Basic Theory of Ordinary Differential Equations Universitext PDF

BASIC THEORY
OF ORDINARY
DIFFERENTIAL
EQUATIONS
Universitext
Editorial Board
(North America):
S. Axler
F.W. Gehring
K.A. Ribet
Springer
New York
Berlin
Heidelberg
Barcelona
Hong Kong
London
Milan
Paris
Singapore
Tokyo
Universitext
Editors (North America): S. Axler, F.W. Gehring, and K.A. Ribet
Aksoy/Khamsl: Nonstandard Methods in Fixed Point Theory

Anderson: Topics in Complex Analysis
Aupetit: A Primer on Spectral Theory
Berberian: Fundamentals of Real Analysis
BoossBleecker: Topology and Analysis
Borkar: Probability Theory: An Advanced Course
BSttcher/Sibermann: Introduction to Large Truncated Toeplitt Matrices
Carleson/Gamelin: Complex Dynamics
Cecil: Lic Sphere Geometry: With Applications to Submanifolds
Chae: Lchcsgue Integration (2nd cd.)
Charlap: Biebcrbach Groups and Flat Manifolds
Chern: Complex Manifolds Without Potential Theory
Cohn: A Classical Invitation to Algebraic Numbers and Class Fields
Curtis: Abstract Linear Algebra
Curtis: Matrix Groups
DiBenedetto: Degenerate Parabolic Equations
Dimca: Singularities and Topology of Hypersurfaces
Edwards: A Formal Background to Mathematics I a/b
Edwards: A Formal Background to Mathematics 11 a/b
Foulds: Graph Theory Applications
Friedman: Algebraic Surfaces and Holomorphic Vector Bundles
Fuhrmann: A Polynomial Approach to Linear Algebra
Gardiner: A First Course in Group Theory
Girding/Tambour: Algebra for Computer Science
Goldblatt: Orthogonality and Spacctime Geometry
Gustafson/Rao: Numerical Range: The Field of Values of Linear Operators and Matrices
Hahn: Quadratic Algebras. Clifford Algebras, and Arithmetic Witt Groups
Holmgren: A First Course in Discrete Dynamical Systems
Howe/Tan: Non-Abclian Harmonic Analysis: Applications of SL(2. R)
Howes: Modem Analysis and Topology
Hsieh/Sibuya: Basic Theory of Ordinary Differential Equations
Humi/Miller: Second Course in Ordinary Differential Equations
Hurwitz/Kritikos: Lectures on Number Theory
Jennings: Modern Geometry with Applications
Jones/Morris/Pearson: Abstract Algebra and Famous Impossibilities
Kannan/Krueger: Advanced Analysis
Kelly/Matthews: The Non-Euclidean Hyperbolic Plane
Kostrikin: Introduction to Algebra
Luecking/Rubel: Complex Analysis: A Functional Analysis Approach
MacLane/Moerditjk: Sheaves in Geometry and Logic
Marcus: Number Fields
McCarthy: Introduction to Arithmetical Functions
Meyer: Essential Mathematics for Applied Fields
Mines/Riehman/Ruitenburg: A Course in Constructive Algebra
(cnnrinued after index)

Po-Fang Hsieh Yasutaka Sibuya
Basic Theory of Ordinary

Differential Equations
With 114 Illustrations
Springer
Po-Fang Hsieh Yasutaka Sibuya
Department of Mathematics School of Mathematics
and Statistics University of Minnesota
Western Michigan University 206 Church Street SE
Kalamazoo, MI 49008 Minneapolis, MN 55455
USA USA
philip.hsieh@wmich.edu sibuya 0 math. umn.edu
Editorial Board
(North America):
S. Axler F.W. Gehring K.A. Ribet
Mathematics Department Mathematics Department Department of
San Francisco State East Hall Mathematics
University University of Michigan University of California
San Francisco, CA 94132 Ann Arbor. Ml 48109- at Berkeley
USA 1109 Berkeley, CA 94720-3840
USA USA
Mathematics Subject Classification (1991): 34-01
Library of Congress Cataloging-in-Publication Data

Hsieh, Po-Fang.
Basic theory of ordinary differential equations I Po-Fang Hsieh,
Yasutaka Sibuya.
p. cm. - (Unrversitext)
Includes bibliographical references and index.
ISBN 0-387-98699-5 (alk. paper)
1. Differential equations. I. Sibuya. Yasutaka. 1930-
II. Title. III. Series
OA372_H84 1999
515'.35-dc2l 99-18392
Printed on acid-free paper.
r 1999 Spnnger-Verlag New York, Inc.

All rights reserved. This work may not be translated or copied in whole or in part without the
written permission of the publisher (Springer-Verlag New York, Inc.. 175 Fifth Avenue, New
York, NY 10010, USA), except for brief excerpts in connection with reviews or scholarly analysis.
Use in connection with any form of information storage and retrieval, electronic adaptation,
computer software, or by similar or dissimilar methodology now known or hereafter developed
is forbidden.
The use of general descriptive names, trade names, trademarks. etc., in this publication, even
if the former are not especially identified, is not to be taken as a sign that such names, as
understood by the Trade Marks and Merchandise Marks Act. may accordingly be used freely
by anyone.
Production managed by MaryAnn Cotton: manufacturing supervised by Jeffrey Taub.
Photocomposed copy prepared from the authors' A5-71EX 2.1 files.
Printed and bound by R.R. Donnelley and Sons, Harrisonburg. VA.
Printed in the United States of America.
9 8 7 6 5 4 3 2 1
ISBN 0-387-98699-5 Springer-Verlag New York Berlin Heidelberg SPIN 10707353

To Emmy and Yasuko
PREFACE
This graduate level textbook is developed from courses in ordinary differential

equations taught by the authors in several universities in the past 40 years or
so. Prerequisite of this book is a knowledge of elementary linear algebra, real
multivariable calculus, and elementary manipulation with power series in several
complex variables. It is hoped that this book would provide the reader with the
very basic knowledge necessary to begin research on ordinary differential equations.
To this purpose, materials are selected so that this book would provide the reader
with methods and results which are applicable to many problems in various fields.
In order to accomplish this purpose, the book Theory of Differential Equations by
E. A. Coddington and Norman Levinson is used as a role model. Also, the teaching
of Masuo Hukuhara and Mitio Nagumo can be found either explicitly or in spirit
in many chapters. This book is useful for both pure mathematician and user of
mathematics.
This book may be divided into four parts. The first part consists of Chapters I,
II, and III and covers fundamental existence, uniqueness, smoothness with respect
to data, and nonuniqueness. The second part consists of Chapters IV, VI, and VII
and covers the basic results concerning linear differential equations. The third part
consists of Chapters VIII, IX, and X and covers nonlinear differential equations.
Finally, Chapters V, XI, XII, and XIII cover the basic results concerning power
series solutions.
The particular contents of each chapter are as follows. The fundamental exis-
tence and uniqueness theorems and smoothness in data of an initial problem are
explained in Chapters I and II, whereas the results concerning nonuniqueness are
explained in Chapter III. Topics in Chapter III include the Kneser theorem and
maximal and minimal solutions. Also, utilizing comparison theorems, some suf-
ficient conditions for uniqueness are studied. In Chapter IV, the basic theorems
concerning linear differential equations are explained. In particular, systems with
constant or periodic coefficients are treated in detail. In this study, the S-N de-
composition of a matrix is used instead of the Jordan canonical form. The S-N
decomposition is equivalent to the block-diagonalization separating distinct eigen-
values. Computation of the S-N decomposition is easier than that of the Jordan
canonical form. A detailed explanation of linear Hamiltonian systems with constant
or periodic coefficients is also given. In Chapter V, formal power series solutions
and their convergence are explained. The main topic is singularities of the first
kind. The convergence of formal power series solutions is proven for nonlinear
systems. Also, the transformation of a linear system to a standard form at a sin-
gular point of the first kind is explained as the S-N decomposition of a linear
differential operator. The main idea is originally due to R. Gerard and A. H. M.
Levelt. The Gerard-Levelt theorem is presented as the S-N decomposition of a
matrix of infinite order. At the end of Chapter V, the classification of the singu-
vii
viii PREFACE
larities of linear differential equations is given. In Chapter VI, the main topics are
the basic results concerning boundary-value problems of the second-order linear
differential equations. The comparison theorems, oscillation and nonoscillation of
solutions, eigenvalue problems for the Sturm-Liouville boundary conditions, scat-
tering problems (in the case of reflectionless potentials), and periodic potentials are
studied. The authors learned much about the scattering problems from the book
by S. Tanaka and E. Date [TD]. In Chapter VII, asymptotic behaviors of solu-
tions of linear systems as the independent variable approaches infinity are treated.
Topics include the Liapounoff numbers and the Levinson theorem together with
its various improvements. In Chapter VIII, some fundamental theorems concern-
ing stability, asymptotic stability, and perturbations of 2 x 2 linear systems are
explained, whereas in Chapter IX, results on autonomous systems which include
the LaSalle-Lefschetz theorem concerning behavior of solutions (or orbits) as the
independent variable tends to infinity, the basic properties of limit-invariant sets
including the Poincar6-Bendixson theorem, and applications of indices of simple
closed curves are studied. Those theorems are applied to some nonlinear oscillation
problems in Chapter X. In particular, the van der Pot equation is treated as both
a problem of regular perturbations and a problem of singular perturbations. In
Chapters XII and XIII, asymptotic solutions of nonlinear differential equations as
a parameter or the independent variable tends to its singularity are explained. In
these chapters, the asymptotic expansions in the sense of Poincare are used most
of time. However, asymptotic solutions in the sense of the Gevrey asymptotics are
explained briefly. The basic properties of asymptotic expansions in the sense of
Poincare as well as of the Gevrey asymptotics are explained in Chapter XI.
At the beginning of each chapter, the contents and their history are discussed
briefly. Also, at the end of each chapter, many problems are given as exercises. The
purposes of the exercises are (i) to help the reader to understand the materials in
each chapter, (ii) to encourage the reader to read research papers, and (iii) to help
the reader to develop his (or her) ability to do research. Hints and comments for
many exercises are provided.
The authors are indebted to many colleagues and former students for their valu-
able suggestions, corrections, and assistance at the various stages of writing this
book. In particular, the authors express their sincere gratitude to Mrs. Susan Cod-
dington and Mrs. Zipporah Levinson for allowing the authors to use the materials
in the book Theory of Differential Equations by E. A. Coddington and Norman
Levinson.
Finally, the authors could not have carried out their work all these years without
the support of their wives and children. Their contribution is immeasurable. We
thank them wholeheartedly.
PFH
YS
March, 1999
CONTENTS
Preface vii
Chapter I. Fundamental Theorems of Ordinary Differential Equations 1
I-1. Existence and uniqueness with the Lipschitz condition 1

1-2. Existence without the Lipschitz condition 8
1-3. Some global properties of solutions 15
1-4. Analytic differential equations 20
Exercises I 23
Chapter II. Dependence on Data 28

II-1. Continuity with respect to initial data and parameters 28
11-2. Differentiability 32
Exercises II 35
Chapter III. Nonuniqueness 41

III-1. Examples 41
111-2. The Kneser theorem 45
111-3. Solution curves on the boundary of R(A) 49
111-4. Maximal and minimal solutions 52
111-5. A comparison theorem 58
111-6. Sufficient conditions for uniqueness 61
Exercises III 66
Chapter IV. General Theory of Linear Systems 69

IV-1. Some basic results concerning matrices 69
IV-2. Homogeneous systems of linear differential equations 78
IV-3. Homogeneous systems with constant coefficients 81
IV-4. Systems with periodic coefficients 87
IV-5. Linear Hamiltonian systems with periodic coefficients 90
IV-6. Nonhomogeneous equations 96
IV-7. Higher-order scalar equations 98
Exercises IV 102
Chapter V. Singularities of the First Kind 108

V-1. Formal solutions of an algebraic differential equation 109
V-2. Convergence of formal solutions of a system of the first kind 113
V-3. The S-N decomposition of a matrix of infinite order 118
V-4. The S-N decomposition of a differential operator 120
V-5. A normal form of a differential operator 121
V-6. Calculation of the normal form of a differential operator 130
V-7. Classification of singularities of homogeneous linear systems 132
Exercises V 137
ix
x CONTENTS
Chapter VI. Boundary-Value Problems of Linear Differential

Equations of the Second-Order
VI- 1. Zeros of solutions
VI- 2. Sturm-Liouville problems
VI- 3. Eigenvalue problems
VI- 4. Eigenfunction expansions
VI- 5. Jost solutions
VI- 6. Scattering data
VI- 7. Reflectionless potentials
VI- 8. Construction of a potential for given data
VI- 9. Differential equations satisfied by reflectionless potentials
VI-10. Periodic potentials
Exercises VI
Chapter VII. Asymptotic Behavior of Solutions of Linear Systems

VII-1. Liapounoff's type numbers
VII-2. Liapounoff's type numbers of a homogeneous linear system
VII-3. Calculation of Liapounoff's type numbers of solutions
VII-4. A diagonalization theorem
VII-5. Systems with asymptotically constant coefficients
VII-6. An application of the Floquet theorem
Exercises VII
Chapter VIII. Stability

VIII- 1. Basic definitions
VIII- 2. A sufficient condition for asymptotic stability
VIII- 3. Stable manifolds
VIII- 4. Analytic structure of stable manifolds
VIII- 5. Two-dimensional linear systems with constant coefficients
VIII- 6. Analytic systems in R2
VIII- 7. Perturbations of an improper node and a saddle point
VIII- 8. Perturbations of a proper node
VIII- 9. Perturbation of a spiral point
VIII-10. Perturbation of a center
Exercises VIII
Chapter IX. Autonomous Systems

IX-1. Limit-invariant sets
IX-2. Liapounoff's direct method
IX-3. Orbital stability
IX-4. The Poincar6-Bendixson theorem
IX-5. Indices of Jordan curves
Exercises IX
CONTENTS xi
Chapter X. The Second-Order Differential Equation

-t2 + h(x) dt + g(x) = 0 304
d
X-1. Two-point boundary-value problems 305
X-2. Applications of the Liapounoff functions 309
X-3. Existence and uniqueness of periodic orbits 313
X-4. Multipliers of the periodic orbit of the van der Pol equation 318
X-5. The van der Pol equation for a small e > 0 319
X-6. The van der Pol equation for a large parameter 322
X-7. A theorem due to M. Nagumo 327
X-8. A singular perturbation problem 330
Exercises X 334
Chapter XI. Asymptotic Expansions 342

XI-1. Asymptotic expansions in the sense of Poincare 342
XI-2. Gevrey asymptotics 353
XI-3. Flat functions in the Gevrey asymptotics 357
XI-4. Basic properties of Gevrey asymptotic expansions 360
XI-5. Proof of Lemma XI-2-6 363
Exercises XI 365
Chapter XII. Asymptotic Expansions in a Parameter 372

XII-1. An existence theorem 372
XII-2. Basic estimates 374
XII-3. Proof of Theorem XII-1-2 378
XII-4. A block-diagonalization theorem 380
XII-5. Gevrey asymptotic solutions in a parameter 385
XII-6. Analytic simplification in a parameter 390
Exercises XII 395
Chapter XIII. Singularities of the Second Kind 403

XIII-1. An existence theorem 403
XIII-2. Basic estimates 406
XIII-3. Proof of Theorem XIII-1-2 417
XIII-4. A block-diagonalization theorem 420
XIII-5. Cyclic vectors (A lemma of P. Deligne) 424
XIII-6. The Hukuhara-T rrittin theorem 428
XIII-7. An n-th-order linear differential equation at a singular point
of the second kind 436
XIII-8. Gevrey property of asymptotic solutions at an irregular
singular point 441
Exercises XIII 443
References 453
Index 462
CHAPTER I
FUNDAMENTAL THEOREMS OF
ORDINARY DIFFERENTIAL EQUATIONS
In this chapter, we explain the fundamental problems of the existence and unique-
ness of the initial-value problem
dy
(P)
dt
= y(to) _ 4o
in the case when the entries of f (t, y) are real-valued and continuous in the variable
(t, y), where t is a real independent variable and y is an unknown quantity in Rn.
Here, R is the real line and R" is the set of all n-column vectors with real entries.
In §I-1, we treat the problem when f (t, y) satisfies the Lipschitz condition in W. The
main tools are successive approximations and Gronwall's inequality (Lemma
In §1-2, we treat the problem without the Lipschitz condition. In this case, ap-
proximating f (t, y) by smooth functions, f-approximate solutions are constructed.
In order to find a convergent sequence of approximate solutions, we use Arzeli -
Ascoli's lemma concerning a bounded and equicontinuous set of functions (Lemma
1-2-3). The existence Theorem 1-2-5 is due to A. L. Cauchy and G. Peano [Peal]
and the existence and uniqueness Theorem 1-1-4 is due to E. Picard [Pi1 and E.
Lindelof [Lindl, Lind2J. The extension of these local solutions to a larger interval
is explained in §1-3, assuming some basic requirement.,, for such an extension. In
§I-4, using successive approximations, we explain the power series expansion of a
solution in the case when f (t, y) is analytic in (t, y). lit each section, examples and
remarks are given for the benefit of the reader. In particular, remarks concerning
other methods of proving these fundamental theorems are given at the end of §1-2.
I-1. Existence and uniqueness with the Lipschitz condition

We shall use the following notations throughout this book:
Y1, f,
y2 f2
y= !y = max{jy,j 1Y21-.. ItYn11.
yn fn
In §§I-1, 1-2, and 1-3 we consider problem (P) under the following assumption.
Assumption 1. The entries of f (t, y) are real-valued and continuous on a rectan-
gular region:
R = R((to,6),a,b) = {(t,yl: It - tot <a, ly - e< b},

where a and b are two positive numbers.
I
2 I. FUNDAMENTAL THEOREMS OF ODES
Since f (t, yam) is continuous on R, I f (t, yul is bounded. Let us denote by M the
maximum value of I f (t, y-) I on R, i.e., M = mac I f (t, y) I. We define a positive
number a by
a if M = 0,
a
min Ca, M if M > 0.
l
To locate a solution in a neighborhood of its initial point, the number or plays an
important role as shown in the following lemma.
Lemma I-1-1. If ff = ¢(t) is a solution of problem (P) in an interval it - tol <
a < a, then I'(t) - 41 < b in It - tol < a, i.e., (t,5(t)) E R((to,co),a,b) for.
It - tol <a.
Proof
Assume that Lemma 1-1-1 is not true. Then, by the continuity of fi(t), there
exists a positive number $ such that
(i) < a,
(ii)
1¢(t) - 41 < b for It - tol <'6,
I4(to+0) -QoI =b or 10(to-fl)-qol
The assumption a < a < a implies (3 < a. Hence, (t,j(t)) E R for It - tot < p.
Therefore, I f (t, (t))I 5 M for it - to! < 0. From 4'(t) = f (t, ¢(t)) and 4(to) = 4o,
it follows that
(1.1.2) fi(t) = co + / t f(s, 0(s))ds for It -tot < fi.

io
Hence,
I4(t) - 4)1 = I1 f(s, i(s))ds < MIt - tol for It - tot < P.
r
This implies that
(1.1.4) 1 (t)-ZbI<-MQ<Ma<Ma<MM=b for It - tol<f.
In particular, 0) - cot < b. This contradicts assumption (ii).

Similarly, we obtain the following result.
Lemma 1-1-2. If y' = 4(t) is a solution of problem (P) in an interval it - tol <
a < a, then I¢(t,) - (t2)I < Mlt1 - t2l whenever t1 and t2 are in the interval
it - tot < a.
1. EXISTENCE AND UNIQUENESS WITH THE LIPSCHITZ CONDITION 3
Remark 1-1-3. Using Lemma 1-1-2, it is concluded that if y = j(t) is a solution

of problem (P) on an interval It - to I < it < a, then (to + & - 0) and (to - ar + 0)
exist.
Now, consider problem (P) with Assumption 1 and the following assumption.
Assumption 2. The function f satisfies a Lipschitz condition on R: i.e., them
exists a positive constant L /such that
If(t, 91) - f (t, W2)1:5 LIlh - 1121
whenever (t,yi) and (t, y2) are on the region R (cf. [Lip]).
Note that L is dependent on the region R. The constant L is called the Lipschitz
constant of f with respect to R.
The following theorem is the main conclusion of this section.
Theorem 1-1-4. Under Assumptions I and 2, them exists a unique solution of
problem (P) on the interval It - to] < a.
In order to prove this theorem, the following lemma is useful.
Lemma 1-1-5 (T. H. Gronwall [Cr]). If
(i) g(t) is continuous on to < t < ti,
(ii) g(t) satisfies the inequality
(1.1.5) 0 < g(t) < K + L g(s)ds on to < t < ti,

to
then
(I.1.6) 0 < g(t) < Kexp[L(t - to))
onto<t< ti.
Proof.
Set v(t) =
Jio g(s)ds. Then, by (1.1.5), ddtt) < K+Lv(t) and v(to) = 0. Hence,
(1.1.7) {exp[-L(t - to)]v(t)} < Kexp[-L(t - to)]
and, consequently,
exp[-L(t - to)]v(t) < {1 - exp(-L(t - to)]).

L
This, in turn, implies that Lv(t) < K{exp[L(t - to)] - 1). Therefore, g(t) <
K + Lv(t) < K exp[L(t - to)]. 0
Proof of Theorem 1-1-4.
We shall prove this theorem in six steps by using successive approximations
which are defined as follows:
Wi(t) _ 4o,
(1.1.9)
k=0,1,2,...
t $k+1(t) = Cq + f (s,Ok(s))ds,
:o
Step 1. Each function 0k is well defined and (t,45k(t)) E R for It - tol < a
(k = 0,1,2,...).
Proof.
We use mathematical induction in Steps 1 and 2. The statement above is evident

for k = 0. Assume that (t, Ok(t)) E R on It - tol < a. Then, I f (t,Ok(t))I < M on
it - toI < a. Hence,
t
IOk+I(t) -Q = f (s, Qsk(s))dsl < MIt - tol < Ma < MM = b.
4 1 -
Thus, (t.0k+1(t)) E R for It - tol < a.

Step 2. The successive approximations given by (1.1.9) satisfy the estimates
MLk
Ilk+1(t) - mk(t)I < (k 1)1It - toIk+1 for It - tol < a, k =0,1,2,... .
Proof.
For k = 0, we have 1m1(t) - y5o(t)I = J r f(s,co)ds` < MIt - tol. Assume that
I
co I
ML
I&(t)(t)I < k1-1 It - tolk for it - tol < a. Then,
14k+I (t) - Qk(t)l =J o

`f (s, k(s)) - f (s k-1(s))} ds
<L I J r
to
t k
AkL
I r It - toIkdsl = (ML1)r It - tolk+1.
+oo
Step 3. The series E (&(t) - 4k-1(t)) is uniformly convergent on the interval
k=1
It - tol<a.
Proof.
By virtue of the result of Step 2, for a given c > 0, there exists a positive integer
N such that
°° M °o (LIt - tol)k M °O (La)k
/ <
E I&(t) - 0k-I(t)I < L k1 <L k!
k=N k=N k=N
Step 4. The sequence {k(t) : k = 0, 1, 2.... } converges to
do(t) + F, I(&(t) - k-1(t))

k=1
uniformly on it - tol < a as k -+ +oo.

Proof.
Use the identity ¢N (t) = do (t) + E ($k (t) - $k_1(t)) and the result of Step 3.
k=1
Step 5. fi(t) given by (I.1.10) is a solution of problem (P) on It - toI < a.

Proof.
The definition
rt
&+1(t) = e + AS, ('1k(s))ds,
Jca
the continuity of f and the results of Steps 3 and 4 imply that
fi(t) = 6o + f f(s, (s))ds.

a
Step 6. fi(t) is the unique solution of problem (P) on It - tol < a.

Proof.
Suppose that y"= fi(t) is another solution of problem (P) on an interval it -toI <
& < a for some &. Lemma I-1-1 and Remark 1-1-3 imply that (t, ti(t)) E R on
It - tol :5&. Note that
fi(t) + f (s, i1(s))ds
on It - tol < &. Hence,
At) - lL(t)I = f' {f(s,.(s)) - f(s,iG(S))} dsl <- L I fo Id(s) - i (s)Idsl
on it - tol < a. Now, by applying Lemma 1-1-5 with K = 0 to the inequality
L f Id(s) - +L(s)Ids
o
on to < t < to+&, we conclude that I¢(t) -tG(t)I = 0 on the interval to < t < to+&.
Similarly, the same conclusion is obtained on the interval to - & < t < to. Steps 1
through 6 complete the proof of Theorem 1-1-4. 0
The following lemma gives a sufficient condition that f (t, y") satisfies Assumption
2 (i.e., the Lipschitz condition).
Lemma I-1-6. If Of (t' y-) (j = 1, 2, ... , n) exist, are continuous, and satisfy the
8yj
condition: I of (t, y j < Lo (j = 1, 2, .. - , n) for (t, y) E I x D, where I is an
1
interval, D is a region, and Lo is a constant, and if V is a convex open set, then

I f(t, y,) - f(t, y2)1 < nLollit - y"2I for t E z, gi E D, and y2 E D.
Proof.
Fix t E 1, and yl and #2 in V. Then, By, + (1 - O)y"2 E D for O < 6 < 1 since D
is convex. Setting g(9) = f (t, 09, + (1 - O)y2), we derive
9(O) = At, 92), 9(1) = f(t,91)-

n
44
do _ 1: (yia - y2 j)' (t, 8y"1 + (1 - B)y'2),
1=1
where yh,1, Yh,2, ... , Yh.n are entries of yh (h = 1,2). Hence,
If(t,91)-f(t,y2)1 =II @(O)dO <nLolii-u21-
This is true for all t E Z, yl E D, and y2 E D. Thus, the lemma is proved.

The following two simple problems illustrate Theorem 1-1-4.
Problem 1. Using Theorem 1-1-4, show that on the interval It - 11 < V"2-, the
initial-value problem
dy 1
- 5)2)'
y(1) - 5
dt (3 - (t - 1)2)(9 - (y -
has one and only one solution.

Answer.
The standard strategy is to find a suitable region
R = {(t,y): It-11 < a, ly-51 < b}

and find the maximum value M of if (t, y) I in R. Then, a = min (a, M)- If
a > f, the problem is solved. In other words, the main idea is to trap the
solution curve in a suitable region R. To do this, for example, use the fact that the
function
,y
f (3 - (t - 1)2)(9 - (y - 5)2)
is continuous and satisfies the Lipschitz condition
Iy' -Y21
If(t,Yi) - f(t,y2)1 < 5
on the region R= {(t, y) : It-11 < f, Iy-51 < 2} (cf. Exercise I-2). Furthermore,
If(t,y)I <- 1
(3_(/)2)(9_22) = 1
5
1
Set a = V25, b = 2, and M = . Then,
5
a = min ( a, ) =min f, 2 min(v ,10) = f.
Therefore, the given initial-value problem has one and only one solution on the
interval It - 11 < f.
Problem 2. Using Theorem I-1-4, show that the initial-value problem
(P) dy = 1 y()
4 = 3
dt (1 + (t - 4)2)(5 + (y - 3)2)'
has one and only one solution on the interval -oo < t < +oc.
Answer.
The function
1
,y
(1 + (t - 4)2)(5 + (y - 3)2)
is continuous and satisfies the Lipschitz condition everywhere in the (t, y)-plane.
Also, If (t, y)I < everywhere. This implies that a = min(a, 5b), if Problem (P)
is considered in a region
R = {(t, y) : It - 41 < a, ly - 31 < b}.
Hence, if a < 5b, Problem (P) has one and only one solution on the interval it - 41 <
a. Now, this solution is independent of a due to the uniqueness. Therefore, we can
conclude that (P) has one and only one solution on the interval -oo < t < +oo.
Remark 1-1-7. It is usually claimed that the general solution of the system
dt = f (t, y-) depends on n independent arbitrary constants. Theorem 1-1-4 verifies
this claim if we define the general solution very carefully, since the initial data (to, Cl
represent n arbitrary constants for a fixed to. However, if the reader looks into this
definition a little deeper, he will find various complicated situations. It is important
for the reader to know that the number of independent arbitrary constants con-
tained in the general solution depends strongly on the space of functions to which
the general solution should belong. For example, if the differential equation
1
(E) tae =
is considered on the set j = {t E R : t A 0), the general solution is y(t) = cl +In Jtj
for t > 0 and y(t) = c2 + In JtJ for t < 0, where ct and c2 are independent arbitrary
constants. If the Heaviside function
H(t) =
{1 if t > 0,
-I if t<0
is used, this general solution is given by
c1 - c2 H(t)
y(t) = + c' + c2 + In JtI.
2 2
Equation (E) may be looked at in terms of the generalized functions (or the distri-
butions of L. Schwartz [Sc]). The transformation y = u + In Jti, changes equation
(E) to
(E') t - = 0.
As a differential equation on the generalized functions, (E') is reduced to
(E") j = cb(t),
where c is an arbitrary constant and b(t) is the Dirac delta function. Integrating
(E"), we obtain u(t) = cH(t) + c, where c is another arbitrary constant. Therefore,
the general solution of (E) is y(t) = cH(t) + c + in Jtt. For more information on
differential equations on the generalized functions, see [Ko] for examples.
The following example might be a little strange. Let us construct solutions of
the differential equation
(E) L + 2y(cost + cos(2t)) = 0
as a formal Fourier series

+'
(S) y(t) _ am eimc
m=-oo
Inserting (S) into (E) we obtain
(R) imam + am-1 + am+1 + am-2 + am+2 = 0 (m E Z),

where Z is the set of all integers. Equation (R) is a fourth-order difference equa-
tion on am. Therefore, solution (S) contains four independent arbitrary constants.
Of course, if (S) is restricted to be a convergent series, three of these four con-
stants should be eliminated. A similar phenomenon may be observed if solutions
of a certain differential equation are expanded in positive and negative powers of
+oo
independent variables such as F, amxm. For such an example, see [Dw].
m=-oo
1-2. Existence without the Lipschitz condition

To treat the initial-value problem (P) without the Lipschitz condition, we need
some preparation.
2. EXISTENCE WITHOUT THE LIPSCHITZ CONDITION 9
Definition 1-2-1. A set.F of functions f (t) defined on an interval I is said to be

equicontinuous on I if for every e > 0, there exists a b(e) > 0 such that ] f(tl) -
f (t2)I < e for every f E .F whenever it, - t2I < b(E), tl E :r and t2 E Z.
For example, if the set F consists of all functions f such that f is continuous
on a closed interval It - tol < a, f is continuous in it - to, < a, and If (t)l < L in
it - toy < a, then.F is erquioontinuous on the interval I = {t : It - tol < a). In fact,
since At 0 - f(t2) = J t l f'(s)ds for t1 E T and t2 E Z, we have ] f (tl) - f (t2)l < E
t2
whenever It1 - t2l < b(c) = L.
Definition 1-2-2. A set F of functions is said to be bounded on the interval Z,
if for every t E I there exists a non-negative number M(t) such that lf (t)I < M(t)
for every t E Z and f E.F.
Lemma 1-2-3 (C. Arzela [Ar]-G. Ascoli [As]). On a bounded interval Z, let.F be
an infinite, bounded, and equicontinuous set of functions. Then, F contains an
infinite sequence which is uniformly convergent on Z.
Proof.
Let Q be the set of all rational numbers contained in the interval Z. Then,
(1) Q is dense in Z; i.e., for every 6 > 0 and every t E Z, there exists a rational
number r(t, b) E Q such that it - r(t, b)1 < b,
(2) Q is an enumerable set; i.e., Q = {r, : j = 1, 2, 3, ... }.
Note that the choice of the rational number r(t, 6) is not unique. We prove this
lemma in two steps.
Step 1. The set .F contains an infinite sequence { fh : h = 1, 2.... } such that
lim ofh(r) exists for every rational number r E Q.
h
Proof.
Choose from F a sequence of subsequences F. 1, 2.... }, j=
1, 2, ... , such that
(i) .Fj +1 C .Fj for every j,
(ii) t lim o f',,e(rl) = c, exists for every j.
To do this, first look at functions in F at t = r1. Using the boundedness of the
set {f(r1) : f E .F}, we choose an infinite subsequence F1 = {f 1.t : t = 1, 2,... }
from .F so that lim f1,t(r1) = cl exists. Next, look at the sequence F1 at t = r2.
1 +00
Using the boundedness of the set { f l,t(r2) : t = 1, 2.... }, we choose an infinite
subsequence F2 = {j 2 t : t = 1,2.... }from F1 so that lim h,t(r2) = c2 exists.
Continuing this way, we can choose subsequences .F,+1 from Y. successively.
Set fh = fh,h (h = 1, 2, 3.... ). Then, ft, E fl for every h >) since fh = fh,h E
.Fh C .Fj if h > j. Therefore, lim fh(rj) = c, exists for every j, i.e., lira fh(r)
- h-+oo h-+oo
exists for every rational number r E Q.
Step 2. The sequence { fh : h = 1, 2.... } of Step 1 converges uniformly on the

interval I.
Proof
For a given positive number f and a rational number r E Q, there exist a positive
number b(e) and a positive integer N(r, e) such that
lfh(t) - fh(r)I <c whenever it - rl < b(e),
Ifh (r) - ft(r)I < E whenever h > N(r, E) and I > N(r, E),
1J(r) - f (t)I < E whenever it - rl < b(e).
Now, observe that
Ifh(t) - ft(t)1 <- LAW - fh(r)I + Ifh(r) - ft(r)I + I h(r) - f!(t)I,
where t E I and r r= Q. Therefore, by choosing r = r (t, b ( 3)) , we obtain
(1.2.1) Ifh(t) - ft(t)I < E

whenever h > N (r I t, 6 (3 bigg)) 3 I and t > N (r (t, b (3 )
.3 )
Since the interval I is bounded, we can cover I by a ite number of open
intervals 11, I2, ... ,1,(,) in such a way that the length of each interval I1 is not
greater than b (11) , where p(E) is a positive integer depending on E. For every t,
choose a rational number rt(e) E 1t f1 Q. For all t E It, set r t,b1 3) I = rt(E)
/E \
and set
1
N(e) = maxE N (r,(.E), 3 )
1
Then, (1.2.1) is satisfied whenever h > N(e) and I > N(E). 0

To construct a solution of the initial-value problem (P) without the Lipschitz con-
dition, approximate the given differential equation by another equation which sat-
isfies the Lipschitz condition. The unique solution of such an approximate problem
is called an E-approximate solution. Using the following lemma, an E-approximate
solution is found.
Lemma 1-2-4. Suppose that f (t, y) is continuous on a rectangular region
R={(t,y): It - tol a, Iy-col<b}.
Then, for every positive number f, there exists a function FF(t, y) such that
(i) F( is continuous for It - toI < a and all g,
(ii) 1 has continuous partial derivatives of all orders with respect to yl, y2,... , y
f o r It - tot < a and all y",
(iii) I FE(t,y)I <c** max If(r,r1)I for It -to l<aand all y",
X
(iv) IFe(t,y)-f(t,y)I <e on
r=
.
Proof.
We prove this lemma in three steps.
Step 1. There exists a function P(t, yj such that

(1) F is continuous for It - tot < a and all y,
(2) IF(t, y)I < c >SERI If (-r, for It - tot < a and all f,
(3)
{f(t on R,
F(t, yJ =
0 for It-tot <a and If-cal>b+1.
The construction of such an f is left to the reader as an exercise. Since f is
uniformly continuous on R, properties (1), (2), and (3) of f imply that
(4) for every positive number e, there exists a positive number b(e) such that
IF(t,f1)-F(t,f2)i <e whenever It - tot < a, Ig, - f2l < b(e).
Step 2. For every positive number b, there exists a real-valued function p(t, 6) of
a real variable t such that
(a) p has continuous derivatives of all orders with respect to for -cc < t < +oo,
(b) p(4, b) > 0 for -co < t < +oc,
(c) p(.,b)=0forI{I>b,
(d) f(.5)de = 1.
,o
The construction of such a function p is left to the reader as an exercise.
Set 0(g,6) = p(yi, b)p(y2, b) ... p(y,,, b). Then
(a') has continuous partial derivatives of all orders with respect to y1, y2,.,. ,y
for ally",
(b') t(#, 6) 0 for all
(c') 0(f, b) = 0 for Iyj > 6,
(d')
f
r +00 ... f +00
'(y, 6)dy1dy2 ... d1M = 1.
Step 3. Set
+- ...
4(t, YJ = 1-00
J+oo (y - 1, b(E))F(t, ff)d ... dnn,
00
where b(e) is defined in (4) of Step 1. Then, F,(t, y-) satisfies all of the requirements
(i), (ii), (iii), and (iv) of Lemma 1-2-4.
Proof
Properties (i) and (ii) are evident. Property (iii) follows from the estimate
I e(t,M1<max{IF(t,10i: all ;t} j 00
< iTMaExRIf(T,nl1.
Property (iv) follows from the estimate

I A, (t, yi - F'(t, Y) I
=l
r+oo
j + (y-;),b(E)){F(t,n7)-F(t,)}dg1...dg
00 l
_ ... r (9 - n,b(E)) {F(t,nj - F(t,y-) } dm ... dn.
E tI ... r- n, b(E))drh ... dnn l = E.

Il ff JW- l1 <_a(E)
This completes the proof of Lemma I-2-4.

Now, we prove the fundamental existence theorem without the Lipschitz condi-
tion.
Theorem 1-2-5. If f (t, y-) is continuous on a region
P- = {(t,y): It-tol < a, I#-col < b}
and M = ma c I f (t, )l, there exists a function fi(t) such that
(i) (to)
(ii) 4'(t) = f (t, 0(t)) on It - toy < a,
when
a if M = 0,
/ l
minla,i ) if M > 0.
Proof.
We prove this theorem in four steps.
Step 1. For every positive number e, construct a function FE (t, y1 which satisfies
all of the requirements given in Lemma 1-2-4. Using Theorem I-1-4, we construct
a function 0 (t) such that
(1)S(,E(to) =_4o,
(2) P. (t, E(t)) on It - toI < o.
Furthermore,
(3) (t,4(t)) E 1 on It - toj < a (cf. Lemma 1-1-1 and Remark 1-1-3).
Step 2. The set F = {¢E : E > 0} is bounded and equicontinuous on It - toI < a.
Proof
Property (3) of functions ¢E implies that F is bounded on It - tol < a and that
lFE(t, jE(t))I < M on It - tot < a. Hence, property (2) of ¢ implies that
I4(t1) - 4(t2)1:5 ljtlt2 iE(s)dsl < Mlt1- t21

if Iti - tol < a and It2 - t0I < a (cf. Lemma 1-1-2). Therefore, for a given positive
number p, we have I0E(t0 - 0E(t2)l < µ whenever Itl - toI < a, It2 - toI < a, and
It1 - t21 < -E. This shows that Jr is equicontinuous on It - tol < a.
Step 3. Using Lemma 1-2-3 (Arzeli -Ascoli) choose a sequence {ee : I = 1, 2,. .. } of
positive numbers such that lim e, = 0 and that the sequence l = 1, 2.... }
r+oo
converges uniformly on It - tot < or as t -- +oo. Then, set
,(t) = lim
+oo
,,(t) on It - to, < a.
Step 4. Observe that
(t) = c"o + fP, (s, ,(s))ds
=ca+ f(s,0,(s))ds+J {f (s,t (s))-f(s,4(s))}ds

eo to
and that
(s)) - f(s,Z (s))Ids < e It - tol on It - to; < a.

Ifca
This is true for all e > 0. Letting a - 0, we obtain
(t} = c"o + f L f (s, 0(s))ds.

u
This completes the proof of Theorem 1-2-5. 0

Example 1-2-6. Theorem 1-2-5 applies to the initial-value problem
dy
(P) xyl/s, y(3) = 0.
dx
However, Theorem I-1-4 does not apply to (P). In fact y(x) = 0 is a solution of
Problem (P) and also
0 for x < 3,
y(x) _ (2(x2_ 9)154
for x > 3
5
is a solution of Problem (P). Note that the right-hand side of the differential
equation (P) does not satisfy the Lipschitz condition on any y-interval containing
y=0.
Two other methods of proving Theorem 1-2-5 are summarized in the following
two remarks.
Remark 1-2-7. Let every entry of an n-dimensional vector f (t, y-) be a real-valued
continuous function of n + 1 independent variables t and y = (yl, ... , y,,) on a
rectangular region R = {(t, y-) : It - tot < a, Iy"-cod < b}. Assume that I f(t,yj 1 < M
on R, where M is a positive number. Set a = min (a, M ) . Since f (t, yI is

uniformly continuous on R, there exists a positive n umber p(e) for every given
positive number a such that I f (tl, y1) - f (t2,1%2)I < E Whenever Itl - t2[ < p(c) and
Iyl - 1%2I 5 p(e) for (xl, WI) E R and (x2, y2) E R. Let 0(e) : to = ro < r1 <
< rn(c) = to + a be a subdivision of the interval to < t < to + a such that
n(c)-1 p(E)
m ax (I rj - rj+ll) < nun P(E), . Set
M
co for t = to,
91(t) = 1/e(rj-1) + J (r)-1,1/c(rj_1))(t - 7-j-1) for rj_1 < t < TI,

(j = 1,2,... ,n(6)).
Then, we can show that
(i) every entry of y, (t) is piecewise linear and continuous in t and (t, &(t)) E R
on the interval to < t < to + a,
(ii) the set {yi : E > 0} is bounded and equicontinuous on the interval to < t <
to +a,
(iii) if we choose a decreasing sequence {c) : j = 1, 2,... } such that Jim E) = 0
-+oo
in a suitable way, the sequence {y',, (x) j = 1,2,. .. } converges to a solution
of the initial-value problem
(1.2.2)
dt = f (t, J- , i(to) = 0o
uniformly on to < t < to + a as j -+ +oo.
For more details, see [CL, pp. 3-5].
Remark 1-2-8. We use the same assumption, M, and a as in Remark 1-2-7. Also,
let i1(t) be a continuously differentiable function of t such that #(to) = co and
(t, #(t)) E R on the interval to - r < t < to, where r is a positive number. For
every e such that 0 < e < r, define
q(t) for to - r < t < to,
c
1%c(t) =
clo+ff(s,il(s_6))ds for to < t < to + a.
o
Then, we can show that

(i) every entry of g, (t) is continuous in t and (t, &(t)) E R on the interval to-r <
t<to+a,
(ii) every entry of 17, (t) is continuous in t on to - r < t < to +a except a jump at
x = to,
(iii) the set {y'E : 0 < c < r} is bounded and equicontinuous on the interval
to -r<t<to +a,
(iv) if we choose a decreasing sequence {c) : j = 1, 2, ... } such that lim e) = 0
)+no
in a suitable way, the sequence {yEI (x) j = 1, 2, ... } converges to a solution
of the initial-value problem (1.2.2) uniformly on xo < x < xo+a as j - +oo.
For more details, see [CL, pp. 43-44] and [Hart, pp. 10-11].
3. SOME GLOBAL PROPERTIES OF SOLUTIONS 15
Remark 1-2-9. If Ax, y-) is assumed to be measurable for each fixed y, continuous
in y for each fixed t, and I f (t, yj I is bounded by a Lebesgue-integrable function when
(t, y") E R, a result similar to Theorem 1-2-5 (Caratheodory's existence theorem)
[Ca, pp. 665-688) is obtained. Similarly, if, in addition, it satisfies an inequality
similar to Lipschitz condition with the Lipschitz constant replaced by a Lebesgue-
integrable function L(t) when (t, y-) E R, we obtain also a result similar to Theorem
1-1-4. (See [CL, p. 43), [Har2, p. 10], and [SC, p. 15).)
1-3. Some global properties of solutions

We can construct a solution to an initial-value problem (P) in a neighborhood
of the initial point by using Theorem 1-1-4 or 1-2-5. To extend such a local solution
to a larger interval of the independent variable t, the following lemma is useful.
Lemma 1-3-1. Assume that
(i) a function f(t, yi) is continuous in an open set D in the (t, y-) -space,
(ii) a function fi(t) satisfies the condition ;s (t) = f(t,a(t)), and (t, Q(t)) E 'D, in
an open interval T = {t : ri < t < r2}.
Under this assumption, if lim (t,, d(t,)) = (r1, ij) E D for some sequence {t,
)too
j = 1,2.... } of points in the interval I, then lim (t, (t)) = (r1, ij). Similarly, if
lim (t,, 0(t,)) _ (T2,771 E D for some sequence {t. : j = 1,2,... } of points in the
j+oo
interval Z, then lim (t, fi(t)) = (r2i
Prof.
We shall prove the part of the lemma which concerns the left endpoint ri of the
interval Z. The other part can be proven similarly.
Let U be an open neighborhood of (ri, rl. We show that (t, 0(t)) E U in an
interval ri < t < r(U) for some r(U) which is determined by U. Assume that the
closure of U is contained in V and that I f (t, y-) I < M in U for some positive number
M. For every positive integer j and every positive number e, consider a rectangular
region R,(e) = {(t,y-) : it - t,I < e, Iy - 4(t,)j < Me}. Then, there exist an e > 0
/ !4fe
and a j1such that (7-1,771 E Rj (e) C U. Note that e =min Ce, I and t, - e < rt .
M
Applying Lemma I-1-1 to the solution y" = d(t) of the initial-value problem
dy = f(t,y-), y(tj) = (tj), we obtain (t,y5(t)) E Rj(E) C U on the interval
r1 < t < tj. Since U is an arbitrary open neighborhood of (r1, i)), we conclude that
lim (ti,0(ti)) _ (ri,7-7) E D.
-+oo
From Lemma 1-3-1, we derive the following result concerning the maximal inter-
val on which a solution can be extended.
Theorem 1-3-2. If
(i) a function f (t, yj is continuous in an open set V in the (t, y-')-space,
(ii) a function fi(t) satisfies the condition 4 (t) = f(t, ¢(t)) and (t, fi(t)) E D, in
an open interval I = {t : r1 < t < r2},
(iii) 4(t) cannot be extended to the left of rl(or, respectively, to the right of r2)
with property (ii),
(iv) jlimo(tj,0(ti)) = (TI,fl) (or, respectively, (r2, r))) exists for some sequence
{tj :j = 1,2....l of points in the interval Z,
then the limit point (Ti, n (or, respectively, (T2, i))) must be on the boundary of V.
Proof
We shall prove this result by deriving a contradiction from the assumption that
(r1,,7) E E) (or, respectively, (,r2, i-11 E D). Applying Lemma 1-3-1 to this situation,
we obtain cr
lim(t,¢(t)) = (rl,r) (or, respectively, (r2ir))). Hence, by applying
Theorem 1-2-5 to the initial-value problem
dy
dt
= f (t, y), y(rl) = n (or, respectively, ff(r2)
the solution 45(t) can be extended to the left of rl (or, respectively, to the right of
r2). This is a contradiction. 0
The following example illustrates how to use Lemma 1-3-1.
Problem 1-3-3. Show that the solution of the initial-value problem
d2 y
2
- 2xyL = y3 - y2 y(xo) = 17, Y (xo) = (,
exists at least on the interval 0 x < xo if xo > 0, q > 0, and (is any real number.
Answer.
Assume that the solution y = 4(x) of the given initial-value problem exists on
an interval I = {x : £ < x < xo} for some positive number . Observe that
L2(.0')2+ 1.63 + 20-2] = 2xo(th')2 > 0
on Z. Hence,
(O'(x))2
+ 14(x)3
3
+ 14(x)-2
2
< 1(0'(xo))2
2
+ 14(x0)3
3
+ 14(x0)-2
2
(= Al > 0)
on Z. Therefore, we have
(4 (x)) z < 2M, 4(x)3 < 3M, 4(x)2 >

21Lt
Now, apply Lemma 1-3-1.

The proof of the following result is left to the reader as an exercise.
Corollary 1-3-4. Assume that f (t, yam) is continuous fort I < t < t2 and all y E R".
Assume also that a function fi(t) satisfies the following conditions:
(a) and d+' are continuous in a subinterval I of the interval tl < t < t2,
(b) (t) = f (t, ¢(t)) in Z.
Then, either
(i) Qi(t) can be extended to the entire interval t1 < t < t2 as a solution of the
differential equation = f (t, y), or
j
(ii) limO(t) = oo for some r in the interval ti < r < t2.
t-r
Using Corollary I-3-4, we obtain the following important result concerning a
linear nonhomogeneous differential equation
dt = A(t)yy + bb(t),
where the entries of the n x n matrix A(t) and the entries of the R"-valued function
b(t) are continuous in an open interval I = It : ti < t < t2}.
Theorem 1-3-5. Every solution of differential equation (1.3.1) which is defined in
a subinterval of the interval I can be extended uniquely to the entire interval I as
a solution of (1.3.1).
Proof
Suppose that a solution y = d(t) of (1.3.1) exists in a subinterval Z' _ It : r1 <
t < r2} of the interval I such that tj < 71 < r2 < t2. Then,
At)I <- Ato)I + jo {A(s)4(s) + b(s)} dsl

in Z', where to is any point in the interval T. Setting
K= (r2 - ri) '15'572

max Ib(t)I,
L= IA(t)I,
we obtain
in T.
o
Thus, the estimate

Kexp[Llt - toll < Kexp[L(r2 - ri)[ <00 in Z'
is derived by using Lemma 1-1-5. Hence, case (ii) of Corollary 1-3-4 is eliminated.
Since the right-hand side of (1.3.1) satisfies the Lipschitz condition, the extension
of the solution y5(t) is unique.
For nonlinear ordinary differential equations, we cannot expect to obtain the
same result as in Theorem 1-3-5. The following example shows the local blowup of
solutions of the differential equation of Problem 1-3-3.
Problem 1-3-6. Show that for any fixed t > 0, there exists a real-valued function
0(x) such that
(1) 0(x) is continuous on an interval 0 < C-x <- 6 for some small positive number
6 which depends on (,
(2) fix) is continuous on an interval 0 < - z < 6,
(3) the function
1 + (: - x)4(x)
y(x) -
F( - x)
satisfies the differential equation
y2
(Eq) dxy - 2xyL y3
ono<x<6.
Answer.
Step 1. Ifwesett=£-randy = , the given differential equation is changed
to t
(1) t2u" - 2tu' + 2u + 9(tu' - u)u = 2t(tu' - u)u + tsu-3 - tut,
where f =
dt
Step2. If wesetu= 1+u'1 and tu' differential equation (1) becomes
=w2,
(2)
_ - 2(1 + wl) + 3w2 - 2(w2 - (1 + wl))(1 + w1)
+ 2r1t(w2 - (1 +wl))(1 + w1)
+ t6 (1 + w1)-3 - (1 + w1)2.
If we set t = 0, wl = 0, and w2 = 0, the right-hand side of (2) vanishes. Now, (2)

can be written in the form
td
!W--'
= W2,
dt
dw2
(3)
t
=2w1+w2-2w1(w2-wl)
dt
+ 2{-lt(w2 - (I + w1))(1 + w1)
+ t6t4(1 + w1)-3 - tt-1(I + wl)2.
Step 3. Write (3) in the following form:
(4) te = f2w +'Pp) + tG(t,w"),
where
_ wl _ 1 0
w- [w2]' ['2 1] F(w) -[2w1(w2-wi)J'
with
G(t,w) = 22-'(w2 - (1 +w1))(1 +w1)

(5)
+ t5e(1 + w1)-3 - t:-'(1 + wl)2.
Two eigenvalues of f2 are -1 and 2, which are distinct. Let us diagonalize SZ by the
transformation
w" = Pv", where P =
Then, (4) becomes
A= [ O1
(6) tv = Av + P-' F(PV) + tP-1G(t, Pv), } .
Set ii= ti. Then, t9 = ti+t2z and F(Pv-) = t2F(Pz). Therefore, (6) becomes
(7) ti' = Aoz + At' zl,
where Ao =
[_2 0I and f (t, z) = tP-'F(Pz) + P-'C(t, tPa). If t > 0 and
if tl1 and tll are small, there exist two positive constants Ko and L such that
If(t, z-) <- tLli - {I and If(t,z)1 <- Ko + tL17.
Step 4. We change (7) to the following system of integral equations:
zi(t) = t'2 fot t fi(r,z(t))dt

(8)
z2(t) = d + t ft r 2f2(t, t(t) )dt,

o
where to 1is a sufficiently small positive number and c is an arbitrary constant, and
z' = (21 and f = 2
J.
We want to construct a bounded solution of (8) on
J
0 < t < to. To do this, first note that, if IZ(t)I < K on 0 < t < to, we obtain
t
Izt(t)I < t-2 fo rI f1(r, zlr))Idr < t-2 t r(Ko + rLK)dr
0
20+3LK < 20+ 3LK,
to
(9) Izs(t)I < Iclt + t f r-21 f2(r, z(r))Idr
< Icit + t
f t
to
r-2(Ko + rLK )dr
t
= Iclt+t (' - 1 )Ko+t (In (12)) LK

< Iclto + K0 + e-1 t0LK.
For a positive number co, fix another positive number K so that co + Ko < 2 and
then fix t0 > 0 so that to < 1, toL < 1, and toKIPI < 1. Here, IPI = max{IP.kl
1 < j, k < 2}, where P = (Pjkj? k=1. Define successive approximations by
t_2Jo
i (t, C) = m > 0-
d +t JJJ( r-2f2(r,-im(r,c))dr
to
It can be shown that the successive approximations converges uniformly for Icl < Co
and 0 < t < to. This establishes the existence of a bounded solution of the integral
equation (8). Thus, we can complete the construction of solution (Sol) of equation
(Eq).
1-4. Analytic differential equations

So far, it has been assumed that the independent variable t and the unknown
quantity y are real. Since a complex number is a two-dimensional real vector, every
result in the previous sections can be extended to the case when the entries of g'
and of the function f (t, g) are complex. Furthermore, if the function f is analytic
with respect to (t, yam), we can construct analytic solutions by means of successive
approximations. Denote by C the set of all complex numbers. The set of all n-
column vectors with complex entries is denoted by C".
4. ANALYTIC DIFFERENTIAL EQUATIONS 21
Theorem 1-4-1. Assume that each entry of a Cn-valued function f (z, yam) is given
by a power series in (z, yl, y2, ... , yn) which converges in a polydisk:
D= ((z, y-) : Iz - zol < a, ly - col < b},
where z, yl, y2, , y" are complex quantities. Assume also that there exist positive
numbers M and L such that
If(z,-)I < /M in D,
l 0z, 91) - 1(z, g2)1 5 Lull - y21 whenever (z,y,) E D (j = 1,2).
Define successive approximations by
1m0(z) = 4o,
om(Z)=co+J f((,0-m-1( ))d(, m>1.

