Ordinary

Differential

Equations

IN THE

Complex

Doaaain

E I N A R H I L LE
 DOVER BOOKS ON
MATHEMATICS

A Concise History of Mathematics, Dirk J. Struik. (60255-9) $7.95

Statistical Method from the Viewpoint of Quality Control, Walter A.
Shewhart. (65232-7) $7.95
Vectors, Tensors and the Basic Equations of Fluid Mechanics,
Rutherford Aris. (66110-5)
The Thirteen Books of Euclid's Elements, translated with an
introduction and commentary by Sir Thomas L. Heath.
(60088-2, 60089-0, 60090-4)
Introduction to Partial Differential Equations with Applications,
E.C. Zachmanoglou and Dale W. Thoe. (65251-3) $10.95
Numerical Methods for Scientists and Engineers, Richard Hamming.
(65241-6) $15.95
Ordinary Differential Equations, Morris Tenenbaum and Harry
Pollard. (64940-7) $18.95
Technical Calculus with Analytic Geometry, Judith L. Gersting.
(67343-X) $13.95
Oscillations in Nonlinear Systems, Jack K. Hale. (67362-6) $7.95
Greek Mathematical Thought and the Origin of Algebra, Jacob Klein.
(27289-3) $9.95
Finite Difference Equations, H. Levy & F. Lessman. (67260-3) $7.95
Applications of Finite Groups, J. S. Lomont. (67376-6) $9.95
Applied Probability Models with Optimization Applications, Sheldon
M. Ross. (67314-6) $6.95
Introduction to the Calculus of Variations, Hans Sagan.
(67366-9) $12.95
Introduction to Partial Differential Equations, Ame Broman.
(66158-X) $6.95
An Introduction to Ordinary Differential Equations, Earl A.
Coddington. (65942-9) $8.95
Matrices and Linear Transformations, Charles G. Cullen.
(66328-0) $8.95
Differential Forms with Applications to the Physical Sciences, Harley
Flanders. (66169-5)
Theory and Application of Infinite Series, Konrad Knopp.
(66165-2) $13.95
An Introduction to Algebraic Structures, Joseph Landin.
(65940-2) $7.95
Games and Decisions: Introduction and Critical Survey, R. Duncan
Luce and Howard Raiffa. (65943-7) $12.95
First Order Mathematical Logic, Angelo Margaris. (66269-1) $7.95
Introduction to Topology, Bert Mendelson. (66352-3) $6.95
Geometry: A Comprehensf/e Course, Dan Pedoe. (65812-0) $12.95
Functional Analysis, Frigyes Riesz and Bela Sz.-Nagy.
(66289-6) $12.95

(continued on back flap)

ft*
Ordinary Differential

Equations in the

Complex Domain

EINAR HILLE

DOVER PUBLICATIONS, INC.

Mineola, New York
 Copyright
Copyright © 1976 by John Wiley & Sons, Inc.
All rights reserved under Pan American and International Copyright
Conventions.
Published in Canada by General Publishing Company, Ltd., 30 Lesmill
Road, Don Mills, Toronto, Ontario.
Published in the United Kingdom by Constable and Company, Ltd., 3 The
Lanchesters, 162-164 Fulham Palace Road, London W6 9ER.

Bibliographical Note
This Dover edition, first published in 1997, is an unabridged and unaltered
republication of the work first published by John Wiley & Sons, Inc., New
York, in 1976 in the Wiley-Interscience Series in Pure and Applied
Mathematics.

Library of Congress Cataloging- in- Publication Data

Hille, Einar, 1894-
Ordinary differential equations in the complex domain / Einar Hille.
p. cm.
"An unabridged and unaltered republication of the work first published
by John Wiley & Sons, Inc., New York, in 1976 in the Wiley-Interscience
series in pure and applied mathematics" — T.p. verso.
Includes bibliographical references (p. ) and index.
ISBN 0-486-69620-0 (pbk.)
1. Differential equations. 2. Functions of complex variables. I. Title.
QA372.H56 1997
515'.35 — dc21 97-70
CIP

Manufactured in the United States of America

Dover Publications, Inc., 31 East 2nd Street, Mineola, N.Y 11501
To the Memory of My Parents
and
the Welfare of My Sons
 Digitized by the Internet Archive
in 2013

http://archive.org/details/ordinarydifferenOOhill_0
PREFACE

A friend recently wrote: "Mathematics is the creation of flesh and blood,

not just novelty-curious and wise automatons. We ought to devote some
part of our efforts to increasing understanding of the observable uni-
verse." This book is a contribution to these efforts. We praise famous
men, men who created beautiful structures and directed the course of
mathematics. A quotation from the Edda may be appropriate: "One thing I
know that never dies, the call after a dead man."
The structures that the masters built are not just beautiful to the eye;
they are also eminently useful. In these days when the need is felt for
"applicable mathematics" and "utilitas mathematica," it is fitting to recall
that few domains of mathematics are so widely applicable as the theory of
ordinary differential equations. This range of ideas is dear to my heart: for
close to 60 years much of my time has been given to the cultivation of
differential equations.
The book deals with ordinary differential equations in the complex
domain. It covers the usual ground, more or less. Here and there features
are introduced that are less canonical. There is a general emphasis on
growth questions: the dominants and minorants of Section 2.7 constitute a
variation of the majorant theme. The Nevanlinna theory of value distribu-
tion plays an important role: it is applied to the Malmquist-Wittich-Yosida
theorem (Sections 4.5 and 4.6) and to Boutroux's investigations (Sections
11.2 and 12.3). The Papperitz-Wirtinger account of Riemann's lectures on
hypergeometric functions and their uniformization by elliptic modular
functions has been rescued from oblivion (Section 10.5). Finally, the
second half of Chapter 12 presents the Emden-Fowler and the Thomas-
Fermi equations, quadratic systems, and Russell Smith's recent work on
polynomial autonomous systems, all matters of some novelty.
The reader is expected to have some knowledge of complex variables, a
subject in which our students are frequently weak: they comprehend
little, and often their knowledge is too abstract and is of the wrong kind.
Elementary manipulative skill is too often atrophied. Hence the second
half of Chapter 1 of this book is devoted to complex analysis. Chapter 1 1
has an appendix on elliptic functions, and modular and theta functions are
vii
 PREFACE
viii

discussed at some length in Sections 7.3 and 10.5. These sections should
help the reader.
Each chapter has a list of references to the literature, and there is a
bibliography at the end of the book. The exercises at the ends of sections
comprise some 675 items.
The book was written at the behest of Harry Hochstadt, who scoffed at
my misgivings and attempts to escape; he has aided and abetted my
efforts, and I owe him hearty thanks. May the book live up to his
expectations. Thanks are also due to numerous friends who have helped
with advice, bibliographical and biographical information, and construc-
tive criticism. Specific mention should be made of L. V. Ahlfors, O.
Boruvka, W. N. Everitt, C. Frymann, Ih-Ching Hsu, S. Kakutani, Z.
Nehari, D. Rosenthal, I. Schoenberg, R. Smith, H. Wittich, C. C. Yang and
K. Yosida. J. A. Donaldson and H. Hochstadt have kindly helped with the
proofreading. Further, I am grateful to Addison-Wesley Publishing Co. and
to the R. Society of Edinburgh for permission to use copyrighted material.
I am also indebted to the Department of Mathematics of the University of
California at San Diego for Xerox copying and to the National Science
Foundation for support (Grant GP 41127). Finally, I owe much to my
family, wife and sons, for encouragement, help, interest, and patience.

Einar Hille

La Jolla, California
February 1976
CONTENTS

Chapter 1. Introduction 1
I. Algebraic and Geometric Structures 1
1.1. Vector Spaces 1
1.2. Metric Spaces 3
1.3. Mappings 5
1.4. Linear Transformations on C into Itself; Matrices 6
1.5. Fixed Point Theorems 9
1.6. Functional Inequalities 12
II. Analytical Structures 17
1.7. Holomorphic Functions 17
1.8. Power Series 21
1.9. Cauchy Integrals 24
1.10. Estimates of Growth 31
1.11. Analytic Continuation; Permanency of Functional
Equations 33

Chapter 2. Existence and Uniqueness Theorems 40

2.1. Equations and Solutions 40
2.2. The Fixed Point Method 44
2.3. The Method of Successive Approximations 47
2.4. Majorants and Majorant Methods 51
2.5. The Cauchy Majorant 57
2.6. The Lindelof Majorant 60
2.7. The Use of Dominants and Minorants 63
2.8. Variation of Parameters 67

Chapter 3. Singularities 76
3.1. Fixed and Movable Singularities 76
3.2. Analytic Continuation; Movable Singularities 81
3.3. Painleve's Determinateness Theorem; Singularities 87
3.4. Indeterminate Forms 96

ix
X CONTENTS

Chapter 4. Riccati's Equation 103

4.1. Classical Theory 103
4.2. Dependence on Internal Parameters; Cross Ratios 106
4.3. Some Geometric Applications 110
4.4. Abstract of the Nevanlinna Theory, I 114
4.5. Abstract of the Nevanlinna Theory, II 121
4.6. The Malmquist Theorem and Some Generalizations 129

Chapter 5. Linear Differential Equations: First and Second

Order 144
5.1. General Theory: First Order Case 144
5.2. General Theory: Second Order Case 148
5.3. Regular-Singular Points 155
5.4. Estimates of Growth 162
5.5. Asymptotics on the Real Line 171
5.6. Asymptotics in the Plane 178
5.7. Analytic Continuation; Group of Monodromy 192

Chapter 6. Special Second Order Linear Differential Equations 200

6.1. The Hypergeometric Equation 200
6.2. Legendre's Equation 208
6.3. Bessel's Equation 211
6.4. Laplace's Equation 216
6.5. The Laplacian; the Hermite- Weber Equation;
Functions of the Parabolic Cylinder 228
6.6. The Equation of Mathieu; Functions of the Elliptic
Cylinder 234
6.7. Some Other Equations 242

Chapter 7. Representation Theorems 249

7.1. Psi Series 249
7.2. Integral Representations 253
7.3. The Euler Transform 258
7.4. Hypergeometric Euler Transforms 265
7.5. The Laplace Transform 270
7.6. Mellin and Mellin-Barnes Transforms 274

Chapter 8. Complex Oscillation Theory 283

8.1. Sturmian Methods; Green's Transform 284

8.2. Zero-free Regions and Lines of Influence 292
CONTENTS

8.3. Other Comparison Theorems 300

8.4. Applications to Special Equations 308

Chapter 9. Linear nth Order and Matrix Differential Equations 321

9.1. Existence and Independence of Solutions 321
9.2. Analyticity of Matrix Solutions in a Star 326
9.3. Analytic Continuation and the Group of Monodromy 328
9.4. Approach to a Singularity 333
9.5. Regular-Singular Points 342
9.6. The Fuchsian Class; the Riemann Problem 353
9.7. Irregular-Singular Points 360

Chapter 10. The Schwarzian 374

10.1. The Schwarzian Derivative 374
10.2. Applications to Conformal Mapping 377
10.3. Algebraic Solutions of Hypergeometric Equations 383
10.4. Univalence and the Schwarzian 388
10.5. Uniformization by Modular Functions 394

Chapter 11. First Order Nonlinear Differential Equations 402

11.1. Some Briot-Bouquet Equations 402
11.2. Growth Properties 408
11.3. Binomial Briot-Bouquet Equations of Elliptic
Function Theory 416
Appendix. Elliptic Functions 427

Chapter 12. Second Order Nonlinear Differential Equations and

Some Autonomous Systems 433
12.1. Generalities; Briot-Bouquet Equations 434
12.2. The Painleve Transcendents 439
12.3. The Asymptotics of Boutroux 444
12.4. The Emden and the Thomas-Fermi Equations 448
12.5. Quadratic Systems 455
12.6. Other Autonomous Polynomial Systems 460

Bibliography 468
Index 471
1

INTRODUCTION

In this chapter we shall list, with or without proof, various facts which
will be used in the following. They will fall under two general headings: (I)
algebraic and geometric structures, and (II) analytic structures. Under I
we shall remind the reader of abstract spaces, metrics, linear vector
spaces, norms, fixed point theorems, functional inequalities, partial
ordering, linear transformations, matrices, algebras, etc. Under II we
discuss analytic functions: analyticity, Cauchy's integral, Taylor and
Maclaurin series, entire and meromorphic functions, power series, growth,
analytic continuation, and permanency of functional equations. This is
quite an ambitious program, and the reader may find the density of ideas
per page somewhat overwhelming. He is advised to skim over the pages in
the first reading and to return to the relevant material as, if, and when
needed.

I. Algebraic and Geometric Structures

1.1. VECTOR SPACES

The term abstract space is often used as a synonym for set or point set,
but the term usually indicates that the author intends to endow the set
with an algebraic or geometric structure or both. If a Euclidean space Rn
serves as a prototype of a space, we obtain an abstract space by
abstracting (= withdrawing) some of its properties while keeping others.
Incidentally, property is an undefined term (we can obviously not use the
definition ascribed to Jean Jacques Rousseau: La propriete c'est le vol!).
We denote our space by X and its elements by x, y, z, We say that the
space has an algebraic structure if one or more algebraic operations can
be performed on the elements, or if a notion of order is meaningful at least
for some elements.
1
2 INTRODUCTION

The set is a linear vector space if the operations of addition and scalar
multiplication can be performed. It is required that the set be an Abelian
group under addition, that is, x + y is defined as an element of X, addition
is associative and commutative, there is a unique neutral element 0 such
that x + 0 = x for all x, and every element x has a unique negative, - x, with
x + (-x) = 0.
To define scalar multiplication we need a field of scalars, which is
almost always taken to be the real field R or the complex field C. For any
scalar a and element x there is a unique element ax; scalar multiplication
is associative, it is distributive with respect to addition, and 1 • x = x,
where 1 is the unit element of the scalars.
We speak of a real or a complex vector space according as the scalar
field is R or C. The elements of X are now called vectors. A linear vector
space which also contains the product of any two of its elements is called
an algebra. The set of all polynomials in a variable t is obviously an
algebra, and so is the set of all functions t »-» f(t) which are continuous at
a point t0.
Consider a set of n vectors x; in X, and let the underlying scalar field be
denoted by F. Then the vectors x, are linearly independent over F if

a,Xi + a2x2 + • • • + a„ x„ = 0 (1.1.1)

implies that all the a's are zero. They are linearly dependent over F if
multipliers a, can be found so that (1.1.1) holds with |a,| + |a2| + - • • +
|a„| >0. Emphasis should be placed on "over F," for restricting F to a
subfield F° or extending it to a larger field F* affects the independence
relations. Thus 1 and 21/2 are linearly independent over Q, the field of
rational numbers, but not over A, the field of algebraic numbers. The
space X is said to be of dimension n if it contains a set of n linearly
independent vectors while any n + 1 vectors are linearly dependent. It is
of infinite dimension if n linearly independent vectors can be found for
any n.
The notion of partial ordering is another form of algebraic structure.
We say that X is partially ordered if for some pairs x, y of X there is an
ordering relation x«sy (equivalently, y^x) which is reflexive, proper,
and transitive, that is, (i) x ^ x for all x, (ii) x ^ y and y ^ x imply x = y, (iii)
x^y, y^z imply x=^z. If X is linear as well as partially ordered, we
should have
implies x + a ^ y+ a for all a, (1.1.2)
implies ax *s ay for a > 0. (1.1.3)
In this case X has a positive cone X+, defined as the set of all elements
x E X such that 0 ^ x. This positive cone is invariant under addition and
METRIC SPACES 3

multiplication by positive scalars. It contains 0, the neutral element,

usually referred to as the zero element. We now have x^yify-xE X+.
The set of real valued continuous functions on the closed interval [0,1],
say C[0, 1], is an algebra. We define its positive cone X+ as the set of
functions t »-»/(0 whose values on [0, 1] are nonnegative. We have / ^ g
if g(t)-f(t) is nonnegative in [0, 1].
A more prosaic example may be helpful: the fowl in a hen-yard are
partially ordered under the pecking order.

EXERCISE 1.1
1. Consider the space of all polynomials P(t) in a real variable which take on
real values. Show that X is an algebra.
2. An order relation P < Q is established in X by defining P as positive if its
values are positive for all large positive values of t. Show that this ordering is a
trichotomy in the sense that for a given P there are only three possibilities: (i)
P is positive, (ii) -P is positive, or (iii) P = 0.
3. An order relation is said to be Archimedean if x < y implies the existence of
an integer n such that y < nx. (The natural ordering of the reals is Archime-
dean.) Show that the order defined in Problem 2 is non-Archimedean
inasmuch as the elements fall into rank classes Rk, where Rk consists of all
polynomials of exact degree k, each Rk is Archimedean, but if / is a positive
element of Rj and g a positive element of Rk with j < k, then f <g and nf <g
for all n. Verify.
4. Prove that \,t,t2, ...,tn are linearly independent over R. What is the
dimension of the space formed by these elements?

1.2. METRIC SPACES

A metric space is one in which there is defined a notion of distance
subject to the following conditions:
D,. For any pair of points P and Q of X a number d(P,Q)^0 is
defined, called the distance from P to Q such that d(P, Q) = 0 iff P = Q.
D2. d{P,Q) = d(Q,P).
D3. For any R we have d(P, Q) ^ d(P,R) + d(R, Q).

These notions go back to the work in the 1890's of Hermann Minkowski

(1864-1909) on what he called the "geometry of numbers." He was chiefly
concerned with the extremal properties of linear and of quadratic forms, for
which he found alternative definitions of distance, adjusted to the problem in
hand. Minkowski did not always require D2. Condition D3 is the triangle
inequality.
4 INTRODUCTION

We say that a linear vector space is normed if the following conditions

hold:

N,. For each x G X there is assigned a number ||x|| ^ 0 such that ||x|| = 0
iff x = 0.
N2. ||ax|| = \a \ ||x|| for each a in the scalar field.
N3. ||x + yH||X|| + ||y||.
A normed linear vector space becomes a metric space by setting

d(x,y) = \\x-y\\. (1.2.1)

In a metric space we can do analysis since the fundamental operation of
analysis, that of finding limits of a sequence, becomes meaningful. If {xn}
is a sequence in the metric space X, we say that xn converges to x0 and
x0 = limx„ if
lim d(x0, x„) = 0. (1.2.2)
n — » °°
We say that {x„} is a Cauchy sequence if, given any e > 0, there exists an
N such that
d(xm, xn)<e for m, n > N. (1.2.3)
If (1.2.2) holds, it follows that {x„} is a Cauchy sequence, but the converse
is not necessarily true, for there may be gaps in the space. A metric space
X is said to be complete if all Cauchy sequences converge to elements of
the space. Euclidian spaces are complete, and so are various function
spaces that will be encountered in the following. The space Q of rational
numbers is not complete.
Various notions of real analysis are meaningful in complete metric
spaces, such as the concepts of closure, open set, closed set, and
e-neighborhood. The Bolzano- Weierstrass theorem need not be valid in a
complete metric space, that is, there may be bounded infinite point sets
without a limit point. Incidentally, "bounded" means that the set can be
enclosed in a "sphere" d(x,0)<R. The topological diameter d(S) of a
subset of X is the least upper bound of the distances d(x, y) for x and y in
S.

EXERCISE 1.2

1. The Euclidean norm ||x||2 of x in C" is [S,"=1 |jcy|2]l/2, where x = (*,, x2, . . . , xn) and
the jc's are complex numbers. Alternative norms are flx^ = 2"=, |jc/| and
||x||oo = sup |jcy|. Show that they are indeed acceptable norms. Between what
limits do they lie if ||x||2 = 1?
2. How do you define an open set in these three normed topologies? Show that a
set open in one of them is also open with respect to the others. Verify that C"
is complete in all three metrics.
MAPPINGS 5

3. Let X = C[0, 1] be the set of all functions, t f{t), continuous in the closed
interval [0, 1]. Define a Cauchy sequence if ||/|| = sup0sSfs5i \f(t)\, and show that
the space is complete.

1.3. MAPPINGS

We shall study mappings from a metric space X into a metric space Y,

both being complete. 'We shall often have Y = X.
The mapping T is a pairing of points x of X with points y of Y, say (x, y).
Here to every x of X is ordered a unique y of Y. To Xi ^ x2 correspond the
two values yi and y2, which may or may not be distinct, in fact every x e X
may be mapped on the same point y0 £ Y. The mapping is onto (a
surjection in the Bourbaki language) if every point of Y is the image y of
at least one x in X. It is (1,1) (read "one to one") if
x, x2 implies yi = T(x0 ^ T(x2) = y2. (1.3.1)
The mapping is bounded if there exists a finite M such that
dlTM, T(x2)] ^ Md(\i, x2). (1.3.2)
This is a generalized Lipschitz condition and implies continuity of T(x)
with respect to x.
If X and Y are linear vector spaces over the same scalar field, and if

TietxXx + a2x2) = a,T(xi) + a2T(x2), (1.3.3)

then T is called a linear transformation. It is bounded iff
</[T(x),0]^Md(x,0). (1.3.4)
The most important case is that in which X and Y are complete normed
linear vector spaces, in which case the spaces are called Banach spaces
after the Polish mathematician Stefan Banach (1892-1945), who termed
them B-spaces. In this case (1.3.4) takes the form

||r[x]HM||x||. (1.3.5)
If T is a linear transformation, then T(0) = 0 and the transformation is
(1,1) if
T(x) = 0 implies x = 0.

If X and Y are B-spaces, the set E(X, Y) of linear bounded transforma-

tions on X to Y is also a B-space under the norm

||r|| = suP||r[x]||, d.3.6)

the supremum being taken with respect to all elements x of X of norm 1 . The
6 INTRODUCTION

algebraic operations in E(X, Y) are defined in the obvious manner by

[T, + T2][x] = r,M + r2[x], (aT)[x] = aT[x]. (1.3.7)
If Y = X, we write E[X] for E[X, X] and note that products are definable in
the obvious manner by

(TlT2)[x]=T1(T2[x\). (1.3.8)
This gives
lir.rdl^llrjinrji. /1.3.9)
It may be shown that E(X, Y) and E(X) are complete in the normad metric,
so they are B-spaces. Also, E(X), which is a normed algebra, actually is a
B-algebra since it is a B-space and satisfies (1.3.9).
If T G E(X) and is (1, 1), there is an inverse transformation T~l such that
T-1[T[x]] = x, Vx, T[r-1[y]] = y ify=T[x]. (1.3.10)
EXERCISE 1.3

1. 1.2:
Show 1. that E(Cn) is complete. Use any of the metrics for C" listed in Problem

2. If T is a linear transformation, verify that T(0) = 0. Here on the right stands

the zero element of Y, while on the left we operate on the zero element of X.
[Hint: T(0 + 0) = T(0).]
3. If T is (1, 1), why does T(x) = 0 imply x = 0 and vice versa?
4. Why is (1.3.6) a norm? Show that it is the least value that M can have in
(1.3.5).
5. Prove (1.3.9).

1.4. LINEAR TRANSFORMATIONS ON C"

INTO ITSELF; MATRICES

The simplest of all linear transformations are those which map C" into itself.
If T is such a transformation, then T is uniquely determined by linearity and
its effect on the basis of Cn. Any set of n linearly independent vectors would
serve as a basis, but we may just as well use the unit vectors

e,=(fc), (1.4.1)
where 8jk is the Kronecker delta, that is, the vector whose jth component is
one, all others being zero. This gives
x = x,e, + x2e2+- • - + xnen, (1.4.2)
if x = (x i , jc2, . . . , xn ) in the coordinate system defined by the vectors e,. Now
TRANSFORMATIONS ON C"; MATRICES 7

T takes vectors into vectors, so there are n 2 complex numbers aik such that
T[ek] = alke, + a2ke2 + - • • + anken, fe = 1, 2, . . . , n. (1.4.3)
The linearity of T then gives

T[x]=T(txkek) = txkT[ek]
or

(1.4.4)
T[*] = 'Z (E «^)e, = y,
from which we can read off the components of the vector y.
The quadratic array
' a2n
an an
021 «22
(1.4.5)

is known as a matrix — more precisely, the matrix of the transformation T

with respect to the chosen basis. We can now write T symbolically as

y = T[x] = si • X, (1.4.6)
where the last member may be considered as the product of the matrix d
with the column vector x, the result being the column vector y.
We have to decide when the mapping defined by T is (1, 1). Here the
condition T[\] = 0 implies that x = 0 now takes the form that the
homogeneous system

X ajkxk =0, / = 1,2, . . . , n, (1.4.7)

must have the unique solution

jc, = x2 = - - - = xn = 0.
This will happen as long as
det (d) * 0. (1.4.8)
In this case the mapping is also onto, since for a given vector y we can
solve the system

k2**-*
=l / = 1,2,...,b, (1.4.9)

uniquely for x = (jc,, x2, . . . , xn). It follows that T has a unique inverse,
8 INTRODUCTION

also an element of E[C"], i.e., a linear bounded transformation of C" into

itself. With this transformation goes a matrix si~\ which we refer to as the
inverse of si. The fact that its elements may be computed from (1.4.9)
shows that the element in the place (/, k) is AkjIA, where Ajk is the
cofactor of ajk in the determinant A = det (si).
We can define algebraic operations and a norm in the set Wn of n-by-n
matrices in terms of which the set becomes a Banach algebra. This
follows from the fact that there is a (1,1) correspondence between the
linear transformations T in E(C) and their matrices. Then to Tt + T2, aT,
and T,T2 correspond
(ajk + bik) = si + ®, (1.4.10)
(aaik) = ad, (1.4.11)

(2 ajmbmk) = si®. (1.4.12)

A number of different but equivalent norms may be defined for Wn. A
suitable one for analysis is

M| = max
i k2= \ \aik\. (1.4.13)
We have then

Since is complete in the normed metric (why?), it follows that %Rn is a

B-algebra.
We have seen that si has an inverse iff det (si) # 0. If this is the case, si
is said to be regular; otherwise, singular. Together with the given matrix
si we consider the family of matrices

where A runs through the complex field C and % = (8jk) is the n-by-n unit
matrix. These matrices are normally regular, but there exist n values of A
si, which \ % - si is singular: the n roots of the characteristic equation of
for

det(A^-5l) = 0. (1.4.14)
The roots Al9 A2, . . . , A„ form the spectrum cr(si) of si. They are known as
characteristic values, latent roots, or eigenvalues. For these values of A
one can find vectors xk in C" of norm 1 such that
si-xk =Xkxk. (1.4.15)
The characteristic vectors xk are linearly independent and may be chosen
so that they form an orthogonal system; in this case the inner product
FIXED POINT THEOREMS

(x,y) = S^/=0 (1.4.16)

for x = \k,y = \m,k^ m. This holds even if (1.4.14) has multiple roots. A
matrix d is singular iff zero belongs to the spectrum.

EXERCISE 1.4

1. Find the elements of sT1 when d is regular.

2. Verify the inequality for the norm of the matrix product.
3. Prove the Hamilton-Cayley theorem, which asserts that the matrix d satisfies
its own characteristic equation.

1.5. FIXED POINT THEOREMS

The Dutch mathematician L. E. J. Brouwer proved in 1912 that a

continuous map of the unit ball in R" into itself must necessarily leave at
least one point invariant. Such a point is known as a fixed point, and an
assertion about the existence of fixed points is known as a fixed point
theorem. We shall prove some theorems of this nature. We start with a
theorem proved by S. Banach in his Krakow dissertation of 1922. It refers
to mappings of a complete metric space by a contraction, i.e., a bounded
transformation of the space X into itself such that
d[T(x),T(y)]^kd(x,y), (1.5.1)
where k is a fixed constant, 0 < k < 1. Such a mapping evidently tries to
shrink the object. Banach's theorem states that there is a point which does
not move.

THEOREM 1.5.1

// T is a contraction defined on a complete metric space X, then there is

one and only one fixed point.

Proof. The triangle inequality plays a basic role here. We start with an
arbitrary point x,GX and form its successive transforms under T:
xn+l=T(xn), n = \,2,.... (1-5.2)
These elements form a Cauchy sequence, and X being a complete metric
space, Jt0 = lim xn exists and is to be proved to be a fixed point— in fact,
the only such point. Now it is sufficient to prove that, given any e >0,
there is an N such that

d(xn,xn+p)<e, n>N, p = l,2,.

10 INTRODUCTION

To this end we note that by the triangle inequality the left member does
not exceed

d(xn, xn+i) + d(xn + 1, xn+2) + • • • + d{xn+p^x, xn+p).

Now, using the contraction hypothesis, we see that

d(xm,xm+i)^kd(xm-uxm)^- • ^km-1d(xux2)
and thus

d(xn, xn+p)^ {kn X 4- kn + • • • + kn+p l)d(xu x2)

<YZjd(xux2).
This expression can be made as small as we please by choosing n large
enough. Hence {jc„} is a Cauchy sequence regardless of the choice of xu
knX
and the limit jc0 exists. Since
xn+\ = T(xn ),
we conclude that

jc0 = \imxn+\ = lim T(xn) = T[\im xn] = T(jc0),

where we have used the continuity of T. It is seen that jc0 is indeed a fixed
point.
Suppose that y0 is a fixed point. Then

d(x0, y0) = d[T(x0), T(y0)] ^ kd(x0, y0).

Since k < 1, this implies that d(x0, y0) = 0 or y0 = x0, so there is one and
only one fixed point. ■
The restriction to contraction operators is a drawback, but it can
occasionally be avoided by observing the following.

COROLLARY

There is a unique fixed point if some power of T is a contraction.

Proof If Tm is a contraction, then there exists a fixed point jc0 such that
Tm(Xo) = x0. We have also
lim(rw,r(jc1) = jc0
for any choice of jc,. Here we set jc, = T(x0) and find that

(Tm)n lT(x0)] = T[(Tmy(x0)] = T(jco).

When n becomes infinite, the first member tends to jc0, so we have
T(xo) = x0 or T admits x0 as a fixed point. Since any fixed point of T is a
FIXED POINT THEOREMS
11
fixed point of Tm and the latter has a unique fixed point, it follows that jc0
is the unique fixed point of T. ■
Vito Volterra (1860-1940), in his discussion of integral equations with
variable upper limits of integration, proved the uniqueness of the solu-
tions. From his work in the 1890's we can distill a fixed point theorem
which is rather useful.

THEOREM 1.5.2

Let X be a B -space. Let z0 be a given element of X, and let S belong to

E(X) and be such that

2l|Sl|<oo.
0 (1.5.3)

Then the transformation

T(x) = z0 + S[x] (1.5.4)

has a unique fixed point x0 given by

x0 = Zo+En=l Sn[z0]. (1.5.5)

It is enough to observe that the series converges in norm by (1.5.3) and
S can be applied termwise to the series and shows that T(x0) = jc0. The
uniqueness proof is left to the reader.
Consider, in particular, Volterra's equation
Jo
/(O = *(*)+ [ K{s, t)f(s)ds. (1.5.6)
Here the kernel K(s,t) and g(t) are known and f(t) is to be found. We
consider the particular case in which the kernel is a function of s alone.

THEOREM 1.5.3

Suppose that g(t)GC[0,a] and K(s)GL(0,a). Then the equation

f(t) = g(t)+ JoPK(s)f(s)ds (1.5.7)

has a unique solution in C[0, a], namely,

f(t) = g(t) + J* K(s) exp £ K(u) du]g(s) ds. (1.5.8)

The verification is left to the reader.
12 INTRODUCTION

COROLLARY

For K(t) = K the equation

(1.5.9)
f(t) = g(t) + K ff(s)ds

has the unique solution

f(t) = g(t) + K Jof0 exp [K(t-s)]g(s)ds. (1.5.10)

EXERCISE 1.5
1. Verify Theorem 1.5.2.
2. Verify Theorem 1.5.3.
3. Verify the corollary.
4. If T[f](t) = K /J f(s) ds, find T*. Is T a contraction? What are the conditions?
For a given interval [0, a] find m so that Tm is a contraction.
5. Let C+[0, a] denote the set of nonnegative functions continuous in [0, a]. The
mapping

is clearly a mapping of X into itself. Is it a contraction? Find Tm, and study its
contractive properties. Find a fixed point by inspection.

1.6. FUNCTIONAL INEQUALITIES

This is a field of increasing importance. We shall consider inequalities of
the form

/(f)*rifl(o. (1.6.1)
We consider a complete metric space X, the elements of which are
mappings from some interval [a, b] in R1. Here T is a mapping of X into
itself, and the problem is to discuss the inequality. Can it be satisfied by
elements / of X? Is it trivial in the sense that it is satisfied by all /'s in X?
If neither of the above is true, characterize the elements of X for which
(1.6.1) holds. Is it so restrictive that it holds for one and only one /? It can
be seen that there are a number of pertinent questions.
The discussion in Section 1.5 leads to several functional inequalities
which are categorical or determinative in the sense that there is a single
element of the space under consideration which satisfies the inequality. If
in (1.5.7) and (1.5.9) we assume g to be identically zero, then / is
identically zero. This suggests
FUNCTIONAL INEQUALITIES 13

THEOREM 1.6.1

Let X be the positive cone of C[0,a], 0<a<oo. Let K(t)GL(0,a),

continuous and nonnegative on the half -open interval (0, a]. 7/fGX and
if for 0 ^ t ^ a
Jo
f(t)< ftK(s)f(s)ds, (1.6.2)
then i is identically zero.
Proof. This functional inequality is very important for the theory of
differential equations (DE's) since it underlies uniqueness proofs based
on a Lipschitz or, more generally, a Caratheodory condition. We shall
obtain the theorem as a consequence of more general theorems, but in
view of its importance it is desirable to give a direct short proof. Set

Jo
F(t)= [ K{s)f(s)ds. (1.6.3)

This is an element of X and F(0) = 0. Furthermore, for 0 < t

F'(t) = K(t)f(t)^K(t)F{t),
so that (1.6.2) implies that

F'(t)-K(t)F(t)^0. (1.6.4)
This we multiply by the positive function exp [-Jo' K(u) du], and the
result is an exact derivative so that

£{F(t)exp[- ^ K(u)du\}*z0.
Since F(0) = 0, this shows that F(t)^0 for 0<t ^a. But we already
know that F(t)^0. To satisfy both inequalities we must have F(t) = 0.
This implies that /(O = 0, as asserted. ■
A uniqueness theorem due to Mitio Nagumo (1926) goes back to the
following functional inequality:

THEOREM 1.6.2

Let X be the subspace of C+[0, a], the elements of which satisfy f(0) = 0,
//mhiof(h)/h = 0. Then, if f GX, and if

f(t)^Pf(sA
Jo !>
then
f(t) = 0. (1.6.5)
14 INTRODUCTION

A proof may be given using (1.6.3) with K(s) replaced by s~\ which is
not integrable down to the origin. The proof is left to the reader.
There are several other uniqueness theorems which also go back to
functional inequalities. We shall not pursue these cases any further but
instead proceed to the use of fixed point theorems in the discussion of
functional inequalities. We have

THEOREM 1.6.3

Let X be a complete metric space which is partially ordered in such a

manner that if {xn} is an increasing sequence in X, so that xn ^ xn+, for all
n, and if limn^ xn = x0 exists in the sense of the metric, then xn ^ Xo for all
n. Let T be an order -preserving mapping of X into X such that Tm is a
contraction for some m. Let f0 be the unique fixed point of T. Then

f^T[f] implies f^f0. (1.6.6)

Proof We say that T is order-preserving if for /,, /2EX

/,*/2 implies TU^^TUil (1.6.7)

Suppose that / E Xo, the subset of X for which the inequality is meaning-
ful. Since Xo contains f0 at least, it is not void. Then

f^T[f]^T2[f]^---^T"[f]^---.
Now Tn [/] tends to the limit /0 as n -» <» for

klim
-» oo Tkm[P(f)] = fo, / = 0,l,...,m-l,

since Tm is a contraction and the limit is the same for all elements of X. It
follows that the increasing sequence {Tn[f]} converges to /0. Since order
is preserved under the limit operation, we have f^fo and the theorem is
proved. ■
If T[f] is defined by (1.6.3), T is order-preserving since the kernel K(s)
is nonnegative and the space X is linear. Here T usually does not define a
contraction, but all powers Tm with a sufficiently large m are contractions
for

Tn[f](t) = K(u) duj 'fis) ds, (1.6.8)

the norm of which goes rapidly to zero as n becomes infinite. This
provides another proof for Theorem 1.6.1.
We can also apply the Volterra fixed point theorem to functional
inequalities.
 15
FUNCTIONAL INEQUALITIES

THEOREM 1.6.4

Let X be a partially ordered B -space such that the positive cone X+ is a

closed point set. Let S be a linear bounded positive transformation on X
to X and such that

nSHSn||<oo.
=1 (1.6.9)

Let % be a given element of X+, and f0 the unique fixed point of

T[f] = g + S[f]. (1.6.10)

Then
f^g + S[f] implies f^f0. (1.6.11)

Proof. That S is positive means that it maps X+ into itself. All the
powers of S are then also positive and S is order-preserving. The
existence of a unique fixed point follows from Theorem 1.5.2. We have
lim„.^c Tn [f] = f0 for any /— in particular, for an / satisfying the first
inequality under (1.6.11). Now we have

Tn[f]^g + S[g] + S2[g] + • • • + Sn~l[g] + Sn [/]. (1.6.12)

This is an increasing sequence which goes to the limit f0 as n becomes
infinite. This, combined with / =s T[f] and the order-preserving properties
of limits, leads to the desired result. ■
We state a couple of applications of this theorem which are of special
importance to the theory of DE's.

THEOREM 1.6.5

Let KeC+(a,b)DL(a,b), and let g and f belong to C+[a,b]. Suppose

that for all t in [a, b]

(1.6.13)
f(t)^g(t) + £K(s)f(s) ds.
Then

f(t)s=g(t) + J' K(s)exp £K(u)dujg(s)ds. (1.6.14)

The proof is left to the reader. We see that if g(t) = 0; then /(f) = 0 and
Theorem 1.6.1 is again obtained. It is worth while stating the case
K(t) = K as a separate result.
16 INTRODUCTION

THEOREM 1.6.6

// K(t)^K and f and g are in C+[a,b], then

(1.6.15)
f(t)*£g(t) + K Pf(s)ds
implies that
(1.6.16)
f(t) ^ g(t) + K f' exp [K(t - s)]g(s) ds.
// g(t) is also a constant, then

f(t)^g + K ff(s)ds (1.6.17)

implies that
(1.6.18)
f(t)^gexp [K(t-a)].
The inequalities listed under Theorems 1.6.5 and 1.6.6 are known as
GronwalVs lemma after the Swedish-American mathematician Thomas
Hakon Gronwall (1877-1932), who found a special case in 1918 when
investigating the dependence of a system of DE's with respect to a
parameter. Gronwall was a pupil of Gosta Mittag-Leffler (1846-1927),
who also taught Ivar Bendixson (1861-1935), Ivar Fredholm (1866-1927),
Helge von Koch (1870-1924), and Johannes Malmquist (1888-1952). All
of them will figure somewhere in this treatise; the first three were among
the teachers of the present author. A theorem proved by Bendixson in
1896 (see Theorem 2.8.2) may be regarded as a forerunner of Gronwall's
lemma.

EXERCISE 1.6
1. Prove Theorem 1.6.2.
2. Prove the following analogue of Theorem 1.6.5 for Nagumo's kernel. Let X
be the space of functions / defined on [0,a] so that (i) /EC+[0, a], (ii)
lim(i0/(O = 0, (iii) lim,i0/(O/' =0, (iv) f(t)/t2EL(0,a). Define a norm in
terms of which X becomes a complete metric space. If / and g belong to X (g
fixed), then

f(t)^g(0+( f(s)s ]ds implies f(t)^g(t) + t( g(s)s 2 ds.

3. Verify (1.6.8).
4. Prove Theorem 1.6.5.
5. Prove Theorem 1.6.6.
6. Prove the following theorem. If / and g E C+[a, b], and if
HOLOMORPHIC FUNCTIONS
17
then

/(*) g(t) + K Jfa sinh [K(t - s)]g(s) ds.

7. Show that the inequality
f(t)^-\-[f(t)V
is absurd for real- valued elements of C[a, b].
8. Show that the inequality 4/(0 =s 3 + [/(f)]4 is trivial for real-valued elements of
C[a,b].
9. Show that 6/(60 ^ 3/(30 + 2/(20 is determinative for / £ C+[0, a], 0 < a ^
+ oo. Show that there are solutions unbounded for approach to zero of the form
r " if p > p0, the positive root of a certain transcendental equation. There are
also solutions of the form - t~q for q < p0. Can such solutions be combined? If
so, how?

II. Analytical Structures

1.7. HOLOMORPHIC FUNCTIONS

There are three essentially different approaches to analytic function

theory, associated with the names of Bernhard Riemann (1826-1866),
Augustin Louis (Baron) Cauchy (1789-1857), and Karl Weierstrass (1815-
1897), respectively. Riemann's approach was largely geometric and
physical (potential theory). His conformal mapping theorem will figure
later, particularly in Chapter 10, but in the main our approach will be a
mixture of that of Cauchy and that of Weierstrass.
The complex variable z — x + iy is represented geometrically by the
points of R2, the euclidean plane, so that to z0 = x0 + iy0 in C corresponds
the point (x0, y0) of R2. We have
z = r(cos0 + / sin0) = reie, (1.7.1)
where

r = +(jc2 + y2)I/2, tan<9=^. (1.7.2)

We write x = Re (z) (read is the real part of z"), y = Im (z) (read "y is
the imaginary part of z"). Furthermore, r = |z| is the absolute value of z,
and 6 = arg z is the argument of z, which is determined only up to
multiples of 27r. We have

|z, + z2| ^ |z,| + |z2|, \az\ = \a\\z\, |z,z2| = |z,| |z2|, (1.7.3)
so that (assuming a higher standpoint) C is a B-algebra, the simplest of the
species.
18 INTRODUCTION

The equation |z - z0\ = r represents a circle in the complex plane with

center at z0 and radius r. The interior and the exterior of this circle are
given by the inequalities |z - z0| < r and |z - z0| > r, respectively. The
former will often be called a circular disk in the following discussion and
is a circular neighborhood of z0. The upper half-plane is given by
Im (z) > 0; the left half -plane, jc < a, by Re (z) < a. If lim |z0 - z„ | = 0, we
say that lim z„ = z0 and {zn} is a Cauchy sequence. The complex plane is a
complete metric space in terms of the norm defined by the absolute value.
Every bounded infinite point set 5 has at least one limit point (theorem of
Bolzano- Weierstrass). A set which contains its limit points is said to be
closed. A set G is open if, whenever z0 E G, there is an e-neighborhood,
|z - z0| < e, the points of which belong to G. We shall call a set S in the
complex plane connected if for any two of its points, z,, z2 say, there is a
polygonal line joining z, with z2 all the points of which are in S. An open
connected set is called a domain and will be denoted by D. The term
region is also used, but we reserve this term for sets which have a
nonempty interior Int (5) ^ 0 and contain some of their boundary points.
So far we have dealt with the finite plane. We extend this by adjoining
an ideal point: the point at infinity. Stereographic projection makes it
possible to visualize this addition. Take a sphere of radius { tangent to the
z-plane at the origin, the "south pole" on the sphere. Join the point z in the
plane with the "north pole" by a straight line. This line has a second
intersection P with the surface of the sphere, and P is taken as the spherical
representative of z. The finite plane is mapped on the sphere omitting the
north pole, and the latter becomes the representative of the point at infinity.
The chordal distance between two points, Zi and z2, in the plane is the length
of the chord joining their representatives, Pi and P2, on the sphere. Chordal
distances are bounded, being equal at most to one.
We have now a sufficient set of notions in terms of which we can define
holomorphic functions. Let D be a domain in the z-plane and consider a
(1, 1) mapping of D into C. This mapping T defines a function z »-» /(z),
where /(z) denotes the image of z. The mapping is continuous at
z = z0 E D if for every e-neighborhood of /(z0), there is a 6-neighborhood of
z0 whose image is restricted to an e-neighborhood of /(z0). We can say that
lim f(zn ) exists and equals /(z0) if lim zn = z0. Or, again, Cauchy sequences in
the z-plane correspond to Cauchy sequences in the w-plane, w =/(z).
We now go one step further. Suppose that /(z) is defined and
continuous in D, and that

lim|[/(z + /i)-/(z)] (1.7.4)

exists and is independent of the manner in which h -^0. Then /(z) has a
HOLOMORPHIC FUNCTIONS
19
unique derivative everywhere in D. Such a function is said to be
holomorphic in D.
The usual rules of the calculus apply. Thus sums, constant multiples,
and products of differentiable functions are differentiable. A quotient is
differentiate at all points where the denominator is not zero. This enables
us to assert that polynomials in z are holomorphic in the finite plane,
while rational functions are holomorphic except at the zeros of the
denominator. In particular, a fractional linear function,

z^az_±b=w ad _bc
cz + d

is holomorphic except at z = - die. The mapping of the extended z-plane

into the extended w-plane is (1, 1), and the family of circles and straight
lines in the z-plane goes into the family of circles and straight lines in the
w-plane. The formula contains three essential constants and is completely
determined if the images of three points are given.
Suppose that z = x + iy and w = u + iv. For the existence of a unique
derivative of w = f(z) it is necessary (but not sufficient) that the Cauchy-
Riemann differential equations hold:

du _ dv du _ dv /t _

This says that the real and the imaginary parts of /(z) have partial
derivatives with respect to the real and imaginary parts of z. If in addition
the partials are continuous, /(z) is differentiable at z = z0 or in D, as the
case may be. As will be seen later, a holomorphic function has derivatives
of all orders. In particular, w(jc, y) = Re [/(z)], v (jc, y) = Im [/(z)] have
second partials with respect to x and y, and these functions are
(logarithmic) potential functions so that Laplace's equation holds:

A d2U d2U A A d2V , d2V n ,inn\

A"saF+a? = 0' A^aP+5? = 0- (L7-7)
We can define the exponential function exp(z) or ez for complex
arguments z in various ways, say by the exponential series or by the
property that it equals its derivative and is one for z = 0. We can also set
ez - exeiy = ex(cos y + / sin y). (1.7.8)
The last member has continuous partial derivatives of all orders, and the
first order partials satisfy the Cauchy-Riemann equations. Hence this
convention defines a function holomorphic in the finite plane which
agrees with ex on the real axis (y = 0). We shall see later that such an
20 INTRODUCTION

extension by a holomorphic function is unique. We see that ez is periodic

with period liri.
Formula (1.7.8) shows that we can define a logarithm by

log z = log (reie) = log r + id, (1.7.9)

and this is a holomorphic function in any simply connected domain
(roughly without holes) which does not contain the origin. The point z = 0
is obviously singular since the real part (i.e., log r) goes to - <*> when r -> 0.
Moreover, the function is not single valued in any domain containing the
origin since the argument is then free to take on infinitely many values for
a fixed z. There are two ways to cope with this difficulty. One method is to
restrict the variability of z by introducing a cut, not to be crossed, say
along the negative real axis, and define log z so that - tt < Im (log z) < tt.
Instead, we can extend the domain of definition to be a surface with
infinitely many sheets, with passage from one sheet to the next along the
negative real axis: up (increasing values of 0) by crossing from positive
values of y to negative, down by going in the opposite direction.
We have of course the same difficulties with roots. The nth root of z
has n determinations. If z = re16, then

zxln = rlln exp ^ (6 + 2tor)i], k = 0, 1, . . . , n - 1. (1.7.10)

Here the extended domain of definition is a Riemann surface with n
sheets, and passage from one to the next is across, say, the negative real
axis. Here we run into visual difficulties. By going around the origin n
times in the positive sense, we are in the top sheet and one more turn
takes us back to the first sheet.

EXERCISE 1.7
1. What is arg (z,z2)?
2. Construct the sum of two vectors, z, and z2, in the complex plane.
3. The vector z = x - iy is known as the complex conjugate of z. Show that
|z|2 = zz. Find arg z.
4. Mark the four points 0, z,, z, + z2, and z2, and draw the parallelogram with
these points as vertices. Prove the parallelogram law (the sum of the squares
of the lengths of the diagonals equals the sum of the squares of the lengths of
the sides).
5. From the definition of an ellipse derive its equation in complex variables.
6. Show that chordal distances define a metric for C.
7. Derive the Cauchy-Riemann equations and Laplace's equation.
POWER SERIES 21

8. Verify that log r is harmonic, i.e., satisfies (1.7.7). Is any power of r

harmonic?
9. Verify that (1.7.5) maps circles and lines into circles and lines.

1.8. POWER SERIES

Power series served as the foundation of the function theory developed

by Weierstrass. He did not invent them, but he perfected their theory. Let
there be given a sequence {an} of complex numbers; form the series

(1.8.1)

Suppose that the series does not diverge

2 "nZn. for all z ^ 0. The basic observa-
tion of Weierstrass is

THEOREM 1.8.1

Suppose that there is a z0 ¥■ 0 such that the sequence {|anz£|} is bounded.

Then the series (1.8.1) is absolutely convergent for |z| < |z0|, uniformly for
|z|^|zo|-5, 8 >0.
This means that the positive reals fall into two classes:

1. Ki contains those numbers r for which {|a„|r"} is a bounded

sequence.
2. K2 contains the remaining positive numbers.
There is a number R which is the supremum of the numbers in class Kx
(equivalently, the infimum of the numbers in K2), and now it is seen that
the series converges absolutely for \z\ < R and diverges for |z| > R. For R
we have the expression

(1.8.2)
— = lim sup |a„|!/".

The sequence {\an |1/n} may have a single limit, in which case we have \/R.
This was the case considered by Cauchy. The sequence may, however,
have more than one limit point, even infinitely many. In any case there is a
largest limit point, and this is the superior limit. The general formula is
due to Jacques Hadamard (1865-1963). This R is the radius of convergence
of the power series. It can take any value in [0, <»].
If R > 0, the sum of the series is evidently a continuous function of z in
the circle of convergence \z\ < R, say f(z). Moreover, we can differentiate
22 INTRODUCTION

term by term to obtain the power series

/,(z)^2
n = 1 nanzn X (1.8.3)

with the same circle of convergence. Furthermore, for |z|^J?0<fl,

|z + h\^R0 we find that

^[/(z + /0-/(z)]-/i(z) 2 n(n-l)|fln|l?r2, (1.8.4)

which goes to zero with \h\. It follows that /(z) is differentiate and
f'{z) = fx(z). It follows also that the sum of the power series is a
holomorphic function of z. Using the same technique, one sees that /(z)
has derivatives of all orders for |z| < R.
The function z >-»/(z) can be expanded in a Taylor series about any
point a in the circle of convergence, and the resulting series, a power
series in z - a, has a radius of convergence Ra :
R -\a\^Ra ^R + \a\. (1.8.5)
Weierstrass obtained this by elementary expansions and rearrangements.
We have

zn = (z - a + a)n = 2 (£)(z - a)V"\ (1.8.6)

Substitution of this in (1.8.1) gives a double series,

^an±(l\z-afan\ (1.8.7)
This series is absolutely convergent for |z - a \+ |a| < R. Now, an abso-
lutely convergent double series can be summed by columns as well as by
rows and, in fact, by any process that is exhaustive. Here (1.8.7) is the
sum by rows. Summing by columns, we get a power series in z - a, where
the coefficient of (z - a)k is

^ 2 n(n - 1) . . . (n - k + \)anan k = (1.8.8)

so that

f(z)=tf-^TT}(z-a)\ (1.8.9)

and this representation is guaranteed to hold for |z - a \ < R - \a\.

Now this is a power series with a radius of convergence Ra which may
very well exceed R -\a\. In the circle of convergence, |z - a \ < Ra, (1.8.6)
defines a holomorphic function of z; this function coincides with /(z) in
POWER SERIES 23

the lens-shaped region of intersection of the two disks, where they are
defined, and the union of the two disks is the domain of definition of a
single holomorphic function which is represented by (1.8.1) in one disk
and by (1.8.9) in the other. We can repeat this process for all points a with
\a \ < R. We obtain a family of power series (1.8.9) which represent the
same holomorphic function in their domains of convergence. The union
of all the disks is the domain of definition that can be reached by direct
rearrangements. There is at least one point on the circle of convergence
|z| = R which stays outside of all disks \z - a\<Ra with \a\<R. This is a
singular point, and every power series admits of at least one singular point
on its circle of convergence.
It is possible for all points on the boundary to be singular. This will
happen iff Ra = R - \a \ for all a with \a \ < R. We then say that z = R is
the natural boundary of f(z). An extreme example of this phenomenon is
furnished by the series

fiz)=thz2n (1.8.10)
with R = 1. Here the series, as well as all the derived series, converge
absolutely on |z| = 1. The unit circle nevertheless is the natural boundary.
Such natural boundaries are typical for so-called lacunary series, where
there are long and widening gaps in the expansion.
The German mathematician Alfred Pringsheim (1850-1941) proved
that, if the coefficients an are nonnegative and infinitely many are
positive, then z = R is a singular point of the function defined by the
series. Since this theorem has important applications to the theory of
DE's, we shall sketch a proof. If the theorem were false, then for an a < R
but close to R the series (1.8.9) would be convergent for a value of z on
the real axis beyond z = R ; i.e., the series

k=0
2 (z - aR)k n=k \fC / an{aRr\
E (?) 0 < a < 1,
for a suitable choice of a near to one would converge for a z = R( \ + S)
for some 8 > 0. Thus

S(i-« + a^s(?)fl.w*
would converge. But this double series has positive terms, so we can
interchange the order of summation, leading to the absurd conclusion that
the series

2 *„[(! + 5)jR]"
24 INTRODUCTION

would converge. It follows that for 0 < a < R we have always Ra = R, so

z = R is a singular point.

EXERCISE 1.8
1. Verify that /(z) and its derived series /,(z) have the same radius of
convergence.
2. How is (1.8.2) obtained?
3. Give examples of power series where R = 0, 1, or °°.
4. Verify (1.8.4).
5. Fill in missing details in the proof of (1.8.9).
6. Why does (1.8.8) hold?
7. For the function z h-> f(z) defined by (1.8.10) the point z = 1 is singular by
Pringsheim's theorem. Show that z = - 1 is singular and that z = / and z = -i
are singular. Extend to 2"th roots of unity, the union of which is dense on the
unit circle.
8. Fill in missing details in the proof of Pringsheim's theorem.
9. In a power series (1.8.1) all the coefficients an, where n is not a multiple of 3,
are zero while a3k ^ 0 with infinitely many larger than zero. What can be said
about singularities on |z| = Rl
10. Why is Ra ^R +|a|?

1.9. CAUCHY INTEGRALS

Our three founders of analytic function theory differed in almost all
respects and not least in their attitude toward publication. Weierstrass,
who was a perfectionist and late in gaining acclaim, published sparingly;
his fame rests mainly on his lectures at the University of Berlin,
1864-1892. Riemann was a man of genius but was shy and plagued by
poor health; a considerable part of his work was published after his death.
Cauchy, on the other hand, overwhelmed the periodicals with his notes.
His first publication on integration between imaginary limits dates from
1825, but he seems to have had the basic ideas as early as 1814. Consider a
function z ^/(z), holomorphic in a simply connected domain D. Set
f(z)= U(z) + iV(zl (1.9.1)
and define the integral

j22F(z)dz = jZ\udx - Vdy] + ij 2[Udy + V dx], (1.9.2)

where the path of integration is a curve joining zx with z2 in D and the
integrals are line integrals in the sense of the calculus. Since U and V
 25
CAUCHY INTEGRALS

satisfy the Cauchy-Riemann equations (1.7.6), Cauchy claimed that the

integral is independent of the path joining Z\ and z2, or, in other words, the
integral along a closed contour C is 0:

L f(z)dz=0. (1.9.3)

There are many questionable points in this argument, and objections go

back to the 1880's; desirable precision and generality were reached
around 1900. It is required that the curve C have an arc length (= be
rectifiable) which requires a representation

z = z(t), O^t^L, (1.9.4)

where L is the length of C and t^z(t) is a continuous function of
bounded variation. The integral then becomes a so-called Riemann-
Stieltjes integral:

L f[z(t)]dz(t), (1.9.5)
which is the limit of Riemann-Stieltjes sums,

2/[*to»[*tt)-*0f-i)]- (1-9-6)
j=0 continuous integrand f and integrator z of
The limit exists for any
bounded variation.
In this setting one proves the theorem for / = 1 and f = z. Then one
observes that (i) C may be approximated arbitrarily closely by a closed
polygon, (ii) a polygon may be triangulated, (iii) the theorem is proved for
a small triangle, and (iv) hence is true for a polygon and an arbitrary
rectifiable curve.
The integral is additive with respect to the path and linear with respect
to the integrand. If D is not simply connected, the integral along C equals
a sum of integrals around the holes that are inside C.
The Cauchy integral is an exceedingly powerful tool. If z *-* f(z) is
holomorphic in a simply connected domain D and on its rectifiable
boundary C, then

1 f f(t) dt ^ f 0 if z is outside C, (19 7)

277-/ Jc t - z l/(z) if z is inside C.
We can differentiate under the sign of integration as often as we please,
and the formal nth derivative represents /(n)(z):

-« J JSfi*^, (L9.8)
26 INTRODUCTION

Thus a holomorphic function has derivatives of all orders, a property

proved for power series in Section 1.8. It has been observed that the
property of defining a holomorphic function of z inside C resides, not in
the factor f(t) in the integral, but in the Cauchy kernel l/(t - z), for we
can replace f{t) by any continuous function F(t) without losing analytic-
ity of the integral, which, however, may not be zero outside of C in this
case. It is the Cauchy kernel which is a holomorphic function of z as long
as z is kept away from the contour of integration C. We can expand the
kernel in powers of z or of 1/z, multiply by f(t), and integrate term by
term, as is usually permitted by uniform convergence of the series. The
linearity in / often implies continuity with respect to /. Among the many
results obtainable by such considerations we list

THEOREM 1.9.1

// z »-> f(z) is holomorphic in D, // the disk |z - a| < R lies in D, then f(z)

can be expanded in the Taylor series

f(z) = 2^(z-a)k, (1-9.9)

k=o K!
the series being absolutely convergent in the disk.

Proof. We have r <~ fF^lU ~X~^ ^ pI^

t-z {t-a)-(z-a) £0(t-a)k+1' J Kl'*M}

which converges uniformly in z and t if |z - a \ ^ R - <5, \t - a \ = R.
Multiplication by f(t) and termwise integration yields (1.9.9) in view of
(1.9.8), where z is replaced by a. ■
In a similar manner one obtains the Laurent series, discovered by Pierre
Alphonse Laurent (1813-1854) in 1843 and known to Weierstrass in 1841.

THEOREM 1.9.2

// zi-»f(z) is holomorphic in an annulus.

0^R,<|z-a|<R2^, (1.9.11)
then

f(z)= | an(z - a)", an = ^ Jc (1.9.12)

where C: |t - a| = r, R, < r < R2.
We shall not give the proof but mention that in addition to (1.9.10) there
is needed
CAUCHY INTEGRALS
= -2 27
1
(1.9.13)
t-z . n =0

which is valid for \z - a\>\t - a\.

Cz-a)n+
The case in which z = a is an isolated singularity is particularly
important. Here R{ = 0, and the negative powers in (1.9.12) constitute the
principal part of the singularity. There are three different possibilities.
1. No negative powers. We define f(a) = a0; the singularity is remov-
able.
2. A finite number of negative powers, an = 0 for n <-m but a_w ^ 0.
This is a pole of order m, and (z - a)mf(z) is holomorphic at z = a.
3. Infinitely many negative powers. Here z = a is an essential singu-
larity. In any neighborhood of z = a the function z »->/(z) assumes any
preassigned value c infinitely often with at most two exceptions (theorem
of Emile Picard).
The property of being holomorphic may be said to be hard to acquire,
but once acquired it persists. It can be expected to survive a passage to
the limit. The simplest case is

THEOREM 1.9.3

Suppose that {fn(z)} is a sequence of functions, holomorphic in a domain

D, which converges uniformly to a function f(z) in D; then f(z) is
holomorphic.
The functions holomorphic in D form a normed algebra under the sup
norm ||/|| = supzeD |/(z)|. Convergence in the norm is uniform con-
vergence in the ordinary sense and the algebra is complete, so the
theorem follows. This case is almost trivial, but we can greatly weaken
the assumptions by using induced convergence. Here is an example.
Instead of assuming convergence in a domain, we can assume it in a
subset from which it spreads to the whole domain.

THEOREM 1.9.4

Let (fn(z)} be a sequence of functions holomorphic in a domain D. Let C

be a simple, closed, rectifiable oriented curve which, together with its
interior, lies in D. Suppose that the sequence {fn(t)} converges uniformly
with respect to t on C. Then there exists a function z »-> f(z) holomorphic
in the interior of C such that fn(z) converges to f(z) uniformly in the
interior of C. Moreover, if S is any subset of the interior of C having a
positive distance from C, and if p is any positive integer, then the
sequence {fnp)(z)} converges uniformly to f(p)(z) in S.
28 INTRODUCTION

We shall not prove this theorem, but we call attention to the fact that
the principle of the maximum (see below) implies that a Cauchy sequence
{fn(t)} on C is also a Cauchy sequence inside and on C.
A zero of /(z) is by definition a point where the function is zero. It is of
order m if the Taylor expansion starts with the term am(z - a)m, am 5* 0.
The zeros of a holomorphic function can not have a cluster point in the
interior of a domain where f(z) is holomorphic except when the function
is identically zero. If z = a is a limit point of zeros of /(z), then in the
Taylor expansion

/(z) = a0+ ax(z - a) + a2(z -a)2+ • • •

the constant term a0 = f(a) is zero by the continuity of the function. But
then

fl(z) = al + a2(z-a) + a3(z-a)2 + ' • •

also has infinitely many zeros with z = a as a limit point. This forces ax to
be zero and so on; all coefficients are zero, and /(z) is identically zero.
This implies that, if two functions /(z) and g(z) holomorphic in a
domain D coincide for infinitely many values of z with a limit point in D,
they are identical in all of D, for h(z) = /(z) - g(z) has infinitely many
zeros and is thus identically zero. This is known as the identity theorem.
Instead of zeros we may of course consider any other fixed value of the
function. We see that limit points of zeros or of a value c are singular
points of the function.
The calculus of residues occupied a central position in Cauchy's work.
Suppose that z = a is an isolated singular point of a function z >->/(z) in
the neighborhood of which f(z) is single valued. There is then an
associated Laurent expansion (1.9.12). The coefficient is the residue of
/(z) at z = a. The reason for the name is that a.x is all that is left when we
form

where C is a small circle, \t - a \ = r, for we can substitute the Laurent

series and integrate termwise since the series is uniformly convergent on
the circle. We have

Jc
f (t-a)n]dt = r "i (""" exp (- niB) d6 (1.9.14)
Jo
and the integral is zero unless n — 0, when it equals 2m. This gives the
residue theorem.
CAUCHY INTEGRALS

THEOREM 1.9.5 29

If f(z) is holomorphic in a simply connected domain D except for isolated

singularities at z = s, s2, . . . , sn, then

(1.9.15)

where r, is the residue of f(z) at z = s} and C is the boundary of D

supposed to be rectifiable.
This is so because the integral along C equals the sum of the integrals
around the small circles surrounding the singularities sh In all these
formulas the integrals are taken in the positive sense.
An important consequence is

THEOREM 1.9.6

// f(z) is holomorphic inside and on C except for poles, then

(1.9.16)

where Zf is the number of zeros, and Pf the number of poles inside C.

Proof. It is assumed that neither zeros nor poles are located on C. The
integrand is then holomorphic inside and on C except for simple poles at
the zeros and poles of f(z). At a zero the residue equals the multiplicity of
the zero, whereas at a pole the residue is the negative of the multiplicity of
the pole. The conclusion then follows from Theorem 1.9.5. ■
The integral in (1.9.16) can be evaluated directly since the integrand is
the derivative of

log /(f) = log |/(0| +/arg/(f).

Here the real part returns to its original value when z returns to the
starting point after having described C once in the positive sense. The
imaginary part, however, does not necessarily return to its initial value
but will differ from it by a multiple of 2tt. From Theorem 1 .9.6 we then get
the so-called principle of the argument :

THEOREM 1.9.7

Under the assumptions of Theorem 1.9.6 the increase in the argument of

f(z) after C has been described once in the positive sense is 2ir(Z{- Pt).
30 INTRODUCTION

A useful addition to Theorem 1.9.6 is given by

THEOREM 1.9.8
Under the assumptions of Theorems 1.9.6 and 1.9.7, suppose in addition
that g(z) is holomorphic inside and on C. Then

dri I g(t) IW dt = S 8(aj) " 2 8(bk)' °'917)

where the summation is extended over the zeros a, and poles bk of f(z),
and each summand is repeated as often as the multiplicity of the zero or
pole requires.
The zeros and poles of f(z) are still simple poles of the integrand, and at
a zero of / of multiplicity fij the residue is jXjg(aj), and similarly at the
poles.
Cauchy's formula (1.9.7) invites some comments. Replace z by z0, and
let the path of integration be the circle t = z0 + reie, where 6 goes from
zero to 2tt. The result is
Z.TT Jo
f(z0) = Y- I*"f(Zo+reie)dd. (1.9.18)
The right-hand side is the mean value in the sense of the integral calculus
over the interval (0, 2tt) of the integrand. This is a basic property of
holomorphic functions but is shared with harmonic (logarithmic) potential
functions. From (1.9.18) we also get

|/(z0)|^ max |/(2)| for|z-z0|=r. (1.9.19)

In this relation the inequality normally holds, equality can hold iff |/(z)|
equals its maximum for all z — to start with, all z on the circle and
ultimately all z in the plane, and /(z) = M eiay where M is the maximum
and a is real, fixed. This is a form of the principle of the maximum. The
principle asserts that the absolute value of a holomorphic function /(z)
cannot have a local maximum unless it is a constant. If in 3-space we plot
the surface

u = \f(x + iy)\2, (1.9.20)

where / is not a constant, if /(jc0 + ry0) = /(z0) # 0, there are paths on the
surface leading from z = z0, u = u0 along which u is strictly increasing,
and also paths along which u is strictly decreasing. The latter type of path
naturally is missing if u0 = 0. We formulate a form of the principle which
is sufficient for our purposes.
ESTIMATES OF GROWTH

THEOREM 1.9.9 31

// f(z) is holomorphic inside and on the rectifiable curve C, and if M(f, C)

is the maximum of |f(z)| on C, then for all z inside C
|f(z)| ^ M(f , C). (1.9.21)

Proof The use of Cauchy's integral below is due to Edmund Landau

(1877-1938), a German mathematician who made profound contributions
to function theory and analytic number theory. We have

1/(2)]* ==M
LTTl Jc y^-dt,
t— Z (1.9.22)
whence
\f(z)\k «s[2mf(z, C)r'L[M(/,C)f.
Here d(z, C) is the distance of C from the point z in its interior, and L is
the length of C. We extract the A:th root and pass to the limit withTlo
obtain (1.9.21). ■

EXERCISE 1.9

1. If C is rectifiable, then 2"=i \z(tj)- z(tj-{)\ ^ L and the sums in (1.9.6) are
dominated by M(f, C)L. Verify.
2. Verify (1.9.7) and (1.9.8). For n = 1 verify that the difference quotient
(l//i)[/(z + h)- f(z)] tends to the formal derivative uniformly if z and z + h
have a distance from C which exceeds a 5 > 0.
3. Fill in details in the proofs of Theorems 1.9.1 and 1.9.2.
4. Use the principle of the maximum to prove that a Cauchy sequence {fn{t)} on
C generates a Cauchy sequence {/„(z)} in the interior.
5. Verify Theorems 1.9.5 and 1.9.6.
6. Show that a harmonic function, not a constant, cannot have a local maximum.

1.10. ESTIMATES OF GROWTH

Given a power series

/(2)=2 anZn (1.10.1)

with radius of convergence R > 0, let M(r, f) be its maximum modulus :

M(r,f)= O«0<2tt
max .|/(re")|. (1.10.2)
Formula (1.9.8) then gives the Cauchy estimates,

\an\^M(r,f)r~n, r < R. (1.10.3)

 32 INTRODUCTION

Here r is arbitrary, so the question of optimizing the estimate arises

If
#<oo, the choice r = [\ -(Mn)]R is often a good one.
The maximum principle shows that for a nonconstant holomor
phic
function M(r, /) is an increasing function of r. For R < oo it may very well
be bounded. This is not the case for R = oo, however, as shown by
Joseph
Liouville (1809-1882). A sharper form of his theorem is

THEOREM 1.10.1

Suppose that R = oo and that there are positive constants A, B, and c such
that
M(r, f) ^ A + Brc (1.10.4)
for all r. Then z >-> f(z) is a polynomial in z of degree not exceedi
ng c.
Proof. By (1.10.3) we have

k|^G4 +Brc)rn,
and this goes to zero as r^oo if „>c. This means that all coefficients
an
are zero for
as asserted. n ■> c, so z ^ f(z) is a polynomial of degree not exceeding c
A power series with R = oo is called an entire function (integral
function
in Great Britain). Such a function may be algebraic or transcendental
,
according as it is a polynomial or not. In the transcendental case
M(r f )
grows faster than any power of r by Liouville's theorem, but there
is
neither a slowest nor a fastest possible growth of the maximum modulu
s
Given an increasing function r^G(r) such that log G(r )/log
r becomes
infinite with r, one can find an entire function whose maximum modulu
s
grows faster than G(r) and an entire function whose maximum modulu
grows slower than G(r), so that s

respectively.
More properties of entire functions will be encountered in Chapters
4
and 5. In addition to entire functions, we shall also ultimately have
to
consider meromorphic functions, i.e., functions having no other
sing-
ularities than poles in the finite domain. Again we have two types,
algebraic and transcendental meromorphic functions. The first is the class
of rational functions, functions which are meromorphic in the extende
d
plane. A function of the second class either has a finite number of poles
plus an essential singularity at z = oo 0r infinitely many poles which have
no limit point in the finite part of the plane.
ANALYTIC CONTINUATION; FUNCTIONAL EQUATIONS 33

In Chapter 4 we shall encounter various generalizations of the max-

imum modulus which are suitable for a study of meromorphic functions.
Here we shall consider just a generalization of Liouville's theorem for an
isolated singular point.

THEOREM 1.10.2

Suppose that z = a is an isolated singularity of f(z), where f(z) is

holomorphic in a punctured disk, 0 < |z - a| < R. Set

Ma(r,f)= max |f(a + reie)|, 0<r<R. (1.10.5)

Suppose there exist positive numbers A, B, and c such that

Ma(r,f)^A + Brc. (1.10.6)

Then z = a is either a removable singularity or a pole of order =^c.
Proof The assumptions imply that there is a Laurent expansion (1 .9. 12),
and we have

\a-k \ =s£ Ma(r,f)rk <(A + Br~c)rk ;

this goes to zero with r if k > c. It follows that z = a is either a removable
singularity or a pole of order not exceeding c, as asserted. ■
Exercises bearing on this section will be found after Section 1.11.

1.11. ANALYTIC CONTINUATION; PERMANENCY OF

FUNCTIONAL EQUATIONS
Analytic continuation is a concept introduced by Weierstrass and basic
for his attack on function theory. A function z f(z) is defined originally
by a power series, say,

/(z)=2 anzn (1.11.1)

with a radius of convergence R. If R = <» and the function is entire, there

is no continuation problem. Also, R =0 is out (we disregard the possibility
that the series may be summable by some method or other). If 0 < R < oo,
there is a continuation problem. We saw in Section 1.8 that f(z) admits of
expansions in a Taylor series,

f{z\a)=^f-^f-{z-a)\
k=o K I (1.11.2)

obtained by direct rearrangement of (1.11.1) after setting z = a +(z - a)

and expanding. The series f(z;a) converges for \z - a\< Ra, where
 34 INTRODUCTION

R - - \a\ « R « i? + |fl|. if |fl|, the point of contac( of , _ a . =

with \z\ = R is a singular point of /(z) and analytic
continuation in the
direction arg z = arg a is not possible. On the other hand,
if R - U < r
the disk \z-a\<R is partly outside |z|<K, and
overhang /(z; a) defines an analytic continuation in the lens-shaped
of /(z). This process is
repeated for all a with \a\<R. If for all such a 's we get R
= R - |fl I n0
analytic continuation is possible and \z\ = R is
the natural boundary of
If on the other hand, Ra >R -|a| for some values of
a, the union of
the disks \z-a\<R. is a simply connected domai
n D, in which our
function is defined by one of the series f(z;a) with
|a | < R. Moreover if a
point z„SD„ it belongs to infinitely many disks \z-a\
corresponding series /(z ; a ) all assign the same value <Ra and' the
to f(z) at z = z0 We
now repeat the process for points a, in Z>, with |a,| s
J?. This gives a set of
power series /(z;a,a,) obtained by double rearra
ngements- of the
original series at z = a with \a\<R, and of the series
/(z; a) at z = a,
This gives a definition of / in the union of all the disks.
Since this set may
be self-overlapping, we are no longer
definition but have to keep account ofassure d of the consistency of the
the steps involved
Suppose that we can find a sequence of points a0,
a„...,a„ such that
power series m terms of z - a, are obtained by repeat
of (1.11.1). Suppose that ed rearrangements

N<K, \al-a0\<Rac, \an-aK.l\<R^_,.

Here (U1.3)

/i(z) = S%(z-nf, \z-a,\<R,.

There is then defined a branch of /(z) at z = a„

obtained by analytic
continuat.on of f(z) using the intermediary points
a„, a„ . . . a„ Accord-
ing to Weierstrass, rte f0fa/(ry o/ 5«cA smes con5//r«/«
an analytic
function f(z). To these regular elements of / are furthe
following: r adjoined the

1. Polar elements (z - a)- 2;=0 f,,(z - ay, one for

each pole
2.
branch Algebra
points.ic elements (z - a)-** 2,% c,(z - a)** for algebraic

JLfT"VV^m'> J, The$e may be regular: 27=0 d,z ', polar:

z 20 d,z ', or algebraic: z"p ^ diz'lklp .
Essential singular points are not considered as belong
ing to the domain
of definition of the function and contribute no functi
onal elements
Logarithmic singularities also are noncontributing.
At a given point z = a there are normally infinitely many
elements,
ANALYTIC CONTINUATION; FUNCTIONAL EQUATIONS
35
regular and/or singular. Henri Poincare (1854-1912) was a famous French
mathematician whose name will be encountered again and again in this
book. At this stage we need a result of his according to which the distinct
elements of /(z) at z = a form a countable set. This we see as follows. An
element at z = a is the end product of a chain of rearrangements. We can
order these first according to the number of steps involved, say n. For
each n we have rearrangements at points a0, au . . . , an. Without loss of
generality we may assume that the a's used have rational coordinates
(these numbers are dense in C). Now the points with rational coordinates
form a denumerable set. Thus we are dealing with a countable number of
countable sets, and such a set is itself countable. Hence the different
determinations at z = a form either a finite set or a countable one.
Leaving these general considerations, we turn to the principle of
permanence of functional equations, which is a very important fact in the
theory of DE's. It is hard to find a definition of the term functional
equation, and the mathematicians who deal with the subject normally
exclude differential and integral equations from consideration, so we
cannot expect much help from that quarter. Before going any further we
have to prove the double series theorem of Weierstrass, which will be
needed here and later.

THEOREM 1.11.1

Suppose that the functions z »-» fn(z), n = 0, 1, 2, . . . , are holomorphic in a

disk |z - a| < R and that the series

|)fn(z)-f(z)
n=0 (1.11.4)

converges uniformly in |z-a|^pR for each p with 0<p < 1. Suppose

also that

fn(z) = k=0
iak,n(z-a)k. (1.11.5)
Then each of the series

2akn-Ak,
n=0 k = 0,l,2,..., (1.11.6)

converges and, for |z - a| < R,

f(z)=2Ak(z-a)k.
k=0 (1.11.7)

Proof. This follows from Theorem 1.9.4, which implies that the sum of
the series (1.11.4) is a holomorphic function /(z), that the series may be
36 INTRODUCTION

differentiated term by term p times, and that the resulting sum of pth
derivatives is the pth derivative of the sum/(z), where p is any integer. In
particular, we have convergence for z = a and this implies the
assertions. M
The name double series theorem refers to the fact that under the stated
conditions the order of summation in the series

2 [2 akAz-a)k] (1.11.8)
may be interchanged.
We shall prove a special form of the principle of permanence of
functional equations, which is sufficient for most of our needs and is
connected with our work on functional inequalities in Section 1.6. There it
was found that, if X is a partially ordered metric space with elements /, if
T is a mapping of X into itself, and if T or one of its powers is a
contraction with fixed point g, then the elements of X which satisfy the
inequality
f^T[f] (1.11.9)
also satisfy
f^g, whereg = 7[g]. (1.11.10)
Now this is a functional equation and obeys the principle of perma-
nence ifthe data are given an analytical form. We consider a function
T(z, w) of two complex variables given by the series

T(z, h0 = y=o
|) k2=0 a,zJV, a00 = 0. (1.11.11)

It is assumed that there exist values of 2 and u>, different from (0, 0), for
which the series is absolutely convergent, say z = a, w = b. These
numbers a and b may be assumed to be positive. The reader should
observe that a double power series does not have a unique radius of
convergence, though there exist pairs of associated radii: if you lower
one, you can increase the other. Our assumption implies that the terms of
the series are bounded for z = a, w = b,

\aik\a'bk ^M. (1.11.12)

By an obvious generalization of Theorem 1.8.1 this implies that the series
(1.11.11) is absolutely convergent for |z| < a, \w\ < b. Next, suppose that
we have found a function

*«-*/(*)- S «m (1.11.13)
m =1
such that (i) there exist a pair of numbers s < a,t <b and (1.11.13) is
ANALYTIC CONTINUATION; FUNCTIONAL EQUATIONS
37
absolutely convergent for \z\^s and \f(z)\ ^ t for such values, and (ii) if
T[z, /(z)] - F(z) for |z| < then /(z) = F(z) or
f(z)=T[z,f(z)]. (1.11.14)
We have then the following principle of permanence of functional
equations:

THEOREM 1.11.2
1/ T satisfies the conditions just stated, and if the solution f(z) as well as
the composite function F(z) = T[z, f(z)] can be continued along the same
path from the origin, then all along the path (1.11.14) holds ; i.e., the
continuation of the solution is the solution of the continuation of the
equation.
Proof. We assume that the continuation involves a chain of disks D0,
Du . . . , Dn, where D0 = [z ; |z| < s] and the center a, of D, lies in D,_i. In
Dj we have functional elements f(z) and Fj(z) = T[z, fj(z)] represented
by convergent power series in z - ah Now, in D0 we have /(z) = F(z), and
this equality holds at z = ax and in some neighborhood thereof. But the
identity theorem then requires that /!(z) = Fi(z) in D, — in particular, at
z = a2 so that /2(z) = F2(z), and so on. ■
The reader should notice that this is not an existence proof. It is by no
means clear that our assumptions are strong enough to guarantee the
existence of a solution. All that is claimed is that, // we have found an
analytic solution and we can continue it analytically together with the
right member of the equation, then it remains a solution.
It should also be observed that (1.11.10) is a rather special case of a
functional equation and that the law of permanence holds in much more
general situations.

EXERCISES 1.10-1.11

1. For a power series (1.10.1) R = 1 and M(r,/)<(1 - r)~c, where c 2*0, find a
realistic estimate for \an\.
2. If z h> (1 - 2 )_c, where c ^ 0, find how fast the binomial coefficients grow.
3. If z»->/(z) is a transcendental entire function and M(r, /) < exp (Arc),
where c and A are positive, find a realistic estimate of the Maclaurin
coefficients, an.
4. If z •— > /(z) has a pole of order m with leading coefficient a-m, how fast does
Ma(r,f) grow?
5. Is z •— > tan 7rz a transcendental meromorphic function? Where are the poles?
Show that they are simple.
38 INTRODUCTION

6. Show that cos z = ^[exp (iz) + exp (- iz)]. Show that for this function the
maximum modulus is attained on the imaginary axis and equals cosh r (the
hyperbolic cosine of r).
7. Show that |sin (jc + iy)\2 = sin2 jc + sinh2 y.
8. If ez = 1 +27 zn/n !, prove by series multiplication or otherwise that ez =
ezae\

9. If P(t) = a + bt + ct2, express 2~=0 P(n)zn ln \ in terms of the exponential

function.
10. Where in the complex plane is sin (x + iy) real, and where purely imaginary?
11. What is the limit, if any, of tan (x + iy) as y -» + oo, y - oo?
12. What is the series expansion for /(z, a) if /(z) = (1 - z) 1 and \a\ < 1? How
does Ra vary with a?
13. Show that in this case the domain D, where the first rearrangements
converge is bounded by a cardioid with the cusp at z = 1 and passing through
z =-3.
14. Prove that the same curve gives the boundary of D, for all the binomial
functions (1 -z)c, where c is not a positive integer.
15. Take c =\, and consider second rearrangements f(z\a,ax). Try to form
some notion of what part of the plane is covered by D2 and show that there
are two distinct elements of the function with center at z = 2.
16. Consider the analytic function z »-> [(1 - z)m - (1 + z)1/2]1/4. Where are the
branch points? Describe the functional elements with centers at the branch
points. What are the elements at infinity? The roots are given all possible
determinations.
17. Two power series have radius of convergence = 1. Express the product of
the two series as a simple series, and determine its radius of convergence.
Can it be greater than one?
18. When are the following series absolutely convergent: (i) ST(zw)", (ii)
2oU + w)n, (iii) 2o2o z'z"?
19. Let T(z, w) = \[z + 2? (z + w)n ]. Let / = T(z, f) be the solution which is 0 at
z = 0. Show that the solution is holomorphic except for two branch points.
Find their locations and natures.
20. State the conditions under which the principle of permanence would apply to
the functional equation involved in Problem 1.6:6.

LITERATURE

As collateral reading for Chapter 1 the author may perhaps be permitted

to refer to three of his own books:

Hille, E. Lectures on Ordinary Differential Equations. Addison-Wesley, Reading, Mass.

1969.
Methods in Classical and Functional Analysis. Addison-Wesley, Reading, Mass. 1970.
LITERATURE
39
Analytic Function Theory. Vols. I, and II, 2nd ed. Chelsea, New York, 1973; 1st ed.,
Ginn, Boston, 1959.
These books will be referred to hereafter as LODE, MCFA, and AFT,
respectively.
For Section 1.1 consult MCFA, Chapters 1 and 2, or LODE, Chapter 1.
The latter can be used also for Sections 1.2-1.6. Matrices and mappings
are discussed in both books.
For Sections 1.5 and 1.6 consult Chapters 5 and 12 of MCFA. For
differential and integral inequalities the basic work is:

Walter, W. Differential- und Integralungleichungen. Springer Tracts on Natural Philosophy,

Vol. 2. Springer-Verlag, Berlin, 1966.
The original Gronwall Lemma occurs in:
Gronwall, T. H. Note on the derivative with respect to a parameter of the solutions of a
system of differential equations. Ann. Math., (2), 20 (1918), 292-296.
See also:

Bendixson, I. Demonstration de l'existence de 1'integral d'une equation aux derivees

partielles lineaires. Bull. Soc. Math. France, 24 (1896), 220-225.

For this question see also Chapter 3 of LODE.

For Part II consult AFT. For Section 1.7 Chapters 1-4 will round out
the presentation. Section 1.8 and Chapter 5 go together. For Section 1.9
consult Chapters 7-9. Functions of bounded variation and the Riemann-
Stieltjes integral are treated in Appendix C. Analytic continuation is
discussed in Chapters 5 and 10 of Vol. II, where also the principle of
permanence of functional equations is treated under different assump-
tions.
The standard treatise on functional equations is:
Aczel, J. Lectures on Functional Equations and Their Applications. Academic Press, New
York, 1964.
2

EXISTENCE AND

UNIQUENESS

THEOREMS

This chapter is devoted to generalities about differential equations (DE's)

and their solutions. The main subject matter will be existence and
uniqueness theorems for analytic DE's: the use of fixed point methods,
successive approximations, majorant methods, the Cauchy majorant, the
majorant of Lindelof, and dominants and minorants. We shall also
consider variations of parameters, internal and external.

2.1. EQUATIONS AND SOLUTIONS

Our first question is, What is meant by a DE and by a solution? A crude

definition of a DE would be a functional equation (also undefined!)
involving derivatives of unknown functions. Such derivatives may be
ordinary or partial, according to whether the number of independent
variables is one or more. We restrict ourselves to functions of a single
variable. This does not exclude the possibility that the equation involves
one or more parameters which occur explicitly in it.
We have now excluded partial DE's from consideration. But not every
functional equation involving derivatives of the unknown with respect to
a single variable (which need not occur explicitly in the equation) is a
bona fide ordinary DE. To clarify the ideas, let us consider the following
equations from the point of view of their being acceptable for study in this
book:
W'(2) = /0(Z) + /1(2)W(Z) + /2(Z)[W(Z)]2, (2.1.1)
w'(z) = z2w(z), (2.1.2)
zw"(z) = [w(z)]\ (2.1.3)
40
 41

EQUATIONS AND SOLUTIONS

w'(z)=w(z + \), (2.1.4)

JO0 (2.1.5)
w'(z)= l'{\ + [w(s)r}ds,
(2.1.6)

The notation is intended to suggest that the functions involved are

analytic functions of a complex variable. The first three equations are
acceptable; the others are not. The first is a nonlinear first order (Riccati)
equation, discussed in Chapter 4, the second is a trivial first order
equation (Chapter 5), and the third is a nonlinear second order equation, a
special case of Emden's equation (Chapter 12). Equation (2.1.4) is
rejected because the unknown is subjected to a translation of the
independent variable besides the differentiation. The remaining equations
are excluded because of the integral operators. Equation (2.1.5), however,
leads to a bona fide second order DE:

w>"(z) = l + [w(z)]2 (2.1.7)

upon differentiation. However, the two equations are not equivalent: all
solutions of (2.1.5) satisfy (2.1.7), but not conversely.
In the following we shall consider only ordinary DE's or systems of
such equations. Up to and including Chapter 4 the equations will be of the
first order and of the form

w'(z) = Flz,w(z)]. (2.1.8)

Here there are two alternatives: F(z, w) is an analytic function of the two
variables z and w, holomorphic in a given dicylinder in C\ or (second
alternative) w(z) is vector valued in C", z is a complex variable, and F is a
holomorphic mapping of C"+1 into C". The latter alternative enables us to
treat nth order DE's as a special case of (2.1.8). Note that a DE is of
order n if the order of the highest derivative entering into the equation is
n.
The notion of what is meant by a solution of a DE has varied
considerably over the three centuries that have elapsed since Isaac
Newton made the first classification of what he called fluxional equations
in 1671. He even invented the method of series expansions with indeter-
minate coefficients. However, since Newton's work was not published
until 1736, long after his death, it had no influence on the development of
the subject. The first published paper on DE's is due to Jacob Bernoulli in
his study of the isochrone (1696); this was rapidly followed by a number
of investigations by Johann and Daniel Bernoulli, Clairaut, Euler,
42 EXISTENCE AND UNIQUENESS THEOREMS

Leibniz, Riccati, and others during the eighteenth century. To these early
founding fathers of calculus and analysis, the problem was to find a
function of two variables {z,w)y-* G{z,w) such that

Gz(z, w) dz + Gw(z, w)dw=0 (2.1.9)

is implied by (2.1.8) and vice versa. This would require

F(^) = -f&5- (2U0)

and the solution would be

G(z, w) = C. (2.1.11)

In (2.1.9) and (2.1.10) the subscripts z and w indicate partial differentia-

tion with respect to z and w, respectively. Since in those days the only
acceptable functions were those built up by composition of a finite
number of what later became known as elementary functions, i.e.,
powers, exponentials, logarithms, trigonometric functions, and their
inverses, the integrable types were soon exhausted. Just as most func-
tions could not be integrated, most DE's were insolvable.
Here Cauchy showed the way out of the difficulty. The founding fathers
had asked too much: they had demanded global solutions in terms of
elementary functions. This was an algebraic straitjacket that was not
suitable and prevented further progress. The typical operation of analysis
is the limit process, which is very powerful but can require a local point of
view.
Let us return to (2. 1 .8), Suppose that (z, w ) *-* F(z, w ) is holomorphic in
a neighborhood of (z0, w0). Cauchy asked, Is there a function of z, say

w = w(z; z0, Wo), (2.1.12)

such that it is holomorphic in some neighborhood of z = z0 and for all z in

the neighborhood, with w0 = w(z0; z0, vt>0),

w'(z;zo, w0) = F[z, w(z;z0, w>0)]? (2.1.13)

Cauchy showed the existence and uniqueness of such a solution. To this
end he developed several ingenious devices. For the real case he worked
out a step-by-step method using linear approximations. For the complex
domain he employed series expansions and the majorant method.
The global problem still remains. Here the method of analytic continua-
tion used by Weierstrass provided at least a partial solution. A local
solution given by a power series at z = z0 is continued analytically along a
 43
EQUATIONS AND SOLUTIONS

path C from z0 to z,. If F[z, w(z)] can also be continued analytically along
C, then by the law of permanence of functional equations (see Section
1.10) the continuation of the solution is the solution of the continuation of
the equation.
This leaves out of consideration the singularities. They are either fixed,
given by the equation, or movable, i.e., dependent on the initial values.
Thus (2.1.2) has only fixed singularities, namely, 0 and oo, whereas (2.1.1)
has both kinds. There the singularities of /0, fu fi and the zeros of f2 are
fixed. The movable singularities are simple poles. This is illustrated by the
special example
w' = \ + w2, (2.1.14)
which is satisfied by w = tan (z - z0) with poles at z = z0 + I(2m 4- \)tt.
Equation (2.1.3) has fixed singularities at z = 0 and oo and movable
singularities, which may be placed anywhere in the plane and are of very
complicated nature.
For a further study of (2.1.8) we must determine the fixed singularities
and the nature of the solutions at such a point. We must also decide why,
when, and where movable singularities occur and what the actual natures
of these singularities are.
Some equations can be characterized by postulating the position and
nature of the singularities together with the relations joining these
elements. This problem was first posed by Riemann for the hypergeomet-
ric equation. There is a Riemann problem for linear DE's, one of the
many Riemann problems. This involves global characterization of equa-
tions and solutions. Usually global information is hard to get; there are,
however, equations for which we can assert that the only movable
singularities are poles, and equations of which the solutions are single
valued.
A solution of (2.1.8) which involves an arbitrary constant is called the
general solution. A solution obtained by giving the constant a particular
value is known as a particular solution. Thus (2.1.14) has the general
solution tan (z - z0), and tan z is a particular solution. This equation has
also two singular solutions, w(z) = i and w(z) = -i. They are not
obtainable from the general solution by specifying the constant z0.

EXERCISE 2.1

Find the fixed singularities of the following DE's, and discuss how the nature of
the singularity varies with the parameters. All the equations are elementary.
1. zw' = aw.
2. (z - a)w' = (z - b)w.
44 EXISTENCE AND UNIQUENESS THEOREMS

3. (z-a)2w" + (z-a)w'-b2w = 0.
4. The functional equation (not a DE)

/'(*) = 3*JoP{l + [/(s)]2}Ws

has
l,/2 = solutions of the form f(z) = Cz (why?). Determine the value of C if
+ l.

Determine the DE of the lowest order which is satisfied by the following conies:

5. b2x2 + a2y2 = a2b2.

6. (x-a2)2 + (y-a)2 = a2 + a\
7. (x-af + (y-bf = 1.
Obtain the envelope of the curve family when it exists and show that it is also
a (singular) solution.
8. Find the general solution of w' = w2-\, and find the singular solutions.
Discuss the movable singularities (nature, frequency, and dependence on the
constant of integration).

2.2. THE FIXED POINT METHOD

Historically this is the most recent method proposed for proving exis-
tence and uniqueness theorems for ordinary DE's. It goes back to a paper
by G. D. Birkhoff and O. D. Kellogg in 1922. They used a fixed point
theorem by L. E. J. Brouwer. That Banach's fixed point theorem
(Theorem 1.5.1) could be applied for the same purpose was pointed out by
R. Cacciopoli in 1930.
We are concerned with the equation

w' = F(z,w), (2.2.1)

where (z, w)*-*F(z, w) is holomorphic in the dicy Under
D:\z = z0\^a, \w-w0\^b. (2.2.2)
It is required to find a function z *-* w (z ; z0, w0), holomorphic in some
disk |z - z0| < r ^ a, such that
w'(z; z0, Wo) = F[z, w(z; z0, w0)],
, v (2.2.3)
w(z0;z0, H>0)= w0.
To apply Theorem 1.5.1 we have to exhibit a complete metric space X
consisting of functions z g(z) and define an operator T which maps X
into itself and is a strict contraction. Also, X has to be chosen so that the
existing unique fixed point is the desired solution. As a first step we
THE FIXED POINT METHOD 45

replace (2.2.1) by the integral equation

/(z)=w0+f F[s,f(s)]ds. (22 A)

If this equation has a unique solution /, then /(z) also satisfies (2.2.1)
including the initial condition. This suggests defining an operator T by

T[g](z)= w0+ f F[s,g(s)]ds, (2.2.5)

where T operates on a space to be defined. We note first that F satisfies
two conditions used in the following, namely,

\F(z,w)\<M, (2.2.6)
\F{z,u)-F(z,v)\<K\u-v\, (2.2.7)
for suitably chosen constants K and M and (z, w), (z, w), and (z, v) in D.
Since D is closed, F is certainly bounded in D and so is Fw(z, w), the
partial of F with respect to w. The Lipschitz condition (2.2.7) is implied
by the boundedness of the partial derivative. We can now state and prove

THEOREM 2.2.1

Under the stated assumptions on F, in the disk D0: |z-z0|<r, where

r< min (a»^[»^)» (2.2.8)

(2.2.1) has a unique holomorphic solution satisfying the initial -value
condition w(z0; z0, w0) = w0.
Proof. Let X be the set of all functions z*-*g(z), holomorphic and
bounded in D0 so that

g(zo)=w0 and ||g - Wo|| f>, (2.2.9)

where

||g-Wo|| = sup|g(z)-w0|. (2.2.10)

Under these assumptions the composite function z ^ F[z, g(z)] exists

and is holomorphic in D0, and its norm is at most M. The holomorphism
follows from an applications of the double series theorem (Theorem
1.11.1) plus analytic continuation. This implies that z »-> T[g](z) is also
holomorphic (the integral from z0 to z of a holomorphic function is
holomorphic); it takes the value w0 at z = z0 and

\\T[g]-wjji*Mr*b
46 EXISTENCE AND UNIQUENESS THEOREMS

by (2.2.8). Hence T maps X into itself. The contraction property follows

from the Lipschitz condition, for we have

\T[g](z) - T[h](z)\ = |£ {F[s, g(s)] - F[s, h(s)]} ds

^ k|£ \g(s) - h(s)\ \ds\\* Kr\\g -h\\.

Since Kr < 1 by (2.2.8), it is seen that

llHg] - T[h]\\ ^ Kr\\g - h\\ = k\\g -hi (2.2.11)

where k - Kr < 1. Also, X is a metric space under the sup norm; it is
complete, for if a sequence {gn} of X is Cauchy, then lim g„(z) = g(z)
exists uniformly in D0. Furthermore, g is holomorphic in D0, g(z0) = w0,
and \\g - w0\\ =s b. Thus gGX and X is complete. All the assumptions of
Theorem 1.5.1 are satisfied, and we conclude that X has a unique fixed
point under T. This function z f(z) satisfies (2.2.4) and hence is our
required solution w(z;z0, w0) of (2.2.1).
By the proof of Theorem 1.5.1 we could start with any element g of X
and obtain the invariant element / as

/ = limr"[g]. (2.2.12)

The convergence of this sequence is comparatively slow. Convergence as

the geometric series 27 kn is all that this method gives. The method of
successive approximations of Picard gives convergence as an exponential
series and also dispenses with the obnoxious condition Kr < 1. ■
The fixed point method also applies to the vector case referred to above
in connection with (2.1.8). We state the result without proof.

THEOREM 2.2.2

Let (z, w) h-> F(z, w) be a mapping of Cn+1 into Cn defined in an (n + 1)-

cylinder :
D: |z - z0| =S a, |wi - w,0| b,, . . . , |w„ - wn0| bn, (2.2.13)
where w = (w,, w2, . . . , wn), and we set (w10, w20, . . . , wn0) = w0. We set
Fi.e.,
= (F,, F2, . . . , Fn). In Cn and Cn+1 we use the maximum coordinate norm,

||w|| = max |Wj|, ||(z, w)|| = max (|z|, |w,|, . . . , |wn|). (2.2.14)
Assume that F(z, w) is holomorphic in D and also that

max ||F(z, w)|| = M, ||F(z, u) - F(z, v)|| ^ K||u - v|| (2.2.15)

SUCCESSIVE APPROXIMATIONS
47
for (z, w), (z, u), and (z, v) in D. Define a disk D0: |z - z0| r, where

r<4tftf-W (2216)
TTien f/iere &risrs uniquely a vector function z »-» w(z; z0, w0), defined and
holomorphic in D0, smc/i f/ia/

and w'(z; z0, wo) = F[z, w(z, z0, w0)] (2.2.17)

w(z0; z0, wo) = wo. (2.2.18)
Note that the derivative of a vector function is the vector whose
components are the derivatives of the components of the given vector.
Similarly, by definition the integral of a vector function is the vector
whose components are the integrals of the components.

EXERCISE 2.2

For the following autonomous equations (i.e., DE's in which the independent
variable does not occur explicitly), find a value of r for which Theorem 2.2.1
guarantees a solution. Take z0 = 0.
1. w'{z) = w(z), W0 5* 0.
2. w'(z)= l + [w(z)]2.
3. [Optimization] What are the best values of r obtainable in Problems 1 and 2?
Would removal of the condition rK < 1 lead to better values?
4. What is the justification for the Lipschitz condition in Theorem 2.2.2?
5. Write a complete proof of Theorem 2.2.2.
6. Reduce w" - zw' + z2w = 0, w(0) = 0, w'(0) = 1 to vector form, and estimate r
by Theorem 2.2.2. Take a = 1, bx = b2 = b.

2.3 THE METHOD OF SUCCESSIVE APPROXIMATIONS

Going back in time to 1890, we find an important memoir by Emile Picard

(1856-1941), introducing the method of successive approximations. This
rapidly became the basic tool for proving the existence of solutions of
DE's as well as other functional equations including various forms of the
implicit function theorem. Basically it is the old method of trial and error,
which dates back to Sir Isaac Newton. There is a profound difference,
however: Picard systematized the trial and estimated the error.
We again consider
w' = F(z,w), w(zo)=w0. (2.3.1)
Here (2, w)*-* F(z, w) is holomorphic in the closed dicylinder

D:\z- z\^a,\w- w0\^b, (2.3.2)

48 EXISTENCE AND UNIQUENESS THEOREMS

and, with 1 K,
\F(z,w)\*M, (2.3.3)
|F(z, u) - F(z, v)\ **K\u-v\ (2.3.4)
for (z, w), (z, u), and (z, y) in D. We introduce a disk,

D0: |z - z0| ^ r with r < min (a, (2.3.5)

Note that the condition rK < 1 is no longer imposed.

THEOREM 2.3.1

Lfader ffte stated conditions (2.3.1) /ias a unique holomorphic solution in

Do.

Proo/. The desired solution must satisfy

w(z) = w0+ f F[s, ds. (2.3.6)

We now define a sequence of approximations wn(z) recursively by

w0(z)=Wo, wn(z)=Wo+\ F[s, Wn-iis)] ds, n>0. (2.3.7)

It should now be shown that the definitions make sense for z in D0. This
involves showing that the approximations exist as holomorphic functions
in D0 and that \wn(z)- w0\ ^ b. Suppose that this has been achieved for
n <m. The implication is that wm-x(z) exists in D0, where it is holomor-
phic and satisfies |wm_i(z)- w0\ ^ b. This in turn implies that the compo-
site function z »-> F[z, wm-x(z)] exists and is holomorphic in D0. It follows
that the integral exists and is a holomorphic function of z in D0.
Furthermore,

F[s, wm-x(s)] ds ^ M\z - Zo\ ^Mr<b, (2.3.8)

I/.:
so that wm(z) is holomorphic in D0 and |wm(z) - w0| ^ b. Thus the ap-
proximations exist for all n, are holomorphic in D0, and take the value w0 at
Z = Z0.
Convergence must now be proved. This follows from the Lipschitz
condition, which has not been used so far. Now we have

Mz) - w0\ ^ M\z - z0| ^ KM\z - z0|,

so that

w2(z) - w,(z) = Jzn

P {F[s, - F[s, wo]} ds
SUCCESSIVE APPROXIMATIONS
49
and

|w2(z) - wi{z)\ ^ k| j \wi(s)- w0\ |ds 11

^X2m|J2|s-20|Ms|
= K2M||z-z0|2.
[To evaluate the integral, set s = z0 + (z - z0)f, where f goes from zero to
one.] This suggests taking as induction hypothesis the assumption
K
(2.3.9)
\wk(z)~ H>*-i(z)|^ Myy|z - Z0|\
which holds for k = 2, giving

|wk+i(z)-wk(z)|= |J {F[s,H>k(s)]-F[s, wk_,(s)]}ds

^X|JjW'c(5)-H'k-i(s)||d5
|5-Zo|k Ids 11
= M
K k+1

Thus (2.3.9) holds for all k, and it follows that the series
oo
W(Z)= W0+ 2 [WnU)~ Wn_,(z)] (2.3.10)

converges absolutely and uniformly in D0. This implies also that w(z) is
holomorphic, at least in Int(D0), and continuous on the boundary.
Furthermore,

lim wn(z) = w(z), lim F[z, vv„_,(z)] = F[z, h>(z)],

and finally

u>(z) = w0+ F[s, w(s)] ds

as desired. This is clearly a solution of (2.3.1).

It remains to show that the solution is unique. This may be proved in a
number of different ways by invoking the Lipschitz condition again.
Suppose that /(z) is a solution; we may assume that it is also holomorphic
in D0. Again we have

/(z) = z0+ JPZn F[s, f(s)]ds.

50 EXISTENCE AND UNIQUENESS THEOREMS

From (2.3.7) we obtain ds,

J ZQ
(2.3.11)
f(z)-wn(z)= f {F[s,f(s)]-F[z, wn-M]}
whence

|/(2) " Wn(z)\ ^ K|£ 1/(5) - W.-t(5)| |<fe

^K2|£ |ds||£|/(f)-Wn-2(0||A|
(2.3.12)
= lH£ |s-z0| |/(5)- w»-2(s)| |ds
Repeated use of the same device finally gives

\f(z)-W„(z)\*Z-^y |£ |, -2„r W0\\dS

Here we may assume that |/(s)- w0\ ^ b, so that

(2.2.13)
|/(Z)-W.(z)|*^fc

and this goes to zero as n becomes infinite. Hence /(z) = w(z) and the
solution is unique, thus proving the theorem. ■
We can get another uniqueness proof from Theorem 1.6.1. If we have
two solutions, w(z) and /(z), their difference satisfies

f(z)-w(z)= P {F[s,f(s)]-F[s,w(s)]}ds, (2.3.14)

and if |/(z) - w(z)\ = g{z) then

g(s)|A
Here we set, with 6 = arg (z - z0), f

5 =Zo+/e''e, Z = Z0 «(*) = *(/), g(z) = h(u)

and obtain the inequality

Jo (2.3.15)
h(u)*sK\" h{t)dt.
Here h(t) is continuous and nonnegative. But by Theorem 1.6.1 the only
solution of this inequality in C+[0, a] is h(t) = 0. This gives g(z) = 0 and
/(z)-w(z).
MAJOR ANTS AND MAJORANT METHODS 51

Finally, the same type of argument applies to the vector case. We can
drop the condition rK < 1 there also. The result will be cited as
Theorem 2.3.2 in the following.

EXERCISE 2.3

1. Find a sequence of successive approximations for the case w' = w, w(0) = 1.

2. Consider w' = l + w2 with w(0) = 0. The successive approximations are
polynomials in z. Find the degree of the «th polynomial. Take w0(z) = z.
3. Prove that the only continuous nonnegative solution of (2.3.15) is identically
zero by the method of successive substitution:

Jo Jo Jo
0^h(u)^K( h(t)dt^K2 (" dt [ g(v)dv ^- • • .
4. C. Caratheodory in 1918 proved existence and uniqueness theorems, replacing
(2.3.3) by \F(z,w)\^MK(\z-z0\) and (2.3.4) by \F(s, r,) - F(s, t2)\
K(\s -z0\)\tl-t2\. Here K(u)^0 and K(u)G C(0, a] U L(0, a). Carry
through the proof, and use Theorem 1.6.1 for uniqueness and the method of
successive approximations.

2.4. MAJORANTS AND MAJORANT METHODS

Given the DE

w'(z) = F[z, w(z)l w(zo)=wQ, (2.4.1)

where F(z, w) is holomorphic in the dicylinder

D:\z- z0\^a,\w-w0\^b, (2.4.2)

it is natural to try to satisfy the equation by a power series

w - Wo = 2 cn(z - z0)n. (2.4.3)

Here the coefficients cn have to be chosen so that the series formally
satisfies (2.4.1). "Formally" means that, if the series is substituted in
(2.4.1) and the right-hand member is written as a power series in z - z0,
then for all k the coefficient on the right of (z - z0)k equals (k + l)ck+1,
which is the coefficient of (z - z0)k on the left. As we shall see, the ck 's
can be determined successively and uniquely.
This method of indeterminate coefficients was discovered by Isaac
Newton in 1671 and published in 1736. The method was rediscovered on
the continent and was commonly used by the analysts of the eighteenth
century. The procedure involves a number of doubtful steps, all of which
go back to the question of whether or not the resulting series is absolutely
52 EXISTENCE AND UNIQUENESS THEOREMS

convergent in some disk D0: |z - z0| < r. Such finicky questions did not
bother the founding fathers, who believed that a well-defined analytic
expression always has a sense. That this is not necessarily true was
discovered much later.
As a case in point we may take the equation

zh>' = (1 + z)h>-z (2.4.4)

with the formal series

2 (-\)nn\zn,
n=0 (2.4.5)

which diverges for all finite values of z. The equation can be solved by
elementary methods, and one solution is

e"j, (2.4.6)
which is actually represented asymptotically by the series.
Divergent series were banished from serious consideration by Abel and
Cauchy in the 1820's (they were resurrected by Poincare in the 1880's).
Cauchy, who had laid solid foundations for analytic function theory as
well as for the theory of DE's, also made the method of indeterminate
coefficients rigorous by his calcul des limites. In this connection limites
should be understood as "bounds" rather than "limits." It is a method of
majorants, a term introduced by Poincare, who also was the first to use
the symbol <^ for the relation between the given function and one of its
majorants.
Let

f(z)=2cnzn (2.4.7)
be a power series with a positive radius of convergence, and let

g(z)=tcnzn (2.4.8)
be a power series with nonnegative coefficients and the radius of
convergence R.

DEFINITION 2.4.1

g(z) is a majorant of f(z), symbolized by

if
f(z)<lg(z), (2.4.9)

lc„I^C„, Vn. (2.4.10)

MAJORANTS AND MAJORANT METHODS 53

This clearly implies that the radius of convergence of (2.4.7) is at least

R. The relation <^ is transitive, i.e.,
f<g and g<h imply f < h. (2.4.11)
LEMMA 2.4.1

// m is a positive integer and if f<g, then

[f(z)r <^ [g(z)r. (2.4.12)

Furthermore,
f(m>(z)<lg(m)(z), (2.4.13)

Jo Jo
[Zf(t)dt<^ fg(t)dt. (2.4.14)
Repeated integration also preserves the majorant relation.
The proof is left to the reader.
The majorant relation is also preserved by some composite mappings.

LEMMA 2.4.2
Let

h(w) = n=0
2 hnwn, H(w) = n=0
2 Hnwn with h(w) <^ H(w), (2.4.15)

the series being convergent for |w| < R0. Suppose that f<g, f(0) = g(0) = 0,
and |g(z)| < R0 for |z| < R. Then
h[f(z)HH[g(z)]. (2.4.16)

Proof Suppose that f and gh j = 1, 2, . . . , ra, are power series, the g, 's
being absolutely convergent for \z\<R, and suppose that

f,<gh / = l,2,...,m. (2.4.17)

Let au a2, . . . , am be arbitrary numbers, and let \a\ ^ Aj. Then

a1fi + a2f2+- • - + + A2g2 + • • • + Amgm. (2.4.18)

Since the right member is a power series in z convergent for \z\< R, the
same holds for the left member.
We now apply this to the case

fs(z) = [f(z)]\ gj(z) = [g(z)]\ a, = hh Aj = H, (2.4.19)

Then for each m

hj{z) + h2[f{z)Y+-'- + hm[i{z)T

< Hlg(z) + H2[g(z)]2 + ••■ + «. [g(z)]m. (2.4.20)

54 EXISTENCE AND UNIQUENESS THEOREMS

Here the right-hand member is the mth partial sum of the series 27 Hnwn
with w replaced by g(z). In the disk |z| «s R - 8 we can find an e >0 such
that \g(z)\ ^ R0- e. Since the series 27 Hn(R0- e)n converges, the series
27 Hn[g(z)]n converges uniformly in \z \^ R0- 8 to a holomorphic func-
tion L(z) with nonnegative coefficients in its Maclaurin expansion. The
series

K(z)=i>„[/(z)r
n=1 (2.4.21)

is also convergent for |z| < R0, and the majorant relation (2.4.20), which
holds for the partial sums, holds also in the limit so that K(z) < L(z), as
asserted. ■
The majorant concept extends to functions of several variables.
Suppose, in particular, that

F(z, w) = 2 J) c&'wk (2.4.22)

j =0 k =0
is a holomorphic function of (z, u>) in the dicylinder
D: \z\^a,\w\^b.
Suppose that

i =0 /c =0
is also holomorphic in D. We say that G is a majorant of F:

F(z, w) <^ G(z, w) if |cjfc |^ Cik, V/, fe. (2.4.24)

We now consider (2.4.1), where for simplicity we take z0 = w0 = 0. Here
the fundamental fact is that a majorant relation for the right-hand sides
implies the corresponding majorant relation for the solutions. We state
and prove

THEOREM 2.4.1

Let F(z, w) be defined by the series (2.4.22), and let G(z, w) be a majorant
of F(z, w) defined by (2.4.23) and (2.4.24). Suppose that
W'(z) = G[z,W(z)], W(0) = 0 (2.4.25)
has a solution

W(z) = 2 C>zs (2.4.26)

convergent for |z| < r. Let

w(z) = E c.zJ (2.4.27)

MAJORANTS AND MAJORANT METHODS 55

be a formal solution of
w'(z) = F[z,w(z)], (2.4.28)
Then w(0) = 0.
w(z) < W(z), (2.4.29)

and the series (2.4.27) is absolutely convergent for |z| < r and is the unique
solution of (2.4.28).
Proof The derivation of (2.4.27) involves a number of operations on
power series which are legitimate iff the series are absolutely convergent.
The series are "formal" if the required operations have been performed
without regard to legitimacy, the justification being given a posteriori
when it is found that the series are indeed absolutely convergent.
We have to form the composite series
oo oo r oo -|/c
(2.4.30)
;=0/c=0 Lp = l J

Here we start by forming the kth powers by using Cauchy's product

formula (valid for absolutely convergent series and leading to absolutely
convergent series). The kth series is multiplied by cjkz\ and the result is
summed for / and k. We then rearrange the result as a power series in z.
This can be justified with the aid of the double series theorem of
Weierstrass, provided that we have absolute convergence. In the result
the coefficient of zn will be a multinomial in the cjk 's and the cp 's, say,
Mn[cjk; cp]. (2.4.31)

Here the big question is, What cjk's occur, and what cp? To the first
question we can reply that a necessary condition for cjk to occur in Mn is
that j + k ^n. This is gratifying, for it means that cn+x depends only on a
finite number of the known coefficients c}ki at most \(n + l)(n +2) in
number. Furthermore, the only cp's that can enter are cu c2, . . . , c„, all of
which have been determined by the time the (n + l)th coefficient is
considered. This means that the coefficients cp can be determined
successively and uniquely. The numerical constants which enter when the
kth powers are formed are positive integers, a fact which is also
important. The series (2.4.27) is a formal solution if
(n + l)c„+i = Mn(cjk; cp) (2.4.32)
for all n. The first three coefficients are

C i — Coo>
Ci — \(c io + CoiCoo)> (2A33)
C3 = 2[C20 + C,,Coo+ C02(Coo)2 + koiC,o + 2(Coi)2Coo].
56 EXISTENCE AND UNIQUENESS THEOREMS

This means that all coefficients cp are ultimately determined in terms of

the coefficients of F(z, w).
If the same procedure is applied to the majorant equation (2.4.25), we
get exactly the same formulas for determining the C's, provided that we
replace the lower case letters in Mn by the corresponding capitals. It is
then clear that

IcI^C, Vn, (2.4.34)

so W(z) is a majorant for w(z) provided that the series (2.4.26) has a disk
of convergence. If we know that this is the case, it follows that (2.4.27)
has a radius of convergence at least as large as that of W(z). And now all
the operations performed to get the coefficients are justified, and the
formal series is an actual solution.
The coefficients cn are uniquely determined so that (2.4.27) is the only
solution which is holomorphic in some neighborhood of z = 0. There is
still the possibility of the existence of a nonholomorphic solution. If this
would be given by a series in terms of fractional powers of z, either the
series itself or some derivative thereof would become infinite as z^>0.
This cannot be a bona fide solution of an equation of type (2.4.1), for a
solution of such an equation must possess derivatives of all orders
continuous at z = 0. Nonanalytic solutions can also be excluded by one of
the uniqueness theorems of Section 2.3. ■
What remains to be done with the series is to find a suitable majorant
equation having an absolutely convergent solution series. This problem
will be taken up in the next two sections.
Before leaving the general majorant problem, let us consider another
variant of the majorant method which leads, not necessarily to a majorant
series, but rather to a dominant function which may give alternative
convergence proofs.
It was seen above that positivity plays an important role in our
problems. If the coefficients cjk are nonnegative, the solution coefficients
are also nonnegative. This means in the first place that, for z real positive,
z = jc >0; then the sequence of partial sums

Sn(x) = f,cjxi (2.4.35)

forms a nondecreasing sequence, so that it is sufficient to show that it is

bounded for some jc. The supremum of the jc's for which the partial sums
are bounded gives the radius of convergence of the series.
In the second place we have the Pringsheim theorem (see the end of
Section 1.8), which states that for a power series with positive coefficients
the point z = R is a singularity of the function defined by the series. If the
THE CAUCHY MAJORANT
57
coefficients are increased, the radius of convergence cannot increase and
the singularity either stays put or moves toward the origin. Both observa-
tions show the importance of having a convenient dominant for F(z, w) at
one's disposal. This question will be examined in greater detail in Section
2.7.

EXERCISE 2.4
1. Show that the function (2.4.6) equals

1 - z"1 + 2!z"2 + (- \)nn \z'n - (- 1)" (n + l)!z ez j r""2 e" dt.

Use integration by parts. Estimate the remainder.
2. Verify Lemma 2.4.1.
3. Take the system w' = z2 + w2, h>(0) = 0, and verify that the formal solution at
the origin is z3 times a power series in z4. What does this imply for the
singularities of the solution?
4. In problem 3 vv'(jc) > [w(jc)]2 if z = x > 0. Deduce from this that w(x) cannot
stay bounded for all x > 0 and that there is a finite oj such that w(x) f +°° as
x t <*>. Furthermore, (a> - x)w(x) < 1 for 0 < x < cj.
5. We have w'(jc) < cj2 + [vv(jc)]2 for 0<x<o). Deduce from this that
arc tan [w(x)la)] < cjx and cj > C^)1'2. Thus the radius of convergence of the
series giving the solution of Problem 3 is at least (W )m.

2.5. THE CAUCHY MAJORANT

We consider the system

w'(z) = F[z,w(z)l w(0) = 0, (2.5.1)

with

Ffew) = nVV, (2.5.2)

Here the double series is absolutely convergent in the dicylinder

D:\z\*za, \w\^b. (2.5.3)

It is desired to prove the convergence of the formal solution

n2= \ cnzn (2.5.4)

by exhibiting a suitable majorant G(z, w) of F(z, w) for which the system

W'(z) = G[z, W(z)l W(0) = 0. (2.5.5)

58 EXISTENCE AND UNIQUENESS THEOREMS

has an absolutely convergent series solution,

W(z)=2n = l Cnz\ (2.5.6)

To this end we note that the series

M(a,b) = 2 2\cik\aibk (2.5.7)

is convergent. In particular its terms are bounded; M(a,b) = M is

trivially a bound so that

|c,k|^M<r'V\ Vj,fc. (2.5.8)

Thus we can take

G(z, w) = M±■Tok=o\a/\b/
2 YgY = (a
. - z)(b
M»» - w)
y (2.5.9)'
Now the system

W'(z) = Mab(a - z) l[b - W(z)Y\ W(0) = 0 (2.5.10)

can be solved explicitly. We have
W'-l WW
b a - z
and

W(z)-jg[W(z)f = -Ma log(l-0

whence

W(z) = b- [b2 + 2abM log (l-^jj' - (2.5.11)

Here the logarithm has its principal value (log u real if u > 0) and
(b2)1'2 = +b. The majorant series is the Maclaurin series for z h-» W(z). It
has positive coefficients, and the series converges for \z \ < R, where R is
the singularity on the positive real axis nearest to the origin. The
singularities of W(z) are (i) singularities of the logarithm and (ii)
singularities of the square root. Here z = a is the only finite singularity of
the logarithm. The square root becomes singular at the point where the
expression inside the square root is zero; this occurs for

+-exp(-dyH (2-5-,2)
since it is < a. Thus we have proved
z =
THE CAUCHY MAJORANT 59

THEOREM 2.5.1

The system (2.5.1) has a series solution (2.5.4), which is absolutely

convergent in the disk |z| < R with R defined by (2.5.12). The coefficients
are uniquely defined by (2.4.32).
The quantity R is defined by (2.5.12) in terms of a, b, and M. These
quantities are not uniquely determined by F(z, w), as observed in Section
1.11 in the discussion of the transformation T(z, w). Here we have two
alternatives. We can regard the result as a pure existence proof: there
exists a disk D0: \z\<r in which the formal series solution is absolutely
convergent and defines an actual solution. Or we can try to find the
optimal choice of a and b (which determine M), i.e., a choice that gives
the largest possible value for R. This is usually a difficult problem, and one
may have to be satisfied with something less than the best result.
As an illustration, let us take

F(z, w) = (1 - z - wyl = 22 ^rr <2-5-13)

Here we take a = p, b = 1 - p, 0<p < 1. The series diverges for this

choice, but the terms are bounded. In fact, we may take M = 1 since

i = [p+(W)r=|o7]^P<(i-P)"-'.
This choice gives

r(p) = p[l-expg-£)],

and now the problem is to maximize r(p) by a proper choice of p. Now

r(p) has a unique maximum, and
0.211 < max r(p) < 0.213. (2.5.14)
How close is this value to the actual radius of convergence of the series?
Since the coefficients of F(z, w) are positive, those of the series solution
will have the same property and z = R is the singularity nearest to the
origin. This means that the essential features of the solution are discerni-
ble in the real domain. Now the solution curve

y=f(x) (2.5.15)
goes from the origin to a point P0 on the line x + y = 1 . Its slope is 1 at the
origin and +oo at P0(xo, 1 - x0). This means that the integral curve is
confined to the triangle

A:x<y<l-x, and R=xQ<i (2.5.16)

60 EXISTENCE AND UNIQUENESS THEOREMS

As we shall see later, the singularity at z = jc0 is a branch point of order 1:

y = 1 - x0 - [2(jc0 - x )]m + • • • . (2.5.17)
As a matter of fact, we can determine R = x0 exactly, for x as a function
of y satisfies the linear DE
dx t

and the solution with the initial value x = 0 for y = 0 equals

x=2-y-2e-\ (2.5.18)
For y = y0 the slope of the curve (2.5.18), i.e., jc'(y0), is 0 since y '(*<>) =
+ oo, and this shows that y0 = log 2, x0 = 1 - log 2 = 0.30685. Since this is
the value of R, the estimate (2.5.14) is rather far off the mark.

EXERCISE 2.5
1. Verify (2.5.18).
2. Find a lower bound for R in the system w' = z2 + w2, w(0) = 0, and compare
your result with that obtained in Problem 2.4:5.
3. It is desired to extend Theorem 2.5.1 to systems (compare Theorem 2.2.2).
Take z0 = w,0 = w20 = • • • = w„0 = 0 and

Fm(z, w) = 2 c(£ fenzV{< • • • w\\ X^m^n.

Suppose that the terms of these series are uniformly bounded by M for
z = a, Wi = • • • = wn = b, where a and b are positive numbers. Show that each
Fm admits of

as a majorant. Use this fact to find a common majorant for the solution series
and corresponding lower bounds for the radii of convergence.

