a dipole, with complex potential Ω(z) = z + (1/z).

restated in terms of harmonic functions: we seek harmonic stream functions Ψ

which vanish on B. There are of course infinitely many different flows which satisfy
this requirement; we only obtain a unique solution if we add other requirements,
such as demanding that the flow become uniform far from the obstacle. Indeed,
the superposition of any two flows round B yields a third such flow.

12.5.2 An Example
As our first example of a flow in the presence of a barrier, reconsider the case of
Ω(z) = z+(1/z). See [12.12]. As we previously discussed, this represents a uniform
flow to the right into which has been inserted a dipole at the origin. On the other
hand it certainly appears that the picture can be interpreted in a second way, as
the flow around a circular obstacle (the shaded unit disc) inserted into a uniform
flow. [Later we shall see that it is no accident that the flow may be interpreted in
these two ways.] That both the top and bottom halves of the unit circle are indeed
segments of the streamline Ψ = 0 is easily seen from the fact that Ω maps them to
the segment −2 ⩽ x ⩽ 2 of the real axis:

Ω(eiθ ) = eiθ + e−iθ = 2 cos θ.

Note the breakdown of the grid at ±1 and check that this corresponds to the fact
that these are the stagnation points of the flow.
While this is a possible flow round C that is uniform at infinity, it is not the only
such flow. First let us discuss this fact in terms of the flow round an obstacle of
arbitrary shape.
600 Flows and Harmonic Functions

For any flow round an obstacle, the boundary curve B is composed of stream-
lines, so B has zero flux. If the obstacle has been inserted into a uniform flow then
there are no singularities outside B (except at infinity), and it follows from the
deformation theorem for flux [(11.9), p. 547] that if B is deformed into any loop
round the obstacle then the flux through it continues to vanish. Loops that do
not enclose the obstacle may be contracted to a point without crossing any sin-
gularities, so their fluxes also vanish. Since all loops have vanishing flux, the flow
is not merely locally sourceless, it is totally sourceless. Put differently, this says
that the flux crossing any curve between two points in the flow is independent
of the path.
It is a different story for the circulation/work round B, for there is no a priori
reason for this to vanish. Let S denote the value of this circulation. By the Defor-
mation Theorem (11.10), p. 547, the circulation round any simple loop enclosing
the obstacle will also equal S, while the circulation round loops that do not enclose
the obstacle will vanish. Put differently, this says that the circulation along a path
between two points in the flow is dependent on the choice of that path. The circu-
lation Φ(z) along a path from a fixed point to a variable point z is a multifunction
of z: if the difference of two paths is a loop winding round the obstacle n times, the
difference of the two values of Φ(z) will be nS.
We may now state (without proof) the uniqueness property of flows round a
given obstacle when it is inserted into a uniform flow of given velocity: for each
value of the circulation S there is precisely one flow. In particular, there is a unique
flow that is totally irrotational, i.e., for which S = 0. For a circular obstacle, this is
the one shown in [12.12].
In the case of the disc it is easy enough to construct each of the flows for which
S ̸= 0. We need only superpose the totally irrotational flow in [12.12] with the flow
of a vortex of strength S at the origin. Figure [12.13] illustrates this flow for a small
value of S. The reader is strongly encouraged to use a computer to verify this figure.
As you gradually increase the value of S, notice how the stagnation points on the
circle move towards each other and finally coalesce at i. At what value of S does
this occur? Verify your empirical answer with an exact calculation. Increasing the
value of S still further, this single stagnation point moves off the circle and up the
imaginary axis, and we obtain the qualitatively different flow shown in [12.14].
Returning to the totally irrotational flow in [12.12], figure [12.15] is an attempt
to illustrate the geometry of its complex potential in greater detail than before. The
top half is essentially just a copy of [12.12], while the bottom half illustrates the
image under the complex potential mapping. Though the figure is intended to be
largely self-explanatory, we make the following observations:

• If we choose to measure circulation and flux along paths emanating from p, then
Ω(p) = 0.
 Flow Around an Obstacle 601

[12.13] For each value of the circulation S around an obstacle, there exists a unique
flow. If S = 0 then the flow is given by the totally irrotational flow in the previous figure.
If S ̸= 0, the unique flow is the superposition of the S = 0 flow with an origin-centred
vortex of circulation S. This figure depicts the case of a small value of S.

[12.14] Increasing S eventually results in a qualitatively different flow. Gradually

increasing the value of S from that of the previous figure, the stagnation points on the
unit circle move towards each other and finally coalesce at i. Increasing the value of S
still further, this single stagnation point moves off the circle and up the imaginary axis,
and we obtain the qualitatively different flow shown here.

• Here the flow is totally sourceless and irrotational, so Ω(z) is a single-valued

function of z.
• The paths from p to z1 and z2 both have the same circulation and flux: Ω is
two-to-one on the boundary of the disc.
• However, no two points lying strictly outside the disc have the same Φ and Ψ,
so we have a one-to-one mapping (with inverse Ω−1 ) between the exterior of the
602 Flows and Harmonic Functions

[12.15] A more detailed view of the complex potential mapping of the irrotational
flow [12.12].

disc and the exterior of the line-segment connecting Ω(±1). [The latter region is
often described as a plane with a cut along this segment.]
• The bottom half of the figure illustrates two routes [black/white arrows] by
which the dot-filled shape may be moved to the brick-filled shape. The top illus-
trates their images under Ω−1 (as defined above). Both routes yield the same
image for the brick-filled shape, but the price we pay is that Ω−1 is not even
continuous (let alone analytic) on the cut: witness the fate of the black shape as
we follow the white route. However, consider the following alternative. Again
following the white arrows, sketch what happens if we analytically continue Ω−1
as we cross the cut. Hint: look at the flow inside the unit circle in [12.12].
• The fact that we have ended up with two different images for the brick-filled
shape reflects the fact that the two routes enclose a branch point at Ω(1). This
 Flow Around an Obstacle 603

[12.16] Electrostatic interpretation of the dual of the irrotational flow around the
disc. A cross section of a copper cylinder inserted into a uniform electric field: the former
streamlines now become equipotentials. The surface of such a conductor must itself be
an equipotential, with the static electric field running perpendicular to it, or else charges
would experience a force acting within the surface, causing them to move.

is implied by the existence at ±1 of stagnation points, i.e., critical points of Ω.

Geometrically, this is evident from the doubling of angles at these points, e.g.,
the angle π at −1 is doubled to 2π at Ω(−1).

We now turn to two more physical interpretations of [12.12]. First, we may view
the two-dimensional flow in [12.12] as being a cross section of a genuine three-
dimensional flow. Suppose we were to make a few thousand photocopies of this
figure and stack them neatly one on top of the other. The shaded discs would fit
together to form a cylinder perpendicular to the stack, and the streamlines would
represent the flow round this cylinder. Of course a real cylinder has ends, and,
when it is inserted into a uniform flow perpendicular to its axis, the flow on a plane
that is close to one of these ends will no longer look like [12.12].
To give the second interpretation we must explain a previous remark, namely,
that the dual of a flow round an obstacle is the electric field which solves an anal-
ogous problem in electrostatics. The dual of the flow in [12.12] is shown in [12.16].
Since C was a streamline of the original flow, it is now an equipotential of the dual
flow.
Suppose we insert a long copper cylinder into a uniform electric field perpen-
dicular to its axis. Almost instantly, the free electric charges within the cylinder
will settle themselves into an equilibrium distribution such that the electric field
becomes steady (“electrostatic”). Figure [12.16] then represents a cross section of
604 Flows and Harmonic Functions

this electrostatic field on a plane that cuts perpendicularly through the conductor
somewhere in the middle. [Again, the field near the ends will be different.]
Here’s why. Just as we require Ψ to be constant on the boundary of an obsta-
cle in a fluid flow, so we require Φ to be constant on the boundary of a conductor in an
electrostatic field. This is equivalent to demanding that the electric field be perpen-
dicular to the boundary of the conductor, and it is not hard to see why the latter
must be true. For if the electric field E = ∇Φ were not perpendicular, it would
have a nonzero component within the surface of the conductor. But this means
that the free charges there would experience a force and would move, contradicting
our assumption that the field has settled into an electrostatic one.
The figure also sketches the equilibrium distribution of negative ⊖ and positive
⊕ charge on the surface of the cylinder. This distribution is such that the charge
density is proportional to the (signed) strength of the electric field at the surface. If
a phase portrait has been divided into squares, as this one has, this implies [why?]
that the charge density is proportional to the density of field lines (per unit length)
leaving the conductor’s surface. To learn more of the physics of electrostatics, see
Feynman (1963); for a geometric approach, see Maxwell (1881).

12.5.3 The Method of Images

To illustrate the different kinds of flow which are possible in the presence of one
and the same obstacle, suppose that we instead inserted the unit disc into the flow
of a dipole located at 2, having dipole moment −(1 + i), and thus represented by
the (undisturbed) complex potential
(1 + i)
Ωu (z) = . (12.14)
z−2
Clearly the disturbed flow must look something like [12.17]: the streamlines
behave like those of Ωu both close to 2 and far from the obstacle, and the unit
circle is composed of streamlines. Later we will use the so-called method of images
to show that (modulo artistic error) this figure is in fact the exact flow. To explain
this method we begin with a much simpler example.
With the lower half-plane as the obstacle, consider the flow of a source of
strength 2π located at (2 + i), for which the undisturbed complex potential is

Ωu (z) = log(z − 2 − i).

Clearly the disturbed complex potential Ωd (which we seek) must have stream-
lines which look something like those in [12.18]: they closely resemble those of
Ωu near the singularity, and the barrier is a streamline. That [12.18] is actually the
exact solution follows from the illustrated fact that equipotentials [dashed] may be
drawn so as to divide this flow into squares. We now describe how this Ωd may be
found by the method of images.
 Flow Around an Obstacle 605

[12.17] The flow resulting from inserting the unit disc into a dipole field.

[12.18] The flow of a source in the presence of a barrier along the real axis.

As we saw in Chapter 5, Schwarz’s Symmetry Principle (p. 286) says that a func-
tion which is analytic on one side of the real line, and which takes real values on that
line (e.g., Ωd ), may be analytically continued to the other side by taking conjugate
points to have conjugate images. This is vividly clear in [12.18]: the grid of squares
606 Flows and Harmonic Functions

[12.19] The Method of Images with a linear boundary. To find the flow depicted in
the previous figure, we superpose the original source with its image (of equal strength)
after reflection across the real axis barrier. More generally, if Ωu (z) is the undisturbed
complex potential of a set of multipoles in the upper half-plane, without any barrier, then
Ω⋆u (z) ≡ Ωu (z) represents the undisturbed flow of the mirror-image multipoles in the
lower half-plane. Then the method of images tells us that in the presence of the real axis
barrier, the disturbed flow is the superposition of the original flow and its mirror
image: Ωd (z) = Ωu (z) + Ω⋆u (z).

in the upper half-plane may be continued to the lower half-plane by reflecting the
grid in the real axis. The complete flow is thus given by [12.19].
Intuitively, we have found that the flow in the presence of the barrier may be
obtained by removing the barrier and instead inserting another source of the same
strength as the original, but located at the mirror image in the real axis. Thus Ωd is
the superposition of these two sources:

Ωd (z) = log(z − 2 − i) + log(z − 2 + i).

Use a computer to verify that this formula produces [12.19].

More generally, if Ωu is the undisturbed complex potential of a superposition of
multipoles (sources, vortices, dipoles, etc.) in the upper half-plane, then Ωd will be
 Flow Around an Obstacle 607

the superposition of the undisturbed flow of this set of multipoles and the undis-
turbed flow of their mirror images. Note that while the mirror image multipoles
will be of the same type and strength as the originals, the direction of their mul-
tipole moments will be different. For example, a dipole in the direction (3 − 2i)
reflects to one in the conjugate direction (3 + 2i). More generally, a multipole with
multipole moment Q reflects to one with moment Q.
Now let us turn this method into a formula. In Chapter 5 we showed that
from a given analytic mapping f(z) we can produce a new analytic mapping f⋆ (z)
according to the recipe

f⋆ (z) ≡ f(z). (12.15)

The physical significance of this new analytic function is easy to see: if Ωu again
represents the complex potential of a superposition of multipoles in the upper half-
plane, then Ω⋆u will be the undisturbed complex potential of their mirror images
in the lower half-plane. The formula for the disturbed complex potential is thus

Ωd (z) = Ωu (z) + Ω⋆u (z). (12.16)

Note that this formula does indeed satisfy Schwarz’s Symmetry Principle: Ω⋆d (z) =
Ωd (z), so the real axis is a streamline. Naturally, if Ωu instead represents a collec-
tion of multipoles in the lower half-plane, then (12.16) is again the solution in the
presence of the barrier.
Essentially the same method may be used to find the disturbed flow in [12.17],
in which the barrier is now a circle rather than a line. Reconsider that figure. We
have drawn the streamlines Ψ = const. at random, but had we instead chosen the
values of Ψ in arithmetic progression (as we did in [12.12]) then it would have been
possible to divide this flow into a grid of infinitesimal squares, with the unit circle
being comprised of edges of these squares. As we saw in Chapter 5, it is again
possible to extend this grid across the barrier (cf. [12.12]), but to do so we must
replace reflection in a line by its analogue for circles, namely, inversion.
Performing this inversion in the unit circle we obtain [12.20], the dipole at 2
inverting to another dipole at (1/2). It is now clear that to find Ωd we should
remove the barrier and superpose the undisturbed flows of these two dipoles. But
what is the undisturbed complex potential of this new dipole at (1/2)?
As we saw in Chapter 5, if reflection in the real line is replaced by inversion in
the unit circle then the recipe (12.15) may be modified to generate a new analytic
function f† given by
 
† 1
f (z) ≡ f . (12.17)
z
The physical significance of this new analytic function is much as before: if Ωu
represents the complex potential of a superposition of multipoles outside the unit
608 Flows and Harmonic Functions

[12.20] The Method of Images with a circular boundary: Milne–Thomson Circle

Theorem. To find the flow depicted in [12.17], we superpose the original dipole with
its image under inversion in the unit circle. More generally, if Ωu (z) is the undisturbed
complex potential of a set of multipoles outside the unit circle, without any barrier, then


Ω† (z) ≡ Ω z1 represents the undisturbed flow of the mirror-image multipoles inside the
unit disc, and vice versa. Then the method of images tells us that in the presence of the
unit circle barrier, the disturbed flow is the superposition of the original flow and its
mirror image: Ωd (z) = Ωu (z) + Ω†u (z).

circle, then Ω†u will be the undisturbed complex potential of their images under
inversion. The analogue of (12.16) is now

Ωd (z) = Ωu (z) + Ω†u (z), (12.18)

which automatically satisfies the symmetry requirement Ω†d (z) = Ωd (z), so

that the unit circle is a streamline. This result is known as the Milne–Thomson
Circle Theorem. If Ωu instead represents a collection of multipoles all of which
lie inside the unit circle, then this Ωd is again the solution in the pres-
ence of the barrier. The reason for the emphasis on “all” will be explained
later.
 Flow Around an Obstacle 609

Let us apply this method to find the disturbed complex potential of [12.20] (and
hence of [12.17]). If Ωu is given by (12.14) then
(1 − i) (1 − i)z
Ω†u (z) = = .
(1/z) − 2 1 − 2z
That this is indeed a dipole at (1/2) may be seen by rewriting it as
 
(1 − i) 1 (1 − i)
Ω†u (z) = − − .
4 z − (1/2) 2
Because constants have no effect on the flow, this is a dipole at (1/2) with dipole
moment (1 − i)/4. Unlike the case of reflection across a line, note that it is not only
the direction of the dipole moment which is affected by the inversion, but also its
magnitude: here the strength of the inverted dipole is one quarter the strength of
the original. Superposing the two dipoles we obtain
(1 + i) (1 − i)z
Ωd (z) = Ωu (z) + Ω†u (z) = + ,
z−2 1 − 2z
and you may use a computer to verify that this formula does yield the flow in
[12.20].
Like [12.19], figure [12.20] has symmetry, but it is of a more subtle kind than
before: by construction, the figure reproduces itself under inversion in the unit cir-
cle. This symmetry can be made to leap from the page by projecting [12.20] onto the
Riemann sphere. As we learnt in Chapter 3, inverting in the unit circle is equivalent
to reflecting the Riemann sphere in its equatorial plane. The flow on the northern
and southern hemispheres should therefore be mirror images of each other in this
plane. Behold figure [12.21]! Note that on the sphere the strengths of the two dipoles
become equal.
We can now see that [12.12] is simply a limiting case of [12.21], for as the southern
dipole moves towards the south pole (0), its reflection moves towards the north
pole (∞), and a solitary dipole at ∞ projects to a uniform flow in the plane.
In Chapter 5 we saw that both reflection in a line and inversion in a circle were
special cases of Schwarzian reflection in a curve; see [5.30], page 290. We now use
this fact to generalize the method of images. For example, suppose that we insert
an ellipse E into a uniform flow. Clearly the flow will look something like [12.22],
and that this is in fact the exact flow is once again apparent from its divisibility into
squares. How did we find it?
The generalization of the formulae (12.15) and (12.17) to the case of Schwarzian
reflection RK in a curve K is

f‡ (z) ≡ f [RK (z)].

For example, if K is the real line (so that RK (z) = z) then f‡ = f⋆ , while if K is
the unit circle (so that RK (z) = 1/z) then f‡ = f† . Thus if Ωu is the undisturbed
610 Flows and Harmonic Functions

[12.21] Projection of the previous figure onto the Riemann sphere. By construction,
the previous figure is symmetric under inversion in the unit circle, which induces reflection
of the Riemann sphere in its equatorial plane. So the flow on the northern and southern
hemispheres should therefore be mirror images of each other, and indeed we see that they
are! Note that the unequal strengths of the original dipoles in the plane project to dipoles
of equal strength on the sphere. Also observe that [12.12] may now be understood as a
limiting case of this figure, for as the southern dipole moves towards the south pole (0),
its reflection moves towards the north pole (∞), and a solitary dipole at ∞ projects to a
uniform flow in the plane.

complex potential of a collection of multipoles on one side of K then the disturbed

complex potential with K as barrier will be

Ωd (z) = Ωu (z) + Ω‡u (z).

This formula automatically satisfies Ω‡d = Ωd , so that K is a streamline.

For example, take the ellipse E in [12.22] to have equation (x/2)2 + y2 = 1. In this
case we have [see p. 291]
h p i
RE (z) = 13 5 z − 4 z2 − 3 .