0
where the path of integration may be taken to 1be the line segment zaz joining zo to
z in the complex z-plane. Set a = min I a, }. Then,
M
(i) For every m, the entries of the C"-valued))) function 4m(z) are power series in
z - zo which are convergent in the open disk A = {z : iz - zoj < a},
(ii) the sequence { ,,,(z) m = 0,1, ... } satisfies the estimates
!5
m
II ,c
,4n+l(Z) - om(Z)I IZ - zojm+l on 0, m = 0, 1, ... ,
(nl + 1),
(iii) the sequence {¢m(z) m = 0, 1,... } converges to
¢(z) = o(z) + (mk(z) - (Z))

k=1
uniformly in 0 as m -+ +oo,
(iv) the entries of the C"-valued function b(z) are power series in z - zo which
are convergent in 0,
(v) the function Q(z) is the unique solution to the initial-value problem
d = f(z,y), y(zo)
The proof of this theorem is left to the reader as an exercise. The reader must notice
that even if a sequence of analytic functions fn(t) of a real variable t converges to a
function f (t) uniformly on an interval Z, the limit function f (t) may not be analytic
at all. For example, any continuous function can be uniformly approximated by
polynomials on a bounded closed interval. In order to derive analyticity of f (t),
uniform convergence must be proved on a domain in the complex t-plane.
Observation 1-4-2. The estimates
MLm ZOlm+1
m+1(Z) - & (Z) 1 :5 (m+ 1)I IZ - in A, n1 = 0, 1,...
+00
imply that there exists a power series Eck(z - zo)k such that
k=0
M
&(z) = E ck(z - z,)' +O(Iz - zol"'+1), m=0,1,2,....
k=O
Observation 1-4-3. Since

m
&(Z) = $O(Z) + E (&(Z)
k=1
00
we obtain (z) - ¢,,,(z) = O(Iz - zol'"+1) and, hence, (z) _ >2Ck(Z - ZO)k.
k=0
Observation 1-4-4. Set
m 00
(Z) = Ck(Z - ZO)k + 1: Cm.k(Z - Zo)k.

k=0 k=m+1
Since cb1 < b in the disk 0, it follows that
1Cm,kI
a (k = m + 1, m + 2,...).
Note that
1 0(z) - ca
4.,k = 2?ri Izo1=P (z - zo) k+1
for any positive number p < a. This, in turn, implies that
,n oo
Ck(Z-ZO)kl :5 b IZ aZOlk
k=O k=m+1
=blzzD 1
I
a
1-IZ
ZIDI
in the disk A.
Example 1-4-5. In the case of the initial-value problem dzy = (1 - z)y2, y(0) = 2,
we obtain
2
O(z) = 2 E(k + 1)zk =
(1 - Z)2
k=O
EXERCISES I 23
by using the successive approximations
Oo(z) = 2 ,
01(z) = 2 + 4z - 2z2,
83 23
02(x) = 2 + 4z + 6x2 - - 6x4 + 4x5
2x4 2825 82zs
¢3(z) = 2 + 4z + 6z2 + 8x3 _
3 3 9
256z7 124z8 68z9 46z10 56z"
+ 63 + 9 9 9 + 9
4z13 2z14
22z12
9 + 9 - 63
Combining Theorem 1-3-5 and Theorem 1-4-1, we obtain the following theorem:
Theorem I-4-6. If all the entries of an n x n matrix A(t) and C"-valued function
b(t) are analytic on an interval Z, then every solution of linear differential equation
(1.3.1) is analytic on Z.
EXERCISES I
I-1. Solve the initial-value problem
d22 + 2y L = 0, y(r) = ,lo, b (r) = 171
The reader must consider various cases concerning the initial data (r, i , m)1).
1
I-2 Show that the function f (x y) _ satisfies the
' (3 - (x - 1)2)(9 - (y - 5)2)
Lipschitz condition
If(x,yl)-f(x,y2)I < IY1Y2I
if
Ix-1I 5 f, Iyi-5I < 2, lye-51 < 2.
1-3. Show that the initial-value problem
dy _ x3 + 3x + 1
sm y(5) = 3,
(P)
dx - 101-y2)
has one and only one solution on the interval Ix - 51 < 7.
1-4. Suppose that u(x) is continuous and satisfies the integral equation
x
u(x) = J sin(u(t))u(t)pdt
0
on the interval 0 < x < 1. Show that u(x) = 0 on this interval if p > 0. What
would happen if p < 0?
u -p
Hint. In case p < 0, the problem is reduced to the initial-value problem du =
sin(u)
dx, u(O) = 0. Thus, the solution u(x) is not identically zero and has a branch point
at x = 0.
1-5. Show that if real-valued continuous functions f (x), g(x), and h(x) satisfy the
inequalities
f(x) ? 0, g(x) < h(x) + f
J0*'C
on an interval 0 < x < xo, then
g(x) < h(x) + f =

exp
[JZ
f (n)dn, } di
ono < x < xo.
Hint. If we put y(x) = f then LY < f(x)h(x) + f (x)y and y(O) = 0.

J
1-6. Let a(t) be a real-valued function which is continuous on the interval I = It:
0 < t < 11. For every positive integer rn, set
fm (t) = I + ta(0) for 0 < t < 1
-m
and
fm(t)= L=0( 1+ma(m))]( l+lt-m/a\m)j

k+1
forkm < t < m
and k = 1, 2, ... , rn - 1. Show that the sequence {fm :M=
1, 2, ... } converges uniformly on the interval Z. Also, find lim f,,,.
m-oo
Hint. For large m, we have 1 + m a (m > 0 and

/
exp [ma(m/ - m21 < 1 + ma \m! < exp [M1 a(m)]
mh
.
for some positive number K. Also

fta(r)dr-Lmn(mj-(t-mja(ml)
` l
E,,,
M -0 lira. Em =
for some positive e,,, such that 0.

m-» too
EXERCISES I 25
1-7. Let f (t, y) be a real-valued function of two independent variables t and y which
is continuous on a region:
D = {(t, y) : 91(t) 5 y <- 92(t), r1 < t < r2}.
Suppose that
(i) 91, g'1, 92, and g2 are continuous on rl < t < T2,
(ii) 91(t) < 92(t) on rl < t < T2,
(iii)
gi(t) < f(t,9i(t)), rl < t < rz,
f 9t> f(t, 92(t),
z
(iv) Tl < t0 < r2 and g1(to) < CO < 92(to)-

Show that there exists a function 0(t) such that
(1) 4 and 0' are continuous on to < t < T2 and (t,4i(t)) E D on to < t
(2) Q'(t) = f (t, O(t)) on to < t < 72 and 0(to) = co.
1-8. Set
f(u) =
{
0
f for
for
- oo < y 5 0,
0 < y < +oo.
Also, let rt(x) be an arbitrary continuous function of x on an interval 0 < x < xo,
where xo is a positive number. Define successive approximations by
rt(r) for in = 0,
tim(x) =
f (ym_l(t))dt for in = 1, 2,
fox
Show that lim ym(x) exists uniformly on the interval 0 < x < xo.
m-+oo
Hint. For this problem, see [Na3, pp. 143-160; in particular Bemerkung 2].
1-9. Let B be a Banach space over the field R of real numbers. Also, let f (t, x)
be a B-valued function of the variables (t, x) E R x B. Consider an initial-value
problem
(IP)
d = f (t, x), X(r) _ .
where (r, {) is a given point in R x B. Assume that f satisfies the following condi-
tions:
(i) f(t,x)iscontinuous onaregion R={(t,x)ERxB:It-rl <a, IIx-t:ll<b},
where a and b are positive numbers and II II is the norm of B,
(ii) f(t.x) also satisfies the Lipschitz condition IIf(t,xl) -f(t,x2)II 5 LIIX1 -X211
whenever (t, xl) E R and (t, x2) E R, where L is a positive constant.
Show that
(1) there exists a positive number Af such that Ilf(t,x)II < M for (t,x) E R,
(2) problem (IP) has one and only one solution on the interval It - rI < or, where
b \
a= min a, I.
b
Hint. See [Huk7].
26 1. FUNDAMENTAL THEOREMS OF ODES
I-10. Let B be a Banach space over the field R of real numbers and let C be a
nonempty compact subset of B. Also, on a bounded interval Z, let F be an infinite
and equicontinuous set of C-valued functions. Show that F contains an infinite
sequence which is uniformly convergent on Z.
I-11. Let B be a Banach space over the field R of real numbers and let C be
a nonempty compact and convex subset of B. Assume that a C-valued function
f (t, x) is continuous on a region
R={(t,x)ERxB:It-rl <-a,llx-t1l <-b},
where r E R, t E B, a > 0, b > 0, and 11 11 is the norm of B. Show that
(1) there exists a positive number M such that IIf(t,z)115 M for (t,x) E R,
(2) the initial-value problem - = f (t, x), x(r) has a solution on the interval
it - rl < or, where a = min ( a, f ).
Hint. Use a method similar to that of Remark 1-2-7. See, also, [Huk7l.
1-12. Let f (t, x") be Re-valued and continuous for (t, x") E Rr+t. Assume also that
E is a nonempty closed set in R"+t. Show that in order that for every (r,)) E E,
there exists a real number a(r, depending on (r, ) such that r < a(-r, {) and
that the intial-value problem
di = f (t, xF), x(r) _ has a solution s = ¢(t, r,
satisfying the condition (t, ¢(t, r, ) E E on an interval r < t < o(r), it is necessary
and sufficient that for every (r, f E E and every positive number e, there exists a
point (a, S) E E such that
0<a-r<e and < E,
where (a, ) generaly depends on (r, {, e).

Hint. The necessity is easy to prove. For sufficiency, use an idea similar to that of
Remark 1-2-7. See, also, [Huk6].
1-13. Assume that f (x, yt, y2,... , y,,) is real-valued, continuous, and continuously
differentiable function of (x, yt, ... , on a region R. = {(x, yl,... , a<x<
b, lyj I < r (j = 1,... , n)), where a and b are real numbers such that a < b, and r is
a positive number. Assume, also, that (1) c is a real number such that a < c < b,
l
(2) 6(x) is a real-valued solution of !° = f (x, y, Ly, ... , fn-Y) on an interval
lz - cl < p, where p is a positive number such that a < c - p < c + p < b, (3)
r (k = 0,1,... n - 1). and (4) f(x,0,... ,0) = 0 on a < x < b.
dx*
Show that if there eixsts a sequence {ch : h = 1, 2, ... } of real numbers such that
Um ch = c and that ¢(ch) = 0 (h = 1, 2, ... ), then O(x) = 0 identically on
h_+00
a<x<b.
EXERCISES I 27
1-14. Let u(x, y) be a continuously differentiable function of two independent vari-

ables (x, y) in a domain D = {(x, y) : y > 0, a + Ay < x < b - Ay}, where A is a
positive number and the two quantities a and b are real numbers such that a < b.
Assume that u(x, y) is continuous on the closure f) of D and satisfies the condition
< AI e +BluF+C
I
on D, where B and C are positive numbers. Set M = max(Iu(x,0)(: a < r < b}.
Show that
u(x, y) 1 5 Me 11 + (eBV - 1)
on D. B
Hint. The last inequality is the Haar inequality. (See [Haa1, [Na5, pp. 51-561, and
[Har2, pp. 140-141]).
1-15. Let H(u, t, x, q) be a function defined on an open set in R4 containing the
point (u, t, x, q) = (0, 0, 0, 0) and satisfy a uniform Lipschitz condition with respect
to (u, q). Let O(x) be a function of class C' satisfying 0(0) = 0 and 0=(0) = 0.
Show that the initial-value problem
ut + H(u, t, x, u=) = 0, u(x, 0) = Q(x)
has at most one solution of class C1 in a neighborhood of x = 0.

Hint. This is Haar's uniqueness theorem of the given partial differential equation
(cf. the references in Exercise 1-14). Upon applying Exercise 1-14, with M = C = 0,
to the difference ti(t, x) = ul (t, x) - U2 (t, x) of two solutions ul(t, x) and U2 (t, x),
the proof follows immediately.
CHAPTER it
DEPENDENCE ON DATA
The initial-value problem
(P) = fit, yam), YAto) = 4
is equivalent to the integral equation
(E) 11(t) = 4 + t f (s, y(s))ds.
The right-hand side of (E) is regarded as a function of (t, to, ca, f, yj. In this chap-
ter, we explain some basic properties of solutions with respect to (t, to, ca, f). A
parameter a is used to represent the variable f. In §11-1, we explain the continuity
of the solution of (P) with respect to (t, to, c"o, e) at (t, r, e, co) when the solution of
(P) is unique for the initial data (r , c', ea) (Theorem 11-1-2). Theorem 11-1-2 was
first proved by I. Bendixson [Benl] for the scalar equation and later by G. Peano
(Pea2] and E. Lindelof [Lind2] for a system of equations. In §11-2, we explain the
differentiability with respect to initial data (to,ca) (Theorem 11-2-1) and with re-
spect to a parameter a (Theorem 11-2-2). The discussion of these topics can be
found in [CL, pp. 22-28, 57-60] and [Har2, pp. 94-100].
II-1. Continuity with respect to initial data and parameters

Let e be a variable which takes values in a set with a separable topology. We
consider RI-valued functions f (t, y, e) and fi(t) under the following assumption.
Assumption 1. For the initial-value problem (P), assume that
(i) f(t, y", e) is continuous in (t, y, e) in an open set Do in the (t, y, e)-space,
(ii) ¢'(t) = f (t, 6(t), co) and (t, ¢(t), eo) E Do on an interval Zo = {t : tt < t <
t2}, where eo is a fixed value of the variable e,
(iii) y ' = ¢(t) is unique in the sense that i f 1r1 = f (t, tl'(t), co) and co) E
Do on a subinterval I of the interval Zo and if y(r) = fi(r) at a point r of
the interval Z, then L,(t) = ¢(t) identically on the interval Z.
We start our discussion with the following lemma.

Lemma II-1-1. Conditions (i) and (ii) of Assumption 1 imply that there exist an
open set A in the (t, y")-space and an open neighborhood llo of co in the a-space such
that
(a) A x Uo C Do,
(b) (t, q',(t)) E A on the interval Zo,
28
1. CONTINUITY WRT INITITIAL DATA AND PARAMETERS 29
(c) I t) I < M on 0 x Uo for some positive number M.

Proof.
Let r E Zo. Since f is continuous in the open set Do and (r, ¢r(r), co) E Do,
there exist an open neighborhood 0(r) of (r, m(r)) in the (t, yl-space and an
open neighborhood U(r) of to in the a-space such that i(r) x U(r) C Do and
I f (t, y, t) - AT, fi(r), to) I < 1 on 0(r) x U(7-). This implies that I f (t, y e) I < 1 +Mo
on U 0(r) x U(r), where M o = m ax I f (r, fi(r), to)I. Since (t, fi(t)) E U 0(r)
rEZ0 TEZo
on the interval Zo and the interval Zo is compact, there exists a finite set (r, : j =
1, 2, ... , N) of points on the interval Zo such that (t, m(t)) E U A(rt) for all
N>1>1
t E Zo. Set
N N
= U,&(r)), Uo = nU(r)), M = 1+Mo.
3=1 J=1
Then, all the requirements of Lemma II-1-1 are satisfied by 0, Uo, and M. 0
The main purpose of this section is to prove the following theorem under As-
sumption 1.
Theorem II-1-2. Let E be an open neighborhood of to and let Z = {t : s1 < t < s2}
be a subinterval of Zo. Assume that
(II.1.1) ,'(t, e) = f (t, >¢(t, t), e) and (t, ir(t, t), e) E Do on Z x E.

Assume also that there exists a real-valued function t(e) such that
(1) t(e) E I fore E E,
(2) lim t(e) = to,
t-w
(3) hm 0(t(e), e) = (to)
Then:-to
Then,
lim i/i(t, t)
it to
uniformly on the interval Z.

Proof.
We shall prove this theorem in two steps. In Step 1, we prove Theorem 11-1-2
assuming that there exists an open neighborhood U of Co such that (t, tb(t, e)) E 0
on Z x (E n U), where 0 is the open set in the (t, y)-space which was determined
by Lemma II-1-1. The existence of such an open neighborhood U of to is proved in
Step 2.
Step 1. We can assume without any loss of generality that U is contained in the
neighborhood Uo of to which was determined by Lemma 11-1-1. Furthermore, using
condition (3) of Theorem 11-1-2, we choose U in such a way that It (t((),e)I is
bounded on U.
30 II. DEPENDENCE ON DATA
Since
(II.1.2) tjJ(t, E) = tp(t(E), E) + / t As, j(s, e), E)ds on Z x £,

JJJt(E)
we obtain
I
-
I!(t(E),E)I+MIt-t(E)I onIx(£fU) and Ii(t,E) AT, E)I
MI t - rI if (t, c) and (r, E) E Z x (£ n U), where M is the positive constant which
was determined by Lemma II-1-1. Therefore, the set F _ E) : E E £ n U} is
bounded and equicontinuous on the interval Z.
Let us derive a contradiction from the assumption that ,JJ(t, e) does not con-
verge to fi(t) uniformly on the interval I as e -+ co. This assumption means that
max I does not tend to zero as e --+ co; i.e., there exists a positive
number p and a sequence (E. : j = 1,2,. .. } such that
Ej E£nU and lim E; = Eo,
(II.1.3)
m Zx 1 (t,Ej) - v(t)I > p.
Here, use was made of the assumption that the topology of the E-space is separable.
Observe that the sequence { (-, j) : j = 1, 2.... } is a subset of Y. Hence,
this sequence is bounded and equicontinuous on Z. Therefore, b y Lemma 1-2-
3 (Arzela-Ascoli), there exists a subsequence v = 1, 2, ... } such that
j., -+ +oo as v - +oo and that tp(t, Ej converges uniformly on I as v -+ +oo.
Set l(t) = lim r%(t, on Z. Since (11.1.2) holds for c = Ef for all v, we obtain
J(t) = $(to) + f f(s,rl'(s),EO)ds on I.

to
Hence, f (t, J(t), Eo) on I and 1J(to) = ¢(to). Now, condition (iii) of As-
sumption 1 implies that j(t) = fi(t) identically on the interval Z. This, in turn,
implies that lim max I (t, E j ) - $(t) I = 0. This contradicts (1I.1.3).
v-+oo LET
Step 2. We shall prove that there exists an open neighborhood U of co such that
UClloand (t,1i(t,E))EAon Ix(EnU).
Choose no > 0 so that the rectangular region
Ro(r) = {(t.: It - rl < ao, (y - (r)I <-Mao}
is contained in A for every r E I (cf. Figure 1).
FIGURE 1.
1. CONTINUITY WRT INITITIAL DATA AND PARAMETERS 31
For every positive number a less than ao, set

R((T, ti7); a) _ {(t, yl: It-.I < a, Il! - i,-I «sa).
Then, ?Z((T, f ); a) C Ro(r) if IT - T 15 ao - a, I i - (r)I < M(ao - a). Fixing a
positive number a less than ao, choose TI, T2,. .. , TN so that
s1 = T1 < 72 < ... < TN = s2,
to = rjo for some jo,
0 < T2+l-T2 < a (j=1,2,...,N-1).
Since lira t(e) = to and lim e) = (to), there exists a neighborhood U1 of
co such that It(e) - tol < ao - a and I>%,(t(e), e) - 4(to)I < M(ao - a) fore E E nU1.
Hence, R((t(e), t%i(t((), e)); a) C Ro(to) C 0i_(t(e),
fore E E n U1. Now, using Lemma I-1-
1, we can verify that (t, fl(t, e)) E R((t(e), e)); a) C i for It - t(e)I < a and
e E EnU1. Also, since lim t(e) = to = rjo, we have It-t(e)I < a for rjo_1 < t < rjo+i
t- !o
and e E E n U1, if the neighborhood U1 of co is sufficiently small.
Using the same argument as in Step 1 for t (t, e) on the interval T,._ j 5 t < T3.+1
and e E E n U1, we obtain lim t(t, e) = fi(t) uniformly on rjo_ 1 < t < 7-,,.+,. In
.1 to
particular,
11m1L(T)o_1ie) _ 0(r)o-1) and lim V(T,o+1, e) = d(rjo+1)

f to (-to
Using the same method in the neighborhoods of and (rio+1,
4(Tjo+i )), we obtain lim
e-co1G(t, e) = Q(t) uniformly on rjo-2 - t<
- rjo and r <
t < rjo+2. Hence, (5p
lim ti(t, e) = q(t) uniformly on To-2 < t <- Tjo+2. Contin-
uing this process, we find an open neighborhood U of co such that U C Uo and
(t,>G(t, e)) E i on I x (E nU) in a finite number of steps. 0
Remark 11-1-3. If f (t, y, () is independent of e, Theorem 11-1-2 yields continuity
of solutions of the initial-value problem f = f (t, y-), 9(to) = 4 with respect to
the initial data (to, Q.
In Theorem II-1-2, it was assumed that (t, y(t, e), e) E Do on Z x E (cf. (II.1.1)).
With regard to this assumption, we can improve Theorem 11-1-2 as follows.
Theorem II-1-4. Assume that Assumption 1 is satisfied. Let r be an open neigh-
borhood of cc). Assume further that there exist functions t(e) and c(e) defined on E
such that
(i) t(e) E Zo and (t(e), c 1c), e) E Do fore E E,
(ii) llm t(e) = to and lim dde) = 0(to).
`-'a cco
Then, we can choose a neighborhood U of eo sufficiently small so that the initial-
value problem
f (t,
dt - y, e), e c)
has on the region lo x (E f1 Ll) a solution y = J(t, e) such that (t, 111(t, E), E) E Do.
Theorem II-1-4 can be proved by using an argument similar to Step 2 of the proof
of Theorem 11-1-2 and the local existence theorem (cf. Theorem 1-2-5). Details are
left to the reader as an exercise.
11-2. Differentiability
In this section, we explain the differentiability of solutions of an initial-value
problem with respect to the initial data under the following assumption.
Assumption 2. For the initial-value problem (P), assume that
(i) f (t, y-) and Lf' (t, y-) (j = 1, 2, ... , n) are continuous in (t, y-) on an open set
A in the (t, yj-space,
(ii) Q'(t) = f (t, and (t, ¢(t)) E ., on an interval To = {t : tl < t < t2}.
Using Lemma 1-1-6, Theorems 1-1-4, 11-1-2, II-1-4, and Remark 11-1-3, we can
show that there exists a positive number p such that the initial-value problem
(11.2.1) fdty = Y-) y(r) = rl
has a unique solution y' = O(t, r, q') on the interval lo if r E Zo and fr) - ¢(r)j <
p. The solution y = m(t, r, ri") is continuous in (t, r, q-) and satisfies the condition
(t. (t, r, E 0 if t E Zo, r E lo, and (r)I < p. The solution t(t, r, III
satisfies the integral equation
m(t, r, q + f t A s , (s, r, gi))ds.
If ¢(t, r, is differentiable with respect to (r, q-), then
- 1('r, q) +
e) +
8f
f t (s,( s, r, ))
(s, $(s, r, ij)) 80(s, 7,1-7) ds,

ds,
I" &R1
where q, is the j-th entry of i), eJ is the vector in IR" with entries 0 except for I at
the j-th entry, and Lf is the n x n matrix whose j-th column is . From this
0Yj
speculation, we obtain the following result.
2. DIFFERENTIABILITY 33
Theorem 11-2-1. Partial derivatives

continuous in (t, r, i)) in the domain
O-r
and
ft(j = 1, 2,... , n) exist and are
(11.2.2) f) : t E lo, t1 < r < t2, Ii - 4(r) I < p}.
Furthermore,
8r
respectively 0is the unique solution of the initial-value
BRA
problem
=
(t,(t, r, n)) z,
(11.2.3)
1AT) = - AT, 4) ( respectively eej).
Proof.
We prove the existence and continuity of by showing that they are the unique
solution of initial-value problem (11.2.3) with the initial-value z-(-r) f(7-, r'). The
existence and continuity of
L can be proved similarly.
d7),
Let +;1(t, r, ij) be the unique solution of (11.2.3) with the initial condition z(r)
- f (r, r)). From Theorems 1-3-5 and 11-1-2 and Remark 11-1-3, it follows that
1i(t, r, n) exists and is continuous in (t, T,i) on the closure D of domain (11.2.2).
Furthermore,
(I1 .2.4) i (t,r,,) _ -f(r,,) + JtOf (s, 0(s,r,rl)) 1'(s,r,q)ds
on '75. To show that 8 r, rl'), we first derive

r
(t,r+h,ff) = ri + f t f(s,G(s,r+h,i))ds,
- rh
f
(11.2.5)
d(t, r, i) = n + f (s, (s, r, i))ds,
r
from (11.2.1) on D assuming that h 54 0 is sufficiently small. Using (11.2.4) and
(H.2-5), we compute
(11.2.6) h) = h [(t, r + h, r1) - 4(t, r, 17)] - +G(t, r,
as follows.
Observation 1. For a fixed (t, r, fl) E D, we have
T j) = I rt I
l`
f (s, $(s, r + h, AS- J(s, 7,70)] ds
- 17,+h
f ( s, $6, r + h,
if h 96 0 is sufficiently small.
Observation 2. Using an idea similar to the proof of Lemma 1-1-6, for a sufficiently
small h, we obtain
f(s,j(s,r+h,n)) -
1 L(s,0 (s,r+h,fl)+(1 -9)¢(s,r,n1)dOl (t,r,rr))
[10
f o r s E 7 o and a fixed (r, i f ) such that (s, r, n E D. Note that (s, O (s, r + h, r') +
(1- 0)4(s, r, n ) ) E A for s E 10 and a f i x e d (r, n) such that (s, r, n') E D, if h 0 0
is sufficiently small.
Finally, applying Observations I and 2 to (11.2.4), (11.2.5), and (I1.2.6), we obtain
g(t, r, h)
1 rr+h
_- h rr
Jl f(r,qj
r
+JtIJ1
(11.2.7)
(s, r + h, J-7) + (1- 0)0(s, r, f )dO

'If -
-p.:+
i9f
Observation 3. The first and the third terms of the right-hand side of (11.2.7)
8
tend to zero as h - 0. Also, since 4r, r, J-7) = n and (s, d(s, r, rte} is bounded
09
on D, there exist a positive constant L and a non-negative function K(h) such that
limh_o K(h) = 0 and that
1g'(t,r,if,h){ <K(h) + LI
I J1g(s,r,il,h)Ids
t for tEZo.
Now, using Lemma 1-1-5, we obtain the estimate
19(t, r, n, h)l 5 K(h) exp[LIt - rIJ on To.

EXERCISES II 35
Thus, we conclude that g(t, r, il, h) - 0 ash -+ 0, i.e., (t, T, t) = rfi(t, r,'.
Upon applying Theorem 11-2-1 to the initial-value problem
f (t, du
d= y, u), = 0, y{ r) = i1, u(r) = e,
we can prove the following theorem.

a
Theorem 11-2-2. Let a be a real variable. Assume that At, y, e), (t, y, e) (,j =
1, 2, ... , n), and 8e (t, y', e) are continuous on an open set Do in the (t, y, e)-space.
Let y = y(t, T, i), e) be the unique solution of the initial-value problem
dy
dt = f (t, y, e),
y(r) _ *!
Then, there exists an open set 11 in the (t, T, 77, e)-space such that the function
00 00
(t, T, ii, e) and are continuous on fl. Furthermore, i = (t, r, 4J, e) is the
unique solution of the initial-value problem
di Of
Z(r) = 0.
dt = zi (t, (t, r, 1/, e), e)l + 8E (t, 4(t, r, n, E). E),
EXERCISES II
II-1. Given an interval I = {x : a < x < b}, show that if f ei is sufficiently small,
the initial-value problem
dy
e cos(x(y - x)), y(a) = a + e
dx
has the unique solution y = O(x, e) on Z. Show also that Ui m O(x, e) = x uniformly
on Z.
Hint. The function cos(x(y - x)) is continuously differentiable on the entire (x, y)-
plane. Furthermore, I oos(x(y - x))I < 1. This shows the existence of the unique
solution. Note that y = x is the unique solution in case e = 0. Complete the proof
by using Theorem 11-1-2.
11-2. Let y = 4 1(x, A) and y =¢2(x,.1) be tvm solutions of the differential equation
2 + A(1 + x2)y = 0
determined respectively by the initial conditions
01(O,A) = 1, A) = 0 and 02(0, A) = 0, a (0, A) = 1.

as
(0,
Show that
(i) 01 and 02 are analytic in (x, A) everywhere in the complex (x, A)-space (i.e.,
in C2),
(ii) aA (1, A0) # 0, if 02(1, 1\0) = 0.
Hint. Theorem II-2-2 implies that z (x, A) is the solution of the initial-value
problem
Z + A(1 + x2)z = -(1 + x2}02(x, A), z(0) = 0, z'(0) = 0.
The method of variation of parameters (cf. (IV.7.9) in Remark IV-7-2 of Chapter

IV; also see, for example, [Cod, pp. 67-68, 122-123] or [Rab, pp. 241-2441) yields
8 (x, A)
01(X"\)
J0 o
Hence, we obtain
8 (1,A0) = 01\1,Ao)fo 020,\0)2(1+t2)dC 0.
Note that 01 and 02 are linearly independent and hence 01(1,Ao) 0.

11-3. Show that if IEI is small, the following boundary-value problem has the unique
solution which is continuous in e:
2 = Esin (100- y(-1) = 0, y(1) = 0.

y2
Hint. Determine O(x, c, E) by the initial-value problem

d2y x
dZ2
= E 5111 ( 100 y2) y(-1) = 0, y'(-l) = c.
Then, the given boundary-value problem is reduced to the equation
(E) 0(1,c,e) = 0.
First of all, 0(1, c, c) is analytic for small itI and #c(. Also, 0(1,c,0) = 2c and, hence,
80(1, c, 0)
= 2 # 0. Therefore, equation (E) has the unique solution c = c(E) such
ac
that c(0) = 0. Furthermore, c(E) is analytic for small lei.
EXERCISES II 37
11-4. Let Ax, y-) be an RI-valued function whose entries are continuously differ-
entiable with respect to (x, yam) in a domain D C IIPn+1 Also, let y = (x, 171 be a
solution of the system fy = f (x, y) such that p(0) and that (x, 5(x, rtj) E D
for 0 < x < 1 and it E Ao, where Do is a domain in R. Note that we must have
(0, rl E D for rj E Ao. Set A = {(1, 1 : if E Do}. Show that
(a) Al is open in lR',
(b) the mapping 4(1,r) : Ao - Al is one-to-one, onto, and differentiable with
respect to i in Do,
(c) if we denote the inverse of the mapping ¢(1,i) by AI -. Ao, then
is also differentiable with respect to S.
11-5. For the differential equation
(1 + x2) LY -_ 2xy = 2.sy2,
find
(i) the unique solution y = O(x, e, ri) satisfying the initial-condition y(£) = q;
(ii) the partial derivative 84(x, , n) ,
of
am(x, £, r7)
(iii) the artial derivative
8 t , n) at x =
(iv) 190(x,
(v) a0(x, 4, rl) at x =

8¢(
Show also that z = 00(. '77) and z = ,17) both satisfy the differential
equation:
8 Bin
(I+ x2) dz
dz
- 2xz = 4xO(x, {, q)z .
II-6. Assume that f (t, x1, x2i ... , xn) is a real-valued, continuous, and continu-
ously differentiable function of (t, xl,... , x,) on an open set A in the (t, x1.... , xn)-
space. Assume also that o(t) is a real-valued solution of the n-th order differential
,0On-1)(t))
equation x(") = f(t,x,x',... ,x(n-1)) and (t,Oo(t),e0(t),... E A on an
interval Zo = It : t1 < t < t2}. Show that there exists a positive number p such
that
(i) the initial-value problem
x(n) f (t, x, x , ... , x(n-1) ), x(r) = SI, x (r) = 52, ... , x(n-1) (r) = n
=
has a unique solution x = {n) on the interval 20 if r E Zo and
t 3-O0-1)(r)I<p (j=1,2,...,n);
isfies the condition t t 'r, 6, )
(ii) the solution x = qi(t, r, ti, ... , {n) is continuous in (t, r, t1, ... , n) and sat-
W) e(t r O("-1)(t r
(j=1,2,...,n);
(iii) the partial derivative exists and is continuous in (t, r, Ft,

t8tn
n) on the domain D = {(t, r, 1, ... , {n) : t E To, t1 < T < t2,
-')(T)I < p (1 = 1,2,...,n)};
(iv) u = Lo (t, r, is the unique solution of the initial-value problem
-1
dnU
dtn `J= axff
- rn ` tt
(t,7,6,... ,Sn)) dtj_1
,u(n-2)(7) u(n_1) =
U(T) = -6, u'(T) ,Ln).
11-7. Assume that f (t, X1, x2) is a real-valued, continuous, and continuously dif-
ferentiable function of (t, xl, x2) on an open set A in the (t, x1, x2}space. Assume
also that O0(t) is a real-valued solution of the second-order differential equation
x" = f (t, x, x') and (t, Oo(t), 00'(t)) E A on the interval Zo = (t : 0 < t < 1).
Set ¢0(0) = a and 0o(0) = b. Denote by t(t,/3) the unique solution of the initial
value-problem x" = f (t, x, x'), x(0) = a, x'(0) = ,B, where lb - 81 is sufficiently
small. Show that (1,b) > 0 if of 0 for t C-
013
11-8. Let g(t) be a real-valued and continuous function oft on the interval 0 < t <
1. Also, let A be a real parameter.
(1) Show that, if Q(t, A) is a real-valued solution of the boundary-value problem
Ez + (g(t) + A)u = 0, u(0) = 0, u(1) = 0,
and if 8 &2 (t, A) is continuous on the region A = {(t, A) : 0 < t < 1, a <
A < b}, then d(t, A) is identically equal to zero on A, where a and b are real
numbers.
(2) Does the same conclusion hold if 0(t,,\) is merely continuous on A?
11-9. Let a(x, y) and b(x, y) be two continuously differentiable functions of two
variables (x, y) in a domain Do = {(x, y) : IxI < a:, IyI < Q} and let F(x, y, z)
be a continuously differentiable function of three variables (x, y, z) in a domain
Do = {(x, y, z) : Ixl < a:, IyI < 3, Izl < 7}. Also, let x = f (t, i7), y = 9(t, ij), and
z = h(t,17) be the unique solution of the initial-value problem
= a(x,y), L = b(x,y), = F(x,y,z),

dt
x(0) = 0, y(0) = n, z(0) = c(+1).
where c(q) is a differentiable function of 11 in the domain 14 = (q: In1 < p). Assume
that (t, rl) = (d(x, y), O(x, y)) is the inverse of the relation (x, y) = (f (t,17), g(t, rt)),
where we assume that O(z, y) and i,1(x, y) are continuously differentiable with re-
spect to (x, y) in a domain Al = {(x, y) : IxI < r, IyI < p}. Set H(x,y) =
h(p(x, y), tp(x, y)). Show that the function H(x, y) satisfies the partial differential
equation
a(x, y) + b(x, y) M = F(x, y, H)
8 y) = c(y).
and the initial-condition H(0,
EXERCISES II 39
Hint. Differentiate H(x,y) with respect to (x,y).

11-10. Let F(x, y, z, p, q) be a twice continuously differentiable function of
(x,y,u,p,q) for (x,y,u,p,q) E 1R5. Also, let
x = x(t, s), y = y(t, s), z = z(t, s), p = P(t, s), q = q(t, s)
be the solution of the following system:
dx OF
- (x+ y, z, p, q),
dy _ OF
flq(T, y, z, p, q),
dt
dz OF OF
dt = pij (x,y,z,P,q) + gaq(x,y,z,P,q),
dp OF OF
dt 8x (T, y, z, P, q) - P 8z (x, y, z, p, q),
dq OF OF
(x, y, z, P, q) - q az (x, y, z. P, q)
dt 8y
satisfying the initial condition
x(O,s) = xo(s), y(O,s) = yo(s), z(O,s) = zo(s),
p(O,s) = Po(s), q(O,s) = qo(s),
where x0 (s), yo(s), zo(s), po(s), and qo(s) are differentiable functions of s on R such
that
dzo(s)=Po(s)- (s)+go(s)dyo(s)
F(xo(s),yo(s),zo(s),Po(s),go(s))=0, ds ds ds
on R. Show that
F(x(t, s), y(t, s), z(t, s), p(t, s), q(t, s)) = 0,
at (t, s) = At, s) 5 (t, s) + q(t, s) (t, s),
flz(t,s)
fls = P(t,s)L(t,s)
as
+ q(t,s)ay(t,8)
as
as long as the solution (x, y, z, p, q) exists.

Comment. This is a traditional way of solving the partial differential equation
F(x, y, z, p, q) = 0, where p = 8 and q = For more details, see (Har2, pp.
131-143].
40 H. DEPENDENCE ON DATA
II-11. Let H(t, x, y, p, q) be a twice continuously differentiable function of

(t, x, y, p, q) in RI. Show that we can solve the partial differential equation
Oz Z 8z\ =0
+ H t, x, y,-,
by using the system of ordinary differential equations

dx 8H dy 8H
dt = 8p' t= q
,
dp 8H (LL dq _ _ 8H
dt 8x ' dt W,
dz 8H 8H du OH
= u + p-p + q8q, _
= --6T.
dt
CHAPTER III
NONUNIQUENESS
We consider, in this chapter, an initial-value problem
dy
dt = f(t,y1,
(P)
without assuming the uniqueness of solutions. Some examples of nonuniqueness are

given in §III-1. Topological properties of a set covered by solution curves of problem
(P) are explained in §§III-2 and 111-3. The main result is the Kneser theorem
(Theorem III-2-4, cf. [Kn] ). In §1II-4, we explain maximal and minimal solutions
and their continuity with respect to data. In §§III-5 and 111-6, using differential
inequalities, we derive a comparison theorem to estimate solutions of (P) and also
some sufficient conditions for the uniqueness of solutions of (P). An application of
the Kneser theorem to a second-order nonlinear boundary-value problem will be
given in Chapter X (cf. §X-1).
111-1. Examples
In this section, four examples are given to illustrate the nonuniqueness of so-
lutions of initial-value problems. As already known, problem (P) has the unique
solution if f'(t, y) satisfies a Lipschitz condition (cf. Theorem I-1-4). Therefore, in
order to create nonuniqueness, f (t, y-) must be chosen so that the Lipschitz condi-
tion is not satisfied.
Example III-1-1. The initial-value problem
dy
(III.1.1) = yt/3 y(to) = 0
dt
has at least three solutions
(S.1.1) y(t) = 0 (-oo < t < +oc),
3/2
(S.1.2) y(t) _ [3 t - to) ]I
, t > to,
0, t < to,
and
3/2
[3(t - to),
2
(S.1.3) y(t) t > to,
0, t < to
41
42 111. NONUNIQUENESS
(cf. Figure 1). Actually, the region bounded by two solution curves (S.1.2) and
(S.1.3) is covered by solution curves of problem (I11.1.1). Note that, in this case,
solution (S.1.1) is the unique solution of problem (P) for t < to. Solutions are not
unique only fort ? to.
Example 111-1-2. Consider a curve defined by
(111.1.2) y = sin t (-oo < t < +oo)
and translate (111.1.2) along a straight line of slope 1. In other words, consider a
family of curves
(111.1.3) y = sin(t - c) + c,
where c is a real parameter. By eliminating c from the relations
c os(t - c), y - t = sin(t - c) - (t - c),

dt =
we can derive the differential equation for family (111.1.3). In fact, since sin u - u
is strictly decreasing, the relation v = sinu - u can be solved with respect to u
to obtain u = G(v) - v, where G(v) is continuous and periodic of period 27r in v,
G(2n7r) = 0 for every integer n, and G(v) is differentiable except at v = 2n7r for
every integer n. The differential equation for family (111.1.3) is given by
dy
(II1.1.4)
dt = 1OS[G(y-t)-(y-t)1.
Since G(2n7r) = 0 and cos(-2n7r) = 1 for every integer n, differential equation
(111. 1.4) has singular solutions y = t + 2mr, where n is an arbitrary integer. These
lines are envelopes of family (111.1.3) (cf. Figure 2).
FIGURE 1. FIGURE 2.
Example 111-1-3. The initial-value problem
(111.1.5) y(to) = 0
dt =
has at least two solutions
(S.3.1) y(t) = 0 (-oo < t < +oo),

1. EXAMPLES 43
and
t,0)2,
4 (t - t > to ,
(S.3.2) y(t)
- 4(t - to)2, t < to
(cf. Figure 3). The region bounded by two solution curves (S.3.1) and (S.3.2) is
covered by solution curves of problem (111.1.5).
FIGURE 3.
Consider the following two perturbations of problem
dy
= + E, y(to) = 0
dy _ y2
y2 + C2 lyl y(to) = 0,
where a is a real positive parameter. Each of these two differential equations satisfies
the Lipschitz condition. In particular, the unique solution of problem (111. 1.6) is
given by
J4(t-to+2VI -E)2 - E, t > to

(S.3.3) y(t) =
-4(t-to-2f)2 + E, t < to
(cf. Figure 4). On the other hand, (S.3.1) is the unique solution of problem (111.1.7).
Figure 5 shows shapes of solution curves of differential equation (111.1.7). Note that
nontrivial solution of (111.1.7) is an increasing function of t, but it does not reach
y = 0 due to the uniqueness.
44 III. NONUNIQUENESS
FIGURE 4. FIGURE 5.
Generally speaking, starting from a differential equation which does not satisfy
any uniqueness condition, we can create two drastically different families of curves
by utilizing two different smooth perturbations. In other words, a differential equa-
tion without uniqueness condition can be regarded as a branch point in the space
of differential equations (cf. [KS]).
Example 111-1-4. The general solution of the differential equation
2
(1II.1.8)
d) +y2=1
is given by
(111.1.9) y = sin(t + c),
where c is a real arbitrary constant. Also, y = 1 and y = -1 are two singular

solutions. Two solution curves (111.1.9) with two different values of c intersect each
other. Hence, the uniqueness of solutions is violated (cf. Figure 6).
This phenomenon may be explained by observing that (111. 1.8) actually consists
of two differential equations:
(III.1.10) dy = 1 - yz and
dy=- l-y2.
dt dt
Each of these two differential equations satisfies the Lipschitz condition for lyI <
1. Figures 6-A and 6-B show solution curves of these two differential equations,
respectively.
y=I
y=-I
FIGURE 6. FIGURE 6-A. FIGURE 6-B.
Observe that each of these two pictures gives only a partial information of the
complete picture (Figure 6).
2. THE KNESER THEOREM 45
We can regard differential equation (111.1.8) as a differential equation

dy
(111.1.11)
dt
=w
on the circle
(111.1.12) w2 + y2 = 1.
If circle (111.1.12) is parameterized as y = sinu, w = cosu, differential equation
(III.1.11) becomes
(III.1.13) d=1 or cos u = 0 .

Solution curves u = t of (111.1.13) can be regarded as a curve on the cylinder
{(t, y, w) : y =sin u, w = cos u, -oc < u < +oo (mod 2w), -oo < t < +oo}
(cf. Figure 7). Figure 6 is the projection of this curve onto (t,y)-plane.
FIGURE 7.
In a case such as this example, a differential equation on a manifold would give
a better explanation. To study a differential equation on a manifold, we generally
use a covering of the manifold by open sets. We first study the differential equation
on each open set (locally). Putting those local informations together, we obtain a
global result. Each of Figures 6-A and 6-B is a local picture. If these two pictures
are put together, the complete picture (Figure 6) is obtained.
111-2. The Kneser theorem

We consider a differential equation
(III.2.1) dt = f(t,y-)
under the assumption that the R"-valued function f is continuous and bounded on
a region
(111.2.2) S2 = {(t, y-) : a:5 t < b , Iy1 < +oo }.
Under this assumption, every solution of differential equation (III.2.1) exists on the
interval Zfl = {t : a < t < b} if (to, y"(to)) E f2 for some to E Zo (cf. Theorem 1-3-2
and Corollary I-3-4). The main concern in this section is to investigate topologi-
cal properties of a set which is covered by solution curves of differential equation
(111.2.1).
2. THE KNESER THEOREM 47
Theorem 111-2-4 ((Kn]). If A is compact and connected, then SS(A) is also

compact and connected for every c E To.
Proof.
The compactness of SS(A) was already explained. So, we prove the connectedness
only.
Case 1. Suppose that A consists of a point (r, t), where we assume without any
loss of generality that r < c. A contradiction will be derived from the assumption
that there exist two nonempty compact sets F1 and F2 such that
(111.2.3) SS(A)=F1uF2, F1nF2=0.

If t'1 E F1 and 2 E F2, there exist two solutions 1 and 4'2 of (111.2. 1) such that
01(r) _ , 01(c) _ 1, and 42(r) _ ., 02(c) _ 6, (cf. Figure 9).
Set
1(r+11) for 05µ5c-r,
h(µ) _
{ &T + JJUJ) for -,r)
Let { fk(t, y-) : k = 1, 2,... } be a sequence of R"-valued functions such that

(a) the functions fk (k = 1, 2, ...) are continuously differentiable on 0,
(b) I fk(t, y1I < M for (t, y) E f2, where M is a positive number independent of
(t, yl and k,
(c) ra+A = f'uniformly on each compact set in Q
k El oo
(cf. Lemma I-2-4). For each k, let &(t,µ) be the unique solution of the initial-
value problem = fk(t, y-), y(-r+ Iµ4) = h(µ). The solution 15k(t, µ) is continuous
for t E Zo and iµl < c - r. It is easy to show that the family k=
1, 2,... ; jµj 5 c - r} is bounded and equicontinuous on the interval Io.
Note that the functions Z(rk(c, µ) (k = 1, 2, ...) are continuous in µ for 1µI <
c - r and that r'k(c,c - r) = ¢'1(c) E F1 and t k(c,-(c - r)) = &c) E F2.
Let d be the distance between two compact sets F1 and F2. Since &(c,µ) is
continuous in p, there exists, for each k, a real number µk such that IµkI < c-r and
d.
distance(!& (c, µk), F1) = Since the family ilk) : k = 1, 2, ... } is bounded
and equicontinuous on the interval Io, there exists a subsequence
j = 1,2.... } such that (i) lim k, _ +oo, (ii) lim µk, = po exists, and (iii)
i--+00 )-.+oo
urn 1Pk, (t, µk1) = 0(t) exists uniformly on To. It is easy to show that
s-+oo
fi(t) = K(AO) + ft f (s, ¢(s))ds,

+ {µol
lµo1 < c - r, distance(¢(c), F1)

=2
Hence, (c) E $.(A) but * (c) 4 F1 U F2. This is a contradiction (cf. Figure 10).
!=c t=c
FIGURE 9. FIGURE 10.

Case 2 (general case). Assume (111.2.3) as in Case 1, set AI = AnR((c) x.Fl) and
A2 = A n R({c} x F2), where {c} x F; = {(c, yl : y" E Fj } (j = 1, 2). Then, the
two sets A, (j = 1, 2) are compact and not empty. Note that A = AI U A2. Since
A is connected, we must have AI n A2 0 0. Choose a point (r, {) E AI n A2. Then,
Sc((T,6) = {Ss((T,S)) nF1}U{S,.((T,6) n12} andS,,((r, ))n.Fi 36 0 (j = 1,2).
This is a contradiction (cf. Case 1). 0
In order to apply the Kneser theorem (i.e., Theorem III-2-4), it is desirable to
remove the boundedness of f from the assumption. To obtain such a refinement
of Theorem 111-2-4, consider differential equation (111.2.1) under the following as-
sumptions.
Assumption 1. A set Ao is a compact and connected subset of the region Q such
that if ¢(t) is a solution of (111.2.1) satisfying
(111.2.4) (to, 0(to)) E Ao for some to E To,
then fi(t) exists on To.

As in Definition 111-2-1, denote by Ro the set of all points (t,yr) E Q such that
y = fi(t) for some solution 4 of differential equation (111.2.1) satisfying condition
(111.2.4). Also, set $S = (9: (c, y-) E Ro) for c E Io.
Assumption 2. The set Ro is bounded.
Now, we prove the following theorem.
Theorem 111-2-5. If a compact and connected subset Ao of 12 satisfies Assump-
tions 1 and 2, then the set Sc is also compact and connected for every c E I.
Proof.
Since Ro is bounded, there exists a positive number M such that
Ro c {(t,yj: tETo, Iyl SM) .

Set
f(t,yl if tETo, Iyj<2M,
9(,t 0= f ( t, if t E To, ly7 ? 2M.
j- y)
3. SOLUTION CURVES ON THE BOUNDARY OF R(A) 49
Then, g(t, g) is continuous and bounded on Q. Using the differential equation

g(t, y), define R(Ao) and SS(Ao) by Definition 111-2-1. Then, we must have
dt =
Ro = 1Z(Ao). Otherwise, there would exist a solution ¢o(t) of differential equation
(111.2.1) such that (to,&(to)) E Ao for to E Zo and (tt,do(tl)) V Ro for tl E -To-
This is a contradiction. The theorem follows from Theorem 111-2-4. 0
Remark 111-2-6. Any kind of shapes can be made by a wire. Therefore, the set
SS(A) needs not to be a convex set even if A is convex.
Using Theorem 111-2-4, we can prove the following theorem, which is a refinement
of Theorem II-1-4.
Theorem 111-2-7. Assume that the entries of an R"-valued function At' In are
continuous in (t, yam) in a domain Do E Rn+l.
Also, let fi(t) be a solution of system

(111.2.1) on an interval a < t < b. Assume that (t,¢(t)) E Do on the interval
a < t < b. Then, for any given positive number e, there exists another solution '(t)
of (111.2.1) such that
(i) (t, e,1(t)) E Do on the interval a < t < b,
(ii) JO(t) - v1'(t)l < e on the interval a < t < b,
(iii) a Y(r) # fi(r) for some r on the interval a < t < b.
Proof.
Using an idea similar to Step 2 of the proof of Theorem 11-1-2, the local existence
theorem (Theorem 1-2-5), and the connectedness of SS(A), we can complete the
proof of Theorem 111-2-7. In fact, subdivide the interval a < t < b (i.e., a = to <
tl < < tk = b) in such a way that [y(t) - ¢(t)1 < e for t, < t :5 tj+l if g(t)
satisfies (111.2. 1) and Ib"(t,) - i(t,)I < 6. Set A. = {EE E R" : l e - (tj)I < b} and
Bl = {y-(t1) : y(to) E Ao}, where g(t) denotes any solution of (111.2.1) with initial-
value y-(to). Since Bl is connected and contains ¢(t1), we must have Bl =
if Bl n AI contains only (t1). In such a case, the proof is finished.
If Bl n Al contains more than one point, then B1 n A, contains a connected set
Sl containing 0(t1) and more than one point. Set B2 = {9(t2) : g(ti) E Si). Since
B2 is connected and contains (t2), we must have B2 = if B2 n A2 contains
only (t2). In such a case, the proof is finished. In this way, the proof is completed
in a finite number of steps. 0
III-3. Solution curves on the boundary of 1Z(A)

We still consider a differential equation
(III.3.1) dt =
under the assumption that the entries of the R°-valued function fare continuous
and bounded on a region
(111.3.2) n = {(t, y-) : a < t < b , lyi < +oc}.
50 M. NONUNIQUENESS
As mentioned in §111-2, under this assumption, every solution of differential equa-

tion (111.3.1) exists on the interval To = {t : a < t < b} if (to,y-(to)) E ft for some
to E I. Define the sets 1(A) and SS(A) for a subset A of ) by Definition III-2-1.
The main concern of this section is to prove the existence of solution curves on the
boundary of R(A). We start with the following basic lemma.
Lemma 111-3-1. Suppose that (1) the set A consists of one point (c1, t) (i. e.,
A = {(c,, f)}), (ii) a < cl < co < c2 < b, and (iii) ij is on the boundary of Sc2(A).
Let B = Then, SI(B) contains at least a boundary point of S0(A).
Proof
A contradiction will be derived from the assumption that SI(B) does not contain
any boundary points of SS (A). Note that SI(B) n 4 (A) 54 0. Set S, = 4 (B) n
S,,. (A), S2 = {y : y' E Se. (B) and Ii 0 SS(A)}. Then, S., (B) = 81 U S2 and
S, n S2 = 0. It is known that S1 is a nonempty compact set. Also, $2 is closed
and bounded, since S,,, (5) is compact and does not contain any boundary points of
SS (A). If S2 is not empty, the set Sc.(B) contains a point in Sc,, (A) and another
point which does not belong to S,, (A). Then, S,.(8) contains a boundary point of
Sc. (A), since S,, (B) is connected. This is a contradiction. Therefore, if it is proved
that S2 is not empty, the proof of Lemma 111-3-1 will be completed (cf. Figure 11).
To prove that S2 is not empty, observe first that Sam({(a2,()}) nS.,,(A) = 0 if
0 SS,(A) (cf. Figure 12).