2.6. THE LINDELOF MAJORANT

Cauchy's calcul des limites goes back to 1839, and for almost 60 years it
remained the only method of majorants. In 1896 Ernst Lindelof (1870—
1940) observed that the best majorant of F(z, w) is obtained by replacing
the coefficients cjk by their absolute values, thus taking

Cjk=\cik\, Vj,k. (2.6.1)

He also found that the whole discussion can then be carried out in the real
domain and the question of convergence can be replaced by questions of
THE LINDELOF MAJORANT 61

boundedness, which are simpler to handle. This shows again the impor-
tance of positivity.
Lindelof's attack on this problem was simple, direct, and based on first
principles, an approach characteristic of his life work. He became the
founder of the school of analysis in Finland and had a number of eminent
pupils. Among them were L. V. Ahlfors and F. and R. Nevanlinna, whose
work will be discussed in Chapter 4.
As usual we set

G(x,y) = 2 2 C&Y (2.6.2)

i =0 k =0
for the majorant of F(x, y). Here the variables are real positive. Consider
the nth partial sum,

Pn(x)= m2=0 Cmx"\ (2.6.3)

of the formal series solution of the problem

y'(x) = G[x,y(x)l y(0) = 0. (2.6.4)

Suppose that the series (2.6.2) is convergent for jc = a, y = b, and set

2 fc=o
i=o 2 Cikaibk=M(a,b)«». (2.6.5)
Next we note that

PU,(x)^2 2 Cjkx'[P„(jc)]\ Vn. (2.6.6)

j=0 k=0
This follows from the structure of the recursion formulas for the
coefficients which define Cp. From the discussion in Section 2.4 we recall
that Cp depends on C, C2, . . . , Cp_, and the Cjk with 0 ^ j + k =s= p. This
means that all the terms on the left will cancel against terms on the right.
The right member would normally involve additional positive terms not
canceled by terms on the left.
Suppose now that
r = min (2-6-7)

Suppose that it is known that for some n

Pn(x)^b forO^JC^r. (2.6.8)
We have then

j[=0 k=0

<Xic»flV *sM(a,b).
j =0 fc=0
62 EXISTENCE AND UNIQUENESS THEOREMS

For 0 x r this gives

Pn+1(jc) ^ xM(a, b) ^ rM(a, b)^b. (2.6.9)
Now
Pl(x) = C00x*£rM(a,b)^b,
so that the estimate for Pn holds for all n.
Now the sequence Pn(x) is nondecreasing for fixed jc, and since it is
bounded for O^x ^ r the partial sums Pn(x) of the formal solution series
converge to a finite limit, i.e., the formal solution is an actual solution and
its radius of convergence is at least equal to r defined by (2.6.7).
Again we have an optimization problem on our hands. How should a
and b be chosen to get the best possible result? Unless other information
is available, it may be advantageous to choose a and b so that
aM{ayb) = b (2.6.10)
and then to maximize with this side condition.
We apply this technique to the system defined in (2.5.13):

w' = (\-z- w)~lw, w(0) = 0. (2.6.11)

Here M(ay b) = (1 - a - b)~\ and condition (2.6.10) gives

--y^f. (2.6.12)
the maximum of which is attained for b = 2m - 1. Then <2max = 3 - 2 • 21/2 =
0.1716.... This is max r(a,b) as obtained by this device, and it is
considerably worse than (2.5.14).
As a second example consider the system

w' = 2z3+w\ w(0) = 0. (2.6.13)

Here
b b

M(a,b) 2a3 + b2'

For a fixed a >0 the maximum of this ratio is \a~2. Hence the optimal
choice is a = la "2 or a = 3"1/3 = 0.693 This estimate is not too bad. The
method of the next section (see Problem 2.6:5) gives as a lower bound for R
21/93-1/67T,/3= 1.316.

In 1899 Lindelof extended his method to give a proof of the implicit

function theorem, which we state without proof. Given the equation

w = F(z, w) = bl0z + S S' bikzjwk, (2.6.14)

where the prime indicates that j + k > 1, it is desired to solve (2.6.14)
 63
THE USE OF DOMINANTS AND MINORANTS

for w. We set

i. Bik=\bjk\. (2.6.15)

THEOREM 2.6.1

Let a and b be positive numbers such that the series (2.6.15) converges for
z = a, w = b. Then the formal series solution of w = F(z, w), w(0) = 0,
converges absolutely for |z| < Rab, where

(2.6.16)

Lindelof showed how to find the exact radius of convergence when

G(z, w) is an entire function of z and w, either a polynomial of degree ^ 2
or transcendental in w. We shall not consider this question; Theorem 2.6.1
is good enough for our purposes. One further remark about Lindelof: his
monograph, he Calcul des Residues et ses Applications a la Theorie des
Fonctions (Gauthier-Villars, Paris, 1909; Chelsea reprint, New York,
1947), is still the best book in this field.

EXERCISE 2.6

1. Determine an admissible r for w' = z2 + w2, w(0) = 0.

2. For the system of simultaneous equations in Problem 2.5:3 find Lindelof
majorants and corresponding lower bounds for the radii of convergence of the
solution series.
Find admissible values of r for each of the following systems:
3. w\ = - w2, h>2 = wu w,(0) = 1, w2(0) = 0.
4. w\ = w2W}, h>2 = - WjVf,, w>3 = -k2WiW2, w,(0) = 0, w2(0) = vt>3(0) = 1,0< k < 1.
5. For the solutions of Problems 3 and 4 show that

[w,(z)]2 + [w2(z)]2^l, *2[w,(z)]2 + [w3(z)]2ssl.

[Remark : These three functions are the elliptic functions of Jacobi or, in
Gudermann's notation, sn (z; k), cn (z; k), and dn (z; k).]
6. Solve the system in Problem 3.
7. What is the simplest differential equation satisfied by eZf? Use a uniqueness
theorem to show that ez 0 for all z.

2.7. THE USE OF DOMINANTS AND MINORANTS

A power series has a finite radius of convergence, R say, iff the function
defined by the series

0 (2.7.1)
Z^f(z)=2fnZ"
64 EXISTENCE AND UNIQUENESS THEOREMS

has no singularity in |z| < R but has a singular point on the rim of the disk
|2|=si?. Naturally this applies also to the solution of the DE
w' = F(z, w) with w(0) = 0, (2.7.2)
but neither the method of Cauchy nor that of Lindelof gives much
information about the distance to the nearest singularity, much less about
its nature.
These defects may be partly remedied for the class of DE's
w' = P(z,w\ w(0) = 0, (2.7.3)
where P(z, w) is a polynomial in z and w with positive coefficients. In this
case the solution

w(z) = 2Cnz"
i (2.7 A)

also has nonnegative coefficients, and the point z = R is a singularity of

w(z). Suppose that

P(z,w) = "Z2Qkziwk, (2.7.5)

where only a finite number of the coefficients are different from zero.
More precisely, we assume the existence of integers m, n, p such that

Cik ^ 0, V/, it; Cfk = 0, />m>l, fc>n>l; (2.7.6)

Cj0 = 0, 0 ^ j < p ^ m, CPo>0; Cmn>0. (2.7.7)
The first result in this setting is now

THEOREM 2.7.1

Under the stated assumptions on P(z, w) the radius of convergence R of

the solution (2.1 A) is finite. The solution tends to +oc as z t R along the
real axis. Moreover, there exists numbers A and B depending only on m,
n, p, Cpo, and Cmn such that
0<A<R<B<oc. (2.7.8)

Proof. We take z = x, w = y as real positive; then with w(z) = y(x) we

have

so that y'(x)>Cp0xp, 0<x<R, (2.7.9)

y(x)>yf^xp+x. (2.7.10)
Next we note that y(jc) increases with x and tends to a limit as x tends
to R. If R should be infinite, then obviously
 65
THE USE OF DOMINANTS AND MINORANTS

limy(jc) = +oo. (2.7.11)

But this is true also if R is finite, for if limx Tr y(x) is finite, equal to y0 say,
then P(z,w) is holomorphic at (R, y0) and the initial- value problem
w'(z) = P[z,w(z)], w(R) = y0
has a unique solution holomorphic in some disk \z - R\< 8. This solution
must coincide with the solution (2.7.2) on the interval (R -8,R) and
hence furnish the analytic continuation of w{z) in the disk \z -R\<8.
This contradicts the fact that z = R is a singularity of w(z).
Thus (2.7.1 1) holds whether or not R is finite. We proceed to prove that
R is finite. Since
y'(x)>Cmnxm[y(x)T, (2.7.12)
we have, for 0 < jc, < x2 < R,

[y(xl)Y~n - ly(x2))]n >^TlCmn(x?+l -x?+1). (2.7.13)

Here the left-hand side is less than the first term, while the right member
becomes infinite with x2. This shows that jR must be finite.
We can now let x2 increase to R. Then for all xx in (0, R)

Rm+]^xr^f^j(cmnrl[y(xdVn-
For the last term we get an upper bound from (2.7.10). The result is of the
form
Rm+l ^x?+l + C(jc,)-<n-1)(p+1), (2.7.14)
where C is independent of jci. This holds for all JCi, 0<*i<jR, and, in
particular, for the value xx = A, for which the right member becomes a
minimum. Here 0 < A < R. Now define B by
Bm+X = Am+1 + C4-("-,)(p+,). (2.7.15)
We have then clearly 0 < A < R < B, and the theorem is proved. ■
We have shown that y(jt)-> + o° as x increases to R. We can actually get
an estimate of the rate of growth. To this end we consider (2.7.13) once
more. We set Xi = jc, and let x2^R. This gives

[y(x)Y-n>-^^Cmn(Rm+1-xm+l)
= (n-\)CmnxZ(R-x).
Here the mean value theorem has been used, and x <xQ< R. Hence

y(jc) < [(n - \)CmnxZrmn-"(R - x)~^ (2.7.16)

66 EXISTENCE AND UNIQUENESS THEOREMS

and

limsupK* -x)lKn-l)y(x)]^[(n - \)CmnRmVKn-'\ (2.7.17)

On the other hand, for a small 8 > 0 we can find an e > 0 such that, for
R - 8 < x < R,

y'(x) < (1 + €)CmnRm[y(x)]n. (2.7.18)

This gives

[y(x)]l n <(\ + e)(n - \)CmnRm(R-x)

and

v(jc)>[(1 + e)(n - \)CmnRmVKn-X)(R - xTlKn-l\ (2.7.19)

whence
liminf (R -xyKn X)y(x)^[(n - \)CmnRmVKn^- (2.7.20)

Combining (2.7.17) and (2.7.20), we obtain

x-+R [(R - xYKn i)y(x)] = [(n - \)CmnRmVKn-x\

lim (2.7.21)
Thus we have proved ■

THEOREM 2.7.2

Under the stated assumptions on P(z, w) the solution w(z) of (2.7.3) has a
singularity at z = R, where

z— R [(R - z)lKn'l)w(z)] = [(n - l)CBlilRmr,/(,l"1), (2.7.22)

lim
at least for radial approach of z to R.

The right-hand members of (2.7.9) and (2.7.12) may be referred to as

minorants of P(jc, y), while the right member of (2.7.18) ranks as a
dominant. We can obviously find other useful dominants and minorants
for this class of equations. Thus P(R,y) is evidently a dominant, and
from
y'(x)*zP[R,y(x)], 0<x<R (2.7.23)
we get

(2J-24)

Since the left side is a function of R alone, say Q(R), we have the
inequality
Q(R)^R. (2.7.25)
VARIATION OF PARAMETERS
67
Now Q(R) is evidently a decreasing function of R, so the equation

Q(r) = r (2.7.26)
has one and only one positive root, r0 say, and this implies
r0^R. (2.7.27)
This inequality is sometimes a useful one.

EXERCISE 2.7

1. Is it essential for the discussion given above that P(0, 0) = 0 and that P(z, 0) is
not identically zero? What modifications in the argument are needed if these
assumptions are dropped?
2. Instead of w(0) = 0, would w(0) = y0>0 be manageable?
3. The restriction of P(z, u>) to be a polynomial in z and w is evidently
undesirable. Can some modification of the method be applicable to the DE

w'(z) = [1 + w2(z)],/2[l + k2w\z)f\ 0 < k < 1,

with u>(0) = 0? This is an elliptic DE satisfied by -i sn(/z;/c).
4. Consider the equation w' = 1 + w2, w(0) = 0. What information is obtainable
from Theorems 2.7.1 and 2.7.2 in this case? Does (2.7.27) give reasonable
results? Note that R = \tt.
5. Apply the methods of this section to the system
w' = 2z3+ w\ H>(0) = 0.
In particular, use y '(x) < 21? 3 + y 3 for 0 < x < R. The gamma function satisfies
T(p)r(l - p) = 7T cosec ptt, which may prove useful.

2.8. VARIATION OF PARAMETERS

As usual we are concerned with an equation

w' = F(z, w), (2.8.1)

where (z, w)i-»F(z, w) is holomorphic in some dicylinder D of C2. We

plan to examine the dependence of the solution on parameters.
Parameters are of two kinds: internal and external. A parameter is
internal if its value characterizes a particular solution or possibly a class
of solutions but does not occur explicitly in the equation.
As an orientation consider the trivial Riccati equation

w' = a+bw\ (2.8.2)

Here a and b are external parameters. Suppose that the solution
68 EXISTENCE AND UNIQUENESS THEOREMS

w(z; Zo, Wo) is under consideration. Here D = C . Now

w(2;0,0)^(f)"2tan[(af,)'«2]. (2.8.3)
For a = 0 the solution is identically zero. For b = 0 it reduces to az. In
both cases (2.8.3) reduces to the indicated solution if we pass to the limit
with the parameter in question, keeping z and the other parameter fixed.
The formula shows that w(z; 0,0) is an analytic function of each of the
parameters once the determination of the square root has been fixed so
that w(z; 0,0) is single valued. Suppose that both square roots are real
positive when a and b are real positive. Now keep z and b fixed, neither
being zero, and consider the solution as a function of a. It is clearly a
meromorphic function of a in the finite complex a-plane. Similar results
hold if we fix a and z and vary b.
On the other hand,

w(z ; zo, wo) = (I)''2 tan \{ab)m{z - z0) + arc tan [ffl2 w0]|. (2.8.4)
This is clearly an analytic function of z0 as well as of w0 when the other
varieties are kept fixed.
There are many other internal parameters. Constants of integration
play a remarkable role in the advanced theory of nonlinear DE's, as will
be seen later. In the present case the general solution of (2.8.2) may be
written as

P (fnPZ- Cp cos pz) (abyn

y v ' (2 8 5)
b (cos pz + C sin pz)
where C is an arbitrary constant of integration. The solution is evidently a
fractional linear (= Mobius) transform of C, and this property is charac-
teristic of Riccati's equation (see Section 4.1). Again we see that the
solution is an analytic function of the internal parameter C for fixed a, b,
and z. Many other types of internal parameters could be shown and will
be presented in later contexts.
Generally speaking, if the case (2.6.2) is at all typical we would expect
the solution of (2.8.1) to be an analytic function of whatever parameters,
external or internal, may enter into the problem of determining a solution.
This is indeed the case. We obtain numerous results along such lines from
Gronwall's lemma; see Theorems 1.6.5 and 1.6.6.
We start by getting information about the dependence of w(z; z0, w0)
on the initial value w0. Here the basic fact is that, if F(z, w) is
holomorphic in D, it has a partial derivative Fw{z, w) with respect to w
which is also holomorphic in D. Shrinking D if necessary, we may assume
VARIATION OF PARAMETERS
69
that the partial derivative is bounded:

\Fw(z,w)\^B for (z,w)GD. (2.8.6)

This implies a Lipschitz condition,

|F(z, w.) - F(z, w2)\ <B\wi- w2\. (2.8.7)

Furthermore, we have

F(z, w,) - F(z, w2) = Fw(z, - w2) + o(| w, - w2|). (2.8.8)

Let us now consider the so-called variational system for the solution
w(z; z0, Wo):

W(z) = Fw [z, w(z ; z0, wo)] W(z), W(z0) = 1, (2.8.9)

with the unique solution

W(z;z0, l) = exp||Z Fw[s, w(s; z0, w0)] dsj. (2.8.10)

Since D is convex, we may take the integrals here and in the following
along a straight line.
Consider two solutions, w(z ; z0, w0) and w(z ; z0, w,), of (2.8.1). Here of
course (z0, w0) and (z0, w0 are in D. Let D0 be a disk in the z-plane with
center at z = z0 in which the two solutions are holomorphic and bounded.
We note first that the two solutions are proximate if \wx- w2| is small.
More precisely, we have

THEOREM 2.8.1

There exists a disk D* in the w-plane such that, if w0 and Wi are both in
D* and z lies in D0, then

|w(z;z0, w,)-w(z;z0, w0)|^ |w,-w0| exp [B|z-z0|]. (2.8.11)

Proof. To simplify the formulas we write Wi(z) and w0(z) for the two
solutions. Then

w,(z)- h>0(z) = h>, - w0+ I {F[s, w,(s)] - F[s, w0(s)]} ds.

By the Lipschitz condition (2.8.7) this gives

|w,(z)- w0(z)|^|w1- w0| + b||z \w1(s)-w0(s)\ dr ,

where r is the arc length on the rectilinear path of integration. We can
70
EXISTENCE AND UNIQUENESS THEOREMS

easily reduce this inequality to the form (1.6.15) by setting

/(f) = |w,(z0 + t eie) - Wo(z0 + t eie)l 6 = arg (z - z0),

g(0 = ki-w0|, and K(t) = B.
We then get (2.8.11) by substituting in (1.6.18). ■
Much more may be said, however.

THEOREM 2.8.2

The solution w(z;z0, w0) is a differentiable function of the complex

variable w0 and hence locally holomorphic. We have

w(z; z<>, w,) - w(z; z0, w0) = W(z; z0, l)(w, - w0) + o(|w, - w0|) (2.8.12)
and
w(z; z0, w0) = exp 1
"Wo f Fw[s, w(s; z0, w0)] dsf.
Uzo J (2.8.13)

Proof We revert to the abridged notation used above. The problem is to

show that
A(z) ^ w,(z) - w0(z) - W(z)(wl - w0) = o(\w, - wJi). (2.8.14)
Now the second member of this equation reduces to

f {F[s, w,(s)]-F[s, WotsH-fw,- w0)Fw[s, w0(5)] ds.

By (2.8.8) this becomes

| Fw[5, Wo(5)][w,(5)- Wo(5)-(w,- Wo)W(s)] ds +o(|w,- W0|).

J 20
Thus we have
(2.8.15)
\A(z)\^o(\w,-w0\) + B (Z \A(s)\ds
This can also be reduced to the standard form (1.6.15) in an obvious
manner and yields the inequality

|A(z)| as o(| w, - w0\) exp [B\z - z0|], (2.8.16)

which implies (2.8.12) and (2.8.13). Note that the difference quotient tends
to W(z ; z0, 1), no matter how u>, tends to w0. The analyticity is implied by
this fact. ■
This theorem was discovered in 1896 by Ivar Bendixson, using a
function-theoretical argument. The use of Gron wall's inequality is of later
date. Theorem 2.8.2 is in a certain sense a forerunner of the in-
equality. The theorem extends to systems of n first order DE's in n
unknowns. Here Jacobians come into play and the formulas become fairly
VARIATION OF PARAMETERS 71

complicated. See Halanay (1966); the first extension was given by G.

Peano in 1897.
We turn now to the dependence of u>(z;z0, w0) on the internal
parameter z0. We have analogues of Theorems 2.8.1 and 2.8.2 for this
case. Since the proofs follow the same lines, the argument will merely be
sketched. We assume that

\F(z, w)|*£M, (z, w)GD. (2.8.17)

Consider two solutions of (2.8.1):

w(z;z,, w0)= w,(z), w(z;z2, w0)= w2(z), (2.8.18)

where z, z,, and z2 are restricted to a disk D0 with center at z = z0, in
which u>,(z) and w2(z) are holomorphic and bounded.

THEOREM 2.8.3
With the stated assumptions

|w(z; z,, w0) - w(z; z2, w0)| ^ M|z, - z2| exp [B 5(z)], (2.8.19)
where 8(z) = min [|z-Zi|, |z-z2|].
Proof {sketch). Suppose that z is nearer to z, than to z2. Then

Wi(z) = u>o+ J F[s, wx(s)] ds,

w2(z) = u>0 + j F[s, w2(s)] ds - j F[5, w2(s)] ds,

where we have used the fact that the integral of F[s, w2(s)] along the
perimeter of a triangle with vertices at z, zu and z2 is zero. Now the last
integral in the displayed formulas is less than M\z{ - z2\ in absolute value.
Using the Lipschitz condition (2.8.7), we get

|w2(z)- w1(z)|^M|z2-z1| + B
| \w2is)-wx(s)\\ds
By (1.6.18) this gives

|w2(z)- w,(z)|*s|z2-z,|exp [B\z -z,|].

Since 8(z) = \z - zx\ in this case (2.8.19) is verified. I

THEOREM 2.8.4

With the stated assumptions, let W(z) = W(z; z0, 1) be the solution of
(2.8.9). Then w(z;z0, w0) is a differentiate function of the complex
72 EXISTENCE AND UNIQUENESS THEOREMS

parameter z0 and
3z0 (2.8.20)
w(z; z0, Wo) = - F(z0, Wo)W(z).

Proo/ {sketch). Consider two solutions of (2.8.1), namely, w0(z) =

w(z ; z0, Wo) and w,(z) = w(z ; z,, w0), where z, z0, z, are in the disk D0. Set
D(z) = w,(z) - w0(z) + F(z0, w0) W(z)(Z] - z0). (2.8.21)
It is required to prove that D(z) = o(|z, - z0|) for any z, in D0. We have

W,(Z)-W0(Z)= I F[s, W,(S)] - J F[s, W0(s)]ds

= f2 {F[s, - F[s, w0(5)]} <fo - P F[5, ds.

Here the first integral is by (2.8.8)

J Fw[s, w0(s)][wi(s) - w0(s)] ds + F,.

The error term £, is small by (2.8.8) and (2.8.19), for F, goes to zero with
\zi - z2\ faster than M|z,-z0
||£
i/, W,(s)- U>0(s)||ds exp [8(s)] \ds

so that Ex = o{\zx - z0|).

Next, note that

W(z)=\+\Z Fw[s, Wo(s)]W(s) ds.

Hence

- ( F[s, w0(s)] ds + F(z0, H>0)(z, - z0)

= - I {F[s, w0(s)] - F(z0, wo)} ds,

the absolute value of which does not exceed

b|JZ1 \w0(s)- w0\\ds\\ *s b|£Z' |<fc||£ |F[f, w0(f)]||^|| ^IB^z.-ZoI2.

Combining the various loose ends, we get

J*o
D(z) = (Z Fw[s9 w0(s)]D(s) ds + E2 with F2 = o(|z, - z0|), (2.8.22)
VARIATION OF PARAMETERS 73

and this implies the differentiability of w{z ; z0, w0) with respect to z0 and
gives the value of the derivative as in (2.8.20). ■
The remaining remarks in this section are devoted to external parame-
ters. We are now concerned with an equation

w ' = F(z, w, A), w (z0, A0) = wo, (2.8.23)

where (z, w, A) F(z, w, A) is hoiomorphic in some tricylinder D in C3.
We may assume that D is centered at the origin and that

F(z, w, A ) = 20 20 20 cykz'VA fc (2.8.24)

convergent in D. It is then natural to take z0 = w0 = A0 = 0, to assume a

power series solution

0 0
w(z,A) = n«mAn, (2.8.25)

and to prove convergence using Cauchy's majorant method. We can find

positive numbers a, b, c, M such that

\c*\a'b'ck*zM9 Vi,j,k.
Hence

G(z, w9 \) = m(\ - 1)"1 (l- £f (l- (2.8.26)

is a majorant of F(z, w, A), and the solution of

W = G(z, W, A), W(0, 0) = 0 (2.8.27)

will be a majorant of (2.8.25). This majorant is easily found to be

W(z, A) = b - [b' + ^f log (l-f )f. (2.8.28)

This is a hoiomorphic function of (z, A ) in a dicylinder in C2, and, in
particular, for a fixed small value of |z | it is a hoiomorphic function of A in
a disk. Thus we have proved ■

THEOREM 2.8.5

The system (2.8.25) with z0 = w0 = A0 = 0 has a unique solution (2.8.25)

which is a hoiomorphic function of (z, A) in any dicylinder in C2 where
W(z, A) is hoiomorphic.
This is a powerful theorem and may, in fact, be considered too
powerful for applications where the external parameters are normally
linear. In such a case the solution would usually be an entire function of A
74 EXISTENCE AND UNIQUENESS THEOREMS

of low order or else a meromorphic function, also of low order. In

conclusion let us remark that expansion in powers of a parameter is
commonly used, for instance, in celestial mechanics, where Poincare put
this useful device on a rigorous basis in his lectures on this subject in the
1890's.
EXERCISE 2.8
1. Complete the proofs of Theorems 2.8.3 and 2.8.4.
2. In (2.8.28) for W(z, A) we have |z| < a and |A| < c. If A is fixed and |A| < c,
find the singularities of W(z, A) as a function of z. Which singularity is nearest
to the origin?
3. The equation (1 - z)w' = Aw, w(0) = 1, VA, has a solution holomorphic in the
unit disk of the z-plane for all A. For a fixed z, |z| < 1, verify that the solution is
an entire function of A. Estimate the maximum modulus, and compare
log M[r; w(A)] with powers of r = |A |.
4. Apply Poincare's method to the equation
w"-(z + A)w =0, w(0)=l, w'(0) = 0, VA.
This involves setting w(z) = w0(z) + ^ wn(z)\n with w0(0)=l, w'0(0) = 0,
wn (0) = w'n(0) = 0, n>0. Here vv'o(z) = zw0(z) and may be assumed to be
known. Furthermore, w"n{z)= wn_,(z), n > 0. For a fixed z =x >0, try to
estimate log M[r; w(A)]. It should grow as rm.

LITERATURE

The presentation is based largely on the author's LODE. Chapter 2 of

LODE has a bearing on Sections 2.1-2.6; Chapter 3 goes with Section 2.8.
In this treatise essentially only the analytic case is considered. Section 2.7
contains some new material, though the basic idea is inherent in earlier
work by the author and others. Further references are:
Banach, S. Sur les operations dans les ensembles abstraits et leur application aux equations
integrates. Fundamenta Math., 3 (1922), 133-181.
Bendixson, I. Demonstration de l'existence de Fintegral d'une equation aux derivees
partielles lineaires. Bull. Soc. Math. France, 24 (1896) 220-225.
Birkhoff, G. D. and O. D. Kellogg. Invariant points in function space. Trans. Amer. Math.
Soc, 23 (1922), 96-115.
Brouwer, L. E. J. On continuous vector distributions on surfaces. Verh. Konikl. Akad. Wet.
Amsterdam, 11 (1909); 12 (1910).
Cacciopoli, R. Un teorema generale sull'essistenza di elementi uniti in una trasformazione
funzionale. Rend. Accad. Lincei Roma, (6), 11 (1930), 98-115.
Cauchy, A. Memoire sur l'emploi du nouveau calcul, appele calcul des limites, dans
l'integration d'un systeme d'equations differentielles. Compt. Rend. Acad. Sci. Paris,
15 (1842); Oeuvres (1), 7 (1898), 5-17.
 75
LITERATURE

Halanay, A. Differential Equations: Stability, Oscillations, Time Lags. Academic Press,

New York, 1966. (Re Theorem 2.8.2.)
Lindeldf, E. Demonstration elementaire de l'existence des integrates d'un systeme
d'equations differentielles ordinaires. Acta Soc. Sci. Fenn., 21, No. 7 (1897), 13 pp.
Demonstration elementaire de l'existence des fonctions implicites. Bull. Sci. Math., (2),
23 (1899), 68-75.
Picard, E. Memoire sur la theorie des equations aux derivees partielles et la methode des
approximations successives. J. math, pures appi, (4), 6 (1890), 145-210.
Poincare, H. Les methodes nouvelles de la mecanique celeste. 3 vols. Gauthier-Villars, Paris,
1892, 1893, 1898.
3

SINGULARITIES

In Chapter 2 our study of the equation

w'(z) = F[z, w(z)], w(zo)=w0

led to the conclusion that there is a unique holomorphic solution
whenever the mapping (z, w)-+F(z, w) is holomorphic at (z0, w0). In this
chapter we shall study what happens at a point where this assumption
does not hold. Such a point is a singularity at least for the solution which
approaches w0 when z-»z0. But there are other possibilities. Solutions
may become infinite as z - > Zo or may not tend to any limit, finite or
infinite. Usually the point at infinity is a fixed singularity. We have to
examine these various possibilities and try to bring some order into the
confusion. Pioneering work in this field was done by Lazarus Fuchs
(1833-1902), whose name will be encountered in various later chapters.
Here we shall be concerned mostly with the work of Paul Painleve
(1863-1933).

3.1. FIXED AND MOVABLE SINGULARITIES

Singularities of DE's are of two types: fixed and movable. Fixed

singularities are external in the sense that they are given position and
nature by the DE, and at least the position should be obtainable in a finite
number of steps without knowledge of the solutions. Movable sing-
ularities, onthe other hand, depend on internal parameters, and normally
such a point can be put anywhere in the complex plane by manipulating an
internal parameter. Shift the parameter and the singularity moves. Thus
the equation
W> = W2 - 7T2 COt2 7TZ (3.1.1)
has fixed singularities at the integers and at infinity. The finite singularities
are simple poles for some solutions, but the general solution has transcen-
dental branch points where infinitely many branches are permuted. The
76
FIXED AND MOVABLE SINGULARITIES
77
movable singularities are simple poles and may be placed anywhere in the
complex plane.
The nature of a movable singularity will depend on the nature of the
equation. Thus the equations

w' + ew=Q and v' + v log2 v = 0 (3.1.2)

are satisfied by

log (jzr^j and v = exp (737), (3-1.3)

respectively. In the first case there is a movable logarithmic branch point;
in the second, a movable essential singularity.
In the following discussion such cases will be excluded. In the first
order case (but not for systems) this can be achieved by assuming that
F(z, w) is a rational function of w with coefficients which are algebraic
functions of z. This restriction makes the number of fixed singularities
finite and permits only algebraic movable singularities. During most of our
work we shall make the further restriction that the coefficients are single
valued; this implies that they may be assumed to be polynomials. We
consider equations of the form

w' = %r^
Q(z,w) (3-1-4)
with

P(z,w) = J=0
^A,{z)w>, Q(z,w)=^Bk(z)w\
Jc=0 (3.1.5)

where the A's and B's are algebraic functions of z.

It is appropriate to provide the reader with a brief abstract of the theory
of algebraic functions. We shall need some of the properties of such
functions for the discussion of the fixed singularities of (3.1.4). For each
algebraic function z A(z) there is a unique irreducible polynomial in
two variables:

V(z, u) = p0(z)un +pl{z)un l + - • • + pn_1(z)w + p„(z), (3.1.6)

such that V[z, A(z)] = 0. Here the p's are polynomials in z without a
common factor. Conversely, every such polynomial V(z, u) defines an
algebraic function. The zeros of A(z) are the zeros of p„(z); the
infinitudes, the zeros of p0(z). The point at infinity may be a singular point
of A(z). Additional singularities may be furnished by the zeros of the
discriminant equation

Do(z) = 0. (3.1.7)
78 SINGULARITIES

If z = z0 is a root of this equation, then the equation

V(zo,m) = 0 (3.1.8)

has multiple roots and z = z0 is an algebraic branch point of A(z). Here

D0(z) is the resultant of eliminating u between the two equations
V(z, u) = 0 and Vu(z, u) = 0. Since the number of multiple roots is at most
equal to the total number of roots, which is n, we see that the number of
singular points of A(z) is finite and is at most n + 1 + degp0.
It should be noted that the algebraic functions of one complex variable
form a field over C: if Ax and A2 are algebraic functions, then so are their
sum, product, and quotients.
We can now start an enumeration of the possible fixed singular points
of the DE (3.1.4). The first subset of such points is formed by the
singularities of the A's and the B's, say,
Si:si,s2, ...,5M,oo. (3.1.9)

This is a finite set. The second subset S2 is the set of points th if any, such
that Q(th w) is identically zero, say,

&:f„*2,...,k (3.1.10)

At z = tj all the algebraic functions B0,BU . . . , Bq are zero, and this

means that the corresponding polynomials Vy(z, 0) are zero, where
Vj(z, u) is the defining polynomial for Bj(z). Now the common roots of a
number of polynomials can be found, for instance by the method of
Bezout. We can affirm, however, that the number of common roots cannot
exceed min, (m;), where m, is the degree of V)(z, 0).
The third subset,

53: eu e2, . . . , eA, (3.1.11)

is formed by the roots of another discriminant D3(z), the vanishing of

which expresses the existence of numbers vv0 such that

P(ehwo) = 0, Q(ehwo) = 0, (3.1.12)

so that PIQ takes on the indeterminate form 0/0 at (z0, w0). Now the
condition for P(z, w) = 0, Q(z, w) = 0 to have a common root is the
vanishing of a determinant with p + q rows and columns, the entries
being the p + q + 2 coefficients A} and Bk with the rows filled out with
zeros. Now D3(z) is an algebraic function of z obtained by expanding the
determinant, and as an algebraic function it satisfies a polynomial.
FIXED AND MOVABLE SINGULARITIES 79

equation V(z, u) = 0. Thus the e's are the roots of the polynomial V(z, 0).
Hence the set S3 is also finite.
In studying the infinitudes of the solutions of (3.1.4), we use the
transformation w = \/v, which transforms (3.1.4) into

, = .i«) = ^l, (3.U3)

where Px and Qx are polynomials in v with the A,'s and IV s as

coefficients. The fixed singularities of this equation are also fixed sing-
ularities of(3.1.4). The sets Si and S2 are unchanged, but the result of
eliminating v between Px = 0 and Qx = 0 is not necessarily the same as
that of eliminating w between P = 0 and Q = 0, since now the common
roots of Pi(z, 0) = 0 and Qi(z, 0) = 0 are relevant. Thus we have possibly
an additional set of fixed singularities,
54: ii,i2,...,ic. (3.1.14)
The set (3.1.14) is a finite one, and the union of the four sets Sx-S4 is a
finite set 5. This is the set of potential fixed singularities of (3.1.4).
A point of the set 5 need not be a singularity of all or any of the
solutions, as is illustrated by the following DE's:
u' = 1 + zmu -zmu\ w' = ijlwIz. (3.1.15)
In both cases z = 0 as a singularity of the coefficients is a potential
singularity. The first equation has the particular solution u = z, which is
holomorphic at z = 0. The general solution has a branch point of order 1
there (order 1 means two branches). The second equation has the general
solution CzM, which is holomorphic at z = 0 iff /jl is a positive integer.
There is a pole if /jl is a negative integer, an algebraic branch point if /x is a
rational number, and a transcendental critical point in all other cases.
Here the nature of the singularity depends essentially on the algebraic
nature of the external parameter ix.
This enumeration of potential fixed singularities is due to Paul Painleve
in the 1890's, and we shall see in the next section how important it is for
the problem of analytic continuation of the solutions of the equation.
Before considering this question, let us mention that Painleve could look
back on a long line of French mathematicians who had contributed to one
phase or another of the theory of nonlinear DE's. Besides E. Picard and
H. Poincare special mention should be made of Charles Briot (1817-1882)
and Jean Claude Bouquet (1819-1885). In the 1850's they were writing a
treatise on elliptic functions (the first of its kind). An elliptic function is a
doubly periodic meromorphic function. They showed that, if / and g are
both elliptic and have the same periods, there exists a polynomial
so SINGULARITIES

P(Z,,Z2) such that

P[/(z),g(z)] = 0.
In particular, if /(z) is elliptic so is /'(z), and the two have the same
periods so that there exists a polynomial DE,

P(W, w') - 2 cjk(w)J(w')k = 0, (3.1.16)

which is satisfied by /(z).and, more generally, by f(z + a), where a is an
arbitrary constant since (3.1.16) is autonomous, i.e., does not contain z
explicitly. Similar equations of the second or higher orders also hold for
elliptic functions.
Not all equations of this type are satisfied by elliptic functions, and
Briot and Bouquet set themselves the task of finding all equations of type
(3.1.16) which have elliptic solutions. We shall take up this problem in
Chapter 11 for first order equations and in Chapter 12 for second order
equations. Briot and Bouquet also studied first order DE's with a fixed
singular point at the origin; see Section 11.1.
At this juncture we merely add that the search for elliptic equations is a
special case of finding equations of type (3.1.16) whose movable sing-
ularities are poles. To this problem we can give at least a partial solution,
which permits eliminating a number of equations from the list of suspects;
see Section 3.3.

EXERCISE 3.1

1. Suppose that z h-> f(z) is a solution of (3.1.1) with a simple pole at z = z0.
Show that the equation also has a solution with a simple pole at z = z0 + n.
Here z0 is an arbitrary complex number, and n is an arbitrary integer. [If z0 is
an integer, there exist two distinct solutions which have a simple pole at
z = z0. The general solution has a transcendental critical point at each integer
involving the powers (z - z0)\ where r is a root of the equation r2 - r = 1.]
2. With the aid of the theory of symmetric functions, prove that the sum of two
algebraic functions is algebraic.
3. Solve Problem 2 for the product.
4. Find the discriminant D0(z) of (3.1.7), given that

V(z, u) = zu' + 2z2u2 + z +2.

5. Determine the fixed singularities of the DE

(z + w)w' - z + w = 0.
Show that only the point at infinity qualifies and that it is a pole of order 2 for
all solutions. Show that there are mobile branch points of order 1. (The
equation has elementary solutions.)
 81
ANALYTIC CONTINUATION; MOVABLE SINGULARITIES

6. In the equation
w' = P„(z) + P,(z)w + • • • + Pn(z)W
the P's are polynomials in z and n > 1. Show that the fixed singular points are
the zeros of P„(z) and the point at infinity.
7. In Problem 6 suppose that the P's are constants. Try to show that, if the
equation is satisfied by an entire function w(z), then w(z) must be a constant.
Show that there are always singular constant solutions, and describe them.
What happens for n = 1?

3.2. ANALYTIC CONTINUATION; MOVABLE SINGULARITIES

We consider again the system

w' = F(z, w) = w{Zo) = Wo> (3'21)

with the unique solution w(z; 20, w0). Here P and Q are polynomials in w
of degree p and respectively, and the coefficients are algebraic
functions of z. Furthermore, (z, w)»-> F(z, w) is holomorphic at (z0, w0).
In particular, z0 does not belong to the set S of fixed singularities.
Starting at z = z0, we continue w(z; z0, w0) analytically along a path C.
We continue F[z, w(z ; z0, w0)] along the same path. This means that z has
to avoid the singularities of the coefficients A, and Bk, as well as points
where C?(z, w) could vanish. If w(z; z0, w0) admits of analytic continua-
tion, so do its powers of finite order; i.e., both P[z, w(z ; z0, vv0)] and
h>(z; Zo, Wo)] can be continued along any path along which
w[z, z0, v^o) can be continued, but their quotient may possibly impose
some conditions on the path C. At any rate, under the stated assumptions
the law of permanence of functional equations (Theorem 1.11.2) permits
us to conclude that the analytic continuation of the solution w(z ; z0, vv0)
satisfies the DE all along C. Moreover, if z = z* is a point on C and if
w(z*; z0, w0) = w*, there exists a disk, its center at z*, in which
w(z;z0, w0) coincides with the local solution w(z;z*, w*).
We now suppose that w(z ; z0, w0) has been continued analytically along
C from z = Zo to z = zu and ask what will happen to the solution as z z,.
There are various possibilities.

1. w(z; z0, H\>)-> Wj and F(z, w) is holomorphic at (z,, w,).

2. F(z, w) is not holomorphic at (z,, w,), but [F(z, w)]"1 is holomor-
phic there and Q(z,, w)^0.
3. As in possibility 2, but Q(z,, w) = 0.
4. P(z, w) and Q(z, w) are holomorphic at (zu w,), but both are zero
there.
82 SINGULARITIES

5. At least one of the functions P and Q is not holomorphic at (z,, w,).

6. As z ->Zi, w -»oo.
7. w(z; z0, Wo) does not tend to any limit, finite or infinite, as z -»Zi.
We list here the results to be proved. In case:
1. The point z = z, is not a singularity.
2. The point is a movable algebraic branch point.
3-5 and 7. The point z = z, belongs to the set S of fixed singular points.
6. The point is a singularity which may be fixed or movable.
Let us start with case 1.

THEOREM 3.2.1

If analytic continuation of a given solution of (3.2.1) leads to a point z,

and a function value w = w, and if F(z, w) is holomorphic at (z,, w,), then
the local solution w(zzi, w,) gives the analytic continuation of the given
solution in the disk where w(z;z,,w,) is holomorphic.

Proof In this case a local solution w(z; z*, w*) exists for all z* on C,
including the point z = z,, where we can take w* = u>,. Now w(z; z,, w,)
is holomorphic in some disk |z - Zi| < r, and this disk contains the end of
C. If z*GC and |z,-z*|<r, then at z = z* we have two solutions,
w(z; z*, w*) and w(z; z,, w,). They are identical provided that they take
on the same value at z =z*. In this case w(z\zu w,) is the analytic
continuation of w(z; z*, w*) and vice versa. Now along C the radius of
convergence of the power series representing w (z ; z *, w *) is a continuous,
positive function of z* and is bounded away from zero. This follows from
Cauchy's formula (2.5.12) if we observe that F(z, w ) is holomorphic at each
point (z*, w*) for z* on C and w* = w(z*; z0, w0), except possibly at the
end point, where w* = wu We may assume that C is rectifiable and that
C:z = z*(f), (3.2.2)
where L is the length of C. This is a compact set in the plane. At each
point z*(0 on C there are corresponding quantities a(t), b(t), and M(0-
These quantities refer to the local expansion of F(z, w) in powers of
z - z*(0 and w - w*(t) and may be assumed to vary continuously with /.
Formula (2.5.12) then asserts that the local solution w[z ; z*(0, w*(t)] has
an expansion which converges at least for

|z-z*(0|<KO, O^r ^L, (3.2.3)

where

(3.2.4)
ANALYTIC CONTINUATION; MOVABLE SINGULARITIES 83

Now we can cover C by a set of disks defined by (3.2.3), and since C is

compact the Heine-Borel theorem ensures the existence of a finite
subcovering corresponding to

t0 = 0<tl<t2<" <tn<L.
We may assume that the subcovering is so dense that the center of the
(J + l)th disk lies in the /th one. This implies that the nth disk contains zu
so that w[z ; z *(*„), w*(tn)] is holomorphic at z — Z\. Since this solution is
the result of n successive rearrangements [see Section 1.11, formula
(1.11.3)] of the solution u>(z; z0, u>0) at the points z*(tj),j - 1, 2, . . . , n, its
value at z = Z\ is wit the postulated limit of the solution for analytic
continuation along C. Thus we have two solutions of the DE at z = zx with
the same value wx there. Hence they are identical, and the theorem is
proved. ■
This theorem is due to Painleve and is often called by his name. We
shall encounter deeper theorems, however, which also bear his name. We
now consider case 2.

THEOREM 3.2.2

// w(z; z0, w0)^ w, as z^z,, // P and Q are holomorphic at (Zi,w,), //

P(z,, wO 7^ 0, and if Q(zb w) has a zero of exact multiplicity k at w = w,,
then z = Zi is an algebraic branch point of order k for w(z; z0, w0).
Proof. The assumptions evidently imply that F(z, w) is not holomorphic
at (Zi, w,), while [F(z, w)]_1 = G(z, w) is holomorphic and not zero at all
points (zi, w). We now consider the solution of the DE,

-^=G(z,w), z(w,) = zi. (3.2.5)

This equation has a unique solution by the existence and uniqueness

theorems, say,

z(w;w„2,) = z1+2n = l c„(H>-H>,)n. (3.2.6)

Since G(z,, w) has a multiple zero at w = w,, a certain number of the

coefficients cn are zero. To make this precise, suppose that G(z, w) is
presented to us in the form

G(z,w)=2 gn(w)(z-z,)\ (3.2.7)

where the functions w gn(w) are holomorphic in one and the same
 84 SINGULARITIES

disk, \w - wx\ < r. For z = z, this reduces to g0(w), so that

G(z,, w) = go(w) = 2 gmo(w w,)m. (3.2.8)
Now by assumption Q(z„ w) has a zero of order k at h> = w„ and, since
P(z„ w,)^0, this implies the same property for G(z,, w). It follows
that
goo = gw = • ' ' = gk-U0 = 0, gk0 5* 0.
We have now from (3.2.5)

2 mcm{w - wT~x = n=0

m~l 2 gH(w)\Lm=0
t cm(w- wrX.J (3.2 9)7
Suppose that in the expansion (3.2.6) the series starts with
the power
(w - w,)p. This means that in (3.2.9) we have on the left (w - w)p~x as the
lowest
Hence power, while on the right (w - w,)k furnishes the lowest power

P = * + (k + \)ck+l = gko*o.
Thus we have

z - z, = (w - wt)k+i 2 c*+1+m(w - w,r. (3.2.10)

Since the power series does not vanish for w = wu we can extract the
(k + l)th root on both sides and obtain

(z - Zl)1/(fc+1) = (w - w.) ± <4+1+m(w> -

m =0 (3.2.1 1)
Here = (cfc+1),/(*+,) ^ 0, and the power
convergence by the double series theorem series has a positive radius of
of Weierstrass. We set
(z-Zl)i/(*+i> = Sj w-wx = t,
and want to solve the equation

s = t 2 <t+i+«r (3.2.12)7
m =0
for r. This we can do with the aid of the implicit function
theorem
(Theorem 2.6.1), the assumptions of which are satisfied.
Hence

t = 2 blS>
i= i
with a positive radius of convergence. Reverting to the old
variables, we
get
= J>((2-2iy«<-'>. (3 2 ,3)

This means that the analytic continuation of w{z; z0>

w0) along C has led
ANALYTIC CONTINUATION; MOVABLE SINGULARITIES 85

to a branch point of order k of the solution where k + 1 branches are

permuted cyclically. ■
Here a question arises: In what sense is this a movable singularity, or
would it have to rate as a fixed one? To elucidate this point let us examine
the conditions for this type of singularity to occur. Give z a value z°, not
belonging to the set S of fixed singularities, and examine the equation

Q(z\w)=tBk(z°)wk=0. (3.2.14)

This is an algebraic equation for w of degree q unless Bq(z°) = 0.

Normally there are q roots w°u w$, . . . , w°q. If vv° is one of these roots, the
event w(z)-» w° as z -»z° gives rise to a branch point for w(z) at z = z°,
and normally the branch point is of order 1 with two permuting branches.
Now the roots of (3.2.14) are continuous (even analytic) functions of z°, so
a small change in z° imposes a small change in w°. This means that there is a
solution w*(z) close to w(z) which has a branch point z°* near to z°, and
normally this branch point is also of order 1. In this sense we can speak of
movable branch points.
We now consider cases 3-5. The point z, is a fixed singularity in each of
the three cases, and almost anything and everything can happen at such a
point. Some examples were given in the preceding section, (3.1.15) and
Exercise 3.1. More examples are to be found in Exercise 3.2.
As for case 6, here we set w = \/v. The transformed equation is

v' = -v2 P^Au\,

Q(z, l/v) v(z)-+0 as z->z.. (3.2.15)
Again we assume that Zi £ S. There are two distinct subcases according to
whether q + 2 - p is >0 or =ss0.
Case 6:1. If q+2>p, (3.2.15) becomes

v. = _v^-^ = §77^- (3-2-16)

0 v > - >
±Mz )
Here P,(z,, 0) = 0 and Q,(z,, 0) = J5q(z,), which is ^ 0 since z, is not a
fixed singularity. For the same reason all the coefficients A, and Bk are
holomorphic at z - zx. Since it follows that the last member of (3.2.16) is
holomorphic at (zu 0), (3.2.16) has a unique solution which takes the value
0 at z = Z\. By inspection one sees that v(z) = 0 is that solution, and this
means that case 6:1 cannot occur.
Case 6:2. Suppose first that q + 2 = p. Then (3.2.16) takes the form

v' = G(z, v), where G(z„ 0) = -^f£ (3-2.17)

86 SINGULARITIES

Now G(z, w) is holomorphic at (z,, 0), so there is a unique solution which

vanishes at z = z,. If Ap(z.) ^ 0, this is a simple zero; but if Ap(z) has a
zero of multiplicity m at z = z,, then t>(z ; z,, 0) has a zero of multiplicity
m + 1 at z = z,. This means a pole of multiplicity m + 1 for w(z ; z0, h>0).
Next, suppose that q+2<p. Then we replace (3.2.16) by

-^=t;p -«-2tf(i;,z), z(0) = z„ (3.2.18)

with obvious notation. Here we may assume H(v, z) to be holomorphic at
(0, z,) and different from zero. In this case Ap{zx) * 0 since z, does not
belong to 5, and this guarantees the holomorphism of H. Now H(0, z,) = 0
iff BJz,) = 0, so our assumption implies that Bq(zx) ^ 0. If this assumption
is not satisfied, the following argument has to be modified along the lines
used in the proof of Theorem 3.2.2. Under the stated assumption, (3.2.18)
has a unique solution of the form

z - z, = vp-q~\c* + cxv + •••), Co * 0. (3.2.19)

Here we extract the (p - q - l)th root and invert the result to obtain

v(z) = 2 d,(z - z,)'7^0, dx * 0. (3.2.20)

Now w(z) is the reciprocal of this, so that

w(z) = (z - z1r1/(p"<,_1) i=0

S - Zi)//(P"<,_1), (3.2.21)
where e0 ^ 0 and the series has a positive radius of convergence. If Bq{zx)
should be zero and, more generally, H(v, z,) has a zero of order k, then in
(3.2.21) p-q-l should be replaced by p - q - 1 + k, as is seen by
imitating the argument used in the proof of Theorem 3.2.2. Thus we have
proved: ■

THEOREM 3.2.3

Equation (3.2.1) has no movable infinitudes if p<q + 2. There are

movable poles if p = q -I- 2. The poles are simple unless accidentally Ap(z)
vanishes at the point in question. If p > q + 2, there are movable branch
points where the solution becomes infinite. Normally p-q-l branches
are permuted at such a point.

The behavior of the solutions of (3.2.1) as z -*o° was studied by Pierre

Boutroux (1880-1922) in a monograph published in 1908. (French
mathematicians are often related: Boutroux was a nephew of Poincare.)
If p = q 4- 1, the integrals are of exponential growth (at least if the
PAINLEVE'S DETERMINATENESS THEOREM; SINGULARITIES 87

coefficients are polynomials, the only case considered by Boutroux). If

p # q + 1, the solutions grow as powers of z, omitting neighborhoods of
the infinitudes when p - q > 1 . We shall take up such questions in
Sections 4.6 and 11.2.

EXERCISE 3.2

1. Discuss the fixed singularities of w' = zm + zmw - w2.

2. The solution of w' = z3 + w\ w(0) = w0, has movable branch points. Find the
order. If w0 > 0, the branch point nearest to the origin lies on the positive real
axis. How does it move if w0 increases?
3. The equation w' = z2 + w\ w(0) = 0, has movable poles (of what order?). If
h>0>0, there are positive poles (why?). How does the pole nearest to the
origin move when w0 increases?
4. What fixed singularities, if any, do the equations in Problems 2 and 3 have?
5. How should w0 be chosen in order for the equation in Problem 3 to possess
solutions with poles on the imaginary axis?
6. Fill in the missing details in the proof of Theorem 3.2.2.
7. Solve the same Problem for Theorem 3.2.3. In particular, carry out a proof
for the case when Bq{zx) = 0 and H(v, z,) has a zero of order k at v = 0.
Show that the following equations have fixed singularities, where for suitable
approach the solutions tend to no limit.
8. w' = z~2w.
9. w' = i(l-z)_1w.
10. w' = w.

3.3. PAINLEVE'S DETERMINATENESS THEOREM; SINGULARITIES

We have to show that case 7 can occur only at a fixed singular point of the
DE. Note that our conclusions have always been based on a pair of values
(zi, h>i), where wx = \\m w(z) as z-»Zi and where the limit is finite or
infinite. If there is no limit, the point of departure for an attack on the
problem -is lost. Painleve was the first to call attention to this difficulty in
1888 and to show that it cannot arise for a DE of type (3.1.4) and (3.1.5) as
long as Zi does not belong to the singular set S. Here the restrictive
assumptions on P and Q are essential, and the result does not hold for
systems of first order equations or equations of second or higher order.

THEOREM 3.3.1

If the solution z w(z; z0, w0) of (3.1.4) is continued analytically along a

rectifiable arc C from z = z0 to z = zu avoiding the set S of fixed
88 SINGULARITIES

singularities, and if z, £ S, then the solution tends to a definite limit, finite

or infinite, as z->Zi.
An equivalent and shorter formulation is

THEOREM 3.3.2

If F(z, w) is a rational function of w with coefficients which are algebraic

functions of z, then the movable singularities of the solutions are poles
and lor algebraic branch points.
Proof Suppose that the analytic continuation has been achieved along C
and that as z approaches z,, not in 5, the solution tends to no limit. It will
be shown that this assumption leads to a contradiction. Mark in the
w-plane the q roots, wx, w2, . . . , wq, of the equation
<2(z„ w) = 0. (3.3.1)
For each k let yk be a small circle with center at wk. Then there exists a
p >0 such that for |z - z,| < p the equation

Q(z,w) = 0
has one and only one root inside each circle yk and none outside the
circular disks. This implies the existence of a constant M such that on the
rim of each circle yk and for |z - zx\ < p we have
P(z, w)
<M. (3.3.2)
Q(z,w)
Furthermore, we take a circle T with center at z = 0 and a large radius. We
can take M so large that (3.3.2) holds also on Y and |z - z,| < p. Let T be
the domain in the w-plane inside T and outside all circles yk. Then (3.3.2)
will also hold for w in T and |z - z,| < p by the principle of the maximum.
As z describes C, we have several possibilities for the behavior of
w(z;z0, Wo). If ultimately w gets outside T and stays outside, then
w(z; z0, Wo) tends to infinity. If the path traced by the solution ultimately
gets inside one of the circles yk and stays inside, the solution tends to the
limit wk. By assumption none of these events takes place. This means that
on the arc C there is an infinite sequence of points {z,} such that
w(Zj ; z0, vv0) = Wj lies in T and z, -» zx. If |z, - z,-| < p, then at the point
(zh Wj) the function F(z, w) is holomorphic and less than M in absolute
value. The local solution w(z; zh Wj), which coincides with the analytic
continuation of w(z; z0, w0), is holomorphic in a disk of positive radius rh
The bounded infinite sequence {W,} has at least one limit point, WQ say.
Now at (zi, W0) there is a local solution w(z; z,, W0), holomorphic in a
PAINLEVE'S DETERMINATENESS THEOREM; SINGULARITIES
89
disk \z - Zi\< r0. This disk contains infinitely many points z, for which
Wj -> W0. It follows that, if the analytic continuation of w(z ; z0, w0) along
C is via the function elements w(z; z,-, Wj), using only those values of j
for which f-» VK0, then lim w(z; z0, h>0) = Then Theorem 3.2.1
shows that the continuation of our solution is holomorphic at z = z, and
has the definite limit W0. This is a finite quantity, and F(z, w) is
holomorphic at (z,, W0). This shows that the hypothesis that the analytic
continuation tends to no limit as z -> Zi is untenable, and the theorems are
proved. ■
Consider, in particular, the equation
(3.3.3)
w'(z) = Po(z) + P1(z)w(z)+. • . + Pm(z)[w(z)r,
where the P's are polynomials in z and m > 1. The coefficients admit
z = oo as the only singularity. Here Q(z, w) = 1, so cases 2-5 cannot occur.
Case 6 will occur and give movable infinitudes. If Pw(zi)^0, the
corresponding expansions take the form

where A = (m - l)Pw(z,). Thus, if m = 2, the singularity is a movable

pole; if m>2, a movable branch point where the solution becomes
infinite and m - 1 branches are permuted when z makes a circuit about
the point. The fixed singularities of this equation are the point at infinity
and the zeros of Pm(z).
The point at infinity is usually a singularity of a higher order: a limit
point of poles if m = 2, of infinitary branch points if m > 2. A fc-fold zero
of Pm(z), say at z = a, permits the existence of solutions holomorphic at
z = a, and also solutions with a branch point where the solution becomes
infinite and the order of infinitude is at most (k + l)/(m - 1). The case
m = 2 is known as the Riccati equation.
We may conclude from the preceding discussion that the equations
whose movable singularities are poles are Riccati equations. In view of the
importance of this fact we shall give a formal proof in the setting of (3.1.4)
and (3.1.5).

THEOREM 3.3.3
Consider the equation
P(z,w)
w' = F(z, w) = (3.3.5)

where P and Q are polynomials in w of degree p and q, respectively, with

coefficients which are algebraic functions of z. 7/ no solution of the
Q(z,w)'
90 SINGULARITIES

equation can have a branch point at a nonsingular point of the equation,

then F(z, w) is a quadratic polynomial in w, that is, (3.3.5) is a Riccati
equation,
w'(z) = A0(z) + A,(z)w(z) 4- A2(z)[w(z)]2. (3.3.6)
Proof. We start by noting that Q(z, w) must be of degree 0 in w.
Otherwise, setting z = z0 with z0 not in S, we find that the equation
Q(z0, w) = 0 has q roots. If w0 is one of the roots, then by Theorem 3.2.2
there is a solution taking on the value w0 at z = z0 and this solution has a
branch point there. Hence we must have q = 0, and we may take
Q(z, w)= 1. Thus the equation is of the form (3.3.3), where the coeffi-
cients are algebraic functions rather than polynomials. Hence we are
dealing with

w' = P(z, w) = A0(z) + A,(z)w + • • • + AP(z)wp. (3.3.7)

If now p > 2, we set w = \/v and obtain

v'(z) = - [A0(z)vp + Al{z)vp~l + • • • + Ap(z)]v2'p.

Here Q\(z, v) = vp~2, which vanishes for v = 0 regardless of the value of
z. Taking a nonsingular value z0 for z, we have Ap (z0) ^ 0 and the
transformed equation has a solution which vanishes at z = z0 and has a
branch point there. Hence the original DE has a solution w(z) = [u(2)]_1
with a branch point at a nonsingular point z0 contrary to the assumption.
Thus we must have p ^ 2. ■
Riccati's equation has a number of interesting properties, which will be
discussed at some length in Chapter 4.
The rest of this section will be devoted to further discussion of
singularities, fixed or movable, of (3.1.16), (3.3.3), (3.3.7), and related
types of equations. The basic tool is the test -power test. This is most
effective in the case of movable singularities but can be used also with
fixed ones, at least for purposes of orientation. The simple idea is that, if a
first order DE has a solution which at some point z, becomes infinite as

a{z-zxya while w'(z) ~ - aa(z - z,)""-1 , (3.3.8)

in both cases up to terms of lower order, then a and a must be so chosen
that terms of highest order cancel when (3.3.8) is substituted in the
equation. Let us try this out on some simple examples where the facts are
known.
For the tangent equation
w' = \ + w2 (3.3.9)
we find a = 1, a = - 1, and in this case z = z, is a movable simple pole
with residue - 1. This is indeed the situation for w(z) = tan (z - z, - \tt).
PAINLEVE'S DETERMINATENESS THEOREM; SINGULARITIES 91

For the Jacobi equation

w' = [(l-w2)(\-k2w2)]m (3.3.10)

we get a = 1, a = ± l//r, values which agree with known facts.
Since these results are rather encouraging, we shall try to apply the
method to equations where we do not know the true situation in advance.
The Briot-Bouquet equation (3.1.16) is a good testing ground [actually
(3.3.10) is a Briot-Bouquet equation]. If we are looking for solutions with
prescribed types of singularities and if the test-power test shows that a
given equation cannot have such singularities, the test is conclusive. On
the other hand, if the test shows that the equation could possibly admit
solutions with such and such singularities, an examination of the approp-
riate series expansion becomes necessary before we can accept the
evidence. Thus in the case of the very simple second order nonlinear
equation
w" = zw2 (3.3.11)
the test for movable singularities gives a = 2, a = 6, but the point zx is not
a double pole. See Section 12.4.
We return to (3.1.16). It contains aggregates of the form

With each such product we associate a straight line

y =(j + k)x + k. (3.3.12)

We are interested in the points of intersection of such lines (see Figure
3.1). Their coordinates are rational numbers and candidates for the value
of a. There are essentially two different cases.

I. All points of intersection are located in the closed left half-plane.

II. Some intersections fall in the open right half plane.

THEOREM 3.3.4

In case I the equation can have no movable infinitudes of type (3.3.8). In

case II, // (x0, y0) is a point of intersection of lines (3.3.12) such that (i)
x0 > 0, y0 > 0, (ii) the vertical line x = x0 has no intersection with the system
(3.3.12) above the point (x0, y0), and (iii) (x0, y0) is the only point of
intersection with these properties, then a = x0 gives the only admissible
movable infinitude of type (3.3.8). In case I an intersection in the left
half-plane will give movable branch points where the solution is zero,
provided that the transformation w = 1/v leads to an equation in which the
condition under II holds.
92 SINGULARITIES

(1. 10)

rti

ft
i

644- <

0 Figure 3.1

Proof. Suppose that case I is present. We substitute the test powers

(3.3.8) in the DE

2wW(H>')fc=0. (3.3.13)
If there exists a solution with an infinitude of this type, it must be possible
to determine a (and a) so that the terms of highest order are of equal
weight and can be made to cancel. This requires, in particular, that there
 93
PAINLEVE'S DETERMINATENESS THEOREM; SINGULARITIES

be two points, (j, k) and (m, n), such that the equation

(/ + k)a + k = (m + n)a + n
is satisfied by a positive number a. This is clearly impossible if all lines
(3.3.12) intersect in the left half -plane.
Suppose now that case II is present. Then there are two lines in the set
(3.3.12) which intersect at a point (jc0, y0) in the open right half -plane.
Suppose that these are the lines

y = (j + k)x + k and y=(ra+n)jt + M.

If they are to intersect, one line must have the larger y -intercept, the other
the greater slope. Suppose that k < n, j > m. Then
n - k
x0 = t-- v — ; rr. (3.3.14)
(j-m)-(n -k)
This is a positive rational number. Hence for a suitable choice of a the
terms

(- x0)kai+kcik(z - Zlru+k)x°~k and (- xo)nam+ncmn(z - ZiT(m+n^-n

are the largest and will cancel. Condition (ii) of Theorem 3.3.4 guarantees
that in the interval (0, jc0) all the other lines of the system lie below the line
y = (j + k)x + k and that possible intersections in this area need not be
considered. Intersections (jc,, yO with jc0 < xx are not excluded, and this is
why condition (iii) is needed. Thus we see that (3.3.13) has movable
infinitudes.
For such an infinitude to be a pole, say of order N, we must have
(n - k)(N + 1) = (J - m)N. This implies the existence of an integer M
such that
/-m=M(N+l), n-k=MN. (3.3.15)
If no such integers M and N can be found and x0 > 0, the singularity is not
polaroid but algebroid. In the case under consideration logarithmic
perturbations cannot occur, so a polaroid singularity is a pole and an
algebroid singularity is algebraic.
Let us now return to case I for a moment. If the transformation w = l/v
is applied to (3.3.13), we obtain after multiplication by t>p, where
p = max (j + 2k),

2(-OkcJko'-|-ak(o')*=0, (3.3.16)
which is of type (3.3.13). Any intersection of lines of system (3.3.12), say
yO with x, < 0, will correspond to intersections of lines of the system

y =(p -j-k)x + k. (3.3.17)

94 SINGULARITIES

In particular, (jc,, y,) will go into a point (- xu y2) in the first quadrant. If
this point satisfies the conditions under case II, then (3.3.16) has a mobile
infinitude at z = zx and (3.3.13) has a solution with a movable zero at z,. If
— jc i is a positive integer, this is a regular zero where the solution is
holomorphic, otherwise an algebraic branch point. This completes the
proof. ■
We have also proved ■

THEOREM 3.3.5

A necessary condition that (3.3. 13) fee satisfied by an elliptic function is that
among the exponents one can find two pairs (j, k) and (m, n) such that the
x0 defined by (3.3.14) is a positive integer which satisfies conditions (ii) and
(iii) of Theorem 3.3.4.
The various cases encountered above are typified by the following
equations:

(1 + w2)w' = 1, no finite infinitudes, movable zeros;

(w')2 = 1 + w\ movable double poles;
w ' = 1 + w 3, movable branch points, order 1 .
Our theorems leave open the question of what happens if the conditions
under case II hold neither for (3.3.13) nor for (3.3.16). Here no singularity
of type (3.3.8) seems to be possible. A case in point is the equation
(3.3.19)
(1 + w)(w')2- w2w' + w2 = 0.
Here we can solve the equation for w' , obtaining two roots. One leads to a
solution asymptotic to e\ the other to z. Here, however, z is an
elementary function of w, having a logarithmic singularity.
The test-power test can be applied to first order equations with variable
coefficients and to equations of higher order. Here it becomes imperative
to test the result by other means. Equation (3.3.1 1) may serve as a warning
in this connection.
In a discussion of (3.3.13) it is useful to observe that w' is an algebraic
function of u>, the singularities of which can be read off from the
discussion at the beginning of Section 3.1. The question of infinitudes of
solutions of (3.3.13) is now reduced to the question of the behavior of w'
for large values of w. This is a classical problem in the theory of algebraic
functions. The method is variously known as Newton's diagram or
polygon, as the algebraic or the analytical triangle, or as the method of
Puisseux (after V. A. Puisseux, 1820-1883). It is closely related to the
method used above, which led to the line (3.3.12). Here instead we plot the
points 0, k) in the (x, y)-plane and draw the least convex polygon which
 95
PAINLEVE'S DETERMINATENESS THEOREM; SINGULARITIES

contains these points. See Figure 3.2, which illustrates the case

(w')5+ w2(w')4 + 2(w')4- w(w'f- w5 = 0. (3.3.20)

The sides of the polygon which face infinity give asymptotic relations of
the form
w'~aw\ (3.3.21)
where - 1//3 is the slope of one of the boundary lines. In the figure = 2
or i The first value corresponds to a solution with simple poles, and the
second to a solution ~ az4 for large values of z. If so desired, we can
elaborate (3.2.21) into a convergent series

but (3.3.21) is sufficient for our present purpose. According to the value of
/3, we have various possibilities:

1. p = 1 + 6, 8 >0, gives
w~[S(z-z1)r,/s. (3.3.22)
This is a mobile infinitude, a pole of order N if 8 = \/N.
2. 0 = 1, w~C<?a2.
96 SINGULARITIES

3. 0< 0 < 1. We have

w~[(l-j8)a(z-z1)],/(1^), (3.3.23)
a mobile branch point where w = 0.
4. /3 = 0, w ~ a(z - z,), a mobile zero.
EXERCISE 3.3

1. Equation (3.3.3) with m>2 has movable branch points so that the general
solution has at least m - 1 determinations. Show by an example that it need
not have more than m - 1 determinations.
2. Apply the test-power test to (2.7.3), and compare the result with (2.7.22).
3. Apply the test to the equation w' = 1 + z2w 3 at the fixed singularity z = 0. Try
to verify the result.
4. Show that the general solution of w' = 1 + zmw2 is single valued on the
Riemann surface of zl/2. Try to generalize to other cases involving algebraic
coefficients.
5. Suppose that the coefficients of a Riccati equation are polynomials in z.
Show that the general solution is single valued.
6. Verify (3.3.2).

Assuming the validity of the test-power test, apply it to the following DE's:
7. (w'f = 4(w - e,)(H> - e2)(w - ei), e, + e2 + e3 = 0.
8. w" = 6w2-\g2.
9. w"=z'xw2.
The equations in Problems 7 and 8 are satisfied by the p-function of
Weierstrass, and the movable singularities are double poles. The equation in
Problem 9 is a special Thomas-Fermi equation, and the movable singularities
look like double poles but are not. The discrepancy shows up in the sixth
term of the proposed Laurent series.
10. In the proof of Theorem 3.3.1 does the assumption that the coefficients A,
and Bk are algebraic functions play an essential role in the argument? Could
it be replaced by a less restrictive assumption without jeopardizing the
conclusion?
11. Consider the equation az + bw(z)
w'(z) = cz + dw(z)

where a, b, c, d are arbitrary fixed constants with ad - be ^ 0. What is the

nature of the fixed singularities? Of the movable singularities?

3.4. INDETERMINATE FORMS

Except for a problem in Exercise 3.3, there has been no mention of case 4.
This type of fixed singularity has been the object of much work for close
INDETERMINATE FORMS 97

to a century and has had a great direct and indirect influence on the
development of mathematics. What started with Poincare in 1878 as a
scrutiny of the properties of functions defined by an ordinary DE
developed in his fertile mind into a theory of the curves defined by a DE
(1879, 1881, 1885, 1886), which laid the foundation of the geometric theory
of differential equations. Moreover, it led him to the creation of what he
called analysis situs, now known as topology.
A fixed singularity z = z0 exhibits case 4 if there is a point w = w0 such
that in the quotient
P^w>
(3 4 1)

both numerator and denominator are holomorphic at (z0, w0), but

P(z0, Wo) = Q(z0, Wo) = 0 so that P/Q becomes an indeterminate form 0/0
at (z0, Wo). We place the singularity at (0, 0), and in this chapter we shall
consider only the case in which both P and Q are linear homogeneous
functions of z and w. More general considerations are postponed until
Chapter 12. Thus

w , = flz+*?w ad_bc7c 0
cz + dw (3 4 2)

Together with this equation, it is occasionally convenient to consider the

system of autonomous equations

z(0 = cz(t) + dw(t), w(t) = az(t) + bw(t), (3.4.3)

where Ms an arbitrary parameter and the dot indicates differentiation
with respect to t (Newton's notation). The word "autonomous" refers to
the fact that t does not occur explicitly in the equation. If [z(t), w(t)] is a
solution of (3.4.3) and p is an arbitrary number, [z(t + p), w(t + p)] is also
a solution. Furthermore,

w(Q = az(t) + bw(t)

z(t) cz(t) + dw(t)' V ''
so [z(0, w(0] is a parametric representation of the solution of (3.4.2).
In Chapter 12 we shall be concerned, at least briefly, with the case in
which

P(z, w) = 2 aikzjw\ Q(z, w) = 2 bikz!wk. (3.4.5)

There is an old tradition in applied as well as pure mathematics of
linearizing a difficult problem in the hope that the linear case will give a
reasonable approximation of the nonlinear problem. The results so
obtained are often misleading, and if the equations involve no linear terms
at all the method fails.
9H SINGULARITIES

After these words of warning let us take a look at (3.4.2), where we

want a solution w(z) that tends to zero with z. The equations are solvable
by elementary means, but it is usually difficult to obtain the desired
information from the elementary solution. There are a couple of cases,
however, in which explicit solutions of w in terms of z are available and
are of interest to our problem. Take c — 1, d = 0 and the equation

w> = a+bj. (3.4.6)

If b 1 , the general solution is

w =-r^—z
o — 1 + Cz\ (3.4.7)

while, if b = 1, the solution becomes

w = z(a log z-C), (3.4.8)
where C is an arbitrary constant. In the second case all solutions go to
zero with z. In the first case this will be true iff Re (b) > 0. If Re (b) ^ 0,
there is one and only one solution that goes to zero with z. These special
cases are of interest both for the nature of the solution and for the
conditions that have to be imposed on the constants a to d.
We now return to the general case in the form of the system (3.4.3). An
equivalent form is the matrix-vector DE
y(t) = My(t\ (3.4.9)
with
v(t)

[;'<?,]• He" i]-

and the derivative of a vector is defined as the vector of the derivatives of
the components. Here the characteristic values and the characteristic
vectors of M become important (see the end of Section 1.4), and the
characteristic equation is
a — A b
c d-k = 0 or \2-{a+d)k+ad-bc = 0. (3.4.11)

Since ad - be ^ 0, no root can be zero, i.e., M is a regular (= nonsingular)

matrix and has a unique inverse M~x. The roots of (3.4.1 1) may be distinct
or equal, and the solutions of (3.4.9) are essentially different in the two
cases. If the roots are distinct, they are numbered Ai, A2 with the
convention that Re (Ai) ^ Re (A2). If the sign of equality should hold here,
it is required that Im (AO > Im (A2). If the roots are equal, the common
value is \{a + d).
INDETERMINATE FORMS
99
Let Hft2 denote the space of all 2-by-2 matrices. We make it into a metric
space by defining a norm for si — (ajk):

= max(|flJ-,| + |flJ-2|). (3.4.12)

In this metric
||JC-H||iC||". (3.4.13)
We define the exponential function by the power series (with % the unit
matrix)

zxp(Mt) = % + ^t+^jt2+-- +f['n + -'-- (3-4-14)

Here the sum of the norms is at most exp (||it|| \t|), so the series is
absolutely convergent in norm. It is a differentiate function of t with
derivative it exp (tM), as is seen by (permissible!) term-by-term differen-
tiation of the series (3.4.14).
It follows that the solution of (3.4.9) with the initial value v0 is

v(0 = exp (itOvo. (3.4.15)

This is not quite the solution that we want. Indeed, we want a solution
which goes to the zero vector when t approaches some limit. What can be
achieved in this direction will depend on the characteristic values of M.
If At 5* A 2, the normal form of the matrix M is

9)
and there exists a nonsingular matrix si such that

M = &-l(2b& or <& = sULsl-1. (3.4.17)

We can transform (3.4.9) into an equivalent equation by multiplying both
sides on the left by si to obtain

M(t) = siMsi-'si\(t). (3.4.18)

Hence, if we set six(t) = u(0, then
u(*) = 2>u(f). (3.4.19)
This has now the solution

u(0 = exp (2) Ouo with u0 = s£v0. (3.4.20)

Now the powers of 2) are found to be

Lo a; J*
100 SINGULARITIES

and this shows that the solution may be written as

u('Ho
^ u0. (3.4.21)

If now the components of u(0 and u0 are, respectively, w,(0, u2(t) and
Uou w02, we have

uM = u(neK^ u2(t) = u02e^'. (3.4.22)

Now, if Re (A,) and Re (A2) are both positive, «,(0 and u2(t) go to zero as
t ->-oo. If both signs are negative, w,(0 and u2(t) go to zero as t -+ + oo.
Since in either case

lim u(0 = 0,
it follows that

lim v(0 = lim si~lu(t) = stl lim u(f ) = 0 (3.4.23)

when t tends to plus or minus infinity, as the case may be. This means that
the system (3.4.3) has a solution that goes to zero and hence also that
there is a solution of (3.4.2) which exhibits case 4 at (0,0). For 0>
Re (A,) > Re (A2) we have
w(z) = Cz[l + o(l)] (3.4.24)
as z -»0. Here C is an arbitrary constant since v0 is an arbitrary vector.
On the other hand, if Re (A,) > 0 > Re (A2), we get a different picture. As
above, we obtain (3.4.22) for ux(t) and u2(t); but now, letting t -» + we
get the limit zero for one coordinate and plus or minus infinity for the
other, and the same is true for f -»-<». We notice, however, that
= A,«,(f ), u2(t) = A2w2(0,
and this gives
A2m2(Owi(O-A,m1(O«2(O = 0,
whence

[«i(0rA2[«2(0]A' = C. (3.4.25)
If the A's are real, A! > 0 > A2, this is a system of hyperbola-like curves in
the («i, w2)-plane, and the coordinate axes are members of the family, the
only members that pass through the origin. We can make an orthogonal
transformation of the coordinates, carrying (uu u2) into (xux2) to get

[*,(f ) + x2(t)]-^[xl(t) - x2(t)p = C„ (3.4.26)

and then the solutions that pass through the origin become

jc2 = jc, and jc2=-jc,. (3.4.27)

INDETERMINATE FORMS 101

We have next the case of a double root, A0 = {{a + d) of (3.4.1 1). Here
we can take as the normal form

\ (3A28)
Again there is a nonsingular matrix L
Ho si such that

M = d~x$d or $ = sUisl-\ (3.4.29)

Here

and this gives

«(».)- [-«■» 'z<$\

whence

m.(0 = (wo. + u02t)ex°\ u2(t) = u02e v. (3.4.31)

If Re (A0) * 0, all solutions go to (0, 0) as t -> + oo or - oo. If Re (A0) = 0, the
solutions will still go to zero if t -> oo along a suitably chosen ray in the
complex f-plane.
We say that two vector-matrix equations are equivalent :
y(t) = Mtfit) equivalent to v(0 = M2\(t) (3.4.32)
if the matrices M.x and M2 are equivalent in the sense of matrix algebra,
i.e., there is a regular matrix such that
SMi, = M2®. (3.4.33)
This means that Mi and M2 have the same characteristic values, and all
equations in the equivalence class have a singularity of type 4 at (0, 0).

EXERCISE 3.4
1. Prove that equivalent matrices have the same spectra.

find characteristic values and vectors.

3. Prove (3.4.13).
4. Verify the statements concerning the convergence properties of (3.4.14) and
the differentiability of the series.
5. Verify (3.4.21) and (3.4.22).
6. Verify (3.4.24).
102 SINGULARITIES

7. If 2ft is an orthogonal matrix (i.e., its row vectors are orthogonal as well as the
column vectors) and 28 was used in passing from (3.4.25) to (3.4.26), show
that <&d\(t) is a solution of a vector-matrix DE equivalent to (3.4.9).
8. Verify (3.4.30) and (3.4.31).
9. Show that (3.4.7) goes into (3.4.8) when b »-» 1 and C is chosen as an
appropriate function of h = b - 1.
10. If t is eliminated between the two equations in (3.4.32) and w = v^t),
z = v2(t), show that a result is obtained similar to (3.4.8).
11. Solve w' = (z/wf, which exhibits case 4 at (0,0).
LITERATURE

As a general reference for this chapter see the author's LODE, Section
12.1.
For the theory of algebraic functions see the author's AFT, Vol. II,
Chapter 12. For details consult:
Bliss, G. A. Algebraic Functions. Colloquium Publications, American Mathematical Society,
New York, 1933.

Painleve's Stockholm Lectures are not easily available but were

published as:
Painleve, P. Lecons sur la theorie analytique des equations differentielles, professees a
Stockholm, 1895. Hermann, Paris, 1897.
See also:
Boutroux, P. Lecons sur les fonctions definies par les equations differentielles du premier
ordre. Gauthier-Villars, Paris, 1908.
For the theory of elliptic functions, see AFT, Vol. II, Chapter 13.
The Briot-Bouquet publications referred to are:
Briot, Ch. and J. CI. Bouquet. Theorie des fonctions elliptiques. 2nd ed. Paris, 1875.
Integrations des equations differentielles au moyen des fonctions elliptiques. /. Ecole
Polytech., 21: 36 (1856).

For the geometric theory of DE's see:

Lefschetz, S. Differential Equations: Geometric Theory. Interscience, New York, 1957; 2nd
ed., 1963 and the literature cited there.
Further references are:

Fuchs, L. Uber Differentialgleichungen deren Integrate feste Verzweigungspunkte besitzen.

Sitzungsber. Akad., Berlin (1884), 699-720.
Poincare, H. Sur un theoreme de M. Fuchs. Acta Math., 7 (1885), 1-32.
4

RICCATI'S EQUATION

In the beginning there was a Riccati — to wit, Conto Jacopo Riccati

(1676-1754). In the Acta Eruditorum (1723, pp. 502-574, and Supplemen-
ting VIII, 1724, pp. 68-74) he treated a DE which Daniel Bernoulli (ibid.,
1725, pp. 473-475) reduced to the form

y'(x) + ay2(x) + bxm =0.

It is closely related to Bessel's equation (Section 6.3).
This is the origin of what later became known as Riccati's equation. Its
unique character in the theory of first order nonlinear DE's and its
importance for first and second order linear DE's became clear much
later. In this chapter we shall discuss various aspects of the theory at
some length. We shall treat the elementary theory, the nature of the
singularities, the dependence of the solutions on internal parameters, and
some geometric consequences of the situation. The equation has some
extremal properties in the class of all nonlinear first order DE's expressed
by the Malmquist-Wittich-Yosida theorem. This calls for a sketch of the
Nevanlinna value distribution theory for meromorphic functions. This
theory plays an important role in the present chapter and later ones.

4.1. CLASSICAL THEORY -1

We take the Riccati equation in the form

y'(z) = /o(z) + /,(z)y (z) + /2(z)[y (z)]2, (4.1.1)

where the coefficients are holomorphic functions of z in the domain under
consideration. The equation is closely related to the linear second order
DE
w"(z) + P(z)w'(z) + Q(z)w(z) = 0, (4.1.2)
for if we set

y(z) = -[/2(z)rV(z)[w(z)] (4.1.3)

103
104 RICCATPS EQUATION

then w(z) will be a solution of

w"(z)-m(2)[/2(z)r,+/1(2)}w'(2) + /^(2)/2(z)>v(z) = 0. (4.1.4)

This is of the form (4.1.2) with

P(z) = -/2(z)[/2(z)] 1 -/,(2), Q(Z) = /o(2)/2(2). (4.1.5)

This is the passage from the Riccati equation to a second order linear
DE. But we can of course reverse the process and get a Riccati equation
corresponding to a given second order linear equation. Suppose that the
given equation is (4.1.2), where P and Q are holomorphic in some
domain. Here we set

y(z) = -^ (4.1.6)
and obtain the Riccati equation

y '(z) = - Q(z) + P(z)y(z) + [y (z)]2. (4.1.7)

It is clear that the correspondence between the two classes of equations
(4.1.1) and (4.1.2) is not one to one. To a given Riccati equation
corresponds infinitely many linear second order equations, and vice
versa.
The case where /0(z) = 0 is a special Bernoulli equation and hence
reducible to a first order linear DE, which is solvable by quadratures. The
equation
y'(2) = /(2)y(2) + g(2)[y(2)r, (4.1.8)
where n ^ 0 and 1 but is not necessarily an integer, is reduced to

v'(z) = (1 - n)[f(z)v(z) + g(z)] (4.1.9)

via the substitution ^ / z (j. ^ Jj n ^ '

v(z) = [y(z)]ln. (4.1.10)
Equation (4.1.8) was proposed by Jacob Bernoulli (uncle of the Daniel
Bernoulli mentioned above) in 1695 and was solved by his brother Johann
and by C. W. Leibniz in 1697, using different methods.
Since the solution v(z;z0, v0) of the linear equation

v'(z) + Pi(z)v(z) = R(z) (4.1.11)

is

i?(z) = u0exp[- jZ PMds^ K(s)exp[JZ Px{t)dt\ds, (4.1.12)

CLASSICAL THEORY tf{Kj if J '{t/ ^ _M5
it follows that the solution of V {-f\ - Jr- „

y'(z) = /1(z)y(z) + /2(z)[y(z)]2 (4.1.13)

is
y(z) = |t;0exp J

This is of the form - J" /2(5)exp[|^/1(0^]^j (4114)

(4.1.15)
v0C{z) + D(zY
i.e., a fractional linear transformation operating on the internal parameter
v0. As will be shown in the next section, the general solution of Riccati's
equation can always be written as a fractional linear (= Mobius) transfor-
mation of an internal parameter with coefficients which are functions of z
and are expressible in terms of particular solutions of the equation.
A Riccati equation is formally invariant under a Mobius transformation
operating on the dependent variable. Take an arbitrary set of four
constants, a, b, c, d, with ad - be ^ 0 and substitute
a + bv
J c + dv (4.1.16)

in (4.1.1). Then v satisfies the Riccati equation

(be - ad)v' = (c2f0+ acjx + a2f2)

+ [2cdf0 + (ad + bc)fi + 2abf2]v
+ (d2f0+bdfl + b2f2)v2. (4.1.17)
In particular, 1/y satisfies

-v'=f2{z) + fx(z)v+U(z)v2. (4.1.18)

We recall from Chapter 3 that the movable singularities of solutions of a
Riccati equation are simple poles. Since the zeros of y are poles of v9 the
solution of (4.1.18), it is seen that the zeros of y(z), not located at a fixed
singularity, are also simple.
A Riccati equation may have a rational function as a solution. Thus

y = ^j satisfies ~y' = z{J-x)+'y2- (4119)

EXERCISE 4.1
1. Verify (4.1.12).
2. Interpret D(z) in (4.1.15).
106 RICCATPS EQUATION

3. Show that the general solution of (4.1.19) may be written as

y =2z(z2-\ym[(z2-\)m-C] '.
Each particular solution with CV 0 has two simple poles on the Riemann
surface of (z2- l)l/2, and in this case z = ± 1 are simple branch points of the
solution. The special case C = 0 is the one figuring in (4.1.19).
4. The equation y' = l-y2 has a solution which becomes infinite as z = x
decreases to zero. Show that Jcy (jc) — > 1 . Show that this solution is positive and
decreases to + 1 as x -> + ». Does the equation have any constant solutions?
5. [H. Wittich] Show that the function z z~m tan zm satisfies the equation

Find the general solution. Discuss the distribution of the poles.

6. [H. Wittich] Show that the equation

y ' = (z3 - 1)V4 + (z3 + 2)(z4 - z)"*y + y 2

has a general solution of the form

y(z) = (z-z-2) tan (iz2 + z", + C).

7. The function z tan z is a solution of y ' = 1 + y2. Apply the transformation
(4.1.10) to this equation, and choose the constants a, b, c, d in such a manner
that the resulting equation becomes v' = 1 + v2. Use this device to obtain the
addition theorem for the tangent function.
8. The equation in Problem 7 has two constant solutions, y = + i and y = - i.
Use the existence and uniqueness theorems to show that tan z assumes the
values 4- / and - i nowhere in the finite plane.
9. Let y,, y2, y3, y4 be four distinct solutions of
y'-/o-/,y-/2y2 = o.
Substitution of these particular values of y leads to a system of four linear
homogeneous equations which are known to have the nontrivial solution 1,
-fo, —fu ~fi- Discuss the resulting determinant condition. Compare Problem
4.2:1.

4.2. DEPENDENCE ON INTERNAL PARAMETERS; CROSS RATIOS

We shall show that the solution of

yf(z) = f0(z) + fl(z)y(z) + f2(z)[y(z)]2, y(z0) = y0, (4.2.1)

is a fractional linear function of y0 with coefficients which are analytic
functions of z. Compare formula (4.1.15) for the special case when
«z)-0.
In the complex plane we mark the fixed singularities of the equation,
DEPENDENCE ON INTERNAL PARAMETERS; CROSS RATIOS 107

say fi, f2, . . . , £m. These points are joined to the point at infinity by cuts
which have no finite points in common. Let T stand for the plane less the
cuts, so that T is a domain. Let z0 G T, and choose y0 arbitrarily. This
gives a unique solution y(z;z0, y0) of (4.2.1). If y0 is finite, this is a
holomorphic function of z in some disk |z - z0| < r. It can be continued
analytically along any rectifiable arc C in T. Since there are no movable
branch points, the continuation is analytic save for poles.
Let z, G C, and consider y(zu z0, y0) = yu where y, is either a definite
finite number or infinity. Thus y(Z\, z0, y0) is analytic save for poles in the
finite y0-plane, so that yi is a meromorphic function of y0. But this holds
also in the extended plane, for y (z ; z0, °°) = [v(z; z0, 0)]', where v is the
solution of (4.1.18). Now a function which is meromorphic in the
extended plane is a rational function, so that

y1 = y(z1;z0,y0) = ^i(yo). (4.2.2)

But here we can interchange the roles of z0 and zx. We consider

y(z;zuyi), (4.2.3)
i.e., the solution which takes the value yt at z = zx. This solution can be
continued analytically along C from z = z, to z = z0, and since every
solution is single valued in T we have

y(z0;z,, yi) = y0. (4.2.4)

Here y(z0;z,,yi) is a meromorphic function of y, in the extended
yrplane, i.e., a rational function of yu so that

y0 = R2(yi). (4.2.5)
Now, if a mapping u -» v of the extended w-plane onto the extended
u-plane is rational and has a rational inverse, it is one to one and is given
by a Mobius transformation

v = ™ + b ad-bc± o. (4.2.6)

Applying this to the case on hand, we get

THEOREM 4.2.1

The solution y(z; z0, y0) is the result of applying a Mobius transformation
to y0:

( y0A(z) + B(z)
y(z,z0,yo)-yoC(z) + D(z). (4.2.7)
Now a fractional linear transformation is determined by three of its
108 RICCATI'S EQUATION

transforms, uniquely up to a common factor. Let the transformation be

given by (4.2.6). Give u four complex values, uu u2, w3, w4, and let the
corresponding values of v be vu v2, v3, v4. We introduce the cross ratio

*(z„ z2, 23, 24) = rI^--T-T,

2i — 24 22 — 24 (4.2.8)
and find after a lengthy but elementary computation that

^(vu v2, v3, u4) = K(ux, u2, w3, u4). (4.2.9)

This implies that the cross ratio of four solutions of (4.2.1) is a constant.
To see this we give y0 four values, y0i, y02, y03, yo4, in (4.2.7) and denote the
corresponding solutions by yu y2, y3, y4, so that

R(yi, y2, y3, yd = $(y0i, y02, y03, yM). (4.2.10)

This means that, knowing three solutions of a Riccati equation, yu y2, y3,
we know all solutions. The general solution may be written as

x( , _ [y»U) ~ y3(2)]y2(2) - C[y2(2) - y3(2)]yi(2)

n } [y.(2) - y3(2)] - C[y2(2) - y3(2)] ' K^'ll)
where C is an arbitrary constant (theorem due to Eduard Weyr, 1875). We
note that the solution is a fractional linear function of the internal
parameter C.
There is another way of tackling this problem which is also quite
instructive. It is based on the representation of the solution of a Riccati
equation in terms of solutions of an associated linear second order DE;
see (4.1.1)-(4.1.5). The discussion will involve several concepts which
belong to the next chapter but undoubtedly are familiar to the reader from
the elementary theory of DE's. Let wx(z) and w2(z) be two linearly
independent solutions of (4.1.4). Then the general solution of this equa-
tion is

w(z)= C,wi(z)+C2w2(z), (4.2.12)

where G and C2 are arbitrary constants. This means that the general
solution of the Riccati equation is given by

1 Cw\{z)+C2w'2{z)
Hz)' Ciwl(z) + C2w2(z)'
Actually the solution does not depend on two arbitrary constants; only
their ratio, CJC2 = C, is significant. Hence

v(2) = l_Oi(z) + H>SU)

DEPENDENCE ON INTERNAL PARAMETERS; CROSS RATIOS 109

In this setting the general solution is a linear fractional transform of C. We

see that the zeros of y(z) are the zeros of

w'2(z)+Cw\(z), (4.2.14)
while the poles are the zeros of
w2(z) + Cwl(z). (4.2.15)
Since both zeros and poles of y(z) are simple, we conclude that the
functions (4.2.14) and (4.2.15) have simple zeros.
In (4.2.13) it is permissible to let C become infinite. The limit exists and
is a solution of the Riccati equation.
We have seen that the general solution of the Riccati equation is a
fractional linear transform of an internal parameter with coefficients
which are functions of z. Conversely, let A(z), B(z), C(z), D(z) be four
functions holomorphic in a domain of the z-plane. Let K be an arbitrary
parameter, and form the family of functions

, , A(z) + KB(z) /A ~
z~vK(z) = C(z) + KD(z). (4.2.16)
These functions satisfy a Riccati equation, for we have

y' = {(A'C - AC) + [{B'C + A D) - {AD' + BC')]K

+ (£'D - BD')K2}(C + DK) \
We solve (4.2.16) for K to obtain

K ~ W^Dy'
Cy-A
Furthermore,

Substitution now gives ic + DKrl= B~Dy

BC-AD'
(AD - BCfy' = (A'C - AC')(B - Dyf
+ [BC -B'C + AD' - A 'D](A - Cy)(B - Dy)
+ (B'-BD')(A-Cy)\ (4.2.17)
which clearly is a Riccati equation.

EXERCISE 4.2
1 . In Problem 4. 1 :8 it was found that a certain determinant had to be zero. Show
that this determinant is the numerator in the derivative of the cross ratio
I£(yi, v2, y3, y4) and hence that the cross ratio is a constant.
110 RICC ATI'S EQUATION

2. The symmetric group on four letters has 24 elements. If the group operates on
the cross ratio of these four, how many different cross ratios are obtained?
3. For what values of C in (4.2.11) are the initial solutions y,, y2, y3 obtained?
Verify the formula.
4. Verify that the only rational functions with a rational inverse are the
fractional linear functions.
5. The Schwarzian derivative

{w,z} = — T—= — I
is a differential invariant of the group of fractional linear transformations
acting on w, i.e.,

{W,z} = {w,z} cw + a
if W=aw*b„ ad-bc*0,
where a, b, c, d are constants. Verify.
Show that the third order nonlinear DE

{w,z} = P(z)
is a Riccati equation for the logarithmic derivative of w'(z).

4.3. SOME GEOMETRIC APPLICATIONS

Cross ratios are a fundamental concept of projective geometry. Let /, and

l2 be two lines in R3, the points on U being given by a parameter 5, and those
on l2 by a parameter t. Let A, B, C, D be any four constants with
AD-BC* 0. The mapping
As +B
(4.3.1)
Cs+D

defines a one-to-one correspondence between the points of the two lines,

and this correspondence is the most general projective relation. Such a
relation is characterized by the fact that the cross ratio of four points on lx
is the same as the cross ratio of the four corresponding points on /2. Thus,
if the points on /, are given by the values 5,, s2, s3, s4 of the parameter s
and if the corresponding points on /2 are given by t = tu t2, t3, t4, then
^(su s2, s„ s4) = f2, ti, U). (4.3.2)
For the Riccati equation (4.2.1) we have seen that the cross ratio of four
solutions, y, yi, y2, y3, is a constant:

R(y, y„ y2, y3) = IKy0, y01, y02, y03) = K, (4.3.3)

where y0, Voi, V02? V03 are the initial values of the solutions at some point z0,
SOME GEOMETRIC APPLICATIONS

not a fixed singularity. By (4.2.11) this gives

v = (yi-y3)y2-K(y2-y3)y>
(yl-y3)-K(y2-y3) ' K*'M}
y
and K is an arbitrary constant. We note that the general solution y(x) is a
rational function — more precisely, a fractional linear function of each of
three arbitrarily chosen particular solutions.
Thus we see that Riccati's equation leads to projectivities. Conversely,
various problems in differential geometry lead to Riccati equations. We
shall discuss two such problems.

The Formulas of Frenet

Let T be a rectifiable curve in 3-space, 5 its arc length, R = R(s) the

radius of curvature, T= T(s) the radius of torsion. Let X, Y, Z be
running coordinates. Let the orientation of the axes be such that an
ordinary right-hand screw placed along the Z-axis will advance in the
positive Z-direction when rotated from the positive X-direction to the
positive Y-direction. Let x,(s), x2(s), x3(s) be the moving trihedral of unit
vectors taken along the tangent, the principal normal, and the binormal of
T at s. The trihedral is supposed to be oriented in the same manner as the
coordinate axes. Let V be given by the vector equation

x = g(5). (4.3.5)

Then the tangent vector is Xi(s) = g'{s), which is of length 1. Furthermore,

g"(s) is the curvature vector. It is not normalized (i.e., of length 1), but
Xiis) = g"(s)||g"(s)||~' is the principal normal. The binormal is a unit vector
orthogonal to the tangent and the principal normal and properly oriented.
The formulas of Frenet we can interpret as simultaneous first order
vector DE's satisfied by the moving trihedral. They are named after Jean
Frederic Frenet (1816-1900), who discovered them (in component form)
in 1847. They were rediscovered by Alfred Serret (1819-1885) in 1850.
The vector equations are as follows:

xM~R(sY X2(S)- R(s)+T(sY 3()" T(sY

Let the components of the vector x,(s) be xik(s), and introduce the
unitary matrix
Xn(s) Xn(s) Xn(s)
°U(s) = x2i(s) x22(s) x23(s) (4.3.7)
_x3l(s) x32(s) x33(s)
112 RICCATI'S EQUATION

We can multiply (4.3.6) on the left or on the right by an arbitrary matrix,

and the resulting vector equations are valid as long as the original
equations are. In particular, we can get the system of three scalar DE's for
the first column of (4.3.7):

= u,
x"(5)=*(77' X2M-~T^)+Tm^ *3l(5)~ T(sy (4'3-8)
Since the sum of the squares of the elements in a column of °U(s) equals
one, we can set
JCn + «21 l+x3t
1 — X31 X\i — IX2\
(4.3.9)
jCn - ix2\ _ 1 + x3i _ _ J_
1-X3i " JCn +IX21 V
From these relations we get

1 - UV . 1 + UV U + V //ii 1 rv\
jcn =u - v, x2i = 1u - v , x3i =u - v • (4.3.10)

If these expressions are substituted in (4.3.8), we obtain the DE

du
ds 2LR(s) T(s)j 2 lR(s) T(s)]

and the same Riccati equation is satisfied by v. Hence u and v are distinct
solutions of the same Riccati equation.
Now if a solution u(s) of (4.3.11) has been found, then

v(s) = -[u(s)V (4.3.12)

is found to satisfy (4.3.11). Here the bar indicates, as usual, the complex
conjugate. Knowing u(s) and v(s), we can compute the first column of
°lt(s). The same DE is found if we apply this procedure to the second or
the third column.
This discussion implies that a rectifiable space curve is completely
determined by initial conditions: (i) for s = s0 the point (jc0, Vo, z0) is on the
curve, and at this point the moving trihedral is given; (ii) the radii of
curvature and of torsion are given functions of the arc length s ; (3) R(s)
and T(s) are differentiate functions of s with bounded derivatives. Then
the existence theorems of Chapter 2 applied to (4.3.11) show that the
matrix °U(5) is uniquely determined and hence also the curve itself. This
observation and this method date back to the Norwegian mathematician
Marius Sophus Lie (1842-1899) in 1882.
SOME GEOMETRIC APPLICATIONS 113

Relations of the form

R = Fl(s), T = F2(s) (4.3.13)

are known as the natural equations of a curve.
The problem of determining a surface in R3 from a knowledge of its
fundamental forms,

E{duf + 2Fdu dv + G(dv)29 (4.3.14)

L (du f + 2mdudv +N (dv )2,
also leads ultimately to the integration of an ordinary Riccati equation.
The formulas are too complicated to be given here.

Asymptotic Lines of a Ruled Surface

As a second differential geometric application of Riccati's equation we

take the problem of finding the asymptotic lines on a ruled surface S in R3.
The surface is given by a vector equation

x = G(w,t>) (4.3.15)
in terms of two parameters, u and v. In the case of a ruled surface this
takes the form

x = G(m) + uH(w). (4.3.16)

Here the system u = C is the family of the rectilinear generators (the
"rulers"). The asymptotic lines are the curves which annihilate the second
fundamental form. In the present case N is identically zero, M is a
function of u alone, and L is quadratic in v. The form reduces to

du (Ldu+2Mdv) = 0.
Setting du = 0, we get the generators u = C as one set of asymptotic
lines, while the second set satisfies the Riccati equation

du = A+Bv
^r + Cv2, (4.3.17)
where the coefficients are functions of u alone. We may assume the
coefficients to be continuous. In this case the right-hand side satisfies a
Lipschitz condition with respect to v, and through each point (w0, v0)
passes a unique solution of (4.3.17). If vu i>2, v3, v4 are four solutions of
(4.3.17), their cross ratio is a constant. This says that, if the four solutions
are real curves, they cut each rectilinear generator in four points having a
constant cross ratio. In turn, this says that the rectilinear generators are
cut by the second family of asymptotic lines in projective point sets.
 114 RICCATI'S EQUATION

In the case of a hyperboloid of one sheet the second set of asymptotic

lines is also a family of straight lines, and it is a known fact that the two
sets of rectilinear generators cut each other in projective point sets.
We refrain from further geometric applications of Riccati's equation.
EXERCISE 4.3

1 . Two vectors in R3 are orthogonal if their dot product x • y = 2? Jtjy, = 0. Find

the components of the binormal vector, knowing that it is orthogonal to the
tangent and principal normal vectors.
-j

4.4. ABSTRACT OF THE NEVANLINNA THEORY, I

In the 1920's the brothers Frithiof and Rolf Nevanlinna, pupils of Ernst
Lindelof, developed an extensive theory of the value distribution of
^entire and meromorphic functions, aimed at sharpening the theorem of
Picard. In 1933 the Japanese mathematician Kosaku Yosida gave an
elegant application of this theory to nonlinear DE's by presenting a new
proof and extensions of a striking theorem by Johannes Malmquist
(1882-1952) dating from 1913. A systematic study of the implications of
the Nevanlinna theory for DE's was undertaken by Hans Wittich, starting
in 1950. We shall give some of the results of Wittich and Yosida in Section
4.6 and in later chapters. It is clear that such a program requires at least a
brief review of the salient points in the value distribution theory. The
basic concepts and formulas are given in the present section; the
proximity function of the logarithmic derivative, the second fundamental
theorem, and the defect relations, in the next section.
It is customary to start with the theorem of Jensen [J. L. V. W. Jensen,
(1859-1925), Danish mathematician and telephone engineer]. We shall
however, go further back and begin with a modified form of Cauchy's
integral which was a useful tool in the Hadamard theory of entire
functions of finite order.
Let 2 G(z) be a function holomorphic in the disk \z | =s= R < oc, and set
G(z) = U(z) + i V(z), where U and V are real potential functions. We
have then
(4.4.1)

(4.4.2)

Here the path of integration is the circle \t\ = R described in the positive
sense, and the bar indicates the complex conjugate. The proof is left to
the reader.
 115
NEVANLINNA THEORY, I

Now let a be a complex number, 0<\a\<R, and form the Mobius

function :

Q(z;a,R) = R£-£. (4.4.3)

It is holomorphic in the extended plane except for a simple pole at
z = R2d l; it maps the whole plane into itself in a one-to-one conformal
manner. Here z = a goes into w = 0, z = R2d 1 into w = <x>. The circle
\z \ = R goes into the unit circle \w\ = 1; the interior (exterior) of the first
circle goes into the interior (exterior) of the second circle.
Suppose now that z »-»/(z) is meromorphic in |z| ^ R and has neither
zeros nor poles at z = 0 or on the circle |z| = R. Let the zeros be
au a2, . . . , a/ with \aa\ = ra, (4.4.4)
and the poles
bub2,...,bk with \bp\ = Pp. (4.4.5)
Here multiple zeros or poles are counted as often as indicated by their
multiplicity. We now construct a function z H(z) which has neither
zeros nor poles in |z| ^ R and for which |//(z)| = |/(z)| on |z| = R. This we
can do by multiplying /(z) by a product of Q-i unctions and their
reciprocals so as to eliminate the zeros and poles of / without affecting
the absolute value on |z| = R. We take

H(z) = f(z) a=l

n [Q(z-aa,R)]1 /3fl
= 1 Q(z;b^R\ (4.4.6)
which clearly has all the desired properties.
Next let
L(z) = log H(z). (4.4.7)
This function is holomorphic in |z| ^ R once we have chosen the value of
the logarithm at some particular point. Since H(z) # 0, <» in the disk, its
logarithm is single valued and hence holomorphic. We now use (4.4.1)
with G(z) = L(z). Here U(z) = log \H(z)l V(z) = arg H(z), and \H(z)\ =
|/(2)| on \z\ = R. Hence

log H(z) = - log i/(0) + (f>lQf dr, (4.4.8)

TTl J t — Z
and by (4.4.6)

TTl J t- Z
+ a=l
2 log Q(z ;aa,R)-j?
0 = 1 log <?(z ; fc„ /?)• (4.4.9)
116 RICCATPS EQUATION

The logarithms in the left member and in the sums have not been defined.
For the time being, this information is immaterial and any ambiguity will
disappear when derivatives are taken with respect to z. At this juncture
we set z = 0 and obtain

log /(0) = - logiJ(O) + ^ J log \f(R e 16 )\ dS

Here we equate the real 4M-

parts fi'
and get-iM
Jensen's
-bi formula,
)

Z7T Jo a = l 'a /3 = 1 P(3

j- P"log|/(KO| d6 = log |/(0)| + £ logf - £ log£. (4.4.10)
Now we are ready for Rolf Nevanlinna. He got his first fundamental
theorem by the following argument. For t real, set

t+ = maxa,0). (4.4.11)
Then

log |/| = logJ |/| - log^jjj

and

|o* log \f(r eid )| dd = m (r, oo; /) - m (r, 0; /). (4.4. 1 2)

The expressions in the right member are the proximity functions of / with
respect to the values oo and 0. If a ^ oo, we set

while
Z7T Jo
m(r,oo;/)= JL r \0g+ \f{reie)\d6. (4.4.14)

To cope with the sums in the right member of (4.4.10) we introduce

enumerative functions. Let n(s, a ; /) be the number of times that f(z) = a
for \z\^s, multiple roots counted with the proper multiplicity. If /(0) ^ a,
we set

N(r,a;f)= Jo
[ n(s,a \f)—S . (4.4.15)

This can also be written as a Riemann-Stieltjes integral,

log {</sn(s, a;/), (4.4.16)

o S
NEVANLINNA THEORY, I 117

and this observation shows that the difference of the two sums in (4.4.10)
equals
N(R,°o;f)-N(R,0;f).
This gives

[m (R, 0; /) + N(R, 0; /)] - [m (R, oo; /) + N(R, oo; /)] = log |/(0)|. (4.4. 17)
If /(z) is meromorphic in the finite plane, R may be taken arbitrarily large.
The expression

T(r;/) = m(r,oo;/) + N(r,o°;/) (4.4.18)

is then defined for all r > 0. It is the characteristic function of / in the
sense of Nevanlinna. It measures the affinity of / for the value oo, for
N(r, oo;/) measures the frequency of the poles and m(r, <»;/) gets its
contribution from the arcs of \z\ = r, where |/(z)| is large. Now (4.4.17)
states that

r(r;i)-r(r;/) = log|/(0)|. (4.4.19)

In other words the affinity of / for the value 0 differs from its affinity for
the value oo by a constant. Replacing / by / - a, a finite, we see that the
affinity for all values is essentially the same:

T(r;f-a)=T(r;f)+0 (log r). (4.4.20)

For fixed a, the deviation is normally O(l), but the larger 0(log r) is
required if /(0) = a.
Similar formulas hold if /(z) is single valued and meromorphic for
0<ro^|z|<oo and is different from zero and infinity on |z| = r0. The
proximity functions are unchanged but with r0 < r. In the enumerative
functions the lower limit of integration is r0 instead of zero. T(r;f) is still
defined by (4.4.18), and (4.4.20) holds.
These various relations constitute Nevanlinna's first fundamental
theorem.
Among the various inequalities satisfied by T(r;f) we list the
following:

T(r; .../.)«^T(r;/i), (4.4.21)

;/(j =)l
) <Sr(r m. )
r; 2 / i + log (4.4.22
\ rn
A Mobius transformation(m applied to a meromorphic function / leads to
another meromorphic function F, and the difference of T(r;/) and
T(r; F) is at most 0(log r).
118 RICCATI'S EQUATION

A meromorphic function / such that T(r;/) = 0(log r) forces / to be

rational. To see this, set

liminf^^c. (4.4.23)

Here c > 0 unless / is a constant. It is desired to prove that c is finite and

equal to a positive integer iff / is a rational function, in which case
"lim inf" becomes "lim."
To perceive this, suppose that / has a total number of m poles in the
finite plane and that /(z) ~ Az" for large values of z. Then n(r, oo; /) = m
for r > pm, so that N(r, oo; /) lies between m log r and m log r + C and
hence lim N(r, oo; /)/(log r) = m. If n >0,then m(r, oo;/) = n log r + O(l),
so that

lim Ty>fKm
log r + w> (4.4.24)

where the term n is omitted if n < 1.

Conversely, suppose that 0 < c < oo, and it is to be proved that / is a
rational function. Let a be an arbitrary complex number or infinity.
Suppose that /(0) ^ a, and consider

N(r,a;f) = Jo
[ n(s,a;f)^>
b \\ n(s9a;f)^
Jrl,k ^
where k is any real number 5*2. The last member is at least

^n(r1/fc,a;/)logr.

Here (k - \)lk^\, and r1/fc sweeps the interval (1, rm) as k goes from two
to infinity. By assumption there exists a sequence {rn}, rn t 00 , such that
TXr* ; /) < 2c log rn. Since every r is ultimately included in some interval
(1, rl!2), it follows that n(r,a ;f)< 4c for all a and all large r. The only
meromorphic function that could and does possess such a property is a
rational function.
In addition to the Nevanlinna characteristic function we have to
consider the spherical characteristic T°(r;f), introduced in 1929 by
Tatsujiro Shimizu in Japan and Lars Ahlfors in Finland independently of
each other. Consider the mapping defined by z i->/(z) of the disk |z| ^ s
on the w-plane; it is the germ of a Riemann surface Rs which covers parts
of the w-plane several times; in fact, n(s, a ; /) measures how many times
the particular value w = a is covered. We make a stereographic projec-
tion of Rs on a sphere of radius J, tangent to the w-plane at the origin. The
area of the projection is
NEVANLINNA THEORY, I 119

where d<o is the surface element in the z-plane. We now define

T°(r;/)= Jo [ A(s;f)%. S (4.4.26)

This function is strictly increasing and its second derivative with respect
to log r is positive, so that T\r\f) is logarithmically convex. Here this
property is evident; the Nevanlinna characteristic has the same property,
but in that case the proof is fairly intricate. The logarithmic convexity is
equivalent to the inequality

T\rcs *-c ; /) cT(r;f) + (1 - c)T(s ; /). (4.4.27)

It turns out that the difference between the two characteristics is a bounded
function of r, so they are interchangeable.
Order and type of meromorphic functions are defined by means of the
characteristic. Thus the order of / is

p(f) = lim sup108^/^, (4.4.28)

and if 0 < p < oo the type is

r(f) = lim sup r pT°(r;f). (4.4.29)

The type is normal if 0<r(/)<oo, minimal if r(f) = 0, maximal if
t(J) = +oo. in the case of entire functions this notion of order agrees with
the classical definition in terms of the maximum modulus,

p(/) = limsup!ogMM£ii). (4430)

The notion of type is also the same (i.e., normal, minimal, maximal), but in
the case of normal type the numerical value will depend on which
definition is used.
For entire or meromorphic functions of finite order p the three integrals

J s'adsn(s,a;f), J\ s~a~ln(s
l ; a; f) ds, (4.4.31)
s'-^Nis, a\f) ds
i,

are equiconvergent and do converge for any a > p. If the roots of the
equation f(z) = a form the set {zn(a)} and \zn(a)\ = rn(a), the con-
120 RICCATI'S EQUATION

vergence of the first integral implies convergence of the series

n2= l [rn(a)] a. (4.4.32)

The exponent of convergence of the series is that number cr such that the
series converges for all a > cr and diverges for all a < a, while a = cr may
yield either convergence or divergence. In the present case cr p, and it
may be shown that cr < p can hold for at most two values of a. These
convergence properties clearly imply that the series

n2= l T0[rn(a);fr (4.4.33)

has a convergence exponent not exceeding one and this holds also if / is
of infinite order. If in the integrals (4.4.31) the factor s~a is replaced by
[T°{s ;/)]"", integrals are obtained for which similar equiconvergence
theorems hold.

EXERCISE 4.4
1. It is desired to prove (4.4.1) and (4.4.2). To this end note that on the circle
|z| = JR the function z*-+G(z) is a Fourier power series. Prove that the
Fourier coefficients are

if^^^t, III
Furthermore, c0+ c0 = 2(7(0), c0- c0 = 2/V(0). Complete the proof.
2. The function z »-> cos z is entire. Determine n (r, 0; /), m (r, 0; /), m (r, oo; /),
and T(r ;f). Find order and type.
3. Solve problem 2 for the logarithmic derivative -tanz.
4. Verify the stated properties of Q(z, a, R).
5. Complete the proof of (4.4.10).
6. Prove the properties of log+ which justify (4.4.21) and (4.4.22).
7. If u F(u ) is a fractional linear transformation on u with constant
coefficients, show that T[r; F(J)] = T(r;f) + 0(log r).
8. Verify the statements made about (4.4.31) and (4.4.32). Use the function ez
to show that the exponent of convergence in (4.4.32) can be less than one for
two values of a.
9. Why does the condition n(r,a ,f)< 4c for all a and all large r imply that a
meromorphic function / is rational?
10. The generalized equiconvergence theorem mentioned at the end of the
section involves the following items:

J" [T\s ; f)]-d,n (s,a;f), J" [ T°(s ; f)] °n (s, a ; /) y ,

NEVANLINNA THEORY, II 121

Jl ■> n = l
[T\s',f)]-°N(s,a,f)^, («);/]}".
To prove equiconvergence, start with the second integral.
11. Show that the Laplacian of log [1 + |/(z)|2] equals 4|/'(z)|2[l + |/(z)|2]"2.
12. Temporarily, let us say that a value w = a is exceptional for a meromorphic
function / if

lim sup N(r, a ; /)[r°(r; /)]"' = v(a) < 1.

It is a Picard value if 77(a) = O^Show that for z *-> J02 exp (- s2) ds there are
three exceptional values, ±\\Ztt, ». Find the values of 17(a).
13. An exceptional value may be asymptotic in the sense that /(z)-» a as z -+0°
in some sector of the z-plane. Show that this is the case for the function in
Problem 12, and find the sectors. (To be an exceptional value, however, it
is neither necessary nor sufficient to be asymptotic.)

4.5. ABSTRACT OF THE NEVANLINNA THEORY, II

In this section we continue the study of the Nevanlinna theory. The main
object will be the proximity functions of the logarithmic derivative of a
meromorphic function, but the second fundamental theorem and the
defect relations will also be presented.
The starting point is (4.4.9):

log /(z ) = - to£/f(0) + -K <£ dt

7Tl J t — Z
+ a2= \
log Q(z> am R)~it
0 = 1 log Q(z, bfi, R). (4.5.1)
The logarithms in the left member and in the sums may be specified by
analytic continuation, but actually the chosen determinations are imma-
terial for our purposes since we are going to take derivatives, thus
removing the ambiguities. We may clearly differentiate the integral with
respect to z under the integral sign and obtain

m_Rr\og]f{R^

/(z) ttJo (t-zf e d(p (4.5.2)

+ y Mi a- \ t(— — + g' )
l\z-aa R2-aaz) , \z - bp R2-bpz/

Here z = re'6 is different from all zeros and poles of z ^->/(z) in the disk,
and r < R. We aim to estimate this expression with the ultimate view of
obtaining an upper bound for the proximity function m(r, 00; /'//). Let d
be the distance of z from the nearest zero or pole, so that |z - aa\^ d,
122 RICCATI'S EQUATION

d. Then
Rz-az

has its greatest value when- aaz

z is real positive, so that
R
R - rR R

The total number of terms in the sums is n (R, 0; /) + n (R, <*>;/), so that the
sums contribute at most

to the estimate. The integral gives at most

21?
-2[m(R,0;f) + m(R,co;f)].
(R-r)
All told, we get
/'(*) 2R

f(z) r^ [m (J*, 0;/) + m (*,»;/)]

+ (j + j^ry) [«* 0; /) + n (JR, oo; /)]. (4.5.3)

Let 8 be any small positive number, d ~ [r°(r;/)] 1 8, and R = r + d.

We have then

^ 2(r + d)d 2[m (r + d, 0; /) + m (r + d, 00; /)]

/(*)
+ 2<T'[n(r + J, 0; /) 4- n(r + d, 00; /)]. (4.5.4)

Here z = re'6, and values of r must be excluded around each pn and rn.
Since d = d(r) may be chosen as a decreasing function of r, the excluded
values may be adjusted so that the total length of the excluded intervals is
dominated by the sum of the convergent series,

(4.5.5)
2 2{[T°(Pn\f)]-l-s+[T°(rn;f)r-s}.
So far, we have been able to keep the discussion perfectly general, but
now we shall assume that z i->/(z) is a meromorphic function of finite
order p. Let a be any number >p. Then

m(r + d,0,f) + m(r + d,oo-f) ^ 2T°(r + d ; /) + 0(log r)

s£ 2(r + </)a + 0(log r) < 3r" (4.5.6)
NEVANLINNA THEORY, II 123

for sufficiently large values of r. This crude estimate is also valid for
[n (r + d, 0; /) + n (r + d, °°; /)], as may be concluded from the convergence
of the first integral under (4.4.31). Assuming r + d <2r and combining,
we get an upper bound
|2f.1+(3+25)a _|_ for(2+S)<* jg^ l+(3+2S)a
This gives

THEOREM 4.5.1

If f(z) is a meromorphic function

f'(z) of finite order p and if y is any
number y > 1 + 3p, then
f(z)
\S\z\y (4.5.7)

for |z| outside of intervals of finite total length.

This estimate is probably very poor, but it is good enough for our main
purpose, which is stated by

THEOREM 4.5.2

Under the assumptions of Theorem 4.5.1

m
(r,oo;^ = 0(/ogr) (4.5.8)
for r outside the exceptional intervals which are of finite total length.
As a matter of fact, all values are admissible when p(f) is finite. For a
function of infinite order we have instead

m
(r,oo; Dj = oflog r) + 0[log T°(nf)]9 (4.5.9)
but now for r outside a set of finite logarithmic measure [i.e., a set E such
that /E (drfr) < oo]. We shall not need these refinements, but we do need

THEOREM 4.5.3

// f is of finite order and if m(r, oo; f) = 0(log r), then m(r, oo; f ) = 0(/og r).
This is so because

m (r, oo; /') = m(V, oo; L fj ^ m^ oo; L^j+ m (r, oo; /) + log 2 = 0(log r).
We have now all the material that is needed for Wittich's proof of the
124 RICCATPS EQUATION

Malmquist-Yosida theorem. But we have also what is required for a proof

of the second fundamental theorem and the defect relations for the case
in which p(f) < oo. Here we can get along without putting restrictions on
fw(r, °°; /)• The following argument is based on Wittich (1955, pp. 17-18).
It is a question of estimating T°(r;f) from above and from below when
/ is a meromorphic function of finite order. We introduce a new set of
enumerative functions,

N,(r,a;f)= Jo
f a ;f)^,
S f(0)*a, (4.5.10)
where the enumeration extends over the multiple roots of the equation
f(z) = a, but a Mold root is counted only k - 1 times. Hence
AT(r, oo; /') = 2N(r, oo; /) - N,(r, oo; /) ^ 2AT(r, oo; /). (4.5. 1 1)
Thus, omitting reference to infinity when no ambiguity is likely, we
have

T°(r;n = N{r;f') + m(r;n = 2N(r;f)-N{r;f) + m(r;n

= 2N(r;/)-N,(r;/) + m(r;^/)

^2N(r;/)-N1(r;/) + m(r;^) + m(r;/) + log2

^ T\r\f) + N(r;f) - N,(r; /) + 0(log r)
and finally

T\r\f')^2T\r',f). (4.5.12)
The estimate from below is harder to get but yields a number of
valuable results. Let au a2, . . . , aq be q (> 1) complex numbers and form
the polynomial

P(w) = fl(w-aj).
i
Here we replace w by /(z) and get a composite meromorphic function
H{z) = P[f(z)]. We have

pfe=l,A- (4-5-13)
Since m(r; ^ m(r; 1//') + m(r;f'lH), it is seen that T°(r;f')^
N(r; l//') + m(r; - m(r; /7tf) + 0(1). By (4.5.13)

m(n £) = m(r; | ^) = | m(r; ^) + O(l) = 0(log r)

NEVANLINNA THEORY, II 125

by Theorem 4.5.2. The first fundamental theorem gives

m(r;if) = N(r ; H) -N(r'>jj) + m(r'>H>)+ 0(\).

Furthermore, we have

m(r;H)= qm(r;f)+ O(l),

so that

m(r;jj) = qN(r;f) - t N(r, a> ;f) + qm (r ; /) +0(1).

It is now seen that

T°(r;f')^ Af(r;1) + qN(r ; /) - £ AT(r, a, ; /) + qm (r ; /) - O (log r)

= N(r;J) + 1 m (r, a, ; /) + O (log r),

or

w(r;1) + 2 m (r, a, ; /) - 0(log r) ^ T°(r ; /' )

*s N(r ; /') + m (r ; /) + 0(log r). (4.5. 14)
We can now dispense with the services of T°(r;f) and are then left
with the inequality

2 m (r, a, ; /) ^ m (r, oo; /) + N(r ; /' ) - iv(r;1) + 0(log r). (4.5. 15)
Here we add 2 N(r, a, ; /) on both sides to obtain

qT°(r ; /) ^ S N(r, a, ; /) + m (r, »; /) + 2N(r, «>; /)

/=i
-N,(r,<»;/)-w(r,ao;I)
or

(<j - 2)Hr ; /) « 21 N(r, a, ; /) - AT,(r, »; /) - m(r, «; /)

-Ar(r,oo;i) + oaogr)
and finally

(<? - 2) TV ; /) =s S1 N(r, a, ; /) - N,(r, »; /) + 0(log r). (4.5.16)

This inequality is the second fundamental theorem of R. Nevanlinna.

126 RICCATPS IQUATION

Here q is any integer > L The case q = 2 is obviously trivial. The a,\ are
any distinct complex numbers, infinity being permitted. An alternative
form of the inequality is

£i m(r, a, ; /) < 2T\r ; /) - N,(r, oo; /) + Oflog r). (4.5. 17)

From (4.5.17) we get

X lim inf 7^.'^ + lim mf r>(r.^ ^ 2. (4.5. 18)

Set
,. . rra(r, «,-;/) . .• N(r,a.;f) e/ . c irY.
lim mf T^/f}- 1 -I'm sup 7^/^ = (4.5.19)

(r oo- /*) =<*>•

lim inf N.r»('r.fl (4.5.20)
These quantities are called the de/ec* of a, and the fota/ algebraic
ramification of /, respectively. "Ramification" refers to the Ricmann
surface defined by the mapping /(z) = w. Formula (4.5.20) then gives the
defect relation,

2 S(«,) + <t>^2, (4.5.21)

for any finite value of q. This implies that fi(a ) 0 for at most a countable
set of values of «. Although meromorphic functions with any prescribed
finite number of defective values are known, no example with an infinite
number is known as yet.
It is clear from (4.5. 19) that the maximal defect that any value can have
is one, and (4.5.21) shows that there can be at most two values of maximal
defectnumber
this one. This
can isbe Picard's
reached.theorem^ The function z »-» tan z shows that
One can also introduce a local ramification index, i)(a):

rJ(a) = lim inf ^V/f' (4'5'22)

This index measures the frequency of algebraic branch points located
over w = a on the Riemann surface. The value a is completely ramified if
the equation /(z) = a has no simple roots. This means that the roots are at
least double roots and the corresponding ramification index * Since the
sum of the ramification indices is at most two, it follows that a meromor-
phic function can have at most four completely ramified values. There are
functions having four completely ramified values: the p-function of
Weierstrass is perhaps the simplest and the best known (see Problem
NEVANLINNA THEORY, II 127

3.3:7). This is the solution of the DE

[w']2 = 4(w -e{)(w -e2)(w - e3), (4.5.23)

where e, t e? I c, 0 and the solution which has a pole at the origin is
chosen. Here the four values eu c, et, and * are completely ramified.
The theorem on complete ramification has an important hearing on
algebraic geometry and uniformization. Here a classical theorem due to
Picard (1887) asserts that an algebraic plane curve does not admit of a
parametric representation in terms of meromorphic functions if its genus
{ deficiency) is p>\. Equivalcntly, an algebraic relation F(z, w) = 0
cannot be uniformized globally by functions meromorphic in the finite
plane if the genus exceeds one. The genus of an algebraic curve is a
positive integer or zero invariant under birational transformations of the
coordinates. An algebraic curve of order N can have at most .'(N - l)x
(N - 2) singularities (nodes and cusps). If there are actually // nodes and c
ordinary cusps, the deficiency is p \(N \)(N 2) n c, and this is
the genus.
For the curves

y2=af[(x
; I a,) and y2=bf\(x
k I - bk), (4.5.24)

where the ay's and bk's are distinct, we have \(N \)(N 2) equal to one
and three, respectively, and // t c is zero in the first case and two in the
second, so that p = I in both cases. If two of the numbers at or two of the
bk should coincide, then p = 0 and parametric representations in terms of
rational functions of / are available. If the a,\ are distinct, a preliminary
afline transformation of the coordinates will reduce the equation to the
form
v2 = 4(u -ex){u -e2)(u e3) (4.5.25)
with ex 4- e2+ 03 = 0, and we can set u = p(t), v = p'(t ).
For the second equation in (4.5.24) the Jacobi normal form,

v2 = (\ u?)(\ -fcV), (4.5.26)

may be more appropriate. Here a fractional linear transformation can be
used to reduce the equation in question to the form of (4.5.26), where k is
an essential parameter depending on the original configuration. After the
reduction a parametric representation u = sn(/), v =en(/)dn(0 in terms
of the Jacobi functions is available. Thus both equations under (4.5.24)
admit of parametric representations by functions meromorphic in the
finite plane.
So much for the elliptic case p = 1. For the hyper elliptic case we have
128 RICCATI'S EQUATION

two alternative normal forms,

2p+l 2P+2
y^aUix-a;)
j= l and y2 = b k\\
= \ (x - bk). (4.5.27)
These curves are of genus p, and any algebraic curve of genus p can be
transformed into one or the other of the normal forms by a birational
transformation, i.e.,

x =Ri(u,v), y=R2(u,v), u=R3(x,y), v=R4(x,y), (4.5.28)

where the jR's are rational functions of the indicated variables. Now, if the
coordinates jc, y of the first curve under (4.5.27) could be represented
globally by, let us say, jc = y = f2(t), where /, and f2 are meromor-
phic functions of t in the finite plane, then

[/2(0]2 = 2fi,[/i(0-a/L (4.5.29)

and the values au a2, . . . , a2p+u <*> must be completely ramified for /,(0 in
order to make f2(t) single valued. But if p > 1, we cannot have 2p +2
completely ramified values. This proves Picard's theorem. The curves in
question may be uniformized; however, x = A,(0, y = A2(t), where A,
and A2 are automorphic functions which are meromorphic but have only
a restricted domain of existence with a natural boundary.
We return briefly to the main theme of this section. We plan to apply the
formulas to functions which are single valued and meromorphic in a
neighborhood of infinity, say, for |z|^# >0. The main difference be-
tween this case and that of R = 0 resides in the definition of the
enumerative and proximity functions, the area function A (s;f) and the
characteristics where now the lower limit of integration is R instead of
zero. We can justify this extension by noting that f(z) may be split into
the sum of two functions /(z) = fx(z) + /2(z), where /, is holomorphic for
|z |^ R and zero at infinity, while f2 is meromorphic in the finite plane. Our
previous considerations apply unchanged to /2, and in forming the
relevant integrals for / with |z| > R the contributions coming from /, are
at most 0(log r).