Thus, with Ωu (z) = z,

h p i
4
Ωd (z) = 3 2z − z2 − 3 ,

and you may use a computer to verify that this formula yields figure [12.22].
Note, however, that everything is not quite as it seems. While Ωu is a uniform
 Flow Around an Obstacle 611

[12.22] The Method of Images with a general boundary via Schwarzian reflection.
To find the illustrated flow around the ellipse, E, we superpose a steady eastward flow
with its Schwarzian reflection RE across E; see [5.31]. More generally, if Ωu (z) is the
undisturbed complex potential of a set of multipoles on one side of a curve K, without
any barrier, then Ω‡u (z) ≡ Ωu [RK (z)] represents the undisturbed flow of the Schwarzian
reflections of the multipoles on the other side of K, and vice versa. Then the method of
images tells us that in the presence of the barrier, K, the disturbed flow is the super-
position of the original flow and its mirror image: Ωd (z) = Ωu (z) + Ω‡u (z). This
subsumes both previous methods of images as special cases, for if K is the real axis then
RK (z) = z and f‡ = f⋆ , while if K is the unit circle then RK (z) = 1/z and f‡ = f† .

flow to the right with speed 1, the behaviour of Ωd for large values of |z| is
a uniform flow to the right with speed (4/3). Of course if we wish the flow
to have unit speed far from the ellipse then we need only multiply this Ωd
by (3/4).

12.5.4 Mapping One Flow Onto Another

We previously established the fact that a conformal mapping sends the streamlines
and equipotentials of steady, sourceless, irrotational flow to the streamlines and
equipotentials of another such flow. This idea has many theoretical and practical
uses.
A theoretical benefit is a fresh insight into the very concept of a complex poten-
tial. Reconsider figure [11.21], p. 570. Applying any conformal mapping f to the
flow on the left yields another steady, sourceless, irrotational flow on the right. The
complex potential may now be defined as the special mapping f = Ω for which the
image flow is uniform with velocity 1. For example, what is the complex potential
of a uniform flow with velocity 1? From the new point of view it is the conformal
mapping which sends this flow to a uniform flow with velocity 1. Thus it is the
identity mapping Ω(z) = z, as it should be!
612 Flows and Harmonic Functions

[12.23] Conformally mapping concentric circles to confocal ellipses. Recall from

[5.22], on page 276, that as z = eit describes the unit circle, C, w = p eit + q e−it
describes the ellipse (x/a)2 + (y/b)2 = 1, where p = (a + b)/2 and q = (a − b)/2. Thus,
2
in the specific example of (x/2)

+ y2 = 1, the conformal mapping we seek of C to E is
1 1
z ÞÑ w = f(z) = 2 3z + z . Circles concentric to C are mapped to ellipses confocal to
E. Far from C, the mapping behaves like 32 z.

To demonstrate the practicality of mapping one flow onto another, let us return
to the problem of finding the flow round an obstacle, restricting our attention to the
case where the obstacle is inserted into a uniform flow. For example, let us rederive
the flow round the ellipse (x/2)2 + y2 = 1 in [12.22]. Suppose that we knew of a
one-to-one conformal mapping w = f(z) from the exterior of the unit circle C in
the z-plane to the exterior of this ellipse E in the w-plane. We already know the
flow round C when it is inserted into a uniform flow of velocity 1, so applying f
to these streamlines yields some flow round E. If we want the flow round E to be
uniform far from E then we must further demand that f(z) behave like a multiple
of the identity far from C: f(z) ≈ cz if |z| is large. Assigning equal values of Ψ to
the original and image streamlines, the image flow far from E will then be uniform,
with speed (1/|c|) and direction c [why?].
Recall figure [5.22], p. 276. As z = eit describes C, w = p eit + q e−it describes
the ellipse (x/a)2 + (y/b)2 = 1, where p = (a + b)/2 and q = (a − b)/2. Thus in
the case of (x/2)2 + y2 = 1, the mapping we seek is

1 1


w = f(z) = 2 3z + z . (12.19)

As illustrated in [12.23], C is mapped to E and circles concentric to C are mapped

to ellipses confocal to E. Far from C the mapping behaves like (3/2)z: you can see
that as the circles on the left grow, their images approach circles that are (3/2) as
big. Thus f should map the flow round C to the flow that results from inserting E
into a uniform flow of velocity (2/3).
 The Physics of Riemann’s Mapping Theorem 613

Let’s go through the details and check that we recover the flow in [12.22]. Since
the flow round C has complex potential Ω(z) = z + (1/z), and since the complex
e
potential Ω(w) at the image w = f(z) of z is defined to be the same as Ω at z, we
e
have Ω(w) = z + (1/z). To express this as an explicit function of w we solve (12.19)
for z and obtain [why the choice of +?]
 p 
z = 31 w + w2 − 3 .

e
Although we could immediately insert this into the formula for Ω(w), we may save
ourselves a little algebra by first noting that (12.19) implies (1/z) = 2w − 3z. Thus
 p 
e
Ω(w) = z + (1/z) = 2(w − z) = 32 2w − w2 − 3 .

Apart from the factor of (2/3), signifying that the velocity is (2/3) far from E
(as anticipated), this is the same formula we previously obtained by Schwarzian
reflection.
As another illustration of this idea, suppose that E were instead inserted into a
uniform flow in the direction of eiϕ . To find the flow round E we need only find
the flow round C when it is inserted into such a flow, and then apply f. Since the
undisturbed complex potential is Ωu (z) = e−iϕ z, the method of images says that
the flow round C is
eiϕ
Ωd (z) = Ωu (z) + Ω†u (z) = e−iϕ z + .
z
With ϕ = (π/4) this flow is illustrated on the left of [12.24]. On the right is
the desired flow round E obtained by applying f to the flow round C. Use your
computer to verify this figure.
In this manner, we may derive the flow round an infinite variety of obstacles:
choose any analytic mapping w = f(z) which is one-to-one outside C and which
e
behaves like cz for large |z|, then Ω(w) = f−1 (w) + [1/f−1 (w)] is the flow round an
obstacle whose boundary curve B is f(C).

12.6 The Physics of Riemann’s Mapping Theorem

12.6.1 Introduction
Recall that Riemann’s Mapping Theorem [p. 204] asserts that any simply connected
region R (other than the entire plane) may be mapped one-to-one and conformally
to any other such region S. Granted this, we saw that no loss of generality results
from taking S to be the unit disc D, and that there must be as many mappings from
R to S as there are automorphisms of D. Later [p. 407] we showed that these auto-
morphisms are the hyperbolic rigid motions, which have three degrees of freedom.
614 Flows and Harmonic Functions

[12.24] Conformally mapping one flow onto another. Using the mapping obtained in
the previous figure, the known flow on the left around the unit disc may be conformally
mapped to the flow around the ellipse on the right. [NOTE: In the text, we confirm that
this new approach yields the same answer that we formerly obtained via the method of
images.] In this manner, we may derive the flow round an infinite variety of obstacles:
choose any analytic mapping w = f(z) which is one-to-one outside C and which behaves
e
like cz for large |z|, so that the flow becomes uniform, then Ω(w) = f−1 (w) + [1/f−1 (w)]
is the flow round an obstacle whose boundary curve B is f(C).

Thus the complete result we seek to understand is this: there exists a three-parameter
family of one-to-one, conformal mappings between R and D.
There are at least two standard proofs of this fundamental result, and most
advanced books on complex analysis include one of these. Despite the fact that
one of the arguments (due to Koebe) is constructive in nature—and therefore, in
principle, comprehensible—we have not yet found a way to present it in a manner
consistent with the aims of this book. However, the interested reader will find an
excellent description of the idea underlying Koebe’s proof in Hilbert (1952), and a
clear description of the technical details in Nehari (1952) or Nevanlinna and Paatero
(2007). To our knowledge, the deepest investigation of this idea is that given by
Henrici (1991).
On a brighter note, the above ideas on flows will enable us to gain considerable
insight into both the existence of such mappings and the fact that they have three
degrees of freedom. We shall do so by reversing the idea of conformally mapping
one flow onto another. That is, by resorting to physical experiment one may obtain
flows, and these may then be used to construct conformal mappings.
 The Physics of Riemann’s Mapping Theorem 615

12.6.2 Exterior Mappings and Flows Round Obstacles

Consider [12.25]. The top part depicts an obstacle R that has been inserted into
uniform flow having velocity 1. If we do not introduce any circulation round R

[12.25] Constraint on the mapping between flows. Divide the flow around R and the
flow around the unit disc D into small, square k-cells. If the illustrated point z is mapped
by f to w = f(z), then if z moves four squares downstream and two squares “up” the
equipotentials, w must do the same. But the sought-after conformal mapping w = f(z)
has a constraint: it must map the stagnation points a and b to the stagnation points a e
e This is because the geometric signature of the critical point b is the angle of (π/4)
and b.
between the streamline and equipotential through b, and this property is preserved by a
conformal mapping.
616 Flows and Harmonic Functions

then, as we have previously remarked, the flow is unique. Next, we have arbitrarily
chosen a point [not shown] from which to measure circulation and flux, i.e., a point
at which the potential Φ and stream function Ψ both vanish. With a small value of k,
we have then constructed the k-flux tubes and k-work tubes, thereby dividing the
exterior of R into small, approximately square, k-cells. As usual, we may imagine
shrinking the value of k to zero, so that ultimately the k-cells are square.
As illustrated, let a and b be the two stagnation points on the boundary of R
(ordered so that the flow passes from a to b), and let S1 and S2 be the two segments
of boundary streamline connecting them. Since the flow is totally irrotational, the
circulations along S1 and S2 must be equal to each other, the common value being
the potential difference [Φ] between a and b:

[Φ] = Φ(b) − Φ(a).

Put geometrically, this says that the number of squares abutting S1 and S2 must be
equal to each other, this number being given by [Φ]/k.
Now insert the unit disc D into a uniform flow with velocity v eiϕ and divide up
the resulting flow into k-cells, using the same value of k as before. Let us employ
e
tildes to denote corresponding entities in the new flow: the stagnation points are a
e e e
and b, the segments of streamline connecting them are S1 and S2 , and the potential
e and Ψ.
and stream functions (relative to an arbitrarily chosen point) are Φ e
The method of conformally mapping points z in the exterior of R to points w =
f(z) in the exterior of D now seems clear: identify “corresponding” squares in the
two grids! Let us control our excitement and think this through. Presumably, by
“corresponding” we mean that once we know the image w of one grid point z then
the image of any other is determined by the pair of grids: if z moves four squares
downstream and two squares “up” the equipotentials, then w does the same. See
[12.25].
However, we cannot arbitrarily choose to map a particular z to a particular w,
for the sought-after conformal mapping w = f(z) must map the stagnation points a and b
e This is because the geometric signature of the critical
e and b.
to the stagnation points a
point b is the angle of (π/4) between the streamline and equipotential through b,
and this property is preserved by a conformal mapping.
Now the snag is this: defining the image of a to be a e if
e (as we must), b will map to b
e
and only if the number of squares abutting S1 is equal to the number abutting S1 . Looking
closely, we see this is not the case in [12.25]: there are fewer squares along S e1 than S1 .
To correct the situation we must slightly increase the speed v of the flow round the
disc. More precisely, v must be chosen so that the potential difference [Φ] e across
D equals the potential difference [Φ] across R. Since we know that [Φ] e = 4v, we
deduce that
A one-to-one, conformal mapping between the two grids is obtained if
and only if the disc is inserted into a flow of speed v = [Φ]/4.
 The Physics of Riemann’s Mapping Theorem 617

Although we hope our use of grids has helped to make the mapping vivid (and
to show that it is both one-to-one and conformal), we should perhaps point out that
the grids are in no way essential to the definition of f. All we are doing is identifying
points by means of circulation and flux: the image w = f(z) of z is defined by

e
Ω(w) e a
= Ω(z) + Ω( e ) − Ω(a) ,

the constant having been chosen so as to ensure that a e = f(a).

Since the choice of the zero point of flux and circulation in each of the flows
has no bearing on the construction of f, we may choose these two points to our
advantage. Henceforth, we will take the zero point in the image flow to be the image of
the zero point in the original flow. In [12.25], for example, this convention says that if
we choose the zero point for the flow round R to be at a (i.e., Ω(a) = 0), then we
must take the zero point for the flow round D to be at a e a
e (i.e., Ω( e ) = 0). The above
equation then simplifies to
e
Ω(w) = Ω(z).

In other words, w and z correspond if and only if their circulation and flux are
equal.
In this case the mapping may be written as f = Ω e −1 ◦ Ω, which we may interpret
e
as follows. Refer to [12.25] and [12.15]. The complex potentials Ω(z) and Ω(w) map
points lying strictly outside R and, respectively, D to points in a plane that is slit
along the real axis from 0 to [Φ]. Thus f may be thought of as first mapping the
exterior of R to the slit plane by means of Ω, then mapping the slit plane to the
exterior of D by means of Ω e −1 .
We now return to the mappings themselves and ask, how “many” of them do we
obtain by means of this construction? We begin by explaining the illusory nature
of some of the apparent freedoms in the construction. First, why not insert R into a
uniform flow of arbitrary speed? Of course we can, but nothing new results from
doing so. For example, if we double this speed then we must also double v, because
we must maintain equality between the number of k-cells along S1 and S e1 . But this
yields the same mapping f as before. Second, in constructing the grid outside D,
why not use a different value of k, say e k, from that used outside R? In order to
maintain equality between the number of squares along S1 and S e1 we would then
e
have to change v to k[Φ]/4k [why?], and this would produce the same grid and
mapping as before.
Clearly, however, we do obtain new mappings by varying the direction ϕ of the
flow round D. If F denotes the particular mapping f corresponding to ϕ = 0, then
the mapping Fϕ corresponding to a general value of ϕ is Fϕ (z) = eiϕ F(z), namely,
F followed by a rotation.
618 Flows and Harmonic Functions

It would seem plausible that still other mappings could be obtained by varying
the direction—let us call it θ—of the flow into which R is inserted. Not so. This fol-
lows from the fact that the flow round R in this case may be obtained by applying
F−1 to the flow round the disc when it is inserted into a uniform flow in the direction
θ. Clarify this in your own mind by referring to figures [12.23] and [12.24]. Thus the
mapping between the two flows is only sensitive to the angle between their direc-
tions: inserting R and C into flows with directions θ and ϕ, respectively, yields the
mapping F(ϕ−θ) .
Since ϕ is the only genuine degree of freedom, the mappings we have con-
structed belong to a one-parameter family. We are therefore missing two degrees
of freedom. Though we will explain this mystery shortly, you may care to think
about it on your own before reading further.
Granted the generosity of Nature in providing such flows round obstacles, we
now have a physical method of determining f, but we lack a mathematical proce-
dure for doing so. This is a very hard problem. However, there does exist an explicit
formula for f in the case that R is bounded by a polygon. This is called the Schwarz-
Christoffel formula; see Nehari (1952) or Pólya and Latta (1974) for good discussions
of the result. We shall not enter into this here, except to say that the advent of high-
speed computers has opened the way to approximating f by approximating the
boundary of a given R with a polygon; see Trefethen (1986) and Henrici (1991) for
this and other algorithmic approaches to the problem.

12.6.3 Interior Mappings and Dipoles

In [12.25], suppose that the speed of the flow round D has been adjusted to [Φ]/4,
yielding a conformal mapping between the exteriors of R and D. If we now perform
inversion in the boundary of D, we obtain the flow [12.26b] inside the unit circle
C. This is the familiar flow of a dipole inserted into the centre of a circular pool of
fluid.
Similarly, if we choose the arbitrarily selected interior point q of R shown in the
previous figure, then perform an inversion4 in a circle of unit radius centred there,
we obtain [12.26a]. Use a computer, or your Peaucellier linkage [see p. 205], to verify
that the boundary of R does indeed invert to the illustrated curve R† . In [12.25], the
streamlines far from R are parallel lines, and the inversion therefore sends these to
small, mutually tangent circles through q. Thus, as illustrated, the inverted flow in
[12.26a] represents a dipole which has been inserted at q into the R† -shaped pool

4
Ideas related to those which now follow may be found in Bak and Newman (2010), Siegel (1969,
p. 148), and especially Courant (1950). Although Riemann himself employed physical reasoning, the
idea of relating his mapping theorem to dipoles seems to have originated with Hilbert (1965, Vol. 3,
pp. 73–80).
 The Physics of Riemann’s Mapping Theorem 619

[12.26] Interior mappings and dipoles. Referring to the previous figure, inversion of R
in a circle centred at the arbitrarily selected interior point q yields [a], while inversion of D
in the unit circle yields [b]. In both cases, we obtain dipoles. In this way, the one-to-one,
conformal mapping between the exteriors of R and C now yields another such mapping
between the interior dipole flows of R† and C: for example, the black T-shape is mapped
to the black T-shape.

of fluid. In fact [exercise] since the original uniform flow had unit speed, the dipole
at q has unit strength.
In this way, the one-to-one, conformal mapping between the exteriors of R
and C now yields another such mapping between the interiors of R† and C. For
example, the stagnation point α maps to the stagnation point α e , q maps to 0, and
†
the black T-shape inside R maps to the one inside C. Although the speed of the
flow grows arbitrarily large as we approach q, there is nothing dramatic about the
behaviour of the mapping near q. For example, imagine sliding the T-shape along
the streamlines towards q, and consider its image.
We may look at this construction rather differently. Suppose that R† is a given,
fixed curve whose interior we wish to map to the unit disc. We may obtain such a
mapping f as follows:

(i) Insert a horizontal dipole of unit strength into the interior of R† at an arbi-
trary interior point p. Divide the interior into k-cells of the flow, and let N
denote the number of them abutting R† .
(ii) At the centre 0 of C, insert a dipole of strength d, and direction ϕ. Divide the
interior of C into the k-cells of this flow, and adjust the strength d until the
number of them abutting C is equal to N. If [Φ] again denotes the circulation
620 Flows and Harmonic Functions

along a segment of boundary streamline connecting the stagnation points α

and β, this condition is met by d = [Φ]/4.
(iii) Except for the fact that the dipole inside R† is now located at a general point p
(instead of the particular point q), we are now back to the situation depicted
in [12.26], and the mapping f is completely determined.

This one-to-one, conformal mapping f from the interior of R† to the unit disc
is, in fact, the most general such mapping. This may be seen from the fact that
the construction has the full three degrees of freedom: two for the location p of the
dipole in R† , and one for the direction ϕ of the dipole in C. It is not hard to see
the geometric significance for f of these physical degrees of freedom: p is the point
which is mapped to the centre of the disc, i.e., f(p) = 0; and ϕ is clearly the twist
of f at p, i.e., ϕ = arg[f ′ (p)].
Note that while the twist of f at p is freely specifiable, the amplification |f ′ (p)| is
not. In fact [exercise, or read on] |f ′ (p)| is simply the strength d of the dipole in C,
and we have seen how the latter is fixed in the course of the above construction.
Put differently, |f ′ (p)| = [Φ]/4.
We can now return to the construction in [12.25] and see why it is missing two
degrees of freedom. To obtain the general mapping between the exteriors of R and
D, we may take this newly constructed general mapping between the interiors of
R† and C, then reverse (i.e., repeat) the inversions which took us from [12.25] to
[12.26].
In the above process of generalization the fundamental step was moving the
dipole at q to an arbitrary interior point p. Since the flow in [12.26b] did not undergo
any fundamental change, inversion returns it to a flow round D like that shown in
[12.25]. However, to send R† back to the boundary of R we must invert in the circle
of unit radius centred at q (which we now take to be the origin), and the flow inside
R† of the dipole at p inverts to the flow outside R of a dipole placed at (1/p). Figure [12.17]
illustrates such a flow when R is a disc.
The construction in [12.25] is now recognizable as the special p = q in which the
dipole outside R has been placed at ∞ rather than at a general point. Here are our
two missing degrees of freedom: there was nothing wrong with choosing to place
one of the dipoles (the one outside D) at ∞, but we then (unnecessarily) insisted
on placing the other dipole there too. In terms of the resulting mapping between
the exteriors, this amounted to insisting that ∞ map to ∞, whereas the general
construction allows us to map any point outside R to ∞ by placing a dipole there.