Let rj be on the boundary of $, (A). Then, there exists a sequence of points
{Sk V Sc, (A) : k = 1, 2, ... ) and a sequence (¢k : k = 1, 2.... ) of solutions of
(111.3.1) such that lira (k = '1, O&2) = (k, and mk(co) 0 SS(A). The family
k+oo
k = 1,2.... } is bounded and equicontinuous on the interval Zo. Therefore,
there exists a subsequence { Jk, j = 1, 2, ...) such that lirn k, = +oo and
j+oo
lim bk, (t) = Q(t) exists uniformly on Zo. The limit function is a solution of
++oo
(111.3.1) such that ¢(c2) = tj. Hence, d(co) E Sc,(B). Since S, (B) does not contain
any boundary points of S,(A), we must have ¢'(co) if $,,(A). This implies that
S2 is not empty. 0
The following theorem is the main result in this section which is due to M.
Hukuhara (Hukl( (see also (Huk2J and [HN31).
3. SOLUTION CURVES ON THE BOUNDARY OF R(A) 51
Theorem 111-3-2. Suppose that

(1) a<cl <c2 <b,
(2) is on the boundary of
Then, there exists a solution of differential equation (M.3.1) such that
(i) (Cl) = E, (C2) = 6,
(ii) is on the boundary of R({(ci,l;))) for cl < t < c2.
Proof.
For a subdivision A : CI = To < Tl < . < Tm-1 < Tm = c2 of the interval
cl < t < c2, there exists a solution ¢o of (111.3.1) such that
(a) O0(CI) _ ., 4fiG(C2)
(p) is on the boundary of R({(cl,O)}), where e = 1,2,... ,m - 1.
(Cf. Lemma 111-3-1 and Figure 13.)
FIGURE 13.
Choose a sequence {ok : k = 1, 2, ... } of subdivisions of the interval cl < t < c2
such that
Ok : C1 = Tk,O < Tk.1 < ... < Tk,2k-I < Tk,2" = C2,
f e
1 Tk,t = C1+2k(c2-cl), e=0,1,...,2k, k=1,2,....
Since Tk+1,21 = Tk,l, we have {Tk,t : e = 0, 1,2,... , 2k} C {Tkrl,e e

2k+1}. Set k =tt jok (k = 1,2,...). Then,
(ak) &(C1) = S, &(C2) = 1,
(Qk) (Tk,t,&(Tk,t)) is on the boundary of where e = 1,2,... ,
2k - 1.
The family {4k k = 1, 2,... } is bounded and equicontinuous on 10. Hence,
there exists a subsequence {&, : j = 1, 2,... } such that lim k, = +oo and that
j +00
lira ¢k, exists uniformly on 10. Then, is a solution of (111.3.1) such that
i-.+oo
(ao) (Cl) (C2)
(Qo) (Tk,t,. k(Tk,t)) is on the boundary of R({(c1i,)}), where f
2k - 1 and k = 1,2,....
Since the set {Tk,t : e = 1,2,... 2k - 1; k = 1,2.... } is dense on the interval
Cl < t < c2, the curve (t,$(t)) is on the boundary of for cl < t <
C2- 0
Remark 111-3-3. In general, the conclusion of Theorem 111-3-2 does not hold
on an interval larger than cl < t < c2.
111-4. Maximal and minimal solutions

Consider a scalar differential equation
(111.4.1) dt = f (t, y),
where t and y are real variables, and assume that f is real-valued, bounded, and
continuous on a region H = {(t,y) : a < t < b, -oo < y < +oo}. Choose c in
the interval I = (t : a < t < b}. Then, Sr ({ (c, r,) }) is compact and connected
for every r E I and every real number r) (cf. Theorem III-2-4). This implies
that there exist two numbers 41(r) and ¢2(r) such that ST({(c,17)}) = { y E
R : bl(r) < y < &2(r)}. Hence, R({(c,n)}) = {(t,y) E R2 : ¢1(t) < y <
02(t), a < t < b} (cf. Figure 14). The two boundary curves (t,4i(t)) and (t,02(t))
of R({(c,rl)}) are solution curves of differential equation (111.4.1) (cf. Theorem
111-3-2). Every solution ¢(t) of (111.4.1) such that 0(c) = n satisfies the inequalities
01(t) < ¢(t) < ¢2(t) on Z. The solution 02(t) (respectively 01(t)) is called the
maximal (respectively minimal) solution of the initial-value problem
dy
(111.4.2)
dt = f (t, y), y(c) = 77
on the interval Z. In this section, we explain the basic properties of the maximal
and minimal solutions. Before we define the maximal and minimal solutions more
precisely, let us make some observations.
FIGURE 14.
Observation 111-4-1. Assume that f (t, y) is real-valued and continuous on a

domain Din the (t,y)-plane. Set lo = {t : a < t < b}. Let 01(t) and ¢2(t) be two
solutions of differential equation (111.4.1) such that (t, ¢1(t)) E D and (t, 02(t)) E D
for t E Zo. Note that we do not assume boundedness of f on D. Set 0(t) =
max {01(t), 02(t)} for t E Zo. Then, Q(t) is also a solution of (111.4.1) on the
interval Zo.
4. MAXIMAL AND MINIMAL SOLUTIONS 53
Proof.
For a fixed to E To, we prove that
(111.4.3) lim do(t) - O(to) = f(to,¢(to)).

t-.to t - to
If ¢1(to) > -02(to), then there exists a positive number 6 such that 4(t) = 01(t) for
It - tol < J. Therefore, (111.4.3) holds. Similarly, (111.4.3) holds if 02(to) > &(to).
Hence, we consider only the case when '(to) _ .01(to) = 02(to). In this case,
0(t) - 0(to) = 0'(t) - 0'(to),
f o r each fixed t E To, we have where j = 1 or
t - to t- to
j = 2, depending on t. Hence, by the Mean Value Theorem on O,(t), there exists
r E To such that r --r to as t - to and 0(t) - ¢(to) = f (r, 0,(r)). Also, 03 (r) _
t - to
¢j(to) + f (a,-Oj (a)) (7 - to) for some a E To such that a to as r - to. Since f is
bounded on the two curves (t, 01(t)) and (t, 02(t)), (111.4.3) follows immediately. 0
Observation 111-4-2. Let f (t, y), D, and To be the same as in Observation
111-4-1. Consider a set F of solutions of differential equation (111.4.1) such that
(t, 0(t)) E D on To for every ¢ E F. Assuming that there exists a real number K
such that 0(t) < K on lo for every 0 E F, set 00 (t) = sup{Q(t) : ¢ E F} for t E To.
Assume also that (t, 00(t)) e D on To. Then, 4'o(t) is a solution of differential
equation (111.4.1) on To.
Proof.
As in Observation 111-4-1, we prove that
(111.4.4) lim 00(t) - Oo(to) = f (to, bo(to))

t-»to t - to
for each fixed to E To. Choose three positive numbers po, p, and S so that
(i) A = {(t, y) : t E To, Iy - 4o(t)I < po} C D,
(ii) we have (t, 4(t)) E A on the interval t - to I < 6, if d(t) is a solution of
I
(III.4.1) such that 0 < 0o(r) - 4i(r) < p for some r in the interval It - tol < b.
There exists a positive number Af such that I f (t, y) I < M on A.
Let us fix a point r on the interval It - tol < b. First, we prove the existence of
a solution ty(t; r) of (III.4.1) such that
(111.4.5) P(r; r) = 0o (r) and yJ(t; r) < 4o(t) for it - tol < b.
To do this, select a sequence {4'k : k = 1, 2,... } from the family F so that
lirao0thk(r) = 40(r). We may assume that (t, hk(t)) E A on It - tol < S for
Y+
k (cf. (ii) above). Then, the sequence {4'k : k = 1, 2, ... )is bounded and
equicontinuous on It - tot < J. Hence, we may assume that lim 4'k(t) = 0(t; r)
k-»+oo
exists uniformly on the interval It - tol < 6. It is easy to show that V'(t; r) is a
solution of (111.4.1) and that (111.4.5) is satisfied.
Set ty(t) = max{t/,(t; r), tJ'(t; to)} for Jt - tol < b. Then, v is a solution of (111.4.1)
such that (1) V)(r) = oo(r) and ty(to) = 00(to), (2) iP(t) < 4'o(t) for It - tol < b,
and (3) (t, v/ (t)) E A for It - tol < 6 (cf. Figure I5).
Y= #o(')
Y s V(1)
FIGURE 15.
O0(T) - ¢040) = t.ti(T) - 0(t0) = f(a,tp(a))
Using this solution tp(t), we obtain
T - to T- to
for some a such that Ja - tol < 6 and that or to as r -" to. Since 10(a) - tt'(td)J <
Mba - t01, (111.4.4) can be derived immediately. 0
Let us define the maximal (respectively minimal) solution of an initial-value
problem
{III.4.6 ) dt = f (t, ii), y(r)
where f is real-valued and continuous on a domain V in the (t, y)-plane and the
initial point (T, 1:) is fixed in D.
Definition 111-4-3. A solution tp(t) of problem (111.4.6-() is called the maximal
(respectively minimal) solution of problem (111.4.6-() on an interval I = it : T <
t < r' } if
(a) tp(t) is defined on I and (t, {1(t)) E D on Z,
(b) if 4(t) is a solution of problem (II14.6-{) on a subinterval r < t r" of Z,
then
0(t) < y(t) (respectively ¢(t) > P(t)) on r < t < r".
The following two theorems are stated in terms of maximal solutions. Similar
results can be stated also in terms of minimal solutions. Such details are left to
the reader as an exercise. In the first of the two theorems, we consider another
(III.4.6-n) !LY
dt
= f(t,y), y(r) = r)
together with problem (111-4.6-{)-

Theorem III-4-4. Suppose that the maximal solution tpt of Problem (IIL4.6-{)
exists on an interval Z = {t : r < t < r'}. Then, there exists a positive number So
such that the maximal solution 0,1 of problem (III 4.6-r1) exists on I for t; < rl <
t + b0. Fjrthermore, tpf(t) < ip,,(t) on I for £ < 17< + 60 and limipR(t) = tb,(t)
uniformly on Z.
Proof
Set 0(p) = {(t, y) : t E Z, yt(t) < y < ot(t) + p}. For a sufficiently small
positive number p, we have A(p) C 1), and, hence, if (t, y)I is bounded on A(p) for
a sufficiently small positive number p.
First, we prove that for a given p > 0, there exists a positive number 6 such that,
if f < q < { + b, every solution 0(t) of problem (111.4.6-1?) defined on a subinterval
r < t < r" of I satisfies
(III.4.7) 0(t) < ,'(t)±p for r < t <7".
Otherwise, there exist two sequences {qk : k = 1 , 2, ... ) and {rk : k = 1, 2, ... } of
real numbers such that (1) qk > {, lim qk = t;, and r < rk < r', (2) 0&-(-r) = nk
k -+oo
and Ok(Tk) = v((rk) + p, and (3) Ok(t) < vf(t) + p for r < t < rk. Furthermore,
there exists a real number r(p) such that rk > r(p) > r (cf. Figure 16).
Y=W4+P
- 1 I
Y Wt
I
I i
t=r t= TA 1=',
FIGURE 16.
Set
max(0k(t), '' (t)), T<t<Tk,
f (t)+p
11)k (t )
,
Tk St ST.
It is easy to show that the sequence {0k W : k = 1, 2, ... } is bounded and equicon-
tinuous on the interval Z. Hence, we can assume without any loss of generality
that
(1) tim Tk = rp exists,
k-.+oo
(2) tim 10k(t) = 0(t) exists uniformly on Z.
Note that
(3) d(r) _ and O(ro) _ 'af(ro) + p,
(4) ro?r(p)>r.
It is also easy to show that 0(t) is a solution of (III.4.6-t) on the subinterval
r < t < ro of Z. Since 104 is the maximal solution of (III.4.6-1:) on 7, we must
have 0(t) < af(t) on the interval r < t < To. This is a contradiction, since
0(ro) = i (ro) + p. Thus, (111.4.7) holds.
Now, for a given positive number p, there exists another positive number 6(p)
such that (111.4.7) holds for any solution ¢(t) of problem (III.4.6-17) if < ' <
t; + 6(p). Hence, using max(4(t), 0((t)), we can extend 6(t) on the interval Z in
such a way that
(111.4.8) ?Gf(t) 5 y(t) < p on Z.
Let F be the set of all solutions 6(t) of (III.4.6-71) which satisfy condition (111.4.8).
Set t[i,r(t) = sup{¢(t) : 0 E F j for t E I. Then, ty,, is the maximal solution of
(III.4.6-r7) on I. Furthermore, 104(t) < 1,1(t) < ot(t) + p if t: < rt < t + 6(p).
Letting p - 0, we complete the proof of the theorem.
Assuming again that f is continuous on a domain D in the (t, y)-plane, consider
initial-value problem (III.4.6-{) together with another initial-value problem
(III.4.9-E) dt = f (t, y) + E, y(r)
where (t, 1;) E V and f is a positive number. We prove the following theorem.
Theorem III-4-5. If the maximal solution tt'(t) of problem (111.4.6-,) on the in-
terval I = {t : r < t < r'} exists, then, for any positive number p, there exists
another positive number e(p) such that for 0 < E < E(p), every solution d(t) of
(11!.4.9-E) exists on I and tp(t) < d(t) < y',(t) + p on the interval I. In particu-
lar, for every sufficiently small positive number f, then exists the maximal solution
tf=, (t) of problem(111.4.9-E) on I and lim tL, = tV uniformly on I.
Proof
The maximal solution tI'(t) satisfies the condition (t,>l,(t)) E D on I (cf. Defini-
tion 111-4-3). Define a function 9(t, y) by
f(t,ty+(t)+p), y>V'(t)+p, t EI,

9(t, y) = f (t, y), do(t) < y < 0(t) + p, t E I,
f (t, d,(t)), y < rif(t), t E I.
Then, g is continuous and bounded on the domain {t E I, -oo < y < +oo}. Hence,
every solution ¢(t, e) of the initial-value problem
= 9(t, y) + E, y(r) _
exists on the interval I.

Note that
(III.4.10) d0(t) = f (t, 0(t)) = 9(t, tl'(t)) < 9(t, tr (t)} + E

dt
on I (cf. Figure 17). Hence,
(111.4.11) t'(t) < p(t, E) on t E I.
We prove that for a given positive number p, there exists another positive number
e(p) such that if 0 < e <_ E(p), we have 0(t, e) < t1(t)+p on I. Otherwise, there exist
a real number r(p) and two sequences {Ek : k = 1, 2.... ) and {rk : k = 1, 2, ... }
of real numbers such that
(1) ek > O and lim Ek = 0,
k +oo
(2) T' > rk ? T(p) > T and lim Tk = To exists,

k+00
(3) 0(7k, ek) = IP(rk) + p,
(4) 0(t, (k) < V,(t) + p for r < t < Tk,
(5) lim 0(t, ek) = 0(t) exists uniformly on 1.
Then, O(t) is a solution of problem on the interval r <_ t < ro < T'. Since
0(ro) = Ty(ro) + p, this is a contradiction. Thus, it was proved that 0(t) S 0(t, e) <
,y(t) + p on I for 0 < e < e(p). Therefore, Theorem 111-4-5 follows immediately. 0
In the proof of Theorem 111-4-5, (111.4. 11) was derived from (111.4.10) (cf. Figure
17). In a similar manner, we can prove the following result.
Lemma 111-4-6. Assume that f (t, y) is continuous on a region 12 = {(t, y)
w_(t) < y < w+(t), t E TO}, where l0 = {t : a < t < b} and
(i) w+ and w_ are real-valued, continuous, and differentiable on T0,
(ii) w- (t) < w+ (t) on lo.
Assume also that
dw+(t)
dt
> f(t,w+(t)) on Zo,
J
dw (t) < Pt"'-(O) on TO
dt
and w_(a) w+(a). Then, every solution 0(t) of the initial-value problem
dy
(111.4.12)
dt = f (t, y), y(a) _
exists on To and w_ (t) < 0(t) < w+(t) on 10 (i.e. (t, y5(t)) E Q). In particular, the
maximal solution 01(t) and the minimal solution 02(t) of problem (111.4.12) on lo
exist and w_ (t) < 02 (t) < ¢, (t) < w+ (t) on Zo
Proof of this lemma is left to the reader as an exercise (cf. Figure 18).
Y = vKW)
v= #(z)
Y=

Using Lemma 111-4-6, we prove the following theorem due to O. Perron (Per2j.
Theorem 111-4-7. Assume that f (t, y) is continuous on a region
fl = {(t, y) (t) < y < w+(t), t E Zo},
whereto = (t:a<t <b) and

(i) w, and w_ are teal-valued, continuous, and differentiable on 4,
(ii) j- (t) 5 w+ (t) on 4.
Assume also that

dw+(t)
dt
> f(t ,W+ (t ) ) on 10,
dw
dt
(t) < f(t,w-( t) ) on 4 -
and w_ (a) < 4 < w+(a). Then, there exists a solution Q(t) of initial-value problem
(111.4.12) on To such that
w_ (t) < Q(t) < W+(t) on To, i.e., (t, ¢(t)) E Q .
Prof.
Let us define a function g(t, y) by
f(t,W+(t)), y ? w+(t), t E To,

g(t, y) = f(t, y), W- W+ (t), t E Zo,
f (t, W_ (t)), y < W- (t), t E lo.
Then, g is bounded and continuous on the region {(t, y) : t E 10, -00 < y < +oo).
Hence, every solution ¢(t) of the initial-value problem
j = g(t, y), y(a)
exists on Za.
Set "I+(t,E) = W.}. (t) + E(t - a) and w_(t,e) = w_(t) - e(t - a), where c is a
positive number and t E To. Then,
+(t, E) dW+(t)
dt
+ ( > f(t,w?(t)} + E > g(t,w+(t,e)),
dt
dw_(t,E) _ d -(t)
- E < f(t,W-(t)) - E < g(t,W_(t,E))-
dt dt
Hence, w_ (t, c) < ¢(t) < w+ (t, e) on Zo (cf. Lemma 111-4-6). Letting e 0, we
complete the proof of Theorem 111-4-7.
Comment III-4-8. The maximal and minimal solutions given in this section are
also defined and explained in [CL, pp. 45-48] and f Har2, p. 251.
111-5. A comparison theorem

In this section, we derive an estimate for solutions of a differential equation by
means of differential inequalities. Let us introduce the basic assumption.
5. A COMPARISON THEOREM 59
Assumption 1. Let t and u be two real variables and let g(t,u) be a real-valued
and continuous function of (t, u) on a domain D in the (t, u)-plane. Also let .y(t)
be the maximal solution of the initial-value problem
du
= g(t,u), u(a) = uo
on an interval lo = It : a < t < b}, where (a, uo) E D and (t, i,b(t)) E D on To.
We first prove the basic theorem given below.
Theorem 111-5-1. Assume that
(1) g(t, u) > 0 on D and uo > 0,
(2) an R"-valued function ¢'(t) is continuously differentiable on a subinterval I =
it: a< t <,r) of 4,
(3) uo,
(4) (t, 14(t) I) E V on Z,
(5)
do(t)
< g on Z.
dt
Then,
P(t) on 1.
Proof.
Let a be a positive number and let 0(t, e) be any solution of the initial-value
problem
du
dt = 9(t, u) + e, u(a) = u0.
If e > 0 is sufficiently small, i/i(t, e) exists on Zo and lim 0(t, e) = P(t) uniformly on
To (cf. Theorem III-4-5). Let us make the following observations:
(I) At)I - Im(s)I <- I ds < f g(o, I m(o) I )d

Jt do
and
(II) 0(t, e) - t'(s, e) t4Y(r,t(a,e)) + e1da,
where t > s. Suppose that l (s)I = r'(s, e) for some s E Z. Then, there exists a
positive number b(e) such that
Ig(a,+G(o, e)) - 9(o, I$(a)I)I 2 for Iv - sl < b(e).

This implies that
Id(t)I < 0(t, e) for s < t < s + b(e),
{ I
> 0(s, e) for s - b(e) < t <s
60 111. NONUNIQUENESS
(cf. Figure 19).
FIGURE 19.
Note that
t'(t,e) - WWI > f [g(a,'p(a,e))

t - g(a, j¢(o)j))do + e(t - s) for t>s
a
and
t)! C j [g(o, v(o, )) - g{O, (o )I)do - e(s - t) for t

t
Thus, it is concluded that I (t) < t'(t,f) on Z. Letting a 0, we complete the

proof of Theorem III-5-1. 0
Example 111-5-2. If an R"-valued function 0(t) satisfies the condition
dj(t) a<t<b
< C + MI¢(t)t for and uo,
dt
we can apply Theorem 111-5-1 to fi(t) with g(u) = C + Mu. Note that the initial-
value problem = C + Mu, u(a) = uo has the unique solution
+G{t) = uoeM(t-a) + M (eai(t-a) - i) .
Hence,
W(t)l <_ M (em('-4) - 1) for a < t < b.

We still assume that Assumption 1 is satisfied by the function g(t, u) and O(t)
and that uo > 0 and g(t, u) > 0 on D. Consider a differential equation
dy
(111.5.1)
dt
= f (t, yI
under the following assumption:

(a) the function f (t, yj is continuous on a domain f2 in the (t, y7)-space,
(0) (t, Iyj) E D if (t, y-) E S2,
6. SUFFICIENT CONDITIONS FOR UNIQUENESS 61
(-v) If (t, y-) I 5 g(t, I yl) for (t, yj E 0.

Then, the following result is a corollary of Theorem 111-5-1.
Corollary 111-5-3. If fi(t) is a solution of differential equation (111.5.1) on a
subinterval a < t < r of the interval Zo such that (t, fi(t)) E 12 for a < t < T and
that jd(a)I S uo, then Im(t)I < v' (t) for a < t < T.
Proof of this result is left to the reader as an exercise.
111-6. Sufficient conditions for uniqueness

In this section, we derive some sufficient conditions for uniqueness by means of
differential inequalities. The basic assumption is given below.
Assumption 1. Let t and u be two real variables and let r(t), w(t), and g(t, u) be
real-valued functions such that
(1) r(t) and w(t) are continuous on an interval I = {t : a < t < b},
(2) r(t)>0 andw(t)>0 on1.
(3) g(t, u) is continuous on the set A = {(t, u) : 0 0 on A and g(t, 0) = 0 on Z,
(5) the problem
du(t) = g(t,u(t)),
(t,u(t)) E A on Z,
dt
(111.6.1)
aim u(t) = 0
has only the trivial solution u = 0 on Z.

Let us consider a problem
d9(t)
(t, y"(t)) E Q on Z,
dt
(111.6.2)
lim 19(01 = 0,
t-.a r(t)
where S1 = {(t, y) : jyj < w(t), t E Z}. The main result of this section is the
following theorem.
Theorem 111-6-1. If the function f (t, y-) is continuous on 1 and if
I f(t, y-)I :5 g(t, Iyi) on Q,
then problem (111.6.2) has only the trivial solution y" = 6 on Z.

Proof
Suppose that problem (III.6.2) has a nontrivial solution j(t) on Z. This means
that (a) ,-a< 6. for some a E Z. Choose a positive number 6 so that f <
min aminbw(t) }.
Let us make the following two observations.

Observation 1. Note that the set A(i3) = {(t, u) : a < t < b, 0 < u < B} is a
subset ofd and that u = 0 is a solution of the differential equation
du
(111.6.3)
dt = g(t, u)
on the interval a < t < b. Using Theorem 111-2-7, we can construct a nontrivial
solution uo(t) of (111.6.3) on the interval a < t < b so that (t, uo(t)) E A(f3) on
a<t<b(cf. Figure 20).
Observation 2. The nontrivial solution uo(t) of (111.6.3) which was constructed
in Observation 1 can be continued on the interval a < t < a so that
(111.6.4) 0 < uo(t) < on a < t < a.

To show this, we first remark that for a given positive number e, any solution t(i(t, c)
(III.6.5) = g(t, u) + e, u(a) = uo(a)
satisfies the condition tfi(t, e) < on any subinterval a < t < a of I if it exists
on this subinterval (cf. Figure 21).
r=a r=b
I I u= 1 ;(r)1
I I u ='Kr' E)
I 1 L -_Ju=P
U TV, El
u= u0(t) I u uo(r)
U=0 u=0
r=b r=a
Now, define a real-valued function G(t, u) by
(g(t, u), u > 0, a < t <a,

G(t, u) 1
to, u < 0, a < t < a.
Then, any solution ty(t, e) of problem (III.6.5) can be continued on the interval
a < t < a as a solution of the differential equation = G(t, u) + e. Since the set
{t(i(-, e) : 0 < e < 1) is bounded and equicontinuous on every closed subinterval of

a < t < or, we can select a sequence {Ek : k = 1,2.... } of positive numbers such
that limEk = 0 and IirW(t,ek) = fi(t) exists uniformly on each closed subinterval
k--0 k-0
of a < t < a. It is easy to show that j)(t) is a solution of the initial-value problem
du
= g(t, u), u(a) = uo(a) which satisfies condition (111.6.4).
Thus, we constructed a nontrivial solution u0(t) of differential equation (111.6.3)
such that (t,uo(t)) E 0 on I and 0 < uo(t) < ]5(t)l on a < t < a. Since
Urn ±L = 0, this contradicts condition (5) of Assumption 1. 0
t-a r(t)
From Theorem III-6-1, we derive the following result concerning uniqueness of
solutions.
Theorem 111-6-2. Suppose that Assumption 1 is satisfied and that
(i) an 1R"-valued function f (t, y-) is continuous on a domain D in the (t, yl -space,
(ii) f satisfies the condition J f (t, 91) - f (f, 92)1 < g(t, Iyi - 92{) whenever t E 1,
Iy, - 921:5 w(t), (t, yi) E V, and (t, y2) E D.
Let Q(t) be a solution of the differential equation
dt = f (t, y) on I = {t : a < t < b}

such that the set Do = {(t, y-) 4(t)1 < w(t), t E 1} is contained in D. Then,
the problem
dylt) on 1,
dt
= f(t,b(t)), (t, y'(t)) E Do
(111.6.6)
lun y(t)r 0
t-a ()
has only the solution (t) itself.
Proof.
Set l(t) = z(t) + .(t). Then, problem (111-6.6) becomes
di(t)
dt = F(t, i(t)) = A t ' z"(t) + fi(t)) - At' fi(t)) on Z,
(t 1="(t)!) E o on Z, and Hill !AL)! = 0.

c-a r(t)
Since I1 (t,y-Q(t))j < g(t, on Do, Theorem 111-6-2 follows from Theorem
II1.6.1. 0
Let us apply Theorem III-6-2 to some initial-value problems.
Example 1 (The Osgood condition (cf. [Os])). Consider a real-valued function
g(t, u) = h(t)p(u) in the case when
(1) h(t) is continuous and h(t) > 0 on an interval 1 = {t : a < t < b},
(2) p(u) is continuous on an interval 0 ,.5 u < K, where K is a positive number,

(3) p(0) = 0 and p(u) > O for 0 < u < K,
(4) r
a+
h(t)dt < +oc,
(}
(5) jo+ p(u) = +oo.
Assume that
(1) an R"-valued function f (t, y) is continuous on a domain V in the (t, yam)-space,
(ii) f satisfies the condition If (t,1%1) - f (t, il2)1 < 9(t, Iii - Y21) = -1721)
whenever t E Z, I171 - g2 l < K, (t, y"1) E D, and (t, y2) E D.
Let ¢1(t) and 42(t) be two solutions of an initial-value problem = f (t, M, y"(a) _
i such that
(a, T11 E D and (t, p1(t)) E D, (t, 2(t)} E D
on the interval Zo = {t : a < t < b}. Then, p1(t) = ¢'2(t) on Zo.

Proof.
It is sufficient to show that Assumption 1 is satisfied by three functions r(t) = 1,
w(t) = K, and g(t, u) = h(t)p(u), and that lim 1&t) 1(t))
= 0. This limit
t-a (t)
condition is evidently satisfied, since ¢2(a) = 41(a) and r(t) = 1. Conditions (1),
(2), (3), and (4) of Assumption 1 are also evidently satisfied. Therefore, it suffices
to prove that condition (5) is also satisfied.
Let u(t) be a real-valued function such that
(A) 0<u(t) <Kon asubinterval r<t<aof I={t:a<t<b},
(B) lira+ u(t) = 0,
t
(C) dutt) = h(t)p(u(t)) on r < t < o.

Then,
1
r( d t)dt = h(t) dt.
I FU (t)) Jo
This contradicts condition (5).
Example 2 (The Lipschitz condition). If we choose h(t) = L and p(u) = u, where
L is a positive constant, then the Osgood condition becomes the Lipschitz condition
(cf. Assumption 2 of §I-1).
Example 3 (The Nagumo condition). Define r(t), w(t). and g(t, u) by
r(t) = (t -a)', w(t) = K, 9(t, u) = t L a u,
where A is a non-negative constant and K is a positive constant. Assume also that

u > 0 and t > a. Then, the general solution of the differential equation = g(t, u)
is given by u(t) = c(t - a)', where c is an arbitrary constant. This implies that
Assumption 1 is satisfied by r(t), w(t), and g(t,u). In particular, choosing A = 1,
we derive the following result due to M. Nagumo (Nal and Na2J. Assume that
(i) f (t, y') is continuous on a domain V in the (t, y-)-space,
191 -921
(ii) f satisfies the condition j f(t, 91) - f (t, 92)1 < _ whenever a < t < b,
III, -Y2I<K,(t,fi)EDand (t,Y2)EV. a dY
Suppose that 1(t) and y+,2(t) are two solutions of the initial-value problem =
dt
f (t, y), g(a) = # such that
(a, 1 E V and (t, 41(t)) E V. (t, fi2(t)) E V
on the interval I= {t: a<t<b}. Then, $1(t) = &t) on 1.

Note that, since 1(a) = $2(a) = i and _WT
j
= (a) = f (a, then
1*1(t) - 02(t)] = 0.
lim
t-a t-a
Remark 111-6-3. The sufficient conditions of Osgood's [Os] and M. Nagumo's
[Nal and Na21 for uniqueness are also explained in [CL. pp. 48-601 and [Hart, pp.
31-351.
Problem 4. Show that the initial-value problem
dy _ {tsin(f. if t 1 0,
dt 0 if t = 0, y(r) = n
has one and only one solution.

Answer.
/ \
Note that t sin I tz I - t sin (L2) = yl t cos (fi) for some y on the interval
between y1 and y2. Therefore, we can use the Lipschitz condition for r - 0, whereas
we can use the Nagumo condition at r = 0.
Remark 111-6-4. In order to apply Theorem 111-6-2 to the initial-value problems
in Examples 1 and 3, it was assumed that f (t, y-) is continuous at the initial point
(a,'qj. However, the reader must notice that problem (111.6.6) is more general
than an initial-value problem. Actually, we can apply Theorem 111-6-2 even when
f (t, yr) is not continuous at t = a. For example, the Nagumo condition (ii) of
Example 3 is satisfied by the function sin 1 t I at a = 0. Therefore, the problem
{!) = sin (i), llm -

t-0 t = 0 I has only the trivial solution y(t) = 0.
EXERCISES III
III-1.-Prove the Kneser theorem under the assumption that

(i) f'(t, yj is continuous on R = {(t, yj : a < t < b, Myj < +oo),
(ii) 111(t, y-)l < K + Llyl on R for some positive numbers K and L.
Hint. Note that if I f (t, yj I < K + LI yj on R for some positive numbers K and
L, any solution of the system 9 = f (t, yJ) exists on a < t < b and satisfies the
estimate jy(t)l < (jy(a)I + K(t - a))eL(I -a) for a < t < b.
111-2. Find the maximal solution and the minimal solution of the initial-value
problem _, y(0) _ -1 on the interval 0 < t < 10.
III-3. Set
for u > 0,
h(u) _
- i for u < 0,
H(v) = of Lh \v n) hn1J for v > 0,

-H(-v) for v < 0,
and
=0 for -1<t<-1+b,
1-4t)
>0 for -I+6<t<0 ,
0 for t = 0 ,
- c(-t) for 0 < t < 1,

where 0 < d < 1. Show that
(a) v = 0 and v = ± 1 (n = 1,2.... ) are solutions of the differential equation
n
dv _ c(t) H'(v) = dH
(E)
a H'(v) ,
where
dv '
(b) solutions of (E) with the initial-values v(-1) = 0 and v(-1) = f 1 are not
n
unique;
(c) for any positive constant c,, the differential equation
172 ISin (r)

dv __ I c(t)
v2 sin (-") +c LH'v) J
vl
satisfies the Lipschitz condition in v for jvi < +oo and Itt < 1.
Also, find the maximal solution and the minimal solution of the solutions of (E)
satisfying the initial-condition v(-1) = 0 on the interval -1 < t < 1.
Hint. See (KSI.
EXERCISES III 67
111-4. Show that if a continuously differentiable R"-valued function b(t) satisfies

an inequality
do(t)
< I40 for t > 0 and o(0) = 0,
dt
then 1¢(t)1 < t2 for t > 0.
Hint. Use Theorem 111-5-1.

111-5. Assuming that an R'-valued function j(t) is continuously differentiable for
0 < t < 1, show that if ¢(t) satisfies the condition
dd(t)
dt
s ia(t)12 for 0 < t < 1 and lim
0+
1d(t)I = 0,
9
then ,(t) = 0 for 0 < t < 1.

Hint. Use Theorem III-5-1.
111-6. Assuming that an R"-valued function fi(t) is continuously differentiable for
0 < t < +oo, show that if a(t) satisfies the condition
d9(t)
< 2tlm(t)I for 0 <- t < +oo,
dt i
l 10(t)i eXpft21 = 0,
then (t) = 0 for 0 < t < +oo.

Hint. Use Theorem 111-5-1.
111-7. Show that the problem
dy(t) = sin
- y(t) lim y(t) = 0
It1 (-0
has only the trivial solution y(t) = 0.
Hint. Note that
/
yt
-sin(/
1M l = Y'- Y2cos1
/
sin( It1 `t1
!I
j,
t1) ltl
for some y between yl and y2.

68 M. NONUNIQUENESS
111-8. Let f (t, i, y) be an R"-valued function of (t, x, y-) E R x R" x R'". Assume
that
(1) the entries of f'(t, i, yl are continuous in the region A 0<t<
a, jxrj < a, jyj < b}, where a, a, and b are fixed positive numbers,
(2) there exists a positive number K such that j f (t, x"j, y- )- f (t, x"2, y -)l < Kjxt -x"2 j
if (t,i),y-) E A (j = 1, 2).
Let U denote the set of all R"'-valued functions u(t) such that ju(t)j < b for 0 <
t < a and that JiZ(t) - u(r)j < Lit - rj if 0 _< t < a and 0 < r < a, where L
is a positive constant independent of u E U. Also, let ¢(t; u+) denote the unique
solution of the initial-value problem 'jj u(t)), x(0) = 0', where u" E U. It is
known that there exists a positive number ao such that for all u E U, the solution
(t, u') exists and u)j < a for 0 < t < aa. Denote by R the subset of Rn+'
which is the union of solution curves {(t, 0(t, 65)) : 0 < t < ao) for all u E U, i.e.,
R = {(t, ¢(t, u)) : 0 < t < aa, u E U}. Show that R is a closed set in R"+t
Hint. See [LM1, Theorem 2, pp. 44-47) and [LM2, Problem 6, pp. 282-283).
111-9. Let f (t, 2, yy be an R"-valued function of (t, a, y) E R x R" x R'". Assume
that
(1) the entries of f (t, f, y-) are continuous in the region 0 = {(t, 2, yj : 0 < t <
a, Jx"j < a, jy"j < b}, where a, a, and b are fixed positive numbers,
(2) there exists a positive number K such that yl- f (t, x2, y-)J < Kjit-i2j
if (t,i1,y)EA (j=1,2).
Let U denote the set of all R'"-valued functions g(t) such that I i(t)J < b for 0 <
t < a and that the entries of u are piecewise continuous on the interval 0 <
t < a. Also, let (t; u) denote the unique solution of the initial-value problem
di = f (t, x, u"(t)), x(0) = 0, where u" E U. It is known that there exists a positive
dt
number ao such that for all u' E U, the solution (t, u) exists and Jd(t, u)j < a for
0 <- t < ao. Denote by 1Z the subset of R"' which is the union of solution curves
{(t, ¢(t, ul)) : 0 < t _< ao) for all u E U, i.e., R = {(t, (t, u)) : 0 < t < ao, u E U).
Assume that a point (r, ¢(r, uo)) is on the boundary of R, where 0 < r < 00 and
!!o E U. Show that the solution curve 0 <- t < r} is also on the
boundary of R.
Hint. jLM2, Theorem 3 of Chapter 4 and its remark on pp. 254-257, and Problem
2 on p. 258).
III-10. Let A(t, x) and f (t, x") be respectively an n x n matrix-valued and R"-
valued functions whose entries are continuous and bounded in (t,x-) E R"+' on
a domain 0 = { (t, x) : a < t < b, x" E R"), where a and b are real numbers.
Also, assume that (r, {) E A. Show that every solution of the initial-value problem
dx
= A(t, x'a + f (t, i), i(r) = t exists on the interval a < t < b.
dt
CHAPTER IV
GENERAL THEORY OF LINEAR SYSTEMS
The main topic of this chapter is the structure of solutions of a linear system
(LP) dt = A(t)f + b(t),

where entries of the n x n matrix A(t) are complex-valued (i.e., C-valued) contin-
uous functions of a real independent variable t, and the Cn-valued function b(t) is
continuous in t. The existence and uniqueness of solutions of problem (LP) were
given by Theorem 1-3-5. In §IV-1, we explain some basic results concerning n x n
matrices whose entries are complex numbers. In particular, we explain the S-N
decomposition (or the Jordan-Chevalley decomposition) of a matrix (cf. Definition
IV-1-12; also see [Bou, Chapter 7], [HirS, Chapter 6], and [Hum, pp. 17-18]). The
S-N decomposition is equivalent to the block-diagonalization which separates dis-
tinct eigenvalues. It is simpler than the Jordan canonical form. The basic tools
for achieving this decomposition are the Cayley-Hamilton theorem (cf. Theorem
IV-1-5) and the partial fraction decomposition of reciprocal of the characteristic
polynomial. It is relatively easy to obtain this decomposition with an elementary
calculation if all eigenvalues of a given matrix are known (cf. Examples IV-1-18
and IV-1-19). In §IV-2, we explain the general aspect of linear homogeneous sys-
tems. Homogeneous systems with constant coefficients are treated in §IV-3. More
precisely speaking, we define e'A and discuss its properties. In §IV-4, we explain
the structure of solutions of a homogeneous system with periodic coefficients. The
main result is the Floquet theorem (cf. Theorem IV-4-1 and [Fl]). The Hamilton-
ian systems with periodic coefficients are the main subject of §IV-5. The Floquet
theorem is extended to this case using canonical linear transformations (cf. [Si4]
and [Marl). Also, we go through an elementary part of the theory of symplectic
groups. Finally, nonhomogeneous systems and scalar higher-order equations are
treated in §1V-6 and §IV-7, respectively. The topics of §§IV-2-IV-4, IV-6, and
IV-7 are found also, for example, in [CL, Chapter 3] and [Har2, Chapter IV]. For
symplectic groups, see, for example, [Ja, Chapter 6] and [We, Chapters 6 and 8].
IV-1. Some basic results concerning matrices

In this section, we explain the basic results concerning constant square matrices.
Let Mn(C) denote the set of all n x n matrices whose entries are complex numbers.
The set of all invertible matrices with entries in C is denoted by GL(n,C), which
stands for the general linear group of order n. We define a topology in Mn (C) by
the norm [A] = max [ask[ for A E Mn(C), where a3k is the entry of A on the j-th
1<j,k<n
all a12 ... aln
row and the k-th column; i.e., A = all a22 a . A matrix A E Mn(C)
and an2 "' ann
69
70 IV. GENERAL THEORY OF LINEAR SYSTEMS
is said to be upper-triangular if ajk = 0 for j > k. The following lemma is a basic

result in the theory of matrices.
Lemma IV-1-1. For each A E Mn(C), there exists a matrix P E GL(n,C) such
that P-1 AP is upper- triangular.
Proof.
Let A be an eigenvalue of A and pt be an eigenvector of A associated with the
eigenvalue A. Then, Apt = A P t and P 1 # 0. Choose n - 1 vectors p", (j = 2, ... , n)
so that Q = [#1A p"nJ E GL(n, C), where the p"j are column vectors of the matrix
A
0
Q. Then, the first column vector of Q-1AQ is Hence, ae can complete the
0
proof of this lemma by induction on n.
A matrix A E Mn(C) is said to be diagonal if a,k = 0 for j k. We denote
by diag(di, d2, ... , d, the diagonal matrix with entries dl, d2, ... , do on the main
diagonal (i.e., d_, = aj.,). A matrix A E Mn(C) is said to be diagonalizable (or
semisimple) if there exists a matrix P E GL(n,C) such that P-1AP is diagonal.
Denote by Sn the set of all diagonalizable matrices in Mn(C). The following lemma
is another basic result in the theory of matrices.
Lemma IV-1-2. A matrix A E Mn(C) is diagonalizable if and only if A has n
linearly independent eigenvectors pt, p2, ... , Pn
Proof
If A has n linearly independent eigenvectors pt, p'2, ... , p,, set P = [P'tpa ... pnJ
E GL(n,C). Then, P'1AP is diagonal. Conversely, if PAP is diagonal for
P = [ 1 6 t h... fl, E GL(n, C), then p1, A, ... , pn 7are n linearly independent
eigenvectors of A.
In particular, if a matrix A E Mn(C) has n distinct eigenvalues, then n eigenvectors
corresponding to these n eigenvalues, respectively, are linearly independent (cf.
[Rab, p. 1861). Therefore, we obtain the following corollary of Lemma IV-1-2.
Corollary IV-1-3. If a matrix A E Mn(C) has n distinct eigenvalues, then A E
Sn.
The set Mn(C) is a noncommutative C-algebra. This means that Mn(C) is a
vector space over C and a noncommutative ring. The set Sn is not a subalgebra of
Mn(C). However, the following lemma shows an important topological property of
Sn as a subset of Mn(C).
Lemma IV-1-4. The set Sn is dense in Mn(C).
Prioof.
It must be shown that, for each matrix A E Mn(C), there exists a sequence
{Bk : k = 1,2,... } of matrices in Sn such that lim Bk = A. To do this, we
k +W
may assume without any loss of generality that A is an upper-triangular matrix
with the eigenvalues At, ... , An on the main diagonal (cf. Lemma IV-1-1). Set
1. SOME BASIC RESULTS CONCERNING MATRICES 71
B k = A + diag[Ek.1, Ek,2, , Ek,n], where the quantities ek,,, (v = 1, 2, ... , n) are

chosen in such a way that n numbers Al + Ek,1, A2 + 4,2, ... , An + Ek,n are distinct
and that lim Ek,&, = 0 for v = 1, 2,... , n. Then, by Corollary IV-1-3, we obtain
Bk E Sn and lim Bk = A.
k-r+oo
For a matrix A E Mn(C), denote by pA(A) the characteristic polynomial of A
with the expansion
n
(IV.1.1) pA(A) = det(AIn - A] = An + ph(A)An-''
h=1
where In denotes the n x n identity matrix. Note that
PA(A) = An + Eph(A)An-h,
A° = In.
h=1
Now, let us prove the Cayley-Hamilton theorem (see, for example, [Be13, pp. 200-
201 and 220], (Cu, p. 220], and (Rab, p. 198]).
Theorem IV-1-5 (A. Cayley-W. R. Hamilton). If A E Mn(C), then its charac-
teristic polynomial satisfies PA(A) = O, where 0 is the zero matrix of appropriate
size.
Remark IV-1-6. The coefficients ph(A) of pA(A) are polynomials in entries ap,
of the matrix A with integer coefficients.
Proof of Theorem IV-1-5.
Since the entries of pA(A) are polynomials of entries a,k of the matrix A, they
are continuous in the entries of A. Therefore, if pA(A) = 0 for A E Sn, it is
also true for every A E Mn(C), since Sn is dense in Mn(C) (cf. Lemma IV-1-4).
Note also that if B = P'1 AP for some P E GL(n, C), then pB(A) = PA(A) and
pB(B) = P-'pA(A)P. Therefore, it suffices to prove Theorem IV-1-5 for diagonal
matrices. Set A = diag(A1, A2, ... , A.J. Then, pA(A) = (A - A1)(A - A2) ... (A - An)
and pA(A) = diag[PA(A1),PA(A2),... ,PA(A.)] = 0.
It is an important application of Theorem IV-1-5 that an n x it matrix N satisfies
the condition N" = 0 if its characteristic polynomial pN(A) is equal to A". If
N" = O, N is said to be nilpotent.
Lemma IV-1-7. A matrix N E Mn(C) is nilpotent if and only if all eigenvalues
of N are zero.
Proof.
If IV is an eigenvector of N associated with an eigenvalue A of N, then Nkp"= Akp'
for every positive integer k. In particular, N"p = A"p Hence, if N" = 0, then
A = 0. On the other hand, if all eigenvalues of N are 0, the characteristic polynomial
pN(A) is equal to A". Hence, N is nilpotent. 0
Applying Lemma IV-1-1 to a nilpotent matrix N, we obtain the following result.
Lemma IV-1-8. A matrix N E .Mn(C) is nilpotent if and only if then exists a

matrix P E GL(n,C) such that P-I NP is upper-triangular and the entries on the
main diagonal of N are all zero. Furthermore, if N is a real matrix, then there
exists a real matrix P that satisfies the requirement given above.
To verify the last statement of this lemma, use a method similar to the proof of
Lemma IV-1-1 together with the fact that if an eigenvalue of a real matrix is real,
then there exists a real eigenvector associated with this real eigenvalue. Details are
The main concern of this section is to explain the S-N decomposition of a matrix
A E Mn(C) (cf. Theorem IV-1-11). Before introducing the S-N decomposition,
we need some preparation.
Let A. (j = 1, 2,... , k) be the distinct eigenvalues of A and let m., (j =
1,2.... , k) be their respective multiplicities. Then, the characteristic polynomial
of the matrix A is given by pA(A) = (A - '\0M'(1\ - A2)m2 ... (A - Ak)m'. Decom-
QU(A
pose 1 into partial fractions 1 = , where, for every j, the
pA(A) pA(A) (A - A,),n,
quantity Q, is a nonzero polynomial in A of degree not greater than m, - 1. Hence,
k
I = EQ,(A) J1 (A - Ah)m". Setting
)=1 h¢j
P,(A) = Qj(A) 1I (A - A,)me

h#1
i= P2(A}.
J=1
Now that this is an identity in A. Therefore, setting
(IV.1.4) Pj(A) = Qj(A)fl(A-AhIn)m" (y = 1,2,....k),

hv&1
we obtain
k
(IV.1.5) I. _ > Ph(A).
h=1
In the following two lemmas, we show that (IV.1.5) is a resolution of the identity in
terms of projections Ph (A) onto invariant subspaces of A associated with eigenvalues
Ah, respectively.
Lemma IV-1-9. The k matrices P, (A) (j = 1,2,... , k) given by (W-1.4) satisfy
the following conditions:
(i) A and P, (A) (y = 1, 2, ... , k) commute.
(ii) (A-),In)'n'Pi(A) =0 (j = 1,2,... ,k),

(iii) P,(A)Ph(A) = 0 ifj h,
k
(iv) >Ph(A) = In,
h=1
(v) Pi(A)2=P,(A) (j=1.2,...,k),
( v : ) P, (A) 0 (j = 1, 2, ... , k).
Proof.
Since P,(A) is a polynomial of A, we obtain (i). Using Theorem IV-1-5, we
derive (ii) and (iii) from (IV.1.4) and (i). Statement (iv) is the same as (IV.1.5).
Multiplying the both sides of (IV.1.5) by P,(A), we obtain
k
(IV.1.6) P,(A) _ >P,(A)Ph(A).
h=1
Then, (v) follows from (IV.1.6) and (iii). To prove (vi), let IT, be an eigenvector of
A associated with the eigenvalue Al. Note that (IV.1.2) implies Ph(A)) = 0 if h 0 j.
Therefore, we derive P,(A,) = 1 from (IV.1.3). Now, since P. (A)#, = P,(A3)p' #
we obtain (vi).
Lemma IV-1-10. Denote by V. the image of the mapping P,(A) : C" -. Cn.
Then,
(1) p'E Cn belongs to V. if and only if P,(A)p= p
(2) Pj(A)p"=0 for all fl E Vh if j 0 h.
(3) Cn = V1 Ei3 V2 e e Vk (a direct sum).
(4) for each j, V, is an invariant subspace of A.
(5) the restriction of A on V, has a coordinates-wise representation:
(IV.1.7) Alv, : AjIj + A,,

where I. is the identity matrix and Nj is a nilpotent matrix.
(6) dime V, = m, .
Proof
Each part of this lemma follows from Lemma IV-1-9 as follows.
A vector IT E V3 if and only if p" = P, (A)q" for some q' E Cn. If p" = P, (A) q, we
obtain P,(A)p= Pj(A)2q'= Pj(A)q =p""from (v) of Lemma IV-1-9.
A vector p" E Vh if and only if ff = Ph (A),y for some q" E Cn. Hence, from (iii)
of Lemma IV-1-9 we obtain P, (A)IF = Pj (A)Ph(A)q"= 0 if 0 h.
(iv) of Lemma IV-1-9 implies p = P, (A)15 + + Pk- (A)p" for every p3 E C",
while (1) implies that P, (A)p E Vj. On the other hand, if p" = )51 + + pk
for some g,E V2 (j = 1,2,... , k), then, by (1) and (2), we obtain P,(A)p" =
Pi(A)p1+...+PJ(A)pk =pj
Ap"= AP,(A)p = P,(A)Ap E V3 for every 15E Vj.
Let n, be the dimension of the space V, over C and let {ff,,t : 1 = 1, 2, ... , n. }
be a basis for V,. Then, there exists an n. X n, matrix N,, such that
(A-)'1In)V1,1PJ,2...p'1.n,J = [Pi,1PJ,2...p3i.nsJNj
as the coordinates-wise representation relative to this basis. This implies that
(A - A,1.)'Pi(A)(P1,1Pi,2...p),nj) _ I
IP1,191,2...Pj,Nj for (t = 1,2,... ).

In particular, from (ii) of Lemma IV-1-9, we derive N, O . Thus, we
obtain
(IV-1-8) ...P = [l ,1P'r,2 ... p'f n,](A) I, + N,),

where 1, is the n, x n1 identity matrix. This proves (IV.1.7).
(6) Let { " l , t : e = 1, 2,... , n, } be a basis for V, (j = 1, 2,... , k). Set
(IV.1.9) Po = (p1,1 . p'2...,
Then, Po E GL(n, C) and (IV.1.8) implies
([V.1.10) Po'AP0 = diagjAllt +N1,A212+N2,...,Aklk+Nk],

where the right-hand side of (IV.1.10) is a matrix in a block-diagonal form with
entries Al h +N1, A212+N2, ... , Aklk+Nk on the main diagonal blocks. Hence,
pA(A) A2)'2 Also, PA(A) _ (A - A1)m'(A -
X2 )12 (A - Ak)mk. Therefore, dimC V) = n, = m, (j = 1,2, ... , k). 0
The following theorem defines the S-N decomposition of a matrix A E Mn(C).
Theorem IV-1-11. Let A be an n x n matrix whose entries are complex numbers.
Then, there exist two n x n matrices S and N such that
(a) S is diagonalizable,
(b) N is nilpotent,
(c) A = S + N,
(d) SN = NS.
The two matrices S and N are uniquely determined by these four conditions. If
A is real, then S and N are also real. Furthermore, they are polynomials in A
with coefficients in the smallest field Q(a,k, A,1) containing the field Q of rational
numbers, the entries ajk of A, and the eigenvalues Al, A2 ... , Ak of A.
Proof
We prove this theorem in three steps.
Step 1. Existence of S and N. Using the projections P,(A) given by (IV.1.4), define
S and N by
S = A1P1(A)+A2Pz(A)+...+At Pk(A), N=A - S.
If P0 is given by (IV.1.9), then
(IV.1.11) Po 1SP0 = diag[A111, A212, ... , Aklk)

and
(IV.1.12) Po 1 NPo = diag[N1 i N2, ... , Nk]
from Lemmas IV-1-9 and IV-I-10 and (IV.1.10). Hence, S is diagonalizable and N
is nilpotent. Furthermore, NS = SN since S and N are polynomials in A. This
shows the existence of S and N satisfying (a), (b), (c), and (d). Moreover, from
(IV.1.4), it follows that two matrices S and N are polynomials in A with coefficients
in the field Q(ajk, Ah).
Step 2. Uniqueness of S and N. Assume that there exists another pair (S, N)
of n x n matrices satisfying conditions (a), (b), (c), and (d). Then, (c) and (d)
imply that SA = AS and NA = AN. Hence, SS = SS, NS = SN, SN = NS,
and NN = NN since S and N are polynomials in A. This implies that S - S is
diagonalizable and N - N is nilpotent. Therefore, from S - S = N - N, it follows
that S-S=N-N=O.
Step 3. The case when S and N are real. In case when A is real, let 5 and N be
the complex conjugates of S and N, respectively. Then, A = S + N = 3° + N.
Hence, the uniqueness of S and N implies that S = 3 and N = N.
This completes the proof of Theorem IV-1-11.
Definition IV-1-12. The decomposition A = S + N of Theorem IV-1-11 is called
the S-N decomposition of A.
Remark IV-1-13. From (IV.1.11), it follows immediately that S and A have the
same eigenvalues, counting their multiplicities. Therefore, S is invertible if and only
if A is invertible.
Observation IV-1-14. Let A be an n x n matrix whose distinct eigenvalues are
A = S + N be the S-N decomposition of A. It can be shown
that n x n matrices P1, P2, ... , Pk are uniquely determined by the following three
conditions:
(i)
(ii) P,P1 = O if j 36 t,
(iii) S = A11P1 + A2P2 + ... + AkPk.
Proof.
Note that
In = P1(A) + P2(A) + ... + Pk(A),
{ Pj(A)Ph(A) = O if j h,
S = A1P1(A) + A2P2(A) + ... + AkPk(A).
k
First, derive that P, S = SP; = \j P,. Then, this implies that .X P1 = >ahPj P1. (A).
h=1
Hence, \jPjPA(A) = \h PiPh(A). Thus, PiPh(A) = 0 whenever j h. Therefore,
it follows that P1 = P1(A) = P2P}(A).
Observation IV-1-15. Let A = S + N be the S-N decomposition of an n x n

matrix A. Let T be an n x n invertible matrix such that if we set A = T-1ST,
then A = diag[A1I1, A2I2, ... , AkIk], Where A1, A2, ... , Ak are distinct eigenvalues
of S (and also of A), I, is the m3 x mj identity matrix, and m3 is the multiplicity
of the eigenvalue A,. It is easy to show that
(i) if we set M = T-'NT, then M is nilpotent, MA = AM, and M =
diag[M1i M2,... , Mk}, where Mj are mj x m j nilpotent matrices,
(ii) if we set Pj = Tdiag[Ej1iE.,2,... ,E,k]T-1, where E,1 = 0 if j 1, while
Ejj = Ij, we obtain
(I =P1+P2+...+Pk, PjPh =0 (.1 h),
. S = A1P1 + A2P2 + ... + AkPk.
Therefore, P. = P, (S) = P? (A) (j = 1, 2, ... , k) (cf. Observation IV-1-14).
The following two remarks concern real diagonalizable matrices.
Remark IV-1-16. Let A be a real nxn diagonalizable matrix and let A1, A2 ... , An
be the eigenvalues of A. Then, there exists a real n x n invertible matrix P such
that
(1) in the case when all eigenvalues A j (j = 1,2,... , n) are real, then P-1 AP is
a real diagonal matrix whose entries on the main diagonal are A1, A2, ... , An,
(2) in the case when all eigenvalues are not real, then n is an even integer 2m and
P-1A fP = diag[D1, D2, ... , DmJ, where A23_1 = a, + ibl, A2J = a) - ibj, and
Dj
abj a,
(3) in other cases, P1AP = diag[D1, D2i ... , Dhj, where A23_1 = a)+ib,, A2.1 =
a, - ib,, and D. = I b, a j for j = 1, 2, ... , h - 1 and Dh is a real diagonal
matrix whose entries on the main diagonal are A, (j = 2h - 1,... , n).
Remark IV-1-17. For any given real n x n matrix A, there exists a sequence
{Bk : k = 1, 2,. ..} of real n x n diagonalizable matrices such that lim Bk = A.
This can be proved in the following way:
(i) let A = S + N be S-N decomposition of A,
(ii) using Remark IV-1-16, assume that S = diag[D1, D2, ... , Dhj, as in (3) of
Remark IV-1-16,
(iii) find the form of N by SN = NS,
(iv) triangularize N without changing S,
(v) use a method similar to the proof of Lemma IV-1-4.
Details of proofs of Remark IV-1-16 and IV-1-17 are left to the reader as exercises.
Now, we give two examples of calculation of the S-N decomposition.
252 498 4134 698
-234 -465 -3885 -656
Example IV-1-18. The matrix A 15 30 252 42
has two
-10 -20 -166 -25
distinct eigenvalues 3 and 4, and
1 2 2
PA(A) = (A - 4)2(A - 3)2,
PA(A) (A 14)2 A 4 + (A 13)2 + A 3
Set
P1(A) = (A - 3)2
-2(A-4)(A-3)2'
P2(A) _ (A - 4)2 + 2(A - 3)(A - 4)2.
Then,
-1 -2 134 198 2 2 -134 -198
1 2 -125 -186 -1 -1 125 186
P1(A) = 0 0 9 12
P2(A) =
0 0 -8 -12
0 0 -6 -8 0 0 6 9
Therefore,
2 -2 134 198
1 5 -125 -186
S = 4P1(A) + 3P2(A) =
0 0 12 12 I'
10 0 -6 -5
250 500 4000 500
-235 -470 -3760 -470
N = A - S =
15 30 240 30
-10 -20 -160 -20
3 4 3
Example IV-1-19. The matrix A = 2 7 4 has two distinct eigenvalues
-4 8 3
A1= 11, A2=1,and
1 ) 21 1 ) , 1 _ 1 _ (A+9)
PA(A)
PA(A) 100(A - 11) 100(A - 1)2*
Hence,
1 (A - 1)2 - (A + 9)(A - 11)
-
100 100
Set
P1(A) (A - 1)2 (A + 9)(A - 11).
_ P2(A) - -
100 100
Then,
1 0 56 28 1 1100 -56 -28
P1(A) =100- 0 76 38 , P2(A)
100
0 24 -38
0 48 24 0 -48 76
Therefore,
10 56 28
1
S = 11P1(A) + P2(A) = 10 0 86 38
0 48 34 1
20 -16 2
N=A - S= 10 20 -16 2
-40 32 -4
In this case, SN = NS = N and N2 = O. Let Vj = Image (PI(A)) (j = 1,2).

Then, by virtue of (1) of Lemma IV-1-10, Pj(A)p" = p" for all IT E Vj (j=1,2).
14 1 0
Furthermore, V1 is spanned by 19 and V2 is spanned by 0 and 1 Set
12 0 -2
14 1 0
Pa = 19 0 1 . Then,
12 0 -2
11 0 0
Po 1 SPo = 0 1 0 Pp 1 NPo = 2 [00 1 -1 .
0 0 1 0 1 -1
It is noteworthy that there is only one linearly independent eigenvector x =
for the eigenvalue A2 = 1.

It is not difficult to make a program for calculation of S and N with a computer.
For more examples of calculation of S and N, see [HKSI.
IV-2. Homogeneous systems of linear differential equations

In this section, we explain the basic results concerning the structure of solutions
of a homogeneous system of linear differential equations given by
(IV.2.1) dt = A(t)y',
where the entries of the n x n matrix A(t) are continuous on an interval I = It :
a < t < b). Let us prove the following basic theorem.
Theorem 1V-2-1. The solutions of (1V2. 1) forms an n-dimensional vector space
over C.
We break the entire proof into three observations.
Observation IV-2-2. Any linear combination of a finite number of solutions of
(IV.2.1) is also a solution of (IV.2.1). We can prove the existence of n linearly
independent solutions of (IV.2.1) on the interval Z by using Theorem I-3-5 with n
linearly independent initial conditions at t = to. Notice that each column vector of
a solution Y of the differential equation
(IV.2.2) dt = A(t)Y
on an n x n unknown matrix Y is a solution of system (IV.2. 1). Therefore, construct-
ing an invertible solution Y of (IV.2.2), we can construct n linearly independent
solutions of (IV.2.1) all at once. If an n x n matrix Y(t) is a solution of equation
(IV.2.2)on an interval I={t:a<t<b}and Y(t)EGL(n,C)for all tE1,then
Y(t) is called a fundamental matrix solution of system (1V.2.1) on.T. Furthermore,
n columns of a fundamental matrix solution Y(t) of (IV.2.2) are said to form a
fundamental set of n linearly independent solutions of (IV.2.1) on the interval T.
2. HOMOGENEOUS SYSTEMS OF LINEAR DIFF. EQUATIONS 79
Observation IV-2-3. Let 4(t) be a solution of (IV.2.2) on Z. Also, let *(t) be a

solution of the adjoint equation of (IV.2.2):
dZ
(IV.2.3)
dt
= -ZA(t)
on the interval Z, where Z is an n x n unknown matrix. Then,
-W(t)A(t)4s(t) + $(t)A(t)4;(t) = 0.
dt
This implies that the matrix %P(t)4i(t) is independent of t. Therefore, W(t)4i(t) =
%P(r)t(r) for any fixed point r E Z and for all t E Z. Note that the initial values
44(r) and %P(r) at t = r can be prescribed arbitrarily. In particular, in the case
when 4?(r) E GL(n,C), by choosing %P(r) = we obtain WY(t)4i(t) = In
for all t E Z. Thus, we proved the following lemma.
Lemma IV-2-4. Let an n x n matrix 4i(t) be a solution of (IV.2.2) on the interval
Z. Then, 45(t) is invertible for all t E I (i.e., a fundamental matrix solution of
(IV.2.1)) if 4i(r) is invertible for some r E Z. Furthermore, 4 (t)-1 is the unique
solution of (IV.2.8) on Z satisfying the initial condition Z(r) = 4i(r)-1.
Observation IV-2-5. Denote by 4?(t; r) the unique solution of the initial-value
problem
-
(IV.2.4) = A(t)Y, Y(r) = In,
dt
where r E Z. Then, 1(t; r) E GL(n, C) for all t E Z. The general structure of
solutions of (IV.2.1) and (IV.2.2) are given by the following theorem, which can be
easily verified.
Theorem IV-2-6. The C"-valued functwn y(t) = 44(t; r)it is the unique solution
A(t)y", y(r) =1,

dt
where it E C. Also, the n x n matrix Y = 4i(t; r)r is the unique solution of the
dY = A(t)Y,
Y(r) = I
dt
where r E
Theorem IV-2-1 is a corollary of Theorem IV-2-6. 0
Remark IV-2-7.
(1) The general form of a fundamental matrix solution of (IV.2.1) is given by
Y(t) = 4i(t; r)r, where r E GL(n, C).
(2) If a fundamental matrix solution is given by Y(t) = 4'(t; r)r, then Y(r) = F.
Hence,
(IV.2.5) 0(t; ,r) = Y(t)Y(r)-1
(t, r E Z)
80 W. GENERAL THEORY OF LINEAR SYSTEMS
for any fundamental matrix solution Y(t). In particular,
(IV.2.6) 1(t;r) _ for t, r, r1 E Z.