EXERCISE 4.5
1. Justify (4.5.5).
2. Justify Theorem 4.4.1.
3. How should (4.5.10) be modified if the equation f{z) = a has a multiple root
at z = 0?
THE MALMQUIST THEOREM GENERALIZATIONS 129

4. If / = tan z, find the limit of the quotient of N(r, °o; /') and N(r, o°; /) as r -> <».
Compare (4.5.11).
5. Does z h-» cos z have any defective values? Are there any completely
ramified values? Compute $ of (4.5.20).
6. Consider z •-» tan z for defective values. Are there any ramified values?
7. Estimate m (r, a ; /) for / = tan z. The values + i and - / require special
treatment.
8. For an entire function, T(r; f) < log M(r; /) + O(l). Why? What happens to
this inequality if / is meromorphic and becomes infinite as r \ R and |z| = R
contains one or more poles?
9. Show that the mean square modulus

is piecewise logarithmically convex and has at most one minimum for

pn < r < pn+). [Hint: Use the Laurent series for /(z) in the annulus under
consideration.]
10. Could the DE (w')2= aTXfcWw - a,) for p > 1 be satisfied by a function
meromorphic in some domain?
11. Given the curve y2 = a(x - a,)(x - a2)(x - a3), how should the constants c, d,
e be chosen so that the transformation x = cs + d, y = et take the curve into
the normal form (4.5.25)?
12. Find a change of variables which takes the second equation in (4.5.24) into
the Jacobi normal form (4.5.26).
13. Determine the genus of the ellipse, and find a parametric representation by (i)
rational functions, and (ii) trigonometric functions.
14. Solve Problem 13 for the hyperbola.
15. The cubic y2 = a(x - b)2(x - c), c ^ b, has a double point at x = b. Find a
rational representation by laying a pencil of straight lines y = t(x-b)
through the double point. ^-^^

4.6. THE MALMQUIST THEOREM AND SOME GENERALIZATIONS

The Malmquist theorem of 1913 may be formulated as

THEOREM 4.6.1

Let R(z, w) be a rational function of z and w. Suppose that the DE

w' = R(z, w)
(4.6.1)
has a solution z w(z) which is a transcendental meromorphic function
with infinitely many poles. Then R(z, w) is a polynomial in w of degree 2,
130 RICCATI'S EQUATION

i.e., the DE is a Riccati equation

w'(z) = Ro(z) + R,(z)w(z) + R2(z)[w(z)]2, (4.6.2)

where the R's are rational functions of z.
Proof. In the main we shall follow Wittich's proof (1954) except for the
trivial but useful Lemma 4.6.1, which may possibly be new in this
connection and appears to be essential for the generalizations.
As in Chapter 3 let

where

P{z,w) = '2Pl(z)w\ (4.6.4)

7=0
Q(z,w)=~2<c=0 Qk{z)w\ (4.6.5)

where the P,'s and the Qk's are polynomials in z.

We have assumed that the postulated solution has infinitely many poles.
The case in which there is only a finite number would lead to the
conclusion that (4.6.2) is actually linear. This was shown by C. C. Yang
(1971) using Yosida's results.
Now the existence of a single distant pole, distant in the sense that
Pp(z)Qq(z) t^O, implies that
p = q + 2 (4.6.6)
and that the pole is simple. For suppose that z = z0 is a distant pole of
w(z) of multiplicity a and leading coefficient c0. Then

R(z, w) = crq(z - Zo)(q-p)aV + 0(DL (4.6.7)

while
w'(z) = -ac0(z - z0)~a_1[l + o(\)]. (4.6.8)
The DE requires that
(p-q-\)a = \. (4.6.9)
Since a is a positive integer, we must have a = 1 and (4.6.6) holds.
We can then reduce the DE to the form

w' = a2(z)w2 + ax{z)w + a0(z) + (4.6.10)

by dividing P(z, w) by Q(z, w). The a's are rational functions of z, and the
degree of P,(z, w) with respect to w is at most q-l. Here a further
THE MALMQUIST THEOREM GENERALIZATIONS 131

simplification can be made. We may assume that a2(z) = 1, for if this is

not the case at the outset, the substitution v (z) = a2(z)w(z) leads to a DE
for v of the same type as (4.6.10), where now the coefficient of v 2 is
identically one. If this simplification has already been made, we are
dealing with the DE

w> = w2 + ax{z)w + a«{z) + ^^y (4.6.11)

We shall now study the solution w(z) in the neighborhood of a pole.
Write

w w w w Q
Our first objective is to find conditions on z and w(z) so that the right
member of this equation is of the form 1 + h(z), where \h(z)\ < \. Now the
coefficients are rational functions which behave as integral powers of z
for large values of |z|. Let R be a large positive number, and consider the
set

S = [z;\z\>R,\w(z)\>\z\e], (4.6.13)
where g will be determined in a moment. Now for z G 5 we can find
numbers a, b, c such that
a2(z)
w(z) [w(z)Y
<lzla <\zr2g,

\[w(z)]2 Q(z, w)\ 11

Now choose g subject to the two conditions
g >max[a,Jfc,k], (4.6.15)
max [Ra-*9Rb-28,Rc 3g] < I (4.6.16)
The principle of the maximum shows that, with this choice of g and for
z G 5, each of the three quotients in (4.6.14) is less than g. Thus for z G 5

-l + /i(z) with|/i(z)|<i (4.6.17)

If now Zy and z2 are in S and can be joined by a line segment also in 5,

then

vv(zi) w(:}—= (Z2[l + h(s)]ds,

(4.6.18)
where the absolute value of the integral lies between { and § of \zx - z2|. In
particular, if z = z0 is a pole of w(z) and if we take z, = z0 and let z be any
132 RICCATTS EQUATION

point in S "visible" from z0, then

(4.6.19)
i\z-zor<\w(z)\<2\z-Zo\-\
It is clear that every pole of w(z) belongs to 5 and also some
neighborhoods of the poles. It is necessary to subject S to a closer
scrutiny. It is an open, unbounded set. It contains a countable number of
maximal components, S, say, i.e., open connected subsets not contained
in any larger open connected subset. Each Sj contains at least one pole,
and if S, is convex, so that every point of Sj is visible from any other point
in Si, then Sj contains one and only one pole. The same conclusion holds if
any two points of Sj may be joined by a path in Sj of length less than twice
the distance between the points, for then the right-hand side of (4.6.18)
cannot be zero.
We now define a polar neighborhood for the point zn as the set
(4.6.20)
U(zn) = {z;\z-zn\<±\zn\-g}.
The problem is to show that U(zn)CS. This requires some delicate
considerations. If all the points of the circle \z - zn | = \zn \~g satisfy
\w(z)\ > |z|*, then by the theorem of the maximum the inequality also
holds inside the circle and the whole disk is in S and so does U(zn). On
the other hand, if the inequality does not hold everywhere on the circle,
we can find a concentric circle of a smaller radius passing through a point
z = t such that \w(z)\ > |z|* in the smaller disk and \w{t)\ = \t\g. Thus the
smaller disk is a polar neighborhood of the pole zn and lies in 5. It is
required to show that it contains U(z„). Here (4.6.18) may be used, where
we take z, = zn and z2 = t. Thus

i*-zn|>§Mor=§ki-*,
and we have to show that \t - zn\ >\\zn\~8. Now \t\ ^ \z„ |+ \t - zn | <
|z„| + |z„r8 so that
\t-Zn\> i{|zn i+ \Zn r*r = ik r {i + \zn
We shall show that the last factor exceeds \ for all g > 0 and all \zn \ > e.
Now

is an increasing function of x for positive g and jc, so it takes its minimum

(1 + *-*-')-*
for e^x^^atx = e. Thus we have to show that

+ >i or 1 + e-*-1 < =(l+|),/g.

Now for 1 «sg the last member exceeds 1 + 3 and we have
for 1 ^ g.
e-«-i<3-i/«
THE MALMQUIST THEOREM GENERALIZATIONS 133

On the other hand, if 0 < g < 1 we have

\\og\>e g ,>log(l + ^ -'-*)

so the required inequality is true for all g > 0 and |z„| 2* e. It follows that
all polar neighborhoods with \zn\^e are subsets of S.
Polar neighborhoods can have no points in common, for if they did then
one pole would be "visible" from the other in 5, and this contradicts
(4.6.18).
Suppose that at a certain point z = z0 we have |w(z0)\ |z0|*+1. This
implies that z0 belongs to a polar neighborhood. By the argument given
above we may show that the disk

D={z;\z-Zo\<i\zo\'g} (4.6.21)
belongs to 5. Then for any z in D

-]—=- f [\ + h(s)]ds +— r-r.

Here the two expressions involving z are holomorphic in D, and the

integral has a simple zero at z = z0. On the rim of D

[l + h(s)] ds >{\z-Zo\ = \\Zolg,

L

while ^(zo)!"1 ^ |zo|~*_1 <i|z0|~g if |z0|>4. We may now invoke the

theorem of Rouche, which says that |w(z)|_I has the same number of
zeros in D as the integral has, i.e., just one simple zero. This in turn
implies that w(z) has a simple pole in D, proving the assertion. See
Lemma 5.6.1.
Outside S we have

\w(z)\<\z\\ (4.6.22)
i.e., outside the polar neighborhoods the solution grows at most as a
power of z. This is in agreement with a general theorem of Pierre
Boutroux (1908) for equations of type (4.6.1) with p > q + 1.
Our next task is a question of polar statistics. We have to estimate the
characteristic function T(r; w), which requires estimating n(r, oo; w) and
m(r, o°; w). The first of these functions gives the number of poles of the
solution z i-* w(z) in the disk |z| < r, while the second receives its main
contributions from those arcs of the circle |z| = r which belong to polar
neighborhoods of w(z). The whole problem reduces to a geometric
problem of counting polar neighborhoods. Since they do not intersect, the
question becomes one of closest packing: How many small disks of
diameter r g can be placed with their centers on a circle of radius r when
134 RICCATFS EQUATION

no overlapping is permitted? The answer is given by the trivial but useful

LEMMA 4.6.1

On the circle |z| = r > R there are (it most

27rr8+,[l+o(l)] (6.4.23)
poles, while in the disk |z|^r the number of poles is at most

4rv [I K>(1)]. (4.6.24)

Proof. The first estimate is obvious, since this is the solution of the
closest packing problem. The disk |z|^r presents a slightly different
problem. The diameter of a polar neighborhood with center at z = z«, with
|z0| = s is K, a decreasing function of s. This means that the smallest of
the small disks are those centered on the rim of the big disk. Thus we get
an upper bound for the packing of the polar neighborhoods by finding how
many disks with diameter r 8 can be packed into a disk of diameter 2r.
Taking the quotient of the areas, we get (4.6.24). ■ This gives

LEMMA 4.6.2

N(r,^;w)<^yr2K,2[l +o(l)|,
for this is the integral of n(s,o°; w)/s.

LEMMA 4.6.3

m(r,oo; w)=0(/og r) and T(r; w) < r '

so that w(z) is of finite order not exceeding 2g + 2.
Proof. The crux of the matter is to estimate the proximity function,

2 77 Ju
m(r,oo;w) = ^- f log* |w(rew)| do. (4.6.25)

Now, for z not belonging to a polar neighborhood, we have \w(z)\ < \z\K
so the contribution to (4.6.25) from such ares is O(logr). It is thus
obvious that major contributions to (4.6.25) can come only from the polar
neighborhoods which happen to have points in common with the circle
|z| = r, and the biggest contributions should come from polar neighbor-
hoods centered on the circle. Suppose that N(z„) is such a neighborhood.
There is a pole at z = z<, with residue - 1. Then N(z<>) subtends an angle w
of opening r * '[ 1 + o(l)] at the origin. If 6 is the angle which the bisector
of a) makes with the radius vector, then on the intersection 7 of N(z0)
 135
THK MALMQUIST THEOREM GENERALIZATIONS

with |z| = r we have

w(z)<2[r\0\\ ', (4.6.26)
so the arc y makes a contribution less than

- (g f 2)r 8 1 log r.

Since no more than 27rrK"l 1 + o(l)] polar neighborhoods can be centered

on |z| = r, we get
m(r, oo; w)<2(f> +2)logr (4.6.27)
and the lemma is proved, since T(r; w) = N(r, oo; w) + m(r, oo; w). ■
We recall that for a meromorphic function of finite order, (4.6.27)
implies
m(r,oo; w') = O(logr). (4.6.28)
Incidentally our estimate for the order, 2# +2, is about twice the correct
order, as we shall see. At this stage all we need is that the order is finite.
We are now ready for the last step in the proof of Theorem 4.6. 1 . Set

so that PQ(lZ) = FU' W(Z)] = F(Z)' (4A29)

F(z) = w'(z) - [w(z)]2 - a,(z)w(z) - a0(z). (4.6.30)

Every pole of w(z) is a zero of F(z), since P, is of lower degree in w than
Q. It is clear that F(z) is holomorphic for all values of z with |z| > K. Now
for |z|>K

m(r, oo; F)^m(r, oo; w') + 3m(r, oo; w) + 0(log r) = O(logr).

Fhe first fundamental theorem gives

N(r' °° ; i) = T(r ' F) + °(l°8 r) = m(r' a)' ^ + °(,0g r)

since N(r, oo;f) = 0. This implies

N(r,a>;^) = 0(1°8r)' (4.6.31)

From this we conclude that F cannot actually depend on w (z ), for if it did
then the infinitely many poles of w are also poles of l/F and N(r, oo; \/F)
would be of the same order of magnitude as N(r, oo; vv). If F is indepen-
dent of w and is a rational function of z alone, it can be incorporated with
a„(z). It follows that the equation is actually a Riccati DE, and the
Malmquist-Wittich-Yosida theorem is proved. ■
136 RICCATTS EQUATION

COROLLARY

// (4.6.1) is not a Riccati equation but has a meromorphic solution, then

the solution is a rational function of Z.
The following result due to Wittich is of interest in this connection.

THEOREM 4.6.2

A meromorphic solution of (4.6.1) which is of order < ' is a rational

function.
Proof. We may assume that the equation is Riccati, say,
w' = K„(z) + K,(z)w + R2(z)w\
since otherwise the assertion follows from the corollary of Theorem 4.6.1.
We may assume that R2(z)^(), for otherwise w(z) must he rational. We
may also assume that vv(z) has infinitely many poles, since otherwise
again R2(z) = 0 and w(z) is rational. With at most a finite number of
exceptions the poles are simple. Next we remove the first order term and
make the coefficient of the quadratic term equal to one. All this can be
done by setting

which gives a result of the form

u' = A(z) + u\ (4.6.33)

where z*-+ A(z) is a rational function of z. We can then find a rational
function R(z) such that v(z)= u(z)- R(z) is meromorphic in the finite
plane with infinitely many poles, all simple and of residue - I. Such a
function must be the logarithmic derivative of an entire function. We
define
y(z) = "E
U)

and find that z E(z) satisfies

E'-IRE' + (A +R2- R')E = 0. (4.6.34)
Now v(z) is of finite order (say by Lemma 4.6.3); hence E must also be of
finite order so that

E(z) = eh(z)P(z).
By a theorem due to Jacques Hadamard h(z) is a polynomial in z and
P(z) is a canonical product for the zeros of E(z), i.e., the poles of v(z)\
THE MALMQUIST THEOREM GENERALIZATIONS 137

P(z) also satisfies a second order linear DE,

P" + 2(h ' - R )P' + (A + R 2 - R ' - 2Rh ' + (h 'f + h ")P = 0. (4.6.35)
Now an entire function which satisfies a linear second order DE cannot
have an order <{, and the function cos(zl/2), which is of order \ and
satisfies zw" + \w' +\w = 0, shows that the lower limit for the order can be
reached. This is under the assumption that the coefficients are rational
functions of z. For such questions consult Section 5.4. Now N(r, oo; w) =
N(r, oo; l/p) + 0(log r), so that w(z) must be of the same order as P(z),
i.e., This proves the theorem. ■
For a Riccati equation with polynomial coefficients

w' = P()(z) + P,(z)w + P2(z)w\ (4.6.36)

H. Wittich has given a direct, elegant proof that the meromorphic
solutions are of finite order. This follows from a use of the spherical
characteristic (4.4.26) with

7rA(r;w)= jf [1^^lj|2]25^ dO, t = sei», (4.6.37)

l«l«r
and by (4.6.36)
|w'|2^|P()|2+--- + |P2|2|wr;
and for / = 0, 1, ... ,4,

so that/4(r; w)<Krk and hence alsoi2T\r\ )2<1 w) is< proved.

(K/k)rk. Thus the order
of the solution is at most /c, and d the+ kassertion It is worth while
observing that k = 2 max [deg |P, |, / = 0, 1, 2] + 2. This estimate is inde-
pendent of that of Lemma 4.6.2. Wittich has also attacked the order
problem from a different angle. The following theorem will be proved in
Section 5.4.

THEOREM 4.6.3
// the Riccati equation

w' = Ro(z) + R,(z)w + w2 (4.6.38)

has rational coefficients, if the limits

Z— »°o
Urn Ro(z)|z| a and Z-*<» R,(z)|z|"fl
Urn (4.6.39)
exist and are neither zero nor infinity, if the equation has a solution w(z)
138 RICCATTS EQUATION

which is single valued and meromorphic outside a circle |z| = R, and if the
solution has infinitely many poles, then w(z) is of finite order p[w], which
is an integral multiple of \, and

p[w] = \ + max [ftia]. (4.6.40)

Note that a and p are integers, and at least one of the inequalities
P > - \,a > -2 must hold. The reason for this restriction is that a
transformation of type (4.1.6), which leads from the Riccati equation to a
second order linear DE, must lead to an equation with an irregular-
singular point at infinity.
Theorem 4.6.2 is of course a corollary of this result.
In his 1933 paper K. Yosida proved several generalizations of Malm-
quist's theorem. We shall state some of these.
THEOREM 4.6.4

// R(z, w) is a rational function of z and w, and if the DE

(w')n = R(z,w) (4.6.41)

has a transcendental meromorphic solution, then R(z, w) must be a
polynomial in w of degree not exceeding 2n. More precisely, p =
n[l + (l/a)].
Here the test-power test shows that
a(p-q-n) = n, (4.6.42)
where, as usual, p is the degree in w of the numerator of R(z, w) while q
is the degree of the denominator. It is seen that a, the order of the poles, is
a divisor of n. Since a is at least one and at most n, one gets

n + 1 ^ p - q ^ 2n. (4.6.43)
On the other hand, Yosida derives an inequality from the Nevanlinna
theory, which in the present case takes the form 2n 2* p. He also finds
2n 35 2n + q, so that q =0.
In principle we can use the argument of Wittich to prove this result.
We shall sketch the case n =2. Before doing so, let us remark that there is
a great difference between the cases n = 1 and n > 1. A Riccati equation
with polynomial coefficients will always have transcendental meromor-
phic solutions with poles clustering at infinity. This need not be true,
however, for equations

(w')2 = 2Pj(2)w'. (4.6.44)

THE MALMQUIST THEOREM GENERALIZATIONS 139

Here the equations

(m;')2 = z*(4m;3-1) (4.6.45)
are instructive. The solution is

p(tT2"+C;0'1)' k*-2; p(logz + C;0,l), k = -2. (4.6.46)

If the solution is to be single valued, k must be a nonnegative even
integer. If k < - 2, the poles do not cluster at infinity even if k is an even
integer.
We now return to the DE (4.6.41) with n = 2. Here we have
(p - q - 2)a = 2 (4.6.47)
with two solutions in terms of integers: (i)a = l,p=q+4, and (ii) a = 2,
P =<J+3.
We consider only the first case, in which all distant poles are simple.
Proceeding as above, we get

(»V = ±^»'+7^y (4.6.48)

where the a,(z) are rational functions of z, and the degree of Pi with
respect to w is at least one unit less than that of Q. Here we assume that
a4(z) s l. If this were not the case at the outset, we could make a change
of the dependent variable by setting

v(z) = [a4(z)]m w(z). (4.6.49)

It should be noted, however, that to preserve the rational character of the
coefficients in the transformed equation, [a4(z)]m must be rational and the
result for (4.6.45) suggests that a4(z) must be the square of a polynomial.
Assuming a4(z) = 1, we can define a number g and corresponding polar
neighborhoods as in the case n = 1. Suppose that |z| > R, and set

S = [z;\w(z)\>\z\*]. (4.6.50)
For |z| > R and z E S we have
a,(z)

|C?[z, w(z)][w(z)T\ 1 1
Choose g s[ow(tzh)a]t4-'

max[aJ -(4-j)g]<0, j = 0,1,2,3 and p - 5g < 0

max[i?"r^,JR^5«]<0.1. (* ' }
140 RICCATI'S EQUATION

(4.6.53)

—2 = l + h(z) with|/i(z)|< (4.6.54)

whence i-i
(4.6.55)
!|z-zor<k(z)|<i|z-Zo|
if z0 is a pole of w(z). The positive square root corresponds
4- to a pole with
residue - 1 ; the negative, to a pole with residue + 1. Both are possible, and
the estimate is the same in the two cases.
We can define polar neighborhoods by (4.6.20), and they will be subsets
of S. Actually we could replace the factor I by a somewhat larger number,
but that fact is immaterial. The polar statistics carry over unchanged,
and we find that w(z) is of finite order, <2g + 2. The proof is then
completed as in the case n = 1. This proves the case n = 2, a = 1. The
procedure for the case a = 2 is the same. In principle the method works
for any value of n and any admissible a, so that p = n [1 + (1/a)].
Furthermore, it is assumed that the coefficient of the highest power of w
is the nth power of a polynomial.
Yosida (1933a) also considered more complicated nonlinear DE's of first
and higher orders. We refer to his paper for further results. See also a
paper by A. A. Gol'dberg (1956). Yosida's results have been extended by
Ilpo Laine (1971-1974) and Chung-Chun Yang (1972). They allow the
coefficients to be transcendental entire or meromorphic functions of
lower order than that of the solution under consideration. (The author is
indebted to Professors H. Hochstadt, I.-C. Hsu, and H. Wittich and to Dr.
Yang for numerous bibliographical references on these questions.) The
methods here used would not seem capable of coping with such a
situation, at least not in its full generality. Meromorphic solutions would
normally be of infinite order, but this difficulty can be surmounted in some
cases at least. In particular, the entire function case is amenable. If the
coefficients are actually meromorphic with infinitely many poles, there
will be infinitely many fixed singularities and the existence of meromor-
phic solutions is doubtful. See Problems 16-22 below.

EXERCISE 4.6
1. Find the value of c0 in (4.6.7) and (4.6.8).
2. The discussion in the text naturally applies to a Riccati equation. Consider
THE MALMQUIST THEOREM GENERALIZATIONS 141

w' = 1 + z2w + zw2. Remove the coefficient z of w2, and find the resulting
coefficients a0 and ax. Find safe bounds for R and g.
3. If h >0, is (1 + h)m< 1 +|h? More generally, if g is a real number, g > 1, is
(l + h)1/8 < 1 + h/g?
4. Verify the calculation that justifies defining a polar neighborhood by (4.6.20).
5. If w' = Po + Piw 4-P2w2 + P3w3, the P's being polynomials, how should the
dependent variable be transformed so that (i) the quadratic term disappears
from the equation, and (ii), in addition, the coefficient of the cubic term
becomes identically one?
6. The equation (w'f = 4w3 has rational solutions. Find them, and find a change
of variables z = cs, w = ay which leaves the equation invariant.
7. Solve Problem 6 for w" = 4w3.
8. Suppose that P0, Pi, P2 are polynomials in z. Show that the equation

w' = P0(z) + P,(z)w + P2(z)w2 - z2u>3

cannot be satisfied by a rational function unless it be a polynomial.
9. Show that the DE

w' =4 + 3z2 + z4-zw3

has a particular rational solution. Show that the general solution has movable
algebraic singularities, and determine their nature.
10. Verify the calculations that led to (4.6.27).
11. Determine A(z) in (4.6.33).
12. Why must v(z) be the logarithmic derivative of an entire function?
13. How is the estimate k ^ 2 max (deg P,) + 2 obtained in the discussion of
(4.6.37)?
14. Verify (4.6.46).
15. If the change of variable (4.6.49) is applied to (4.6.48), find the transformed
equation.
The equation
w>'=/(z)(l + u>2)
is integrable by quadratures. Discuss the nature of the singularities if /(z) equals
16. e\
17. z\k*-\.
18. z1.
19. 2(1 + z2)'.
20. sec2 z.
21. Which of the preceding equations possess solutions that are single valued in
a neighborhood of infinity and have the point at infinity as the only cluster
points of its poles?
142 RICCATPS EQUATION

22. If /(z) is an entire function, show that

T(r; w)^exp[rT2(r;f)],
using (4.6.37), for instance.

LITERATURE

Sections C.l and 12.1 of the author's LODE have a bearing on this
chapter.
Frenet's formulas and the moving trihedral are discussed in Sections
15.4 and 15.5 of another book of the author's:
Hille, E. Analysis, Vol. II, Ginn-Blaisell, Waltham, Mass., 1966.
For connections between the Riccati equation and differential geometry
see p. 186 (Vol. 1) and pp. 54 and 1196 (Vol. 3) of:
Strubecker, Karl. Differentialgeometrie. 3 vols. Walther de Gruyter, Berlin, 1964.
The Nevanlinna theory is discussed at some length in Chapter 14 of
Vol. II of the author's AFT. See also:
Hayman, W. Meromorphic Functions. Oxford University Press, 1964.
Nevanlinna, Rolf. he theoreme de Picard-Borel et la theorie des fonctions meromorphes.
Gauthier-Villars, Paris, 1929; Chelsea, New York, 1973.
Sario, Leo and Kiyoshi Noshiro, Value Distribution Theory. Van Nostrand, Princeton, N.J.,
1966.
The Malmquist theorem is discussed in:
Bieberbach, Ludwig. Theorie der gewdhnlichen Differentialgleichungen. Grundlehren, No.
86. Springer- Verlag, Berlin, 1953, pp. 86-101.
Wittich, Hans. Neuere Untersuchungen iiber eindeutige analytische Funktionen. Ergebnisse
of Math., N.S., No. 8. Springer-Verlag, Berlin, 1955, pp. 73-80; 2nd ed., 1968.
Further references for Section 4.6 are:
Boutroux, Pierre. Lecons sur les fonctions defines paries equations differentiates du premier
ordre. Gauthier-Villars, Paris, 1908, pp. 47-54.
Gol'dberg, A. A. On single-valued solutions of first-order differential equations. Ukrain.
Mat. Zurnal, 8:3 (1956), 254-261. (Russian; NRL Translation 1224, 1970.)
Hille, Einar. Finiteness of the order of meromorphic solutions of some non-linear ordinary
differential equations. Proc. Roy. Soc. Edinburgh, (A) 72:29 (1973/74), 331-336.
Laine, Ilpo. On the behaviour of the solutions of some first order differential equations. Ann.
Acad. Sci. Fenn., Ser. A, I, 497 (1971), 26 pp.
Admissible solutions of Riccati differential equations. Publ. Univ. Joensuu., Ser. B, 1
(1972).
Admissible solutions of some generalized algebraic differential equations. Publ. Univ.
Joensuu., Ser. B, 10 (1974), 6 pp.
LITERATURE 143

Malmquist, Johannes. Sur les fonctions a un nombre fini des branches definies par les
equations differentielles du premier ordre. Acta Math., 36 (1913), 297-343.
Sur les fonctions a un nombre fini de branches satisfaisantes a une equation
differentielle du premier ordre. Acta Math. 42 (1920), 433-450.
Wittich, Hans. Ganze transzendente Losunger algebraischer Differentialgleichungen. Math.
Ann., 122 (1950), 221-234.
Zur Theorie der Riccatischen Differentialgleichung. Math. Ann., 121 (1954), 433-450.
Einige Eigenschaften der Losungen von w' = a(z) + b(z)w + c(z)w2. Arch. Math., 5
(1954), 226-232.
Zur Theorie linearer Differentialgleichungen im Komplexen. Ann. Acad. Sci. Fenn.,
Ser. A, I, Math., 379 (1966), 19 pp.
Yang, Chung-Chun. On deficiencies of differential polynomials. I, II. Math. Zeit., 116 (1970),
197-204; 125 (1972), 107-112.
A note on Malmquist's theorem on first-order differential equations. NRL-MRC
Conference on Ordinary Differential Equations, edited by L. Weiss, Academic Press,
New York, 1972, pp. 597-607. Also Yokohama Math. J., 20 (1972), 115-125.
On meromorphic solutions of generalized algebraic differential equations. Annal. Mat.
pura appi, (IV) 91, (1972), 41-52.
Yosida, Kosaku. A generalization of a Malmquist's theorem. Japan. J. Math., 9 (1933),
253-256.
A note on Riccati's equation. Proc. Phys.-Math. Soc. Japan, (3) 15, (1933), 227-237.
On the characteristic function of a transcendental meromorphic solution of an
algebraic differential equation of the first order and the first degree. Proc. Phys.-Math.
Soc. Japan, (3) 15, (1933), 337-338.
On algebroid solutions of ordinary differential equations. Japan. J. Math., 10 (1934),
119-208.

ADDENDUM
The modified Weierstrass DE

(W)2 = [f'(z)]\w - ex)(w - e2)(w - e3)

is satisfied by w(z) = p[f(z); 2cou 2o>3]. The reader is urged to discuss the
nature of the singularities of w(z) and the existence of meromorphic
solutions if /(z), e.g., is a power of z. The case f'(z) = (z2 - 1)~,/2, wi = \,
w3 = tti, is treated by Stevens B. Bank and Robert P. Kaufman in a
forthcoming paper. Here the finite fixed singularities are only apparent,
w (z ) is single valued and meromorphic in the finite plane, of order zero, and
T(r ; w) = 0[log2 r]. Verify! S. Bank has many important papers on
nonlinear DE's but seems to pass by the Malmquist theorem cycle.
5

LINEAR DIFFERENTIAL

EQUATIONS: FIRST

AND SECOND ORDER

In this chapter and the next four we shall be concerned with linear DE's.
It is appropriate to start with the first and second order cases which have
been considered here and there in preceding chapters. The second order
case is the most important for the applications and exhibits most of the
phenomena that we aim to study in the nth order case. It also presents
fewer algebraic difficulties.
Our previous existence theorems apply at nonsingular points. All
singularities are fixed; they are the singularities of the coefficients and the
zeros of the coefficient of the highest derivative plus the point at infinity.
The singularities are of two types: regular and irregular. In the case of a
regular singular point, expansions valid in a full neighborhood of the point
will be given. They involve ordinary power series with multipliers which
are fractional powers and/or logarithms. Similar expansions involving
Laurent series could be given at the irregular singularities, if any, but they
are normally of little use in studying the analytical properties of the
solution. They will not be discussed here. Instead we devote considerable
space to questions of growth and asymptotic integration. There is also a
brief mention of the group of the equation.

5.1. GENERAL THEORY: FIRST ORDER CASE

An nth order DE is linear and homogeneous if it is the sum of terms of
the form Pk{z)win~k\z) equated to zero; thus
P0(z)w(n\z) + P1(z)w(",)(z) + • • • + P„(z)w(z) = 0. (5.1.1)
It is often convenient to write L(z, w) or simply L(w) for the left
144
GENERAL THEORY: FIRST ORDER CASE 145

member. The linearity implies that the sum of two solutions is a solution,
and a constant multiple of a solution is also a solution, for

L(0, + C>2) = C,L(w,) + C2L(w2) = 0.

The general solution of (5.1.1) involves n arbitrary constants and is a
linear homogeneous function of the constants. The initial-value problem

w(zo) = Wo, w'(zo)=w,, w(n_,)(z0) = Wn-i (5.1.2)

has a unique solution when z = z0 does not belong to the fixed sing-
ularities ofthe equation.
The equation
L(w) = Q(z) (5.1.3)
is said to be linear and nonhomogeneous. Its solution is of the form
W(z)+ V(z), where W(z) is the general solution of (5.1.1) and V(z) is a
particular solution of (5.1.3). Any particular solution will do unless some
selective condition has to be satisfied. Thus in the case of

w' + w=z (5.1.4)

there is one and only one solution which is a polynomial, namely, z - 1,
and there is one and only one solution which is zero for z = 0, namely,
e~z+z-l.
Let us consider the first order case which figured in Section 4.1. The
equation
w'(z) = F0(z) + F1(z)w(z) (5.1.5)
can be solved by "quadratures," i.e., by explicit integration. To achieve
this, set w = uv. Then
(uv)' = uv' + vu' = F0 + Fiuv.
A condition may be imposed on one of the factors. Take

v' = Fxv so that v(z) = exp £J Fx(s) dsj (5.1.6)

and vu ' = F0. Hence

u{z) = C + j F0(f)exp[-J F,(s)<fc] dt

Suppose that it is desired to find h>(z; z0, w0), the solution that takes on
the value w0 when z = z0. This gives

w(z;z0, Wo) = exp F,(s)dsj

x |h>0+ jZ F0(Oexp[-J' F,(s)£fc]dfJ. (5.1.7)

146 FIRST AND SECOND ORDER EQUATIONS

This expresses the fact that w(z; z0, vv0) is the sum of the solution of the
homogeneous equation which is w0 at z = z0 plus the particular solution of
the nonhomogeneous equation which is zero for z = z0.
This discussion has been of a somewhat formal nature, but if 2 ^ F0(z)
and z h-> F,(z) are holomorphic with a common domain of holomorphy D,
to which Zo belongs, every step makes sense and is justified. The
integrations may be performed along any path C in D, and the resulting
solution is locally holomorphic. It is holomorphic in the large if D is
simply connected. The solution cannot have any movable singularities, as
is obvious from its form. The fixed singularities are those of F0 and F,
plus the point at infinity.
We shall scrutinize in more detail the function v(z) defined by (5.1.6).
Suppose that F,(z) is single valued and holomorphic except for isolated
singularities, one of which is located at z = a. Here we have a Laurent
series

Fl(z) = f,F>n(z-ay. (5.1.8)

The question is now, What happens to w(z; z0, w0) in a neighborhood of

z = a. Here the first question is: What happens when z describes a closed
path T in the positive sense around z = a, leaving all other singularities, if
any, on the outside? Suppose that we start at a poinl z = z, on T with a
function element u0(z), and we return to z, with a function element t>,(z),
possibly distinct from v0(z). But both elements are solutions of the
equation v' = Fxv, all the solutions of which are constant multiples of an
arbitrarily chosen solution, not identically zero. It follows that there is a
number co such that vx(z) = cov0(z) and co^ 0. Set

where, to start with, the determination of the logarithm is left open. This
means that co is determined only e2mp =modulo 1. The power (z - a)p is
multiplied by co

as z describes T, and
V(z) = (z-a)pv(z)
is single valued in the neighborhood of z = a and as such may be
represented by a Laurent series convergent in a disk 0 < |z - a \ < R, so that

v(z) = (z-ayfJbn{z-a)\ (5.1.10)

Here there are essentially two distinct possibilities. If z = a is actually

GENERAL THEORY: FIRST ORDER CASE 147

a singular point of F,(z), some powers with negative exponents occur in

(5.1.8) and there are two possibilities: (i) a finite number, and (ii) infinitely
many, of such powers. In the first case F,(z) has a pole; in the second, an
essential singular point. We have a similar situation for V(z). We now
want to know what condition Fi(z) must satisfy in order that V(z) tends
to a finite definite limit as z -» a. If this is the case, then

v(z) = (z-a)p[bo+bl(z-a) + - • •], (5.1.11)

where now p is a definite real or complex number. To find what condition
this imposes on Fu we note that v(z) is the logarithmic derivative of F,(z),
so that

tj( \ = n ( w P ■ bi + 2b2(z - a) + - •
F,(z) [log v(z)] =--+b^biiz _a) + b2(z
_a)2 +
= (z - a) l[c^ + c0(z - a) + ci{z - a)2 + - • •]. (5.1.12)
Thus F, must have a simple pole at z = a with residue c_, = p.
If this condition is satisfied, we say that z = a is a regular -singular point
of (5.1.6). The first study of singularities of linear DE's was made by
Lazarus Fuchs in 1866 and was followed by the work of G. Frobenius
(1849-1917) in 1873. Fuchs referred to such a singularity as a Bes-
timmtheit s stelle \ later German terminology is schwach singuldre Stelle,
and the term stark singuldre Stelle is used for an irregular-singular point.
For (5.1.6) such a singularity would correspond to a pole of higher order
than the first of F,(z) or an essential singular point. A point of the second
kind is z = 0 for the equation

w' = z'2w with w(z) = Cexp(-iy (5.1.13)

The origin is evidently an essential singular point of the solution which
goes to zero as z-»0 in the sector |argz|<57r, and to infinity in
|arg (- z)| < {tt. Furthermore, w(z) takes on every value # 0 and oo in
6 -sectors centered on the imaginary axis. This is typical for irregular
singularities of finite rank or grade, terms which will be defined in Section
5.4. The DE

w' = ezw withw>(z) = Cexp(e2) (5.1.14)

has z = oo as an irregular-singular point of infinite rank and exhibits a
correspondingly more complicated behavior at infinity. Instead of sectors
we have now to consider half-strips.
The equation (5.1.6) belongs to the Fuchsian class if there is only a
finite number of singular points, each of which is a regular-singular point
including the point at infinity. This forces F,(z) to be a rational function
148 FIRST AND SECOND ORDER EQUATIONS

with simple poles, say, P(z)

(5.1.15)
F,(z) =
fl(2-<ii)
In order that the point at infinity also be regular-singular, the degree of
P(z) cannot exceed n - 1. Let the partial fraction expansion of F,(z) be
F.(*) = 2t3
so that
n
v(z) = Y[(z-ai)\ (5.1.16)

EXERCISE 5.1
1. When is v(z) as defined by (5.1.16) single valued?
2. What does it look like at z = oo? What conditions should the exponents satisfy in
order that v{z) be single valued in a neighborhood of z = oo?
3. Verify the statements concerning the behavior of the solutions of (5.1.13).
Show also that in the disks |z - / 2 " | < 2~"~2, n = 0, 1, 2, ... , the solution takes
on every value except zero and infinity infinitely often [Gaston Julia, 1924].
4. Discuss limits and value distributions for the solutions of (5.1.14). There are
two Picard values.
5. Suppose that Fi(z) has two simple poles only, one at z = a, the other at z = b.
Suppose that F0(z) is a polynomial in z of degree not exceeding two. Find the
nature of w(z; z0, w0) at z = a.
6. Discuss the equation w' = i(z - \)[w. Here z = 1 is a regular-singular point.
7. Explain the conditions for belonging to the Fuchsian class.

5.2. GENERAL THEORY: SECOND ORDER CASE

It was observed in Section 4.1 that there is a close connection between
Riccati's equation,
y'(z) = F„(2) + F,(z)>>(z) + F1(z)[y(z)]\ (5.2.1)
and the linear second order DE,

w"(z) + P{z)w'(z) + Q(z)w(z) = 0. (5.2.2)

Setting

(5.2.3)

we proceed from (5.2.1) to a DE of type (5.2.2), while setting

GENERAL THEORY: SECOND ORDER CASE 149

w(z)

takes (5.2.2) into a Riccati equation. As a matter of fact the theories of

these two types of equations are essentially equivalent.
Now Riccati's equation has no movable branch points, but it does have
movable poles. The first property clearly carries over to the second order
linear case, but what happens to the movable poles in the transit? Suppose
that y(z) has a pole at z = a, not one of the fixed singularities, i.e., the
zeros of F2, the singularities of F0, Fu F2, and the point at infinity. Thus
F2(a)^0, so that

(z - a) t1 + a^z
= ~ r2\a) tt^\ a^z - )2 + • • •]• (5.2.5)

This follows from setting y = l/v and finding the solution of the corres-
ponding Riccati equation (4.1.16), which has a zero at z = a. Now F2(z) is
holomorphic at z = a and has a convergent power series expansion there.
It follows that

- J F2(t)y(t) dt = j [(t - a)'1 + bo+b^t - 0) + • • •] dt

= \og (z-a) + P,(z-fl),
where P\(u) is a power series in u. Since w(z) is the exponential function
of this expression, it is of the form

w(z) = c(z - a)P2(z - a),

where P2(u) is a power series in u and P2(0) = 1. Convergence follows
from the double series theorem of Weierstrass. Thus a pole of y(z)
becomes a zero of w(z), and this function is holomorphic at z = a.
With a slight change of notation, suppose that w,(z) and w2(z) are
solutions of the DE

P0(z)w"(z) + P,(z)w'(z) + P2(z)w(z) = 0, (5.2.6)

and that w2(z) is not a constant multiple of w,(z). It is assumed that z is
restricted to a domain D, where the coefficients P, are holomorphic and
Po(z)#0. We have then two more equations:

P0(z)w'{(z) + P,(z)w;(z) + P2(z)w,(z) = 0,

Po(z)w'2'(z) + P,(z)w2(z) + P2(z)w2(z) = 0.
These equations, together with (5.2.6), form a system of three linear
homogeneous equations which are satisfied by P0(z), Pi(z), P2(z) not all
150 FIRST AND SECOND ORDER EQUATIONS
= 0
W2 w2
identically zero. It follows that the determinant of the system
W2 W i
w w w
(5.2.7)

w'l w2\
must be identically zero. Next, w'2\
w\ thew2coefficients P, are proportional to the
minors of the first row, i.e.,H>2 w2
wA w\\
(5.2.8)
P0:P,:P2 =

This shows that it must be possible to find

w'l a system w,, vv2 such that the
first minor is not zero and is a holomorphic functionw'l of z in D. Such
systems were introduced by Fuchs in 1868 under the name fundamental
systems.
For the purpose of orientation, suppose that the first minor is identi-
cally zero so that
wxw2 — w2w[ = 0.
w"
We may assume that wx ^ 0; then division by w] gives

or w2/wx = c, a constant. In this case w2(z) is a constant multiple of w,(z),

>
i&y<[see (1.1.1) and below]. Thus for
so that Wi and w2 are linearly dependent
a fundamental system we must choose linearly independent solutions of
(5.2.6). Now, if z0 is not a singular point, the initial conditions:

w,(z0) = 0, w2(z0) = 1,
(5.2.9)
w\(z0)= 1, w2(zo) = 0,
determine uniquely a pair of solutions which cannot be constant multiples
of each other. We shall see in a moment that the first minor is nowhere
zero in D if it is not zero at z = z0.
If w,, w2 form a fundamental system, the DE can be recovered from
(5.2.8) by expanding the determinant (5.2.7) according to the elements of
the first row. This gives

wx w'[ - w2w"
Pi _ Wx P2 _ w'xW2- W2w"x (5.2.10)
P0 W2 - W2W\ Po WxW2~ W2W\
Here the numerator of the first quotient is the derivative of the de-
nominator. Let W(z\wx,w2) denote the denominator. It is called the
Wronskian after the Polish mathematician known as Wronski, whose real
name appears to have been Josef Maria Hoene (1778-1853). From
GENERAL THEORY: SECOND ORDER CASE 151

-PJP0= W'lW it follows that

(5.2.11)
W(Z)=W(z0)^[-fzo^dt\
a formula due to Niels Henrik Abel (1802-1829). It occurs in one of his
five memoirs in Vol. 2 (1827) of Crelle's Journal (= Journal der reine und
angewandte Mathematik, dubbed /. d. unangewandte Math, by Crelle's
hecklers). In (5.2.11) we may integrate along any rectifiable path in D. If
W(z0) ^ 0, the same is true of W(z) everywhere in D. On the other hand,
if W(zo) = 0, then W{z) is identically zero in D. Thus we see that (5.2.9)
determines a fundamental system.
Going back to (5.2.7), we assume that w,, w2 form a fundamental
system and that w is any solution of (5.2.6). Then the fact that the
determinant is identically zero implies that the first row is linearly
dependent on the second and third rows. Hence there exist two constants,
C, and C2, such that
w(z) = Cxw{{z) + C2w2(z). (5.2.12)

We say that n functions, F,, F2, . . . , F„, are linearly dependent if n

constants, C,, C2, . . . , C„, exist, not all zero, such that

CxFx{z) + C2F2(z) + • +CFn(z)^0 (5.2.13)

in the common domain of definition of the functions. They are linearly

independent if such a relation can hold iff all the C's are zero. In this
terminology any three solutions of a linear second order DE are linearly
dependent, while two solutions which form a fundamental system are
linearly independent.
The solutions are locally holomorphic in any domain D in which the
quotients Pi/P0 and P2/Po are holomorphic. We come now to the question
of what happens in a neighborhood of a singular point, and start with the
problem of analytic continuation along a closed path T surrounding an
isolated singularity z = a. We start at z = Z\ with a fundamental system
wu w2. Analytic continuation along T back to z, will lead to a system wf,
wf which is still fundamental. Since any three solutions are linearly
dependent, there are constants Cjk such that
w f = Cnw, + C12w2,
? = C21w, + C22w2. (5.2.14)

These equations must be solvable for Wi, w2 so that

C\\C22 C\2C2\ 0.
152 FIRST AND SECOND ORDER EQUATIONS

This means that the matrix

1 (5 215)
S
is nonsingular. This matrix MMgoverns £
the analytic continuation of the
fundamental system (w,, w2) along the path I\ Using the notation of the
matrix calculus, we may write
v* = itv, (5.2.16)
where v and v* are the column vectors

Here the transit matrix M obviously depends on the choice of the

original fundamental system. However, changing to another fundamental
system (w0, Woo) merely replaces it by a similar matrix, i.e., a matrix it*
such that there exists a nonsingular matrix 2ft with it* = ^it^T1, for if
(w0, Woo) is a fundamental system there must exist a nonsingular matrix si
such that

h-'U^hl Lw2J
l_w>ooJ or h'W'M.
Lw2J LwooJ
In the second equality we operate on the left by it to obtain

Lw2J Lw?J LWqoJ

Here we operate with si on the left in the last two members to obtain

LwfJ LWooJ
|[M,° LW$oJ
]=fH, ii
4*11=^
as the result of the circuit applied to (w0, Woo). Here the matrices it and
sO/Ls^"1 are clearly similar in the sense of matrix theory.
By (1.4.14) the characteristic equation of it is

det[it- w«] = 0, (5.2.18)

where % is the 2-by-2 unit matrix. This gives

(o2 - (Cn + C22)w + C,,C22 - C12C21 = 0. (5.2.19)

Since
sUid~l -a)% = &Msi^ - si(o%si"x = si(M - co%)si'{
and the determinant of a product of matrices equals the product of the
determinants of the factors, and since det (si~*) = [det (si)] \ the charac-
teristic equations of similar matrices are identical. This shows also that
GENERAL THEORY: SECOND ORDER CASE 153

the characteristic roots of (5.2.18) are numerical invariants associated

with the path V. Since det (M) 7* 0, the roots are different from zero.
There are two cases: (i) the roots w,, a>2 are distinct, and (ii) the
equation has a double root. In case (i) we can find two linearly indepen-
dent solutions, Wo and Woo, which are multiplied by a constant when z
describes T, for among the matrices similar to M there is also the diagonal
matrix

9) is clearly similar to M, and we can find a regular matrix si such that

<£> = d-]Md. (5.2.21)

We start at z = z1 with the fundamental system w,, w2, and apply

successively the matrices si, M, and si~l. The result is a fundamental
system w0, wiyo, and a circuit along T takes it into

(5.2.22)
LwooJ LW2W00J

Although the fundamental system which transforms in this manner is not

uniquely determined, the multipliers are unique and are the two roots of
(5.2.19).
We can now discuss the analytical nature of the multiplicative solutions
along the lines pursued in Section 5.1 in the discussion of v(z). We set

<>.-*£?■
^

again leaving the logarithms arbitrary up to multiples of 2777. It is found

that
w0(z) = (z - ap V,(z), woo(z) = (z - apV2(z). (5.2.24)

Here Vx and V2 are single valued in the neighborhood of z = a and may

be represented by convergent Laurent series in z - a. It will be deter-
mined later under what circumstances the Laurent series may be replaced
by ordinary power series.
Let us first settle the case in which the characteristic equation has a
double root w,. A solution w0(z) can obviously be found, which is
multiplied by w, when z describes T. Take any solution Woo which is linearly
independent of vv0, and suppose that describing V takes Woo into

Woo = /3w0+ ywoo.

154 FIRST AND SECOND ORDER EQUATIONS

Thus the fundamental system vv0, Woo has the transit matrix

IP yJ
with respect to the path T. Its characteristic equation is
(0\ co 0
(w, - (o)(y -o)) = 0.
P y -co
Since Mo and M are similar matrices, we must have y = w, and

It follows that the quotient Woo/w0 is transformed into

COPi Wop
w0

as z describes V. This means that the quotient has an additive period pi to ,.

Now the function

has the same property, since the logarithm increases by 2m along the
circuit. Hence

Wo
5s- Ct) i h
f Z7Ti 108(2 -«> = ^>.

where V2(2) is single valued in a neighborhood of z = a and may be

expanded in a convergent Laurent series in terms of z - a. Thus in the
case of a double root of (5.2. 19) a fundamental system can be found of the
form
w0(z) = (z-apVi(z),

woo(z) = (z - a)*!L co\ =±7

v3(z) + -£ im J
V,(z) log (z - fl)l (5.2.25)
where
V3(z)= V,(z)V2(z)

and the V's are convergent Laurent series in z - a.

EXERCISE 5.2
1 . If the Wronskian of two locally holomorphic functions F, and F2 vanishes in a
set 5 with a limit point in the common domain of holomorphism of F, and F2,
show that the Wronskian vanishes identically in D and F2= CFU C a
constant.
REGULAR-SINGULAR POINTS 155

2. A 2-by-2 matrix z M(z) is holomorphic in a domain D iff the four entries

mjk(z) are holomorphic there. Suppose that det [M(z)] vanishes in a set S with
the properties as in Problem 1. Show that M(z) is algebraically singular in D,
i.e., has no inverse anywhere in D.
3. If cos z 1/2 and sin z 1/2 form a fundamental system for a linear second order DE,
find the equation.
4. Solve Problem 3 for the system z1/2, z 1/2 log z.
5. The DE's in Problems 3 and 4 have z = 0 and z = <» as only singularities. Find
the transit matrix for a circuit about the origin in the positive sense. Find the
characteristic equations and its roots. The singularities are obviously regular
singular. How would you choose the p's?

5.3. REGULAR-SINGULAR POINTS

In Section 5.2 it was shown that at an isolated singularity z = a of Pi/P0

and/or P2/P0 it is possible to find a fundamental system of the form
(5.2.24) if the roots of the characteristic equation are distinct, and of the
form (5.2.25) if they are equal.
The next problem is to determine if the singularity is regular singular or
irregular singular. In the first case the functions z V)(z) are holomor-
phic at z = a for a suitable choice of the p, 's. This problem can be solved
with the aid of formulas (5.2.10), which express the coefficients of the DE
in terms of a fundamental system. The question is now to find necessary
and sufficient conditions at z = a for the presence of a regular-singular
point at z = a.

THEOREM 5.3.1

A point z = a is a regular-singular point of the DE

w"(z) + P(z)w'(z) + Q(z)w(z) = 0 (5.3.1)

iff (i) at least one of the coefficients P and Q has an isolated singularity at
z = a, (ii) P(z) either is regular or has a simple pole at z = a, and (iii) Q(z)
either is regular or has a pole of order =s 2.
Proof One can proceed directly from (5.2.10), but a slight modification
saves labor. We write the fundamental system in the form

w,(z) = (z-ap Vx(z\

(5.3.2)
w2(z) = (z - ap V2(z) + A(z- ap V,(z) log (z - a)
with the following conventions: (i) A = 0 if p, ^p2 (mod 1), while (ii) if
pi = p2, then A = 8 = (/3/a>i)/(27ri) and V2 = Vs in our previous notation.
156 FIRST AND SECOND ORDER EQUATIONS

Now

Q(z) = -^-P(z)^, (5.3.4)

where the second relation follows from the fact that wx is a solution of
(5.3.1). In (5.3.3) and (5.3.4) the p's are now definite complex numbers and
the V's are ordinary power series with Vx{a)= V2(a) = 1. Now

— = A log (z - a) + (z - a)p>-p> V4(z),

where V4= VjVi and V4(a)= 1 if A =0. In any case the derivative
equals

w'd~ + (P2 " P.)U " aP ^[\ + 0(\z - a |)].

Next z
2 d (w2
(^) = A(z-af'-1[l + 0(\z-a\)]
+ (p2 - p,)(z -ap+p^[\ + 0(\z - a |)].
Now, if A = 0, p, ^p2, the logarithm of the last member is

(p, + p2-l)log (z-a) + 0{\z-a\),

and its derivative equals

P'z+!2Q"1 + 0(D = -P(z). (5.3.5)

This expression is also valid if A ^ 0 and p, = p2.
Next we have

Q(z)=-^-P(2)^=-4^-PU)A

where P(z) is given by (5.3.5) and

^ = _£i_+0(i).
w, z - a
We take p, ^ 0 for the time being. Then

w',(z) = p,(z - a)p'-f[l + 0(|z -

log w',(z) = (p, - 1) log (z - a ) + O(l),
so that

^ = £iz! + 0(1).
w i(z) z - a
REGULAR-SINGULAR POINTS 157

Combining and substituting in (5.3.4), we get

Q(z) = (z~^aj2 + °(|z " c = P,P2' (5'36)

which is seen to be true also for px = 0. This proves that the conditions of
the theorem are necessary and shows also that

and lim(z-fl)P(z) = -(p, + p2-l) 0.3. /)

lim (z - a)2 Q(z) = pxp2
are necessary conditions.
To prove sufficiency we write the DE in the form

(z-af w"(z) + (z-a)p(z)w'(z) + q(z)w(z) = 0, (5.3.8)

following the precept of G. Frobenius. Here p(z) and q(z) are holomor-
phic in some neighborhood of z = a. The quadratic equation
f0(s) = 5(5 - l) + p(a)5 + q(a) = 0 (5.3.9)
is known as the indicial equation at z = a, and by (5.3.7) its roots are px
and p2. It is required to show that (5.3.8) has at least one solution of the
form (z - a)pV(z - a) and two such solutions if px and p2 are incongruent
(modulo 1). To simplify the notation we take a = 0. We also assume that
the singularities of p(z) and q(z) lie outside the unit circle. If this is not
the case at the outset, an affine transformation z az + b will lead to the
desired situation.
We have thus the equation

z2w"(z) + zp(z)w'(z) + q(z)w(z) = 0 (5.3.10)

with

(5.3.11)
n=0 q{z)=^qnzn.
p(z)=f,PnZ\ n=0
Here we substitute

w(z)= n2=0 c„zn+p, (5.3.12)

where p and the coefficients cn are to be determined. Here various formal
processes are involved, as we know from the discussion in Section 2.4.
The series (5.3.12) is differentiated twice termwise, the formal first
derivative is multiplied by the series for p(z), using Cauchy's product
theorem, and similarly the series for q(z) and for w(z) are multiplied; the
three series are then combined, terms with equal powers of z are
collected, and their coefficients are equated to zero. All this is justified if
we are actually dealing with absolutely convergent series. We have to
prove the convergence a posteriori; nothing in our previous work tells us
158 FIRST AND SECOND ORDER EQUATIONS

that the series are convergent, although this is a very plausible guess. First
we have to find the series, however. The indicated procedure leads to an
infinite set of conditional equations:

Co/o(p) = 0,
c,/0(p + l) + c0/i(p) = 0,
c2/0(p + 2) + c,/,(p + 1) + C0/o(p) = 0,

cj0(p + *) + ck_,/,(p + * - 1) + • • • + Co/Up) = 0,

Here /0(p) is given by (5.3.9) and

fi(p) = ppi + qh 0<y. (5.3.14)

We may assume that c0= 1. Then the first condition under (5.3.13)
requires that /0(p) = 0, so that p has to be one of the roots of the indicial
equation (5.3.1 1). The latter has two roots, p, and p2, where the numbering
is such that Re (pO ^ Re (p2). After p has been chosen, (5.3.13) determine
successively c,, c2, . . . , ck, . . . , provided that /0(p + k) ^ 0 for all > 0.
This will certainly be the case for p = p2, the root with the greater real
part. For p = p, there are four possibilities: (i) if p2-pi is not zero or a
positive integer, the procedure works also for p = p,; (ii) if p2 - p, = /c > 0,
we come to an impasse unless accidentally

ck-,/i(p, + * - 1)+ • • + ca/*(p,) = 0, (5.3.15)

in which case ck may be chosen arbitrarily and all coefficients cn with
n > k are uniquely determined in terms of the preceding coefficients: in
this case we obtain a one-parameter family of solutions corresponding to
p = p,, the parameter being ck ; (iii) p2 - p, = > 0, and (5.3.15) does not
hold; in this case and in (iv) p, = p2 only one solution is obtainable in this
manner and some other device is needed to produce a second linearly
independent solution. Let us first settle the convergence question.
Suppose that in case (i) or (ii) we have computed a formal solution. It is
desired to prove that it is convergent in some punctured neighborhood of
the origin. Here the assumed properties of p(z) and q(z) come into play.
Since these functions are holomorphic in a disk \z \ < r, where r > 1, we
can find constants K > 0, jR > 1 such that for all n

\pn\< K(n + \ylRn, \qn\< K(n + \)lR~n. (5.3.16)

Furthermore, it is possible to find a d > 0 such that

(5.3.17)
|/0(p + n)|>d(n + l)2.
REGULAR-SINGULAR POINTS
159
This is to hold for p = pi and p2 for all n > 0 in case (i) and for n > k in
case (ii). The assumption (5.3.16) implies that

I/, (p + k)\ < + l R -if j > o, * * 0, (5.3. 18)

where cr = max (|p,|, |p2|). Now

|c| |/0(p + n)| ^ |/i(p + n - 1)| + • • • + |co| |/„(p)|.

Suppose that an integer m and a constant M have been found such that

\c^jfjR-\ j=0,l,...,m-l. (5.3.19)

Such a choice can always be made, but here we want to show that if m
and M satisfy certain conditions, (5.3.19) holds also for j = m and hence,
by induction, for all /.
In fact, from (5.3.17)-(5.3.19) it follows that

cm\< M fl o-4-m 1a + m - 1 _1 cr + 1]
d(m + 1)2 12 m 3 m-1 m + 1 1 J'
Here the fraction (a + k)lk decreases as k increases, so for k between 1
and m it has its largest value (cr + 1) for k = 1. Hence

d(m + l)2 * l2 + 3+ +m + lj-

Here the quantity between the braces is <log(wt + 1). Hence

|cm|<
1 1 £^d(m + 1) to8 m(m+D
+ 1 MR (53.20)

and (5.3.19) will hold also for j = m provided m is chosen so that

K(a + l)log(m + 1) ^ x
d m + 1 (5.3.21)

Now d, X, cr are fixed numbers, so that this condition can always be

satisfied. Once m has been chosen, we can find an M such that (5.3.19) is
satisfied for j < m. A finite number of conditions are involved so that M
can be found. Once this has been accomplished, complete induction
shows that (5.3.19) holds for all j. Thus the formal solution is absolutely
convergent for 0 < z < R, and the formal solution is then the actual
solution.
In case (i) we have then two convergent series solutions, and they are
160 FIRST AND SECOND ORDER EQUATIONS

linearly independent for

W)=zP2"W)=z"""VAz)' (53-22)
where the V/s are absolutely convergent power series and V,(0) = 1. If
Re (px) < Re (p2) the quotient tends to zero with z, while if Re(p,) =
Re (p2) there is no limit. In neither case can the quotient be a constant, so
the two solutions are linearly independent.
The same argument applies in case (ii), the only difference being that
the integer m of (5.3.19) must also satisfy m > k.
In cases (iii) and (iv), p2 - pi = k, a nonnegative integer. Here we use a
method for reducing the order of a linear DE by one unit. In our case,
n = 2, this means that the reduced equation is of order 1 and the
discussion in Section 5.1 applies. We know one solution, h>2(z), corres-
ponding to p = p2. For the moment let w(z) denote the known solution,
and set

w = uj vdt, (5.3.23)
where v is to be found. Then

w' = u' j v dt + uv, w" = u " j v dt + 2u ' v + uv ' .

Thus v satisfies the equation

v(z) u(z) z
so that

\ogv = -2 log u-J^df

or

v(z) = C[u(z )r2exp [-\Z^f-dt\. (5.3.25)

Here we take u(z) = w2(z) = zP2V2(z), and note that the second factor
in (5.3.25) equals

zp°V(z) with-p0 = p, + p2-l. (5.3.26)

Here V and V2 are convergent power series, and V(0) = 1. In the case of
V(z) we use the fact that exp [P(z)] is a convergent power series if P(z)
has this property, this by virtue of the Weierstrass double series theorem.
Hence

v(z) = z-2w^' V0{z) = z'k~x V0(z),

REGULAR-SINGULAR POINTS 161

where V0(z) is also a convergent power series and Vo(0) = 1 . Suppose that

v(z) = z~k~1 J) bnz\ (5.3.27)

Then

| v{t)dt = ^-^zn k + bk logz,

where the prime indicates that k. Hence

w,(z) = w2(z) n=o

2 TTV
" K z" " + ^^(z) log z
or
w,(z) = z" V,(z) + /w2(z) log z. (5.3.28)
This qualifies as the second solution in the fundamental system where
w2(z) is the first component. It should be noted that normally bk^ 0. If
bk = 0, case (ii) is present instead of case (iii). Furthermore, if k = 0, so
that the indicial equation has a double root, b0 j6 0. Thus logarithms
always occur in case (iv). In addition, it should be noted that w2(z) is a
multiplier of logz. This completes the proof of Theorem 5.3.1. ■

EXERCISE 5.3

1. The equation z2w" - 3zw' + 4w = 0 has z2 as a solution. Find a second linearly

independent solution, using (5.3.23).
2. If a and b are constants, what is the indicial equation of z2w" + azw' + bw =
0? If the roots p, and p2 are distinct, then

zPl-zp2
P\- Pi
is a solution. What is the limit of this solution as p2-» Pi? Is the limit a solution
of the limiting equation, and if so why?
3. The equation z2w" + (z2 - k2 + \)w = 0 is closely related to Bessel's equation.
Find the indicial equation at z = 0 and its roots. They differ by an integer if 2k
is zero or an integer. If k = 0 or a positive integer, a logarithm is present in the
general solution, but not if k = n + \. The latter statement is trivial if n = 0,
k = \. Verify it for n = 1, k = i, given that (z cos z - sin z)/z is a solution.
4. Prove that when the indicial equation has a double root the second solution
always involves a logarithm. Suppose that z describes a closed circuit in the
positive sense about the origin (an isolated regular-singular point). Find the
transit matrix.
5. Verify (5.3.16).
6. Justify (5.3.17).
7. Apply the transformation (5.3.23) to a third order linear DE.
162 FIRST AND SECOND ORDER EQUATIONS

5.4. ESTIMATES OF GROWTH

Our study of irregular-singular points will proceed along three different

lines which complement each other. The singularity is placed at infinity.
In this section we try to estimate the rate of growth of a solution when
rates of growth for the coefficients are known, the equation being

w" + P(z)w' + Q(z)w = 0. (5.4.1)

In Section 5.5 we shall consider asymptotics on the real line for real
coefficients, and in Section 5.6 we shall use a method of asymptotic
integration in sectors of the complex plane.
The coefficients are supposed to be single valued and holomorphic for
O^R < |z| <oo, and at least one of the functions

zP(z) and z2 Q(z) (5.4.2)

has a singularity at infinity. It will be assumed that the solution w(z) under
consideration is single valued for |z| > R and that it behaves as an entire
function of z for |z| > R. With the notation borrowed from the Nevan-
linna theory, we define the order of w(z) to be

p(w) = hmsup &lQgr (5.4.3)

where R < r and
ITT Jo
T(r; w) = m(r, oo; f) = 4~ f ^log" \w(reie)\ d6. (5.4.4)
Furthermore, when w(z) ranges over the solutions of (5.4.1), we define
the grade of the singularity of the equation as

g = supp(w). (5.4.5)
Note that the grade is not the same as what is usually referred to as the
rank of an irregular singular point: the rank is an integer (or infinity),
whereas the grade may very well be a fraction. In fact, for an nth order
linear DE the grade must be an integral multiple of I In, as first observed by
the Swedish mathematician Anders Wiman (1865-1959) in 1916.
Although the conditions on P(z) and Q(z) stated above guarantee that
z=oc is an irregular-singular point, they do not exclude, for instance, a
polynomial solution. Thus the equation

w"-zQ(z)w' + Q(z)w = 0 (5.4.6)

is satisfied by w(z) = z for any choice of Q(z).
An early attempt to estimate the grade of the irregular singularity for
the case of an nth order linear DE with polynomial coefficients is due to
ESTIMATES OF GROWTH 163

Helge von Koch (1870-1924) in 1918, followed a year later by Oscar

Perron (1880-), who removed an e from the estimate. If a personal note
may be permitted, the author heard von Koch present his solution in a
course on infinite determinants at the University of Stockholm in 1917-
1918. It was infinitely complicated. We shall sketch a proof here; the
general theorem for nth order equations will be given in Section 9.4.

THEOREM 5.4.1

Suppose that
P(z) = zp V,(z), Q(z) = zq V2(z) (5.4.7)
where we set p = - oo /f P(z) = 0. The functions Vi and V2 are assumed to
be holomorphic for |z| > R and to tend to finite limits ^ 0 as z -> °°. Set
g0=l + max (pjq). (5.4.8)
Then the grade of w(z) is at most g0, and there are bounds B = B(w), and
K independent of w, such that

|w(r e")| *s B(w) exp (Kr8*). (5.4.9)

Proof. The idea of the proof is to reduce the second order linear DE to a
linear first order vector-matrix equation, to take norms, and to cast the
result in a form to which Gronwall's lemma applies Theorem 1.6.5. To
carry out this program we set

wx = zaw, w2 = z^w', (5.4.10)

where a and p are real numbers to be chosen. The result of the
substitution is

w\ = %, + za~l3w2,

W2 = - z"-aQ(z)wl + [|- P(z) jw2.

This may be written as a vector-matrix equation,

\'(z) = d(z)\(z), (5.4.11)

with obvious notation. The norm of si(z) is taken to be

\\d(z)\\ = max [\an(z)\ + |a12(z)|, |a21(z)| + |a22(z)|]

and

||v(z)|| = max [|>v,(z)|,|w2(z)|].

The problem is now to choose a and /3 in such a manner that \\si\\ is as
small as possible for large values of |z|, taking into account the assump-
164 FIRST AND SECOND ORDER EQUATIONS

tion (5.4.7) on the behavior of P and Q for large |z|. It is clear that the
behavior of an(z) = za~& and a2i(z) = - zli~aQ(z) is decisive. They be-
come approximately equal if a-p=\q^-\. Some simplification is
gained by choosing a = {q, (3 = 0. The final choice leads to the matrix

the norm of which is 0(rY), y = max (p, \q) and r = |z|.

From the resulting equation we then get

v(z) = v(z0) + f 5«5)v(s)ds. (5.4.13)

Here £ < |z0| < \z\, arg z0 = arg z, 0 ^ arg z0 < 2tt, and |z0| = r0. After integ-
rating radially from z0 to z, we can find constants C and M such that

||v(z0)||< C, \\d(z)\\^Mr\ z = rel\ (5.4.14)

Then if ||v(z)|| = f(r) we have the inequality

f(r)<C + M[ tyf(t)dt, (5.4.15)

which is of the form (1.6.13). This implies that

/(r)<C + C J Mty exp (J^ Muy duj dt,

which simplifies to

fir) < C exp (r^+1 - r?+l)]. (5.4.16)

Since y + 1 = g0 and max [|h>i(z)|, |w2(z)|] < /(r), it follows that

|w(r^ie)|<B(w)r-q/2 exp(— rK«), (5.4.17)

where
B(w) = C Qxp(-—
V gorgA
/

This is (5.4.9) in a sharper form. If the factor r~q'2 is omitted, we obtain

an estimate for |w'(z)|. This proves Theorem 5.4.1. ■
This is a best possible result in various ways. Thus, if the coefficients P
and Q are polynomials, we have always g = gQ and g0 is the normal order
of a transcendental entire solution. Note that there may be a polynomial
solution in addition to the transcendental solution. Thus consider (5.4.6),
where Q(z) is a polynomial of exact degree q. Here g0 = q +2, and using
the order reducing transformation (5.3.23) with u = z it is seen that the
order of the transcendental solution is indeed q + 2.
 165
ESTIMATES OF GROWTH

The following observation illustrates one sense in which the result is

optimal:

THEOREM 5.4.2

// g = \k, k a positive integer, then there exist a second order linear DE for
which z = oo fa an irregular-singular point of grade g and a solution w(z)
such that
lim|z| « log|w(z)|>0 (5.4.18)
for a radial approach to infinity, a finite number of directions excepted.
Proof. It suffices to produce the example

w(z) = {[exp (zK) 4- exp (-z8)], (5.4.19)

which is clearly of order g and satisfies the DE

w'-g2z2g-2w =0. (5.4.20)

Here 2g - 2 = k - 2 is an integer and p = - 1 , \q = g - 1 , so that g =

1 + max (p, \q). The exceptional directions in (5.4.18) are arg z =
{mlk)TT, m = 1, 2, . . . , k. These rays divide the plane into k sectors in the
interior of which either Re(z*) or Re(-z*) goes to +oc. ■
The orders of the transcendental entire functions which satisfy second
order linear DE's with polynomial coefficients have been investigated by
K. Poschl (1953) and H. Wittich (1952). With our previous notation, let
p 2* q + 1. Then all transcendental solutions are of the order 1 + p = g0. If
p ^ [q, all transcendental solutions are of the order 1 + \q = g0. Deviation
from this pattern can occur only if \q < p *s q. Here g0 = \ + p, and there
are always solutions of this order; under certain circumstances, however,
a lower order q - p + 1 may also be present.
In the case last mentioned the solution of lower order admits w = 0 as a
Picard value. If this is understood in the strict sense to mean that
h>i(z) 0 for all z, then w,(z) = exp [B(z)], where B(z) is a polynomial,
say of degree d. Taking this observation as our point of departure, we can
construct infinitely many linear second order DE's which have exp [B(z)]
as particular solutions, while the order of the general solution is > d. Let
C(z) be any polynomial of degree m. Then w,(z) satisfies the equation

w"-[B'(z)+ C(z)]w' - [B"(z) - B'(z)C(z)]w = 0. (5.4.21)

A linearly independent solution is given by

w2(z) = exp[B(z)]Jo exp

f C(s)ds-B(t)]dt. (5.4.22)
166 FIRST AND SECOND ORDER EQUATIONS

If now m + 1 > d, the order of w2 is m + 1. Here p = ra, q = m + d - \ >

m, so that
\q = s(m + d — l)<m = p<m+d-\ = q.
Thus the stated conditions are satisfied. Furthermore, g0 = m + 1 and
p(w) = go except for constant multiples of Wi(z).
Using Theorem 5.4.1, we can now give the proof of Theorem 4.6.3,
omitted in the preceding chapter.
Proof of Theorem 4.6.3. Given the Riccati equation

w' = R0(z) + Ri(z)w + w\ (5.4.23)

where R0 and are rational functions of z such that

R0(z) = O(\z\a\ Rl(z) = 0(\z\p) asz-^oo, (5.4.24)

it is supposed that (5.4.23) has a solution w(z), single valued and
meromorphic for \z \> R ^0 and with infinitely many poles clustering at
infinity. It is to be shown that w(z) is of finite order, p(w), which is a
positive multiple of \— more precisely,
p(w)= 1 + max (/Ma). (5.4.25)
We restrict ourselves to the case in which R0 is a polynomial of degree a
and Ri a polynomial of degree (3.
Set
W
» = ~W>
so that
W' + RfcW + Ro{z)W = 0 (5.4.26)
and - w is the logarithmic derivative of W. Now W is an entire function
of z of finite order, the order being g0, the right member of (5.4.25). We
know that there is only one possibility for a lower order: if \a < p ^ a,
there is a solution of (5.4.26) of order a - (3 + 1 < g0. We may disregard
such a solution, however, since it has zero as a Picard value and its
logarithmic derivative can have only a finite number of poles. Thus
p[ = g0. We have now

m(r, oo; w) = ra(V, oc;-^^ = 0(log r)

and
N(r, oo; W) = N(r, 0; W) = T(r ; W) + 0(log r) = O(rgo),
so that

as asserted. ■ T(r; w)=0(rg°),

ESTIMATES OF GROWTH 167

We return now to Theorem 5.4.1. If P or Q is an entire transcendental

function instead of a polynomial, we may expect the transcendental
solutions of (5.4.1) to be of infinite order. This is indeed the case.
Equation (5.4.6) shows that such a situation does not preclude the
presence of a polynomial solution.

THEOREM 5.4.3

If in (5.4.1) either P or Q is an entire transcendental function while the

other is a polynomial, then every transcendental solution of (5.4.1) is an
entire function of infinite order. This is not necessarily true, however, if
both P and Q are entire.
Proof. Suppose that Q is transcendental while P is a polynomial. The
proof is an application of the Nevanlinna theory. If the assertion were
false, there is a solution w(z) which is an entire function of finite order so
that T(r; w)= 0(rp) and T(r; w) = m(r,oo; w). Then

Q(z) = - —W- P(z) —W = - ^

W —W- P(z) —w , (5.4.27)

and a consideration of proximity functions gives

m (r, oo ; Q) m[r, «>; ^ + 2m[r, °°; ^) + m (r, oo; P) (5.4.28)

with an error of at most 0(log r). Here each of the terms on the right is
0(log r), and this leads to a contradiction since T(r; Q) = m(r, oo; Q)
goes to infinity faster than any constant multiple of log r.
To get a counter example for the case in which both P and Q are entire
transcendental functions, we use (5.4.21). If B(z) is a polynomial in z of
degree d and if C(z) is entire transcendental, both B'(z) + C(z) and
B"(z) - B'(z)C(z) are entire and Wi(z) is of finite order while w2(z) is of
infinite order. ■
The following theorem due to H. Wittich (1948) may be regarded as a
converse of Theorem 5.4.3.

THEOREM 5.4.4

In (5.4.1) suppose that P and Q are entire functions, and suppose that the
equation has a fundamental system w,(z), w2(z), where w, and w2 are
entire functions of order p, and p2, respectively. Then P and Q are
polynomials.

Proof. Let W(z) = W(z ; w,, w2) be the Wronskian of the system. This is
168 FIRST AND SECOND ORDER EQUATIONS

an entire function of order p = max (p,,p2). But

P(z) = -^[log W(z)\.

It follows that

m(r,o); p) = m^oc;^^ = 0(log r). (5.4.29)

Now P is an entire function, so this gives T(r; P) = 0(log r) and P must
be a polynomial. From (5.4.27) we then obtain the fact that Q is also a
polynomial. ■

THEOREM 5.4.5
For the DE
w" = Q(z)w, (5.4.30)
where Q(z) is an entire function, if it is known that the equation has a
solution w(z) which is an entire function of finite order, then Q(z) is a
polynomial.
Proof We have

m (r, oo; Q) = m(r, oo; = m(r, oo; ^ = 0(log r),

which again implies that Q is a polynomial. ■
In his investigations Wittich has made extensive use of the so-called
central index and the corresponding maximal term in the Maclaurin series
of an entire function,

E(z)= 2 anz\ (5.4.31)

n 0
Let r = |z| >0 be fixed. Since the terms |a/ |ri tend to zero as j -*oo, there
is a largest member of the sequence {|«/|ry}. If there are several contes-
tants for this distinction, the one with the highest index / is designated as
the maximal term, m(r) = m(r;E), and its index j is the central index
v(r) = v(r ; E). These notions were introduced by A. Wiman (1914-1916)
and amplified by Georges Valiron in 1923. Wiman showed that there is a
close relation between the maximal term and the maximal modulus of an
entire function. Let M(r; E) = max |E(re,<p)|; then
m(r; E)<M(r; E)<m(r; E)[logm(r; E)]U2'\ 0< 8, (5.4.32)
and

v(r\ E) < [log m(r; E)]m'\ (5.4.33)

where the first part of (5.4.32) holds for all r, while the second half
ESTIMATES OF GROWTH 169

together with (5.4.33) holds for r outside an exceptional set of finite

logarithmic measure, i.e., a set where the total variation of log r is finite.
These relations extend also to the derivatives of E(r). Thus, if vj(r; E) is
the central index of Ea\z) and M}{r) = M(r; E(/)), then, for r outside sets
of finite logarithmic measure,

vM = Hr)[\ + eMI 0 < e,(r) = 0[v(r)lM]9 (5.4.34)

^jM(r)[l + o(l)]. (5.4.35)

We shall not prove any of these statements; the interested reader is
referred to Wittich (1955, pp. 9-10) and the literature therein cited.
At this juncture we call attention to another of Wiman's observations.
In the case of the DE (5.4.30), where Q(z) is a polynomial of degree d, it is
found that

Jo
log M (r; w) ~ v(r\ w) ~ [' [Q(t)]m dt (5.4.36)
up to terms of lower order. Suppose now that we are given an nth order
linear DE
ww + Pt(z)wlH-n+-' - + Pn(z)w =0, (5.4.37)
where the coefficients are polynomials in z (or, more generally, functions
holomorphic for 0<R<\z\<°° with poles at infinity). Consider the
algebraic equation

Wn + P,(z) Wn~' + • • • + Pn(z) = 0. (5.4.38)

If the P's are polynomials, this equation defines an algebraic function
normally with n determinations
Wi(z)9 W2(z),...,Wn(z) (5.4.39)
at infinity. According to Wiman, the corresponding functions

exp[J W,(s)ds],...,exp[J W„(s)ds] (5.4.40)

are approximate solutions of (5.4.37). Details were worked out in an
Uppsala dissertation (1924) by Mogens Matell.

EXERCISE 5.4

1. Verify the statements concerning (5.4.21). In particular, prove that the order
of w2(z) is infinite when C(z) is an entire transcendental function, assuming
that the coefficients of the Maclaurin series are nonnegative. Prove that in
170 FIRST AND SECOND ORDER EQUATIONS

this case

w2(r) > [C(r)] 1[exp [J' C(t) dt]- 1J.

2. Consider the equation w" = z2w with w(0) = 1, w'(0) = 0. Find its Maclaurin
series. Compute v{r; w) from the series, and show that log M{r;w) =
\r2 + 0(\ogr).
3. Let Q,(0 and Q2(t) be positive and continuous for 0^ t < oc. Consider the
three DE's
M"(0- QMu(t) = 0, t/'(0" Q2(*MO = 0,
w"(t)-[Q>(0 + Qi(t)]w(t) = 0,
with m(0) = t;(0) = w(0) = 1, u'(0) = a, v'(0) = b, w'(0) = c, where a, b, c are
positive and c < a + b. Show that
u(t)v(t)>w{t), r >0.
4. Determine other solutions of (5.4.6) besides w = z.
5. In Theorem 5.4.4 what bounds can be obtained for the degrees of the
polynomials P and £>?
6. In (5.4.21) take B(z) = \z2, C(z) = z3 and determine the functions Wx{z) and
W2(z) of (5.4.38). Compare exp LP W,(0 dt] and exp [f W2(0 dr ] with the
known solutions of the DE.
7. For the equation zw " + \w ' + \w = 0 we have R =0, p = q = - \, and go = |.
Show that cos z1/2 is a particular solution which is of order j. Find the general
solution.
In the following six problems the function t ^G(t) is positive and
continuous on [0, o°). It is desired to study the solutions w(f) of the DE
w"(t)identity
the = G(t)w(t), where w(t) is real and not identically zero. A useful tool is

[w(t)w'(t)]ba = j [w'(t)f dt + j' G(t)[w(t)f dt.

8. Verify the identity.
9. Show that w(t)w'(t) can be zero at most once in [0,
10. From the point (0, 1) draw the graphs of the solutions corresponding to
w(0) = 1, h>'(0) = a, -oc < a < oc. Show that no two graphs can have a point
in common except for the starting point. Show that any solution with a 5= 0
becomes infinite with t.
11. Suppose that a <0. Show that the solutions separate into three mutually
exclusive classes: (i) solutions with a positive minimum, which go to infinity
with f; (ii) a solution which goes to a finite limit L, 0=5 L < 1; (iii) solutions
which have a positive zero and go to - oc when t -> oo. There is only one solution
in class (ii), and its graph separates those of class (i) from those of class (iii).
12. Show that, if there is a solution such that lim w(t) = 1 and lim tw'(t) = 0,
then rG(OeL(0,«>).
13. Show the existence of a solution such that w'(t) and [G(t)]U2w(t) are both
L2(0,oo).
ASYMPTOTICS ON THE REAL LINE 171

5.5. ASYMPTOTICS ON THE REAL LINE

Before starting to study this section, the reader should familiarize himself
with Problems 5.4:8-13. In this section we plan an inquiry into the
behavior for large positive values of t of the real solutions of the DE

w"(t) = G(t)w(t), (5.5.1)

where G(t) is positive and continuous in [0, oo). The results of Exercise
5.4 indicate that the general solution becomes infinite with t but that an
exceptional solution tends to a finite limit as t -* + oo. In this setting the
exceptional solution is known as a sub dominant, while the general
solution, which involves two arbitrary constants, is termed the dominant.
Normally a subdominant tends to zero as t -> + oo. The case in which
tG(t)G L(0, oo) is exceptional. Here the dominant satisfies for some a the
conditions

lim-^7^=a^0, limw;(0 = a, (5.5.2)

while, if suitably normalized, the subdominant satisfies

limw0(0=l,
t-»oo limrv^(O
r — ►<» = 0. (5.5.3)
Conversely, the existence of such solutions implies that tG(t)E L(0, oo).
Actually this is true under the weaker condition that G(t) ultimately
keeps a constant sign.
We now assume that tG(t)f£L(Qf<x>). Define w,(0 by w,(0) = 1 and
w>',(0) = 0. Then w>, is increasing, and (5.5.2) is replaced by

lim^y^ = +oo, HmH>'1(0 = + a>. (5.5.4)

We now define a subdominant solution by

w0(t)=wl(t) f [w^Vds. (5.5.5)

the integral clearly exists, since wx(t) ultimately exceeds at for any
positive a. Furthermore, w0(t) is decreasing since

w'M = w\(t) |" [w,(5)r2 ds - [w1(t)V

< | wi(s)[w,(5)r2 ds - [wMV = 0. (5.5.6)

This implies that wQ(t) decreases to a nonnegative limit as r-> + oo. The
limit cannot be positive, for then tG(t) would be L(0, oo), contrary to
assumption.
172
FIRST AND SECOND ORDER EQUATIONS

The subdominant solution also has the property that w'0(t) and
[G(t)]mw(t) both are L2(0,oo). It is clear that no dominant solution can
satisfy this condition. On the other hand, the identity

W
o(Ow0(0- w0(a)w'o(a)= j' [w'0(s)]2ds +j G(s)[w0(s)f ds (5.5.7)
shows that the left member tends to the limit - w0(a)w'0(a) as t -* + <». It
follows that each of the integrals must then tend to a finite limit
< - w0(a)w'o(a) as t ^oc, and this proves the stated integrability proper-
ties.
We now set

M(t, w) = Jo
i'[w'(s)]2 ds, N(t, w) = JoP G(s)[w(s)]2 ds, (5.5.8)
L(r, w) = M(t, w) + N(t, w). (5.5.9)
As a curiosity we note that

M(t, w) = JV(r, w) iff w(t) = w(0) exp |J' [G(s)]m ds}. (5.5.10)
For a solution of (5.5.1) this can hold iff G(t) is a constant. For a general
continuous and positive G(t) it can hold only asymptotically. Actually,
for dominant solutions and under mild restrictions on G it holds in the
weak sense

N{t,—w)(=1.
hm^T^ (5.5.11)

Let G(t), in addition to being positive and continuous, also be strictly

increasing. The classification of the solutions given in Problem 5.4:11
implies the existence of (i) solutions with a zero, (ii) solutions with an
extremum (maximum or minimum), (iii) dominant solutions with
w(t)w'(t) t6 0 for O^t, (iv) subdominant solutions.
We now transfer the problem to a study of the Riccati equation

y'(f) = G(t) - [y(t)]2 with y(t) = ^g. (5.5.12)

The integral curves of this equation also exhibit tetrachotomy. If w(t)
belongs to class (i) and has a zero at / = a, then the y -curve of (i) has a
vertical asymptote for t = a and y(t) < 0 to the left of the asymptote and
y(O>0 to the right of the asymptote. If w(t) belongs to class (ii) and
w'(to) = 0, then y(t) < 0 for t < t0, positive for t0 < t. These y -curves form
class (ii). If w(t) belongs to class (iii), the y -curves of (iii) have positive
ordinates for all t 2* 0. Finally, there is one and only one y -curve of class
(iv):
ASYMPTOTICS ON THE REAL LINE 173

C0:y = y0(0, (5.5.13)

which corresponds to the subdominant solution w0(t) of (5.5.5). This
curve lies in the fourth quadrant and separates the curves of class (i) from
those of class (ii). All y -curves below C0 have vertical asymptotes; they
escape from the lower half -plane by going to - °° as t t a and reappear at
the upper end of the asymptote in the upper half-plane, where they spend
the rest of their existence. See Figure 5.1.

c0

Figure 5.1

We shall examine the y -curves for large values of t, aiming to show that
y0(t) is approximately - [G(t)]m for large values of t, while for all other
y-curves y is approximately +[G(f)]1/2- In a (t, y)-plane we mark the
curves
T+: y = [G(t)V'\ T :y=- [G(t)]il2. (5.5.14)
These curves, together with the positive /-axis, divide the plane into four
regions, D,-D4. Figure 5.1 gives a schematic view of the regions and some
of the integral curves. The figure brings out the fact, already clear from
the preceding discussion, that all curves, with the sole exception of C0,
ultimately enter D3, where y'(t)> 0. The curves must stay there, for Y+ is
an entrance boundary where the slope is zero. Since G(t) is strictly
monotone, an integral curve cannot cross from D3 into D4. This means
that ultimately
y(t)<[G(t)]m. (5.5.15)
174 FIRST AND SECOND ORDER EQUATIONS

We shall estimate the difference between the two sides

D(t) = [G(t)]m-y(t). (5.5.16)

THEOREM 5.5.1

// G(t) is continuous, positive, and strictly increasing, and if y(t) is a

solution of (5.5.12), increasing and positive for 0^c<t, then

D(s) ds < \ log G(t) - \ log y(c), (5.5.17)

8 £ D(s)[log G(s)] 2 ds < [log y(c)r i/y(c)> 1, (5.5.18)

//m supD(t)^Uim sup*^. (5.5.19)
Proof We have

y'(0 = G(0-[y(0]2 = {[G(r)]1/2 + y(0}I>(0>2y(0O(0

so that

£ D(5) ds < | £ ^ ds = | log y{t)-{ log y(c)

<HogG(0-Hogy(c),
as asserted. Next, if y(c)> 1,

1 __ f_jWL_>? f" D(t)dt r D(t)dt

logy(c) Jc y(0[logy(0]2 1 [logy(r)]2 °JC [logG(0]2'
and this is (5.5.18).
For the proof of (5.5.19) we add the assumption that G(t) has a
continuous derivative everywhere; the existence of a derivative almost
everywhere follows from the assumed strict increase. Under this assump-
tion D(t) is differentiate and satisfies a Riccati equation

D'(t) + 2[G(t)]mD(t)- [D(t)f = \G'{t)[G{t)Ym. (5.5.20)

Now 0<D(t)<[G(t)]m, since y(t) satisfies this inequality. It follows
that
[G(t)V,2D(t)-[D(t)]2>0,
so that
D'(t) + [G(t)yl2D(t)<\G'(t)[G(t)]m.
The left member of this inequality becomes an exact derivative upon
multiplication by

E(f) = exp |£[G(s)],/2 ds].

ASYMPTOTICS ON THE REAL LINE 175

Integration gives

D(t)E(t)-D(c)<±j 'J^dE(s)
^sup^[E(0-l], (5.5.21)
whence

D(t)<D(c)[E(t)V + i sup ^±{\-[E(t)]

If here G'(s)IG(s) is unbounded, its supremum is +°° and (5.5.19) is

trivially true. On the other hand, if the fraction is bounded, we note that c
can be taken as arbitrarily large, E(c) = 1, and E(t) \ °° with t. Now the
supremum is arbitrarily close to the lim sup, and again (5.4.19) follows and
Theorem 5.5.1 is proved. ■
Various comments and corollaries are in order.

COROLLARY 1
If G is increasing and has a continuous derivative, if

lim G'(t)[G(t)r3/2 = 0, (5.5.22)

t-»oo
and if w(t) is a dominant solution of (5.5.1), then

(im[G(t)r1/2^=l. (5.5.23)
Proof. If y(t) is the corresponding solution of (5.5.12), the problem is to
prove that
y(t)[G(t)]m = 1- D(t)[G(t)]112
tends to one as t becomes infinite, and this is equivalent to showing that
D(t)[G(t)Tm tends to 0. Now from (5.5.21) one obtains that

D(t )[G(t)TV2E(t) < D(c)[G(t)] m + \[G{t)Ym J" dE(s)

< D(c)[G(t)Tm + \ [ iQ^fi dE(s), (5.5.24)

since G(s) is strictly increasing. Then (5.5.22) implies that the integral is
o[E(t)] and thus proves (5.5.23). ■

COROLLARY 2

Under the assumptions of Corollary 1 (5.5.11) holds.

Proof. Apply the rule of L'Hospital to the quotient M/N. The quotient
176 FIRST AND SECOND ORDER EQUATIONS

of the derivatives is

G(t)[w(t)r
[w'(Q]2
and this tends to one by (5.5.23). ■

COROLLARY 3

Under the original assumptions on G(t) and for all dominant solutions
with w(c)>0, w'(c)>0

w(t) > [w(c) w'(c)],/2[G(t)r1/4 exp £ [G(s)]1/2 ds. (5.5.25)

Proof. This is a consequence of (5.5.17), for we have

whence ^ = [G(011/2-D(f),

log^= (' [G(s)]mds- f D(s)ds

> J' [G(s)]m ds -HogG(t) + ilogy(c),

and this implies (5.5.25). ■
An asymptotic relation of the form

w(f) = [1 + o(\)][G(t)] m exp [G(s)]1/2 ds} (5.5.26)

holds if, for instance, G has continuous first and second order derivatives
and

[G(0] 1/4£-2{[G(t)] V4}GL (0,oc). (5.5.27)

This follows from an application of the Liouville transformation. See the
next section.
The subdominant is harder to pin down, but we have the following
results.

THEOREM 5.5.2

If G(t) is continuous, positive, and strictly increasing, then the subdomin-

ant solution of (5.5.12) has these properties:

1. y0(t)<-[G(t)],/2 for all t.

2. l + tG^l^tyoWr'eL^,^).
ASYMPTOTICS ON THE REAL LINE
177

3. limjup y0(t)[G(t)r1/2 = -l.

4. The measure of the set where y0(t)[G(t)]~1/2 < - 1 - e is finite for
every e >0.
Proof. Assertion 1 is that the graph C0w'Q(of
t) y = y0(0 lies in the region D,.
It is certainly located in the lower half-plane for
Wo(tY
yo(t)
where w0(t) is defined by (5.5.5) and has a negative derivative. In D2 the
integral curves have positive slopes and are increasing so that such a
curve ultimately crosses the positive f-axis and gets into the upper
half -plane. All this is out of the question for C0. Hence y = y0(0 stays in
Di, and point 1 is verified.
Set y0(0 = -z0(t), and note that
W) = [zo{t)f-G{t),
whence

Jo [Zo(s)] Jo I [Z0(S)] J
Both integrands are positive, and as t tends to infinity, the left member
tends to the finite limit [zo(0)]_l. Thus the right member must also tend to a
finite limit, as asserted in point 2. Now a bounded continuous positive
function in L (0, o°) can exceed a given e >0 only on a set of finite measure.
The factor 1 + [G(0]I/2|yo(0|~! lies between 1 and 2, so it must be the other
factor, 1 - [G(t)]l/2|yo(Or\ that is small on the average. This is expressed
by points 3 and 4. ■

COROLLARY
We have

'W[G(A)=-1, (5-5-28)
and the set of values t for which the quotient is less than - 1 - e is of finite
measure.

EXERCISE 5.5
1. Suppose that w0(t)>0 is a subdominant solution of (5.5.1), and form

w(t)= w0(t) I [w0(s)Y2ds.

Show that w'(f)>0 and that w(t) isJoa dominant solution.
178 FIRST AND SECOND ORDER EQUATIONS

2. If t »-* f(t ) is positive and continuous together with its first and second order
derivatives, find the equation of type (5.5.1) which is satisfied by exp [f(t)].
3. Suppose that G and H are continuous increasing functions and 0< G(t)<
H(t). Consider the systems

w"=G(t)w, W"= H(t)W, with 0^ w(0)= W(0), 0 ^ w'(0) < W'(0).

Show that w(t)< W(t) for 0< t.

4. The function 1 1-> [G{t)] 1/4 exp {& [G(s)]]l2 ds} satisfies the DE w" = Q(t)w.
Find Q(t) if G(t) is positive and continuous, together with its first and
second order derivatives.
5. If 5[G'(t)f>4G{t)G"(t), show that (5.5.1) has a dominant solution
satisfying

[Hint: Use Problems 3 and 4. For the opposite inequality see (5.5.25).]
6. If Q is defined as in Problem 4 and if G(r)=r\ O^fc an integer, find
Q(t)- G(t). Show that the condition of Problem 5 is satisfied, and give the
corresponding estimates for dominant solutions.
7. Let t h-> F(t) be positive and continuous. The equation w"(t) + F(t)w(t) = 0
is said to be oscillatory in (a, b) if every real solution has at least one zero in
(a,b). If c >0, is every solution of w" + ct~2w =0 oscillatory in (0, oo)?
8. If F(O>M>0 for all f >0, prove that any interval of length irM~m
contains at least one zero of any given real solution.
9. Why is it necessary to specify "real" in these cases?
10. Given the equation y' = G(t)-y\ O^t, let T be the curve V: y = [G(t)]m.
Here G is supposed to be positive, continuous, strictly increasing, and
unbounded. Discuss the integral curves in a way similar to that used for
(5.5.12).
11. If a solution exists for t >a >0 but becomes infinite as t decreases to a,
show that y(t){t - a)]l2 tends to a limit, and find this limit.
12. Show that each solution is ultimately less than [G{t)]m. Set y(f) =
[G(t)]m-D(t), and show that D(t)(=L(a,™) for large a.
13. If linw G'(t)[G(t)Ysn = 0, show that

limy(r)[G(f)rl/3=l.

5.6. ASYMPTOTICS IN THE PLANE

We now return to the case of analytic coefficients and will be concerned
with a linear second order DE with the point at infinity as an irregular-
singular point normally of finite grade. The method to be used is a
reduction of the equation to a quasi-Bessel form with the aid of the
ASYMPTOTICS IN THE PLANE 179

transformation of Liouville (Joseph Liouville, 1809-1882), dating from

1837. This has a strong smoothing effect and leads to a transformed
equation which is a perturbed sine equation. It applies to formally
self -adjoint equations:

[K(z)w'(z)]' + G(z)w(z) = 0. (5.6.1)

We can use it for the equation

y" + Pi(z)y' + Q,(z)y =0, (5.6.2)

where the coefficients are polynomials in z or, more generally, functions
holomorphic for |z|>P except for poles at infinity. Condition (5.4.2) is
supposed to be satisfied; it is the necessary condition on the coefficients in
order for z = °° to be irregular singular. It is convenient, however, to
reduce the equation to a form where the first derivative is missing. This is
done by setting

which leads to y (z) = w(z) exp [- X-jZ P,(s) ds], (5.6.3)

w"(z)+Q(z)w(z) = 0, (5.6.4)
where
Q(z) = QAz)-\[PAz)f-\P\{z). (5.6.5)
Note that, if Pi and Qx are polynomials, so is Q. Moreover, the grade of
the singularity of (5.6.4) is the same as that of (5.6.2), namely,

g = 1 +max(p,J<?i) (5.6.6)
with obvious notation. Note that the multiplier of w(z) in (5.6.3) is of the
order p i + 1 g.
New variables are introduced in (5.6.4) by setting

Z = j [Q(s)V2 ds, W = [Q(z)],/4w. (5.6.7)

The result is of the form

W"+[\-F(Z)]W = 0, (5.6.8)
where differentiation is with respect to Z. Various expressions are
available for F(Z). Setting

[G(z)]]l4 = H{Z), (5.6.9)

we get

F(Z) = ^j^. (5.6.10)

180 FIRST AND SECOND ORDER EQUATIONS

In terms of the original variable z we have

1 Q"{z) 5 [Q>(z)Y 56
F(Z)~4[Q(z)]2 16 [Q (z )] 3' (5-6' U)
where now differentiation is with respect to z. An alternative expression is

F(Z) = -[Q(z)Yl/4£-2{[Q(z)Tm}. (5.6.12)

Compare (5.5.27).
As an illustration consider the case in which Q(z) is a polynomial of
degree ra, say,
Q(z) = a0zw+---. (5.6.13)
Then for |z|>jR

Z(z) = (aoyn l [1 + o(l)]. (5.6.14)

2 (m/2)+
Here R is the absolute value of the zero of Q(z) furthest away from the
origin. The remainder in (5.6.14) is an ordinary power series in 1/z without
a constant term if ra is odd, while if ra is even, ra = 2/c, the (k + l)th term
can be a constant times z~k 1 log z, the other terms being negative powers
of z. For simplicity it will be assumed here and in the following discussion
that no logarithms occur. Then Z(z) will have a pole at z = <» of order
k + 1 if ra = 2k, and a branch point if ra is odd. Furthermore,

F(Z) - - m(^+4)z"m"2[l + 0(z1)]. (5.6.15)

Now, if there is no logarithmic term in (5.6.14), we can invert the series
and expand z in fractional powers of Z. Set

u = [Km + 2)a0 1/2Z]2/(m+2). (5.6.16)

Then
z = m[1 + 0(m ')], (5.6.17)
and, finally,

F(Z) = - y++2? Z "[1 + 0(Z-2/(m+2))]. (5.6.18)

The important thing here is that Z2F(Z) is bounded at infinity and that
Z 2 is integrable out to infinity. This motivates a study of the solutions of
the DE (5.6.8) under suitable assumptions on F(Z), which are patterned
on the particular case just considered. Although not the most general that
could be handled, they are sufficient for the cases of interest here.
Hypothesis F. It is assumed that z^F(z) is an analytic function
such that:
ASYMPTOTICS IN THE PLANE 181

1 . F(z ) is holomorphic in a domain D extending to infinity and located

on the Riemann surface of log z. D contains a sector 5: -2tt < arg z <
2 77, R < |z |. For 0 such that - 2tt ^ 0 ^ 2tt we denote by D0 that part of S
where, if z G S, the ray z + re'6 is in S.
2. linwrF(reie) = 0.
3. For each z in De the integral

| |F(z +reie)|dr (5.6.19)

exists, and the set of all such integrals is bounded for all admissible values
of z and 0, the least upper bound being M.
This hypothesis holds if, e.g., F(z) is a rational function of z having its
poles inside the circle |z| = R and having a zero at infinity of at least the
second order. It is also satisfied by the function Z »-> F(Z) obtained in
(5.6.18) from (5.6.13) by the Liouville transformation.
The investigation proceeds by a number of steps.

THEOREM 5.6.1

Let w(z) be a solution not identically zero of the DE

w"(z) + [1 - F(z)]w(z) = 0, (5.6.20)

where F satisfies Hypothesis F and w(z) is defined in D0. Then there exists
a solution w0(z) of the sine equation

w"(z) + w(z) = 0 (5.6.21)

such that for all z = x + iy in D0

|w(x + iy) - w0(x + iy)| ^ M(y){exp [J |F(s + iy)| ds j- 1 }, (5.6.22)

where M(y) = maxs |w0(s + iy)|.
Proof. Choose a point x0> R on the real axis in D0. Let z be any point in
Do. There is then a 6, - < 6 < \tt, such that x0 and z both belong to De.
We can then join x0 and z by a path T composed of, at most, two straight
line segments, one parallel to the real axis, the other parallel to the ray
arg z = 6. See Figure 5.2. No matter how 6 is chosen subject to the stated
conditions, we have

j |F(s)||ds|=s2M,
and by condition F(3) this holds uniformly for z in D0.
 FIRST AND SECOND ORDER EQUATIONS
/

20 <
\ r

wk \

Figure 5.2

Next, there is a unique solution w,(z) of (5.6.21) such that

Wi(xo) = w(xo), wi(xo) = w'(xo). (5.6.23)

We have then, if T is oriented from x0 to z,

w(z)=wl(z) + j sin(z-t)F(t)w(t)dt. (5.6.24)

This is verified by twofold differentiation with respect to z combined with
the initial conditions (5.4.23). Let D0(c) be that part of D0 whose distance
from the real axis is at most c >0. Suppose now that |wi(z)|*sK,
|sin(r-z)|^K for t and z in D0(c). Then

|w(z)|-s|w,(z)|+ Jr |sin(r-z)||F(0||w(0||^|
or

|w(z)|^K+x|r |F(Olk(0||^|.

If now t is expressed in terms of arc length on T, it is seen that the

inequality is one to which Gronwall's lemma applies (Theorem 1.6.5), so
that
 183
ASYMPTOTICS IN THE PLANE

|w(z)|^K exp^KJ |F(0|Mf|J^K exp (2XM). (5.6.25)

Thus \w(z)\ is uniformly bounded in D0(c).

Let us now form

w(z)- j sin(r - z)F(t)w(t) dt = w0(z). (5.6.26)

Here the integral is taken along the line t = z + r, 0 «s r < <». The integral
exists by F(3) plus (5.6.25), which is valid as long as z is confined to D0(c).
It follows that w0(z) is defined and holomorphic in Do and bounded in any
D0(c). Twofold differentiation shows that w0(z) is a particular solution of
(5.5.21). Thus

w(z)=w0(z) + sin(t-z)F(t)w(t)dt. (5.6.27)

This is a singular Volterra integral equation which has a unique

solution, namely, the solution of (5.6.20), from which we started. Here we
have

\w(x + iy)-w0(x + iy)\ ^ j \F(s + iy)\ \w(s + iy)\ dy,

and this implies (5.6.22) by Gronwall's lemma.

The Volterra equation may be solved by the method of successive
approximations, and the estimate

\wn(z)-wn-l(z)\^M(y)^U \F(t)\dtJ (5.6.28)

shows that the approximations converge rapidly to w(z). H
In this manner we get a correspondence between the solutions of
(5.6.20), on the one hand, and those of (5.6.21), on the other. It is a
one-to-one correspondence which is type preserving. Now (5.6.21) has
solutions of three different types:

1. Solutions which go exponentially to zero in the upper half -plane,

type eiz.
2. Solutions which go exponentially to zero in the lower half-plane,
type e 'z.
3. Oscillatory solutions, type sin (z - z0).
Once a particular w0(z) has been chosen, it is to be expected that the
corresponding solution w(z) will inherit the asymptotic properties of
n>o(z), and this is indeed the case.
184 FIRST AND SECOND ORDER EQUATIONS

THEOREM 5.6.2

Equation (5.6.20) has a unique solution E+(z) asymptotic to eiz in the

sector - it < arg z < 2tt in the sense that

|E+(z)e ,z-l|^exp[J |F(t)| |dt|]- 1, (5.6.29)

where the path of integration is arg (t - z), equal to
0 for z£ Do, ^7r/or zGD^, and it for z E D,,.
We recall that, e.g., D^n is that part of S for which z E S implies that
the points z + ir E S for any positive r.

Proo/. Set
E+(z)eJz = U(z). (5.5.30)
1,
Here C/(z) satisfies the integral equation

U(z) = 1 + ^7 f [e2i(t"z)- l]F(t)l/(r) df, (5.6.31)

where the path of integration is arg (f - z) = 0. Since the kernel satisfies

^[e2i(t-z)-l]
Gronwall's lemma applies and gives

L/(z) - l|^exp
J |F(z + r)|dr]-l (5.6.32)
for all z in D0. This is the first inequality listed above. It shows that U(z)
is bounded in D0 since the integral cannot exceed M, by F(3).
For the next step it is necessary to swing the path of integration through
an angle of 90°. This is permitted for z E D0 H D^n. The integral taken
along a quarter-circle with center at z and radius p plus two radii, one
parallel to the real and the other to the imaginary axis, is obviously zero.
The integrand is bounded uniformly with respect to p and is o(p_1) as
p -»oo} by F(2). It follows that the integral along the curvilinear part tends
to zero as p & and

The resulting integral equation is valid for all z in D^!2. Since now the
kernel is <^ in absolute value along the new path of integration, we have
the preliminary inequality
r"
r-
ASYMPTOTICS IN THE PLANE 185

|[/(z)-l|«i£+' |F(t)||l/(0||<ft|.
and by Gronwall's lemma

|L7(2)-l|«exp[i£+' |F(r)||df|]-l. (5.6.33)

This is actually a sharper inequality than what is stated in the theorem for
this case.
Again we can turn the path of integration through an angle of 90°,
provided that z G D^n Pi D^. The discussion follows the same lines as
above, and for any z G D„ we have a new integral equation,

U(z) = 1+^7 jZ [e2Utz)-\]F(t)U(t)dt, (5.6.34)

valid for all z in Dw, and the resulting estimate

117(2) - 1| * exp [jZ \F(f)\ \dt |]- 1.

This is the last assertion. (5.6.35)
No further turns of the path of integration are permitted, since the
kernel of (5.6.31) is not bounded in the lower half-plane. This completes
the proof. I
The theorem asserts that

\U(z)-l\^eM - 1 (5.6.36)
for z in Do U D^n U and M is the bound postulated by F(3). This
estimate is rather poor if M is large, but it may be sharpened by trimming
the domain along the edges. This can obviously be done in such a manner
that M can be replaced by a smaller number. Set

I(z, 0) = J \F(z + reid)\dr, O^d^ir.

By suitable trimming we can obtain the fact that in the resulting region D +
min 7(z, 6) < log 2.

If this has been done, then for z£D +

\U(z)-\\<\ and l/(z)^0. (5.6.37)
The same method can be used to prove that

|E-(z)eiz-l|^exp[[ \F{t)\ \dt |]- 1 (5.6.38)

1S6 FIRST AND SECOND ORDER EQUATIONS

in the sector - 2tt < arg z < tt for suitable choice of the path of integra-
tion. Again, by trimming the sector along the edges, we may obtain a
domain D in which

\E ~{z)eiz - 1|< 1 and E'(z)^0.

This result will be referred to as THEOREM 5.6.3. The proof is left to the
reader.
The case in which w0(z) is oscillatory requires several steps.

THEOREM 5.6.4

Equation (5.6.20) has a solution w,(z) which is asymptotic to sin (z-z0),

z0 = x0 + iy0, in D0 so that

[cosh (y - yo)]"1 |w,(z) - sin (z - z0)| ^ exp ^ |F(z + s)| ds j- 1. (5.6.39)

There is also a solution w2(z), which is asymptotic to sin (z — z0) in D„ in
the sense that

[cosh (y - y0)] 1 |w2(z) - sin (z - z0)| ^ exp

j |F(z-s)|dsj-l. (5.6.40)
Normally w,(z)^ w2(z). Both solutions are oscillatory, wx in the right
half -plane, and w2 in the left. Furthermore, w,(z) may be continued
analytically into as well as into D-„, and the continuations may be
expected to be oscillatory in both regions. In particular, w,(z) is oscilla-
tory in Dw if y0 is sufficiently large positive, and in D ^ if y0 is sufficiently
large negative.
Remark. Here we say that a solution is oscillatory in a region which
extends to infinity if it has infinitely many zeros in the region. The next
theorem will provide estimates of the number of zeros.
For the proof we need the theorem of Rouche (Eugene Rouche,
1832-1910), which may be stated as follows:

LEMMA 5.6.1

Let the rectifiable simple closed curve C bound a simply connected

domain D. Let f , g, h be functions holomorphic in D U C such that (i)
h = f + g, and (ii) |f(z)| > |g(z)| everywhere on C. Then f(z) and h(z) have
the same number of zeros in D.
The proof is left to the reader.
Proof of Theorem 5.6.4. Formula (5.6.39) is read off directly from
ASYMPTOTICS IN THE PLANE 187

(5.6.22) since in the present case M(y) = cosh (y - y0). The existence and
properties of w2(z) follow from the analogue of Theorem 5.6.1 for the
region D^. A proof is obtained by replacing z by - z, which preserves the
form of the equation.
The analytic continuation of Wi(z) from D0 to the adjacent regions D„
and D „ clearly exists. Let us consider D^. By the analogue of Theorem
5.6.1 there exists a solution w0(z) of (5.6.21) to which the continuation is
asymptotic in DT. Here w0(z) may be expected to be oscillatory in D^.
Now Wo(z) cannot be a constant multiple of e'\ for this would make wx(z)
bounded in the upper half-plane, contradicting (5.6.39). We can exclude
the possibility that w0(z) is a constant multiple of e~,z if w,(z) is
oscillatory in the intersection of D0, and D+, and this would be the
case if y0 is large positive. Similar considerations hold for the continua-
tion of Wi(z) into and for the continuations of w2(z). H
We shall now discuss the zeros of

w,(z) = sin (z - z0) + R(z) (5.6.41)

in the right half-plane, where

\R(z)\ cosh (y - y0){exp ^ \F(z + s)| dsj - 1 J

for z G Do. Let 8 be given, 0 < 8 < \tt, and let H8 be a half-plane
Re (z) > 80, such that z£D0 and

exp[|jF(2 + s)|di]<1+-^A (5.6.42)

is satisfied. We assume that z0 and Hs are such that z0- 8 G Hs, while
z0 - 7r + 8 does not. We now introduce a set of square boxes £>„.«, where
|x - x0- nir\ < 8, \y - y0\ < 5, z0 = x0+/yo, (5.6.43)
and n =0, 1,2,
We aim to prove that each box contains one and only one zero of Wi(z)
and that there are no other zeros in the half-plane H8. On the boundary of
Q„,« we have |sin (z - z0)| ^ sin 5, the lower bound being reached for
z=Zo+rnr±8. On the other hand, the estimate for \R(z)\ on the
boundary reaches its maximum at the four vertices of the square, and
there \R(z)\ < sin 8. This shows that the conditions of Lemma 5.6.1 are
satisfied by

/(z) = sin(z-z0), g(z) = R(z)

on the perimeter of (?„,«. Now sin (z - z0) has a simple zero in the square,
so w,(z) has one and only one zero in each square.
188 FIRST AND SECOND ORDER EQUATIONS

Next, it is required to show that there are no zeros in H8 outside the

squares. We have

|sin (z - 2„)|2 = sin2 (x - x0) + sinh2 (y - y<>).

Thus, if |y - y0| > 6, we get |sin (z - z0)| > sinh |y - y0|, while

|K(z)|<cosh(y - y0) ^^8 < cosh (y - y0) tanh 8.

Since sinh |y - y0| > cosh (y - y0) tanh 8 for |y-y0|>5, it is seen that
|sin (z - z0)| > |-R(z)|, so that Wi(z) cannot be zero outside the strip
jy - y0| < 8 for z in H5. The rectangles between the squares remain for
examination. But here we have |sin (z - z0)| > sin 8 and \R(z)\ < sin 8 in
any one of the rectangles. Thus we have proved

THEOREM 5.6.5

The solution w,(z) is oscillatory in the half -plane Hs defined above. It has
one and only one zero in each of the squares Qn,s, n = 0, 1, 2, . . . , and no
zeros in Hs outside the squares.
Let us now return to our starting point, the DE (5.6.4), to which we
applied the Liouville transformation and obtained (5.6.8), where F(Z) is
given by any one of the three expressions (5.6. 10)— (5.6. 12). In the case
where Q(z) is a polynomial of degree m as in (5.6.13), we have to
scrutinize the conformal mapping defined by the Abelian integral
rz ~(m/2)+1

Z(z) = J [Q(s)Vnds=(aoY/2Z'Zm+2 [l+o(l)]. (5.6.44)

If m is odd, the remainder is an ordinary power series in 1/z ; if m is even,
m =2/c, there may be a term of the form z~kllogz. The function
zt->Z(z) maps partial neighborhoods of z = °° onto partial neighbor-
hoods of Z = oo. To apply the preceding theory we must determine what
regions in the z-plane correspond to approximate half-planes in the
Z-plane.
To illustrate let us consider the DE

w"(z) + (a4- z4)w(z) = 0, 0 < a. (5.6.45)

A singular boundary value problem was considered by Edward Charles
Titchmarsh (1899-1963) in 1946 for this equation. It involved subdomi-
nant solutions, and Titchmarsh conjectured that such a solution had only
real (finite in number) and purely imaginary (infinitely many) zeros. This
can be proved as follows (Hille, 1966). Here for \z\> a and up to an
additive constant,
ASYMPTOTICS IN THE PLANE 189

Figure 5.3. At the four points a, - a, ai, — ai the curve tangents make angles of 1 20° with each
other. The six curve bundles have asymptotes that make 30°, 90°, 150°, 210°, 270°, and 330°
angles with the horizontal.

"half-planes" corresponding to D, and D2 are indicated in Figure 5.4.

These two "half-planes" are connected along the real Z-axis from
la3B(§, J) to +oo, where B{p,q) is the Euler beta function. There is a
subdominant solution Ek(z) for each Dk, and the solution Et(z) is

E,(z) = z"1 exp Hz3)[l + o(l)] (5.6.48)

in Di. By Theorem 5.6.2 this solution has the same asymptotic form in the
three domains D6, Du and D2 and has only a finite number of zeros, if any,
190 FIRST AND SECOND ORDER EQUATIONS

Figure 5.4

in any right half-plane, x > b > 0. The situation is similar for the other
domains, Dk. Oscillatory solutions have their zeros normally in six strings
asymptotic to the separation curves of the Dk's. The number of zeros (all
simple) of absolute value in a string is (3tt)~V3[1 + o(l)]. Thus for a
solution which is dominant in all six regions we have

N(r, 0; w) = ^- r3[l + o(l)]. (5.6.49)

On the other hand, a subdominant solution normally lacks two strings,

those associated with the boundary lines of Dk, so that

N(r, 0; Ek) = r3[l + o (1)]. (5.6.50)

It should be noted that (5.6.45) is invariant under the substitution z - z.

The Titchmarsh functions Tm(z) are the solutions of the singular bound-
ary value problem

Tm (x )GL2(- oo, «>), m = 1,2,3,... (5.6.51)

and correspond to a sequence of spectral values a = am, where 0 < am <
am+x t 00 . The integrability condition shows that Tm(z) must be subdomi-
nant in Dx and in D4. Furthermore, Tm(z) is an even or odd function of z
according as m is an even or odd integer. The symmetry forces the two
remaining strings to lie on the imaginary axis outside the interval
(- iam, + iam). Problem 5.4:9 shows that Tm can have no zeros on the real
axis outside the interval (- am, + am ). That there are no zeros off the axes
is indicated in Problem 8.4:20. We have
ASYMPTOTICS IN THE PLANE 191

N(r, 0; Tm) = ^- r3[l + o(l)]. (5.6.52)

For Tm the value 0 has the defect |, while for an ordinary Ek the defect is
only s.

EXERCISE 5.6
1. Verify (5.6.6).
2. Verify (5.6.7)-(5.6. 12).
3. Verify (5.6.13)-(5.6.18).
4. Show that

z = x + iy, y * 0.

y = 0, x >0.

5. Discuss tf\z + rei0\ 2 dr.

6. Gronwall's lemma was invoked repeatedly in the proofs of Theorems 5.6.1
and 5.6.2. Show that it applies, and verify the inequalities.
7. If F(z) = z 2, how is the "trimming" to be done in (5.6.38) for the corres-
ponding solution? Here U(z) is a power of z times a Bessel-Hankel function
of order \Vl.
8. Prove the theorem of Rouche. [Hint: h = /(l + glf). If z describes C, where is
\+glf located?]
9. Verify the various inequalities between |sin(z - z0)| and |R(z)| used in the
proof of Theorems 5.6.4 and 5.6.5.
10. If Q(z) = z, determine the domains Dk. Show that the equation is invariant
under a positive rotation about the origin through an angle of 120°. What does
this imply for the solutions? If D\ is the sector containing the positive real
axis, how does a dominant (subdominant) solution behave in Di? Determine
N(r,0; w) and N(r,0; Ek). What is the defect of the value 0 for Ekl
11. Verify the formulas relating to the Titchmarsh equation. How does the beta
function enter the problem?
12. What is the asymptotic form of E4 in D4?
13. If k is a positive integer, the DE

w"(z) - [/cV 2 + k(k - l)zk_2]w(z) = 0

exhibits 2k sectors. Show that there is a solution for which w = 0 is a Picard
value.

14. Discuss Q{z) = am -zm. There are m +2 sectors and corresponding sub-
dominants. For m > 1 there may be values of a for which a subdominant
lacking more than two strings exists. If m is even, there is a Titchmarsh
problem.
192 FIRST AND SECOND ORDER EQUATIONS

5.7. ANALYTIC CONTINUATION; GROUP OF MONODROMY

Consider the DE
w"(z) + P(z) w'{z)+Q(z) w(z) = 0, (5.7.1)
where P and Q are single-valued analytic functions of z.

DEFINITION 5.7.1

// z = z0 is a point where P and Q are holomorphic, we define the

Mittag-Leffler star A(z0;P,Q) to be the set of all points z, in the finite
plane such that P(z) and Q(z) are holomorphic at all points of the straight
line segment joining z0 with zu including the end points.
Thus, if z2 lies on the boundary of A(z0; P, Q), either z2 is singular for at
least one of the functions P and Q or analytic continuation from z0 in the
direction of z2 encounters a singularity before reaching z2. As an example
take P(z) = (1 - z)~\ Q{z) = (1 + z)"2, z« = 0. Here A(0; P, (?) has as its
boundary the line segments (-°°, -1] and [+l,oo), and only the points
z = - 1 and z = + 1 are actual singularities.
Suppose now that w,(z), w2(z) is a fundamental system of (5.7.1) at
z = z0, and form the Wronskian
w,(z) h>2(z)
W(z \ Wi, u>2) (5.7.2)
w',(z) w'2(z)
which is different from zero at z = z0 and in some neighborhood thereof.
We introduce also the Wronskian matrix

V(z;^,^2)=[lv,(2) W,(2)J. (5.7.3)

This is an element of Wl2, the algebra of 2-by-2 matrices; it is nonsingular
at s = z0, and we shall see that it is nonsingular in A(z0; P, Q).
We shall need some additional notation and terminology. An n-by-n
matrix si(z) = (aik(z)) is an element of the matrix algebra 3ft„. We often
find it convenient to say the a matrix si has the property P if the elements
ajk have the property. We shall use this convention if P stands for
analyticity, continuity, differentiability, holomorphy, integrability,
measurability, or positivity.
We introduce a metric in Wln by defining a norm of si, by

:max[|aJ-I| + |fli2| + . • - + \ain\]. (5.7.4)

With this definition the norm of a product does not exceed the product of
the norms. Under this norm Wln is a complete metric space and is actually
ANALYTIC CONTINUATION; GROUP OF MONODROMY 193

a Banach algebra. The algebra has a unit element, the matrix (8ik ) of norm 1 .
A simple calculation shows that the Wronskian matrix satisfies
°W'{z) = ®{z)°W{z), (5.7.5)
where ^ is the matrix

(5.7.6)

with determinant Q(z). This could be expected from (5.2.1 1). Now 3F(z) is
holomorphic in A(z0; P, (?) by the definition of the star and the definition
of holomorphy for a matrix. It is to be expected that the property will
carry over to the solution of (5.7.5), and this is indeed the case.

THEOREM 5.7.1

// °W(z) is the solution of (5.7.5) with initial value °W(zo), a nonsingular

matrix, at z = z0, then the unique solution is holomorphic in A(z0; P, Q).
Proof. This follows from the argument used in Section 2.3, the method
of successive approximations, with minor changes. First let A0(z0) be a
star domain with respect to z = z0 contained in A(z0; P, Q) but bounded
and closed. Then there exist constants p and M such that

p(z)||^M, |z-Zo|^p, zGA0.