12.6.4 Interior Mappings, Vortices, and Sources

Having had the inspiration to construct mappings by means of multipoles, it is
natural to wonder if we may simplify this dipole method by employing the most
 The Physics of Riemann’s Mapping Theorem 621

primitive of all multipoles, the source. However, if we continue to think in terms of

fluid flows then this idea cannot work. For if we place a source inside the region we
wish to map, the fluid flowing out of the source and into the region has nowhere
to go! The only way out is to also insert a sink of equal strength5 , thereby creating
a doublet. But this is really no improvement over our earlier dipole construction,
for a dipole is merely a limiting form of a doublet.
However, as we shall explain shortly, the failure of this attempted use of
sources is merely a consequence of our thinking in terms of fluid flowing within a
strangely-shaped pool; by thinking in terms of electric fields or heat flows, we can
use sources to construct mappings.
Continuing with the fluid interpretation for the time being, there does exist a
simple alternative to the dipole method, but in place of sources we must use vor-
tices. Let B be the boundary curve of the simply connected region which we wish
to map to the interior of the unit circle C. To construct the mapping, insert a vortex
into B at an arbitrary interior point p, and insert another vortex into C at 0. See
[12.27]. Dividing each flow into k-cells, as before, conformal mappings leap from
the page. Let’s examine the details.

[12.27] Mapping interiors with vortices instead of dipoles. Place a source of strength
S at p inside B, and a source of strength S e at the centre of C. Dividing each flow into
small, square k-cells, we readily obtain a conformal mapping from one interior to the
other, with p mapping to 0. But this is only possible if the number N of k-cells abutting
B equals the number N e abutting C. This implies that the two vortices must be of equal
strength: S = Nk = Nk e = S. e Finally, we may specify which boundary point a e of C
corresponds to the boundary point a of B. Now the mapping is completely tied down,
and, for example, the black T-shape is mapped to the black T-shape.

5
Of course we could insert several sinks whose strengths summed to that of the source, but that
would be messier still.
622 Flows and Harmonic Functions

Firstly, in defining the correspondence between the grids we are forced to map
p to 0. Secondly, having chosen the strength S of the vortex in B, the strength S e of
the one in C must be chosen so that (as illustrated) the number N e of k-cells abut-
ting C is equal to the number N abutting B. This is the same idea as in the dipole
construction, but the answer in this case is much simpler: put S e equal to S.
The explanation is as simple as the answer. Concentrate on one of the vortices,
say the one in B. Its strength S is, by definition, the circulation along any simple
loop round p, and we may conveniently take this to be one of the streamlines. The
equipotentials cut this streamline into N segments, each having circulation k. Thus
S = Nk, and likewise S e = Nk.
e The condition S e = S follows immediately.
Finally, observe that we have not yet pinned down the mapping. In both [12.25]
and [12.26] there were stagnation points which we were forced to identify, thereby
tying down the mapping. However, here there are none, and we must do the job by
hand. A common procedure goes like this. We know that the streamline B maps to
the streamline C, and we may now insist that a particular point a on the first maps
to a particular point a e on the second.
If we wished to be more definite, we might do the following. Consider the
heavily-dashed equipotential in B which exits p travelling due east, and choose
a to be its intersection with B. See [12.27]. If we now choose a e = eiϕ then the
mapping f between the two grids is completely tied down. For example, the black
T-shape in B maps to the black T-shape in C. The only advantage of specifying a
and a e in this particular way is that ϕ then has a simple interpretation in terms of f,
for it is clearly the twist of f at p: ϕ = arg[f ′ (p)].
As with the dipole method, we have obtained the full three-parameter family of
mappings: two for the point p that f maps to the centre of C, and one for the twist
of f at p.
Now let us turn to other physical interpretations of this construction. Taking
the dual of the flows in [12.27], we obtain [12.28]. The equipotentials have become
streamlines, and the streamlines (B and C in particular) have become equipoten-
tials. By the same reasoning as before, the strengths of the two sources manifest
themselves geometrically as k times the number of streamlines emanating from
each, and these must be set equal to each other in order to construct the conformal
mapping between the two grids.
Let us digress briefly. Since the strength of a source is measured by the number
of k-flux tubes emanating from it, a conformal mapping sends a source to another
source of equal strength. Obviously, the same goes for a sink. Returning to [12.26],
we can understand why the strength d of the dipole inside C is the amplification
of f at p. The dipole of unit strength at p may be thought of as a doublet in which
the source and sink are an infinitesimal distance ϵ apart, and each has strength
(1/ϵ). The conformal mapping f sends this to another doublet at 0: the strengths of
 The Physics of Riemann’s Mapping Theorem 623

[12.28] The source construction dual to the vortex construction of the previous
figure. In order to make physical sense, here we must replace the fluid flow with an
electrostatic field, the sources being electric charges. By the same reasoning as before,
the strengths of the two charges manifest themselves geometrically as k times the number
of electric field lines emanating from each, and these must be set equal to each other in
order to construct the one-to-one conformal mapping between the two grids.

the source and sink are preserved, but their separation (and hence the strength of
dipole) is amplified by |f ′ (p)|, as was to be shown.
Returning to the physical interpretation of [12.28], we have already observed
that such pictures of sources make no sense when thought of as fluid flows. How-
ever, they do make perfect sense when thought of as electrostatic fields. Imagine
that the dark regions in [12.28] represent cross-sections of blocks of copper through
which we have bored “cylindrical” holes with cross-sections given by B and C.
Now imagine that p and 0 are the cross-sections of two long, very thin, uniformly
(and equally) charged wires running down these holes. The electrostatic fields gen-
erated by these wires automatically have B and C as equipotentials, so [12.28]
faithfully represents these fields, with the streamlines being the electric field lines.
We may also interpret [12.28] in terms of heat flows. Imagine that the white
shapes bounded by B and C have been cut from a sheet of heat-conducting metal,
and imagine that the dark regions are filled with ice, thereby maintaining B and C
at constant temperature (i.e., potential). If heat is introduced at a steady rate at p
and at 0 then the flow of heat will eventually settle down to the one in [12.28], with
the dashed equipotentials being the isotherms.
It must be observed, however, that a point source of heat is a much less physical
concept than its electrostatic analogue, which may be realized (as we have said)
by a very thin charged wire. The reason is that the potential function Φ becomes
624 Flows and Harmonic Functions

[12.29] Using flows to construct the most general automorphism of the disc. Return-
ing to [12.27], suppose we choose B = C to be the unit circle. Then the vortex flow shown
here in [a] may be viewed as a one-to-one conformal mapping to the flow in [b], i.e., it is
an automorphism of the disc. Recall that we previously used the Symmetry Principle to
find the formula for the most general such automorphism: (3.52), on page 201. We can
now rederive this formula by finding the flow in [a], which we accomplish by means of
the method of images, in the next figure.

arbitrarily large as we approach a source. In electrostatics this does not present

any difficulties, but in the heat-flow interpretation −Φ represents temperature,
so the metal in the vicinity of such an imagined source of heat would vaporize!
For an excellent discussion of such physical distinctions, see Maxwell (1881), p. 51
onwards.
While many readers may be unfamiliar with electrostatics, few will be unfamiliar
with heat. Thus, ignoring the above objection, we shall persist in expressing ideas
mainly in the language of heat rather than electrostatics.

12.6.5 An Example: Automorphisms of the Disc

Let us explicitly carry out the construction in [12.27] for the case B = C, and thereby
re-obtain the automorphisms of the unit disc from a fresh point of view.
Figure [12.29] illustrates the construction in this case. Previously we did not
specify the strength S of the two vortices because the choice had no effect on the
resulting mapping, but let us now choose S = −2π. Taking the zero of flux and
circulation in [12.29b] to be at 1, the complex potential for [12.29b] is then
e
Ω(w) = i log w.
Now consider the flow in [12.29a]. We may find its complex potential by the
method of images6 . That is, we superpose the real vortex at p with its reflection

6
However, note that we are not using (12.18). See Ex. 14 to understand why.
 The Physics of Riemann’s Mapping Theorem 625

[12.30] Using the method of images to construct the most general automorphism of
the disc. To obtain the flow in [12.29a], we superpose a vortex at p with an equal and
opposite vortex at its reflection, 1/p.

in C, namely, a fictitious vortex of equal and opposite strength at (1/p). See [12.30].
The complex potential of [12.29a] is thus

Ω(z) = i log(z − p) − i log(z − 1/p) − γ

 
z−p
= i log −δ (12.20)
pz − 1
where γ and δ are constants.
In [12.29b] we chose to measure circulation and flux from a boundary point, and
we now choose to do the same in [12.29a]. In terms of the above equation this is
equivalent [exercise] to demanding that the constant δ be an arbitrary real number.
To pin down the mapping we must now choose a and a e. Instead of doing this
as we did in [12.27], let us this time choose them to be the two boundary points
from which we have elected to measure circulation and flux in the two flows. As
previously discussed, the mapping is then effected by equating circulation and flux:
e
Ω(w) = Ω(z).

Solving for w, we do indeed recover the familiar automorphisms of the disc:

 
iδ z−p
w = f(z) = e . (12.21)
pz − 1
626 Flows and Harmonic Functions

Of course we have not done anything really new, for we have used the method of
images, and this is merely a disguised form of the Symmetry Principle by means of
which the result was originally obtained. Nevertheless, we hope you have found it
instructive—and delightful!—to be able to understand these automorphisms from
a new, physical point of view.

12.6.6 Green’s Function

We now return to [12.28] and to the heat-flow approach to the construction of
conformal mappings.
Figure [12.31] is essentially a copy of [12.28], but with a few added details which
we will need shortly. We supply heat at the constant rate 2π to the point p of the
region R while holding the temperature all around the boundary B at the constant
value 0. After the heat flow has settled down, the temperature in R will be a well
defined (except at p) harmonic function Gp (z) called the Green’s function of R with
pole at p. Note the new, special sense of the word “pole”.
As an example, let us obtain the Green’s function for the unit disc. In [12.29a]
we considered the flow of a vortex of strength −2π inserted at p. And because we
chose a point on C as the zero of circulation and flux, the stream function vanished
on the boundary streamline. The dual of this flow is therefore precisely what we are
after: a source of strength 2π at p, with the boundary at zero potential. By (12.20),
the complex potential is
 
z−p
Ω(z) = Φ(z) + i Ψ(z) = log + iδ,
pz − 1
where δ is a real constant. Thus the temperature in the disc is given by
z−p
Gp (z) = −Φ(z) = − ln . (12.22)
pz − 1
Note that Gp (z) = − ln |f(z)|, where f(z) is any of the one-parameter family of
mappings to the unit disc such that f(p) = 0. We shall see that this is true quite
generally.
Also note that this Green’s function has a very interesting symmetry property:
Gp (q) = Gq (p).
Thus, with the boundary packed with ice, the steady-state temperature at q due to
a point source of heat at p is the same as the temperature at p when the source is
instead at q. Remarkably, we shall see that this symmetry holds true for the Green’s
function in a region of arbitrary shape!
The Green’s function is a powerful tool in several areas of mathematics and
physics. For the time being, we will concern ourselves primarily with its relation-
ship to the conformal mappings w = f(z) from the interior of B to the unit disc. In
the next section we shall discuss another important application.
 The Physics of Riemann’s Mapping Theorem 627

[12.31] The Green’s Function Gp (z) of a region R is defined to be the harmonic equi-
librium temperature distribution when heat is supplied at constant rate 2π to the point p
while holding the temperature all around the boundary B at the constant value 0. We may
then construct a harmonic dual Hp (z), that is, a harmonic function whose level curves are
the paths along which the heat flows (orthogonally through the isotherms Gp = const.)
from p to the boundary. Thus knowledge of Gp is sufficient for the construction of the
complete complex potential, Ω(z) = −[Gp (z) + i Hp (z)], and the conformal mapping of
R to D is then explicitly given by w = f(z) = eΩ(z) = e−G e−iH .

Returning to the general case in [12.31], suppose Gp is known. As previously

described, we may then construct a harmonic dual Hp (z), that is, a harmonic func-
tion whose level curves are the paths along which the heat flows (orthogonally
through the isotherms Gp = const.) from p to the boundary. Thus knowledge
of Gp is sufficient for the construction of the complete complex potential, Ω(z) =
−[Gp (z) + i Hp (z)].
Consider the behaviour of Gp in the immediate vicinity of p. Physical intuition
leads us to expect, irrespective of the temperatures assigned to B, that the flow out
of p will almost be like that of an isolated source, so that Ω(z) ≈ log(z − p). Thus
if z = p + ρ eiθ then

Gp (z) ≈ − ln ρ, (12.23)

for small values of ρ.

Similarly, Hp (z) ≈ −(θ + ϕ), where ϕ is a constant. The freedom in choosing ϕ
is equivalent, as we shall see, to the freedom of choosing which point of B maps to
a particular point of C. If we choose ϕ = 0 then the value of Hp at a typical point
q has a particularly simple interpretation. See [12.31]. Following the flow of heat
back from q to p, the angle at which it enters p is −Hp (q).
628 Flows and Harmonic Functions

Returning to (12.23), the precise version is that Gp differs from − ln(ρ) by a

function gp (z) which is harmonic throughout R:

Gp (z) = − ln ρ + gp (z).

Since Gp vanishes on the boundary, the values of gp on B are determined directly

by the shape of B and the location of p within it:

gp (z) = ln ρ.

The problem of constructing Gp is therefore equivalent to the problem of finding

the function gp that takes these values on B and is harmonic throughout the inte-
rior. This is an example of the type of “Dirichlet problem” to be discussed in the
next section. The solution of this problem also gives us Hp , for we may construct
a harmonic dual hp of gp , and then Hp = −θ + hp .
To construct a conformal mapping w = f(z) to the unit disc D such that f(p) = 0,
we may (as previously explained) equate the complex potential Ω(z) of the flow
e
inside R with the complex potential Ω(w) = log w in D of the heat source at 0, with
the boundary at zero temperature. Thus

w = f(z) = eΩ(z) = e−G e−iH . (12.24)

As illustrated in [12.31], the dashed isotherm at temperature G(q) is mapped to the

dashed circle of radius e−G(q) , and the streamline H = H(q) = −θ entering p at
angle θ maps to the ray entering 0 at angle θ. Thus f has zero twist at p. This is the
significance for f of our previously choosing ϕ = 0; in general arg[f ′ (p)] = ϕ.
Now that we possess the mapping f, any harmonic temperature distribution T (z)
on R may be conformally transplanted to a harmonic function Te(w) on D (and vice
versa) by assigning equal temperatures to corresponding points of the two regions:
Te[f(z)] ≡ T (z). In particular, the values of T on B are transplanted to C.
We can use this to understand the following. If we know a conformal mapping
w = f(z) from R to D such that f(p) = 0, then the Green’s function of R with pole
at p is

Gp (z) = − ln |f(z)|.

To see this, consider the temperature distribution Gep (w) in D obtained by trans-
planting Gp with f. We know that this conformal transplantation preserves the
harmonicity of the original, that it sends the source at p to an equal source at
f(p) = 0, and that the vanishing temperature on B is transplanted to vanishing tem-
perature on C. Thus this temperature distribution G ep (w) = Gp [f−1 (w)] in D must
be the Green’s function with pole at 0, and we already know that this is − ln |w|.
Done.
Exactly the same reasoning yields the following generalization. Let J(z) be a
one-to-one conformal mapping of R to some other simply connected region S with
 The Physics of Riemann’s Mapping Theorem 629

boundary Y. Then J conformally transplants the Green’s function Ga (z) of R with

pole at a to the Green’s function of S with pole at J(a). In particular, the streamlines
of the flow in R map to the streamlines of the flow in S. In this sense, the concept of
the Green’s function is conformally invariant.
This result immediately yields the following generalization of (12.6.6) to the case
where the pole of the Green’s function of R is an arbitrary point s, rather than f−1 (0).
From (12.22), we know the formula for the Green’s function in the disc when the
pole is moved from 0 to f(s). Conformal transplantation by means of f−1 carries
this pole to the desired point s of R, so

f(z) − f(s)
Gs (z) = − ln .
f(s) f(z) − 1
As a bonus, notice that this general formula establishes the previously claimed
“symmetry property” of the Green’s function:

Gs (z) = Gz (s).

For a more common approach to the symmetry property, see Ex. 15.
We end this section with a result which we will need later. The analogue for heat
flows of the velocity of a fluid flow is the heat flow vector H; in the present case,
H = −∇G. Let us call its magnitude Q ≡ |H| the local heat flux; this is the analogue
of fluid speed, and it represents the heat flux (per unit length) across a short line-
segment at right angles to the flow. Since B is an isotherm, H is orthogonal to B,
and so the local heat flux at a boundary point z may be expressed as
BG
Q(z) = − , (12.25)
Bn
where n measures distance in the direction of N, the outward unit normal vector
to B (see [12.31]). We can now state the result:

At a boundary point z, the local heat flux is equal to the amplification of

(12.26)
f : Q(z) = |f ′ (z)|.
For example, using (12.21), the result predicts that the local heat flux at the
boundary of the unit disc is given by [exercise]
1 − |p|2 1 − |p|2
Q(z) = |f ′ (z)| = = . (12.27)
|z − p|2 ρ2
This formula will play a central role in the next section. Of course Q(z) may instead
be calculated directly by substituting (12.22) into (12.25), but this is a little easier
said than done [exercise].
The general result (12.26) can be understood very intuitively. See [12.31]. With an
infinitesimal value of k, consider the shaded k-flux tube emanating from p which
hits B at z, and let its width there be ϵ. Its image under f is a k-flux tube emanating
630 Flows and Harmonic Functions

from 0 and hitting C at w = f(z). [Remember, f was originally defined to have this
property!] Let e
ϵ be the width of this image tube at w. Since the segment of B at z of
length ϵ is amplitwisted by f ′ (z) to the segment of C at w of length e
ϵ,
e
ϵ
|f ′ (z)| = .
ϵ
Next, recall that the width of a k-flux tube at any given point is equal to k divided
by the local heat flux [previously fluid speed] at that point. Since the local heat flux
is constant on C, its value at w is simply the ratio of the strength of the source at 0
to the perimeter of C, and by construction this ratio is (2π/2π) = 1. Since the local
heat flux at z is Q(z), we see that

e
ϵ = k/1 and ϵ = k/Q.

Thus
e
ϵ k
|f ′ (z)| = = = Q(z),
ϵ k/Q
as was to be shown.