(3) In the case when A(t) is a scalar (i.e., n = 1), we obtain easily
(IV.2.7) 1(t; r) = exp I JA(s)ds)]

t
t 1
In the general case, we define exp I J A(s)ds] by
V,r
+M
exp[B(t)] = In
T11.
B(t)"', where B(t) = Jt A(s)ds.
m=1
However, generally speaking, (IV.2.7) holds only in the case when B(t) and
B'(t) = A(t) commute. In particular, 4(t; r) = exp[(t - r )AJ if A = A(t) is
independent of t. In §IV-3, we shall explain how to calculate exp((t - r)AJ,
using the S-N decomposition of A. Also, (IV.2.7) holds in the case when
A(t) is diagonal on the interval I. A less trivial case is given in Exer-
cise IV-9. It is easy to see that B(t) and A(t) commute if A(t) is a 2 x 2
upper-triangular matrix with an eigenvalue of multiplicity 2. For exam-
ple, the matrix A(t) = [coStt I satisfies the requirement. In this
(jsin t o sin r
case, 4b(t; r) = exp I J t A(s)ds = exp sint - sin r, )
1
exp(sin t -sin r) 10
t 1 r
1 J1
if
(4) det Y(t) = det Y(r) exp trA(s)ds) if Y(t) satisfies (IV.2.2), where det A
and trA are the determinant and trace of the matrix A. This formula is
known as Abel's formula (cf. (CL, p. 28]).
Proof.
Regarding detY(t) as a function of n column vectors {y'1(t),... of
Y(t), set det Y(t) = 7(g, (t), ... , y (t)). Then,
(IV.2.8)
d det Y(t)
dt
- [:
L P(... , A(t)ym(t), ... ).
M=1
Denote the right-hand side of (IV.2.8) by G(y'1(t),... Then, 9 is

multilinear and alternating in y1(t),... ,j,,(t). Furthermore, 9 = trA(t) if
Y(t) = I,,. Therefore, 9 = tr A(t) det Y(t). Solving the differential equation
ddet Y(t)
= tr A(t) det Y(t), we obtain Abel's formula. 0
dt
3. HOMOGENEOUS SYSTEMS WITH CONSTANT COEFFICIENTS 81
IV-3. Homogeneous systems with constant coefficients

For an n x n matrix A, we define exp[A] by
(IV.3.1) exp[A] = In + E h Ah.

h=1
It is easy to show that the matrix exp[A] satisfies the condition exp[A + B] =
exp[A] exp[B] if A and B commute. This implies that exp[A] is invertible and
(exp[A])-1 = exp[-A]. Thus, we obtain a fundamental matrix solution Y = exp[tA]
of the system d = Ay with a constant matrix A by solving the initial-value
problem
(IV.3.2) = AY, Y(0) = In.

This, in turn, implies that the unique solution of the initial-value problem
(IV.3.3) d = Ay, y(T) = P

is given by y = exp[(t - r)A]p, where p E C' is a constant vector.
In this section, we explain how to calculate exp[tA] for a given constant matrix
A, using the S-N decomposition of A. Assume that an n x n matrix A has k distinct
eigenvalues Al, A2, ... , Ak. Let A = S + N be the S-N decomposition of A. Also,
let Pj (A) (,j = 1, 2,... , k) be the projections defined in §IV-1 (cf. (IV.1.4)). Then,
k k
(IV.3.4) In = > PJ(A), S= AjPJ(A), N = A - S,
J=1 )=1
and
(IV.3.5)
h j
PJ(A)Ph(A) (3, h = 1, 2, .... k).
0 (A) if h =
36 j
The two matrices S and N commute.
Denote by V, the image of the mapping PJ(A) : C" Cn (cf. Lemma IV-I-10).
It is known that Sp = \,)5 for pin VJ . Hence, Sl p" = \1)5 and
+00
exp[tS]p' = 1 + E (A2t)n }ic earh=1
n-i h
On the other hand, exp(tN] = In + A Nh since N is nilpotent. Therefore,
h=1
exp[tA)# = exp[t(S + N)]p' = expitN] exp[tS]p" = e'\'` exp[tN]P

(IV.3.6) [In n- i th
= eA'` + p" for fl E Vj.
n=1
Applying (IV.3.4) and (IV.3.6) to a general p E C", we derive

n-10
(IV.3.7) exp[tA]p = C.Njt
In + E hl Nh P, (A)p' for 9E C".
j=1 h=I
Thus, we proved the following theorem.

Theorem IV-3-1. The matrix exp[tA] is calculated by formula
k n-1 th
(IV.3.8) exp[tA] _ > ea+` In + F, T! Nh Pj (A).
=l h=1
Since the general solution of the differential equation
(IV.3.9)
is given by (IV.3.7), the following important result is obtained.

Theorem IV-3-2.
(i) If R(,\,) < 0 for j = 1, 2, ... , k, then every solution of (IV. 3.9) tends to 06 as
t -' +00,
(ii) if R(Aj) > 0 for some j, some solutions of (IV.3.9) tend to 00 as t - +00,
(iii) every solution of (IV.3.9) is bounded for t > 0 if and only if R(AE) < 0 for
j = 1, 2, ... , k and NP, (A) = 0 if R(A)) = 0.
Now, we illustrate calculation of exp[tA] in two examples. Note that in the case
when A has nonreal eigenvalues, we must use complex numbers in our calculation.
Nevertheless, if A is a real matrix, then exp[tA] is also real. Hence, at the end of
our calculation, we obtain real-valued solutions of (IV.3.9) if A is real.
-2 1 0
Example IV-3-3. Consider the matrix A = 0 -2 0 . The characteristic
3 2 1
polynomial of A is pA(A) _ (A -1)(A+2)2. By using the partial fraction decomposi-
1 (A + 2)2 - (A + 5)(A - 1) (A + 2)2
tion of we derive 1 = 9 . Setting P1(A) =
PA(A), 9
+ 5)(A - 1),
and P2(A) we obtain
9
0 0 0 1 0 0
P1(A) = 0 0 0, P2(A) = 0 1 0
1 1 1 -1 -1 0
Set
-2 0 0 0 1 0
S = PI(A) - 2P2(A) = 0 -2 0 , N=A-S= 0 0 0.
3 3 1 0 -1 0
Note that N2 = 0. Hence,

exp[tA] = e'[ 13 + tN ] PI(A) + e-21 [ 13 + tN ] P2(A)
e-2t to-2t
0
0 e-2c 0
t)e-2t
et - e-2t et - (1 + et
The solution of the initial-value problem dt = Ay, y(0) = y is y(t) = exp[tA]i).

To find a solution satisfying the condition Jim y(0) = 0, we must choose it so that
t +m
P1(A)ij = 0. Such an iJ is given by it = P2(A)c", where c is an arbitrary constant
vector in C3.
0 -1 1
Example IV-3-4. Next, consider the matrix A = 1 0 -1 . The charac-

-1 1 0
teristic polynomial of A is PA(A) = A(A2 + 3) = A(A - if)(A + i\/3-). Using the
1
partial fraction decomposition of , we obtain
PA (A)
1= (A- if)(A+ if) - A(A+ if)- sA(A- if}.

Setting
P1(A) = 3(A2 + 3), P2(A) I A(A + if), P3(A) A(A - if),

we obtain
1 1 1 1 1 1-if 1+if
-2
Pi(A) = - 1 1 1 , P2(A) = -- 1+if -2 1-if
3 1 1 1 6 1-if 1+iv -2
and P3(A) is the complex conjugate of P2(A). If we set
S = (if)P2(A) - (if)P3(A),
then S = A. This implies that N = 0. Thus, we obtain
exp(tA] = P1(A) + e,.''P2(A) + e-;J1tP3(A)

= P1(A) + 23t (e'3tP2(A)) .
Using
(e'' t(1 + if)) = cos(ft) - \/3-sin(ft),
2 (e'-1t(1 - if)) wa(ft) + f sin(ft),
we find
a(t) b(t) c(t)
exp[tA] = 1 c(t) a(t) b(t)
3
b(t) c(t) a(t)
where
a(t) = I + 2cos(ft),
b(t) = 1 - {cos(ft) + f sin(ft)} ,
c(t) = 1 - {wa(ft) - f sin(ft))
Remark IV-3-5. Fhnctions of a matnx In this remark, we explain how to define
functions of a matrix A.
I. A particular case: Let A0, I,,, and N be a number, the n x n identity matrix,
and an n x n nilpotent matrix, respectively. Also, consider a function f (X) in a
neighborhood of A0. Assume that f (A) has the Taylor series expansion (i.e., f is
analytic at A0)
(h)
f h?Ao)(A
f(A) = f(A0) + 1 - Ao)h.
h=1
In this case, define f (,\o1 + N) by

=
f(h)PLO) .A = n-1 f(h)(AO) NA.
f(AoI. + N) = f(Ao)I. + A f(Ao)II +
h.t h.
h=1 h=1
n-1 (h)
Since N is nilpotent, the matrix >2f is also nilpotent. Therefore, the
h=1
characteristic polynomial pf(A0,+N)(A) of f(AoI + N) is
Pf(aol-N)(A) = (A - 1(A0))".
II. The general case: Assume that the characteristic polynomial PA(A) of an n x n
matrix A is
PA(A) = (A - .l1)m1(A - A 2 ) ' - 2 ... (A - Ak)mr.
where A1,... , Ak are distinct eigenvalues of A. Construct P, (A) (j = 1, ... , k), S,
and N as above. Then,
A = (A1In + N)P1(A) + (A21n + N)P2(A) + ... + 0k1n + N)Pk(A).
Therefore,
A' = (A11n + N)1P1(A) + (A2In + N)'P2(A) 4- ... + (Ak1n + N)'Pk(A)
for every integer P.

Assuming that a function f (A) has the Taylor series expansion
f(A) = f(A3) 0f(h)(A,)(A-A,)''

00
h!
h=1
at A = A., for every j = 1, ... , k, we define f (A) by
f(A) = f (A11, + N)Pk(A) + f (A21n + N)P2(A)

(IV.3.10)
+ ... + f (Akin + N)Pk(A).
Since P2(S) = P,(A) (cf. Observation IV-1-15), this definition applied to S yields
f(S) = f(AIIn)P1(A) + f(A21.)P2(A) + -' + f(AkII)Pk(A)

and f (A) - f (S) has a form N x (a polynomial in S and N). Therefore, f (A) - f (S)
is nilpotent. Furthermore, f (S) and f (A) commute. This implies that
f(A) = f(S) + (f(A) - f(S))

is the S-N decomposition of f (A). Thus,
Pf(A)(A) = pf(s)(A) = (A - f(A1))m' ... (A - f(Ak))m
Example IV-3-6. In the case when f (A) = log(A), define log(A) by
log(A) = log(A1In+N)Pk(A) + log(A21n+N)Pk(A) + ... + log(Akln+N)Pk(A),
where we must assume that A is invertible so that A, # 0 for all eigenvalues of A.

Let us look at log(AoI,, + N) more closely, assuming that A0 0 0. Since
/
log(Ao + u) = 1000) + log t 1 + -o) = logl o) +
+O° m+1
(-I)m
\\ m=1
Io
we obtain
n-1 (-1)m+1
log(.1oIn + N) = log(Ao)ln +
m=1
m
It is not difficult to show that exp[log(A)J = A. In fact, since
(log(A))m = (log(A11n + N))mPk(A) + (log(A21n + N))mp2(A)

+ ... + (log(Akln + N))mPk(A),
it is sufficient to show that exp[1og(Aoln + N)J = A01,, + N. This can be proved by
using exp[log(Ao +,u)] =A0 +,u.
Observation IV-3-7. In the definition of log(A) in Example N-3-6, we used

log(A,). The function log(A) is not single-valued. Therefore, the definition of
log(A) is not unique.
Observation IV-3-8. Let A = S + N be the S-N decomposition of A. If A is
invertible, S is also invertible. Therefore, we can write A as A = S(II + M), where
M = S' N = NS-1. Since S and N commute, two matrices S and M commute.
Furthermore, M is nilpotent. Using this form, we can define log(A) by
log(A) = log(S) + log(Ih + M),
where
log(S) = Iog(A1)P1(A) + log(A2)P2(A) + . + log(Ak)Pk(A)

and
(1)m+1
log(IR + Al) = E - m Mm.
M=1
This definition and the previous definition give the same function log(A) if the same
definition of log(A,) is used.
3 4 3
Example N-3-9. Let us calculate sin(A) for A = 2 7 4 (cf. Example
-4 8 3
IV-1-19). The matrix A has two eigenvalues 11 and 1. The corresponding projec-
tions are
0 14 7 -14 -7
25 25 25 25
P1(A) = 0 L9 L9 0
P2(A) = 2s 50
12
25 6
50 -12 19
0 0
25 25 5 25
Define S = 11P1(A) + P2(A) and N = A - S. Then N2 = 0. Also,
( sin(11 + x) = -0.99999 + 0.0044257x + 0.499995x2 + 0(x3 ),
t sin(1 + x) = 0.841471+0.540302x-0.420735x 2 + 0(x3).
Therefore,
sin(A) _ (-0.9999913 + 0.0044257N)P1(A) + (0.84147113 + 0.540302N)P2(A)
1.92207 -1.8957 -0.407549
1.0806 -1.42252 -0.591695
-2.1612 0.845065 0.1834
It is known that sin x has the series expansion

+00
sin x = 1 (2h +)1)I
T2h+1
4. SYSTEMS WITH PERIODIC COEFFICIENTS 87
Therefore, we can also define sin(A) by
sin(A) -
-- + (2h(-1)h
+ 1)!
A2h+1
However, this approximation is not quite satisfactory if we notice that
_1h
sin(11) = -0.99999 and 112h+1 = -117.147.
(2h + 1)!
h=O
IV-4. Systems with periodic coefficients

In this section, we explain how to construct a fundamental matrix solution of a
system
dy
(IV.4.1) = A(t)y
dt
in the case when the n x n matrix A(t) satisfies the following conditions:
(1) entries of A(t) are continuous on the entire real line R,
(2) entries of A(t) are periodic in t of a (positive) period w, i.e.,
(IV.4.2) A(t + w) = A(t) for t E R.
Look at the unique n x n fundamental matrix solution 4'(t) defined by the initial-
value problem
dY = A(t)Y,
(IV.4.3)
.it
Y(0) = In
Since 4i '(t + w) = A(t + w)4S(t + w) = A(t)4'(t + w) and +(0 + w) = 4t(w), the

matrix 41(t +w) is also a fundamental matrix solution of (IV.4.3). As mentioned in
(1) of Remark IV-2-7, there exists a constant matrix r such that 44(t +w) = 4i(t)r
and, consequently, r = 4?(w). Thus,
(IV.4.4) 41(t + w) = 4(t)t(w) for t E R.
Setting B = w-' log(4?(w)J (cf. Example IV-3-6), define an n x n matrix P(t) by
(IV.4.5) P(t) = d'(t) exp(-tBJ.
Then, P(t + w) = ((t + w) exp(-(t + w)BI = 4i(t)0(w) exp(-wB) exp(-tBJ =

4'(t) exp(-tBJ = P(t). This shows that P(t) is periodic in t of period w. Thus, we
proved the following theorem.
Theorem IV-4-1 (G. Floquet [Fl]). Under assumptions (1) and (2), the funda-
mental matrix solution 4i(t) of (IV.4.1) defined by the initial-value problem (IV.4.3)
has the form
(IV.4.6) 4'(t) = P(t) exp[tB],
where P(t) and B are n x n matrices such that
(a) P(t) is invertible, continuous, and periodic of period w in t,
(Q) B is a constant matrix such that 4'(w) = exp[wB].
Observation IV-4-2. As was explained in Observation IV-3-8, letting 4'(w) _
S + N = S(I + M) be the S-N decomposition of 4(w), we define log($(w)) by
log(4?(w)) = log(S) + log(1 + M), where
log(S) = log(A1)P1(4'(w)) + log(a2)P2(4'(w)) + - + log(Ak)Pk('6(w))-
and
n- 1 l)m+l
log(!! + M) = E (- m Mm.
M=1
In the case when A(t) is a real matrix, the unique solution 4 (t) of problem (IV.4.3)
is also a real matrix. Therefore, the entries of 4'(w) are real. Since S and N are real
matrices, the matrix M = S-'N is real. Therefore, log(I + M) is also real. Let
us look at log(S) more closely. If \j is a complex eigenvalue of 4'(w), its complex
conjugate A3 is also an eigenvalue of 4'(w). In this case, set A,+1 = A3. It is easy
to see that the projection Pj+1(4?(w)) is also the complex conjugate of P,(4'(w)).
However, if some eigenvalues of 4'(w) are negative, log[S] is not real. To see this
more clearly, rewrite log[s] in the form
log[S] _ E log[A,]Pi(4?(w)) + E log1A3]P'(4'(w))-
A, <0 other j
The matrix log(A,]P,(4'(w)) is a real matrix, while the matrix

log[aj]Pis
other j
E
A, <0
not real. Therefore, log[S] is not real. To rectify this sit-
uation, let us look at S2. By virtue of the relations given in Lemma IV-1-9, we
obtain
S2 =
(,\J)2P+ _y2P
A, <0 other j
Notice that
log(S2] _ E log[(Ai)2]Pj(4'(w)) + 2 E log[a3]P,(4(w))
af<0 other 7
is a real matrix. Therefore, we can find a real matrix
log[4'(w)2] = log[S2J + 2log(In + MI.
Now, observe that 4'(,)2 =' (2w) and 4'(t + 2w) = 4'(t). (2w). Thus, setting
(IV.4.7) B = _L log{4'(w)2] and P(t) = 4'(t) exp[-tB],
we obtain the following theorem.
4. SYSTEMS WITH PERIODIC COEFFICIENTS 89
Theorem IV-4-3. In the case when the matrix A(t) is real, the matrix 4s(t) has
the form
P(t) exp[tB],
where P(t) and B are n x n real matrices such that
(a) P(t) is invertible, continuous, and periodic of period 2w in t,
(p) B is a constant matrix such that 4?(w)2 = exp[2wB].
dP(t)
Remark IV-4-4. In the case when 4?(t) = P(t) exp[tB], we have dt
A(t)P(t) - P(t)B. Therefore, (IV.4.1) is changed to
-
dz- = Bz
(IV.4.8)
dt
by the transformation y" = P(t)!. As noted above, P(t) is periodic in t of period

w (or, respectively, 2w when A(t) is real), the behavior of solutions of (IV.4.1) is
derived from the behavior of solutions of (IV.4.8).
Definition IV-4-5. The eigenvalues p1.... , p of the matrix B are called the
characteristic exponents of equation (IV.4.1). The eigenvalues pi = exp[wpjl, ... ,
p= of the matrix 4?(w) (or, respectively, when A(t) is real, the eigen-
values pi = exp(2wpi],... , P = exp[2wµn] of 4?(w)2) are called the multipliers of
equation (IV.4.1).
The following basic result is a corollary of Theorem IV-3-1.
Corollary IV-4-6. All solutions of equation (IV.4.1) tend to 0' as t -. +oo if and
only if Ipj I < 1 for all j (or R(pj) < 0 for all j).
Observation IV-4-7. Let be an eigenvector of 45(w) associated with a multiplier
p. Then, the solution d(t) = t(t)p" of (IV.4.1) satisfies the condition ¢(t + w) =
pi(t). The terminology multiplier came from this fact. (In the case when A(t) is
real, choosing an eigenvector p" of 4;(w)2 associated with a multiplier p, we obtain
(t + 2w) = pi(t) if d(t) = 4i(t)p.)
Example IV-4-8. Consider the initial-value problem
dY = A(t)Y,
(IV.4.9) Y(0) = 12,
-it-
where
cost 1
A(t) =
0 cost]
The solution of (IV.4.9) is given by

r I t
41(t) = exp I A(s)dsexp(sin t) t]
l fo L
(cf. (3) of Remark IV-2-7). Therefore,
$(27r) = I 21 B= 2-. log[4i(2a)].

0 1 1
Since p.(\) = (a - 1) 2, Pt(4i) = 12, we obtain

2,r
S = 12i M = [0
20 , and log[4s(27r)j = log[I2 + M) = M = I 0
0
J
This gives B= 1 M= I 0 01. Therefore,
It I = exp(sin t)I2
P(t) = i(t) exp [ 0 Ot, = 4i(t) l
and fi(t) = P(t) exp[tBj = exp(sin t) exp I 0 t ]. Moreover, the characteristic

exponents are 0, 0, and the multipliers of (IV.4.9) are 1, 1. Note that in this case,
as both of two multipliers are positive, it is not necessary to take 4i(2a)2 and period
41r to find B.
IV-5. Linear Harniltonian systems with periodic coefficients

This section is divided into two parts. In Part 1, we explain the structure of
solutions of a linear Hamiltonian system. The basic facts are found in the theory
of real symplectic groups and their Lie algebra. If we denote by J the real (2n) x (2n)
matrix I 0O 1, where In is the n x n identity matrix, the Lie algebra of the

real symplectic group Sp(2n, R) of order 2n is the set of all real 2n x 2n matrices
of the form JH, where H is a real (2n) x (2n) symmetric matrix. A linear system
T = A(t)f on y' E R2" is a linear Hamiltonian system, if there exists a real
(2n) x (2n) symmetric matrix H(t) such that A(t) = JH(t). Consequently, the
fundamental matrix 0(t) of this system belongs to Sp(2n,R) if 0(0) = I. We
explain all these elementary facts of Hamiltonian systems in Part 1. We also prove
that if G E Sp(2n, R) and G = S+N = S(I2"+M) is the S-N decomposition of G,
then S and 12" + M belong to Sp(2n, R). In Part 2, we refine the Floquet theorem
(cf. Theorem IV-4-1) for the linear periodic Hamiltonian systems in such a way
that the periodic transformation y = P(t)i given in Remark IV-4-4 is canonical.
This means that P(t) belongs to Sp(2n, R). The important consequence is that the
dx
linear system = BE of Remark IV-4-4 is a Hamiltonian system.
Part 1
dp b-KH dq
Consider a linear Hamiltonian system
dt
= (t, y')
dt
= -!2c (t, y-), where
8P 8q"
and j are unknown vectors in R", y = [q1, ?{(t, y7 = 2y"TH(t)y with a real
on 8H
ON Bpi an Ni
(2n) x (2n) symmetric matrix H(t), and Here,
8p 8q.
5. LINEAR HAMILTONIAN SYSTEMS 91
yT is the transpose of the column vector Y. Since
8 (t, y-) = [0, In[ H(t)y, (t, y-) = [1n, O[ H(t)y,
the Hamiltonian system can be written in the form
(IV.5.1) = JH(t)y,
where y is a unknown vector in R2n, H(t) is a real (2n) x (2n) symmetric matrix,
and
O In
(IV.5.2) J= [ -In O
The matrix J has the following important properties:
(IV.5.3) JT = -J, J-1 = -J,
where JT is the transpose of J.
Let fi(t) be the unique real (2n) x (2n) matrix such that
do(t)
= JH(t))(t) , 4(0) = 12n.
Lemma IV-5-1. The matrix 4(t) satisfies the condition 4(t)T J4(t) = J, where
t(t)T is the transpose of t(t).
Proof
Differentiate O(t)TJ4(t) to derive
I(t)TH(t)(-J)J4(t) +
a
I(t)T H(t)41(t) - O.
Then, 1(t)TJ4(t) = 4(0)T Jo(0) = J. 0

The converse of Lemma IV-5-1 is shown in the following lemma.
Lemma IV-5-2. Assume that a real (2n) x (2n) matrix G(t) satisfies the following
conditions:
(i) the entries of G(t) are continuously differentiable in t on an interval 1,
(ii) G(t)T JG(t) = J for t E Z.
Then, G(t) is invertible and H(t) _ -Jd d(t)G(t)-I is symmetric for t E I.
Proof.
Computing the determinant of both sides of condition (ii), we obtain [det G(t)12
= 1. Therefore, G(t) is invertible. To show that H(t) is symmetric, differentiating
both sides of G(t)T JG(t) = J, we obtain
[dG(t)]TJG(t) + G(t)T J fi(t) = O.

Since [G(t)-']T = [G(t)T]-1, we further obtain
[G(t)-']T [dG(t)]T(- J) = JdG(t )G(t)_1.
Observing that the left-hand side of this relation is the transpose of the right-hand
side, we conclude that H(t) is symmetric. 0
Observation 1. If GT JG = J, then (G-' )T JG-' = J and GJGT = J. In fact,
GT JG = J implies that (G-' )T JG-' = J. Then, taking the inverse of both sides
of the first relation, we obtain the second relation (cf. (IV.5.3)).
Definition IV-5-3. A (2n) x (2n) real invertible matrix G such that GT JG = J is
called a real symplectic matrix of order 2n. The set of all real symplectic matrices
of order 2n is denoted by Sp(2n, R).
It is easy to show that Sp(2n, R) is a subgroup of GL(2n, R) (cf. Observation 1).
Hence, Sp(2n, R) is called the real symplectic group of order 2n.
Observation 2. If we change Hamiltonian system (IV.5.1) by a linear transfor-
mation
(rV.5.1) becomes
- {P(t)-1JH(t)P(t) - P(t)-1 dt)J i.

dt
Furthermore, if P(t) E Sp(2n, R) for t E Z, this system becomes
(IV.5.5)
d J [p(t)TH(t)P(t) - JdP( 'P(t) i.
Since P-(t)P(t) = 12 and P(t)T E Sp(2n,R), we obtain
I P(t)_1
dP(t)
dt +
dP(t)
di P(t) =dlzd = 0,
1 P(t)-'J = JP(t)T.
dP(t)
t
Observe that -J P(t) is symmetric since P(t) E Sp(2n, R) (cf. Lemma
IV-5-2). This implies that (IV.5.5) is a Hamiltonian system.
Definition IV-5-4. Transformation (IV.5.4) is called a canonical transformation
if P(t) E Sp(2n, R).
Observation 3. For a given constant matrix G in Sp(2n, R), let us construct the
S-N decomposition G = S + N, where S is diagonalizable, N is nilpotent, and
SN = NS. Since G is invertible, S is invertible (cf. Remark IV-1-13). If we set
M = S-' N, then M is also nilpotent, and we derive the unique decomposition
(IV.5.6) G = S(12 + M), SM = MS.
The identity GTJG = J implies JGJ-' = (GT)-l. Using (IV.5.6), we can write
this last relation in the form
(ST)-1 (I2, + MT)-'.
(JSJ-') (J(I2n + M)J-1) =
Set Si = JSJ-1, M1 = J(I2n + M)J-' - 12,,, S2 = (ST)-', and M2 =
(12n + MT) -1 - 12n. It is not difficult to show that S. = 1, 2) are diagonal-
izable and that Af. (j = 1,2) are nilpotent. Furthermore, S,llf, = MSS, (j = 1, 2).
Hence, the uniqueness of S-N decomposition implies that S1 = S2 and 12n + M1 =
12n + M2. Therefore,
(IV.5.7) JSJ-' = (ST)-', J(I2n + Af)J-' = (12n + AfT)-1.
Thus, we proved the following lemma.
Lemma IV-5-5. If we write a symplectic matrix G of order 2n in form (IV.5.6)
by using the unique S-N decomposition of G (where Af = S- IN), the two matrices
S and 12n + M are also symplectic matrices of order 2n.
Remark IV-5-6. Let J be the 2n x 2n matrix defined by (111.5.2) and let H be a
2n x 2n symmetric matrix. Then, it is important to know that if A is an eigenvalue
of JH with multiplicity m, then -A is also an eigenvalue of JH with multiplicity
m. In fact, (JH)T = -HJ = HJ-'. Hence, J-'(JH)TJ = -JH, where AT
denotes the transpose of A.
Remark IV-5-7. Let G be a 2n x 2n symplectic matrix. Then, it is important
to know also that if A is an eigenvalue of G with multiplicity m. then is also an
eigenvalue of G with multiplicity m. In fact, JGJ-' = (CT)-'.
Remark IV-5-8. Assuming that the set of 2n x 2n symplectic matrices with 2n
distinct eigenvalues is dense in the symplectic group Sp(2n, R), it can be shown
that if I (and/or -1) is an eigenvalue of a symplectic matrix G, its multiplicity is
even. Also, det G = 1. In fact, slim, 1. It is also known that det A is equal
to the product of all eigenvalues of A.
Remark IV-5-9. Assuming that Sp(2n, R) is a connected set, we can show that
det G = 1 for all G E Sp(2n, R). In fact, (det G)2 = 1 and the 2n x 2n identity
matrix is in Sp(2n, R).
Part 2
Now, assume that the real symmetric matrix H(t) of system (IV.5.1) is periodic
in t of a positive period w, i.e., H(t + w) = H(t). Then, applying Theorem IV-4-3
and Remark IV-4-4 to system (IV.5.1), we conclude that the unique (2n) x (2n)
matrix solution 0(t) of the initial-value problem d dtt) = JH(t)4y(t), 4i(O) _
12n can be written in the form 0(t) = P(t) exp[tB], where P(t) and B are real
(2n) x (2n) matrices such that P(t) is invertible, P(t+2w) = P(t) for all t E R, and
B is a constant matrix such that 4i(w)2 = exp[2wB]. Furthermore, system (IV.5.1)
is changed to
dzF
(IV.5.8)
dt
= Bi
94 W. GENERAL THEORY OF LINEAR SYSTEMS
by the transformation
(IV.5.9) y = P(t)f.
The main concern of Part 2 is to show that a matrix P(t) can be chosen so
that the transformation (IV.5.9) is canonical. Note that Y = exp[tB] is the unique
solution of the initial-value problem = BY, Y(O) = I2,,. Therefore, if JB is
symmetric, the matrix exp[tB) is symplectic. Hence, if JB is symmetric, the matrix
P(t) = 4i(t) exp[-tB] is symplectic, since 4i(t) is symplectic. This implies that in
order to show that the transformation (IV.5.9) is canonical, it suffices to show that
we can choose P(t) so that JB is symmetric.
Hereafter we use notations and results given in §IV-4. For examples,
k
(1V.5.10)
I2. = E P,('t(w)), 0 (j # t),
,=1
Pj(4'(w))2 = P
Taking the transpose of (IV.5.10), we obtain
k
l2n = yP,(.6(w))T, 0 U 36t),
(IV.5.11) ,=1
\P(.t(w))T )2 = Pj(.0(W))T.
Note that pAr(A) = PA(A) for any square matrix A. Also,
log[4?(w)2] = log[S2] + 2Iog[I2r, + MI,

log[S2] _ log[(,\,)2]Pi(I(w)) + 2 E log[Aj]Pj('(w)),
ai <0 other ,
1 m+l
1og[I2n + M] = E (-)
m=1
Mm,
and
B= log[4i(w)2].
In order to prove that JB is symmetric, it suffices to prove that J log[S2] and

J log[I2i + M] are symmetric. To do this, let us first prove the following lemma.
Lemma IV-5-10. For the symplectic matrix 4i(w), we have
(IV.5.12) Pj(0(w)T)JPe(0(w)) = 0 if A,Ae 34 1,
where Aj (j = 1, 2,... , k) are distinct eigenvalues of 4i(w).

Proof.
Upon applying (IV.5.6) to G = 0(w), write 4'(w) in the form -O(w) = S(I2n+M).
Then, S is symplectic, i.e., ST JS = J = 12,,JI2n (cf. Lemma IV-5-5). Using the
formulas
t=1=1
k k
12. = > Pt(4 (w)), 12n = P
(IV.5.13) k k
S= ST = EA,P,(,p(w))T,
r=1 J=1
rewrite the identity ST JS = J = I21 JI2,, in the following form:

k k
Pi(qD(w))TJPt(.D(w))
(IV.5.14) A31
,,t=1 ,.t=1
Fixing a pair of indices j and t and multiplying both sides of (IV.5.14) by Pj (4ti(w))T
from the left and by Pt(4i(w)) from the right, we derive
A, AtP,(4 (w))T JPt(4 (w)) = P3(4'(w))TJPt(4 (w))
Hence, (IV.5.12) follows. 0

Observation 4. Let us prove that J log[S2] is symmetric. To do this, by using
(IV.5.12) and (IV.5.13), we write J log[S2] and Iog[(S2)T](-J) in the following
form:
I2nJ log[S2] = log[(At)2]Pi(.O(w))TJP

A,ae=1
log[(S2)T](-J)I2n = log[(A,)21
A,A =1
Now, choose 1og[(A, )2] for j = 1, 2,... , k in such a way that
log[(At)2] = - log[(A,)2] if A,A, = 1.
Then,
1og[(S2)Tj(-J) _ log[(A1)2]P,(.t(w))TJPt(+(w))
= J log[S2].
Note that for any square matrix A, we have [log(A)IT = log(AT). Therefore,
[J log(S2)]T = - log[(S2)T]J. Hence, J log[S2] is symmetric.
Observation 5. Let us prove that J log[I2 + M) is symmetric. To do this, we
write first [12n + M]TJ[I2n + M] = J in the form J[I2n + M] = (12n + MT)-'J.
Then, write [12n + MT J -1 in the form [12,, + MT J-1 = 12. + M. Using (IV.5.7), it
is not difficult to show that M is nilpotent and JM = MJ. Hence,
JMm = (Kf)mJ for m= 1,2,....

Therefore,
J 00 (-lmm+l MM _ (-lmm+l
(IV.5.15) km J
=
m=1 m=1 /
and
J 1og[I2,, + MI = (log(12n + 1GIJ) J

= (log [(I2 + MT)-1])
]J.
J = - (log [12" + Mn)
Thus,
[J log(I2 + M)]T = log(I2i + 111T)(-J) = J log(I2n + M).
This proves that J log[I2,, + M] is symmetric. Here, use was made of the fact that
if 1 + y(x), then
1 _+X =
log I log[1 + y(x)J = - log(1 + x]

1 + xJ =
as a power series in x. L
Thus, we arrived at the following conclusion.
Theorem IV-5-11. In the case when the matrix H(t) of system (IV.5.1) is real,
symmetric, and periodic in t of a positive period w, there exist (2n) x (2n) real
matrices P(t) and B such that
(a) P(t) E Sp(2n, R) for all t E R,
(b) P(t+2w)=P(t) for alit ER,
(c) B is a constant matrix such that JB is symmetric,
(d) the canonical transformation y" = P(t): changes system (IV.5. 1) to L = BE.
For more information of exponentials in algebraic matrix groups, see (Mar] and
(Si4].
IV-6. Nonhomogeneous equations

In this section, we explain how to solve an initial-value problem
(IV.6.1) di = A(t)9 + b'(t), y(r) = ij,
where the entries of the n x n matrix A(t) and the C"-valued function b(t) are
continuous on an interval Z = {t : a < t < b} and r E Z, if E C". As it was
6. NONHOMOGENEOUS EQUATIONS 97
explained in §IV-2, let the n x n matrix 4>(t; r) be the unique fundamental matrix
solution of the homogeneous system
dy
(IV.6.2)
dt
= A(t)y
determined by the initial-value problem
dY = A(t)Y,
(IV.6.3) Y(r) = In,
dt
where r E I. System (IV.6.2) is called the associated homogeneous equation of
(IV.6.1). We treat the solution of (IV.6.1) by using the knowledge of the funda-
mental matrix solution 4)(t; r) of (IV.6.2) (cf. §§IV-2 and IV-3).
Consider the transformation
(IV.6.4) y' = 4)(t; r)z.
Then, the problem (IV.6.1) is changed to
(IV.6.5)
dz _
dt
r)z(r) = 4.
In fact, differentiating both sides of (IV.6.4), we obtain
di*
A(t)4)(t; r)l + b(t) = r)z""+ 4 (t; r) .
Since the unique solution of problem (IV.6.5) is given by
z= i + jt(s;rY1(s)ds,
t
(TV.6.6) l
(t; r)r1 +
Jr
r_
y" _ (t;r)Iit + j4(s;T)_1(s)ds]
t $(t; s)b(s)ds.
Here, use was made of the fact that 0(t; s) = 4'(t; r)O(s; r)-1 (cf. (2) of Remark
IV-2-7; in particular (IV.2.6)).
In a similar way, using (IV.6.6), we can change a nonlinear initial-value problem
d = A(t)17 + 9(t y), y(r) = i

to an integral equation
(IV.6.7) y = 4 (t; 7-)17 + t '(t; s)g(s,7(s))ds.

Jr
The function g(t, y-) is the nonlinear term which satisfies some suitable condition(s).
In the case when the matrix A is independent of t, formulas (N.6.6) and (IV.6.7)
become
(IV.6.8) g" = exp[(t - r)A]it + / t exp[(t - s)A]b(s)ds

r
and
(IV.6.9) exp[(t - r)A]>; + exp[(t - s)A]9(s,y1s))ds,

Jr
respectively.
Example IV-6-1. Let us solve the initial-value problem
(IV.6.10) dt = Ay + 6(t), y(0)
where
-2 1 0 [2] #p= 1
A= 0 -2 0 , b(t)= 0 , 1
3 2 1 t 0
The matrix A was given in Example IV-3-3. As computed in §IV-3, a fundamental

matrix solution of the associated homogeneous equation is
e-2t to-21
0
0) = exp(tA) = 0 e-2t 0
t)e-2t
et - e-2t et - (1 + et
Therefore, the solution of (IV.6.10) is
r t 1 1 + to-2t
exp[tA] j i0 + exp[-sA]b(s)ds } e-2t
111 fo 111
-9-t+5et-(l+t)e
IV-7. Higher-order scalar equations

In this section, we explain how to solve the initial-value problem of an n-th order
linear ordinary differential equation
ao(t)ucn) + al(t)u(n-1) + ... + an-1(t)u + an(t)u = b(t),

(IV.7.1) U(n-1)(,r)
u(r) = 'h, u'(r) = Q2, ... , =
where the coefficients a°(t), a1(t), ... , and the nonhomogeneous term b(t) are
continuous on an interval I = {t : a < t < b}. In order to reduce (IV.7.1) to
7. HIGHER-ORDER SCALAR EQUATIONS 99
a system, setting yj = u and 112 = u', ...I yn = u(n-1), we introduce a vector

yl
I. Then, problem (IV.7.1) is equivalent to
tin
d11
(IV.7.2) = A(t)f + b(t),
dt
where
0 1 0 0 ... 0 1
0 0 1 0 ... 0
A(t) =
an0
an(t) an-1(t) -3(t) _a 2(t) al(t)

ao(t) ao(t) do(t) ao(t) ao(t)
0
0
b(t) =
0t)
b(t)
ao(
Assuming ao(t) 54 0 on Z, let 4?(t) = [ 1(t) 2(t) ... hn(t)J be a fundamental

matrix solution of the associated homogeneous equation
d
at = A(t)y
(IV.7.3)
of (IV.7.2). Using 4i(t), we can solve problem (IV.7.2) (cf. §IV-6). The first
component of the solution of (IV.7.2) is the solution of problem (IV.7.1). Also,
the first components of n column vectors of the matrix $(t) give the n linearly
independent solutions of the associated homogeneous equation
(IV.7.4) ao(t)u(n) + al (t)u(n-1)

+... + an(t)u = 0
of (IV.7.1).
Example IV-7-1. Let us solve the initial-value problem
(IV.7.5) u"' - 2u" - 5u' + 6u = 3t, u(0) = 1, u'(0) = 2, u"(0) = 0.
To start with, set
111
111=u, 112=u',
y2=u",
and y= 1122
IY31
Then, problem (IV.7.5) is equivalent to
dy" = Ay" + b'(t),

(IV-7.6) y(0) = #I,
dt
where
10 1 0 0 1
(IV.7.7) A= 0 0 1, b(t) = 0 I, and [2].

-6 5 2 3t 0
Three eigenvalues of the matrix A are 1, -2, and 3. The corresponding projections
are
-6 -1 1 1 3 -4 1
P1(A) _ -6 -6 -1 1 I , P2(A) -6 8 -2
15
-6 -1 1 12 -16 4
1-2 1 1
P3 (A)
= 10
-6 3 3
-18 9 9
and N = 0. Therefore,
4)(t; 0) = exp(tAJ = etP1(A) + e-2tP2(A) + e3tP3(A)

et -6 -1 _2t 3 -4 -2
=-6 -6 -1
-6 -1
1
1
1
e -615
12
8
-16
1
-2 + 10e -6 3 3.
4
3t
-18
1
9 9
1
Thus,
t
4D(s;0)-1 6(s) ds
1 1 - (t + 1)e-t 1 1 + (2t - 1)e2t 1 1 - (3t + 1)e '3t

1-(t+1)e-` +-20-2(1+(2t-1)e2tJ +- 3(1-(3t+1)e-39
,
1)e-t 1)e-3t1
1 - (t + 4[1 + (2t - 1)e2t] 911 - (3t +
and, consequently,
1 1 [1_(t+1)e_tl 1 1 + (2t - 1)e2t

- 1 - (t + 1)e' + - -2(1 + (2t - 1)e2CJ
20
0 1 - (t + 1)e-t
4[1 + (2t - 1)e2L]
- 1 - (3t + 1)e-3t
1 30t + 25 - 17e'2t + 50et + 2e3L
+ 3[1 - (3t + 1)e-3t]
= 2(15 + 17e-2t
+ 25et + 3e3t)
1)e-3t] 60
9(1 - (3t + 1 2(-34e-2t + 25et + 9e3t)
The first component of g(t) gives the solution of (IV.7.5).

7. HIGHER-ORDER SCALAR EQUATIONS 101
Remark IV-7-2. In the case of a second-order linear differential equation
(IV.7.8) ao(t)u" + al (t)u' + a2(t)u = b(t),
a fundamental matrix solution of (IV.7.3) has the form 4P(t) = 0' (t) ¢2(t) ,
.01(t) 02(t)
where 01 (t) and ¢2 (t) are two linearly independent solutions of the associated ho-
1 .02(t)
mogeneous equation (IV.7.4). Since 4i(t)-1 = -02(t) where
W(t) -011(t) 01(t)
W(t) = det(4i(t)), the first component of the formula
y"(t) _ (t}it + -t(t) t 4>(s)-16(s)ds

J
gives the general solution of (IV.7.8) in the form
u(t) = r)1451(t) + 77202(t)

(IV.7.9) 42(s)
- 41(t) It ao(s)W(s)b(s)ds + 02(t) rt 01(s)
ao(s)W(s)b{s)ds,
where ill and q2 are two arbitrary constants. This is known as the formula of
variation of parameters (see, for example, [Rab, pp. 241-2461). Moreover,
(IV.7.10) W(t) =W(r)exp

f do (s)
ds
J
as in (4) of Remark IV-2-7.

Remark N-7-3. Write (IV.7.10) in the form
(IV.7.11) 01(t)02(t) - dl (t)t2(t) = c2 exp [ - t s)ds

L ft ao(s) I
,
where C2 is a constant. Regarding (IV.7.11) as a differential equation for 02(t), w

obtain
1 al(s)
02(t) = C101(t) +C201(t)
J t 01(T )2 [ -J t
ao(s)ds }dry
where cris anoth1er constant. Therefore, 1(t) and ¢1(t)

f 1(r)2
x exp [ - J t _(s)ds] d7- form a fundamental set of solutions of ao(t)u"+al(t)u'+
s
a2(t)u = 0.
Remark IV-7-4. Let 01 (t) and 02(t) be linearly independent solutions of a third-
order linear homogeneous differential equation
(IV.7.12) ao(t)u"' + a1(t)u" + a2(t)u' + a3(t)u = 0.

102 IV. GENERAL THEORY OF LINEAR. SYSTEMS
Also, let 0(t) be any solution of (IV.7.12). Then, (4) of Remark IV-2-7 implies that
(IV.7.13)
0
10'
01
401
02
452 =c3exp
[ft ao(s)ds]
'P 4i oz
where c3 is a constant. Write (IV.7.13) in the form
(IV.7.14) Au(t)o" + A1(t)O' + A2(t)4 = c3exp

[-Ic al(s)ds] ao(s)
,
where
461 02 01 02 0'2
As(t)
COI
A1(t) = 101 ¢2 4 A2(t) - 1,i '02" ' 40 oz
A fundamental set of solutions of the homogeneous part of (IV.7.14) is given by

{01,02}. Therefore, using (IV.7.9), we obtain
0(t) = C101 (t) + c202(t) + C3 I-01(t)
- J ao(s)ds]dr}
I 02
Ao
(T3 eXp I -
J
T ao(s) I dT
+ 02 (t) I ` Ao(r)2 e L
where cl and c2 are constants. This implies that 01(t), 02(t), and
02(r)
Aa(T)2
exp
[f'
i)
ao(s)dsj
1(r)
dr - -02(t) I Ao(r)2 exp [ - f ao(s)
)
ds] dr
form a fundamental set of solutions of (IV.7.12).
EXERCISES IV
IV-1. Let A bean n x n matrix, and let A1, A2, ... , Ak be all the distinct eigenvalues
of A. On the complex A-plane, consider k small closed disks Of = {A : IA - A2 I <
r)} (j = 1,... , k}. Assume that Aj n At = 0 for j # 1. Set
A)-1dA (j = 1,... ,k),

f P] 2'rf d
ail=rj(AI,, -
and N=A-S,
where the integrals are taken in the counterclockwise orientation. Show that
(i) Pl + ... + Pk = In,
(ii) Pi Pt = O if U'41),
(iii) A = S + N is the S-N decomposition of A.
EXERCISES IV 103
Hint. Note that
1 {(u1n - A) -1 - (7-In - A) -1 } = (oln - A) -1 (rln - A) -1 .

r-o
Also, if p(A) is a polynomial in A with constant coefficients, then
k
p(A) = 1j
J=12rri
p(A)(Aln - A)'1dA.
For general information, see [Ka, pp. 38-43).

IV-2. Under the same assumptions as in Exercise IV-1, show that if f (A) is analytic
in a neighborhood of each eigenvalue AJ, and an n x n matrix B is in a sufficiently
small neighborhood of A, then f (B) is well defined and Biim f (B) = f (A).
Hint. If f (A) is analytic in a neighborhood of each eigenvalue A,, then
f (A) _ 1
J=12rri
IV-3. Let A = S + N be the S-N decomposition of an n x n matrix A. Show that,

as functions of A, the matrices S and N are not continuous.
Hint. Use Lemma IV-1-4.
IV-4. Let A be an n x n matrix with n eigenvalues Al, ... , An. Fix a real number
r satisfying the condition r > R[AJ] (j = 1,... , n). Show that
j+00
(F) exp[tA] =
2T
io)In - A)-'do
00
for t E R.
Remark. This result means that exp[tA] is the inverse Laplace transform of (sln -
U1
A) 1. For example, if A = [ ], then (s12 - A) -1 =
s2 + 1
[_i 1 ]. Hence,
taking the inverse Laplace transform, we obtain exp(tA) = cos t sin t

For-
cost I-sint I.
mula (F) is useful in the study of exp[tA] of an unbounded operator A defined on
a Banach space (cf. [HUP, Chapter XII, pp. 356-386]).
IV-5. Show that J(A2 +T 21" )-'dr= A-1 arctan(tA-1) if every eigenvalue of
an n x n matrix A is a nonzero real number.
m
IV-6. For an n x n matrix A, show that
In
lim (In + A ) = eA.
+00 M
IV-7. Find the solution of the following initial-value problem
2 - 2A + A21r = etAe b7(0) = n, V(0) = S,
where A is a constant n x n matrix, {6, ,j,S are three constant vectors in C", and
y E C' is the unknown quantity.
Hint. If we set y = eÀU, the given problem is changed to
Al =E, u"(0) =r1', u(0)
dt2
IV-8. Find the solution of the following initial-value problem
2 +A21 =D, 9(0) = 40 , !Ly (0) = 41

where E C" is an unknown quantity, A is a constant n x n matrix such that
det A = 0, and {rjo, >7j } are two constant vectors in C".
Hint. Note that cos(xA) and sin(xA) are not linearly independent when det A = 0.
sin u
If we define F(u) = then the general solution of the given differential equation
u
is
j(x) = cos(xA)c"1 + xF(xA)c2.
IV-9. Find explicitly a fundamental matrix solution of the system dy = A(t)y

if there exists a constant matrix P E GL(n,C) such that P-"A(t)P is in Jordan
canonical form, i.e.,
P-1A(t)P = diag[A, (t)I1 + N1iA2(t)12 + N2,... ,.1k(t)Ik +Nkj,
where for each j, Ap(t) is a C-valued continuous function on the interval a S t S 6,
I1 is the n, x n j identity matrix, and N, is an nl x n, matrix whose entries are 1
on the superdiagonal and zero everywhere else.
Hint. See [GH].
/r 3 0 -1
IV-10. Find log (I 11
]).cos([ 1 2 -1 j).anii
\` ` 0 1 1
252 498 4134 698 1).

-234 -465 -3885 -656
arctan
15 30 252 42
-10 -20 -166 -25
IV-11. In the case when two invertible matrices A and B commute, find
log(AB) - log(A) - log B
if these three logarithms are defined as in Example IV-3-6.
EXERCISES IV 105
IV-12. Given that

0 0 1 0 yi
0 0 0 1 I/2
A = 2 3 0 0 ' 1/3
4 -2 0 0 1/4
find a nonzero constant vector ii E C4 in such a way that the solution l(t) of the
d9
A9, 9(0)
satisfies the condition lim y"(t) = 0.

t -+oo
IV-13. Assume that A(t), B(t), and F(t) are n x n, m x m, and n x m matrices,
respectively, whose entries are continuous on the interval a < t < b. Let 4i(t) be
an n x n fundamental matrix of - = A(t)4 and W(t) be an in x in fundamental
matrix of Z = B(t)u. Show that the general solution of the differential equation
on an ii x m matrix Y
dY
(E) = A(t)Y - YB(t) F(t),
_WT
is given by
Y(t) = 4;(t) C 4,(t)-1 r`
where C is an arbitrary constant n x m matrix.