Set °lf(z0) = °W0, and define the successive approximations by

°Wn(z) = W0 + r ^(s)Wn^(s)ds. (5.7.7)

This makes sense in A0 and shows that the approximations are holomor-
phic in A0. Moreover, the usual estimates and the classical induction
argument show that

\\W„ (2)- V ..,(2)11* \\W 0||^j(pAf)",

so that the series

Wo + inMz)-^,^)]
1 (5-7.8)
converges uniformly in A0 to an element of 2ft2 which is holomorphic in
A0. Since clearly
J 20
<W(z) = W0+ P ^(s)°W(s)dst (5.7.9)

it is seen that °lV(z) is a solution of (5.7.5) satisfying the initial condition.

Uniqueness is proved in the usual manner. The extension to A(z ; P, Q) is
obvious. ■
194 FIRST AND SECOND ORDER EQUATIONS

The same method may be used to show that the system

T'(z) = -T(z)^(z), T(z0) = To (5.7.10)

has a unique solution holomorphic in A(z0; P, Q). We shall use this fact to
prove
THEOREM 5.7.2

The solution °W(z; z(>, °if0) of (5.7.5) is a regular (= nonsingular) element

of Wl2 iff °lf o is regular.
Proof The following argument is due to W. A. Coppel. Let us first take
WQ = %2, the unit element of Tl2. Then

W(z) = &(z)°W(z), °IV(zo) = «2,

T'(z) = -T(z)F(z), Y(zo) = %.
Here °lf(z) and T(z) are holomorphic in A(z0; P, 0), and
T(z) °T(z) + T'(z) °lf(z) s 0. (5.7. 1 1)
Now, if a matrix product has a derivative which is identically zero, the
product is a constant matrix and

y(z)°W(z)^%2
since this holds for z = z0. Thus °lf (z) has T(z) as left inverse everywhere
in A(z0; P, Q). On the other hand, since W(z0) = %2 the matrix W(z) must
have a two-sided inverse in some neighborhood of z = z0, and in this
neighborhood

[°lf(z)]-' = T(z), T(z) W(z) = °W (z) T(z) = ^2. (5.7.12)

Now the last relation is a functional equation satisfied by T(z) and °W(z),
and, by the law of permanency of functional equations, it must hold as
long as °W(z) and T(z) can be continued analytically along the same
path — in particular, everywhere in A(z0; P, Q). ■
If the initial matrix is °W0, a nonsingular matrix, we have
°W(z ; z0, Wo) = <W(z ; z0, %) *W0, (5.7. 13)
for both sides are solutions of (5.7.5) and have the same initial value at
z = z0. Here both factors on the right have inverses; hence the left
member does also.

COROLLARY 1

A fundamental system defined at z = z0 remains a fundamental system

everywhere in A(z0; P, Q).
ANALYTIC CONTINUATION; GROUP OF MONODROMY 195

This of course agrees with our previous remarks concerning the

Wronskian (determinant).

COROLLARY 2
For any choice of points z and z, in A(z0; P, Q)

W(z; z0, %) = W (z; z„ %2) W (z,; z0, ^2). (5.7.14)

This follows immediately from (5.7.13).
We can now tackle the problem of analytic continuation of the
solutions of the DF (5.7.5). It is clearly sufficient to solve this problem for
°W(z ; z0, %). We shall drop the subscript 2 and simply write % for the unit
element. Now let Zi E A(z0; P, Q), Zi ^ z0, and set °W{z\\ Zo,%) = Wi.
There is then a unique solution W(z ; Zi, %) which is defined in A(zi; P, Q)
and where

W(z ; z0, g) = W(z ; z„ «)<T, (5.7.15)

for z G A(z0; P, (?) fl A(z,; P, Q). The right member of this equation gives
the analytic continuation of W (z ; z0, %) in A(z,; P, Q). Given a finite set of
points {zk} with
zkGA(zk-,;P,Q),
we obtain a sequence of analytic continuations of the form

W(z-zk,%)°Wk^.-.°W2<Wu (5.7.16)
which represent the original solution in \(zk ; P, Q) for = 1, 2, . . . , m
for analytic continuation along the polygonal line with vertices at
Zo, Z\, . . . , Zm.
Suppose that z0G A(zm ; P, Q). Thus, after describing a closed polygon
with vertices z0, z,, . . . , zm, z0, we arrive at z = z0 with the value

which may very well be distinct from the original value %. Since each °Wk
is regular, so is the product. Consider now the set of all closed rectifiable
curves C beginning and ending at z = z0 along which °W(z ; z0, ^) may be
continued analytically. Here we may limit ourselves to polygonal lines,
which, however, may be self-intersecting. This is no restriction. On the
other hand, we shall assume that ^(z) has only a finite number of
singularities, poles or essential singularities. This is a restriction, but since
the most interesting case is that in which ^(z) is a rational function, i.e., P
and Q are rational scalar functions, the limitation is not serious.
C is to be given an orientation. If C is a simple closed polygon, the
Jordan curve theorem asserts that C separates the plane into an interior
of C and an exterior. These parts have C as a common boundary.
196 FIRST AND SECOND ORDER EQUATIONS

Moreover, any polygonal line which joins a point in the interior with a
point in the exterior must intersect C at least once — in any case, an odd
number of times. If z = a is a point not on C and if z describes C once,
arg(z - a) has an increment which is a multiple of 2ir, say, lirn. The
integer n is known as the index or winding number of C with respect to a.
Here n = 0 if a is exterior to C. If C is simple and if a lies in the interior
of C, then n = + 1 or - 1, depending on how we describe C. We say that
the orientation of C is positive if n= + \; otherwise, negative. To
distinguish, we denote the curve with the positive orientation by C and let
- C stand for the corresponding curve with the opposite orientation.
If C is self-intersecting, we use a parametric representation z = z(t),
0< t < 1, z(0) = z(l) = z0. Here z(t) is continuous and piecewise linear.
This induces an orientation of C in which z(f 0 precedes z(r2) if r, < r2. This
will be accepted as the positive orientation of C provided that it agrees with
the positive orientation of the least simple subpolygon of C that begins and
ends at C, where the positive orientation is defined by the index. If the two
disagree, we change the orientation of C.
Now with every curve C there is associated a transit matrix W(C) such
that analytic continuation along C carries

°lf (z ; z0, %) into °lf (z ; z0, %) °W (C), (5.7. 17)

while continuation along - C carries

°W{z\ z0, %) into °W (z; z0, %)[°W(C)Y\ (5.7.18)

The set of all these transit matrices forms a group ©, known as the
group of monodromy of the equation. It is clear that the unit matrix is an
element of @. It may be shown that, if a path C can be contracted to z0
without passing over a singularity, W(C) = %. Furthermore, for every
matrix °lf(C) there is an inverse matrix [¥(C)]"' = °W(- C). If we describe
first C, and then C2, °W(z ; z0, %) will be multiplied on the right by
<W(C2W(Cl), where order is essential. The number of elements of © is
countable (possibly finite) if the number of singularities is finite. With
each path, there is associated a continuum of equivalent paths T for which
°W(C) = °lf (O. Two paths are equivalent if C, - C2 bounds a region
containing no singular points. Thus the paths break up into a countable
(possibly finite) number of equivalence classes.
The assumption that the DE has only a finite number of singularities
implies that the group is finitely generated, i.e., °W (C) can be written as an
aggregate of positive and negative powers of the generators (at most one
generator for each singular point), where the order is essential and powers
ANALYTIC CONTINUATION; GROUP OF MONODROMY 197

of the same generator cannot be collected. Finally the group is indepen-

dent of the choice of z0. There is only one abstract monodromy group. The
various realizations are isomorphic.

EXERCISE 5.7

1. Determine the star A(0; P, Q) if P = (1 - z3)"1 and Q(z) = tanh z.

2. If norms are defined by (5.7.4), verify that the norm of a product is at most
equal to the product of the norms. Is ||,s4n|| necessarily equal to Ml"?
3. Let {dm} be a sequence of matrices in %Jln. What is the relation between
convergence in the sense of the metric and convergence of the entries am ,jtk ?
4. Consider the matrix DE (5.7.10). What second order linear DE do t>,(z) and
v2(z) satisfy?
5. Show that, if d is a matrix in 2ft„ and if \\si - %\\ = k < 1 , then si has a two-sided
inverse given by the convergent matrix series % + 27 {% - si)".
6. Try to justify the appeal to the law of permanency of functional equations
made in the proof of Theorem 5.7.2. The law is valid for scalar functions, but
here we are concerned with matrices.
7. Consider the equation

W(z) = d(\ - z)~] °W(z), where si =

A fundamental solution (algebraically regular) can be found as a linear

combination of (z - l)fl and (z - l)"fl with matrix coefficients. Find the latter
(not unique).
8. Show that the group of monodromy of the equation in Problem 7 has a single
generator. If a is an integer, the solution is single valued and the group
reduces to the identity (the unit element). If a is not an integer, the group is
cyclic (all elements are powers of the generator). The group is finite iff a is
rational. Verify.
9. In Problem 7 replace si by the - b
nilpotent amatrix

Find a solution matrix and the group. (A matrix is nilpotent if its powers are
the zero matrix from a certain point on.)
10. Suppose that (5.7.5) has singularities at z = 0, 1, °°, and nowhere else. Take
z0 = i, and lay loops around zero and one. Let °W(Cj) = °Whj = 1,2. Find a
path C such that

Draw a figure. Such double loops play an important role in the integration of
linear DE's by definite integrals. See the next chapter.
198 FIRST AND SECOND ORDER EQUATIONS

LITERATURE

This chapter corresponds to Appendix B of the author's LODE, sup-

plemented with various selections from Chapters 6, 7, and 9. All these
have extensive bibliographies to which the reader is referred. Also see, in
particular, Chapters 4 and 5 of:
Coddington, E. A. and Norman Levinson. Theory of Ordinary Differential Equations.
McGraw-Hill, New York, 1955.

Section 5.1 is partly based on pp. 66-90 of:

Schlesinger, Ludwig, Differentialgleichungen. 2nd ed. Sammlung Schubert, Vol. 13.
Goschen, Leipzig, 1904.

For Sections 5.2 and 5.3, pp. 91-121 of the same treatise served as
inspiration, together with pp. 108-138 of:
Bieberbach, Ludwig, Theorie der gewohnlichen Differentialgleichungen. Grundlehren, No.
66. Springer-Verlag, Berlin, 1953.
Some references for Section 5.4 are:
Koch, H. von. Un theoreme sur les integrates irregulieres des equations differentielles
lineaires et son application au probleme de Integration. Ark Mat., Astr., Fys., 13, No.
15 (1918), 18 pp.
Matell, M. Asymptotische Eigenschaften gewisser linearer Differentialgleichungen. Uppsala,
1924, 67 pp.
Perron, O. Uber einen Satz des Herrn Helge von Koch iiber die Integrale linearer
Differentialgleichungen. Math. Zeit., 3 (1919), 161-174.
Poschl, Klaus. Zur Frage des Maximalbetrages der Losungen linearer Differential-
gleichungen zweiter Ordnung mit Polynomkoeffizienten. Math. Ann., 125 (1953), 344-
349.
Valiron, G. Lectures on the General Theory of Integral Functions. Edouard Privat, Toulouse,
1923; Chelsea, New York, 1949.
Wiman, Anders. Uber den Zusammenhang zwischen dem Maximalbetrage einer analytis-
chen Funktion und dem grossten Glied der zugehorigen Taylorschen Reihe. Acta
Math., 37 (1914), 305-326.
Uber den Zusammenhang zwischen dem Maximalbetrage einer analytischen Funktion
und dem grossten Betrage bei gegebenem Argument der Funktion. Acta Math., 41
(1916), 1-28.
Wittich, Hans. Ganze transzendente Losungen algebraischer Differentialgleichungen. Gott.
Nachrichten, (1952), 277-288.
Neuere Untersuchungen iiber eindeutige analytische Funktionen. Ergebnisse d. Math,
N. S., No. 8. Springer-Verlag, Berlin, 1955; 2nd ed., 1968.
Zur Theorie linearer Differentialgleichungen im Komplexen. Ann. Acad. Sci. Fenn.,
Ser. A, I, Math., 379 (1966).
Section 5.5 is based essentially on LODE, Chapter 9. The results go
back to:
LITERATURE
199
Wiman, Anders. Uber die reellen Losungen der linearen Differentialgleichungen zweiter
Ordnung. Ark. Mat., Astr., Fys., 12, No. 14 (1917), 22 pp.
A survey of the field is given by:
Coppel, W. A. Asymptotic solutions of second order linear differential equations. MRC
Technical Summary Report, No. 555, March 1965.
Section 5.6 is based on Section 7.4 of LODE. It goes back to work done
by the author in 1920-1924. The use of Gronwall's lemma is a late
interpolation. For the Titchmarsh problem see Chapters VII and VIII of:
Titchmarsh, E. C. Eigen function Expansions Associated with Second Order Differential
Equations. Clarendon Press, Oxford, 1946. (Page 147 for the conjecture.)
Section 5.7 is related to LODE, Sections 6.1 and 6.2.
6

SPECIAL SECOND

ORDER LINEAR

DIFFERENTIAL

EQUATIONS

E. Kamke published in 1944 a magnificent monograph on DE's entitled

Differentialgleichungen, Losungsmethoden und Losungen. Of a total of
666 pages, he devoted over 200 to special second order linear equations
and listed some 450 equations. Our projected coverage is much more
modest. There are some equations which must be included in any
account. Others are mentioned because at some time or other the author
had occasion to study them; the reader who is not satisfied with the
selection here is referred to Kamke.
We include the hypergeometric equation, the equations of Bessel,
Laplace (as a sample of the confluent hypergeometric case), and
Legendre, the equations of elliptic and of parabolic cylinders, and some
odds and ends.

6.1. THE HYPERGEOMETRIC EQUATION

No other DE has such a venerable history as the hypergeometric
equation: Euler, Gauss, Kummer, and Riemann all contributed to the
theory, and fresh contributions are still being offered. It is an equation
that appears in many situations, and the theory has a high degree of
formal elegance. It is connected with conformal mapping, difference
equations, continued fractions, and automorphic functions, to mention
only a few.
Equations of the Fuchsian class with only two singularities lead to
200
THE HYPERGEOMETRIC EQUATION 201

trivial Euler equations. Three singularities lead to the hypergeometric

case. Here we have a choice between the normal form of Gauss:

z(\-z)w" + [c -(a +b + \)z]w'-abw = 0, (6.1.1)

where the singular points are 0, 1, a> and the solutions are expressible in
terms of the hypergeometric series,

F(a,b,c;z)=l+i(a.j"(g"2", (6.1.2)
where (p)„ = p (p + 1) (p +2) • • (p + n - 1), and the alternative: the
Riemann P-function,
a b c
A, A2 A3 z 2 (A, +/!*)= 1. (6.1.3)
Ml ^2 ^3
The P-function has three singular points, a, b, c ; the roots of the indicial
equations are A,, i±x at z = a, A2, pt2 at z = b, and A3, /i3 at z = c, with the
consistency relation requiring that the sum of the roots be one. The
coefficients of the equation are uniquely determined by these data. This
form of the equation dates from 1857. Like most of the things that
Riemann touched, this led to a Riemann problem: if the singular points
and the monodromy group are given, does there always exist a DE of the
Fuchsian class with these singularities and this group? Answers to this
question were given by G. D. Birkhoff (1884-1944), David Hilbert
(1862-1943), and Josef Plemelj (1873-1967). See Section 9.4.
We return to the normal type of Gauss. Although Gauss published an
exhaustive discussion of the hypergeometric series in 1812, his work on
the DE itself appeared only after his death.
The DE has three singular points at z = 0, 1, and oo with the correspond-
ing roots of the indicial equations

0,\-c;0,c-a-b;a9b. (6.1.4)
Note that logarithmic singularities are to be expected

at z = 0 if c is an integer,
at z = 1 if a + b - c is an integer, and
at z = oo if a -b is an integer.
At z = 0 and for c not equal to zero or a negative integer,
w0i(z) = F(a,b,c;z) (6.1.5)
is the holomorphic solution. The series converges for |z| < 1, i.e., up to the
nearest singularity. We shall see later what happens on the unit circle. The
202 SPECIAL SECOND ORDER EQUATIONS

Mittag-Leffler star A(0; wol) of this solution is the finite plane less the line
segment [ 1, °°). Analytic continuation is possible along any path that does
not pass through any of the singular points. A closed path, starting and
ending at z =0 and surrounding z = 1 , would normally lead to a determina-
tion of K>0,(z) which has a singularity at z = 0 because the second solution
at the origin has a singular point there. To find the second solution, we
may use the fact that there is a set of 24 transformations of the variables
which carry a hypergeometric equation into such an equation with a
different set of parameters. These 24 solutions were found in 1834 by
Eduard Kummer (1810-1893). One of these transformations is w = z' cu,
which gives a hypergeometric equation for u with parameters

fl-c + l, fc-c + 1, 2-c. (6.1.6)

This gives the second solution at z = 0 as

w()2(z) = zl ' F(a -c + \,b -c + l,2-c;z). (6.1.7)

Here c is not an integer, and the two solutions are obviously linearly
independent.
It is of some interest to detour at this point with a view to discussing
very briefly the corresponding matrix equation of the Fuchsian class with
singularities at z = 0, 1, <». It is

W(z)= [j + y^j^z), (6.1.8)

where sA and % are constant matrices. This equation is formally invariant
under the transformations of the (inharmonic group 21 with the six
elements
1 1 z z - 1(6.1.9)
z' 1-z' z-1'
Each of these transformations leads to an equation of type (6.1.8) though
the matrix "parameters" are changed according to fairly simple laws.
Actually the group 21 enters also into the transformation of the coefficient
"vector" (a 2-vector with matrix entries). The point of departure is a set
of six 2-by-2 matrices which form a group under matrix multiplication.
They are as follows:

h: :]■ *-[.: ~i\ h~; :]■ (6.1.10)

THE HYPERGEOMETRIC EQUATION 203

There are various identities:

^ = ^ = %y (<f$f = ($yf = %, (5<$(5 = yc5y. (6.1.1 1)

We shall also need a notion of conjugacy. If °U G 91, set
on* = $&>?rGVL3'&)?r. (6.1.12)
Then, if °1X £ 21 and

we define

^(z) = ^Td' H»J = Urf + iia} (6113)

Here a, b, c, d have values in the set - 1, 0, 1, while si, are the matrix
coefficients of (6.1.8). We then have

THEOREM 6.1.1

The transformations

s = 0U(z), [f] = vg] (6.1-14)

carry (6.1.8) into

T+rbr(s)- (6I15)
The proof is left to the reader.
Thus if we know one solution of (6.1.8) we know six. Unfortunately the
transformation theory of the classical DE is not quite that simple. Of the
six anharmonic transformations on the independent variable, only the
identity and s = 1 — z lead to hypergeometric equations. Thus the reflec-
tion s = 1 — z gives a new equation with parameters

a, b, a + b + 1 - c.
This gives a fundamental system at z = 1:
wn(z) = F(a, b9 a + b + 1 - c ; 1 - z), (6.1.16)

w,2(z) = (1 - z)c"a"F(c - a, c - fc, c + 1 - a - 1 - z). (6.1.17)

For the other solutions it is necessary to proceed as in (6. 1 .7) by setting
w =z\\-zYu (6.1.18)
and choosing 8 and e properly. We list the following cases:

1. Solutions at infinity: set s = 1/z, 8 = - a, e = 0. The new parameters

204 SPECIAL SECOND ORDER EQUATIONS

are
a, a - c + 1, a - b + 1,
and the solution

w.i(z) = z~aF^a, a - c + 1, a - b + (6.1.19)

Now (6.1.1) does not change if a and b are interchanged so that a second
solution at infinity is given by

w~,2(z) = z~bF(b, b - c + 1, b - a + 1; j). (6.1.20)

One of the solutions breaks down if a - b is an integer. If not, they are
obviously linearly independent and form a fundamental system at infinity.
These series converge for |z|>l, while the series (6.1.16) and (6.1.17)
converge for |z - 1| < 1. We can also get solutions converging for |z - 1| >
1. Of more interest are solutions converging in a half -plane, either
Re(z)<^ or Re(z)>i
2. Set 5 = (z - 1)1 z, 8 = - a, e = 0. The parameters are a, a - c + 1,
a + b + 1 - c, and the solutions are

w1,3(z) = z'aF^a, a-c + l,a + b + l- c; ^^)» (6.1.21)

w,,5(z) = z-bF^, fc-c + l,a + b + l- c; ^J^)- (6.1.22)

The numbering refers to the fact that both solutions are holomorphic at
2 = 1, where they have the value 1. This means that
whi(z) = w,,3(2) = w1>5(z), (6.1.23)

one of Kummer's 24 relations. The solutions with subscripts 1, 3 and 1, 5

give analytic continuation of wM(z) in the half -plane Re(z)>|. Similar
analytic continuations may be obtained for wU2(z) in the half -plane:

wUz) = zb-c(\-z)c-a'bF^c -b, \-b, 1 + c -a -b;1"^-)' (6.1.24)

w,,6(z) = za_c(l - z)c-a-bF(c -a,\-a,\ + c-a-b\ (6.1.25)

giving another Kummer relation, wU2 = wiA = Wi,6. For w0,i and w0,2 we
can get representations in the half -plane Re (z) < 5 in a similar manner.
Thus all the solutions of (6.1.1) are expressible in terms of
hypergeometric series in one of the six "anharmonic" z-variables multip-
lied by a power of z or/and a power of 1 - z. The parameters are linear
functions of a, b, c, and so are the exponents of the powers. It is clear that
THE HYPERGEOMETRIC EQUATION 205

the solutions are analytic functions of the parameters, as well as of z. This

implies further that in the transit formulas such as

uvi(z) = A w,,i(z) + B w,,2(z) (6.1.26)

the coefficients A, B are also analytic functions of a, b, c — in fact,
meromorphic functions, for we can differentiate (6.1.26) with respect to z
and set z = { in the result as well as in (6.1.26). We have now two linear
equations for A and B with locally holomorphic coefficients. Actually
this gives

A = %<»"■" Wi (6.1.27)
in terms of the Wronskians of the solutions evaluated at z = \. Now a
Wronskian can become zero iff the two solutions become linearly
dependent for some particular choice of the parameters; it becomes
infinite (since z = I is not a singular point) if one of the solutions becomes
identically infinity. This representation shows that A is a meromorphic
function of the parameters and tells us something about the zeros and
poles of A. Thus A has poles at the points c = 0, — 1, — 2, . . . because w0\
has poles at such points. Furthermore, c — a - b = 0, — 1,... are poles
because vvn and wi2 become linearly dependent. The positive integers
congruent to c —a — b drop out of consideration since wu(i) becomes
infinite. Finally, if c - a or c - b is zero or a negative integer, then w0i
and wi2 become linearly dependent, so these values should be zeros of A.
All these observations will be found to be correct once we get the explicit
formula for A. We note, however, that for Re (c - a - b) > 0 we have
A =lim
z — ► 1 F(a, b, c ;z)=f(a, b, c) (6.1.28)
as follows from (6.1.26), since u>,,(z)-* 1 and u>i2(z )-►(). This function of
three parameters enters into all the basic transit relations; hence the
finding of /(a, b, c) becomes paramount.
Let us observe in passing that the hypergeometric series (6.1.2) is
absolutely convergent for |z| < 1 for any choice of the parameters except
c equal to zero or a negative integer. The series is absolutely convergent
for |z | = 1 provided Re (c - a - b ) > 0. It converges but not absolutely for
|z| = 1, z 5* 1 , as long as - 1< Re (c - a - b ) ^ 0 and diverges on |z | = 1 if
Re(c -a -b)^~\.

LEMMA 6.1.1

For Re (c - a - b) > 0 we have

(6.1.29)
 SPECIAL SECOND ORDER EQUATIONS

Proof. We note that the gamma quotient on the right is the simplest
mcromorphic function having poles and zeros at the points indicated in
the discussion above, but this is no proof. We shall use a device which is
worked out in much more detail in Sections 7.3 and 7.4. The gamma
quotient suggests bringing in the Euler beta function:

I .-o-.r'*-™,

valid for Re(jc)>0, Re(y)>0. If now

K(n) j?anu\ W\<U (6.1.31)

then for |z|< 1

0(z) = j tx W-tY lf>(zt)dt

nX «>
anzn Jof tx+n \\-ty 1 dt

»> a I (X + V + H )
by uniform convergence. Here we set

x = a, y = c - a, g(u) = (\-a)
Since

we get =„?.m:"'

,r1, (1)„ I (c + n)
r(c -fl)r(u)
= j^— j F(a, fr, c ; z),
so that

Here the convergence of the integral requires that Re(«)>0,

Re (c - a ) > 0. If in addition Re (c* - a - b ) > 0, we can let z -> 1 from the
left, and the Lebesgue theorem on dominated convergence shows that the
resulting integral converges. Thus we get

/(«,/>,<•)= r(c) f'r \\-ty a " 1 at

He) rXfl)l(c -a -b)
r(c-a)T(a) r(c-b)
THE HYPERGEOMETRIC EQUATION 207

which is the required result. In (6.1.33) and the following formula we may
interchange a and b. ■
We have found the expression for A in (6.1.26). We can now also find
B. If in this relation z decreases to zero along the real axis, then w0,i -» 1,
h>i,i — » f(a, b, a + b - c + 1) while w,,2— »/(c-a, c - b, c-a-fr + 1),
and simplification gives

_ T(c)T(a + b -c)
B~ — r^Wib) — ' (6L34)
again a meromorphic function of a, b, c. This is rather important, for
A{ayb,c) and B(a,b,c) are defined everywhere in C3, certain lower
dimensional varieties excepted. We can then resort to the law of
permanency of functional equations to conclude that the expressions for
A and B are independent of the previously assumed restrictions on the
parameters.
The same method may be used to find the other transit relations. All
that one needs is convergent series expansions valid in a domain with two
singular points on the boundary. Here the half-plane solutions are useful.
A second method based on integral representations of one type or another
will be given in Chapter 7.

EXERCISE 6.1
Verify the following relations:
1. (dldz)F(a, b,c;z) = (ablc)F(a + 1, b + 1, c + 1; z).
2. (\-z)a =F(a,b,b;z).
3. arc sin z = zF(U J; z2).
4. arc tan z = zF( l,i, I; -z2).
5. lim
a -»oo F(a, 1, l;z/a) = <?2.
6. The elliptic normal integrals of Jacobi are

K= j\\-s2)m(\-k2s2rmds, iK' = J1/k(l-s2)1/2(l-/c2s2r1/2ds,

where k ^ - 1 , 0, + 1 , ». Show that
K =WF(i 1, 1 ; fc2), K ' = WF& -J, 1 ; 1 - fc2).
7. Prove the invariance of (6.1.8) under the anharmonic group.
8. Verify that the matrices (6.1.10) form a group under matrix multiplication,
and verify (6.1.9).
9. Prove Theorem 6.1.1.
10. Determine the set of self-conjugate matrices in the set (6.1.8).
11. The characteristic values of the anharmonic matrices are roots of unity.
What roots occur, and how are they reflected in (6.1.9)?
 SPECIAL SECOND ORDER EQUATIONS

12. Construct a fundamental system with convergent series expansions in

|z - 1| > 1. Express these solutions in terms of w„A and vvoo,2.
13. Verify (6.1.23) and (6.1.27).
14. If a = 1, b = 1, c =\, show that w02(z) = w,2(z) = w^(z) is an algebraic
function. Find this function. Are there any other algebraic solutions of the
equation?
15. With the parameters as in Problem 14, show that q{z) = w>02(z)/h>oi(z) maps
the upper z-half-plane on the interior of a curvilinear triangle (possibly
degenerate) bounded by straight lines and circular arcs. Describe the
triangle, and find the sides and the angles.
16. The equation w2 + 2w - z = 0 has a root that goes to zero with z. Express the
root as a hypergeometric function.
17. Show also that the equation w3 + 3w-z=0 is satisfied by
w =\zF&U;-k2),\z\<2.
18. What are the singularities of the function z^w(z) of the equation in
Problem 17, and what is their nature? Use a transit relation to express the
solution in a neighborhood of z = 2i.
19. Generalize to arbitrary trinomial equations.
20. Show that the hypergeometric equation may be written as

z(# + a)(xJ + b)w = tJ{xJ - 1 + c)w, xJ = z-f.

6.2. LEGENDRE'S EQUATION

This is the DE

(1 -z2) w"(z)-2zw'{z) + a{a + 1) w(z) = 0 (6.2.1)

or

Tz [(1" z2) S + a (a + 1)vv = °- (6'2-2)

It is named after the French mathematician Adrien Marie Legendre
(1752-1833), who considered the case a = n, an integer, in 1785. Here the
equation has a polynomial solution, the nth Legendre polynomial, P„(z).
The substitution u=\(\-z) reduces the equation to hypergeometric
form with parameters a + 1, - a, 1. The roots of the indicial equations are
0,0; 0,0: a + 1,-a.
Thus at both z = - 1 and + 1 logarithms must appear.
The holomorphic solution at z = + 1 is

Pa(z) = F[a + 1, -a, 1;|(1 - z)]. (6.2.3)

If a is an integer, a = n, the hypergeometric series breaks off and a
LEGENDRE'S EQUATION 209

polynomial of degree n results. The equation is unchanged if z is replaced

by - z and a + \ by - (a + I). In particular,

Pn(z) = (-\)nPn(-z).
The second solution at z = + 1, which involves a logarithm, may be
written as

p.,<z) - RE) lot bp - »>] + g (* l,)'?;}^'

as the reader may verify.

A rather useful analytic continuation of w0i(z), i.e., (6.2.3), is furnished
by the representation

Pa(z) = - z)r, aF(a + 1, - a, l;f^{) (6.2.5)

valid in the right half -plane, the analogue of Wi,3(z).
Formula (6.2.4) involves a pair of finite sums,

It is seen that each individual difference is 0(j~2), so that as n -><» each

sum converges to a finite limit when a is not an integer. It is not surprising
that the limits are expressible in terms of the logarithmic derivative of the
gamma function:

where C is the Euler-Mascheroni constant 0.57721 This gives

limD„(fl) = ^(a + l) + ¥(-a)-2C = D(a).

This leads ultimately to a transit formula

Pa(z) = sin (Tra )[D (a )Pfl (- z ) + Pa,,(- z )]. (6.2.8)

The solutions Pa(z) and Pa,i(z) are linearly independent and form a
fundamental system. At infinity we have the fundamental system

Qa(z) and Q-,-a(z), (6.2.9)

where

Q"<z> = rr^+VW'-°F(k +i,ia + l.fl +i;2-2). (6.2.10)

210 SPECIAL SECOND ORDER EQUATIONS

The transit formula from Pa to the Q's has the elegant form
Pa(z) = tan (ira)[Qa(z) - Q-x-a(z)l (6.2.1 1)
Here it is assumed that 2a is not an integer.
The rest of this section is devoted to the properties of Legendre
polynomials. The set {P„(z)} is an orthogonal system for the interval
(-1, 1), and
wn(t) = (n + -2),/2P„(0 (6.2.12)
is an orthonormal system. Moreover, this sytem is complete in
L2(- 1, + 1). This property may be defined in several different but equival-
ent ways: (i) the only element of L2(-l,l) whose Legendre-Fourier
coefficients

/„ = jy f(t)Pn(t)dt (6.2.13)
are zero for all n is the zero element; (ii) the finite linear combinations of
the Pn's are dense in L2(-l, 1); and (iii) the closure relation

n=0
i("+!)|M2= J-l
f \f(sfds (6.2.14)
holds for all / in L2(-l, 1).
If z is not real and located in (-1,1), then t y-*(z-t)~x belongs to
L2(-l, 1), and the corresponding Legendre-Fourier coefficients,

are obviously analytic functions of z. That the integral equals Q„(z) was
proved by F. Neumann (1848). We have then

-J-=
Z 2 {In + \)Pn(t)Qn(z),
l n=0 (6.2.16)

to start with, in the sense of L2-convergence. But the terms of the series
tend to zero so rapidly that there is absolute convergence not merely for
— 1 2* t ^ + 1 but also in the complex plane for t inside an ellipse with foci
at - 1 and + 1 and passing through z. Moreover, any function holomorphic
inside such an ellipse may be represented by a series of Legendre
polynomials absolutely convergent inside the ellipse.

EXERCISE 6.2
1. Prove that
BESSEL'S EQUATION 211

2. Prove the orthogonality relations

J' Pm(t)Pn(t)dt = 8mn(n+irl.

3. If y(t) is a real-valued solution of (6.2.1), prove that, if
F(0 = (1 - t2)[y'(t)]2 + a(a + \)[y(t)]\ y(l) = 1,
then - 1< y (0 < + 1 on (0, 1). In particular, this holds for a = n, y(r) = P„(f).
4. Show that Pn{t) has only real simple zeros all in (-1, 1).
5. Show that Pa(z) >0 for 1< x < oo, a real.
6. Let a =-| + /ci, and consider P i/2+k.(z). These are the conical harmonics
(German: Kegelfunktionen) of E. H. Heine (of Heine-Borel theorem fame).
Show that these functions are real for z = t > — 1 and positive on (-1, 1).
They arise in potential-theoretical problems involving cones.
7. Let the z-plane be cut along the real axis from -1 to +1, and set
u = z + (z2 - l)l/2, where the square root is real positive when z is real and > 1.
Show that the curve |u| = r > 1 is the ellipse passing through the point z and
having its foci at - 1 and + 1. Show that

limjup|Pn(z)|1/n =|k|.
8. Use these results to discuss the convergence of the series (6.2.16).
9. Let f(t) be holomorphic in an ellipse with foci at - 1, +1. Verify the stated
convergence properties of the Legendre series of f(t). Show that on the
ellipse of convergence there is always a singular point.
10. Verify (6.2.4).
11. Verify (6.2.5).
12. Since Re(c - a -b) = 0 for Legendre's equation, Pa(x) cannot tend to a
finite limit as x decreases to — 1. Instead it becomes infinite as a multiple of
log[i(l -x)J. Determine the multiple and derive (6.2.8).
13. Present a similar discussion at z = oo with a view toward finding (6.2.11).
14. For fixed z the Legendre function Pa(z) satisfies a linear second order
difference equation with respect to a :
(a + l)Pa+1(z)-(2a + l)zPa(z) + aPflI(z) = 0.
Verify for integral values of a, knowing that P0(z)= 1, Pi(z) = z.
15. All polynomials P2r. + i(z) evidently have z = 0 as a root. Show that no number
can be a common root of P„(z) and P„+)(z), and that, if z = z0 is a common
root of P„(z) and P„+2(z), then z() = 0.

6.3. BESSEL'S EQUATION

Few scientists have been able to contribute to as many fields as Gauss did,
but it was as an astronomer that he made his living and the few pupils that
212 SPECIAL SECOND ORDER EQUATIONS

he had were trained as astronomers and became good ones. But it was no
accident that they also gained fame for mathematical contributions.
August Ferdinand Mobius (1790-1868) is a case in point: one volume of
his collected works is on astronomy, the remaining three deal with
mathematics. Friedrich Wilhelm Bessel (1784-1848) is another. In 1824 he
studied the perturbations in the orbits of the planets and in this connec-
tion developed the theory of the DE
z2w" + zw' + (z2-a2)w = 0, (6.3.1)
the parameter a being zero or a positive integer. G. Schlomilch (1857) and
R. Lipschitz (1859) proposed to name the equation after Bessel. Special
instances of (6.3.1), however, were known long before Bessel. In this
connection we can mention the contributions of Jacob Bernoulli (1703),
Daniel Bernoulli (1725, 1732), L. Euler (1764), J. L. Lagrange (1769), and
J. Fourier (1822). These were investigations into applied mathematics,
involving such varied topics as oscillations of heavy chains, vibrations of
circular membranes, elliptic motion, and heat conduction in cylinders.
There is no doubt that Bessel functions have enjoyed wide popularity
and been of great use. During the 250 years that have elapsed since the
start, the literature has grown enormously. G. N. Watson's Theory of
Bessel Functions, the modern standard treatise on the subject, has 36
pages of bibliography.
Equation (6.3.1) has a regular- singular point at z - 0 with roots of the
indicial equation equal to a and -a. The point at infinity is irregular
singular of grade 1 = 1 + sup (- 1, 0).
The roots of the indicial equation differ by 2a, and as long as 2a is not
an integer, the two functions Ja(z) and J-a(z) form a fundamental system
with

J°(2> = gor(fc + i)V(l)+fc + i)(l) • <6-3-2>

Here the series converge for 0 < |z| < oo.
If a=n+^, the roots of the indicial equation differ by 2n + 1 and
logarithmic terms could be expected but are actually missing. For n = 0
we have
/ 2 \1/2 / 2 \1/2
/i/2(z)=^— J sin 2, J-m(z) = (~) cos z (6.3.3)
with no logarithms. It may be shown that

Jn+m(z) = An (2) cos 2 + Bn(z) sin 2, (6.3.4)

where An and Bn are polynomials in z~m.

On the other hand, if a is zero or an integer, logarithms do appear, and
 213
BESSEL'S EQUATION

in addition to the Bessel function Ja(z) of the first kind we have also to
consider Bessel functions of the second kind. Various definitions are
commonly used; they differ only by constant multiples of Jn(z). If a is not
an integer, set

Ya(z) = cosec (7ra)[Ja(z) cos (ira) - J-a(z)]. (6.3.5)

This is obviously a solution of (6.3.1). We now let a -» n and find that a
limit exists, denoted by Y„(z). It is a Bessel function of the second kind of
order n. A tedious calculation gives

+ % i0kv. * >ni2 io* (63-6)

where C, as usual, is the Euler constant.
The function (a, z)»-> Ja(z) satisfies a number of functional equations
with respect to a and/or z. Note, in particular, the second order difference
equation
zJa+l(z) - 2aJa(z) + zJa-x{z) = 0. (6.3.7)
We also note
J'a(z) = iUa-xiz) - J. + ,(Z)]. (6.3.8)
The difference equation shows that the sequence {Jn(x)} for x real >0
forms a so-called Sturmian chain; i.e., if there is an n and an x0 > 0 such
that /„ (jc0) = 0, then Jn-i(x0)Jn+i(xo) < 0. In particular, two consecutive /„ 's
cannot have a positive root in common.
Equation (6.3.1) is formally real if z is purely imaginary and a is real,
for if we set z = iy, 0 < y, the result is

2 d2W , dW , 2 , 2x n t £ <vv
y -^2 + y-fi-iy +«)^=o. (6.3.9)
For a real, two real solutions of this equation are L(y) and I-a(y), where

(6-3-10)
w-Znk + ma(-vY+2+k k + iy
and this differs from Ja(iy) by a constant multiplier. These solutions form
a fundamental system for (6.3.9) unless a equals zero or a positive
integer, in which case logarithms enter and a second solution may be
found along the lines indicated above.
There is a Laurent series expansion

exp [k^ -jjj = 2 h{z)t\ (6.3.11)

214 SPECIAL SECOND ORDER EQUATIONS

Here the left member may be considered as a generating function of the

system {/„(z)}. The left member is the product of exp {hzt) and
exp(-|zt '). To derive (6.3.11) one expands the factors in exponential
series, multiplies termwise, and rearranges according to powers of r. The
expansion was found by Schlomilch and implies a number of important
relations and inequalities.
Take first z = iy, t = - i, which gives

ey=/o(y) + 2^/„(y)
n=1 (6.3.12)

convergent for all y. If y > 0, all terms are positive and we get
h(y)<e\ L(y)<\e\ (6.3.13)

If y < 0, replace y by |y | in the inequalities. Next take t = e 10 to obtain

eizsine =2 Mz)eine (6.3.14)

convergent for all z and real 0. It follows that X. (z) is a Fourier coefficient,

Z77" J tt
Jn(z) = ^- P eiUsin9-nd)de. (6.3.15)

Here we can take the "formally real part" to obtain

TT Jo
Jn(z) = —{ cos(z sin0 -n$)dd, (6.3.16)

which happens to be Bessel's definition of /„(z). Note that the "formally

imaginary part" is zero.
Now, if z = x > 0, it is seen that
|/„U)|<1, Vjc,V«. (6.3.17)

Next observe that the function 0 »-> eiz sin0 G L2(- tt, tt), so that for
z = x + iy one obtains

J f TT
\eizsine\2dd = JTTT e-2ysinedS =27r/0(2iy) = 27rIo(2y).

The closure relation for the orthogonal system {enie} then gives

M2y) = i|X,(x + /y)|2. (6.3.18)

This yields

\Mx + iy)| ^ e|y|, \Jn(x + iy)| ^ 2 1/2 eM, 0<n. (6.3.19)

Note that (6.3.18) suggests the identity
BESSEL'S EQUATION 215

i
[/o(z)]2 + 2 2[/„(z)]2 = l (6.3.20)

for all z. By (6.3.18) the relation is true for z real, y = 0, and the series
(6.3.20) is a uniformly convergent series of holomorphic functions, so the
sum is holomorphic. Now, if a holomorphic function is identically one on
the real axis, it is identically one in its domain of existence — in this case,
the finite plane.
Integral representations of Bessel functions will be considered later.
Here we have only to mention Bessel functions of the third kind,
introduced by Hermann Hankel (1839-1873), a brilliant mathematician
and classical scholar, in 1869. Besides the Bessel functions, his name is
attached to an integral for the reciprocal of the gamma function and a
class of determinants. He codified the field postulates and wrote a history
of early mathematics; most of this work was published after his untimely
death.
To get the proper background for the Bessel-Hankel functions we have
to go back to Section 5.6. The substitution

w{z) = z~mW{z) (6.3.21)

takes (6.3.1) into

W"(z) W(z) = 0, (6.3.22)

an equation which satisfies Hypothesis F. From the discussion in Section

5.6 we know of the existence of a solution E«(z), which is asymptotic to
eiz in the sector - tt < arg z < 2tt, and a solution E~a{z) asymptotic to e~,z
in the sector - 2tt < arg z <tt. These solutions correspond to the Bessel-
Hankel functions Ha\z) and H(a2)(z); we have

H(l\z) = Cx{a)z~mEl{z\ m\z) = C2{a)z~mE-{z\ (6.3.23)

and the constants are chosen so that

Ja(z) = \[H«\z) + Hfl (6.3.24)

Thus Ja(z) must be an oscillatory solution of (6.3.1), and it may be shown
that

Ja(z)={~^) cos tz ~^(« +^)] + 0(|zT3/2). (6.3.25)

EXERCISE 6.3

The first four problems deal with the expansion of the Cauchy kernel (t - z)~l in
terms of Bessel functions Jn(z). This is in analogy with (6.2.16) and one of the
many expansions of this kind.
216 SPECIAL SECOND ORDER EQUATIONS

1. Prove that

ezu = Uz ) + 2 h (z ){[u + (u 2 + I)1'2]" + [u - (u 2 + l)1/2]n },

and show that the series is absolutely convergent for all z and u. Prove that
the coefficient of J„(z) is actually a polynomial in u of degree n.
2. For Re (0 > 0 show that

On(t) = \ [e 'u{[u +(u2+ l),/2]fl +[u+(u2+ 1)1/2]" } du

exists as a polynomial of degree n in 1/r.
3. Show that the series
Oo(0/o(z) + 2 n2= \ On(t)Jn(z)

converges absolutely for |z| < |r| and uniformly for |z| =s a < b ^ \t\.
4. Multiply both sides of the identity in Problem 1 by e~,u, and integrate with
respect to u from zero to infinity. By Problems 2 and 3 the right member
equals the series in Problem 3, so that

—I —— Z
= O0(t)Mz ) + 2 n2= l On (t)Jn (z ).

Show that the procedure is justified, say for 0 < z < t, and then extend it to
0< |z| < |r| by analytic continuation.
5. Carl Neumann found in 1867 the expansion

f(z) = 2ajn(z), an =^-jc f(t)On(t)dt,

where z ^/(z) is holomorphic in a disk |z| < r, and C is the circle \t\ = p <r.
Verify.
6. Otto Schlomilch in 1857 studied expansions in terms of Bessel functions, of
which the following is typical:

20 anJo(nx), 0 < jc < 7r.

Show that the series converges if \an\ is bounded and, further, the series
27 n~m\an - an+i\ converges. When does the series converge for x = 0?
7. Given (6.3.25), where the remainder is <C(a)x~m, find intervals on the
positive x-axis in each of which Jn(x) has a single zero and outside the union
of which there are at most a finite number of zeros.

6.4. LAPLACE'S EQUATION

Pierre Simon Laplace (1749-1827) in his treatise Theorie analytique des
probabilities of 1817 considered a DE which carries his name and may be
LAPLACE'S EQUATION 217

written as

(a2 + b2x) ^4 + (a, + bxx) ^ + (a0 + b0x)w = 0, (6.4.1)

where the a's and b's are given numbers. We set

a2+b2x=z (6.4.2)
and obtain

Dz(w) = zw"(z) + (Co + cz)w'(z) + (do + d,z) w(z) = 0. (6.4.3)

This equation has z = 0 and z = o° as singular points, the former regular
singular, the latter irregular singular. The indicial equation at the origin is

r(r - 1) + c0r =0, r, = 0, r2 = 1 - c0. (6.4.4)

If Co is not an integer, one solution is an entire function of z, while the
other is zl c° times an entire function. Omitting some trivial cases, we find
that the point at infinity is irregular singular with grade
g = 1 + max (0,0)= 1. (6.4.5)
It is desired to study this DE for large values of z. More precisely, we
want asymptotic representations of the solutions in sectors of the plane,
and the method developed in Section 5.6 is not satisfactory in this fringe
case. Laplace was faced with this problem, and he invented what was
later known as the Laplace transform to cope with the difficulty.
Actually, a number of integrals go under this name. They are of the form

j ezsF(s)ds, (6.4.6)
where the weight function F(s) and the path of integration C have to be
chosen so that (6.4.6) becomes a solution of (6.4.2). Assuming the right to
apply the differential operator Dz under the sign of integration, we get

where DWc 6 "F(5)d5] = / Dz(ezs)F(s) ds, (6.4.7)

Dz(ezs) = (zs2 + c0s + d0+ dxz)ezs

ds
= (s2 + c,5 + d,)-^-(c2S) + (do+c0s)ezs.
Substitution in (6.4.7) and integration by parts leads to

[(s2 + c,5 +d,)F(s)e2S]c- £ {(s2 + c,s +di)F'(s)

- [do - d + (Co - 2)s]F(s)}ezs ds. (6.4.8)
218 SPECIAL SECOND ORDER EQUATIONS

This will reduce to zero, provided that two conditions are met:
1. F is chosen as a solution of

(s2 + clS + d,)F'(s) = -[C - d0 + (2- c0)s]F(s). (6.4.9)

2. The path of integration C can be chosen so that

[(s2 + c,s+d,)F(s)ezs]c =0. (6.4.10)

The first condition is sufficiently clear, but it leads to two distinct cases
according as the equation

52 + c,5 +d, = 0 (6.4.11)

has (i) distinct or (ii) equal roots.
In case (i) suppose that the roots are Si and s2. Then

F(s) = (s-sOa(s-s2f (6.4.12)

for a definite choice of a and /3.
In case (ii) suppose that the double root is s0 = -\c\. Then

F(s) = (5 - SoT exp (- j^), (6.4.13)

again for some choice of the parameters.
Condition 2 leads to various possibilities. The integrated part should
vanish in the limit. If the roots are distinct, three different types of paths
are available:
1. Arcs joining two singular points.
2. Simple loops starting at one singularity, surrounding another
singularity, and returning to the starting points.
3. A double loop surrounding S\ and s2.

Using type 1, we have three choices: (l:i) from s = Si to s = s2, (l:ii)

from s i to <», and (l:iii) from s2 to <». Here (l:i) requires that Re (a) > - 1,
Re(/3)>-l. If this condition is satisfied, the integrated part,
G(s,z) = (s -s,)a+,(s -s2f +VS, (6.4.14)
vanishes at both ends of C, and we have the solution

wo(z) = j Si)a(s - s2)V2 ds, (6.5.15)

which is an entire function of z. To use choice (l:ii) we must still have
Re (a) > - 1, and we must approach the point at infinity along a line such
that Re (zs ) -> - <». For a given z ^ 0 such a direction can always be found.
LAPLACE'S EQUATION 219

For such a choice, G(s, z) will again vanish at the end points and a
solution is obtained. The situation is similar for choice (l:iii).
A simple loop from infinity and surrounding s = sx is shown in Figure
6.1. It is supposed that the line arg (s - sx) = coi is permissible at infinity.

Figure 6.]

The loop starts at infinity, follows one edge of the line, makes a 8 -circuit
in the positive sense about s = Si, and retreats to infinity along the other
edge. The asymptotic representation of this integral at infinity will be
derived.
A double loop is shown in Figure 6.2. When s describes C, it surrounds
each of the points Si and s2 twice but in the opposite direction, so that
G(s, z) returns to its original value and the integrated part vanishes.

Figure 6.2
220 SPECIAL SECOND ORDER EQUATIONS

In case (ii) of condition 1 the weight function F(s) is given by (6.4.13).

There is a straight line through s = s0 along which /3(s - s0) is purely
imaginary, and the exponential function is bounded to one side of the line
and unbounded to the other. There are essentially two different paths at
our disposal, both shown in Figure 6.3. One starts and ends at 5 = s0,
where it has a cusp. It is a closed path which passes from the half-plane

Figure 6.3

where the exponential function is bounded, through the half-plane where

it is unbounded, and then back again. This solution is an entire function.
The second choice starts at s = s0 in the bounded half-plane and proceeds
to infinity in a permissible direction. We shall not pay any further
attention to case (ii).
Returning to case (i) and the double loop integral around s, and s2, we
set s2- Si = j and s = Si + jr. Then
eiztta(l-tf dt,
W0(Z) = j (6.4.16)
L
xes\z
where C0 is a double loop around zero and one. Hence

(6.4.17)

Here the double loop integral

■ exists for all values of a and (3. It is then
r
enough if we evaluate the integrals, say for Re (a ) > - 1 , Re (/3 ) > - 1 , and
use analytic continuation. Under the stated conditions the integrals may
be explicitly evaluated, and if
ta+k(\
p = r(a + k + i)r(j3 + i) (6.4.18)
; r(a +j3 +fc +2)
LAPLACE'S EQUATION 221

the corresponding double loop integral is

(1 - e-2ma)(\ -e~2^)Jk = J°k. (6.4.19)

To see this, suppose that we start the integration at t = 8 with a value of
the integrand which is real positive if a and /3 are real positive. Here
0 < 8 < 1 - 8 < 1. Proceed along the real axis to r = 1 — 5, and there make
a circuit in the negative sense around t = 1. The new determination of the
integrand at t = 1 — 8 equals e ~2m(3 times the old determination. Now
proceed on the real axis to t = 8, and there make a negative circuit around
t = 0. Return to t = 8 with e~2^a+^ times the initial determination. Now
proceed along the real axis to t = 1 — 8, followed by a positive circuit
about t = 1. The new determination at t = 1 - 5 is e~2ma times the value
encountered the first time when this point was reached. Proceed to t = 5,
and follow with a positive circuit around t = 0. This means returning to
the starting point and the original value of the integrand. It follows that
+ o(6),

Jl = [1 - e-2lrip+ e-2^a+P) - e~2lTia] j

where the remainder represents the four circuit integrals. Letting 8
we get (6.4.19). Draw a figure and follow the steps! It follows that

wo(z) = (1 - e-2"ia)(\ - e-2^)T(p + l)r+/3+ V>z

w^ r(a+fc + l) (jz)k
\?or(a+/3+/c+2)TT- (6-4-2°)
At this stage we introduce the Pochhammer-Barnes notation,

pF£?(a1, a2, . . . , Op ; bi9 b2, . . . , bq ; z) = ^ (6.4.21)

With this notation we have

vvo(z)-j (1 e )(1 e
) r(a+j8+2)
XeviFi(a + l;a + j8 +2; jz), (6.4.22)
where j = s2 — Si. If one of the quantities a, /3, a + j3 + 1 is a negative
integer, the right-hand side is to be replaced by its limit as the parameter
approaches the critical value. Formula (6.4.22) holds also for (6.4.15)
provided that the factor (1 - e'2iria)(\ - e~2^) is omitted.
We shall find the asymptotic behavior of this solution after that of the
simple loop integral from infinity to one of the points 5 = Si or s = s2 has
been found.
At this stage we need Hankel's integral (1864) for the reciprocal gamma
222 SPECIAL SECOND ORDER EQUATIONS

function:

^ = ^1 euuadu. (6.4.23)

Cut the M-plane along the negative real axis, and define u~a =e"a,og"?
where the imaginary part of log u lies between - tt and + it. The contour
of integration is a simple loop surrounding the negative real axis. It
follows the lower edge of the cut from — °° to — 5, 0 < 5, a positive 6-circuit
around the origin, and then along the upper edge of the cut back to — <».
The integral exists for every finite value of a, and using, for instance, the
theorem of Morera one proves that the integral defines an entire function
of a. That this function is 1/T(a) is proved along lines now familiar to the
reader.
Let us consider

Wl(z) = ezs(s-s])a(s-s2fds, (6.4.24)

where L is a simple loop to be defined. Mark a ray from s = Si to infinity

such that the ray does not pass through s = s2 and along the ray
Re (zs)-» — oo as s -»<». L is now a simple closed loop surrounding the ray
in the positive sense. Start at infinity with the value 0, and follow the
appropriate edge of the ray until a point is reached at a distance 8 from si,
where 8 < Make a positive circuit around s = Si, and return to infinity
along the other edge of the cut. As above, set s = Si +jt. This maps the
loop L onto a loop Li and
A,
w,(z) = j^V'2 f eiztta (1 - tf dt. (6.4.25)

The loop Li surrounds t = 0, it does not pass through t = 1, and the

integral out to infinity exists. Expand (1 — 0" m binomial series

k=o Kl
This series converges for \t\ < 1, but for the time being we disregard this
restriction and consider the formal series,

k=0 K\ J Li eizttk+adt
£t-£M

Here we make the change of variable jzt = u. The resulting integral in the
fcth term of the formal series becomes

(jzya k ljc euuk+adu,

LAPLACE'S EQUATION 223

where C is the loop occurring in the Hankel integral (6.4.23). The value of
the integral then is

Ti
1 (— 2™
k — a), = (- sin (7rQ:)r(a + k + 1),
where we have used the symmetry formula for the gamma function,

r(x)T(\-x) = Sin -,
-rJL7TX
and the periodicity properties of sin ttx. Furthermore,

r(a + l + /c) = r(a + l)(a + l)k.

This gives finally

w,(z) = 2/ sin (Tra)r(a + l)j*+Va~V'z 2F0(<* + 1, - j3 ;-j^j. (6.4.26)

The series naturally diverges, but it is an asymptotic representation in the
sense of H. Poincare (1886). He defined a formal power series to be
asymptotic to F(u):

0
F(M)~ianH" (6.4.27)
for approach to the origin along a ray or in a sector if, for all rc,

lim u ~n [f(u) - 2 M "]= 0. (6.4.28)

To prove that (6.4.26) has such asymptotic properties it is necessary to
go back to the point where divergent series entered the argument. We can
always write

{\-tf = k^{^P^tk
=o K\ +Rn{t). (6.4.29)
From the obvious relation

(1-^-1 = -0 Jf(i-,y-' ds
we can obtain an expression for Rn(t) by repeated integration by parts.
The result may be written as

n ! Jo
Rn(t) = n f (t - 5)n(l -sf"-' ds

n !
|o (1 - /•)"(! - rt)" "'dr.
224 SPECIAL SECOND ORDER EQUATIONS

Now, if t = Re'6, then 0^0 since Li does not go through t = 1. Further-

more,
min |1 - rt \ = |sin 0|,
and there exists an M independent of n such that

|| (l-r)n(\-rtf~nl dr <M\cosec6\n+l+w j (\-r)ndr.

Hence

Rn(t) = tJ*)ii tn+lPn(t) and \Pn(t)\ ^ M|cosec e\n+l+^. (6.4.30)

In the integral (6.5.25) we now replace (1 - tf by (6.4.29). Here the

finite sum gives an expression involving a partial sum of the 2F0-series, so
we have to estimate the remainder using (6.4.30). Here

|| eizttaRn(t) dt\^~^n\ || eizttn+1+aPn(t)dt

M|cosec 6\n+l+w || eizttn+l+a dt\

Here we can choose Lx so that jzt becomes real negative. Thus jzt =
-\jz\ \t\ and Ir"! < A|f|'a|. The integral is then dominated by
|-n-2-|a|
2A Jo exp[-|;2||t|]|tr+,+w|dt| = 2Ar(n+2 + |«|)|/Z|-
This gives

< M,|cosec 6 \n+l+lpl ^ T(\p |+ n )

|£ eiztCRn(t) dt
Set xr(|a| + n +2)|/2 r-2Ha|. (6.4.31)

2i sin (aiT)T(a + l)f = K. (6.4.32)

Then

|K-Xjz
I)-a-1e^2w1(2)-2^(a
o K! + l)k(-/3)k(-j2)k|

< M,|cosec 6 1" ^ r(|i8 1+ n )T(\a |+ n + 2)\jz \—2-M. (6.4.33)

After multiplication by \zj\n the right-hand member still goes to zero as
z ->oo? so the 2F,-series is indeed asymptotic to X_1(j2^0,", e~s,2u>i(z). It is
seen that w^z) goes to infinity in the half-plane Re (siz) > 0 and to zero in
Re (siz) < 0 when z -»<».
Let w2(z) be the solution corresponding to a simple loop about s = s2
 225
LAPLACE'S EQUATION

from and to infinity. This solution has the same type of asymptotic
expansion as (6.4.26), where now Si is replaced by s2, j by - j, and a and /3
are interchanged. Thus

w2(z) ~ 2i sin (pTTWip + l)(-/r (-jz)

(6.4.34)
x2F0(-a, 0 + 1 es*z

We can now also get the asymptotic representation of u>0(z), defined by

(6.4.16). Since the solutions are meromorphic functions of a and /3, we
may assume that Re (a) > - 1, Re (/3) > - 1, so that the loop integrals may
be replaced by straight line integrals as in Figure 6.4. The three straight

Figure 6.4

lines are closed by an arc of a circle of large radius. The integral along the
closed contour is zero, and for suitably restricted values of z the integral
along the arc goes to zero as the radius becomes infinite. This implies that

(6.4.35)
mm;
If now the straight line integrals are expressed in terms of the loop
integrals, we get the relation between w0(z), w,(z), and w2(z) and obtain
the asymptotic formula for w0(z) from (6.4.26), (6.4.34), and (6.4.35). The
unpleasantly difficult details are left to the diligent reader.
To illustrate we consider Bessel's equation
z2w" + zw' + (z2-a2)w =0, (6.4.36)
226 SPECIAL SECOND ORDER EQUATIONS

which is reduced to the form (6.4.2) by setting w = zav. Thus

zi/' + (2a + l)v' + zv = 0, (6.4.37)
so that Co = 2a + 1, Ci = 0, d0 = 0, di = 1. The equation for F(s) becomes

(s2 + l)F'(s) = (2a - l)sF(s), (6.4.38)

so that
F(s) = (s2+l)a"1/2 (6.4.39)
and

V(z) = j ezs(s2+\)a-U2 ds. (6.4.40)

Here C may be a loop integral or a straight line integral. The latter choice
is admissible if Re (a) > -i There are then three choices: an integral from
s = i to infinity or from - 1 to infinity or from - i to + i. The first choice
leads to an asymptotic expansion,

(-2i)a+,/2r(a +i)z—meiz2Fo(a +|,-a (6.4.41)

Multiplication by z" gives the Bessel-Hankel function Ha\z), the second

choice leads to H(a2)(z), and the integral from —i to +/ gives Ja(z), in all
three cases up to a multiplicative constant.
Assuming z = x to be real positive and taking only the first terms of the
two asymptotic series, we obtain an expression which implies that

Ja(x)~(^) C0S[*-iiKa+5)] (6.4.42)

in the first approximation.
The equations studied in this section go under the name of confluent
hypergeometric equations. The hypergeometric equation

x(l-x)^|+[c -{a +b + \)x]^+aby =0 (6.4.43)

has three singular points, 0, 1, ». If we now set x = z/b, a. transformed
equation is obtained with the singular points 0, b, <». Here we let
and find that the limiting equation may be written as

zw"(z) + (c - z)w'(z) + aw(z) = 0 (6.4.44)

with the solution

,F,(a ; c ; z) = lim 2F, (a, fr, c ;0 (6.4.45)

A second solution may be obtained by applying the same process to the
solution w02(x) of (6.4.43). It is
 227
LAPLACE'S EQUATION

zl~ciFx(a-c + l;2-c;z). (6.4.46)

Equation (6.4.44) is the typical confluent hypergeometric equation.

It is clear that, if a =-n, a negative integer, then ,Fi(a;c;z) is a
polynomial of degree n in z. We write
(6.4.47)

and refer to L(„a)(z) as the nth (Abel-) Laguerre polynomial of order a. If

a = 0, the superscript is usually dropped and L„(z) is the nth Laguerre
polynomial, named for Edmond Laguerre (1834-1886), a brilliant
Frenchman who enriched algebra, analysis, and complex geometry.
These polynomials have a number of important and interesting proper-
ties. We list only two, and ask the reader to verify them. The expansion

(6.4.48)

may be regarded as a generating function of the polynomials. Further-

more, we have the orthogonality property

(6.4.49)

The set {e x/2Ln(x)} is a complete orthonormal system for L2(0, °°).

EXERCISE 6.4
1. Determine the values of a and j8 in (6.4.12).
2. Why is w0(z) of (6.4.15) an entire function of z?
3. Consider the equation zw" + w =0, which is a Laplace equation satisfied by
Bessel functions. Evaluate the loop integral of Figure 6.3 for this case
(calculus of residues or otherwise). Express it as a Bessel function.
4. Verify (6.4.20).
5. Verify (6.4.22).
6. Verify the Hankel integral (6.4.23).
7. Verify (6.4.34).
8. Find the asymptotic representation for w0(z).
9. Express the integral

in terms of Ja{z).
10. Explain how (6.4.42) is obtained.
11. Carry through the derivation of (6.4.46).
12.
Prove that ezxF\{c -a,c;-z) = iF,(a, c ; z).
(^228^ SPECIAL SECOND ORDER EQUATIONS
13. Show that

n \Ln(z) = ez-£-r(zn
az e z)
and obtain similar formulas for the nth polynomial of order a.
14. Show that the zeros of Laguerre polynomials are real positive.
15. Try to prove (6.4.48). Note that the right member is holomorphic in \t \ < 1 for
any value of z and may be expanded in a power series for |f|< 1 with
coefficients which are functions of z. Discuss the coefficients.
16. Prove (6.4.49). [Hint: Try using (6.4.48).]
17. In Bessel's equation (6.4.36) set t = - 2iz, w = za enu. The resulting equation
for u as a function of t is of form (6.4.44). Find the parameters.

6.5. THE LAPLACIAN: THE HERMITE-WEBER EQUATION;

FUNCTIONS OF THE PARABOLIC CYLINDER

The DE's of classical mathematical physics usually arise from a boundary

value problem involving one or an other of the partial DE's: Laplace's
equation, Poisson's equation, the wave equation, or the heat equation.
Frequently the problem suggests a change of variables which affects the
Laplacian.
The two-dimensional Laplacian is

^—rm^u. (6.5.1)
The basic formula for the change of variables is given by

LEMMA 6.5.1

Let U(x, y) = U(z) be defined and twice continuously differ entiable in a

domain D of the z-plane, and let D be the one-to-one image of a domain
Do under the conformal mapping
z = f(u), u = s + it, (6.5.2)
where f(u) is holomorphic in D0. Then

AstU = |f'(u)|2AxyU. (6-5.3)

Proof. Straightforward computation shows that

^=^=0[(SHS)>0[
(1H!O]
THE LAPLACIAN: THE HERMITE- WEBER EQUATION 229

Here each of the last two brackets is zero since x and y are harmonic
functions of s, t. By the Cauchy-Riemann equations each of the first two
brackets equals so that (6.5.3) holds. ■
In particular, if U is a harmonic function of x, y, then U[f(u)] is a
harmonic function of s, t.
There are three problems in potential theory involving cylinders and
axial symmetry. The cylinders are circular, parabolic, and elliptic, and
they give rise to the functions of the circular cylinder, the parabolic
cylinder, and the elliptic cylinder, respectively. The reader has already
encountered the circular cylinder functions; they are the Bessel func-
tions, but at this juncture it may be desirable to justify the name "circular
cylinder functions."
It is required to find a solution of the Poisson equation [Simeon Denis
Poisson, 1781-1840]:
Axy U + U = 0, (6.5.5)

which is a function of r and 6 alone, z = rel6, twice continuously

differentiate for |z|2 = x2 + y2 < R2, and which satisfies a given boundary
condition on |z| = R ; for instance, U(reie) converges to a given function
/(0) in the L2(0, 27r)-metric as r | R. The problem calls for polar
coordinates, and then (6.5.5) becomes

d2U 1 dU 1 d2U
lrr+7JF+r1lF+u-0' (6-5-6)
Here it is possible to separate the variables. The equation has solutions of
the form U(r eie) = f(r)g(6). If we perform the indicated differentiations
and divide the result by f(r)g(d), it is seen that

r2 f"(r) + rf'(r) + (r2 - c ) f(r) = 0, (6.5.7)

g"(d) + cg(0) = 0, (6.5.8)
where c is a constant, but not an arbitrary constant, for f(r)g(0) must
satisfy (6.5.6) and be twice continuously differentiate. This requires g to
be periodic with period 2tt, so that c must be the square of an integer,
c = n2. Then / must satisfy

r2f"{r) + rf'(r) + (r2 - n 2)/(r) = 0, (6.5.9)

i.e., f(r) is a Bessel function of integral order n. The continuity condition
forces the solution to be a constant multiple of X, (z). Thus we see that
(cnenie +c-ne~nie)jn(r) (6.5.10)
is a solution of (6.5.6) for arbitrary integral values of n and arbitrary
230 SPECIAL SECOND ORDER EQUATIONS

constants c„, c~n. Since the problem is linear, the general solution is

Co/o(r) + nJ)
= l Jn(r)(cn enie + C-n e~nie). (6.5.11)
This should be L2 with respect to 0 uniformly in r for 0 ^ r ^ R, and since

/n(r) = (irr^[l + 0(n1)L

the L2-condition is satisfied iff

i(^)W"[|c„|2 + |c-„|2] (6.5.12)

is convergent. Here the boundary values must be given by the L2-series,

f(6)~ c0/o(i?)+ 2 /n(i?)(c„^"ie +c_ne me). (6.5.13)

On this note we leave the functions of the circular cylinder.
Our main task in this section is to study the corresponding problem for
a domain bounded by a parabolic cylinder,

y2 = 4R(R-x) (6.5.14)
as a right-hand boundary. The domain bounded by the parabola and cut
along the negative real axis is mapped conformally on a vertical strip,
0<s <(2R)m, by
z =f(s + it) = \u2 = \{s2-t2) + ist. (6.5.15)
We now take the partial DE in the form

A,y U- U =0, (6.5.16)

so that the transformed equation is

A« U - (s2 + t2)U = 0. (6.5.17)

Here again the variables may be separated. The equation has solutions of
the form f(s)g(t) and

f"(s) - (c + s2) f(s) = 0, g"(t) + (c - t2) g(t) = 0, (6.5.18)

so the equations are almost identical. The solutions are the functions of
the parabolic cylinder.
We take the equation in the normal form

w"(z) + (c -z2) w(z) = 0, (6.5.19)

where z is a complex variable and c an arbitrary constant. This DE is also
THE LAPLACIAN: THE HERMITE-WEBER EQUATION 231

known as the Hermite-Weber equation. Charles Hermite (1822-1901)

considered it in 1865; Heinrich Weber (1842-1913), in 1869. It may be
reduced to a confluent hypergeometric equation by the substitution

which leads to t=z\ u=exp(h2)w, (6.5.20)

1IF + 6 " 0 ft + l(c " 1} u = 0 (6 5 21)

with solutions

iFiG(l-c),i;f] and f1/2iF,(l-k,§;0. (6.5.22)

The case in which c is a positive odd integer is particularly important. If
c = 4m + 1, then
.F,(- m,!;0

is a polynomial in t of degree m. If c = 4m +3, then

r1/2,F,(- m,i;0

is t 1/2 times a polynomial in t of degree m. This means that, if c = In + 1,

then (6.5.19) has a solution of the form

En(z) = exp (-k2) H„(z), (6.5.23)

where H„(z) is the nth Hermite polynomial, which is of degree n and is an
even or odd function of z according as n is an even or odd integer. We
have the generating function

exp (2zt -t2)=i Hn(z)—.

n=o n ! (6.5.24)
and the representation

H„(z) = (-1)" exp(z2)^pr[exp(-z2)]. (6.5.25)

Not all authors use the same notation for the Hermite polynomials; the
main variation is that z2 is replaced by lz2, as in E. T. Whittaker and G. N.
Watson, A Course in Modern Analysis (1952).
The set {En(z)} forms an orthogonal system for the interval (-00,00).
This follows from the DE. If m and n are distinct integers, then

E:(x) + (2m + l-x2)Em(x) = 0,

£';(x) + (2n + 1-jc2)E„U) = 0.
Here we multiply the first equation by E„(x), the second by Em(x), and
subtract to obtain

-p [Em {x )E 'n{x )-EH(x )E Ux )] = 2(m-n )Em (x )En (x ).

232 SPECIAL SECOND ORDER EQUATIONS

This is integrated from -<» to +00. It follows from (6.5.25) that the left
member is integrable and the integral is zero:

j Em(x)En(x) ds = I exp(-x2)Hm(x)Hn(x)dx =0 (6.5.26)

if m 7^ n. It may be shown that

J exp(-x2)[H„U)]2 dx = 77- 1/2 2" n!. (6.5.27)

The same technique gives

En(x)E'n(x) + j [E'n(s)f ds + j [s2-2n-l][En(s)f ds = 0. (6.5.28)

The identity implies that

En(x)E'n(x)<0 forx22*2n + l. (6.5.29)

In particular, the product cannot be zero for such values of jc, and this
extends to the interval (-00, -(2n + 1)1/2] since the product is an odd
function of jc. Thus all real zeros of Hn(x) belong to the interval
(-{In + l)1/2,(2n + 1)1/2). Formula (6.5.25) implies that
Hn+i(z) - 2zHn(z) + 2nH„_,(z) = 0. (6.5.30)
This, in turn, implies that the set {Hn(x)} forms a Sturmian chain. Hence,
if Hn(xo) = 0, then Hn+l(xo)Hn->(x0) < 0. Since H0(x) = 1 and H2k+,(0) = 0
for all k, we conclude that sgn H2k(0) = (- l)k for all k. Furthermore, it is
known that Hn [(2n + 1)1/2] > 0 for all n. From these facts we conclude step
by step or by induction that all the zeros of Hn(z) are real and that the
zeros of Hn+i(x) alternate with those of Hn(x). The verification is left to
the reader.
The asymptotic behavior of the solutions of (6.5.19) for large values of z
may be concluded from the results for confluent hypergeometric equa-
tions via (6.5.21) or by the method of asymptotic integration of Section
5.6. The results will be reviewed briefly. We have

Z(z)= f (c -s2)mds =iiz2 + \ic log z + 0(z"2). (6.5.31)

Jo
The lines y = ±x divide the plane into four sectors, in each of which there
is a subdominant solution. For the sector |arg z | < \tt this takes the form

E|(z) = z(c-1)/2 exp (~\z2)[\ +o(l)], (6.5.32)

and the same formula represents £3(2) in the sector \tt < arg z < lir. Here
E+\(z) has no distant zeros in |arg z| < W, while Et(z) has no distant zeros
in |arg (- z)| < \tt. For these two solutions, which are entire functions, the
THE LAPLACIAN: THE HERMITE-WEBER EQUATION 233

defect of the value 0 is normally \. The zeros of E\{z) satisfy lim arg z„ =
lir or -!tt, and N(r, 0; Et) = (l/7r)r2 + O(r). The value c = In + 1 is
exceptional.
E„(z), which Here Ej"(z)
has only and El(z)
n zeros, so the coincide
defect ofand
zerobecome multiples
is one and zero isofa
Picard value. A similar situation holds when c is a negative odd integer,
for then
H„(/z)exp(k2) (6.5.33)
is a solution for c = -In - 1 and the subdominants Ej(z) and Et(z) are
linearly dependent. No other values of the parameter c can give rise to
such phenomena. We note that the set {E„(z)} corresponds to a Titch-
marsh problem for (6.5.19).

EXERCISE 6.5
1. Show that a solution of the problem

A„ U-(s2 + t2)U = 0, 0<s <(2£)1/2, lim U(s,t) = F{t),

where F(t)E L2(— °°, °°) and convergence is in the L2-metric, is of the form

112= 1 cn exp \i(t2-s2)]Hn(s)Hn(it).

2. If w(z, a) is a solution of (6.5.21), show that w(iz, - a) is also a solution.

3. If c = - 1, show that w(z) = exp (k2) JT exp (- sz) ds is a subdominant in the
sector |argz|<|7r, and find subdominants for the other sectors.
4. Find the DE satisfied by Hn (z), and show that it is confluent hypergeometric.
5. Prove (6.5.27).
6. Verify (6.5.29).
7. Find H2k(0) and tf2k+1(0), and show that they are 0(k~m). (Hint: The formula
for tt as the limit of factorials due to John Wallis may come in handy.)
8. Prove the statements in the text regarding the zeros of Hn(x).
9. Since there are n zeros of Hn(x) in the interval (-A, A), where A =
(In + 1)1/2, the average distance between consecutive zeros is ~23/2m"1/2.
Show that the minimum distance is about 7r(2n)_1/2.
10. Define two quadratic differential forms by

Fn (x ) = [E'n(x )f + (In + 1 - jc 2)[En (x )]2,

OAx) = ^^HEnMt
Show that the ordinates |En(xk)| at the consecutive extrema of En(x) form an
increasing sequence in [0, (In + 1)1/2] and that (In + 1 - xiyn\E„(xk)\ is a
decreasing sequence.
11. Use the results of Problems 7 and 10 to show that on any fixed interval
(-b, b) the set nm\En(x)\ is uniformly bounded.
 SPECIAL SECOND ORDER EQUATIONS

THE EQUATION OF MATHIEU; FUNCTIONS OF THE

ELLIPTIC CYLINDER
We have left for consideration the elliptic cylinder. Let the axis of the
cylinder be perpendicular to the z-plane at the origin, and let the
intersection of the cylinder with the z-plane be the ellipse

The foci of E are at the points z = - 1 and z = + 1 . We cut the domain

bounded by E along the real axis from the left end point of the major
axis to the left focus and from the right focus to the right end point of the
major axis. The cut domain is the conformal image under the mapping
z = sin u, u = s + it (6.6.2)
of the rectangle

-127r<s<y, -c<t<c, c =\og[R+(R2+\)m]. (6.6.3)

We have
z = x + iy, x = sin s cosh t, y = cos s sinh t
and
|/'(m)|2 = cos2s + sinh2 r.
Consider now the partial DE,
Axy U + K[/ = 0, (6.6.4)
where K is a positive parameter. The transformed equation is

AstU + K (cos2 s + sinh2 1 ) U = 0. (6.6.5)

Here again the variables may be separated, and the equation admits of
solutions of the form f(s)g(t), which satisfy the equations

/"(s) + (C + K cos2s)/(s) = 0, (6.6.6)

g"(0 + (- C + K sinh2 t)g(t) = 0. (6.6.7)
We shall consider the first equation in the normal form

w"(z) + (a + b cos2z)w(z) = 0. (6.6.8)

This is known as Mathieu's equation after Emile Leonard Mathieu
(1835-1890), who in 1868 studied the vibration of an elliptic membrane. It
leads to another equation of type (6.6.4) after the time coordinate has
been eliminated from the wave equation. Mathieu's name is also attached
to a class of finite groups investigated by him.
Equation (6.6.8) contains two parameters, a and b. In the physical
THE EQUATION OF MATHIEU 235

problem just mentioned, a and b must be real, and it is also seen that the
solution must be periodic with tt as a period or a half-period to be
physically significant. This raises the question of the existence of periodic
solutions of a DE with periodic coefficients. We shall have something to
say about this general problem in the next section; here we shall present
some aspects of the problem as posed by Mathieu's equation.
Equation (6.6.8) does not change if z is replaced by z + kir, where k is
an integer. This means that if w(z) is a solution so is w(z +/c7r). The
problem facing us is to decide under what conditions on the parameters a
and b there exists a solution w(z) with period kir,

w(z + kir)=w(z). (6.6.9)

For the physical problem it is enough to treat the cases k = 1 and 2.
In passing, we note that if w(z) is a solution so is w(-z) and
J[w(z)+ w(-z)] and s[w(z) - w(-z)].
The first is an even function of z ; the second, an odd function. This means
that the search for periodic solutions may be restricted to the subsets of
even and of odd solutions. A further simplification arises from the fact
that the existence of even or odd periodic solutions is equivalent to the
fact that certain boundary value problems have solutions. Now the
existence of such solutions follows either from the general theory of such
problems or from function-theoretical considerations. The second alter-
native isappropriate in a treatise on DE's in the complex plane and will be
used here.
Suppose that w(z) is an even solution of primitive (= least) period 2tt.
Then w'(z) is odd and of period 2tt. We have w'(0) = 0 and w'(-tt) =
w'(tt) by the periodicity, while w'(-tt) = -w'(tt) since the function is
odd. Thus w'(tt) = 0. Hence our solution must satisfy the two boundary
conditions,

w'(0) = 0, w'(tt) = 0. (6.6.10)

Suppose conversely that a solution of (6.6.8) satisfies (6.6.10). In the DE
(6.6.8) the factor (a + b cos 2z) is even, so a solution with w(0) = c ¥=■ 0,
w'(0) = 0 is an even function of z. By the same token w(z) is an even
function of z - tt. Now w(tt - s) = w(ir + s) implies successively that
w (277 + z) = vv(-z) = w(z) or w(z + 27r) = w(z),
so that our solution does have the period 27r.
Suppose instead that w(z) is an odd solution of primitive period 2tt.
Then w(0) = 0 and w(- tt) = - w(7r), since the solution is odd, while, on
the other hand, w(-tt) = w(tt) by the periodicity. Hence w(tt) = 0, and
236 SPECIAL SECOND ORDER EQUATIONS

our solution satisfies the conditions

w(0) = 0, w(tt) = 0. (6.6.11)

Conversely, if w(z) is a solution satisfying these conditions, then the
argument given above shows that w{z) is an odd function of z as well as
of z - 77. From w{tt - s) = - w{tt + s) we infer that with s = it - z we
have w(z) = - w(2tt - z) = w(z - 2tt) and, finally, w{z + 2tt) = w{z).
Thus (6.6.11) Implies that w(z) is odd and has the period Itt.
Next we consider the case in which w(z) is odd and has the period tt.
This implies, in the first place, that w(0) = 0. Since w{-z) = -w{z) and
w(z + 7r) = w(z), this gives w{-\tt) = w{\tt) by the periodicity and
w(-57r) = - w(j7r) by the oddness, so that w{\tt) = 0. Thus
w(0) = 0, h(57t) = 0. (6.6.12)
Conversely, suppose that a solution satisfies these conditions. Then we
see that w(z) is an odd function of z as well as of z -\tt. Now from
w(jtt — s) = — w(jir + s) we infer that w(z + tt) = — w(— z) = w(z). Thus
(6.6.12) implies that w(z) is odd and has period tt.
Finally, if w(z) is even and has period tt, we show that

w'(0) = 0, w'(r7r) = 0; (6.6.13)

and conversely, if these equations hold, then w(z) is even and has period
TT.
If solutions of these four boundary value problems exist, they may be
expanded in Fourier series, and corresponding to the four conditions we
have the expansions

(i) n=0
2 c \,n cos {In + l)z, (ii) n=0
2 c \,n sin {In + l)z,
(6.6.14)
(iii) n=0
Xc3,nSin2nz, (iv) 2 cJ,„cos2nz.
n=0
These four types of periodic solutions are denoted by
(i) ce2fc+,(z), (ii) se2k+i(z),
(d.d.Ij)
(iii) se2»c(z), (iv) ce2k(z),
and the notation is such that

lim cem (z) = cos mz, lim sem (z) = sin mz

when b ->Q.
In the applications b is given and a has to be determined so that one or
another of the four sets of boundary conditions is satisfied.
THE EQUATION OF MATHIEU
237

We shall indicate a function-theoretical argument to prove the exis-

tence of infinitely many values of a such that for a given b one of the
boundary value problems, e.g., (6.6.11), has a solution. The existence of a
solution of (6.6.9) such that

w(0) = 0, w'(0)=l,
is obvious and this holds for all values of a and b. Let this solution be
w(z; a, b); it satisfies the Volterra integral equation

w(z) = z-j (z -s)(a +b cos 2s)w (s) ds. (6.6.16)

The problem is now to show that for a given value of b ^ 0 it is possible to
find values of a such that w(tt ; a, b) = 0. To do this we need, in the first
place, an estimate of the rate of growth of w(ir; a, b) as a function of a
for fixed b. It is clearly an entire function of a. We take z = x as real
positive, and set |w(jc)| = u(x). This gives

u(x)^x +(|a| + |fc|) (x-s)u(s)ds. (6.6.17)

From (6.6.17) we get, for example by successive substitutions,

and finally u(x)^(\a\ + \b\ym sinh [(|a| + |fr|),/2x]

\w(<ir;a9b)\*z{\a\ + \b \)~1/2 sinh [(|a |+ \b |)1/2tt]. (6.6. 18)

It follows that w (77 ; a, b ) for fixed b is an entire function of a of order \
and normal type. An entire function of order at most \ either reduces to a
polynomial and the order becomes 0, or has infinitely many zeros {an(b)}
with an exponent of convergence ^5, i.e.,

n2= l \an(b)\-~ (6.6.19)

converges for any o- >i The corresponding solutions w(z; an(b)9b) are
the functions se2n+i(z) which satisfy (6.6.12). The probability of the
function w(tt; a, b) reducing to a polynomial in a would seem to be (and
actually is) zero, but the author knows of no elementary argument that
establishes this. With this reservation we have proved what we set out to
do. We have, furthermore,

THEOREM 6.6.1

// b is real, the characteristic values of the boundary value problems

(6.6.10)-(6.6.13) are real and the functions ce2n+i(x), se2„+i(x), se2n{x), and
ce2„(x) form an orthogonal system for L2(0, tt).
238 SPECIAL SECOND ORDER EQUATIONS

Proof. Consider one of the functions wm(x) corresponding to a = am.

Then, for z = x real,
w^(x) = -(am+b cos2x)wm(x). (6.6.20)
Suppose that am is not real, so that wm(x) is complex valued. Multiply
both sides by the conjugate of wm(x), and integrate from 0 to it. Thus

Jo
I wm(x)w'^(x)ds = - JoJ (am +b cos2x)|wm(jc)|2djc. (6.6.21)
Integration by parts gives

[wm(x)wUx)]o = Jo
f \wL(x)\2 dx - Jof (am + b cos 2jc)|w„,(jc)|2 dx. (6.6.22)
Here the integrated part vanishes in both limits by virtue of the boundary
conditions. Since the first integral is real positive, the second integral must
be real, as is the case iff am is real.
A glance at (6.6.14) shows that sem(jc) and ce„(jc) are orthogonal on
(0, 77 ) regardless of the values of m and n. It remains to prove the
orthogonality of sem(jc) and se„(x), m^n, and of cem(x) and ce„(jc).
Suppose that wm(x) with characteristic value am satisfies (6.6.20), and
let wn(x) with characteristic value an satisfy

w"(x) = ~(an +b cos2jc)w„0c), m^n. (6.6.23)

Multiply (6.6.20) by wn(x) and (6.6.23) by wm(x), subtract, and integrate
from 0 to 7T, obtaining

Jo [wn(x)w'^(x)-wm(x)w'n(x)]dx=(an-am)j wm{x)wn(x) ds.

The integral on the left can be carried out and gives

[w„ (x )w U x ) - wm (x )w 'n(x )]«,

which is zero by the boundary conditions. Since am # an we see that, as
asserted,

i. wm(x)wn(x) dx = 0.

The periodic solutions of (6.6.8) and, in particular, the functions ce„(z)

and sem(z) are known as Mathieu functions of the first kind. There are also
associated Mathieu functions which are essentially the solutions of
(6.6.7). Suppose that a and b are real. Then for any integer k or 0 (6.6.8) is
formally real on all the lines

z=\kTT + iy, -oo<v<+oo, (6.6.24)

THE EQUATION OF MATHIEU 239

the transformed equations being

(6.6.25)
^TT~[a +(-1)"/? cosh2y]w = 0.
The solutions of these equations are the associated Mathieu functions.
Their behavior depends on the sign of b and the parity of k. If b > 0, k
even, the real solutions of (6.6.25) go hyperexponentially to infinity or,
exceptionally, to zero. If b > 0, k odd, the real solutions exhibit very
rapid, strongly damped oscillations. If b < 0, the parity of k should also
be changed in this description.
To study the asymptotic behavior of these solutions or, equivalently,
the behavior of w(z) in the complex plane, we use the transformation of
Liouville. Set

W(Z) = (a+b cos 2z),/4w(z). (6.6.26)

The lines x = (k + \)tt, together with the real axis, divide the complex
plane into infinitely many half-strips,

Dk: [z\z =x +iy,(k -Xt)tt <x <(k +5)tt,0 <y], (6.6.27)

Dk: [z|z = x + iy,(fc-5)7r < x <(k +\)tt9 y <0].
In each of these half-strips there is a unique normalized subdominant
furnished by the discussion in Section 5.6. Let the subdominant in Dj be
denoted by E(z). We know that its asymptotic form for z restricted to Dl
is, for b > 0,

E(z) = exp [-{\b)m e'iz + \iz][Co + 0(eiz)l Co * 0. (6.6.28)

The symmetry properties of (6.6.8) show that JE(-z) is subdominant in
Do, and, more generally, the subdominants are given by

E(z — kir) in Dl and E(kir-z) in Dk,

for E(z-kir) is clearly subdominant in Dk and, since the normalized
subdominant is unique, the conclusion follows.
These properties have important consequences. E(z) is an entire
function defined in the whole finite plane. It goes to zero in Do and to
infinity in the adjacent half-strips, D f , and D\. This implies, among other
things, that E(z) and E(z + 7r) are linearly independent; in particular,
E(z) cannot have the period tt.
The solutions E(z) and E(- z) are linearly independent. To prove this,
set
l2[E(z)+E(-z)] = C{z), i2[E(x)-E(-z)] = S(z).
Here C (z ) is even while S (z ) is odd. If E (- z ) were a constant multiple of
240 SPECIAL SECOND ORDER EQUATIONS

E(z), then S (z) would be a constant multiple of C(z), and this is absurd.
The possibility that E(z) itself is even or odd is excluded by (6.6.32)
below.
From these properties it follows that for no values of a and b can there
exist two linearly independent solutions with the period tt or the period
2tt. To have two solutions with period tt is impossible, for this would
require E(z) to have the period tt — an impossibility. If the solutions had
period 2tt but not tt, one solution could be taken as ce2„+i(z) while the
other was se2n+i(z). Then these solutions would have the property
/(z + tt) = -/(z), and ultimately this would lead to E(z) and E{-z) being
linearly dependent — another absurdity.
To complete the discussion we must justify (6.6.28), and this requires a
discussion of the mapping z »-»Z(z) defined by (6.6.26). The discussion
can be restricted to the mapping of the half-strip Do. See Figure 6.5 for

plane
plane

Z(-±TT)

Figure 6.5.

the geometry of the mapping. We assume a > b > 0 and z0 = 0. Then there
are two branch points of Z(z) on the vertical boundaries, given by

D [a + (a2 - b2)m].
±2-77 + i\ log \- (6.6.29)

Then Z = Z(z) maps the interior of Do on the interior of a region U+

which contains an upper half-plane and is bounded by five line segments,
two vertical and three horizontal. There are four vertices, two corres-
ponding to the points z = ±\ir, and two with interior angles of 270°
corresponding to the branch points.
Using the symmetry principle of H. A. Schwarz, we can continue Z(z)
THE EQUATION OF MATHIEU 241

analytically across each of the five boundary lines of Do. Let us enter Df
by crossing the right vertical boundary of Do, above the branch point if
a > b > 0. The image of D| under this determination of Z(z) is a domain
U7, the mirror image of U+ with respect to the horizontal line segment
furthest to the right. Similarly we get an image U~i of D_i by reflecting U+
in the horizontal line segment furthest to the left.
In the three domains Dj, Dl", and D-u the function z *-+ Z(z) has the
same asymptotic character,

Z(z) = i{hb)m e~iz + O(l), (6.6.30)

and the corresponding function F(Z) is found to satisfy Hypothesis F of
Section 5.6, so the asymptotic integration theory applies. We have

F(Z) = -\Z~2 + 0(Z~4). (6.6.31)

It follows that the subdominant solution E(z) in Do is indeed given by
(6.6.28), as asserted. The estimate for the remainder holds uniformly for
-§7r + e jc ^ \tt - e, 0 < y. In particular, on the lines x = ± it we have
E (± 7T + iy ) = exp [{\b )m e y - \y ][± iC0 + O (e y )]. (6.6.32)
Finally, in Do we have

E(z) = exp [(\b)meiz -|iz][C, + 0(e iz)] (6.6.33)

as y -> — oo. Thus E(z) is a dominant solution in Do.

EXERCISE 6.6
1. Are there any Mathieu functions having the primitive period 37r?
2. Why do the boundary conditions (6.6.10)-(6.6.13) imply Fourier series of
type (6.6.14)?
3. If b is real, show that the Fourier coefficients are real.
4. It is desired to estimate the rate of decrease of the Fourier coefficients. Show
that

^ ce2n(s) e2iks dsj< M(t) e~2kt, 0<k,

j
where M(t) = max |ce2*(z)| in the rectangle with vertices at 0, it, it + it, it.
[Hint: Integrate along the perimeter of the rectangle.]
5. A reasonable estimate of M(t) is C(n) exp [(\b)me' - \t\ How should t be
chosen so that the estimate in Problem 4 will be as small as possible for a
given k >0, and what is the resulting estimate?
6. Using this estimate and Stirling's formula, show that the Fourier coefficients
in (6.6.14) go to zero as

C(n)(\b)k+m(2k +\Tm[nik +!)]-'.

242 SPECIAL SECOND ORDER EQUATIONS

7. Show that cem(z) and sem(z) are either real or purely imaginary on each of
the lines x = (k + \)tt.
8. Describe the map obtained when z crosses one of the other three boundaries
of Dd.
9. Find the explicit expression for F(Z) in terms of z.
10. Why can an entire function of order \ have no Picard values?
11. Suppose that a > b >0 and that w(x) is a real solution of
w"(x) + (a + b cos2jc)w(jc) = 0.
Show that w(x) has infinitely many real zeros, and try to estimate the
number in the interval [0, R].

6.7. SOME OTHER EQUATIONS

Mathieu's equation is a special case of a linear DE with simply periodic

coefficients. We have seen that such an equation may have a periodic
solution for a given value of b and a suitable choice of a. What we have
not seen is that the equation always has solutions of the form e~"P(z),
where c is a constant and P(z) is periodic — this regardless of the values
of a and b. This is a special case of a theorem discovered in 1883 by
Gaston Floquet (1847-1920) for systems of linear DE's with periodic
coefficients. We shall prove this theorem in matrix form. The proof
requires more matrix theory than was developed in Section 1.4, however,
so the first item on the agenda is matrix-valued analytic functions. In
particular, we need the fact that a regular matrix has a logarithm.
If si G 3ft„, the algebra of n-by-n matrices, then a matrix <3l E Wln is
called a logarithm of si if
exp(2&) = 5*. (6.7.1)
Now exp (S&) has an inverse, exp (— 2&), so we see that the existence of a
logarithm of si requires that si have an inverse, i.e., be regular. This
condition is also sufficient, but to show this fact a more elaborate
argument is required. It is fairly simple if the characteristic values of si
are simple, more complicated if there are multiple roots. Suppose that
s = Sj is a i>-fold root of the minimal equation (this may be less than the
multiplicity as a root of the characteristic equation); then there exist two
matrices, % and Nh The former is an idempotent, and the latter a
nilpotent. They commute and we have

{%,f = %h %%k =%k%,= 0, j * k, (6.7.2)

= X,% = Xh * 0, X; = 0. (6.7.3)
Suppose now that f(s) is an analytic function which is holomorphic on the
SOME OTHER EQUATIONS 243

spectrum of si, i.e., for s = sh V j. Then f(sl) can be defined, and we have

fW = £ [/(s,)*, +/'(s,yf,/l! + - • •+r-|(s,Vrr1/(i' - I)!]. (6.7.4)

In particular, this formula is valid for/(s) = log 5 when 54 is regular. Thus
a regular matrix has a logarithm. The reader who wants an exercise in
matrix calculus is invited to prove that this definition of the logarithm
satisfies condition (6.7.1). We can now state and prove the theorem of
Floquet.

THEOREM 6.7.1

Let ?F(z) be an n-fry-n matrix of functions fjk(z) holomorphic in the strip

S:-o°<x<o°, — b < y < b, and periodic with the real period oj. Then the
matrix DE
W(z) = &(z)W(z) (6.7.5)
has a fundamental solution matrix of the form

°W(z) = 9>(z) exp (<€z), (6.7.6)

where SP(z) is holomorphic in the strip S and of period o>. // si is the
transit matrix from W(z) to W(z + o>), so that W(z + o>) = W(z)si, then si
is nonsingular and % = (log si)/(o.

Proof. The matrix 'W(z) is a fundamental solution matrix if it is a

solution matrix and nonsingular. If W(z) has this property, every other
solution matrix is of the form °WH(z) = °W(z)M, where jit is a constant
matrix and °W0(z) is fundamental iff M is nonsingular. Since the DE is
unchanged if z is replaced by z + o>, it follows that °W(z + oj) is fundamen-
talfiT 'W(z) is fundamental. In this case the matrix si is uniquely defined
by W(z + a>) = W(z)si and si is nonsingular. Now si has a logarithm. We
can obviously assume that (6.7.5) has a solution of the form (6.7.6), but the
problem here is to determine what properties % and must have in view
of the periodic character of >.
Now exp (Jit, + M2) = exp (jit,) exp (M2) iff Mi commutes with jtt2. We
now set % = (Mod) log s4 and observe that

exp [<€(z + a))] = exp (<€z ) exp (<€&>) = exp (%z)si.

Furthermore,
W(z +(o) = <P(z +o>)exp [%(z +co)] = 9P(z + a>) exp (<€z)j4.
Comparing with °W(z + co) = W(z)si, we see that (6.7.6) is a solution if
°3>(z + ay) = 9?(z)y as asserted. ■
Mathieu's equation is a special case of what is known as Hill's
244 SPECIAL SECOND ORDER EQUATIONS

equation,

w"(z) + (J) ak cos 2Jcz) w(z) = 0, (6.7.7)

named after the American astronomer G. W. Hill (1838-1914), who
encountered such equations in lunar theory in 1877. In his study he was
led to the use of infinite determinants, a field in which he was the pioneer.
Hill's heuristic considerations were put on a firm basis by Henri Poincare
in his revision of celestial mechanics in the 1890's. G. Mittag-Leffler
succeeded in having Hill's paper reproduced in Acta Mathematica, Vol. 8
(1886), and he suggested to one of his pupils, Helge von Koch, that the
latter work in this new field. It was a labor of love, and in the 35 years of
his scientific life von Koch produced general theories of infinite determin-
ants and numerous applications to linear equations in infinitely many
unknowns and to the theory of linear DE's.
In the early papers von Koch considered determinants of the form
D = det (8jk + ajk), (6.7.8)
where
E5>*| (6-7.9)
is convergent. Such determinants are absolutely convergent in a sense
familiar from the theory of infinite products. The minors formed by the
elements with subscripts =s= n form a convergent sequence the limit of
which is by definition the value of the determinant. All the elements of
this sequence are dominated by

n(l + 2M)» (6-7.10)

which is then a bound for D. It is also a bound for all minors of D. The
linear system
2k (5/k + aik)xk = y,, j = 0, 1, 2, ... , (6.7.11)
has a unique solution in the space of bounded sequences iff D ¥■ 0, while
the homogeneous system is solvable if D = 0.
Several problems posed by Mathieu's equation and the more general
Hill's equation lead to linear equations in infinitely many unknowns and
hence to infinite determinants. Thus by the Floquet theory Hill's equation
ought to have solutions of the form

eC2^cne2ni% (6.7.12)
and here we have the problem of determining c and the c„ 's. If the series
is substituted in (6.7.7) and the absolutely convergent series is multiplied
SOME OTHER EQUATIONS / 245

out, one gets a system of homogeneous equations of the form

(c+2m)2c„ + m S
= — oo amcn m=0, n = . . . , -2, - 1, 0, 1, 2, . . . . (6.7.13)

If here the nth equation is divided by [(c + 2ni)2 + a0], the resulting
system has an absolutely convergent determinant, D(ic) say. It follows
that c has to be determined from the equation
D(ic) = 0. (6.7.14)
In this case it turns out that D(ic) is an elementary function of c, and c
has to satisfy
sinh2 \ttc = C,

where C is a specific constant. Once c is determined, the c's are

uniquely determined in terms of c0 and the cof actors of D(ic).
Another problem that leads to similar considerations is to determine the
values of a in terms of b if, say, se„(z) is to be a solution of Mathieu's
equation and the corresponding Fourier coefficients. We refrain from
further details.
Linear DE's with doubly periodic coefficients have also been studied.
There does not seem to exist a strict analogue of the theorem of Floquet,
the difficulty being that the equation now has infinitely many fixed
singularities at the poles of the coefficients, and these may be irregular
singular or, if regular singular, may lead to branch points possibly
involving logarithms.
The linear second order case leads to Lame's equation, named after
Gabriel Lame (1795-1870). Important Lame~equations^arrise in Newtonian
potential theory when elliptical coordinates are introduced in the Lapla-
cian. Consider a family of confocal central quadrics:

x2 v2 z2
a+s b + s c+5

Through a given point (x, y, z) in 3-space pass three quadrics of the

system corresponding to the three roots r, s, t of (6.7.15), regarded as a
cubic equation in s. The roots are real; if 0 < a < b < c, then - c2 < r <
-b2 < s <-a2 <t <oo. The three roots are the confocal coordinates of
(x, y, z). The transformed equation has solutions ol the form j\r )g {s^)h (t).
The factors satisfy a linear second order DE of the Fuchsian class. These
equations can be written in several different forms, algebraic or elliptic.
One algebraic form is

dju,}_( L_ 1 _1 \du_ n(n + l)s + C r*7i^

d52 + 2\a2 + s b2 + 5 c2 + s)ds "4(a2 + s)(fr2 + s)(c2 + 5)M' { '
246 SPECIAL SECOND ORDER EQUATIONS

where n is the degree of the ellipsoidal harmonic in (jc, y, z) and C is a

constant. This equation has four singular points, — a2, - b2, — c2, and o°, all
regular singular. The indicial equations have roots 0, \ at the first three
points and -\n, i(n + 1) at infinity. A simple linear translation will shift the
finite singularities into new positions, e,, e2> e3, with sum 0. In fact,

= -a2 + 5(a2 + b2 + c2),. .. ,ei = -c2 + j(a2 + b2 + c2).

We can now introduce the Weierstrassian p-function corresponding to the

equation {v'f = 4(v - ex){v - e2)(v - e3) as a new independent variable
and obtain the elliptic form of Lame's equation:

w"(z) = [n(n + l)p(z) + £]w(z), (6.7.17)

where B is a parameter. Here the singular points are the double poles of
the p-function, points congruent to zero modulo period. The roots of the
indicial equation at such a point are n + 1 and - n. Thus at every
singularity there exists a holomorphic solution corresponding to the root
n + 1. Since the difference of the indicial roots is an odd integer, we may
expect the second solution in the local fundamental system to involve a
logarithm. Actually, however, this does not happen. Suppose that the
holomorphic solution is denoted by Wi(z); then a linearly independent
solution w2(z) is given by

(6.7.18)

In the neighborhood of the singularity, say at u = 0, the integrand is an

even function of u, so no logarithms can arise from the integration of the
negative powers in the Laurent expansion. This means that the solutions
are all locally meromorphic and can be continued analytically along any
path in the plane which avoids the singular points.
The solution is normally not a doubly periodic function, but for
particular values of B and n one solution may be doubly periodic. A case
in point is furnished by the equation
w" = [^ + 2p(z)]w, (6.7.19)
where the solutions are as follows:
1/2 and
[p(z)-ei] [p(z) - e,]1/2[£(w + *>•) + eiz], (6.7.20)
the first of which is doubly periodic whereas the second is not. Here the
Weierstrass zeta function is -/ p(t) dt.
LITERATURE 247

EXERCISE 6.7
1. A possible way of defining log si is by the logarithmic series

i n converge?
For what values of si does the series
2. Give examples of idempotents and nilpotents in the matrix algebra 2ft3.
3. Suppose that °W(z) is a fundamental solution of (6.7.5). Show that multiplica-
tion on the right of °W(z) by a constant matrix yields a solution which is
fundamental iff the multiplier is nonsingular.
4. Verify that exp (si + <3l) = exp (si) exp CM) if the matrices si and % commute.
Try to find two noncommuting matrices in $Jl2 for which the addition formula
breaks down.
5. Suppose that D = det (8jk + ajk) is an infinite determinant with nonnegative
elements satisfying the convergency condition (6.7.9). Form the sequence of
principal minors {M„}, where only elements with subscripts ^ n occur in M„.
Show the convergence of the sequence, using the bound (6.7.10).
6. Verify that the roots of the indicial equations for (6.7.16) and (6.7.17) have
the stated values.
7. Verify that (6.7.18) defines a solution of (6.7.17). If Wi(z) is holomorphic at
z = 0, verify that w2(z) is single valued at z = 0.
The equation
(1 + z2)2w"(z) + aw(z) = 0
is known as Halm's equation [J. Halm, Trans. Roy. Soc. Edinburgh 41
(1906), 651-676].
8. Show that z = ±i are regular-singular points, and find the roots of the
indicial equation. What is the nature of the point at o°?
9. Find the general solution by the substitution

t = arctan z, v = (1 + z2)~mw.
10. Let z = x be real. Find the solution that tends to one as (i) *->-<», (ii)
x -> + oo. When are these solutions identical?

LITERATURE

This chapter corresponds to Chapter 7 of the author's LODE, which has

an extensive bibliography. Special reference should be made to:
Whittaker, E. T. and G. N. Watson. A Course in Modern Analysis. 4th ed. Cambridge
University Press, London, 1952.

The general theory of second order linear DE's starts in Chapter X, where
Section 10.6 is devoted to the DE's of mathematical physics and where,
following investigations of Felix Klein and Maxime Bocher (1894), a
248 SPECIAL SECOND ORDER EQUATIONS

certain second order DE with five singular points is singled out as the
common source of the six equations of Bessel, Hermite, Lame, Legendre,
Mathieu, and Stokes, through confluence and choice of parameters. The
program is carried out in Chapter XVIII. Chapters XIV-XVII deal with
the hypergeometric case, Legendre's equation, the confluent
hypergeometric case, and Bessel functions. There are a large number of
exercises.
Much information is to be found in the following books:
Ince, E. L. Ordinary Differential Equations. Dover, New York, 1944.
Kamke, E. Differentialgleichungen : Losungen und Losungemethoden. Vol. 1, 3rd ed.
Chelsea, New York, 1948.
See also the Bateman papers and:
Erdelyi, A., W. Magnus, F. Oberhettinger, and F. Tricomi, Higher Transcendental Func-
tions. 3vols. McGraw-Hill, New York, 1953-1955.
The standard treatise on Bessel functions is:
Watson, G. N. Theory of Bessel functions. 2nd ed. Cambridge University Press, London,
1962.

There is a large literature on confluent hypergeometric equations. In

addition to Chapter XVI of Whittaker and Watson (loc. cit), there are
several books by Tricomi. See in particular:
Tricomi, F. G. Fonctions hypergeometriques confluentes. Gauthier-Villars, Paris, 1960.
The discussion in Sections 6.5 and 6.6 is based partly on papers by the
author in Ark. Mat., Astr., Fys., 18, 26 (1924) and Proc. London Math.
Soc. (2), 23, (1924).
For the Floquet theory see the author's paper:
Floquet, G. Sur les equations differentielles lineaires a coefficients periodiques. Ann. Sci.
Ecole norm., Sup. (2), 12 (1883), 47-89.
For the theory of infinite determinants see:
Hill, G. W. On the part of the motion of the lunar perigee which is a function of the mean
motions of the sun and the moon. Acta Math., 8 (1886), 1-36.
Koch, H. von. Sur les determinants infinis et les equations differentielles lineaires. Acta
Math., 16 (1892), 217-295.
7

REPRESENTATION

THEOREMS

In this chapter we shall be concerned with the question of analytic

representations of solutions of DE's not necessarily linear of the second
order. Here the classical expansions in Taylor or Laurent series, with or
without a multiplicative power and/or a logarithm, come to mind and are
fairly obvious. Some problems lead to expansions in fractional powers for
algebraic or algebroid solutions: these are perhaps less trivial. But for
nonlinear equations we have also to allow expansions of one of the types

n2=0
anz\ m=0
2 n=0
2 amnzm+\ m2=0 n2=0 amnzK»(\ogz)n.
Such series will be called psi series in the following discussion; the third
type, in particular, is a logarithmic psi series.
Integral representations of the form

K(z,s)f(s)ds

are particularly important for the linear theory. The kernel K(z,s) =
(s - z)~a is associated with the name of Euler; es\ with Laplace. Double
and multiple integrals have also been pressed into service.

7.1. PSI SERIES

In the theory of nonlinear first-order DE's we may encounter equations

with solutions having infinitely many algebraic branch points, all but a
finite number of which are of the same order. The trivial equation

w' = w(w2-l) with w(z) = C(C2-e2zym (7.1.1)

is a case in point. There are branch points, all of order 1, at the points
249
250 REPRESENTATION THEOREMS

zn = log C 4- niri, where n is any integer. For the class

w' = R(z, w), (7.1.2)
where R is a rational function of z and w, the numerator a polynomial in
w of degree p and the denominator one of degree q, the typical branch
point is of the type
c(z-a) ~,/(p -q-l), (7.1.3)
with at most a finite number of exceptions.
For more spectacular singularities we have to consider second order
nonlinear equations. The deceptively simple special Emden equation,
zw"=w\ (7.1.4)
has a whole gallery of peculiar solutions and series expansions. Here
z = 0 and o° are the fixed singularities. The function 2/z is a solution with a
simple pole at the origin and a simple zero at infinity. This elementary
solution is embedded in two one-parameter families of solutions of the
form

where ^2+2 c^z""] and ±{2 + J) dnbnznT}, (7.1.5)

^} = l(Vl7±3).
The first series converges for large values of |z|; the second, for small
values. Here a and b are the arbitrary parameters. These are rather
simple types of psi series.
At the origin we also encounter logarithmic psi series. There are
expansions of the form

ip„(logi)z", (7.1.6)
where Pn(s) is a polynomial in s of degree [|(n + 1)]; furthermore, Po(0)
and Pi(0) are arbitrary constants.
All this happens at the fixed singularities. Naturally the mobile ones are
no better. Suppose that z = a is such a point, say a real positive. The
test-ratio test shows that w(z) becomes infinite as 6a(z - a)~2[\ + o(l)],
but the singularity is not a pole of order 2. Lacking a better term, we may
call it a pseudopole of (apparent) order 2. At such a point we have
logarithmic psi series of the form

n2=0 P„[log(z-fl)](z-ar2, (7.1.7)

PSI SERIES 251

where now P„(s) is a polynomial of degree [In] and P6(0) is arbitrary.

This survey indicates that psi series are indispensable for the theory of
nonlinear DE's. Such series are known from the theory of systems of the
type
dxx dx2 dxk ,
-xrxi="-x-k=du (7-L8)
where the Xj's are functions of X\, x2, . . . , xk, t holomorphic for small
values of the variables and vanishing for jci = x2 = • • • = xk = 0. See the
monograph of H. Dulac (1934) and pp. 315-347 of J. Horn (1905).
Psi series give rise to a number of general questions in the mind of an
analyst. Here are some:
1. What analytic functions may be represented by a psi series of a
given type?
2. Given a psi series, what is its domain of absolute (nonabsolute,
uniform) convergence?
3. Does the function represented by the series necessarily have a
singularity on the boundary of the domain of convergence?
It is easy to ask questions, but answering them is another matter. To the
basic question 1 an obvious and trivial answer is: certain classes of
solutions of selected nonlinear DE's. We obviously bypass the heart of
the question, namely, What analytic properties of the solutions demand
representations of this type? The author does not know the answer.
In regard to question 2, a little more can be said. Given a series

i>„z\
0 (7.1.9)

where
numbersthesuch
A„'s that
form a strictly increasing sequence of nonnegative real

log n (7.1.10)

with n, an obvious extension of the Cauchy-Hadamard criterion shows

that if

limsup|a„|,/A" = /x, (7.1.11)

the series (7.1.9) converges absolutely for

\A<K (7.1.12)

and the terms are unbounded for |z| > l//ot. Nonabsolute convergence may
252 REPRESENTATION THEOREMS

possibly take place for a z with |z| = The function represented by the
series is normally infinitely many valued but becomes single valued on the
Riemann surface of the logarithm of z.
If z^f(z) admits of a representation of type (7.1.9), it is unique, i.e.,
there is no nontrivial representation of zero by a series of type (7.1.9).
This is proved in the same manner as for an ordinary power series by
showing successively that all coefficients must be zero.
The same type of argument applies to series of the form

m2=0 ni=0 amnzm+\ (7.1.13)

where the A„'s satisfy the conditions listed above. If the series has
bounded terms for some value of z ^ 0, say z = a, it is absolutely
convergent for |z|<|a|; hence the domain of absolute convergence is
again of the form |z| < R, -«> < arg z < +°°, for if

\amn\\a\m+x»<M and |z| = r|a|, (7.1.14)

where 0 < r < 1, then
\amn\\z\m+k»<Mrm+K» (7.1.15)
and

m2=0 nir"+A"=(l-r)-'|>\
=0 n =0 (7-1-16)

Here the A-series is convergent by the A-root test by virtue of (7.1.10).

Again the representation of a function by a series of type (7.1.13), if at
all possible, is unique, at least if no A„ is an integer, for if now /(z) is
identically zero and is represented by a series of type (7.1.13), each of the
series 2"=i amnzx" must be a null series and this requires that all coeffi-
cients amn be zero. If the A„'s can be integers, the representation is not
necessarily unique.
We come finally to series of the form

SF^log^z", (7.1.17)
where the functions Fn(s) are entire functions of order (3 < 1. More
precisely, there should exist positive numbers a, j3, A, and B, independent
of n, such that

\Fn(reie)\<(A + r)an exp(MBr/3), 0< j3 < 1, (7.1.18)

for all n. Such a condition is obviously satisfied if the Fn 's are polynomials
in 5 of a degree which is dominated by a linear function of n. To discuss
INTEGRAL REPRESENTATIONS

such a series we make the change of variables

s = log - or z = e .
The series then becomes i 1

n=0

SFn(that
Set z = x + iy, s = m + iu, and note ns.interval 0 < x < 1 corresponds
s)e-the
to 0< u <°°. The series (7.1.19) will be absolutely convergent if

(A + r)a exp (Br** - r cos 0) < 1, 6 = arg s,

or

a log (A + r) + Br'3 < r cos a (7.1.20)

This inequality determines a domain D in the s-plane. Any ray arg s = 0
with -\tt < 6 < + \tt has the property that its distant part belongs to D.
The transformation z = e~s maps D conformally on a domain S spread
out on the Riemann surface of log(l/z). Each sheet of the surface
contains a disk \z \ < rn which belongs to S. Here r„ goes to zero with \l\n\.
If a function f(z) admits a representation by a series of type (7.1.18), it
is unique, for if the sum of the series is identically zero, then each
function Fn(s) must be identically zero; this requires that the Taylor
coefficients of each F„(s) be zero so only a trivial null series can exist.
Question 3 remains. For series of type (7.1.9) condition (7.1.10) is
necessary for the existence of singularities on the boundary of the domain
of convergence; however, it is not clear whether this condition is
sufficient, and for the two other types of psi series no information appears
to be available.

EXERCISE 7.1

1. Try to derive (7.1.6). Set log (1/z) = s.

2. Verify the claims of uniqueness for the series (7.1.9), (7.1.13), and (7.1.17).

7.2. INTEGRAL REPRESENTATIONS

We encountered several integral representations in Chapter 6. Here a

general theory will be given, and as preparation some definitions and
identities are needed. We shall be concerned with linear nth order DE's.
The variables are complex, though with suitable restrictions the notions
are meaningful also for real variables and are particularly useful for
254 REPRESENTATION THEOREMS

boundary value problems. Let the DE be

L(w)=j^Pk(z)w{k) = 0, (7.2.1)
k =0
where the coefficients are locally analytic functions. The expression

L*(w)=j?(-\)k[Pk(z)wr
k =0 (7.2.2)
will be called the adjoint of L(w). We observe in passing that in the
theory of boundary values the coefficients Pk are normally replaced by
their conjugates in defining the adjoint. We shall not do so here. L(w) is
self-adjoint if L(w) = L*(w), and anti-self-adjoint if L*(w) = -L(w).
We have by partial integration

f U^Vds
J =(-l)p J
f UV(p)ds + j+k=P-i
2 (-iyU™VQ\ (7.2.3)
This gives

j vL(u)ds = J [2 Pk™(k)] ds = 2 | u(k\Pkv) ds

= "Zu(k l\PkV)-yZ j uik-l\pkvy ds

= "Zuikl\pkv)-"Zuik-2\Pkvy

+ 2 \ u^2\PkV)"ds. (7.2.4)
After n integrations by part we have

j vL(u) ds = j uL*(v) ds + L*(u, v), (7.2.5)

where L*(m, v) is a bilinear form in the m's, the iTs, and their derivatives.
We have

L*(u,v)="2m = 1 j+k 2=m — 1 (-iy(Pmt00Vk). (7.2.6)

The corresponding relation

uL(m) — uL*(v) = L*(m, u) (7.2.7)

is known as Lagrange's identity after J. L. Lagrange (1736-1813), to

INTEGRAL REPRESENTATIONS 255

whom several other identities are also accredited. The expression

j [vL(u)-uL*(v)]ds =[L*(u,v)t (7.2.8)

is known as Green's formula after Cambridge don George Green (1793-
1841), whose important contributions to potential theory, electricity, and
magnetism were overlooked during his lifetime.
We now take up the problem of solving a linear DE (7.2.1) by means of
a suitable integral of the form

t)F(t)dt, (7.2.9)

where the kernel K(z, 0 is given, subject to certain conditions to be

stated later, whereas F(t) and the path of integration C have to be
determined. It is assumed that K(z,t) is locally holomorphic and that the
differential operator L = Lz may be taken under the integral sign so that

L2[J*c K(z,t)F{t)dt] = Jc Lz[K(z,t)]F(t)dt.

It is further assumed that a differential operator M = M, exists such that

Mt[u] = kj?gk(tWk)(t)
=0 (7.2.10)

and has the following properties: (i) the functions t gk(t) are holomor-
phic in some neighborhood N(C) of the path of integration C in the
f-plane, except possibly at the end points of C if C is not a closed arc; and
(ii) there exists a kernel Kx{z, t) which is locally holomorphic in the two
variables and such that

Lz[K(z,t)] = Mt[Kl(z,t)l (7.2.11)

and this gives

jc Lz[K(z,t)]F(t)dt = jc Mt[Kx{z,tWit)dt. (7.2.12)

Here we can use Lagrange's identity. If M?(w) is the adjoint of Mt(u)
and M?(w, v) is the corresponding bilinear form, then

Lz(w) = £ K,(z, f )M?[F(t)] d* + j^[MnKlyF)]dt (7.2.13)

If now F(t) is a solution, not identically zero, of the DE

M*(m) = 0, (7.2.14)
256 REPRESENTATION THEOREMS

and if the second integral in (7.2.13) vanishes for a suitable choice of C,

then the function w(z) of (2.7.9) is a solution of the DE (7.2.9).
Among the special kernels K(z, 0 used in this connection, the follow-
ing are the most important:

(z-tY~l Euler,
ezt Laplace,
K(zt) Mellin,
(-z)1 Barnes.
Various observations are in order concerning the choice of the path of
integration C. Here C may be an open arc or a closed curve. Suppose that
C is open and goes from t = a to t = b, where one or both points may be
infinity. If M?(Ki, F)\t=a is to be zero, this requires that t =a be a
singular point of (7.2.14), for
n «m n—m —1

M?(X„F)= m=0 CH )] p=0

2 ^prlKx(z,t 2 (-l)p(gP+m+1F)(p),
and if the partials of Xi with respect to t actually depend on z, then
M*(K1,F)|t=a can be identically zero iff
n—m =1
2 (-l)p(gp+m+,F)"" = 0 (7.2.15)
holds for t = a and m = p=0 0, 1, 2, . . . , n - 1. If g„(a) ^ 0, this requires that
F(a) = F'(a)= • • • = F(n~l\a) = 0 and F(O = 0 contrary to the assump-
tion. If the path C extends to infinity, so that b = oo, a similar argument
shows that f = oo is a singularity of (7.2,14) if it is required that Mf(Ki, F)
vanish as f -> oo.
If C is a closed contour instead, C must surround a singular point of K
or of F, for if the integrand in (7.2.9) is single valued and holomorphic
inside C, the integral is identically zero and is of no interest for our
problem. If the integrand is not single valued, it is essential that the
integrand return to its original value after t has described C. Here double
loops may be useful (see Figure 6.2).
Double integrals have also been pressed into service for the representa-
tion problem. Here naturally the associated DE is a partial DE, say,

A(5'°^ + B(5'°lf+C(s't)l7 + D(s'OM =0- (7-2-16)

We denote the differential operator by Ms,t. Suppose that two kernels,
K(z, 5, 0 and K*(z, s, t), exist such that
Lz[K(z, s,t)] = Ms,t[K*(z, 5,01- (7.2.17)
INTEGRAL REPRESENTATIONS

Furthermore, let F(s, t) be a solution of the partial DE

(7.2.18)
M1Av) = ^fzT(Av)-^-(Bv)-^-(Cv) + Dv = 0
Then
(z) = J
/
w (7.2.19)
Jc, Jc2 K*(z, s,OF(s,t)ds eft
is a solution of (7.2.1) provided that the paths C\ and C2 can be chosen so
that M*(K*,F) vanishes in the limits. Details are left to the reader.
Finally, in connection with the Barnes integral,

(7.2.20)

it should be observed that the functional equation to be satisfied by F(t) is

normally not a DE but a linear difference equation instead. Conversely,
certain classes of linear difference equations with polynomial coefficients
are Mellin transforms of functions which satisfy a linear DE. See Section
7.6.

EXERCISE 7.2
1. Verify (7.2.6).
2. Verify (7.2.13).
3. Find when (7.2.19) is a solution of (7.2.1).
4. In forming a solution of type (7.2.9) it is not necessary that F(t) be the general
solution of (7.2.14) — any particular solution will do. On the other hand, it may
be advantageous to consider expressions like

where the A 's are constants, the F's are solutions of (7.2.14), and the path Ck
is adjusted to the solution Fk. Under what conditions will this expression be a
solution of (7.2.1)?
5. Consider the special hypergeometric equation
(\-z2)w"-(a + b + \)zw' -abw = 0,
and form the double integral

Show that this converges for Re (a) > 0, Re (b) > 0, Re (z) < 1 and defines a
solution of the equation in this half -plane. Show that the solution is a constant
multiple of w0i(z). Reconstruct the argument which led to this integral.
258 REPRESENTATION THEOREMS

7.3. THE EULER TRANSFORM

Quite an extensive theory is built around the Euler transform; Ludwig
Schlesinger (1864-1933) devotes 120 pages to this topic in his monumental
Handbuch der Theorie der linearen Differ entialgleichungen, Vol. 2, 1897.
Incidentally, he coined the term Euler transform in imitation of "Laplace
transform," introduced by Poincare. Here we are forced to squeeze an
abstract of the theory and some of its applications into a fraction of the
space that Schlesinger had at his disposal.
We write the transform

(t -zf lF(t)dt = w(z), (7.3.1)

where £ is an arbitrary parameter at our disposal, and the problem is how

to choose the path C, the function F, and £ so that w(z) becomes a
solution of the linear DE:
n
L(w)= kX-0 Pk(z)w(k) = 0. (7.3.2)

The Pk 's are polynomials in z of a degree at most equal to the integer p.

Further restrictions will be introduced later.
We start by forming

Ld(t ~zf x] = k2=0 (- Dkk \Pk(z)C^k(t -zf~k'\ (7.3.3)

where Q-i,k is the /cth binomial coefficient of £ - 1. Here we expand Pk(z)
in powers of t — z:

P*(z)= i2?^ (-!)"« -2)".

This gives

k=om=o
Mtt-z)«-i] = s m
£ (-irwQ-,,fc^^( f-z)*
We set

^■(-iTlcttifcrtQ (7.3.4)
with

C(£j,fc,p) (g-lXl-2)---(f-fc)
« + pXf + p-D---(£ + p-J)
THE KULKR TRANSFORM
259

and obtain

M(t-z)4 ']=S/i(0^[(t-z)"" ']• (73.5)

We then set

M(o)=2/j(0»0>(0. (7-3-6)
I o
This gives

LA(t-zY *] = Ml[(t-z)i " ']. (7.3.7)

Lifting a term from the theory of Abelian integrals, Schlesinger refers to
this formula as the theorem on interchange of argument and parameter,
for on the left side t is the argument and z the parameter, while on the
right side z is the argument and t the parameter.
It follows that the associated function F of (7.3.1) should be a solution
of the linear DE

M1(v) = 0, (7.3.8)
which is of order n + p.
The singular points of (7.3.1) are the zeros of the polynomial P„(z),
while those of (7.3.6) and hence also of (7.3.8) are the zeros of /„ ,,,(*)•
Now (7.3.4) shows that tn(t) is a constant multiple of P„(0> so the two
equations have the same singular points.
Suppose then that F(t) is a solution, not identically zero, of (7.3.8). It
follows that w(z), defined by (7.3.1), is a solution of (7.3.2) provided that
the path C can be chosen so that

j ^Mm-zf " ',F(t)]d*=0, (7.3.9)

for in that case we have

Lz[j (t-zY lF(t)dt] = f (t-z)€+p lM*\F(t)]dt =0.