12.7 Dirichlet’s Problem

12.7.1 Introduction
Consider a steady heat flow within a metal plate whose faces are insulated. Other
than at singularities, the temperature T (z) is then a harmonic function, and the
(locally) sourceless heat flow vector field is H = −∇T .
Let us measure the temperature around the circumference C of a circle of radius
R, the interior of which is free of sources and sinks, and the centre of which we may
conveniently choose to be the origin. As z = R eiθ moves round C, we may express
the measured temperature as a function of the angle: T = T (θ). We hope that it
may seem physically plausible that these values actually determine the tempera-
ture at any interior point a. Indeed, if a = 0 then we know [from Gauss’s Mean
Value Theorem] that the temperature at the centre of C is simply the average of the
temperatures on C:
Z
1 π
T (0) = ⟨T ⟩ = T (θ) dθ. (12.28)
2π −π
Eventually we will discover that the generalization of this result to the case a ̸= 0
is given by
Z  
1 π R2 − |a|2
T (a) = T (θ) dθ. (12.29)
2π −π |z − a|2
 Dirichlet’s Problem 631

Writing a = r eiα (r < R), and appealing to the cosine formula [exercise], this is
usually written as
Z  
1 π R2 − r2
T (a) = T (θ) dθ.
2π −π R2 + r2 − 2Rr cos(θ − α)

This is called Poisson’s formula, and the quantity in square brackets is called the
Poisson kernel, which we shall write as Pa (z).
Formula (12.29) says that T (a) is a weighted average of T on C, the temperature
of each element of C contributing to T (a) in proportion to its weight Pa (z). Notice
that Pa (z) dies away inversely with the square of the distance between a and the
element of C, so that if one element is twice as far from a as another, its influence
on the temperature at a is only one quarter as great. If a = 0 then all parts of C have
equal influence (for all are equally far from a) and you can see that we do recover
(12.28).
Poisson’s formula is connected with the following important and difficult
issue. Instead of dealing with a pre-existing harmonic function, Dirichlet’s prob-
lem demands that we arbitrarily (but piecewise continuously) assign values to the
boundary of a simply connected region R and then seek a continuous harmonic
function in R which takes on these values as the boundary is approached.
In the case of the disc, H. A. Schwarz demonstrated that not only does the
solution to Dirichlet’s problem exist, but it is explicitly given by (12.29). If we
are handed the piecewise continuous values T (θ) on C then we may construct
a function T (a) in the interior according to Poisson’s recipe. Schwarz’s solution
then amounted to showing that T (a) is automatically harmonic, and that as a
approaches a boundary point at which T (θ) is continuous, T (a) approaches the
given value T (θ). Let us begin to explain all this.7

12.7.2 Schwarz’s Interpretation

In 1890 Schwarz8 discovered a lovely geometric interpretation of formula (12.29)
which deserves to be far better known than it is:
To find the temperature at a, transplant each temperature on C to the
point directly opposite to it as seen from a, then take the average of the (12.30)
new temperature distribution on C.
Schwarz deduced this from Poisson’s formula, itself derived by computation.
We shall instead demonstrate his result directly and geometrically, only then
producing the Poisson formula as a corollary.
The example in [12.32] illustrates the beauty of (12.30). In [12.32a] half of C is
kept at 100 degrees with steam, while the other half is kept at 0 degrees with ice.

7
Much of the following material previously appeared in Needham (1994).
8
Schwarz (1972b).
632 Flows and Harmonic Functions

[12.32] In 1890 Schwarz discovered a beautiful geometrical interpretation of Pois-

son’s formula. [a] Transplant each temperature on C to the point directly opposite to it
as seen from a, thereby obtaining the new temperature distribution shown in [b]. Then
the temperature at a is simply the average of the new temperature distribution on C.

Being close to the cold side, we would expect a to be cool. Figure [12.32b] shows the
new temperature distribution obtained by projection through a. It is now vividly
clear how the distant hot semicircle is ‘focused’ through a onto a much smaller arc,
yielding a low average temperature on C and hence a low temperature at a itself.
To begin to establish (12.30), recall the conformal invariance of harmonic func-
tions: if T (z) is any harmonic function and h(z) any conformal mapping, then T (z⋆ )
is automatically harmonic, with z⋆ = h(z).
Suppose now that h(z) maps the disc to itself . If z = R eiθ lies on C then so does
⋆
z⋆ = R eiθ , and since we suppose that we have measured the temperature all round
C, we therefore know the temperature T (θ⋆ ) at z⋆ . Having the values of T (θ⋆ ), we
may now compute the integral in (12.28) for the harmonic function T [h(z)] to obtain
Zπ
⋆ 1
T (0 ) = T (θ⋆ ) dθ , (12.31)
2π −π

in which it should be stressed that the averaging is still taking place with respect
to the angle of z, not its image z⋆ .
We may interpret (12.31) as follows: the temperature at 0⋆ is the average of
the new temperature distribution on C obtained by transplanting the temperature
measured at each z to the new location z⋆ . We are now half way to Schwarz’s result.
To find the temperature at a we must find a conformal mapping of the disc to itself
such that 0 is sent to a, then take the average of the new temperature distribution.
But viewing the disc as the Beltrami–Poincaré model of the hyperbolic plane,
we are already very familiar with such mappings! Peek at [12.33b], to which we
 Dirichlet’s Problem 633

[12.33] Hyperbolic geometry proof of Schwarz’s interpretation of Poisson’s formula.

According to (12.31), if we can find a conformal mapping of the disc to itself that sends 0
to a, sending boundary points z to z⋆ , then we may find the temperature at a by averaging
these transplanted temperatures at z⋆ . But [b] reminds us that we have already met such
a transformation in hyperbolic geometry! In the Beltrami–Poincaré disc, let m be the
midpoint (in the hyperbolic sense) of the line-segment 0a, then the half-turn z ÞÑ z⋆ =
Ma (z) of the hyperbolic plane about m interchanges 0 and a: 0⋆ = a (as we desire) and
a⋆ = 0. As explained in the text, [a] allows us to prove that in fact z⋆ lies at the end of the
Euclidean chord through a and z, as illustrated, thereby proving Schwarz’s result.

shall return in a moment. If m is the midpoint (in the hyperbolic sense) of the line-
segment 0a, then the half-turn9 z ÞÑ z⋆ = Ma (z) of the hyperbolic plane about m
interchanges 0 and a: 0⋆ = a (as we desire) and a⋆ = 0. Thus to establish (12.30) we
need only demonstrate the illustrated fact that if z lies on C then z⋆ lies at the end of
the (Euclidean) chord B passing from z through a. In Chapter 3 we derived a formula
for Ma (z), so we could easily obtain this result by calculation; however, we prefer
a direct geometric approach.
First we need a simple result which is explained in [12.33a]. Consider the family
N ≡ {circular arcs passing from z to z⋆ },
where for the moment z⋆ may be thought of as any given point of C. The figure
shows three members of N: the arc A through 0, the Euclidean chord B, and the
hyperbolic line L. [Recall that in terms of hyperbolic geometry the members of N
consist of the equidistant curves of L.] The result we need is this:
The Euclidean chord B is the unique member of N such that L bisects the
(12.32)
angle contained by A and B.
Since each member of N is uniquely determined by the direction in which it
emerges from z, and since the radius z0 is tangent to L at z, this is equivalent to

9
See (3.53), p. 201.
634 Flows and Harmonic Functions

the following: if tz and t0 are tangent to A at z and 0, respectively, then the black
angle tz0 and the shaded angle z⋆ z0 are equal. The proof is immediate from the
figure and is left to the reader.
Now turn your attention to [12.33b], in which z⋆ is the image of z after a half-
turn about m. To finally establish Schwarz’s result we must prove the illustrated
fact that a lies on the chord B. To do so, consider the image A⋆ of A. Since the half-
turn interchanges the pair z and z⋆ and the pair 0 and a, A⋆ must be a member of N
passing through a. But since L⋆ = L, the conformality of the mapping also says that
the angle between A⋆ and L must equal the angle between A and L. We conclude
from (12.32) that this arc A⋆ through z, a, and z⋆ is none other than B. Done10 .

12.7.3 Dirichlet’s Problem for the Disc

Our example in [12.32] was a trifle hasty. For the moment, Schwarz’s result merely
says how the interior values of a given harmonic function in the disc may be found
from the values on C. But in [12.32] we blithely assumed that we could also use it to
construct such a function in the disc, given arbitrary piecewise continuous bound-
ary values. In other words we assumed Schwarz’s solution of Dirichlet’s problem
for the disc, as outlined in the introduction. We now justify this.
Figure [12.34a] shows a approaching a boundary point z; also shown are the
images (C⋆1 and C⋆2 ) under projection through a of the two small arcs (C1 and C2 )
adjacent to z. If the given boundary values are continuous at z then T is essen-
tially constant on C1 ∪ C2 , and so the new temperature distribution is likewise
almost constant on C⋆1 ∪ C⋆2 . As required, the constructed function T (a) therefore
does approach T (z) as a approaches z.
Although Dirichlet’s problem makes no demands on the behaviour of T (a) as a
approaches a boundary point at which T is discontinuous, it is easy to see (though
not to calculate!) what actually happens. Suppose that the boundary temperature
jumps from T1 to T2 as we pass from C1 to C2 . If a arrives at z while travel-
ling in a direction making an angle βπ with C, then [exercise] T (a) approaches
[βT1 + (1 − β)T2 ]. This result is relevant to the representation of discontinuous
functions by Fourier series.
It now only remains to show that the constructed function is indeed harmonic.
First we shall pause to recover Poisson’s formula in its classical form. We begin by
noting that (12.31) may be re-expressed [why?] as
Z
1 π
T (a) = T (θ) dθ⋆ . (12.33)
2π −π
In order to put this into the same form as (12.29), we now require dθ⋆ in terms of
dθ.
Consider [12.34b], which shows the movement R ∆θ⋆ of z⋆ resulting from a
movement R ∆θ of z. These arcs are ultimately equal to the chords t and s, so

10
This argument is perhaps conceptually clearer than the more elementary one in Needham (1994).
 Dirichlet’s Problem 635

[12.34] Solving Dirichlet’s Problem with Schwarz’s geometric construction. [a] Sup-
pose that the boundary temperature jumps from T1 to T2 as we pass from C1 to C2 .
If a arrives at z while travelling in a direction making an angle βπ with C, then T (a)
approaches [β T1 + (1 − β) T2 ]. [b] To derive Poisson’s formula from (12.33), con-
sider the movement R ∆θ⋆ ≍ t of z⋆ resulting from a movement R ∆θ ≍ s of z.
So (∆θ⋆ /∆θ) ≍ (t/s). But t and s are corresponding sides of two similar h i triangles
[shaded], so (t/s) = (σ ′ /ρ). And since (σ ′ /ρ) ≍ (σ/ρ), we obtain dθ
⋆
dθ
= σ ρ . Finally,
h i h i
ρ σ = ρ ′ σ ′ = (R + r)(R − r) = (R2 − r2 ), so dθ
⋆ 2 2
dθ
= σρ =
R −r
ρ2
= Pa (z). Done.

that (∆θ⋆ /∆θ) is ultimately equal to (t/s). But t and s are corresponding sides of
two similar triangles [shaded], so (t/s) = (σ ′ /ρ). Finally, since (σ ′ /ρ) is ultimately
equal to (σ/ρ), we obtain
 
dθ⋆ σ
= .
dθ ρ
Thus (12.33) becomes
Zπ  
1 σ
T (a) = T (θ) dθ . (12.34)
2π −π ρ

Consequently, to derive Poisson’s formula we need only show that [σ/ρ] is the
Poisson kernel Pa (z). This was precisely how Schwarz, working in the opposite
direction, originally deduced his result from Poisson’s formula.
Since ρ σ = ρ ′ σ ′ is constant, we may evaluate it for the dotted diameter through
a to obtain ρ σ = (R2 − r2 ). Thus we do indeed find that
   2 
σ R − r2
= = Pa (z).
ρ ρ2
636 Flows and Harmonic Functions

[12.35] Geometric proof that Pa (z) is harmonic, and the derivation h i of Schwarz’s
formula. Since the angle at E is a right angle, we have Pa (z) = σ = |z+a| cos γ
|z−a| =

ρ
Re z−a . Since Pa (z) is the real part of an analytic function, it is indeed harmonic. Let
z+a

S be a harmonic dual of T , so that f = T + iS is an analytic function. This function is

uniquely defined
Rπ (up
to a constant), and so it must be given by Schwarz’s formula:
1 z+a
f(a) = 2π −π z−a T (θ) dθ, for this is analytic and has T (a) as its real part. Thus
Schwarz was able to resurrect the complete analytic function f everywhere inside C just
from the ashes of its real part on C!

As an interesting consequence of the geometric interpretation of the Poisson kernel,

we see that (with z fixed) the level curves of Pa are the circles which are tangent to
C at z (i.e., horocycles), with Pa = 0 being C itself.
Returning to the issue of harmonicity, we see that if we permit ourselves differ-
entiation under the integral sign of (12.34), then it is sufficient to show that [σ/ρ] is
a harmonic function of a. To see that it is, consider [12.35].Since the angle at E is a
right angle, we have
   
σ |z + a| cos γ z+a
= = Re .
ρ |z − a| z−a

Because it is the real part of an analytic function of a, [σ/ρ] is automatically

harmonic, and we are done.
This line of reasoning yields a bonus result. Let S be a harmonic dual of T , so
that f = T + iS = −(complex potential) is an analytic function. This function is
uniquely defined (up to an additive imaginary constant) and so it must be given
by
Z  
1 π z+a
f(a) = T (θ) dθ,
2π −π z − a
 Dirichlet’s Problem 637

for this is analytic and has T (a) as its real part. This result is called Schwarz’s formula,
and it enables us to resurrect the complete analytic function f everywhere inside C
just from the ashes of its real part on C.

12.7.4 The Interpretations of Neumann and Bôcher

If we specify arbitrary piecewise continuous temperatures T (x) along the edge (the
real axis) of the upper half-plane, then there is another formula due to Poisson that
yields the temperature at any point a = X + iY (Y > 0):
Z  
1 ∞ Y
T (a) = T (x) dx . (12.35)
π −∞ (X − x)2 + Y 2
We shall explain this result by reinterpreting (12.33) in terms of elementary hyper-
bolic geometry. The transition from (12.29) to (12.35) will then be seen as nothing
more than a transition between the Beltrami–Poincaré disc and upper half-plane
models of the hyperbolic plane. First, however, let us obtain still another geometric
interpretation of Poisson’s formula.
For simplicity, let us employ the unit circle. Consider [12.36]. Let the arc K be
heated to unit temperature while the rest of C is kept at zero degrees. By Schwarz’s
result, the temperature at a is T (a) = (K⋆ /2π), while the temperature at the centre
of the circle is T (0) = (K/2π).
Next, imagine yourself standing at a looking out at a vast number of thermome-
ters placed at equal intervals along the circle. As you turn your head through a full
revolution—remembering to turn your feet!—let ⟨⟨T ⟩⟩a denote the average (over all
directions) of the temperatures you see. For example, the average ⟨T ⟩ occurring in
Gauss’s Mean Value Theorem is ⟨⟨T ⟩⟩0 .
In our case ⟨⟨T ⟩⟩a = (λ/2π), where λ is the angle subtended by K at a. But we see
from the figure that11
1
λ= 2 (K⋆ + K) ,
so ⟨⟨T ⟩⟩a = 12 [T (a) + T (0)]: the average of the boundary temperatures as they appear to
you is equal to the average of the temperature where you are and the temperature at the
centre. It is then easy to see that this is still true if we instead have many arcs at dif-
ferent temperatures, and ultimately a general piecewise-continuous temperature
distribution. Thus Poisson’s formula may be re-expressed as what we shall call
Neumann’s formula:
T (a) = 2 ⟨⟨T ⟩⟩a − T (0).
This result is due to Neumann (1884); we merely rediscovered it, as did Duffin
(1957) from another point of view. For an interesting generalization, see Perkins
(1928).

11
Incidentally, this means that the isotherms are the arcs of circles through p and q.
638 Flows and Harmonic Functions

[12.36] Neumann’s geometrical interpretation of Poisson’s formula. Let the arc K be

heated to unit temperature while the rest of C is kept at zero degrees. By Schwarz’s
result, the temperature at a is T (a) = (K⋆ /2π), while the temperature at the centre of
the circle is T (0) = (K/2π). Standing at a, turn your head through a full revolution,
and let ⟨⟨T ⟩⟩a denote the average (over all directions of your head) of the temperatures
you see. In our case ⟨⟨T ⟩⟩a = (λ/2π), where λ is the angle subtended by K at a. But
we see from the figure that λ = 12 (K⋆ + K) , so ⟨⟨T ⟩⟩a = 12 [T (a) + T (0)]: the average
of the boundary temperatures as they appear to you is equal to the average of the
temperature where you stand and the temperature at the centre. This is still true if
we instead have many arcs at different temperatures, and ultimately a general piecewise-
continuous temperature distribution. Thus Poisson’s formula may be re-expressed as what
we shall call Neumann’s formula: T (a) = 2 ⟨⟨T ⟩⟩a − T (0).

Figure [12.37] is intended to make this result vivid.Turning one’s head succes-
sively through the same small angle marked • one would see the thermometers
located at the white dots on the boundary. The average of their temperatures is
then a good approximation [exact as • → 0] to ⟨⟨T ⟩⟩a , and hence to the average of
the temperature where we stand and the temperature at 0. Note how the white dots
become crowded together on the part of the boundary nearest us. As anticipated,
this part of the boundary therefore has the greatest influence on the temperature
where we stand.
To obtain our third and final interpretation of Poisson’s formula, imagine that the
disc is the Beltrami–Poincaré model of the hyperbolic plane, and that you are once
again standing at the point a looking out to K, which is now infinitely far away
on the horizon. How big does K appear to you in this distorted geometry? To a
Godlike observer looking down on this model of the hyperbolic plane, the straight
 Dirichlet’s Problem 639

[12.37] Visualizing Neumann’s geometrical interpretation of Poisson’s formula.

Turning one’s head successively through the same small angle marked • one would see
the thermometers located at the white dots on the boundary. The average of their temper-
atures is then a good approximation [exact as • → 0] to ⟨⟨T ⟩⟩a , and hence to the average
of the temperature where we stand and the temperature at 0. Note how the white dots
become crowded together on the part of the boundary nearest us. As anticipated, this
part of the boundary therefore has the greatest influence on the temperature where we
stand.

lines along which light travels to you now appear to be arcs of circles orthogonal
to C, and so you see the angular size of K as being

hyperbolic angle = λ + (• + ⊙) .

But we see in the figure that

(• + ⊙) = 1
2 (K⋆ − K),

and hence we obtain the following remarkable fact:

1
hyperbolic angle = 2 (K⋆ + K) + 12 (K⋆ − K)
= K⋆
= 2π T (a).