IV-14. Given the n x n linear system
dY
= A(t)Y - YB(t),
dt
where A(t) and B(t) are n x n matrices continuous in the interval a < t < b,
(1) show that, if Y(to) -I exists at some point to in the interval a < t < b, then
Y(t)-i exists for all points of the interval a < t < b,
(2) show that Z = Y-1 satisfies the differential equation
dZ
= B(t)Z - ZA(t).
dt
IV-15. Find the multipliers of the periodic system dt = A(t)f, where A(t) _
[cost simt
-sint cost
Hint. A(t) = exp It 101 0]]

l`
IV-16. Let A(t) and B(t) be n x n and n x m matrices whose entries are real-
valued and continuous on the interval 0 < t < +oo. Denote by U the set of all
R'-valued measurable functions u(t) such that ,u(t)j < 1 for 0 < t < +oo. Fix a
vector { E R". Denote by ¢(t, u) the unique R"-valued function which satisfies the
at
initial-value problem = A(t)i + B(t)u(t), i(0) where u E U. Also, set
R = {(t, ¢(t, u)) : 0:5 t < +oo, u E R}. Show that R is closed in Rn+i.
Hint. See ILM2, Theorem 1 of Chapter 2 on pp. 69-72 and Lemma 3A of Appendix
of Chapter 2 on pp. 161-163j.
IV-17. Let u(t) be a real-valued, continuous, and periodic of period w > 0 in t on
R. Also, for every real r, let ¢(t, r) and t1'(t, r) be two solutions of the differential
equation
d2y
(Eq) - u(t) y = 0
dt2
such that
o(r, r) = 1, 0'(r, r) = 0, and 0(7-,T) = 0, t;J'(T,T) = 1,
where the prime denotes d Set
C(r) = 10'(0.'r)
0(0, r) V(0, r) 1
,
i,I (0,r) J
M(r) _ 0(r+w,r) Vl(r+w,r)

0(r+w,r) tG'(r+w,r)
(I) Show that
M(r) = C(r)'1M(O)C(r).
(11) Let A+ and A_ be two eigenvalues of the matrix M(r). Let I K+(r)J and
be eigenvectors of M(r) corresponding to A+ and A_`, respectively.
K _(r)
how that
(i) K±(r) are two periodic solutions of period w of the differential equation
dK
dT
± K2 + u(r) = 0,
(ii) if we set
Qf{t,r) = exp [J t Kt(() d(l
r
these two functions satisfy the differential equation (Eq) and the conditions
13f(t+w,r) = A*O*(t,r).
EXERCISES IV 107
Hint for (I). Set 4'(t,r) _

(t (t' ) I
JJJ -
Then, ((t, r) is a fundamental
matrix solution of the system
and y = 4'(t, r)c is the solution satisfying the initial condition y(r) = c'. Note that
C(r) = 4'(0, r) and M(r) = 4'(r + w, r). Therefore, 4'(t, r) = 4'(t, 0)C(r) and
4'(t + r) = 4'(t, r)M(r). Now, it is not difficult to see 4'(w, r) = 4'(w, 0)C(r) _
4'(0, r)M(r) = C(r)M(r) and 4'(w, 0) = M(O).
Hint for (II). Since eigenvalues of M(r) and M(0) are the same, the eigenvalues of
M(r) is independent of r. Solutions 77+ (t) of (Eq) satisfying the condition Y7+ (t +
w) = A+n+(t) are linearly dependent on each other. Hence, W+(t) = K+(t) is
independent of any particular choice of such solutions q+(t). In particular K+(t +
W) = 17'+ (t + w) = K+(t). Problem (II) claims that the quantity K+(r) can be
17+(t + w)
found by calculating eigenvectors of M(r) corresponding to A+. Furthermore,
A+ = exp I o
o
K+(r)drJJ I I.
The same remark applies to A_. The solutions 0±(x, r)
of (Eq) are called the Block solutions.
IV-18. Show that
(a) A real 2 x 2 matrix A is symplectic (i.e., A E Sp(2, R)) if and only if det A = 1;
(b) the matrix 01 N is symmetric for any 2 x 2 real nilpotent matrix N.
l 0J
Hint for (b). Note that det(etNJ = I for any real 2 x 2 nilpotent matrix N.
IV-19. Let G, H, and J are real (21n ) x (2n) matrices such that G is symplectic,
H is symmetric, and J = I 0n . Show that G-1JHGJ is symmetric.
IV-20. Suppose that the (2n) x (2n) matrix 4'(t) is the unique solution of the
initial-value problem L'P = JH(t)4', 4'(0) = 12n, where J = I 0n n and H(t)
l T J
is symmetric. Set L(t) = 4'(t)-1JH(t)4'(t)J. Show that d4'dt) = JL(t)4'(t)T,
where 4'(t)T is the transpose of 4'(t).
CHAPTER V
SINGULARITIES OF THE FIRST KIND
In this chapter, we consider a system of differential equations
(E) xfy = AX, Y-),
assuming that the entries of the C"-valued function f are convergent power series
in complex variables (x, y-) E Cn+I with coefficients in C, where x is a complex inde-
pendent variable and y E C" is an unknown quantity. The main tool is calculation
with power series in x. In §I-4, using successive approximations, we constructed
power series solutions. However, generally speaking, in order to construct a power
00
series solution y"(x) = E xmd n, this expression is inserted into the given differ-
M=1
ential equation to find relationships among the coefficients d, , and the coefficients
d,,, are calculated by using these relationships. In this stage of the calculation, we
do not pay any attention to the convergence of the series. This process leads us
to the concept of formal power series solutions (cf. §V-1). Having found a formal
power series solution, we estimate Ia',,, i to test its convergence. As the function
x-I f (x, y) is not analytic at x = 0, Theorem 1-4-1 does not apply to system (E).
Furthermore, the existence of formal power series solutions of (E) is not always
guaranteed. Nevertheless, it is known that if a formal power series solution of (E)
exists, then the series is always convergent. This basic result is explained in §V-2
(cf. [CL, Theorem 3.1, pp. 117-119) and [Wasl, Theorem 5.3, pp. 22-251 for the
case of linear differential equations]). In §V-3, we define the S-N decomposition
for a lower block-triangular matrix of infinite order. Using such a matrix, we can
represent a linear differential operator
(LDO) G[Yj = x + f2(x)lj,
where f2(x) is an n x n matrix whose entries are formal power series in x with
coefficients in Cn. In this way, we derive the S-N decomposition of L in §V-4 and a
normal form of L in §V-5 (cf. [HKS]). The S-N decomposition of L was originally
defined in [GerL]. In §V-6, we calculate the normal form of a given operator C by
using a method due to M. Hukuhara (cf. [Sill, §3.9, pp. 85-891). We explain the
classification of singularities of homogeneous linear differential equations in §V-7.
Some basic results concerning linear differential equations given in this chapter are
also found in [CL, Chapter 4].
108
1. FORMAL SOLUTIONS OF AN ALGEBRAIC DE 109
V-1. Formal solutions of an algebraic differential equation

We denote by C[[x]] the set of all formal power series in x with coefficients in
w 00
C. For two formal power series f = E amxm and g = b,,,x', we define
m=0 m=0
f = g by the condition an = b", for all m > 0. Also, the summf + g and the
00 00
product f g are defined by f +g = (am +b,,,)xm and f g = (F am-nbn )x"',
m=0 m=0 n=0
00
respectively. Furthermore, for c E C and f = >2 a,,,xm E C[[x]], we define cf
m=0
00
by cf = >camx'. With these three operations, C[[x]] is a eommutattve algebra
m=0
over C with the identity element given by E bmxm, where bo = 1 and b,,, = 0
m=0
if m > 1. (For commutative algebra, see, for example, [AM].) Also, we define the
derivative of f with respect to r by dx _ E(m+ 1)a,,,+,x' and the integral
m=0
"0 00
rX r
( f (x)dx by f f (x)dx = airi x"' for f = E a,,,xm E C[[xj]. Then, C[[x]]
0 o m=1 m=0
is a commutative differential algebra over C with the identity element. There are
some subalgebras of C[[x]j that are useful in applications. For example, denote by
C{x) the set of all power series in C[]x]] that have nonzero radii of convergence.
Also, denote by C[x] the set of all polynomials in x with coefficients in C . Then,
C{x} is a subalgebra of C[[x]] and C]x] is a subalgebra of C{x}. Consequently, C[x]
is also a subalgebra of C[[x]].
Let F(x, yo, y, , ... , yn) be a polynomial in yo, y,..... yn with coefficients in
C[[x]]. Then, a differential equation
(V.1.1) F ,...dxn
dy
y,,d"y
=0
f
is called an algebrutc differential equation, where y is the unknown quantity and x
is the independent variable. If F ` x, f, ... ,
dxn
= 0 for some f in C[[x]], then
such an f is called a formal solution of equation (V.1.1). In this definition, it is not
necessary to assume that the coefficients of F are in C{x}.
Example V-1-1. To find a formal solution of
(V.1.2) xLY +y-x=0,

set y = Famx'. Then, it follows from (V.1.2) that ao = 0, 2a, = 1, and
m>0
(m + 1)am = 0 ( m > 2 ). Hence, y = 2x is a formal solution of equation (V.1.2).
110 V. SINGULARITIES OF THE FIRST KIND
In general, if f E C{x} is a formal solution of (V.1.1), then the sum of f as a

convergent series is an actual solution of (V.1.1) if all coefficients of the polynomial
FareinC{x}.
Denote by x°C([x]j the set of formal series x°f (x), where f (x) E C([x]], a
is a complex number, and x° = exp[a logx]. For 0 = x°f E x°C[[x]], 0 =
x°g E x°Cj[x]j, and h E C(jx]], define 0 + iP = x(f + g), hO = x°hf, and
xL = ox°f + x°x Then, x°C([xfl is a commutative differential module over
.
the algebra C((x]]. Similarly, let x°C{x} denote the set of convergent series x°f (x),
where f (x) E C{x}. The set x°C{x} is a commutative differential module over
the algebra C{x}. Furthermore, if a is a non-negative integer, then x°C([x]] C
C[[x]]. If F(x, yo,... , y,) is a formal power series in (x, yo,... , yn) and if fo(x) E
xC[[x]j, ... , fn(x) E xC([xj], then F(x, fo(x),... , fn(x)) E C[[xJj. Also, if
F(x, yo,. .. , yn) is a convergent power series in (x, yo,... , yn) and if fa(x) E xC{x},
... , fn(x) E xC{x}, then F(x, fo(x), ... , f,,(x)) E C{x}. Therefore, using the no-
tation x°C[[x]]n to denote the set of all vectors with n entries in x°C[(x]], we can de-
fine a formal solution fi(x) E xC[[x]]n of system (E) by the condition x = Ax, i)
in C[[x]j . (Similarly, we define x°C{x}n.) Also, in the case of a homogeneous sys-
tem of linear differential equations xfy A(x)g, a series d(x) E x°C[[xj]n is said
to be a formal solution if x = A(x)e in x°C[[x]jn. Now, let us prove the following

theorem.
Theorem V-1-2. Suppose that A(x) is an n x n matrix whose entries are formal
power series in x and that A is an eigenvalue of A(O). Assume also that A + k are
not eigenvalues of A(O) for all positive integers k. Then, the differential equation
(V.1.3) x!Ly = A(x)g
has a nontrivial formal solution fi(x) = xa f (x) E x"C[[x]]n

Proof.
00
Insert g = xa E x'nam into (V.1.3). Setting A(x) _ y2 xmAm, where An E
m=0 m=0
Mn(C) and Ao = A(O), we obtain
00 Co
xa (A + m)xm= xa LtAm_i.
Therefore, in order to construct a formal solution, the coefficients am must be
determined by the equations
m-h
,1ao = Aoa"o and (A + m)am = Aodd,n + E Am-hah (m > 1 ).
h=0
1. FORMAL SOLUTIONS OF AN ALGEBRAIC DE 111
Hence, ado must be an eigenvector of Ao associated with the eigenvalue A, whereas

m-h
ii',,, = ((A + m)1. - A0)-1 Am-hah form? 1. 0
h
For an eigenvalue Ao of A(O), let h be the maximum integer such that Ao + h
is also an eigenvalue of A(O). Then, Theorem V-1-2 applies to A = A0 + h. The
convergence of the formal solution ¢(x) of Theorem V-1-2 will be proved in §V-2,
assuming that the entries of the matrix A(x) are in C(x) (cf. Remark V-2-9).
n
Let P = Eah (x)&h (b = x d be a differential operator with the coefficients
h=0 /
ah(x) in C([x]]. Assume that n > 1 and an(x) 54 0. Set
ab
P[x'] = x' fm(s) xm,
m=np
where the coefficients f,,, (s) are polynomials in s and no is a non-negative integer
such that 0. Then, we can prove the following theorem.
Theorem V-1-3.
(i) The degree of f,,,, in s is not greater than n.
(ii) If zeros of fno do not differ by integers, then, for each zero r of f,,, there
exists a formal series x' (x) E xr+IC[[x]] such that P[xr(1 +O(x))] = 0.
Proof
(i) Since 6h(x'] = shx', it is evident that the degree of f,,,, in s is not greater than
n.
+o0
(ii) For a formal power series d(x) = E ckxk, we have
k=1
+oo
P[x'(I + b(x))] = P(x'] +
L. Ck P[x'+k J
k=1
+oo +oo +oo
= x'
{mnof"
(s)xm + > Ckxk
k =1
(frn(s
m=no
+ k)xm/
}
= x' fno(s)x +
+oo
((,fm(s) +
m -no
k=1
Ckfm-k(s + k)) x"`m
J1
xa+no AK(
+ + E Ckfno+m-k(s + k)) x"' .
M=1 k=1 J
In order that y = x''(1 +b(x)) be a formal solution of P[y] = 0, it is necessary and

sufficient that the coefficients cm and r satisfy the equations
m
fno (r) = 0, fno+m (r) + E Ckfno+m-k(r + k) = 0 (m > 1).
k=1
It is assumed that f,,,, (r + m) 54 0 for nonzero integer m if r is a zero of

Therefore, r and c,,, are determined by
m-1
1
(r) = 0 and cm = -7,( ckfn+m-k(r + k) (m > 1).
r + m)
Convergence of the formal solution x'(1+¢(x)) of Theorem V-1-3 will be proved

at the end of §V-7, assuming that the degree of in s is n and the coefficients
ah(x) of the operator P belong to C{x}. The polynomial f.,, (s) is called the indicial
polynomial of the operator P.
Remark V-1-4. Formal solutions of algebraic differential equation (V.1.1) as de-
fined above are not necessarily convergent, even if all coefficients of the polynomial
00
F are in C{x}. For example, y = Y (-1)m(m!)xm+i is a formal solution of the
m=0
2y
differential equation 2d + y - x = 0. Also, the formal solution x' (I + fi(x)) of
2
Theorem V- 1-3 is not necessarily convergent if the degree of f,,. (s) in s is less than
n. In order that f = >a,,,xm be convergent, it is necessary and sufficient that
m=0
la,,, I < KA' for all non-negative integers m, where K and A are non-negative num-
00
bers. Also, it is known that some power series such as (m!)mxm do not satisfy
any algebraic differential equation. The following result'n--R
gives a reasonable neces-
sary condition that a power series be a formal solution of an algebraic differential
equation.
Theorem V-1-5 (E. Maillet [Mail). Let F(x, yo, yi, ... , yn) be a nonzero polyno-
mial in yo, yi, ... , y, with coefficients in C{x}, and let f = E amxm E CI[x]] be
m=0
a formal solution of the differential equation F(x, y, Ly, ... , dny

dxn = 0. Then,
/
there exist non-negative numbers K, p, and A such that
(V.1.4) I am I < K(m!)PA' (m > 0).

We shall return to this result later in §XIII-8.
Remark V-1-6. In various applications, including some problems in analytic num-
ber theory, sharp estimates of lower and upper bounds of coefficients am of a formal
solution f = F,00axm of an algebraic differential equations are very important.
m =9
For those results, details are found, for example, in [Mah], [Pop], [SS1], (SS2], and
[SS3]. The book [GerT] contains many informations concerning upper estimates of
coefficients lam I.
2. CONVERGENCE OF FORMAL SOLUTIONS 113
V-2. Convergence of formal solutions of a system of the first kind

In this section, we prove convergence of formal solutions of a system of differential
equations
(V.2.1) X fy = AX, Y-),
where y E C' is an unknown quantity and the entries of the C"-valued function f
are convergent power series in (x, y7) with coefficients in C. A formal power series
00
(V.2.2) E x' , ,,, E xC([xfl" (c,,, E C" )
rn-t
is a formal solution of system (V.2.1) if
(V.2.3) X = f(x,0)
dx
as a formal power series.

To achieve our main goal, we need some preparations.
Observation V-2-1. In order that a formal power series (V.2.2) satisfy condition
(V.2.3), it is necessary that f (O, 0) = 0. Therefore, write f in the form
f (x, y-) = o(2-) + A(x)g + F, y~'fv(x),

Iel>2
where
(1) p = (pl,... ,p") and the p, are non-negative integers,
(2) jpj = pi + + pn and yam' = yi' yn^, where yi,...are the entries
of g,
(3) fo E xC{x}" and f, E C{x}",

(4) A(x) is an n x n matrix with the entries in C{x}.
Note that
fo(x) = Ax, 6) and A(x) _ Lf(x,0).
Setting
A(x) = F, x'A"s Ao = 0,0) I

M=0 LZ
where the coefficients A", are in M"(C), write condition (V.2.3) in the form
xd'o
dx
= Ao + f (x, ¢) - A0 .
Then,
(V.2.4) "t,, = A0E + rym for m = 1, 2, ... ,

where
(J(x))p
f (x, 0) - AoO = o(x) + IA(x) - Ao1 fi(x) + fp(x)

IpI>2
00
1: xm7m
M=1
and Im E C n. Note that ry'm is determined when c1 i ... , c,,,_ 1 are determined and
that the matrices mI" - A0 are invertible if positive integers m are sufficiently
large. This implies that there exists a positive integers mo such that if 61, ... , C,,,a
are determined, then 4. is uniquely determined for all integers m greater than mo.
Therefore, the system of a finite number of equations
(V.2.5) mEm = A0 + (m = 1,2,... ,mo)
decides whether a formal solution ¢(x) exists. If system (V.2.5) has a solution
{c1 i ... , c,,,a }, those mo constants vectors determine a formal solution (x) uniquely.
Observation V-2-2. Supposing that formal power series (V.2.2) is a formal solu-
N
tion of (V.2.1), set ¢N (x) _ x'6,. Since
m=1
Ax, (x)) - f (x, N(x)) = A(x)(d(x) - N(x)) + F, [d(x)p - N(x)D] ff(x),

Ipl>2
it follows that AX, $(x)) - f (X, dN(x)) E xN+C[Ix]]". Also,

xd¢N(x) = xdd(x) - xdoN(x) E xN+1C((x((n.
dx dx dx
xdON(x)
Hence, AX, 4V W) - dx
E xN+1C{x}". Set
(V.2.6) 9N,0(x) = zddN(x)

dx
Now, by means of the transformation y" = z1+ y5N(x), change system (V.2.1) to the
system
dz
(V.2.7) xdx = JN(x,z
on i E C", where
xdjN(x)
9N(x, l = f (x, z + dN(x)) - dx
= 9N,O(x) + f(x,Z+dN(x))f r- f(x,dN(x))
= 9N,0(x) + A(x)z" + L. [(+ $N(x))" - N(x)'] f,,(x)
InI?2
As in Observation V-2-1, write gN in the form
9N(x, = 9N,O(x) + BN(x)z + L xr 9N,p(x),

Ip1?2
where
(1) 9N,o E xN+1C{x}" and g""N,p E C{x}",
(2) BN(x) is an n x n matrix with the entries in C{x},
(3) the entries of the matrix BN(X) - Ao are contained in xC{x}.
Observation V-2-3. System (V.2.7) has a formal solution
00
(V.2.8) ON(x) = O(x) - ON(x) _ x'"cn, E xN+1C1jx11 n.

m=N+1
The coefficients cm are determined recursively by
me,, = Aocm + '7m for m = N + 1, N + 2, ... ,
where
9N(x,ILN(x)) - AO1GN(x) = f(x,$(x)) - Ao1 N(x) - dx
= f (x, fi(x)) - AOcN(x) - xdVx)

= 9N,O(x) + [BN(x) - Ao1 i'N(x) + E (ZGN(x))p 9N,p(x)
Ip1>2
ou
=E m=N+1
xm-'1'm
Note that the matrices min - Ao are invertible for m = N + 1, N + 2, ... , if N is

sufficiently large.
Observation V-2-4. Suppose that system (V.2.1) has an actual solution q(x)
Taylor expansion ¢(x) _ j

such that the entries of i(x) are analytic at x = 0 and that ij(0) = 0. Then, the
dim
(0) of q(x) at x = 0 is a formal solution of
(V.2.1). Furthermore, is convergent and E xC{x}".

Keeping these observations in mind, let us prove the following theorem.
Theorem V-2-5. Suppose that fo(x) = f (x, 0) E xN+1C{x}n and that the ma-
trices mIn - A0 (m > N + 1) are invertible, where Ao = (0, 0). Then, system
69
(V.2.1) has a unique formal solution
00
(V.2.9) t(x) = E X'n4n E x^'+1C((x}]n.
m=N+1
Furthermore, ¢ E xN+1C{x}n.
Remark V-2-6. Under the assumptions of Theorem V-2-5, system (V.2.1) pos-
sibly has many formal solutions in xC[[x)J". However, Theorem V-2-5 states that
there is only one formal solution in xN+1C[[xJJn
Proof of Theorem V-2-5.
We prove this theorem in six steps.
Step 1. Using the argument of Observation V-2-1, we can prove the existence and
uniqueness of formal solution (V.2.9). In fact,
f (x, 4) - Ao¢ = o(x) + JA(x) - Ao}fi(x) + E (_)P_)

IpI?2
00
xmrym.
m=N+1
This implies that rym = 0 for m = 1, 2, ... , N. Hence, (m > N + 1) are

uniquely determined by
mc,,, = Ao6,,, + rym for m=N+1,N+2,....

Step 2. Suppose that system (V.2.1) has an actual solution q(x) satisfying the
following conditions:
(i) the entries of rl(x) are analytic at x = 0,
(ii) there exist two positive numbers K and 6 such that
jy(x)I < KIxjN+1 for IxI < 6.
00
xm
Then, the Taylor expansion E ml a.,.7m'1
(0) of #(x) at x = 0 is a formal solution
m=N+1
of (V.2.1) (cf. Observation V-2-4). Since such a formal solution is unique, it follows
that $(x) = E -
xm d1ij
(0). Because the Taylor expansion of if(x) at x = 0 is
m=N+1
convergent, the formal solution ¢ is convergent and E xN+1C{x}".
Step 3. Hereafter, we shall construct an actual solution f(x) of (V.2.1) that sat-
isfies conditions (i) and (ii) of Step 2. To do this, first notice that there exist three
positive numbers H, b, and p such that
(I) HIxIN+1 IxI < 5
If(x,o)I 5 for
and
(II) If(x,91) - f(x,y2)I 5 (IAoI + 1) 1111 - y2I

for Ix) < b and Iy, I <p (j = 1, 2).
Hence,
(III) If(x, y-)i <_ HIxl "+1 + (IAoI + 1)Iyl for Ixl < b and Iyl <_ p.
Using the transformation of Observation V-2-2, N can be made as large as we

want without changing the matrix A0. Hence, assume without loss of any generality
that
IAoi + 1 1
(V.2.10)
N+1 < 2
Also, fix two positive numbers K and b so that
H + NAol 1]K
(V.2.11) K> i and K6N+1 < p
Step 4. Change system (V.2.1) to an integral equation
:
(V.2.12)
J
f n+(f ))
Define successive approximations
)b(x) = 0 and 1?k+1(x) = IZ for k=0,1,2,....

Now, we shall show that
lnk(x)I < KIxIN+1 for jxl <b and k=0,1,2....
Since this is true for k = 0, we show this recursively with respect to k as follows.
First if this is true for k, then
I i k(x)1 < Kjxl N+l < K&N+I < p for jxj < b.
Hence,
I'.F(c nk (o) < {H + IIAol + 1jK}j£VN for 6.
Therefore,
H + IAoI 1]K
II7k+1(x)I < IxIN+I < K 1xIN+1 for jxj < S.
Step 5. Set 1
++1k(x)I
I1'Rk+1 -- nkll = maxiInk+iIZI InI < b}.
Then, since
I ik+1(x) - ik(x)I = fo
we obtain
j
IInk+1 - llkII N ++1 171k - '1k-11I 2lI'7k - 1k-l II
This implies that
lim ilk(x) _ #e+1(x) - #((X)
k-.+00 xN+1 xN+1
e=0
exists uniformly for Ixl < J.
Step 6. Setting
#(x) = xN+1 ("liM xJ`'+1) = ku> nik(x),
it is easy to show that n'(x) satisfies integral equation (V.2.12). It is also evident
that i7(x) is analytic for IxI < 6. Thus, the proof of Theorem V-2-5 is completed. 0
Now, finally, by using the argument given in Observations V-2-2 and V-2-3, we
obtain the following theorem.
Theorem V-2-7. Every formal solution E xC[[zjJ" of system (V.2.1) is conver-
gent, i.e., E xC{x}".
Remark V-2-8. In general, (V.2.1) may not have any formal solutions. However,
Theorem V-2-7 states that if (V.2.1) has formal solutions, then every formal solution
is convergent.
Remark V-2-9. If yi(x) = xA f (x) E XAC[[xfj" is a formal solution of a linear
system X!LY = A(x)y, the formal power series f is a formal solution of xd _
[A(x) - A!, }i. Therefore, f E C{x}" by virtue of Theorem V-2-7. This proves
convergence of the formal solution constructed in Theorem V-1-2.
V-3. The S-N decomposition of a matrix of infinite order

In §V-4, we shall define the S-N decomposition of a linear differential operator.
As a linear differential operator will be represented by a lower block-triangular
matrix of infinite order, we derive, in this section, the S-N decomposition of such
a matrix
All 0 0 ... ... ...
A21 A22 0 ... ... ...
Amt Am2 ... Am,n 0 . .
where for each (j, k), the quantity AJk is an n, x nk constant matrix. Set
Al = All,
All 0 O .. 0
A21 A22 O .. 0
Am = (m > 2).
Am l Am2 ... ... Am,
Form > 2, we write Am - Am-! 0 Set Nm = m nt. Then, A. is an

[ B. Amm J l-1
N", x Nm matrix, while Bm is an nm x N,"_ 1 matrix. Also, let Amm = Smm +Nmm
3. THE S-N DECOMPOSITION OF A MATRIX OF INFINITE ORDER 119
and Am = Sm + Aim be the S-N decompositions of Amm and Am, respectively,

where Smm is an nm x nm diagonalizable matrix, Sm is an Nm x Nm diagonalizable
matrix, Nmm is an nm x nm nilpotent matrix, Aim is an N. x Nn nilpotent matrix,
SmmNmm = NmmSmm, and SmAm = NmSm. The following lemma shows how
the two matrices Sm and Aim look.
Lemma V-3-1. The matrices Sm and Aim have the following forms:
S1 = s11, S"'
=
rI Sm_1 0l (m > 2),
J L Cm Smm
[Aim_i
N, = Ni1, Aim = (m > 2),
fm O 1
Nmm
where Cm and fm are am x N,_1 matrices.
Proof
Consider the case m > 2. Since the matrices Sm and Aim are polynomials in Am
with constant coefficients, it follows that
Sm = [B m1 and Aim = I
fm1 01,
ILM0
where 13m-1 and Dm-1 are Nm-1 x Nm_1 matrices, Cm and fm are nm x Nm_1
matrices, and µm and vm are nm x nm matrices. Furthermore, Dm_1 and Vm are
nilpotent. Also, Bm-1Dm-1 = Dm-1Bm-i and tlmvm = 1.m µm. Hence, it suffices
to show that Bm-1 and µm are diagonalizable.
Note that since Sm is diagonalizable, Sm has Nm linearly independent eigen
vectors. An eigenvector of Sm has one of two forms k and [?.] , where p is an
PJ
eigenvector of Bm_1i whereas q is an eigenvector of µm. Therefore, if we count
those independent eigenvectors, it can be shown that Bm-1 has N,, -I linearly in-
dependent eigenvectors, while Pm has nm linearly independent eigenvectors. Note
that Nm = Nm_ 1 + nm. This completes the proof of Lemma V-3-1. 0
Lemma V-3-1 implies that the matrices Sm and Nn have the following forms:
S1 = SI1,
S11 0 0 ... 0
C21 S22 0 ... 0
Sm = (m>2)
Cmi Cm2 ... ... Smm
and
N1 = N11,
Nil 0 D ... 0
Y21 JN22 0 0
Aim = (m>2),
Fm 1 Fm2 ... ... Nmm
where C,k and Fjk are n3 x nk matrices.

Set
S11 0 0
C21 S22 0
S =
Cml Cm2 "' Smm 0
N11 0 0
f f21 N22 0 "'
N=
1m 1 Fm2 ... .Wmm 0
Then,
(V.3.1) A=S+N and SN=NS.

We call (V.3.1) the S-N decomposition of the matrix A.
V-4. The S-N decomposition of a differential operator

Consider a differential operator
(V.4.1) C(yl = x + f2(x)Y,
w
where f2(x) _ > xt1 and 1 E Let us identify a formal power series
t=o
ao
a1
xt dt with the vector p = a2 E C. Then, the operator C is represented
t>o
by the matrix
flo 0 0 0 ...
f2l In + f2o 0 0 ...
f12 l1 21, + f2o 0 ...
A =
11m f4.-l fZn-2 ... f21 mI, + fZa 0 .
5. A NORMAL FORM OF A DIFFERENTIAL OPERATOR 121
where I is the n x n identity matrix. Let A = S + N be the S-N decomposition
of A. Since A =
L
j) 11
A1, where (i = R2 and I,> is the oo x oo identity
matrix, the matrices S and N have the forms

So 0 0 O ... ...
al I. + So 0 0 ... ...
a2 al 21n + So 0 ... ...
(V.4.2) S=
am am-l am-2 at min + SO 0
and
No O O
V1 No 0
V2 V1 No
(V.4.3) N=
Vm Vm-1 Vm-2 ... VI No 0 ...
respectively, where the a. and v, are n x n matrices and S2o = So +No is the S-N
decomposition of the matrix S2o.
00 00
Set a(x) = So + E xtat and v(x) = No + trot. Then, S represents a

t_1 tvl
differential operator Lo[y1 = x + a(x)y, while N represents multiplication by

v(x) (i.e., the operator: y" -+ v(x)y-). Since A = S + N and SN = NS, it follows
that
(V.4.4) L[gf = Lo[Yj + v(x)y" and Lo[v(x)y1 = v(x)Lo[yl-

We call (V.4.4) the S-N decomposition of operator (V.4.1). We shall show in th
next section that Lo[yj is diagonalizable and v(x)n = O.
V-5. A normal form of a differential operator

Let us again consider the differential operator
(V.5.1) L[yj = X f + Q(x)U,

00
where f2(x) = Ex1fle and Sgt E WC). In §V-4, we derived the S-N decomposi-
t=o
tion
(V.5.2) L[y-] = Lo[yl + v(x)y" and Co[v(x)yl = v(x)Lo[yj.
where
(V.5.3) Co[YI = x + a(x)y
and
Co
a(x) = So + 00 xtat and v(x) = No + xtvt

t=1 t_1
(cf. (V.4.4)). Notice that N = So + No is the S-N decomposition of S2o and

that the operator Go[ul and multiplication by v(x) are represented respectively by
matrices (V.4.2) and (V.4.3).
Set
0 0 0
I. +,% 0 0
Sm = al 21 + So 0
am_ 1 am-2
Since So is diagonalizable, there exist an invertible n x n matrix P0 and a diagonal
matrix Ao such that
SoPo = PoAo, NO = diag[al, A2, ... , . \n)-
Note that A, (j = 1, 2,... , n) are eigenvalues of So. Hence, A. (j = 1, 2,... , n) are

also eigenvalues of flo. For every positive integer m, Sm is diagonalizable. Hence,
Po
Pi
there exists an (m + 1)n x n matrix P.. = such that SmPm = PmA0. If
Pm
m is sufficiently large, we can further determine matrices Pt for t > m + 1 by the
equations
t
(V.5.4) (tl + S0)Pt + E ahP1_h = PtAo.
h=1
Equation (V.5.4) can be solved with respect to Pt, since the linear operator Pt -+
(tl + So)P1 - PtAo is invertible if m is sufficiently large. In this way, we can find
A
P 1
an oo x n matrix P = I such that SP = PAa. Set P(x) = >x'P,. Then,

Pm t=D
entries of P(x)-1 are also formal power series in x and
(V.5.5) Co[P(x)j = P(x)Ao.

Define two differential operators and an n x n matrix by
X [U1 = P(x)-iC(P(x)uZ, /Co(u1 = P(x)-'Co(P(x)Uj,

{ vo(x) = P(x)-'v(x)P(x),
respectively. Then 1C(u'j = X 0(u'] + vo(x)u'. Observe that
.- - _. _.. _, f _. , l du' \ _ . _. 1
(V.5.6)
=x du +Aou".
This shows that the operator Co(y1 is diagonalizable. Furthermore,
dvo P(x)-'Lo(P(x)vo(x)1
dxx) + Aovo(x) = = P(x)`Co[v(x)P(x)1
(V.5.7) x
= P(x)-'v(x)Lo(P(x)1 = vo(x)P(x)-'Co(P(x)1 = vo(x)Ao
This shows that the entry i,k(x) on the j-th row and k-th column of the matrix
vo(x) must have the form v,k(x) = 7jkxAk-a', where ryjk is a constant. Since Lt(x)
is a formal power series in x, it follows that
(V.5.8) vjk(x) = 0 if Ak - )j is not a non-negative integer.
Observe further that the matrix &. (x) can be written in the form
'711 'Yin
vo(x) = x-^°rx^°, where r=
'7ni Inn
Hence, for any non-negative integer p, we have vo(x)P = x-A0rPxAo, where
zÔ = exp ((log x)Ao] = diag(za' , Z112'... , za^ I.
On the other hand, since No is nilpotent, vo(x)P can be written in a form i (x)P =
X'pQP(x), where mp is a non-negative integer such that lim mP = +oo and
P+oo
the entries of the matrix Qp(x) are power series in x with constant coefficients.
Therefore, L0(x)P = 0 if p is sufficiently large. This implies that the matrix r is
nilpotent. Since v(x) = P(x)vo(x)P(x)-1, we obtain v(x)^ = O.
Thus, we arrive at the following conclusion.
Theorem V-5-1. For a given differential operator (V.5.1), let 0o = So +No be
the S-N decomposition of the matrix N. Then, there exists an n x n matrix P(x)
such that
(1) the entries of P(x) are formal power series in x with constant coefficients,
(2) P(O) is invertible and SoP(0) = P(0)Ao, where A0 is a diagonal matrix whose
diagonal entries are eigenvalues of Q0r
(5) the transformation

(V.5.9) P(x)u
changes the differential operator (V.5.1) to another differential operator
(V.5.10) P(x)-1G(P(x)ui = KOK + vo(x)u,

where
dil
Ko(61 = x + Aou, vo(x) = x-Ôrx^0,
and
r=
Ynl Inn
with constants "ljk such that
(V.5.11) Vj k = 0 if Ak - Aj is not a non-negative integer.
Furthermore, the matrix r as nilpotent.

Remark V-5-2. It is easily verified that the matrix P(x) is a formal solution of
the system
xd) = P(x)(Ao + vo(x)) - Sl(x)P(x).
Since the entries of vo(x) are polynomials in x, the power series P(x) is convergent
if f2(x) is convergent (cf. Theorem V-2-7). Therefore, in such a case, v(x) is
convergent and, hence, o(x) is convergent.
Observation V-5-3. Choose integers l1, ... , to so that A.,+f j = Ak +fk if Aj -Ak
is an integer. Then, vo(x) = x4rx-L, where L = diag(e1, f2, ... , en]. If we set
7t{V1 = x'! 1C(xf and ho{v") = x-LICo(x' i , it follows that
dv'
?{o(vl = x'L fXLTf + zLLv + Aoz' 7} = xdx + (Ao + L)v
Hence,
(V.5.12) R{v = x + (Ao + L + r)v.

Note that
Ao + L = diag(A1 + P1, A2 + e2, ... , An + enj
and
(V.5.13) (Ao + L)r = r(A0 + L).

It was already shower in §V-4 that r is nilpotent. Thus, the following theorem is
proved.
Theorem V-5-4. The transformation y = P(x)xLiT changes the system
(V.5.14) G[yi = xdy + fl(x)yl = 0
to
d1U
(V.5.15) VV! = x + (Ao + L + r)u = 0.
Observation V-5-5. The matrix

n-1 h
(-1)h[logx[ rh
o{x) = x-Ao-L-r = x-Ao-L I +
h!
h-1
is a fundamental matrix solution of system (V.5.15). Note that if (A3+tj)-(Ak+tk)

is an integer, then A, + ej = Ak + tk in Ao + L.
Remark V-5-6. A fundamental matrix solution 4D(x) of system (V.5.14) is given
by 4>(x) = P(x)x4o(x), which can be written in the form
n-1
45(x) = P(x)xLx-M-L In +
IL
F {-1)h (lo x h rh
h=1
h!
Since L - A0 - L = -A0, the matrix 44(x) can be also written in the form
n-1 h
4i(x) = P(x)x-ne
In + (-1)h (lohx) rh
h=1
I n-i h
However, x-Ao and I nr >(-1)h [lohx[ rh do not commute, whereas x-A°-L
h=1
n-I [log x[ h
and In + E(-1)h h!
rh commute.
h=1 I
Remark V-5-7. The methods used in §§V-3, V-4, and V-5 are based on the
original idea given in [GerL[.
Example V-5-8.
In order to illustrate the results of this section, consider the differential equation
of the Bessel functions
(V.5.16) + (z2 - a2)y = 0,

za (zT)
where a is a non-negative integer. If we change the independent variable z by
x = z2, (V.5.16) becomes
d 2y=0.
x lxdx
This equation is equivalent to the system
(V.5.17) x + Q(x)y = 0, y=
0 -1 0 -1 0 0
where 11(x) = x - a2 =10+x11i c o = a2 and 0i = 1
0 0
4 4 4 0
To begin with, let us remark that the S-N decomposition of the matrix 11 is
given by S2o = So + No, where
O if a = 0, N if a = 0,
and
80 =
{Q if a > 0, N0 - to if a>0.
a 0
Also, set Ao = 2 a Note that two eigenvalues of Q are

0
2
Now, we calculate in the following steps:
Step 1. Fix a non-negative integer m so that m > 2 -

(_a) = a.
Step 2. Find three 2(m + 1) x 2(m + 1) matrices A,,,, Sm, and Nm (cf. §§V-4 and
V-5).
Po
P1
Step 3. Find a 2(m + 1) x 2 matrix Pm = , where the Pr are 2 x 2 matrices
Pm
such that SmPm = Pn&.
m in
Step 4. Find two 2 x 2 matrices Mm(x) = No + >2xtvt and Qm(x) _ >xtPt
t=1 t-o
(cf. §§V-4 and V-5).
Step 5. We must obtain Qm(x)-1Mm(x)Qm(x) = vo(x) +O(xm+i), where
Po InoP0 if a = 0,
vo(x) r0 a(a)xn
if a>0,
0 0 J
where a(a) is a real constant depending on a (cf. Theorem V-5-1).

After these calculations have been completed, we come to the following conclu-
sion.
+00
Conclusion V-5-9. There exists a unique 2 x 2 matrix P(x) = >xtPe such that
1=o
(1) the matrices Pl for t = 0,1, ... , m are given in Step 8,
(2) the power series P(x) converges for every x,
(8) the transformation y = P(x)u" changes (V.5.17) to
dii
(V.5.18) xaj + (Ao + vo(x)) i = 0'.
We illustrate the scheme given above in the case when a = 0. First we fix m = 2.
Then,
0 -1 0 0 0 0 0 0 0 0 00
0 0 0 0 0 0 0 0 0 0 00
1 1
0 0 1 -1 0 0 1 0 0 0
4 2
A2 = S2 = 1 1
1 0 0 1 0 0 0 1 0 0
4 4 4
0 0 0 0 2 -1 3 3 1 1
2 0
32 32 4 2
0 0 0 0 2 - 1 3 1 1
0 2
4 16 32 4 4
and
0 -1 0 0 0 0
0 0 0 0 0 0
1 1
0 -1 0 0
4 2
N2= 1
0 0 0 0 0
3 3 1 1
0 -1
32 32 4 2
0 0 0
16 T2- 9
1192 -320
64 -128
Calculating eigenvectors of S2, we find a 6 x 2 matrix P2 = -16 16
such
-32 48
0 1
1 0
that S2P2 = 0. Note that, in this case, Ao = 0.
Set
1 1 3 3
4 2 32 32
0 1 1 3
4 16 32-
M2 (x) = 1 o + xvl + x2v2
and
192 -320 P1 -16 16 ] [0 1

P2
(V.5.19)
PO = [ 64 -1 8] ' = [ -32 48 , = 01 '
Q2(r) = P° + xPl + x2P2.
]
Then, Qz-' 11f2(x)Q2(x)
-1 2
+ O(x3). Note that
1 5
1,° 12 634 [192 [--21 4
[0 64 -128] = 2].
01]
64 64
Thus, we arrived at the following conclusion.

00
Conclusion V-5-10. There exists a unique 2 x 2 matrix P(x) Zx'Pr such
e=o
that
(1) the matrices Pt for i = 0, 1, 2 are given by (V.5.19).
(2) the power series P(x) conueryes for every x,
(3) the transformation y" = P(x)u" changes the system
0 -1
dy
(V.5.20) xdx + X 0 Y = 0'
4
to
d6
(V.5.21) x3i + [_1 2, u" = 0.
-
Furthermore, the transformation y" = P(x)Po iv changes (V.5.20) to
(V.5.22) x + 10 0
" = 0.
In the case when a=1, wefixm=2. Then

1 0 -1 0 0 0 0 0 -1 0 0 0 0
1
0 0 0 0 0 0 0 0 0 0
4
1
0 0 1 -1 0 0 1 -1 0 0
8 4
A2 = 132= 3
0 -4 1 0 0
16
1
8 4
1 1
0 0
0 0 0 0 2 -1 1 1 1 1
2 -1
16 16 4
0 0 0 -4 2 J
3 1 3 1 1 2
64 16 16 8 4
and 4
0 0 0 0 0 01
0 0 0 0 0 0
1 1
0 0 0 0
8 4
Ar2 = 1 1
8 0 0 0 0
16
- 1 1 1 1
00
16 16 8 4
-3 1 1 1
0 0
64 16 16 8
384 32
192 -16
Calculating eigenvectors of S2, we find a 6 x 2 matrix P2 = _72
-124 such
2 0
5 1
1
0
that S2P2 = P2Ao. Note that, in this case, A0 = 1 2 1 . Set
1
L0 2
1 1 1 1
8 4 16 16
vl = 1 1 v2 = 1
3
16 8 64 16
M2(x) = xl/1 + x2v2,
and
Po
84 62 -48 -12
1'2 = [2 1j
(V.5.23) [19 3 J' Pl - [-72 -14] , ,
Q2(x) = Po + xP1 + T2 p2.

x
0
Then, Q2-'M2(X)Q2(X) = 48 + Q(x3). Thus, we arrived at the following
0 0
conclusion.
+oo
Conclusion V-5-11. There exists a unique 2 x 2 matrix P(x) = >xtPt such
e=o
that
(1) the matrices Pt for P = 0, 1, 2 are given by (V.5.23),
(2) the power series P(x) convet es for every x,
(3) the transformation y = P(x)u changes the system
-1
(V.5.24) xdx + V = 0
0
to
(V.5.25) x- +
2 48I u = U.
0 1
2
For a further discussion, see IHKS[. A computer might help the reader to calculate
S2i N2, and P2 in the cases when a = 0 and a = 1. Such a calculation is not difficult
in these cases since eigenvalues of A2 are found easily (cf. §IV-1).
V-6. Calculation of the normal form of a differential operator

In this section, we present another proof of Theorem V-5-1. The main idea is to
construct a power series P(x) as a formal solution of the system
dP(x)
= P(x)(Ao + vo(x)) - Q(x)P(x)
(cf. Remark V-5-2).

Another proof of Theorem V-5-1.
To simplify the presentation, we assume that So = A0 = diag[,ulI,, µ2I2, ... ,
Aklk[, where pi, p2, ... , µk are distinct eigenvalues of no with multiplicities m1,
respectively, and the matrix I, is m, x m j identity matrix. Since AOA(o = NoAo,
the matrix No must have the form No = diag(N01,No2, ... ,Nok[, Where Nor is an
00
mi x m1 nilpotent matrix. Let us determine two matrices P(x) = I,a + E xmP,,,
M=1
6. CALCULATION OF THE NORMAL FORM 131
00
and B(x) = Sto + E xmBm by the equation
m=1
xad (in+mm) (Oo

+ *n=1 xmBm)
Cep + m=1 (In + m 'nPm)
This equation is equivalent to

m-1
mPm = PmS2o - S20Pm + Bm - 1m + ( PhBm-h - fl.-hPh) (m > 1).
[=1
Therefore, it suffices to solve the equation
(V.6.1) mX + (A0 +N4)X - X(Ao+No) - Y = H,

where X and Y are n x n unknown matrices, whereas the matrix H is given. If we
write X, Y, and H in the block-form
X11 ... Xlk [Y11 ... Xlk H11 ... Xlk
Xkl Xkk Ykl ... Ykk Hkl ... Hkk
where XXh, YYh, and H,h are m, x mh matrices, equation (V.6.1) becomes
(m + P) - µh)X)h + NO,Xjh - X,hNOh - Yjh = Hjh,

where j, h = 1,... , k. We can determine XJh and Y,h by setting Y,;h = 0 if
m + u, - Ah 4 0, and X,,h = 0 if m + p, - µh = 0. More precisely speaking, if
m + p - Ph 96 0, we determine X3h uniquely by solving
(m + p, - µh)Xlh + NojXjh - X3hNoh = H)h.

If M + Aj - Ah = 0, we set Y,h = H,h. In this way, we can determine P(z) and
B(x). In particular, go + B(x) has the form x ^°I'xA0, where r is a constant
d
n x n nilpotent matrix. Furthermore, the operator x +A0 and the multiplication
ds
operator by No + B(x) commute. D
The idea of this proof is due to M. Hukuhara (cf. [Si17, §3.9, pp. 85-891).
Remark V-6-1. In the case of a second-order linear homogeneous differential
equation at a regular singular point x = a, there exists a solution of the form
+00
01(z) = (x - a)aE c, (x - a)n, where the coefficients c,, are constants, co 0 0,
n=O
and the power series is convergent. If there is no other linearly independent solu-
tion of this form, a second solution can be constructed by using the idea explained
in Remark IV-7-3. This second solution contains a logarithmic term. Similarly,
a third-order linear homogeneous differential equation has a solution of the form
+00
01(x) _ (x - a)° c, (x - a)" at a regular singular point x = a. Using this solu-
n=0
tion, the given equation can be reduced to a second-order equation. In particular, if
there exists another solution ¢(x) of this kind such that ¢1 and ¢2 are linearly in-
dependent, then the idea given in Remark IV-7-4 can be used to find a fundamental
set of solutions.
In general, a fundamental matrix solution of system (V.5.14) can be constructed
if the transformation y" = P(x)xLV of Theorem V-5-4 is found. In fact, if the
definition of fo(x) of Observation V-5-5 is used, P(x)xL2-no-L-r is a fundamental
matrix solution of (V.5.14). The matrix P(x) can be calculated by using the method
of Hukuhara, which was explained earlier.
V-7. Classification of singularities of homogeneous linear systems

In this chapter, so far we have studied a system
(V.7.1) x dt = A(x)y, y E C",

where the entries of the n x n matrix A(r) are convergent power series in x with
complex coefficients. In this case, the singularity at x = 0 is said to be of the first
kind. If a system has the form
(V.7.2) xk+1 = A(x)y", y" E C",
where k is a positive integer and the entries of the n x n matrix A(x) are convergent
power series in x with complex coefficients, then the singularity at x = 0 is said to
be of the second kind
In §V-5, we proved the following theorem (cf. Theorem V-5-4).
Theorem V-7-1. For system (V.7.1), there exist a constant n x n matrix A0 and
an n x n invertible matrix P(x) such that
(i) the entries of P(x) and P(x)-1 are analytic and single-valued in a domain
0 < jxi < r and have, at worst, a pole at x = 0, where r is a positive number,
(ii) the transformation
(V.7.3) y' = P(x)i7
changes (V.7.1) to a system
(V.7.4) du = Aou".
Theorem V-7-1 can be generalized to system (V.7.2) as follows.

7. CLASSIFICATION OF SINGULARITIES OF LINEAR SYSTEMS 133
Theorem V-7-2. For system (V.7.2), there exist a constant n x n matrix A0 and
an n x n invertible matrix P(x) such that
(i) the entries of P(x) and P(x)-1 are analytic and single-valued in a domain
V = {x : 0 < lxi < r}, where r is a positive number,
(V.7.5) y = P(x)i
changes (V.7.2) to (V.7.4).
Proof.
Let 4i(x) be a fundamental matrix of (V.7.2) in D. Since A(x) is analytic and
single-valued in D, fi(x) _ I (xe2"t) is also a fundamental matrix of (V.7.2). There-
fore, there exists an invertible constant matrix C such that 46(x) = 4i(x)C (cf. (1)
of Remark IV-2-7). Choose a constant matrix Ao so that C. = exp[2rriAo] (cf. Ex-
ample IV-3-6) and let P(x) = 4(x)exp[-(logx)Ao]. Then, P(x) and P(x)-1 are
analytic and single-valued in D. Furthermore,
dP(x) = dam)
exp(_(logx)Ao] - P(x)(x-'Ao)
= x-(k+1)A(x)P(x) - P(x)(x-'Ao).
This completes the proof of the theorem. 0

An important difference between Theorems V-7-1 and V-7-2 is the fact that the
matrix P(x) in Theorem V-7-2 possibly has an essential singularity at x = 0.
The proof of Theorem V-7-2 immediately suggests that Theorem V-7-2 can be
extended to a system
dg
(V.7.6) = F(x)y,
dx
where every entry of the n x n matrix F(x) is analytic and single-valued on the
domain V even if such an entry of F(x) possibly has an essential singularity at
x = 0. More precisely speaking, for system (V.7.6), there exist a constant n x n
matrix Ao and an n x n invertible matrix P(x) satisfying conditions (i) and (ii) of
Theorem V-7-2 such that transformation (V.7.5) changes (V.7.6) to (V.7.4).
Now, we state a definition of regular singularity of (V.7.6) at x = 0 as follows.
Definition V-7-3. Let P(x) be a matrix satisfying conditions (i) and (ii) such
that transformation (V.7.5) changes (V.7.6) to (V.7.4). Then, the singularity of
(V.7.6) at x = 0 is said to be regular if every entry of P(x) has, at worst, a pole
atx=0.
Remark V-7-4. Theorem V-7-1 implies that a singularity of the first kind is a
regular singularity. The converse is not true. However, it can be proved easily that
a regular singularity is, at worst, a singularity of the second kind. Furthermore, if
(V.7.2) has a regular singularity at x = 0, then the matrix A(0) is nilpotent. This
is a consequence of the following theorem.
Theorem V-7-5. Let A(x) and B(x) be two n x n matrices whose entries are
formal power series in x with constant coefficients. Also, let r and s be two positive
integers. Suppose that there exists an n x n matrix P(x) such that
(a) the entries of P(x) are formal power series in x with constant coefficients,
(b) det(P(x)] ,E 0 as a formal power series in x,
16 -1
(c) the transformation y = P(x)ii changes the system x' = A(x)y to x' 2i _
X
B(x)il.
Suppose also that s > r. Then, the matrix B(O) must be nilpotent.
Proof
Step 1. Applying to the matrix P(x) suitable elementary row and column opera-
tions successively, we can prove the following lemma.
Lemma V-7-6. There exist two n x n matrices
+oo +oo
T(x) _ xmTm and S(x) _ x"'Sm,
m=0 m=o
and n integers 1\1,,\2, ... , an such that

(i) the entries of n x n matrices T,,, and Sm are constants.
(ii) det To 0 0 and det So 0 0,
(iii)
T(x)P(x)S(x) = A(z) = diag(x''',zA2,...
(iv)A1<A2<...<An .
The proof of this lemma is left to the reader as an exercise.

d6
Step 2. Change the two systems x'd = A(x)y and x' = B(x)u, respectively,
to xr d = C(x)a and x' d

= D(x)ii by the transformations z' = T(x)y and
ii = S(x)v. Then, F = A(x)v. Furthermore, if the matrix D(0) is nilpotent, the
matrix B(0) is also nilpotent.
Step 3. Look at D(x) = x'-'A(x)-1C(x)A(x) - x'A(x)-ld ). This shows

clearly that if s > r, the matrix D(0) is nilpotent.
The following corollary of Theorem V-7-5 is important.
Corollary V-7-7. Assume that conditions (a), (b), and (c) of Theorem V-7-5 are
satisfied. Also, assume that A(O) and B(O) are not nilpotent. Then, r = s.
Definition V-7-8. If a singularity of the second kind is not a regular singularity,
this singularity is said to be irregular. In particular, if the matrix A(O) of (V.7.2)
is not nilpotent, the singularity of (V.7.2) at x = 0 is said to be irregular of order
k. Also, a regular singularity is said to be of order zero.
In order to define the order of singularity at r = 0 for all systems (V.7.2), the
independent variable x must be replaced by x'1P with a suitable positive integer p.
7. CLASSIFICATION OF SINGULARITIES OF LINEAR SYSTEMS 135
For example, for a differential equation
(V.7.7) ah(x)bhl/ = 0,
h=0
assume that the coefficients ah are convergent power series in x and that an (x) # 0.
Set y; = 6'-1y (j = 0,... , n - 1). Then, (V.7.7) becomes the system
f 0 1 0 0
0 0 0 ... 0 yl
(V.7.8) 6y = 91
/i! =
0 0 0 . 1
yn
a0 an-1
an as
If we change (V.7.8) by the transformation y" = diag[l, x-O, '.T-(n-1)0
Iii,
then
(V.7.9)
x°6u
0 1 0 ... 0
0 0 0 0
+ x°diag[l, a, 2a,... , (n - 1)a) u'.
0 0 0 ... 1
-xna ao
a.
-x' an-1
an
ah(x)x(n-h)O
Set bh(x) = . Choose a non-negative rational number a so that bh(X)
an(x)
(h = 0,... , n - 1) are bounded in a neighborhood of x = 0. Also, if a must be
positive, choose a so that bh(0) 0 0 for some h. Then, the matrix on the right-hand
side of (V.7.9) is not nilpotent at x = 0 if or > 0. Hence, we define a as the order of
the singularity of (V.7.7) at x = 0. System (V.7.2) can be reduced to an equation
(V.7.7) (cf. §XIII-5).
The following theorem due to J. Moser [Mo) concerns the order of a given sin-
gularity-
Theorem V-7-9. Let A(x) = x-a E x'A be an n x n matriz, where p is an

=o
integer, A are n x n constant matrices, A0 y6 0, and the power series is convergent.
Set
Im(A) = max (P_ 1 + n,0) ,
µ(A) = minlm(T-1AT-T-1dx)J
Here, r = rank(Ao) and min is taken over all n x n invertible matrices T of the
form T(x) = x-9 E x°T,,, where q is an integer, T are constant n x n matrices,
and the power series is convergent. Note that det T(x) 0 0, but det To may be 0.
Assume that m(A) > 1. Then, we have m(A) > p(A) if and only if the polynomial
P(A) =
vanishes identically in A, where In is the n x n identity matrix.

The main idea is to find out p(A) in a finite number of steps. Using the quantity
p(A), we can calculate the order of the singularity at x = 0. The criterion for
m(A) > p(A) is given in terms of a finite number of conditions on the coefficients
of A(x). There is a computer program to work with those conditions. In particular,
in this way, we can decide, in a finite number of steps, whether a given singularity
at x = 0 is regular.
Notice that IP(x)I can be estimated at z = 0 for the matrix P(x) of Theorem V-7-
2, using similar estimates for fundamental matrix solutions of (V.7.2) and (V.7.4).
Therefore, an analytic criterion that a singularity of second kind at x = 0 be a
regular singularity is that in any sectorial domain V with the vertex at x = 0,
every solution fi(x) satisfies an estimate
c K[xj' (x E V)
for some positive number K and a real number m. These two numbers may depend
on 0 and V. In fact, Theorem V-7-2 and its remark imply that a fundamental matrix
solution of (V.7.6) is P(x)xA0. The matrix P(x) has, at worst, a pole at x = 0 if
and only if !P(x)I < Kjxjr in a neighborhood of x = 0 for some positive number
K and a real number p (cf. [CL, §2 of Chapter 4, pp. 111-1411). Another criterion
which depends only on a finite number of coefficients of power series expansion of
the matrix A(x) of (V.7.2) was given in a very concrete form by W. Jurkat and D.
A. Lutz [JL] (see also 15117, Chapter V, pp. 115-1411).
Now, let us look into the problem of convergence of the formal solution which
was constructed in Theorem V-1-3. Let
n
P = Eah(x)Jh
h=0
be a differential operator with coefficients ah(x) in C{x). Assume that n > I and
a,(x) 0 0. Defining the indicial polynomial of the operator P as in §V-1,
we prove the following theorem.
Theorem V-7-10.
(i) In the case when the degree of f b in s is equal to n, if we change the equation
P[y1 = 0 to a system by setting yt = y and yj = 61-1 [y) (j = 2, ... , n), the
system has a singularity of the first kind at x = 0.
EXERCISES V 137
(ii) In the case when the degree of fn ins is less than n, if we change the equation
P[y] = 0 by setting y1 = y and y, = b'-1 [y] (j = 2,... , n), the system has an
irregular singularity at x = 0.
Proof
n 00
Look at P(xd] _ >ah(x)bh[x°] = xe= x" Ltahi Tn=r+o

fm(S)xm and set
ah(x) = xnobh(x). Then, the functions bh(x) are analytic at x = 0. Furthermore,

if the degree of is n, we must have bn(0) # 0. Therefore, in this case we can
n-1
write the equation P(y) = 0 in the form bo[y] = -bn(Ebh(x)bh(yJ. Claim (i)
h=0
follows immediately from this form of the equation.
If the degree of fn0 is less than n, we must have bn(0) = 0 and b,(0) - 0 for
some j such that 0 < j < n. To show (ii), change Yh further by zh = x(h-1)ayh
with a suitable positive rational number a so that the system for (z1, ... , zn) has
a singularity of the second kind of a positive order (cf. the arguments given right
after Definition V-7-8, and also §XIII-7).
From Theorem V-7-10, we conclude that the formal solution x''(1 + O(x)) of
Theorem V-1-3 is convergent if the degree of f,, (s) in s is n. The following corollary
of Theorem V-7-10 is a basic result due to L. Fuchs [Fu].
Corollary V-7-11. The differential equation Ply] = 0 has, at worst, a regular
singularity at x = 0 if and only if the functions ah(x) (h = 0, ... , n - 1) are
x
analytic at x = 0.
Some of the results of this section are also found in [CL, Chapter 4, pp. 108-137].
EXERCISES V
V-1. Show that if A is a nonzero constant, H is a constant n x m matrix, and N1

and N2 are n x n and m x in nilpotent matrices, respectively, then there exists one
and only one n x m matrix X satisfying the equation AX + NIX - XN2 = H.
V-2. Show that the convergent power series
(a+1)...(a+m-1Q(Q+1) Q+m-1) x,,

y = F(a,x) = 1+a
m=1
x(1-x)!f 2 + +1)J
Y
- c3y = 0,
where a, /3, and -y are complex constants.
Comment. The series F(a, Q, ry, x) is called the hypergeometric series (see, for
example, [CL; p. 1351, 101; p. 159], and [IKSYJ).
V-3. For each of the following differential equations, find all formal solutions of
c
the form x'' I 1 + > cmxm] . Examine also if they are convergent.
L m=1
(i) xb2y + aby + /3y = 0,
(ii) b2y + aby + $y = rxmy,
where b = xa, the quantities a, 3, and 7 are nonzero complex constants and m
is a positive integer.
V-4. Given the system
(E) dt = (N + R(t))il, N = [0 and R(t) = t-3 I 0 0J ,

O] ,
show that
(i) (E) has two linearly independent solutions
, where (t) o m!(m+1)!'

and
J2(t) = 1612 (t
(t)
)1 , where 02(t) = t + +00 bmt-'n - ¢1(t)logt,
m-1
with the constants bm determined by b_1 = 1, bo = 0, and
2m + 1
m(m + 1)bm - b,,,-, = (m = 1,2,... ),
m!(m+ 1)!
(ii) 1 lim t-1(Y(t)-e IN) = O for the fundamental matrix solution Y(t) = [if1(t)12(t)J,
(iii) the limit of e-'NY(t) as t -+ +oo does not exists.
V-5. Let y be a column vector with n entries and let Ax, y-) be a vector with n
entries which are formal power series in n + 1 variables (x, yj with coefficients in
C. Also, let u' be a vector with n entries and let P(x, u) be a vector with n entries
which are formal power series in n + 1 variables (x, u") with coefficients in C. Find
the most general P(x, u") such that the transformation u + xP(x, u7 changes
dy" du"
the differential equation = f (X, y-) to = 0.
ds
Hint. Expand xP(x, u) and f (x, u+xP(x, u)) as power series in V. Identify coeffi-
d[xP(x, V-)]
cients of with those of Ax, u'+zP(x, iX)) to derive differential equations
which are satisfied by coefficients of xP(x, u).
EXERCISES V 139
V-6. Suppose that three n x n matrices A, B, and P(x) satisfy the following
conditions:
(a) the entries of A and B are constants,
(b) the entries of P(x) are analytic and single-valued in 0 < [xj < r for some
positive number r,
dii
(c) the transformation y' = P(x)u changes the system xfy = Ay" to xjj = Bu.
Show that there exists an integer p such that the entries of xPP(x) are polynomials
in x.
Hint. The three matrices P(x), A, and B satisfy the equation xd) = AP(x) -
+"0
P(x)B. Setting P(x) = >2 xmPm, we must have mPm = APm - PmB for all
m=-oo
integers m. Hence, there exists a large positive integer p such that Pm = 0 for
ImI ? P.
V-7. Let ff be a column vector with n entries {y,... , and let A(ye) be an n x n
matrix whose entries are convergent power series in {yj, ... , yn} with coefficients
in C. Assume that A(0) has an eigenvalue A such that mA is not an eigenvalue of
A(0 for any positive integer m. Show that there exists a nontrivial vector O(x)
with n entries in C{x} such that y" = b(exp[At]) satisfies the system L = A(y-)y.
Hint. Calculate the derivative of (exp(At]) to derive the system Axe =

for .
N
V-8. Consider a nonzero differential operator P = >2 ak(x)Dk with coefficients
k=O
ak(x) E C([x]], where D = dx. Regarding C([x]] as a vector space over C, de-
fine a homomorphism P : C[[x]] -. C[[x]j. Show that C((xjj/P(C[[x]j] is a finite-

dimensional vector space over C.
Hint. Show that the equation P(y) = x"`'4(x) has a solution in C([xjj for any
O(x) E C[[x]] if a positive integer N is sufficiently large. To do this, use the indicial
polynomial of P.
N
V-9. Consider a nonzero differential operator P = >2 ak(x)dk with coefficients
k=0
ak(x) E CI[x]j, where b = x2j, n > 1, and 54 0. Define the indicial poly-
nomial as in Theorem V-1-3. Show that if the degree of is n and
if n zeros {A1, ... , A, } of do not differ by integers, we can factor P in the
following form:
42(x))...(6
P = a+,(x)(b - &1(x))(b - - 0n(x)),
where all the functions ¢, (x) (j = 1, 2, ... , n) are convergent power series in x and
4f(O) = Aj.
Hint. Without any loss of generality, we can assume that an(x) = 1. Then, An +
n-1
Eak(O)Ak = (A - A1)... (A - An). Define constants {yo... , In-2} by A' ' +
k=0
n-2
,YkAk
= (A - A2) ... (A - An). For v] (X) E xC[[xI] and ch(x) E xC([x]l (h =
k=O
0,. . . , n - 2), solve the equation
( n-2
(C) P = (6 - Al - 2(7h+Ch(x))6h
h=0
If we eliminate O(x) by Al + V(x) = 1`n-2 + ct-2(x) - an-1(x), condition (C)

becomes the differential equations
6(CO(x)1 = a0(x) + (-Yn-2 + Cn-2(x) -a. -1(x))(1'0 +40(x)),

1 b[ch(x)] = ah(x) + (1'n-2 + Cn-2(x) - an-1(x))(7h +Ch(x)) - (1h-1 + Ch-1(x)),
where h=I,...,n-2.
n
V-10. Consider a linear differential operator P = E atbt, where ao,... , an are
t=o
complex numbers and 6 = x. The differential equation
(P) P[y1= 0
is called the Cauchy-Euler differential equation. Find a fundamental set of solution
of equation (P).
Hint. If we set t = log x, then b = dt

V-11. Find a fundamental set of solutions of the differential equation
(6-06-8)(6-a-8)[yl =xn'Y'
where 6 = xjj and m is a positive integer, whereas a and /3 are complex numbers
such that they are not integers and R[al < 0 < 8t[01.
V-12. Find the fundamental set of solutions of the differential equation
(LGE) f(1-x2)LJ +a(a+1)y=0
at x = 0, where a is a complex parameter.