Various types of paths of integration C are at our disposal. They all
involve the singular points of the DE's, i.e., the zeros of P„(z) and the
point at infinity. Let the distinct zeros of P„(z) be a,, a2, . . . , av. Mark
these points and the point z in the complex plane. Denote by [jc, y] an
integral from t = x to t = y, denote by [jc, y) a simple loop starting at t = x
and surrounding t = y in the positive sense, and, finally, let (jc, y) stand
for a double loop surrounding jc and y. Here x and y take on values from
the set S: aly a2, . . . , a„, z, <». Double loop integrals always exist. The
260 REPRESENTATION THEOREMS

existence of other types will depend on the nature of the singularities,

whether regular or irregular, and in the first case the roots of the
corresponding indicial equation. In all cases care must be taken to ensure
that Mf[(f - zf+p~\ F(t)] returns to its original value after the path has
been described.
It should be noted that in order to form the Euler transform (7.3.1) we
do not need the general solution of (7.3.8). Any particular solution will do,
since the variety of choices for C provides enough freedom to enable one
to find fundamental systems of solutions of (7.3.1). Furthermore, the
parameter £ is at our disposal and can occasionally be used to simplify the
work.
Further simplifications are available if the equation is of the Fuchsian
type or, merely, the point at infinity is regular or regular singular. As will
be seen in Section 9.6, the point at infinity will have the desired property
iff the degrees of the coefficients

P„, Pn-l, . • • , Po
are dominated by a strictly decreasing sequence. Hence, if the degree of
Pn is exactly p, the degree of Pk is at most p - n + k. But then for j < n

By (7.3.4) this implies that the functions

/0(0,/l(0,...,/n-l(0
are all identically zero, and the DE (7.3.6) takes the form

Mt(u) = i=o
f,fn+i(t)^j[u(n\t)],
at (7.3.10)
so the equation is of order p with respect to the nth derivative of u.
Further simplifications will take place if the equation is of the Fuchsian
type, but we do not intend to pursue the general theory any further. In the
rest of this section applications to Legendre's equation are given as
illustrations. The general hypergeometric case will be discussed in the
next section in some detail.
Given the Legendre equation,

Lz(w) = (\-z2)w"-2zw' + a(a + l)w =0, (7.3.11)

which is of the Fuchsian class, it is desired to find solutions of the form
(7.3.1) for a suitable choice of the parameter £. Here it is found that M,(w)
is simply

M,(M) = (l-r>" + 2(£- \)tu' + [a(a + !)-£(£- l)]w =0.

THE EULER TRANSFORM 261

If we choose £ so that the multiplier of u is 0, which means either £ = a + 1

or - a, we get for £ = — a
Mt(u) = (1 - t2)u"-2(a + \)tu' = 0, (7.3.12)
and the adjoint equation becomes

Af?(u) = (1 - r V + 2(a - l)tv' + 2au = 0. (7.3.13)

This expression is an exact derivative and

M?(w) = [(1 - r V + 2atv]' = 0. (7.3.14)

a particular solution of which is our F(t):

F(t) = (\-t2)a. (7.3.15)

Thus, if the path C satisfies the appropriate conditions,

w(z) = j (t-zya \\-t2)adt (7.3.16)

is a solution of Legendre's equation. If Re (a) > - 1, we can take [-1, 1]
for C. If a is not an integer, we cut the plane along the real axis from - 1
to + 1 ; the integral is taken along the upper edge of the cut with
(1 - t2)a = 1 for t = 0 and (t - z)~a~x defined in an obvious manner. Now

(r-z)—1 = (-z)-a-1(l-|) U \ |z|>l,

and here the second factor may be expanded in a binomial series.
Multiplication by (l-r2)fl and term-by-term integration are permitted.
The result is of the form (-z)~a_1 times a power series in z-2, so the
integral is a constant multiple of Qa(z). To get a representation of this
function valid for all values of a, one may use a double loop (- 1, 1), but
actually it is sufficient to use a lemniscate-shaped contour surrounding
t = + 1 in the positive sense and t = — 1 in the negative. The right half of
the circuit multiplies the integrand by e2lTia, and the left half by e ~27Tia, so
that we return to the starting point, say t = 0, with the original value of the
integrand.
If a = n, a nonnegative integer, one may take for C the circle
\t| = r > 1. For 2 inside the contour of integration, Cauchy's formula for
the nth derivative, together with Problem 6.2:1, shows that

(-2)^1t^7)^=p"(2)- (73-17)
The case a = — \ is important on account of its connection with the
theory of elliptic functions and modular functions. It is convenient to
262 REPRESENTATION THEOREMS

make a change of variables, so that the integral is taken to be

j [t(i-t)(t-z)]'mdt. (7.3.18)
Here all paths of integration are permissible. If s is an arbitrary point in
the z-plane, then [1,5] is an elliptic integral of 5 of the first kind in the
terminology of Legendre, who, however, spoke of elliptic functions in his
Traite des fonctions elliptiques of 1825. Two years later N. H. Abel and C.
G. J. Jacobi (1804-1851) showed independently of each other that,
while the integral

jS [r(l - t)(t - z)Ym dt = u (7.3.19)

is an infinitely many-valued function of s, the inverse function s = s(u) is
a single-valued meromorphic and doubly periodic function of u. For this
class of functions the name elliptic function is commonly accepted, while
(7.3.19) has become known as an elliptic integral It is of the first kind in
the sense that it is everywhere bounded. Integrals of the second kind are
obtained, for instance, by inserting a factor t in the integrand. Such
integrals become infinite with s. The inverse function is single valued; it is
not doubly periodic but has additive periods. The third kind has logarith-
mic singularities.
The periods of (7.3.19) are of the form 4mK +2m'K', where

2K= J [t(t - l)(t - z)Tm dt, (7.3.20)

UK' = j [t(t - \)(t - z)Ym dt. (7.3.21)

Here K = K(z) and K' = K'{z) are holomorphic functions of z, real
positive for z real between zero and one. More precisely,

K(z) = l7rF(U,l;z), (7.3.22)

K'(z) = K(l - z) = |ttF(U, 1; 1 - z). (7.3.23)
That the periodicity modules satisfy a Legendre equation for a = -\ was
first proved by Legendre.
The quantity

T=^Yi (7.3.24)

has a positive imaginary part, for any z ^ 0, 1, oo. This is obvious for the
case here considered, where z is real and between zero and one, but it is
THE EULER TRANSFORM 263

also generally true. Set

q=e™. (7.3.25)
Since Im (t) > 0, it is seen that \q \ < 1 , so that series expansions in powers
of q converge rapidly. Such expansions are basic for the theory of theta
functions, introduced by Jacobi in 1829. A typical theta series is

fl3(z, q ) = 1 + 2 2 q "2 cos 2nz, (7.3.26)

which is an entire function of z. All the theta functions have a bearing on
the theory of heat conduction — #3, in particular, with the so-called ring of
Fourier. Jacobi proved further that z1/4 is the quotient of the theta null
series,

#3(0, q)
2"4 = fM- (7.3.27)

In the Jacobi theory z 1/2 is known as the modulus of the corresponding

elliptic functions, usually denoted by k. The function defined by (7.3.27),
known as the elliptic modular function, is a single-valued function of t
defined and holomorphic in the upper r-half-plane. It is obtained by
inversion of the quotient of two fundamental solutions of Legendre's
equation. It is invariant under the so-called modular group, a group of
Mobius transformations of the upper half-plane into itself. Fuchs called
attention to the fact that similar considerations apply to more general
equations of the Fuchsian class. Poincare elaborated this question into an
extensive theory of automorphic functions: single-valued functions,
invariant under a group of Mobius transformations, and normally with a
natural boundary invariant under the group. Poincare, always generous in
naming the problems he considered after the author whose work had
inspired him to delve deeper into the matter, termed these functions
fonctions Fuchsiennes. Felix Klein (1849-1925) protested against that
name, claiming that Fuchs had not considered such functions whereas
he (Klein) had. Poincare responded by calling the next class of automor-
phic functions which he discovered fonctions kleiniennes, since Klein had
not considered them!
Incidentally, Picard built the proof of his theorem on the properties of
the elliptic modular function (7.3.27): the existence of an entire function
omitting the values 0, 1, oo would imply the existence of an entire function
whose values have a positive imaginary part — an obvious impossibility.
All this goes to show the importance of Legendre's equation and of the
Euler transform. The latter also adds to the understanding of hyperelliptic
and more general Abelian integrals. Thus, if the integer p > 1 and if the
264 REPRESENTATION THEOREMS

a's are 2p distinct complex numbers, then

jS [(t - fll)(t - a2) • • • (t - a2p-x)(t - a2p)(t - z)]"2 dt (7.3.28)

is a hyperelliptic integral of s of the first kind and of genus p. It is
infinitely many valued, and so is the inverse. The periodicity modules are
integrals of the form [ah aj+i], j = 0, 1, . . . , 2p - 1, where a0 = z. They
satisfy a linear DE of order 2p and of the Fuchsian class. The interested
reader is referred to Schlesinger's Handbuch, Vol. 2, pp. 489-524, for
further details.

EXERCISE 7.3

1. Verify (7.3.3)-(7.3.5).
2. Verify (7.3.10) and preceding matters.
3. Verify (7.3.12)-(7.3.15).
4. Verify (7.3.22) and (7.3.23).
5. What is the DE satisfied by (7.3.18)?
6. Verify that

£ (t -z)~a~\\-t2y dt
is a function of a times Qa(z). Find the multiplier.
7. Consider (7.3.16) for a not an integer and a > -5. Let C be a simple loop
starting and ending at t = - 1 and surrounding t = 1 and t = z once in the
positive sense. Assuming that |z - 1| < 2, expand the integral in powers of
1 - z. Show that it is a constant multiple of Pa(z), and find the multiplier.
8. Show that F(U, l;z)^0 for z real, 0< z < 1.
9. Consider the quotient
K(z)
K(\-z)
r(z) = i
Show that it maps the upper z-half-plane conformally upon a curvilinear
degenerate triangle above the circle |t - {\ = \ and between the lines Re (t) =
0 and Re (r) = 1. There are three horn angles.
10. The theta function #3(z, q ) is defined by (7.3.26). The other Jacobi thetas are

#,(z, q) = 2 2 (- D" <7(2"+,)2/4 sin {In + l)z,

#2(z, q) = 2 2 <z(2n+1)2/4 cos (2n + l)z,

#4(z, (? ) = 1 + 2 2 (- 1)" q "2 cos 2nz.

hypergeometric euler transforms 265

Show that all four functions are entire functions of z of order 2 for fixed
<?,kl<i.
11. Estimate T(r, F) for #3(z, q) if q is real, 0<q < 1.
12. Show that the thetas have period it or 2tt. Show that di(z + itt, <?) =
</ 1 e :' rh(z, q ), and find similar relations for the other thetas.
13. Show that the quotient of two thetas is an elliptic function.
14. The theta null series are obtained by setting z = 0 in the expression for
dj(z, q), j = 2, 3, 4. They are analytic functions of t, defined in the upper r-
half-plane. If t is purely imaginary, try to estimate the rate of growth of
ft3(0,<7) as q->l.

7.4. HYPERGEOMETRIC EULER TRANSFORMS

Let us consider the hypergeometric DE

z(l -z)w"+[c -(a + b + \)z)w' -abw = 0. (7.4.1)

Here we know that the solutions are representable as Euler transforms.
The object of the present section is threefold: (i) to derive the integrals by
a more elementary method, (ii) to identify the various integrals with the
known series solutions, and (iii) to use the integrals to obtain some of the
transit formulas.
For these problems the Euler beta function is essential. For Re (u ) > 0,
Re (v) > 0 we have

jV-(l-Sr'dS=^. (7.4.2)
If instead of an integral [0, 1] we use a double loop (0, 1), we get

f i)
J<o, s" '(1 - sf-1 ds = (1 - e 27Ti")(l -e 2"iv)Tr\u)I}Vl
1 {u -r v ) (7.4.3)
This becomes indeterminate if u or v or both are negative integers or
zero. We get the correct value by replacing the right-hand member by its
limit as u or v approaches the value in question. In either case it is finite,
and it is zero if both u and v approach nonpositive integers.
As in the proof of Lemma 6.1.1, we start by finding an integral
representation for the hypergeometric series F(a, b, c ; z) when |z|<l
and c is not a negative integer or zero. We note that for \zs \ < 1

(l-zs)- = i^(zs)".
For a fixed z, |z| < 1, this converges uniformly with respect to s on the
interval [0, 1], and the series may be multiplied by an integrable function
266 REPRESENTATION THEOREMS

of s and integrated term by term between zero and one. We choose this
function as su \\ - s)v~\ where u and v are to be disposed of later and
for the moment only satisfy Re(w)>0, Re(u)>0. This gives succes-
sively

JO
f1s"W - sYW - zsya ds=2n=0 ^
\ l)nzn JOf su+n-\\ - sT ds

= ^ (fl)wr(n + w)r(i?) „
„40(l)n Y{u+v+n)Z
_Y(u)T(v)y (a ),(!!),
T(u + v) £o(l)n(u+v)n

Here we choose u = b, v = c — b to get

T(c-b)T(b)
s
\\ - sY b \\ - zs)'a ds = YicV'" F(a' b' C ; Z>' (7A4)
where we have to assume that Re (b) > 0, Re (c - b) > 0. If these condi-
tions are not satisfied, the integral [0, 1] must be replaced by a double loop
(0, 1) with corresponding modification of the multiplier.
The result is not yet of the Euler transform type but becomes of the
canonical form by the change of variable t = lis. This gives

r-C(t _ iy-b-i(t _ zya dt = T(c *)r(fr) F(fl> bf c ; z). (7.4.5)

The conditions on the parameters are unchanged. Now this representation

is valid in the whole z-plane cut along the positive real axis from one to
plus infinity.
We can now consider the integrals

(t-l)c-b-\t-zradt. (7.4.6)
[*,y] = JV
Here we have six possible choices for [x, y], valid under different
conditions on the parameters. The list below gives the different choices of
path and the corresponding conditions on the parameters:

[0, 1] Re (a -c)>-l,Re(c -b)>0;

[0,z] Re (a - c) > - 1, Re (a) < 1;
[-oo,0] Re (a -c)>-l,Re(k)>0;
[l,oo] Re(c)>Re(fc)>0;
[l,z] Re(a)< l,Re(b)<Re(c);
[z,oo] Re(a)< l,Re(b)>0.
HYPERGEOMETRIC EULER TRANSFORMS
267
We start our survey with a brief look at the integral [0,1]. For large
values of z it is clearly 0(z"), which suggests that it is a constant
multiple of Wooi(z). It is easy to confirm this by series expansion, and the
multiplier is obtained by multiplying the integral by z° and passing to the
limit with z, thus obtaining

,<c— b-.,r(a-c + l)r(c-fr)

[0, 1] = gw— » v" r(a - b + i) VV~ l(2)- (7A7)
The DE is not altered if a and b are interchanged, and this shows that
an interchange of a and b in an analytic expression for a solution
preserves the property of being a solution. This procedure does not
always give a new solution. In the present case, however, it does, and we
have

th~c(i-ty~a~\t-zyb dt

^c-q-»-i)r(fr-c + i)r(c-q)
e r«>-a + i) w"Az)- (7A8)
Our next item is the integral [0, z]. Here a change of the variable of
integration, setting t = zs, gives

, z] = zl c I sa~c(sz - l)c-b-\s - \ya ds,

and the integral is a holomorphic function of z at the origin. This shows
[0
that [0, z] is a constant multiple of w0,2(z). Letting z -»0 in the integral
gives the multiplier and shows that

^(c-a-b-D r(a - c + 1)T(1 - a)

[0, z] = e-*— f(2-fl) Wo'2(z)- (7A9)
The fact that the integral [- <», 0] tends to a finite limit as z -> 1 and
appears to be holomorphic there would indicate that [- oo, 0] is a constant
multiple of wn(z). This is indeed the case; and

[-^H-£«,,(2). (7.4,0)
The case [1,<»] served as our starting point, so we can pass to the
integral [l,z]. Here the substitution r-l = s(z-l) gives (l-z)c_a b
times a function holomorphic and different from zero at z = 1. The usual
steps give

[l,z] = flrttc^''r(c-a)-fc + l))",'3(2)- (7A11)

268 REPRESENTATION THEOREMS

Our final item is [z, <»]. Here the transformation t = z/s shows that
[2,00] is a multiple of Woo,2(z). Thus

wi(c-«-«,-i)r(b)r(l - a)
fr.-]-'™- -"lV-a + (7A,2)
Again, by interchanging a and b in the integral, we get an alternative
representation of Woo,i(z). This finishes the discussion of the six (eight)
individual integrals.
We can also obtain transit formulas from the integral representations.
For this see Figure 7.1. There are six paths of integration for z nonreal.
They divide the f-plane into four triangular regions, of which three are
degenerate. If Re(b)>0, integrals along arcs of the circle \t\ = R go to
zero as R -»<». By the usual argument in complex integration it follows
that the sum of the integrals along each of the four triangular boundaries
is zero. This gives four sets of homogeneous linear equations between six
unknowns, each involving three terms. We can write the set as follows:
h = [0, 1],

J. + 0 +I3 + I4 + O +0 =0, I2 = [0, z],

1,-/2 + 0 +0 +J5 + 0 =0, J3 = [-°o,0], (7.4.13)
0 +/2 + /3 + 0 +0 +I6 = 0, l4=[l,00]f
0 +0 +0 + I4-I5 + h = 0. J5 = [l,z],
h = [Z, 00].

In the 4-by-6 matrix all 4-by-4 minors are zero, but there are 3-by-3
minors which are not zero. There are obviously only three independent
equations. This means that we get relations between three solutions, each
of a different fundamental system. These are not transit formulas in the
ordinary sense, nor do they readily lend themselves to the obtaining of
ordinary transit formulas.
Before leaving the applications of the Euler transformation to the
hypergeometric equation, let us note that (7.4.5) yields the value of
/(a, b, c) when Re (c - a - b) > 0. We can then let z -> 1 to obtain

Kt - 1) dt = j^— j /(a, b, c).

Here the left member equals

T(b)T(c -a -b)
T(c-a)
HYPERGEOMETRIC EULER TRANSFORMS 269

so that

f^b,c) = nc_a)nc_by (7.4.14)

in agreement with (6.1.29).
For a representation of hypergeometric functions based on uniformiza-
tion by theta functions, see Section 10.5.

EXERCISE 7.4
1. Assuming (7.4.2), prove (7.4.3) when both are meaningful and use analytic
continuation.
2. If (7.4.4) is replaced by a double loop (0, 1), how should the right member be
changed?
3. The integral [1, °°] of (7.4.5) satisfies the hypergeometric DE by construction.
Why do the other integrals [jc, y] satisfy? The three integrals involving z in
one of the limits may require special consideration.
4. Verify the limits on the parameters needed for the existence of the
integrals [x, y].
5. Discuss (7.4.7).
6. Try to find some justification for the interchange principle used in deriving
(7.4.8). Permanency of functional equations?
 7 ^ REPRESENTATION THEOREMS
^ 2
7. Obtain (7.4.9) by series expansion.
8. Obtain (7.4.10) and (7.4.11) by series expansion.
9. Verify (7.4.12).
10. Find a 3-by-3 minor, not zero, of the 4-by-6 matrix (7.4.13).

7.5. THE LAPLACE TRANSFORM

This transform was encountered in Section 6.4 in the second order case.
Here we shall be concerned with a linear nth order DE with polynomial
coefficients

Mw) = k=0
2 Px(z)w(k\ Pk(z) = p=0
2 Cp*zp, (7-5-D

where Cmn = 1, so that P„(z) is of exact degree m while the other Pk's
may be of lower degree. It is desired to find solutions of this equation of
the form

w(z) = j eztF(t)dt (7.5.2)

for a suitable choice of F and of the path of integration C. Here

Lz{ezt)=^Pk(z)tkezt. (7.5.3)
Jc=0
This equals

k=Op=0
LAezt)=j? 2 Cpktkzpezt

p=0 (11 fc=0

where

Mt(u) = p=0
2 Qp(0"(p) with QP(0 = fc=0
2 Cpkt*. (7.5.4)
The adjoint is

Mf(M) = 2 (- 1)'-^- [<J(Om] (7.5.5)

with the associated bilinear differential form

[/, g] = 2 2 (- 1)' [Qi (Og] ^fU (/). (7.5.6)

Here we set

/ = «", g = F
THE LAPLACE TRANSFORM
271

and apply Lagrange's identity,

FLZ (ezt)-e Z,M*(F) = J- M*(e z\ F),

to obtain

L2||c eztF(t) dt} = jc eztM*[F] dt + £ ^ [M*(ezf, F)] dt.

If now F is chosen as a solution of

M*(w) = 0, (7.5.7)
and if C can be chosen so that

jcj{[M*(ez\F(t)]dt=0, (7.5.8)
then (7.5.2) will furnish a solution of (7.5.1) of the desired type. We note
that it is not necessary to find the general solution of (7.5.7); any solution,
the simpler the better, will do as long as it is not identically zero.
The path C has to be adjusted to the singularities of the adjoint
equation. These are the zeros of

Qm(t)= 2 Cmptp (7.5.9)

plus the point at infinity. The use of the
p=0 Laplace transform is normally not
advisable when m ^ n; only when the order of M* is less than that of L
can one expect a simplification of the integration problem. If m < n, the
point at infinity cannot be regular singular; in fact, we may expect some
transcendental entire solutions.
The case considered by Laplace in 1812 involved m = 1, so that the
adjoint equation is solvable by quadratures. Let the DE of Laplace be

2 (C,kz + Cofc)w(k) = 0. (7.5.10)

Here

2 Cipt",
Qi(0 = p=0 2 Copt",
Qo(t) = p=0 (7.5.11)
so that

"(f)=ofeexp[/oSH- (7-5i2)
Here Q0/Qi should be written as the sum of partial fractions, and there the
relation between the degrees of Q0 and Q, becomes important. Let the
272 REPRESENTATION THEOREMS

degrees of Q0 and Q, be q0 and qu respectively. If q0- q\ = q, the partial

fraction expansion will be of the form

i=o+2 i=oP 2= i Mt-o.r,

§48=2^
vMU <?^o, (7.5.13)

where Ii(r - afp = Q.(0- This gives

^ds = C + ±J^ti+^tbiAog(t-ai)

+ i=\
yZ^bjP(p-\r\t-aiypf
p=2
whence

u(t) = iU(t -aif"-'

=i

L 0 + D-r^' + j=lp=2
xexpficiO' J (7-5-14>
iE bip(P-l)"'('-ai)'"i-
Thus, if either q ^ 0 or Qi = 0 has multiple roots, n(0 will have (isolated)
essential singularities at infinity and/or at the multiple roots.
Here the simplest case is obviously q = 0, and Qi = 0 has only simple
roots. In this case we have

U(t) = fl(t-ai)br\ (7.5.15)

i= i
where is the residue of Q0/Qi at t = a,. With this choice of F(t) we can
take for C one of the line segments [ah ai+x] or the corresponding double
loop (ah cij+i). The first choice is permissible if both residues, b, and bj+i,
have positive real parts. The double loop is always permissible. We can
also consider simple loops around the points ah where the loop begins and
ends at r = oo and the path approaches infinity in such a manner that
Re(zO-»-00- This gives a fairly large choice of paths of integration.
We note that (7.5.1) has only one finite singularity, z = -C0„/Ci„, and
this point is obviously a regular singularity. The corresponding indicial
equation has the roots

0, l,2,...,n-2

and an nth root which is ordinarily not an integer. To the n - 1 integral

roots correspond n - 1 linearly independent transcendental entire func-
tions, while to the nth root corresponds a power of z times an entire
function. All these entire functions are of order 1 . For exceptional values
of the parameters Cpk polynomial solutions may occur.
THE LAPLACE TRANSFORM 273

The Laplace transformation is effective and useful when the order of

the irregular singularities is at most unity. Here the order is measured by
the maximal rate of growth of the maximum modulus for approach to the
singularity. If the order exceeds one, the Laplace transform can no longer
cope with the problem. Here George David BirkhofF (1884-1944) showed
the way to surmount the difficulties. In a number of papers starting in
1909, he proposed the generalized Laplace transform

jcexp Ups)[2 z'Ms)] ds = w(z), (7.5.16)

which he studied in some detail. Here the expected order of the
singularities exceeds p - 1 but is at most p. The p functions «,(s) and the
path C are to be determined.
Let the DE be of order p:

Lz{w) a 2 PJ(z)w0) = 0, (7.5.17)

where the coefficients are polynomials
j=0 in z and

degP^fcj^G-DP (7.5.18)
with equality for at least one value of j.
Substitute (7.5.16) in (7.5.17). The result is a conglomerate of terms of
the form

jc exp (zps)sizHp-l)+m+kum(s) ds. (7.5.19)

Here j goes from 0 to p - 1, and so does m while k ^ fc,. Each such term is
now subjected to integration by parts at most p - 1 times in order to
reduce the exponent j(p - 1) + m + k to at most p - 1. The idea is that

jc U(s, z) ds exp (zps) = [U(s, z) exp (zps)]c

[ , p , 317(5,2) J

Here dU(s,z)lds as a polynomial in z is of degree p units less than

j(p - l) + m and because of the assumptions on kj not more than
p - 1 integrations by parts are needed to lower the exponent of z to at
most p - 1. The various integrated parts are supposed to be eliminated by
the choice of C. Each time we integrate, um(s) will be differentiated. The
net result of the reduction process is of the form

jcexp (2P5)|^2 z'£f(Ko, Mi, ... , mp_,; s)J ds, (7.5.20)

274 REPRESENTATION THEOREMS

where each L, is a linear homogeneous differential expression in the m's

and their derivatives up to at most order p - 1. The coefficients are
polynomials in s. To get a solution we solve the simultaneous DE's
Lj(u0, Mi, . . . , uP-u s) = 0, j = 0, 1,2, . . . ,p - 1, (7.5.21)
and choose C so that the integrated parts vanish in the limit. Rectilinear
paths may be permissible, as well as loops and double loops. One can
integrate to infinity, provided that the line of integration is such that
Re (zps)-^-°° along the path.
Birkhoff works with matrix equations of the form
(7.5.22)

Here si(z) = (ajk(z)) is a p-by-p matrix where the diagonal elements are
polynomials of degree p and the other elements are of degree p — 1 at
most. Birkhoff claimed that a DE having an irregular singularity of rank p
at infinity could be reduced to this normal form by suitable analytic
transformations. Gaps in Birkhoff's proof were later filled in by H. L.
Turrettin in 1955 and 1963.

EXERCISE 7.5

1. Verify (7.5.4)-(7.5.6).
2. In (7.5.15) let u(t) = (r3 - 1)1/2 dt and consider

Discuss the various choices for C. Here is the solution; where is the DE?
3. Supply additional details in the discussion of (7.5.19).
4. terms
Apply of
Birkhoff's method to the equation w" - 4z2w = 0, which is integrable in
Bessel functions.
5. Find sufficient conditions so that the DE L, =0 involves only uh
6. If Lj = 0 involves only u„ when are the singularities of the equation regular
singular?

7.6. MELLIN AND MELLIN-BARNES TRANSFORMS

In the 1890's the Finnish mathematician Hjalmar Mellin (1854-1933)

devoted much attention to linear DE's involving the differential operator
(7.6.1)

which possesses the property of having the powers of z as characteristic

MFXLIN AND MELLIN-BARNES TRANSFORMS 275

functions in the sense that

$zza=aza (7.6.2)
for any constant a. The powers of the operator are defined recursively by

d;=d,d;_1. (7.6.3)
Thus dnz(za) = anza.
Now let F(u) and G(w) be polynomials in u with constant coefficients,
let n be a positive integer, and consider the DE

L2(w) = znF(dz)w + G(d2)w = 0. (7.6.4)

Let H(u) be any polynomial and K(z) any solution of the DE

znF(dz)u - H(#2)w = 0; (7.6.5)

then K(zt) satisfies the partial DE

{z " [F(#2 )] + G(#Z)}17 = {G (fl, ) + t "ff (dt )} [7, (7.6.6)

so that
L:[K(zt)] = Mt[K(zt)]
with obvious notation. Hence the integral

w(z) = | K(zfMOd* (7.6.7)

is a solution of (7.6.4), provided that v(t) is a solution of

Mf(u) = 0, (7.6.8)
where M? is the adjoint operator of Mf, and provided that the correspond-
ing associated bilinear form can be nullified. We refrain from further
details.
Let t h-> /(f) be defined and continuous for 0 =ss t ^ oo, and suppose that
tpf(t) G L(0, oo) for - 1< p < w < oo. Then the Mellin transform

SW[/](s) = |o f"7(t) dr ^ d(5) (7.6.9)

exists for 0< Re (s) < w + 1 and is a holomorphic function of 5 in this
strip and also bounded on vertical lines. The inverse transformation is

^ 1[t?1(t)==2^r/ v(s)t~Sds> 0<c<o) + l. (7.6.10)

Here the integral may have to be taken as a Cauchy principal value or,
possibly, in the (C, l)-sense. The Mellin transform has a number of
interesting properties, some of which are presented as problems below.
276 REPRESENTATION THEOREMS

The most elegant application of the Mellin transform to DE's was made
in 1908 by the Cambridge don Ernest William Barnes (1874-1955), later
the Right Reverend the Lord Bishop of Birmingham, who studied the
hypergeometric DE by this method. By Problem 6.1:24 the equation may
be written as

z(#2 + a)(#2 + b)w = #2(#2 + c - l)w. (7.6.11)

This is now to be satisfied by an integral of the form

w(2) = 2^[ jcv(s)(-z)sds, (7.6.12)

where C is a vertical line (with possible deviations to avoid singularities).
Here v (s ) has to be determined so that w (z ) satisfies (7.6. 1 1). Assuming the
right to apply the operator #2 and its powers under the sign of integration,
we obtain

ds.
jc(s + a)(s + b)v(s)(-z)s+l ds = j^sis + c - \)v(s)(-z)s
In the right member we replace s by t + 1, and write again s instead of t.
This gives

| [(5 + ci)(5+fe)u(5) + (5 + l)(s+c)i;(5 + l)](-2r+,d5 = 0. (7.6.13)

This is legitimate, and the shift of the line of integration back to C is
allowed, provided that for the appropriate v(s) we have

lims2iKs)(-zr =0. (7.6.14)

This replaces the bilinear form condition for the previously examined
transforms. Equation (7.6.14) should hold uniformly in finite vertical
strips as |t|-»o°, s = a + it, for fixed z.
Thus (7.6.12) will be a solution of (7.6.11) if (i) v(s) is a solution of the
difference equation
, . t. (S+a)(.S+fr) , .
p(8 + 1) = -(s + 1)(»+c)0(j)' a6-15)
and (ii) condition (7.6.14) is satisfied for the solution of (7.6.15) that has
been selected.
The difference equation (7.6.15) evidently is solvable in terms of gamma
functions. A solution of

u(s + l) = (s +d)u(s) is r(s+d). (7.6.16)

MELLIN AND MELLIN-BARNES TRANSFORMS
277
The equation

M(5 + 1) = -^J (7.6.17)

has various solutions, but T(-s) is the most suitable for our purposes.
Combining, we see that

vis) = ns+r^Sc)+b)n-s) (7.6.18)

is a solution of (7.6.15).
We have now to verify condition (7.6.14). Here we need to use various
properties of the gamma function. We have

r(f)r(1-<>=»

which shows that

^+iT)i = (^)"2<(2ir)"2exp(--iTi>-
Furthermore,

lim -^4~F7~r = 1

for s tending to infinity in such a manner that its distance from the
negative real axis becomes infinite. If —z = r e,e, |0| < tt, and s = a + it,
then

\(-zare
Combining, we see that there = r"e-\ m and M depending on a, b>
Y\ numbers
c, and x = sup \cr\ so that

|52i;(5)(-zr|^M(|r|+ir^ exp[-(7r-|0|)M]; (7.6.19)

and if |0| < tt - e, 0 < e, and r R < oo, the right member of (7.6.19) goes
to zero as |t|-»<», uniformly with respect to r, 0, and a. Thus (7.6.14) is
satisfied, and a solution of (7.6.11) is given by

Here it is assumed that the points - a and - b lie to the left of the path of
integration. If this is not so at the outset, the line may be deformed locally
to achieve it.
278 REPRESENTATION THEOREMS

Here w(z) is a multiple of w0i(z),

w(z) = r("|^) F(a> ^ c; 2) (7 6 21)

This can be verified for |z| < 1 by contour integration. If in (7.6.21) the
integral is taken along a rectangle with vertices at s=—h±iT and
-\ + N ±iT, the integral is the sum of the residues inside, i.e.,

^ T(a +n)T(b + n) „
h T(l + n)r(c +n) 2 •
Here we can let first T and then N go to infinity. For |z| < 1 the sum of the
residues converges to (7.6.21), and the integrals along t = ±T and
a = —\ + N go to zero as T and N become infinite. The details are left to
the reader.
From (7.6.20) we can also find the expression of w0i(z) in terms of
w=oi(z) and u>oo2(z), assuming |z|>l and a-b not congruent to zero
(modulo 1). This requires another contour integration argument. This time
we integrate along the boundary of a rectangle with vertices at 5 = —\±iT
and s = —i — N±iT. Here it assumed that the poles of T(a + s) and
T(b + s) are avoided by the contour, so that all poles are ultimately inside
the contour. The integral is the sum of the residues. At s = — a — n the
residue is

r(c - a - n)Y(n + 1)
Interchanging a and b, we get the residue at 5 = - b - n. We let T and N
tend to infinity; this makes the integrals along r = ±T and or = - \ - N go
to zero, while the sum of the residues converges. The various gamma
functions of the type T(u - n) are reduced to

T(u-n) = (-\)n+l sin utt Y{u + 1 + n)'

For n = 0 the same formula is used to eliminate the sine functions in the
final result. After much simplification we emerge with the final result:

T(a)T(b) ^ , , r(q)r(q - b) , ,_acv , L±1

— Y^yJ-F(a,b,c;z)= r(fl _c) F(a, a - c + 1, a - b + 1 ; z l)

+ ~}a) (-z)bF(b, b - c + 1, fr-a +l;z~1). (7.6.23)

Barnes could also obtain the transit formula expressing w0i(z) in terms
of w,,(z) and w,2(z), but for this the reader is referred to the presentation
in Whittaker and Watson, A Course in Modem Analysis, Chapter XIV.
MKIXIN AND MELLIN-BARNES TRANSFORMS 279

The Barnes integral representations for hypergeometric functions have

been extended to the abstract case of (6.1.8) by Hans Wilhelm Burmann
(1971), who also derived the transit formulas which the present author
was unable to find in his discussion in 1968.
The use of the Mellin-Barnes transform is not limited to the
hypergeometric equation. In fact, any equation of the form (7.6.4) can be
handled in this manner, though the corresponding difference equation for
v(s) may become puzzling. We shall illustrate both facts by Bessel's
equation. This may be written as

(d2-a2)w +z2w = 0, (7.6.24)

and the corresponding difference equation is

u(s+2) = -,(5+2
^ + a)(s
"w^. +2- a)v (7.6.25)
Here the span is two, but it may be reduced to one by setting 2s = r,
v(2s) = u(t); this gives

rt + V-tt + iXto-t-iy (76"26)

a solution of which is

2, T(ia-t)
u(t) = 2
Thus T(r + l+k)*

v{s) = 2° realty (7-6-27)

and finally

LlTl J l/2-ioo 'rff

I {2S ++ 2(l
'J+n+ 1) (- z y ds- (7A28)
w (z > = dri f 2
Contour integration shows that this is e',7TlJa(z). The formula does not
seem to add much to our knowledge of Bessel functions.
We have seen that the use of the Barnes transform in the theory of
linear DE's leads to a difference equation for the weight function v(s).
Conversely, certain classes of linear difference equations can be solved
by the use of the inverse transformation. This mode of attack was
initiated and developed by Niels Erik Norlund in a number of papers in
the 1910's. We shall illustrate by considering the difference equation
satisfied by the Legendre functions. See Problem 6.2: 14. We write it as

(a +2)X(a +2) -(2a +3)zX(a + l) + (a + l)X(a) = 0, (7.6.29)

280 REPRESENTATION THEOREMS

and try to solve it by an integral

X(a) = | ta lY(t)dt. (7.6.30)

It is found that Y(t) has to be a solution of

t(t2-2zt + l)Y'(t) + (zt " 1)Y(0 = 0. (7.6.31)

We can take
Y(t)=t(t2-2zt + 1)_,/2. (7.6.32)
We set

{ = z - (z2 - 1),/2, | = z + (z2 - 1)1/2, (7.6.33)

where the square root is real positive for z = x real and > 1. Let C^ be a
closed curve surrounding t = f and t = 1/f in the positive sense, leaving
f = 0 on the outside. Let C2 be the ray from t = 0 to t = f. Then for
Re(a)>-1

p-(z)--M, (.'-£■%»• <"-34)

EXERCISE 7.6
1. Justify (7.6.6).
2. Why is (7.6.9) bounded on vertical lines in the strip?
3. Justify (7.6.10).
4. Prove (7.6.15), (7.6.18), and (7.6.19).
5. Give a detailed proof of (7.6.23).
6. Prove that for arg (- z ) < tt

ZTTl J-i/2-i« r(s+a)n-s)(-zYds.

r(a)(\-z)a
7. Fill in the missing details in the discussion leading to (7.6.28).
8. Prove (7.6.34) and (7.6.35). Is it clear that the integrals define solutions of
Legendre's equation (6.2.1) as functions of z? If so, why is the first integral
holomorphic at z = 1 and the second integral goes to zero as z becomes
infinite? What is the limit of £?
9. According to (6.3.7), the Bessel function Ja (z ) satisfies the second order linear
difference equation
zX(a+2)-2aX(a + l) + zX(a) = 0.
Try to solve this equation by an integral of type (7.6.30). Determine Y(s) and
suitable paths of integration C.
LITERATURE
281
LITERATURE

For psi series see:

Dulac, M. Points singuliers des equations differentielles. Memorial des Sci. Math, No. 61.
Gauthier-Villars, Paris, 1934.
Hille, E. On a class of series expansions in the theory of Emden's equation. Proc. Roy. Soc.
Edinburgh, (A) 71:8 (1972/73), 95-110.
A note on quadratic systems. Proc. Roy. Soc. Edinburgh, (A) 72:3 (1972/73), 17-37.
Horn, J. Gewohnliche Differentialgleichungen beliebiger Ordnung. Sammlung Schubert, 50.
Goschen, Leipzig, 1905.
Smith, R. Singularities of solutions of certain plane autonomous systems. Proc. Roy. Soc.
Edinburgh, (A) 72:26 (1973/74), 307-315.
General surveys of integral representations of solutions of linear DE's
are found in Chapter VIII of:

Ince, E. L. Ordinary Differential Equations. Dover, New York, 1944;

and on pp. 90-100 of
Kamke, E. Differentialgleichungen: Losungsmethoden und Losungen. 3rd ed. Chelsea, New
York, 1944.
Euler and Laplace transforms are well presented by Ince, loc. cit. A
detailed account of Laplace transforms is given also on pp. 401-435 of
Vol. 1 and of Euler transforms on pp. 405-534 of Vol. 2 of:
Schlesinger, L. Handbuch der Theorie der linearen Differentialgleichungen. Teubner, Leip-
zig, 1896-1897.
Further references for Sections 7.3-7.5 are:
Birkhoff, G. D. Singular points of ordinary differential equations. Trans. Amer. Math. Soc,
10 (1909), 436-470.
Equivalent singular points of ordinary linear differential equations. Math. Ann., 74
(1913), 134-139.
Laplace, P. S. Theorie analytique des probabilities. Vol. 1, 1812.
Poincare, H. Sur les integrates irregulieres des equations lineaires. Acta Math., 8 (1885),
295-344.
Turrettin, H. L. Convergent solutions of ordinary linear homogeneous differential equations
in the neighborhood of an irregular singular point. Acta Math., 93 (1955), 27-66.
Reduction of ordinary differential equations to the Birkhoff canonical form. Trans.
Amer. Math. Soc, 107 (1963), 485-507.

See also pp. 198-209 and 266-272 of the author's LODE and the literature
quoted therein. The discussion of the Birkhoff -Laplace transform there in
Section 6.9 is based on:

Miller, J. B. Solution in Banach algebras of differential equations with irregular singular

points. Acta Math., 110 (1963), 209-231.
2S2 REPRESENTATION THEOREMS

The Mellin transform is discussed in Ince, loc. cit. For the Barnes-
Mellin transform, see Chapter XIV of:
Whittaker, G. T. and G. N. Watson, A Course in Modern Analysis. 4th ed. Cambridge
University Press, London, 1952.
The original papers are:
Barnes, E. W. A new development of the theory of the hypergeometric functions. Proc.
London Math. Soc. (2), 6 (1908), 141-177.
Burmann, H. W. Die hypergeometrische Differentialgleichung iiber Banachalgebren.
Habilitationsschrift, Gottingen, 1971; Math. Zeit, 125 (1972), 139-176.
Mellin, Hj. Uber die fundamental Wichtigkeit des Satzes von Cauchy fur die Theorien der
Gamma- und der hypergeometrischen Funktionen. Ann. Sci. Soc. Fenn., 20 (1895), No.
12, 115 pp.
Tables of Mellin transforms are given in:
Colombo, S. and J. Lavoine. Transformations de Laplace et de Mellin. Memorial des Sci.
Math., No. 160. Gauthier-Villars, Paris, 1972.
The standard treatise on difference equations is:
Norlund, N. E. Vorlesungen iiber Differenzenrechnung. Grundlehren, No. 13. Springer-
Verlag, Berlin, 1924.
8

COMPLEX

OSCILLATION THEORY

We shall be concerned with second order linear DE's and, more specific-
ally, with the oscillatory behavior of solutions in the complex domain
when the coefficients of the DE are analytic functions. This is essentially
a question of the distribution in the plane of zeros and extrema of the
solution. In a few cases boundary value problems make sense in the
complex plane. This is in contrast to the behavior of real solutions on the
real line, where since the days of Joseph Liouville (1809-1882) and
Jacques Charles Francois Sturm (1803-1855), effectively from the 1830's,
boundary value problems and oscillatory properties of the solutions have
been the objects of much research.
That there could be interesting oscillation problems for the complex
plane was not realized until much later. The pioneer was Adolf Hurwitz
(1859-1919), who in 1889 examined the zeros of the Bessel function Ja(z)
for real values of a. This was followed by work on the hypergeometric
equation by E. B. Van Vleck in 1902, by Hurwitz again in 1907, and by P.
Schafheitlin in 1908. Their methods were essentially function theoretical.
The Bessel functions or, rather, their logarithmic derivatives were consi-
dered by Pierre Boutroux in his profound investigation of the Painleve
transcendents in 1913-1914. The displacement of the poles (zeros of
Bessel functions) with changing initial conditions served as prototype for
the much more profound problem posed by the Painleve functions.
From 1918 to 1924 the present author re-examined these problems,
using a variety of methods (Sturmian, Green's transform, transformation
of Liouville, asymptotic integration, variation of parameters), and applied
these
tions. methods to Legendre's, the Hermite-Weber, and Mathieu's equa-
The Schwarzian derivative and univalence were brought to bear on the
problem by Z. Nehari (1949) and by his pupils P. Beesack and B.
Schwartz, while C-T. Taam used comparison theorems in 1952.
283
284 COMPLEX OSCILLATION THEORY

STURMIAN METHODS; GREEN'S TRANSFORM

We shall be concerned mainly with second order linear DE's of the form
w"+G(z)w = 0, (8.1.1)
where G(z) is holomorphic except for poles in the finite plane. Our
general object is to obtain information about the distribution of the zeros
and extrema of solutions of such an equation. Here extremum denotes a
zero of the derivative.
Our first observation is that complex oscillation theory is essentially
nonoscillation theory: such a solution cannot have zeros on a certain
curve or in a domain of the complex plane or in a neighborhood of a given
point. But rarely do we obtain positive information regarding the exis-
tence of zeros. This negative character has been apparent from the very
beginning of the theory. Thus Hurwitz proved that Ja(z) can have no
complex zeros if a > — 1. There is an old result of A. R. Forsyth (1900),
according to which a solution of (8.1.1) with a zero at a nonsingular point
z = a can have no other zero in the domain

\z-a\2\G(z)\<4. (8.1.2)
No information on the rate of growth of \G(z)\ or on arg G(z) can
exclude solutions of the form

P{z)eQU\ (8.1.3)
where P and Q are polynomials in z. We can assert, however, that (8.1.1)
cannot have two linearly independent solutions of this form unless G(z)
is a constant, for the Wronskian is a constant, and if P eQ and Res are the
solutions, then

eQ+s[PR' - P'R + PR(S' - Q')] = C, (8.1.4)

leads to the stated conclusion.
Here Q + S and the expression between brackets are polynomials.
The latter expression cannot be zero for any value of z. Thus it must be a
constant. Now, if eQ+s is a constant, the exponent must also have this
property. From Q + S =K we get S' = -Q', and this gives
RP' -PR' -2QPR = C. (8.1.5)
Here there are several possibilities. If both P and R are constants, there
exists a constant a # 0 such that Q' = a and the solution is
Aeaz +Be~a\ G(z) = -a\ (8.1.6)
 285
STURMIAN METHODS; GREEN'S TRANSFORM

If P is a constant but not R, then

P0R'(z)-2Q'(z)PoR(z) = K.

If Q' is not identically zero, this is impossible, for the degree of R ' is one
unit less than the degree of R and hence at least two units less than that of
Q'R. If, on the other hand, Q' =0, there exists a constant b such that
R\z) = b and the solution is
a+bz withG(z) = 0. (8.1.7)
Finally, if neither P nor R is a constant, one sees that is
impossible. If Q' =0, then
PR'-RP' = C. (8.1.8)

Since P and R are at least of degree 1, PR ' and RP' must be of the same
degree, and this requires that P and R be of the same degree, say k. From
(8.1.8) it follows that the zeros of P and R are simple and they have no
zeros in common. On the other hand, we have
T> " J?"

This shows that any zero of P must be a singular point of the DE. Since
there are no finite singularities, P" must vanish at all the zeros of P. But
P" has only k - 2 zeros and cannot vanish at the k zeros of P. Hence P" is
identically zero, so that P(z) is linear. This shows also that G(z) = 0. The
assertion is now proved: G(z) must be a constant.
In a general way we can say that the modulus of G(z) controls the
frequency of the oscillation, while the argument of G(z) governs the
orientation and twist of the zeros. This vague statement will be made
more precise later.
Keeping in mind the negative character of the expected information, let
us see how the venerable methods of Sturm can be adjusted to explora-
tions in the complex plane. The prototype is the identity (5.5.7), i.e.,

[w(t)w'(tm - £* [w'(t)f dt - £2 G(t)[w(t)f dt = 0, (8.1.9)

which shows that for G(t)>0 in the interval (a, b) the equation

w"-G(t)w =0 (8.1.10)
can have no nontrivial real solution for which w(t)w'(t) vanishes more
than once in (a, b). Here (8.1.9) is obtained from (8.1.10) by multiplying by
w(t), integrating from tx to r2, and integrating once by parts.
Consider now (8.1.1). Here [w(t)]2 and [w'(t)f are normally not real,
2X6 COMPLEX OSCILLATION THEORY

nonnegative, whereas |w(z)|2 and |w'(z)|2 are. This suggests multiplication

of (8.1.1) by w(z). The interval (a, b) is now to be replaced by an oriented
rectifiable arc C in the complex plane. It is assumed that C has piecewise,
continuously turning semitangents. Usually C will be a polygonal line or
simply a line segment. The next hurdle to overcome is the meaning of the
derivative of the conjugate along C, the derivative taken in the direction
of the local positive semitangent. Here we use the parametric representa-
tion of C in terms of arc length, t = r(<x), and the definition of the
Riemann-Stieltjes integral to obtain

and the integral may be written as

r\wVfdi.
Finally we have the identity

[wJT)w'(zm- Jzi
P|w'(ON' + Jzi
P G{t)\w(t)\2dt = 0. (8.1.11)

known as Green's transform. Here Zi and z2 are any two points on the arc
C such that Zi precedes z2 in the positive orientation of C. We shall give a
number of applications of this identity, starting with some simple cases.
First we need some conventions. A domain D in the complex plane will
be called vertically convex if every vertical line which has a nonvoid
intersection with D intersects D in an open line segment. Similarly, D is
horizontally convex if horizontal lines, which do intersect D, intersect in
an open line segment.

THEOREM 8.1.1

Suppose that (i) G(z) is holomorphic in a domain D symmetric with

respect to the real axis and vertically convex, (ii) the real axis intersects D
in the interval (a, b) and G(z) is real for z = x, a < x < b, (iii) P(x, y) =
Re [G(x + iy)] is positive in D, and (iv) w(z) is a nontrivial solution of
(8.1.1) which is real on (a,b). Then w(z)w'(z) has only real zeros in D, if
any.

Proof. We set G(z) = P(x, y) + iQ(x, y). The complex zeros, if any,
would occur in conjugate complex pairs. Suppose that z2 = xx + ivi is a
zero with y, > 0. Then xx-iyx would also be a zero of w(z)w'(z). We take
Zi = Xi - iyi, z2 = Xi + iy, in (8.1.1.11) and integrate along the vertical line
STURMIAN METHODS; GREEN'S TRANSFORM 287

joining these two points. We have

(8.1.12)

By assumption the real part of this is positive, and we have a contradic-

tion. Since G(x) = P(x) is positive in (a, b), there may be real zeros. If by
any chance b - a> 7r[min P(jc, 0)]~1/2, there are certainly solutions with
zeros in (a, b). ■
The same method gives

THEOREM 8.1.2

Let G(z) be holomorphic in a vertically convex domain D, let P(x, y) =

Re [G(x + iy)]>0 in D, and let w(z) be a nontrivial solution of (8.1.1).
Then no two zeros of w(z)w'(z) can lie on the same vertical line in D. The
same result holds if Q(x, y) = Im [G(z)] keeps a constant sign in D.
Another vertical line result of some interest is

THEOREM 8.1.3

Let G(z) be holomorphic in a vertically convex domain D, symmetric to

the real axis, on which G(z) is supposed to be real. Let Q(x, y) keep a
constant sign in the upper half of D, the lower boundary of which is the
interval (a, b) of the real axis. Let w(z) be a nontrivial solution of (8.1.1),
real on the interval (a, b) so that w(x)w'(x) is real on (a, b), where it keeps
a constant sign the same as that of Q(x, y) above the axis. Then w(z)w'(z)
has no complex zeros in D.

Proof. * If the theorem were false, there would be a zero x0 + iy<> with
y0>0, and the following expression must vanish:

w(xo)w'(x0)-i |w'(Jc0+/y)|2dy -i (p + /Q)|w(x0 + iyf dy.

The assumptions imply that the real part of this cannot be zero, so the
assumption of complex zeros evidently leads to a contradiction. ■
An important consequence of Theorem 8.1.2 is that in a vertically
convex domain D where either P(x, y)>0 or Q(x, y) keeps a constant
sign the zeros of w(z)w'(z) can be ordered in a finite or infinite sequence
according to increasing values of the abscissa.
Similar results hold for horizontally convex domains D in which either
P(x,y)<0 or Q(x, y) keeps a constant sign. We shall denote as
THEOREM 8.1.4 the result obtained by replacing in the wording of
2SS COMPLEX OSCILLATION THEORY

Theorem 8.1.2 the expressions "vertically convex," "P(x, y)>0," and

"vertical line" by "horizontally convex," "P <0," and "horizontal line,"
respectively.
Analogous results hold for solutions real on a segment of the imaginary
axis. Here is one such result, the proof of which is left to the reader.

THEOREM 8.1.5
Let G(z) be holomorphic in a horizontally convex domain D symmetric
with respect to the imaginary axis on which G(z) is real negative. Suppose
that P(z) < 0 in D. Let w(z) be a nontrivial solution of (8.1.1) which is real
on that part of the imaginary axis between ci and di which lies in D. Then
w(z)w'(z) may have zeros between ci and di on the imaginary axis but
nowhere else in D.

Note that w'(z) is purely imaginary on the imaginary axis.

We note further that in a horizontally convex domain where either
P(x, y)<0 or Q(x, y) keeps a constant sign we may order the zeros of
w(z)w'(z) according to increasing ordinates, for under the stated assump-
tions there can be at most one zero on each horizontal line. Now, if the
domain is convex and hence both horizontally and vertically convex, we
have two orderings of the zeros, one by increasing abscissas and the other
by increasing ordinates. Are these orders consistent? The answer is that
either the two orders agree or one is the reverse of the other. This follows
from

THEOREM 8.1.6

Suppose that G(z) is holomorphic in a convex domain D where #, <

arg G(z) < #2 and #2 - #, < it. Let z = a and z = b be two zeros in D of
w(z)w'(z). Then either arg(b-a) or arg(a-b) lies between -|#2 and

Proof. This follows from (8.1.11), where z, = a, z2 = b, t = a+seie,

0 = arg {b - a), and r = \b - a\. We have then that

is real positive. The argument of this expression, however, lies between

#, + 20 and #2 + 20, so that
#, + 20 <0< #2 + 20,
from which the assertion follows. ■
We can now compare the two orderings of the zeros of w(z)wf(z) in D.
STURMIAN METHODS; GREEN'S TRANSFORM 289

If Q(jc, y)>0 in D, we have

#i = 0, #2 = it, whence -\tt < arg (b - a) < 0.
If the zeros in D have been ordered by increasing abscissas, and if

Zn = Xn + lVn, Z„ + i = Xn + 1 iV„ + 1

are "consecutive" zeros of w(z)w'(z), then

Xn<x„+i, yn>yn+l (8.1.13)

by the inequality proved for the argument of z„+i — z„. In this case, the
order by increasing ordinates is opposite to the order by increasing
abscissas. If instead Q(x, y)<0, then 0 < arg (b - a) < \tt and
x„<jc„+i, y„<y„+i, (8.1.14)
so the two orderings are consistent.
For the next theorem we need the notion of a curve being an indicatrix,
a notion that will reappear again and again in the following. It is supposed
that G(z) has no other singularities than poles in the finite plane and that
z = z0 is not singular. Consider

Z(z) = Jzo
f [G{t)]m dt = X(z) + iY(z). (8.1.15)

In general, this is an infinitely many-valued function. At the zeros of G(z)

the function Z(z) has in general algebraic branch points, while the poles
give algebraic and/or logarithmic branch points. The determinations of
Z(z) are of the form
±Z(z) + a,

where a is a number depending on the path leading from z0 to z. It will be

zero if Z(z) has at most one algebraic branch point and no logarithmic
ones.
For the following it is essential to consider the curve net

X(z) = C„ Y(z) = C2. (8.1.16)

Curves of the first family will be referred to as anticlinals, and those of
the second as indicatrices for, as we shall see, they indicate the direction
to nearby zeros. These curves should be traced on the Riemann surface of
Z(z). Actually the net is the same on all sheets of the surface, but the
labeling of individual curves may differ from one sheet to another.
Through a point Zi, which is neither a zero nor a pole of G(z), passes one
and only one curve of each family. The situation is different at poles and
zeros of G(z).
2<W COMPLEX OSCILLATION THEORY

For the discussion immediately following, the crucial property of the

indicatrix is the value of its slope. At a point Zi, neither a zero nor a pole
of G(z), the local indicatrix
Y(z)= Y(z,)
has a unique tangent and its slope is tan 0i, where

0, = -|argG(z,) (modTr). (8.1.17)

With this background we can state a convexity theorem:

THEOREM 8.1.7

Suppose that D is a convex domain where G(z) is holomorphic and ¥■ 0.

In D suppose that di < arg G(z) < #2, where #2 — #i < tt. Furthermore, no
indicatrix shall have a point of inflection in D. Let w(z) be a nontrivial
solution of (8.1.1) with a finite number of zeros and extrema in D. Then
these points form the vertices of a convex polygon.
Proof. Without loss of generality we may assume that #i = 0 so that
Q(x, y ) > 0 in D. If this is not so at the outset, we can bring it about by a
suitable change of independent variable: a rotation about the origin will
preserve convexity, indicatrices, and the nonexistence of points of
inflection.
Now the absence of points of inflection means that each individual
indicatrix has a curvature of the same sign, and the continuity of G(z) and
the condition G(z) ^ 0 mean that the sign of the curvature is the same for
all indicatrices in D. There are then only two possibilities:
1. The indicatrices are concave upward and lie above the local curve
tangent.
2. The indicatrices are concave downward and lie below the curve
tangent.
It is enough to discuss one of these cases, and we take the first.
Mark the zeros of w(z)w'(z), say Zi, z2, . . . , zn with zk = xk + iyk, where
now
xk<xk+u yk>yk+u k = 1,2, . . . , n - 1.
Suppose that the indicatrices through the points zk are plotted as
L: Y{z)=Y{zk\
and let Tk be the tangent of Ik at z = zk. Then Tk divides D into two
complementary parts: Dt above Tk and Dk below and on Tk. It is claimed
that no point in one of the regions Dk can be a zero or extremum of our
STURMIAN METHODS; GREEN'S TRANSFORM 291

solution u>(z). If this were false, we could find an r > 0 and a 0, 0k - tt ^

0 ^ 0k9 0k = -\ arg G(zfc), such that

- Jo
[ \w'{zk + seiefds + e2ie Jo[ G(zk + s eie)\w(zk + 5 eie)\2 ds = 0.

Here we take the imaginary part and obtain

j sin [20 + y(s)]g(s)\w(zk + s ei9)\2 ds = 0, (8.1.18)

where

g(s) = \G(zk + 5 eie% 7(5) = arg G(zk + s eie).

Now (8.1.18) is clearly not true as long as sin [20 + 7(5)] keeps a constant
sign. For it to change its sign, there must be a value s where 20 + 7(5) = 0
(mod tt ). At such a point, according to (8.1.17), the line would have to be
tangent to the local indicatrix, so a question arises: Is it possible to draw a
tangent from z = zk to an indicatrix in Dkl Now all these indicatrices turn
their concave side toward zk so no tangent can be drawn, and this means
that (8.1.18) cannot hold for points in Dk. From this fact we conclude that
no zeros or extrema of w (z) can be located in U Dk and all points z, are in
fl Dk. Now this implies convexity of the polygonal line Zi, z2, . . . , z„, Zi
unless all z/s lie on a straight line so that the polygon degenerates. This
can happen iff G(z) is a constant. Excluding this fairly trivial case, we
have a nondegenerate polygon and the angle between the sides joining
Zk-i with zk and zk with zfc+i is less than 77, so all interior angles at the
vertices are < tt. This proves the convexity. ■
Now in the chain Zi, z2, . . . , zn some points are zeros, whereas others
are extrema. Do they alternate? This question was posed by the author
over 50 years ago and remains unanswered. In some cases, however, it is
possible to answer the question in the affirmative by an argument based
on variation of parameters. We shall encounter some such instances later.
In the statement of the convexity theorem the condition regarding no
points of inflection is puzzling and leads naturally to the question of how
one can verify such a condition. Now the system of indicatrices is the set
of integral curves of a certain linear homogeneous first order DE, and it is
well known that for such DE's one can determine various loci where this
or that phenomenon occurs. The points of inflection present such a
phenomenon, and for the family Y(z) + C we get the inflection locus

lm{G'(z)[G(z)Tm} = 0. (8.1.19)

This says that the derivative of [G(z)]~1/2 is real along the locus. It is a part
292 COMPLEX OSCILLATION THEORY

of the locus

lm{[G'(z)]2[G(z)T3} = 0, (8.1.20)
the other part of which is the inflection locus of the anticlinals.

EXERCISE 8.1

1. Verify the assertions concerning solutions of the form P eQ. Find an expres-
sion for G(z) when such a solution occurs.
2. A convex domain is horizontally and vertically convex. Is the converse true?
3. Use (8.1.11) to show that, if G(z) = a2, a constant, and if w(z) is a nontrivial
solution with a zero at z = z0, then all its zeros and extrema lie on a straight
line through z0. Find the line.
4. How should oo be chosen, \co\ = 1, so that the rotation z = cot leads to a DE
with the properties used in the proof of Theorem 8.1.7?
5. If Gm = U + iV, show that the indicatrices satisfy the DE U dy + V dx = 0,
while the anticlinals satisfy U dx - V dy = 0.
6. Use this fact to verify (8.1.19).
7. Why does (8.1.19) imply (8.1.20)? Find the inflection locus of the anticlinals.
8. If G(z) = z - a, sketch the indicatrices and find their inflection locus.
9. In Problem 8 take a = 0 and the solutions determined by (i) Wi(0) = 0,
wi(0) = 1, (ii) w2(0) = 1, w'2(0) = 0. Find where their zeros are located.

8.2. ZERO-FREE REGIONS AND LINES OF INFLUENCE

Various zero- free regions have been encountered in Section 8.1. A region
S(z0) in the complex plane is a zero-free region with respect to a solution
w(z) satisfying a condition at z = z0 if w(z)w'(z) ^ 0 in S(z0). There are
various ways of constructing such regions.
We may consider (8.1.1) as the equation of motion of a "particle" w(z)
when z describes a path C and as a consequence w(z) describes an orbit
T in the w-plane. A zero of w(z) encountered by C corresponds to a
passage of T through the origin. A zero of w'(z) is another story. If
w(z) = R(z)ei&(z\ (8.2.1)
then

w'(z) = [R'(z) + iR(z)e'(z)] ei€Kz\ (8.2.2)

If this is to be zero at z = zu we must have

R'(Zl) = e'(zi) = 0, (8.2.3)

and normally these are simple zeros where R' and @' change their signs.
ZERO-FREE REGIONS AND LINES OF INFLUENCE 293

The orbit T has a cusp for this value of z, and both branches of T lie to the
same side of the cusp tangent.
Now it is clear that between consecutive passages through the origin of
F there must be a point where R(z) has a maximum, but there is no reason
why this should correspond to a cusp and a zero of w'(z). Let us now note
the equations of motion. We have

w "(z ) = {R "(z ) - R (z )[©'(z )]2 + i [2R '(z )0'(z ) + R (z )©"(z )]} e ,0(2).
(8.2.4)

If z = r e ie and the differentiations are with respect to r for fixed 0, the

DE takes the from

R "(r) - R (r)[0'(r)]2 + i [2R'(r)0'(r) + 1? (r)0"(r)]

+ g(r)eiY(r)+2ieR(r) = 0, (8.2.5)
where

= g(r)eiY(r).and obtain the real-valued

Here we separate reals andG(z)imaginaries
equations of motion:

R "(r) + {g(r) cos [y (r) + 2d] - [®'(r)f}R (r) = 0, (8.2.6)

{®'(r)[R (r)]2}' + g(r) sin [7(r) + 20][K (r)]2 = 0, (8.2.7)
where the second equation has been multiplied by R(r) to bring it into
self-adjoint form.
These equations are satisfied by

R(r) = \w(zQ+rei6)\ and @(z ) = arg w (z0 + r e ie )

for small positive values of r. The functions are continuous together with
their first and second order derivatives. As r increases, various things
may happen. We mark on the ray arg (z - z0) = 6 the points where one of
the following events takes place: the point is (i) a zero of G(z), (ii) a
singularity of G(z), (iii) a zero of w(z), (iv) a zero of w'(z).
Let Zi = ri e'e be the first such point. The first two categories may be
considered as known, but disregarding these points, we need a lower
bound for an r, corresponding to categories (iii) and (iv). We suppose
w(z) to be fixed by the conditions

w(zo) = 0, w'(zo) = 1.
Equation (8.2.6) leads to the observation that

g(r) cos [y(r) + 26]- [0'(r)]2 ^ g(r),

and this, in turn, leads to the following question. Equation (8.2.6) is
294 COMPLEX OSCILLATION THEORY

satisfied by R(r) in the interval (0, r.) and R(0) = J?(r,) = 0. Consider the
equation

v"(r) + g(r)v(r) = 0, u(0) = 0, u'(0)=l. (8.2.8)

Suppose that the least positive zero of v(r) is r0. If

we have g,(r) = g(r) cos [y(r) + 26]- [O'(r)]2,

R(r0)v'(ro)-R'(ro)v(ro) + f\g(s)- gi(s)]v(s)R(s) ds = 0.

Here u(r0) = 0, v'(r0) < 0, and the bracket is nonnegative. It follows that
R (r) > 0 for 0 < r < r0. This gives

THEOREM 8.2.1

Let v(r) = v(r, 0) be the solution of

v"(r) + |G(z0 + r eie)|v(r) = 0, v(0) = 0, v'(0) = 1. (8.2.9)

Let the least positive zero of v(r, 6) be ro(0). Then the solution w(z) of
(8.1.1) with w(zo) = 0, w'(z0)=l, has no zeros on the open interval
(z0, zo + ro(0)eie).
COROLLARY

Let G(z) be holomorphic in the disk |z - Zo| < R and satisfy |G(z)| =ss 7r2R~2
there; then the solution w(z) with w(z0) = 0, w'(z0) = 1 has no other zeros
in |z - z0| < R.
Here we have the comparison equation

with i>"(r)+(f)2iKr) = 0 (8.2.10)

v(r) = sin r^j

and the only zero in (- Ry R ) is r = 0.
An alternative formulation of the corollary is:

If G(z) is holomorphic in |z - z0| < R and if |G(z)| M there, then w(z)

has no zeros in the punctured disk,

0<|z-zo|< min (R, 7rM~1/2). (8.2.11)

This should be compared with Forsyth's bound, given in (8.1.3).
 295
ZERO-FREE REGIONS AND LINES OF INFLUENCE

We obtain similar results when w(z) has an extremum at z = z0, but

here the zero-free region is smaller.

THEOREM 8.2.2

If G(z) is holomorphic in |a - z0| < R and if |G(z)| *s M there, then the

solution of (8.1.1) with w(z0) = 1, w'(zo) = 0 has no zeros in
|z - z0| < min (R, \irWm). (8.2.12)
The comparison equation is still (8.2.10), but now the solution is
cos (f r),

and the least positive zero is \ttR. Theorems 8.2.1 and 8.2.2 are due to
Choy-Tak Taam (1952).
So far the zero-free regions have been obtained using comparison
theorems, a classical Sturmian device.
The conclusion to be drawn from (8.2.7) is of a different nature, as we
have already seen in the proof of Theorem 8.1.7. We consider a solution
for which w(z0)w'(zo) = 0. This implies @'(0).R(0) = 0. We now integrate
(8.2.7) with respect to r, keeping 0 fixed:

e'(r)[R(r)f + ^ g(s)sm[y(s) + 2e][R(s)f ds = 0. (8.2.13)

This shows that we cannot have S'(r)[R(r)f = 0 as long as
sin [7(5) + 20]
has a constant sign in 0 < s < r. Let us now mark the first point,
Zi = Zi(r, 0) t6 z0, where either G(z) is singular or

argG(zo+reie) + 20 =0 (mod tt). (8.2.14)

We have then

THEOREM 8.2.3

A solution o/(8.1.1) for which w(zo)w'(z0) = 0, where G(z) is holomorphic

at z0, admits the set

So(zo) = {z;z = Zo + a[z,(r,0)-Zo],O<a =s 1,O=S0 <2tt} (8.2.15)

as a zero- free star in which w(z)w'(z) * 0.
Condition (8.2.14) is familiar to us from the proof of Theorem 8.1.7. It
asserts that at the point z, in question the ray arg (z - z0) = 6 is tangent to
296 COMPLEX OSCILLATION THEORY

the local indicatrix Y(z) = Y(zi). If z0 is not a zero of G(z), then there is
one and only one indicatrix through z0; and if in addition z0 is not on the
indicatrix inflection locus (8.1.19), then in some partial neighborhood of
z = z0 the indicatrices turn their concave side toward z0. Let T0 be the
tangent of the indicatrix Y(z) = Y(z0). Then one side of T0 belongs to
So(zo), whereas the other does not (compare the situation encountered in
the proof of Theorem 8.1.7). Let us call the two sides the "on side" and
the "off side," respectively. On the off side of T0 we can draw tangents to
the indicatrices from z0. From this we get the following

COROLLARY
The boundary of S0(z0) is the locus of the nearest points oftangency between
the pencil of straight lines through z = z0 and the system of indicatrices plus
part of the tangent To, and possibly also parts of lines beyond the singular
points.

As observed above, this picture will have to be modified if z = z0 is a

zero of G(z) or belongs to the indicatrix inflection locus.
We turn to the construction of zero-free regions based on the use of the
corresponding Green's transform (8.1.7). We start by finding a zero-free
star. Suppose that w(z) is a solution of (8.1.1) such that w(z0)w'(zo) = 0,
where, as usual, z0 is not a singular point of the DE. We integrate along
the ray arg (z - z0) = 0 to obtain

f K(0|2 ds - Jof g(s) eiy(s)+2ie\w(t)\2 ds.

e'Vz)w'(z) = Jo (8.2.16)
Here

t=zo+se'0, z=z0+re'6,
g(s) = \G(t)\, 7(s) = argG(0.
The first integral on the right is positive; the second one has an argument
which lies between

miny(s) + 20 and max 7(5) + 20,

provided the difference between the maximum and the minimum is < tt.
Let us mark on the ray arg(z - z0) = 0 the first point where (i) G(z) is
singular or (ii) arg G(z) + 20 = 0 (mod tt) or (iii) max arg G(z)
— min arg G(z) = tt. Here the maximum and the minimum refer
to the points on the ray from z = z0 to the point in question. Let the
marked point correspond to s = r(6), and define

S(z0) = [z; z = z0 + 5 ex\ 0 < s ^ r(0), 0 ^ 0 < 2tt]. (8.2.17)

ZERO-FREE REGIONS AND LINES OF INFLUENCE 297

This is a star containing the star S0(z0), and the inclusion may be proper.
We have

THEOREM 8.2.4

If w(z) is a solution of (8.1.1) such that w(z0)w'(z0) = 0, where z0 is

nonsingular, then w(z)w'(z) ^ 0 in S(z0).
The proviso that z0 be nonsingular may be disregarded if w(z) is a
holomorphic solution at z = z0 or w(z)w'(z) tends to the limit 0 as z -+z0
and the integrals in (8.2.16) exist for the lower limit 0.
This is only one of the ways in which the restrictions involving the use
of Green's transforms may be relaxed. We are not restricted in the
discussion to the use of rays; any oriented rectifiable arc C will do if it
leads to a definite vector field in a sense to be clarified. We start by taking
a solution w(z) defined at the nonsingular point z0 by the initial parameter

VV(Zo)
^^ = A. (8.2.18)

This may be zero, infinity, or a complex number. In the third case set
at = arg A. Consider the identity

w(2)w'(z)= w(zo)w'(zo) + jjw'(t)\2Jt- jcG(t)\w(t)\2 dt. (8.2.19)

The first term on the right either is zero or is a nonzero vector whose
argument is co. It is assumed that C is piecewise smooth and has a
continuously turning positive semitangent except for a finite number of
points. Let the slope angle of the semitangent be 0(z). With the first
integral is then associated the set of vectors
{e~ifl(2);z GC), (8.2.20)
while
{expi [argG(z)-7r + 0(z)];z GC} (8.2.21)
goes with the second integral. The union of these two sets and the vector
e"° is the vector field of C with respect to w(z). The vector field is definite
if all three components are confined to an angle of opening < tt, and in this
case C is called a line of influence of the solution w(z) with respect to
Z = Zo.
Thus C is an li (-line of influence) if there exists a # such that
(i) #<w<# + 7r, (ii) # <-0(z)<# + 77, z GC, and further
(iii) # < arg G(z) - tt + 6(z) < # + tt, z G C.
If the vector field is definite, the sum of the three vectors in the
right-hand member of (8.2.19) cannot be zero, so w(z)w'(z) cannot vanish
298 COMPLEX OSCILLATION THEORY

at the upper end point of C. In the three inequalities we can replace "<"
by provided "<" holds either in (i) or in a subset of C of positive
measure in (ii) and (iii).
The totality of points z which can be reached by an li from z = z0 with
respect to the given solution w(z) is defined as the domain of influence
DI (w, z0) of w(z) with respect to z0. It is clearly a zero-free region.

THEOREM 8.2.5

The solution w(z) specified by condition (8.2.18) satisfies w(z)w'(z) ^ 0 in

DI (w, zo).
The search for lines of influence can be standardized to some extent.
Various curve nets present themselves as possibilities, and one can
always try polygonal lines, where the sides are parallel to the coordinate
axes.
As an example take the equation

w" + zw = 0, zo = a > 0, A real. (8.2.22)

Let C be the line from a to b, b real > a, followed by the vertical from
z = b to z = b + ly, 0 < y. This is an li. For a> = 0 or tt, the set (8.2.20)
consists of two vectors, one of argument -\tt and the other 0, while the
set (8.2.21) has one vector along the negative real axis and the others in
the fourth quadrant. The vector field is then definite, so the solution can
have neither zeros nor extrema X\ + iyi with a =s 0 < yx. Note that the
segment (a, b) is not an li. It is recommended that the reader plot the
vector field.
Another suitable net is formed by the curves

G(z) = j G(t)dt =G,(z) + iG2(r), (8.2.23)

G,(z) = C„ G2(z) = C2. (8.2.24)
This is the G-net; the two sets are orthogonal trajectories of each other,
and through a point z which is neither a zero nor a pole of G(z) passes
one and only one curve of each family. If we integrate along a member of
the G, -family, then

-jc G(0|w(f )|2 dt = - ijc\w(t)\2 dG2 (8.2.25)

is purely imaginary, whereas the integral along a G2-curve gives

(8.2.26)
|jw(0|2 dG„
ZERO-FREE RECIONS AND LINKS OF INFLUENCE 299

which is real. These are useful paths of integration as long as the vectors
[e "' z6C) give a definite vector field in conjunction with (8.2.25) or
(8.2.26), as the case may be.
As an illustration of the use of this net let us take the Hermite-Weber
equation in the form

w"(z) + (c2 - z V(z) = 0, 0 < c. (8.2.27)

Here

G,(x, y): c2x - \x* + xy2= Cu

' a ,\ (8.2.28)
G2(jc,y):c2y-x2y +iy3 = C2.
Let us consider a solution of (8.2.27) which on the real axis is real and on
the imaginary axis satisfies a restriction on the initial parameter,

to the effect that its argument a> = <o(iyn) will not fall in the first quadrant
for y<>>0 or in the fourth quadrant for y«<0. If these conditions can be
satisfied for all y0, there exist two domains, D and D, in which
w(z)w'(z) 0. Here D is bounded by the hyperbola x2 - \y2 = c2, x < -c,
0 < y, the segment (- c, 0) of the real axis and the positive imaginary axis.
D is the conjugate domain.
We can integrate along a G2-curve from t = iy0 to t = z G D, where
y0>0. We have then
Jiyo
^{z~)w'(z) = A(/yo)|w(iy0)|2+ f \w'(t)\2Tt
J«'yo
- ( \w(t)\2 dG,(t). (8.2.30)
Here A(/y<>) does not lie in the first quadrant, the first integral lies in the
third quadrant, and the third term is real negative. The resultant of the
three vectors cannot be the zero vector, no matter what value A(/y0) has,
as long as it is not in the first quadrant. This proves the assertion for D.
The same type of argument works if yo<0.

EXERCISE 8.2
1. Verify (8.2.2M8.2.7).
2. Why does a zero of w'(z) correspond to a cusp of the orbit, and why do the
two branches lie to the same side of the cusp tangent?
3. At a zero r0 of R(r) this function changes its sign. Does this mean that the
absolute value of w(z) becomes negative? At such a point @(r) is discontinu-
300 COMPLEX OSCILLATION THEORY

ous. Show that there are finite left- and right-hand limits which differ by an
odd multiple of tt. Show that @'(r0 + 0) = 0'(rfl - 0) = 0.
4. Verify Theorem 8.2.3.
5. Let D be a convex domain in the z-plane, where G(z) is holomorphic and
¥ 0. Take z0 E D, and consider the pencil of straight lines through z0. Find the
locus of tangency of the lines with the indicatrices in D. Show that this locus
is a part of Im [(z - z0)2G(z)] = 0, and show that the rest is the tangent locus
of the anticlinals.
6. For the equation zw" - w =0 and z<» = 1 + i, determine the star S(z0) with
respect to a solution for which w(zo)w'(z0) = 0.
7. The equation in Problem 6 has a regular-singular point at the origin where
there is a solution which is an entire function of z. Show that it is possible to
define the star 5(0) for this solution, and determine it.
8. For the same equation take a solution real on the positive real axis and a
point z = a, 0 < a, and w(a) >0, w'(a) < 0. Take the path z = a + r e'\ 0 < r,
0 =s 6 < 2ir. Determine the vector field of the path. Are there any limitations
on r?
9. In (8.2.30) show that the first integral is a vector in the third quadrant, as
asserted.
10. Verify that the condition imposed regarding A(iy0) in the discussion of
(8.2.30) is satisfied by E„(z), where c2 = la + 1.

8.3. OTHER COMPARISON THEOREMS

The results to be proved or stated in this section are due to Z. Nehari and
his school, P. R. Beesack and B. Schwarz, as well as V. V. Pokarnyi and
C. T. Taam. They are mostly nonoscillation theorems with a bearing on
the theory of univalence of analytic functions (cf. Chapter 10). We start
with a result obtained by Taam (1953), for which we give a proof due to
Beesack (1956) involving a generalization of Green's transform and
another quadratic integral identity.
Consider the equation

w"+G(z)w = 0 (8.3.1)
and the corresponding Green's transform,

[^vv'(z)]!;- Jfa
\w'(t)\2 dt + JIa G(t)\w(t)\2 dt =0. (8.3.2)

Here a and b are two zeros of w(z)w'(z) in a convex domain D where

G(z) is holomorphic, t = a + r eie, and 0 < r r, = \b - a |. Hence

Jo Jo Jo
[Vf dr= I'* e2'°G\w\2dr= f (g. + /g2)|w|2 dr.
OTHER COMPARISON THEOREMS 301

This gives

Jo
(r'\w'\2dr= Jo
I'* gi\w\2dr, Jo
[l g2\w\2dr = 0.
Next let A and ti be real numbers, A2 + /a2 0. Multiply the first equation
by A and the second by /a, and add to obtain

Jo' (Ag, + iJLg2)\w\2 dr = \j^ |w'|2 dr. (8.3.3)

This is Beesack's modification of Green's identity. The second required
identity we state as a lemma:

LEMMA 8.3.1

Let F(r) be real and continuous in the interval (a, b), and let (t - a)2 x
(t - b)2F(t) be bounded in (a, b). Suppose that the Riccati equation
Y'(t) = F(t) + [Y(t)]2 (8.3.4)
has a solution Y(t) continuous together with its first derivative and such
that (t - a)(b - t)Y(t) is bounded in (a, b). If y(t) is a piecewise smooth
function and y(a) = y(b) = 0, then

[ {[y'W]2 - F(t)[y(t)]2} dt = £ [y'(t) + Y(t)y(t)]2 dt. (8.3.5)

Proof (sketch). The integral in the left member exists since the assump-
tions on y(t) imply that \y{t)\^K(b-t)(t-a), so that F(t)[y(t)]2 is
bounded and continuous; further left- and right-hand derivatives exist
everywhere and are bounded and equal with a finite number of excep-
tions. If the right-hand integral exists, it equals

j\y'(t))2dt+2 j*y(t)y'(t)Y(t)dt+ [Y(t)y(t)f dt. (8.3.6)

The first integral exists and is also present in the left member of (8.3.5).
An integration by parts, using the vanishing of Y(t)[y(t)]2 in the limits,
reduces the second integral to

- ja [y(t)f Y'(t) dt = -f* [y(r)]2F(r) dt - f* [y(t)Y(t)f dt.

Here the first integral on the right exists, and the second integral cancels
the third integral in (8.3.6). Hence the right member of (8.3.5) exists and
equals the left member, so the lemma is proved. The important thing here
is that the right member is nonnegative and can be zero iff the integrand is
identically zero. ■
302 COMPLEX OSCILLATION THEORY

THEOREM 8.3.1
Let G(z) be holomorphic in a domain D containing the line segment
z = zo + re,0,O^r^R. Set
e2ieG(z0 + rei9) = g,(r, 6) + ig2(r, 0). (8.3.7)
Let Q(r) be continuous in (0, R), and suppose that the DE

v"(r) + Q(r)v(r) = 0 (8.3.8)

has a solution with a zero at r = 0 and no zeros in (0, R). Suppose that real
numbers A, \x can be found, A 5= 0, A2 + ju,2 > 0, such that
Ag,(r, 0) + iLtg2(r, 0) ^ AQ(r), 0 =s r ^ R. (8.3.9)
Then any nontrivial solution of (8.3.1) with w(z0) = 0 is unequal to zero on
the segment (z0, z0 + R eie ) unless g2(r, 0) = 0. Even if g2(r, 0 ) should be = 0,
the statement holds, provided that A > 0. Moreover, if strict inequality holds
in (8.3.9) at a single point, then w(z) has no zeros in (z0, Zo + R e,e).
Proof Suppose that w(z,) = 0, Zi = z0+r, eie, 0< r, < K. Now (8.3.3)
with (8.3.9) gives

A J \w'\2dr ^A J Q(r)|w|2dr.
Since A may be zero, this case must be settled before we can proceed. If
A = 0, then \x # 0, and (8.3.9) shows that either g2(r, 6) has a constant sign
opposite to that of i±, or else g2(r, d) = 0. In the first case, (8.3.3) has a
negative left member while the right one is zero — a clear contradiction.
From this it follows that w(z) can have no zero in (z0, z0 + R ei0). The
second alternative, g2(r, 6) = 0, cannot arise if strict inequality holds in
(8.3.9) for at least one value of r and hence in some interval.
Suppose now that strict inequality holds nowhere and g2(r, 6) = 0; then
for A >0 (8.3.9) shows that g,(r, 6) = Q(r) and the desired conclusion is
immediate. Suppose now that g2(r, 0)^0 and that A >0. Note that

and set J*Vf dt *z £ Q(r)\w\2 dr,

\w(z0+reie)\ = W(r).
We observe that W'(r) exists for 0< r < r,. Moreover, by (8.2.1) and
(8.2.2), we see that W'(r) = R'(r), so that

|W(r)|*s|w'(z)|, z =zQ+rei(
OTHER COMPARISON THEOREMS 303
and

We can now usejLemma

' [W'(r)f8.3.1
dr^j1 Q(r)[W(r)f
to obtain dr.
the opposite inequality(8.3.10)
and a
contradiction. Take

a=0, b=rl9 y(t)=W(t), F(t) = Q(t), Y(t) = -^.

Since v(t) has no zeros in (0,/?), Y(t) satisfies the Riccati equation

Y'(t)=Q(t) + [Y(t)f
in (0, R), and t(R - t)Y(t) is bounded. Furthermore, W(t) is smooth and
W(0) = W(ri) = 0. Since the assumptions of the lemma are satisfied, we
get

j\w'(t)f dt = Q(t)[W(t)f dt

+ f ' [W'(t)+Y(t)W(t)]2 dt

Jo
» f ' Q(t)[W(t)f dt, (8.3.11)
and here equality can hold iff Jo

W'(t)+Y(t)W(t) = 0, W(t) = Cv(t).

If this should happen, W(t) * 0 for 0< t < R. Finally, if W(t) is not a
constant multiple of v(t), the resulting inequality (8.3.11) contradicts
(8.3.10). It follows in either case that w(z)^0 on (z0, z0 + R eie), as
asserted. ■
As special cases of this theorem we list the following:

1. Q(r) = i47T2R-\
2. A = l,^i =0, Q(r) = gi(r, d).
3. A =O,g2(r,0)^O on (0,i?), /ut =-sgng2.
4. A = 1, /Lt =0, Q(r) = gl(r,0)<O.
Here case 1 is related to the Corollary of Theorem 8.2.1 and to Theorem
8.2.2. The result is due to Z. Nehari (1949). We state it here as

THEOREM 8.3.2

Let G(z) be holomorphic in |z - z0| < R and satisfy |G(z)| ^ Itt'R"2. Then
no solution of (8.3.1) can have more than one zero in the disk.
304 COMPLEX OSCILLATION THEORY

Since we have g,(r, 0) ^ \e2i0G{zo + r eie)\ ^ \ir2R 2, we can take A =

1 , fx = 0 and the majorant equation

v" + W2R-2v = 0.
There are real solutions with equidistant real zeros, and the distance
between consecutive zeros equals the diameter of the disk.
Case 2 follows from (8.2.6) since

g(r) cos [y(r) + 20] - [O'(r)]2 ^ gl(r, 6).

This case is essentially covered by Theorems 8.2.1 and 8.2.3. Case 3
follows from (8.2.6) and the proof of Theorem 8.2.1, while case 4 is
essentially Theorem 8.2.2.
Beesack has combined comparison theorems with fractional linear
transformations to obtain

THEOREM 8.3.3

Let G(z) be holomorphic in the unit disk, |z| < 1 , and let M(r) be the
maximum modulus of G(z), i.e., M(r) = max |G(r e,8)|. Suppose that
(1 -r2)2M(r) is nonincreasing in (0, 1) and that the DE
v" + M(r)v = 0 (8.3.12)

has a solution without zeros in the interval (- 1, 1). Then no solution of

(8.3.1) can have more than one zero in |z|< 1.
Proof. Suppose that there is a solution of (8.3.3) with two zeros, a and b
in the unit disk. By a suitable Mobius transformation we can bring it about
that a transformed DE has two zeros in the interval (- 1, 1). If a and b are
already real, the next step is to be omitted. If not, we observe that there is
a unique circle T passing through a and b and orthogonal to the unit
circle. We may assume this circle to be symmetric to the imaginary axis,
for if this is not so at the outset we can use a rotation to achieve it. This
means replacing G(z) by a)2G(o)z), where |o>| = 1. It does not change the
maximum modulus or any of the hypotheses.
Suppose now that V intersects the imaginary axis at z = pi, 0 < p < 1.
The transformation

1 + piz (8.3.13)

_ z - Pi'
will then map the arc of Y in the unit disk on the interval - 1 < f < + 1 , and
the points z = a and z = b go into points on the interval (-1, 1), say
OTHER COMPARISON THEOREMS 305

-Kfi<f2<l. The Mobius transformation (8.3.13) takes (8.3.1) into

y* y> + (l~p2)2 G(^iP) Y=0 (8 3 14)

where Y(£) = w(z) and the primes denote differentiation with respect to
f. Here we can remove the first derivative by setting «(£) = (1 - ip£)Y(Z),
which leads to the DE

d.c + (i - mr \\ - ifur ~ ( }
Here we are interested in real values of f between - 1 and + 1 . We have
then

1- m
(+ {p2 + j
"ll+/32fV '

Now (8.3.15) has a solution with two zeros in the interval (-1,1). We are
thus led to the comparison
dC equation

d2U (l-j32)2 r/.^^M*

and an auxiliary hypothesis which will be eliminated afterward.

Hypothesis H. No solution of the DE (8.3.16) can have more than one
zero in the interval (—1,1) for any value of ft 0< < 1.
Our discussion has shown that the DE (8.3.15) has a solution with two
zeros in (— 1, 1), and if Hypothesis H holds this contradicts Theorem 8.2.1,
which allows a solution of (8.3.15) to have at most one zero in (- 1, 1).
To complete the proof of Theorem 8.3.3 we have to show that
Hypothesis H is equivalent to the condition that (1 - r2)2M(r) is nonin-
creasing. We shall show that
(l-PT
(1+/3V)

for 0=Sj3<l,0=sr<l.To prove M[( thelTF p)"]"M

equivalence of the(r) (83-17)
two conditions,
assume first that (l-r2)2M(r) is nonincreasing. Set x2 =
(j32 + r2)(l + /32r2) '. Then r2 ^ x2 ^ 1 and
(l-x2)2M(;c)^(l-r2)2M(r),
and this reduces to (8.3.17). Conversely, suppose that (8.3.17) holds, and
let r and x be such that 0^r^x<\. Define fi2 so that x2 =
(fi2+ r2)(l + 2r2) 1 ; then ft r satisfy condition (8.3.17). The inequality now
306 COMPLEX OSCILLATION THEORY

reduces to

(1 - r2)2M(r) (1 - x2)M(x), (8.3.18)

and since r this proves the nonincreasing property of the product. This
completes the proof of Theorem 8.3.3. ■

COROLLARY

No solution of (8.3.1) can have more than one zero in |z| < 1 if (1 - r2)2M(r)
is nonincreasing in (0, 1) and if the DE

y" + M(r)y = 0 (8.3.19)

has a solution which does not vanish in - 1 < r < 1.
In other words, all the solutions of (8.3.1) are univalent in the unit disk.
This result is due to Nehari (1954), who pointed out that it includes three
sharp special cases: (i) M(r) = ^772, (ii) M(r) = 2(1 - r2) ', (iii) M(r) =
(1 - r2) 2. The first and the third case had been proved earlier by Nehari
(1949). The second case is contained in a theorem stated by V. V.
Pokornyi (1951). Beesack showed that the entire theorem of Pokornyi is
covered by Theorem 8.3.3.
To apply the theorem one must be able to refer to a majorant equation
y" + M(r)y =0 with the desired properties. For case (i) we can refer to
the corollary of Theorem 8.2.1. For case (iii) we note that the equation

w" + (l -z2) 2w = 0 (8.3.20)

has the general solution

w(z) = (1 - z2),/2(c, + C2 logf^), (8.3.21)

which has at most one zero in (- 1, 1) and no zeros if Ci = 1, C2 = 0. Here
we cannot introduce a constant multiplier a > 1 before the factor
(1 - z2)"2 without getting solutions with infinitely many zeros in (- 1, 1).
The general Beesack-Pokarnyi theorem involves
r2"V-" (l-rT\ 0«A<l,
M(r)-|2' -*<l-rT\ 1«A<2, (83-22)
for 0 r < 1. The majorant equations are fairly complicated, and we refer
to the original article for details.
Beesack has some interesting results concerning zeros in a neighbor-
hood of a regular-singular point, a problem broached by the present
author in 1925, using the Liouville transformation and the Green trans-
form. Beesack proves
OTHER COMPARISON THEOREMS 307

THEOREM 8.3.4

Let G(z) be holomorphic except for a double pole at z = 0. Suppose that

|G(z)| ^ \\z\~2 + A \z\'1 - A 2 (8.3.23)

for all z and some fixed A ^ 0. If z = a ^ 0 is a zero o/ w(z), a solution of
(8.3.1), then w(z) can have other zeros neither in the half-plane |a + r e,e | ^
r nor in fne disk |z - a| < |a|.
For details the reader is referred to the original paper (Beesack, 1956).
Beesack considers a number of other problems such as bounds for the
distance between zeros, nonoscillation in sectors, and strips or circles.
Some of these may be reducible to cases already settled by the use of
conformal mapping. As a sample we consider a special result due to B.
Schwarz (1955).

THEOREM 8.3.5

Let G(z) be holomorphic in the right half-plane, where it satisfies

|G(x + iy)| ^x~2. (8.3.24)

Then no solution of (8.3.1) can have more than one zero in the right
half-plane.
Proof. The transformation

'=YT? vv(z) = (i + o^(0

maps the right z-half-plane into the unit disk of the r-plane and carries
(8.3.1) into

Y"(t ) + 4(1 + ty4G(j^j Y(t) = 0. (8.3.25)

Here the coefficient of Y(t) does not exceed in absolute value

(1-1* I2)"2, (8.3.26)

so we are back to the corollary of Theorem 8.3.3. ■
We end this discussion with a result of the present author, which is
stated without proof.

THEOREM 8.3.6

Let G(z) be holomorphic in a sector S: \arg z| ^ a < ir. Let G(x + iy) be
308 COMPLEX OSCILLATION THEORY

integrable out to infinity for fixed y and satisfy

|zJ G(t)dt|^p <i z6S, (8.3.27)

where the path of integration is parallel to the positive real axis. Then
(8.3.1) has a solution without zeros in S.

For proof see Hille (1948), pp. 695-698.

In view of this result one might be inclined to believe that all solutions
have only a small number of zeros in S. That this is not the case is shown
by an example used by Beesack for a similar purpose:
w» + e-2zw = 0, (8.3.28)

which is satisfied by any Bessel function of order 0 in the variable e~\

Any sector S with a < \tt will do. We choose a solution of (8.3.28) with a
large positive zero x0. This solution also has complex zeros Xo + kiri, k =
0, ± 1, ±2, . . . . The larger jc0, the more zeros are in S. There is no solution
with infinitely many zeros in S, but there is obviously no finite upper bound
on the number of zeros that a solution can have in 5.

EXERCISE 8.3
1 . Verify the various statements about the circle in the proof of Theorem 8.3.3.
2. Verify (8.3. 14)-(8.3. 16).
3. Check the equivalence proof in Theorem 8.3.3.
4. Let w" + a(\ - z2) 2w = 0 with a real >1. Show the existence of solutions
with infinitely many zeros in (-1, 1).
5. Verify (8.3.25) and (8.3.26).
6. Give an example of a function satisfying (8.3.24).
7. Answer Problem 6 for (8.3.27).
8. Determine the half-plane \a + r ei0\** r.
9. Show that the function Y(t) = A -\t~x satisfies Y'(t) = [Y(t)f + F(t) with
F(0 = ir2 + Ar*-A2 so that F(t) and Y(t) satisfy the hypotheses of
Lemma 8.3.1. How does this lead to condition (8.3.23)?
10. Show that the inequality \G(a + r e'e)\ \r 2 + Ar 1 - A2 is satisfied if
\a+re,0\^r. Conclusion? How does the condition |z-a|<|a| lead to a
zero-free region?

8.4. APPLICATIONS TO SPECIAL EQUATIONS

The methods developed in the preceding sections will now be applied to
some special equations. Some of the equations treated in Chapter 6 are
considered again here from this point of view.
APPLICATIONS TO SPECIAL EQUATIONS 309

The Hypergeometric Equation

The equation

z(l - z)u"+[c -(a + b + l)z]u'- abu = 0 (8.4.1)

is not in a form suitable for our consideration and is replaced by the
invariant form

w" + \l{\-a])z-2 + {\-a22)(z -ir2 + (a? + a? + a^-a?-l)

xz_1(z-l)"> =0. (8.4.2)
Here

w{z) = zp{\-zYu{z), p =!(!-«,), cr=Kl-«2), (g

a = 1(1 - a, - a2 + a3), b = ^(\ - ax- a2~ a3), C = 1 - a ,.

In 1907 A. Hurwitz got results for u0\(z) = F(a, b, c; z). Here we shall aim
at Mooi(z) = (-z)~aF(a, a-c + l,a-Hl;z"') and large neighborhoods
of infinity. For this choice of w(z) we must assume that w(z)->0 as z
becomes infinite. If the coefficients are real, as we suppose in the
following, we must have a — b = a3 > 1.
We can now use Green's transform and integration to infinity along
vertical or horizontal lines. If the coefficient of w in (8.4.2) is denoted by
G(z) = P(x, y) -i- iQ(x9 y), z = jc + iy, we have

[w(0w'(0]:-Jz |w'(0|25i + J2 (P + /Q)|w|2df =0.

Now w(z) goes to zero as zl~a+b as z becomes infinite. This implies that
the integrated part goes to zero and the first integral exists. The second
integrand is o (z 2), so again the integral exists. If now the integral is taken
along a vertical line in the upper half-plane, the sign of Q(x, y) along the
path of integration is decisive for the success of this mode of attack. We
have

Q(x, y) = -2y{A*|z|-4 + B(x - l)|z - 1|~4 + C{x -i)\z\~2\z - 1|-2}, ^

A = k(\- a2), B = J(l - ol\), C = \(a] + a\ - a\ - 1).
To fix the ideas, suppose that all three coefficients, A, B, C, are of the
same sign or are zero; then Q(x, y) keeps a constant sign in the four
quarter-planes
x^O, yssO; x =s 0, y 3=0; x 5* 1, y 3=0; x^l,y^0.
These quadrants are thus zero-free regions for

(-z)"aF(a, a-c + l,a-fc + 1; z"')

310 COMPLEX OSCILLATION THEORY

under the stated conditions on A, B, C. The strip 0 < x < 1 is in doubt. If,
however, A ^ B ^ 0, C 2* 0, the right half of the strip is zero-free.
Here we get better results by integrating radially from a hypothetical
zero to infinity. Now

-J" |w'|2 dr + J~ G{reie) e2ie\w\2 dr = 0.

Then the imaginar y part of G(re'e) e2,e is
sin 6{2B\r - e~ie\'\r + cos e) + Cr~l\r - e~ie\-2}. (8.4.5)
If B and C have the same sign and if 0 < S < \tt, the expression is not zero
and we see that there can be no zeros in the open first and fourth
quadrants. With these fragmentary results we leave the hypergeometric
equation.

Legendre's Equation
Here more satisfactory results are obtainable. We start by considering

Pa(z) = F[a + l, -a, 1; iCl -z)L

which satisfies the equation

Tz [° " z2)^f] + a {a + l)w = °* (8 4'6)

It is assumed that either a is real 2= — I or a — — 5+ mi, 0< m. In either
case Pa(z) is real for z real, z > - 1. We know that P„(z) is a polynomial
of degree n, all the zeros of which are real and located in the interval
(— 1, 1). For a nonintegral value of a the z-plane should be cut along the
real axis from -00 to - 1 in order to deal with a single-valued function. We
shall show that in the cut plane Pa(z) has only real zeros; for a real - \ =s a
the zeros lie in the interval (- 1, 1), and for a — —5 + im they lie in (1, 00)
and are infinitely many. In the real case the number of zeros is (a], where
this symbol denotes the least integer > a but less than a -I- 1. To show the
reality of the zeros we use Green's transform — more precisely, the
zero-free star S0(l).
We have then, for w = Pa(z), z = 1 + r eie, and supposing that 1 + R eie
is a zero of PaP'a, we obtain

Jo (2re ie + r2)\w'\2 dr + a(a + \) j |w|2dr = 0, (8.4.7)

the imaginary part of which is

2 sin 0 j r\w'\2 dr.

APPLICATIONS TO SPECIAL EQUATIONS 311

This is zero iff 0 is congruent to zero (modulo tt). Thus the zeros under
consideration are real. If 6 = 0 and a 2* 0, the left member of (8.4.7) is
positive and no zeros can belong to (1, o°). If 6 = tt, R < 2, a(a + 1) < 0,
we have again a contradiction. If a is real, there can be no zeros on the
edges of the cut, for on the cut the imaginary part of Pa(z) is
± iriPa (- z) 5* 0. One way of establishing the number of zeros in (— 1, 1) of
Pa(z), — i< a, is to find the change in the argument when z describes a
great semicircle in the upper half -plane. For the case a =— i+mi, the
Liouville transformation maps (8.4.6) on a perturbed sine equation. Since

Z(z) = (i + m2),/2 log [z + (z2 - 1)1/2], (8.4.8)

the number of zeros of Pa(x), a = -\ + mi, in the interval (1,A) is
approximately

-(Km2)"2
IT log 2 A (8.4.9)

We turn now to the function

Qfl(2) = -ilr(^.t1) (2zTx-aF(ha +Ua + 1, a +§; z~2) (8.4.10)

where a is real, -\<a. To make the function single valued we cut the
plane along the real axis from — °o to 4- 1. We can restrict ourselves to the
closed right half -plane since if z0 is a zero so is -z0. Here we use the
Green's transform and integration along rays from the origin, starting at a
hypothetical zero z0 and going to oo. From the behavior of Qa(z) for large
values of z we see that Qa(z)Q'a(z)(\ - z2)-»0 as z->o°. Furthermore,
|Q'a(z )|2(1 - z2) and |Qa(z)|2 are integrable out to infinity. We thus get

-f (e Tl6 -r2)\w'f dr + a(a + \)\ \w?dr = 0. (8.4.11)

and here the imaginary part is

-sin 26 ( \w'\2dr,

which is zero iff sin 26 = 0 or 6 is a multiple of {tt. For 6 =\tt and a ^ 0

this leads to a contradiction, as shown by (8.4.11). For 6 = 0, r0> 1, a >0
we again get a contradiction. We conclude that for a ^ 0 the function
Qa (z ) can have no zeros in the cut z-plane. For - \ < a < 0 we are in doubt
concerning the existence of real zeros > 1. But for such values of a the
coefficients of Qa(z) in the representation (8.4.10) are all positive, so that
there can be no zeros on the interval (l,o°). Thus we have shown that
Qa(z) for a >— \ has no zeros in the cut plane.
312 COMPLEX OSCILLATION THEORY

Bessel's Equation
We start by proving Hurwitz's theorem, stating that Ja(z) for a > - 1 has
only real zeros. Suppose contrariwise that z = c is a nonreal zero of Ja(z).
The function Ja(ct) in its dependence on t satisfies the DE

(tu')' + (c2t-a2tx)u = 0. (8.4.12)

Here we multiply by ii and integrate from t = h >0 to t = 1. An in-
tegration byparts gives

[tuu'VH-j^ t\u'(tfdt + (c2t-a2t1)\u(t)\2dt=0. (8.4.13)

Here u(\) = Ja(c) = 0 by assumption, so the equation reduces to

-hu(h)u'(h)- j t\u'(t)\2 dt + j (c2t - a2t~l)\u(t)\2 dt = 0.

Now, if a complex number is zero, its conjugate is also zero and so is their
difference. In forming the latter, the first integral drops out and in the
second
left withintegral the term involving a2t~l also drops out; therefore we are

-h[V(h)u'(h)-u(h)u'(h)] + (c2-c2) j t\u(t)\2 dt = 0. (8.4.14)

Here the first expression is

F(a)\c\2a(c2- c2)h2a+2[\ + 0(h2)l (8.4.15)

and this goes to zero with h for any a >-l. Here F(a) is an analytic
real-valued function of a of no importance for the following. Hence we
are left with

(c2-c2) J"Qt\u(tfdt =0.

The integrand is 0(f2a+1), and hence the integral exists. Since the integral
cannot be zero, the only way for the result to hold is that

c2 = c2 or either c = c or c=—c.
In the first case the zero would be real; in the second, purely imaginary.
The second alternative cannot hold, however, if a>-l, for in the
expansion

JAiy ) = exp fr* ) i r(fc + 1)r('fc+ a + |)(f

)"" (8.4. .6)
all the coefficients are positive. This proves Hurwitz's theorem.
APPLICATIONS TO SPECIAL EQUATIONS 313

Hurwitz also discussed the case of a < - 1, where in addition to the real
zeros there are complex zeros. If - n — \<a<-n, there are n pairs of
conjugate complex zeros symmetrically located with respect to the origin.
If n is odd, there is one and only one pair of purely imaginary zeros.
Next we take up the complex zeros of a solution Ca(z) of BesseFs
equation such that C(z) is not real on the real axis (or a complex number
times a real solution). Suppose that z = Zi = Xi + iyi is a zero of C(z) in
the first quadrant. To study the implications of such a zero we consider
the DE

w" + [\-(a2-\)z 2]w = 0, (8.4.17)

which is satisfied by z,/2C(z). Here there are three cases according as
a2 < J, = J, or > J. The case a2 = \ does not pose any problem, since here
(8.4.17) is a pure sine equation and the solution which has a zero at z = Z\
is C sin (z — Zi) with the zeros Z\ ± mr.
Let a2 multiply (8.4.18) by w, and integrate along the line y = y,.
For x 25 0 there can be no other zeros on the horizontal line through
z = Zi, for if x2 = jc2+ iy\ were a second zero, then

-j 2\w'\2dx + j (P + iQ)\w\2 dx =0, (8.4.18)

where

P (jc, y ) = 1 - (a 2 - l,)\z \ \x 2 - y 2), Q(Xyy) = (a2- \)\z\'42xyy (8.4. 19)

and y = y{. The imaginary part of (8.4.18) is then

2yy(a2-\) j 2 x\w(x +iy,)|2dx*0

for x2 2= 0. In the same manner it is proved that no two zeros in the first
quadrant can have the same abscissas.
Suppose now that z2 = x2 + ry2, jc, < x2, is a second zero of Ca(z) in the
first quadrant. Here yi < y2 or y, > y2 according as a2 — \ < 0 or >0. This
follows from Theorem 8.1.6, for the sign of Q(x, y) is the same as that of
a2 - i Thus

0<argG<7r, -57r<-UrgG<0 \i a2>\,

-7r<argG<0, 0<-UrgG<|7r ifa2<4,
whence the result follows.
In any case, if the zeros z„ = xn + iy„ are ordered according to
increasing abscissas, x„ <x„+i, the ordinates form a bounded monotone
sequence of positive numbers which tend to a positive limit as n -»<».
Furthermore, lim (xn+i - xn) = it. This follows from the discussion in
314 COMPLEX OSCILLATION THEORY

Section 5.6, since Hypothesis F is satisfied. If w(z) is an oscillatory

solution of (8.4.17), there is, by Theorem 5.6.1, an associated asymptotic
sine function, C sin (z - z0), such that lim (z„ - z0- nir) = 0, and here z0
cannot be real, for there is a one-to-one correspondence between
solutions Ca(z) and z0, reduced modulo ir, and real values of z0
correspond to solutions real on the real axis.
It is natural to ask whether the convexity theorem, Theorem 8.1.7,
applies in the present case to the zeros in the first quadrant. Now the
argument of G(z) lies between 0 and tt if a2 - \ > 0 and between - tt and
0 if m = a2-k<0. This is satisfactory, but the inflection locus of the
indicatrices places restrictions on what part of the quadrant is admissible.
By (8.1.19) this curve is

Im{G'(z)[G(z)r3/2} = 0, (8.4.20)
which forms a part of the combined inflection locus

Im{[G'(z)]2[G(z)r3} = 0. (8.4.21)
The sign of m affects the system of indicatrices in various ways. Thus the
origin, which is a topological singularity of the system, is a center
(= vortex) of the system if m >0 but a star point if m <0. The curve
(8.4.21) is a sextic which factors into xy = 0, i.e., the axes, times the
quartic

(3 jc 2 - y 2)(\x 2 - y 2) - 2m (x 2 - y 2) + m 2 = 0. (8.4.22)
The case m > 0 is sketched roughly in Figure 8.1. For m < 0 the graph is
obtained by interchanging x and y, i.e., a rotation of the figure through an
angle of 90°. If m > 0, there are double points with real tangents at
z=m,/2 and z=-mi/2. The lines y=±31/2x and y=±3~1/2x are
asymptotes. Now in the upper half-plane the two branches of (8.4.22)
whose asymptotes make the angles 30° and 150° with the horizontal
belong to the inflection locus of the indicatrices. They divide the upper
half-plane into three curvilinear sectors in each of which the convexity
theorem holds. In the middle sector the indicatrices are concave
downward; in the other sectors, concave upward. It may be shown that a
solution Ca{z) which has infinitely many zeros in the first quadrant can
have at most a finite number in the fourth quadrant.
We still have to consider the Bessel-Hankel functions which
correspond to the solutions e'z and e ~iz of the associated sine equation. It
is enough to consider the first case, Ha\z), which is z"1/2 times a
subdominant solution of (8.4.17). If z0 = x0 + iVo is a zero, then, using the
Green's transform and integration along the vertical line from z0 to
APPLICATIONS TO SPECIAL EQUATIONS 315
\\
\ \\\
\ \

\ \
\ \
\ \ \

/ /
/ /
/ /

-(m)/2\ 0
/ \ 1 \

Am)*

Figure 8.1

2o + i°°, we see that

f Q\w\2dy=0,
Jyo f [\w'\2 + P\w\2]dy = 0.
Jyo

In the upper half-plane the first of these identities holds only on the
imaginary axis, while the second one cannot hold in the part of the plane,
visible from /o°, where P ^0. The curve P = 0 is the lemniscate

U2 + y2)2 = m(x2-y2), m=a?-i

The foci of this curve are at z = ±(km)m if m >0 and at ±i(i\m\)m if
m < 0. For m > 0 this means no zeros in y 0 or in the parts of the third
and fourth quadrant to the left of the line x = - m 1/2 and to the right of the
line x = m 1/2, respectively. For m < 0 the segment of the imaginary axis
between 0 and i'|m|1/2 remains in doubt, while the rest of the upper
half-plane is zero-free. The regions in the lower half-plane are now
bounded by x = ±\(6\m\Y'2. Here we leave Bessel's equation.

The Hermite-Weber Equation

This equation is written in the form

w" + (2a + 1 -z2)w =0, (8.4.23)

Mb COMPLEX OSCILLATION THEORY

where a is real and >— i We recall the existence of four subdominant

solutions, Ea(z), E tJ ,(-/z), Ea(-z), E-a-x(iz), for the four sectors
bounded by the lines y = ±x.
We note first that a solution which is either even or odd cannot vanish
anywhere in the two sectors that contain the imaginary axis. This follows
from a construction of the zero-free star S(0), which is left to the reader.
Part of this result is true also for a solution whose initial parameter A (0) is
real. If A (0) > 0, the left halves of these sectors are zero-free. If, however,
A(0)<0, we get the sectors \tt < 6 < \tt and -\tt < 6 <-{ir. The first
result is practically the best of its kind, but for A(0)<0 it is less than
satisfactory, for the Liouville transformation shows that the indicatrices
have the lines y = ±x as asymptotes, and this means that an oscillatory
solution has infinitely many zeros forming four strings such that the
distance of the nth zero from the asymptote goes to zero as n becomes
infinite. See also (8.2.27)-(8.2.30) and the adjacent discussion.
Let us now take up the zeros of E«(z). Here we see first that the right
half -plane can contain no zeros of Ea(z)E'a(z) except real ones on the
interval [0, (2a + l)l/2]. We use the Green transform and integrate along
horizontal lines, applying the fact that E(1(z)E'a(z) goes exponentially to
zero as z in the sector |arg z| ^ \tt - e. Since Ea(0)E'a(0) < 0, we get in
addition no zeros in \tt < 6 «s \tt and the symmetric sector in the lower
half-plane. A more painstaking examination would enable us to extend the
zero-free regions, but it is not worth the effort. As a matter of fact, the
solution E a \{—iz) is easier to study from this point of view. Here there
are no zeros in the upper half -plane. The zeros in the third quadrant lie in
a region bounded by the hyperbola

x2-y2 = a + l2, 0>jc, 0>y,

by the cubic

x'-3xy2-3(2a + \)x + 2(2a + 1)3/2 = 0, 0>jc, 0> y,

and a segment of the real axis from -(a +!)l/2 to -{2a + 1),/2.
Normally a solution of (8.4.22) has four strings of zeros, one in each
quadrant. Let S7 stand for the string in the jth quadrant. Then Ea(z) lacks
the strings S, and S4, while E_a-i(- iz) lacks Si and S2, and so on. If a is a
nonnegative integer, Ea(z) is a constant multiple of Ea(- z) and
consequently lacks all four strings. Similar results hold for E-a-\(-iz)
when a is a nonnegative integer.
In the case of this equation is it possible to study the composition of the
strings by the method of variation of the parameter, in this case variation
of a. We recall that the string is the string of zeros of w(z)w'(z), so that it
contains both zeros and extrema of w(z), and one problem is to decide
APPLICATIONS TO SPECIAL EQUATIONS 317

whether or not they alternate. To fix the ideas, let us start with a = —5,
i.e., the DE

w"-z2w = 0, (8.4.24)
and the solution, which is zero at z = 0. This solution is z times a power
series in z4 with positive coefficients. Hence on the four rays
arg z = {2k + 1)577 it is of real character (i.e., a complex constant times a
real-valued function). If we set z =exp(;j7n)f, (8.4.3) becomes

jjp+t2v = 0, v(t) = exp(-Wi)w(t).

It follows that the solution is oscillatory on this ray, as well as on the three
other rays. Since there is a zero at the origin, the nearest zeros of
w{z)w'{z) on the four rays are extrema of w(z). Hence on all four rays
the pattern is E - Z - E - • • • , where E stands for an extremum and Z for
a zero. Now the zeros and extrema are analytic functions of a — in
particular, continuous functions, and they are also continuous functions
of the initial parameter A(0). It follows that, if we vary either of these
parameters, the configuration will change continuously, and since no two
zeros or extrema can have the same (nonzero) abscissas or ordinates, the
pattern is preserved unless A(0) tends to a limit for which two or four
strings are missing, the corresponding points having moved off to infinity.
New zeros are fed into the system, one at a time, from infinity. Extrema,
on the other hand, appear on the real axis by coalescence of a pair of
complex extrema at a point ±{2a + 1)1/2, where they separate and march
off in opposite directions. It is quite a fascinating study, which should be
pursued by any reader who is curious to know where zeros and extrema
come from and go to.

The Equation of Mathieu

In the normal form

w"(z) + (a + b cos2z)w(z) = 0, (8.4.25)

we assume 0 < b < a. The Liouville transformation will then map the
half-strips D+n and Dn of (6.6.26) conformally on domains containing an
upper or a lower half -plane, as in Figure 6.5. The equation (8.4.25) is real
in nature on each of the lines x = \kir. If k is odd, there are oscillatory
solutions with infinitely many zeros and extrema on the line and the
extrema alternate with the zeros. Variation of an initial parameter may
push one of these points off the line. Then they will all go and to the same
side of the line. According as the string is left of the line, on it, or to the
318 COMPLEX OSCILLATION THEORY

right of it, we speak of /-, r-, or d-strings (laevus, rectus, or dexter). The
Mathieu functions, ce„(z) and se„(z), have all their complex zeros on the
lines x =(k - {)tt, so they are all r-strings. Zeros and extrema alternate,
and the point nearest to but not on the real axis is an E or a Z according as
z = (k - \)ir is a Z or an E. In addition to the complex zeros, there are also
infinitely many real zeros. If the string 5^ is a d-string and if w(z) is real
on the real axis, then S„ is also a d-string, but to what extent these facts
influence the nature of the other strings is unknown so far. The author
(1924) has given a complete description for the case a < b <0.

EXERCISE 8.4
1. From (8.4.1) derive (8.4.2) and (8.4.3).
2. Verify (8.4.4) and (8.4.5). Why is it permissible to integrate to infinity?
3. In studying the variation of the argument of Pa{z) the following path is
considered: (i) real axis from 1 to R, (ii) semicircle |z| = R, 0 s£ arg z ^ ir, (iii)
real axis from -R to - 1 - e, (iv) semicircle |z + 1| = e, (v) real axis from
- 1 + e to 1, avoiding the zeros by small semicircles in the upper half -plane.
The total variation should be zero. Find the variation along (i)-(iv) with an
error < tt, and use this to determine the number of zeros of Pa(z) in (— 1, 1).
4. Verify (8.4.8) and (8.4.9).
5. Show that it is permissible to use infinity as the upper limit in (8.4.11).
6. Verify (8.4.12) and the proof of Hurwitz's theorem.
7. Show that for a >\ the function x*-+ Ja{x) has no zeros in the half -open
interval (0, mI/2], m = a2 -\.
8. Suppose that m > 0 and that z„ is a distant zero of Ca (z) in the first quadrant.
Set m\zH r=P«, j3„ = arcsin P„. Show that (i) |arg G(z„)| ^ j8„, (ii) /3„+1 < j3„,
(iii) 27 j8„ < <».
9. Discuss the shape of the indicatrices of (8.4.17) in a neighborhood of the
origin.
10. Derive (8.4.21) from (8.4.20), and determine the various parts of the inflection
locus.
11. Verify the statements about the zeros of the Hankel-Bessel function.
12. Discuss the zero-free regions of a solution of (8.4.23) for which the initial
parameter A(0) is real.
13. Discuss the zero-free regions of Ea(z).
14. Show that E-a-\{- iz) is of real nature on the imaginary axis, and show for
this solution A (0) = bi, b > 0.
15. Discuss the zero-free regions of E-a-x(-iz).
16. Verify the results stated for (8.4.24).
17. Suppose that for this equation the initial parameter A (0) > 0. For A (0) = +oo
there is a zero at the origin. How does this zero move when A (0) decreases
from +oc? How are the four strings pushed?
LITERATURE 319

18. When A (0) = 0, there is a triple extremum at the origin. What happened to the
zero? Indicate how the triple extremum was generated.
19. Discuss the singular boundary value problem

w "(x ) + (2a + 1 - jc 2) w (x ) = 0, w(x)6L2(-«',<»).

20. In a discussion of the singular boundary value problem

w"(x) + (b -x4)w(x) = 0, w(jc)GL2(-oo,oo),

E. C. Titchmarsh (1899-1963) conjectured that the solutions would have real
zeros in the interval (- bm, bV4), infinitely many purely imaginary zeros, and
no other complex zeros. Verify that this is so by discussing the equation in
the complex plane. Verify the existence of six sectors bounded by the rays
arg z = i(2k - \)tt, in each of which there is a subdominant solution that goes
to zero in the interior of the sector as r~lQxp(-\r}) as r-><*> and
consequently lacks the two adjacent strings. The solutions of the boundary
value problem correspond to those values of b for which the subdominants
in the sectors |arg z| < \tt and |arg (-z)\< \ir are linearly dependent. In this
case four strings are missing, *and only the strings in the direction of the
imaginary axis are left. Fill in the details.
21. Show that, if w(z) is a solution of (8.4.25) which is oscillatory in Do, is real
on the real axis, and has a single zero or extremum in (-{it, \tt), then the
string in Do is /-, r-, or d-, according as the real zero or extremum < 0, = 0, or
>0.

22. Discuss the nature of the string in Do when that of the string in Do" is known.
23. Can a singular boundary value problem

-jpr + (a - b cosh 2t)v =0, v(t) G L2(-°o, <*>),

have a solution? Here b is a given positive number.

LITERATURE

See, in particular, Chapter 11 of the author's LODE, which has an

extensive bibliography. See also Chapter 21 of:
Ince, E. L. Ordinary Differential Equations. Dover, New York, 1944.
For the papers mentioned in the text see:
Beesack, P. E. Non-oscillation and disconjugacy in the complex domain. Trans. Amer.
Math. Soc, 81 (1956), 211-242.
and B. Schwarz. On the zeros of solutions of second order linear differential equations.
Can. J. Math., 9 (1956), 504-515.
Boutroux, P. Recherches sur les transcendentes de M. Painleve et Tetude asymptotique des
equations differentielles du second ordre. Ann. Ecole Norm. Super., (3), 30 (1913),
255-375; 31 (1914), 99-159.
320 COMPLEX OSCILLATION THEORY

Forsyth, A. R. Theory of Differential Equations. Part I, Vol. II. Cambridge University Press,
London 1900, pp. 275-276.
Hurwitz, A. Uber die Nullstellen der Bessel'schen Funktion. Math. Ann., 33 (1889), 246-266.
Uber die Nullstellen der hypergeometrischen Funktion. Math. Ann., 64 (1907),
517-560.
Nehari, Z. The Schwarzian derivative and schlicht functions. Bull. Amer. Math. Soc, 55
(1949), 545-551.
On the zeros of solutions of second-order linear differential equations. Amer. J. Math.,
76 (1954), 689-697.
Some criteria of univalence. Proc. Amer. Math. Soc, 5 (1954), 700-704.
Univalent functions and linear differential equations. Lectures on Functions of a
Complex Variable. University of Michigan Press, Ann Arbor, 1955, pp. 148-151.
Pokarnyi, V. V. On some sufficient conditions for univalence (Russian). Doklady Akad.
Nauk SSSR, N.S., 79 (1951), 743-746.
Schafheitlin, P. Uber die Nullstellen der hypergeometrischen Funktion. Sitzungsber.
Berliner Math. Gesell. 7 (1908), 19-28.
Schwarz, B. Complex non-oscillation theorems and criteria of univalence. Trans. Amer.
Math. Soc, 80 (1955), 159-186.
Taam, C-T. Oscillation theorems. Amer. J. Math., 74 (1952), 317-324.
On the complex zeros of functions of the Sturm-Liouville type. Pacific J. Math., 3
(1953), 837-845.
Schlicht functions and linear differential equations of the second order. /. Rat. Mech.
Anal., 4 (1955), 467-480.
On the solutions of second-order linear differential equations. Proc. Amer. Math. Soc,
4 (1953), 876-879.
Van Vleck, E. B. A determination of real and imaginary roots of the hypergeometric series.
Trans. Amer. Math. Soc, 3 (1902), 110-131.
Mention should also be made of six papers of the author, results of
which are cited in the text:
Hille, E. On the zeros of Legendre functions. Ark. Mat., Astr., Fys., 17, No. 22 (1923), 16 pp.
On the zeros of Mathieu functions. Proc London Math. Soc, (2), 23 (1924), 185-237.
On the zeros of the functions of the parabolic cylinder. Ark. Mat., Astr., Fys., 18, No.
26 (1924), 56 pp.
A note on regular singular points. Ark. Mat., Astr., Fys., 19A, No. 2 (1925), 21 pp.
Non-oscillation theorems. Trans. Amer. Math. Soc, 64 (1948), 234-252.
Remarks on a paper by Zeev Nehari. Bull. Amer. Math. Soc, 55 (1949), 445-457.
9

LINEAR nth ORDER

AND MATRIX

DIFFERENTIAL

EQUATIONS

This chapter is devoted to the theory of linear DE's of order n, normally

greater than 2. For many purposes it is more convenient to consider an
equivalent matrix equation. The passage from an nth order linear DE to a
first order matrix DE or vice versa is elementary. Among the various
problems to be considered are (i) existence and independence of
solutions, (ii) existence and analyticity in a star, (iii) the group or
monodromy, (iv) global representation and the Riemann problem, (v) rate
of growth for approach to a singular point, (vi) regular-singular points,
(vii) the Fuchsian class, (viii) irregular singularities.

9.1. EXISTENCE AND INDEPENDENCE OF SOLUTIONS

We start with an nth order linear DE,

P0(z)w{n\z) + Pl(z)win-l\z) + • • • + Pn(z)w(z) = 0, (9.1.1)

where the P, 's are holomorphic in the domain D under consideration and,
furthermore, P0(z) ^ 0. Some DE's are conveniently expressed in terms
of the differential operator:

0/(z) = z/'(z), #n = W1), (9.1.2)

with a corresponding DE

dnw + Q,(z)dn_lw + • • + Qn(z)w = 0. (9.1.3)

321
322 LINEAR nth ORDER EQUATIONS

We can also replace the nth order DE (9.1.1) by a matrix DE:

W'(z) = s£{z)W(z), sd(z) = [aik(z)l W(z) = [wjk(z)l (9.1.4)

Here sd and W are elements of the Banach algebra Wln of n-by-n
matrices, and the entries aik(z) are holomorphic in D, while the wjk's are
locally holomorphic there.
To pass from (9.1.1) to (9.1.3) we should note that

zTXz) = W - W - 21) ...[#- (k - l)/]/(z), (9.1.5)

where I denotes the identity operator, and k = 1, 2, . . . , n. We now divide
the equation by P0(z), and in the term involving the kth derivative we
write
w{k\z) = z-kzkw(k\z).
Using (9.1.5), we get a result of the form

k2= \ BH-dz){K& - !)(# " 21) . . . [d - (k - \)I]w + Bn(z)w = 0, (9.1.6)

where the B's are holomorphic functions of z in D, except possibly for

the origin, if z = 0 belongs to D. Here we expand the operator
polynomials and collect terms in order to obtain (9.1.3).
1,
The passing from (9.1.1) to (9.1.4) may be achieved in various ways.
The simplest one is to set

w = wu w' = w2, w("-1)=uv (9.1.7)

We form the column vector w, whose components are Wi, w2, . . . , w„, and
the vector-matrix DE,
w'=s£(z)w, (9.1.8)
where s£(z) = (ajk(z)) and
flu+i(z)= 1 J = 1,2,.
aUk(z) = 0, k^j + 1

an,k{z) = ~lfj^, k = l,2,...,n.