The temperature where you are is simply proportional to how big K looks! [The
result can also be obtained directly by appealing to the conformal and circle-
preserving nature of the hyperbolic half-turn Ma (z) considered earlier.]
Reinterpreting (12.33), we now see that dθ⋆ is simply the hyperbolic angle sub-
tended at a by the element of C: the temperature of each element of C contributes to
640 Flows and Harmonic Functions

the temperature at an interior point in proportion to its hyperbolic size as seen from that
point. Much as we did in the Euclidean case, let ≺ T ≻a denote the average of the
temperatures you see on the horizon of the hyperbolic plane as you turn your head
through a full revolution while standing at a. We have thus found what we shall
call Bôcher’s formula:
T (a) = ≺ T ≻a . (12.36)
This result (exceeding even the beauty of Schwarz’s) is due to Bôcher (1898), Bôcher
(1906). We have chosen to present (12.36) as a consequence of Schwarz’s result, but
at the end of the section we shall see that it can be understood in a much simpler
way.
The analogue of [12.37] is now [12.38]. Standing at the same point as before, and
again turning one’s head successively through the angle •, the figure shows the
new locations of the thermometers we see on the boundary. The average of their
temperatures is then a good approximation [exact as • → 0] to ≺ T ≻a , and hence to
the temperature where we stand. Note how the white dots again become crowded
together on the part of the boundary nearest us, so that this part of the boundary
has the greatest influence on the temperature where we stand.
From the vantage point of (12.36), the distinction between (12.28) and (12.33)
evaporates. Every point of the hyperbolic plane is on an equal footing with every
other, it is merely that the hyperbolic angle dθ⋆ happens to coincide with the more
familiar Euclidean angle dθ when a = 0.
Formulated in this way, we may carry the result over to the upper half-plane
model for hyperbolic geometry. [The full justification for this transition will be
explained at the end of the section.] The horizon is now the real axis and ‘straight
lines’ are now (for our Godlike observer) semicircles meeting the real axis at right
angles. The temperature where you stand is now the average (as • → 0) of the
temperatures at the white boundary points in [12.39].
Figure [12.40] analyses this in greater detail. It shows both the hyperbolic angle
∆θ⋆ and the Euclidean angle ∆θ subtended at a by the element ∆x of the horizon.
Thinking of ∆x as sufficiently small that T (x) is essentially constant on it, the con-
tribution to the temperature at a is (1/2π) T (x) ∆θ⋆ . Integrating along the entire
horizon we obtain
Z
1 x=∞
T (a) = T (x) dθ⋆ . (12.37)
2π x=−∞
In order to put this into precisely the same form as (12.35), we need to find
(dθ⋆ /dx). We shall do this via an attractive and rather surprising fact: The non-
Euclidean angle ∆θ⋆ is exactly double the Euclidean one ∆θ, even if ∆x is not small. To
see this, concentrate on the semicircle meeting the axis at p. The angle between the
dotted tangent at a and the vertical is clearly double that between the chord ap and
the vertical. The result then follows immediately.
 Dirichlet’s Problem 641

[12.38] Bôcher’s hyperbolic interpretation of Poisson’s formula. Let ≺ T ≻a denote

the average of the temperatures you see on the horizon of the hyperbolic plane as you
turn your head through a full revolution while standing at a. Then we have what we shall
call Bôcher’s formula: T (a) = ≺ T ≻a . This figure visualizes this average: turning your
head successively through the small angle •, the white dots show the locations of the
thermometers you see on the boundary as light travels along hyperbolic straight lines to
your eye. The average of their temperatures is then a good approximation [exact as • → 0]
to ≺ T ≻a , and hence to the temperature where you stand. Note how the white dots again
become crowded together on the part of the boundary nearest to you, so that this part of
the boundary has the greatest influence on the temperature where you stand.

[12.39] Bôcher’s hyperbolic interpretation of Poisson’s formula in the upper half-

plane. The hyperbolic interpretation of the previous figure carries over without change
to the Beltrami–Poincaré upper half-plane model of the hyperbolic plane. The horizon is
now the real axis and hyperbolic straight lines are now semicircles meeting the real axis
at right angles. The temperature where you stand is now the average (as • → 0) of the
temperatures at the white boundary points.
642 Flows and Harmonic Functions

[12.40] Geometric proof that Bôcher’s formula is equivalent to Poisson’s half-plane

formula. To establish that (12.37) is equivalent to (12.35), we must find (dθ⋆ /dx) ≍
(∆θ⋆ /∆x). The first step in evaluating this is to recognize that even if ∆x is not small (as
we shall ultimately need it to be), the non-Euclidean angle ∆θ⋆ subtended at a by
∆x is exactly double the Euclidean one, ∆θ. To see this, concentrate on the semicircle
meeting the axis at p. The angle between the dotted tangent at a and the vertical is clearly
double that between the chord ap and the vertical. The result then follows immediately.

[12.41] Conclusion of the geometric proof that Bôcher’s formula is equivalent to

Poisson’s. The small shaded triangle is constructed to be right-angled, and it is thus
ultimately similar to the large shaded triangle as ∆θ Ñ 0. Thus (ξ/∆x) ≍ (Y/Ω).
Also, since ξ is ultimately an arc of a circle of radius Ω, it is ultimately equal to Ω ∆θ.
Thus as ∆θ Ñ 0, Ω∆x ∆θ
≍ ∆xξ
≍ Ω Y
. We can now combine
h this with
i the result of the
dθ⋆  dθ
 Y  Y
previous figure to obtain dx = 2 dx = 2 Ω2 = 2 (X−x)2 +Y 2 . Inserting this into
(12.37), we do indeed obtain (12.35).

Now consider [12.41]. The small shaded triangle is constructed to be right-

angled, and it is thus ultimately similar to the large shaded triangle as ∆θ Ñ 0.
Thus (ξ/∆x) is ultimately equal to (Y/Ω). Also, since ξ is ultimately an arc of a
circle of radius Ω, it is ultimately equal to Ω ∆θ. Thus as ∆θ Ñ 0,
 Dirichlet’s Problem 643

Ω ∆θ ξ Y
≍ ≍ .
∆x ∆x Ω
We can now combine this with the result of the previous figure to obtain
     
dθ⋆ dθ Y Y
=2 =2 = 2 .
dx dx Ω2 (X − x)2 + Y 2
Inserting this into (12.37), we obtain (12.35).
While the precise form of the above argument may be new, the basic idea of
transferring Bôcher’s result from the disc to the half plane was given by Osgood
(1928). For a different approach to (12.35), see Lange and Walsh (1985). For more
on all three of the interpretations thus far obtained, see Perkins (1928).

12.7.5 Green’s General Formula

If R is a simply connected region of arbitrary shape there exists a generalization of
Poisson’s formula (due to Green) for finding the temperature at any point a inside
R in terms of the values T (z) on the boundary B. As before, let Ga (z) be the Green’s
function of the region when the heat source is placed at the point a, so that local
heat flux at the boundary point z is given by
BGa
Qa (z) = −
.
Bn
With the aid of Qa we may now determine T (a). Here is the remarkable Green’s
formula:
I
1
T (a) = Qa (z) T (z) ds , (12.38)
2π
B

where ds is an element of arc length along B. Thus Qa now plays the same role as
the Poisson kernel did in (12.29). Indeed, we previously calculated Qa for the unit
disc, and we now recognize the result (12.27) as the Poisson kernel Pa .
Although formula (12.38) is valuable both in theory and practice, we should
point out that it is less explicit than Poisson’s formula, for to find Qa we must first
find the Green’s function. But as we previously explained, the problem of finding
Ga is itself a Dirichlet problem: to construct
Ga (z) = − ln ρ + ga (z),
we must find the harmonic function ga with boundary values ga (z) = ln ρ. For-
mula (12.38) says that if we can just solve this particular boundary value problem
then we can solve them all.
To begin with, imagine that T (z) is a given harmonic function in R whose value
T (a) at an interior a we wish to determine from the boundary values. The idea
behind our explanation of (12.38) is very simple12 . The Green’s function Ga enables

12
For a beautiful physical explanation of (12.38) in terms of electrostatic energy, see Maxwell (1873)
or, better still, Maxwell (1881, Ch. III).
644 Flows and Harmonic Functions

[12.42] Geometric proof of Green’s formula. If we know the Green’s function of

R, then, as previously explained, we can use it to conformally map R to the disc D:
z ÞÑ w = f(z) = e−G e−iH , and f(a) = 0. The harmonic function T (z) in R may be
transplanted to the harmonic function Te(w) ≡ T [f−1 (w)] in D, the boundary values on
B becoming the boundary values on C. But the average of these boundary tempera-
tures on C is the temperature at theHcentre, and this isHprecisely what we want, because
Te(0) = T (a). Therefore, T (a) = 2π
1 e ′
C T (w) dθ = 2π B T (z) |f (z)| ds. Finally, recall the
1

result (12.26): the amplification |f ′ (z)| equals the local heat flux Qa (z). This concludes
the derivation of Green’s formula, (12.38).

us to construct a conformal mapping f (and an inverse mapping f−1 ) between R and

the unit disc D such that a corresponds to 0. Figure [12.42] essentially reproduces
[12.31] in which this was explained.
Using the mapping z ÞÑ w = f(z), the harmonic function T (z) in R may be trans-
planted to the harmonic function Te(w) ≡ T [f−1 (w)] in D, the boundary values on B
becoming the boundary values on C. But the average of these boundary tempera-
tures on C is the temperature at the centre, and this is precisely what we were after,
because Te(0) = T (a).
Expressing this idea symbolically, we have
I
1
T (a) = Te(w) dθ.
2π
C
If the element dθ of C at w = f(z) is the image of the element ds of B at z then
dθ = |f ′ (z)| ds,
so
I
1
T (a) = T (z) |f ′ (z)| ds.
2π
B
Finally, recall the result (12.26): the amplification |f ′ (z)| equals the local heat flux
Qa (z). This concludes the derivation of formula (12.38).
 Dirichlet’s Problem 645

This argument also explains the stronger result that (12.38) solves Dirichlet’s
problem for R. Using f to conformally transplant the given boundary values from B
to C, we know that Poisson’s formula allows us to construct the solution to Dirich-
let’s problem in D. Transferring this solution back from D to R with f−1 , we have
found the harmonic function T in R, and its value at a must then be given by (12.38).
You can now understand why we lavished so much attention on the special case
of the disc.
We end this section with the observation that Green’s formula (12.38) possesses
the beautiful geometric interpretation shown on the LHS of [12.43]. Just as in
[12.38], one imagines standing at a and turning one’s head successively through
the small angle •. But now suppose that light travels along the illustrated streamlines
of the heat flow H = −∇Ga associated with the Green’s function. We would then see
the thermometers at the illustrated points on the boundary. The general formula
(12.38) says that the average of these temperatures (as • → 0) is the temperature
where one stands! Bôcher’s interpretation is clearly just a special case13 .
The explanation essentially reiterates the derivation of (12.26). Let zθ be the
boundary point we see when we look in the direction θ. Green’s formula says that
the temperature T (zθ ) of the element ds contributes to the temperature at a in pro-
portion to Qa (zθ ) ds, which is the flux of H through ds. Now follow the shaded flux
tube back to the source at a, and let dθ be its angular width there. Since 2π of flux
emerges symmetrically from a, the flux Qa (zθ ) ds emitted into our tube is equal to
dθ. Thus (12.38) may be re-expressed as
I
1
T (a) = T (zθ ) dθ, (12.39)
2π B
namely, as the average of the boundary temperatures T (zθ ) over all directions.
Done.
We have presented [12.43] as a geometric interpretation of (12.38), but we may
instead use it to simplify and illuminate our derivation of that formula. The key
observation is that (even without passing to the limit of vanishing •) the average in
[12.43] of the observed temperatures on B is conformally invariant. As before, let J(z)
be a one-to-one conformal mapping of R to some other simply connected region
S with boundary Y. Just as we did with f, let us choose J so that the directions of
curves through a are preserved (i.e., arg [J ′ (a)] = 0). Let wθ ≡ J(zθ ) be the image
on Y of zθ on B.
By the conformal invariance of the Green’s function, the image of the streamline
leaving a at angle θ is the streamline leaving J(a) at the same angle. Thus wθ is not

13
If we define the distance between two infinitesimally separated points of R to be the hyperbolic
distance between their images in D, then R becomes a (non-standard) conformal model of the hyper-
bolic plane, and the geodesics emanating from a are the streamlines of [12.43]. The two results may
then be viewed as identical.
646 Flows and Harmonic Functions

[12.43] The geometric meaning of Green’s formula. Green’s formula (12.38) possesses
the beautiful geometric interpretation shown on the left. Just as in [12.38], one imagines
standing at a and turning one’s head successively through the small angle •. But now
suppose that light travels along the illustrated streamlines of the heat flow H = −∇Ga
associated with the Green’s function. We would then see the thermometers at the illus-
trated points on the boundary. The general formula (12.38) says that the average of these
temperatures (as • → 0) is the temperature where one stands! Bôcher’s interpretation is
clearly just a special case. Indeed, if we define the distance between two infinitesimally
separated points of R to be the hyperbolic distance between their images in D, then R
becomes a (non-standard) conformal model of the hyperbolic plane, and the geodesics
emanating from a are the streamlines. The two results may then be viewed as identical.

only the image of zθ , it is also the boundary point which an observer at J(a) sees
when he looks in the direction θ. But, by definition, the temperature at each point
zθ on B is transplanted to wθ on Y, so the observer at J(a) sees exactly the same
temperatures on Y as the original observer at a saw on B. Done.
Passing to the limit of vanishing •, the conformal invariance of this average may
expressed as
I I
1 1
T (zθ ) dθ = Te(wθ ) dθ.
2π B 2π Y
Figure [12.43] illustrates the particular case where J = f is the previously con-
structed function which maps R to D and a to 0. The virtue of this special case
is that the conformally invariant average may now be evaluated. By Gauss’s Mean
Value Theorem, the average of the transplanted temperatures Te(wθ ) ≡ T (zθ ) on C
is the temperature Te(0) ≡ T (a) at the centre:
I I
1 1
T (zθ ) dθ = Te(wθ ) dθ = T (a).
2π B 2π C
 Dirichlet’s Problem 647

Thus, returning to R and passing to the limit of vanishing •, the average of the
observed temperatures is the temperature where you stand. Finally, the argument
associated with [12.43] shows that (12.39) is equivalent to (12.38).
As illustrated in [12.44], this idea of a conformally invariant hyperbolic average
lends unity to much of what we have done. Top centre is a depiction of Gauss’s

[12.44] Hyperbolic geometry unifies all methods of solving Dirichlet’s Problem. For
the Beltrami–Poincarites who inhabit the hyperbolic plane, the conformally invariant aver-
age that Gauss calls for at the centre of the disc [see top centre] is utterly indistinguishable
from the Poisson disc-formula [top left] at an arbitrary point. In both cases, the Beltrami–
Poincarite turns her head successively through the small angle •, and both figures show
the locations of the thermometers she sees on the boundary as light travels along hyper-
bolic straight lines (geodesics) to her eye. The average of their temperatures is then a
good approximation [exact as • → 0] to the temperature where she stands. But she
could also not detect any change if we transplanted her and the boundary temperatures
to the Beltrami–Poincaré upper half-plane: the temperature would still be the average
over all directions. Finally, if we conformally transplant both the boundary temperatures
and the hyperbolic metric to a general region R [top right], then it becomes a (non-
standard) conformal model of the hyperbolic plane. All four figures may then be viewed
as hyperbolically indistinguishable.
648 Flows and Harmonic Functions

theorem: as • → 0 the average of the temperatures at the white boundary points

is equal to the temperature at the centre of the disc. Applying an automorphism
to this picture yields the visual form of Poisson’s formula for the disc; applying
a Möbius transformation from the Beltrami–Poincaré disc model to the half-plane
model yields the visual form of Poisson’s formula for the half-plane; and applying
a more general conformal mapping yields the visual form of Green’s formula.
 Exercises 649

12.8 Exercises

1 (i) Show that the dual of a dipole is a dipole.

(ii) Think of a dipole as the limiting form of a doublet [a source and sink of
equal strength]. Sketch the dual of a doublet, and thereby make geometric
sense of part (i).
b is its harmonic dual, show that ΦΦ
2 If Φ is a real harmonic function, and Φ b is
b
also harmonic. [Hint: Consider the complex function (Φ + i Φ ).]
3 Use Gauss’s Mean Value Theorem to show that a harmonic function cannot
have a local maximum.
4 Find the generalization of (12.7) in three dimensional space.
5 (i) If f is analytic, show that ∇|f| = ( f f ′ )/|f|, and explain how this agrees
with part (i) of Ex. 19, p. 425.
(ii) Show that if ∆ is the Laplacian, then ∆|f|2 = 4 |f ′ |2 . Try deriving this by
brute force.
6 Let f(z) be analytic. By applying ∇ in turn to (f + f) and to (f f), show that if
either the real part or the modulus of f is constant, then f itself is constant.
7 (i) By thinking of z and z as functions of x and y, use the chain rule of partial
differentiation to show (at least formally) that

Bz = 1
2 ∇ and Bz = 1
2 ∇.

(ii) Deduce (at least formally) that an analytic function f depends on z but not on
z:

Bz f = f ′ and B z f = 0.

8 Let J denote the Jacobian of a transformation z ÞÑ w. Referring to the previous

question, show that the determinant is given by det(J) = |Bz w|2 − |B z w|2 .
9 From Ex. 7(i) we see that ∆ = ∇∇ = 4 Bz B z . Use this fact to solve the following
problems:
(i) Show that Φ = [1 − (x2 + y2 )]−1 satisfies ∆Φ = 4Φ2 (2Φ − 1).
(ii) Solve ∆F = ez for F by formally integrating with respect to z and then
with respect to z. Deduce that R = 14 ex (x cos y + y sin y) is a solution of
∆R = ex cos y. Verify this by calculating ∆R explicitly.
(iii) Show that if f is the most general harmonic function in the plane, then
f(z, z) = p(z) + q(z), where p and q are arbitrary analytic functions.
650 Flows and Harmonic Functions

10 As usual, let b
ξ be the unit tangent to a curve of curvature κ, and let e
κ be the
curvature of the image under the analytic mapping f(z). Also let s and e
s denote
arc length of the original and image curves. Finally, let Ψ = (1/|f ′ |) be the
stream function of the complex curvature, K = −i∇Ψ.
(i) Use (5.31), p. 272 to show that B s̃ e b ].
κ = Ψ Bs [κΨ + K ξ ·
(ii) Show that Bs bξ = iκbξ and Bs Ψ = ξ ·
b (iK).
(iii) Deduce that

B s̃ e · b.
κ = Ψ2 Bs κ + Ψ (Bs K) ξ

· ·
[Hint: Remember (or prove) that (ia) (b) + (a) (ib) = 0.]
(iv) Recall Ex. 18, p. 296, in which we saw how Newton attempted to define
the “angle” Θ between two touching curves as the difference of their
curvatures: Θ ≡ (κ1 − κ2 ). Although this is not quite conformally
 
invariant, show that Θ2 /Bs Θ is conformally invariant. This geometri-
cally meaningful generalization of the concept of angle is called Kasner’s
Invariant.
11 With the same notation as in the previous exercise, let p measure distance in
the direction ib ξ perpendicular to the curve. By substituting K = −i∇Ψ into
(5.31), p. 272, show that the image curvature is given by the tidy formula,

e
κ = Bp Ψ + κ Ψ.