EXERCISES V 141
Remark. Differential equation (LGE) is called the Legendre equation (ef. [AS, pp.
331-338] or [01, pp. 161-189]).
V-13. Show that for a non-negative integer n, the polynomial Pn(x) = 2nn!
1
x do [(x2 - 1)'] satisfies the Legendre equation
[(1-x2) ] +n(n+1)Pn=0.
d
Show also that these polynomials satisfy the following conditions:
(1) Pn(-x) = (-1)nPn(x), (2) Pn(1) = 1, (3) JPn(t)j ? 1 for {xj > 1,
+00
E Pn(x)tn, and (5) lPn(x)I < 1 for Jxj < 1.
(4) 1 - 2xt + t = n=o
Hint. Set g(x) = (1 - x2)n. Then, (1 - x2)g'(x) + 2nxg(x) = 0. Differentiate this
relation (n + 1) times with respect to x to obtain
x2)g(n+2)(x)
(1,- - 2(n + 1)xg(n+1)(x) - n(n + 1)g(n)(x)
+ 2nxg(n+1) (x) + 2n(n + 1)g(")(x) = 0
or
(1 -X 2)g(n+2)(x) - 2ng(n+1)(x) + n(n + 1)g(n)(x) = 0.
Statements (1), (2), (3), and (4) can be proved with straight forward calculations.
Statement (5) also can be proved similarly by using (4) (cf. [WhW, Chapter XV,
Example 2 on p. 303]). However, the following proof is shorter. To begin with, set
F(x) = Pn(x)2 x2)P''(x))2.

n(n + 1){1 -x2)((1 -
2
Then, F(±1) = Pn(±1)2 = 1, p(X) _ (P"(x))
n(n + 1)
,and F(x) = Pn(x)2 if (x)
0. Therefore, for 0 < x < 1, local maximal values of Pn(x)2 are less than F(1) = 1,
whereas, for -1 < x < 0, local maximal values of Pn(x)2 are less than F(-1) = 1.
Hence, Pn(x)I < 1 for JxI < 1 (cf. [Sz, §§7.2-7.3, pp. 161-172; in particular,
Theorems 7.2 and 7.3, pp. 161-162]. See also (NU, pp. 45-46].)
The polynomials Pn(x) are called the Legendne polynomials.
V-14. Find a fundamental set of solutions of (LGE) at x = oo. In particular, show
what happends when a is a non-negative integer.
V-15. Show that the differential equation b"y = xy has a fundamental set of
solutions consisting of n solutions of the following form:
(logx)k-1-A
OA(x) (k = 1, ... , n),

A-o (k - 1 - h)!
where the functions Oh(x) (h = 0,... , n -1) are entire in x, 00(0) = 1, and Oh(O) _
0 (h = 1, ... , n -1). Also, find O0 (x).
V-16. Find the order of singularity at x = 0 of each of the following three equa-
tions.
(i) x5{66y - 362y + 4y} = y,
(ii) x5{56y - 3b2y + 4y} + x7{b3y - 55y} = y,
(iii) x5y"' + 5x2y" + dy + 20y = 0.
V-17. Find a fundamental matrix solution of the system
dbi
x2 = xyi + y2,
dx
dy2
x2 = 2xy2 + 2y3,
dx
x2 dx3 = x3y! + 3y3.
V-18. Let A(x) be an n x n matrix whose entries are holomorphic at x = 0. Also,

let A be an n x n diagonal matrix whose entries are non-negative integers. Show
that for every sufficiently large positive integer N and every C"-valued function
fi(x) whose entries are polynomials in x such that those of xA¢(x) are of degree N,
there exists a C"-valued function f (x; A, ¢, N) such that
(a) the entries off are polynomials in x of degree N-1 with coefficients depending
on A, N, and ¢,
(b) f is linear and homogeneous in tb,
(c) the linear system
xAdd9 =
z
A(x)3l +
has a solution #(x) whose entries are holomorphic at x = 0 and il(x) is linear
and homogeneous in ',
(d) fi(x)) = O(xN+i) as x 0.
Hint. See [HSW].
V-19. Let A(x) and A be the same as in Exercise V-18. Assuming that n >
trace(A), show that the system xA dz = A(x)y has at least n - trace(A) linearly
independent solutions holomorphic at x = 0.
Hint. This result is due to F. Lettenmeyer [Let]. To solve Exercise V-19, calculate
f (x; A, ¢, N) of Exercise V-18 and solve f(x; A, 4, N) = 0 to determine a suitable
function 0.
00
V-20. Suppose that for a formal power series ¢(x) _ > cx' E C[]x]], there ex-
m=0
/dl
ist two nonzero differential operators P = E ak (x) ` I k
d
and Q = > bk (x) (d) k
k=0 rk=10
with coefficients ak(x) and bk(x) in C{x} such that P[0] = 0 and Q I = 0. Show
J
that 0 is convergent. l
EXERCISES V 143
Hint. The main ideas are

(a) derive two algebraic (nonlinear) ordinary differential equations
(E) F(x,v,v',... ,v(")) = 0 and G(x,v,v',... ,v(' ) = 0
for v = from the given equations P[¢) = 0 and Q ['01 = 0.

(b) eliminate all derivatives of v from (E) to derive a nontrivial purely algebraic
equation H(x, v) = 0 on v. See [HaS1] and [HaS2[ for details.
CHAPTER VI
BOUNDARY-VALUE PROBLEMS OF LINEAR

DIFFERENTIAL EQUATIONS OF THE SECOND-ORDER
In this chapter, we explain (1) oscillation of solutions of a homogeneous second-

order linear differential equation (§VI-1), (2) the Sturm-Liouville problems (§§VI-
2-VI-4, topics including Green's functions, self-adjointness, distribution of eigen-
values, and eigenfunction expansion), (3) scattering problems (§§VI-5-VI-9, mostly
focusing on reflectionless potentials), and (4) periodic potentials (§VI-10). The
materials concerning these topics are also found in [CL, Chapters 7, 8, and 11],
[Hart Chapter XI], [Copl, Chapter 1], [Be12, Chapter 61, and [TD]. Singular self-
adjoint boundary-value problems (in particular continuous spectrum, limit-point
and limit-circle cases) are not explained in this book. For these topics, see [CL,
Chapter 9].
VI-1. Zeros of solutions

It is known that real-valued solutions of the differential equation L2 + y = 0 are
linear combinations of sin x and cos x. These solutions have infinitely many zeros on
the real line P. It is also known that real-valued solutions of the differential equation
2 - y = 0 are linear combination of e= and a-x. Therefore, nontrivial solutions

has at most one zero on the real line R. Furthermore, solutions of the differential
equation2 + 2y = 0 have more zeros than solutions of L2 + y = 0. In this

section, keeping these examples in mind, we explain the basic results concerning
zeros of solutions of the second-order homogeneous linear differential equations. In
§§VI-1-VI-4, every quantity is supposed to take real values only.
We start with the most well-known comparison theorem concerning a homoge-
neous second-order linear differential equation
(VI.1.1) + 9(x)y = 0.
dx2
Theorem VI-1-1. Suppose that

(i) g, (x) and g2(x) are continuous and g2(x) > 91(x) on an interval a < x < b,
d 2 01 (x) d2622x)
(ii) + 91(x)o1(x) = 0 and + g2(x)02(x) = 0 on a < x < b,
(iii) S1 and 6 are successive zeros of 01(x) on a < x < b.
Then. q2(x) must vanish at some point £3 between t:l and C2.
144
1. ZEROS OF SOLUTIONS 145
Proof.
Assume without any loss of generality that Sl < 1;2 and 01(x) > 0 on 1;1 < x < £2.
Notice that 41(1::1) = 0, (1:2) < 0. A contradiction
1 (l:l) > 0, 01(S2) = 0, and
will be derived from the assumption that 02(x) > 0 on S1 < z < 1;2. In fact,
assumption (ii) implies
_ 01 ( x ) d202 (x) =
02(x) d 20, (7)
VI .1. 2 [ 92 (x) - 91 (x)10 (x)0 2 (x)
1
and, hence,
(VI - 1 - 3 ) 0202) 1L( 6) - 02 V 1 )!!LV ) =

1
rE2
E1
(x) - 91 (x)j¢ (x)02 (x)dx .
1
The left-hand side of (VI.1.3) is nonpositive, but the right-hand side of (VI.1.3) is
positive. This is a contradiction. 0
A similar argument yields the following theorem.
Theorem VI-1-2. Suppose that
(i) g(x) is continuous on an interval a < x < b,
(ii) 0i(x) and 02(x) are two linearly independent solutions of (VI.1.1),
(iii) 1;1 and 6 are successive zeros of 01 (x) on a < x < b.
Then, 02(x) must vanish at some point 3 between t;1 and 1;2.
Proof.
Assume without any loss of generality that Sl < £2 and .01 (x) > 0 on S1 < x < 1;2.
Notice that 01(11) = 0, 1 (S1) > 0, 01(12) = 0, and X1(1;2) < 0. Note also that
if 02(1) = 0 or 02(6) = 0, then 01 and 02 are linearly dependent. Now, a
assumption that 02(x) > 0 on fi < x < 2
contradiction will be derived from the assumption
a- (x) - p1(z) d2 02 (x) = 0 and, hence,
In fact, assumption (u) implies 02(x)
dS2 dx2
(VI.1.4) 02(2) 1(6) - 4'201) 1(W =0.
Since the left-hand side of (VI.1.4) is positive, this is a contradiction. 0
The following result is a simple consequence of Theorem VI.1.2.
Corollary VI-1-3. Let g(x) be a real-valued and continuous function on the in-
terval Zo = {x : 0 < x < +oo}. Then,
(a) if a nontrivial solution of the differential equation (VI.1.1) has infinitely many
zeros on I , then every solution of (VI.1.1) has an infinitely many zeros on
Zo,
(b) if a solution of (VI.1.1) has m zeros on an open subinterval Z = {x : a < x <
3} of Zo, then every nontrivial solution of (VI.1.1) has at most m + 1 zeros
on Z.
Denote by W(x) = W(x;01,02) = I the Wronskian of the set of

02(x)
functions (01(x), 02(x)}. For further discussion, we need the following lemma.
146 VI. BOUNDARY-VALUE PROBLEMS
Lemma VI-1-4. Let g(x) be a real-valued and continuous function on the interval
To = {x : 0 < x < +oo}, and let 771(x) and M(x) be two solutions of differential
equation (VI 1.1). Then,
(a) W (x; 772 , ri2 ), the Wronskian of {rli (x), 712(x)}, is independent of x,
(b) if we set t;(x) =
771(x),
then
4(x) = c , where c is a constant.
772(x) dx 772(x)2
Also, if 77(x) is a nontrivial solution of (VI.1.1) and if we set w(x) = then
we obtain
dw(x)
(VI.1.5) + w(x)2 + g(x) = 0 .
Proof.
(a) It can be easily shown that
d 772(x)
771(x)
= 771(x) 772(x) -g(x) ! 712(x) =0
dx I77i(x) 772(x) 77i'(x) 772(x) 771(x) 712(x)
(b) Note that

dl;(x} =- 1 I rl' (x) 712(x) !
and that
dw(x) if'(x)
dx 712(x)2 771(x) 772(x)1 dx :l(x)
1 = -g(x) - w(x)2.
12
((xx))
The following theorem shows the structure of solutions in the case when every
nontrivial solution of differential equation (VI.1.1) has only a finite number of zeros
on To= {x:0<x<+oo}.
Theorem VI-1-5. Let g(x) be a real-valued and continuous function on the inter-
val I. Assume that every nontrivial solution of differential equation (VI.1.1) has
only a finite number of zeros on T. Then, (VI.1.1) has two linearly independent
solutions 777(x) and 712(x) such that
(a) lim 771(x) = 0,
x-+oo 712(x)
(b)
+00 dx +00 dx
(VI.1.6) x)2 = +oo and 1 2
< +00,
1.0 771( Jxo 7r (x)
where xo is a sufficiently large positive number,

(c) the solution 712(x) is unbounded on To.
Proof.
(a) Let (1(x) and (2(x) be two linearly independent solutions of (VI.1.1) such that
(, (x) > 0 (j = 1, 2) for x > xo > 0. Using (b) of Lemma VI-1-4, we can derive the
1. ZEROS OF SOLUTIONS 147
(1(x)
following three possibilities (1) =limo(2(x) = 0, (ii) =U +oo, and (iii)
Urn ( 1( x )>0 . Set

x-+oo (2 (x)
(i(x) in case (i), (2(x) in case (i),
711(x) _ (2(x) in case (ii), 772(x) = (i(x) in cage (ii),
(i(x) - 7'(2(x) in case (iii), 1(2(x) in case (iii).
Then, (a) follows immediately.

(b) Conclusion (b) of Lemma VI-1-4 implies that
1 _Id(Y12(x) and
1 _-1d >)1(x)
,
ql (x)2 ^ c dx . q1(x)) 712(x)2 c dx \ 2(x)
where c is the Wronskian of q1 and rte. Hence, (VI.1.6) follows.
+oc
(c) If q2 is bounded, we must have j()2 -= +00. 0
The following theorem gives a simple sufficient condition that solutions of (V1.1.1)
have infinitely many zeros on the interval Zo.
Theorem VI-1-6. Let g(x) be a real-valued and continuous function on the in-
terval lo. If f g(x)dx = +oo, then every solution of the differential equation
0
(VI.1.1) has infinitely many zeros on Zo.
Proof.
Suppose that a solution 17(x) satisfies the condition that 77(x) > 0 for x >
xo > 0. Set w(x) = !L.(x) Then, w(x) satisfies differential equation (VI.1.5).
77(x)
Hence, lien w(x) = -oo. This implies that lim 17(x) = 0 since 17(x) = r7(xo)
:-+co :-+oo
x exp Vo This contradicts (c) of Theorem VI-1-5. O
J
The converse of Theorem VI-1-6 is not true, as shown by the following example.
Example VI-1-7. Solutions of the differential equation
(VI.1.7)2 + 1
A
z2
= 0,
where A is a constant, have infinitely many zeros on the interval -oo < x < +oo if
and only if A > 4
Proof
Look at (VI.1.7) near x = oo. To do this, set t = to change (VI.1.7) to

x
r
(VL1.8) 462 + 26 + 1 +
L t, 0,
where b = t. Note that t = 0 is a regular singular point of (VI.1.8) and the

indicial equation is 4s2 + 2s + A = 0 whose roots are s = - 2 f 4 - A. These
roots are real if and only if A < 4 This verifies the claim. 0
+00
Note that J dx
0 1 + x2 2*
The following theorem shows some structure of solutions in the case when
+00
ig(x)ldx < +oo.
0
Theorem VI-1-8. Let g(x) be a real-valued and continuous function on the inter-
f+C0
valA. If 1 jg(x)ldx < +oo, then there exist unbounded solutions of differential
0
equation (VI. 1. 1) on Ia.
Proof.
The assumption of this theorem and (VI.1.1) imply that lim dq(x) = -Y exists
:-+oo dx
for any bounded solution n(x) of (VI.1.1). If 7 0, then i(x) is unbounded. Hence,
lim odd) = 0. Therefore, calculating the Wronskian of two bounded solutions
z
of (VI.1.1), we find that those bounded solutions are linearly dependent on each
other. This implies that there must be unbounded solutions. 0
Remark VI-1-9. Theorems VI-1-1 and VI-1-2 are also explained in ]CL, §1 of
Chapter 8, pp. 208-211] and [Hart, Chapter XI, pp. 322-403]. For details concern-
ing other results in this section, see also [Cop2, Chapter 1, pp. 4-33] and [Be12,
Chapter 6, pp. 107-142].
VI-2. Sturm-Liouville problems

A Sturm-Liouville problem is a boundary-value problem
dd-x (p(x) Ly) + u(x)y = f (.T),

(BP)
y(a) cos a -I p(a)y'(a) sin or = 0,
y(b) cos Q - P(b)y(b) sin )3 = 0,
under the assumptions:

(i) the quantities a, b, a, and $ are real numbers such that a < b,
(ii) two functions u(x) and f (x) are real-valued and continuous on the interval
Z(a,b) = {x: a < x < b},
(iii) the function p(x) is real-valued and continuously differentiable, and p(x) > 0
on 1(a, b).
In this section, we explain some basic results concerning problem (BP).
2. STRUM-LIOUVILLE PROBLEMS 149
Let 4(x) and O(z) be two solutions of the homogeneous linear differential equa-
tion
(VI-2.1) (p(x)A) + u()y = 0

-dx
such that
(VI.2.2) m(a) = sins, p(a)d'(a) = cosa, ,p(b) = sin f3, p(b)t'(b) = cosfi,
respectively. Then, these two solutions satisfy the boundary conditions
(VI.2.3) ¢(a) cos a - p(a)4,'(a) sin a = 0, i (b) cos /3 - p(b)0'(b) sin /3 = 0.
The two solutions 4(x) and /;(x) are linearly independent if and only if
(VI.2.4) 4(b) cos Q - p(b)©'(b) sin /3 # 0 or '(a) cos a - p(a)rb'(a) sin a # 0.
The first basic result of this section is the following theorem, which concerns the
existence and uniqueness of solution of (BP).
Theorem VI-2-1. If the two solutions 4(x) and ip(x) of (VI.2.1) are linearly
independent, then problem (BP) has one and only one solutwn on the interval
I(a, b).
Proof
Using the method of variation of parameters (cf. Remark IV-7-2), write the
general solution y(x) of the differential equation of (BP) and its derivative y'(x)
respectively in the following form:
(VI.2.5)
6
VW = CIOW + C20 W + 4(x) 'j'(_)f (_) 4 + i,&(x)1= o(f)f
J P(f)WW a
Y' (X) = Clo'(x) + C21k'(x) + 0'(x) I 4 (x) J A,

s P(A)W JJJu P(A)W V)
where cl and c2 are arbitrary constants and W (x) denotes the Wronskian of
Now, using (VI.2.3) and (VI.2.4), it can be shown that solution
(VI.2.5) satisfies the boundary conditions of (BP) if and only if cl = 0 and c2 =
0. 0
Observation VI-2-2. It can be shown easily that p(x)W(x) is independent of x.
Observation VI-2-3. Under the assumption that the two solutions 40(x) and tfi(x)
of (VI.2.1) are linearly independent, the unique solution of (BP) is given by
rb
(VI.2.6)
y(x) = d(x)
Jx P(A)W (O 10(x) I i (- ))- W(f )
rb owfwdC + V,'(x) r= 4(W W 4.
Y 'W = O(x)
1r W Ja P(OW (O
Setting
4(x)1() if x < C <b ,

p(OW (c)
(VI.2.7) G(x, ) =
O WOW if a < t< x,
p(f)W(c)
we can write (VI.2.6) in the form

b b aG
y(x) =
J
G(x, )f y '(x) =
8x (x, Of (Ode.
J
The function G(x,l;) is called Green's function of problem (BP). The following
theorem gives the characterization of Green's function.
Theorem VI-2-4. The function G(x, C) given by (VI. 2.7) satisfies the following
conditions:
(i) G(x, l:) is continuous with respect to (x, e) on the region D = {(x,) a < x <
b, a< £<b},
(ii) 8 (x, l:) is continuous with respect to (x, C) on the region D - a<
C < b},
(iii) at ({, l;), we have
8 (tr + 0. 0 - ( - 0, 0 = for a< < b,

p(0
(iv) as a function of x, G(x, l:) satisfies homogeneous linear differential equation
(VI. 2. 1) if x 0 4,
(v) as a function of x, G(x, 4) satisfies the boundary conditions of (BP), i. e.,
G(b,Z:)cosf3-p(b)8G(b,{)sinf3=0
for a:5 < b.
Furthermore, the function G(x, f) is uniquely determined by these five conditions.
Proof.
It is evident that the function G(x, l;) satisfies these five conditions. Suppose
H(x,t) satisfies the five conditions. Conditions (iv) and (v) imply that
Cl(l;)4(x) for a < x < l;,

{ C2(t;)V(x) for {<x<b
for (x,l;') E D, where Ci(1;) and C2(C) must be determined by conditions (i), (ii),
and (iii). This means that C,({) and must be determined by
0 and
-PW
2. STRUM-LIOUVILLE PROBLEMS 151
OW and C2({) = 4(f)

This yields Cl(a;) = p(OW(O Since {(x,e) : a < x < }
p(OW(O - -
x<t<b}and {(x,t): 1;<x< b}{(x,£): a<t<x}for
(x, ) E V, we obtain H(x,1;) = G(x, l;). 0
The second basic result co//ncerns self-adjointness of problem (BP). Define a dif-
ferential operator L[y] _ (p(x) I + u(z)y and a vector space V(a, b) over the
real number field 1R by \ /
V(a, b) _ { f r C2(a, b) : f (a) cos a - p(a) f'(a) sin a = 0,
f(b)cos5 - p(b)f'(b) sin,3 = 01,
where C2(a, b) denotes the set of all real-valued functions which are twice contin-
uously differentiable on the interval I(a, b) = {x : a < x < b). Define also an
inner product (f, g) for two real-valued continuous functions f and g on I(a, b) by
(f,9) = fa
Since
(f, 4191) =
(()) +
4
[
J6f()
= p(b)f (b)9 (b) - p(a)f (a)9'(a) + f a

b
dl;
for f E V(a, b) and g E V(a, b), we obtain
(f,L[g) - (L(f1,9)
={p(b)f(b)9'(b) - p(b)g(b)f'(b)} - {p(a)f(a)9 (a) - p(a)9(a)f'(a)}.
Also, since
( f (a) cos a - p(a) f'(a) sin a = 0, f (b) cos Q - p(b) f'(b) sin O= 0,
t 9(a) cos a - p(a)9'(a) sin a = 0, g(b) cos, - p(6)9'(b) sin /3 = 0
for f E V(a. b) and g E V(a, b), we obtain
p(a)f(a)9'(a) - p(a)g(a)f(a) = 0, p(b)f(b)9'(b) - p(b)g(b)f(b) = 0.
Thus, we proved the following theorem.

Theorem VI-2-5 (self-adjointness). The operator L has the following property:
(f, 0191) = (L(f], 9) for f E V (a, b) and g E V(a, b).

Observation VI-2-6. In Theorem VI.2.1, it was assumed that the two solutions
¢(x) and '(x) are linearly independent. Consider the case when this assumption
is not satisfied. So, assume that ¢(x) and ?P(x) are linearly dependent. This means
that
0(a) cos a - p(a)0'(a) sin a = 0 and ¢(b) cos ji - p(b)5'(b) sin fi = 0.
In other words,
(VI.2.8) lr[o) = 0 and 0 E V(a, b).
The boundary-value problem (BP) can be written in the form
(VI.2.9) G[yl = f and y E V(a, b).
Using (VI.2.8) and (VI.2.9), we obtain
(f, 0) = (C [y], 0) = (y, x[01) = 0,
if there exists a solution y of problem (VI.2.9). The converse is also true, as shown
in the following theorem.
Theorem VI-2-7. Assume that O(x) satisfies condition (VI.2.8). Then, if a real-
valued continuous function f (x) on T (a, b) satisfies the condition
(VI.2.10) (f. 0) = 01
problem (BP) has solutions depending on an arbitrary constant.
Proof.
Using the method of variation of parameters, write the general solution y(x) of
the differential equation of (BP) and its derivative y(x) respectively in the following
form:
(VI.2.11)
b
y(x) = cjW(x) + c2p(X) + 0(x)j + (x)J __ 11

P(A)W f ) o P(A)W (E)
?% (z) = J p_ {
r P(f)ww
+ F' (r)
u P(t)
1-v _),
where u(x) is a solution of the linear homogeneous differential equation (VI.2.1)
such that 0 and p are linearly independent, two quantities ca and c2 are arbitrary
constants, and W(x) denotes the Wronskian of {0(x),p(x)}. Note that p(x)W(z)
is independent of x and that
(VI.2.12) p(a)cosa-p(a)u'(a)sina,-40, p(b)cosQ-p(b)p'(b)sinfl,40.
From (VI.2.10) and (VI.2.11), we derive
b
y(a) = ctm(a) + c2p(a) + 0(a)
Jc P(4)W(E)
r° A,
y'(a) = ca0'(a) + c2 (a) + 0'(a)
a -N)
y(b) = ciO(b) + c2p(b), y'(b) = c,O'(b) + c2p'(b)
The condition that y E V(a,b) and (VI.2.12) imply that ca is arbitrary and c2 =
0. 0
3. EIGENVALUE PROBLEMS 153
Remark VI-2-8. The materials of this section are also found in (CL, Chapters 7
and 11] and [Har2, Chapter XI].
VI-3. Eigenvalue problems

In this section, we consider the eigenvalue problem
(P(x) dx ) + u(x)y = Ay,

(EP) y(a)coso - p(a)y'(a)sina = 0,
y(b) cos /3 - p(b)y'(b) sin /3 = 0,
under the assumptions:

(i) the quantities a, b, a, and 0 are real numbers such that a < b,
(ii) u(x) is a real-valued and continuous function on the interval Z(a, b) = {x :
a<x<b},
(iii) the function p(x) is real-valued and continuously differentiable, and p(x) > 0
on I(a, b).
The quantity A is the eigenvalue parameter.
Let 6(x,,\) and ti(x, A) be two solutions of the homogeneous linear differential
equation
(VI.3.1) (px) + u(x)y = Ay

such that
¢(a, A) = sin a, p(a)4'(a, A) = cos a,
(VI.3.2) { ,O(b, A) = sin /3, p(b)t,V(b, A) = cos f3,
respectively. Then,
-O(a, A) cos a - p(a)6'(a, A) sin a = 0,
(VI.3.3)
tp(b, A) cos /3 - p(b) 0'(b, A) sin /3 = 0.
These two solutions 6(x, A) and tp(x, A) are analytic in A everywhere in C. Also,
they are linearly independent if and only if
f 0(b, A) cos (3 - p(b)¢'(b, A) sin /3 i4 0 or

(VI.3.4)
t'(a,A)cosa - p(a)ty'(a,A)sina # 0.
Therefore, we obtain the following result.
Theorem VI-3-1. In order that A be an eigenvalue of (EP), it is necessary and
sufficient OW ,\ satisfies the equation
(VI.3.5) ¢(b, A) cos i3 - p(b)cp'(b, A) sin 0 = 0
or, equivalently,
+'(a, A) cos a - p(a)0'(a, A) sin a = 0.
Observation VI-3-2. All roots of (VI.3.5) are simple, since, if A is a root of

(VI.3.5), we can prove
80(b, A) 8¢'(b, A)
8A
cos j3 - P(b)
8A
sun f # 0
84(x, A)
in the following way. Let A be a root of (VI.3.5). Then, notice that z =
8A
is a solution of the differential equation
CP(x)L1 + u(x)z = Az + ¢(x,A).
This implies that

(VI.3.6) r_
' , A)2 de,
A) dC + p(x) f =
Z(X) = c16(x, A) + c2p(x) - O(x, A)
J
1n Ja P(O1't'(O
z'(x) = clOx, A) + C21L (x) - 4'(x, A) f pW¢(f, A)dC + p '(x) = w(), A)2
PwW (f)
dd,
j P(O-N)
where p(x) is a solution of homogeneous linear differential equation (VI.3.1) such
that 4(x, A) and p(x) are linearly independent, two quantities ct and c2 are con-
stants, and W(x) denotes the Wronskian of {¢(x,A),p(x)}. Note that p(x)W(x) is
independent of x and that
(VI.3.7) p(a) cos a - p(a)p'(a) sin a 0, p(b) cos,3 - p(b)1 (b) sin Q # 0.
Formulas (VI.3.6) imply that

(VI.3.8)
z(a) = clo(a,A) + c2p(a), z'(a) = ci4'(a,A) + c2,u (a),
fb
z(b) = cj¢(b, A) + c2p(b) - 4(b, A) p(b)
Jn Pwww
z'(b) = c1 ¢'(b, A) + c2p'(b) - 0'(b.,\) j + p'(b) rP(u)W
w o P(4)W (0
Since (VI.3.3) is true for all values of A, it follows that z(a) cos a - p(a)z(a) sin a
0. Hence, from (VI.3.7) and (VI.3.8), we conclude that c2 = 0. Therefore,
z(b) cos - P(b) z'(b) sin,3 = (,u(b) cos$ - P(b) (b) sin 0) ° 0(Td # 0.
Ia a P(A)W w
0
Remark VI-3-3. If A = 0 is not an eigenvalue of (EP), we can define Green's

function G(x,t) of problem (BP of §V-2 so that (EP) is changed equivalently to
the integral equation y(x) = A G(x,
To prove that all eigenvalues of (EP) are real, it is convenient to extend Theorem
VI-2-5 (self-adjointness) to complex-valued functions f and g. Define a differential
operator C[y] = ((x)) + u(x)y and a vector space U(a, b) over the complex
number field C by
U(a,b) _ If E C2 (a, b; C) : f (a) cos a - p(a) f'(a) sin a = 0,

f (b) cos a - p(b) f'(b) sin,3 = 0},
where C2(a, b; C) denotes the set of all complex-valued functions which are twice
continuously differentiable on the interval I(a, b) = {x : a < x < b}. Also, define an
inner product (f, g) for two complex-valued continuous functions f and g on 2(a, b)
by
b
(f, 9) = f (09(041
Jn
where g(1;) denotes the complex conjugate of g(f ). Since
(f,'CWi) = 1
a
f \p(e) <J + u(C)9(C)] dC
= P(b)f(b)9'(b) - p(a)f(a)9'(a) + e [-Pwfv)9'(o + uh)fw9(-j <'

J
for f E U(a, b) and g E U(a, b), we obtain
(f, C(91) - (C[f ], 9)

= {p(b)f(b)9'(b) - p(b)9(b)f'(b)} - {p(a)f(a)g(a) - p(a)9(a)f'(a)}.
Also, since
f (a) cos a - p(a) f'(a) sin a = 0, f (b) cos Q - p(b)f'(b) sin /3 = 0,

. 9(a) cos a - p(a)9'(a) sin a = 0, 9(b) cosQ - p(b)91(b) sin p=0
for f E U(a,b) and g E U(a,b), we obtain
p(a)f(a)9'(a) - p(a)9(a)f'(a) = 0, p(b)f(b)9'(b) - p(b)9(b)f'(b) = 0.
Thus, we extended Theorem VI-2-5 as follows.

Theorem VI-3-4 (self-adjointness). The operator C has the following property:
(f, C[9]) = (C[f], 9) for f E U(a, b) and g E U(a, b).
Theorem VI-3-4 implies the following conclusion.

Theorem VI-3-5. All eigenvalues of (EP) are meal.

Proof.
Let A be an eigenvalue of (EP) and let y(x) be an eigenfunction corresponding
to A. We may assume that (y, y) = 1. Then,
(y,Cjy]) = (y, Ay) = X(y,y) = X,

(hjyl,y) = (Ay,y) = A(yy) = A,
and Theorem VI-3-4 imply that J = A. 0
In order to explain distribution of eigenvalues on the real line IR, look at the basic
comparison theorem (Theorem VI-1-1) more closely. Let us rewrite the differential
equation
(VI.3.9) (P(X) dy) + g(x)y = 0
in the form
p(x) = (1 - p(x)g(x))rsin(O)cos(9)
and
(VI.3.10) p(x)dO = I + (p(x)g(x) - 1) sin 2(a)
by setting
y = r sin(O), p(x) Li = r cos(B).

The right-hand side of (VI.3.10) is independent of r. Therefore, (VI.3.10) is a first-
order nonlinear differential equation on 0. Our main concern is the behavior of
solutions of (VI.3.10).
Remark VI-3-6. Set w = ! ((x) L) . Then, differential equation (VI.3.9) be-
comes p(x) dx + w2 + p(x)g(x) = 0. This equation is further changed to (VI.3.10)

by the transformation w = cot(8).
Remark VI-3-7. If the function g(x) is continuous on the interval 2(a, b), then
every solution 0(x) of (VI.3.10) exists on T(a,b).
Now, we prove the following basic lemma.
Lemma VI-3-8. Assume that
(1) four functions gl(x), g2(x), pi (x), and p2(x) are continuous on the interval
2(a, b),
(2) 92(x) ? gi(x) and pi(x) ? P2(x) > 0 on T(a,b),
!
(3) 01(x) (j = 1,2) are solutions of the differential equations
Pi (x) = 1 + [pi (x)gg (x) - 1J sin2(01) (j = 1, 2),

respectively.
Then,
02(x) > 61(x) on I(a,b) if 02(a) > 01(a).
Proof.
Set
9(x) - 1 stn2(02) - sin2(0)
(1x,61,62()1p() l
/ l 02 01
1 WS2(02)
h(x,02) _ (92(x) - 91(x))Sin2(62) + 1
P2(x) P1 (x)
Then, the function w(x) = 02(x) - 01(x) satisfies the differential equation
dw
- A(x,01(x),02(x)) w = h(x,02(x)) > 0.
dx -
Therefore,
w(x)eXP
L-J a
L I
r
w(a) + x h(s,02(s))eXP [- f ds > 0
Ja
on T(a, b). 0
Remark VI-3-9. The proof given above also showed that if pi (x) > p2(x) > 0 on
Z(a, b), then we obtain
(a) 02(x) > 61(x) on I(a,b) if 02(a) > 01(a),
02(x) > 01(x) for a < x 01(a) and g2(x) > g1(x) on a < x < b,
( y) in the case when g2(x) > g1(x) on a < x < b, if 01(a) = 0 and 61(b) = 7r and
if 0<02(a)<rr,then 02(b)>a.
Also, y=rsin(0)>0 if and only if 0<0<r(mod 2rr)and p(x) = 1 > 0 at
0 = 0 (mod jr). Theorem VI-1-1 follows from (y).
Setting 9(x,,\) = u(x) - A, apply Lemma VI-3-8 to problem (EP).
Lemma VI-3-10. Let 0(x, A) be the unique solution to the initial-value problem
Ax) = 1 + (P(x)(u(x) - A) - 1)sin2(0), 0(a,A) = a

on the interval Z(a, b), where u(x) is continuous on Z(a, b), p(x) is continuous and
positive on I(a, b), A is a real parameter, and at is a fixed real number such that
0 < a < rr. Then, for every c E 2(a, b) such that c > a,
(i) 0(c, A) is a continuous and strictly decreasing function of A for -oo < A <
+00,
(ii) lira 0(c, A) = +oo,
a
(iii) lim 0(c, A) = 0.
A +00
Proof.
We prove this lemma in three steps.
Step 1. To prove (i), apply Lemma VI-3-8 to pi(x) = 12(x) = p(x), 91(x) =
g(x, A1), and 92(x) = g(x, A2). If A2 < A1, then g2(x) > 91(x) on Z(a, b). Hence,
9(x, A2) > 9(x, A1) on a < x < b, since 9(a, A2) = 9(a, Al) = a.
Step 2. To prove (ii), choosing a real number m and a positive number P so that
u(x) > m and p(x) m -A and P > p(x) > 0 for x E I(a, b), -oo < A < +oo
and a > 0, it follows from Lemma VI-3-8 that
9(c, A) > 9o(c, A) for a<c<b and - oo < A </+oo.
Observe that v = P tan(Oo) satisfies the differential equation = 1 + ( MP A ) v2.
Hence, tan(9o(x, A)) = tan ( mP A (x - a) f for A < m. Note that

P(m - A)
tan(9o(x, A)) = 0 if and only if tan I VmP A (x - a)) = 0, i.e.,
m-A
tan(8o(x,A)) = 0 at (x - a) = na for n=1,2....
P
and that 9o(x, A) is strictly decreasing with respect to A if x > a (cf. Step 1).
Furthermore, since tan(9o(x, m)) = x a, we must have 0 < 9o(x, m) < 2 for
P
a < x. Therefore,
9 m-P
()2) = ntr for n = 1, 2, ... .
Hence,
r
9 I x, m -
(.)2) > na for n = 1, 2, ... .
x-a
acc
Thus, we conclude that lim 9(c, A) = +oa if c > a.
Step 3. To prove (iii, first notice that 0(c, A) > 0 for -oo < A < +oo, since
P(IX)
0(a, A) = a > 0 and = > 0 if 0 = 0. Choose a positive number M so that
2(e)
lu(x) sin2(9) + 1 I <M on I(a, b).
Ax)
si n
Then,
dO < M - A sin2(0) on Z(a, b).
For any positive number 6 such that 0 < a < r - 6, fix another positive number
A(6) so that
M - A sin2(6) < -
c-10a
for --
6<0< r - b and A > A(6).
-
This implies that dO < 0 if 9 = 6. Hence, 9(x, A) > 6 for a < x < c if 9(c, A) > 6,
whereas9(c,A) <a-10<r-10<0if 9(x, A)>6fora<x<c. Thisisa
contradiction. Therefore, 0 < 9(c, A) < 6 for A > A(6). Thus, we conclude that
lim 9(c, A) = 0. 0
+00
Now, using Lemma VI-3-10, we prove the following theorem concerning problem
(EP).
Theorem VI-3-11. Assume that u(x) is continuous on the interval I(a.b) = (x:
a < x < b} and that p(x) is continuously differentiable and positive on I(a,b).
Then,
(1) problem (EP) has infinitely many eigenvalues:
AI >A2>...>An >...,
(2) lira An = -oo,
n
(3) every eigenfunction Q,,(x) corresponding to the eigenvalue An has exactly n-1
zeros on the open interval a < x < b.
Proof.
Assume without any loss of generality that 0 < a < r and 0 < 0 < r. Define
9(x, A) by the initial-value problem
(VI.3.11) p(x) = 1 + (p(x)(u(x) - A) - 1)sin2(9), 0(a, A) = a,

Tx-
and, then, define r(x, A) by
(VI.3.12) p(x)d. = (1 - p(x)(u(x) - A))rsin(9) cos(9), r(a,A) = 1.
Then,
(VI.3.13) y = i(x, A) = r(x, A) sin(9(x, A))

is a nontrivial solution of the differential equation
d
(VI.3.14) + u(x)y = Ay
ds (P(x)±y)
d.T
that satisfies the condition y(a)cos (a) - p(a)LY(a)sin (a) = 0. Note that from
(VI.3.11), (VI.3.12), and (VI.3.13), it follows that p(x) A) = r(x, A) cos(9)(x, A).
The eigenvalues of (EP) are determined by the condition 0(b,,\) = f3 ( mod sr). Since
8(b, A) is strictly decreasing as A +oo and since 8(b, A) takes all positive values
(cf. Lemma VI-3-10), the eigenvalues A. are determined by
0(b,,\.) = Q + (n - 1)a, n = 1,2,... .
Observe that
(I) ¢(x, A) = 0 if and only if 8(x, A) = 0 (mod a),
8=0(mod
(II) 2 =p(x) >0if a).
This implies that 4(x, A) has exactly k zeros on the open interval a < x < b if and
only if
kar < 0(b,,\) < (k + 1)7r.
Observe also that
(n - 1)-,r < ,3 + (n - I)a < nn.
Hence, the eigenfunction 45(x, has n - I zeros on the open interval
a<x<b. 0
Observation VI-3-12. As shown earlier, if u(x) > m and p(x) n'rr for n = 1, 2, ... .
b-a
This implies that
/
(VI.3.15) an>m - P( bn,)z
- a
n = 1, 2,....
Observation VI-3-13. If u(x) p > 0 on I(a.b), determine 81(x, A)
by the initial-value problem
p dx = 1 + (p(p sin2(81), 81(a, A) _ ?r.
Then, since u(x) - A p > 0 for x E T(a, b), -oo < A < +oo,
and a < s, it follows from Lemma VI-3-8 that
0(c,,\) < Oi(c,A) for a<c<b and -00 < A < +oo.
Observe that v = p tan(81) satisfies the differential equation L = 1 + (1f -P A) v2.

Hence,
tan(81(x, A)) = 1 tan (x - a) for a < µ.

Pµ- P
Note that tan(91(x,A)) = 0 at (x - a) = nrr for n = 1,2,... and

AP2
that 01(2, A) is strictly decreasing with respect to A if x > a. Furthermore,
7r < 91(x, p) <
2 for a < x since tan(01(x, {e)) = x p a . Therefore,
91 (X, (-na \2
p xa I = (n + 1)ir for n = 1, 2, ... .
Hence,
P ( nTr )2)
< (n + 1);r for n = 1, 2,...
x-a
This implies that

2
nrr
(VI.3.16) An+2 < p - p for n = 1,2,... .
G-a
If Al and A2 are determined by
(VI.3.17) 01(b,A1) = Q and 01(b,A2) = 0 + it,
then Al < Al and A2 < A2. From (VI.3.15) and (VI.3.16) we derive the following
result concerning distribution of eigenvalues of (EP).
Theorem VI-3-14. For n > 3, eigenvalues An satisfy the following estimates:
< Iml if An < 2

P '
- a)
(i)
n2 n2 Cb
and
2 2
rr 1 1
P(b- a) < IµI + 4PCb-a) C1
n2 + n2 n n
_An>
if n2 -P Wa
where in, p, P and p are real numbers such that
P > p(x) > p > 0 and p > u(x) > m for x E Z(a, b).
Remark VI-3-15. Lemma VI-3-8, Lemma VI-10, and Theorem VI-3-11 are also
found in (CL, §§1 and 2 of Chapter 8] and [Hart, Chapter XI].
VI-4. Eigenfunction expansions

Let us define a differential operator L[y] = dx (p(z)) + u(x)y and the vector
!LV
space V(a, b) over the real number field R in the same way as in §VI-2. Also, as
in §VI-2, define the inner product (f, g) for two real-valued continuous functions f
b
and g on Z(a, b) _ {x : a < x < b} by (f, g) = It is known that the
J
operator L has the following property:
(f,L[g]) = (L[f],g) for f E V(a,b) and g E V(a,b)

(cf. Theorem VI-2-5). Define a norm of a continuous function on X(a, b) by 11f 11 =
(f, f). Then, 1(f, g)f < Ilf II llgll. The following theorem is a basic result in the
theory of self-adjoint boundary-value problems.
Theorem VI-4-1. Let Al and A2 be two distinct eigenvalues of the problem
(EP) L[y] = Ay, y E V(a, b).
Let r11(x) and r}2(x) be eigenfunctions corresponding to Al and A2, respectively.

Then, (rll,rh) = 0.
Proof.
Note that (r71,L[r12]) = A2(rh,g2) and (Ljrll1,rh) = Al(rl1,rn). This implies that
A2(r11, r12) = AI (r11, r12). Therefore, (r1i, r12) = 0 since Al 0 A2 0
It is known that problem (EP) has real eigenvalues
Al > A2 > ... > An > ... ( lien

n-++oo
An = -co).
Let r)1(x),-.
*2(x), . , be the eigenfunctions corresponding to Al, A2,...
such that 11,11 = 1 (n = 1, 2, ... ). Theorem VI-4-1 implies that (' h,''m) = 0 if
h # k. From these properties of the eigenfunctions rjn, we obtain
k k
11f
- >(f,vh)T1hII2 = Ilfll2 - E(f.rlh)2 > 0
h=1 h=1
and hence
+oo
E(f,7,lh)2
< IIffl2 (the Bessel inequality)
h=1
for any continuous function f on Z(a, b).
In this section, we explain the generalized Fourier expansion of a function f (x)
in terms of the orthonormal sequence {rl (x) : n = 1,2,... }. As a preparation, we
prove the following theorem.
4. EIGENFUNCTION EXPANSIONS 163
Theorem VI-4-2. Let µo not be an eigenvalue of (EP) and let f (x) be a contin-
uous function on Z(a, b). Then, the series
(f, 1h)t1h(x)
h=1 Ah - Jb
is uniformly convergent on the interval Z(a, b).

Proof
Define a linear differential operator Go by £o[yl = G[yl - poy. Then, there exists
Green's function of the boundary-value problem
(BP) Go [y1 = f (x), y E V (a, b)
such that the unique solution of problem (BP) is given by

b
y(x) =
J.
(cf. §VI-2; in particular, Observation VI-2-3). Define the operator G[f[ by

b
G(f 1 = (G(x, ), f) = G(x, Of WA
J
for any continuous function f(x) on 1(a, b). Since Go[g[f]l = f, we obtain
(I,G[.1) = (4[9[f)), 9191) = (9[f], 4191911) = (9[fl,9)
for any continuous functions f (z) and g(x) on Z(a, b). Also, since Go[rlh1 = c[nh1
Ahr1hUO
A01% = (Ah - l'o)llh, we obtain 9['7h] = . Hence,
(f, rlh)ghI
lz lz
< K (f, '1017h =K (f, ih)2,
h=l, h=11
where K is a positive constant such that J!G(x, ) Il < K on 1(a, b). Finally,
I 12
the Bessel inequality implies that Jim (f' nh)nh(x) l = 0 uniformly on

li,ls-+oo h=l - 110
,j
T(a, b). 0
Now, we claim that the limit of the uniformly convergent series of Theorem
VI-4-2 is equal to 9[f].
Theorem VI-4-3. For any continuous function f (x) on Z(a, b), the sum
+00
(h,nh) is equal to 9[f].
Proof
Set 91 (f ] = 9 If, - 1: (f, TO 17h . Then, it suffices to show

h=1=1
Ah - PO
(VIA. 1) lim IIGe[f111 = 0
for every continuous function f (x) on Z(a, b). We prove (VIA.1) in four steps.
Without any loss of generality, we assume that Ah - uo < 0 for h > t . Also, note
that
(VI.4.2) (f,9e[9]) = (94f1,9)
for any real-valued continuous functions f and 9 on Z(a, b).
Step 1. It can be shown easily that Gr[>jh] = chrjh, where
for h = 1,2,... ,e,
Ch =
for h > t.
Let y be an eigenvalue of 9t and let ¢(r) be an eigenfunction of 9t associated with

(0, 17017h .
y. Then, ¢ E V(a,b), 11011 # 0, and y¢ = 9[41 - Also, (VI.4.2)
h=1 An - go
implies that (0,9471h]) = (91'101,170. Hence, ch(¢, nh) = y(¢, nh) Consequently,
(¢, rjh) = 0 if y 0 ch. Therefore, (¢, rjh) = 0 for h = 1, ... , I and y¢ = G[¢] if
y 96 0. Hence, either y = O or y = ch (h > P).
Step 2. In this step, we prove that
(VI.4.3) sup 1190JI1 = sup 1(911f1.f)I-
uni=1 nfa=1
In fact, the inequality I(9e[f),f)1 5 11Ge[f]IIIIfII implies sup 1I9t[f111 >-

ufu=1
sup 1(91[f],f)I Also, since (Ge[f ±9],f ±9) = (911f],f)+(9t[9],9)±2(9t[f],9),
Of H=1
we obtain
4(91[f],9) = (91 [f+9) ,f+9) - (9e[f-9],f-9)
< ( sup l(Ge[w], w)I (IIf + 9II2 + If - 9112)
lH'°t=1
=2 sup I(9e[wj,w)l (11f112+119112).

N++ N=1
Suppose that 9t[f] 0 and IIfIII = 1, and set g = Gt[f] Then, IIGe[f]II -
IIGt[f]II'
sup I (Gt[w], w)I. This shows that sup 119t[f]II sup I (Gt[f]+f )I Thus, the
IIwIl=1 11!11=1 I1I11=1
proof of (VI.4.3) is completed.
Step 3. In this step, we prove that - sup I (Gt If], f) I is an eigenvalue of Gt. To
IlfI1=1
do this, set c = sup I(Ge[f], f)I. Note first that c > I(Gt[ne+I],ne+i)I = Ice+II >
Ill 11=1
0. Also, note that c = sup (911f], f) or -c = inf (Gt[f], f). Suppose that
Iif11=1 [if U=1
c = sup (9t If], f ). Then, there exists a sequence f (j = 1, 2.... ) of continuous
11111=1
functions on T(a, b) such that lim (Gt[ fi], fi) = c and IIf, II = 1(j = 1, 2, . . . )
Since {Gt[ fj] : j = 1,2.... } is bounded and equicontinuous on T(a, b), assume
without any loss of generality that lim Gt[ f)] = g E G(a, b) uniformly on T(a, b).
-.+00
Let us look at
II2
(VI-4.4) IIGt[fjI - cf, = I1Gt[fj]II2 + c2 - 2c(Gt[f,], f,) <- 2c2 - 2c(Ge[fjl , fi).
This implies that lim - cfj II = 0 and, hence, Um 119 - c f j II = 0. Thus,

)+oo 11 9t [f, I j- +00
Gt[g] = cg. Also, from (VI.4.4), it follows that IIgII2 = c2 > 0. Therefore, c must be
an eigenvalue of Gt. However, since all eigenvalues of Gt are nonpositive, this is a
contradiction. Therefore, -c = inf (91 [fit f ). We can now prove in a similar way
Illil=l
that
1
-c = ct+1 =
),ttl - µ0
Step 4. Since sup IIGt[f]II = 1 , we conclude that lien sup I191[f]II = 0

11111=1 !Lo - t+1 t-'+0°1!11!=l
and the proof of (VI.4.1) is completed.
The following theorem is the basic result concerning eigenfunction expansion.
+00
Theorem VI-4-4. For every f E V(a, b), the series f = 1: (f, nh)nh converges to
h=1
f uniformly on T(a, b).
Proof.
Set g = .Co[f], then f = g[g]. Therefore,
+00 +C0
f= (9+17 h)17h = 1: (GO[J]+nh)nh
h=1 Ah - {b h=1 Ah - 110
+00 +00
(1 £ofrlh])r)h
_ ah _ _ >(f, rlh)rlh
h=1 h=1
For every continuous function f on 1(a, b) and a positive number e, there exists
an f, E V(a, b) such that 11f - fe II <_ e. Since
t t t
f - (f, 77h)rh = f -
/
ff + fe - 1: (ff,17h)nh + 1: (ff - f, 17077h,
h=1 h=1 h=1
we obtain lim I f -
t-.+00
(f , 7nh )T7h = 0. Therefore, f = 0 if (f, Tjh) = 0 for h =
l
h=1
1, 2,.... This proves the following theorem.
Theorem VI-4-5. If a continuous function f on 1(a, b) satisfies the condition
U, 17h) = 0 f o r h = 1, 2, ... , then f is identically equal to zero on the interval
1(a, b).
Remark VI-4-6. Also, we have

+00
j11112 = >(f, rth)2 (the Parseval equality).
h=1
Furthermore, this, in turn, implies that if f (x) and g(x) are continuous on 1(a, b),
then
h=1
Example VI-4-7. Using the eigenvalue problem = Ay, y'(0) = 0, y(a) = 0,

dX2
we construct an orthogonal sequence {cos(nx) : n = 0, 1, 2.... } of eigenfunctions.
00
2 + Ean cos(nx) of a function f (x), where
This yields the Fourier cosine series a-°
n=1
2 r*
an = f(x) cos(nx)ds. By virtue of Theorem V1-4-4, this series converges
a oJ
uniformly to f (x) on 1(0, 7r) if f (x) is twice continuously differentiable on 1(0, zr)
and f'(0) = 0 and f'(Tr) = 0.
Example VI-4-8. Using the eigenvalue problem d = Ay, y(O) = 0, y(7r)
= 0, we construct an orthogonal sequence {sin(nx) : n = 1,2,...l of eigenfunc-
00
tions. This yields the Fourier sine series >bn sin(nx) of a function f (x), where
n=1
bn =
ao f/ f (x) sin(nx)ds. By virtue of Theorem VI-4-4, this series converges
uniformly to f (x) on 1(0, ir) if f (x) is twice continuously differentiable on 1(0, ir)
and f (0) = 0 and f (a) = 0.
Example VI-4-9. If f (x) is twice continuously differentiable on Z(-ir, ir), f (-vr) _
f (ir), and f'(1r) = f'(-7r), set fe(x) = 2 if (x)+f (-x)] and fo(x) [f (x)-f (-x))
Then, fe satisfies the conditions of Example VI-4-7, while fo(x) satisfies the condi-
tions of Example VI-4-8. Thus, we obtain the Fourier series
cc
ao + [an cos(nx) + bn sin(nx)],
n=1
r
where an = * f (x) oos(nx)ds and bn = ! J f (x) sin(nx)ds. The Fourier
7r J
series converges uniformly to f (x) on Z(-7r, 7r).
Remark VI-4-10. The Fourier series of Example VI-4-9 can be constructed also
from the eigenvalue problem L2 = Ay, y(-7r) = y(7r), y'(-7r) = y'(ir). Including
this eigenvalue problem, more general cases are systematically explained in [CL,
Chapters 7 and 11]. For the uniform convergence of the Fourier series, it is not
necessary to assume that f (x) be twice continuously differentiable. For example,
it suffices to assume that f (x) is continuous and f'(x) is piecewise continuous on
1(-7r, 7r), and f (7r) = f(-7r). For those informations, see [Z].
Remark VI-4-11. The results in §§V1-2, VI-3, and VI-4 can be extended to the
case when p(x) is continuous on 1(a, b), p(x) > 0 on a < x < b, and p(a)p(b) =
0 under some suitable assumptions and with some suitable boundary conditions.
Although we do not go into these cases in this book, the following example illustrates
such a case.
Example VI-4-12. Let us consider the boundary-value problem
d f_dy . _
(VI.4.5)
y(x) is bounded in a neighborhood of x = 0, and y(l) = 0,
where u(x) and f (x) are real-valued and continuous on the interval 1(0, 1).
Step 1. Consider the linear homogeneous equation
(X2idv)
(VI.4.6) + u(x)v = 0
d
on the interval 1(0,1). Since 1 and logs form a fundamental set of solutions of
the differential equation ± (xdx = 0, a general solution of (VI.4.6) can be
constructed by solving the intergral)equation
v(x) = c1 + C2 109X + (logs)10 u(C)V(4)dC,

10 0
where c1 and c2 are arbitrary constants. In this way, we find two solutions O(x) and
-i(x) of (VI.4.6) such that O(x) and 4'(x) are continuous on 1(0,1), lim O(x) = 1,
and lien x¢'(x) = 0, whereas tI'(z) and tY(x) are continuous for 0 < x < 1,
r_0+
lim (O(z) - logs) = 0, and urn (xo'(x) - 1) = 0. In general, this step of analy-
0+ z-+0+
sis is very important. The behavior of solutions at x = 0 determines the nature of
eigenvalues of the given problem. Denote also by p(x) the unique soltion of (VI.4.6)
such that p(1)=0andp'(1)=1.
Step 2. Assuming that 0(l) # 0, set
¢(x AO for 0< x < e <1 ,

WW
G(x,
¢(F)p(x)
for
where W(x) is the Wronskian of 10,p) and xW(x) = 0(1). Then, G(x,t) is
Green's function of problem (VI.4.5). The unique solution of (VI.4.5) is given
i
by y(x) = 1
G(x, )2dxd{ < +oo.