Other transformations are possible and may be preferable for particular
problems. We shall encounter such cases in the following.
From the vector-matrix equation we can proceed to a pure matrix DE in
the following manner. Take a point z0 in D, and determine n solutions of
(9.1.8) by the initial conditions

w(kiXzo) = 8i+uk, j =0, 1 w — 1; fc = l,2,...,n. (9.1.10)

These conditions determine uniquely n distinct solution vectors of (9.1.8)
EXISTENCE AND INDEPENDENCE OF SOLUTIONS 323

in some neighborhood of z = z0. The solution vectors wk(z) are linearly

independent as elements of C". We take these n column vectors as
column vectors of a matrix W(z), and this matrix satisfies (9.1.4).
Equation (9.1.1) has the property that, if w0(z) and u>oo(z) are two
solutions and G, C2 any two constants, then C\ w0(z) + C2Woo(z) is also a
solution, and similarly for the vector-matrix equation. For the matrix DE
we have a more general situation: whenever °W0(z) is a solution and M is
an n-by-M matrix, then °W0(z)M is also a solution. This follows from the fact
that
[°Wo(z)MY = W0(z)M = si(z)[°W0(z)M],
since matrix multiplication is associative. It is normally noncommutative,
so that Msi(z) 7* si(z)M and M°W0(z) is normally not a solution.
A solution matrix °Wo(z) of (9.1.4) is called fundamental if every
solution of the DE is of the form °W0(z)M.

THEOREM 9.1.1

A solution Woiz) of (9.1.4) is fundamental iff ^(z) is a regular element of

the matrix algebra, i.e., has a unique inverse, and this is the case iff
oj(z) = det pf0(z)] * 0.
Proof. Suppose that for some z = z0ED the determinant (o(z0)9£0.
Since w(z) is a holomorphic function of z in D, there is a neighborhood
N(zo) of z0 in which o>(z)^0. It follows that \°W(z)]~l exists and is
holomorphic in N(z0). Now let °lV\z) be any nontrivial solution of (9.1.4),
and set <W(z0) = Mo. Here it is assumed that °W(z) is defined and
holomorphic in N(z0). Then

<Wo(z)[<Wo(zo)]10^(zo) (9.1.11)
is a solution of (9.1.4) holomorphic in N(z0), and at z = z0 it takes the
value °lf(z0). Now there is one and only one solution of (9.1.8) with these
properties, namely, °W(z), so we must have °W(z) represented by (9.1.11);
i.e., the arbitrary solution °W(z) admits of a representation of the form
°W0(z)M. Thus the condition co(z) ^ 0 is sufficient to make 1Vo(z) funda-
mental.
It is also necessary, for if every solution of (9.1.4) is of the form
°W0(z)M, this applies also to the solution whose initial value at z = z0 is %,
the unit matrix of the algebra Then

Wo(z0)jM,o = % det rWoUo)] det [M0] = 1 . (9. 1 . 12)

Since w(z) is holomorphic at z = z0, we must have w(z0) ^ 0, so that the
condition is also necessary. ■
 LINEAR nth ORDER EQUATIONS

theorem suggests that a solution matrix defined in the domain D is

regular either everywhere in D or nowhere. That this is indeed the case
follows from the explicit representation of (o(z), a formula which for the
nth order linear DE goes back to N. H. Abel.
We shall consider a general matrix equation (9.1.8) not necessarily
obtained from a linear nth order DE. The matrix s£(z) = [ajk(z)] shall be
holomorphic in D; i.e., the ajk shall be holomorphic in the sense of
Cauchy. We recall that the trace of a matrix is the sum of the diagonal
elements:

trd = 2fltf ifd = [ajk]. (9.1.13)

We have now 7= 1

THEOREM 9.1.2

Let W(z) be a solution of (9.1.8), defined, holomorphic, and regular in

some neighborhood N(z0) of z = z0 E D. Then

det [°W (z)] = (o(z) = w(z0) exp {J* tr [d(t)] dtj. (9.1.13)
The function a)(z) defined by this formula is locally holomorphic in all of
D, and the formula is trivially valid also if W{z) is singular at z0.
Proof. Take z in N(z0), and choose 8 so small that all points z + h with
|fc|<5 are also in N(z0). We have then

°W(z + h) = W(z) + hW(z) + 0(/i2) = °W(z) + hd(z)°W(z) + 0(h2)

= [% + hs#(z)]W(z)+0(h2)
= {% + hs£{z)+ 0(h2)[W(z)]~l}W(z),
where we used the algebraic regularity of the solution. We now form
determinants to obtain

co(z + h) = + h^(z)+ O(h2)]co(z),

whence

^ Mz + h) - co(z)] = ^ {«)[% + fi^(z)] + 0(h2) - l}w(z).

Here
u>\% + /i^(z)]= 1 + ft tr[^(z)] + 0(h2).
It follows that
<o'(z) = tr [d(z)]co(z),
and formula (9.1.13) results.
The function co(z) defined by the formula evidently exists everywhere
 325
EXISTENCE AND INDEPENDENCE OF SOLUTIONS

in D and is locally holomorphic there. It is holomorphic in D if D is

simply connected or, more generally, if cu(z) is single valued. If °W(z0) is a
singular matrix, so that co(z0) = 0, the formula still defines a function
holomorphic in D, namely, the zero function.
It is not clear from the results obtained so far that (o(z), be it ^ 0 or =0,
gives the determinant of °W(z) in all of D or in the domain of existence of
°lf(z), whichever is larger. The main difficulty here is that our solution is
defined only locally, namely, in N(z0). It is essentially a question of the
analytic continuation of the matrix-valued function z »-> °W(z), on the one
hand, and the Cauchy holomorphic function, z cu(z), on the other. In
both cases we may appeal to the law of permanency of functional
equations. The matrix function °W(z) satisfies the DE (9.1.8) in N(z0). The
matrix si(z) exists and is holomorphic in D. Now start from z = z0 along a
path L in D. The analytic continuation of d(z) exists everywhere along L.
Since d(z) is holomorphic everywhere on L, the existence theorems
show that (9.1.8) has analytic solutions at all points of L. If z = z, is such a
point, and if the analytic continuation of °W(z) along L gives the value
°lf(zi) at z = Zi, the analytic continuation of °W(z) along L agrees at z = zx
with the local solution, which takes on the value °lf (zj) at z = zx. It follows
that °W(z) exists in all of D and is a solution of (9.1.8). It is locally
holomorphic and will be holomorphic in D if it is single valued.
On the other hand, the function z^wfz) defined by (9.1.13) exists and
is locally holomorphic in D. If a)(z0) ¥■ 0, we have

det[°!4/Xz)] = w(z), zGN(zo). (9.1.14)

Now, if M G Win, then a = det [sd] defines a mapping of the algebra Wfln
onto the space C of complex numbers. The mapping function
a=f(d)

is called a functional — in the present case a multiplicative functional,

since
/(^) = f(d)fm>
Here also (9.1.14) is a functional equation to which the permanency
principle applies. Both sides of the equation may be continued analyti-
cally everywhere in D. It follows that, if both sides are continued
analytically along the same path L in D, the determinant of the continua-
tion of W(z) equals the continuation of the determinant, so that (9.1.14)
holds everywhere in D.
The case co(z0) = 0 remains. Here we are faced with two possibilities:
either a)(z) = 0, or else there are points in D where w(z)?£0, say,
o>(zi) # 0. If this is the case, we can replace z0 and N(z0) by z, and N(z,),
326 LINEAR nth ORDER EQUATIONS

respectively, in the preceding argument. It then follows that o>(z)^0

along all paths in D, (9.1.14) holds, and the assumption a)(z0) = 0, together
with w(zi)^0, is untenable.

COROLLARY

The solution matrix °W(z) is algebraically regular either everywhere in D or

else nowhere.

EXERCISE 9.1

1. Write the equation z2(# - a/)[# + (a + \)I]w = - I)w in more conven-

tional form.
2. Find a matrix DE equivalent to Legendre's equation.
3. Show that the solution °W(z)M is singular whenever M is.
4. Discuss any doubtful points in the proof of Theorem 9.1.2.
5. Show that any positive integral power of det (si) is a multiplicative functional
on aw..

9.2. ANALYTICITY OF MATRIX SOLUTIONS IN A STAR

We consider the matrix DE

<W'(z) = d(z)<W(z). (9.2.1)

Here d(z) is assumed holomorphic in the finite plane save for a finite
number m of poles. The point at infinity may be an essential singulary.
Let z = z0 be a point where si(z) is holomorphic, and let A(\s4; z0] be the
Mittag-Leffler star of d(z) with respect to z0. A point z = z, belongs to the
star if si(z) is holomorphic at all points z = az0 + (1 - a)zu 0 ^ a ^ 1. The
boundary of the star is composed of at most m rays joining the poles with
the point at infinity so that the prolongations of the rays intersect at z0. We
have now

THEOREM 9.2.1

Equation (9.2.1) with initial condition °W(zo) = % the unit matrix, has a
unique solution holomorphic in A[^;z0].

Proof. Let A0 be a star with respect to z = z0, which is a subset of

A[<s#; z0], bounded and bounded away from the boundary. Then there are
two numbers such that

\\d(z)\\ s£ M, |z - z0| *s R for z e A0. (9.2.2)

Here the matrix norm may be taken as
ANALYTICITY OF MATRIX SOLUTIONS IN A STAR 327

i k£= i \aik(z)l
\\d(z)\\ = max (9.2.3)

and similarly for other matrices in Wn. Then \\%\\ = 1. The matrix DE is
equivalent to the matrix integral equation of the Volterra type,

W(z) = % + P sl(tyW(t)dt, (9.2 .4)

and this may be solved by the method of successive approximations. Thus

we obtain a sequence {Wm(z)} of functions defined and holomorphic in
A[d; Zo]. For zGA0 the usual estimates give

\\Wm+l(z)-WM)\\^(-^f, (9.2.5)
which shows that the sequence converges absolutely and uniformly in A0
to a limit function °W(z). By the arbitrariness of the choice of A0, it follows
that W(z) exists in all of A[sd; z0] and is holomorphic in any star domain
like A,,. ■
We denote this solution by °lf(z;z0,^). Similarly, °lf (z ; z0, Wo) stands
for the solution of (9.2.1), which at z = z0 takes the value °WQ. By (9.1.11)
we have
°lf (z ; zo, W0) = W(z ; z0, %)°W». (9.2.6)
This formula and its generalizations play an important role in the theory.
Suppose that z = zx is a point in A[^; z0] and that W(zu zx, %)= ^i.The
matrix function z »-> d(z) has a Mittag-Leffler star with respect to zh
which is denoted by A[s£; z,].

THEOREM 9.2.2

If ze z0] D A[s£; zj, ffren

°lV(z; zo, %) = ^(z; z., %W(zu z0, <£). (9.2.7)

Proo/. The second factor on the right is simply °W and at z = z, both

sides take on the value hence the solutions are identical. ■
Actually the right-hand side of this formula gives the analytic continua-
tion of °W(z; Zo, in z,]. Moreover, there is no need to stop at this
point. Suppose that z = z2 lies in A[^; zj. We can then obtain the analytic
continuation of °W(z; z0,%) in A[,s4; z2] by forming

°lf(z ; z0, %) = W(z ; z2, %W(z2; z„ %)°W(Zl; z0, %), (9.2.8)

and this process can be repeated as often as we please. This is the basis
328 LINEAR nth ORDER EQUATIONS

for the method of analytic continuation, which will be discussed more

fully in the next section.
Before taking up these matters, let us look at the matrix DE
T(z) = Y(zmz), Y(z0) = «. (9.2.9)
Here Sfc(z) has the same analyticity properties as &(z). The method used
above can be applied also to the present case and proves the existence
and analyticity of Y(z) in A[2&; z0]. Here we have

Y(z ; z0, To) = T0r(z ; z0, «). (9.2.10)

The analogue of (9.2.8) now proceeds with left-hand multipliers.
Consider the equations

W(z) = d(zW(z), W(z0) = %

We have T(z) = -T(z)d(zl V(zo) = *. (' " }

Y(zW(z) = %, (9.2.12)
as shown by W. A. Coppel (private communication). This shows that T(z)
is the left inverse of °W (z) in A[s£; z0]. On the other hand, from °W(z0) = %
it follows that °W(z) has a two-sided inverse in some neighborhood of
z = z0. In this neighborhood

[W(z)V = T(z), Y(zW(z) = °W (z)T(z) = %. (9.2.13)

We now appeal to the law of permanency of functional equations, which
shows that the relation must hold in the common domain of analyticity of
y(z) and W(z), particularly in \[d\ z0].

EXERCISE 9.2
1. Carry through the proof of Theorem 9.2.1 in detail.
2. Fill in details in the discussion of (9.2.7) and (9.2.8).
3. Obtain the analogue of (9.2.8) for (9.2.9).
4. Prove (9.2.12) and (9.2.13).
5. 2. is a nilpotent element of Wl„ if some power of 2, is the zero matrix. Prove
that the singular DE TW'(z) = °W(z) has the zero matrix as its only solution.

9.3. ANALYTIC CONTINUATION AND THE GROUP OF

MONODROMY
In Section 9.2 we presented a method of analytic continuation. The
solution °W(z ; z0, %) of the equation
W (z) = si(z)°W(z) (9.3.1)
ANALYTIC CONTINUATION GROUP OF MONODROMY
329
can be extended consecutively in a sequence of Mittag-Leffler stars,
A[d;z0], A[^;z,], A[d; z2], A[d;z„], (9.3.2)
where each zJ+i is a point in the preceding star, \[si; z}]. There is no need
of considering analytic continuation along an arbitrary rectifiable oriented
arc C since such an arc may be approximated arbitrarily closely by a
polygonal line.
Suppose now that after n steps we find that z0 lies in A[s£; zn], and at
z = z0 the continuation takes on the value °lf 0. Here °W0 may be equal to %
but also it may very well be different. If this is the case, the analytic
continuation along the closed polygon IT: z0, z,, z2, . . . , z„, z0 multiplies
the solution matrix on the right by °W0:

<W(z ; z0, %) W(z ; z0, %W0. (9.3.3)

We see that with each closed polygon n with z0 as beginning and end, and
not going through any of the singular points au a2, . . . , am, there is a
unique associated transition matrix °lfn, and the continuation of
°lf(z;z0,^) along II carries it into the multiple °W (z ; z0, It is
necessary to assign an orientation to the path II. If II does not intersect
itself, then by the Jordan curve theorem II separates the plane into two
parts, an interior and an exterior. If z = c belongs to the interior, II has a
winding number of either + 1 or - 1 with respect to z = c. If the winding
number is + 1, then II will be said to have the positive orientation ; if — 1,
the negative orientation. The winding number is by definition (\I2tt) times
the change in the argument of z - c as z describes II. A path II described
in the opposite sense will be denoted by - II, and (9.3.3) is replaced by

W(z ; z0, %) — ^ °W(z ; z0, %)(WnTl. (9.3.4)

It should be noted that a transition matrix is necessarily algebraically
regular, since °W(z ; z0, ^) is regular and regularity is preserved under
analytic continuation.
So much for the case in which n is a simple closed polygon; if this is not
so, several possibilities arise. If II returns to z0 several times before
ending there, each subpolygon formed by consecutive passages of z
through z0 has to be treated separately. We can thus restrict ourselves to
polygons beginning and ending at z = z0 with no intermediary passages
through z0. Such a polygon may still intersect itself either at a vertex or
between two vertices. In any case there is a least subpolygon 111 with z0 as
beginning and end point. This will be a simple polygon for which an
orientation is given by the previous rules; this orientation is assigned to all
of II. In this manner all polygons with z0 as a vertex are oriented. They
form the set II(zo).
330 LINEAR nth ORDER EQUATIONS

We now consider the set X of all transition matrices Wn. We say that
two paths, Ilo and IT,, are topologically equivalent with respect to the DE
(9.3.1), the initial point z = z0, and the initial value % if all the members of
the family of closed polygons,

na=an0 + (l-a)n„ O^a^l, (9.3.5)

are admissible paths in the sense that none of them passes through one of
the singular points a, of the equation. If this is the case, the path n0 can be
deformed continuously into II, without striking a singular point. In the set
II(zo) we introduce the notion of an equivalence class : two paths belong to
the same class iff they are topologically equivalent in the sense defined
above. For each equivalence class we select in an arbitrary manner one of
its elements as the representative of the class. Let P(z0) be the set of
equivalence classes. To each equivalence class corresponds a unique
transition matrix, since all the paths of an equivalence class produce the
same transition matrix. The set of all transition matrices, one for each
equivalence class of paths, is a group of matrices under the operation of
matrix multiplication. This group © is the group of monodromy of (9.3.1).
It is justifiable to say the group of the equation, for, as we shall see later,
groups based on different initial points z0 are isomorphic.
At the present juncture it is perhaps more important to show that we
are really dealing with a group. Now the unit matrix is evidently an
element of the set. For every matrix °lf 0 in the set there is an inverse
matrix (W0)~\ for if °lf 0 corresponds to the path n0 its inverse corresponds
to - Ilo. Finally the product of two transition matrices, °W , and °lf2, is a
transition matrix, for if goes with the path II, and °W2 with the path n2,
then °W ,°lf 2 (in this order) goes with the path n2 followed by the path II,.
This shows that © is indeed a group.
The group is finitely generated ; i.e., there is a finite number of elements,
the generators, such that every element of the group is the product of
generators in some order. Factors are repeated in the product any (finite!)
number of times, but since matrices normally do not commute, it is not
permissible to combine equal matrices if they are separated by other
matrices. Thus

We may assume without loss of generality that arg (a, - z0) ^ arg (ak - z0),
for j ^ k, and that the singularities are numbered so that arg (a,- - z0) is an
increasing function of /. We now take a set of simple loops Lh [z0, a,-) in
our previous notation, starting at z = z0 and surrounding z = a, once in
the positive sense, leaving all other singularities on the outside. We may
take Lj as composed of a line segment described twice in opposite senses
ANALYTIC CONTINUATION GROUP OF MONODROMY 331

Figure 9.1

plus a small circle with center at ah See Figure 9.1. With each L, there is
associated a unique transition matrix We assume if /V k.
These matrices form the set of m generators ®u ©2, . . . , 06 m.
Suppose now that the path n belongs to Il(zo). There exists then an
equivalent path made up of linear combinations of the loops Lh We write
such a combination as

kxLh + -k2Lh + -... + -kpLip. (9.3.6)

Here the order is essential, as is indicated by the dot following the plus
sign. The /c's are integers, positive if the loop is to be described in the
positive sense and negative if described in the negative sense. The
formula is to be understood in the following manner. First the loop Lh is
to be described kx times, then the loop Lh described k2 times, and so on.
The subscripts /„ take values from the set 1, 2, . . . , m, and consecutive
loops are distinct but may be repeated later. To this representation of the
path n corresponds a representation of the transition matrix,

^n = (%)fc^(%_1)fc-'...(%IA (9.3.7)
This representation serves to establish the fact that the group is finitely
generated, since every transition matrix admits of a representation in
terms of the @'s. It is not claimed that (9.3.7) is minimal or unique, since
there may be relations between the generators. If the generators are of
332 LINEAR nth ORDER EQUATIONS

finite order, a power of a generator equals the unit matrix; more generally,
some product of generators equals the unit matrix.
Finally we have to settle the dependence of the group on the initial
point z0. Suppose that z0 is replaced by some other point zu nonsingular
for the equation. Then we can proceed as above and get a group @(z,)
which is normally different from ®(z0). But there exists a regular matrix
M such that
®(zi) = M-l®(z0)M, (9.3.8)
so the two groups are isomorphic and hence abstractly the same. To see
this we join z0 and z, by a rectifiable arc C, a straight line segment if
possible. Describing this arc from zx to z0 carries

°W(z\zu%) into °W(z',Zo^)M

for some regular matrix M, which may depend on the arc. Now let z
describe the loop L,. We return to z0 with the new determination
W(z;Zo^)%M,
and then return to zx along C. We return with the determination

W(z;zly%)M-l%M.
Now C +• Lj +• (- C) is a loop Lf with respect to z = zx. This shows that
the transition matrices

M-^JL, / = l,2,...,m,
form a set of generators for the group ®(z,). Thus the group ©(zO is the
transform of the group ®(z0) under the transformation M. The two groups
are isomorphic, since for all j and k

M~l%®kM = M^iMM'®, M. (9.3.9)

Thus the underlying abstract group is unique.

EXERCISE 9.3
1. Consider the equation in %Jl2:

W(z) = d(z-\y1W(z) with^=[^ q].

Show that

\(% +j^)(z - \y -\si){z - ir-

is a fundamental solution.
2. Show that the group of the equation in Problem 1 has a single generator
© = exp (27n ,s4), where the exponential function is defined by the exponential
series. Sum the series. If a is an integer, the solution is single valued and the
APPROACH TO A SINGULARITY 333

group reduces to the identity %. If a is not an integer, the group is cyclic and is
finite iff a is rational.
-u
3. In Problem 1 replace si by the nilpotent amatrix

Find a solution matrix and the group.

4. Suppose that (9.3.1) has the singular points 0, 1, °°, and no others. Take z0 = \
and lay loops L0 and L, around 0 and 1. Let the corresponding transition
matrices be ©0 and Find a path corresponding to the transition matrix
©^^©F'OJo. Draw a figure.

9.4. APPROACH TO A SINGULARITY

For the study of this problem it is desirable to allow the matrix equation to
have somewhat more general coefficients. In the DE

W'(z) = d(zW(z) (9.4.1)

the entries ajk(z) shall be analytic but not necessarily single valued. The
singularities shall be isolated, however, and we place one at the origin. Let
a solution °W(z) of (9.4.1) be determined by its initial value at z = r0, and
let the interval (0, r0) be free of singularities. We intend to approach the
origin — to start with, along the positive real axis. It turns out that the
behavior of |fW(r)|| for r j 0 is essentially determined by the rate of
growth and the integrability properties of ||«s£(r)|| for r j 0. The basic result
is

THEOREM 9.4.1
For r | 0 we have

II^MI exp {- jrF°H(t)|| dt) ^ |plf(r)|| ^ ||W(r0)|| exp \\d(t)\\ dt}. (9.4.2)
Proof. LetO<a<b<rQ, and note that

<W(b)-'W(a)= JCa d(t)°W(t) dt, (9.4.3)

whence

Ja
inf(fo)iisinf(a)ii+ [V(/)iiiiin/)M(,

and by Gronwall's lemma (Theorem 1.6.5) this implies that

\\°W(b))\ « |rMf(a)|| exp {|V(Oll<fc}.

334 LINEAR nth ORDER EQUATIONS

In particular, for a = r, b = r0 we get

||1f(r„)||exp {- f'°\\d(t)\\dt] ^pr(r)l

which is the lower bound in (9.4.2).
To get the upper bound we make a change of variables. We set

r = r0-x, °lf(r) = d{r) = ®(x)

to obtain
W(x) = -<&(xmx) (9.4.4)
and

q^(jc) = ^(0) - JoP <&(sms) ds. (9.4.5)

This implies that

Jo
||qy(x)H ||^(0)||+ fxp(5)||||^(5)|| ds,
and again by Gronwall's lemma

exp{Jj&(s)||</s}.
Changing back to the old variables, we obtain the upper bound in
(9.4.2). ■

COROLLARY
If as riO

irW(r0)|| exp {p ||4(t)|| dt} 0, (9.4.6)

then °W(r) = 0.
Similar results hold if z tends to zero in a sector

5: a <argz < p, 0<|z|^#,

in which s&{z) is holomorphic. We can apply Theorem 9.4.1 to each ray,
arg z = 0, in S. This gives

pf(R ew)\\ exp {- j* \\d(t e'°)\\ dt] s ||Hf(r e")\\

=s (R e")\\ exp {|* \\d(t e")\\ dt}. (9.4.7)

Similar relations hold for the approach to infinity in a sector or in a
strip. One can usually find a compact subset of (a, /3) on which \fW(R e")\\
APPROACH TO A SINGULARITY 335

is bounded away from zero and infinity, and on such a subset it may be
possible to find majorants and minorants for \\d(r el6)\l independent of 0,
and thus obtain simpler estimates of \\°W(r eie)\\.
Before discussing several such cases, let us consider briefly approach
to the point at infinity in a strip. This becomes significant when si(z) is an
entire periodic function of z so that the solution °W(z) is an entire function
of z of infinite order. Without loss of generality we may assume that the
period equals one and study the bounds of \\°W(x + iy)|| for O^x 1 as
y -» + oo. If si(z) is of finite order, a reasonable assumption on the norm of
si(z) is
aeky ^\\sl(x + iy)\\*zbeky (9.4.8)
for some k 2^0 and some positive numbers a and b and for O^jc ^ 1,
0 ^ y. We take an initial matrix °W(x0) for an x0 in [0,1]. The general result
is given by

THEOREM 9.4.2

If si(z) is holomorphic in the closed half-strip 0^x^1,0^ y, and °lVXz)

is a solution of (9.4.1), then

\\W00\\ exp j- £ \\si(x + it)||dt] |fW(x + iy)||

irW(x)|| exp {£ \\si(x + it)|| dt J. (9.4.9)

The proof is left to the reader. It follows, mutatis mutandis, the
argument used above for Theorem 9.4.1.
Suppose, in particular, that the norm of s&(z) is bounded by (9.4.8) and
that

min 11^0011 = A, m<xx\\°W(x)\\ = B forO^x < 1.

Then f or k > 0

A exp [~^(eky - 1)] ^\\W(x +/y)||^B exp \j{eky - l)j, (9.4.10)

while for k = 0 we have

Ae~hy ^\\W(x +iy)\\**Beby. (9.4.11)

Let us return to the case in which the singular point is at the origin. The
preceding theorems show that the case in which the norm of d(z) is
integrable down to the origin plays an exceptional role. We proceed to an
analysis of this case. The basic result is
336 LINEAR nth ORDER EQUATIONS

THEOREM 9.4.3

// \\sl(t)\\ e L(0, R), then the equation

°W'(t) = d(t)°MT(t), 0<t^R, (9.4.12)

with the boundary condition <W(t)^% as 1 1 0, has a unique solution.
Proof. Apply the method of successive approximations to the integral
equation

°W(t) = % + j si(sW(s)ds (9.4.13)

with

°W0(t) = %, °lfn(r) = <g + J* slisWn-^ds.

Here the integrability condition implies the existence of °Wi(t). The
existence of the higher approximations and the convergence proof are left
to the reader. ■

COROLLARY

Under the assumption that \\si(t)\\ E L(0, R), all solutions of (9.4.12) tend
to a finite limit as 1 1 0, and this limit is the zero matrix iff W(t) = 0.

Proof. Consider the solution °W(t; R, %). It is a fundamental solution,

and so is the solution °W(t;0,%) constructed above. It follows that each
one is a multiple of the other, giving

°W(t ; R, %) = Wit; 0, %)[°W(R; 0, %)Y\

since this is true for t = R. Hence

lim °W(t; R, %) = [°W{R ; 0, (9.4. 14)

Since °W(t\ R, °W0) = 'Wit ; R, %)°W0, the existence of a limit holds for all
solutions. ■
The extension to the complex plane simplifies if one assumes the
existence of a uniform dominant for the integral. This is what is done in

THEOREM 9.4.4

Suppose that sl(z) is holomorphic in the sector S defined above and there
exists a function r»-^.K(r) belonging to C(0, R] H L(0, R) such that

||^(reie)||^K(r),|0|<a. (9.4.15)
Then (9.4.1) has a unique solution <W(z;0,<g) such that
APPROACH TO A SINGULARITY 337

UmW(reie;0,%)
riO = % (9.4.16)

uniformly in S. Moreover, every solution of (9.4.1) tends to a finite limit as

r 4 0, and the limit is independent of 6 and exists uniformly with respect to
e.
The proof is left to the reader. It follows the pattern of previous
arguments. The uniform convergence follows from the estimate

|pMr(z; 0, %) - %\\** exp {£" K(r) drj - 1. (9.4. 17)

As the last item under the integrable case we list

THEOREM 9.4.5

// ||^(t)|| G L(R, oo), then every solution of (9.4.12) tends to a finite limit as
1 1 oo and the limit is the zero matrix iff °W(t) = 0.
The essential step in the proof is to show that the integral equation

°W(t) = %-jt sl(s)°W(s)ds (9.4.18)

has a unique solution W(t ; oo, <£), which tends to % as t \ oo. This solution
is fundamental, and every solution of (9.4.12) is a multiple thereof, the
multiplier figuring on the right and being an element of Wln.
Suppose now that ||^(z)|| is not integrable down to zero. The
possibilities are legion, but outstanding cases correspond to &(z) having a
pole at the origin. Here there is a great difference between a simple pole
and a multiple one. For a simple pole we have.

THEOREM 9.4.6

Let 2ft(z) = zs£(z) be holomorphic in a disk |z| < R, and let 2ft(0) ^ 0, so
that si(z) has a simple pole at z = 0 with residue 2ft(0). For \arg z| < a we
can find positive numbers b(a), c(a), and M such that

b(a)\z\M ^ \\W(z)\\ ^ c(a)|z|-M. (9.4.19)

Proof. This follows from (9.4.7). For R sufficiently small and |z| = r ^ R
we have

M(z)H!p(0)||i,

so we can take M = lp(0)||, b(a) = R M inf ||°MT(r eie)\\, and c(a) =

RM sup\\°W(reie)\\ for 0 restricted to the interval (-a, a). ■
338 LINEAR nth ORDER EQUATIONS

An estimate of this type was first given in 1913 by G. D. Birkhoff

(1884-1944), who was inspired by ideas of Alexander Mikailovic
Liapoonov (1857-1919). Similar estimates based on product integration
were presented by L. Schlesinger at about the same time.
We have similar results at infinity.

THEOREM 9.4.7

If Sft(z) = zsi(z) is holomorphic at z = o° and if S&(°°) # 0, then z = °° is a

regular-singular point of (9.4.1) and the solutions satisfy estimates of the
form
b(a )|z|"M ^ \\W (z)|| ^ c(a )|z|M, (9.4.20)
where M ^ ||Sft(°°)|| and \arg z| < a.
We shall return to regular-singular points in the next two sections. But
first we have to consider growth properties at an irregular-singular point,
which it is convenient to place at infinity. It is also convenient to consider
the situation in a sector and to allow a more general rate of growth of
||^(re'e)|| so as to permit fractional powers of r. The basic result follows
immediately from Theorems 9.4.1 and 9.4.2 by a suitable modification of
the argument.

THEOREM 9.4.8

// si(z) is holomorphic in a truncated sector S\arg z| < a, R < \z\ < °o, if

\\sd(ree)\\^Mr^\ 0</ll, (9.4.21)

and if °W(z) is any solution of (9.4.1) in S, then there are positive numbers
b(a) and c(a) such that

b(a) exp [~^) ^ irW(z)ll ^ c(a) exp /). (9.4.22)

The proof is left to the reader.
We now consider the H. von Koch-O. Perron theorem, an extension of
Theorem 5.4.1. We start with a lemma needed for the proof.

LEMMA 9.4.1

Let mi, m2, . . . , mn be n given real numbers, and set

l«j«n j
m = mox-. (9.4.23)

Then for any choice of n real numbers a,, a2, . . . , an, the largest of the
2n — 1 numbers,
APPROACH TO A SINGULARITY 339

mi, m2 - (an - an-i), m3 - (an - an-2), (9.4.24)

mn-(an-a,),

is not less than m and there is at least one choice of the a's for which the
maximum equals m.
Proof. Let M be the infimum of the maximum of (9.4.24) for any choice
of the a's. It is clear that M 2* mi. Now, if j 2*2, the larger of the two
numbers
m, - (an - an-s+\) and a„-J+2 - a^-j+i
is at least equal to
^[m, - (an - an-j+2)],
for if both were less than this number, adding the resulting inequalities
would give 0 < 0. Next, if j s= 3, the larger of this number and
a„-j+3 — a„-,+2 at least equals
|[m, -(a„ - a„-i+3)].
Again, if both were less than this bound, the first number plus half the
second would give 0 < 0, and so on. In this manner we show that M 2» m, /j
for j = 1, 2, . . . , n.
Suppose now that

and that, if there are several possible values for /c, we take the least one.
Next, we choose

Thus fl,=6'-l)m, j = 1,2 ft. (9.4.25)

m, - (an - an-i+i) = m, - (j - l)m ^ m,

i.e., a choice for which M = m. ■
This is now to be applied to the nth order linear DE

w(n)(z) + 2 PMw^Xz) = 0, (9.4.26)

where the Pj's are holomorphic for R < |z| < °° and have at most a pole at
infinity. The order of P,(z) at infinity is to be defined. If P;(z) = 0, we set
the order m, = — oo; otherwise the order m, is defined so that
PKz^z^U), (9.4.27)
where Pi(z) is holomorphic at infinity and different from zero there.
340 LINEAR nth ORDER EQUATIONS

Setting

g = 1 + max yj, (9.4.28)

we have now

THEOREM 9.4.9

Every solution of (9.4.26) satisfies

|w(z)| ^ B exp (K|z|8) (9.4.29)

for |z| large and suitably chosen constants B and K.
Proof. We change the nth order DE (9.4.26) into a vector matrix DE by
setting
w. = z-o-»)»wo-« j = 1, 2, . . . , n, (9.4.30)
where m is defined by (9.4.23). If w(z) is a column vector with the u>,'s as
entries, then w(z) satisfies the equation

w'(z) = ^(z)w(z), (9.4.31)

where the matrix si(z) has the elements

aji(z) = -(j -1) j, a,-,i+i(z) = zm, j = 1,2, . . . , n - 1,

a*(z) = -z"(" i)mP„_j+I(z), j = 1,2, . . . , n - 1,(9.4.32)

anAz) = ~(n-\)j-Px(z\
all other elements being zero. We may assume that m > — 1.
The order at infinity of the elements of si(z) in the first n - 1 rows is at
most m, and of the elements above the main diagonal exactly m. The
order of fl„j(z) is
m„-i+i - (n - j)m ^ m
by lemma 9.4.1. This holds also for j = n. Hence we can find M and R so
that a,k(z) is holomorphic for R < |z| = r < <», |arg z| < 77, and

Mz)|^Mrm, (9.4.33)
so that
\\sl(z)\\^nMrm. (9.4.34)
Now let z0 be chosen with R < |z0|, let the initial value w(z0) be given,
and integrate radially from z0 to z to obtain

w(z ) = w(z0) + JZQ si(s)w(s)ds

APPROACH TO A SINGULARITY 341

so that

JZQ
||W(2)H l|w(Zo)||+ fS ||^(5)|| ||w(5)|||ds|

« ||w(zo)|| + nM f' |s |m i|w(s)|| ld5 1.

In passing, let us observe that it is immaterial here what norm we use for
the vectors: U, l2, or L. By Gronwall's lemma the inequality gives

||w(z)|| « ||w(z„)|| exp [\zr< - |2„r+l]}, (9.4.35)

and this implies (9.4.29) if we take

YYl t 1
K =_nM B -max||w(roeio)||exp(-Xrom+1). ■
Lemma 9.4.1 implies that the choice of multipliers in (9.4.30) is the best
possible one in the sense of producing the lowest estimate for ||w(z)||. It
should be noted that the estimate for the norm of the solution vector is
also an upper bound for |u>i(z)| and that w,(z) = w(z) is a solution of
(9.4.26).
We call g the grade of the irregular-singular point of (9.4.26) at infinity,
as well as of the essential singularity of the entire functions which satisfy
the equation. It is evidently a positive rational number p Ik with 1 ^ k ^ n,
and simple examples show that the minimal value 1/n can be taken on.
There is no difficulty in constructing a DE of given order and polynomial
coefficients and a preassigned grade g. There is, however, a Riemann-
Birkhoff problem which does not seem perfectly trivial: given the order of
the equation and the degrees of the polynomial coefficients so that the
number g is also given, find an equation for which the grade of the
solutions is exactly g.

EXERCISE 9.4
Fill in missing details or give proofs of the following:
1. Corollary of Theorem 9.4.1.
2. Theorem 9.4.2, including formulas (9.4.10) and (9.4.11).
3. Theorem 9.4.3.
4. Theorem 9.4.4, particularly (9.4.17).
5. Theorem 9.4.5, including (9.4.18).
6. Theorem 9.4.6.
7. Theorem 9.4.7.
342 LINEAR nth ORDER EQUATIONS

8. Theorem 9.4.8.
9. Lemma 9.4.1.
10. Theorem 9.4.9.
11. Transform the DE w" + (1 - z2)w = 0 into vector-matrix form. Find the
estimate of |fW(z)||, and determine g and K. The DE has a particular solution
which is an exponential function. Find it, and compare the estimate with the
actual rate of growth. How good is your value of Kl
12. Given the family of DE's
w"' + Azw" + Bz2w' + Cz3w =0,
Show that the solutions are of grade 2. If a is given, a # 0, show that there
are values of A, B, C for which eazl is a solution. Here A is uniquely
determined, while B and C have to satisfy a linear equation.

9.5. REGULAR-SINGULAR POINTS

A point z = a is a regular singularity of the matrix DE

W(z) = si(zW(z) (9.5.1)

if (z — a)s&{z) is a holomorphic matrix function at z = a:

(9.5.2)
(z-a)^(z)=j^sik(z-a)\
k =0 -singular
where s£0 is not the zero matrix. The point at infinity is a regular
point of (9.5.1) if zsi(z) is holomorphic at infinity and

k=0
zd(z)=2&kz-\ (9.5.3)
where may be the zero matrix.
These conventions are of course of much more recent date than the
original definitions for linear nth order DE's, which go back to Lazarus
Fuchs (1866) and G. Frobenius (1873). Suppose that

w(nXz) + P,(z)vv(" ^z) + - • - + P„(z)vv(z) = 0, (9.5.4)

where the coefficients are meromorphic in some neighborhood of z = a
and at least one has a pole. We can put (9.5.4) into matrix form by setting

w>i = w, vv2 = (z - a)u>\ vv3 = (z — a)2w", (9 5 5)

The resulting matrix si(z) = (ajk(z)) then has the entries:

OhU) = 0 - 0(2 - aY\ au+i(z) = (z- a) ', j = 1, 2, . . . , n - 1;

anAz) = -(z-a)nkPn-k+l(z), k = 1,2,..., n-1; (9.5.6)
ann(z) = (n - \)(z - a)1 - Pt(z);
RECTANGULAR-SINGULAR POINTS 343

all other entries being zero. This matrix will have a simple pole at z = a iff

(z-a)kPk(z) (9.5.7)
is holomorphic at z = a, for k = 1, 2, . . . , n. Hence this condition is
necessary in order that z = a be a regular-singular point. This does not
imply that the point is a function-theoretical singularity of any of the
solutions. They may all be holomorphic there, and this implies that the
corresponding solution matrix is a holomorphic function of z. The point
z = a is, however, an algebraic singularity: the matrix °W(a) turns out to
be singular. For (9.5.4) it is customary to speak of an inessential
singularity. We shall see that the Wronskian vanishes at z = a.
We have similar results at infinity. Suppose that the coefficients of
(9.5.4) are holomorphic for large values of \z \ except possibly for poles at
z = oo. We can reduce to the matrix case via the substitution (9.5.5)
provided that we take a = 0. The matrix s£(z) is the same at infinity as at
zero. The matrix zsi(z) will be holomorphic at infinity iff each of the
functions
zkPk{z) (9.5.8)
is holomorphic there. The matrix is clearly not the zero matrix (why?).
The singularity may be inessential in the sense that all solutions are
holomorphic at z = oo. We have proved ■

THEOREM 9.5.1

The point z = a is a regular -singular point, possibly inessential, if

(z - a)kPk(z) is holomorphic at z = a for each k. The point at infinity is a
regular -singular point, possibly inessential, if zkPk(z) is holomorphic at oo
for each k.
We shall make a study of the solutions in a neighborhood of a
regular-singular point, which, for convenience, we place at the origin. It is
also convenient to rewrite the equation in the form

zHwiH\z) + 2 QU^-V^z) = 0, (9.5.9)

where i= i

Qi(z) = ziP,(z). (9.5.10)

By assumption the coefficients Qj(z) are holomorphic at z = 0.
If every Q, (z) is a constant ch the DE becomes an Euler equation with
solutions of the form zr, where r is a root of the algebraic equation
n
r(r - \)(r - 2) • • • (r - n + 1) + 2 c,r(r - 1) • • • (r - n + j + 1) = 0. (9.5. 11)
344 LINEAR nth ORDER EQUATIONS

If this equation has multiple roots, say r = r0 is a Jc-tuple root, an aggregate

of the form

zr°[Co+C,logz + • • - + Ckl(\ogz)kl] (9.5.12)

appears in the solution. These facts were of course well known to Fuchs
and could serve as a pattern. Equation (9.5.9) is a perturbed Euler
equation, and perhaps the solution could be obtained by a similar
perturbation of the solution of the Euler equation. Be that as it may,
Fuchs substituted a series

w(z)= J) akzr+k (9.5.13)

k =0
in the equation and proceeded to a determination of the coefficients. Let us
denote the left member of (9.5. 1 1) by f(r). Fuchs found that for r in (9.5. 13)
one must take a root of the equation

/(r) = 0, (9.5.14)
which he referred to as determinierende Fundamentalgleichung and which
in English is known as indicial equation. He found that, if the roots of this
equation are distinct and do not differ by integers, it is possible to
determine n linearly independent convergent series solutions. The case of
integral differences of the roots and, in particular, of equal roots was
settled by Frobenius 7 years later. A number of arguments are available in
the literature all of them more or less "corny" (slang or Milton). What I
shall give here is not the corniest; it is adapted to the matrix case and is
closely related to arguments given by J. A. Lappo-Danilevskij (1896-1931)
for the matrix case and by the present writer for DE's in Banach algebras.
It involves the notion of the commutator operator in the matrix algebra
Tln, which is next on our agenda.
Let si be an arbitrary matrix in Wn, distinct from the zero and the unit
matrices. We define three operators on Tln into itself; left-hand multipli-
cation bysi, right-hand multiplication by si, and the commutator, which
is the difference between the two. Thus

L*% = sl%, R*% = %sl9 C*% = L*%-R*% = (9.5.15)

These are obviously bounded operators, and

||M = W = IMI, ||CJ|*2N|. (9.5.16)

Next, we have to consider the spectra of these operators and of si. The
spectrum of si, written as ar(si), is the set of numbers \j which are roots
of the equation
det[A^-s£] = 0 (9.5.17)
 345
RECTANGULAR-SINGULAR POINTS

or, equivalently, the values of A for which the matrix k% - si is singular.

The spectrum of the operator is the set of values A for which the
matrix equation
\% = (9.5.18)
cannot be solved with respect to % for all A similar definition holds for
o-(IU):

LEMMA 9.5.1
a(L3l) = a(Rsl) = a(si).
Proof. If A G a (si), then X% - si is a singular matrix, so (9.5.18) can be
solvable for % only when ^ belongs to a linear subspace of singular
elements. This shows that A G cr(L^) and, by the same token, A G cr(Rrf).
On the other hand, if A G a (Lai), then \ % - si cannot have an inverse, for
if it did one would have (% = {\si-%) ^ for all ^, which is absurd. It
follows that A G a(si), and the lemma is proved. ■
The spectrum of calls for more elaborate considerations. Here the
buffer condition of N. Jacobson is basic. In his terminology Wln is a prime
ring. What we need is

LEMMA 9.5.2

If 3£ and °U are nonzero elements of 3ftn, then there exist elements 3£ such
that VtWty is not the zero element.

Proof. This is trivial if ffity is not the zero element, for then any multiple
of % will do as a buffer. If OTJ = 0, then % and as divisors of zero are
both singular. But we can find an entry xjm in the matrix 8£ and an entry ypk
in which are not zero. Hence the matrix 3£, which has a one in the entry
(m, p ), will do, for then 2I£2E<& has the entry xjmypk ^ 0 in the entry (j, k ). ■
LEMMA 9.5.3

Let a and /3 belong to a(sl), and let % and be nonzero matrices such
that si% = a%, °Usi = pa!); then either M = 0 or a - p G o-(C^) and
= (a -
Proof. This follows immediately from
= = (a ■ (9.5.19)
Since
cj&m) = si^T*) - %z<tosi = (a -py%&&,
it follows from the buffer condition that the spectrum of is the
difference set of the spectrum of si:
346 LINEAR nth ORDER EQUATIONS

LEMMA 9.5.4

cr(C^) = cr(si) — cr(si), i.e., the set of n2 differences of the spectral values of
si.

Here each difference is repeated as often as it occurs. In particular,

A = 0 is at least an n-fold spectral value of the commutator, and if A is a
spectral value so is — A. Various properties of the characteristic values
and elements are found in Exercise 9.5.
We shall need further analytic matrix theory for our considerations.
They center around the resolvent of si, i.e., the inverse of k% - si, which
is denoted by 2ft(A;s£). It is a rational matrix function with poles at the
roots of (9.5.17). These numbers have several names: latent roots,
characteristic values, and eigenvalues are the most common. The equa-
tion is known as the characteristic equation. The Cayley-Hamilton
theorem asserts that si satisfies its own characteristic equation. Actually
it may satisfy an equation of lower degree, known as the minimal
equation. The distinct roots of the characteristic equation satisfy the
minimal equation, but multiple roots may occur with a lower multiplicity
in the minimal equation. Suppose that A = a is a characteristic value
which is a root of multiplicity fx of the minimal equation. Then at A = a
the principal part of the resolvent is

Op G) G) ^ _1
n - a)
A — a + (A
IT— (A — a)w
V + ' * * + ti (9-5'20)

Here is an idempotent, also known as a projection, i.e., = <3>, & is a

nilpotent such that ^ 0, and ^ = 0. The resolvent vanishes at A = <*>
and equals the sum of its principal parts:

P r Ob. G). GiVT1 "I

;-) = 2 (rt? + • • • + (a^t ]■ (9-5-21)
The number of projections equals the number of distinct roots. Here

+ 0>2 + • • • + 9>p = %, ®i®k = 8ik®h

known as a resolution of the identity. Furthermore,

si = A,2?>1 + A2g>2 + - • - + ap^p +a, + a2 + - • - + ap.

Only multiple roots A, of the minimal equation contribute to the 2,-sum. It
should be noted that

&} % = % = 8ik %, % %=0 if jV k. (9.5.22)

With the aid of these formulas an exponential function of zsl becomes
RECTANGULAR-SINGULAR POINTS 347

exp(z^) = 2 [», +%z +%z2+ • • ■ + ^ I z^'] eV. (9.5.23)

Replacing z by log w, we obtain an expression for the power u*:

u*= 2 + a' log w + ^ (log u )2 + • • • + ^ j},(log u )t 1j

u\
(9.5.24)
We have also to consider the resolvent of C^, 2ft(A ; C*). This is an
operator-valued rational function of A with poles at the points of the
difference set of a(si), i.e., the points a, - fik where ay and fik E cr(si).
That these points are actually poles of the resolvent follows, for instance,
from results due to Yu. L. Daletzky (1953) and S. R. Foguel (1957). Since
CJC = si%-%siG Wn, when £ does, 9fc(A ; C*)% G Wln whenever it exists,
i.e., except for A in the difference set of a (si). This makes it rather
plausible that Sft(A ; C*) has no other singularities than poles, and since it
vanishes at infinity it is a rational operator-valued function of A. Daletzky
gave an explicit representation of Sft(A ; from which this can be
proved rigorously. It is also evident from the results of Foguel, from
which one can read off that the rational character of Sft(A ; si) implies the
same property of ; C^). Moreover, Foguel proved that, if A0 is a pole
of 2&(A ; CI*), its order equals max (ray + nk — 1) if A0 = a, — j3k and ay is a
pole of 2ft(A ; si) of order m} while (3k is a pole of order nk ; the "max"
refers to the fact that A0 may be expressible in several ways as a difference
of spectral values of si. See Exercise 9.5 for some further details. For the
following discussion we need only the fact that the spectral singularities
of 2ft(A ; Csi) are poles.
We return now to (9.5.1) with a regular-singular point at the origin. By
analogy with (9.5.9), where we attempted the series solution (9.5.13), we
shall here attempt a matrix series solution of the form

°W(z)= 2 <4kzk+*. (9.5.25)

Here the coefficients %k and si (all elements of Wln) are to be determined.

The power zA is defined by (9.5.24), and the exponent k + si is to be read
as k% + si whenever needed. We have zsi(z) = SS(z) represented by
(9.5.3), where 2fc0 * 0. Substitution of (9.5.25) in the DE gives

2 %m(m% + si)zm+sA=fi
m=0 ®kzk m=0
fc=0 2 <gTOzm+J*

m=0 \fc=0 /
348 LINEAR nth ORDER EQUATIONS

assuming convergence in norm of the series involved. This must of course

be justified a posteriori.
The term of lowest order on the left is ^QsizA\ on the right, ^^z*. The
convenient choice of ^0 is = though later on we have to replace this
by a scalar multiple of %. This shows that

The equations for the determination of the coefficients then become

=
<«.». si= a.<«
o, (95 26)
m%m -WMm -%A)= 2 98™-!, ^n,
k=0 m = 1,2,3

Here the first equation is satisfied by taking = %, as already mentioned.

In trying to solve the second set of equations for %m, we are confronted by
the spectral properties of the operator C»0. If the spectrum contains no
positive integers, all is plain sailing and the coefficients %m may be
obtained successively by applying the operators 2ft(ra;Cao). If, on the
other hand, there are positive integers in o-(C»0), such a procedure breaks
down. Now there will be positive integers in the spectrum of Cao iff there
are spectral values of which differ by an integer. This is the matrix
version of the difficulty which beset Fuchs in his discussion of the linear
nth order DE. There it was the roots of the indicial equation which, when
they differed by integers, caused trouble. It is advisable to separate the
two cases and give first the analogue of Fuchs's result.
THEOREM 9.5.1

Suppose that (9.5.1) has a regular- singular point at z = 0, where si{z) has
a simple pole with residue 2S0. Suppose that no two spectral values of S&0
differ by a positive integer. Then the equation has a unique solution of form
(9.5.25) with %0 = %, and the series converges in norm in the largest
punctured disk, 0 < |z| < R, in which sd(z) is holomorphic. The solution is
fundamental for 0 < |z| < R.
Proof. We have already seen that the coefficients %m are uniquely
determined, so all that is to be done is to prove convergence of the series.
This can be done by the usual device of a majorant series. Set

B(z)=|>fcz\ bk=\\®k\\. (9.5.27)

By (9.5.16) we have
(9.5.28)
||Cj^2p0|| = 2fr0;
RECTANGULAR-SINGULAR POINTS 349

and since
m -1

we get
m —1
(m -2Mft-.il* 2 fc.-*lft»ll.

which is nontrivial for m > 2b0. Let r be chosen so that 0 < r < R. Then

/c=0
*=[£(r)-&0] 0=s/cmax
«m -1 (K||rk).
We now choose N > B(r) + bQ, and see that for m > N

K, \\rm ^ max fe*||rk) s M(r).

This shows that for all k

\\%k\\rk ^M(r), (9.5.29)

or the terms of the formal solution series are bounded in norm for \z\ = r.
Since r is arbitrary, 0 < r < R, a well-known theorem due to Weierstrass
shows that the series converges in norm for 0^ \z \ < R and the formal
solution is an actual solution. This proves the theorem. ■
We have now to consider the case in which spectral values of differ
by an integer. We treat this by a suitable modification of the method of
Frobenius. We shall prove

THEOREM 9.5.2

Suppose that the spectrum of C<a0 contains the positive integers

n,, n2, . . . , nk and no others, where

1 n, < n2 < • • • < nk < 2b0,

to which correspond poles of Sft(A; C^) of orders m,, m2, . . . , mk, respec-
tively, and set

m, + m2 + • • • + mk = p. (9.5.30)
Then there are p + 1 power series such that

j2dogz)j
=0 m5)= 0 %mizm+^ (9.5.31)
is a solution of (9.5.1). The series converge in norm for 0 < |z| < R, and the
solution is fundamental in the punctured disk.
350 LINEAR nth ORDER EQUATIONS

Proof. There exists a positive 8 such that each disk |A-n|<Sforn>0

contains at most one spectral singularity of Cgs0. For n = nu n2, . . . , nk
there is a pole of R(\;C^0) at the center of the disk, and for all other
values of n there is no point of the spectrum in the disk.
Now take a complex number a with 0 < |cr| < 8, and consider the series

W(z, a) = mJ)
=0 %m(*)z^+m+°. (9.5.32)
Since a 0, this cannot be a solution of the matrix DE, but we may
choose the coefficients %m{a) so that °lf(z, a) satisfies a nonhomogeneous
linear equation such that the perturbation term is small in norm when \a\
is small. To this end, let us form

[$ -®(z)]W(z,a), d =z^'
and substitute the relevant series, multiply out, and rearrange according
to ascending powers of z. This gives

m=0
2 L /=0
hm(<r)(®o+m +*)-f,® J
m-i%(*)]z^+m+° .
For m = 0 the coefficient of the power is
+ ^0(0-)^0 - ^O^O(O-).
We choose
^a) = a"% (9.5.33)
with p given by (9.5.30). For m >0 we choose ^>m(cr) so that the
coefficient of z*°+m+<r is zero. This calls for
m 1
(m + a)%m{o-) - Wm (o-) + ^m (o-)^o = 2 am_,-%(o-). (9.5.34)

Since the resolvent exists for 0 < \a\ < 8, we get

<<Uo-) = &(m + o-; Cao)[2 9L-,-%(<r)], m = 1, 2, 3, . . . . (9.5.35)

These coefficients can be determined successively, and if the resulting
series converges it satisfies the equation

dZ
z^-°W(z, a) = mzW(z, a) + crp+,z^+f\ (9.5.36)
The time has come to examine the analytical nature of the coefficients
%m(cr) as functions of cr. For 1 < nx we observe that 2ft(l + cr; C^o) is a
holomorphic function of cr in |cr| < 8 so that
RECTANGULAR-SINGULAR POINTS 351

%(a) = &(1 + a; C<*)[®i<rp] = aP(3i{\ + or; Cao)[&,]

is holomorphic and has a zero of order p at o- = 0. The same result holds
for m = 2, 3, . . . , nx — 1. Something new happens for m = nu where we
have „ _,

Here the operand has a zero at a = 0 at least of order p, but the resolvent
operator has a pole of order mx. This shows that ^nX^) is holomorphic in
the disk and has a zero of order at least p - mx at the origin. For
«i ^ m ^ n2 — 1, the resolvent is holomorphic and operates on a
holomorphic function of a which has a zero of order ^p - mx at the
origin. Thus these coefficients are holomorphic for a in the disk and have
a zero at the origin of order at least p - mx. At m = n2 the resolvent has a
pole of order ra2 at the origin, and it operates on a matrix-valued function,
holomorphic in a and with a zero of order p - mx at a = 0. The result is a
coefficient ^(o") holomorphic in the disk and with a zero of order
-ra1-ra2. In this manner we proceed, until finally for m = nk the
orders of poles and zeros just balance and for m ^ nk the coefficients
^>m(a) normally do not vanish at the origin.
It is desired to show that the series

(9.5.37)
2 ^m((r)z
converges for \z \ < R, \a \ < 6, uniformly on compact subsets. To this end
it suffices to show that the sequence of terms is bounded for \z\ = r <R,
uniformly with respect to a for |<r| ^ e < 8. This can be proved by the
method used to prove (9.5.29). For |cr| ^ e we have

k<m [^.(o-)^"],
(m - e - 2b0)\HU<r)\\rm ^ [B(r) - b0] max

and for m > N ^ B (r ) 4- b0 + e we obtain

\\^m(or)\\rm ssmax
|cr|«€ max
fc=SN \\^k(<r)\\rk = Af(r, e).
This shows that the terms of the series (9.5.37) are bounded, uniformly
with respect to z and a for |z| *s r < R, \cr\ ^ e < 8. This in turn implies
convergence of the series and shows that W (z, <r) satisfies (9.5.36). The
uniform convergence with respect to z and a implies that °W (z, a) is a
holomorphic function of both variables and, moreover, that the series
(9.5.37) may be differentiated termwise as often as we please with respect
to z as well as with respect to cr. Furthermore, we have mixed partials and
352 LINEAR nth ORDER EQUATIONS

relations of the form

i^^^i&i^^ (9-5-38)
The question of how we can utilize these various facts in the search for
a solution of (9.5.1) now arises. We can let a -»0 in the series for W(z, a).
The uniform convergence of the series, as well as of its partials, shows
that there is a limit and that this limit satisfies the DE. The chances are,
however, that the limit is identically zero, and, if not, it is likely to be a
nonfundamental solution. Here we resort to the time-honored device of
differentiation with respect to a parameter, a in the present case. In fact,
this device was used by Frobenius in his memoir of 1873.
The perturbation term o-p+!za°+0' vanishes for a = 0, together with its
partials with respect to a of order ^p. Differentiating (9.5.36) p times
with respect to cr and using (9.5.38), we get

Here we let a tend to zero,

" dpthe perturbation term vanishes, and it is seen
that

°lf(z,cr)lJ<7=0 =W(z) (9.5.39)

is a solution (and actually a fundamental solution) for z # 0.
Obviously (?) °lf(z) has the structure predicted in (9.5.31). In fact, each
time we differentiate a powerda'z^0+m+(T with respect to cr a factor log z
appears. Since p
\im^-v(apz^+n = p\z\
it is seen that
<60O = pW*0 (9.5.40)
and a regular matrix. The first logarithmic factor appears when the term
with index nx is differentiated; fordo-m = n2 a further factor will appear and
(log z)p will correspond to m = nk. No higher powers of log z will appear.
Our solution °Mf(z) is fundamental since the leading term p !za° is a regular
matrix for z ^ 0. This completes the proof of Theorem 9.5.2. ■

EXERCISE 9.5
1. Verify (9.5.6).
2. Verify the corresponding results at oo.
3. Derive the indicial equation (9.5.11).
4. How is (9.5.16) obtained, and what does it mean?
THE FUCHSIAN CLASS; THE RIEMANN PROBLEM 353

"2
5. Show that, if C*% = \%,%*0, then % is nilpotent.
6. If A, and A2 are nonzero points in the spectrum of C^, and if C<ffi = A, 8? and
= A2<&, then either » = 0 or A, + A2 G o-(C^) and = (A, + A2)M.
Given the matrices
0 0 2 0 0
0 1 0 0 0 0
0 1 1 0 0 2
si =
find (i) the characteristic and (ii) the minimal equations of these matrices.
7. 2& =
Obtain the corresponding resolutions of the identity.
8. What form does exponential formula (9.5.23) take for these special matrices?
9. Derive (9.5.23).
10. da dp,
Verify Daletzky's formula,

(2tt/)2 JFl Jr2 A -a +0

where the contours of integration surround the spectrum of si once in the
positive sense. [Hint: Operate with A^-C^ to verify that
(\%-C3lW(\;C«)<% = % for all Then replace % in the formula by
\%-s£% + des£ to get the inverse formula. Remember further that
(a%-s$)9l(a;sdW = %.]
11 For the matrix % defined in Problem 6 the operator Ca has a positive integer
in its spectrum. What does Foguel's theorem say about the multiplicity of
this pole? Try to verify this result with the aid of Daletzky's formula.
12. Verify (9.5.26).
13. Why does (9.5.29) imply convergence?
14. Verify (9.5.36).
15. Supply missing details in convergence proofs.
16. Discuss the differentiation process to get (9.5.31).
17. Are there any coefficients %* which have to be zero?
18.
Solve z°W'(z) = (\% + siz)°W(z), where A is a complex number and si G 2TC„.
19. Find the resolvent of si and zA if

20. With this si study the commutator Show that (C^)3 = 4Cd, and use this to
find the resolvent &(A ; Cd).

9.6. THE FUCHSIAN CLASS; THE RIEMANN PROBLEM

A linear DE belongs to the Fuchsian class if all its singularities are regular
singular points. This implies that there are only a finite number of such
354 LINEAR nth ORDER EQUATIONS

points and that the point at infinity is either regular or regular singular. For
the case of a matrix DE this tells us right away the equation is of the form

W'(z) = 12 :rr^} W(z)' (9-61>

where the a/s are distinct complex numbers, and the SS/s are elements of
Wln, not necessarily distinct.
In the nth order linear case the situation is less obvious. Theorem 9.5.1
gives us, however, the structure at each of the singularities; and, piecing
these items of information together, we get the following picture. Set

P(z) = (z - a,)(z - a2) • • • (z - am), (9.6.2)

and let Qp{z) denote a polynomial of degree not exceeding p; then an
equation of the class must have the form

+ ,,, + ter)w(z) = a (9,63)

This form is necessary as well as sufficient for all the singularities to be
regular singular.
In the matrix case the point at infinity is not a singular point if
$, + 2&2+- • •+»„, = 0, (9.6.4)
the zero matrix. For the nth order linear case the condition is much more
involved. The standard method of studying this case is to use an inversion
z = lis. If the transformed DE in s has the origin as a regular point, then
by definition the original DE admits the point at infinity as a regular point.
The calculations become more complicated the larger n is, but for n = 3
the work is manageable. Set

„M-M) „ / -X - Q2m-2(Z) v <?3W-3U) ,0,.v

P.U)- p(z) , PM- [p(z)]2, p3(z)- [p(2)]3. (9.6.5)
Then the transformed equation is

,("6__1_ /J\] ^/2w

d53 + U 52P,U/J ds2

Here the three coefficients have to be holomorphic at s = 0. This

THE FUCHSIAN CLASS; THE RIEMANN PROBLEM 355

condition places rather severe restrictions on the coefficients. Thus

Pl(2)=|+«+...)
p2(2) = p-p+---, (9.6.7)

pAz)=0(z-%
For the (^-polynomials this implies that

Qm-,U) = 6zm-' + bz'-2+---,

, (9'6'8)
Q2m_2(z) = 6z2-2 - 2bz2-3 +■■■
while Qim i(z) actually has to reduce to Q3m-6(z), which presupposes
m > 1.
For (9.6.1) it is possible to obtain a representation of the solutions valid
in the entire domain of analyticity. Indications of such a possibility occur
in papers of Fuchs and of Poincare, and the representation was worked
out in detail by J. A. Lappo-Danilevskij. The work of this brilliant author
was collected and published by the Akademia Nauk after his death.
Mark the points au a2, . . . , am in the complex plane and denote by S
the universal covering surface of the extended plane, punctured at
z = au a2, . . . , am and infinity. On this surface each of the functions
log(z - cij) is holomorphic, and it is the least surface with this property.
Let b be a point on the surface, and define the family of hyperlogarithms
Lb(ah, ah, . . . , aip; z), where

-1,
— = log -rb —- cij
Lb{ai\z)=\Jb -s —- dj ds (9.6.9)
Lb(ah, . . . , aip, aip+l ; z) = Jb| Lb(ah, ...,ajp;s) S - Qi

The subscripts /,, . . . ,/p take on the values 1, 2, . . . , m. These functions

are zero at z = b and are single valued and holomorphic on S. If C is any
rectifiable path on 5 from b of length L and if its distance from the
branch points of S is 6, then

M«h, • • • > a* ; 2)1 < (f (9.6.10)

This is clearly true for p = 1 and is easily verified by complete induction.

The representation theorem of Lappo-Danilevskij (1928) reads as follows:
356 LINEAR nth ORDER EQUATIONS

THEOREM 9.6.1
For any z on S

Hf(z; b, *) = * + 2 {2 • • • g)jpLb(ajp) ajp_„ . . . , a„; z)}, (9.6.1 1)

vt>/iere f/ie second summation is extended over all ordered products of p
factors formed from the matrices SS2, • • • , 2&m.
Proof For the convergence proof we start by estimating the sum for a
fixed value of p. We have

2 • • • %pU(ah, aip „ . . . , a„; z)

This shows that the series

*^is )'2ll»i^
(fconvergent and that

\\W(z; b, %) - %\ s exp [| | p,||] - 1. (9.6.12)

That this function W(z; b, %) is a solution of (9.6.1) is shown by the
following argument. By (9.6.9)

Lb{ajp, . . . , ai2, ah; z) = ^ \a Lb(aip, . . . , ah\ z).

The series can obviously be differentiated term by term, giving

°r(z; b, %) = 2 — }— %x%2 ■ • • tohLb(ah, . . . , ah; z).

In this multiple series we collect the terms for which /, has the fixed value
k. It is seen that the coefficient of (z — ak)~l equals $fokW{z ;b,%), so that
the series (9.6. 1 1) satisfies (9.6. 1) and has the initial value % at z = b. ■
At this juncture it is appropriate to make some remarks on the Riemann
problem for DE's of the Fuchsian class. The original Riemann problem of
1857 concerned his P-f unction. It is required to determine a function
\a b c ]
z^P\a (3 y z (9.6.13)

such that (i) it is analytic throughout the whole plane except for the
singular points a, b, c; (ii) any three determinations of this function are
linearly dependent; and (iii) at each of the singular points a, b, c there are
two distinct determinations,
THE FUCHSIAN CLASS; THE RIEMANN PROBLEM 357

(z - z0)p'/,(z; z0), (z - z0Y2f2(z; z0),

where /, and f2 are holomorphic at z = z0 and ^ 0 there. Here for

z0=a, Pi = a, p2 = a',
z0=b, p, = 0, p2 = p', (9.6.14)
z0=c, pi = y, p2 = y'.
These conditions determine the P-function uniquely, provided that

a + a' + j8 + P' + y + y' = 1.

This satisfies a linear second order DE reducible to the hypergeometric
equation. The six special determinations will be of the prescribed type if
none of the differences a - a', p - P', y-y'is zero.
For the early history of the Riemann problem, see Section 10.5. It did
not appear in print until the publication of Riemann's Gesammelte Werke:
Nachlass in 1876, and for almost 30 years afterward the Riemann
problem slumbered in peace. In 1905 it caught the attention of David
Hilbert (1862-1943), who may have been inspired by a lecture given by
Wirtinger at the 1904 Heidelberg Congress on part of the Riemann
Nachlass. Be that as it may, Hilbert found the problem amenable to the
new theory of integral equations. He first solved as a preparation the
problem of constructing inner and outer matrix functions with respect to a
given analytic matrix si(z) and a given simple closed rectifiable curve C
such that J>(z) = sd(z)°U(z) on C. Here 3 (z ) is holomorphic inside C, while
°ll(z) is holomorphic outside. sd(z) is holomorphic in a ribbon containing
C, and C has to satisfy fairly severe restrictions. These, however, made
no difference to the original Riemann problem, which Hilbert solved for
the case of 2-by-2 matrices. The problem was then to determine a DE of
the Fuchsian class, knowing the location of its singular points and the
generators of its group of monodromy.
Soon afterward (1906, 1907) Josip Plemelj (1873-1967) took up both
problems in Wln with simpler methods and fewer restrictions on C. George
David Birkhoff (1884-1944) tackled the problems in 1913; moreover, he
generalized them so as to deal also with DE's having irregular-singular
points. He also considered Riemann problems for linear difference and
g-difference equations.
A new mode of attack was found by Lappo-Danilevskij in 1928. For a
DE of the Fuchsian class his general representation theorem (Theorem
9.6.1) gives the value of the solution at every nonsingular point of the
Riemann surface S. We are dealing here with the matrix DE (9.6.1) and
the solution °lf (z; fr, %). The representation is in terms of a power series in
m noncommuting matrix variables with complex coefficients. The
358 LINEAR nth ORDER EQUATIONS

generators of the monodromic group are the values of W at z =

bub2, . . . ,bm, where bj is congruent to b modulo a closed circuit L„
beginning and ending at z = b (in the plane) so that Ly surrounds the
singular point z = a, and no other singularity. In this setting the problem
becomes one of inverting m simultaneous power series in m unknowns —
obviously not a trivial task.
If the generators of the group © are denoted by %, %, . . . , <gm, we have
the system of simultaneous equations

%=% + 2m% + 2 {E 2& A ' • • ®,PLb(aipt . . . , ah; &,■)},

/ = l,2,...,m. (9.6.15)
Note that for each / there is a single first degree term. Its multiplier is the
increment of Lb(ay;z) when z describes the loop L,; all the other
Lb(ak ; z) are assigned the increment zero when L, is described if k ^ /. To
make the system more amenable to a function-theoretical argument, we
set
%-% = Mh / = l,2,...,m. (9.6.16)
Equation (9.6.15) then becomes

% = (m, - 2 {S • • • %PU{aipi . . . , ah; 6y)}), (9.6.17)

where j takes on the values from 1 to m.
It pays to rewrite this equation in vector form. Set

V = (T„ Y2, . . . , Tm) with ||V|| = max Ml, (9.6. 18)

so that the vectors M and B are well defined. We also define a
vector-valued function of vectors F(B), whose kth component is the
matrix-valued function

2 {2 ^ A ••• (a* • • • , fl/. ; )}• (9.6. 19)

In this manner we obtain the functional equation

B = ^t[M-F(B)]. (9.6.20)

Here F(B) satisfies a Lipschitz condition,

||F(B) - F(C)|| ^ (1/N)[exp (2mLN/8)- 1 - 2mLN/8] \\B - C||, (9.6.21)

where N is an upper bound for the matrix vectors to be considered. We
denote the multiplier by K(N).
THE FUCHSIAN CLASS; THE RIEMANN PROBLEM 359

We now define a complete metric space 3c and a contraction operator T,

whose unique fixed point is the desired solution. We consider all vectors
V = (T,, . . . ,Tm) subject to the following conditions. Let

T[V]=^t[M-F(V)]. (9.6.22)
Let No be the positive root of the equation
K(N) = 277, (9.6.23)
and let Ni be a fixed positive number < N0.
We have then

LIT K (N,)||U - V|| ^ k ||U - V||,

||T(U) - T(V)|| ^ J— (9.6.24)

where k <\. Thus the mapping V T(V) is a pure contraction. Let 3E be

the space of all vectors V of norm <NU If Ve£, then by (9.6.22)

||T[V]|| ^ =!-
L7T [||M|| + ||F(V)||] ^ ±
LTT [||M| + k\\\\\].

Here ||V||<N,, so ||T[V]|| < Nx provided that

||M||<(27r -k)Nu (9.6.25)

If M satisfies (9.6.25), then V e 1 implies that T[V] G £. Since the space £
is complete (why?), there is a unique fixed point and (9.6.20) has a unique
solution.
It remains to prove the Lipschitz condition (9.6.21). Consider the terms
of weight p in (9.6.19), i.e., disregarding scalar factors, terms of the form

<&i<&k . . . &ip. (9.6.26)

Here the first factor can be selected in m different ways, and the choice
does not affect the choice of the following factors. Thus we have rap
terms of weight p. Furthermore, the difference

can be estimated by writing &jk = %k + 9)/k so that \\%k ||«s ||B - C|| = D.
Since pJk|| and \\%k\\ < N, we have for D < N,

11^, • • • %P - % ■ • • %A = + %Z - - • (%P + %h ■ - ■ %A
^(N, + D)p - Ni =D[(N, + D)P ,+(N1 + D)p_2N, + - • - + Nr']
< Nr'^"1 + 2P~2 + • • • + 1]D = Nprl2pD.
360 LINEAR nth ORDER EQUATIONS

Thus in the norm of F(B) - F(C) the terms of weight p contribute at most

(2m)"AT"-'^T(J)P ||B-C||.
Summing for p from 2 to oo, we get (9.6.21).
We have thus proved ■

THEOREM 9.6.2

For ||M|| < (277 - k)Ni the functional equation (9.6.20) has a unique
solution B = (3#i, @t2, • • • , 38m). To this solution corresponds a DE (9.6.1)
with coefficients and having ^ = % + Mh j = 1, 2, . . . , m, as generators
of its group of monodromy.
This result seems to be the best obtainable by this method. Every
transit matrix is regular, but not every regular matrix can figure as a
transit matrix for a particular DE of the Fuchsian class. Here the evidence
is somewhat conflicting. On the one hand, L. Bieberbach (1953, p. 251)
claims that the matrices

°aj, a*l, together with c*l,ac*l, (9.6.27)

cannot be transit matrices of a hypergeometric equation. On the other
hand, for m = 2, n = 2, Lappo-Danilevskij seems to have proved that the
functional equation (9.6.20) is solvable for B in terms of M no matter how
M is chosen. The solution is no longer necessarily unique and need not be
single valued (Lappo-Danilevskij, Memoires, 1936, Vol. Ill, pp. 77-97).
EXERCISE 9.6
1. Verify (9.6.3).
2. Verify condition (9.6.4).
3. Obtain the transformed equation (9.6.6), and verify condition (9.6.8).
4. Verify (9.6.10).
5. Expand Lb(ah, aj2, . . . , ajp ; z) in powers of z - b.
6. Prove that

Lb(ah, ah,..., aip ; z) = k2= 1 Lb(ah, ...,aik; c)Lc(aik+], ...,aip; z).

9.7. IRREGULAR-SINGULAR POINTS

Let us now consider a matrix linear DE with an irregular singularity which
we place at infinity. Thus the equation is
IRREGULAR-SINGULAR POINTS 361

z^'(z) = zp 2 Mp-'mz), (9.7.1)

where the Mfs are constant matrices in %Jln and M0 ^ 0. The series is
supposed to be convergent in norm for large values of |z|. Also, p is a
positive integer, the rank of the singularity. It is desired to find a solution
in Wln holomorphic in some sector abutting at infinity. We shall take the
equation in the Birkhoff canonical form,

z°W'(z) = zp 2 djZ-'Wte). (9.7.2)

In two papers (1909, 1913) G. D. Birkhoff
J=0 attempted to show that it is
always possible to reduce (9.7.1) to the form (9.7.2) by a change of
dependent variable. In both cases he put

°lf = 9&(z)<&, (9.7.3)

where the matrix S&(z) is holomorphic in some neighborhood of infinity.
In the first case he assumed det 2&(oo) ^ 0. In the second case

mz) = z8fl®iz-' (9.7.4)

convergent for R |z| < ». Here det 2fc0 may = 0 but det $Kz) should ^ 0
for R |z|. Furthermore, g is an integer.
Birkhoff 's argument was questioned; in 1953 F. R. Gantmacher and in
1959 P. Masani gave counterexamples for p = 0 (the regular-singular
case). In 1963 H. L. Turrittin came to the rescue. He showed that it is
always possible to find an equivalent matrix,

d(z) = z"5>,z-', (9.7.5)

with a finite q which, however, may have to exceed p. If p = 0, Birkhoff 's

first method would have given q - 1 while the second one could have
given q = p = 0. More generally, when p > 0 and the second type of
transformation is used, Turrittin could show that q = p is obtainable
provided that the characteristic roots of si0 are distinct. When there are
multiple roots the situation is still obscure. Turrittin could prove, how-
ever, the following theorem, which we state without proof.

THEOREM 9.7.1

// n = 2, (9.7.1) may be reduced to the form (9.7.2) with q = p by a

transformation of the second type, provided that the asymptotic solution
of (9.7.1) does not contain fractional powers of z"1.
362 LINEAR nth ORDER EQUATIONS

This is true for p = 0, and Turrittin's proof is essentially induction on

the rank.
Turrittin has also proved that it is always possible to reduce the rank
from p to 1 at the price of increasing the order of the matrix algebra from
n to np.
Fractional powers may very well appear. Thus the equation

zw(z)=r?
l_2 ?
4 °W(z) (9.7.6)

is of rank 1, and a solution matrix is given by

rexp(z,/2) exp(-z,/2) ]
hz-m exp (zm) -\z~m exp (- zm)\' ^ ,J)
Formal, (possibly asymptotic) solutions have figured in the study of
irregular-singular points since 1872, when Ludwig Wilhelm Thome (1841—
1910) introduced what he called normal series. For a second order linear
DE such a solution would be of the form

eP(zV(c0+c1z-, + - • •)• (9.7.8)

Here P(z) is a polynomial in z of degree p, the rank of the singularity, p is
a constant, and the coefficients q have to be determined. This can be done
step by step if the coefficients of the DE are rational functions of z. The
e'P(z)
expansion is normally divergent, but the
Z-pw (z) series Xo cjz~i represents

asymptotically in the sense of Poincare.

Now we recall that, when n = 2, the grade of the singularity in the
sense of Section 5.4 is an integral multiple of \. If in the present case the
grade is an odd multiple of \, the asymptotic behavior is sharply changed.
Now P(z) will be a polynomial in z,/2, and the series will proceed by
powers of z~1/2. Such variants of the normal series, known as subnormal
series, date back to the 1885 Paris thesis of Eugene Fabry (1856-1944),
whose name is also attached to a theorem on noncontinuable gap series.
For the matrix case (9.7.2) we have the following result, which will be
stated without proof:

THEOREM 9.7.2

// the characteristic roots of d0 are all distinct, say s,, s2, . . . , sn, there
exist n scalar polynomials pk(z) and n exponents ax, <72, . . . , o~n such that
°lf (z) = X(z0{ . . . , ep'(z)+CT" ,oe \ . . . } (9.7.9)
IRREGULAR-SINGULAR POINTS 363

is a fundamental solution of (9.7.2). Here

Pk(z) = -P skzp + • • • + sk,1zp-1 + • • • + sk,k_,z. (9.7.10)

%{...} is a diagonal matrix with the stated element in the kth place. The
curvilinear sectors are bounded by infinite branches of the curves

Cjk:^[Pj(z)] = ^[Pk(z)]. (9.7.11)

For each sector there is a matrix series N(z) in powers of 1/z which is
asymptotic to the function.
If there are multiple characteristic roots of d0, the representation
changes in nature. There is a polynomial Pk for each of the distinct roots,
but this may be a polynomial in a fractional power of z. The X(z) series
may proceed after negative fractional powers. Finally, instead of the
diagonal matrix 2) we obtain a diagonal block matrix. We omit further
details.
The divergent series X(z) may be replaceable by a convergent factorial
series,

F(z) = a0+ ~,
S z(z
, +^ c), . .^1 < - TTv
.[z + (m l)c] c>0- (9-7-12)

Such a series has a half -plane of absolute convergence, Re(z) > ora, and in
addition possibly a half-plane of ordinary convergence, Re(z) > o-0, where
0 =ss cra - (To ^ c. Such a representation of F(z) presupposes that F(z) can
be represented by a Laplace integral,

Jo
F(z)=\ e ztG(t)dt + a0, (9.7.13)
with

G(t)= m2= 1 aM-e-ct)m, (9.7.14)

where the series converges in a domain D extending to infinity, the
conformal image under the mapping v = e'ct of the disk \v - 1| < 1 and
having the lines lm(0 = ±\ttc~x as asymptotes. If F(z) is not bounded in a
half-plane but has this property in a sector, we may get a valid series
representation by replacing z by zk in (9.7.12) and (9.7.13) for some
positive integer k.
The use of factorial series for the purpose of obtaining convergent
representation in a partial neighborhood of an irregular singularity dates
back to Jakob Horn (1867- J 946), who considered such representations in
1915. In 1924 he also examined binomial series for the same purpose.
Such representations were again studied in great detail by the Russo-
364 LINEAR nth ORDER EQUATIONS

American mathematician W. J. Trjitzinsky in 1935. He showed that not all

formal series arising in formal integration of a linear DE at an irregular-
singular point could be replaced by factorial series in some variable z8.
In this part of the discussion we have dealt with series having scalar
coefficients. We get similar results if the coefficients are n -by-n matrices.
Turrittin (1964) considered linear DE's of the form

VWM&zU + l7.^ + m)H>' (9J-'5)

where p = 0 or 1. The sf s are constant matrices, and the lead matrix d0
has, for instance, the form

s&o = (8ii(pi3>i +&)). (9.7.16)

Then (9.7.15) has a fundamental solution of the form

where the series converges in norm in some half-plane if the series in the
right member of (9.7.15) has such a half-plane of convergence. Here
p = 0, the J>,'s are unit matrices, one for each distinct root p, of the
characteristic equation of si0, and the sum of the orders of the J^'s is
equal to n. Finally the $,'s are square matrices of the same order as
with zeros and ones running down the first subdiagonal, all other elements
being zero.
Mention was made in Theorem 9.7.1 of sectors in the plane abutting at
infinity, in which the solution has different asymptotic behavior. This is
Stokes's phenomenon, named after the Cambridge don George Gabriel
Stokes (1819-1903), who also made important contributions to optics and
hydrodynamics. We have already encountered this phenomenon for
second order linear DE's — more precisely, in our study of Bessel's, the
Hermite-Weber, and Laplace's equations.
Stokes's phenomenon leads to further difficult questions — in particular,
the problem of finding the relation between different asymptotic solu-
tions. The general theory of equations of type (9.7.1) tells us that, if °Wx{z)
and °W2(z) are distinct fundamental solutions, °W2(z) = °Wx(z)%, where C is
a nonsingular constant matrix. This applies also to the Stokes problem.
Let us, with Turrittin (1964), consider (9.7.2), the Birkhoff canonical
form. Suppose that the neighborhood of infinity is covered by m sectors
Si, S2, . . . , Sm, where the numbering is counterclockwise and 5, contains
the end of the positive real axis. Two adjacent sectors, S, and 5J+, (with
Sm+i = SO, have a narrow strip in common. For each sector S, a
fundamental solution TV, will be given, and the problem is to determine
 365
IRREGULAR-SINGULAR POINTS

the Stokes multiplier %:

WJ+1(z) = W,(z)%. (9.7.18)
In principle, though but seldom in practice, this is possible since St and
Sj+i intersect. If z = z0 is a point common to both sectors, we find that

% = [°^i(2:o)]-%+i(2o). (9.7.19)
Equation (9.7.2) has z = 0 as its only finite singularity, and this is a
regular-singular point (why?). It follows that the equation has a funda-
mental solution °lfo(z) which involves powers of z and possibly also
powers of log z, each multiplied by an entire function. This means that
°lf0(z) is defined for all finite values of z^ 0. The domain of definition of
°W0 thus contains each of the sectors Sh We have now Stokes's matrices
such that Wy(z) = °lf0(z)%o, and we have % = (%0r'%+i,o. Again this is
more easily said than done. In some cases, however, we have seen that
the divergent asymptotic series may be replaced by convergent factorial
series, so at least the determination of the Stokes matrices will involve
convergent expansions.
We shall now discuss the application of the Laplace transformation to
the case p = 1 and the BirkhofT canonical form

z°IT(z) = [d0z + d.W (z), (9.7.20)

where si0 and six are n-by-n constant matrices and a solution in Wln is
desired. We set

°W(z) = j %{t)eztdt, (9.7.21)

where the weight function °\L(t) and the path of integration are to be
determined. Substitution and integration by parts gives

(t$ -sdo)ezt\c = fc[tf ~sio)<JU'(t) + (sd^J>)(U(t)] ezt dt

To satisfy this we can find a solution of

(tJ -stoWV)* + S)°U(t) = 0 (9.7.22)

and a path C such that the integrated part

(t<?-s#o)ezt\c=0. (9.7.23)
Now, if t does not belong to the spectrum of &0, the matrix (t$- d0) has
a unique inverse, the resolvent $l(t ; d0), and the auxiliary DE becomes

= - $l(t ; sd0)(sd , + $)<U{t). (9.7.24)

To satisfy (9.7.23) the path C would normally surround a part of the
366 LINEAR nth ORDER EQUATIONS

spectrum of si0 and extend to infinity in a direction such that

\imezt = 0. (9.7.25)
This condition can be satisfied for any given value of z ^ 0.
The singular points of (9.7.24) are the point at infinity and part or all of
the spectrum of si0. Normally the whole spectrum would be singular, but
the factor s&i + $ may possibly annihilate the projection operator & and
hence also the nilpotent 21 that belong to a particular spectral value. See
formula (9.5.21) and the following ones. We omit the trivial case six = -3,
where the solution is elementary and may be read off by inspection. If R
is so large that the spectrum of si0 lies in the disk |f | < R, we can expand
the coefficient of °IX(0 in (9.7.24) in powers of t~x to obtain

t<W(t) = - |af, + $ + ]| jC(j*i + $)t~m j°U(0. (9.7.26)

Referring to Section 9.5, we see that, if the difference set

[y ; y = a - j8, a G <r(s4,), 0 E o-(s£,)]

does not contain any positive integers, (9.7.26) has a solution

<U(t) = (fi + 2 (9.7.27)

where the series converges for |f| > r(^0), the spectral radius of si0, i.e.,
the radius of the least circle with center at t = 0 which contains the
spectrum of &0. Furthermore, a(&i) is the spectrum of si{ and the power
r-'""* is defined as in (9.5.27).
We can use this solution to form the integral (9.7.21). Suppose that
Re(z) > 0, and take for C a loop surrounding the spectrum of s£0 and the
negative real axis. More precisely, let C be the semicircle \t\ = r(&0)+ 1,
Re(0 > 0, and the two lines lm(0 = ± [r(s£0) + 1], Re(0 ^ 0, described in
the positive sense so that arg t goes from - tt to + it. The integral will then
exist, for on C

|pl(0||^M|f|b, 11*11-1 (9-7-28)

and
\\t$-M IMOII exp [Re(zO]-0.
To derive the series expansions for this solution, we need to define the
reciprocal gamma function of a matrix d £ Wln. In the scalar case |T(A
is representable by a power series convergent for all A, an infinite product,
and Hankel's integral (6.4.23). In each of these expressions it is permis-
sible toreplace the scalar A by a matrix si in Wn and thus obtain a definition
of (\IT)(s£). The infinite product shows that (l/r)(^) is a singular element of
IRREGULAR-SINGULAR POINTS 367

$Jln iff one of the elements si,si + 3>, ... ,si + n3>, .. .is singular. It also
gives relations like

±11
(d) i - = -7T sin (77%s4), (9.7.29)

where the sine function is defined by the power series or the infinite
product. It is also seen that

^f^ + « = f(4 (9.7.30)

Hankel's integral is what is particularly needed here. It gives

(9Z31)
f^dbl
where C is a loop surrounding the spectrum of d and the negative real
axis; t ~^ is defined as in (9.5.27). It should be noted that the three
definitions of 1/r given above are consistent. Moreover, the extension of
1/r from C to Wln is unique in the sense that this is the only definition that
leads to matrix functions with components holomorphic in the sense of
Cauchy.
Let us now substitute the series (9.7.27) in the integral (9.7.21) and
integrate termwise, as is permitted by the uniform convergence. The
result is

W(z) = 2m 2 %A [Ax + (k + \)3>]z^+k. (9.7.32)

This is an entire function of z of order 1 multiplied by z*\ It is a

fundamental solution unless six + 3 is singular. The order is obtained by a
direct estimate of the terms of the series, using (9.7.31). Although it is a
solution, the representation gives no idea of its asymptotic behavior.
To explore this we have to examine the spectrum of sl0. The method to
be used presupposes that the spectrum has certain properties to be stated,
and we shall see later what simpler assumptions will ensure that these
properties hold. We assume the following:
51. The spectrum of si0 contains a value S\ which is a simple pole of

52. If 0*1 is the corresponding idempotent, the difference set of the

spectrum of ^(si, + $) contains no integer ^ 0.
We have then

<3i{t ; d0)(d, + *) = ^'(^' + ^) + m2=0 Sr + 'Cstf, + ^)(s, - t)m.

* Sl (9.7.33)
368 LINEAR nth ORDER EQUATIONS

Here = = 0. This follows from (9.5.21), for we have now

I — S\ j =2

Since SP,^,- = 0 f or 1, we see that the series in (9.7.33) is annihilated by

and this will hold also if the factor (s^ + J>) is omitted. Since this holds
for all t for which the series converges, it also holds for t = su which
means that 0>,<3& = 0. Since commutes with 3&, we have also <3&0>, = 0.
Now the auxiliary DE (9.7.24) may be written as

°H'(0= +
L t — 5i J) am+,(^, + ^)(s,-OmK(0.
m=o J (9.7.34)
This shows that t = sx is a regular-singular point of the DE, and condition
S2, by Theorem 9.5.1, ensures that the equation has a solution of the form

°ltff ) = mE=0 9L(f - Sl)"*,(*,+*)+m, $0 = 4. (9.7.35)

Here the series converges in the same disk as the resolvent series (9.7.33).
Now (9.7.24) has the point at infinity as a regular-singular point. The finite
singularities are poles at most n in number. It follows that the solution
(9.7.35) may be continued analytically to infinity along any path that
avoids the finite singularities. Here the part of the path outside a large
circle is at our disposal and may be chosen so that Re(zt)-*-™.
We can take C as a loop in the resolvent set of si0, i.e., the set where
t§ - si0 is regular. If Re(z) > 0, we can take the infinite part of the loop to
be two horizontal lines going to the left. This will ensure that Re(zf ) -> - °°,
and condition (9.7.23) is ensured, so that the corresponding integral
(9.7.21) is a solution.
To get the asymptotic representation of this solution we proceed as in
the discussion of Laplace's equation in Section 6.4, formulas (6.4.24)-
(6.4.34). In other words, we substitute the series (9.7.35) in the integral
(9.7.21) and integrate termwise. Since the series that enters in the integral
diverges for all large values of t, the series obtained by integration will
also diverge, but it gives an asymptotic representation of the solution in
the sense of Poincare.
We start by assuming 5, to be real positive. This is not as restrictive as it
may seem, for (9.7.21) admits of transformations which preserve the form
but shift the "parameters" and dx. The substitution
W(z) = (%(z)e(3°zz(i> (9.7.36)
has this property and replaces si0 and sii by

slQ-Po$ and di-pj, (9.7.37)

IRREGULAR-SINGULAR POINTS 369

respectively. Here the /3's are arbitrary complex numbers. The spectral
values of MQ - 0to$ are Sj - /30. Here /30 is at our disposal and may be
chosen so that the singularity under consideration is real positive. The
transformation obviously carries simple poles into simple poles. Consider
now one of the integrals obtained by the process,

j"(t-SCly^+ *)+keztdt.
We take Re (z) > 0 and deform into a loop surrounding the line segment
(-00, Si) of the real axis. We set t = Si + s/z to obtain

z»1(*i+*>-fc-1 £si2 j s~9%isAi+*)+k es ds

= 27rij; [g>,(s&, + 4) - k^z^^-^1 es'\

where we use (9.7.31) and, if necessary, a suitable rotation of the path of
integration. Multiplying by £ftk and summing for /c, we obtain the formal
series

°W(z) ~ 2iri k=0

2 1 + ^) - Wlz*'^^-*-1 es'2. (9.7.38)
The series is likely to be divergent, for the factor involving the reciprocal
gamma function grows as a factorial. The series is asymptotic in the sense
of Poincare and represents the solution in the right half-plane. In the first
approximation it behaves as
2^lW esxZ (9.7.39)
The proof that this is indeed an asymptotic representation follows the
same general lines as the argument in Section 6.4 for the special scalar
case. It is long and involved and will be omitted here.
The same type of argument applies to the left half-plane; the same
series is obtained, but normally it represents a different solution. The
solutions may be expected to be singular elements of since <3i\{sl\ + 3)
is a singular matrix and hence also the reciprocal gamma function. If there
are several spectral values of «s40 satisfying conditions Si and S2, each will
give an asymptotic representation of a particular integral, making the
appropriate change of the coefficients 2&m, the idempotent SP,, and the
spectral value 5, in (9.7.35). If, in particular, the n spectral values of ,s40
are all distinct and hence simple poles of <3l{t; sio), then we have n linearly
independent solutions which may be combined to form a fundamental
system. In this case the solution (9.7.32) will be the sum of solutions of
known asymptotic behavior.
Suppose now that the matrix has distinct characteristic roots
370 LINEAR nth ORDER EQUATIONS

Si, s2, . . . , sn. We can then transform si0 into a diagonal form. There
exists a regular matrix jit such that MsloMT1 = 9)0 = 2>(. . . Sj . . .), and
^(z) = M°W(z)M satisfies
z<3/'(z) = (9>0z + <€,)<3/(z), <€, = «AU4,JT!. (9.7.40)
We have now
+ £) = + (9.7.41)
where ^ is the matrix which has a one in the place (j, j) and zeros
elsewhere. If % = (cJk), the right member of (9.7.41) is the matrix whose
jth row is
- 8jk ~ Cjk,

while all other elements are zero. The characteristic values of this matrix
are 0, repeated n — 1 times, and — 1 — c„. It is permitted to assume that the
last number is not an integer or zero, for if this should turn out not to be
the case we could use a transformation of type (9.7.36), now affecting only
sii. The preceding discussion applies and gives n linearly independent
solution vectors for the right half-plane and n for the left. The elements of
the jth column vector behave asymptotically as constant multiples of
z*e* ft = c„.
Equations of type (9.7.20) are of considerable importance for the set of
applicable DE's where the associated matrix algebra is normally Wl2. This
includes the confluent hypergeometric case. For n >2, K. Okubo (1963)
has done valuable work, but nontrivial problems remain unsolved in this
field.
For the case p > 1 we refer to Section 6.9 of LODE and the memoirs of
Turrittin, in particular.

EXERCISE 9.7

In 1913 G. D. Birkhoff examined as a by-product the case p = 1, n = 2 and, in

particular, the canonical form of the second order DE
zw" + (z - a)w' + bzw = 0.
Let a and /3 be two roots of the equation
\2 + (a - 1)A + b =0,
where a - (5 is not zero or an integer.
1. Show that the equation has the solutions

za,F,(a,a -j3;z) and zp,F,(/3, /3 - a; z).

2. If a and a - /3 are real positive, estimate the maximum modulus of the first
solution.
LITERATURE
371

3. Prove the existence of asymptotic solutions of the form

2F0(a, j8 ; - j) and ezZ-\FQ(\ - a, 1 - 0

4. Show that, if the series (9.7.12) converges for z = x0>0, it converges for
every z with Re(z)>x0.
5. Why is the width of the strip of nonabsolute convergence at most c?
6. Use the properties of the gamma function to show that, if the series
converges for z = jc0 > 0, then

am=o{cm exp[(y°-l)logm]].
7. Verify the form of G(t) in (9.7.14).
8. With G(t) as in (9.7.14) find the Laplace transform of G'(t).
9. Why is the origin a regular-singular point of (9.7.20)?
10. Show that the point at infinity is a regular-singular point of (9.7.24).
11. Verify (9.7.32).
12. Try to verify the asymptotic character of (9.7.38).
13. Take Wl2 and the matrices

<■-[*» MS?}
where /3 is real positive. Determine the spectrum of sio, and find the residue
idempotents, the spectrum of <3>(sli + $), and the leading terms in the
asymptotic expansions.
Solve Problem 13 if

»• *-c :i *-c -ii

LITERATURE

The relevant parts of the author's LODE are Appendix B and Chapter 6.
See also Chapters 3-5 of:
Coddington, E. A. and Norman Levinson. Theory of Ordinary Differential Equations.
McGraw-Hill, New York, 1955.
Also see Chapters 15-20 of:
Ince, E. L. Ordinary Differential Equations. Dover, New York, 1944.
This treatise has a brief discussion of the monodromic group on p. 389.
In connection with Theorem 9.4.6 see:
372 LINEAR nth ORDER EQUATIONS

Birkhoff, G. D. A simplified treatment of the regular singular point. Trans. Amer. Math.
Soc, 11 (1910), 199-202.
For Lemma 9.4.1 and Theorem 9.4.9 see:
Koch, H. von. Un theoreme sur les integrates irregulieres des equations differentielles
lineaires et son application au probleme de l'integration. Ark. Mat., Astr., Fys., 13, No.
15, (1918), 18 pp.
Perron, O. Uber einen Satz des Herrn Helge von Koch iiber die Integrale linearer
Differentialgleichungen. Math. Zeit., 3 (1919), 161-174.
The discussion in Section 9.5 is an adaptation to the matrix case of
Section 6.6 of LODE, and this goes back to three papers of the author.
See:
Hille, E. Linear differential equations in Banach algebras. Proceedings of the International
Symposium on Linear Spaces. Jerusalem, 1960, pp. 263-273.
Remarks on differential equations in Banach algebras. Studies in Mathematical
Analysis and Related Topics: Essays in Honor of George Polya. Stanford University
Press, 1962, pp. 140-145.
Some aspects of differential equations in B-algebras. Functional Analysis: Proceedings
of the Conference at University of California, Irvine, 1966, edited by B. R. Gelbaum.
Thompson Book Co., Washington, D.C., 1967, pp. 185-194.
For a fuller discussion of the commutator operator, including the
integral of Yu. L. Daletskij and the theorem of S. R. Foguel, see Section
4.6 of LODE.
There is an extensive literature on the Fuchsian class and the Riemann-
Hilbert and Riemann-Birkhoff problems. The matrix theory is developed
in great detail in:
Lappo-Danilevskij, I. A. Memoires sur la theorie des systemes des equations differentielles
lineaires. 3 vols. Trav. Inst. Phys.-Math. Stekloff, 1934, 1935, 1936; Chelsea, New York,
1953.

Though not perfect, a convenient, readable presentation of the

Riemann-Hilbert and Riemann-Birkhoff problems is found in:
Birkhoff, G. D. The generalized Riemann problem for linear differential equations and the
allied problems for linear difference and q-difference equations. Proc. Amer. Acad. Arts
Sci., 49, No. 9 (1913), 521-568.
This paper has a good account of the earlier work on these problems. Of
later papers we mention:
Gamier, R. Sur le probleme de Riemann-Hilbert. Compositio Math., 8 (1951), 185-204.
Helson, H. Vectorial function theory. Proc. London Math. Soc, (3), 17 (1967), 499-504.

Helson extends Birkhoff's equivalence problem for analytic matrices to

analytic operator functions operating in a Hilbert space.
References for Section 9.7 are:
LITERATURE 373

Birkhoff, G. D. Singular points of ordinary differential equations. Trans. Amer. Math. Soc,
10 (1909), 436-470.
Equivalent singular points of ordinary linear differential equations. Math. Ann., 74
(1913), 134-139.
A simple type of irregular singular point. Trans. Amer. Math. Soc, 14 (1913), 462-476.
Fabry, E. Sur les integrates des equations differentielles a coefficients rationnels. Thesis,
Paris, 1885.
Gantmacher, F. R. The Theory of Matrices. Vol. 2, Chelsea, New York, 1959; Russian
original, Moscow, 1953.
Horn, J. Integration linearer Differentialgleichungen durch Laplacesche Integrale und
Fakultatenreihen. Jahresb. Deut. Math. Ver., 24 (1915), 309-329.
Laplacesche Integrale, Binomialkoeffizientenreihen und Gammaquotientenreihen.
Math. Zeit., 21 (1924), 85-95.
Hukuhara, M. Sur les points singuliers des equations differentielles lineaires. II. J. Fac. Sci.
Hokkaido Imp. Univ., 5 (1937), 123-166. III. Mem. Fac. Sci. Univ. Kyusyu, A2 (1942),
127-137.
Masani, P. On a result of G. D. Birkhoff on linear differential systems. Proc. Amer. Math.
Soc, 10 (1959), 696-698.
Okubo, K. A global representation of a fundamental set of solutions and a Stokes
phenomenon for a system of linear ordinary differential equations. J. Math. Soc. Japan,
15 (1963), 268-288.
Thome, L. W. Zur Theorie der linearen Differentialgleichungen. /. Math., 74 (1872); 75, 76
(1873); and many later volumes. Summaries in 96 (1884) and 122 (1900).
Trjitzinsky, W. J. Analytic theory of linear differential equations. Acta Math., 62 (1934),
167-227.
Laplace integrals and factorial series in the theory of linear differential and difference
equations. Trans. Amer. Math. Soc, 37 (1935), 80-146.
Turrittin, H. L. Stokes multipliers for asymptotic solutions of certain differential equations.
Trans. Amer. Math. Soc, 68 (1950), 304-329.
Convergent solutions of linear homogeneous differential equations in the neighbor-
hood of an irregular singular point. Acta Math., 93 (1955), 27-66.
Reduction of ordinary differential equations to the Birkhoff canonical form. Trans.
Amer. Math. Soc, 107 (1963), 485-507.
Reducing the rank of ordinary differential equations. Duke Math. J., 30 (1963), 271-274.
Solvable related equations pertaining to turning point problems. Symposium on
Asymptotic Solutions of Differential Equations and Their Applications. Publication No.
13, Mathematical Research Center, U.S. Army, University of Wisconsin, John Wiley,
New York, 1964, pp. 27-52.
10

THE S CHWAR Z IAN

The Schwarzian derivative enters into several branches of complex

analysis. It plays an important role in the theory of linear second order
DE's, in conformal mapping, and in the study of univalence and of
uniformization. We shall have something to report on all these aspects.
First a historical note is in order: the first "Schwarzian" seems to have
occurred in a paper by Kummer listed in a Liegnitz Gymnasialprogramm
of 1834. The paper, "De generali aequatione differentiali tertii ordinis,"
was reprinted in Crelle, Vol. 100, 1887. As we shall see in Section 10.5, the
operator was also known to Riemann as early as 1857.

10.1. THE SCHWARZIAN DERIVATIVE

Hermann Amandus Schwarz (1843-1921) was Weierstrass's most brilliant

pupil and ultimately became his successor at the University of Berlin. In
the late 1860's he was engrossed in the problem of mapping circular
polygons conformally. This led him to his famous principle of symmetry
and also, in 1869, to the differential operator which later became known as
the Schwarzian derivative or the differential parameter or, simply, the
Schwarzian. What he wanted was a differential operator invariant under
the group of all fractional linear transformations, also known as the
projective group.
by We denote the Schwarzian of w with respect to z by {w, z}. It is defined

(10.1.1)

This is obviously

(10.1.2)

374
THE SCHWARZIAN DERIVATIVE
375

and may also be written as

w' 2 \w' / '

The primes denote differentiation with respect to z. There are of course
other differential operators invariant under the projective group, but in
1969 the late Israeli mathematician Meira Lavie showed that such an
operator of order n operating on a function /, under mild restrictions on /,
is expressible rationally in terms of {/, z} and its derivatives of order
s£ n - 3. There are various considerations which lead to this operator, but
in a treatise on ordinary differential equations perhaps the following mode
of approach is the most natural one.

THEOREM 10.1.1

Let yi and y2 be two linearly independent solutions of the equation

y" + Q(z)y = 0, (10.1.3)

which are defined and holomorphic in some simply connected domain D in
the complex plane. Then

w(z) = £g (10.1.4)
satisfies the equation

{w,z} = 2Q(z) (10.1.5)

at all points of D where y2(z) ^ 0. Conversely, if w(z) is a solution of
(10.1.5), holomorphic in some neighborhood of a point z0ED, then one
can find two linearly independent solutions, u(z) and v(z), of (10.1.3)
defined in D so that

w(z) = Hg, (10.1.6)

and if v(z0) = 1 the solutions u and v are uniquely defined.
Proof. We may assume that the Wronskian of y, and y2 is identically
one. Then
w'(z) = [y2(z)]\
so that

w"(z)= y2(z)
w'(z) y2(z)
376 THE SCHWARZIAN

and

l_w'(z)J = _2zM£)+2r>*)i2
2 y2(z) + ZU(z)J = 2Q(2)+i
z^z) + 2 r^)T
[w'(z)i '
from which the first assertion follows.
Next suppose that a solution of (10.1.5) is given by its initial values
vv(zo), w'(zo), w"(z0) at a point z0 in D. Here we may assume that
w'(zo) t6 0 since otherwise Q(z) could not be holomorphic at z = z0. We
can now choose two linearly independent solutions, u(z) and v(z), of
(10.1.3) so that at z = z0 the quotient Wi(z) = u(z)lv(z) has the initial
values vv(zo), w'(zo), w"(z0), the same as w(z). If v(z0)= 1, the solutions
w(z) and u(z) are uniquely determined, and we must have w(z) = Wi(z).
The details are left to the reader. ■
Since (10.1.5) is of the third order, its general solution should involve
three arbitrary constants. Such a solution can be found via (10.1.3). If
yi(z) and y2(z) are linearly independent solutions, if a,b,c,d are con-
stants such that ad - be = 1, then ayi(z) + by2(z) and cyi(z) + dy2(z) are
also linearly independent, and their quotient

cy,(z) + dy2(z) = ^
ayfltb/f\ tt°.1.7)
is a solution of (10.1.5) involving three arbitrary constants. This situation
suggests the truth of

THEOREM 10.1.2

The Schwarzian is invariant under fractional linear transformations

acting on the first argument

The proof is a matter of a computation and is left to the reader. Thus

the Schwarzian is a differential invariant of the projective group. A
change of independent variable obeys the law

{wyz} = {w,t}(^)\{t,z}. (10.1.9)

We may apply a Mobius transformation to the second argument instead
of the first. The result is

f az + b] (ad - be)2 , , ,1A 1 im

(10-110)
[w'7FTd\ (cz + dY =Kz}-
APPLICATIONS TO CONFORMAL MAPPING
377
The second factor on the left is the square of the derivative of the second
argument with respect to z.
We have expressed the general solution of (10.1.5) in terms of solutions
of (10.1.3). The converse is also possible, for if yu y2 are solutions of
(10. 1 .3) of Wronskian equal to one, and if w = y Jy 2, then, as we have seen,

y2(z) = [w'{z)V\

and yi(z) = w(z)y2(z) = [u>'(z)]"1/2h>(z), so that the general solution of

(10.1.3) is

Ay,(z) + By2(z) = [w\z)Tm[Aw(z) + B]. (10.1.11)

In the domain D where Q(z) is holomorphic, the solutions of (10.1.5)
are evidently meromorphic with simple poles (why?) and these are
movable singularities.

EXERCISE 10.1
1. If {w, z} is identically 0, show that w is a linear fractional function of z.
2. If /(z) is holomorphic at infinity, show that z4{/, z} is holomorphic there.
3. In the second half of the proof of Theorem 10.1.1, determine the initial
values of u(z) and v(z) subject to the stated conditions.
4. Prove Theorem 10.1.2.
5. Prove (10.1.9).
6. If f(z) = 4z(l + z)~\ show that {/, z} = - 6(1 - z2) 2 and in the unit disk z, * z2
implies /(zi) ¥■ f(z2), so that / is univalent there.
7. If /(z) = log[(z-a)/(z-&)], a*b, find {f,z} = 2Q{z).
8. What is the general solution of y" + Q(z)y =0 with Q as in Problem 7?
9. Prove that {w, z} = -{z, •
10. Prove (10.1.9).

10.2. APPLICATIONS TO CONFORMAL MAPPING

The conformal mapping problem was formulated by Bernhard Riemann in

the (later) famous dissertation of 1851. He based his existence proof on
Dirichlet's principle in the calculus of variations. Weierstrass showed that
the existence of a minimizing function could not be concluded in this way,
and the mapping problem remained an open question until 1913, when
rigorous proofs were given by Caratheodory, Koebe, and Osgood. In the
meantime numerous attempts were made either to provide an existence
proof under less general assumptions or to find explicit solutions for
important special cases. In both attempts Schwarz was a pioneer.
378 THE SCHWARZIAN

Particularly important for the later development was his solution of the
problem of mapping the upper half -plane conformally on the interior of a
circular polygon (1869). It revealed a close connection between mapping
problems and linear DE's of the Fuchsian class, which in its turn led to
Henri Poincare's discovery of automorphic functions, i.e., analytic func-
tions F(z) such that
F(Sz) = F(z), (10.2.1)
where S belongs to a given group © of fractional linear transformations
on z. Special examples already known before Poincare's work in the
1880's are doubly periodic functions and elliptic modular functions.
Under suitable assumptions regarding ®, there exists a circular polygon
n, the interior of which acts as a fundamental region of the group, in the
sense that the images of this region under the elements of the group cover
without overlapping the domain of existence of z •-> F(z). Moreover, the
inverse of the mapping function of the fundamental region is an auto-
morphic function under the group.
With this background information, let us proceed to the specific
mapping problem. There is given a simple, closed polygon II in the finite
w-plane with vertices A,, A2, . . . , A„ numbered in such a way that the
interior of II lies to the left of an observer proceeding from w = Aj to
Aj+i, where An+X = Ax and j = 1, 2, . . . , n. The sides of the polygon are
circular arcs or straight line segments. At w = Aj the two adjacent sides
make an angle of asir with each other in the interior of II, where 0 < as < 2
and aj ¥■ 1.
It is possible to map the upper half of the z-plane conformally on the
interior of II, and the mapping function is unique up to a fractional linear
transformation which leaves the upper half-plane invariant. Here the
boundaries are in a one-to-one correspondence. To the vertex Aj corres-
ponds apoint aj on the real axis, and, without loss of generality, we may
suppose that
fli < a2 < ■ • • < a„.
Actually it is possible to fix three of these points, say the first three, at
preassigned points by a suitable choice of the mapping function. The
remaining points are then uniquely determined. Let this choice of the
mapping function be z f(z).
Whereas /(z) is normally rather complicated, its Schwarzian {/, z} is a
simple rational function, with double poles at the points ajy which
vanishes at least to the fourth order at infinity. This implies that /(z) is the
quotient of two linearly independent solutions of a second order linear
DE of the Fuchsian class. More precisely, we have
APPLICATIONS TO CONFORMAL MAPPING 379

THEOREM 10.2.1

If the a's and a's satisfy the above restrictions, then there exist real
numbers ft such that

where

Ei& = 0,2i (2/3^+1 -a?) = 0,

(10.2.3)
2[j3ja? + (l-a?)ai] = 0.

Proof (sketch). The correspondence between the upper z-half -plane and
the interior of II is one-to-one. This implies that f'(z) ^ 0 for Im (z) >0
and, thus, that {/, z} exists and is holomorphic in the upper half -plane. On
the open intervals (ah ai+i) the mapping function /(z) is continuous and
takes values on the arc (Ah A,+i) of II. The original formulation of
Schwarz's principle of symmetry states that, if a function z^g(z) is
holomorphic in a domain D in the upper half -plane bounded below by a
segment (a, b) of the real axis on which the boundary values are real and
continuous, then g admits of an analytic extension across (a, b) into the
symmetric domain D and g(z) = g(z). The extended symmetry principle
is obtained by applying a Mobius transformation to the w-plane. The
principle then states that /(z) can be continued into the lower half -plane
across any of the line segments (a„ flj+i), and its value at z is obtained
from its value at z by reciprocation in the arc (Ah AJ+i). Thus, if the
center of the circle is at w = b and its radius is R, then

f{z) = b+— f (10.2.4)

The values of /(z) in the lower half -planef(z)-bobtained by crossing (ah aJ+i)
fill out the interior of a polygon 11, similar to n and obtained from it by a
reciprocation in the arc (Ah Ai+i). We can obtain the analytic continuation
of /(z) across each of the arcs (Ah Ai+i), including (A„, Ai), and the result
is a set of n polygonal regions attached to the n sides of II. Each of these
regions is the conformal image of a copy of the lower z-half-plane. The
derivative f'(z) ^ 0 in each of these half-planes, and, since the correspon-
dence of the boundaries is one to one, f'(z) also ^ 0 in each of the open
intervals (ah ai+i). This means that {/, z} is holomorphic for each continua-
tion.
This reciprocation process can now be repeated for each of the n lower
380 THE SCHWARZIAN

half-planes. From the jth lower half-plane we can cross n - 1 intervals

(ak,ak+\) with k^j. This gives rise to n - 1 polygons attached to the
previously free sides of the polygon IX. Over the z-plane a Riemann
surface is generated, normally with infinitely many sheets, each sheet
consisting of an upper and a lower half-plane joined along one of the line
segments (ah aj+i). The values of /(z) at conjugate points are related by a
reciprocation in one of the circular arcs. If z describes a closed path in the
plane surrounding one or more of the points ah the mapping function will
normally not return to its original value but /(z) is locally holomorphic in
the finite plane except at the points ah It is also holomorphic at infinity,
which on the sphere is an interior point of the arc of the great circle which
joins an with a\. The Schwarzian {/, z} is holomorphic everywhere in the
finite plane, the points a} excepted. It is also holomorphic at infinity, since
/(z) has this property. Moreover, by Problem 10.1:2, z4{f, z} is holomor-
phic at infinity, and this requires that {/, z} have a zero at infinity at least
of order 4.
We have to show that {/, z} is holomorphic in the whole plane, again
excepting the points a, — in fact, that {/, z} is a rational function of z with
poles at z = a,, j = 1, 2, . . . , n. First let us note that {/, z} is real on each of
the intervals (ah ai+i) regardless of the determination of the mapping
function. This follows from Theorem 10.1.2, for the values of / in the
interval (ah ai+\) lie on some circular arc, and we can find constants
a, b, c, d with ad — be = 1 such that

$f^F(2) (10.2.5)
is real on the interval under consideration. We recall that a Mobius
transformation maps circles and straight lines into circles and straight
lines. In particular, we can find a transformation which carries the circular
arc that is the carrier of the values of /(z) into the real axis. But if F is real
on (ah so are all its derivatives and hence also {F, z} = {/, z}.
We come finally to a discussion of the singular points ah which
correspond to the vertices A, of the original polygon. Here we know that
the adjacent arcs meet at w = Ah forming an angle c^tt with each other.
For the discussion of the local properties of {/, z} at z = it would be
simpler if the sides were straight line segments rather than circular arcs.
Again this can be achieved with the aid of a transformation of type
(10.2.5), which does not change the Schwarzian. We take

(,0-2-6)

Here vv0 is the vertex under consideration, the original A, or one of its
 APPLICATIONS TO CONFORMAL MAPPING 381

homologues in the repeated reciprocation process. The two circles

intersect at w0 and at wu and cj is so chosen that one of the arcs (w0, Wi) is
mapped on the positive real axis, while the other is mapped on a ray which
makes the angle with the positive real axis. The corresponding
mapping function F(z) is then of the form

F(z) = (z-ajpHi(z), (10.2.7)

where Hj(z) is holomorphic at z = a} and different from zero there, and
the power series expansion of Hj(z) in powers of z - a, has real
I coefficients. This gives

F (z) Z - Gj
where normally bj 5* 0 and is real. This shows that in a neighborhood of
z = cij we have

tt*>-«*>-I?F^i+I^ + ^)' <10-2-8)

where hj(z) is holomorphic at z = ah and its only singularities are the
points ak with j. From this we conclude that {/, z} is holomorphic in
the extended plane except for double poles at the points ah
It remains to prove that {/, z}, which is a rational function of z, actually
equals the sum of its principal parts. To this end we form the difference

D^=tfz>4|,^-|,A- (10-2-9)
This difference is holomorphic at each of the points ah regardless of the
determination of the mapping function considered. It is also holomorphic
at infinity. This means that D(z) is a constant, and since D(z) is zero at
infinity, it must be identically zero.
Now {/, z} vanishes at least to the fourth order at infinity, and if the
right-hand side of (10.2.2) is expanded in powers of 1/z, conditions (10.2.3)
express the fact that the terms in l/zm are missing for m = 1, 2, 3.
If {/, z} is given by (10.2.2), the equation

y"(z) + M/,z}y(z) = 0 (10.2.10)

is clearly of the Fuchsian class. Its only singularities are the points
au a2, . . . , an, and they are regular-singular points. ■

EXERCISE 10.2
1. Find the general form of a fractional linear transformation which leaves the
upper half-plane invariant. What is the image of the real axis?
382 THE SCHWARZIAN

2. Consider

(1-f2) 1/2(4-f2),/2 dt,

where the radical equals +\ at t = 0. Show that this function maps the upper
half-plane conformally on the interior of a rectangle in the upper half-plane
with two vertices on the real axis.
3. Find the Schwarzian of the mapping function in Problem 2.
4. Show that the inverse of the mapping function in Problem 2 is a doubly
periodic elliptic function, and find integrals representing the primitive
periods, one real, the other purely imaginary.
5. Make a similar study of the mapping function

which is associated with an equilateral triangle.

6. Generalize so as to obtain a mapping function for a general circular polygon.
7. Consider the function /(z) = 2z(l + z2)-1. Show that it maps the upper
half-plane conformally on the whole w-plane cut along the line segment
(-1, + 1). This is a mapping of the upper half -plane on the exterior of a
degenerate polygon, namely, the double line from - 1 to + 1. Here there are
two angles, each of opening 27r in the "interior" of the degenerate polygon.
8. Show that

with log z = log r + id, 0 =££ 6 ^ tt, maps the upper half-plane conformally on
the interior of a circular polygon bounded by two tangent circles, one inside
the other. This is a case with <xx = a2 = 0 (horn angles).
9. Find the Schwarzian of this mapping function.
10. Find the general solution of the equation

y"(z) + Mz}y=0.
1 1 . Find the general solution of
{w,z} = {f,z},
where again / is defined as in Problem 8.
12. Construct a mapping function for a domain outside of two externally tangent
circles.
13. Let y,(z) = F(U, l;z), y2(z) = iF(U, 1; 1 - z). Verify that the quotient
w(z) = y2(z)/y,(z) satisfies

(the Schwarzian equation for the modular function).

ALGEBRAIC SOLUTIONS OF HYPERGEOMETRIC EQUATIONS 383

14. Show that w(z) maps the upper half-plane conformally on a "triangle"
bounded by two parallel vertical lines and a half-circle joining the end points
of the lines on the real axis, the triangle lying in the upper half-plane. There
are three horn angles. The inverse of the mapping function is an automorphic
function invariant under the group generated by the transformations

z ->z + 1, z ->-.z

10.3. ALGEBRAIC SOLUTIONS OF HYPERGEOMETRIC EQUATIONS

The problem of finding when a hypergeometric DE has algebraic solutions
was the point of departure of Schwarz's investigations, which led to so
many profound results. This is of course only a special case of the more
general problem of determining when an equation of the Fuchsian class of
any order has algebraic solutions. The results are more striking and easier
to grasp, however, in the hypergeometric case, to which we restrict
ourselves.
There are really two problems:
1 . When does a hypergeometric DE admit of an algebraic solution?
2. When are all solutions algebraic?
We start with the first problem.
1. Let the equation

z(l - z)w" + [y - (a + p + l)z]w' - a/3w = 0 (10.3.1)

have an algebraic integral w(z). This would normally not be single valued,
but we may assume that the various functional elements that belong to the
same center z = z0 have a constant ratio. This implies that the logarithmic
derivative w'(z)/w(z) is single valued, and being an algebraic function of
z it is in fact rational:

w(z)
^j^ = R(z). (10.3.2)

There can obviously be no logarithmic terms in the expansions of w(z) at

the singular points 0, 1, ». If w(z) should be holomorphic at z = 0 and 1,
then w(z) must be a polynomial, and the indicial equation at infinity
shows that one of the parameters a or /3 is then required to be a negative
integer or zero. Suppose that a=-n. Then the solution is w(z) =
CF(-n, /3, y ; z). Various special cases occur in Exercise 10.3.
Suppose that w(z) is not a polynomial. We can then find rational
numbers a and b such that

w(z) = za(z-\)bp(z)y (10.3.3)

384 THE SCHWARZIAN

where p(z) is a polynomial, for the rational function R(z) can only have
simple poles and it vanishes at z = ». Hence

(10.3.4)
Z Z - 1 &Z-CI}
Indeed a multiple pole would imply an irregular singularity, which is
absurd. Integration gives

w(z) = za(z-\)bf\(z-ain

and here the exponents must be positive j = i integers, since otherwise there
would be additional singularities. It is clear that a and b must be rational
if w(z) is to be algebraic. Again the indicial equations impose some
conditions on a, b and the degree N of p(z). We leave these considera-
tions to the reader.
2. Suppose now that all solutions are algebraic and that Wi(z) and
w2(z) are linearly independent solutions. Their quotient
w2(z)
W(z) = ^r\ (10.3.5)

is then also algebraic and satisfies the Schwarzian equation

1\2 2 1x2 2,2

{W'Z] = — + W^?+ 2z(l-z) ' (103-6)

where the relations between the two sets of parameters are

A = 1-7, /ui=y-a-j3, v = a - fi, (10.3.7)

and
a = Kl - A — 11 + v)9
P = \{\-\- H-v), (10.3.8)
y= 1 - A.
The general solution of (10.3.6) is then also algebraic, and it has the form
aW(z) + b
ad-bc = \. (10.3.9)
cW(z) + d'
If we start at some point z = z0 where W(z) is holomorphic and describe a
closed path returning to z0, all determinations of the solution are of the
form (10.3.9). Now, if the solution is to be algebraic, only a finite number
of determinations can exist at z = z0. This implies and is implied by the
group of monodromy of (10.3.1) being finite. To each element of the group
ALGEBRAIC SOLUTIONS OF HYPERGEOMETRIC EQUATIONS 385

of monodromy corresponds a linear fractional transformation

w — >cs +—7.
d (10.3.10)

Now (and this is the crucial observation), a group of linear fractional

transformations can be interpreted as a group of rotations of the sphere
via the stereographic projection of the plane on the sphere. Here the word
"rotation" can be taken in the sense of either Euclidean or non-Euclidean
geometry. Finite groups of rotations of the sphere are of the following
types: (i) the identity, (ii) cyclic groups, (iii) dihedral groups, (iv) the
tetrahedral group, (v) the octahedral group, and (vi) the icosahedral group.
It would take us too far afield to attempt a rigorous proof of these
statements, but we shall try to indicate the various considerations
involved and will enumerate the sets of parameters a, /3, y for which the
hypergeometric DE has all-algebraic solutions.
The point of departure is Schwarz's triangular functions s(A, n,v ;z),
i.e., the solutions of (10.3.6). From the discussion in Section 10.2 we know
that any determination of such a function maps the upper half-plane into a
circular triangle with angles A 77, i±tt, vtt. The mapping is one to one and
conformal if the (real) parameters satisfy the inequalities |A| «s 2, ^ 2,
IH^2. If all solutions of (10.3.1) are to be algebraic, there can be no
logarithmic terms in the expansions at the singular points. This implies, in
particular, that 1 - 7, y - a - f$, and a - /3 cannot be zero; i.e., the
parameters A, /ll, v are not zero, so no horn angles can be present. The
monodromy group of (10.3.1) now appears as the group of transforma-
tions which are composed of an even number of reflections in the three
circles that carry the circular arcs bounding the triangle in question. We
may assume that these circles are distinct. If all three should coincide, the
group is the identity. The case with two coincident circles and a third
distinct from them is also trivial. Therefore let all three circles be distinct.
We now map the plane on the surface of a sphere by stereographic
projection. Then the three circles in the plane are mapped on three circles
on the sphere, and the group of monodromy becomes a group of
collineations leaving the sphere invariant. Our three circles on the sphere
are obtainable as intersections of three planes with the surface of the
sphere. Here there are three separate cases according as the planes
intersect in a point outside, on, or inside the sphere. In the first case the
three circles in the plane have a common orthogonal circle which is left
invariant by the monodromy group, and the latter becomes a group of
motions in the non-Euclidean plane (the Lobatsevskij geometry as
realized by Poincare). In the second case the common point of the three
circles can be sent to infinity by a suitable linear transformation. The
386 THE SCHWARZIAN

monodromy group then becomes a group of motions in the Euclidean

plane. In the third case the point of intersection of the three planes can be
sent to the center of the sphere by a suitable collineation. Now the
monodromy group appears as a group of rotations of the sphere. The
exceptional cases in which two or all three circles coincide can be brought
under this heading as a group of rotations of the sphere. Here the finite
groups of rotations of a sphere are well known. They may be cyclic or
leave a dihedron on one of the regular polyhedra invariant. Note that the
cube and the octahedron have the same group, and so also is the case with
the dodecahedron and the icosahedron.
The finite groups of motion of the non-Euclidian geometry are cyclic.
This is seen as follows. Linear fractional transformations are classified
according to the nature of their fixed points and the motion around these
points. There are loxodromic, hyperbolic, parabolic, and elliptic transfor-
mations. Only the last type can occur if the group is to be finite.
Moreover, the group is generated by a single element (5. If there is an
element of the group, say which does not commute with <S, it may be
shown that (S^©-1^1 has a fixed point on the orthogonal circle, and this
element of the group is either hyperbolic or parabolic, which is nonadmis-
sible. A similar argument applies to a group of motions in the plane: if a
group is not cyclic, it must contain a translation and this violates the
assumption that the group is finite.
Once the finite rotation groups of the sphere are known, one can
proceed to further study of the problem on hand. The mapping of the
z-plane on the s-plane is obtained by analytic continuation of the function
which maps the upper half-plane conformally on the basic circular
triangle with angles \tt, /jltt, vtt, the continuation being achieved by
repeated reflections in the sides of the original and the image triangles.
The resulting Riemann surface over the s-plane must cover it a finite
number of times if the solutions of (10.3.6) are algebraic. If the s-plane has
nowhere more than one covering, the mapping is one-to-one and z is a
single-valued function of 5. The values of A, /x, v for which such a
single-valued inverse exists are listed in the table.