12 Consider the image under an analytic mapping f of a source of strength S

located at p.
(i) Show geometrically, then algebraically, that if p is not a critical point of f
(i.e., f ′ (p) ̸= 0) then the image is another source of strength S at f(p).
(ii) Show geometrically, then algebraically, that if p is a critical point of f of
order (n − 1) then the image is again a source, but now of strength (S/n).
13 Repeat the investigation of the previous question in the case of a dipole located
at p.
14 Show that applying the Milne-Thomson formula (12.18) to a vortex at a point
p inside the unit circle yields two new vortices: at 1/p and at 0. Explain this
with the aid of a picture of the flow on the Riemann sphere.
15 (i) Show that if u and v are harmonic then X ≡ (u∇v − v∇u) is sourceless.
(ii) Prove (12.38) by taking u = T and v = Ga , then equating the flux of X out
of B with the flux out of an infinitesimal circle centred at a.
(iii) By taking u = Ga and v = Gb , prove the symmetry property of the Green’s
function: Ga (b) = Gb (a).
 Exercises 651

16 Let T (z) be the temperature distribution in the unit disc if the top semicircle
(Im(z) > 0) is kept at temperature +(π/2) and the bottom semicircle is kept at
temperature −(π/2). Show that
 
1+z
T (z) = Arg .
1−z
17 Let T (z) be a non-negative temperature distribution in the disc |z| ⩽ R. Writing
|a| = r, use Poisson’s formula (12.29) to derive Harnack’s inequality:
   
R−r R+r
T (0) ⩽ T (a) ⩽ T (0).
R+r R−r
18 Use Harnack’s inequality [previous exercise] to prove the following analogue
of Liouville’s Theorem: If T is harmonic in the whole plane and is bounded from
above (or below), then T is a constant.
19 Substitute the disc Green’s function (12.22) into (12.25), thereby confirming
the formula (12.27) for the heat flux at the boundary of the disc.
20 (i) Use the method of images to find the Green’s function for the upper half-
plane.
(ii) Use this to show that Green’s general formula (12.38) does indeed yield
the Poisson half-plane formula (12.35).
21 Use the idea behind the method of images to show that if 0 < Re (p) < 1 then
the Green’s function of the half-disc Re (z) ⩾ 0, |z| ⩽ 1 is
z−p z+p
Gp (z) = − ln + ln .
pz − 1 pz + 1
Check this by getting the computer to draw the level curves of Gp (z).
BIBLIOGRAPHY

Ahlfors, L. V. 1979. Complex analysis: an introduction to the theory of analytic functions

of one complex variable, International series in pure and applied mathematics, 3d
ed edition. New York: McGraw-Hill.
Altmann, S. L. 1989. Hamilton, Rodrigues, and the quaternion scandal. Mathematics
Magazine, 62(5):291–308.
Arianrhod, R. 2019. Thomas Harriot: A Life in Science. Oxford University Press.
Arnol’d, V. I. 1989. Mathematical Methods of Classical Mechanics, volume 60 of
Graduate Texts in Mathematics, 2nd edition. New York: Springer-Verlag.
Arnol’d, V. I. 1990. Huygens and Barrow, Newton and Hooke. Basel: Birkhäuser Ver-
lag. Pioneers in mathematical analysis and catastrophe theory from evolvents to
quasicrystals, Translated from the Russian by Eric J. F. Primrose.
Bak, J. and D. J. Newman 2010. Complex Analysis, Undergraduate texts in mathe-
matics, 3rd ed edition. New York: Springer.
Beardon, A. F. 1979. Complex Analysis: the Argument Principle in Analysis and
Topology. Chichester: Wiley.
Beardon, A. F. 1984. A Primer on Riemann Surfaces, volume 78. Cambridge: Cam-
bridge University Press.
Beardon, A. F. 1987. Curvature, circles, and conformal maps. The American Mathe-
matical Monthly, 94(1):48–53.
Beltrami, E. 1868a. Essay on the interpretation of non-Euclidean geometry. In
Stillwell (1996). American Mathematical Society.
Beltrami, E. 1868b. Fundamental theory of spaces of constant curvature. In Stillwell
(1996). American Mathematical Society.
Bivens, I. C. 1992. When do orthogonal families of curves possess a complex
potential? Mathematics Magazine, 65(4):226–235.
Bloye, N. and S. Huggett 2011. Newton, the Geometer. Newsletter of the European
Mathematical Society, (82):19–27.
Boas, R. P. and H. P. Boas 2010. Invitation to complex analysis, MAA textbooks, 2nd
ed edition. Mathematical Association of America.
654 Bibliography

Bôcher, M. 1898. Note on poisson’s integral. Bull. Amer. Math. Soc., 4(9):424–426.
Bôcher, M. 1906. On harmonic functions in two dimensions. Proceedings of the
American Academy of Arts and Sciences, 41(26):577–583.
Bohlin, K. 1911. Note sur le problème des deux corps et sur une integration nouvelle
dans le problème des trois corps. Bull. Astr., 28:113–119.
Braden, B. 1985. Picturing functions of a complex variable. The College Mathematics
Journal, 16(1):63–72.
Braden, B. 1987. Polya’s geometric picture of complex contour integrals. Mathemat-
ics Magazine, 60(5):321–327.
Braden, B. 1991. Visualization in teaching and learning mathematics: a project, volume
no. 19, chapter : The Vector Field Approach in Complex Analysis. Mathematical
Association of America.
Brieskorn, E. and H. Knörrer 2012. Plane Algebraic Curves, Modern Birkhäuser
classics. Basel: Birkhäuser.
Carathéodory, C. 1937. The most general transformations of plane regions which
transform circles into circles. Bulletin of the American Mathematical Society,
43(8):573–579.
Carathéodory, C. 1964. Theory of Functions of a Complex Variable, 2nd english ed
edition. New York: Chelsea Pub. Co.
Chandrasekhar, S. 1995. Newton’s Principia for the Common Reader. Oxford: The
Clarendon Press, Oxford University Press.
Courant, R. 1950. Dirichlet’s Principle, Conformal Mapping, and Minimal Surfaces,
volume 3. New York: Interscience Publishers.
Coxeter, H. S. M. 1946. Quaternions and reflections. The American Mathematical
Monthly, 53(3):136–146.
Coxeter, H. S. M. 1967. Proceedings of the International Conference on the Theory of
Groups, chapter : The Lorentz Group and the Group of Homographies, Pp. 73–77.
New York: Gordon and Breach Science Publishers.
Coxeter, H. S. M. 1969. Introduction to Geometry, 2nd edition. New York: Wiley.
Coxeter, H. S. M. and S. L. Greitzer 1967. Geometry Revisited, volume 19. Random
House.
Davis, P. and H. Pollak 1958. On the analytic continuation of mapping functions.
Transactions of the American Mathematical Society, 87(1):198–225.
Davis, P. J. 1974. The Schwarz Function and its Applications, volume no. 17. Mathe-
matical Association of America.
 Bibliography 655

de Gandt, F. 1995. Force and Geometry in Newton’s Principia. Princeton, NJ: Princeton
University Press. Translated from the French original and with an introduction
by Curtis Wilson.
do Carmo, M. P. 1994. Differential Forms and Applications, Universitext. Berlin:
Springer-Verlag.
Dray, T. 2015. Differential Forms and the Geometry of General Relativity. CRC Press,
Taylor & Francis Group.
Duffin, R. J. 1957. A note on Poisson’s integral. Quarterly of Applied Math, 15:109–111.
Earl, R. 2019. Topology: A Very Short Introduction, Very short introductions. Oxford
University Press.
Echols, W. H. 1923. Some properties of a skewsquare. The American Mathematical
Monthly, 30(3):120–127.
Eves, H. 1992. Fundamentals of Modern Elementary Geometry. Boston: Jones and
Bartlett.
Feynman, R. P. 1963. The Feynman Lectures on Physics. Reading, Mass.: Addison-
Wesley Pub. Co.
Feynman, R. P. 1966. Nobel lecture: ”The Development of the Space-Time View of
Quantum Field Theory”. Physics Today, August:31–34.
Feynman, R. P. 1985. QED: the Strange Theory of Light and Matter. Princeton, N.J.:
Princeton University Press.
Feynman, R. P. 1997. Surely you’re joking, Mr. Feynman!: Adventures of a Curious
Character, Norton paberback edition. New York: W.W. Norton.
Finney, R. L. 1970. Dynamic proofs of Euclidean theorems. Mathematics Magazine,
43(4):177–185.
Ford, L. R. 1929. Automorphic Functions, 1st edition. New York: McGraw-Hill book
company, inc.
Frankel, T. 2012. The Geometry of Physics: an Introduction, 3rd ed edition. Cambridge:
Cambridge University Press.
Fulton, W. 1995. Algebraic Topology: A First Course, volume 153. New York: Springer-
Verlag.
Gauss, C. F. 1827. General Investigations of Curved Surfaces, Translated from the Latin
and German by Adam Hiltebeitel and James Moreh ead, 1965 edition. Raven
Press, Hewlett, N.Y.
Goodstein, J. R. 2018. Einstein’s Italian Mathematicians: Ricci, Levi-Civita, and the Birth
of General Relativity. AMS.
656 Bibliography

Greenberg, M. J. 2008. Euclidean and non-Euclidean Geometries, 4th edition.

New York: W.H. Freeman.
Guicciardini, N. 1999. Reading the Principia. Cambridge: Cambridge University
Press. The debate on Newton’s mathematical methods for natural philosophy
from 1687 to 1736.
Guicciardini, N. 2009. Isaac Newton on Mathematical Certainty and Method, Transfor-
mations. Cambridge, Mass.: MIT Press.
Henrici, P. 1991. Applied and Computational Complex Analysis, volume 3 of Wiley
classics library, wiley classics library ed edition. New York: Wiley.
Hilbert, D. 1902. The Foundations of Geometry [Second English Edition of Tenth German
Edition, translated by Leo Unger, Revised and Enlarged by Dr. Paul Bernays, published
1971]. Chicago: The Open Court.
Hilbert, D. 1952. Geometry and the Imagination. New York: Chelsea Pub. Co.
Hilbert, D. 1965. Gesammelte Abhandlungen. Bronx, N.Y.: Chelsea Pub. Co.
Hildebrandt, S. and A. Tromba 1985. Mathematics and Optimal Form. New York:
Scientific American Library.
Hoggar, S. G. 1992. Mathematics for computer graphics, volume 14. Cambridge:
Cambridge University Press.
Hopf, H. 1946–1956. Differential Geometry in the Large: seminar lectures, New York Uni-
versity, 1946 and Stanford University, 1956 [Second Edition, 1983], volume 1000 of
Lecture notes in mathematics. Berlin: Springer-Verlag.
Horn, B. K. P. 1991. Relative orientation revisited. J. Opt. Soc. Am. A, 8(10):
1630–1638.
Jones, G. A. and D. Singerman 1987. Complex Functions: An Algebraic and Geometric
Viewpoint. Cambridge: Cambridge University Press.
Kasner, E. 1912. Conformal geometry. In Proceedings of the International Congress of
Mathematicians, Vol. 2.
Kasner, E. 1913. Differential-geometric Aspects of Dynamics, American Mathematical
Society. American Mathematical Society.
Katok, S. 1992. Fuchsian Groups, Chicago lectures in mathematics series. Chicago:
University of Chicago Press.
Klein, F. 1881. On Riemann’s Theory of Algebraic Functions and their Integrals: A Sup-
plement to the Usual Treatises ; translated from the German by Frances Hardcastle.
Mineola, NY: 2003, Dover Publications.
Körner, T. W. 1988. Fourier Analysis. Cambridge: Cambridge University Press.
 Bibliography 657

Lanczos, C. 1966. Discourse on Fourier series. Edinburgh: Oliver and Boyd.

Lange, R. and R. A. Walsh 1985. A heuristic for the poisson integral for
the half plane and some caveats. The American Mathematical Monthly, 92(5):
356–358.
Ludvigsen, M. 1999. General Relativity: A Geometric Approach. Cambridge: Cam-
bridge University Press.
Markushevich, A. 2005. Theory of Functions of a Complex Variable, number pt. 11 in
AMS Chelsea Publishing Series, 2nd edition. AMS Chelsea Pub.
Marsden, J., U. Marsden, M. Hoffman, U. Hoffman, and T. Marsden 1999. Basic
Complex Analysis. W. H. Freeman.
Maxwell, J. 1873. A Treatise on Electricity and Magnetism, A Treatise on Electricity
and Magnetism 2 Volume Paperback Set. 2010, Cambridge University Press.
Maxwell, J. 1881. An Elementary Treatise on Electricity, Cambridge Library Collection
- Physical Sciences. 2011, Cambridge University Press.
Milnor, J. 1982. Hyperbolic geometry: the first 150 years. Bull. Amer. Math. Soc.
(N.S.), 6(1):9–24.
Misner, Thorne, and Wheeler 1973. Gravitation. San Francisco: W. H. Freeman
[A marvellous new hardback edtition was published by Princeton Univer-
sity Press in 2017. It contains a new introduction by Charles Misner and Kip
Thorne, discussing exciting developments in the field since the book’s original
publication.].
Mughal, A. and D. Weaire 2009. Curvature in conformal mappings of two-
dimensional lattices and foam structure. Proc. R. Soc., A(465):219–238.
Needham, T. 1993. Newton and the transmutation of force. Amer. Math. Monthly,
100(2):119–137.
Needham, T. 1994. The geometry of harmonic functions. Mathematics Magazine,
67(2):92–108.
Needham, T. 2014. Visual Differential Geometry and Beltrami’s Hyperbolic Plane.
In The Art of Science: From Perspective Drawing to Quantum Randomness, R. Lupac-
chini and A. Angelini, eds., Pp. 71–99. Springer International Publishing.
Needham, T. 2021. Visual Differential Geometry and Forms: A Mathematical Drama in
Five Acts. Princeton, N.J.: Princeton University Press.
Nehari, Z. 1952. Conformal mapping. New York: Dover Publications.
Neumann, C. 1884. Vorlesungen über Riemann’s Theorie der Abel’schen Integrale.
Leipzig : B. G. Teubner.
658 Bibliography

Nevanlinna, R. H. and V. Paatero 2007. Introduction to Complex Analysis, 2nd edition.

Providence, R.I.: AMS Chelsea Pub./American Mathematical Society.
Newton, I. 1670. The Mathematical Papers of Isaac Newton, volume III. Cambridge:
1981, Cambridge U.P.
Newton, I. 1687. The Principia: Mathematical Principles of Natural Philosophy, 1999
edition. Berkeley, CA: University of California Press. A new translation by I.
Bernard Cohen and Anne Whitman, assisted by Julia Budenz, Preceded by “A
guide to Newton’s Principia” by Cohen.
Nikulin, V. V. and I. R. Shafarevich 1987. Geometries and Groups, Universitext. Berlin:
Springer-Verlag.
Nitsche, J. C. C. 1989. Lectures on Minimal Surfaces. Cambridge: Cambridge Univer-
sity Press.
Ogilvy, C. S. 1990. Excursions in Geometry. New York: Dover Publications.
O’Neill, B. 2006. Elementary Differential Geometry, Revised 2nd edition. Amsterdam:
Elsevier Academic Press.
Osgood, W. 1928. Lehrbuch der Funktionentheorie. Teubner.
Penrose, R. 1978. The geometry of the universe. In Mathematics today: twelve informal
essays, L. A. Steen, ed., Pp. 83–127. New York: Vintage Books.
Penrose, R. 2005. The Road to Reality. New York: Alfred A. Knopf Inc. A complete
guide to the laws of the universe.
Penrose, R. and W. Rindler 1984. Spinors and Space-Time. Cambridge: Cambridge
University Press.
Perkins, F. W. 1928. An intrinsic treatment of Poisson’s Integral. American Journal of
Mathematics, 50(3):389–414.
Poincaré, H. 1985. Papers on Fuchsian Functions. New York: Springer-Verlag.
Pólya, G. 1954. Mathematics and Plausible Reasoning. Princeton, N.J.: Princeton
University Press.
Pólya, G. and G. Latta 1974. Complex Variables. New York: Wiley.
Prasolov, V. V. 1995. Intuitive Topology, volume v. 4. American Mathematical Society.
Reech 1858. Journal de l’École Polytechnique, 27:169–178.
Richeson, D. S. 2008. Euler’s Gem: The Polyhedron Formula and the Birth of Topology.
Princeton, N.J.: Princeton University Press.
Schutz, B. F. 1980. Geometrical Methods of Mathematical Physics. Cambridge: Cam-
bridge University Press.
 Bibliography 659

Schutz, B. F. 2022. A First Course in General Relativity, 3rd edition. Cambridge:

Cambridge University Press.
Schwarz, H. 1972a. Gesammelte Mathematische Abhandlungen, volume II,
Pp. 144–171. Chelsea Publishing.
Schwarz, H. 1972b. Gesammelte Mathematische Abhandlungen, volume II, P. 360.
Chelsea Publishing.
Shapiro, H. S. 1992. The Schwarz Function and its Generalization to Higher Dimensions,
volume v. 9. New York: Wiley.
Shaw, W. T. 2006. Complex Analysis with Mathematica. Cambridge, UK: Cambridge
University Press.
Siegel, C. L. 1969. Topics in Complex Function Theory, volume 25. New York: Wiley-
Interscience.
Sommerville, D. 1958. The Elements of Non-Euclidean Geometry. Dover.
Springer, G. 1981. Introduction to Riemann Surfaces, 2nd ed edition. New York:
Chelsea Pub. Co.
Stewart, G. W. 1993. On the early history of the singular value decomposition.
SIAM Review, 35(4):551–566.
Stewart, I. and D. O. Tall 2018. Complex Analysis: The Hitch Hiker’s Guide to the Plane,
second edition. Cambridge University Press.
Stillwell, J. 1992. Geometry of Surfaces, Universitext. New York: Springer-Verlag.
Stillwell, J. 1994. Elements of Algebra: Geometry, Numbers, Equations. New York:
Springer-Verlag.
Stillwell, J. 1996. Sources of Hyperbolic Geometry, volume 10 of History of Mathematics.
Providence, RI: American Mathematical Society.
Stillwell, J. 2010. Mathematics and its History, third edition. New York: Springer.
Thorne, K. S. and R. D. Blandford 2017. Modern Classical Physics: Optics, Fluids,
Plasmas, Elasticity, Relativity, and Statistical Physics. Princeton University Press.
Thurston, W. P. 1997. Three-Dimensional Geometry and Topology, volume 35. Prince-
ton, N.J.: Princeton University Press.
Trefethen, L. N. 1986. Numerical Conformal Mapping. Amsterdam: North-Holland.
Vitelli, V. and D. R. Nelson 2006. Nematic textures in spherical shells. Phys. Rev. E,
74:021711.
Waerden, B. L. v. d. 1985. A history of algebra: from al-Khwˉarizmˉı to Emmy Noether.
Berlin: Springer-Verlag.
660 Bibliography

Wald, R. M. 1984. General Relativity. Chicago: University of Chicago Press.