Jo 0
Step 3. It is easy to prove the self-adjointness of problem (VI.4.5). Hence, all

eigenvalues are real, and the orthogonality of eigenfunctions follows. By virtue of
(VI.4.7), eigenfunction expansions can be derived in the exactly same way as in
§VI-4.
Step 4. We can also derive Theorem VI-3-11 in the exactly same way as in §VI-3.
To do this, consider the differential equation
(VI.4.8) 0-0 + u(x)y = )y.

For every value of A, there exists a unique solution O(x, A) such that lim O(x, A) = 1
xo+
and lim x¢'(x, A) = 0 (cf. Step 1). It is easy to see that p(x, AT and x*'(x, A)
X--o+
are continuous in (x, A) for 0 < x r
B x, A )) = A) We fix B(0, A) = . Using the same method as in
4(x,
§VI-3, we can verify that ¢(1,A) is strictly decreasing and takes all values be-
tween 0 and -oo. Eigenvalues of problem (VI.4.8) are determined by the equations
9(1, A) = mar (m = 1, 2.... ).
Theorem VI-3-14 cannot be extended to the present case, since there is no posi-
tive lower bound of x on 1(0,1).
VI-5. Jost solutions

So far, we have studied boundary-value problems on a bounded interval on the
real line R. Hereafter, we consider the scattering problem, which is a problem on
the entire real line. To explain the essential part of the problem, let us consider
a homogeneous linear differential equation Ly + ((I - u(x))y = 0, where ( is a
complex parameter and u(x) is a real-valued continuous function of x such that
u(x) = 0 for all sufficiently large values of ext. It is evident that e_ttx and e'(' are
two linearly independent solutions of this equation if 1xI is large. Therefore, if a
5. JOST SOLUTIONS 169
positive number M is sufficiently large, the solution y = e-'(= for x < -M becomes
a linear combination a(()e-'t=+b(()e't= of two solutions a-K= and eK' for x > M.
The main problems are (i) the properties of a(() and b(() as functions of ( and (ii)
construction of u(x) for given data {a((), b(()). Keeping this introduction in mind,
let us consider a differential operator
(VI.5.1) G = - D2 + u(x),
where D = and u(x) is real-valued and belongs to C°°(-oo, +oo) such that
aj +o°
(VI.5.2) (1 + jxj)lu(x)jdx < +oc.
J
We study the differential equation
(VI.5.3) Gy = (2y,
where (is a complex parameter. First, we construct two basic solutions which are
called the Jost solutions of (VI.5.3).
Theorem VI-5-1. There exist two solutions f+(x, () and f_ (x, () of (VI.5.S) such
that
(i) ft are continuous for
(VI.5.4) -oc < x < +oo, 0,
(ii) ft are analytic in ( for £ (() > 0,
(iii)
e:
°
jf+(x, () - e'`=I <_ C(x)P(x) 1 +- 1(I
e-K=I e"x
f- (x, () - < C(-x)P(x)1 +
I(1
for (VI.5.4),, where q = Q3'((). C(x) is non-negative and nonincmasing,
t00
P(x) = I (1 + IrI)Iu(r)Idr, and O(x) = (1 + Irj)ju(r)jdr.
J o0
Proof.
Step 1. Construction of f+: Solve the integral equation
sin(((s - r)) u(r)y(r,
y(x, () = e"z ()dr
by setting
yo(x,() = eK=,
+00 sin((((- r)) u(r)ym(r,
ym+1(x, () ()dr (m > 0),
x
and
00
f+(x, () = E Ym(x, ()
m-o
Step 2. Proof of (i) and (ii): To prove (i) and (ii), it suffices to prove that if
e(x) - ( 21x1 + 2 for x < 0,

2 for x > 0,
then
(VI.5.6) Iym(x> }l < mj

(C(x)P(x))me-1,
m= 0,1,2....
for (V1.5.4).
Inequality (VI.5.6) is true for m = 0. Assume that (VI.5.6) is true for an m;
then,
r+ Isin(((x - r)) I
Iym+i(x,C)J < mi
Z
Note that
sin(((x - r)) a z-r
2i(
(1 - e-2i(z-7)() = eft(Z-r)
1 e-2it=dz.
Hence,
sln(((x - r)) I
< e-q(z-T)(r - x) for r1 > 0.
C
Therefore,
+00
I ym+1(x,C)I <
(C(x)me-nx
f (r - x)tu(r)IP(r)mdr.
z
Since
j +00
(T - x)Iu(T)IP(r)mdr < C(x)f.
+00 (1
C(x)P(x)m+
+ I rI )I u(r)I p(r)mdr
m+1
we obtain
(C(x)P(x))m+ie-nZ
I ym+1(x, ()I <

(m 1)!
Remark VI-5-2. At the last estimate, the following argument was used:
(a) r-x<r<1+rforx>O,
(b)
+00
= ( +
J
< 2Ixl1 I u(r){P(r)mdr + 2fz TI u(r)I P(T)mdT
Jz z z z
for x < 0.
5. JOST SOLUTIONS 171
Step 3. Proof of (iii): First, the estimate
'o I
I f+(x, 0- I <- E n,, (O(x)v(x))me-''z
l
m=1
(VI.5.7) 00
< C(x)p(x) e-"z
I i (C(x)p(x))m-1 I
m=1
follows from (VI.5.6). However, this estimate is not enough to prove (iii). So, let
us derive another estimate for large I(I. To do this, we shall prove that
(VI.5.8) I ym(x, ()I rn!

(k) m e-?=,
m=0,1,2....
or(x) = fx
" iu(r)jdr.
Inequality (VI.5.8) is true for m = 0. Assume that (VI.5.8) is true for an m. Then,
+ao
Iym+1(x, ()1 - M! I(I 1 Jz
I sin(((x - r))IIu(r)I e-"'a(r)mdT.
Note that
sin(((x - r))e-+,zl = e 2 11 - e-2it(z-7)1 e-' if 3(() 2! 0.
Hence,
+00 1 m+1
1 e-'7= a-'
Iym+1(x,()I m! iu(r)io(r)mdr = (m+ 1)! (!!-(X) .
KIm+1I
J. 1(i}
This establishes (VI.5.8). Therefore,
v(x)
(VI.5.9) If+(x, () - ei(zl < = -))M-,l
(i
1
M! OWC-1 I e-qz.
The proof of Theorem VI-5-1 for f+ (x, () can be completed by using (VI.5.7) and
(VI.5.9).
For f-(x,(), change x by -x to derive
(VI.5.10) -LY + u(-x)y = (2y.

The solution f+ for (VI.5.10) is the solution f- for (VI.5.3). 0
Remark VI-5-3. The solutions f± are uniquely determined by the conditions

given in Theorem VI-5-1. Also,
d j+00 mi(((x
(VI.5.11) (x, - r))u(r)f+(r, ()dr
and
Id'
(x,() - i(e{C2 + f cos(((x - r))u(r)eK=dr
(VI.5.12) -
< C(x)p(x)a(x)
1+I(I
for (VI.5.4).
Remark VI-5-4. If u(x) = 0 identically on the real line, then f f (x, () = et't'.
If u(x) = 0 for Ixl _> M for some positive constant M, then f+(x, () = e'(= for
x > M, while f- (x, () = e-K= for x < -M.
VI-6. Scattering data
For real (, 1+ (x, () = f+ (x, -(), where f denotes the conjugate complex of f .
Both f+ and f+ are solutions of Gy = (2y and
I f+(x, f) - e ZI < C(x)P(x)

1 + ICI
dl (x, + i(e-+ r+
r))u(r)e(rdrl C() + (x)
dx 16
for -oo < x < +oo and -oo < f < +oo. Let W(f,gj denotes the Vlwronskian of
{f,g}. Then,
W If+, f+l - I f+(x, () -2i( (-oo < < +oo).

f+, (x, () J
This implies that I f+, f+} is linearly independent if 36 0. This, in turn, implies
that the solution f_ is a linear combination of I f+, 1+}. Set
(VI.6.1) f-(x,() = a(()f+(x,() + b(f)f+(x,()
It is easy to see that
i i f+(x+ f-(x,()
(VI.6.2) a(() =
and
(VI.6.3) b(() = Wif+,f-j = I f+(x,() f .(x,()
I.
The function a(() is analytic for %() > 0 and

(VI.6.4) a(() = 1 + 0((-1) as (-' 00 on `3(() > 0,
whereas the function b(() is defined and continuous on -oo < ( < +oo. Further-
more, b(() = O(Ifl-1) as I(I ---. +oo. To simplify the situation, we introduce the
following assumption.
6. SCATTERING DATA 173
Assumption VI-6-1. We assume that

Iu(x)I < Ae-kj=j (-oo < x < +oo)
for some positive numbers A and k.
The boundary-value problem
(VI.6.5) Gy = Ay, y E L2(-oo,+oo)

is self-adjoint, where L2(-00, +oo) denotes the set of all complex-valued func-
r+00
tions f (x) satisfying the condition `f (x) I2dx < +oo. The self-adjointness
J
of problem (VI.6.5) can be proved by using an inner product in the vector space
L2(-oo, +oo) in the same way as the proof of Theorem VI-3-4. Therefore, eigenval-
ues of problem (VI.6.5) are real. Furthermore, all eigenvalues are negative. In fact,
if t 96 0 is real, then A = £2 > 0 and the general solution ci f+ + c2f+ is asymptoti-
cally equal to cle`4=+c2e-'4= as x --. +oc. If t = 0, then f+r(x,0) is asymptotically
equal to 1 as x +oc. Moreover, another solution f+J z is asymptotically
f+
equal to x as x +oc. Therefore, all eigenvalues are A = (ir7)2 < 0. Furthermore,
all eigenvalues of (VI.6.5) are determined by
(VI.6.6) a(iry) = 0.
This implies that all zeros of a(() for 3(() > 0 are purely imaginary.
Under Assumption VI-6-1, ft(x, () are analytic for 9'(() > - 2. Hence, a(()
has only a finite number of zeros for 3(() > 0 (cf. (VI.6.4)). Let S = irlj 0 =
1, 2, ... , N) be the zeros of a(() for (() > 0. Then, f+(x, ir73) are real-valued and
f+(x, ins) E L2(-oo, +oo). Set
c) _ r+ao 1 (j = 1, 2, ... , N),

(VI.6.7) f+(x,n,)2dr
(-oc < { < +oo).
Observation VI-6-2. Every eigenvalue of (VI.6.5) is simple, i.e.,

da(C)
;f 0 if a(() = 0.
4
Proof.
From a(t;) = 0, it follows that
ìa(() i d W[ f+,
d( 2S d( f-[ - 2 2 W [f+, f-I = 2Zr (u'[f+(, f-] + N'[f+, f-<l),
where ft denotes 4f . Also, two relations
`f+S + of+C = (2f+C + uf+ and
imply that
d d
d w(f+(, f-1 = 2-f+f- and W(f+, f-cl = -2f+f_
Since a(() = 0, there exists a constant d(() such that f_(x,() = d(()f+(x,C).
Therefore,
W(f+t, f-) =
-2S+00
+ f+f-dr and W(f+, f-tl = -2C f o0

f+f-dr.
Thus, we obtain
+oo +oo f+dr = -id(()
da(()
-i f+f-dr = -id(()f # 0.
d( 00 oo Ci
Observation VI-6-3. The quantities a(() and b({) satisfy the following relation:
(VI.6.8) Ia(t)12 - Ib(E)12 = 1 for - oo < +00.
Proof.
From f_(x, -{), it follows that
aW = a(-4), b(() = b(-(), and W(f-,f-l = 2i,.

Therefore, (VI.6.8) follows from
f- = af+ + b(af- - bf_) = a(f+ + bf-) - Ib12f_

_ aaf- - Ibl2f_ = {1a12 - Ibl2}f .
Observation VI-6-4. The formula
(- iq, exp [-L f+°° log(1 -

(VI.6.9) a(() _ 2ri
i1h ao (- S J
for %() > 0 shows that the quantities q, and r({) determine a(().
Proof
Set f (() = a(()j Then,

7. REFLECTIONLESS POTENTIALS 175
(i) f (() is analytic for $(() > -2 and (94 0,

(ii) ((f(() - 1) is bounded for 3(() > -2,
(iii) f (() & 0 for Qr(() >- 0 and (54 0.
Set also F(() = log(f (()). Then, O(log(()) near = 0 and F(() _
0((-1) as ( -, 00 on 3(() > 0. Observe that f(() - 1 = O((-'). From this
observation, it follows without any complication that
1 /'}Oc 2log if (01
F(() - Trif_. for £(() > 0
-(
(cf. Exercise VI-18). Now, (VI.6.9) follows from If(t) = Ia(()j and
loBja(()f2 = log(Ta(f)22/
Definition V1-6-5. The set {r((), (n,,n2,... ,nN), (cl,c2,... ,CN)} is called the
scattering data associated with the potential u(x).
VI-7. Reflectionless potentials

The function is called the reflection coefficient. If this coefficient is zero, the
potential u(x) becomes a function of simple form. Let us look into this situation.
Observation VI-7-1. If r(() = 0, then b(() = 0. This means that f_(x,t:) =
a(t) f+(x, £) = a(t) f+(x, -t) (cf. (§VI-6) ). Therefore, using the relation
f-(x, -C)
f+(x,0) = for 3(() !5 0,
a(-()
we can extend f+ for all ( as a meromorphic function in (. Furthermore, if irlj (y =
1, 2,... , N) are zeros of a(() in 3'(() > 0, then -inj (j = 1, 2,... , N) are simple
poles of f+(x, () in ( and
Ftesidueoff+at -inj = - f-(x,ir,) = C)f-(x,irli) -

= -icj f+(x, ins).
a<(inj) id(irtj)
Observation VI-7-2. Set gJ (x) = e-" cj f+(x, iii ). Then, from the fact that
e_,<=
f+(x, () - I as I(I - +oo, it follows that
(VI.7.1) f+(x,0) = e'S= 1 - i (+9,(x)ins

=1
Setting ( = int in (VI.7.1), we obtain

e2ne= N 1
(VI.7.2) 9t(x) + nt+n'gj(x) = 1 (t=1,2,...,N).
C-1
J=1
Observation VI-7-3. If we set F(x,() = e-'t= f+(x, (), then -F" - 2i(F'+uF =
'` 91 +(x) , it follows that
0. Since F(x, ) = 1 - i-1
N
(VI.7.3) u(x) = 2E gj,(x).
.7_1
Observation VI-7-4. Let us solve (VI.7.2) for the g,(x). First, set
hi = gie'''= (j = 1,2,... ,N).
Then, (VI.7.2) becomes
c
(VI.7.4) ht + h = cte-''i: (t
J=1 rJt + I?j
Write the coefficient matrix of (VI.7.4) in the form IN + C(x), where IN is the
N x N identity matrix. Since
N +op N 2
E 717t = E 7 e-n,,= dx
)t°1 'b + 171 L i p
and 77, (j = 1, 2,... , N) are distinct, the matrix IN + C(x) is invertible for -oo <
x < +oo. Set L(x) = det(IN + C(x)). Then, manipulating with Cramer's rule, we
dL
can write g, (j = 1, 2,... , N) and in the following forms:
da
(VI.7.5) 9j (x) = A(X) (j = 1, 2,... , N)
and
N
dL(x)
(VI.7.6)
dx
J=1
Thus finally, from (VI.7.3), it follows that

N
2E g,.;(x)
u(x) = = -2 a(x)
or
(VI.7.7) u(x) = -2z(log(0(x))}.

7. REFLECTIONLESS POTENTIALS 177
This implies that u(x) is a rational function of exponential functions and satisfies
Assumption VI-6-1 of §VI-6. To see this, the following remarks are also useful:
(a) we have the identity
N N
gg(x) 2 ?1 e2n'x9s(x)2,
j=1 j=1 i
(b) g j (-oo) = = limo g . (x) ( j = 1, 2, ... , N) exist and

N 1
E g,(-oo) = 1 (e = 1, 2,... , N).

1=1171+n,
To show (a), derive
e2nez N 2cl[ e2arxgt(x)
cr 9t(x) +
=1 1h + n
g, (x) = - ((= 1, 2, ... , N)
from (VI.7.2). Multiplying both sides by ge, adding them up over f, and interchang-
ing the orders of summation, we obtain (a). To show (b), calculate the inverse of
the coefficient matrix of (VI.7.2). Using Cramer's rule on (VI.7.4), it can be shown
that he(x) -, 0 exponentially as x -+ oo. Also, (b) implies that e2°,, xg,(x)2 is expo-
nentially small as x -oo. Therefore, (a) implies that u(x) satisfies Assumption
VI-6.1.
e2+
gz 1
Example VI-7-5. In the case N = 1, if we determine g(x) by g(x) _
g /
1, then u(x) = 2g'(x). Since g(x) = c' we obtain u(x) = -- e2'rzg(x)2.
C
Qx 2c e+7x +
Also, FF17e-nx g(x) = 1 implies that g(x) = 2 e-n=
(211
In
xsech(11(x + p)), where p = `c , and, hence,
211
(VI.7.8) u(x) = -2112 sech2(g(x + p)).

4
In particular, formula (VI.7.8) yields u(x) = -8 sech (2 (x + In (s) ) I in the

case when N = 1, n1 = 2, and c1 = 5. On the other hand, a\straightforward calcu-
lation using formula (VI.7.7) yields u(x) _ (5+4)2. Also, u(s) = -2sech(x)
is the reflectionless potential corresponding to the data r1= 1 and c = 2.
Example VI-7-6. If N = 2, ill = 2, rh = 3, c1 = 5, and c2 = 2, we obtain
40e42(16 + 135e2-- + 600e6x + 2160e10s + 3600e12=)
U(X)
(1 + 20e4x + 75e6. + 60ebox)2
by calculating (VI.7.7).
Remark VI-7-7. If the reflection coefficient r(t) = 0 and if there is no eigenvalue

for the problem
d2 y
- + u(x)y = Ay, y E L2(-oo, +oo),
2
then u(x) = 0 identically for -oo < x < +oo. In fact, f+(x,() is analytic in
everywhere in the (-plane. Furthermore, l e-`<= f+(x, () -11 -+ 0 as - oo. Hence,
f+(x,() = e'(=. This shows that u(x) = 0.
Remark VI-7-8. If u(x) 0 but u(x) = 0 for }xI _> M, where M is a positive
number, then the reflection coefficient r(S) # 0. In fact, if r(C) = 0, then u(x)
is analytic for -oo < x < +oo. Hence, u(x) = 0. This is a contradiction. This
remark reveals that the problem posed at the beginning of §VI-5 is not as simple
as it looks.
VI-8. Construction of a potential for given data

If scattering data {0, (71,... 7x), (cl,... ,cN)} are given, formula (VI.7.3)
gives the corresponding reflectionless potential u(x) as it is shown in Examples VI-
7-5 and VI-7-6. However, in these two examples, we assumed implicitly that such
a potential exist. Since the existence of u(x) has not been shown yet, it must be
proved. To do this, for given data
7i>0, 72>0, ..., 7N>0 and c1>0, c2>0, ..., cN>0,
define g, (j = 1, 2,... , N) by (VI.7.2) and define f+ and u by (VI.7.1) and (VI.7.3).
Hereafter, the first thing that we have to do is to show that y = f+ is one of the
Jost solutions of Ly = (2y.
Observation VI-8-1. Let gi (t = 1, 2,... , N) be determined by (VI.7.2) and u(z)
be defined by (VI.7.3). Then,
(VI.8.1) gi + 27r9t - ugt = 0 (=1,2,... , N).
Proof
Write (VI.7.2) in the form
N
(VI.8.2) E ae, (x)9, (x) = 1 (t = 1, 2,... , N)
3=1
and differentiate both sides with reaspect to z. Then,
(VI.8.3) Eat, (x)9f (x) + 2t, 9j(x) = 0 (t = 1, 2,... , N)

}=1
and, hence,
N N
Eata(x)9j'(x) + 27t 11 - 1 g, (x) =0 (t = 1,2,... ,N).
1=1 i=1 7e+7)
8. CONSTRUCTION OF A POTENTIAL FOR A GIVEN DATA 179
From (VI.8.2), it follows that
N N
E ajj(x){g,(x) + 2gegj(x)} - 237tE g, (x) = 0
j=1 j=1 ge +T7j
(e = 1, 2, ... , N)
or
N N N
g,
Eatj(x){gf(x) + 2gtgj(x)} - 2>9,(x)+ 2E 711+17., 92(x) = 0
3=1 j=1 j=1
(e = 1, 2, ... , N).
Differentiating again, we obtain
N 2o[s
E ata (x){gj"(x) + 2r7eg, (x) } + 2ge (9e(x) + 2gtgt(x)) - u(x)
,=1
N
+ 2E gj 9'(x) = 0 (e = 1,2,... N)
3=1 ge+g,
or
N N
Eat,j(x){g,"(x) + 217eg,(x)) - Eaea(x)u(x)9J(x)
,=1 f=1
N e2nas
+ aea(x)2%g,'(x) + 2171 C27jt
ct ) = 0 (e = 1,2,N).
... ,9e(x)
,=1
Thus, we derive
N
F'ae,,(x){9,(x) + 217,g,(x) - u(x)g,(x)}
j=1
N
2171 E atj (x)9, (x) + 217t ct 91(x) =0 (t = 1, 2, ... , N),
,=1
and (VI.8.3) implies that
N
F'at,1(x){9;'(x) + 2s7,g,,(x) - u(x)g,(x)} = 0 (e= 1,2,... ,N).
,=1
Therefore, (VI.8.1) follows.

Observation VI-8-2. If we further define f+ by (VI.7.1), then £f+ = (2f+.

Proof.
(e-"=f+)11 + u(e_" f+)
-iJ=1
-i N
g1'
g;,
( ,=1
+ 2i(g - ttgJ
+i
+ 2rligl' - ugJ = 0.
- 2
N
9J
E
j=1
(+ tnJ
Observation VI-8-3. Since
N x
g., (x) en,"
f+(x,it7t) = e-fez 1 -
711+17.;
_ -gt(x)
ct
E L2(-oo,+oo),
J=1
N numbers -1l,2 (j = 1, 2,... , N) are eigenvalues.
N
Observation VI-8-4. Set a (cf. (VI.6.9) with r 0) and
J=1
N
g3 (x)
f-(x,() = a(()f+(x, -() =
a(S)e-Cz
1 + iE J=1(-ig,
(cf. Observation VI-7-1). Then, £f_ _ (2f_ . Furthermore, since
E gJ (-oo) = 1 (e = 1, 2, ... , N)
J=1171+17J
(cf. (VI.7.2)), we obtain
(VI.8.4) a(() = 1 - i1 ____ )
j=1(+nJ
In fact, this follows from the fact that both sides of (VI.8.4) are rational functions
in ( with the same zeros, the same poles, and the same limits as ( , oo.
Observation VI-8-5. The functions f±(.x, () are the Jost solutions of Ly = (2y.
Proof.
Note first that lim f+(x. ()e-'t= = 1 (cf. (VI.7.1)). Also, we have
s+oo
lim f+(x,()e-`<x = 1 - gg(-o) = a(() (cf. (VI-7.1) and (VI.8.4)).

i=1
This implies that
lim f+(x,-()e"x = a(()a(-() = 1.
lim f-(x,()e'S= = a(() x---oo
t--oo
O
9. DIFF. EQS. SATISFIED BY REFLECTIONLESS POTENTIALS 181
Observation VI-8-6. Note that f_ (x, ) = a(() f+(x, -{) = f+(x, ) for ( =
C real. This means that 0. Hence, r(C) = 0. Thus, we conclude that u(x) is
reflectionless.
Remark VI-8-7. For the general r((), the potential u(x) can be constructed by
solving the integral equation of Gel'fand-Levitan:
+00
K(x, r) + F(x + r) + + F(x + r + s)K(x, s)ds = 0,
0
where
F(x) = 1R J +oo
2cL` cle-2n,=
J=1
Find K(x, r) by this equation. Then, the potential is given by
(x,0).
u(x) 8x
If we set r({) = 0 in this integral equation, we can derive (VI.7.2) and (VI.7.3).
Details are left to the reader as exercises (cf. (Ge1LJ ).
V1-9. Differential equations satisfied by reflectionless potentials

Suppose that u(x) is a reflectionless potential whose associated scattering data
are given by 0, (r11, r12, ... flN ), (cl, c2,... , cN) },where 0 < ril < ]2 <
< rily. It was proven that one of the Jost solutions is given by
N
f+(T,() = e`t= 1 - j_1i (+ 9j (T)
Furthermore,
N
f+(x, -() = e`= 1 - g1 (r)
-( + trj7
is also a solution of
dX2
- (u(x) - A)y = 0, where J = (2.
Therefore,
N
(VI.9.1) P(x, A) = 11 ! 1+11- f+(T,C)f+(x, -()
3=1
A
)
3
ddx
(E)
2 dz3
+ 2(u(x) - A) dP + )P = 0.
Let
+oo
(S) P = E an (PO 3 0)
n=0
be a formal solution in powers of A-1 of differential equation (E). Then,
0 = P [-- + 2(u(x) - A)P + u'(x)P

, + (4)2 + (u(x)
P2 - A)P2
J
and, hence,
A(P2)' = I - PZ it + (4)2 + u(x)P2 }, .
Therefore,
(i) the coefficients pn can be determined successively by
po = ca, where co is a constant,

n n n nPtPn+1-t
2C Pn+l = 2 E PtFri-1 + 4E Ptpn-t +
uE pip.-t - r
t=o t=o c=o 1=1
(n = 0,1,2,...),
(ii) the coefficients pn are polynomials in u and its derivatives with constant
coefficients,
(iii) the formal power series (S) is uniquely determined by the condition P = 1 for
u = 0,
()lv i f we denote the unique
q P of (iii) by G u, A ) =
+00
Y G"( u) ( Go = 1), then the
(
An
n=0
general formal solution of (S) is given by P(u, A) = P(0, A)G(u, A).
Example VI-9-1. A straight forward calculation yields
u 3u2 - u" 10u3 - lOuu" - 5(u')2 + u(4)
Gl = Z'
G2 = 8 G3 = 32
Observe that P(x, A) of (VIA 1) is a polynomial in of degree N. Therefore,

the potential u(x) satisfies a differential equation
(VI.9.2) GN+1(u) + a1GN(u) + a2GN-1(u) + ... + QNG1(u) = 0
for some suitable constants a,, a2, ... , 0N. More precisely speaking, it is shown
above that P(u, A) = P(0, A)G(u, A). Compute the coefficients of A-(N+1) on both
sides of this identity. In fact,
N N
e2,7ixgj(x)2
u(x) = 2 g}(x) 4 1` 171
C.,
i=1 1=1
10. PERIODIC POTENTIALS 183
implies that u = 0 if and only if gl = 0 (1 = 1, 2, - , N). This, in turn, implies

that
P(0, A) = 11 C1 +
A
and
G(u, A) = f+(x, C)f+(x, where A=C2.
Hence, ao = 1 and
N +
1 + EN =
2
3=1 1 =1
Example VI-9-2. The function u(x) = -2q2 sech2(17(x /+ p)) is a reflectionless

potential corresponding to the data {r7, c}, where p = 2 In
-
1c)
(cf. Example VI-
7-5). In this case, G2(u)+172G1(u) = 0. Also, G3(u)+(n1+7)2 2(u)+rI14Gi(u) _
0 in the case when N = 2 and {r(C) = 0, (i71, n2), (cl, c2)} are the scattering data.
In particular, G3 (U) + 13G2(u) +36G1(u) = 0 if i1 = 2 and '12 = 3.
Remark VI-9-3. The materials of §§VI-5-VI-9 are also found in (TDJ.
VI-10. Periodic potentials

In this section, we consider the differential equation
(VI.10.1) 2 + (A - u(x))y = 0
under the assumption that
(I) u(x) is continuous for -co < x < +oo,
(II) u(x) is periodic of period 1, i.e., u(x + e) = u(x) for -oo < x < +oo.
The period a is a positive number and A is a real parameter. Denote by 01(x, A)
and 02(x, A) the two linearly independent solutions of (VI.10.1) such that
1(o
(VI.10.2) 01(0,A) = 1, A) = 0, 02 (0, A) = 0, d2(0,A) = 1.
Set
01(e, A) 02(t, A)
.D(A)
= O1 (t"\) (e A)
The two eigenvalues Z1(A) and Z2 (A) of the matrix are the multipliers of the
periodic system
d
(VI.10.3)
dx [Y2J = [u(--O-A 0, [y21
(cf. Definition IV-4-5). It is easy to show that det [1(A)] = 1 for -oo < A <
+oo. Hence, the two multipliers Z1(A) and Z2(A) are determined by the equation
Z2 - f (,\)Z + 1 = 0, where
2(t,A)
(VI.10.4) f(A) =
Note that f (A) is continuous for -oo < A < oo (cf. Theorem 11-1-2). We derive
first the following conclusion.
Lemma VI-10-1. The two multipliers Z1(A) and Z2(A) of system (VI.10.3) are
f(,\) + f (A)2 - 4, ,
2l
2 l f (A) - f (A)2 - 4, ,
where f (A) is given by M. 10.4). Therefore,

(i) if I f (A) I > 2, two multipliers are real and distinct, i.e..
ZI(A) > 1, 0 < Z2(A) < 1 if f(A) > 2,

-1 <Z1(A) < 0, Z2(A) < -1 if f(A) < -2,
(ii) if I f (A) I < 2, two multipliers are complex and distinct, and
I Zt (A) I
=1 and I Z2(A) 1 = 1,
(iii) if f (A) = 2, then Z1(A) = Z2(A) = 1.
(iv) if f(A) _ -2, then Z1(A) = Z2(A) = -1.
ICCl2
Observation VI-10-2. If f (A) = 2, let be an eigenvector of 4;(A) associated
J
with the eigenvalue 1. This means that cl and c2 are two real numbers not both
zero and that
c141(t, A) + C202(1, A) = c1, Cl d ' (t, A) + c2 d-i (t, A) = C2 -
Set 6(x,,\) = cl o1(x, A) + c202(x, A). Then, d(x, A) is a nontrivial solution of

(VI.10.1) such that
.0(t, A) = 0(0,A) and LO (t, A) = (0, A).
This implies that O(x, A) is a nontrivial periodic solution of (VI.10.1) of period t.

Observation VI-10-3. If f (A) = -2, it can be shown that equation (VI.10.1) has
a nontrivial solution O(x, A) such that
0(t, A) = -0(0,A) and (t, A) = - (O, A)

This implies that ¢(x, A) satisfies the condition

¢(x + e, A) = -&, A) for - oo < x < +oo.
Observation VI-10-4. Cconsider the eigenvalue problem
(VI.10.5) 2 + (A - tL(x) y = 0, y(o) = 0, y(e) = 0.
Applying Theorem VI-3-11 (with -A instead of A) to problem (VI.10.5), it can be
shown that (VI.10.5) has infinitely many eigenvalues
(VI.10.6) µI < i2 < <µ<
which are determined by the equation
(VI.10.7) t2(£, A) = 0.
The eigenfunction 02 (x, has exactly n -1 zeros on the open interval 0 < x < e.
Since
.2 (0, A) = 1 > 0, it follows that
dO2 <0 for n is odd,
dx (e'µ") I > 0 for n is even.
Furthermore, condition (VI.10.7) implies that det t I (A)] = A) = 1.

1
Thus, f (A) = 01(e, A) + follows from (VI.10.4). Therefore,
Z V. A)
f(A) > 2 if (e, A) >0 and f(A) :5-2 if (e, A) < 0,
>2 if n is even,
f (µ") { < -2 if n is odd
(cf. (VI.10.7)). Here, use was made of the fact that jaj+ 1a1,
ifa76 0.
Observation VI-10-5. If
(VI.10.8) A < min{u(x) : -oo < x < +oo},
then f (A) > 2. In fact, dX21 (x, A) = (u(x) - A)01 (x, A) > 0 as long as 01(x, A) > 0
dol
and, hence, (x, A) > 0. Thus, we obtain s1(£, A) > 1. Similarly, 2 (P, A) > 1
if (VI.10.8) is satisfied. In this way, a rough picture of the graph of the function
f (A) is obtained (cf. Figure 1).
At A2 44
3 A3= f=2
T
3
I l4
-F
W lA
Ul U2 U3=N3U4 f= -2
u5 U6
FIGURE 1.
Actually, we can prove the following theorem.
Theorem VI-10-6. Assuming that u(x) is continuous and periodic of period t > 0
on the entire real line R, consider three boundary-value problems:
(A) 2 + (A - u(x)) y = 0, y(0) = 0, y(t) = 0,
($)2 + (A - u(x)) y = 0, W) = y(0), dx (t) _

y
(0),
(C) d2 + (A - u(x)) y = 0, y(t) = -y(0), dY(t) _ L(0).

Then, each of these three problems has infinitely many eigenvalues:
µI <µ2</A3...<An<...,
(VI.10.9) A0<\1<A2<A3<a4 <...<A2n_I<\2n <...,

V1 <v2 <v3 :5 L14 <... <V2n-1 <V2n <... ,
respectively. Furthermore, if we denote by an(x), and 7n(x) the eigenfunc-
tions associated with the eigenvalues µn, An, and vn, respectively, then
(1) A0 < v1,
(2) v2n-1 :5 92n-I 1'2n < A2n-1 !5 µ2n < A2n < V2n+1 < µ2n+1 < i/2(n+l) for
n = 1,2,...,
(3) 02n_1(x) and /32n(x) are linearly independent if A2,,-1 = A2n,
(4) 72n-1(x) and 72n (x) are linearly independent if v2n-1 = v2n,
(5) 3D(x) does not have any zero on the interval 0 < x < t,
(6) 02._1(x) and /32,,(x) have exactly 2n zeros on the interval 0< x < t,
(7) 72n-1(x) and 72n (x) have exactly 2n - 1 zeros on the interval 0 < x < t.
Proof.
We prove this theorem in five steps. Note first that (VI.10.9) is obtained from
Figure 1.
Step 1. If ¢(x, A) is a solution of differential equation (VI.10.1), then (x, A) is
a solution of the initial-value problem
(VI.10.10)
i
w(0, A) = (0, )L),
d2w
+ (A - u(x)) w + O(x, A) = 0,
dx2
i (0,,\) =
TA_ ( dx
(cf. §lI-2). Therefore, using the variation of parameters method, we obtain
(VI.10.11) _
-1(x, A) _ (x, A) fT 01(t, A)02(t, A)dt ¢2(x, A)
f01(t"\)2dt'
d z (x,,\)
= 01(x"\) 102(t, A)2dt - 02(X,,\) f -01(t, A) 02(t, A)dt
and, hence,
da (dd-x (x, A)) =

1(x, A) f ; m2(t, A)2dt - (_, A)
la
mi (t, A)yi2(t, A)dt.
df
Therefore,
dA
= rr Q(01 (t, A), ¢2(t, A))dt follows from (VI. 10.4), where
0
Q(Y1,Y2) = -02(e,A)Y2 + dTl(e,A)Y2 + [1(t1A) - (e,A)Jr1Y2.
Step 2. The discriminant of the quadratic form Q is f(,\)2 - 4. In fact,

2
xd
1(t"\) +4 l (e, A)o2(t, A)
A)]
2
dO2
2 (e, A)] + 401(e, A) (e, A) - 4 = f(,\)2 - 4.
Note that
41(e, ) (e, A) - ' (e, A) = det [ I (A)] = 1.
Thus, Q(01 (X, A), ¢2(x, A)) does not change sign for 0 < x < e if If (A)f < 2. It
follows that
(I) df(A) 0 if If(A)f < 2.

dA
Step 3. Also, it can be proven that
df(A)
<0 if A < pl and f(,\)2 = 4,
d,\
(-1)1 d('\) < 0 if pl < A < pi+1 and f(A)2 = 4.
To prove this, notice that Q is a perfect square if f(,\)2 = 4 and that 02(1,.\) # 0
if A 0 pj for every j. Hence, dA\) ¢2(e, A) < 0. Also, using Lemma VI-3-11 (with
-A instead of A), we obtain
02 (1, A) > 0 if A < p1,

A) > 0 if p., < A < pj+1.
Thus, (1) is verified.

Step 4. Let us prove that
dzf
(a)
J< 0 if f (a) = 2 and d(\) = 0,
dal df(A)
>0 if f (A) _ -2 and
d,\
= 0.
First observe that Q is a perfect square if (f(A)l = 2. Hence,
71K (A)
=± f 0
t
[c101(t, A) + c202(t, \)12 dt = 0,
where cl and c2 are some real numbers. Therefore, cl = 0 and c2 = 0, since 01 and
02 are linearly independent. This means that
(IV) 02(f,,\)=0, ' (e, A) = 0,

UX ax
Since det 1, we conclude that
12 if f (a) = 2 and d (a) = 0,

fi(A) =
(V)
- I2 if f (A) = -2 and -(A) = 0,
df
where 12 is the 2 x 2 identity matrix. Furthermore,
z e
2 (a) = Jo - (t, A)0 (t, A)2 + da ( d ') (e, a)2
+ d ) (e, A) } m (t, A)m2(t, A)1 dt.

d
Also, if f (A) = 2 and (A) = 0, it follows from (VI.10.11) and (IV) that
d (e, a) =
I
z
{e, ) = j42(s,A)2ds,
d t
0i (s, A)2 ds,
0
da ( f 0t 0r (s, A)02(s, A)ds.

Hence,
t(0f(t,
d (A) _ - f tf A)02{s, A) - 0e(s, A)02(t,A)j2dtds < 0
if f (A) = 2 and I (A) = 0. Similarly,
d2f
2 (A) = ft f t [Ol (t, A)02(s, A) - 41(s, A)4 2(t, A)J2dtds > 0
if f (A) = -2 and df (A) = 0. Note that, if A2,,-1 = A2,,, it follows from (VI. 10.2),
(IV), and (V) that 01(x, A) and 02(x, A) are two linearly independent solutions for
problem (B). Similarly, if w2r_1 = v2,,, it follows from (VI.10.2), (IV), and (V) that
01(x, A) and 62(x, A) are two linearly independent solutions for problem (C). Thus,
(3) and (4) are verified.
Step 5. The functions (respectively -yn(x)) have an even (respectively odd)
number of zeros on the interval 0 < x < 1, since /3, (0) = Qn(1) and 7n(0) Yn(1)
As can be seen in Figure 1,
112n-i < A2n-1 A2n < 12n+l-
Therefore, by virtue of Theorems VI-3-11 and VI-1-1, we conclude that fit, 1 and
per have more than 2n - 1, and less than 2n + 2, zeros on the interval 0 < x < 1.
Hence, they have exactly 2n zeros there. Similarly, since
(VII) /12n-2 < V2n-1 V2n < 112n,
and -t2n have exactly 2n - I zeros on the interval 0 < x < e. The function
,8a does not have any zero on the interval 0 < x < t since Ao < 111. Thus, (5), (6),
and (7) are verified. Finally, (2) follows from (VI.10.9), (VI), and (VII).
Definition VI-10-7.
(I) The set {A f(A)2 < 4} is called the stability region of the differential
:
equation (VI.10.1).
(11) The set {A : f(A)2 > 4} is called the instability region of the differential
equation (6210.1).
Example VI-10-8. A periodic function u(x) is called a finite-zone potential if the

d2y
function f (A)' - 4 of the differential equation + (A - u(x))y = 0 has a finite
dx2
number of simple zeros (i.e., all other zeros are double). For example, consider
the case when u(x) = a, where a is a constant. The differential equation becomes
d? 2
dx+ (A - a)y = 0 and, hence,
cos(v1'"T-_ax) if A > a,
01 (X, A) = 1 if A =a,
cosh( a --Ax) if A < a
and
( sin(v1_T_-ax)
if A>a,
A-a
02(x, A) = x if A =a,
sinh(x) if A < a.
Therefore,
12 cos(\/-), --at) if A >a,

AA) = A) + (e, A) _ 2 if A=a,
2 cosh( a --At) if A < a.
Thus, we conclude that in this case, f (A)2 -4 has only one simple zero a (cf. Figure
2).
FicURE 2.
The materials in this section are also found in [CL, Chapter 8j.
EXERCISES VI
VI-1. Assume that u(x) is a real-valued continuous function on the interval 20 =

{x : 0 < x < +oo} such that u(x) > rno for x >_ xo for some positive numbers "to
and x0. Show that
(1) every nontrivial solution of the differential equation
day
(E)
&2
- u(x)y = 0
has at most a finite number of zeros on Z0,
(2) the differential equation (E) has a nontrivial solution 1)(x) such that
hm q(x) = 0.
Hint.
(1) Note that if y(xo) > 0, then y"(xo) > 0. Hence, y(x) > 0 for x > x0 if
1/(xo) > 0.
(2) It is sufficient to find a solution 0(x) such that 0(x) > 0 and 0'(x) < 0 for
x>xo.
EXERCISES VI 191
VI-2. For the eigenvalue-problem
(EP) L2 + u(x)y = Ay, Y(()) = y(1), y(0) = y'(1),
where u(x) is real-valued and continuous on the interval 0 < x < 1,

(1) construct Green's function,
(2) show that (EP) is self-adjoint,
(3) show that (EP) has infinitely many eigenvalues.
VI-3. Let A, > A2 > - > An > - be eigenvalues of the boundary-value problem
d2y
+ u(x)y = Ay, y(a) = 0, y '(b) = 0,
where u(x) is real-valued and continuous on the interval a < x < b. Show that
there exists a positive number K such that
2
n + ir
K for n = 1, 2, 3, ... .
n2 b-a n
VI-4. Assuming that u(x) is real-valued and continuous on the interval 0 < x <
+oo and that lim u(x) = +oo, consider the eigenvalue-problem
r+oo
dx2 - u(x)y = Ay, y(0) cos a - y'(0) sin a = 0,

z
lim y(x) = 0.
where a is a non-negative constant. Show that

(a) there exist infinitely many real eigenvalues AI > A2 > . . . such that lim An =
n-.ioo
-00,
(b) eigenfunctions corresponding to the eigenvalue An have exactly n -1 zeros on
the interval 0 < x < +oo.
Hint. Let A1(b) > A2(b) > . . . be the eigenvalues of
dx2 - u(x)y = Ay, y(0) cos a - y'(0) sin a = 0, y(b) = 0,
where b > 0. Define Am by lim A,(b). See JCL, Problem 1 on p. 2541.

6 +oo
VI-5. Show that if a function O(x) is real-valued, twice continuously differen-

tiable, and ¢"(x) + e-'O(x) = 0 on the interval 10 = {x : 0 < x < +oo} and
if J
J0
O(x)2dx < +oo, then O(x) is identically equal to zero on I .
VI-6. Using the notations and definitions of §VI-4, show that
+oo
(f,C(f)) = j>n(f,17n)2
n=1
b
= {u(x)f(x)- P(x)f'(x)2}dx + P(b)f(b)f'(b) - P(a)f(a)f'(a),

j
if f E V(a, b).
VI-7. Assume that u(x) is real-valued and continuous and u(x) < 0 on the interval
I(a, b), where a < b. Denote by yh(x, A) the unique solution of the initial-value
problem 2 + u(x)y = Ay, y(a) = 0, y(a) = 1, where A is a comlex parameter.
Show that
(i) O(b, A) is an entire function of A,
(ii) ¢(b, A) 54 0 if A is a positive real number,
(iii) ¢(b, A) has infinitely many zeros A,, such that 0 > A0 > Al > A2 > and
-
- - \ b r- al/
2
llm A" ) .
n +oo n2 \
VI-8. Find the unique solution O(x) of the differential equation
Ix ) = y such that is analytic at .x = 0 and 0(0) = 1. Also, show that
(i)\\\ m(x)
is an entire function of x,
(ii) ¢(x) A 0 if x is a positive real number,
(iii) ¢(x) has infinitely many zeros an such that 0 > \o > a1 > A2 > . and
lim 1n = -oc,
n-+00
(iv) nx)dx = 0 if n 96 m.
J0
VI-9. Show that the Legendre polynomials
2"n1 d"
Pn(x) !dx"((x2 - 1)'J (n = 0,1,2,...)
=
satisfy the following conditions:

(i) deg P,, (x) = n (n = 0,1,2,...),
(ii) 1 Pn(x)Pm(x)dx = 0 if n 0 m,
J
(iii) j P n (x)2dx = 1 (n = 0,1, 2, ... ),
n z
xkPn(x)dx = 0 for k = 0,... , n - 1,

(iv)
J
(v) Pn(x) (n > 1) has n simple zeros in the interval Ixl < 1,
EXERCISES VI 193
(vi) if f (x) is real-valued and continuous on the interval jxj < 1, then
2
1 N
lim f
N-+oo 1
f (x) - E (n+ Z) (f, PP)Pn(x)) dx = 0,
n=O
where (f, Pn) = f 1
f (x)PP(x)dx,
(vii) the series +00

F, (n + 21 (f, PP)Pn(x) converges to f uniformly on the interval
n=0
jxj < 1 if f, f', and f" are continuous on the interval jxj < 1.
Hint. See Exercise V-13. Also, note that if f (x) is continuous on the interval
jxj < 1, then f (x) can be approximated on this interval uniformly by a polynomial
in x. To prove (vii), construct the Green function G(x,t) for the boundary-value
problem
((I -x2)d/ +aoy=f(W),
y(x) is bounded in the neighborhood of x = ±1,
1 1
where ao is not a non-negative integer. Show that f f G(x, )2dx< < +oo.
J! 1J 1
Then, we can use a method similar to that of §VI-4.
VI-10. Assume that (1) p(x) and p'(x) are continuous on an interval Zo(a,b) _
{x : a < x < b}, (2) p(x) > 0 on Zo, and (3) u(x, A) is a real-valued and continuous
function of (x, A) on the region Zo x ll = {(x, A) : x E Zo, A E It} such that
lim u(x, A) = boo uniformly for x E Zo. Assume also that u(x, A) is strictly
A-too
decreasing in A E R for each fixed x on I. Denote by O(x, A) the unique solution
of the intial-value problem
(P(x)L ) + u(x,.\)y = 0 y(a) = 0, y(a) = 1.
Show that there exists a sequence {pn : n = 0, 1, 2, ... } of real numbers such that
(i) pn < 1An-1 (n = 1, 2, ... ), and lim pn = -oo,
n .+oo
(ii) 4(b, pn) = 0 (n = 0,1, 2, ... ),
(iii) ¢(x, .\) 0 on a < x po, and q(x, A) (n > 1) has n simple zeros
,-
on a<x<bforpn<\<,un-1,
(iv) strictly decreases from +oo to -oo as A decreases from pn_1
to An-
VI-11. Assume that p(x) and u(x) are real-valued and continuous on the interval
1(0,1) = {x : 0 <- x < 1} and that p(x) is also continuous on 7(0,1), p(x) > 0
for 0 < x < 1, p(0) = 0, and p'(0) < 0. Show that the differential equation
dx
(px ) d
dx)
+u(x)y = 0 has a fundamental set {0, ii} of solutions on the interval
0 < x < 1 such that
(i) slim 4(x) = 1 and Zli o p(x)O'(x) = 0,
(ii) (i(z) - 0 and =li m- p(x)ii (x) = 1.

sl
VI-12. Set
o 11 F )
for 0<x<
for 2<x 1.
Denote by 01(x, A) and 42(x, A) the two unique solutions of the differential equation
y
+ (A - u(x))y = 0 satisfying the initial conditions 4i(0, A) = 1, X1(0, A) = 0,
42(0, A) = 0, and (0, A) = 1, where A is a real parameter. Sketch the graph of

the function f (A) = 4(1, A) + (1, A).
VI-13. For the scattering data {r({) = 0, (1,2,3), (1,1,1)}, find the potential u(x)
and the Jost solutions ft(x,().
VI-14. Calculate the scattering data for each of u(x + 1) and u(-x), assuming
that {r({), (271, ... ,T)N), (cl, ... , cN)} are the scattering data for u(x) and that
u(x) satisfies a condition Eu(x)I < Ae-k1=I for some positive numbers A and k.
Hint. Let f f(x, () be the Jost solutions for u(x). Two quantities a(() and
are given by
a(S) = and 1
2C I f+(x,C) f_(x,<) I (X,0 f-' (x:0

respectively.
The Jost solutions f o r u(x + 1 ) are e ' ft (x + 1, (). Therefore, the scattering
data for u(x + 1) are
(171,... ,nN), (e-2"'c1,... ,e-2"r'cN)}.
The Jost solutions for u(-x) are ff(-x, (). Hence, the scattering data for u(-x)
are
b(l)'
a(C)
(171,...,17N)I
(
1
cia<(iv71)2
,...,- 1 l}
CNa<(LT,N)21 J
Note that
f-(x,iil3) = i a<(iul,)f+(x,iu,)
EXERCISES VI 195
VI-15. Let u(x) be real-valued, continuous, and periodic of period t > 0. Also,
for every real , let O(x, {, A) be the solution of the differential equation
(Eq) 2 + (A - u(x))y = 0
satisfying the initial conditions 0({, {,A) = 0 and {G'(£, t, A) = 1. Let
A = Al (0 < A2W < µ3(f) < ...

be all roots of ii({ + t, , A) = 0 with respect to A. Show that
Am(t) - A
( + [', , A) = Q +00 2
m=1
e
Hint.
Step 1. Let us construct O(x, , A) for negative A. To do this, change (Eq) to the
integral equation
sinh(p(x - )) + 1 r sinh(µ(x -
p p f
where it = v > 0. Since e-;`(=-() sinh(µ(x - )) is bounded as p - +oo, it can

be shown that
e-v(x-E)+G(x, such{µ x- f )) + O( A 1
, A) = e-al=-E)
on the interval 0 < x - < t as A -e -oo. Thus, we derive
lim t'(4+f,.,A) = 1.
(sinh(e))
Step 2. There exists an entire function R(A) of A such that
sin(Pf) for A > 0,

R( A ) = J a
sinh(CX) for A <0.
This function R(A) has the following factorization:
R(A) = t
-_,
m=1 \1
mr
where Cm =
t
Step 3. If the function V)((+1, 1, A) is entire in .\ and ''(+t, 1,,\) = O(exp(1 JaJ))
as A -+ oo, we can write 1'(( + 1, t,,\) in the following form:
lp(f + t, (, a) = c
[i(A]
where c is independent of A. Here, we used the fact that cn,J is bounded
as m +oc (cf. Theorem VI-3-14).
Step 4. Note that
_((+t,t,X) - 'c' + [0-W
_- a
(1)
R(A) ( )H cam, - A ]
Note also that
c -a cm -A
This implies the uniform convergence of the infinite product on the right-hand side
Z'((+ 1, t, _ = 1 = c
of (1) for -oo < A < 0. Thus, lim
a-.-OO R(a) l
Remark. For expressions of analytic functions in infinite products, see, for exam-
ple, (Pa, pp. 490-5041.
VI-16. The functions are defined in §VI-9. Calculate G4(u) and G5(u).
I
VI-17. Use the same notations as in Theorem VI-10-6. Show that if J u(t)dt = 0,
0
ft
then Ao < 0. Also show that if / u(t)dt = 0 and = 0, then u(x) = 0 identically
Jo
on-oc<x<+00.
Hint. Since j3o(z) # 0 on -oc < x < +oo, set w(z) = !Lx). Then, w'(x) +
w(x)Z + A - u(x) = 0 (cf. 1A'IW, Theorem 4.4, p. 621).
V1-18. Set F(() = log f (() for à( > 0, ( # 0, where
(i) f (() is continuous for QJ( > 0, ( 0,
(ii) f (() is analytic for `£( > 0,
(iii) ((f (() - 1) is bounded for !a( > 0,
(iv) f (() # 0 for !'( > 0, (0 0.
Denote by SR the semicircle {(: 1(I = R,0 < arg(< r) which is oriented counter-
clockwise, where R is a positive number. Show that
1 1 F(z) dz = f F(() (`1( > 0,1(1 < R),
(a) 2ri IR 2Ri 0 (`'(<0,1(1<R),
(b) F(() = tai j F'(() d{ + 7 f- g for a(> 0,

where fl() is the complex conjugate of F().
CHAPTER VII
ASYMPTOTIC BEHAVIOR OF
SOLUTIONS OF LINEAR SYSTEMS
In this chapter, we explain the behavior of solutions of a homogeneous linear

system = A(t)y as t +oo in the case when the coefficient matrix A(t)
has a limit A0 as t -+ +oo. The purpose is to show how much information we
can glean from the limit matrix Ao. We are interested in the exponential growth
of solutions and the asymptotic behavior of solutions. In order to measure the
exponential growth of a function, we use Liapounoff's type numbers which was
originally introduced by A. Liapounoff in [Lia]. Liapounoff's type numbers are
explained in §VII-1. Also, we explain Liapounoff's type numbers of solutions of a
homogeneous linear system in §VII-2. (Liapounoff's type numbers are also found
in L. Cesari [Ce, pp. 50-551.) In §VII-3, assuming that t-too lim A(t) = Ao, we
calculate Liapotimoff's type numbers of solutions in terms of the eigenvalues of A0
(cf. [Huk3j, [Hart, Chapter X1, and [Si9]). In §VII-4. we explain how to derive the
asymptotic behavior of solutions by diagonalizing the given system. A theorem of
M. Hukuhara and N. Nagumo gives the original motivation (cf. Theorem VII-4-1).
The main result is Theorem of N. Levinson (cf. Theorem VII-4-2). (Generalizations
and refinements of Theorem of Levinson are found, for example, in [Bell], [Bel2],
[CK], [Cop1]. [Dev], [DK], [El], [E2], [Gi], [GHS], [HarL1], [HarL2], [HarL3], [HW1],
[HW2], [HX1], [HX2], and [HX3].) In §VII-5, the Theorem of Levinson is applied
to a system whose matrix has a limit as t -+ +oc and its derivative is absolutely
integrable on the interval 0 _< t < oo. The topics of §§VII-4 and VII-5 are also
found in [CL, §8 of Chapter 3, pp. 91-97]. In §VII-6, we explain how we can reduce
some problems such as the differential equation dt2 + {X + h(t) sin(at)}rj = 0 to
the Theorem of Levinson, even if the derivative of h(t)sin(at) is small but not
absolutely integrable on the interval 0 < t < oo. The main idea is to apply the
Floquet theorem (Theorem IV-4-1) to z + {1 + e sin(ot)}rt = 0 to eliminate
the periodic parts of coefficients so that we can use the Theorem of Levinson (cf.
[HaS3]; see also [HarLl], [HarL2], [HarL3]).
VII-1. Liapounoff's type numbers

In order to measure the exponential growth of a function, let us introduce Lia-
pounof''s type numbers.
Definition VII-1-1. Let f (t) be a C"-valued function whose entries are continu-
ous on an interval Z = {t : to < t < +oo}. Let us denote by A the set of all real
197
198 VII. ASYMPTOTIC BEHAVIOR OF LINEAR SYSTEMS
numbers a such that exp[-at] f (t) is bounded on the interval Z. Set
+oo ifA =0,

(VII.1.1) 1 (f) = inf{o : a E A} i f A j4R and A 3 40 ,
-00 ifA=R.
The quantity A (f) is called Liapounoff's type number off at t = +oo.

Note that if a E A and a <,6, then a E A.
Example VII-1-2.
J+ (0) = -00, A (exp[-t2j) _ -00, A (tm) = 0 (for all constants m),
{ A (exp[ort]) = a, A (exp[t2j) _ +00.
Here, it was assumed implicitly that to > 0 if m < 0.