A M V

i 1 1 Identical group
1 1/n 1/n Cyclic group (n positive integer)
1/2 1/2 1/n Dihedron (n positive integer)
1/2 1/3 1/3 Tetrahedron
1/2 1/3 1/4 Cube (octahedron)
1/2 1/3 1/5 Icosahedron (dodecahedron)
ALGEBRAIC SOLUTIONS OF HYPERGEOMETRIC EQUATIONS 387

It is now possible to handle the general situation in which the solutions

are algebraic but s(z) does not have a single-valued inverse. Suppose that
the general integral of (10.3.1) is algebraic, and consider the correspond-
ing Schwarzian equation
{s,Z} = fl(Z), (10.3.11)
where R(Z) is rational. If the general solution of (10.3.1) is algebraic, so is
the general solution of (10.3.11). Let s = A(Z) be an algebraic solution of
(10.3.11). Together with these equations, consider an equation of type
(10.3.6), say,
{s,z} = r(z). (10.3.12)
It is assumed that this equation has the same group as (10.3.11) and that
the parameters A, /t, v coincide with one of the sets listed in the table. It
has then an algebraic integral s = s(z) with single-valued inverse. Thus
5 (z) is an automorphic function of the monodromy group of (10.3.1 1). We
can then form
z=z[A(Z)] = Q(Z), (10.3.13)
which is algebraic and single valued and hence is a rational function. This
implies that (10.3.1) is obtainable from (10.3.12) by making the substitu-
tion z = Q(Z). From these considerations we obtain

THEOREM 10.3.1

All algebraically integrable DE's of the Fuchsian class and second order
can be obtained by taking an equation of type (10.3.6) with A, /ll, v taken
from the table and setting z = Q(Z) in the equation, where now Q(Z) is an
arbitrary rational function.
We have to restrict ourselves to these brief indications of this beautiful
theory, discovered and developed by H. A. Schwarz around 1870.

EXERCISE 10.3

1. Why is a single-valued algebraic function necessarily rational?

2. What light does (10.3.3) throw on the roots of the indicial equations and on
permissible values of a and bl
Among the hypergeometric equations having polynomial solutions, the following
should be noted:

3. a = - n, j3 = n + 1, y = 1: Legendre polynomial in t =\(\ - z).

4. a = -m, /3 = p + n, y = a, t = 1(1 - z): Jacobi polynomial.
5. For p =0, q = \ we obtain the nth Cebysev polynomial of the first kind:
T„(z) = 2,-"F[-n, nA'Ml-z)].
 THE SCHWARZIAN

hat orthogonality properties do these polynomials have on the interval

(-1,1)?
7. Show that for all polynomials Pn(x) = x" + • • • of degree n and for x
restricted to (-1,1) the maximum of the absolute value of P„(x) is a
minimum for P„ = T„.
8. A "dihedron" is a double pyramid with the base a regular polygon of n sides.
The top and the bottom vertices are at a vertical distance 1 from the center of
the polygon. The dihedral group are the rotations about the diameter joining
the top and bottom vertices. What is the order of the group?
9. What are the orders of the octahedral and the icosahedral groups?
10. Why are the groups of the cube and of the octahedron the same?
11. Answer the question of Problem 10 for the dodecahedron and the
icosahedron.
12. In principle it would seem to be possible for the planes which have circular
intersections with the sphere to have a line in common instead of a single
point. Why can this situation not arise?
13. Find the values of a, /3, y which correspond to the parameters A, /a, v of the
table.

10.4. UNIVALENCE AND THE SCHWARZIAN

Suppose that /(z) is holomorphic and univalent in a domain D in the

sense that Zi # z2, both in D, implies that /(zi)#/(z2). A necessary
condition for univalence is that f'(z) ¥■ 0 in D. Now, if / is holomorphic
and univalent in D, then {/, z} exists and is holomorphic in D. Under
special assumptions on D, it is possible to find other necessary or
sufficient conditions on {/, z} which imply univalence.
There are two domains which are usually considered in problems
involving univalence, namely, the interior K and the exterior Y of the
unit circle. Let (5 denote the class of functions

z^>f(z) = z + a2z2+aiz*+- • • (10.4.1)

which are univalent in K, and let U denote the functions

z ^> F(z) = z + A0 + Aa~* + A2z~2 + • • • (10.4.2)

which are holomorphic and univalent in Y. There is a one-to-one
correspondence between the elements of the two classes, for if

/(z)G(5, then J 6U (10.4.3)

and vice versa. If F(z) = [/(l/z)]~\ then

Ax = a22-a3. (10.4.4)
UNIVALENCE AND THE SCHWARZIAN 389
This will be useful below. We shall also need

LEMMA 10.4.1

If F(z)eU, then |A,| 1.

This follows from Gronwall's area theorem,

nE= l n|A„|2^l. (10.4.5)

Actually this inequality also shows that the inequality |A,| ^ 1 is the best
possible, since

z + Ao+e'V'eU
for any choice of 6. For a proof of the lemma consult Exercise 10.4.
The basic result here is a theorem proved by Z. Nehari (1949).

THEOREM 10.4.1

If f(z) is univalent in K, then

|{f,z}|^6(l-|z|2r2 (10.4.6)
and 6 is the best possible constant.

Proof. Take an arbitrary point z = a in K, and consider the Mobius

transformation,

1 + az = a + (1 - \a |2)(z - az2 - a2z3 +•••) = « + w,

Z = ^4- (10.4.7)

which maps X in a one-to-one manner on itself. It follows that /(Z) is

univalent in K. It does not belong to (5, but

g(2) = /^)(\-\al) = z + a2Z' + fl3z3 + ' ' * (10A8)

does. The assertion of the theorem follows from a comparison of a2 and
a3. By Taylor's theorem

/(a + w) = f(a) + f'(a)w +\f"(a)w2 + if"(a)w3 + • • • .

Here we substitute the power series for w, expand the powers of w, and
collect terms. By the Weierstrass double series theorem this is allowed for
390 THE SCHWARZIAN

small values of z. We obtain

a2=_"+lfS)(l_|a|2)' (10.4.9)

«'-%ja-i.iHfg<i-i.rr.
Thus

fl2-a3 6(1 |fl|) l/'(fl)-2l

7wJ J
-l{\-\a\2f{f,a}.
By (10.4.4) and Lemma 10.4.1 this expression is at most one in absolute
value. Since a is arbitrary in K, inequality (10.4.6) follows. ■
That the constant 6 is the best possible follows from Problem 10.1:4,
according to which z t-»/(z) = z(l + z)~2 is an element of © and {/, z} =
-6(1 -z2)-2.
COROLLARY

If f (z) is holomorphic and univalent in the upper half -plane y > 0, then

|{f,z}|^|y-2, z = x + iy, (10.4.10)

and the constant \ is the best possible.

Proof. This follows from (10.1.9). The upper z-half-plane is mapped on

the interior of the unit circle in the s-plane by

s.
z + i

Univalence of /(z) in the upper z-half-plane implies and is implied by

univalence of the transformed function in the disk K, and a simple
calculation gives (10.4.10). The function z ^ z2 is univalent in the upper
half -plane, and {z2, z} = §z 2.
A simpler necessary conditionf"(z)\ for univalence in K is

(10.4.11)
fU)l l-|z|
Here also the constant is the best possible, for

tt , z . 12f"(z)
• 2z-4
/(2) = (TT77 glves 7(z-) = T^
and the numerator goes to — 6 when z goes to — 1 .
These various conditions are necessary for univalence in the domain
UNIVALENCE AND THE SCHWARZIAN 391

under consideration. Sufficient conditions are normally obtained by

lowering the constant figuring in the inequality. Thus Theorems 8.3.3 and
8.3.4 and the comments thereon show that

|{/,z}|^2(l-|2|T2, |z|<l, (10.4.12)

and

|{/,z}|^b>-2, z=x+iy, 0<y, (10.4.13)

are sufficient for univalence of / in the indicated domain. Condition
(10.4.12) was found to be of importance in the theory of quasi-conformal
mapping by L. Ahlfors and G. Weill in 1962.
Again by (8.3.19) and comments

|{/,z}|^!7r2, |z|<l, (10.4.14)

is a sufficient condition, as shown by Nehari. In all these cases the
sufficient condition is the best of its kind in the sense that for any larger
numerical constant there are nonunivalent functions satisfying the condi-
tion in question. We also note the Pokarnyi-Beesack condition:

|{/,z}|^2(3-a)(l-|z|2r, l^a^2, |z|<l (10.4.15)

for univalence in the unit disk. See also (8.3.22).
S. Friedland and Z. Nehari (1970) have generalized this range of ideas
by proving that, if R (r) is a nonincreasing nonnegative function for which

then the condition J' R(r)dr^\9 (10.4.16)

|{f, z}\ ^ 2R(r)(\ - r2T\ \z\<r, (10.4.17)

is sufficient for univalence of / in the unit disk. This condition is aiso
sufficient if R(r)(\ - r2)"1 is nonincreasing in [0, 1).
Other relations between the Schwarzian and univalence were explored
by Rolf Nevanlinna in 1932. The results may be regarded as a partial
converse of his second fundamental theorem as expressed by (4.5.19). It
is a construction of meromorphic functions with preassigned defects. He
posed and solved the following problem:

Given q (^ 2) points ai, a2, . . . , aq and q positive integers mi, m2, . . . , mq

with X? nrij = p, find all analytic function z(w) with the following properties:

The function w»-»z(w) is meromorphic in the w-plane punctured at a,,

a2, . . . , aq.
2. z(w) is univalent.
392 THE SCHWARZIAN

3. Over each of the points a, there are m} distinct logarithmic elements

and possibly also regular elements.
4. The range of z(w) is a simply connected domain D.
If q = 2, we must have ra, = ra2 = 1, p = 2, and

w - a2+ B
z(w) = A log^-^ (10.4.18)

with arbitrary constants A and B. For q > 2 the problem is solvable iff
2ra, p for each j. In this case the range of z(w) must be the finite plane,
so a unique single-valued inverse function w(z) is defined in the finite
plane. Furthermore, there exists a polynomial P(z) of degree p -2 such
that

(w,z} = 2P(z). (10.4.19)

Conversely, any choice of such a polynomial leads to functions z(w) with
such a set of properties.
Thus we have

*<z> = wfv (10A20)

where yi and y2 are two linearly independent solutions of the DE

y"(z) + P(z)y(z) = 0. (10.4.21)

The discussion in Section 5.4 shows that y{ and y2 are entire functions of z
of order [\p] and of the normal type of this order. Thus w(z) is a
meromorphic function of the same order. Any choice of yi and y2 subject
to the stated conditions will generate a function w(z) with the desired
properties. The corresponding numbers a, are simply the values of C for
which the solution

y,(z)-Cy2(z) (10.4.22)
is subdominant in one of the p sectors at infinity (see Section 5.6). If for
C = a, the solution is subdominant in m, sectors, the inverse function
z(w) has ra, distinct logarithmic elements over w = a,. If p >2, there are
in addition always infinitely many regular elements at a,. The condition
ra, =s£ \p expresses the fact that the solution can be subdominant in at most
half of the sectors. For the equation

y"(z) = n[nz2n~2 + {n -2)zn'2]y(z) (10.4.23)

we have p = 2n, and the solution ez" is subdominant in n sectors. Thus
the bound ra, ^ \p is sharp.
UNIVALENCE AND THE SCHWARZIAN 393

The value of q ^ p and the integers m, are uniquely determined by the

coefficients of P(z). On the other hand, the points a, depend on the choice
of yi and y2. In this case passing from one fundamental system to another
involves a fractional linear transformation of the w-plane and the points
a}. If p = q = 3, the a^s can be assigned prescribed values and P(z) may
be chosen as -z. For p = 4 we have q = 3 or 4, and (10.4.21) may be
reduced to the Hermite-Weber equation involving a single arbitrary
parameter. The exceptional case q = 3 arises iff the equation is

y"(z) + (2n + l-z2)y(z) = 0,

where n is an integer. For nonintegral values the cross ratio of the four
critical points as is a simple function of the parameter.
For further details we refer the reader to Nevanlinna's paper.
EXERCISE 10.4
1. Verify (10.4.3) and (10.4.4).
2. The area theorem may be proved as follows. Let F(z) G 11 and r > 1; then
the circle |z| = r is mapped by F(z) on a simple closed analytic curve C. The
area enclosed by C can be obtained by the methods of the calculus, say by
using polar coordinates. A somewhat more sophisticated expression is

A{r) = \ [Uj£- Vj^} dd, F(z) = U(reid) + iV(re»).

Show that this integral is the imaginary part of

F(z~)F'(z)dz.
3. Show that

A(r) = ir[r2-j£n\An\2r-2n]>0
for all r > 1, and use this to prove (10.4.5).
4. Verify (10.4.7)-(10.4.9).
5. Verify the corollary.
6. How is (10.4.11) obtained?
7. Show that f(z) is univalent in the upper half -plane if
[jc2 + (y + \)2f\{f,z}\^2, z=x+iy, 0<y.
8. Show that another sufficient condition is

y[x2 + (y + \)2]\{f,z}\^4.
9. For q = 2, z(w) is given by (10.4.18). Verify that this function has the desired
properties. Find the corresponding equation (10.4.21). Show that P(z) =
-A 2, independently of au a2, and B. Determine these quantities if
394 THE SCHWARZIAN

y, = sinh (z/2A), y2 = cosh (z/2A ). What happens if yi, y2 = exp (± z\2A )

instead?
10. For what values of A (>\tt2) can there exist a function which is holomorphic
and univalent in the unit disk where the supremum of \{f, z}\ = A?

10.5. UNIFORMIZATION BY MODULAR FUNCTIONS

In 1904 the Austrian mathematician Wilhelm Wirtinger (1865-1945)

addressed the Third International Mathematical Congress at Heidelberg
regarding Riemann's lectures on the hypergeometric series and their
importance. Riemann lectured twice on this topic at Gottingen, in the
Wintersemesters of 1856-1857 and 1858-1859, and lecture notes exist of
both courses. As could be expected, the two courses abounded in new
ideas, most of which were to bear fruit much later.
In the first course the P-function was introduced, represented by Euler
transform integrals, which were to appear in print in a posthumous paper
by Jacobi in 1859. In the second course Riemann solved the special
Riemann problem by constructing a system of functions which transform
in a given manner when z makes a circuit around given singular points,
two or three in number. This was preceded by a brief exposition of the
composition and reduction of linear substitutions, terra incognita to most
mathematicians of the time. He also used the Schwarzian; he probably did
not know of Kummer's paper of 1834, and in any case this was 10 years
before Schwarz used it. Riemann treated conformal mapping of cur-
vilinear triangles on the sphere (he seems to have been the first to use
stereographic projection) or in the plane, the mapping function being the
quotient of two linearly independent solutions of the DE. He noticed that
the independent variable z can be a single-valued function of the quotient,
and that what later became known as the monodromy group implies that z
is invariant under a group of linear fractional transformations of the
quotient. Riemann used the symmetry principle long before Schwarz. In
his lectures he carried out the mapping of the fundamental region for the
elliptic modulus k2 effected by the quotient of the periods of the elliptic
integral of the first kind. In addition, he determined some cases in which
the mapping properties indicate that the solutions are algebraic functions
of z.
Wirtinger's address must have been a splendid lecture. The printed
paper, however, puts rather heavy demands on the reader and suffers
from a lack of formulas and bibliographical references.
We shall take up here an investigation dating from 1888 by Erwin
Papperitz (1857-1938), who carried out part of the Riemann program of
which he could have had no direct knowledge. He considered the
UNIFORMIZATION BY MODULAR FUNCTIONS 395

hypergeometric DE for the P-function which involves six parameters, the

sum of which is one. The quotient of two linearly independent solutions is
a solution of the equation

{^} = (l-AV(A-+^-^-l)Z+(l-^V) (1051)

where the left member is the Schwarzian with respect to z of the integral
quotient s(A, pt, v\ z). Here the three parameters are the differences
A = a - a', pi=/3-/3', v = y -y'.
Felix Klein proved in 1878 that, if a new variable Jc2(t) = z is intro-
duced in the equation, 5 (A, fx, v: z) becomes a single-valued function of
t = iK' IK, where
K = \<nF{\, i, 1; Ac2), K ' = \irF{h, 1, 1; 1 - k2)
by (7.3.22) and (7.3.23). Actually, rik2) is an s-function corresponding to
the parameters 0, 0, 0. The inverse function k2(r) is single valued. This
theorem also occurs in Riemann's second course, unknown to Klein and
to Papperitz. The latter gave a simple direct proof for Klein's theorem
based on the properties of the group © operating on t which leaves
z = Jc2(t) unchanged. This group is a subgroup of the modular group
which contains all mappings of the upper T-half-plane into itself of the
form

T^7^> ad-bc = l, (10.5.2)

where a, b, c, d are integers. Here $R is generated by

S(r) = t + 1, U(t) = -K
T
and the subgroup @ is generated by

S2(t) = t+2, rj5^(T) = _r_ (1Q53)

and (US)\t) = t. Here a and d are odd integers, while b and c are even
and of course subject to the condition ad - be = 1. The fundamental
region B of © is depicted in Figure 10.1, which also shows the image of B
in the z-plane under the mapping

k2(r) = z. (10.5.4)
We have

Int (B) = {t ; |t + J| > \, \r - \\ > |, - 1< Re (t) < 1, 0 < Im (t)}. (10.5.5)
To complete B we add that part of the boundary where Re(T)^0.
396 THE SCHWARZIAN

(0)

(0) (0)

Incidentally this important figure occurs in Riemann's lecture notes for

the second course.
That B is indeed the fundamental region for © follows from the fact
that for any choice of z in its plane there is one and only one root of
(10.5.4).
Let us add some comments on the Klein-Riemann problem. For a given
value of r we have a unique value of z but normally infinitely many
determinations of s(A, ix,v\z). The latter are related via elements of $Jl.
Any determination of s can be obtained from a fixed other determination
in one of two ways: either analytically by letting z describe a suitable
closed path in the z-plane surrounding one or more singular points of
(10.5.1), or by applying a Mobius transformation belonging to © operating
on r. Now to each completed circuit in the z-plane corresponds an
element V of © such that

k2[V(r)] = k2(r). (10.5.6)

Thus, if there should be a value of t such that to it would correspond two
distinct values, S\ and s2, of s[A, /jl, v; /c2(t)], then we can carry one into
the other, say s2 into Si, by one of the two operations indicated above.
Here we must have V(t) = t. Papperitz concludes that this is impossible,
UNIFORMIZATION BY MODULAR FUNCTIONS
397

since @ does not contain any periodic transformations. Both U and US

are periodic, since U2 = I and (USf = I, but neither U nor US belongs to
©, though both belong to 9ft.
Actually we have here a solution of the uniformization problem, since
k2(r) = z and s [A, /x, v \ k2(r)] are single-valued functions of t.
Papperitz goes much further, however. We denote the right member of
(10.5.1) by R(A, ijl,v;z), so that the equation becomes

{s,z} = R(\,ix,v;z). (10.5.7)

But we have also

{r, z} = R (0, 0, 0; z), z= k\r\ (10.5.8)

as observed above. Now we introduce t as a new variable in (10.5.7) with
the aid of (10.1.9), obtaining

{s,z} = {s,T}(£j + {T,z}.

It follows that 5 (A, /ll, v; z) as a function of r satisfies the DE

{5, r} = {R [A, fi9 v ; /c2(r)] - R [0, 0, 0; k 2(r)]}[^^]2. (10.5.9)

We denote the right member of this equation by F(A, pt, v\ r).
The analogue of (10.1.3) is now

^2 + iK(A,jLL,isz)y =0. (10.5.10)

If yi and y2 are two linearly independent solutions of this equation, their
quotient is a solution of (10.5.7). We can introduce t as a new independent
variable in this equation. We do not state the form of the resulting
equation; instead we note the DE
d2Y
dr 2 +£F(A,/u,*;t)Y = 0. (10.5.11)

If Y is any solution of this equation, then

1/2
Y (10.5.12)

is a solution of (10.5.11).
For the F-function we get by substitution

A2(l - k2) - ii2k\\ -k2)+ v2k2 (dk2\2

F(A'^;T) = \k\x-k2) Ur)-
398 THE SCHWARZIAN

Here we need the theta null functions, for which see Exercise 7.3. We
have

* di(0,T)' 1 * di(0,T)" (10.5.13)

Furthermore, it may be shown that

4~ log /c2 = 77/#:(0, t). (10.5.14)

Combining, we find that

F(A, /li, v ; t) = |7r2(A2d3^: - /x2#44#2 + i/2d5dj), (10.5.15)

where we have omitted the (0, r) for the sake of brevity.
The further substitution

q =exp(7rir) (10.5.16)
leads to the equation

0+^-WF(A''t',';T)y = O- (10-517)
Here F(. . .) may be expanded in positive powers of q convergent for
\q\< 1; this means for t in the upper half-plane. Indeed, the formulas in
Exercise 7.3 show that the theta null functions are power series in q with
a factor q1'4 added in the case of #i and -&2. Since we are concerned with
fourth powers, we see that F is a power series in q. Actually

=K
Z7TF(A, M, v ; t) = U 2 + 4 „J)
=i[A2(- q )n + (- 1)>2 + ^2]n3 t-^.
i— q
(10.5.18)
It follows that the indicial equation of (10.5.17) is

p(p - l) + p -U2 = 0 with roots i\ and-U.

Hence, if A is not an integer, equation (10.5.17) has two linearly
independent solutions,
Y, = qA/2G(A, ix, v,q),

<10-519)
y2^-G(-a,^^).
Here G is a power series in q, convergent for \q \ < 1, i.e., Im (t) >0, and
G(A, ix, v,Q) = 1. The coefficients are rational functions of A, /x, v and
G(A, /ll, i/, q) = G(A, v, ix, -q)
identically.
UNIFORMIZATION BY MODULAR FUNCTIONS 399
Since

by (10.5.13) and (10.5.14) we get two solutions of the hypergeometric

equation (10.5.10):

, „ L ,2
y,(T) = (Wi)"2{^i)2<JA'2G(A,M^,q), (10.5.21)

y2(r) = (™)"2{^j q A'2G(-A, „, <,).

Here the theta quotient is, aside from a factor q1/2, a holomorphic
function of q in \q\ < 1 for which we have the infinite product

v2 n a - <,2")2[n (i + q2-Ki - <j2-')]4[n u + <?2"-')]". <10-5-22)

Here all four products are absolutely convergent for \q\< 1, and, since the
last factor does not become zero, the quotient is holomorphic for \q \ < 1.
As t-> + i"<», q^>0, and we see that
y,(T) = q(,+A)/2[l + 0(^)],
rm<™
y2(r)^-1l + OU)].
Now we have
z=/c2(r)=16q[l + 0(^)].
Combining these formulas, we see that, if we set

a =Kl + A), a'=Kl-A),

then yi and y2 must be constant multiples of the two branches of the
P-function which are prescribed at the origin. These two determinations
of P(. . .) are then represented by formulas (10.5.21) in the whole z-plane,
cut as indicated in the right half of Figure 10.1.
We can get similar representations for the two branches of P at z = 1
and at z = oo. For this purpose it is necessary to invoke more theory of
theta functions, namely, how these functions transform if t is replaced by
T(t), where T is an element of the modular group. To study the solutions
of (10.5.1) in a neighborhood of z = 1, we take T = U, i.e., r ->- 1/r. We
write #j(0, - 1/r) = 0, and have the basic transformation formula

^ =(-ir) ",/2t96_j. (10.5.24)

This is now used to express F(A, /ll, v ; r) as a function of the 6 's. If we set
Q=exp(-f),
400 THE SCHWARZIAN

we can expand F(A, /jl, v\t) in powers of Q. In the resulting equation

(10.5.17) q is replaced by Q, and the corresponding indicial equation has
the roots i/ut and -J/ll. There will be solutions analogous to (10.5.19) with q
replaced by Q and A by /li. In the analogue of (10.5.21) the new theta
quotient will also admit of a product expansion of type (10.5.22). Here we
have

k2(r) = 1 - 16Q[1 + O(Q)], (10.5.25)

which shows that we are dealing with the branches prescribed at z = 1.
We refrain from further elaboration of this theme.
The function 5 [A, n, v ; /c2(r)] has the real r-axis as a natural boundary.
Papperitz examines in great detail the behavior of the function as t
approaches a rational real value t along a path not tangent to the real axis.
The s-function becomes infinite for such an approach, but the rate of
growth depends on the arithmetical nature of r. Again we have to abandon
the pursuit.

EXERCISE 10.5
1. Show that the modular group is generated by S and U.
2. Show that the subgroup <& is generated by S2 and US2U.
3. Try to show that the equation k\r) = z has a solution in B which is unique.
4. Verify (10.5.8).
5. Verify that (10.5.12) is a solution of (10.5.10).
6. Assuming the validity of (10.5.13) and (10.5.14), verify (10.5.15).
7. Verify (10.5.17).
8. Prove the absolute convergence of the infinite product in (10.5.22).
9. Verify (10.5.23) and the expression for /c2(r).
10. Discuss Q and (10.5.25).

LITERATURE

We refer to Appendix D of the author's LODE.

The subject matter of Section 10.3 is treated at great length on pp.
40-147, Vol. II: 2, of:
Schlesinger, L. Handbuch der Theorie der linearen Differentialgleichungen. Teubner, Leip-
zig, 1898.
See also pp. 232-245 of:
Bieberbach, L. Theorie der gewdhnlichen Differentialgleichungen. Springer-Verlag, Berlin,
1953.
The references in the Handbuch are in the list of contents but lack titles
LITERATURE 401

and year; those of Bieberbach are very scanty, not to say, in some cases,
lacking.
For Section 10.2 see also Section 17.6 of Vol. II of the author's AFT.
Other references are:
Beesack, P. E. Non-oscillation and disconjugacy in the complex domain. Trans. Amer.
Math. Soc, 81 (1956) 211-242.
Friedland, S. and Z. Nehari. Univalence conditions and Sturm-Liouville eigenvalues. Proc.
Amer. Math. Soc, 24 (1970), 595-603.
Hille, E. Remarks on a paper by Zeev Nehari. Bull. Amer. Math. Soc, 55 (1949), 552-553.
Lavie, M. The Schwarzian derivative and disconjugacy of nth order linear differential
equations. Can. J. Math., 21 (1969), 235-249.
Nehari, Z. The Schwarzian derivative and schlicht functions. Bull. Amer. Math. Soc, 55
(1949), 544-551.
Some criteria of univalence. Proc Amer. Math. Soc, 5 (1954), 700-704.
Univalent functions and linear differential equations. Lectures on Functions of a
Complex Variable. University of Michigan Press, Ann. Arbor, 1955, pp. 148-151.
Nevanlinna, R. Uber Riemannsche Flachen mit endlich vielen Windungspunkten. Acta
Math., 58 (1932), 295-373.
Pokarnyi, V. V. On some sufficient conditions for univalence (Russian). Doklady Akad.
Nauk SSSR, N.S., 79 (1951), 743-746.
Robertson, M. S. Schlicht solutions of W" + pW = 0. Trans. Amer. Math. Soc, 76 (1954),
254-274.
Schwarz, H. A. Uber einige Abbildungsaufgaben. J. reine angew. Math., 70 (1869), 105-120;
Ges. Math. Abhandl., 2 (1890), 65-83.
References for Section 10.5 are:
Klein, F. Ueber die Transformation der elliptischen Functionen und die Auflosung der
Gleichungen funften Grades. Math. Ann., 14 (1878).
Papperitz, E. Ueber die Darstellung der hypergeometrischen Transcendenten durch ein-
deutige Functionen. Math. Ann., 34 (1889), 247-296.
Wirtinger, W. Riemanns Vorlesungen uber die hypergeometrische Reihe und ihre Be-
deutung. Sitzungsberichte der 3ten Internal. Math. Kongress, Heidelberg, 1904. 19 pp.
The notation for theta functions varies; Papperitz used #0 where we
have #4. Also, he wrote SU where we have US. Papperitz has another
paper fairly closely related to the paper here discussed.
Papperitz, E. Untersuchungen uber die algebraische Transformation der hypergeometris-
chen Functionen. Math. Ann., 27 (1886), 315-356.
11

FIRST ORDER

NONLINEAR

DIFFERENTIAL

EQUATIONS

We shall consider some problems relating to first order nonlinear DE's in

this chapter. Second order nonlinear DE's will be discussed in Chapter 12.
In both cases the theory is too extensive and has too many at least
seemingly unrelated features to be covered in the space available; we
have to be satisfied with a glance here and a nibble there. Before
proceeding any further in this chapter, the reader is advised to review the
first three sections of Chapter 3, where he or she will encounter a
discussion of fixed and mobile singularities, the problem of continuing a
solution analytically, and various results obtained by Paul Painleve in
respect to such questions. In the present chapter investigations by
Charles Briot and Jean Claude Bouquet and by Pierre Boutroux will
occupy the foreground.

11.1. SOME BRIOT-BOUQUET EQUATIONS

In Section 3.1 mention was made of equations of the form

2Cikvv'(vv')k =0 (11.1.1)
with constant coefficients encountered by the brilliant French research
team of Briot and Bouquet in their work on elliptic functions. We shall
consider such equations again in this chapter, but at this juncture we shall
examine another class of equations also considered by Briot and Bouquet
402
SOME BRIOT-BOUQUET EQUATIONS 403

(1856). We take
zw' =pz+qw + F(z, w), (11.1.2)
where
F(z,w) = 2 2 c*zV\ (11.1.3)

a function holomorphic in some dicylinder |z| < Rly |w| < K2. Here z = 0,
u> = 0 is a fixed singularity of the equation and is class 4 in the Painleve
classification (see Section 3.2). Here we have

P(z, w) = pz+qw + F(z, w), Q(z, w) = z. (11.1.4)

Both P and Q are holomorphic at (0,0), where both are zero.
Now the first question which the DE (11.1.2) poses is whether or not
there are solutions holomorphic at z = 0. We prove

THEOREM 11.1.1

If q is not a positive integer, then (11.1.2) has a unique solution z»-> w(z)
which is holomorphic at z = 0 and w(0) = 0.
Proof. Let us first note that, if there is a solution holomorphic at z = 0,
then zw'(z) is also holomorphic and obviously zero there. If we replace z
by zero in the right member of (11.1.2), the result is a power series in w
starting with the term qw. It follows that a holomorphic solution at z = 0
must have the initial value w(0) = 0. Suppose that such a solution exists,
and set
w(z) = 2aizi. (11.1.5)
Substitution gives i= i

J'j^ja^
= l =pz+q^Ja>zi
J = l +^j+k2*2
2 c^i^
\n = lanzn\/ . (11.1.6)
Expanding, rearranging, and collecting terms, we get
(1 -q)ax =p,
(2-q)a2 = C20 + Cndi + c02(ai)2, ^ ^1)
(n -q)an = Mn(cjk ; au . . . , an_,),
where Mn is multinomial in the indicated quantities, and j + k ^ n. Since q
is not a positive integer, the a's may be found successively and uniquely.
Convergence of the formal series is proved by a Cauchy majorant
argument involving an implicit function rather than a DE. Consider an
equation of the form
BY = G(z, Y), (11.1.8)
404 FIRST ORDER NONLINEAR EQUATIONS

where B is a suitably chosen number, and

pz + qw + F(z, w)<G(z, w). (11.1.9)
We can take
n
B min|n -q\ = 1, (11.1.10)

G(z,w) = m(i-£) '(l-j) -M. (11.1.11)

Here 0 < a < tfi, 0 < b < K2, and M > max (Bb, M0), where

Mo = max \cik\a'bk, Ci0 = p, c0i = q. (11.1.12)

Equation (11.1.8) is now a quadratic in Y, and the root which is zero for
z = 0 is

y(z)=-^-r1+(a-z) La"lM^B)2J 1' (1LL13)

where the roots are real positive for 2=0. This function is holomorphic
for

Let us now set

Y(2)= n2= 1 A„2" (11.1.15)

and determine the A's so that (11.1.8) is satisfied by the series (11.1.15)
identically. We have

BAn = M„(CJk;A1,...,An_1). (11.1.16)

Here Mn is the same multinomial in the Cjk, and the Am 's and all numerical
coefficients are positive. We have \cjk\^ Cjk for all j and k. Since

BAi = Co, BA2 = CM+ C„A, + C02(A,)2,

we have clearly \ai\ ^ Au \a2\ ^ A2, and by complete induction
|fl„|^A„, Vn. (11.1.17)
It follows that the series (11.1.5) converges absolutely for \z \ < a and is a
solution of (1 1.1.2), holomorphic at the origin. Since the coefficients a, are
uniquely determined, this is the only solution holomorphic at the
origin. ■
It is not, however, the only solution that is zero for 2=0. This is shown
by the simple example

zw' = pz + qw, 0<q. (11.1.18)

 405
SOME BRIOT-BOUQUET EQUATIONS

Here the general solution is

-— 3-rz+C2<} ifq^l,
Q + 1pz log z
Cz if q = 1 .
If q = 1 , p ^ 0, there is no solution holomorphic at z = 0, whereas if q is a
positive integer > 1 , there are infinitely many solutions holomorphic at the
origin and vanishing there.
We shall now examine the excluded case of q a positive integer and
F(z, w)#0. We start with q = 1.

THEOREM 11.1.2

The equation
zw' = pz + w + F(z, w) (11.1.19)
has no solution holomorphic at z = 0 if p ^ 0.

Proof. If w(z) = 2o a^z' is a holomorphic solution, it must vanish for

z = 0, so that

zw'(z)- w(z) = 2 (J - OO^ = 0(z2).

j=2
Similarly F(z, w) = 0(z2), so we must have p = 0 for the existence of a
holomorphic solution. As we have seen above, there may exist
nonholomorphic solutions which vanish at z = 0. ■
We still have the problem of finding the general analytic, nonholomor-
phic solution at the origin. The special cases considered above make it
plausible that the general solution admits of a representation in terms of
psi series (see Section 7.1), a double series in z and zq if q is not a
positive integer and a logarithmic psi series if q is a positive integer. The
existence of such expansions was proved in 1921 by Johannes Malmquist,
whose work on the Riccati equation occupied us in Section 4.6. We state
the result without proof.

THEOREM 11.1.3

In a neighborhood of z = 0 the general solution of (11.1.2) is given by a

convergent psi series of the form

w(z) = 2 2 amncnzm+nq (11.1.20)

if q is not a positive integer and by

(11.1.21)
w(z) = 2 2amnZm+nq(c + h/og z)n
406 FIRST ORDER NONLINEAR EQUATIONS

in the integral case. Here c is an arbitrary constant and h is a fixed

constant.

Although we shall not prove this theorem (see Problem 14 below,

however), we shall have more to say about the existence of holomorphic
solutions in exceptional cases. We start with

THEOREM 11.1.4

The equation
zw' = w + F(z,w) (11.1.22)
has infinitely many solutions holomorphic at the origin. Here w'(0) is
arbitrary.
Proof. We assume the existence of a solution given by the power series
(11.1.5). The coefficients can be determined from (11.1.7), but since q = 1
the leading coefficient ax is arbitrary. Once it has been chosen, all other
coefficients are uniquely determined. The convergence proof carries over,
and (11.1.10) shows that B = 1. The majorant requires M to be chosen so
large that a\ai\^M. The rest of the proof carries over without
change, fli
If q is a positive integer > 1 , we can make a change of dependent
variable so as to get an equation of the same type but with a lower value
of q. We set in (11.1.2) \-q
w=t^- + zv. (11.1.23)
The result is

z V = (q- \)zv + f[z, (y^- + i>)z].

Here the F-term is divisible by z (actually by z2), so we obtain

zv' = p,z + (q - l)v + F,(z, v), (11.1.24)
an equation of the same form as (11.1.2) but with q replaced by q - 1. If
q > 2, we repeat this type of substitution and after q - 1 steps we have an
equation with q = 1. Here there are two possibilities.
1. In the final equation the coefficient pq \ = 0. In this case Theorem
11.1.4 applies and yields infinitely many solutions holomorphic at the
origin. This gives, of course, infinitely many holomorphic solutions for the
original equation (11.1.2). This case will arise iff the final coefficient
= 0 — something that is not very likely to happen.
2. In the final equation pq-i ^ 0. There are now no solutions which are
holomorphic at the origin.
SOME BRIOT-BOUQUET EQUATIONS 407

To round off the discussion, let us consider a very simple equation of

the Briot-Bouquet type with an abominable solution. We take

z2w' =pz+qw, q^O. (11.1.25)

The solution is elementary:

.(z) = exp(-f)[p/exp(f)f+c].
If p = 0, the solutions have an essential singularity at the origin. If p ^ 0, a
logarithm is present and the solution takes the form

W(2) = exp(-f)[P(.og2-^-^-..) + c].

Thus there are no solutions which are holomorphic in a domain containing
the origin.

EXERCISE 11.1
1. Is there any value of p for which the equation

zw' = pz + 3w + w2
has solutions holomorphic at the origin?
2. In utilizing the lower bound (11.1.14) for the radius of convergence of the
series (11.1.5), the choice of a and b is usually quite arbitrary. In the case
where the right member of (11.1.2) is a polynomial in z and w, this is
particularly the case and a problem of optimization presents itself. Try to find
a reasonable bound for the equation zw' =\{z + w) + (z + w)2.
3. The equation zw' = aw + z", where n is a positive integer and a ^ n, has a
unique solution holomorphic at the origin. Find this solution explicitly. What
happens when a-*n!
4. Show that, if w(z) is an entire function satisfying
w' = P(w),

where P(w) is a polynomial in w of degree >2 having constant coefficients,

w(z) is a constant. Specify the constants which can give solutions.

In the following we consider some nonlinear matrix DE's, mostly generalized

Riccati equations:
<3/'(z) = si{z) + mzWz) + ^{z)%{z Mz),
and the associate:
T'(z) = <€(z) + T(z)9&(z) + Y(z)j4(z)Y(z).
Here all matrices shall be elements of the algebra 2ft„ of n-by-n matrices, and si,
% shall be holomorphic except for poles in the finite plane which form a finite
set S of fixed singularities of the equations.
408 FIRST ORDER NONLINEAR EQUATIONS

5. Show that, if °U(z) has an inverse in the domain of consideration, T(z) =

-[^(z)]-1 is a solution of the associate equation.
6. The substitution °H(z) = -[^(z)Y'°W'(z)[<W(z)Yx carries the first Riccati
equation into a linear second order matrix equation for °W(z). Find this DE.
7. The movable singularities of ^(z) are poles located at the points where °lV(z)
is singular, and similarly for T(z), the poles of which occur where W(z) is
singular, ^(z) and T(z) may have poles in common. There is no limitation on
the order of the poles. Try to verify these assertions.
8. Show that the equation <2/'(z) = - [^(z)]2 is satisfied by all resolvents
$!(z, si) = [zJ> - jtf ]-1 wherever the inverse exists. Here si is any constant
matrix in Wln, regular or singular.
9. If 2& is any regular matrix in 3Jln, find the solution of the equation in Problem
8 which takes the value ^ at z = z0.
10. Show that the movable singularities of the equation in Problem 8 are poles
and the sum of their orders is n.
11. If is an idempotent matrix, a a nilpotent one, 9>2 = 9>, 9>a = £<3> = 2,, a" = 0,
show that a solution is given by
^z_1 + az-2 + - • • + &n~VB.
12. The equation (k - l)°U/(z) = - [°U(z)]\ k> 1, may be reduced to the form of
the equation in Problem 8 by setting °llk~1 = What is now the nature of the
movable singularities?
13. With the notation of Problem 11, but with n replaced by m n and z by
z-a, the resulting matrix 2ft(z) is a solution of Problem 8. Find the
expression for [2/Uz)]1/2. What DE does it satisfy?
14. If q in (1 1.1.2) is neither a positive integer nor a real negative number and if
y(z) is the holomorphic solution at z = 0, then the solution (11.1.20) is
obtainable by setting w(z) = y(z) + zqv(z). If

and zv' = vJJ'Zamnzm+nqvn (*)

^Z^\amn\am+nQbn =M<«>, |<j|<Q, integer,

then (*) has a solution d(z) = 1 2 cmnzm+nq where the cmn are uniquely
determined once Coo has been chosen and the series for v (z) has as a majorant
the solution V(z) with V(0) = Coo of the algebraic equation

H( V - Coo) = MV{(\ - zla) \\ - zQVl{aQb)V ~ !}•

Here H is suitably chosen and Coo> |coo|- Verify!

11.2. GROWTH PROPERTIES

We shall now consider the equation
GROWTH PROPERTIES 409

with

P(z, w) = j=0
2 P,(z)w', Q(z, w) = k=0
2 Qk(z)wk, (11.2.2)
where P, and Qk are polynomials in z. This is the equation which figured
in Chapters 3 and 4. The fixed singularities of the equation are the zeros of
PP(z) and Qq(z). These are finite in number, and R is so chosen that all
the finite fixed singularities will lie in the disk |z| < R. The nature of the
movable singularities will depend on the relation between the orders p
and q. We start with the case

p=<Z + l. (11.2.3)
The discussion in Section 4.6 [see in particular (4.6.7) and (4.6.8)] shows
that there can be no movable infinitudes. On the other hand, there may be
movable branch points where the solution approaches a finite limit.
Dividing P(z, w) by Q(z, w), we obtain an expression of the form

/ Pq + l(z) i O / \ , "S*l(Z, W) /11 *\ A\

w =-qMw+So(z)+q(^~y (lL2-4)
where S0(z) is a rational function of z, and Si(z, w) is a polynomial in z
and w of degree at most q — I in w. In the following it will be assumed that
the degree of Pq+i is at least equal to that of Qq. We let 2*0 be the
difference in the degrees and set

E(z) = j E0(s)ds =f-f^i + '''> (H.2.6)

where
Pq+l(z) = azm + • • • , Qq(z) = bzn + • • • . (H.2.7)
Here E0(z) is clearly a rational function, and all its poles are inside the
circle |z| = R. We may assume R so large that the arcs
Re[E(z)] = 0 (H.2.8)

consist of 2pt 4- 2 separate arcs which proceed from the circle |z| = R to
infinity with no finite intersections. The arcs divide the z-plane beyond the
circle into 2/ll +2 sectors Sh where Re [E(z)] keeps a constant sign. We
number these cyclically Si, S2, . . . , S2m+2, where Re [E(z)] is positive or
negative in Sh according as j is even or odd.
We have now

^-w-^+-rSr1i-
w(z) vv(z) w(z)Q(z, w) di.2.9)

Let z = z0 be a point in S2j for some j. Further restrictions will be

410 FIRST ORDER NONLINEAR EQUATIONS

imposed on z0 when they become meaningful. Let g be a positive number

which will also be restricted later. Let S*2j be the set
Sli = {z;sG S2i, |w(z)| > \z\8}. (11.2.10)
We shall see later that S~2j is not void; in fact, it occupies the greater part
of S2i.
We now proceed to an estimate of the right member of (11.2.9),
assuming that z GS*2j. For \z\> Ri> R,
So(z) = c0zkl[l + o(l)], (11.2.11)
so that

j£g(<izr, (n.2.12)
where a > max (0, ki + 1).
The second fraction on the right in (1 1.2.9) is more intricate. The degree
of 5i(z, w) as a polynomial in w is at most q - 1; to fix the ideas, suppose
that it is exactly q - 1 and that

and also S,(z, w) = F0(z) + F,(z)w + • • • + F,_1(z)wq"1

max deg [Fj] = k2.
Then for large values of z, say for |z| > Ru

R|^L<k(2)n2r<2- (n.2.13)
where (3 > k2.
We now choose g so that

g>a+2, 2g>/3+4, (11.2.14)

and z0GS*2h The right-hand side of (11.2.9) is now less than 2|z|"2 for
z G S2i. There is then a neighborhood of z0 where (1 1.2.10) is satisfied. We
now integrate (11.2.9) along a straight line from z0 to z in S*2j. Here we
may assume that F(zo) = 0. Thus

log53-E(z)='I(z'2°)'
where the right-hand side is bounded and less than 4^0^ in absolute value
if |z|>|z0|. It follows that
w(z) = w(zo) exp [F(z)]H(z). (11.2.15)

If |z0|>4, the factor H(z) is bounded and lies between e and e 1 in

absolute value. To start with, this holds in a neighborhood of z0 where
GROWTH PROPERTIES 411

|w(z)| we
holds > |z|8.
mustOnhave
the equality.
boundary But
of the subsetalso
we have of equality
S*2j wherein this inequality
(11.2.15). This
requires that
|w(z0)| |exp [E(z)]\\H(z)\ = \z\g.
Here w(z0) is a fixed number, and |H(z)| lies between e and e~l. It follows
that the boundary of S*2i must be asymptotic to the boundary of S2j. Since
w(z) becomes infinite as an exponential function when z recedes to
infinity in the interior of the even-numbered sectors, Boutroux refers to
this case as one of exponential growth. The analysis gives no information
about what happens in the odd-numbered sectors.
Here we have supposed that pi = m - n 2* 0. If this is not the case, the
solutions may very well be bounded at infinity. An example is given by
(11.1.25) with p = 0, i.e.,

z2w' = qw with w(z) = C exp ^-

Even if p # 0 in (1 1.1.25), the logarithmic rate of growth is not impressive.
We state the proved results as

THEOREM 11.2.1

If P = Q + 1 and if = deg Pq+i - deg Pq 2= 0, then all solutions of (11.2.1)

are of exponential growth in the even -numbered sectors S2):
w(z) = exp [E(z)]H(z, z0), (11.2.16)
where H(z, z0) is bounded away from zero and infinity, provided that

\exp [E(z)]| > C|z|g. (11.2.17)

If /lc < 0, the solutions may be bounded in the neighborhood of infinity.
We now turn to the case in which p ^ q + 1 . There are actually two
cases according as p < q + 1 or p > q + 1. In the first instance there are no
mobile infinitudes, while in the second there are such. See Section 4.6,
where the case p = q + 2 is treated.
We start with

Case 1: q = p + m, m > 1.
Here

where

j=0 \*p/ k=0 \Wp+m/

412 FIRST ORDER NONLINEAR EQUATIONS

We set

R0(z) = nPpU,L R(z)= f Ro(s)ds. (11.2.19)

Here P0(z) is a rational function of z, all the poles of which are inside the
circle |z| = i?. If necessary, we make R(z) single valued by radial cuts
from the poles to infinity. We suppose that

degPP -degQp+m = p 3=0. (11.2.20)

To simplify matters still further, we restrict ourselves to the case in which
deg P, ^ deg PP, deg Qk ^ deg QP+m (11.2.21)
for all j, k under consideration.
Equation (11.2.18) may now be written as

wmw'-R0(z) = R0(z) l^y. (11.2.22)

Let g be a positive integer to be disposed of later, and consider the set

Let & = {z,R <|z|>(z)|>|z|8}. (11.2.23)

[\PP(z)y \QP+m(z)\) k =0, 1, . . . ,p + m - 1

and z be restricted to |z|>2P. Since the ratios tend to constant limits
when z it is seen that M is finite. Hence for z G ©

\S - T\ < 2M[\z\~8 + \z\-28 + • • • + |z|-(p+™>«]

< 2M|z T8 [1 - |z T8 ]_1 < 2M(|z |8 - 1)~ 1 < 4M|z |~8
if |z|8 >2. Furthermore,

|1 + T| 5* 1 - \T\ > 1 - 2M[|z|8 - I]-1 >i

if \z\8 >4M+ 1. It follows that the right member of (11.2.22) for z in ©
does not exceed

$M\Ro(z)\\z\~g. (11.2.25)
Suppose that z0 E © and z is a point of © visible from z0. We integrate
(11.2.22) along the line segment from z0 to z and obtain

m 1+ T{[w(z)r+,-[w(zo)r+,} = o[|zr+i-8] 01.2.26)

so that

w(z) = 0[\z\^+l-8),(m+l)). (11.2.27)

GROWTH PROPERTIES 413

This estimate will contradict our basic hypothesis (11.2.23) unless

m + 1

Here we can take equality so that g = (pt + l)/(m +2) and

H>(z)=0[|z|(M+1)/(m+1)]. (11.2.28)
This estimate holds in ©. Outside © the inequality (11.2.23) is reversed.
Thus we have proved

THEOREM 11.2.2

If Q = P + m, 1 < m, and i/fhe degrees of the polynomials Pj and Qk satisfy

the conditions (11.2.20) and (11.2.21), rften as z becomes infinite the
solutions satisfy
|w(z)|<C|zr+1)/(m+,). (11.2.29)
Boutroux refers to this case as one of rational growth. We shall say that
the equation is of the power type.
We turn now to

Case 2: p = q + ra, Km.

Here we have movable branch points where the solutions become
infinite. By (4.6.9) at such a point

w(z) = a(z - z0r1/(m_1)[l + o(l)], (11.2.30)

where a depends on ra, Pq+m(z0), and Qq(z0). We divide P(z, w) by
Q (z, w ) and obtain

w' = A0(z)wm + • • • + Am(z) + ~^2-^. (11.2.31)

Q(z, w)
Here the A 's are rational functions of z, and Pi(z, w) is a polynomial in z
and u> of degree at most q - 1 in w. Compare Section 4.6, where m =2.
Since w(z) is not single valued, we have to apply a system of radial cuts
from the fixed singularities and from the branch points of the solution.
The cuts extend to infinity.
We have now

(w) L Aow A0(w) A0(w)Q(z, w)J

and proceed as in Section 4.6, the object being to make the right-hand side
of (11.2.32) into an expression of the form A0[l + h(z)]y where \h(z)\ <! in
a subset of the distant part of the cut plane. Let g be a constant to be
414 FIRST ORDER NONLINEAR EQUATIONS

determined, and set

S = [z;R <\z\,\w(z)\>\z\8]. (11.2.33)
This set S obviously contains the infinitudes in the domain under
consideration. To find the desirable limitations on g we proceed as in
earlier discussions of this nature. We can find numbers ai9 a2, . . . , am+i
such that for z E S

—lAL^i^-* j = 1,2,. ..,m, (11.2.34)

\Pi(z,w)\
\w\m\A0\\Q(z,w)\
It is now assumed that g is so large that (i) all the exponents in the right
members of these inequalities are negative, and (ii)

Rarjg <2(m 1+ l)
Then for z E S

^L
[w(z)J = Ao(z)[\+h(z)], (11.2.35)

where |/i(z)|<i Here we assume that

Ao(z) = = Az»[\ + o(l)], il > - 1, (11.2.36)

so that

£
A0(a)ds =^T(z^,-Zo^1)[l + o(l)].
Thus for z0 and z in S, z visible from z0 in S, we get

Km-l)|A(z)-A(zo)|<|[w(z)]1-m-[w(zo)],"m|
<!(m -l)|A(z)-A(z0)|. (11.2.37)

In particular, if z0 is an infinitude of w(z), the term [w(z0)]1_m drops out,

and we have an inequality for [w(z)],~m.
We can now introduce infinitudinal neighborhoods of the form

Un = [z; \z-zn\< K|z„rm-,)g-a (11.2.38)

It is possible to choose constants K and R0 so that these neighborhoods
do not overlap. The argument is analogous to that given in Section 4.6 for
m = 2 but considerably more complicated. The only case of interest is
that in which there are infinitudes clustering at infinity. We can then find a
point z0 where
iw(z0)i>iz0r.
 415
GROWTH PROPERTIES

If all the points on the circle |z -z0| = |z0|_g satisfy the inequality |w(z)| >
|z|g, we may assume that the inequality holds also inside the circle. This
means that the disk |z - z0| < \z0\~g belongs to S. If this is not the case, we
can find a concentric circle with a smaller radius passing through a point
z = t such that \w(z)\ > |z|g in the smaller disk and \w(t)\ = |r|g. Then the
smaller disk belongs to S, and we can use (11.2.37) with z = t. Here

\A(t) - A (zo)| < 2|A |[|z0| + \t - Zo\T\t - z0|

<2|A|[|zoMzorr|f-z0|
= 2|A||zor[l + |zorg-1r|r-z0|.
From (11.2.37) we then get

Zo
ir-zoi>(3|A|)-iizor[i + iz0rg-,r
t

Here the factor [1 + |z0|~s_I]~M exceeds \ when |z0| exceeds some number
depending on g and fx. The same is true for the last factor, since the

| maximum
can find an ofR0\tlz0\ = 1 + |z0|"g
depending on g,and
ra, |zor('"_1)g is small.
fx such that This
for |z0| > JRomeans that will
(11.2.38) we

hold with K >[12|A|(m - 1)]_1. If K and R0 have been chosen in this

manner, then Un, the infinitudinal neighborhood of the infinitude at z„,
belongs to S and by (11.2.37) these neighborhoods cannot overlap. We
have thus proved ■

THEOREM 11.2.3

If P = Q + m> 1 < m, and if deg Pq+m - deg Qq = n > - 1, then the solutions
of (11.2.1) have (movable) infinitudes where w(z) becomes infinite as
indicated by (11.2.30). With each infinitude z = zn there is an infinitudinal
neighborhood given by (11.2.38). These neighborhoods do not overlap in
the same sheet of the Riemann surface on which the solution is single
valued. The diameter of Un goes to zero as n-»o°. Outside these neighbor-
hoods w(z) satisfies an inequality of the form |w(z)| < |z|8. The infinitudes
are branch points where m - 1 branches commute.
This is again a case of equations of the power type, in the modified
terminology of Boutroux.

EXERCISE 11.2

1. For the equation w' = z3-w3 we have p = 3, q = 0. Show that the finite
singularities are algebraic branch points where (z - z0)1/2w>(z)-» ±2'm as
416 FIRST ORDER NONLINEAR EQUATIONS

2. Show that the equation in Problem 1 has a solution which tends to +<» when
z = x 1 0. Show that this solution is positive for jc > 0, has a single positive
minimum, and becomes infinite with x.
3. More precisely, show that w(x) = x -\x~2 + 0(x 5).
4. Find a suitable g for this equation, and estimate K and R0 in (11.2.38).
5. Try to estimate the rate of growth of \w(z)\ as z->o°, avoiding the branch
points. Problem 3 suggests 0(|z|) for the growth. Is this right or wrong?
6. Another equation of Boutroux's power type is w' = z2w - w\ This is a
Bernoulli equation and hence reducible to a linear equation and solvable by
elementary methods. The solution with w(0)= 1 is

>v(z) = exp(k3)[l+2 Jo exp(fs3)ds]

Here the branch points are given by the zeros of the second factor. Define six
sectors, Sk = [z ; \{2k - 3)tt < arg z < \{2k - 1)tt], k = 1, 2, . . . , 6. Show that
the branch points lie in Si, 53, S5 and that the sequences of the infinitudes
approach the boundary lines of the sectors. Show that w(z) goes exponen-
tially to zero in S2, S4, S6 and w(z)/z->l in the interior of the remaining
sectors, omitting small neighborhoods of the branch points.

Next we present two DE's with elementary integrals which are of the exponential
type.
7. Take

with w' ,= — 2+l

zw2w
rz -.1/2
w(z) = exp (iz2) I expHs2)ds +C| .
Verify. The lines y = ±x divide the plane into four sectors. Determine the
behavior of the solutions in each sector. Where are the branch points?
Determine their nature, and find roughly how many there are in the disk |z| < r
as r-»oo.
8. Take

W =z~W +wZ+z+Z With vv(2) = 2[-1±(21°gCz + 1)1/2]-

Verify. Determine the behavior of the solutions for large values of z. Where
are the branch points, and what is their nature?

11.3. BINOMIAL BRIOT-BOUQUET EQUATIONS OF ELLIPTIC

FUNCTION THEORY
We consider equations of the form

(wT =P(w), (11.3.1)

where P(w) is a polynomial in w of degree m and with constant
BINOMIAL BRIOT-BOUQUET EQUATIONS 417

coefficients. It is assumed that P(w) is not the fcth power of a polynomial

where k is a divisor of m. We want to determine the equations of this type
which have a single-valued solution and, in particular, the equations which
are satisfied by elliptic functions (= doubly periodic meromorphic func-
tions). There are a number of trivial cases to put aside.
For n = 1 we know that the equation must be a Riccati equation. There
are essentially three subcases:

1:1. P(w) = A (w - a) with the solution

w(z) = a +(w0-a)exp[A(z -z0)], (11.3.2)
which is single valued and periodic and is an entire function.
1:2. P(w) = A(w - a)(w - b), a^b, with a solution which is periodic
and meromorphic but not elliptic.
1:3. P(w) = A(w - a)2, with a solution which is rational with a single
pole.
For n = 2 we have a greater variety and even encounter elliptic
functions. By Yosida's extension of Malmquist's theorem, m is at most 4
if the solution is to be single valued and meromorphic. For m = 1, the
solution is a quadratic polynomial. For m =2, P(w) = A(w — a)(w — b),
a^b, the solution is simply periodic and meromorphic. For m =3, we
can discard the DE's
(w'f=w3 and (w')2 = (w - a)(w - bf. (11.3.3)

The first solution

has the one is satisfied by a rational function 4(z - C)~2, while the second

w(z) = b +{exp [(b - a)1/2(z - z0)] - 1 + (w0- by1'2}'2-

If the cubic polynomial P(w) has three distinct roots, a linear transforma-
tion on z and on w will bring the equation into the form

(w')2 = 4(u> -e,)(w -e2)(w - e3), ei + e2 + e3 = 0, (11.3.4)

the general solution of which is p(z - z0; g2, g3), the Weierstrass p-
function where g2 and g3 are symmetric functions of the roots eh In all
three cases there are double poles.
We can also have m =4; the poles are now simple. Here the case
P(w) = (w - a)(w - bf is trivial, a rational function of z with a single
pole. The case with two double roots reduces to subcase 1:2. If P(w) has
four distinct roots, we can use a linear transformation to reduce the
equation to the Jacobi normal form:

(w')2 = (1 - w2)(l - k2w2), (11.3.5)

where the modulus k^±\. The solution is the Jacobi sine amplitude,
418 FIRST ORDER NONLINEAR EQUATIONS

sn (z - z0; k) in the Gudermann notation. Christopher Gudermann (1798—

1851), among other distinctions, was the teacher of Weierstrass.
Actually Yosida's extension of Malmquist's theorem is not needed to
prove that m =s 4 when n =2. This follows from the test-power principle
of Section 3.3. See (4.6.47), which here becomes

(p -2)a = 2, (11.3.6)
a a positive integer. If a = 1, then p = 4, and if a = 2, then p = 3; these
are the only possibilities, in agreement with the statements made above.
We can continue in this manner to test the various possibilities for the
nonexistence of mobile branch points, but it is helpful to derive some
general results. We start with a binomial equation
P(w)
(wy=R(w), R(w) = ^^y (11.3.7)
where P and Q are polynomials in w without a common factor, P of
degree p and Q of degree q. Here we find

THEOREM 11.3.1

A necessary condition in order for (11.3.7) to have no mobile branch

points is that R(w) = P(w), a polynomial of degree «s 2n.
Proof. If this is not true, Q(w) will admit at least one factor, say
Q(w) = (w - c)kQi(w), where Qi(c) # 0 and k is a positive integer. For
small values of \w - c\ we have then

w' = (w -cyk/n^ai(w -cy

or
j=0
^Zbjiw-cy^'-w' = \
and

whence by inversion

w-c = ic,(z-2or"l+"1.
Thus, if Q(z) is not identically i =one,
i there are mobile branch points.
If Q(w) = 1, we make the inversion w = \/v and obtain

(-inv'T =v2nP^.
BINOMIAL BRIOT-BOUQUET EQUATIONS 419

Here the right member must be a polynomial in v, to avoid branch points,

and this is the case iff deg(P)=s2n. ■
Furthermore, the factors of P(w) are subject to severe restrictions.
Suppose that (w - a)k is a factor of P(w) so that

P(w) = {w- a)kP,(w), P,(a) * 0.

Then for small values of |w - a\
CO
(w - a)' = (w -a)k/n2 _a)J> Co 5^0.

Here we set w — a = u " to obtain

nun lu' = 2, CjUk+in. (11.3.8)

There are three different possibilities. The first case is k = n — 1, and

then

MM'(2) = Co +2 Cjll* Co 5*0,

so that j= i

u (z ) = 2 fr, (z - Zo)j, w (z ) = a + [ w (z )]".

i= i
This function is holomorphic in a neighborhood of z = z0 where w (z0) = a.
The second case is k ^ n, and then
MM
,)"
'(z) = EQ"',+'+U"
1=0 u(z) = 0, so that w(z) = a.
with the unique holomorphic solution
Finally, suppose that k < n - 1. Then
boM"~k-1 + fr,w2',"'c-, + --- = z'(M),
so that

Z - Zo
= „-<(J^+-A-
\n - fc 2m -„» + ...)
/
or
(z-Zo)1/("-k)=M(do+d,Mn +•••)•
This gives

«(z) = d„-(2 - Zo)lK-"\l + 2 ef(z - Zo^"""0],

'"' . , (11.3.9)
w(z) = a + do~"(z - 2„)""""|1 + 2 /,(z - Zo)'"'*"""].
420 FIRST ORDER NONLINEAR EQUATIONS

This solution has a branch point at z = z0 unless

n

an integer 2*2. This gives

k =VJz^n^\n, (11.3.10)
P
where p is a divisor of n. Thus we have proved ■

THEOREM 11.3.2

A factor (w — a)k leads to solutions with movable branch points unless

either k 5? m - 1 or k = [(p - l)/p]n, where p ^ 2 is a divisor of n.
Suppose now that the degree of P(w) is exactly 2m, and let (w - af be
a factor of P(w) with Jc ^ n. The equation P(w) = 0 can have only one
root distinct from a, for if there were two other roots, then

P(w) = (w - a)k(w - b)V - c)m.

Here k ^ n, and by (11.3.10) both / and m must be at least \n, so the
degree of P(w) would exceed 2m. This would lead to movable branch
points and hence is not admissible. Therefore I =2n - k and m = 0.
There are two possibilities: (i) / = n - 1, and (ii) I =[(p - l)/p]n, where
p ^ 2 is a divisor of n. The second alternative leads to a contradiction, for
we would have
n _ n

n-(n-k)~ k9
an integer > 1, and this is impossible since k ^ n.
The resulting equation is thus

(w')n =(w -a)n+l(w -b)n \ (11.3.11)

The general solution of this equation is a rational function of z of degree n
and with simple poles. This follows from setting

w-a=^, (11.3.12)
leading to the equation

(W')n =(-l)B[l+(a -b)Wf-\ (11.3.13)

which has polynomial solutions. If in (13.3.12) we replace a by b, we
obtain an equation of the form

(W')n = A(W-c)n+\ (11.3.14)

which also has polynomial solutions.
BINOMIAL BRIOT-BOUQUET EQUATIONS 421

We can take k = n, but this offers very few admissible equations. In this
case with (w - a)n as a factor, the other factors of P(w) are either
(w-b)n or (w -b)nl\w - c)n'\
The first choice works iff n = 1 and the corresponding equation is subcase
1:2. Larger values of n are ruled out, for the corresponding equations are
reducible. The second choice is admissible iff n = 2; this leads to the
equation
(W')2 = (w -a)(w -b), (11.3.15)
which was mentioned above. The solutions are simply periodic entire
functions.
We have left the case in which the exponents are of the form (11.3.10).
Here the problem is to find positive fractions of the form j/n which add
up to 2. There are only four possibilities:

2,2,2,2, M — Z, 3,3,3, It— 3,

4,4,2, n = 4\ 6,3,2, n — 5.
The corresponding equations are as follows:

(w'f = C(w - a)(w - b)(w - c)(w - d), (11.3.16)

(w'f = C(w - af(w -bf(w - cf (11.3.17)
(w')4 = C(w - a)\w - b)\w - c)\ (11.3.18)
(w'f = C(w - a)\w - b)\w - c)\ (11.3.19)
This exhausts the list of binomial equations of the form (w')n = P(w),
where P is a polynomial of degree 2n and the solution has no branch
points and is hence single valued. We shall discuss some of these
equations below in more detail.
First let us observe, however, that there are also admissible binomial
equations where the degree m of P(w) is less than In. We have seven
admissible equations at the outset, namely, subcase 1:2, i.e.,

w' = C(w -a)(w -b), (11.3.20)

(11.3.11), and (11 .3.15)— (1 1.3.19). From each of these seven equations we
can get more admissible equations by suppressing one of the factors in the
right member. The suppression is achieved with the aid of the transforma-
tion (11.3.12), the power of which we have already seen in carrying
(11.3.11) into (11.3.13). We take a more general case. Suppose that we are
given

(wT =C0(w -a)J(w -b)k(w -c)', j 4- k + I = 2n, (11.3.21)

and make the change of dependent variable defined by (11.3.12). The
422 FIRST ORDER NONLINEAR EQUATIONS

result is
(W'T =K(W -B)k(W -C)\ (11.3.22)
where

K = (- l)"c0(a - bf{a - c)', B = (b - a)"1, C = (c- a)\

This proves the assertion.
Of the seven original equations only four, (1 1.3. 16)— (1 1.3. 19), lead to
elliptic functions. So do the reduced equations. Quite generally the
equations are formally invariant under fractional linear transformations
on w, and this can be used to bring the equations into a desirable normal
form. For (11.3.16) we have already shown the Jacobi normal form
(11.3.5). In this setting the Weierstrass normal form (11.3.4) can be
regarded as a reduced form of (11.3.16) followed by a translation
XV > W + D to ensure that the sum of the roots of the cubic polynomial
is zero. The factor 4 can be introduced by a dilation of the independent
variable z »->• kz.
All the equations (1 1.3.16)— (1 1.3. 19) have solutions which are rational
functions of a suitably chosen p-function and its derivative. We shall
study (11.3.17) and (11.3.19) from this point of view. We start with

(w'f = (w - a)\w - b)\w - c)\ (11.3.23)

where a, b, c are distinct complex numbers. We remove the third factor
on the right by setting w - c = 1/ W. The result is
(W'f = K(W -A)2(W -Bf (1 1.3.24)
with

K=-(c -a)\c - b)\ A = (a - c) \ B = (b - c)~\

The solutions of this equation have triple poles, and so has the derivative
of the p-function. Our next step calls for the differential equation satisfied
by p'. Now differentiating
(p')2 = 4p3-g2p-g3,
we obtain, after cancellation,

p" = 6p2-|g2. (11.3.25)

The case in which g2 = 0 is known as equiharmonic. Here the roots
eue2, e3 form the vertices of an equilateral triangle, and so do the points
0, 2ft>i,2w3, where 2wi,2w3 form a pair of primitive periods of p. In this
case we have

P" = 6p2, (p')2 = 4p3-g3. (11.3.26)

We eliminate p between these two equations and obtain
BINOMIAL BRIOT-BOUQUET EQUATIONS 423

(P")3 = ¥[(P')2 + g3]2. (11.3.27)

This shows that p'(z;0, g3) satisfies the equation
{v'f = 2i{v2 + g,)\ (11.3.28)
This can be brought to the form (11.3.24) by a linear transformation of
both variables. The result is that

mp'(z+/c:0,g3) + KA +J8) with m m2g3 = -i(A - Bf

is a solution of (11.3.24), and to get the solution of (11.3.23) we have
simply to note that w = c + 1/ W, which is clearly also an elliptic function
but with simple poles.
We now consider (11.3.19), which we first reduce to the form

(W'f = K(W - B)\ W - Cf. (11.3.29)

Here we set

and get W = B + U3

36Ul\U'f = KUl2(U* + B- C)\

We cancel the factor U 12 and extract a cube root to obtain

(U')2 = lKm{U' + B -C), (11.3.30)

which is satisfied by an elliptic function of the type ap(/3z + y, 0, g3). This
shows that all nonconstant solutions of (11.3.19) are fractional linear
transforms of p3(/3z + g ; 0, g3).
Equation (11.3.18) is integrable in terms of the Jacobi sine amplitude.
Details are left to the reader.
We end this discussion by showing that every elliptic function satisfies
a Briot-Bouquet equation. This is a by-product of a result, also proved by
Briot and Bouquet, to the effect that any two elliptic functions with the
same periods are algebraic functions of each other. In other words, if /(z)
and g(z) are elliptic functions of periods Icoi and 2w3, there exists a
polynomial F(Zi, Z2) in two variables with constant coefficients such that

F[f(z),g(z)] = 0. (11.3.31)

Now, if w(z) is an elliptic function with periods 2o)i,2o>3, so is w'(z).

Hence we have a DE

F[w(z), w'(z)] = 0. (11.3.32)

This equation has a somewhat special structure.
424 FIRST ORDER NONLINEAR EQUATIONS

THEOREM 11.3.3

The differential equation satisfied by an elliptic function of order m is of

the form

(w')m + 2 PJ(w)(w')mH = 0, (1 1.3.33)

where Pj(w) is a polynomial in j=>w with constant coefficients and of degree
2j. Furthermore, Pm i(w) = 0.
Proof. For properties of elliptic functions used in the following see the
Appendix of this section, and for further details Chapter 13 of Vol. 2 of
AFT.
The order of an elliptic function f{z) is the number of roots of the
equation f(z) = a in the period parallelogram. It is independent of a.
Suppose now that the DE satisfied by w(z) is

2 J>j(w)(w')n"i =0. (11.3.34)

That n has to be equal to j=0m will become apparent from the proof.
Our first task is to show that P0(w) must be a constant. If not, the
equation Po(w) = 0 has at least one root, say w = a. This implies by a
well-known property of algebraic functions that the equation F(Zi, Z2) =
0 has a solution Z2 = G(Zi) which becomes infinite as Zi -> a. But this is
impossible, for in our case Zi = w(z), Z2 = w'(z) and they are cofinite,
i.e., if one is finite so is the other. This shows that P0(w) must be a
constant which we may take to be equal to one.
Next we have the degrees of the coefficient polynomials. These can be
estimated in one way or another by the infinitary behavior of the solution.
Here a difficulty arises from the fact that an elliptic function may have
poles of various degrees so that the order of the poles is not unique, a
phenomenon which we have not encountered before. That it can happen
is shown by the elliptic function

p'(z) + [p(z) - exy\ p(o)0 = eu (1 1.3.35)

which is of order 5 and has a triple pole at the origin, a double pole at
z = o)i, and congruent points modulo period. We can use the test-power
test of Section 3.3 to study this matter, but for a complete discussion we
have to invoke the method of Puisseux or a similar argument.
Suppose that z = z0 is a pole of w(z) of order a so that

w(z) ~ a(z - Zo)~a, w'(z) ~ -aa(z - z0)"a_1.

We denote the degree of P;(w) by p, and substitute in (11.3.34). It is seen
 425
BINOMIAL BRIOT-BOUQUET EQUATIONS

that we have to balance expressions of the form

bo(z - 2o)-n(,+a), •••, bAl- Z0yP^n-Dil+a\ ....

Suppose initially that the first term is the heaviest. There will then be an
integer j such that

n(\ + a) = pia + (n -j)(l + a), (11.3.36)

whence

j<Pj=^j^2j. (11.3.37)
If the terms with subscripts 0 and j are the only ones which are of
maximum order, we find that pk < 2k for all other subscripts.
The situation is less satisfactory when the first term is not of maximum
order. Then there are two integers, j and fc, such that

n(l + a)<pJa +(n - j)(l + a) = pka + (n - k)(\ + a),

and the inequalities give
a + 1 . a + 1 .
P>>—^-h pk>—^-k.
Thus, if the theorem is correct, this contingency could arise only for
multiple poles. The equality and the other inequalities do not provide
enough information to give useful upper bounds for the p's. Here we get
more help from the method of Puisseux, which was also sketched at the
end of Section 3.3.
In an (x, y)-plane we mark the points (j, k) corresponding to the terms
w'(w')k actually present in the DE. Next we plot the least convex
polygon II which encloses these points. Here II has points on the positive
axes; the point X = (p„, 0) is furthest away on the x-axis, and the point
Y = (0, n) furthest on the y-axis. We are interested in that part of II which
faces infinity. It is a broken line segment joining X and Y and lying above
or on the straight line segment (X, Y). See Figure 11.1.
Now the equation F(ZU Z2) = 0 has all its infinitary branches of the
form
Z2= CZ(,a+1)/a[l + o(l)], (11.3.38)
since at a pole w'(z) = C[w(z)](a+1)/a[l + o(l)] and w(z) has only polar
infinitudes. Here, as observed in connection with (3.3.18), the negative of
the reciprocal of the exponent is the slope of one of the boundary lines of
n. The polygon lies above or on the line XY of the equation

y-n=-£x. (11.3.39)
426 FIRST ORDER NONLINEAR EQUATIONS

Now, if the point (Ph n - j) lies above the line, it is seen that

Pn n
But n also lies below or on the line through Y parallel to the side of 11 of
largest slope. This is the line

y-n =--^
CX-0 ' I frX, (11.3.40)

where a0 = inf a, the infimum of the order of the poles of w(z). This
shows that the point (ph n - j) lies below or on the line (1 1.3.40), so that
«o . a0 + 1 .
n-—— -TP^n-j
Oto -r 1 or Pj<^-ao-j^2j,

and this now holds for all j. Note that no side of II can have slope - 1
(why?).
Next, the number of poles is m = 2 ah where the a's are the numbers
read off from the slopes - a l(a + 1) of the sides of n, starting from Y. On
the other hand, the geometry of the polygon shows that 2 a, = n, so that
n = m.
It remains to show that Pm_,(w) = 0. We have
APPENDIX: ELLIPTIC FUNCTIONS 427
" Pm(w)

equal to the sum of the reciprocals of the roots of (1 1.3.34) regarded as an

algebraic equation for w'. Thus m i

pm(w) = ftw'W
y —
where Zi, z2, . . . , zm are the points in the period parallelogram where w(z)
assumes the value w, an arbitrary complex number. Now
z;(w),

and by a basic property of elliptic functions

is independent of w and hence a constant. Its derivative with respect to w

then is zero or Pw-i(w) = 0, as asserted. This completes the proof.

Appendix

ELLIPTIC FUNCTIONS

An elliptic function is a single-valued doubly periodic function whose

only singularities in the finite plane are poles. That /(z) is doubly periodic
means the existence of two complex numbers, 2o>i and 2o>3, such that

f(z + 2a),) = /(z), f(z + 2o)3) = /(z). (A.l)

The ratio o>3/a)i = t is necessarily nonreal, and it is customary to assume
its imaginary part to be positive. If the ratio is real, the function is simply
periodic if t is a rational number, and a constant if r is irrational. There
are no triply periodic functions except constants. The numbers
2rao)i + 2na>3, m, n — 0, ± 1, ±2, . . .
are periods of /(z). The parallelogram with vertices at

0, 2o>i, 2a)2 — 2o) i -f 2a)3, 2(i)3

is known as the period parallogram of /(z). The series

2 2l2ma)1 + 2na)3|-s, (A.2)

428 FIRST ORDER NONLINEAR EQUATIONS

where the prime indicates that (ra, n) ^ (0, 0), is convergent for s > 2 and
diverges for s ^2.
It follows that the series

z "3 + 2 2 - 2mo>, - 2no>3)"3 (A.3)

is absolutely convergent for all values of z different from one of the
points 2mwi + 2nw3. It clearly admits these numbers as periods, and it has
no other finite singularities than triple poles at the periods. Thus it is an
elliptic function and, in fact, minus one half of the derivative of the
Weierstrass p-function. The latter has the representation

2 ~2 + 2 x k* - y2 - - r2]> (a.4)
where com,„ = 2maji + 2n&>3.
Thus we know that there are elliptic functions. In studying their
properties Briot and Bouquet found Cauchy's integral theorem very
useful. In applying the latter, it is convenient to assume that there are no
poles on the boundary of the period parallelogram. We have then

JLJ£g*-0. (A.5)
where the integral is taken along the boundary. The integral is zero by the
double periodicity. On the other hand, the value of the integral is the sum
of the residues of the poles inside the contour. This says that the sum of
the residues is zero. One consequence thereof is that the order of /(z)
must exceed one, the order ra being the sum of the orders of the poles.
Consideration of the integral

2m J f(z)-a
shows that the function assumes every value a in the period parallelo-
gram m times if multiple roots are counted with their proper multiplicities.
Similarly the integral

^i\^lW-adZ=^C'-^P' (A-6)
equals a period of /(z). Here the c's are the m roots of the equation
f(z) = a and the p's are the m poles of /(z). This implies that 2 c, is
independent of a, a fact used above.
We have more to say about the p-function. We expand (A.4) in powers
of z. Since
p(-z) = p(z), (A.7)
APPENDIX: ELLIPTIC FUNCTIONS 429

the expansion will involve only even powers of z:

Now p(z) = z_2 + A0 + A,z2 + A2z4+- • • .

1 1 _ 2z 3z2 4z3 5z\

7
(z — CO)\2 O) 2 — 0) 3 + (x) 4 H 0)5" ' (O 6 • * * * >
so that

Ao = 0, Ai = 3 2 2 (<«Vn) 4» A2 = 5 2 2 6'
Next, we define

so that g2 = 60 2 2 t^)"4, g3 = 140 2 2 (aw)"6, (A.8)

P(z)=p + 2^g2Z2 + 2lg3Z4+- • • ,

P'U) = -p + *>g2Z + |g3z3 + - • • .

Our object is to derive the DE satisfied by p(z). To this end we look for a
combination of p and p' which is nonsingular at z = 0. Here
[p'(z)]2-4[p(z)]3 looks promising. We have

[p'(z)]2 = p-ig2V^g3 + 0(z2),

4[p(z )]3 = p + ?g2 p + |g3 + O (z 2),

the difference of which is

- 82 p - g3 + O (z 2) = - g2p(z ) - g3 + O (z 2).
It follows that

F(z) = [p'(z)]2 - {4[p(z)]3 - g2p(z) - g3}

is not singular at z = 0. Since F(z) is obviously doubly periodic, it follows
that F(z) has no singularities and is hence a constant. Since F(0) = 0, it is
seen that F(z) = 0, and we conclude that p(z) is a solution of the DE
(vv')2 = 4>v3-g2w-g3, (A.9)
as expected.
If we set z = (oh j = 1, 2, 3, in (A. 3), it is seen that all terms cancel, so
that
p'(o>I) = p'(w2) = p'(w3) = 0.
The corresponding values of p are denoted by e., e2, e3. Their sum is zero.
It is possible to recover the periods from the values of g2 and g3 via the
430 FIRST ORDER NONLINEAR EQUATIONS

theta functions. For us, however, it is more natural to use the DE. For the
sake of simplicity suppose that the e's are real and that e3< e2< ex. We
have then

[4(5 - ex)(s - e2)(s - e,)]'1'2 ds,

w,-f " (A- 10)

<*>> = JJ-co3 [4(5 - et)(s - e2)(s - e,)Ym ds,

where the argument of the integrand is zero in the first integral and \tt in
the second. These formulas are valid for all values of ex and e3 with
e2 = - ex - e^ and define analytic functions of eu e3. We mention (but have
no need of the fact) that p(z) has an algebraic addition theorem in the
sense that/(w + v) is a rational function of f(u),f'(u),f(v), and/'(u).
The Jacobi elliptic functions can be defined as the solutions of the
system of nonlinear first order equations:

w\(z) = w2(z)w3(z), Wi(0) = 0,

w2(z) = -w3(z)wl(z), w2(0)=l, (A.ll)
w3(z) = -k2Wi(z)w2(z), w3(0)=l.
Here k is an external parameter known as the modulus. The solutions are
analytic functions of z as well as of k. The complementary modulus is
defined as
k' = +(\-k2y12. (A.12)
By elementary manipulations we obtain

[u^z)]2 + [w2(z)]2 = 1, fcV.U)]2 + [w3(z)]2 = 1. (A.13)

Combining with the DE's, it is seen that Wi(z) satisfies the DE
(w'f = (1 - w2)(l - k2w2). (A.14)
We have Wi(z) = sn (z, k), w2(z) = cn (z, k), w3(z) = dn (z, k) in the Guder-
mann notation. The periods are 4K and 2/K', where

[(1-52)(1-/C252)]1/2 ds,
(A. 15)
[(i + 52)(i + kV)r,/2 ds.

As observed earlier, they are hypergeometric functions of k2. The points

iK' and 2K + iK' are poles of sn z with residues \/k and - I Ik, respec-
tively. The poles are simple, and the sum of the residues in the period
parallelogram is zero, as it should be.
This brief survey should give the reader some idea of the properties of
elliptic functions.
LITERATURE 431

EXERCISE 11.3

1. Solve w' = A(w - a)(w - b) and (w'f = A2(w - a)(w - b).

2. Solve (w'f = (w- a)(w -bf.
3. How is the equation (w'f = (w - a)(w - b)(w - c)(w - d) reduced to the
Jacobi normal form?
4. Solve (11.3.13) and (11.3.14).
5. Solve (11.3.18).
6. Transform Jacobi's and Weierstrass's normal forms into each other.
7. Use the result of Problem 6 to express the p-function in terms of a sine
amplitude. Keep good track of the parameters.
8. Derive (11.3.28), and reduce it to the form (11.3.24).
9. If the order of an elliptic function is m, what are the bounds for the order of
the derivative? Give examples.
10. Suppose that weights are assigned to the terms of equations of type (1 1.3.24)
or higher order in such a manner that w is of weight 1, w' of weight 2, . . . ,
w(k) of weight k + 1. The weight of a product is to be the sum of the weights
of the factors; the weight of a sum, equal to the maximal weight of any term
(if there is a single dominating term, otherwise *s the maximum). Show that
Theorem 11.3.1 expresses the fact that no term can have a heavier weight
than that of the leading term (w')m. This is a special case of a theorem of Jean
Chazy (1911) for DE's with fixed critical points.
11. p(2z) is a rational function of p(z). The function [p(z)]4[p'(z)] 2 has the same
poles as p(2z), but the principal parts do not coincide. How should the
expression be modified in order to represent p(2z)?
12. Given p(z + a) + p(a), where a is not congruent to zero modulo period,
construct an expression involving p(z), p'(z), p(a), and p'(a) with the same
poles and principal parts, so that the difference is a constant and the addition
theorem results.
13. Prove the formulas (A. 13).
14. Find the DE's for w2(z) and w3(z).
15. Prove that wx{z) becomes infinite as z tends to iK' along the imaginary axis,
if 0<fc < 1.
16. Show that sn(K)= 1 and sn (IK) = 0.

LITERATURE

The author's LODE, Section 12.3, and AFT, Vol. II, Chapter 13, have a
bearing on Section 11.3.
There is an extensive literature relating to Section 11.1. We cite:
Briot, Ch. and J. CI. Bouquet. Recherches sur les proprietes des fonctions definies par des
equations differentielles. J. Ecole Imp. Polytech., 21 (1856), 133-197.
Dulac, H. Solutions d'un systeme d'equations differentielles dans le voisinage de valeurs
singulieres. Bull. Soc. Math. France, 40 (1912), 324-383.
432 FIRST ORDER NONLINEAR EQUATIONS

Points singuliers des equations differentielles. Memorial des Sci. Math., No. 61.
Gauthier-Villars, Paris, 1934.
Hukuhara, M. Sur la generalisation des theoremes de M. T. (sic !) Malmquist. /. Fac. Sci.
Univ. Tokyo, 6 (1949), 77-84.
, T. Kimura, and T. Matuda. Equations differentielles du premier ordre dans le champ
complexe. Publ. Math. Soc. Japan, 7 (1961).
Malmquist, J. Sur les points singuliers des equations differentielles. Ark Mat., Astr., Fys., 15,
Nos. 3, 27 (1921).
Sansone, G. and R. Conti. Non -linear Differential Equations (translated by A. H. Diamond).
Pergamon Press, Macmillan, New York, 1964.
Trjitzinsky, W. J. Analytic Theory of Non -linear Differential Equations. Memorial des Sci.
Math., No. 90, Gauthier-Villars, Paris, 1938.
Valiron, G. Equations fonctionnelles : Applications. 2nd ed. Masson, Paris, 1950, pp. 268-274.
For Section 11.2 the main source is Chapters 1 and 2 of:
Boutroux, P. Lecons sur les fonctions definies par les equations differentielles du premier
ordre. Gauthier-Villars, Paris, 1908.

The use of infinitudinal neighborhoods, implicit in Boutroux's work, was

strongly influenced by the work of Hans Wittich; see Section 4.6.
Section 11.3 goes back to:
Briot, Ch. and J. CI. Bouquet. Integration des equations differentielles au moyen des
fonctions elliptiques. J. Eccle Imp. Polytech., 21: 36 (1856).
12

SECOND ORDER

NONLINEAR

DIFFERENTIAL

EQUATIONS AND

SOME AUTONOMOUS

SYSTEMS

This chapter also considers an assortment of loosely related topics which

center on nonlinearity, usually of higher order than the first. Some
attention will be paid to Briot-Bouquet equations of second and higher
order. At the end of the nineteenth century Painleve determined all
nonlinear second order DE's with single-valued solutions. He found some
50 different types, all but 6 of which could be reduced to equations with
already known solutions. Six types led to new transcendental meromor-
phic functions; we shall discuss briefly the simplest of these types. Some
60 years ago Boutroux showed that the solutions of this equation are
asymptotic in a certain sense to the Weierstrass p-function. Here again
our discussion will just skim the surface. We shall take up the Emden and
the Thomas-Fermi equations, which lead to rather peculiar solutions of
the psi-series type. Finally we shall consider quadratic and higher order
polynomial autonomous systems. They may be regarded as nonlinear first
order equations, but some are closely related to the second order Emden
DE. Here work by the author and by the British mathematician Russell
Smith provides most of the material. These investigations have appeared in
the Proceedings of the Royal Society of Edinburgh ; the Society has
graciously permitted the use of this material here.
434 SECOND ORDER NONLINEAR EQUATIONS

12.1. GENERALITIES; BRIOT-BOUQUET EQUATIONS

For first order equations Painleve proved that mobile singularities are
necessarily algebraic (see Section 3.3), provided that the coefficients
satisfy certain restrictions. This property no longer holds, however, for
second or higher order equations, again as demonstrated by Painleve.
Here we can have movable essential singularities, as shown by the
equation

C,exp(z-C2) \ (12.1.1)

Here there is a movable essential singularity at z = C2. Similarly the DE

(a - l)(w')2 - aww" = 0 with w(z) = C,(z - C2)a (12.1.2)

has a movable singularity at z = C2 which is a transcendental critical point
unless a is a rational number. We can also find second order equations
with movable logarithmic branch points or movable cluster points of
poles.
The situation becomes considerably more complicated for third order
equations. Here we may have movable natural boundaries. To show this,
we may use the properties of the modular function. Consider the w-plane,
and mark the points 0, 1, <». Let S be the universal covering surface of the
w-plane punctured at these points. This surface can be mapped confor-
mally on any simply connected domain of the z-plane. In particular, we
may map S on the upper z-half -plane. Let the mapping function be z(w).
Its inverse w(z) is the modular function. Then z(w) satisfies

{z,w} = R(w), (12.1.3)

where {z, w} is the Schwarzian and R(w) is a rational function of w with
poles at w = 0 and w = 1. The Schwarzian is invariant under any
fractional linear transformation applied to the first argument. Hence for
any choice of constants a, b, c, d with ad - be = 1 the function
r7i . az(w) + b 1 A.
z~Z{w) = 77(^d (l2-L4)
is a solution of (12. 1. 3). This function maps the real axis in the z-plane on a
circle C, and the upper z-half -plane goes into the interior or the exterior of
C, depending on the choice of the constants. Since they are at our
disposal, C is an arbitrary circle in the z-plane. By Problem 10.1:9

{w,z} = -R(w)(w')\ (12.1.5)

and this is a third order nonlinear DE satisfied by the functions
 435
GENERALITIES; BRIOT-BOUQUET EQUATIONS

for any choice of the constants a, b, c, d satisfying ad - be = 1. The

domain of existence of w(Z) has the circle C as a natural boundary. This
means that the domain of existence of the solutions varies with the three
essential constants of integration.
Returning to the second order case, we start with some remarks in
connection with Section 11.3. There it was observed that, if two elliptic
functions / and g have the same periods, then one is an algebraic function
of the other. This was used to show that each elliptic function w(z)
satisfies a first order nonlinear DE because w(z) and w'(z) are coelliptic
and have the same periods. But the same evidently holds for w(z) and
w"(z) or, for that matter, for w(z) and wik\z) for any k. We sketch a
proof of an analogue of Theorem 11.3.1 for the second order case.

THEOREM 12.1.1

The second order nonlinear DE satisfied by an elliptic function w(z) of

order m is of the form

(w'T + 2 Pj(wXwT"J = 0. (12.1.7)

i= i
Here pj = deg (Pj) =s 3j and n ^ m.

Proof That the coefficient of the highest power of w" is a constant

which may be taken to be one is proved in the same manner as in the first
order case. The statements about p, can be proved in various ways. If the
theorem of Chazy is accepted (see Problem 11.3:10), then

3n 2s p, + 3(n - j) or ps =s= 3j, Vj.

We can also use the test-power test if
n (a + 2) ^ p}a + (n - j)(a + 2), Vj
with the same result.
For the inequality between m and n we need the method of Puisseux.
Let p = max p„ and let X = (p, 0), Y = (0, n). There is a least convex
polygon n that encloses all the points (j, k) for which there are terms
w'(w")k in the equation. As usual, we are interested only in that part of
the polygon which joins X and Y and faces infinity. The various sides of
this part have slopes, going from the left to the right,

Oi + 2* a2 + 2' a„+2
436 SECOND ORDER NONLINEAR EQUATIONS

and
a,<a2<- • •<<*„, (12.1.8)

by the convexity of the polygon. The a's are the different orders of the
poles in the period parallelogram of w(z). Their sum is m. The polygon n
lies on or below the line

and above or on the line

y-n = ~JX> (12.1.10)

which joins X and Y. All the terms involved in P,(w)(w")n~' give rise to
lattice points interior to II with the possible exception of (ph n - j), which
may be a vertex of II or lie on one of the sides. From the fact that the
point (phn - j) lies on or below the line (12.1.9) it follows again that
Pi ^ 3j, and now for all j.
Suppose that all poles are of the same order k^\. The part of the
polygon that faces infinity now reduces to a single straight line from Y to
X having the equation

y-n = -y^2x- (12.1.H)

It intersects the jc-axis at x = p = (k +2)n/k. This must be an integer.
Hence there are three possibilities: (i) k = 1, (ii) k = 2, (iii) k divides n.
The line passes through the lattice points

[(k+2)j,n-kjl j =0,l,...,p (12.1.12)

Here the last point must be (p, 0), so that k divides n. This makes sense as
long as n ^ k. But if k > n, we must have k = 1 or 2. Here the case k = 1,
n < k is clearly excluded, so we are left with k = 2, n = 1. This means an
equation of the form
w" = cio+aiw +a2w2 (12.1.13)
with constant a's, which is satisfied by c0+ dp(c2z) for suitable choice of
the constants. We cannot allow a third order term, since this would lead to
simple poles whereas we want double poles. Any higher power would lead
to branch points.
Let us return to the case of k =s n and k a divisor of n. Now m, the
order of w(z), equals the number p of distinct poles in the period
parallelogram multiplied by their common order k, so that m = pk. On the
other hand, the lattice points (12.1.12) with j >0 are n/k in number and
are connected with the algebraic meaning of the Puisseux diagram. The
GENERALITIES; BRIOT-BOUQUET EQUATIONS 437

equation F(ZU Z2) = 0 has n/k solutions of the form

Z2 = Z{r2),k(a, +2 axiZ7,/k), (12.1.14)

A = 1, 2, . . . , n/k, at infinity. Now there are p distinct poles in the period
parallelogram, and if

w(z) ~ A (z - z0r\ w"(z) ~ Ak(k + l)(z - ZoTk-\

we should have

a*A2/k = k(k + 1), A = [fc(fcflx+1)]fc/2-

This implies that n/k ^ p and hence n =s m. It is seen that both m and n
are divisible by k and n^k.
Here it is assumed that all poles have the same order k. Now, if the
polygon has v sides facing infinity and the enumeration is such that the
orders a, satisfy (12.1.8), we can essentially repeat the argument for each
side. If there are ttj poles of order ctj and k, lattice points on the jth side,
we have kj ^ tt, for each j and
V V

so the theorem is proved. ■

As illustrations we take up some special cases. We have already
considered the Weierstrass p-function, which leads to a DE with k = 3,
m = 2, n = 1. For the elliptic functions defined by

(w')4 = (W _ a)3(n, _ k)3(w; _ cf (12.1.15)

we find (w")2 = P2(u>), where P2 is of degree 6. Here k = 1, m = 4, n = 2.

Similarly for the functions satisfying

{w'f = {w-a)\w-b)\w-c)\ (12.1.16)

we find that (w")3 = P3(w), where P3 is a polynomial of degree 9, not a

perfect cube; in fact, P3(w) = (w - b)(w - c)2[Q(w)]3, where Q is a
quadratic polynomial. Here k = l, m=6, n=3. Finally for

(w')3 = (w - a)2(w - bf (12.1.17)

we obtain

(w")3 = S(w - a)(w - + ^)]3

with k = m = n = 3.
We can extend the discussion to equations of higher order than the
438 SECOND ORDER NONLINEAR EQUATIONS

second and prove analogues of Theorem 12.1.1 for equations

[w(kT + X P/(w)[w(k)]n"i = 0. (12.1.18)

We observe only that deg (P,) = j =p,i ^ (fc + l)j. For the binomial equations
[w(kT =P(w) (12.1.19)
the discussion is still not too complicated, but we refrain from further
excursions in this direction.
Equations of the form
w" = R(z, w) (12.1.20)
have been studied from various points of view. If the problem is to find
necessary conditions for the existence of solutions holomorphic in the
distant part of the plane save for movable poles, we can start with the
test-power test. If

where P and Q are polynomials in z and w, and if P is of degree p, Q of

degree q in w, the substitution w (z ) ~ A (z - z0)~a leads to the condition
(p -q - \)a = 2. (12.1.21)
If a is to be an integer, there are two possibilities: (i) a = 2, p = q + 2, and
(ii) a = 1, p = q +3.
Though it is not obvious that we must have q = 0 in order to exclude
branch points, we shall assume this to be the case. Thus we have the
equations

w" = i Mz)w\ (12.1.22)

where p = 2 or 3. Paul Painleve, who

j=0 made a general study of the problem
of finding nonlinear second order DE's without movable branch points, did
find two equations of type (12.1.22) which define new transcendental
meromorphic functions, one corresponding to p = 2 and the other to p =3.
See Sections 12.2 and 12.3.
It should be observed, however, that our conditions are merely
necessary for the existence of poles, not sufficient. This is shown in a
glaring manner by the special Emden equation
zw"=w\ (12.1.23)
where the solutions have movable logarithmic pseudopoles. This case and
related ones will be discussed in Section 12.4 and following.
THE PAINLEVE TRANSCENDENTS 439

EXERCISE 12.1
1. Verify (12.1.1) and (12.1.2).
2. How should the constants a, b, c, d be chosen in order that the circle C will be
I* -211 = 17
3. Why does (12.1.8) hold?
4. Why do the points (a, n - j), a integer <ph lie inside the polygon II if
(Ph n nes inside or on n?
5. Find the second order DE satisfied by Jacobi's sine amplitude.
6. Find the first order DE satisfied by p'(z;0, 4). Here, to simplify, we take
g2 = 0, g3 = 4. Show that the second order equation satisfied by this function is
{w"y = 2iw\w2 + 4)y
so that k = m = n = 3.
7. Verify the conclusion for (12.1.15).
8. Solve Problem 7 for (12.1.16).
9. Solve Problem 7 for (12.1.17).
10. Verify that in (12.1.18) we have p, ^ (fc + 1)/. The task is trivial on the basis of
Chazy's theorem, so do it the hard way.
11. Verify (12.1.21).

12.2. THE PAINLEVE TRANSCENDENTS

Consider DE's of the form

w" = F(z, w, w'), (12.2.1)
where F is rational in w and w', and analytic in z. In 1887 Emile Picard
raised the question of when such an equation has fixed critical points. Here
"critical points" includes branch points and essential singular points for the
sake of convenience. The problem was solved by P. Painleve in a series of
papers starting in 1893. R. Gambier made important contributions from
1906 to 1910, and R. Fuchs (the son of Lazarus Fuchs) in 1905. Extensions
to third order equations were made by J. Chazy (1911) and Rene Gamier
(1912-16). The general method, due to Painleve, consists in finding two
necessary conditions (this is fairly easy) and then verifying that these
conditions are also sufficient. This part of the argument involves consider-
ing a large number of separate cases and showing that the corresponding
equations either are reducible to equations known to have the desired
property or define new transcendents. This part is very laborious.
Painleve's basic device is his a-method. He introduced a parameter a in
the equation such that for a = 0 the resulting equation can be seen by
inspection to be with or without movable critical points. If now the
440 SECOND ORDER NONLINEAR EQUATIONS

equation has single-valued solutions for small \a\ with the possible
exception of a = 0, then a = 0 can be no exception. Using this artifice,
Painleve showed that F must actually be a polynomial in w' of degree *s 2
so that the equation is of the form

w"= L(z, w)(w')2 + M(z, w)w' + N(z, w), (12.2.2)

where L, M, N, are rational in w. If now z = c is not a fixed singular point,
he showed that the equation

W" = L(c, W)(W')2

must have fixed critical points. This will be the case if L = 0, but there are
also five other possibilities, obtained by applying logarithmic differentia-
tion to the Briot-Bouquet equations of Section 11.3 and replacing the
constants by functions of z. If L = 0, then M must be linear in w, while N is
a polynomial in w of degree ^3. If L^O, M and N take the forms
m(z, w) n(z, w) /i^o^
M(Z'H,)=DU^)' N(2'W) = D(^j- (12-2J)
Here D(z, w) is the least common denominator of the partial fractions in
L(z, w) and is a polynomial of w of degree d, 2 ^ d ^ 4, while m and n are
polynomials in w of degrees ^d + 1 and =sd+3, respectively. This
exhausts the necessary conditions and leads to 50 possible types. Only 6 of
these lead to new classes of functions, the Painleve transcendents. These
equations are as follows:

(1) w" = 6w2 + z,

(2) w" = 2w3 + zw + a,

(3) w" = J-(w')2-i

w z w'+i(aw2
z + /?) + cw3 + —u> ,

(4) w" = -i-(w')2 + 5vv3 + 42vv2 + 2(22_a)H;+^5

, w w(w + 1)
z w-1

2\w w-1 w-z J \z z-1 w-z/

+ w(w-lXw-z)r +te + cJ^+

z (z - 1) L vv (w-1) <f(w-z
^01 ) J
Equation (6) is the master type and was added to the list by R. Fuchs. The
other equations may be obtained from it by passage to the limit and
THE PAINLEVE TRANSCENDENTS 441

coalescence. Equations (4) and (5) were added by Gambier. There are no
critical points at all in the first three cases. If the substitution z = e ' is made
in (4) and (5), the solutions become single-valued functions of t. For type (6)
0, are fixed critical points.
We shall discuss equation (1), which is the simplest and has very
interesting properties. Here we start by showing that the only algebraic
singularities are movable double poles. There is a unique solution
u>(z; z0, Wo, Wo), which satisfies the equation and is holomorphic in some
neighborhood of z = z0, where it takes on the value w0 while its derivative
equals w0. It follows that, if z = z0 is to be an algebraic singularity, at least
one of the quantities w(z) and w'(z) must become infinite when z
approaches z0. Now the equation shows that the three quantities w(z),
w'(z), w"(z) are cofinite, so for an algebraic branch point we must have
both lim w(z) and lim w'(z) infinite.
This being the case, we may postulate a formal series solution

w(z)= n2=0 an(z-Zo)\ (12.2.4)

where the a 's form a strictly increasing sequence tending to °o and a0 < 0.
Such an expansion will exist at an algebraic branch point but also if a
nonlogarithmic psi series is hidden. The argument to be given will show,
however, that the a 's have to be integers so that at least some pathological
phenomena cannot occur. Since we are not going to give Painleve's proof
of the meromorphic character of the solutions, and psi series will occur in
Section 12.4 and later, some reassurance is due the reader.
Substitution in the equation gives

n=0 |_n=0
2 oLn (ctn - \)an (z - z0)a"~2 = 6[ J + z0 + (z - z0).
2 Mz - Zo)a"l
Here the terms of lowest order must cancel. Thus

so that a0(a0- l)a0(z - Zo)"0"2 = 6al(z - z0)2a°

a0 = — 2, a0 = 1.
This is in agreement with our expectations. After these terms have been
canceled, the lowest term on the left is

a,(a, - l)ai(z - Zo)"1-2, a,>-2,

while on the right we have two terms,

12a,(z -z0r"2 + z0.

If z0 5* 0, the two terms on the right are of the same order iff a, = 2, and for
442 SECOND ORDER NONLINEAR EQUATIONS

this value the term on the left is also constant. This gives a, = 2, ax = -joZ0y
which is true also if z0 = 0.
At the next stage we should have

a2(a2 - \)a2(z - z0p~2 = 12a0(z - Zop-2 + (z - z0),

which gives
a 2 = 3, a2 = — i
For n = 3 we should have

a3(a3 - l)a3(z - z0P~2 = 12a3(z - z0p~2,

and this shows that a3 = 4 and leaves a3 = h arbitrary. We are thus going to
get a one-parameter family of solutions. In this manner we prove
successively that all exponents an are integers. Moreover we have
an ^ n + 1 for n > 0. The expansion of the solution starts out

w (z) = (z - z0r2 - it)Zo(z - z0)2 - i(z - zof + h (z - z0)4

+^>zl{z -Zo)6 + boZ0(z -z0)7+ • • (12.2.5)
This formal series should now be proved convergent in some neighbor-
hood of z = z0. We start by choosing a number M > 1 such that

|zo|<10M, |h|<M3. (12.2.6)

A look at the series (12.2.5) shows that we have then
\ak\<M\ 0^k^6. (12.2.7)
Now the coefficients ak are determined from the recurrence relation

[an(an - D-12R =6^ a,afc, (12.2.8)

where j and k are positive integers < n whose sum j + k =s n. The sum on
the right in (12.2.8) contains at most n - 1 terms because for a given j,
1 j n — 1, there is at most one fc in the same range such that a^ak is
actually present in the sum. The bracket in the left member equals
(n - 3)(n +4) at least, since an ^ n + 1. If (12.2.7) holds for the j's and /c's
under consideration, we have

Mn^Mn for 6^/i.

(n-3)(n+4)
Since the inequality holds for n < 6, it holds for all n. This shows that the
formal series (12.2.5) converges for

and represents a solution ofO the -z0\<M'1

<|zPainleve equation (1) in this punctured
disk.
THE PAINLEVE TRANSCENDENTS 443

From this point on we shall accept Painleve's result that the solutions of
the equation
w" = 6w2 + z (12.2.9)
are transcendental meromorphic functions with double poles. Hans
Wittich (1953) considered these functions from the point of view of the
Nevanlinna theory. The following remarks have a similar trend but veer off
in a slightly different direction. To fix the ideas, let z0 ^ 0 and consider the
solution w(z;z0, 0,0), which is holomorphic in some neighborhood of
z = z0. Multiply the equation by 2 w ' , and integrate from z0 to z along a path
which avoids the poles. In view of the initial conditions the result is
JZO
[w'(z)f = A[w(z)f + zw{z)- f w(t)dt. (12.2.10)

Here we divide by 4[w(z)]3 to obtain

4[w(z)f 4
\^-\{Lw{t)dt/[w{z)A- (l2-2-n)
Consider a set S, defined as follows:

5 = {z ; \w(z)\ > |z|1/2, \J(z)\ < 1}, (12.2.12)

where J(z) = [w(z)]3 J*/0 w(t) dt. All poles of w(z) belong to S. It is clear
that the first condition is satisfied at a pole. Since /(z)-»0 as z tends to a
pole, the second condition is also satisfied. If z G 5, both fractions in the
right member of (12.2.1 1) are less than \ in absolute value, so the right-hand
side is of the form 1 + h0(z), where \h0(z)\ < \. Next we extract the square
root and obtain

2[w(z)fl2=l + h{zl lh<z>l<J forzeS- (12.2.13)

If now Zi and z2 are two points in S which may be joined by a line segment
in S, integration gives
Jz\
[w(z1)r1/2-[w(z2)r1/2= f 2 [\ + h(t)]dt. (12.2.14)
If now z2 = z„, a pole of w(z), we obtain the inequality

M|z - z„T < k(z)| < li\z - z„T2 (12.2.15)

for any point z which is "visible" in S from z„.
We can now introduce polar neighborhoods,

Un ={z;|z-z„|<!|z„r1/2}. (12.2.16)
That the constant j is admissible is proved in the same manner as was used
in Section 4.6, and the proof may be left to the reader. No two polar
neighborhoods can intersect, as is shown using (12.2.14).
444 SECOND ORDER NONLINEAR EQUATIONS

We can then use the polar statistics of Section 4.6 with g = i This gives
n(r, oo; w) = 0(r3), m(r, oo; vv) = 0(log r), and
T(r; w) = 0(r3). (12.2.17)
This estimate is surprisingly good for, as will be shown in the next
section, w(z) is of order §. We have found the estimate p(w) = 3 for the
solution w(z ; z0, 0, 0). Although the assumption w(zo) = 0, w'(zo) = 0
makes the formulas neater, the same result holds if this assumption is
dropped.
Wittich has proved that the Nevanlinna defect 5(a) = 0 for all values of
a. The value a = » is completely ramified for all solutions, but no other
value can have this property.

EXERCISE 12.2

1. Show that a solution of (12.2.9) can have a triple zero at z = z0 iff z0 = 0.

2. Let 2 =x >0, and take a solution w(z ; 0, w0, w,) with w0>0, (w,)2 > 4(w0)3.
Show that w(x) becomes infinite for some x = xx with 0< jc, <(w0)"1/2.
3. Let w(z) be a solution of (12.2.9), z0 not a pole of w(z), and form
£(z) = exp [— fz0 J70 w(0 ds]. Show that £(z) is an entire function of z, and
express w (z ) in terms of E (z ) and its derivatives. Show that a pole of w (z ) is a
zero of E(z). This function E(z) plays a role in the Painleve theory analogous
to that of the sigma function in the theory of the Weierstrass p-function.
4. Consider a Painleve equation of type (2). Show that its movable singularities
are simple poles. Find the Laurent expansion at one of the poles, and give a
convergence proof for the expansion. Take a = 1.
5. Show that the neighborhoods defined by (12.2.16) are actually contained in S
and hence do not overlap.

12.3. THE ASYMPTOTICS OF BOUTROUX

In 1913-1914 Boutroux published a memoir (180 pp. !) in which he examined

the properties of the Painleve transcendents from the point of view of
complex function theory and also developed a method of asymptotic
integration for a class of second order nonlinear DE's including those of
Painleve. We shall discuss briefly his results for equations of type (1), to
which he gave the form
w" = 6w2-6z. (12.3.1)
This may be obtained from (12.2.9) by dilation of the variables.
The equation is unchanged under the rotation
2 it i
z cuz, w -+ <jo\', a) = exp — r-. (12.3.2)
THE ASYMPTOTICS OF BOUTROUX 445

This implies that, if w{z) is a solution, so is

It also shows that (12.3.1) admits of the fifth roots of unity as distinguished
directions, so that a study of the equation in one of these directions gives
global information.
This observation suggests the following change of variables:

Z=iz5'\ W = zll2w, (12.3.3)

and the resulting equation
W 4 W
W"-6W2 + 6=-^ + ^j2' (123-4)
This equation is asymptotic to the elliptic equation
WS-6WS + 6 = 0, (12.3.5)

which is satisfied by p(Z - Z0; 12, g3) for any choice of Z0 and g3. This is
analogous to the situation encountered in Section 5.6, where the sine
equation
YS+Yo = 0 (12.3.6)
was found to be asymptotic to the generalized Bessel equation

Y" + [1-F(Z)]Y = 0 (12.3.7)

under suitable assumptions on F(Z).
The first integral of (12.3.5) is

(W'o)2 = 4Wl-l2W0-g3 (12.3.8)

with integrals as stated above. Normally the solution is doubly periodic
with an array of poles at the points
Z0 + 2raa>i + 2ncu3,
where the periods wi,w3 may be computed from the value of g3. The case
where the cubic in the right member of (12.3.8) has equal roots, g3 = ±8, is
exceptional. Here solutions degenerate into simply periodic solutions:

1 + 3[sinh 31/2(Z - Z0)]~2 or - 1 + 3[sin 31/2(Z - Z0)]-2 (12.3.9)

with a single string of poles instead of a double array. This behavior of
W0(Z) is reflected as asymptotic properties of W{Z).
Now take a solution W(Z) of (12.3.4) given by its initial values, 170 and 17,,
at a distant point Z0, set

D=(t?1)2-4(t?0)3+12t?0^±8,
446 SECOND ORDER NONLINEAR EQUATIONS

and let W0(Z) be the solution of

(W'of = 4(W0f - 12W0 + D, (12.3.10)

which has the same initial values as W(Z) at Z = Z0. This is an elliptic
function of Z. Let a, b, c be the distinct roots of the cubic
4s3 -12s +D =0.
At this stage it is convenient to work with the inverse functions Z( W) and
Zo(W). Now let W describe a loop Li, starting and ending at 170 and
surrounding a and b but not c. At W = tj0 both inverses start with the value
Z0, and at the end point of L, we have

Z0( W) = Z0 + 2a>„ Z(W) = Z01,

where 2o>i is the period

2oj!= f (4s3-12s+D)"2ds
of W0(Z;12,D). Moreover,

|Z(W)-Z0(ttO| (12.3.11)
is small for all W on Li. In the same manner we can let W describe a loop
L3 surrounding b and c but not a. Here the terminal values are,
respectively,

Zo + 2w3 and Z,0, 2w3 = J (4s3 - 12s + D)_1/2 ds.

These operations may now be repeated at Z0i and Zi0, where in each case
the initial values, the parameter D, the roots of the cubic, and the loops are
given their local values. It follows from (12.3.1 1) that these values change
by small amounts only when we go from Z0 to Z10 or Z01. Suppose that at Z01
we use the L3-operation and at Zl0 the Li-operation. For the class of DE's
asymptotic to (12.3.5) distinct though nearby points would normally be
obtained. But in the case under consideration we get the same point, Zn.
Now the four points
Zo, Z01, Zn, and Z10
form the vertices of a curvilinear quadrilateral Qn, the sides of which are
traced by different determinations of Z(W) when the loops are traced.
This construction can now be repeated at the new vertices and leads to an
array of lattice points {ZJk}, obtained by j applications of the Li-operation
and k applications of the L3-operations. The four points
Zj+i,k, Zj +i,ic+i, and Z^+i
THE ASYMPTOTICS OF BOUTROUX 447

form the vertices of the curvilinear quadrilateral Qi>k. These quadrilaterals

form approximate period parallelograms. Each QUk contains a single
double pole of W(Z) and three zeros of W'(Z). Furthermore,
\W(Z)-Wo(Z; 12, (12.3.12)
is small everywhere in Qj k except for small neighborhoods of the poles. We
have
Dj,k = [W(Zj>fc)]2-4[W(Zj,k)]3 + 12W(Zi,fc).
Here j and k are positive; but if the imaginary part of Z0 is large, Zitk can be
defined for negative values of j, and if the real part is large, for negative
values of k.
It should be noticed that the approximation is local and shifts from one
quadrilateral to the next. It is true, however, that
lim Dj k and lim DJ k
exist. Furthermore,

lim (Zj +i,k - Z},k )

j—*oo and lim
k— »<» {Zuk + , - Zi>k )
exist, and the first is real, the second purely imaginary.
If Dj,k keeps away from the critical values ± 8, the poles of W(Z) form a
net {ZUk} which at a distance from the origin is approximately rectangular.
The number of poles inside a big circle \Z\ = R satisfies AR 2 < n{r) < BR2.
Now, if D has one of the critical values ± 8, Boutroux shows that there is an
ultimate string of poles approximately horizontal or vertical, beyond which
there are no poles and W(Z) tends to + 1 or - 1 as Z ->» in the pole-free
half-plane.
Now the mapping Z =iz5'4 takes Z-half -planes into sectors. There are
five such sectors,

kiir < arg z < (k + l)lir, k = 0, 1, 2, 3, 4. (12.3.13)

Since the order of W(Z) is 2, that of w (z) will be §. Normally there will be a
curvilinear net of poles of u>(z) in each of the sectors (12.3.13). There are,
however, infinitely many truncated solutions where an ultimate string of
poles exists in one sector or such a string extends through two adjacent
sectors. Such a truncated solution ordinarily lacks one fifth of the normal
allotment of poles. But it is not merely shorn of poles: every value a is
taken on with the same lowered frequency. There are, in addition, five
triply truncated solutions related by the transformation (12.3.2). Such a
solution has a net of poles in only one of the five sectors; each of the
remaining sectors has at most a finite number of poles, and there

limz"1/2u>(z) = + l. (12.3.14)
448 SECOND ORDER NONLINEAR EQUATIONS

EXERCISE 12.3
1. Show that (12.3.14) is consistent with (12.2,12).
2. Verify (12.3.4).
3. The equation w" = 2wi -2zw + a is reducible to Painleve's type (2). How?
Show that the transformation

Z=\zm, W = z'mw
transforms the equation into
W 1 W a
W" = 2W'-2W --f- + ^ + f-

4. The (Z, W)-equation in Problem 3 has, as asymptotic elliptic equation,

W"0 = 2Wl-2W0 or (W'o)2=W4o-2Wl + D.
Express the solution of the second equation in terms of a Jacobi elliptic
function. What are the critical values of D for which the solutions become
degenerate, and what are the degenerate solutions?
5. Show that the order of the Jacobi functions is two and hence that the order of
w(z) is three.
6. Discuss Nevanlinna defects and complete ramifications for w(z).

12.4. THE EMDEN AND THE THOMAS-FERMI EQUATIONS

The equations encountered so far in this chapter have fairly simple
solutions. We shall now consider a class of equations which presents a
more complicated picture: the Emden-Fowler and Thomas-Fermi equa-
tions. First we have something to say about the background of these
equations.
In 1907 R. Emden published a book entitled Gaskugeln, in which he
developed a theory of gas spheres with applications to cosmology and
stellar dynamics. In this theory the equation

y"(z) = -xl-m[y(x)]m, Km <3, (12.4.1)

plays a fundamental role. Further developments are due to R. H. Fowler in
1914 and 1931.
In the following we shall attach Emden's name to the equation
y"(x) = xl-m[y(x)]m, (12.4.2)
an equation with much more interesting and spectacular properties. Set
P + 2 . ^ ^
m =- P , 1 < p < °°, (12.4.3)
 449
THE EMDEN AND THE THOMAS-FERMI EQUATIONS

(i so that the equation becomes

w'\z) = z-2,p[w(z)]°}+2)lp. (12.4.4)

In the physical theory the variables are normally real positive, but the
various series which satisfy the equation (and which have been developed
largely by physicists) call for complex variables.
The special case p = 4 is known as the Thomas-Fermi equation. This DE
was encountered in 1927 by the American physicist L. H. Thomas, then
working at Cambridge, England, on the potential field around an atom, and,
independently, by the Italian Enrico Fermi, later of atom bomb fame, but in
1927 still working in Rome. Fermi encouraged two of his Italian mathemati-
, cal colleagues, A. Mambriani and G. Scorza Dragoni, to provide rigorous
proofs of the existence of solutions of a singular boundary value problem
which he had encountered in the work. The Thomas-Fermi equation
w„ = z-mwm (12.4.5)
has quite an extensive literature. Additional references are given below
and at the end of the chapter.
Equation (12.4.4) is algebraic iff p > 1 is an integer, when it may be
written as

(w")p =z~2wp+2.
Since it does not belong to any of the Painleve classes, its movable
singularities cannot be poles, though there are algebraic branch points.
The only rational case occurs for p =2:
w" = z'lw2, (12.4.6)
which repeatedly has served as a bugbear in earlier chapters.
If (12.4.4) has a solution w(z) and if k is any constant, then

kx pw(kz) (12.4.7)
is also a solution. This can be used to place movable singularities where it
is convenient, to normalize boundary conditions, etc.
The equation has an elementary solution,

we{z) = [p{p-\)]pl2zx-p, (12.4.8)

which normally is singular at both the fixed singularities, zero and infinity.
This solution may be embedded in two systems of solutions, one valid
in a neighborhood of infinity, and the other in a neighborhood of zero. The
first system has the form

(12.4.9)
Mz){l + 2 A„c"z"<'].
450 SECOND ORDER NONLINEAR EQUATIONS

Here c is an arbitrary parameter, and a is the negative root of the

equation
s2-(2p - \)s-2(p -1) = 0, (12.4.10)
i.e.,
(T = p-{-[(p+hY-2]m. (12.4.11)
The series converges for large values of \z |. At z = 0 we have the second
embedding system,

vve(z){l + 2 B„cnz"Tj, (12.4.12)

where r is the positive root of (12.4.10). The series converges for small
values of |z|. For the Thomas-Fermi case these series are associated with
the names of C. A. Coulson and N. H. March (1950).
We shall sketch how the series (12.4.9) is obtained. Set

w(x) = we(x)v(x). (12.4.13)

Here v(x) is monotone and converges to + 1 as jc ->a>. Furthermore, v(x)
satisfies the DE

x2v"-2(p - \)xv' +p(p -\)v=p(p- lKP+2)/p, (12.4.14)

which has the solution v = 1. To get other solutions we set v = 1 + u. The
resulting equation can be brought to the form

x2u"-2(p - l)xu'-p(p - \)u = F(«), (12.4.15)

where F(u) is holomorphic for \u \ < 1, and its Taylor series at the origin
starts with a term in u2. Now the homogeneous DE
x2U"-2(p - \)xU'-p(p - l)L/ = 0
is an Euler equation with the two linearly independent solutions
jcct and jct.
To get the system (12.4.9) we expand the solution of (12.4.15) in powers of
xa. The system (12.12) is obtained by expanding u in powers of x\
Convergence
method. is shown by a suitable application of Lindelof's majorant

There is no real-valued integral which tends to +°c with jc. It is not

known whether there are other unbounded integrals at infinity. At zero we
have a large number of additional integrals besides (12.4.12). For p > 2 the
boundary value problem

limy(jc)=l,
x 10 \\my\x)
x 10 = a (12.4.16)

has a unique solution given by a series of the form

 451
THE EMDEN AND THE THOMAS-FERMI EQUATIONS

y(x) = 1 + ax + 2i c„jcm", (12.4.17)

where the /jl 's form an increasing sequence taken from the semimodule
kx + 2k2[l - (l/p)L where kx and fc2 are nonnegative integers. For p = 4,
the Thomas-Fermi case, this series is known as the Baker series (E. B.
Baker, 1930). The coefficients cn are rational functions of p, and the set of
coefficients has infinitely many poles. In particular, the values
P=2k= 2kl

are poles, and these are precisely the values for which the initial value
problem limx i , y(x)= 1 does not have a unique solution and a logarithmic
psi series appears. This series is of the form

y(x) = %Pn(logx)xn,k, P=TfH' (12.4.18)

Here Pn(t) is a polynomial in t of degree [(n + l)l(k + 1)], P0(f) = y0>0,
and Pk+i(0) = b is arbitrary. The series is convergent for small values of
|jc|. Actually there is a solution which is holomorphic at the origin and has
the form

w(z) = f,bkz2k+l.
0 (12.4.19)
This solution is unique.
So much for the fixed singularities. The point where a solution vanishes
is obviously a singularity, and since the initial-value problem

w(zo) = 0, w'(z0) = b (12.4.20)

has a unique solution, we have movable branch points. The corresponding
expansion is
w(2) = 2a„(z-z0)A";
i (12.4.21)

ax = b, Ai = l, A2 = (3p+2)/p, and generally {A„} is an increasing se-

quence taken from the same semimodule as the /i's.
But we have also movable infinitudes, and if 2p is an integer they
involve logarithmic psi series. Of the real-valued solutions only those
whose graphs lie below or on that of the elementary solution (12.4.8) do
not become infinite somewhere. The following case is typical.