Westfall, R. S. 1980. Never at Rest. Cambridge: Cambridge University Press. A
biography of Isaac Newton.
Whyburn, G. 1955. Introductory Topological Analysis, number v. 783, no. 22 in
Introductory Topological Analysis. University of Michigan Press.
Whyburn, G. 2015. Topological Analysis, Princeton Legacy Library, second edition.
Princeton University Press.
Woodward, L. M. and J. Bolton 2019. A First Course in Differential Geometry: Surfaces
in Euclidean Space. Cambridge University Press.
Yaglom, I. M. 1988. Felix Klein and Sophus Lie: evolution of the idea of symmetry in the
Nineteenth Century. Boston: Birkhäuser.
INDEX

2-spinors, 179 angular excess, 305–307

≍–notation, xvi additivity of, 374
in hyperbolic plane, 357–359
Abel’s convergence theorem, 78 on general surface, 312, 314, 383
absolute convergence, 76, 77 on sphere, 316–317
acceleration, 299, 478 angular momentum, 274, 279
Ahlfors, L. V., 175, 230 animation, computer, 332, 576
algebra annular Laurent series, 501–505
Cotes’ work on, 25 anticonformal mapping, 147
cubics, 5, 51, 66 antiderivative, 462–464
cyclotomic polynomials, 57 antipodal mapping, 321, 533
Descartes’ Factor Theorem, 26 antipodal points
Fundamental Theorem of, 26, 27, 404 and inversion, 167, 205, 206
quaternions, 49, 329–333, 374, 375 as intersection of great circles, 318
algebraic branch point, 102 Borsuk-Ulam Theorem, 425
algebraic multiplicity, 394–396, 398 definition of, 167
amplification, 218, 224 non-uniqueness of great circle between, 317
amplify, to, 218, 219, 224 orthogonal in C2 , 183
amplitwist stereographic images of, 167
as verb, 219 used to define axis of rotation, 318
defines analytic mappings, 223 used to find sphere rotation formula, 326
definition of, 218, 224 antipodal triangle, 316
German precedent for, xxvii Apollonius, xv
of exp(z), 262–264 Apollonius, circles of, 197
of log(z), 252–253 approximation
of za , 260–262 of (1 + ϵ), 93
of complex inversion, 229, 241 of exp(z), 91–93, 112
of inverse mapping, 255 of area under curve, 430–436
of power series, 259 of complex integral, 437–440
rules for calculating, 254–256 of power series, 78, 80, 122, 127, 128, 257
terminology illustrated, 219 of vector field locally, 553, 554
analytic continuation, 281 Archimedes, 121
preservation of identities, 286 area
uniqueness, 284–286 as integral of z, 448–450, 550–551
via reflections, 286–292 hyperbolic, 357–359
analytic function, 223–225 non-Euclidean, 306
and Riemann sphere, 230 of polygon, 33
equivalent definitions of, 227–230 signed, 31
angle of parallelism, 347, 381 spherical, 316–317
662 Index

Area Theorem, Bieberbach’s, 483 hyperbolic, 304

areal speed, 274, 299 parallel, 304
Argand, J. R., 2, 3 spherical, 304
Argument Principle
for polynomials, 393–394 Bôcher’s formula, 641
for rational functions, 414–416 Bak, J. and Newman, D. J., 506, 618
generalized, 415 basis vectors, 330
topological, 399–401 Beardon, A. F.
Arianrhod, R., 161 and Pick’s curvature formula, 298
arithmetic and Schwarzian derivative, 297, 481
equivalence of symbolic and geometric, and topological approach to infinite
8–11 differentiability, 485
Arnol’d, V. I. Beltrami, E.
as champion of Newton’s geometrical conformal “Poincaré” model, 359–363
method, xviii discovery of hyperbolic models, xii
dual force fields, 276, 277, 280 hemisphere model, 367–369
Kasner–Arnol’d Theorem, 280 hyperbolic plane, 339–342
astronomy, 156 interpretation of non-Euclidean geometry,
asymptotic 313
h-lines, 346 projective “Klein” model, 367
triangle, 357–359 pseudosphere model , 307
atlas, geographic Beltrami–Poincaré disc model, 359–363
and the metric of a surface, 322 motions of, 363–367
definition of, 158 Beltrami–Poincaré models of the hyperbolic
for pseudospherical surfaces, 341 plane, 342
stereographic and Mercator maps, 294 Beltrami–Poincaré upper half-plane, 349
attractive fixed point, 187 Beltrami–Poincarites, xii, 350
automorphism Bendixson’s Formula, 533
definition of, 199 Bernoulli, 498
of disc, 199–204, 211, 212 James, 68, 126
connections with non-Euclidean John, 132
geometry, 315, 367 Bieberbach’s Area Theorem, 483
uniqueness, 201 bilinear mappings, 137
via physics, 624–626 binary rotation, 331
of upper half-plane, 212 Binomial Theorem
connections with non-Euclidean and the series for log(z), 113
geometry, 349 and the series for ez , 92
average convergence and branch points, 106–110
complex, 483 Euler’s derivation of, 129
explanation of Cauchy’s Formula, 486–487 for negative integer powers, 128, 129
interpretation of Green’s formula, 645 Newton’s and Euler’s use of, 82
interpretation of Laplacian, 586 using complex numbers, 56
interpretation of Poisson’s formula, 631 Bivens, I. C., 592, 595
of harmonic function, 487, 581 Bôcher, M., 640
over circle, 121–124 Bohlin, K., 277
over polygon, 118–121 Bolyai, J., 307, 347
axiom Bolyai-Lobachevsky Formula, 347
 Index 663

geometric proof, 348 polar form, 247–248

Bombelli, R., 1, 4 used to derive log(z), 249–252
Borsuk-Ulam Theorem, 425 via the Jacobian, 237–238
Braden’s Theorem, 575 Cauchy-Hadamard Theorem, 84
Braden, B., xxi, 549, 556, 575 Cayley, A., 332, 426
branch cut, 104, 105, 109, 110, 131 celestial mechanics, 274–280
branch of multifunction, 103–106 central dilation, 45
branch point, 101 centroid, 134, 135, 487
algebraic, 102 of regular n-gon, 116, 135
as obstacle to power series, 107 of triangle, 53
at infinity, 164 of triangle constructed via equilateral
logarithmic, 102 triangles, 54
order of, 102 used to explain Gauss’s Mean Value
related to critical point, 233 Theorem, 115–121
Brouwer’s Fixed Point Theorem, 404–405, 424 chain rule of calculus, 254, 297
Burnett-Stuart, G. Chandrasekhar, S., 276
author’s debt to, xxx and ≍–notation, xvi
on Newton’s Principia, 23
calculus (real) characteristic equation, 181, 422
and complex arithmetic, 22–25 circle at infinity, 341, 343
Newton’s three versions of, xxiv circle of convergence, 78, 299
Cardano, G., 1, 4, 5 circle of curvature, 270, 300, 336, 435
Cardano-Tartaglia formula, 51 circles, touching, 153–154, 208
cardioid, 125 circulation of vector field, 541
Cassini, G., 67 Stokes’s theorem for, 547
Cassinian curve, 67–70, 125, 126, 576 vanishing of, xxi, 541, 545, 549, 557
Cauchy’s Formula, 485 work interpretation of, 539–542
derived via averaging, 486–487 classification
derived via Cauchy’s Theorem, 488–489 of Euclidean motions, 38–41
physical explanation of, 556 of hyperbolic motions, 348–357
used to derive existence of Laurent series, of Möbius transformations, 184–192, 194
502–505 of singular points, 518
used to derive existence of Taylor series, of singularities, 416–418, 503
491–493 of spherical motions, 317–321
used to derive infinite differentiability, of surfaces, 526
489–490 complex curvature
Cauchy’s Theorem additional curvature formula (in Ex. 11), 650
and antiderivatives, 464 advanced properties, 590–597
as nonlocal manifestation of local and Kasner’s Invariant (in Ex. 10), 650
amplitwist, 442 applied to dual central force fields, 279–280
for a general loop, 472–475 applied to the geometry of harmonic
geometric explanation of, 467–471 functions, 590–594
informal introduction to, 430 curl of, 595
used to derive Deformation Theorem, 453, curvature of streamlines and equipotentials
454 in terms of complex potential, 595
Cauchy–Riemann equations definition of, 271
explained visually, 245–247 elementary properties, 270–274
664 Index

complex curvature (Continued) Courant, R., 618

further geometry of, 595–597 Coxeter, H. S. M., 48, 137, 375
geometric interpretation of, 271 critical point
geometric interpretation of real and at infinity, 164
imaginary parts, 272 defined for non-analytic case , 399
interpretation in terms of rotation and definition of, 164, 231, 534
expansion of infinitesimal shapes, 273 degrees of crushing, 231–232
of various mappings (in Ex. 17), 295 geometric significance, 421, 425
VCA introduced both the concept and the Lucas’ Theorem, 294
name, xxi order of, 232
complex gradient operator, 588–589, 593, 649 physical significance, 571, 615, 616, 650
complex inversion, 139–150 relation to branch point, 233
amplitwist of, 229, 241 use in counting preimages, 394–396
as rotation of Riemann sphere, 163 cross product, 31
integral of, and winding number, 442–446, discovered via quaternions, 330
551–553 cross-ratio, 174–177
of hyperbola into lemniscate, 211 hyperbolic distance in terms of, 382
complex numbers, 1–8 projective invariance, 209
complex potential, 569–571 six possible values of, 210
confocal conics, 67, 100, 612 crossing rule for winding numbers, 388
conformal invariant, 296, 650 crosspoint, 518
conformal mapping, 147, 227–230 index of, 519
congruence, 34 cube roots, 100–106
conjugate cubic
harmonic function, 577 Cardano-Tartaglia solution of, 51
transpose matrix, 183 geometry of Cardano-Tartaglia formula, 66
conjugate transpose, 183 role in discovery of complex numbers, 5
conjugation mapping Viète solution of, 66
analytic fake of, 289 curl of vector field, 543
area as integral of, 447–450, 550–551 geometric formula for, 545, 574, 593
as reflection of Riemann sphere, 162 vanishing of, xxi, 545, 549, 557
connected set, 103 curvature, 210, 269
conservation of energy, 274, 541 applied to celestial mechanics, 278–280
conservative vector field, 541, 567–569 of harmonic equipotentials, 590–594
contour, 429 principal, 312
convergence related to divergence and curl, 543–545
absolute, 76, 77 related to Schwarz function, 299
circle of, 78, 299 related to Schwarzian derivative, 298
disc of, 75–78 transformation by analytic function,
radius of, 71, 73, 74, 78 267–270, 650
tests for, 83–86 used to estimate error of Riemann sum,
uniform, 79 434–436
convex hull, 116, 134, 294 cyclotomic polynomial, 57
convex region, 295
cosine, 94–100 Davis, P. J. and Pollak, H., 289
Cotes’ result, 25–30 Deformation Theorem
Cotes, R., 25–27, 29 for analytic function, 453–454
 Index 665

for flux, 547 Dirichlet’s problem, 630–648

for work, 547 disc of convergence, 75–78
Hopf’s, 388–393 disconnected set, 103
degree of a mapping, 390, 391 distributive property, 6, 10
degrees of freedom divergence of vector field, 543
in Riemann’s Mapping Theorem, 204, 613 Gauss’s theorem on, 547
of dipole mapping, 620 geometric formula for, 544, 574, 593
of disc automorphisms, 199–201 vanishing of, xxi, 545, 549, 557
of Green’s function, 627 dot product, 30
of Singular Value Decomposition, 236 discovered via quaternions, 330
of upper half-plane automorphisms, 212 in C2 , 182
using flows to construct mappings, 617–618 double pole, 417
Desargues, G., 175 doublet, 518
Descartes’ Factor Theorem, 26 doughnut
determinant and lemniscate, 70
of Jacobian, 240, 398, 421, 649 and Riemann surfaces, 230
of Möbius transformation, 178, 180 is called a torus, 526
differential geometry spiric sections of Perseus, 70
angular excess interpretation of curvature, vector field on, 527
312 dual
Beltrami’s use of in hyperbolic geometry, force laws, 280
307 harmonic, 577
conformal curvature formula, 380 Duffin, R. J., 637
constant negative curvature of
pseudosphere, 335–336, 357 Earl, R, 531
Further Reading, 373 earthquake, location of, 38
Gauss’s 1827 breakthrough, 311 Echols, W. H., 483
metric of a surface, 323 eigenvalue, 180–182
parallel transport, 383 interpretation of multiplier, 192–194
dilative rotation, 45, 50 eigenvector, 180–182, 192–194
dipole Einstein, A.
as fusion of source and sink, 558 and non-Euclidean geometry, 307, 379
at infinity, 560 General Relativity, 303
complex potential proof of circular Georg Pick’s influence on, 411
streamlines, 572 Lorentz transformations as Möbius
dipole moment of, 559, 560 transformations, 139
dual of, 578, 649 Penrose’s 2-spinors, 179
fused to create quadrupole, 559 spacetime interval, 138
image of, 604, 607 Special Relativity, 137, 138
index of, 519 electricity, flow of, 513–515
inserted into uniform flow, 572 electrostatic field
on Riemann sphere, 525, 609 as equilibrium state, 603
Pólya vector field, 557 behaves like ideal fluid, 516
superposition of multipoles, 561, 562 dual of, 603
used to construct Riemann mapping, energy contained in, 570, 643
618–620 Further Reading, 604
vanishing integral of, 557 interpretation of Green’s function, 623
666 Index

electrostatic field (Continued) Feynman, R. P.

method of images, 604–611 bet by, 496
of doublet, 516, 518 Nobel lecture, xxiii
potential of, 567 on complex numbers in quantum
elliptic integral, 68, 131, 243 mechanics, xxvii
elliptical orbits, 274–278 path integrals, 438
energy visualizing circulation, 541
conservation of, 274, 541 fiducial vector field, 528
electrostatic, 570, 643 Finney, R. L., 19
total, 274, 280 fixed point, 171
used to define potential function, 567 at infinity, 172
work, 539–543 attractive, 187
epicycloid, 478 Brouwer’s Theorem, 404–405, 424
equidistant curve repulsive, 187
angular definition of hyperbolic distance, flows, 513–518
361 around obstacles, 598–614
definition, 353 dual, 577–580
in Beltrami–Poincaré disc, 361 on curved surfaces, 525–531
equipotentials, 567 flux of vector field, 516, 537
l-work tube, 567, 568 k-flux tube, 563
harmonic, 590–594 Gauss’s theorem for, 547
equivalence relation, 36 local density of, 543
Erlangen Program, 37 vanishing of, xxi, 539, 545, 549, 557
Euclidean geometry, 303 foci of Cassinian curves, 68
motions of, 38–41 force fields, 274, 515
similarities of, 44–48 conservative, 541
spatial complex numbers?, 48–50 dual, 280
Three Reflections Theorem, 42 linear and inverse-square, 274–278
through the eyes of Felix Klein, 33–38 Ford, L. R., 186, 210
Euler characteristic, 527 Fourier series
Euler’s formula, 11–15 complex unification with Taylor series,
Euler’s Polyhedral Formula, 534 86–88
Euler, L., 501 examples derived from power series, 130,
and Binomial Theorem, 82 134
and birth of complex numbers, 2 reprsentation of discontinuous functions,
and principal curvatures, 312 634
discovery of Euler’s formula, 11 Fourier, J.
proof of Binomial Theorem, 129 discovery of Fourier series, 86
sum of reciprocals of squares, 498, 509 Fundamental Theorem of Algebra, 26, 27, 404
Euler-Lhuilier Formula, 534 Fundamental Theorem of Calculus, 458–462
exponential function, 88–94
extended complex plane, 157 Gauss’s Divergence Theorem, 547
extrinsic geometry, 309 Gauss’s Mean Value Theorem
and Dirichlet’s Problem, 630
Faraday, M., 515 for power series, 124, 646
k-flux tube terminology, 563 proof of, 581
Ferro, Scipione del, 5 related to Laplacian, 586
 Index 667

second proof of, 587 Harnack’s inequality, 651

statement of, 487 Harriot, T.
Gauss, C. F., 184 angular excess of spherical triangle, 316
complex plane, 2, 3 conformality of stereographic projection,
differential geometry, 308, 311, 312 161
non-Euclidean geometry, 306, 307 Hawking, S., 161
quaternions, 50 heat flow
rotations of sphere, 315 and Green’s function, 623, 626–648
Gaussian curvature, 311–313 around an obstacle, 598
formula in terms of metric, 380 steady state, 515
of hyperbolic plane, 357 temperature as potential function, 567
of pseudosphere, 335–336 helix, 298
of sphere, 316 Henrici, P., 614, 618
surfaces of constant, 313–315 Hexlet, Soddy’s, 209
Gaussian integers, 54 Hilbert, D., 342, 372, 618
genus of surface, 526 holes, 103, 523, 526
geodesic, 308 holomorphic, 223
geometry homogeneous coordinates, 178–180, 183, 193,
Erlangen Program, 37 376
extrinsic, 309 homographic transformations, 137
intrinsic, 310 Hopf’s Degree Theorem, 388–393
non-Euclidean, 33–38, 303–315 Hopf, H., 529
on curved surface, 307–309 horizon, 341
projective, 37, 175, 179 horocycle, 351
via complex arithmetic, 18–22 horosphere, 372
glide reflection, 58 Hurwitz polynomial, 422
gradient operator, 565–567 Huygens
complex, 588–589, 593, 649 influence on Newton, xv
gravitational orbits, 274–280 Huygens, C., 334, 357
Green’s formula, 643 hydrodynamics, 567
Green’s Function, 626–630, 650, 651 hydrogen atom, 506
Green’s General Formula, 643–648 hyperbolic functions, 95–100
group, 37 hyperbolic geometry, 305, 307, 313, 315
absolute unit of length, 306
Hamilton, W. R., 49, 329, 330, 332, 333 angle of parallelism, 347
harmonic family of curves, 590–594 angular excess, 357–359
curvature test for, 591 angular excess proportional to area, 306
definition of, 590 Beltrami–Klein model, 367, 377, 378
geometric test for, 591 Beltrami–Poincaré disc model, 359–363
harmonic function Bolyai-Lobachevsky Formula, 347
conformal invariance of, 583–586 Bolyai-Lobachevsky Formula (geometric
conjugate, 577 proof), 348
Dirichlet’s Problem for, 630–648 construction for distance, 361
dual of, 577 curvature of, 307
maxima and minima of, 407 equidistant curve, 353, 361
potential function, 569 hemisphere model, 367–369
stream function, 567 horizon, 341
668 Index

hyperbolic geometry (Continued) inversion in a circle, 139–142

horocycle, 351 geometric constructions of, 144–147, 205
horosphere, 372 illustrative applications of, 153–157
hyperbolic plane, 339–342 mechanical construction of, 205
lines and reflections, 342–347 preservation of angles, 147–149
maximum area of triangle, 306 preservation of circles, 142–144
motions in terms of reflections, 353–357 preservation of symmetry, 149–150
motions of, 348–353 inversion in a sphere, 150–153, 205
Poincaré upper half-plane model, 342 local geometry of, 243
Schwarz–Pick Lemma, 412 relation to stereographic projection, 161
triangles have negative angular excess, 306 role in hyperbolic geometry, 367–373
unifies solutions of Dirichlet’s Problem, 647 involutory transformation
via tractrix and pseudosphere, 333–339 and classification of z ÞÑ (1/z), 211
hyperbolic Möbius transformation, 174, 187 definition of, 141
geometric inversion as, 143
ideal points, 341
hyperbolic half-turn of the disc, 202, 203,
identities
365
hyperbolic, 95
using matrix representation of Ma , 212
preservation of, 286
irrotational vector field
trigonometric, 16–17
definition of, 542
identity matrix, 178
generated by a given potential, 581
Identity Theorem, 79
has harmonic potential function
identity transformation, 19, 36 if sourceless, 567
imaginary number, 2 potential function, 567
incompressibility, 516 unique flow round obstacle, 600
index of singular point
used to construct conformal mapping, 616
Bendixson’s Formula, 533
vanishing curl and Stokes’s theorem,
defined in the plane, 519 545–547
defined on a curved surface, 528 isogonal mapping, 148
Index Theorem, 523–525 isolated singularity, 503
of Pólya vector field, 548 isometric circle, 210, 242
Poincaré’s method of finding, 522–523 isometry, 35
Poincaré–Hopf Theorem, 526, 529–530 isotherm, 623, 627–629, 637
induced mapping, 162
infinitesimal, 23 Jacobian matrix
infinity, point at, 157 and topological multiplicity, 398
“inside” (of closed curve), 386–387 definition of, 217–218
integral, complex describes the local transformation, 235
as limit of Riemann sum, 437–440 determinant via Bz and B z , 649
parametric evaluation of, 466–467 eigenvalues and the SVD, 421
path-independence of, 442, 462–464 in non-analytic example, 422
integration by parts, 481 in three dimensions, 421
interior (of closed curve), 386–387 local area magnification is
interval of convergence, 71 (amplification)2 , 240
intrinsic of z ÞÑ (1/z), 241
geometry, 310 used to derive Cauchy–Riemann equations,
invariant curve, 42 237
 Index 669