THEOREM 12.4.1

Let y(x) be a solution of (12.4.4) with y(a) = b, y'(a) = c, where a, b, c are

positive numbers such that

(p + l)c2 ^ pa~2/pb2(p+,)/p. (12.4.22)

452 SECOND ORDER NONLINEAR EQUATIONS

Then there exists an X with a<X such that /jmxTX y(x) = +00 and

xi-i/P<ai-i/p + (p_1)^P±iy/2b-1/p, (12.4.23)

and for a < x < X
y(x) < [p(p + 1)]P/2X(X - x)-p. (12.4.24)
Proof (sketch). Since

2y"(t)y'(t) = 2t-2,p[y(t)r+2)lpy'(t),
integration from t = a to t = x gives

Ja
[y'(x)]2 -c2 = l\X r2/p [y(OJ(p+2)V(0 dt.
Here we integrate by parts to obtain

[y '(x )]2 = c 2 - a 2lph 2(p +,)/p + jc -2,p [y (x )]2(p+1)/p

By virtue of (12.4.22) we have

(12.4.25)
[y\s))2>-^s-2/p[y(s)Fp+»lp, a^s,
and since y'(s)>0 we get
(12.4.26)
Integration gives [y(*)r1-,V(s)>(^2rr)'V'''\

{^-1/p-[y(x)rl/p}>(^T)1/2^T(xi-,/p-«i,/p).
Here the left-hand side is bounded for all x > a, whereas the right side is
not. It follows that there is a finite X > a to the left of which the
inequalities hold, but not to the right. This shows that limx ^xy(x) = +°°.
If in (12.4.25) we integrate from x to X instead, we get an inequality; and
if we apply the mean-value theorem of the differential calculus to the right
member, (12.4.24) results. The solution ceases to be real for x > X. The
constant in the right member of (12.4.24) is the best one possible. ■
We come now to the expansions at a movable infinitude. In view of
(12.4.7) we can place this infinitude anywhere we choose. It simplifies the
work somewhat to take X = 1 . We now set
x = l-e\ t<0, (12.4.27)
THE EMDEN AND THE THOMAS-FERMI EQUATIONS 453

in the DE and obtain

Y"- Y' = e2t(l-e'r2lpYl+2lp, (12.4.28)
where the primes denotes differentiation with respect to t. We postulate
the existence of a solution of the form

nO=E Pn(t)e(n-p\ (12.4.29)

where the P„'s are polynomials in t which may possibly reduce to

constants. The degree and coefficients of Pn are to be determined. If the
series (12.4.29) has a domain of absolute convergence extending to -<»,
we can differentiate the series, substitute the result in (12.4.28), multiply
by ept, expand (1 - et)~2lp in powers of e', collect terms, and equate the
coefficient of ent to zero for all n. For n = 0 we get
PS(f)-(2p + l)«(f) + p(p + DPo(0 = [Po(*)]1+2/p,
which has a unique polynomial solution, namely, the constant

Po(0 = [p(p + l)]p/2,

which we recognize as the factor occurring in the right member of
(12.4.24).
We are finally led to an infinite system of linear nonhomogeneous
second order DE's for the successive determination of the P„'s. These
equations are of the form

Kit) + (2n -2p- \)P'n(t) + (n + \)(n - 2p - 2)Pn (t) = Qn(t), (12.4.30)

where Qn is a multinomial in P,, P2, . . . , Pn-i, 0<n. Here it makes a
difference whether or not 2p is an integer. Suppose that 2p is not an
integer, and that for n < k the coefficients Pn have been found to be
specific constants. This means that the Qn 's with n ^ k are constants, and
(12.4.30) then gives for n = k the unique polynomial solution

Pk(t) = (k + l)(k-2p -2Y

again a well-determined constant since the denominator cannot vanish for
any value of k. The solution is then of the form k-p

k=0
Y«)=±Pkelk-»; y(x)=2k=0 ftd-*)
This solution has a branch point at x = 1 which is algebraic iff p is
rational.
If 2p is an integer, the situation changes radically. Then the coefficients
454 SECOND ORDER NONLINEAR EQUATIONS

are constants only for n <2p +2. For n = 2p + 2 (12.4.30) takes the form
PWO + (2p + D^WO = Q2p+2(0.
Here the right member is a constant, and the equation has the polynomial
solution n
P2p+2(t) = j^\t + B,
where B is an arbitrary constant. For 2p+2<n<4p+4 the (?„'s are
first degree polynomials, and so are the P„'s. For n =4p +4 there is a
change, since Q4p+4(t) involves the square of P2p+2 and hence PAp+A(t) is a
quadratic polynomial and so on. Each time that n passes a multiple of
2p +2 the degree of Pn(t) increases by 1 unit.
The resulting series (12.4.29) may be shown to have a domain of
absolute convergence extending to -oo in the f-plane or a punctured
neighborhood of z = 1 on the Riemann surface of log (1 - z). The various
proofs are nasty, and considerations of space require their omission.
We formulate the results as

THEOREM 12.4.2

The equation (12.4.4) has movable singularities. If 2p is not an integer

> 2, then there is a solution at z = 1 of the form

w(z) = k=0
2Pk(l-z)k~"p. (12.4.31)
// 2p is an integer >2, then the solution is given by the logarithmic psi
series

w(z) = n=0
2 Pn[bg (1 - z)](l - z)n p, (12.4.32)

where Pn(t) is a polynomial in t of degree [n/(2p + 2)]. Here P2p+2(0) = B is

an arbitrary constant, and Pn(t) is a polynomial in B as well as in t.

EXERCISE 12.4
1. Verify (12.4.7).
2. Verify (12.4.8).
3. Prove that the function x h-> v(x) defined by (12.4.13) has the stated property
of monotony and existence of limit.
4. Verify (12.4.14) and (12.4.15), and find the leading term of F(w).
5. Show that the graphs of two solutions of (14.4. 14) cannot intersect and hence
that a solution whose graph contains a point (a, b) above the graph of
y = we(x) stays above the latter and is monotone decreasing to its limit.
 455
QUADRATIC SYSTEMS

6. Find the first two coefficients and exponents in Baker's series.

7. Show that the coefficients cn are rational functions of p and that the point
2kl(2k - 1) is a pole for some coefficients.
8. Discuss the series (12.4.12). Show that this solution exists also for p < 1,

9. Verify doubtful points in the proof of Theorem 12.4.1.

10. Solve Problem 9 for Theorem 12.4.2.

12.5. QUADRATIC SYSTEMS

The term quadratic system was introduced by the Australian mathemati-

cian W. A. Coppel in 1966 for a pair of autonomous DE's of the form
x = Q1(jc,v), y = Q2(x,y), (12.5.1)
where Q, and Q2 are quadratic polynomials in x and y with constant
coefficients, and the dot indicates differentiation with respect to the
variable t. We shall take the equations in the normal form

x=xL](x, y), y = yL2(x, y), (12.5.2)

where Lx and L2 are first degree polynomials,

Lx = a0 + clxx + a2y, L2 = fr0+ bxx + b2y. (12.5.3)

The coefficients and the variables are assumed to be real.
Coppel's paper was essentially devoted to the topological nature of the
integral curves, singularities, closed cycles, etc. Although such questions
will be considered in the following, our main task will be the existence and
nature of the movable singularities. In particular, we shall consider the
question, Are there solutions which are singular when t -» t0, an arbitrary
finite value?
Quadratic systems are not artificial constructs; as a matter of fact,
Coppel lists no fewer than 13 problems in applied mathematics that lead
to such systems, either directly or as the result of applying a suitable
transformation to a nonlinear DE, usually of the second order. Thus
quadratic systems arise in the theory of the struggle for life of two
competing species (Lotka, 1925; Volterra, 1931). The second order
nonlinear DE
w>" = z'-lH>k (12.5.4)
reduces to the quadratic system

jc = l-x + y, y=j + kx-y (12.5.5)

456 SECOND ORDER NONLINEAR EQUATIONS

via the substitution

W' ;Wk . .
x — z —W
, y=z'—W 7, / = log 2.

Here Emden's equation corresponds to j = 1 - 2/p, k = 1 + 2/p. The more

general Emden-Fowler equation of astrophysics,
[z2w']' + zqwn =0, (12.5.6)
corresponds to a0 = ax = a2 = - 1, bo = <? + 1, b\ = n, b2 = 1. The equation
z2w"=w", n integer >1, (12.5.7)
has been encountered by Kurt Mahler (private communication).
The system (12.5.2) has one and only one integral curve passing through
any given point (jc0, y0), with normally four exceptions, namely, the points
where x and y vanish simultaneously. There are four straight lines which
determine these points: the coordinate axes X and Y and the lines L, = 0,
L2 = 0. We denote the points by O, U, V, and W. See Figure 12.1. In
addition, there are normally three infinitary topological singularities, Xoc,
Yx, and F™, which are the end points of the axes and the line
F = L2-L, = 0. (12.5.8)
For our present purposes it is convenient to restrict the discussion to
the class F, for which the layout is given schematically by the figure; i.e.,
U lies on the positive X-axis, V lies on the positive F-axis, O is the
origin, and W lies in the third quadrant. In addition it is required that both
Li and L2 be positive at the origin. The assumptions on the critical points
imply that O, U, V, W are the vertices of a nonconvex quadrilateral. For
this case Coppel has proved that the critical points are either three saddle
points and one nonsaddle point or the other way around. For the class F
the origin is always a node, but the other phase portraits, in the sense of
Poincare (see S. Lefschetz, 1963), may vary. Analytically the class F is
determined by the inequalities
a0>0, a,<0y a2>0, b0>0, fc, >0, b2<0. (12.5.9)
The four lines X, Y, L,, L2 divide the plane into 11 sectors Si-Sn, in
each of which the four functions

x(0, y(o, m, Ht) (12.5.10)

keep a constant sign and at least one signature is changed when a
boundary line is crossed. We are chiefly interested in solutions such that
x(t) and/or y(t) becomes infinite as t approaches a finite value. Such a
solution has a graph which approaches one of the infinitary critical points
Xx, Yx, or Foe. The last case is the most interesting one.
QUADRATIC SYSTEMS 457

Figure 12.1

THEOREM 12.5.1

Take a point (x0, y0) in Si, and define a trajectory T = [x(t), y(t)] with
x(0) = x0, y(0) = y0. As t increases from 0, the trajectory T stays in S, and
tends to infinity as t tends to a finite value to. On T the radius vector r(t) is
monotone increasing, while the argument 0(t) is ultimately monotone and
tends to a limit 0O as t^t0 and

tand^^—r-.
a2- b2 (12.5.11)

The line ¥ or a parallel thereof is an asymptote of T.

Proof. From the system (12.5.2) we obtain

rr= jc2L,(jc, y) + y2L2(x, y), (12.5.12)

0(0 = { sin [20(O][L2(x, y) - L,(jc, y)]. (12.5.13)
458 SECOND ORDER NONLINEAR EQUATIONS

In Si both L, and L2 are positive, and so are the four functions of

(12.5.10). It follows that x(t), y(t), and r(t) are strictly increasing as long
as T stays in S,. But T cannot get out of S, without crossing one of the
lines L, and L2, and the curve tangent would have to be vertical on L, and
horizontal on L2; neither alternative agrees with the geometry of the
situation. It follows that x(t), y(t), and r{t) tend to +00 as t increases
toward a limit which may possibly be +00.
We now look at (12.5.12). Here the sine factor >0 in S, and bounded
away from zero. L2-L, is zero on the line F, negative above F, and
positive below. If (jc0, y0) lies on F, then T has to remain on F. Note that
the slope of F is tan 0O. If (jc0, y0) lies in Si below F, it has to stay below,
and 0(0 is increasing and tends to a limit 6X ^ 0O. On the other hand, if
(jc0, y0) lies above F in S,, there is a possibility for T to cross F, provided
that at the crossing one can have a curve tangent passing through the
origin. At most one crossing is permissible. Thus 0(0 is certainly
ultimately monotone. To prove that the limit of 0(0 is 0O, we note that
0(0 is integrable, and the factor sin [20(0] is also integrable so L2-L,
must be an integrable function of t along T. Now L2- Lx is proportional
to the distance of (jc, y ) from F, so the distance function must be bounded;
and since it is ultimately monotone, it must tend to a limit ^0. If t0 is
infinite, the limit must be zero.
We have left to prove that tQ is finite. To this end we observe that
(12.5.12) implies that r 2f is bounded away from zero and infinity.
Integration gives an inequality,

mt <[r(0)] x-[r{t)V <Mt.

When t \ t0, this yields mt0< [r(0)] -1 < Mt0, so r0 is finite. Integration of
[r(s)]~2r'(s) from s = t to s = t0 gives
m(U-t)<[r{t)V<M(U-t),

suggesting a simple pole of r(r), jc(0, and y(t) at t = t0. This proves the
theorem. ■
Similar arguments may be given for solutions which are asymptotic to
the X-axis or the Y-axis.
It is time to start the analytical discussion and consider the various
series expansions. Here we must first examine the case where there are
solutions with a singularity of jc(0 and y(0 which is a simple pole. To
assure this a certain parameter q should not be a positive integer. But in
addition to the polar solution we obtain also nonlogarithmic psi series.
Since the system is autonomous, we can place the singularity anywhere.
The choice / = 0 simplifies the exposition.
QUADRATIC SYSTEMS 459

THEOREM 12.5.2
The system will have a solution of the form

x(t) = at1 + n2= 0 antn,

(12.5.14)

y(t) = /3r' + n=0

2/3nt°,
iff

(a,-bO(a2-b2)
(a,b2-a2b,)
is rcof a positive integer. Then the coefficients are uniquely determined, and

a= a,b2-a2bi
t2~\, P= aib2-a2bi (12.5.16)
In addition, there are solutions of the form

O 0mncr1+nq,
x(t,c) = iia (12.5.17)

y(t, c) = 2 2 bmncnr-I+nq,
vvnere c is an arbitrary parameter.
We omit the proof. The existence of additional solutions in psi series
was noted by Russell Smith. Such nonlogarithmic psi series also arise in
the discussion of the solutions whose trajectories are associated with Xx
and Yoo. Here the exponents are taken from the set

M = {/- ^; j = - 1, 0, 1, . . . , k = 0, 1, 2, . . .}.
If q is a positive integer we have

THEOREM 12.5.3

// q is a positive integer and if

y(t). t decreases to zero, then

is a solution which becomes infinite when

z(t) = 2oVn[tog(i)]t"-', (12.5.18)

where Vn(s) is a vector-valued polynomial in s of degree =s[n/q]. There
 SECOND ORDER NONLINEAR EQUATIONS

exists a K > 0 such that the series

2 Vn(s)e(, n)s (12.5.19)

n=0
converges in norm for s real and greater than the positive root of the
equation K(s + 2) = eqs.
Since the Emden equation may be transformed into a quadratic system
(12.5.5), the convergence proofs for Theorems 12.5.2 and 12.5.3 will also
constitute convergence proofs for the expansions listed in Theorem
12.4.2.

EXERCISE 12.5
1. Verify the reduction of (12.5.4) to the system (12.5.5).
2. What are the coordinates of U, V, Wl
3. Verify that the stated properties of L, and L2 imply (12.5.9).
4. Discuss the question of T escaping from Sx. Discuss also the geometry of the
intersection of T with F = 0.
5. Fill in missing details in the proof of the limit of 6{t) and for the finiteness of
fo.
6. Set up the recurrence relations for the coefficients of (12.5.14). How does the
fact that q is not a positive integer come into play? Verify (12.5.17).

12.6. OTHER AUTONOMOUS POLYNOMIAL SYSTEMS

This section is devoted to a discussion of higher polynomial systems, in

the main reproducing recent results of Russell Smith. In the system of
DE's
x=P(x,y), y = Q(x,y) (12.6.1)
P and Q are polynomials in jc, y with constant coefficients, and the dot
denotes differentiation with respect to a variable t. Both P and Q are of
degree n > 2. The problem is again to study the movable singularities,
their nature, and the associated series expansions.
We write P and Q as the sum of homogeneous polynomials

P(x, y) = 2 P*(x, y), <?(*, y) = 2 qAx, y), (12.6.2)

where pv and qv are homogeneous of degree v. The leading polynomials pn
and qn contribute most to the behavior of the solutions. Smith defines two
accompanying polynomials tt(z) and 7r*(z):
tt(z) = zpn(\,z)-qn(\,z),
(12.6.3)
Tr*(z) = zqn(z,\)-pn(z,\).
OTHER AUTONOMOUS POLYNOMIAL SYSTEMS 461

They are related by the identity

tt*{z) = -z"+1tt(z '). (12.6.4)

Since n > 2, there are no longer single-valued solutions having simple
poles. We do have algebraic solutions, however, provided that a certain
parameter A is not a positive integer. For n = 2, A reduces to the
parameter q of (12.5.15).
Suppose that tt(z) has a root z = (3 for which p„(\, P) ^ 0. Then set

A = j'lPL u=[(n- l)p„(l, p)(to-t)VKn l\ (12.6.5)

THEOREM 12.6.1

I/A is not a positive integer, then (12.6.1) has one and only one solution of
the form

x = ul(\ + g ajUj), y = u-^l + J) bjUJ), (12.6.6)

where the power series have positive radii of convergence. If A >0, then
(12.6.1) has also a family of solutions

x = u_1{l + P,[u, uA (c + h log u)]},

, ' 5 (12.6.7)
y = u '{l+P2[u,uA(c + h/og u)]},
where c is an arbitrary constant and Pi(u, v), P2(u, v) are convergent
power series without constant terms, whose coefficients are independent of
c. Here h = 0 unless A is a positive integer, and then h is a rational
function of A and the coefficients of P and Q.
Proof (sketch). We use Pk(u, v) as generic notation for a convergent
power series in u, v without a constant term. Smith uses several changes
of variables and inversions to reduce the system (12.6.1) to a form to
which known results on Briot-Bouquet equations will apply (see Section
11.1). The first substitution is f = 1/*, z = y/jc, which leads to

f=-*'4i)' f=*4i)-*4?) »"»

Here we take £ as an independent variable to get

dt_ -1 dz zP(\l&zl€)-Q(\l(,zl€)
dz £2p(ii(,zitr dt (pvi&zio - UZD*;
The virtue of this system is that the second equation involves only two
variables, z and £ Given any solution of this equation, we can substitute it
into the first equation and integrate to obtain a solution r(£), z(£) of
462 SECOND ORDER NONLINEAR EQUATIONS

(12.6.9). By inverting t = t(£ ), we get a solution z(0 of (12.6.8), from

which we can obtain a solution x(t),y(t) of (12.6.1) via the relation
jc = l/£y =zl(.
We substitute (12.6.2) into (12.6.9) and use the homogeneity to get

dt -|""2
+ £2p„ 2+- + (12.6.10)
* = P„+fp„-.
Ck = (Zp„ -gn)+f(ZPH-l-gn-l)+- • • + g"(zp0-qfo) (12611)

in which p„ are understood to mean p,(l,z), g,(l,z). The polynomial

(zpn - q„) in (12.6.11) equals 7r(z), whose root is /3. We substitute
z = j8 + f in (12.6.11), and expand its right-hand side in powers of f and f
by Taylor's theorem to get

^% = Af + Mf + *(f,f), (12.6.12)
where
_ tt'(/3) _j8p-i(l,j3)-q-i(l,i8)
p„(l,j6)' M Pnd./S)
and the power series £(£, £) involves all terms of the Taylor expansion of
degree >1. The same procedure applied to (12.6.10) gives

fr-£p)[l+PM'()] (12-6-13)
with the usual convention on the series. Equations (12.6.12) and (12.6.13)
form a system equivalent to (12.6.9). Here we use the results on
Briot-Bouquet equations as given in Section 11.1. Theorem 11.1.3 applied
to (12.6.12) gives
l = Pte,gx(c + h log 01 (12.6.14)
provided that A > 0. Here h = 0 unless A is a positive integer. In the
power series P4(w, v) = 2j>fc ajku'vk we have am = 0, a0i = 1, and the ajk are
independent of c. When substituted in (12.6.13), this expression for f
gives

-p„(l, 0) = r~2{l + P5[£ ^A(c + A log {)]}. (12.6.15)

When this is integrated with respect to £ the result is of the form

(n - l)p„(l, P)(to~ t) = r '{1 + P6[f, fA(c + h log £)]}. (12.6.16)

Something will be said later in justification of this step. Here t0 is a
constant of integration. Next we take the (n - l)th root on both sides of
OTHER AUTONOMOUS POLYNOMIAL SYSTEMS 463

(12.6.16) and expand (1 + P6)1/(" ,} in powers of P6 by the binomial

theorem. The powers of P6 are expanded in powers of f and of
f A(c + h log £); the terms are collected and the result is
u = f {1 + P7[f, f A (c + A log f )]} (12.6.17)
with u as defined above. By inversion we express f as a function of u to
obtain
f = u{\ + P8[w, wA(c + h log ii)]}, (12.6.18)
and now jc = 1/f gives the first equation under (12.6.7). Next we note that

y = r'z = Cl(P +0 = rlW + Mf, £ A (c + log f )]},

where we have now to substitute the value of f in terms of u as given
by (12.6.18). ■
This completes the proof of Theorem 12.6.1 except for some comments
on the admittedly very heavy analytical machinery. At various stages we
are dealing with double series of the form

v= 2 S ciksi+p[sx(c + h log s)]\ (12.6.19)

j *
On such series essentially three different operations are performed: (i)
substitution of one double series into another of the same type, expan-
sion, rearranging, and collecting terms; (ii) term wise integration and
rearrangement and collecting terms; and (iii) inverting (12.6.19), i.e.,
solving for 5 in terms of v. All these operations preserve the form of the
series and preserve convergence for small values of the variables.
Ad (i). This follows from repeated application of the double series
theorem of Weierstrass.
Ad (ii). Here the formulas

Jo L + m=l
I* sj(c + h log s)kds = Txi+l\hk J
2 gm{c + h log x)m\
(12.6.20)
r = (- D'J !(/ + \yk \ gm = (- \)mhk m Kj + ir(m - l)!]"1
carry the burden of the proof plus again the Weierstrass double series
theorem.
Ad (iii). This is too complicated for inclusion here, so the reader is
referred to the original paper of Smith.
Smith remarks that one can interchange the roles of the symbols jc and
y in (12.6.1); this gives a new system but of the same type. In the new
system the polynomials playing the roles of pn and tt in the old are simply
464 SECOND ORDER NONLINEAR EQUATIONS

the original qn(x, y) and 7r*(z). Here /3 is replaced by a, the root of

7r*(z) = 0, now with the condition qn(a, 1)^0. The new parameter A and
the independent variable are now

A*=
<?*(«, 1) u* = [(ii - l)<?„(a, l)]1*-". (12.6.21)

In terms of these quantities we get expansions for x and y of the types

(12.6.6) and (12.6.7) with A and u replaced by A* and u*, respectively.
The reader is encouraged to formulate the result in detail. It is referred to
below as THEOREM 12.6.2.
If a is a nonzero root of 7r*(z) = 0, then p = a 1 is a nonzero root of
7r(z) = 0 by virtue of (12.6.4), and calculation shows that the solution
given by Theorem 12.6.2 is essentially the same as that of Theorem 12.6.1.
It is only when a = 0 that something new is obtained.
At this juncture the question arises in (12.6.1) whether the two
theorems obtained exhaust the finite singularities. Here Smith proves the
following results, which we shall state without proof.

THEOREM 12.6.3

Suppose that (i) [pn(x, y)]2 + [qn(x, y)]2 > 0 for all real (x, y) * (0, 0), (ii) at
least one of the polynomials tt(z), 7t*(z) has a real root, and (iii) all the
real roots of 7r(z), 7t*(z) are simple roots. Then the real solutions of
(12.6.1) can have singularities at t = to only of the types given by Theorems
12.6.1 and 12.6.2.

Here the crux of the argument involves the existence of real roots of
77 (z), 7r*(z). These are real polynomials of degree n + 1, so a real root
must exist if n is even. If n is odd, n + 1 even, it is possible that there are
no real roots. In this case Smith shows that the argument function must
tend to ±oo. This is the result:

THEOREM 12.6.4

Suppose that neither of the polynomials tt(z), 7t*(z) has a real root. If a
real solution x(t), y(t) has a singularity at t = to, then both the functions
x(t), y(t) oscillate as t^U and their limits of indetermination are ±°°.
The possibility is not excluded that the solutions tend to infinity with t,
even spirally. With this remark we end the discussion of Dr. Smith's basic
paper.
 465
LITERATURE

EXERCISE 12.6
1. Verify (12.6.4).
2. Interpret A and the condition on pn(l, B).
3. Show that for n = 2 we have A = q as given by (12.5.15), and find B and
P2(l,/3).
4. Verify (12.6.8).
5. Verify (12.6.12).
6. Find the trajectories if r2 = x2 + y2 and P, Q are given by (i) yr2, -xr2, (ii)
yr4 + \xr2, \yr2 - xr\ and (iii) \x + yr4, \y - xr\

LITERATURE

Section 12.4 of the author's LODE has a bearing on the first three sections
of this chapter. The higher order elliptic Briot-Bouquet equations are
scarcely considered in the literature. Advanced treatises on elliptic
functions list the higher derivatives of p(z) in terms of p(z) and p'(z).
Since pak\z) and [p(2fc+1)(z)]2 are polynomials in p(z), these expressions
give examples of higher order elliptic Briot-Bouquet equations. See, for
instance, the first collection of formulas in:
Tannery, J. and J. Molk. Elements de la theorie des fonctions elliptiques. Vols. I-IV.
Gauthier-Villars, Paris, 1893-1902; 2nd ed., Chelsea, New York, 1972.
For an excellent and detailed presentation of the Painleve transcen-
dents and the 50-odd equations discussed see Chapter 14 of:
Ince, E. L. Ordinary Differential Equations. Dover, New York, 1944.
See also:

Boutroux, P. Recherches sur les transcendents de M. Painleve et l'etude asymptotique des

equations differentielles du second ordre. Ann. Ecole Norm. Super., (3), 30 (1913),
255-375; 31 (1914), 99-159.
Chazy, J. Sur les equations differentielles du troisieme ordre et d'ordre superieur dont
l'integrale generale a ses points critiques fixes. Acta Math., 34 (1211), 317-385.
Fuchs, R. Compt. Rend. Acad. Sci. Paris, 141 (1905), 555.
Gambier, B. Sur les equations differentielles du second ordre et du premier degre dont
l'integrale est a points critiques fixes. Acta Math. 33 (1910), 1-55.
Gamier, R. Ann. Ecole Norm. Sup., (3), 27 (1912).
Compt. Rend. Acad. Sci. Paris, 162 (1916), 937; 163 (1916), 8, 118.
Horn, J. Gewdhnliche Differ entialgleichungen beliebiger Ordnung. Sammlung Schubert, Vol.
50. Goschen, Leipzig, 1905, Chapter 14.
Painleve, P., Memoire sur les equations differentielles dont l'integrale est uniforme, Bull.
Soc. Math. France, 28 (1900), 201-261.
Sur les equations differentielles du second ordre et d'ordre superieur dont l'integrale
generale est uniforme. Acta Math., 25 (1902), 1-82.
466 SECOND ORDER NONLINEAR EQUATIONS

Picard, E. Remarques sur les equations differentielles. Acta Math., 17 (1893), 297-300.
Valiron, G. Equations Fonctionnelles. Applications. Masson, Paris, 1950. Chapter 8.

The literature on Emden's and the Thomas-Fermi equations is vast. A

few of the papers of more mathematical interest are:
Baker, E. B. Applications of the Thomas-Fermi statistical model to the calculations of
potential distributions in positive ions. Phys. Rev., 36 (1930), 630-647.
Coulson, C. A. and N. H. March. Moments in atoms using the Thomas-Fermi method. Proc.
Phys. Soc, (A) 63 (1950), 367-374.
Emden, R. Gaskugeln: Anwendungen der mechanischen Wdrmetheorie auf Kosmologie und
meteorologischen Probleme. Teubner, Leipzig, 1907.
Fermi, E. Un metodo statistico par la determinazione di alcune proprieta dell'atome. Rend.
Accad. Naz. Lincei, CI. Sci. Fis. Mat. Nat., (6), 6 (1927), 602-607.
Fowler, R. H. The form near infinity of real continuous solutions of certain differential
equations of the second order. Quart. J. Pure Appl. Math., 45 (1914), 289-350.
Further studies of Emden's and similar differential equations. Quart. J. Pure Appl.
Math., Oxford Ser. 2 (1931) 239-258.
Hille, E. Some aspects of the Thomas-Fermi equation. J. Analyse Math., 23 (1970), 147-170.
Aspects of Emden's equation. /. Fac. Sci. Tokyo, Sec. 1, 17 (1970), 11-30.
On a class of non-linear second order differential equations. Rev. Union Mat.
Argentina, 25 (1971), 319-334.
On a class of series expansions in the theory of Emden's equation. Proc. Roy. Soc.
Edinburgh, (A) 71 (1972/73), 95-110.
Karamata, J. Sur l'application des theoremes de nature tauberienne a l'etude des equations
differentielles. Publ. Inst. Math. Beograd., 1 (1947), 93-107.
and V. Marie. On some solutions of the differential equation y"(x) = f(x)y*(x). Rev.
Fac. Arts Nat. Sci., Novi Sad, 5 (1960), 415-422.
Mambriani, A. Sur un teorema relativo alle equazioni differentiali ordinarie del 2a ordine.
Rend. Accad. Naz. Lincei, CI. Sci. Fis. Mat. Nat., (6) 9 (1929), 142-144, 620-625.
Marie, V. and M. Skendzic. Unbounded solutions of the generalized Thomas-Fermi
equation. Math. Balkanica, 3 (1973), 312-320.
Rijnierse, P. J. Algebraic solutions to the Thomas-Fermi equation for atoms. Thesis, St.
Andrews University, 1968, 127 pp.
Thomas, L. H. The calculation of atomic fields. Proc. Cambridge Phil. Soc, 13 (1927),
542-548.
For quadratic systems we can restrict ourselves to:
Coppel, W. A. A survey of quadratic systems. /. Diff. Equations, 2 (1966), 295-304.
Dulac, H. Points singuliers des equations differentielles. Memorial des Sci. Math., No. 61.
Gauthier-Villars, Paris, 1934.
Hille, E. A note on quadratic systems. Proc. Roy. Soc. Edinburgh, (A) 72 (1972/73), 17-37.
Lefschetz, S. Differential Equations : Geometric Theory. 2nd ed. Wiley-Interscience, New
York, 1963.
Lotka, A. J. Elements of Physical Biology. Williams and Wilkins, Baltimore, 1925.
Volterra, V. La lutte pour la vie. Gauthier-Villars, Paris, 1931.
LITERATURE
467

For higher polynomial autonomous systems we refer to Dulac, loc. cit.,

and also to:

Dulac, H. Solutions d'un systeme d'equations differentielles dans le voisinage de valeurs

singulieres. Bull. Soc. Math. France, 40 (1912), 324-383.
Ince, E. L. Loc. cit., Chapter 13.
Malmquist, J. Sur les points singuliers des equations differentielles. Ark. Mat., Astr. Fys., 15,
Nos. 3, 27 (1921).
Sansone, G. and R. Conti. Non -linear Differential Equations (translated by A. H. Diamond).
Pergamon Press. Macmillan, New York, 1964.
Smith, R. A. Singularities of solutions of certain plane autonomous systems. Proc. Roy. Soc.
Edinburgh, (A) 72:26 (1973/74), 307-315.
 BIBLIOGRAPHY

I Bellman, R.chard. A Survey of the Theory of Boundedness,

Behavior of Solutions of Linear and Non-linear Differential Stabilit y, and Asymptotic
and Difference Equations
Office of Naval Research, NAVEXOS P-596. Washington,
D.C. (1949).
Stability Theory of Differential Equations. McGraw-Hill,
New York 1953
7heaoti LhdWir
theoretische n Grundlage dargestellt. gewdhn,ich<»
Springer- Differentialgleichungen
Verlag, Berlin 1953 auf funktionen-

B6cher _ Maxime
dtfferen nelles. lineaire
Leconss etsurtears
lesdevelop
methoments
deTdTmoderne
^turm s.dans 'lar-Villa
Gauthie theoriersdesPari
equatio
ngns
vtl^H^w- ^T, Differe"'ia"^forma,ionen
Verlag der Wissenschaften. BerlijiJ.967. 2:^r3^ VEB'-Demsc11er ucuiM-ner

BOU%%^!ZTJeTu^aliqUeS- VO' IV: F°nC,ions dW """able reelle. Chapter

(~~ P^JT d,fferen,'e"eS- ActUalit6s ^ientifiques et indus.rielles, .132. Hermann,
^premier ordre. uauthi
cZwer-Vil
Zu'lars, 'T™
Paris, 1908. <>"' * *"""i0"S "^entieUes du
^Zmie^Zle

Briot^Charles and Jean Claude Bouquet. 'Theorie des

fonctions elliptiques. 2nd ed. Paris,
CCesari, Lamberto. Asymptotic Behavior and Stabili
ty Problems in Ordinary Differential
... Equations. Ergebn.sse d. Math., N.S., No. .6, Spring er- Verlag, Berlin, 1959

C'Sw-HilC N^^^inSOn- The°" °< ^erential Equations.

(^oppel, W. A. StabM^dAlyln~ptotic Behavior of Differential Equation
s. Heath, Boston,
_. Disconjugacy. Lecture Notes in Mathematics. No.
220. Springer-Verlag. Berlin, 197.
CcauZ^
Emden R. Gaskugeln: Anwendungen der mechanischen
Wdrmetheorie auf Kosmologie
und meteorolog,schen Probleme. Teubner, Leipzig 1907 kosmologie
110ns. 3 vols.
^onsAi McGraw-Hill,
To^Mcr8, New York, a"d
F ^teUinm- R TriC°mi- higher Transcendental Func-
1953-1955.

^Tondt.^oT1 6 VO'S- Cambndge Universi,y Press'

468
 BIBLIOGRAPHY

6
^ftartman, Philip. Ordinary Differential Equations. Wiley^New York, 1964.
" Hayman, W. K. Meromorphic Functions. Cjarendfln Prp^jlxforrl 1964.
"^Hilb, Emil. Lineare Differentialgleichungen im komplexen Gebiet. Encyklopddie der
mathemetischen Wissenschaften. Vol. 11:2. Teubner, Leipzig, 1915.
Hille, Einar. Analytic Function Theory. 2 vols. Ginn, Boston, 1959, 1962; Chelsea, New
York, 1973.
Analysis. Vol. II. Ginn-Blaisdell, Waltham, Mass., 1966, Chapter 20.
Lectures on Ordinary Differential Equations. Addison-Wesjey , Reading, Mass., 1969.
Methods in Classical and Functional Analysis. Addison-Wesley. Reading, Mass., 1970.
Horn, Jakob. Gewdhnliche Differentialgleichungen beliebiger Ordnung. Sammlung Schubert,
Vol. 50. Goschen, Leipzig, 1905.
Hukuhara, M., T. Kimura, and T. Matuda. Equations differ entielles ordinaires du premier
ordre dans le champ complexe. Publications of the Mathematical Society of Japan, No.
p_ 7, 1961.
\ Ince, E. L. Ordinary Differential Equations. Dover, New York, 1944.
Kamke, E. Differentialgleichungen: Losungen und Losungsmethoden. Vol. 1, 3rd ed.
Chelsea, New York, 1948.
Klein, Felix. Vorlesungen iiber die hypergeometrische Funktion. Springer-Verlag, Berlin,
1933.
Lappo-Danilevskij, J. A. Theorie des systemes des equations differentielles lineaires. Chelsea,
^ New York, 1953.
Lefschetz, Solomon. Differential Equations : Geometric Theory. 2nd ed. Wiley-Interscience,
New York, 1963.
Massera, J. L. and J. J. Schaffer. Linear Differential Equations and Function Spaces.
Academic Press, New York, 1966.
Nevanlinna, Rolf. Le theoreme de Picard-Borel et la theorie des fonctions meromorphes.
Gauthier-Villars, Paris, 1929.
Painleve, Paul. Legons sur la theorie analytique des equations differentielles, profesees a
Stockholm, 1895. Hermann, Paris, 1897.
Poincare, Henri. Les methodes nouvelles de la mechanique celeste. 3 vols. Gauthier-Villars,
Paris, 1892, 1893, 1899.
Sansone, G. and R. Conti. Non -linear Differential Equations (translated by A. H. Diamond).
Pergamon Press, Macmillan, New York, 1964.
Schlesinger, Ludwig. ffaTCdhuch def^ffiebrie der linearen Differentialgleichungen. Vols. 1,
2:1, 2:2. Teubner, Leipzig, 1895, 1897, 1898.
Einfuhrung in die Theorie der Differentialgleichungen mit einer unabhangigen Veriab -
len. 2nd ed. Sammlung Schubert, Vol. 13. Goschen, Leipzig, 1904.
Titchmarsh, E. C. Eigenfunction Expansions Associated with Second Order Differential
Equations. Clarendon Press, Oxford, 1947.
Tricomi, Francesco. Fonctions~h~ypergeoriT$triques confluentes. Gauthier-Villars, Paris, 1960.
Equazioni differenziale. Einudi, Turin, 1965.
Trjitzinsky, W. J. Analytic Theofy^of~Non -linear Differential Equations. Memorial des
Sci. Math., No. 90. Gauthier-Villars, Paris, 1938.
Valiron, Georges. Equations fonctionnelles : Applications. 2nd ed. Masson, Paris, 1950.
 470 BIBLIOGRAPHY

Walter, Wolfgang. Differential- und Integralungleichungen. Springer-Verlag, Berlin, 1964.

Watson^ G. N. Theory of Bessel Functions. 2nd ed. Cambridge University Press, London,
'
1962.
Weiss, Leonard. Ordinary Differential Equations: 1971 NRL-MRC Conference. Academic
Y~ Press^ New York, 1972. — ■~
^_WJiittaker, E. T. and G. N. Watson. A Course in Modern Analysis. 4th ed. Cambridge
University Press, London, 1952. " ~-
Wittich, Hans. Neuere Untersuchungen tiber eindeutige analytische Funktionen. Ergebnisse
td. Math., N.S., No. 8. Springer-Verlag, Berlin, 1955; 2nd ed., 1968.
ida, Kosaku. Lectures on Differential and Integral Equations. Wiley-Interscience, New
York, 1960.
INDEX

To facilitate the use of the Index it is partly analytical. Thus, e.g., Bessel's equation is listed
under BESSEL and Differential Equations, Linear but not under Equations.

ABEL, Niels Henrik (1802-29), 52, 151 series 451, 466

elliptic functions, 262 BANACH, Stefan (1892-1945), 5
-ian integrals, 263 algebra, 6
Wronskian, 324 fixed point theorem, 9
ACZEL, Janos, 39 space, 5
Addition, 2 BANK, Steven S., 143
of matrices, 8 BARNES, Ernest William (1874-1955), 276
of operators, 6 integral for Bessel function, 279
theorems for: integral for hypergeometric function,
elliptic functions, 430-1
tangent function, 106 276-8 - notation, 221
Pochhammer
Adjoint, 254 Basis, 6
anti, 254 Bateman papers, see ERDELYI; MAGNUS;
self, 254 OBERHETTINGER; and TRICOMI
Affinity, 117 BEESACK, Paul R., 283, 300-1, 304, 306-8,
AFT (Analytic Function Theory), 39 320, 391,401
AHLFORS, Lars V., viii, 61 BELLMAN, Richard, 468
characteristic function, 118-9 BENDIXSON, Ivar (1861-1935), 16, 74
Schwarzian and conformal mapping, 391 theorem, 16, 39, 70, 74
Algebra, 2 BERNOULLI:
Banach, 6 Daniel, 41, 103-4, 212
Analysis situs (topology), 97 Jakob, 41, 104
Anticlinal, 289 ^ Johann, 41, 104
Argument, 17 equation, 104
interchange of, and parameter, 254 BESSEL, Friedrich Wilhelm (1784-1848):
principle of, 29 equation, 103, 211-6, 279, 312-7
)tics:
quasi, 178
loutroux, 444- functions, 1st kind, 212
the plane, 178-191 2nd kind, 213
m the real line, 171-178 3rd kind (Hankel), 215
.cubic equation, 178 Bestimmtheitsstelle, 147
Autonomous: BEZOUT, 78
equation, 47, 80 method, 78
polynomial system, 455-465 BIEBERBACH, Ludwig, 198, 360, 400,
468
BAKER, E. B: Binormal, 1 1 1

471
 472 INDEX

LiLLRKHOFF, George David (1884-1944), Hadamard formula, 21

44, 74 extension, 251
' canonical form, 361 , 378 integral, 24, 31
generalized Laplace transform, 271-4 theorem, 25
regular singular point, 337-8, 372 kernel, 22, see also Kernel
v Riemann problem, 201, 357, 370-2 majorant, 57-60
^second order linear DE, 370-1, 373 Riemann equations, 19
BLISS, Gilbert Ames, 102 .sequenced ■
BOCHER, Maxime, 247, 468 \CAYLE
;AYLEY, Arthur (1821-95)
BOLZANO, Bernhard (1781-1848): J
Weierstrass theorem, 4 Center jlfon - theorem
vortex, 314 9
BORUVKA, Otakar, viii, 468 CESARI, Lamberto, 468
Boundary: CEBYSEV, Pafnutij Livovic (1821-94):
of domain 9 (D), 18 polynomials, 387-8 388
entrance, 173 minimax property,
natural, 23 Characteristic function:
movable, 434 of rational function, 118
BOUQUET, Jean-Claude (1819-85), 79, 402 of R. Nevanlinna, 117
Briot- elliptic functions, 79-80, 102, 418- spherical of Shimizu and Ahlfors, 118-9
428,431,435-9,468 CHAZY, Jean, 431, 439, 465
1st order equations singular at the origin, CLAIR AUT, Alexis-Claude, 41
402-416 Closure, 4
BOURBAKI, N., 468 relation, 210
BOUTROUX, Pierre (1880-1922), 86, 102, CODDINGTON, Earl A., 198, 371, 468
135,402,411,432-3,468 COLOMBO, Serge, 282
asymptotics, 283, 319, 444-8, 468 Commutator, 344
solutions, of exponential type, 86, 312 Daletzskij's formula, 347, 353
of power type 87,413, 415-6 Foguel's theorem, 347
Branch point: resolvent, 347, 353
algebraic, 34 spectrum, 344
movable, 83-5 Comparison theorems, 300-8
logarithmic, 77 Component, maximal, 132
transcendental, 76, 79 Condition:
BRIOT, Charles (1817-82), see Bouquet Carathdodory, 13
LBROUWER, L. E. J.: • Lipschitz, 5
(fixed point theorem, 9, 44, 74 Cone:
BURMANN, Hans-Wilhelm, 279, 282 ,t__pjositive conal harmonics, 211
^Conformal Mapping:
CACCIOPOLI, R.,44, 74
Calcul des limites, 51, 74 (application of Schwarzian to, 377-383
\of circular polygons, 378-383
Calculus of residues, 28-30 Wjnfldular functions, 397-8
Lindelof, 83 Conjugacy, 203
CARATHEODORY, Constantin (1873- of complex numbers, 20
1950): Connectedness, 18
condition, 13 simple, 20
conformal mapping, 377 \ CONTI, R.,432, 469
/ uniqueness theorem, 51 LXontinuation, analytic, 33-8
^\CAUCHY, Augustine-Louis, Baron de A of solution, 81-96, 192-4, 327-333
/ (1789-1857), 17, 24,42,52
C-calcul des limites, 51, 74 v Contraction, 189
{Continuity,
 INDEX 473

singular, 328
(j;xed point theorem, 9-11 nonlinear, 41
( Convergence: order of, 4 1
absolute, 21
equi-, 119-121 solution, 41
exponent of, 1 20 general, 43
holomorphy preserving, 27-8 particular, 43
induced, 27 singular, 43
in norm, 99 Differential equations, linear, 144-373
radius of, 21 first order, 144-8
uniform, 27 matrix DE, 322-373
L^onvexity: second order, 148-153
horizontal, 286 nth order, 321-373
theorem, 290-1 complex oscillation theory, 283-320
vertical, 286 /rr. integral representation, 253-283
[Coordinates, confocal,\£4£/ special second order, 200-248
0)PPEL, W. A., 194, 199, 328, 425, 466, Bessel, 211-216, 278, 312-5
468
COULSON, C. A., 450, 466 Floquet,247242-3
Halm,
CRELLE, August Leopold (1780-1855), Hermite-Weber, 228-238, 315-7
Journal, 151 Hill, 244
(Curvature: hypergeometric, 200-8, 265-270, 308-
(radius of, 111 310,399-400
Rector, 111 confluent, 226
kCut, 20 Laguerre, 227
binder functions:
Lame\ 245-6
circular (= Bessel), 229-230 Laplace, 216-228
elliptic, 234-242 Legendre, 208-211, 260-4, 279-280,
tabolic, 230-3
310-5 234-242,317-8
Mathieu,
DALETZKIJ, Yu. L., 347, 353, 372 Titchmarsh, 188-191, 319
Defect, 126 Differential equations, nonlinear, 40-143,
relation, 126 374-467
Dependence: autonomous 1st order Briot-Bouquet, 80,
linear, 2, 151 91-6, 102,416-432
linear in-, 2, 151 autonomous 2nd order Briot-Bouquet,
Derivative, 19
Determinant, 7 435-9 polynomial systems, 460-7
autonomous
infinite, 248 autonomous quadratic systems, 455-460
Diameter, topological, 4 Emden and Thomas-Fermi DEs, 448-455
Dicylinded, 44 existence and uniqueness of solutions, 40-
^Difference equation, 257, 277, 279, 281 75
I of Bessel functions, 213, 280 first order Briot-Bouquet DEs singular at
Uif Legendre functions, 211, 274 z = 0, 402-416
Difference set of spectrum, 346, 367 Painleve transcendents, 439-448
^_PjfJ«i^ntiaJ equation = DE, 40 asymptotics of Boutroux, 444-8
linear, 144
Riccati DE's, 103-143
homogeneous, 144 generalizations, 138-143
nonhomogeneous, 145 singularities, 76-102
self-adjoint, 179, 254 Differential forms, quadratic, 211, 233
anti, 254 Differential operators:
 INDEX

incar L, L*, M, M*, 254-6 natural, of curve, 118

# = zd/dz, 203,276,321-2 \ Equi-convergence, 1 19, 120, 121
Dimension, 2 \ Equivalence of:
Discriminant equation, 77 \1 matrices,
m 101
Disk(s):
circular, 18 l^paths,
ERDELYI,330 A.:
closest packing of, 133-4 Bateman papers, 248, 468
Distance, 1 EULER, Leonard (1707-83):
chordal, 18 Bessel's DE, 212
postulates, 1 beta function, 109, 206, 265
Domain, 18 DE, 343
boundary of, 18 Mascheroni constant, 209
Dominant, 56, 63-67, 171-6 transform, 258-265
sub-, 171-7 hypergeometric, 265-270
Dot:
Legendre DE, 260-1
Newton's symbol, 97 Existence theorems, 40-75
product (= inner product), 9, 114 concept of solutions, 40-4
DULAC, Henri, 251, 281, 431-2, 467-8 dominants and minorants, 63-7
fixed point method, 44-7
Eigenvalue, 8 indeterminate coefficients, 41-51
Elements, 1
majorants,
calcul des51-7
limites, 52
algebraic, 34
infinitary, 34 Cauchy majorant, 57-60
logarithmic, 392 Lindelof majorant, 60-3
negative, 2 successive approximations, 47-51
neutral, 2 Caratheodory condition, 51
regular, 34 variation of parameters, 67-74
zero, 3 Exponential function, 19
Elliptic functions, 79, 80, 91, 102, 427-431 of matrices, 99
DEs satisfied by, 126-7, 143, 416427, Exterior, 195
434-9 Extremum, 284
of Jacobi, 91, 127, 207, 262
of Weierstrass, 96, 126-7, 143 FABRY, Eugene (1856-1944):
EMDEN, R.: subnormal series, 362, 373
equation, 433, 448-453, 456, 466 FERMI, Enrico, 449, 466
Gaskugeln, 448, 466, 468 Thomas-F. equation, 96, 433, 446-453
Equation(s): Field:
^ Cauchy-Riemann, 19 complex, 2
characteristic, of matrix, 8 real, 2
defining algebraic function, 77 definite vector, 297
difference see under Difference equation Fixed point, 9
differential, see above under Differential theorems:
equation; Differential equation, Banach, 9
linear; Differential equation, non- Brouwer, 9
linear; or under the appropriate Volterra, 1 1
author
applied to DEs, 44-7
discriminant, 77-8 applied to functional inequalities, 14-6
\ indicial, 157, 344 Fluxion:
J Laplacian, 19 al equations, 41
I minimal, of matrix, 244, 346 FLOQUET, Gaston (1847-1920):
INDEX 475

theorem, 245, 248 generating of :

FOGUEL, R. S.: Bessel functions, 213
theorem, 347, 372 Hermite polynomials, 231
Forms: Laguerre polynomials, 227
indeterminate, 96-102 Legendre polynomials, 210
quadratic differential, 211, 233 Harmonic, 21
FORSYTH, A. R.: Holomorphic, 17
zero-free region, 284-5, 320, 468 hypergeometric, 200-8, 260-270, 276-280,
FOURIER, Joseph (1768-1830): 309-311, 383-8, 394401
coefficients, 214, 241 f(a,b,c), 205-6, 269
Legendre coefficients, 210 meromorphic, 32
ring, 262 value distribution, 114-129
Thdorie analytique de la chaleur, 212 modular, 263, 394-401
FOWLER, R. H., 456, 466 potential, 19
FREDHOLM, Ivar (1866-1927), 16 rational, 32
FRENET, Jean-FrCdCric (1816-1900): theta,
formulas, 111-3 null, 263-5
265
FRIEDLAND, S., 391 transcendental, 32
FROBENIUS, Georg (1849-1917), 147, Functional, 325
342, 344, 349, 352 equation, 37
FUCHS, Lazarus (1833-1902), 76, 102, inequalities, 12-7
147-8, 263,342, 344,355,439 multiplicative, 325
Bestimmthetsstelle, 147
-ian class, 147, 353-360 GAMMBIER, B., 439, 441, 465
Fundamentalgleichung, determinierende, Gamma function:
344 Hankel's integral, 222, 366
fundamental system, 148 logarithmic derivative of, 209
indicial equation, 157, 344 reciprocal, 222
fonctions Fuchsiennes, 263 of matrices, 366-7
FUCHS, R., 439, 440, 465 GANTMACHER, F. R., 361, 373
Function(s): GARNIER, R., 439, 465
algebraic, 32, 77-8, 102, 383-8 Riemenn-Hilbert problem, 372
analytic, 34 Gaskugeln, 448, 466
automorphic, 128, 263 GAUSS, Carl Friedrich (1775-1855), 200-1,
Bessel, 211-6, 278, 312-5 211
beta, 189, 200 hypergeometric series, 201
branch of, 34 Generators:
characteristic, see above under Character- of group, 196
istic function rectilinear, 113
confluent hypergeometric, 226 Geometry:
cylinder-, see above under Cylinder of numbers, 3
functions of solutions of DEs, 97
elliptic, see above under Elliptic Genus = deficiency, 127
functions Picard's - theorem, 127-8
entire, 32 elliptic case, 127
exponential, 19
of matrices, 99 G-net,hyperelliptic
298 case, 128
fractional linear = Mobius, 19 GOL'DBERG, A. A., 140, 142
gamma, reciprocal, 222 Grade of singularity, 162, 341
of matrices, 366-7 GREEN, George (1791-1841):
 INDEX
476

formula, 255 spheroidal, 211

transform, 286-292 HARTMAN, Philip, 469
GRONWALL, Thomas Hakvin (1872-1932): HAYMAN, W. K.,469
area theorem, 389, 393 HEINE, E. H.:
Borel theorem, 211
T~______lemma, 15-6, 39
"GROUP: Kegelfunktionen, 211
1 anharmonic, 202 HELSON, Henry, 372
conjugate element, 203 Thermite, Charles (I822-1900), 231
^ generators, 330 1 polynomials, 231
finitely generated, 196, 330 ' Weber DE, 228-233, 315-7
isomorphic, 332 HILB, Emil, 469
modular, 395 HILBERT, David (1862-1943):
monodromic, 192-8, 330-3 Riemann problem, 201, 357
motions in the plane, 383 HILL, G. W.:
projective, 374 equation, 244
rotations of the sphere, 384 infinite determinants, 244, 248
cyclic, 383 HILLE, Einar, 38-9, 188, 308, 318, 320,
dihedral, 383 372
icosahedral, 383 psi series, 249-253, 281
octahedral, 383 **- zeros of Hermite-Weber and Mathieu func-
tetrahedral, 384 tions, 248
Growth: ^HOCHSTADT, Harry, viii, 140
for approach to singularity, 333-342 HORN, Jakob, 251, 281, 363, 373, 469
estimates of, 31-2, 162-170, 333-343 Horn angles, 362-3
exponential, 411 Hyperboloid, 114
Liouville's theorem, 32 Hypothesis F, 180
power type, 413-8 HSU, Ih-Ching, viii, 140
properties, 420-6 HUKUHARA, M., 432, 469
v. Koch-Perron theorem, 163-4, 198, 333- HURWITZ, Adolph, 283-4, 309, 312, 320
341, 372
in vertical strip, 335 Idempotent, 242, 367
GUDERMANN, Christoffer (1798-1851), Identity theorem, 28
63,430 Implicit function theorem, 63, 75
1NCE, E. L., 248, 281, 320, 371,465,469
HADAMARD, Jacques (1865-1963), 21 j Index = winding number, 196
Cauchy- formula, 21 L Indicatrices, 289
extension, 251
entire functions of finite order, 138 /inequality:
functional, 12-17
version of Cauchy 's integral, 114, 120 categorical = determinative, 12
HALANAY, A., 75 Caratheodory, 13, 51
HALM, J.: Lipschitz, 13
equation, 247 Nagumo, 13, 16
HAMILTON, Sir William Rowan (1805-65): use of fixed point theorem, 14-17
- Cayley theorem, 9 Gronwall's lemma, 15-6
HANKEL, Hermann (1839-73): Integral:
Bessel function, 215 Barnes, 276-280
integral, 222 Cauchy, 24-31
y Harmonics: elliptic, 262, 267
conical, 211 hyperelliptic, 264
solutions of Laplace's equation, 21 integrand, 25
 477
INDEX

integrator, 25 generating function, 227

line, 24 LAINE, Ilpo, 140, 142
Riemann-Stieltjes, 25 LAME, Gabriel (1795-1870):
Interior, 18, 195
Isochrone, 41 DEs, doubly
with 245-6 periodic solution, 246
LANDAU, Edmund (1877-1938), 31
JACOBI, Carl Gustav Jacob (1804-51), 394 CLAPLACE, Pierre-Simon (1749-1827):
elliptic functions, 63, 91, 127, 262
theta functions, 263-5 -|DE,
partial216-223
DE, 19
JACOBSON, N.: integral, 217
buffer condition, 345 The"orie analytique des probability, 216,
prime ring, 345 281
JENSEN, J. L. V. W.: transform, 270-8, 217
formula, 114, 116 transformation, 228
JORDAN, Camille (1838-1922): Ijaplacian, 19
curve theorem, 195, 329 change of variables in, 228
JULIA, Gaston, 148 Laplace-Birkhoff transform, 365-371
EaTpO DANILEVSKIJ, J. A., 344, 469
KAMKE, E., 200, 248, 281 , 469 representation theorem, 356, 372
KARAMATA, Jovan, 466 Riemann problem, 356-360, 373
JCELLOGG, Oliver Dimon, 44 LAURENT, Pierre- Alphonse (1813-54):
Kernel: series, 26
Cauchy, 25-6 LAVIE, Meira, 375,401
representation by Bessel functions, LAVOINE, J., 282
215-6 Law, parallelogram, 20
representation by Legendre functions, LEFSCHETZ, Solomon, 102, 456, 466, 469
lv 210 LEGENDRE, Adrien-Marie (1752-1835):
I of integral equation, 1 1 elliptic integrals, 262
V>of integral transform, 255 equation, 208-211, 260-4, 310-1
KIMURA, T., 432, 469 polynomials, 208-211
KLEIN, Felix (1849-1925), 247 LIEBNIZ, Gottfried Wilhelm (1646-1716),
automorphic functions, 263 42, 104
fonctions kleiniennes, 263 LEVINSON,469 Norman (1912-75), 198, 371,
hypergeometric functions, 461
uniformization of, 395, 401 LIAPOONOV, Alexander Mikailovic, 338
KOCH, Helge von (1870-1924), 16 LIE, Marius Sophus (1842-99), 112
infinite determinants, 244-5, 248 Frenet formulas, 111-2
order estimate, 163, 198, 338-341, 372 LINDELOF, Ernst (1870-1940):
KOEBE, Paul (1882-1945), 377 calculus of residues, 63
KRONECKER, Leopold (1823-91): implicit function theorem, 63, 75
6,6 majorant, 60-3, 75
KUMMER, Ernst Eduard (1810-93), 200 Lines:
"Schwarzian," 374 asymptotic, 113
twenty-four solutions, 202 of influence, 297

LAGRANGE, J. L. (1736-1813): ^LIOUVILLE,

I theorem, 32 Joseph (1809-82), 32
Bessel functions, 212 L-^, transformation, 179
identity, 250 / LIPSCHITZ, Rudolf (1832-1903), 5
LAGUERRE, Edmond (1834-86): J Bessel functions, 212
polynomials, 227 J condition, 5
 INDEX
478

LOBATSEVSKIJ, Nikola] Ivanovo (1773- Matrix(ces), 7

1856): algebraic operations on, 8
geometry, 385 analytic functions of, 243, 347
A. J., 455, 466 exponential functions of, 99
Y^LPTKA,
Locus: logarithm, 244, 247
of inflection, 291-2 power, 347 equation, 8, 98, 346
of tangency, 300 characteristic
LODE (Lectures on Ordinary Differential characteristic values, 8
Equations), 39 characteristic vector, 8
Logarithm, 20 equivalent, 99
hyper, 355 Hamilton-Cayley theorem, 9
logarithmic derivative of gamma function, idempotent, 346
209 inverse, 8, 197
logarithmic derivative of meromorphic minimal equation of, 244, 346
function, 123 nilpotent, 197
estimate thereof, 1 23 norm of, 8
estimate of proximity function, 123 of linear transformation in Cn, 7
A of regular matrix, 244, 247 property P, 192
Loop: regular, 8
double, 197,219 resolvent of, 346
simple, 218 similar, 152
singular, 8
MAGNUS, W.: space 90?n, 8
Bateman papers, 248, 268 spectrum, 8
T_JV1AHLER, Kurt, 456 trace, 324
J MALMQUIST, Johannes (1888-1952), 16, transit, 152, 196
/ 103 unit, 99
Briot-Bouquet DE, 405, 432 Matrix differential equation, 193
theorem, 113, 129-137, 143 first order in 2fln, 322-371
generalizations, 138-143,417 fundamental solution of, 327
Majorant, 52 analytic continuation, 326-8
, 1-7 monodromic group, 126-8, 330-3
Ceatuhcohdys',s 5 57-60 Wronskian, 150-1, 324-8
Lj mLindAelNoIf, s, 60-3
MAMBRI A., 449, 466 hypergeometric, 202-3
with periodic coefficient matrix, 243-4
,
Mapping d,5 of resolvent, 408
bounde 5
conformal, 377-383, 365-7, 395-6 vector, 163,469
.
onto (= surjection), 15 MATUDA,:T
order preserving, 14 Maximautmi, 407-8
MARCH, N. H.,450, 466 LRicc
MARIC, V., 466 principle of, 30
MASANI, P., 361, 373 MCTA (Methods), in Classical and Functional
MASSERA, J. L.,469 Analysis 39
L modulus, 31
MATELL, Mogens, 169, 198 Mean:
r*MATHIEU Emile (1835-90), 234 square modulus, 129
equation, 234-242 value, 30
functions, 236 MELLIN, Hjalmar (1854-1933):
associated, 336 integrals, 256
boundary value problem, 235-8 transform, 274-280, 282
INDEX 479

MILLER, John Boris, 281 diagram, 94

Minimal equation, 244 fluxional equations, 41
MINKOWSKY, Hermann, 3 indeterminate coefficients, 41 i
Minorant, 63-7 trial and error, 47 J
MITTAG-LEFFLER, Gosta (1846-1927), NORLUND, Niels Erik:
16 difference equations, 279-280, 282
, star, 192 Norm, 4
MOBIUS, August Ferdinand (1790-1868): of matrix, 8
function, 115 Normal:
linear fractional function, 19, 212 bi, 111
Modular: principal, 1 1 1
function, 263, 395 series, 362
group, 395 type of meromorphic function, 118
Modulo (up to multiples of), 146 NOSHIRO, Kiyoshi, 142
Modulus:
of a Jacobi elliptic function, 263, 430 OBERHETTINGER, F. Bateman papers,
complementary, 430 248,468
maximum, 31 OKUBO, K. A., 370, 373
mean square, 129 Optimization of majorant, 39
MOLK, J., 465 Order(ing):
Multiplication: Archimedean, 3
matrix, 8 of branch point, 79
left right: of DE,41
commutator, 344 of elliptic function, 424, 428
by a matrix, 394 of entire function, 119
spectra, 345 of meromorphic function, 119
scalar, 2 partial, 2
proper,
reflexive,2 2
NAGUMO, Mitio:
inequality, 13, 16 transitive, 2
uniqueness theorem, 13 solution, 162
Negative, 2 of zeros of 195,
ww', 329
268-9
NEHARI, Zeev, 288, 300 Orientation,
number of zeros in disk, 303, 306, 320 OSGOOD, William Fogg, 377
Schwarzian and univalence, 320, 391,
401 Packing, closest, 133
Neighborhood, 4,18 PAINLEVE, Paul (1863-1933), 76, 79, 87,
infinitudinal, 414 102,439-440, 465,469
polar, 132, 139 theorem on:
NEUMANN, Carl (1832-1925), 213 algebraic branch points, 83-5
NEUMANN, F., 210 determinateness, 87-90
NEVANLINNA, Frithiof, 61 infinitary solutions, 85-7
NEVANLINNA, Rolf, 61 transcendents, 43844
abstract of theory, 114-129 asymptotics (Boutroux), 444-8
characteristic function, 117 PAPPERITZ, Erwin (1857-1933):
enumerative function, 116, 124 uniformization of hypergeometric func-
fundamental theorems, 117, 125 tions by modular functions, 394-
partial converse, 391, 401 401
- proximity function, 116 Parallelogram:
NEWTON, Sir Isaac (1642-1727): law, 20
 INDEX

GD
period, 427 Point:
Parameter: fixed, 9
analytic dependence on, 73-4 at infinity, 18
expansion in powers of a, 74-5 singular, 76
external, 67 star, 314
initial, 297 nscendental critical, 79
interchange of argument and, 259 JPOISSON, Simeon-Denis (1781-1840):
internal, 67 equation, 229, 230, 234
matrix, 202 POKARNYI, V. V., 300, 306, 391, 401
transformation of, 203 . Pole, 27
variation of, 67-74, 314-5 \ Polynomial:
Part: CebySev, 387
imaginary, 17 minimax property, 388
principal, 27 Hermite, 231
real, 17 Jacobi, 387
PEANO, Guiseppe (1858-1932), 71 Laguerre, 227
Period: Legendre, 210, 387
additive, 154 POSCHL, Klaus:
of elliptic functions, 427, 430 growth of solutions, 165, 198
of expomemtial functions, 20 Potential:
parallelogram, 427 logarithmic, 19
approximate, 447 Newtonian, 245
PERRON, Oskar: Postulates:
order estimate, 163, 198, 338, 372 distance, 3
norm, 4
^^JJhase portraits, 456
/ PICARD, Emile (1856-1941): 1 Power:
^■"i fractional, 20
/ genus theorem, 127
method of successive approximations, 47- ) with matrix exponent, 347
51, 75 I type (rational by Boutroux), 87, 413, 415-
second order nonlinear DE's with only 416
fixed critical points, 439, 466 Principle:
L valutheorem, 27, 126, 263 argument, 29
e, 126 maximum, 28, 30-1
PLEMELJ, Josip (1873-1967): permanence of functional equations, 37
Riemann-Hilbert problem, 201, 357, PRINGSHEIM, Alfred (1850-1941), 23
372 theorem, 23
POCHHAMMER, L.:
/Product:
rn es no ta ti on , 231
ri Ba
POINCARE, Henri (1854-1912), 35, 53, j dot- = inner,
of linear 8
transformations, 6
,
97 10 2, 35 5 of matrices, 8
analysis situs, 97 Projection:
asymptotic power series, 52, 57, 223 * idempotent, 346
automorphic functions, 128, 263, 378
celestial mechanics, 75-5, 469 Projereog
Iste 141 18, 118, 394
ty, c,
ctivirphi
countability of elements, 35
curves defined by DE's, 97 operty, 1
infinite determinants, 244 Pof moalter,ix, 192
majorants, 52 Pseudop 250
noneuclidean geometry, 388 psi n 110 405, 408, 454, 459-
,
tio249-253,
rela468
phase portraits, 456 tseries,
 INDEX 481

logarithmic, 249, 451, 454, 460-3 Cauchy, equations, 19

PUISSEUX, V. A. (1820-83): hypergeometric integrals, 304
method of, 94, 434 lectures, 394
P-function, 201, 394
Quadratures, 104, 141 problem, 43, 356-360, 372, 394
Quadrics, central confocal, 247 surface, 20
Stieltjes integral, 25
sums, 25
j Radius:nvergence,
I of co 21 Riemann-Birkhoff problem, 341
1 associated radii, 36 Riemann problem, 43, 356-360
1 of curvature, 111 history, 357, 394
] of torsion, 111 ROBERTSON, M.S., 401
RAINIERSE, Petrus Joannes, 466 Residue(s), 28
Ramification: calculus of, 28, 63
completely ramified value, 126 theorem, 28-9
index, 126 Resolvent of:
total algebraic, 126 commutator, 353
Rank of irregular singularity, 162 matrix, 346
Rearrangement of power series: Resultant, 78
direct, 22-3 Roots:
repeated, 34-5 characteristic, 8
Reciprocation, 374 latent, 8
Reduction of order of DE, 160 nth, 20
Region, 18 of trinomial equation, 208
fundamental, 378 ROUCHt, Eugene (1832-1910), 186
zero-free, 292-300 theorem, 186
Removal of first derivative, 179 ROUSSEAU, Jean-Jaques, 1
Representation:
asymptotic, 52, 57, 220-6, 369 SANSONE,G.,432,467,469
by double integrals, 256-7 SARIO, Leo, 142
Euler transforms, 258-265 SCHAFFER, Juan Jorge, 469
hypergeometric, 265-270 SCHAFHEITLIN, P., 283
Laplace transforms, 270-4 _ SCHLESINGER,
469 Ludwig (1864-1933), 198,
Mellin and Mellin-Barnes transforms,
274-280 algebraic hypergeometric functions, 400
by integrals, 253-282 Euler transform, 258, 264, 281
Lappo-Danilevskij's, theorem, 386 Laplace transform, 281
by psi series, 249-253 ^regular-singular points, 338
theorems, 249-282 SCHLOMILCH, Otto, 212, 216
\ Resolution of the identity, 346 ^J^CHWARZ, Binyamin, 283, 300, 307, 320
RICCATI, Conto Jacopo (1676-1754), 42, SCHWARZ, 1921):Hermann Amandus (1843-
103
Acta Eruditorum, 103 on algebraic hypergeometric functions,
DE, 89-90, 103-143
geometric applications, 110-4 383-8 mapping, 377-383
on conformal
Malmquist's theorem, 129-138 differential invariant (Schwarzian), 374-
generalizations, 138-140, 143 401
^matrix, DE's, 407-8 symmetry principle, 240, 379, 394
RIEMANN, Bernhard (1826-66), 17, 24, triangular functions, 385, 395
200 Schwarzian, 374-401
 INDEX
482

applications to: topological, of curve system, 456

algebraic solutions, 383-8 SKENDZIC, M., 466
conformal mapping, 377-383 SMITH, Russell, 433, 467
uniformization, 394-400 autonomous polynomial systems, 460-5
univalence, 388-393 quadratic system, psi series, 459
derivative, 110, 374 Solution:
invariants of projective group, 374-5 existence and uniqueness of, 40-75
La vie, 375 history, 40-44
Kummer, 374, 394 dependence on parameters, 67-74
Riemann, 374, 394 expansion in powers of a parameter, 74
SCORZA-DRAGONI, Guiseppe, 449 general, 43
Separation of variables in Laplace-Poisson oscillatory, 178, 182
equations, 229, 230, 234 particular, 43
Sequence, Cauchy, 4 singular, 43
Series: matrix, 323-6
^asymptotic, 52, 57, 223, 362 algebraically regular, 326
Baker, 451, 456 transit matrix, 329
binomial, 363 r Space:
Coulson-March, 450, 466 abstract, 1
divergent, 52 I Banach, 5
factorial, 363-4 complete, 4
metric, 3
Fourier power, 120
hypergeometric, 201 normed, 4
lacunary, 23 vector, 1
Laurent, 26
normal, 362 commutator, 345
matrix, 8
subnormal, 362 rSpectrum of:
\ power, 214 multiplication operators, 345
\psi, 249-256
SERRET, Alfred (1819-85), 111 I Mittag-Leffler, 328
Set, 1
closed, 4 point, 314
| Star:
open, 4 '\ Statistics:
polar, 134-5, 140
Sheet, 20 STIELTJES, Thomas Jan (1856-94):
paro:ameters, 368-9 Riemann, integral, 25
ShifMIZ
rSHI rix ujir
t ofU,matTats
spherical characteristic, 118 f~ STOKES, George Gabriel (1819-1900), 248
X, phenomenon, 364-5
Sine equation, asymptotic, 183 String of zeros, 190
Singularities: d, l,r, 318
algebraic, 34 STRUBECKER, Karl, 142
on circle of convergence, 23 Structure:
essential, 27 algebraic and geometric, 1
fixed, of DE,43, 76-81 analytical, 17
irregular, of DE, 144, 147 STURM, Jaques-Charles-Francois (1805-55),
logarithmic, 34 283
matrix, 8 Sturmian chain, 213, 233
movable, of DE, 43, 76-81 , 86-7 Sturmian methods, 284-292
pole, 27 Surface:
regular, of DE, 144, 147, 155-161, 342- fundamental forms of, 113
363 Riemann, 20, 118
INDEX
&
ruled, 113 Trichotomy, 3
Symmetry principle, 240, 379, 394 •TRICOMI, Francesco:
System: Bateman papers, 248, 468
autonomous polynomial, 460-4 confluent hypergeometric functions, 247,
469
autonomous quadratic, 455-460
fundamental, 150 Trihedral, moving, 111
orthogonal, 210, 227 TRJITZINSKI, W. J., 364, 373, 432, 469
orthonormal, 210 TURRITTIN, Hugh L.:
Birkhoff problem, 274, 281, 361, 372
TAAM, Choy-Tak, 283, 300, 320 factorial matrix coefficients, 364
Tangent: Stokes phenomenon, 364-5
of curve, 111
function, 48 Type:
exponential growth, 411
locus, 300 of meromorphic function of finite order,
vector, 111 119
TANNERY, J., 465 power, 413
Testpower test, 90
Theorem: Uniformization:
Bolzano-Weierstrass, 4, 18 by automorphic functions, 128
double series, 35-6 by meromorphic functions, 127
ifixed point, 9-12 by modular functions, 400
lamilton-Cayley, 9 by rational functions, 1 29
leine-Borel, 211 Uniqueness of:
Identity, 28 fixed point, 9-12
>ainleve\ 83, 87 solution of DE, 40-47
'ringsheim, 23 /Uni vale nee:
Theta functions, 263-5 classes <& and U , 388
null, 265, 398-400 ^/in a half-plane, 390, 393
THOMAS, Llewellyn Hilleth, 449, 466 use of the Schwarzian, 388-394
Fermi DE, 448-453
THOM£, Ludwig Wilhelm (1841-1910): VALIRON, Georges, 432, 469
normal series, 362, 375 central index, maximal term, 168, 198
TITCHMARSH, Edward Charles (1899- Value:
1963), 469 absolute, 21
singular boundary value problem, 188- asymptotic, 121
191, 199, 319, 320 distribution, 114-129
Trace, 324 eigen, 8
Transformation: exceptional, 121
Cn into Cn, 6 VAN VLECK, E. B., 283, 320
inverse, 6 Variation:
Kummer, 202 bounded, 25
linear, 5 of parameters, 314-5
Liouville, 174 variational system, 69
Mobius (fractional linear), 19 I Vector, 2
order preserving, 15 I characteristic, 8
positive, 14 field, 297
Triangle: Ispace, 2
analytical, 94 \ tangent, 111
axiom, 3
functions, 385, 395 TEdRRpoiA,
TVOLfixe Vito
nt the , m,11 1 1
ore
 INDEX
484

integral equation, 1 1 approximate solutions of linear DE, 169,

struggle for life, 455, 466 198
central index, maximum term, 168
WALLIS, John, 233 estimates of w'/w, 175-7, 199
WALTER, Wolfgang, 39, 470 grade, 162
jWATSON, G. N.: Wjnding number (index), 329
M Bessel functions, 212, 248 WIRTINGER, Wilhelm (1865-1945):
( Whittaker, Modern Analysis, 231, 247, Heidelberg lecture, 394
\ 282,470 WITTICH, Hans, 103, 106, 140, 143
WEBER, Heinrich (1849-1913), 231 growth of solutions of 2nd order linear
.— - Hermite equation, 228-233 DE, 165-8, 198
I WEIERSTRASS, Karl (1815-97), 17, 21, Malmquist theorem, 129-138
24, 26, 33, 34, 35, 36,42 Painleve" transcendents, 443-5
analytic continuation, 33-5 ^7 1853):
\WRONSKI (Hoene, Josef Maria) (1778-
double series theorem, 35-36
pe-function, 96, 126-139 kvronskian, 150-1
power series, domain of convergence, 3 Wronskian matrix, 192
zeta-function, 246
WEILL, G., 391 ANG, Chung-Chun, 130, 140, 143
WEISS, Leonard, 143,470 YOSIDA, Kosaku, 103, 114, 130, 138, 140,
NWEYR, Eduard, 108 143,417, 470
_XoAs
^Ty 1S
(WHITTAKER, E. T.:
1 Watson, 231, 247, 281, Zero, element, 3
I 470 free region, 292-300
WIMAN, Anders (1865-1954): of function, 28
 A CATALOG 3 OF SELECTED
DOVE R BOOKS
IN SCIENCE A ND MATHEMATICS

m
 A CATALOG OF SELECTED

DOVER BOOKS
IN SCIENCE AND MATHEMATICS

QUALITATIVE THEORY OF DIFFERENTIAL EQUATIONS, V. V. Nemytskii

and V.V. Stepanov. Classic graduate-level text by two prominent Soviet mathe-
maticians covers classical differential equations as well as topological dynamics
and ergodic theory. Bibliographies. 523pp. 5% x 8& 65954-2 Pa. $14.95
MATRICES AND LINEAR ALGEBRA, Hans Schneider and George Phillip
Barker. Basic textbook covers theory of matrices and its applications to systems of
linear equations and related topics such as determinants, eigenvalues and differ-
ential equations. Numerous exercises. 432pp. 5% x 8^4. 66014-1 Pa. $10.95
QUANTUM THEORY, David Bohm. This advanced undergraduate-level text
presents the quantum theory in terms of qualitative and imaginative concepts,
followed by specific applications worked out in mathematical detail. Preface.
Index. 655pp. 5% x 8H. 65969-0 Pa. $14.95

ATOMIC PHYSICS (8th edition), Max Born. Nobel laureate's lucid treatment of
kinetic theory of gases, elementary particles, nuclear atom, wave-corpuscles, atomic
structure and spectral lines, much more. Over 40 appendices, bibliography. 495pp.
53/8x8'/2. 65984-4 Pa. $12.95
ELECTRONIC STRUCTURE AND THE PROPERTIES OF SOLIDS: The
Physics of the Chemical Bond, Walter A. Harrison. Innovative text offers basic
understanding of the electronic structure of covalent and ionic solids, simple
metals, transition metals and their compounds. Problems. 1980 edition. 582pp.
&A*9H. 66021-4 Pa. $16.95
BOUNDARY VALUE PROBLEMS OF HEAT CONDUCTION, M. Necati
Ozisik. Systematic, comprehensive treatment of modern mathematical methods of
solving problems in heat conduction and diffusion. Numerous examples and
problems. Selected references. Appendices. 505pp. 5% x m. 65990-9 Pa. $12.95
A SHORT HISTORY OF CHEMISTRY (3rd edition), J R. Partington. Classic
exposition explores origins of chemistry, alchemy, early medical chemistry, nature
of atmosphere, theory of valency, laws and structure of atomic theory, much more.
428pp. 5% x m. (Available in U.S. only) 65977-1 Pa. $11.95
A HISTORY OF ASTRONOMY, A. Pannekoek. Well-balanced, carefully rea-
soned study covers such topics as Ptolemaic theory, work of Copernicus, Kepler,
Newton, Eddington's work on stars, much more. Illustrated. References. 521pp.
53/8x8'/2. 65994-1 Pa. $12.95
PRINCIPLES OF METEOROLOGICAL ANALYSIS, Walter J. Saucier. Highly
respected, abundantly illustrated classic reviews atmospheric variables, hydro-
statics, static stability, various analyses (scalar, cross-section, isobaric, isentropic,
more). For intermediate meteorology students. 454pp. 6% x 9^. 65979-8 Pa. $14.95
 CATALOG OF DOVER BOOKS

RELATIVITY, THERMODYNAMICS AND COSMOLOGY, Richard C. Tol-

man. Landmark study extends thermodynamics to special, general relativity; also
applications of relativistic mechanics, thermodynamics to cosmological models.
501pp. 5% x 8'/2. 65383-8 Pa. $13.95

APPLIED ANALYSIS, Cornelius Lanczos. Classic work on analysis and design of

finite processes for approximating solution of analytical problems. Algebraic
equations, matrices, harmonic analysis, quadrature methods, much more. 559pp.
5Xx8£ 65656- X Pa. $13.95

INTRODUCTION TO ANALYSIS, Maxwell Rosenlicht. Unusually clear, acces-

sible coverage of set theory, real number system, metric spaces, continuous
functions, Riemann integration, multiple integrals, more. Wide range of problems.
Undergraduate level. Bibliography. 254pp. 5% x m. 65038-3 Pa. $8.95

INTRODUCTION TO QUANTUM MECHANICS With Applications to Chem-

istry, Linus Pauling & E. Bright Wilson, Jr. Classic undergraduate text by Nobel
Prize winner applies quantum mechanics to chemical and physical problems.
Numerous tables and figures enhance the text. Chapter bibliographies. Appen-
dices. Index. 468pp. 5% x m. 64871-0 Pa. $12.95
ASYMPTOTIC EXPANSIONS OF INTEGRALS, Norman Bleistein 8c Richard A.
Handelsman. Best introduction to important field with applications in a variety of
scientific disciplines. New preface. Problems. Diagrams. Tables. Bibliography.
Index. 448pp. 5% x 8V2. 65082-0 Pa". $12.95
MATHEMATICS APPLIED TO CONTINUUM MECHANICS, Lee A. Segel.
Analyzes models of fluid flow and solid deformation. For upper-level math, science
and engineering students. 608pp. b\ x 8'/2. 65369-2 Pa. $14.95

ELEMENTS OF REAL ANALYSIS, David A. Sprecher. Classic text covers

fundamental concepts, real number system, point sets, functions of a real variable,
Fourier series, much more. Over 500 exercises. 352pp. b% x m. 65385-4 Pa. $1 1.95

PHYSICAL PRINCIPLES OF THE QUANTUM THEORY, Werner Heisenberg.

Nobel Laureate discusses quantum theory, uncertainty, wave mechanics, work of
irac, Schroedinger, Compton, Wilson, Einstein, etc. 184pp. 5% x 8!4.
60113-7 Pa. $6.95

INTRODUCTORY REAL ANALYSIS, A.N. Kolmogorov, S.V. Fomin. Trans-

lated by Richard A. Silverman. Self-contained, evenly paced introduction to real
and functional analysis. Some 350 problems. 403pp. 5% x 8'/2. 61226-0 Pa. $10.95
PROBLEMS AND SOLUTIONS IN QUANTUM CHEMISTRY AND PHYSICS,
Charles S. Johnson, Jr. and Lee G. Pedersen. Unusually varied problems, detailed
solutions in coverage of quantum mechanics, wave mechanics, angular momen-
tum, molecular spectroscopy, scattering theory, more. 280 problems plus 139
supplementary exercises. 430pp. 6'/2 x 9'4. 65236-X Pa. $13.95
 CATALOG OF DOVER BOOKS

ASYMPTOTIC METHODS IN ANALYSIS, N.G. de Bruijn. An inexpensive,

comprehensive guide to asymptotic methods— the pioneering work that teaches by
explaining worked examples in detail. Index. 224pp. 5% x 8V2. 64221-6 Pa. $7.95
OPTICAL RESONANCE AND TWO-LEVEL ATOMS, L. Allen and J.H. Eberly.
Clear, comprehensive introduction to basic principles behind all quantum optical
resonance phenomena. 53 illustrations. Preface. Index. 256pp. 5% x 8V2.
65533-4 Pa. $8.95
COMPLEX VARIABLES, Francis J. Flanigan. Unusual approach, delaying
complex algebra till harmonic functions have been analyzed from real variable
viewpoint. Includes problems with answers. 364pp. 5% x 8J4. . 61388-7 Pa. $9.95
ATOMIC SPECTRA AND ATOMIC STRUCTURE, Gerhard Herzberg. One of
best introductions; especially for specialist in other fields. Treatment is physical
rather than mathematical. 80 illustrations. 257pp. 5% x m. 601 15-3 Pa. $6.95
APPLIED COMPLEX VARIABLES, John W. Dettman. Step-by-step coverage of
fundamentals of analytic function theory — plus lucid exposition of five important
applications: Potential Theory; Ordinary Differential Equations; Fourier Trans-
forms; Laplace Transforms; Asymptotic Expansions. 66 figures. Exercises at
chapter ends. 512pp. b% * 8J4. 64670-X Pa. $12.95
ULTRASONIC ABSORPTION: An Introduction to the Theory of Sound Absorp-
tion and Dispersion in Gases, Liquids and Solids, A.B. Bhatia. Standard reference
in the field provides a clear, systematically organized introductory review of
fundamental concepts for advanced graduate students, research workers. Numerous
diagrams. Bibliography. 440pp. 5% x m. 64917-2 Pa. $1 1.95
UNBOUNDED LINEAR OPERATORS. Theory and Applications, Seymour
Goldberg. Classic presents systematic treatment of the theory of unbounded linear
operators in normed linear spaces with applications to differential equations.
Bibliography. 199pp. 5% x 8!4. 64830-3 Pa. $7.95
LIGHT SCATTERING BY SMALL PARTICLES, H.C. van de Hulst. Compre-
hensive treatment including full range of useful approximation methods for
researchers in chemistry, meteorology and astronomy. 44 illustrations. 470pp.
53/8x8>/2. 64228-3 Pa. $11.95
CONFORMAL MAPPING ON RIEMANN SURFACES, Harvey Cohn. Lucid,
insightful book presents ideal coverage of subject. 334 exercises make book perfect
for self-study. 55 figures. 352pp. 5% x 8*4. 64025-6 Pa. $1 1.95

OPTICKS, Sir Isaac Newton. Newton's own experiments with spectroscopy,

colors, lenses, reflection, refraction, etc., in language the layman can follow.
Foreword by Albert Einstein. 532pp. 5% x m. 60205-2 Pa. $1 1.95
GENERALIZED INTEGRAL TRANSFORMATIONS, A H. Zemanian. Gradu-
ate-level study of recent generalizations of the Laplace, Mellin, Hankel, K.
Weierstrass, convolution and other simple transformations. Bibliography. 320pp.
5% x 8!4. 65375-7 Pa. $8.95
 CATALOG OF DOVER BOOKS

THE ELECTROMAGNETIC FIELD, Albert Shadowitz. Comprehensive under-

graduate text covers basics of electric and magnetic fields, builds up to electromag-
netic theory. Also related topics, including relativity. Over 900 problems. 768pp.
5%x8Va. 65660-8 Pa. $18.95
FOURIER SERIES, Georgi P. Tolstov. Translated by Richard A. Silverman. A
valuable addition to the literature on the subject, moving clearly from subject to
subject and theorem to theorem. 107 problems, answers. 336pp. 53/s x 8&
63317-9 Pa. $9.95
THEORY OF ELECTROMAGNETIC WAVE PROPAGATION, Charles Her-
ach Papas. Graduate-level study discusses the Maxwell field equations, radiation
from wire antennas, the Doppler effect and more, xiii + 244pp. 5% x 8%.
65678-0 Pa. $6.95
DISTRIBUTION THEORY AND TRANSFORM ANALYSIS: An Introduction
to Generalized Functions, with Applications, A.H. Zemaniah. Provides basics of
distribution theory, describes generalized Fourier and Laplace transformations.
Numerous problems. 384pp. 5% x 8^2. 65479-6 Pa. $1 1.95
THE PHYSICS OF WAVES, William C. Elmore and Mark A. Heald. Unique
overview of classical wave theory. Acoustics, optics, electromagnetic radiation,
more. Ideal as classroom text or for self-study. Problems. 477pp. 5% x 8'/2.
64926-1 Pa. $12.95
CALCULUS OF VARIATIONS WITH APPLICATIONS, George M. Ewing.
Applications-oriented introduction to variational theory develops insight and
promotes understanding of specialized books, research papers. Suitable for
advanced undergraduate/graduate students as primary, supplementary text. 352pp.
5% x 8'/2. 64856-7 Pa. $9.95
A TREATISE ON ELECTRICITY AND MAGNETISM, James Clerk Maxwell.
Important foundation work of modern physics. Brings to final form Maxwell's
theory of electromagnetism and rigorously derives his general equations of field
theory. 1,084pp. b% x m. 60636-8, 60637-6 Pa., Two-vol. set $23.90
AN INTRODUCTION TO THE CALCULUS OF VARIATIONS, Charles Fox.
Graduate-level text covers variations of an integral, isoperimetrical problems, least
action, special relativity, approximations, more. References. 279pp. 5% x 854.
65499-0 Pa. $8.95
H YDRODYNAMIC AND H YDROMAGNETIC STABILITY, S. Chandrasekhar.
Lucid examination of the Rayleigh-Benard problem; clear coverage of the theory of
instabilities causing convection. 704pp. 5% x 8!4. 6407 1-X Pa. $14.95
CALCULUS OF VARIATIONS, Robert Weinstock. Basic introduction covering
isoperimetric problems, theory of elasticity, quantum mechanics, electrostatics, etc.
Exercises throughout. 326pp. 5% x 8H>. 63069-2 Pa. $8.95
DYNAMICS OF FLUIDS IN POROUS MEDIA, Jacob Bear. For advanced
students of ground water hydrology, soil mechanics and physics, drainage and
irrigation engineering and more. 335 illustrations. Exercises, with answers. 784pp.
6K*9W. 65675-6 Pa. $19.95
 CATALOG OF DOVER BOOKS

NUMERICAL METHODS FOR SCIENTISTS AND ENGINEERS, Richard

Hamming. Classic text stresses frequency approach in coverage of algorithms,
polynomial approximation, Fourier approximation, exponential approxima-
tion, other topics. Revised and enlarged 2nd edition. 721pp. 5% * 8i4.
65241-6 Pa. $15.95

THEORETICAL SOLID STATE PHYSICS, Vol. I: Perfect Lattices in Equilib-

rium; Vol. II: Non-Equilibrium and Disorder, William Jones and Norman H.
March. Monumental reference work covers fundamental theory of equilibrium
properties of perfect crystalline solids, non-equilibrium properties, defects and
disordered systems. Appendices. Problems. Preface. Diagrams. Index. Bibliog-
raphy. Total of 1,301pp. 5% * m. Two volumes. Vol. I 65015-4 Pa. $16.95
Vol. II 65016-2 Pa. $14.95

OPTIMIZATION THEORY WITH APPLICATIONS, Donald A. Pierre. Broad-

spectrum approach to important topic. Classical theory of minima and maxima,
calculus of variations, simplex technique and linear programming, more. Many
problems, examples. 640pp. 5% x 8V2. 65205-X Pa. $14.95

THE CONTINUUM: A Critical Examination of the Foundation of Analysis,

Hermann Weyl. Classic of 20th-century foundational research deals with the
conceptual problem posed by the continuum. 156pp. 5h x 8& 67982-9 Pa. $6.95
ESSAYS ON THE THEORY OF NUMBERS, Richard Dedekind. Two classic
essays by great German mathematician: on the theory of irrational numbers; and on
transfinite numbers- and properties of natural numbers. 1 15pp. 5% x 8&
21010-3 Pa. $5.95

THE FUNCTIONS OF MATHEMATICAL PHYSICS, Harry Hochstadt. Com-

prehensive treatment of orthogonal polynomials, hypergeometric functions, Hill's
equation, much more. Bibliography. Index. 322pp. 5% x 8V2. 65214-9 Pa. $9.95
NUMBER THEORY AND ITS HISTORY, Oystein Ore. Unusually clear,
accessible introduction covers counting, properties of numbers, prime numbers,
much more. Bibliography. 380pp. 5% x 8V2. 65620-9 Pa. $9.95
THE VARIATIONAL PRINCIPLES OF MECHANICS, Cornelius Lanczos.
Graduate level coverage of calculus of variations, equations of motion, relativistic
mechanics, more. First inexpensive paperbound edition of classic treatise. Index.
Bibliography. 418pp. 5% x 8'/2. 65067-7 Pa. $12.95
MATHEMATICAL TABLES AND FORMULAS, Robert D. Carmichael and
Edwin R. Smith. Logarithms, sines, tangents, trig functions, powers, roots,
reciprocals, exponential and hyperbolic functions, formulas and theorems. 269pp.
5^8'^. 60111-0 Pa. $6.95
THEORETICAL PHYSICS, Georg Joos, with Ira M. Freeman. Classic overview
covers essential math, mechanics, electromagnetic theory, thermodynamics, quan-
tum mechanics, nuclear physics, other topics. First paperback edition, xxiii +
885pp. 53/- x m. 65227-0 Pa. $21 .95
 CATALOG OF DOVER BOOKS

HANDBOOK OF MATHEMATICAL FUNCTIONS WITH FORMULAS,

GRAPHS, AND MATHEMATICAL TABLES, edited by Milton Abramowitz and
Irene A. Stegun. Vast compendium: 29 sets of tables, some to as high as 20 places.
1,046pp. 8 x 10& 61272-4 Pa. $24.95
MATHEMATICAL METHODS IN PHYSICS AND ENGINEERING, John W.
Dettman. Algebraically based approach to vectors, mapping, diffraction, other
topics in applied math. Also generalized functions, analytic function theory, more.
Exercises. 448pp. 5% x m. 65649-7 Pa. $10.95

A SURVEY OF NUMERICAL MATHEMATICS, David M. Young and Robert

Todd Gregory. Broad self-contained coverage of computer-oriented numerical
algorithms for solving various types of mathematical problems in linear algebra,
ordinary and partial, differential equations, much more. Exercises. Total of
1,248pp. 5% x 854. Two volumes. Vol. I 65691-8 Pa. $14.95
Vol. II 65692-6 Pa. $14.95

TENSOR ANALYSIS FOR PHYSICISTS, J.A. Schouten. Concise exposition of

the mathematical basis of tensor analysis, integrated with well-chosen physical
examples of the theory. Exercises. Index. Bibliography. 289pp. 5% x 8'/2.
65582-2 Pa. $8.95

INTRODUCTION TO NUMERICAL ANALYSIS (2nd Edition), F.B. Hilde-

brand. Classic, fundamental treatment covers computation, approximation, inter-
polation, numerical differentiation and integration, other topics. 150 new prob-
lems. 669pp. b% x 854. 65363-3 Pa. $15.95

INVESTIGATIONS ON THE THEORY OF THE BROWNIAN MOVEMENT,

Albert Einstein. Five papers (1905-8) investigating dynamics of Brownian motion
and evolving elementary theory. Notes by R. Furth. 122pp. 5% x 854.
60304-0 Pa. $4.95

CATASTROPHE THEORY FOR SCIENTISTS AND ENGINEERS, Robert

Gilmore. Advanced-level treatment describes mathematics of theory grounded in
the work of Poincare, R. Thorn, other mathematicians. Also important applications
to problems in mathematics, physics, chemistry and engineering. 1981 edition,
ferences. 28 tables. 397 black-and-white illustrations, xvii + 666pp. 6Vs x 9'/4.
67539-4 Pa. $17.95

AN INTRODUCTION TO STATISTICAL THERMODYNAMICS, Terrell L.

Hill. Excellent basic text offers wide-ranging coverage of quantum statistical
mechanics, systems of interacting molecules, quantum statistics, more. 523pp.
5%x 854. 65242-4 Pa. $12.95

STATISTICAL PHYSICS, Gregory H. Wannier. Classic text combines thermo-

dynamics, statistical mechanics and kinetic theory in one unified presentation of
thermal physics. Problems with solutions. Bibliography. 532pp. 5% x 854.
6540 1-X Pa. $12.95
 CATALOG OF DOVER BOOKS

ORDINARY DIFFERENTIAL EQUATIONS, Morris Tenenbaum and Harry

Pollard. Exhaustive survey of ordinary differential equations for undergraduates in
mathematics, engineering, science. Thorough analysis of theorems. Diagrams.
Bibliography. Index. 818pp. 5% x m. 64940-7 Pa. $18.95
STATISTICAL MECHANICS. Principles and Applications, Terrell L. Hill.
Standard text covers fundamentals of statistical mechanics, applications to
fluctuation theory, imperfect gases, distribution functions, more. 448pp. 5% x SlA.
65390-0 Pa. $11.95

ORDINARY DIFFERENTIAL EQUATIONS AND STABILITY THEORY: An

Introduction, David A. Sanchez. Brief, modern treatment. Linear equation,
stability theory for autonomous and nonautonomous systems, etc. 164pp. 5% x 8'4.
63828-6 Pa. $6.95

THIRTY YEARS THAT SHOOK PHYSICS: The Story of Quantum Theory,

George Gamow. Lucid, accessible introduction to influential theory of energy and
matter. Careful explanations of Dirac's anti-particles, Bohr's model of the atom,
much more. 12 plates. Numerous drawings. 240pp. 5% x 8'/2. 24895-X Pa. $6.95

THEORY OF MATRICES, Sam Perlis. Outstanding text covering rank, non-

singularity and inverses in connection with the development of canonical matrices
under the relation of equivalence, and without the intervention of determinants.
Includes exercises. 237pp. 5% x W. 66810-X Pa. $8.95

GREAT EXPERIMENTS IN PHYSICS: Firsthand Accounts from Galileo to

Einstein, edited by Morris H. Shamos. 25 crucial discoveries: Newton's laws of
motion, Chadwick's study of the neutron, Hertz on electromagnetic waves, more.
Original accounts clearly annotated. 370pp. 5% x 8& 25346-5 Pa. $10.95
INTRODUCTION TO PARTIAL DIFFERENTIAL EQUATIONS WITH AP-
PLICATIONS, E.C. Zachmanoglou and Dale W. Thoe. Essentials of partial
differential equations applied to common problems in engineering and the
physical sciences. Problems and answers. 416pp. 5% x 8& 65251-3 Pa. $1 1.95

BURNHAM'S CELESTIAL HANDBOOK, Robert Burnham, Jr. Thorough guide

to the stars beyond our solar system. Exhaustive treatment. Alphabetical by
constellation: Andromeda to Cetus in Vol. 1 ; Chamaeleon to Orion in Vol. 2; and
Pavo to Vulpecula in Vol. 3. Hundreds of illustrations. Index in Vol. 3. 2,000pp.
6% x 9'/4. 23567-X, 23568-8, 23673-0 Pa., Three-vol. set $44.85
CHEMICAL MAGIC, Leonard A. Ford. Second Edition, Revised by E. Winston
Grundmeier. Over 100 unusual stunts demonstrating cold fire, dust explosions,
much more. Text explains scientific principles and stresses safety precautions.
128pp. 5% x 8'/2. 67628-5 Pa. $5.95
AMATEUR ASTRONOMER'S HANDBOOK, J.B. Sidgwick. Timeless, compre-
hensive coverage of telescopes, mirrors, lenses, mountings, telescope drives,
micrometers, spectroscopes, more. 189 illustrations. 576pp. 5% x 8'/4. (Available in
U.S. only) 24034-7 Pa. $11.95
 CATALOG OF DOVER BOOKS

SPECIAL "FUNCTIONS,
mous Russian work treatingN.N.
moreLebedev.
importantTranslated by RichardwithSilverman.
special functions, Fa-
applications
to specific problems of physics and engineering. 38 figures. 308pp. 5% x 854.
60624-4 Pa. $9.95
OBSERVATIONAL ASTRONOMY FOR AMATEURS, J.B. Sidgwick. Mine of
useful data for observation of sun, moon, planets, asteroids, aurorae, meteors,
comets, variables, binaries, etc. 39 illustrations. 384pp. 5% x 8^. (Available in U.S.
only) 24033-9 Pa. $8.95
INTEGRAL EQUATIONS, F.G. Tricomi. Authoritative, well-written treatment
of extremely useful mathematical tool with wide applications. Volterra Equations,
Fredholm Equations, much more. Advanced undergraduate to graduate level.
Exercises. Bibliography. 238pp. 5% x m. 64828-1 Pa. $8.95

POPULAR LECTURES ON MATHEMATICAL LOGIC, Hao Wang. Noted

logician's lucid treatment of historical developments, set theory, model theory,
recursion theory and constructivism, proof theory, more. 3 appendixes. Bibli-
ography. 1981 edition, ix + 283pp. 5% x 8!4. 67632-3 Pa. $8.95
MODERN NONLINEAR EQUATIONS, Thomas L. Saaty. Emphasizes practical
solution of problems; covers seven types of equations. ". . . a welcome contribution
to the existing literature. . . /'—Math Reviews. 490pp. 5% x m. 64232-1 Pa. $11.95
FUNDAMENTALS OF ASTRODYNAMICS, Roger Bate et al. Modern approach
developed by U.S. Air Force Academy. Designed as a first course. Problems,
exercises. Numerous illustrations. 455pp. b% x 8'/2. 60061-0 Pa. $9.95
INTRODUCTION TO LINEAR ALGEBRA AND DIFFERENTIAL EQUA-
TIONS, John W. Dettman. Excellent text covers complex numbers, determinants,
orthonormal bases, Laplace transforms, much more. Exercises with solutions.
Undergraduate level. 416pp. 5X x 8K2. 65191-6 Pa. $10.95
INCOMPRESSIBLE AERODYNAMICS, edited by Bryan Thwaites. Covers theo-
retical and experimental treatment of the uniform flow of air and viscous fluids past
two-dimensional aerofoils and three-dimensional wings; many other topics. 654pp.
5Xx8& 65465-6 Pa. $16.95
INTRODUCTION TO DIFFERENCE EQUATIONS, Samuel Goldberg. Excep-
tionally clear exposition of important discipline with applications to sociology,
psychology, economics. Many illustrative examples; over 250 problems. 260pp.
5X x m. 65084-7 Pa. $8.95
LAMINAR BOUNDARY LAYERS, edited by L. Rosenhead. Engineering classic
covers steady boundary layers in two- and three-dimensional flow, unsteady
boundary layers, stability, observational techniques, much more. 708pp. 5% x 8'/£
65646-2 Pa. $18.95
LECTURES ON CLASSICAL DIFFERENTIAL GEOMETRY, Second Edition,
Dirk J. Struik. Excellent brief introduction covers curves, theory of surfaces,
fundamental equations, geometry on a surface, conformal mapping, other topics.
Problems: 240pp. 5% x m. 65609-8 Pa. $8.95
 CATALOG OF DOVER BOOKS

ROTARY-WING AERODYNAMICS, W.Z. Stepniewski. Clear, concise text covers

aerodynamic phenomena of the rotor and offers guidelines for helicopter per
formance evaluation. Originally prepared for NASA. 537 figures. 640pp. 65* x 954.
64647-5 Pa. 515.95

DIFFERENTIAL GEOMETRY, Heinrich W. Guggenheimer. Local differential

geometry as an application of advanced calculus and linear algebra. Curvature,
transformation groups, surfaces, more. Exercises. 62 figures. 378pp. b% x 854.
63433-7 Pa. $9.95
INTRODUCTION TO SPACE DYNAMICS, William Tyrrell Thomson. Com-
prehensive, classic introduction to space-flight engineering for advanced under-
graduate and graduate students. Includes vector algebra, kinematics, transforma-
tion of coordinates. Bibliography. Index. 352pp. 5% x 8'/2. 651 13-4 Pa. $9.95
A SURVEY OF MINIMAL SURFACES, Robert Osserman. Up-to-date, in-depth
discussion of the field for advanced students. Corrected and enlarged edition covers
new developments. Includes numerous problems. 192pp. 5% x 854.
64998-9 Pa. $8.95
ANALYTICAL MECHANICS OF GEARS, Earle Buckingham. Indispensable
reference for modern gear manufacture covers conjugate gear-tooth action, gear-
tooth profiles of various gears, many other topics. 263 figures. 102 tables. 546pp.
5%x8!4. 65712-4 Pa. $14.95
SET THEORY AND LOGIC, Robert R. Stoll. Lucid introduction to unified
theory of mathematical concepts. Set theory and logic seen as tools for conceptual
understanding of real number system. 496pp. 5% x 8*4. 63829-4 Pa. $12.95
A HISTORY OF MECHANICS, Rene Dugas. Monumental study of mechanical
principles from antiquity to quantum mechanics. Contributions of ancient Greeks,
Galileo, Leonardo, Kepler, Lagrange, many others. 671pp. 5% x 854.
65632-2 Pa. $14.95
FAMOUS PROBLEMS OF GEOMETRY AND HOW TO SOLVE THEM,
Benjamin Bold. Squaring the circle, trisecting the angle, duplicating the cube:
learn their history, why they are impossible to solve, then solve them yourself.
128pp. 5% x 854. 24297-8 Pa. $4.95
MECHANICAL VIBRATIONS, J. P. Den Hartog. Classic textbook offers lucid
explanations and illustrative models, applying theories of vibrations to a variety of
practical industrial engineering problems. Numerous figures. 233 problems,
solutions. Appendix. Index. Preface. 436pp. 5% x 854. 64785-4 Pa. $11.95
CURVATURE AND HOMOLOGY, Samuel I. Goldberg. Thorough treatment of
specialized branch of differential geometry. Covers Riemannian manifolds, topol-
ogy of dif ferentiable manifolds, compact Lie groups, other topics. Exercises. 3 1 5pp.
5%x 854. 64314-X Pa. $9.95

HISTORY OF STRENGTH OF MATERIALS, Stephen P. Timoshenko. Excel-

lent historical survey of the strength of materials with many references to the
theories of elasticity and structure. 245 figures. 452pp. 5% x 854. 61 187-6 Pa. $12.95
 CATALOG OF DOVER BOOKS

GEOMETRY OF COMPLEX NUMBERS, Hans Schwerdtfeger. Illuminating,

widely praised book on analytic geometry of circles, the Moebius transformation,
and two-dimensional non-Euclidean geometries. 200pp. b% x SV*.
63830-8 Pa. $8.95
MECHANICS, J. P. Den Hartog. A classic introductory text or refresher. Hundreds
of applications and design problems illuminate fundamentals of trusses, loaded
beams and cables, etc. 334 answered problems. 462pp. 5% x m. 60754-2 Pa. $10.95
TOPOLOGY, John G. Hocking and Gail S. Young. Superb one-year course in
classical topology. Topological spaces and functions, point-set topology, much
more. Examples and problems. Bibliography. Index. 384pp. 5% x 814.
65676-4 Pa. $10.95
STRENGTH OF MATERIALS, J. P. Den Hartog. Full, clear treatment of basic
material (tension, torsion, bending, etc.) plus advanced material on engineering
methods, applications. 350 answered problems. 323pp. 5% x 854. 60755-0 Pa. $9.95
ELEMENTARY CONCEPTS OF TOPOLOGY, Paul Alexandroff. Elegant,
intuitive approach to topology from set-theoretic topology to Betti groups; how
concepts of topology are useful in math and physics. 25 figures. 57pp. 5% x 8&
60747-X Pa. $3.95
ADVANCED STRENGTH OF MATERIALS, J. P. Den Hartog. Superbly written
advanced text covers torsion, rotating disks, membrane stresses in shells, much
more. Many problems and answers. 388pp. b% x m. 65407-9 Pa. $10.95
COMPUTABILITY AND UNSOLVABILITY, Martin Davis. Classic graduate-
level introduction to theory of computability, usually referred to as theory of
recurrent functions. New preface and appendix. 288pp. 5% x 8!4. 61471-9 Pa. $8.95
GENERAL CHEMISTRY, Linus Pauling. Revised 3rd edition of classic first-year
text by Nobel laureate. Atomic and molecular structure, quantum mechanics,
statistical mechanics, thermodynamics correlated with descriptive chemistry.
Problems. 992pp. 5% x m. 65622-5 Pa. $19.95
AN INTRODUCTION TO MATRICES, SETS AND GROUPS FOR SCIENCE
STUDENTS, G. Stephenson. Concise, readable text introduces sets, groups, and
most importantly, matrices to undergraduate students of physics, chemistry, and
engineering. Problems. 164pp. b% x m. 65077-4 Pa. $7.95
THE HISTORICAL BACKGROUND OF CHEMISTRY, Henry M. Leicester.
Evolution of ideas, not individual biography. Concentrates on formulation of a
coherent set of chemical laws. 260pp. 5% x 854. 61053-5 Pa. $7.95
THE PHILOSOPHY OF MATHEMATICS: An Introductory Essay, Stephan
Korner. Surveys the views of Plato, Aristotle, Leibniz 8c Kant concerning proposi-
tions and theories of applied and pure mathematics. Introduction. Two appen-
dices. Index. 198pp. 5% x 8!4. 25048-2 Pa. $8.95
THE DEVELOPMENT OF MODERN CHEMISTRY, Aaron J. Ihde. Authorita-
tive history of chemistry from ancient Greek theory to 20th-century innovation.
Covers major chemists and their discoveries. 209 illustrations. 14 tables. Bibliog-
raphies. Indices. Appendices. 851pp. b% x 8& 64235-6 Pa. $18.95
 CATALOG OF DOVER BOOKS

DE RE METALLICA, Georgius Agricola. The famous Hoover translation of

greatest treatise on technological chemistry, engineering, geology, mining of early
modern times (1556). All 289 original woodcuts. 638pp. 634 x 1 1
60006-8 Pa. $18.95

SOME THEORY OF SAMPLING, William Edwards Deming. Analysis of the

problems, theory and design of sampling techniques for social scientists, industrial
managers and others who find statistics increasingly important in their work. 61
tables^ 90 figures, xvii + 602pp. 5% x m. 64684-X Pa. $15.95

THE VARIOUS AND INGENIOUS MACHINES OF AGOSTINO RAMELLI: A

Classic Sixteenth-Century Illustrated Treatise on Technology, Agostino Ramelli.
One of the most widely known and copied works on machinery in the 16th century.
194 detailed plates of water pumps, grain mills, cranes, more. 608pp. 9 x 12.
28180-9 Pa. $24.95

LINEAR PROGRAMMING AND ECONOMIC ANALYSIS, Robert Dorfman,

Paul A. Samuelson and Robert M. Solow. First comprehensive treatment of linear
programming in standard economic analysis. Game theory, modern welfare
economics, Leontief input-output, more. 525pp. 5% x 8& 65491-5 Pa. $14.95

ELEMENTARY DECISION THEORY, Herman Chernoff and Lincoln E. Moses.

Clear introduction to statistics and statistical theory covers data processing,
probability and random variables, testing hypotheses, much more. Exercises.
364pp. 5% x 8V2. 65218-1 Pa. $10.95

THE COMPLEAT STRATEGYST: Being a Primer on the Theory of Games of

Strategy, J.D. Williams. Highly entertaining classic describes, with many illus-
trated examples, how to select best strategies in conflict situations. Prefaces.
Appendices. 268pp. 5% x m. 25101-2 Pa. $7.95

CONSTRUCTIONS AND COMBINATORIAL PROBLEMS IN DESIGN OF

EXPERIMENTS, Damaraju Raghavarao. In-depth reference work examines
orthogonal Latin squares, incomplete block designs, tactical configuration, partial
geometry, much more. Abundant explanations, examples. 416pp. 53/« x g%.
65685-3 Pa. $10.95

THE ABSOLUTE DIFFERENTIAL CALCULUS (CALCULUS OF TENSORS),

Tullio Levi-Civita. Great 20th-century mathematician's classic work on material
necessary for mathematical grasp of theory of relativity. 452pp. 53/« x 8&
63401-9 Pa. $11.95

VECTOR AND TENSOR ANALYSIS WITH APPLICATIONS, A.I. Borisenko

and I.E. Tarapov. Concise introduction. Worked-out problems, solutions, exer-
cises. 257pp. 5% x 8V4. 63833-2 Pa. $8.95
 CATALOG OF DOVER BOOKS

THE FOUR-COLOR PROBLEM: Assaults and Conquest, Thomas L. Saaty and

Paul G. Kainen. Engrossing, comprehensive account of the century-old combina-
torial topological problem, its history and solution. Bibliographies. Index. 110
figures. 228pp. 5% x 854. 65092-8 Pa. $6.95
CATALYSIS IN CHEMISTRY AND ENZYMOLOGY, William P. Jencks.
Exceptionally clear coverage of mechanisms for catalysis, forces in aqueous
solution, carbonyl- and acyl-group reactions, practical kinetics, more. 864pp.
5ft *8£ 65460-5 Pa. $19.95
PROBABILITY: An Introduction, Samuel Goldberg. Excellent basic text covers set
theory, probability theory for finite sample spaces, binomial theorem, much more.
360 problems. Bibliographies. 322pp. b% x m. 65252-1 Pa. $9.95
LIGHTNING, Martin A. Uman. Revised, updated edition of classic work on the
physics of lightning. Phenomena, terminology, measurement, photography,
spectroscopy, thunder, more. Reviews recent research. Bibliography. Indices.
320pp. b% x 8>4. 64575-4 Pa. $8.95
PROBABILITY THEORY: A Concise Course, Y.A. Rozanov. Highly readable,
self-contained introduction covers combination of events, dependent events,
Bernoulli trials, etc. Translation by Richard Silverman. 148pp. 536 x 8'4.
63544-9 Pa. $6.95
AN INTRODUCTION TO HAMILTONIAN OPTICS, H. A. Buchdahl. Detailed
account of the Hamiltonian treatment of aberration theory in geometrical optics.
Many classes of optical systems defined in terms of the symmetries they possess.
Problems with detailed solutions. 1970 edition, xv + 360pp. 5% x 8&
67597-1 Pa. $10.95
STATISTICS MANUAL, Edwin L. Crow, et al. Comprehensive, practical
collection of classical and modern methods prepared by U.S. Naval Ordnance Test
Station. Stress on use. Basics of statistics assumed. 288pp. 5% x 854.
60599-X Pa. $7.95
DICTIONARY/OUTLINE OF BASIC STATISTICS, John E. Freund and Frank
J. Williams. A clear concise dictionary of over 1 ,000 statistical terms and an outline
of statistical formulas covering probability, nonparametric tests, much more.
208pp. 5% x 854. 66796-0 Pa. $7.95
STATISTICAL METHOD FROM THE VIEWPOINT OF QUALITY CON-
TROL, Walter A. Shewhart. Important text explains regulation of variables, uses
of statistical control to achieve quality control in industry, agriculture, other areas.
192pp. 5% x 854. 65232-7 Pa. $7.95
THE INTERPRETATION OF GEOLOGICAL PHASE DIAGRAMS, Ernest G.
Ehlers. Clear, concise text emphasizes diagrams of systems under fluid or
containing pressure; also coverage of complex binary systems, hydrothermal
melting, more. 288pp. 6!4 x 9W. 65389-7 Pa. $10.95
STATISTICAL ADJUSTMENT OF DATA, W. Edwards Deming. Introduction to
basic concepts of statistics, curve fitting, least squares solution, conditions without
parameter, conditions containing parameters. 26 exercises worked out. 271pp.
5% x 8'/2. 64685-8 Pa. $9.95
 CATALOG OF DOVER BOOKS

TENSOR CALCULUS, J.L. Synge and A. Schild. Widely used introductory text
covers spaces and tensors, basic operations in Riemannian space, non-Riemannian
spaces, etc. 324pp. 5% x SV*. 63612-7 Pa. $9.95
A CONCISE HISTORY OF MATHEMATICS, Dirk J. Struik. The best brief
history of mathematics. Stresses origins and covers every major figure from ancient
Near East to 19th century. 41 illustrations. 195pp. 5% x m. 60255-9 Pa. $7.95
A SHORT ACCOUNT OF THE HISTORY OF MATHEMATICS, W W. Rouse
Ball. One of clearest, most authoritative surveys from the Egyptians and Phoeni-
cians through 19th-century figures such as Grassman, Galois, Riemann. Fourth
edition. 522pp. 5% x m. 20630-0 Pa. $1 1.95
HISTORY OF MATHEMATICS, David E. Smith. Nontechnical survey from
ancient Greece and Orient to late 19th century; evolution of arithmetic, geometry,
trigonometry, calculating devices, algebra, the calculus. 362 illustrations. 1,355pp.
5% x m. 20429-4, 20430-8 Pa., Two-vol. set $26.90
THE GEOMETRY OF RENE DESCARTES, Rene Descartes. The great work
founded analytical geometry. Original French text, Descartes' own diagrams,
together with definitive Smith-Latham translation. 244pp. 5% x SVz.
60068-8 Pa. $7.95
THE ORIGINS OF THE INFINITESIMAL CALCULUS, Margaret E. Baron.
Only fully detailed and documented account of crucial discipline: origins;
development by Galileo, Kepler, Cavalieri; contributions of Newton, Leibniz,
more. 304pp. 5% x 8!4. (Available in U.S. and Canada only) 65371-4 Pa. $9.95
THE HISTORY OF THE CALCULUS AND ITS CONCEPTUAL DEVELOP-
MENT, Carl B. Boyer. Origins in antiquity, medieval contributions, work of
Newton, Leibniz, rigorous formulation. Treatment is verbal. 346pp. 5% x 8'/£.
60509-4 Pa. $9.95

THE THIRTEEN BOOKS OF EUCLID'S ELEMENTS, translated with introduc-

tion and commentary by Sir Thomas L. Heath. Definitive edition. Textual and
linguistic notes, mathematical analysis. 2,500 years of critical commentary. Not
abridged. 1,414pp. 5% x m. . 60088-2, 60089-0, 60090-4 Pa., Three-vol. set $31.85
GAMES AND DECISIONS: Introduction and Critical Survey, R. Duncan Luce
and Howard Raiffa. Superb nontechnical introduction to game theory, primarily
applied to social sciences. Utility theory, zero-sum games, n-person games,
decision-making, much more. Bibliography. 509pp. 5% x m. 65943-7 Pa. $12.95
THE HISTORICAL ROOTS OF ELEMENTARY MATHEMATICS, Lucas
N.H. Bunt, Phillip S. Jones, and Jack D. Bedient. Fundamental underpinnings of
modern arithmetic, algebra, geometry and number systems derived from ancient
civilizations. 320pp. 5% x m. 25563-8 Pa. $8.95
CALCULUS REFRESHER FOR TECHNICAL PEOPLE, A. Albert Klaf. Covers
important aspects of integral and differential calculus via 756 questions. 566
problems, most answered. 431pp. 5% x m. 20370-0 Pa. $8.95
 CATALOG OF DOVER BOOKS

CHALLENGING MATHEMATICAL PROBLEMS WITH ELEMENTARY

SOLUTIONS, A.M. Yaglom and I.M. Yaglom. Over 170 challenging problems on
probability theory, combinatorial analysis, points and lines, topology, convex
polygons, many other topics. Solutions. Total of 445pp. 5% x 8H. Two-vol. set.
Vol. I 65536-9 Pa. $7.95
Vol. II 65537-7 Pa. $7.95

FIFTY CHALLENGING PROBLEMS IN PROBABILITY WITH SOLU-

TIONS, Frederick Mosteller. Remarkable puzzlers, graded in difficulty, illustrate
elementary and advanced aspects of probability. Detailed solutions. 88pp. 5% x 854.
65355-2 Pa. $4.95

EXPERIMENTS IN TOPOLOGY, Stephen Barr. Classic, lively explanation of

one of the byways of mathematics. Klein bottles, Moebius strips, projective planes,
map coloring, problem of the Koenigsberg bridges, much more, described with
clarity and wit. 43 figures. 210pp. b% x m. 25933-1 Pa. $6.95

RELATIVITY IN ILLUSTRATIONS, Jacob T. Schwartz. Clear nontechnical

treatment makes relativity more accessible than ever before. Over 60 drawings
illustrate concepts more clearly than text alone. Only high school geometry needed.
Bibliography. 128pp. 6% x 9%. 25965-X Pa. $7.95

AN INTRODUCTION TO ORDINARY DIFFERENTIAL EQUATIONS, Earl

A. Coddington. A thorough and systematic first course in elementary differential
equations for undergraduates in mathematics and science, with many exercises and
problems (with answers). Index. 304pp. 5% x m. 65942-9 Pa. $8.95

FOURIER SERIES AND ORTHOGONAL FUNCTIONS, Harry F. Davis. An

incisive text combining theory and practical example to introduce Fourier series,
orthogonal functions and applications of the Fourier method to boundary-value
problems. 570 exercises. Answers and notes. 416pp. 5% x 854. 65973-9 Pa. $1 1.95

AN INTRODUCTION TO ALGEBRAIC STRUCTURES, Joseph Landin.

Superb self-contained text covers "abstract algebra": sets and numbers, theory of
groups, theory of rings, much more. Numerous well-chosen examples, exercises.
247pp. 5% x 8!/2. 65940-2 Pa. $8.95

Prices subject to change without notice.

Available at your book dealer or write for free Mathematics and Science Catalog to Dept. GI,
Dover Publications, Inc., 31 East 2nd St., Mineola, N Y. 1 1501. Dover publishes more than 175
books each year on science, elementary and advanced mathematics, biology, music, art,
literature, history, social sciences and other areas.
(continued from front flap)

Matrices and Linear Algebra, Hans Schneider and George Phillip

Barker. (66014-1) $10.95
Modern Algebra, Seth Warner. (66341-8) $17.95
Advanced Calculus, David V. Widder. (66103-2) $11.95
Linear Programming and Economic Analysis, Robert Dorfman, Paul A.
Samuelson and Robert M. Solow. (65491-5) $14.95
An Introduction to the Calculus of Variations, Charles Fox.
(65499-0)
Applied Analysis, Cornelius Lanczos. (65656-X) $13.95
Topology, George McCarty. (65633-0) $8.95
Lectures on Classical Differential Geometry (Second Edition), Dirk
J. Struik. (65609-8) $8.95
Challenging Mathematical Problems with Elementary Solutions,
A.M. Yaglom and I.M. Yaglom. (65536-9, 65537-7) Two-volume
set $14.90
Asymptotic Expansions of Integrals, Norman Bleistein and Richard A.
Handelsman. (65082-0) $12.95
Some Theory of Sampling, W. Edwards Deming. (64684-X) $15.95
Statistical Adjustment of Data, W. Edwards Deming. (64685-8) $8.95
Introduction to Linear Algebra and Differential Equations, John W.
Dettman. (65191-6) $10.95
Calculus of Variations with Applications, George M. Ewing.
(64856-7) $8.95
Introduction to Difference Equations, Samuel Goldberg.
(65084-7) $7.95
Probability: An Introduction, Samuel Goldberg. (65252-1) $8.95
Group Theory, W. R. Scott. (65377-3)
An Introduction to Linear Algebra and Tensors, M.A. Akivis and V.V.
Goldberg. (63545-7) $5.95
Non-Euclidean Geometry, Roberto Bonola. (60027-0) $10.95
Introduction to Nonlinear Differential and Integral Equations, H.T.
Davis. (60971-5) $11.95
Foundations of Modern Analysis, Avner Friedman. (64062-0) $7.95
Differential Geometry, Heinrich W. Guggenheimer. (63433-7) $8.95
Ordinary Differential Equations, E.L. Ince. (60349-0)
Lie Algebras, Nathan Jacobson. (63832-4) $8.95

Paperbound unless otherwise indicated. Prices subject to change

without notice. Available at your book dealer or write for free catalogues
to Dept. 23, Dover Publications, Inc., 31 East 2nd Street, Mineola, N.Y.
11501. Please indicate field of interest. Each year Dover publishes over
200 books on fine art, music, crafts and needlework, antiques,
languages, literature, children's books, chess, cookery, nature,
anthropology, science, mathematics, and other areas.
Manufactured in the U.S.A.
ORDINARY DIFFERENTIAL EQUATIONS
IN THE
COMPLEX DOMAIN

EINAR H1LLE

Few domains of mathematics are as widely useful as the theory of ordinary dif-
ferential equations. And while every book on the topic contains at least some
material on the behavior of the solutions in the complex domain, this highly
regarded graduate-level text is unique in that it was the first book in English to
deal exclusively with the subject.

Moreover, Professor Hille's book incorporates much material not available in

other treatises and offers full and extensive treatments of existence theorems,
representation of solutions by series, representation by integrals, theory of majo-
rants, dominants and minorants, questions of growth, applications of the modem
value distribution theory and more. Included are investigations by Boutroux,
Briot-Bouquet, Malmquist, Painleve, R. Smith, Wittich and Yosida.

For students not conversant with the required complex analysis, review materi-
al has been supplied throughout the text. In addition, each chapter has a list of
references to the literature, and there is a bibliography at the end of the book.
The exercises at the ends of sections comprise some 675 items.

Unabridged Dover (1997) republication of the work published by John Wiley &
Sons, New York, 1976. Preface. References at chapter ends. Bibliography.
Index. 17 text figures. 496pp. 5% x 8%. Paperbound.

ALSO AVAILABLE
The Qualitative Theory of Ordinary Differential Equations: An Introduction,
Fred Brauer and John A. Nohel. 320pp. 5% x 8/2. 65846-5 Pa. $8.95
Modern Elementary Differential Equations: Second Edition, Richard Bellman
and Kenneth L. Cooke. 240pp. 5% x 8/2. 68643-4 Pa. $8.95
Ordinarv Differential Equations, Edward L. Ince. 558pp. 5% x H%. 60349-0 Pa.
SI 3.95

Free Dover Complete Mathematics and Science Catalog (59065-8) available

upon request.

ISBN D-4flh-bclt2D-Q
90000