Kasner’s Invariant, 650 lemniscatic integral, 131, 242

Kasner, E., 280 self-intersects at right angles, 125
“angle” between touching curves, 296 lemniscatic integral, 68, 131, 243
dual force fields, 277 Liebmann, H., 137, 321
Kasner–Arnol’d Theorem, 277, 280 limit rotation, 350
Kepler’s equation, 71 linear transformation
Klein, F. geometry of, 235–237
Erlangen Program, 33–38 Jacobian, 217
and amplitwist terminology, xxvii Möbius transformation as, 179
Beltrami-Klein model of the hyperbolic matrix representation of, 177
plane, 342 of spacetime, 138
classification of Möbius transformations, linear-fractional transformations, 137
184 Liouville’s Theorem, 409–411, 506, 651
monograph on Riemann surfaces, xxvii, 62, Liouville, J., 243
530 Lobachevsky, N., 307, 347
projective model of hyperbolic plane, xii local degree, 397
Koebe, P., 614 logarithm, 110–112
Kubrick, S. and analytic “rigidity”, 249–252
2001: A Space Odyssey, xv and Mercator map projection, 294
as integral, 465–466
Lambert, J. H., 306
definition of general powers, 113–115
Lanczos, C., 86
power series for, 112–113
Lange, R. and Walsh, R. A., 643
principal value of, 111
Laplace’s equation, 293
visual differentiation of, 252–253
Laplacian operator, 293
logarithmic branch point, 102, 164
complex gradient form of, 589
logarithmic power series, 113, 132, 133
conformal invariance of, 584–586
loop
geometric interpretation of, 586–587
“inside” of, 386–387
in terms of Bz , 649 simple, 386
Laurent series, 456, 477 winding number of, 385
and poles, 493–494 Lorentz transformations, 138
annular, 501–505 classification of, 172
classification of singularities, 503 loxodromic Möbius transformation, 174, 187
Legendre polynomials, 506 Lucas’ Theorem, 294
Legendre, A., 534
Leibniz, xv Mˉadhava, of Sangamagrˉama, 113
Leibniz’s rule, 127 Macdonald’s Theorem, 424, 425
Leibniz, G., 1 matrix
series for π, 113 conjugate transpose, 183
lemniscate identity, 178
as Cassinian curve, 69 Jacobian, 217
as image of hyperbola under inversion, 211 normalized, 177
as spiric section of Perseus, 70 of Möbius transformation, 177–178
as streamline and equipotential unitary, 183
of quadrupole, 576 Maximum-Modulus Theorem, 405–406
definition of, 68 Maxwell, J. C.
James Bernouilli’s standard lemniscate, 126 k-flux tube terminology, 563
670 Index

Maxwell, J. C. (Continued) related to maxima and minima, 126, 406,

and Macdonald’s Theorem, 426 407, 424
energy interpretation of k-cell, 569 moments, multipole, 560, 607
geometric approach to electrostatics, 604 momentum, 541
inversion is the only conformal mapping of angular, 274, 279
3D space, 243 Morera’s Theorem, 471
mathematically exploited Faraday’s field Morse Theory, 536
lines, 515 motions
physical explanation of Green’s formula, direct, 39
643 Euclidean, 38–41, 373
physically distinct interpretations of hyperbolic, 348–353, 369–373
potential, 624 isometry, 35
Mercator map projection, 294, 377 opposite, 39
Mercator, G., 294 rigid, 35
Mercator, N., 133 spherical, 317–321, 373
meromorphic function, 417 multifunctions, 100–110
Merton College, Oxford, xxx, 209 multiplicity
method of images, 604–611 algebraic, 394–396, 398
metric of a surface, 323 topological, 397–403
Midpoint Riemann Sum, 432 multiplier of Möbius transformation, 186
Milne–Thomson Circle Theorem, 608 calculation of, 191–192
Minding, F., 321 eigenvalue interpretation of, 192–194
minimal surfaces, 579 geometric meaning of, 189–190
Minimum-Modulus Theorem, 405–406 multiply connected set, 103
Möbius transformations, 137 multipole, 560
as automorphisms of the disc, 199–204 at infinity, 560
as matrices, 177–178 interpretation of Laurent series, 561
superposition of multipoles, 561
basic results on, 168–177
multipole expansion, 561
classification of, 172
multipole moments, 560, 607
decomposed into 2 or 4 reflections, 194–199
mythology, 440
decomposed into simpler mappings, 139
in Einstein’s Relativity Theory, 138–139 natural boundary, 299
in non-Euclidean geometry, 315, 325–329, Needham
349–353, 356, 360, 363–367, 372 Claudia, xxxi
normalized, 170 Faith, xiii
of upper half-plane, 211, 349 Guy, xxxi
visualized and classified, 184–194 Hope, xiii
modular functions, xx Mary, xxxi
modular group, xx Rodney, xxxi
modular surface, 63, 425 Needham, T., 276, 278, 280, 631
and Cassinian curves, 70, 125 Neumann’s formula, 637
near essential singularity, 417 Newton
of sin(z), 95 ultimate equaility, xv–xix
of polynomial, 69, 70 embraced geometrical methods in 1680s,
of powers, 63 xvi
poles as “volcanoes”, 74, 417 Principia, xviii
 Index 671

priority battle with Leibniz, xv phase portrait, 514

shunned his 1665 calculus in the Principia, photons, 206
xv Pick’s curvature formula, 298
synthetic method of fluxions, xvi Pick’s Theorem, 411
Nikulin, V. V. and Shafarevich, I. R., 33 Pick, G., 411
non-commutative multiplication, 49, 330 influence on Einstein, 411
non-singular Möbius transformation, 139 planetary motion, 274–280
normalized Möbius transformation, 170 Poincaré, H.
Nyquist Stability Criterion, 422 and motions of hyperbolic 3-space, 315
hyperbolic motions as Möbius
octupole, 560
transformations, 315
orbits in force fields, 274–280
motions of hyperbolic 3-space, 372
order of branch point, 102
rediscovery of Beltrami hyperbolic
order of critical point, 232
models, xii, 342, 359
orthogonal trajectory, 125
Poincaré–Hopf Theorem, 526, 529–530
osculating plane, 298
point at infinity, 157
Osgood, W. F., 643
Poisson’s formula
Pappus, xv for disc, 631
parabolic Möbius transformation, 172, 190 for upper half-plane, 637
Parallel Axiom, 304 polar line of a rotation, 320
parallel transport, 383 pole, 415, 416
parametric evaluation, 466–467 order of, 417
partial fractions, 25, 455 pole of a rotation, 320
Pascal’s triangle, 129 Pólya vector field
Peaucellier, A., 205 area as flux, 550–551
Penrose, R. Braden’s Theorem, 575
2-spinors, 180 Cauchy’s formula, 555–556
and complex numbers in physics, xxvii Cauchy’s Theorem, 549–550
and Möbius transformations in relativity complex integration, 547–549, 574
theory, 137 definition of, xxi, 547
author’s debt to, xiv, xxx index of, 548
labelling of lights rays with complex integral of zm , 556–560
numbers, 206 named by and advocated for
on Differential Geometry, 373 by Bart Braden, xxi
on quaternions, 333 non-analytic example, 564
Perelman, G., xx of dipole, 557
periodic function physical meaning of contour integral, 549
cos z, 98 relation to complex potential, 571
sin z and cos z, 94 winding number as flux, 551–553
exponential function as, 89 polynomials
Fourier series representation, 86 approximating power series, 78–79
saw tooth function, 134 Argument Principle for, 393–394
square wave function, 134 Cassinian curves via, 68–70
Perkins, F. W., 637, 643 cubic, 5, 51, 66
Perseus, 70 cyclotomic, 57
perspective drawing, 209 Descartes’ factor theorem, 26
perspectivity, 36 Fundamental Theorem of Algebra, 26, 404
672 Index

potential function, 567–569 recurrence relations, 55, 56

power series reflection
convergence of, 71–78, 83–86 glide, 58
differentiation of, 257–259 in a circle, 141
for exp(z), 83 in a general curve, 289, 290
for log(1 + z), 112–113 in a hyperbolic line, 345
Laurent, 493–494, 501–505 in a sphere, 151
manipulation of, 81–83 in a spherical line, 318
polynomial approximation of, 78–79 Three Reflections Theorem, 42
related to Fourier series, 86–88 reflection property of conics, 99, 100
Taylor, 87, 491–493 reflections, reduction to
uniqueness, 79–80 Euclidean motions, 41
used to find residues, 497–498 hyperbolic motions, 353
powers of z Möbius transformations, 198
fractional, 100–106 spherical motions, 319
general, 113–115 Relativity Theory, Einstein’s, 137–139, 379
positive integer, 64–66 repulsive fixed point, 187
visual differentiation of, 260–262 residue
primitive root, 57 evaluation of real integrals, 495–496, 506,
principal branch 507
of arg(z), 105 evaluation of series, 508
of log(z), 111 evaluation of series, 498–501
√
of 3 z, 105 finding using power series, 497–498
of z(a+ib) , 114 formula for a simple pole, 494
used in Binomial Theorem, 113, 130 general formula for, 494
principal curvatures, 312, 335, 545 Residue Theorem, 457, 477, 494
Principia, Newton’s, xxv, 67, 276, 434 Riemann sphere, 159
projections, group of, 36 behaviour of functions at ∞, 163–164
projective geometry, 37, 175, 179 classification of
pseudosphere, 307, 334 Möbius transformations, 173
̸= hyperbolic plane, 339 conformality, 160, 164, 165, 230
area of, 357 effect of inversions on, 162, 206, 325–326,
building your own, 379 374
conformal map of, 336–339 flows on, 560, 609–610, 650
constant curvature of, 313, 335–336, 357 rotations of, 182–184, 325–333
edge of, 340 symmetry of exp(z), 206
interpretation of limit rotations, 352 transferring functions to, 162–163
radius of, 313 Riemann sum, 430
pseudospherical surfaces, 333, 336, 342 accuracy of, 434–436
Ptolemy, C., 156, 158, 207 complex, 436–440
superiority over trapezoidal rule, 436, 478
quadratics, 68
using midpoints, 432
quadrupole, 559, 560, 576
Riemann surfaces
superposition of multipoles, 561, 562
and critical points, 233
quantum mechanics, xxvii
and doughnuts, 230
quaternions, 49, 329–333, 374, 375
are two dimensional, 62
radius of convergence, 71, 73, 74, 78 Further Reading, 530
 Index 673

omitted from this book, xxvii Siegel, C. L., 618

representation of multifunctions, 466 similarities, 44–48
Riemann’s Mapping Theorem, 204, 613–630 direct or opposite, 46
Riemann, B. simple loop, 386
and birth of complex analysis, 3 simple pole, 417
and Macdonald’s Theorem, 424 simple root, 395
discoverd n-dimensional Riemannian simply connected set, 103
geometry, 308 sine, 94–100
doctoral thesis, 204 singular point, 518
his physical approach to types of sector surrounding, 532
Riemann surfaces, xxvii Singular Value Decomposition, 236
introduction of Riemann sphere, 158 2-dimensional, 236, 237
rigid motion, 35 3-dimensional, 243
rigidity of analytic functions, 249–252 eigenvalues, 421
robotic vision, 332 singular values (of linear transformation), 237
Rodrigues, O., 49, 320 singularities
roots classification of, 503
algebraic multiplicity of, 394–396, 398 essential, 417
Descartes’ Factor Theorem, 26 of pole type, 415, 416
Fundamental Theorem of Algebra, 26, 404 order of, 417
of unity, 29 sinks, 516–518
primitive, 57 index of, 519
Rouché’s Theorem, 403 soap films, 579
simple, 395 Soddy, F., 209
topological multiplicity of, 397–403 Sommerville, D. M. Y., 237
rotations source-vortex
as complex functions, 19 explanation of Cauchy’s Formula, 556
as two reflections, 319, 350 role in Laurent series, 561
composing, 20 source-vortex combination
of space, 317–321, 325–333 role in Laurent series, 562
Rouché’s Theorem, 403 superposition of multipoles, 561, 562
sourceless vector field
Saccheri, G., 304, 305 and harmonic duals, 581
saddle points, 425, 518 definition of, 539
Schwarz function, 289, 299, 300 has harmonic potential
Schwarz’s formula, 637 if conservative, 569
Schwarz’s Symmetry Principle, 291, 605, 607 has harmonic stream function if irrotational,
Schwarz, H. A., 289, 631 567
Schwarzian derivative, 297, 481 of complex curvature of analytic mapping,
and curvature, 298 592
Schwarzian reflection properties if irrotational, 545–547
becomes reflection near the curve, 300 stream function of, 563
definition of, 289–292 strength proportion to density
in an ellipse, 299 of k-flux tubes, 564
used to generalize method of images, 609 used to prove Green’s formula, 650
used to generalize symmetric circles, 301 sources, 516–518
Shaw, W. T., xix, 139, 180 Pólya vector field of (1/z), xxi
674 Index

spacetime preservation under inversion, 149–150

Lorentz transformations, 138 Symmetry Principle, 168
Penrose’s stereographic construction, 206 disc automorphisms via, 200–201
spherical geometry, 313, 315 Schwarz’s, 291
absolute unit of length, 306 used to decompose Möbius transformations,
angular excess, 316–317 195
angular excess proportional to area, 306
Tartaglia, N., 5
conformal map, 321–325
temperature as potential function, 567, 624
curvature of, 307
Three Reflections Theorem, 42
definition of, 305
Thurston, W. P., 336, 361, 372, 379
motions of, 317–321
Topological Argument Principle, 399–403
rotations as Möbius transformations,
topological multiplicity, 397–403
325–329
topology
rotations as quaternions, 329–333
Euler characteristic, 527
triangles have positive angular excess, 306
Euler-Lhuilier Formula, 534
spiric section of Perseus, 70
genus of surface, 526
stagnation point, 571
Hopf’s Degree Theorem, 389
Star Trek
Poincaré–Hopf Theorem, 526
The City on the Edge of Forever, xiii
universal cover, 340
Guardian of Forever, xiii
torus, 526
Steiner, J., 208
and lemniscate, 70
stereographic projection
and Riemann surfaces, 230
as inversion in a sphere, 161
spiric sections of Perseus, 70
conformality of, 160, 164, 165, 230
vector field on, 527
formulae for, 165–168
tractrix, 333
in hyperbolic geometry, 367–373
translations
in spherical geometry, 324–329
as complex functions, 19
of antipodal points, 167
composing with rotations, 20
Penrose’s spacetime construction, 206
transplanted function, 583
preservation of circles, 161
transpose (of matrix)
Stewart, I. and Tall, D.,xix, 4
simple geometrical interpretation of, 237
Stiefel Vector Field, 534
transpose, conjugate, 183
Stiefel, Eduard, 534
Trapezoidal Rule, 432
Stillwell, J., xxvii
triangle inequality
Stokes’s Theorem, 547
Braden’s Theorem, 575
stream function, 563–565
generalized, 8
streamlines, 514
integral version of, 440–441
k-flux tube, 563
trigonometry, 16–17
strength of a source, 516
triple pole, 417
subgroup, 41
twist, 218, 219, 224
superposition of multipoles, 561, 562
as verb, 218, 224
surface
geometry on, 307–309 Ultimate Equality
of constant curvature, 313–315 ≍–notation, xvi
vector field on, 525–531 adoption of ≍–notation
SVD, 237, 243, 421 by Chandrasekhar, xvi
symmetric points, 141 basic properties of, xvi
 Index 675

definition, xvi sourceless, 539, 545–547, 563, 564, 567–569,

Newton’s use of in the 581, 592, 650
Principia, xv, xviii Stiefel, 534
used to calculate (tan θ) ′ , xvii, xxv vectorial operations, 30–33, 330
used to study complex differentiation, 245 velocity, 12
velocity of spiral motion, 23 Viète’s cubic solution, 66
ultra-parallel h-lines, 350 vortex, 555
uniform convergence, 79 index of, 519
uniqueness
Wallis, J., 2, 4, 481
of analytic continuation, 284–286
Wessel, C., 2, 3
of Möbius transformations, 170
Whyburn, G. T., 429
of power series, 79–80
winding number, 385
unitary matrix, 183
and Cauchy’s Formula, 488–489
universal cover, 340
and Residue Theorem, 475–477, 494
upper half-plane
as flux, 551–553
interpretation of cross-ratio, 176
as integral of (1/z), 446
Möbius transformations of, 211, 349
crossing rule for, 388
model of hyperbolic geometry, 339–342
work
Poisson’s formula for, 637
l-work tube, 567
valence, 396 circulation interpretation of, 541
vector local density of, 543
sourceless, 547 Stokes’s Theorem, 547
vector field, 511 used to define potential function, 567
conservative, 541, 567–569 vanishing of, xxi, 541, 542, 545, 549, 557
fiducial, 528
Y&Y, Inc., xiii, xxiv
irrotational, 542, 545–547, 567, 581, 600, 616
on a curved surface, 525–531 Zeitz, P., xxix, 129, 154