The following lemma can be proved easily.
Lemma VII-1-3. Let A (f) be Liapounoff's type number of a C' -valued function
f (t) whose entries are continuous on an interval Z = It: to < t < +oo}. Then,
(i) exp I - (. (f) + e) t] f(t) is bounded on I if e > 0, and unbounded if e < 0,
whenever A (f) # -oo,
(ii) A fi < max{A (f,) : j = 1,2,... m
(tii) A f, = A(ft), tfA(f1) > A(fi)forj=2....,m,

1<f<m
(iv) f1, f2, ... , f, are linearly independent on the interval if A if 1) , ... , A (fm)
are mutually distinct,
m
(V) A(f1f2...fm) <- FA(f,) in the case when f1i... , fm are C-valued junc-
tions.
(vi) A (P(t)f) = A (f), if the entries of an n x n matrix P(t) and the entries
of its inverse P(t)-1 are bounded on the interval Z.
The following lemma characterizes Liapounoff's type number.
Lemma VII-1-4. If A(f) is Liapounoff 's type number of a function f(t) at t =
+oo, then
log If (s) I
a(f) = lim sup
t-.+oo lt<s<oo s
2. LIAPOUNOFF'S TYPE NUMBERS OF A LINEAR SYSTEM 199
Proof.
Since there exists a positive constant K such that I f (s)I < Keia(f)+ì' for e > 0
and large values of s, it follows that
logIf(s)I <A(f)+e + logK 5.1U) +e+logtK for t < s.
Hence,
log If(3)1
< a(f)
lim sup
t-+m s
It<5<+m
Also, for a fixed positive number a and any positive integer in, there exists a large
log m +
value of sm such that If (sm)1 and lim sm = +oo. Hence,
sm
e < log If (sm)1 Therefore, we obtain

Sm
log If(s)
lim sup
t-+oo<a<+ao s
I
> (f )
Thus, Lemma VII-1-4 is proved.
VII-2. Liapounoff's type numbers of a homogeneous linear system

In this section, we explain Liapounoff's type numbers of solutions of a homoge-
neous linear system
dy
(VII.2.1) = A(t)9
dt
under the assumption that the entries of the n x n matrix A(t) are continuous and
bounded on an interval T = It : to < t < +oc}. Let us start with the following
fundamental result.
Theorem VII-2-1. If y" = fi(t) is a nontrivial solution of system (VII. 2. 1) and if
f A(t)e < K on the interval I = It : to < t < +oo} for some non-negative number
K, then
(VII.2.2)
Proof.
Let the n x n invertible matrix 4P(t) be the unique solution of the initial-value
dY
problem 7 = A(t)Y, Y(to) = In, where In is the n x n identity matrix. Then,
dZ
4i(t)-1 is the unique solution of the initial-value problem = -ZA(t), Z(to) _
dt
I (cf. Lemma IV-2-4). Since 1, we obtain
1 + KJ I4(t)Idt,
to
1 +K/ I t
,I dt
for t r- T.
Therefore,
)4i(t)I < exp[K(t - to)) and I < exp[K(t - to)] for t E T
(cf. Lemma 1-1-5). Observe that fi(t) = 4D(t)j(to) and b(to) From
the definition of the norm of a matrix, it follows that
exp[-K(t - to)] < 1 (t)I < 1 (to)j exp[K(t - to)] for t E Z.
Therefore, (VII.2.2) follows, since I¢(to)I 0 0.

Set A = {a Q is a nontrivial solution of (VII.2.1)}. Then, (iv) of Lemma
VII-1-3 implies that A is a nonempty subset of R which contains at most n numbers.
Let Al > A2 > ... > Am (1 < m < n) be all of the distinct numbers in A. Set
(VI1.2.3) Vj = {j : is a solution of (VII.2.1) such that A (p) < A-}
for j = 1,2.... Tn. Then, (ii) and (vi) of Lemma VII-1-3 imply that V, is a vector
space over C. Set
(VII.2.4) 1i = dim Vj (j = 1,2,... ,m).
The following lemma states that there exists a particular basis for each space V,
which consists of y, solutions whose type numbers are equal to A,.
Lemma VII-2-2. For system (VI1.2.1) and y, given by (VII. 2-4), it holds that
(_) yi = n,
(ii) ym < -Ym-1 < ... < y,
(ii:) for each j, there exists a basis for V, that consists of y, linearly independent
solutions d,,- (v = 1,2,... , y,) of (VIL2.1) such that A (p,,°) = A3.
Proof
It is easy to derive (i) and (ii) from the definition of V, and the definition of
the numbers at, ... , am. To prove (iii), let y,,,, (v = 1, 2, ... . y,) be a basis for Vj.
Assume that
A, for v = 1,...t,
(
for v=f+1...... ).
It follows from (ii) of Lemma VII-1-3 and definitions of V, and A, that t > 1. Set
for v = 1.... , t,
+ for v = t + l.... , y2.
Then, , , (v = 1, ... , y,) satisfy all the requirements of (iii).
2. LIAPOUNOFF'S TYPE NUMBERS OF A LINEAR SYSTEM 201
Observation VII-2-3. The maximum number of linearly independent solutions

of (VII.2.1) having Liapounoff's type number A,, is rye.
Definition VII-2-4. The numbers Al, A2, ... , Am are called Liapounoff's type
numbers of system (VIL2.1) at t = +oc. For every j = 1, 2,..., m, the multiplicity
of Liapounoff's type number A, is defined by
for j=1,2,... in- 1,

(VII.2.5) hj _ for j = in.
7m
The structure of solutions of (VI1.2.1) according with their type numbers is given
in the following result.
Theorem VII-2-5. Let {(1, ¢2r ... , iin} be a fundamental set of n linearly in-
dependent solutions of system (VII.2.1). Then, the following four conditions are
mutually equivalent:
(1) for every j, the total number of those 4t such that A (y5t) = A, is h, (cf.
(VII. 2.5)),
(2) for every j, the subset {¢t : A (Qt) < A, } is a basis for V,,
(3) A ctt > A) if the constants ct are not all zero,
(4) A = max {A (dt) : ct 0} for every nontrwzal linear combi-

n
nation FC, it of e Wn}
t=1
Proof.
Assume that (1) is satisfied. Then, the total number of those ¢t such that
m
A (3t) < A, is Fht = ryr = dims Vj. Hence, (2) is also satisfied. Conversely,
t=j
assume that (2) is satisfied. Then, the total number of those t such that A (fit) <
Aj is equal to dime V) = -,. Hence, (1) is also satisfied (cf. (VII.2.5)).
Assume that (2) is satisfied. Then, if A ct4 A, for some ct, it
follows that ctc E V,, and hence
clot = tt
for some c e. Since {¢j : t = 1,2,... , n} is linearly independent, all constants cc

must be zero. Thus, (3) is satisfied.
n
Assume that (3) is satisfied. Write a linear combination I:ct4t in the form
t=1
m
cj¢t. Then, A CWI = A), if ci4r 96 0. Hence, (4)
j=1 a(ft)=A, a(mt)=a1 a(Ot}=a,
follows from (iii) of Lemma VII-1-3.
Finally, assume that (4) is satisfied. Then, every solution d of (VII.2.1) with
A (j) < A, must be a linear combination of the subset {¢t : A ((rt) S Aj }. Hence,
(2) is satisfied. 0
Definition VII-2-6. A fundamental set (01,02, ... , On } of n linearly independent
solutions of system (VIL2.1) is said to be normal if one of four conditions (1) - (4)
of Theorem VII-2-5 is satisfied.
Since V. C Vm-1 c . C V2 C V1, it is easy to construct a fundamental set
of (VII.2.1) that satisfies condition (1) of Theorem VII-2-5. Thus, we obtain the
following theorem.
Theorem VII-2-7. If the entries of the matrix A(t) are continuous and bounded
on an interval Z = It : to < t < +oo}, system (VIL2.1) has a normal fundamental
set of n linearly independent solutions on the interval Z.
Example VII-2-8. For a system = Ay" with a constant matrix A, Lia

pounoff's type numbers A1, A2, ... , Am at t = +oo and their respective multiplicities
h1, h2, ... , hm are determined in the following way.
Let IL I , p2, ... , Pk be the distinct eigenvalues of A and M1, m2, ... , Mk be their
respective multiplicities. Set v, = R(pj) (j = 1, 2,... , k). Let Al > A2 > > Am
be the distinct real numbers in the set {v1i v2, ... , vk }. Set h, _ mr for
t=at
dy
j = 1, 2, ... , m. Then, A1, A2, .... Am are Liapounoff's type numbers of = Ay"
at t = +oc and h1, h2, ... , hm are their respective multiplicities. The prom of this
result is left to the reader as an exercise.
Example VII-2-9. For a system
dy
(VII.2.6)
dt = A(t)y
with a matrix A(t) whose entries are continuous and periodic of a positive period c..'
on the entire real line R, Liapounoff's type numbers )k1, A2, ... , Am at t = +oo and
their respective multiplicities h1, h2, ... , hm are determined in the following way:
There exists an n x n matrix P(t) such that
(i) the entries of P(t) are continuous and periodic of period w,
(ii) P(t) is invertible for all t E R.
3. CALCULATION OF LIAPOUNOFF'S TYPE NUMBERS 203
(iii) the transformation y = P(t)z changes system (VII.2.6) to
(VII.2.7)
with a constant matrix B.

Therefore, (vi) of Lemma VII-1-3 implies that systems (VII.2.6) and (VII.2.7) have
the same Liapounoff's type numbers at t = +oo with the same respective multiplic-
ities. Liapounoff's type numbers of system (VII.2.7) can be determined by using
Example VII-2-8. Note that if pl, p2,. .. , p" are the multipliers of system (VII.2.6),
then p, log[p,] (j = 1, 2,... , n) are the characteristic exponents of system
(VII.2.6), i.e., the eigenvalues of B, if we choose log[p3] in a suitable way. Hence,
R(µ1) log[[p2I] (j = 1, 2,... , n).
Those numbers are independent of the choice of branches of log[p,].
Example VHI-2-10. For the system =I the fundamental set {et

J [b],
eu [0] I is normal, but the fundamental set
I {[] [J}
-e is not normal.
2t ,
VII-3. Calculation of Liapunoff's type numbers of solutions

The main concern of this section is to show that Liapounoff's type numbers of
a system J = B(t)y" at t = +oc and their respective multiplicities are exactly the
same as those of the system = Ay" with a constant matrix A if iimoB(t) = A.
dt t-+0
It is known that any constant matrix A is similar to a block-diagonal form
diag[pllm, + W.421,., + Rf2i ... , µkl,,, + Mk],
where µl, ... , Pk are distinct eigenvalues of A whose respective multiplicities are
m1, ... , mk, is the mj x mj identity matrix, and M1 is an mj x m, nilpotent
matrix (cf. (IV.1.10)).
Consider a system of the form
(VII.3.1) 10 = A2y1 + EBjt(t)Ut (j = 1,2,...

t=1
where y2 E C'-,, A. is an n, x n) constant matrix, and B2t(t) is an n, x nt matrix

whose entries are continuous on the interval Za = {t : 0 < t < +oo}, under the
following assumption.
Assumption 1.
(i) For each j, the matrix A. has the form
(VII-3.2) Aj = )1In, + E. + N., (j = 1,2,... ,m),
where Al is a real number, In, is the nJ x nl identity matrix, E. is an n. x nj

constant diagonal matrix whose entries on the main diagonal are all purely
imaginary, and Nj is an nj x n, nilpotent matrix,
(ii) Am < )1m-1 < ... < 1\2 < Al,
(iii) N,E, = E?NJ (j = 1,2,... ,m),
(iv) t limp B,t(t) = 0 (j,t = 1,2,... ,m).
The following result is a basic block-diagonalization theorem.
Theorem VII-3-1. Under Assumption 1, there exist a non-negative number to
and a linear transformation
(VII-3.3) y , = z ' , + I: T , (t)zt (j = 1,2,... ,m)

tj
with ni x nt matrices T,t(t) such that
(1) for every pair (j, t) such that j f, the derivative dt tt (t) exists and the
entries of Tat and dj tt are continuous on the intervalI = (t : to < t < +oo)
(2) lim T;t(t) = O (j # e),
t-+00
(3) transformation (VII.3.3) changes system (VI1.3.1) to
(VII.3.4) t Lzj A,+B,,(t)+EBjh(t)Th,(t)I (j=1,2,...,m).
Proof.
We prove this theorem in eight steps.
Step 1. Differentiating both sides of (VII.3.3), we obtain
dzl T,tdz"t dTit5t

+
dt dt t#? df
m
(VII.3.5) = A) % + E Tj tat
t0r t=1#t
Bjt t + Tt V
Aj 4' Bjj +
h*
BjhThj i, +
t AjTjt + Bjt +
hit
BjhTht zt,
where j = 1, 2,... , m. Note that

in
r
n+ 1
2, = F BjnTnv zv
e=1 Vol v=1 h#v
Rewrite (VII.3.5) in the form
-[At+Fj }+ET,t{ Lz' -[At+Fe]zt}

{ JJJJJJ t#i l 1
_ [AJTJt + Bit + BlhThe - Tlt (At + Ft)

dtt Ztf
t#, h#t
where j = 1, 2,... , m, and
F t = B» + B)hTh) M)
h#j
Define the Tit by the following system of differential equations:
(VII.3.6) T (.l r e)
dtt h#t
Then,
{ -[A,+F,]x",}+Tj dt -[At+Ft]zt}=Q (j=1,2,...,m).

This implies that we can derive (VII.3.4) on the interval I if the Tit satisfy (VII.3.6)
and condition (2) of Theorem V1l-3-1, and to is sufficiently large.
Step 2. Let us find a solution T of system (VII.3.6) that satisfies condition (2) of
Theorem VII-3-1. To do this, change (VII.3.6) to a system of nonlinear integral
equations
(VII.3.7) T,t(t) = exp[(t - s)A,]U,t(s,T(s))exp[-(t - s)A1]ds,

J
for j 0 e, where the initial points rat are to be specified and
U,t(t,T) = B11(t) + E Bjh(t)Tht - Tjt Btt(t) + E Bth(t)Tht

h961 h#t
Using conditions given in Assumption 1, rewrite (V11.3.7) in the form

rt 1
(VII.3.7') Tit(t) = J exp 2(a1 - t)(t - s)] s,s,T(s})ds,
where j t and
Wjt(r,s,T)
(VII.3.8)
= exp [(A, - At)r] exp[rEj j exp[rNj[UJt(s, T) exp[-rEtj exp[-rN,].
On the right-hand side of (VII.3.7'), choose the initial points rat in the following
way
(VII.3.9) ra t
Step 3. Let us prove the following lemma.
( +00
j`
to
if Aj > At
if A < A 1t..,. (i.e., j < t),
(ie j > t)
Lemma VII-3-2. Let a be a positive number and let f (t, s) be continuous in (t, s)
on the region Z x Z, where Z = {t : to < t < +oo}. Then,
(VII.3.10)
III ft"
t exp[-a(t - s)] f (t, s)dsl < - max l f (t, s)to
-s<t y
J
for each fixed t E T. Also, if I f (t, s)' is bounded for to < t < s < +oo, then
(VII.3.11) rt
J+00
exp[a(t - s)j f (t, s)dsl <_ 1 sup { I f (t, $)I : to < t < s < +oo}
a
I
for each fixed t E Z.

Proof
In fact, estimates (VII.3.10) and (VII.3.11) follow respectively from
rt
exp[-a(t - s)]ds = a {1 - exp[-a(t - to)]} <
Jto
t
exp[a(t - s)]ds = - 1a . O
+oo
Step 4. Let us estimate s, T) fort < s < +oo if A j > at, and for to < s < t
if A_, < at. If A, > At, r < 0, and s > t, it follows that
Irlm'+mi-2)
i1',t(r,s,T)I < Ko(l + exp 12(A, - A,) 7-1 IVit(s,T)I
for some positive number 1Co. It can be shown èasily that there exist a positive
number X and a non-negative valued function p(t) such that
I + Ir1m'*m`_2 exp [(A1 - At)r < X, ] for r < 0,
B, (s) Q(t) for t < s < +oo and all pairs (p, q),
t-+00
lim p(t) = 0.
Hence, if Aj > At, r < 0, and s > t, we obtain IW,t(r,a,T)I < /C2,3(t) {1 + ITI2}
for some positive number 1C2, where ITI = maxIT,t[. Similarly, the estimate
t#j
I Wjt(r, s, T) I < 1C219(to) { 1 + ITI2) is obtained by choosing a positive number 1C2
sufficiently large, if A., < at, r > 0, and to < s < t.
Step 5. In a manner similar to Step 4, we can derive the following estimates:
IWj1(T,s,T)
- Wj1(r,s,T)I
K3$(t) {1 +ITI + ITI} IT - TI if A, > A1, T < 0, s > t,
1 K3$(to){1+IT{+ITI}IT-T{ if A<A1, T>0, to<s<t

for some positive number K.
Step 6. Define successive approximations on the interval I as follows:
TO ye(t) = 0,
TP1t;)t(t) =
f [ exp
i..
[(A, - A,)(t - s)j 9,1(t - s, s, Tp(s))ds,
where j 54 a and p = 1, 2, .... Suppose that
(VU.3.12) I Tp;j,(t) j < C on the interval I = ft: to < t < +oo}

for some positive number C. Then, Lemma VII-3-2 and Step 4 imply that
2 K2i3(t){1 + C2} if A) > A,,

I TP .iat(t) I <_ A'2- Al
K21300){I +C2} if A, < '\1
Al Aj
on the interval Z. Hence, choosing to so large that
2 K2/3(t){1 + C2} < C on Z,

IA, - A,I
we obtain
ITp+1,je(t)j < C on the interval I
from (VII.3.12).
Step 7. Suppose that (VII.3.12) holds for p = 0, 1, 2,.... Then,
lTp+i.t(t) - Tp,1(t)I S 2 aup ITP(s) - TP-1(s)I <_ C on Z,

WP
where ITpI = i ITpael if to is so large that K30(t){1 + 2C} < on Z. Since
P
Tp.e(t) _ {TQae(t) - TQ_lae(t)}, it can be shown easily that lim Tpe(t) _
Tje(t) exist for all (j, 1) such that j # f uniformly on the interval T. The limit
Tj1(t) satisfies integral equation (VII.3.7).
Step 8. In this final step, we prove that the bounded solution Tjt of (VII.3.7)
satisfies condition (2) of Theorem VII-3-1. It easily follows from Steps 6 and 4 that
1
t lim o exp [(A3 - at)(t - s)1 ir4jt(t - s, s,T(s))ds = 0

1+ 00
if > X1.
For (j, f) such that ), < at, write the right-hand side of (VII.3.7') in the following
form:
t s)] IV1t(t - s, s,T(s))ds
Jto exp 12 (A
Q 11
fexp L(A - a)(t - s)] - s, s,T(s))ds

or
r 1
+ t exp 12 (At - at)(t - s)] bt%tt(t - s, s, T(s) )ds
Jo l
for any a such that to < or < t. Observe that
fexp - At)(t - s)] Wjt(t - s,s,T(s))dso
< exp L2 (.1j - '\t)(t - a)] 12 (a) - at)(a - s)J 14 t(t -- s, s, T(s))ds
I 1'. a exp `
2KO(t°)
< {l + C2} exp 1(aI - AE)(t - a) J ,
At - A) 1:5
if JT(t)[ < C on I (cf. (VII.3.10)). Observe also that
f exp L2(,\j - at)(t -

j
s)]ji)t(t - s,s,T(s))dsl A i3(a) {1 +C2}
if T(t)j < C on I (cf. (VII.3.10)). Hence,
jexp
t 1(A3 - a)(t - s)] - s, s, T(s))dso
2K1\1 211-
< C2} {(to)exP {(A, - At)(t - )1 +
J
if T(t) I < C on T. Therefore, letting a and t tend to +oo, we obtain

rt
lim I exp 12 (A, - At)(t - s)] Yt,jt(t - s, s, T(s))ds = 0.
t-+3o to L
This completes the proof of Theorem VII-3-1.

In order to find Liapounoff's type numbers of system (VII.3.4), it is necessary
to establish the following lemma.
Lemma VII-3-3. If N is a nilpotent matrix, then for any positive number e such
that 0 < e < 1, there exists an invertible matrix P(e) such that
P(e)-INP(e)I < Xe
for some positive number IC independent of e.
Proof.
Assume without any loss of generality that N is in an upper-triangular form with
zeros on the main diagonal (cf. Lemma IV-1-8). Set A(e) = diag[1, e, ... , e"-1].
Then, A(e)-'NA(e) = [ ek-Jz,k ] (a shearing transformation). Hence, as 0 < e <
1 and j < k, we obtain IA(e)-1NA(e)I < e(N(. 0
Let us find Liapounoff's type numbers of system (VII.3.4), i.e.,
dzj
a. = IA + B,,(t) + z, (7 = 1,2,... ,m).
h*j
Theorem VII-3-4. A system of the form

dz
= (AIn+E+N+B(t)Ja
dt
has only one Liapounoff's type number ,\ at t = +oo if
(i) A is a real number,
(ii) In is the n x n identity matrix, E is an n x n constant diagonal matrix whose
entries on the main diagonal are all purely imaginary, N is an n x n constant
nilpotent matrix, and EN = NE,
(iii) the entries of the n x n matrix B(t) are continuous on the interval I = {t :
to < t < +oo},
(iv) t lim B(t) = 0.
+00
Proof.
Set i = eaèÈU. Then,
(VII.3.13) dt = [N + C(t)) u, C(t) = e-ÈB(t)etE

Let p be any Liapounoff's type number of (VII.3.13) at t = +oo. Then, (µ( <
sup(N + C(t)( (cf. Theorem VII-2-1). This is true even if to -+ +oo. Since
tez
lim C(t) = 0, we conclude that 1µ1 < (N(. Using Lemma VII-3-3, it can be
t+00
shown that any Liapounoff's type number p of (VII.3.13) at t = +oo is zero. This,
in turn, completes the proof of Theorem VII-3-4. 0
Remark VII-3-5. For a nilpotent matrix N, the two matrices N and eN are
similar to each other for any nonzero number e. To prove this, use the Jordan
canonical form of N.
From the argument given above, we conclude that Liapounoff's type numbers of
system (VII.3.4) are at, ... ,.1,,, and their respective multiplicities are n1, ... , r4,,.
Thus, the following theorem was proved.
Theorem VII-3-6. Liapounof's type numbers of a system = B(t)y at t =

+oo and their respective multiplicities are exactly the same as those of the system
dy
dt= Ay with a constant matrix A if the entries of the matrix B(t) are continuous
on the interval Io = {t : 0 < t < +oo} and lim B(t) = A.
t-.+ao
Applying Lemma VII-1-4 to the solutions of the system = B(t)y', we obtain

the following corollary of Theorem VII-3-6.
Corollary VII-3-7. If an n x n matrix B(t) satisfies the conditions
(i) the entries of B(t) are continuous on the interval To = {t : 0 < t < +oo},
(ii) lim B(t) = A exists,
t-.
dy log 1-0(t)I
then, for every nontrivial solution ¢(t) of the system = B(t)y, lim
dt t--+oo t
p exists and p is the real part of an eigenvalue of the matrix A.
Remark VII-3-8. The conclusion of Theorem VII-3-1 still holds even if condition
(iv) of Assumption 1 of this section is replaced by IB)k(t)j < f (t) (j, k = 1, 2,... , m)
on the interval Zo, where f (t) satisfies
rP
(VII.3.14) sup(1 +p- t)-1 J f(s)ds - 0 as t +oo.
p>t t
To see this, let us assume that a positive-valued function f (t) satisfies condition
tP
(VII.3.14). Set h(t) = sup ((1 + p - t)-1 ( f (s)ds) for a fixed positive number
p>t \\ ft
t. Also, set E(t) = suph(r). Then, lim E(t) = 0, and E(t2) < E(t1) if t1 < t2.
r> t t--+oo
Now, it is sufficient to prove the following lemma.
Lemma VII-3-9. For any positive numbers t, to, and c, it holds that
(a) rt e-c(t-a) f(s)ds < (I + -) E(to),

to
f + ec(t-s) f(s)ds < 1+ cE(t).
lim
t-.+oo
f t e-`(t-`) f (s)ds = 0.
o
Proof of (a).
Set p(t) = f e-`(t-°)f

t (s)ds. Then, 0'(t) = -cb(t) + f (t) and, hence,
I.
0(r) - 0(o) = -c / T th(s)ds + / T f (s)ds
0 0
for to < a < r. Suppose that there exists a positive number 6 such that
b(r) = C1 + - +6 E(to), 0(a) = C- + 6 E(to),
and
0(s) > (!± b E(to) for a < s < r.
Then,
E(to) + c l! + 6 E(to)(r - a) < E(a)(1 + (r - o)).
This is a contradiction. \
Proof of (b).
Set ,L(t) = f et-f (s)ds for 0 t T for a fixed T > 0. Then, "(t) _
r
cO(t) - f (t) and, hence, w(r) - ?P(t) = cJ ;(s)ds J f (s)ds for t < r < T.
r
Suppose that there exists a positive number 6 such that t
><i(r) = I ! + 6) E(t), y(t) = (1 + 1 +6) E(t),
and
1l'(s) > (!+o)E(t) for t < s < r.
Then,
E(t) + c (1+ 6) E(t)(r - t) < E(t)(1 + (r - t)).
This is a contradiction. This, in turn, proves that
T
f (s)ds < C1 + c j E(t) for T > t.
J
+430
Since lim E(t) = 0, the integral + e'('-")f (s)ds exists and
t-+o ft
j+00
ec(t-e)f(s)ds < 1+ 1 f E(t).
e \ c
The proof of (c) is left to the reader as an exercise. 0

For the argument given above, see [Harl).
VII-4. A diagonalization theorem

Liapounoff's type number of a solution is useful information since it provides
some idea about the behavior of the solution as t --. +oo. However, it is not quite
enough when we look for a more specific information. For example, Liapounoff's
type number of the second-order differential equation
x
dt2 + {1 + R(t)}r, = 0
is 0 and its multiplicity is 2 if the function R(t) is continuous on the interval 0 < t <
lim R(t) = 0 (cf. Theorem VII-3-6). However, this does not imply the
+oc and t-+"0
boundedness of solutions as t --+ oc. 0. Perron [Per3[ constructed a function R(t)
satisfying the two conditions given above in such a way that solutions of (VII.4.1)
are not bounded as t +oc. Looking for better information concerning equation
(VII.4.1), M. Hukuhara and M. Nagumo [HNI[ proved the following theorem.
Theorem VII-4-1. Every solution of differentud equation (VII.4.1) is bounded as
1
+C0
t +x, if + jR(t)ldt < +oc.
0
Proof.
r+x
First, fix to > 0 in such a way that a jR(t)ldt < 1. Write a solution 0(t)
SU
of (VII.4.1) in the form
b(t) = q(to) cost - to) + 0'(to) sin(t - to) - J t R(s)¢(s) sin(t - s)ds.
to
Choose a positive number K so that 4(to)j + kd'(to)j < K and choose another
KQ
positive number At so that M > 1 > K. Then,
K + At J t JR(s)jds < At for to < t < tt

to
if jb(t)j < At for to < t < t1. Hence, 14(t)I < M for to < t < +oc. 0
In this section, we explain the behavior of solutions of a system of linear dif-
ferential equations under a condition similar to the Hukuhara-Nagumo condition.
Precisely speaking, we consider a system of the form
(VII.9.2) [A(t) + R(t))1i

d
under the following assumptions.
4. A DIAGONALIZATION THEOREM 213
Assumption 2. Assume that A(t) is an n x n diagonal matrix

(VII.4.3) A(t) = diag(A1(t), A2(t), ... , An (t)],
R(t) is an n x n matrix whose entries are continuous on the interval Zo = {t : 0 <
t < +00)1 and
(VII.4.4) IR(t)Idt < +oo.
J
Set
(VII.4.5) A,k(t) = al(t) -)tk(t) and D-,k(t) = R(AJk(t)) (j,k = 1,2,... ,n).
Concerning the functions A. (t) (j = 1, 2,... , n), the following is the main assump-
tion.
Assumption 3. The functions Al (t), A2(t), ... , A,, (t) are continuous on the inter-
val 4. Furthermore, for each fixed j, the set of all positive integers not greater
than n is the union of two disjoint subsets P.t and P32, where
(i) kEP) I if
lim I Djk(r)dr = -oo and Dik(r)dr < K for 0 < s < t
t-+oo 0 Li
for some positive number K,
(ii) k E Pj2 if
`t
f/a
Dik(r)dr < K for s > t > 0
for some positive number K.

Remark VII-4-2. Assume that the functions A1(t), ... , An(t) are continuous on
the interval Z o and that lim A, (t) = pj (j = 1, 2, ... , n) exist. Then, the functions
t +oo
A1(t),... ,An(t) satisfy Assumption 3 if the real parts of pl,... p, are mutually
distinct. The proof of this fact is left to the reader as an exercise.
The main concern in this section is to prove the following theorem due to N.
Levinson.
Theorem VII-4-3 ([Levil]). Under Assumptions 2 and 3, there exists an n x n
matrix Q(t) such that
(1) the derivative dQ(t) exists and the entries of Q and Q are continuous on
d
the interval Z0,
(2) t1 1 Q(t)
(VII.4.6) y' = V. + Q(t)] E
changes system (VII.4.2) to
d"
(VII.4.7) dt= A(t)i
on the interval Z0, where I is the n x n identity matrix.
Prool
We prove this theorem in six steps.
Step 1. Differentiating both sides of (VII.4.6), derive
a + (In + Qj dr = [A(t) + R(t)J (I + Q) z.

Then, it follows from (VII.4.7) that Q should satisfy the linear differential equation
(VII.4.8) dt _ (A(t) + R(t)1[I,, + Q] - (In + Q] A(t)

or, equivalently,
(VII.4.9) dQ = A(t)Q - Q A(t) + R(t) (In + Qf .

The general solution Q(t) of (VII.4.9) can be written in the form
(VII.4.9') Q(t) ='Nt)C41(t)-1 + f0(t)O(s)-1R(s)Ì,(s),I,(t)-1ds,
where C is an arbitrary constant matrix, 44(t) is an n x n fundamental matrix of

Al
dt = [A(t) + R(t)]4>, and 'F(t) is an n x n fundamental matrix of = (cf.
Exercise IV-13). Thus, the general solution Q(t) of (VII.4.9) exists and satisfies
condition (1) of Theorem VII-4-3 on 10. Therefore, the proof of Theorem VII-4-3
will be completed if we prove the existence of a solution of (V1I.4.9) which satisfies
condition (2) of Theorem VII-4-3 on an interval I = {t : to < t < +oo} for a
large to.
Step 2. Now, construct Q(t) by using equation (VII.4.9) and condition (2) of
Theorem VII-4-3. To do this, let 4?(t. s) be the unique solution of the initial-value
problem
dY = A(t)Y,
Y(s) = In.
dt
Then, (VII.4.9) is equivalent to the following linear integral equation:
(VII.4.10) Q(t) = J 1 (t, s)R(s) f 1 + Q(s)145(t, s)'1ds,
where
4?(t, s) = diag(Fl (t, s), F2(t,1s), ... , Fn(t, s)],
F, (t, s) = exp [ I A,(r)d; { (j = 1,2,. .. ,n).

JJJ
Step 3. Letting q,k(t) and rik(t) be the entries on the j-th row and the k-th
column of Q(t) and R(t), respectively, write the integral equation (VII.4.10) in the
form
[ft
(VII.4.10') ggk(t) =
J
exp Aik(r)drr)k(s) + F ds,
11
A=1
where j, k = 1, 2,... , n and the .1jk(t) are defined by (VII.4.5). Note that
t
( J(Ajk(r)dr}
t
lexp I
I = exp I J Djk(T)dr] < eK
ll a
if
0< s< t for k E Pj1 and t< s< +oo for k E Pj2,
where j, k = 1, 2,... , n and the Djk(t) are defined by (VII.4.5). The initial points
rjk are chosen as follows:
to if k E P1I,
rjk = { +00 if k E Pj 2
for some to > 0.

Step 4. Define successive approximations by
gojk(t) = 0,
gPJk(t) = j exp l rja(S)gp-I;hk(s) ds,

t t Ajk(r)dTJ rk(S) +
LJ h=1
where p = 1, 2, .... Then, we obtain I gpuk(t) I < eK (1 + nC]J r(s)ds on the

to
interval I={t:to <t<+oc} if

(VII.4.11) Igp_I,jk(t)I < C on the interval 2,
where r(t) = max(j,k) Irjk(t)I. Using assumption (VII.4.4), choose to so large that
r+oo
e'11 + nCJ r(s)ds < C.
Jto
Then, from (VII.4.11), it follows that Igpjk(t)I < C on the interval Z.
Step 5. Similarly, if (VII.4.11) holds for p = 1, 2, ... , we obtain
P
Igp+Iuk(t) - gpjk(t)l < (2} C
on the interval I = {t : to < t < +oo} if to is chosen sufficiently large. Hence,

plim gpUk(t) = gjk(t) exists uniformly on the interval Z, and the limit satisfies
integral equation (VII.4.10').
Step 6. Let us prove that the bounded solution qjk(t) satisfies condition (2) of
Theorem VII-4-3. If k E Pj2, then Tjk = +oo. Hence,
lim q)k(t)
1im f exp f f AJk(T)d-rl [rik(s)+rih(s)chk(s)](is = 0.

LJ h=1
In the case when k E PJ1, note that
t
q)k(t) = f exp [f>tik(r)dr] [r)k(s)+r)h(s)hk(s)} ds
h=1
and
IrJ//
ft
v
exp IJ \Jk(r)dr]
L s
rik(S) +
l1 A=1
r)h(S)hk(S)
o [ft ll
=f
.\)k(T)dr] [r)k(S) +> rJh(S)hk(s)J ds
exp h=1
for any a such that to <l1a < t. Observe that
exp [ft ak(T)dr] fIrk(s)+
= eXp Lf t Dk(r)dTJ
f
s
\.k(T)dr] {r)k(S) +
exP
rJh(s)hk(s)
[fC.Jk(r)dr] [rJk(s) +
h-1
n
1
dsl
rJh(S)ghk(S)ds
h=1
r)h(s)hk(s)l1 ds
JI
( t +oo
< exp J f D)k(r)dr] eK (1 + nCj ft) r(s)ds
l o
if (q)k(t)(< C on Z. Observe also that
I
f t
exp I f t
AJk(T)dr] [r)k(s) +
h=t
rJh(s)hk(s) ds
< eK(1 + tnC} f 0

+co
r(s)ds
if lq)k(t)E < C on I. For a given positive number fix a positive number a so large
ft-
that eK (1 + nCj r(s)ds < Since
t-»+0
lim f t D)k(r)dT = -oo if k E P)
for a fixed a, we obtain
t
exp f DJk(r)dr] eK [1 + nC]
L o
1
f+ to
r(s)ds <
e
for large t. Therfore, for any positive number e, there exists such that Iq,k(t)( <
e for t > t(e). This complete the proof of Theorem IV-4-3.
Remark VII-4-4. The n x n matrix W(t, s) = [In + Q(t)j$(t, s) is the unique

solution of the initial-value problem
dY = [A(t) + R(t)]Y,
Y(s) = In + Q(s),
dt
where di(t, s) is the diagonal matrix defined in Step 2 of the proof of Theorem
VII-4-3 (cf. (VII.4.10)). Since det[%P(t,s)] i4 0 for large t, the matrix WY(t,s) is
invertible for all (t, s) in Yo x Yo (cf. Exercises IV-8). This, in turn, implies that
the matrix In + Q(t) is invertible for all t in Io.
Remark VII-4-5. Theorem VII-4-3 has been shown to be the basis for many
results concerning asymptotic integration (cf. [El, E2] and [HarLl, HarL2, HarL3]).
Remark VII-4-6. Using the results of globally analytic simplifications of matrix
functions in [GH[, H. Gingold, et al. [GHS] shows some results similar to Theorem
VII-4-3 with Q(t) analytic on the entire interval Io under suitable conditions.
Remark VII-4-7. Instead of (VII.4.4), [HX1] and [HX2] assume only the inte-
grability at t = oc for above (or below) the diagonal entries of R(t) and obtain the
results similar to Theorem VII-4-3. More results were obtained in [HX3] applying
a result of [Si9j.
The following example illustrates applications of Theorem VII-4-3.
Example VII-4-8. Let us look at a second-order linear differential equation
(VII.4.12) d 22 + p(t)rt = 0.
If we set
A(t) = 0 1l
1= do ,
0J
-P(t)
dt
equation (VII.4.12) becomes the system

dy
(VII.4.13)
dt
= A(t)y
The two eigenvalues and corresponding eigenvectors of the matrix A(t) are
'\t(t)
= i p(t)112,
911 _ [i p(t)1/2]
1
A2(t) = -t p(t)'/ pct) _ [-i pt)1/2J
Set Po(t) _ [i P(t)'/2 -ip(t)'/2J.

Then,
dPo(t) - ip'(t) [0 01
POW-1 =
-i [ip(t)'/2 11
1 ,
dt 2p(t)'/2 1 1 '
2p(t) '/2 jP(t)1/2
and
P(to pt)ti21
0
Po(t)-'A(t)Po(t) _ 1/2 _i
[i
The transformation y = Po (t) i changes system (VII.4.13) to
- dP t)j z.
- P0(t)-1 {A(t)Po(t)
Using the computations given above, we can write this system in the form
(VII.4.14) ]
.ii = {t P(t) 1/ 2 [0 01] - 49t) [ 11 11
Suppose that
(1) a function p(t) is continuous on the interval Zo = {t : 0 < t < +oo},
(2) there exists a positive number c such that p(t) c > 0 on the interval Zo,
(3) the derivative p'(t) of p(t) is absolutely integrable on Io,
Then, Theorem VII-4-3 applies to system (VII.4.14) and yields the following theo-
rem (cf. IHN2]).
Theorem VII-4-9. If a function p(t) satisfies the conditions (1), (2), and (3)
given above, every solution of equation (IV.4.12) and its derivative are bounded on
the interval Zo.
The proof of this theorem is left to the reader as an exercise. Note that condition
(3) implies the boundedness of p(t) on the interval 1 .
If p(t) is not absolutely integrable, set
(VII.4.15) z = [I2 + q(t)E]t,

where I2 is the 2 x 2 identity matrix, q(t) is a unknown complex-valued function,
and E is a constant 2 x 2 unknown matrix. Then, the transformation (VII.4.15)
changes system (VI I.4.14) to
(VII.4.16)
dt = [Iz + q(t)E]-1 { [ip(t)h12A0 - 4p(t) A11 [12 + q(t)E] dtt) E} u,
J
where
Ao = Al =
Anticipating that
(i) q(t)I and Ip'(t)I are of the same size,
(ii) Jq'(t) I and Ip"(t)I are of the same size,
choose q(t) and E so that two off-diagonal entries of the matrix on the right-hand
side of (VII.4.16) become as small as Ip'(t)I2 + [p"(t)I. In fact, choosing
q(t) = 8-ip(t)312
p'(t) , E = [01 -11
0J
S. SYSTEMS WITH ALMOST CONSTANT COEFFICIENTS 219
we obtain
(12 + q(t)E]' = 1 + I9(t)2 (12 - q(t)E(
and
q(t)E]-1 { [1(t)1/2
(12 + Ao - 4p(t) A11 (12 + q(t)E1 - dt) E}
1 jip(t)1/2A0 _ p'(t) EAo - p'(t) Al
1 + q(t) 2 8p(t) 4p(t)
+g(t}AoE+E(p,p',p") - q(t)
dtt)12}
I lp(t)1/2Ao- p,(t)12+E(p,p,p')-q(t)dq(t)12 ,
1 -r q(t) 2 I 4p(t) dt
where
E2 = -12, EAo = -AoE = 0 Il
11 0
and E(p, p', p") is the sum of a finite number of terms of the form
12
a (P (t) + $ p"(t)
p(t)h/2
with some rational numbers or and /3 and some positive integers h. Applying The-
orem VII-4-3 to system (VII.4.16), we can prove the following theorem.
Theorem VII-410. Suppose that
(1) a function p(t) is continuous on the interval Zo = {t : 0 < t < +oo},
(2) there exists a positive number c such that p(t) > c > 0 on the interval ID,
(3)
+00
(Ip(t)12 + Ip'(t)I } dt < +oo.

0
(4) limop'(t) = 0.
Then, every solution of equation (VII.4.12) is bounded on the interval 7o.
The proof of this theorem is left to the reader. Condition (3) of Theorem VII-4-10
does not imply the boundedness of p(t) on the interval Io. For example, Theorem
VII-4-10 applies to equation (VII.4.12) with p(t) = log(2 + t).
VII-5. Systems with asymptotically constant coefficients

In this section, we apply Theorem VII-4-3 to a system of the form
d9
(VII-5.1)
where A is a constant n x n matrix and V(t) is an n x n matrix whose entries and

their derivatives are continuous in t on the interval Zo = It : 0 < t < +oo} under
the following assumption.
Assumption 4. The matrix A has n mutually distinct eigenvalues p 1, u2, ... , An,
and the matrix V(t) satisfies the conditions
+00
t
lim V(t) = 0
+ Cc
and + [V'(t)[dt < +oo.
o
Let A1(t), 1\2 (t) .... , an(t) be the eigenvalues of the matrix A+V(t). Then, these
are continuous on the interval I. Furthermore, it can be assumed that
lim ),(t) = µj
t-+, (j = 1,2,... ,n).
Choose to > 0 so that A) (t),.\2(t), ... , an(t) are mutually distinct on the interval
I={t:to<t<+oo}. Set
F(t, A) = det[AII - A - V(t)].
Then, F(t, \j(t)) = 0 on II for 1 = 1,2,... ,n, and
Also,
j (t, A, (t)) + as (t, \,(t)) A (t) = 0 (j =1,2,... n)
OF
(t, \j (t)) # 0 (j = 1,2,... ,n),
on I.
8A
8F (t,
A, (t ))
since III, µ2i ... , p, are mutually distinct. Observe that .X (t) is
aF
8a (t, af(t))
linear homogeneous in the entries of the matrix V'(t). In this way, we obtain the
following lemma.
Lemma VII-5-1. Under Assumption 4, the derivatives of the eigenvalues of the
matrix A + V (t) are absolutely integrable over the interval I = {t : to < t < +oo},
i. e.,
JA(t)dtl
w
< +oo (j = 1,2,... ,n).
An eigenvector p, (t) of the matrix A + V(t) associated with the eigenvalue as(t)
can be constructed in the following manner. Observe that the characteristic poly-
nomial of A + V(t) can be factored as
F(t, \) = (A - \1(t))(A -1\2(t)) ... (A - \.(t))

Hence, by virtue of Theorem IV-1-5 (Cayley-Hamilton), we obtain
(VII.5.2) (A+V(t)-.11(t)In)(A+V(t)-.\2(t)In)...(A+V(t)-an(t)IJ) = 0
5. SYSTEMS WITH ALMOST CONSTANT COEFFICIENTS 221
on Zo. Furthermore, if t > to,
J1 (A+V(t) - Ah(t)In) # 0,
h 0j
lim fl(A + V(t) - Ah(t)In) = H(A - phln) 36 0,
t-+oo
h#l h#j
for j = 1, 2,... , n. From (VII.5.2), it follows that
[A + V(t)] 1(A+ V(t) - Ah(t)In) = A,(t) [I(A + V(t) - Ah(t)In)

h#J h#)
Hence, choosing a suitable column vector j,(t) of H(A+V(t)-Ah(t)In), we obtain

h#)
p3(t) 1-1 0 and [A+V(t)]p,(t) = \j(t)jYj(t)

on the interval I = It : to < t < +oo} if to > 0 is sufficiently large. Furthermore,
lim j5, (t) = 4,

t+oo (.7 = 1,2,... n)
are eigenvectors of the matrix A associated with the eigenvalues p. (j = 1, 2,... , n),
respectively. Observe that the entries of the vectors pF,(t) (j = 1,2,... , n) are
polynomials in the entries of V(t) and A1(t),... ,An(t) with constant coefficients.
Hence,
+00
Ip"j'(t)I dt < +oc (j = 1,2,... ,n).
Jta
Thus, we proved the following lemma.
Lemma VII-5-2. Under Assumption 4, them exists a non-negative number to
such that
(1) the matrix A + V (t) has n mutually distinct eigenvalues A1(t)...... n (t) on
the interval I = It : to < t < +oo},
(2) the eigenvalues Al(t), ... , An(t) are continuously differentiable on I and
lim Aj(t) = p, (j = 1,2,... ,n),
(3) the matrix A + V(t) has n eigenvectors p1(t),... ,#n(t) associated with the
eigenvalues \I(t) .... ,An(t), respectively, such that lim 73 (t) = 4j (j =
t-+oo
1, 2, ... , n) are eigenvectors of the matrix A associated with the eigenvalues
p) (j = 1,2,... ,n),
(4) the derivatives of the eigenvalues A,(t) (j = 1,2,... , n) and the derivatives
o f the eigenvectors p ' , (t) (j = 1, 2, ... , n) with respect to t are absolutely inte-
grable on the interval Z.
Set
Po(t) _ [ 91 (t) #2 (t) ... ( t ) [[ ,, Q = [ 91 92 ... 4n [ ,

A(t) = diag[A1(t), A2(t), ... , An (t)[, M = diag[pl, p2, ... , pn]-
Then,
lim Po(t) = Q, lim A(t) = M,
t-+oo
Po(t)-'[A + V(t)]Po(t) = A(t), Q-'AQ = M,
f +oo I Po(t)_ 1
dP°(t) I
dt < +oo.
to dt
Observe that the transformation y"= P°(t)z changes system (V11.5. 1) to
dl t)1 i= [A(t) - Po(t)-'d t)]
dt = POW-1 [[A + V(t)]P0(t) - Y.
Applying Theorem VII-4-3 to this system, we obtain the following theorem.

Theorem V11-5-3. Under Assumption 4, if the eigenvalues A1(t)...... n (t) of
the matrix A + V (t) satisfy all requirements given in Assumption 3 (cf. § VII--4) on
the interval .7 = {t : to < t < +oo}, a fundamental matrix solution of (VII.5.1) will
be given by
1(t) = P°(t)[In + H(t)]exp I / ( A(s)ds],
where H(t) is an n x n matrix whose entries are continuously differentiable on the
interval I andt+00
lim H(t) = O.
Remark VII-5-4. 1
r:
(a) As t -. +oo, the matrix exp [ - J A(s)ds] 4i(t) approaches the matrix Q,
(b) we can prove a result similar to Theorem VII-5-3 even if system (VII.5.1) is
replaced by
d = (A+V(t)+R(t)]y,
where the matrix A + V (t) satisfies all the requirements of Assumption 4 and
the entries of the matrix R(t) are absolutely integrable for t > 0, i.e.,
f+00 JR(t)jdt < +oo.

0
Observation VII-5-5. Let us look into the case when the matrix A has multi-
ple eigenvalues. To do this, consider system (VII.3.1) under Assumption 1 given
in §VII-3. By virtue of Theorem VII-3-1, system (VII.3.1) is changed to system
(VII.3.4) by transformation (VII.3.3). Furthermore, (VII.3.4) is changed to
(j = 1, 2, ... m),
(VII.5.3)
dt [ N 3 + R 3 (t)) u 3
where
R3(t) = e-'E, B33(t)+E B3h(t)Ttj(t) etE, (j=1,2,...,m)

h #)
z3 = explt(A3I., + E, )]u3 (.) = 1, 2, ... , m).
Therefore, we look into the following two cases.
5. SYSTEMS WITH ALMOST CONSTANT COEFFICIENTS 223
Case VII-5-5-1. If N is an n x n nilpotent matrix such that N' = 0 and R(t)

is an n x n matrix whose entries are continuous on the interval 0 < t < +oo and
satisfy the condition
(VII.5.4) + jt2(T_1)IR(t)Idt < +oo,
we can construct a fundamental matrix solution 45(t) of the system

dy
(VII.5.5) = (N + R(t))y'
dt
such that lim
t -.+00
e-' N4(t) = I, , where I is the n x n identity matrix. In fact, the
dii
transformation y" = etNii changes system (VII.5.5) to = e-tNR(t)e1NU. Since
leftNI < Kjtjr-1 for some positive number K, the integral equation X(t) = In
e-R(r)erNX(r)dT can be solved easily.
j
Case VII-5-5-2. Let N be an n x n nilpotent matrix such that Nr = 0 and let
R(t) be an n x n matrix whose entries are continuous on the interval 0 < t < +oo
and satisfy the condition
+00
(VII.5.6) tr-1IR(t)jdt < +oo.
10
Exercise V-4 shows that this case is different from Case VII-5-5-1. As a matter of
fact, in this case, we can construct a fundamental matrix solution 'P(t) of system
(VII.5.5) such that
(VII.5.7)
t-+00
lira t-(k-t) (41(t) - etN) c" = 0', whenever Nkc_ = 0,
where k < r. We prove this result in three steps.

Step 1. First of all, if an n x n matrix T(t) satisfies condition (VII-5.7), then
We N'5
c' = 0 if 41(t)c = 0 for large t. In fact, noice first that lim e tr if
f!
etNt:
NPe = 0 for p = F + 1, ... , r - 1. = 0 if a(t)e = 0
Also, note that lim
t-+oo tktkl
and Nkcc = 0. Therefore, N'lE = 0 if k = r. Hence, Nr-2c' = 0. In this way,
we obtain c" = 0. Thus, we showed that if the matrix %P(t) satisfies the differential
equation (N + R(t))* and condition (VII.5.7), then %P(t) is a fundamental
matrix solution of (VII.5.5).
Step 2. To prove the existence of such a matrix W(t), we estimate integrals of
ske(t-")NR(s) with respect to s, where 0 < k < r - 1. Observe that
r1
ske(t_e)N = Sk(t - S)hNh
h1
h=0
Let us look at the quantity sk(t - s)h. Note that 0 5 k 5 r - 1 and 0 5 h < r - 1.
Case 1. k + h < r - 1. In this case, since sk(t - s)h = sk+h C t _ 11, define
- s
Sk(th'
8)h N°R(s)ds = / t 8k(th! S)h NhR(s)ds.
Jt
(hp) Sk+pth-p
Case 2. k + h > r. In this case, look at sk(t - 8)h = D-1)p
p_Q
Subcase 2(i). k + p < r - 1. In this case, sk+pth-p = Sr-I 1 t", where

\s
r-1-u=k+p,u+v=h-p. Since
v = h - p - p = (h-p) + (k+p) - (r-1) = (k+h) - (r-1) = k - (r-1-h),
we obtain 0 < v < k. Now, in this case, define
Irt
h t
j Sk+pth-PNhR(S)dS
= J+ao p) sk+Pth-PNhR(s)ds.
Subcase 2(ii). k + p > r. In this case,

Sk+pth-p = Sr-1(tlµt", where r-l+p=k+p, v -p=h -p.
Since
v = h - p + p = (h-p) + (k+p) - (r-1) = (k+h) - (r-1) = k - (r-i-h),

we obtain 0 < v < k. Also, note that h-p<(r-1)+(k-r)=k-1. Now, in
this case, define
=
; k+Pt h- PNhR{s)d s =
-3
k
h
)
r
t
o -
kh $ k+h-
Pt P Nh R(s)d$, where t > to.
From the definition given above, it follows that
t Ske(t-s)NR(S)d8 = etN / t Ske sNR(s)ds,

f', v
rt
lim I / ske(`'")NR(s)ds = O (k = 0,... , r - 1),
t+00 F. q
Setting
r
U(t) = f t ske(`-e)NR(s)ds,
6. AN APPLICATION OF THE FLOQUET THEOREM 225
we obtain
dU(t)
= NU(t) + tkR(t).
dt
Step 3. Let us construct n x n matrices Uk (t) (k = 0, 1, 2, ... , r - 1) by the integral
equations
ft
Uk(t) = Nk + 1 ske(t-")NR(s)Ui(s)ds
Then,
lira
t-++00
Uk(t) = Nk
Observe that
tkUk(t) tkNk
rt e(t_s)NR(s)skUk(s)ds.
k! k! n k!
Hence, setting
tkUk(t)
k!
k=0
we obtain r-1
k r
1(t) = eIN +
t
e(t-,)NR(s) s
ds.
Y' r k!!
k=o n
This implies that
do(t)
_ (N + R(t))W(t).
dt
It can be easily shown that
L(t)c = etNc' - Nk"
lim Urn
t-+oo tk t-+00 tic
if Nk+16= 0. Thus, the construction of W(t) is completed.

Case VII-5-5-2 was given as Exercise 35 in [CL, p. 1061.
VII-6. An application of the Floquet theorem

The method discussed in §VII-5 does not apply directly to the scalar differential
equation
d-t2 + {1 + h(t)sin(at)} ri = 0
when h(t) is a small function such as t112,

i , l01 t, and sin(tt/2) since the
log
g t g
derivative of h(t) sin(at) is not absolutely integrable over the interval 0 < t < +oo.
In this section, using the Floquet theorem (cf. Theorem IV-4-1), we eliminate
periodic parts of coefficients so that Theorem VII-4-3 applies. Keeping this scheme
in mind, consider a system of the form
djj
(VII-6-1) = A(t,h(t))y
under the following assumption.
Assumption VII-6-1.
(1) The entries of an n x n matrix A(t, ej are continuous in (t, e) E R x A(r)
and analytic in f E A(r) for each fixed t E LR, where FE C'" unth the entries
et, ... , A(r) = IF: Ie < r}, and r is a positive number.
(2) The entries of A(t, e) are periodic in t of a positive period w.
(3) The entrees hl,... , h,,, of a C"t -valued function h(t) are continuous on the
interval T(to) = {t : to < t < +oo} for some non-negative number to and
lim h(t) = 6.
t+00
Let us consider the following two cases.
Case 1. The function h(t) is supposed to be continuously differentiable on T(to)
and satisfy the condition
+00
(VII.6.2) J Ih(t)Idt = +oo and f + Ih'(t)Idt < +oo.
I.
Case 2. The function h(t) is supposed to be twice continuously differentiable on

T(to) and satisfy the condition
lim
t -+00
+00 +00
(VII.6.3) r= +00, / = +oo,
0
L00I;(t)ldt
Ih"(t)Idt < +oo, J + Ih'(t)I2dt < +oo
J0 0
sin(f)
For example, the function h(t) _
satisfies (VII.6.3).
I satisfies (VI1.6.2), whereas h(t) =
f
Observation in Case 1. In Case 1, we use the following lemma.
Lemma VII-6-2. If a matrix A(x, ej satisfies conditions (1) and (2) of Assump-
tion VII-6-1, them exist n x n matrices P(t,e) and H(t") such that
(i) the entries of P(t, ) are continuous in (t, e) E R x A(f) and analytic in
FE A(f) for each fixed t E R, where f is a suitable positive number.
(ii) P(t + w, ej = P(t, t-) for (t, F) E R x ©(f),
(iii) P(t, fj is invertible for every (t, t) E R x A(f),
(iv) the entries of H(t) are analytic in FE A(f),
2ai
(v) any two distinct eigenvalues of H(0) do not differ by integral multiples of -,
w
(vi) a P(t, a exists for (t, e) E R x L(f) and given by
(VII.6.4) P(t, eM = A(t, t)P(t, e) - P(t, e)H(e)
for (t, e) E R x L(f).

Proof
In order to prove this lemma, construct a fundamental matrix solution $(t, e)
of the differential equation d = A(t, e)y by solving the initial-value problem
dX = A(t, e-) X, X (O) = I, where In is the n x n identity matrix. The en-
dt
, bog((w
tries of 4!(w, ej are analytic in A(r). Define H(e) by H(ej = and
P(t,') = 4i(t, cl exp(-tH(E)]. Then, (VII.6.4) follows.
The most delicate part of this proof is the definition of H(). Details are left for
the reader as an exercise (cf. [Sill).
Changing system (VI I.6.1) by the transformation
(VII-6-5) W = P(t, i(t))u
we obtain the following theorem.

Theorem VII-6-3. Transformation (VII.6.5) changes system (VII.6.1) to
(VII.6.6) di =H(i(t)) - P(t,h(t))-1

hj(t)aP(t,h(t)) }u.
1<j<m
Proof
In fact,
dil
= P(t, h(t))-1 {A(t(t))P(t&(t)) - dt [P(t, h(t))] } u"
iiT
(VII.6.7) = P(t, i(t))-1( A(t, h(t))P(t, h(t)) - (t, K (t))
1 < <m
Since H(e) is given by (VII.6.4), equation (VII.6.6) follows from (VII.6.7).

Observe that
r+oo
(IV.6.8) P(t, h(t))-' h(t)) dt < +oo
fro 1<j<m j
under assumption (VII-6.2). Also, observe that H(i(t)) does not contain any pe-
dH(i(t))
riodic quantities and dt is absolutely integrable. Therefore, if eigenvalues
of H(6) satisfy suitable conditions, the argument given in §VII-5 applies to system
(VII.6.6).
Observation in Case 2. Set

Bo (f) =
(VII-6-9) 8P
B1(t, E, iZ) _ -P(t, E)`1 u 8e (t, E1,
I<j<m
where µ E C' and the entries of µ are u1, P2, ... , µm. Then, differential equation
(VII.6.6) can be written in the form
(VII.6.10) dt = {Bo(h(t)) + B,
Notice that the entries of the matrix B, (t, E, µ) are periodic in t of period tv and
that B1 (t, E, 6) = 0. Furthermore, any two distinct eigenvalues of Bo (6) do not
27n
differ by integral multiples of -. Set D(r) = IE + lµI < r}.
In Case 2, the following lemma is used.
Lemma VII-6-4. There exist n x n matrices P, (t, e, {'t) and HI (E, µ) such that
(:) the entrees of P1 (t, E, ui) are continuous in (t, E, u') E R x D(r) and analytic in
(E, )i) E D(r) for each fixed t E Ii£, where r is a suitable positive number,
(i:) P, (t + w, F,)!) = P, (t, e, it) for (t, F, tt+) E 1 x D{r
(iii) P1(t, E, 6) = O for (t, i) E R x 0(r),
(iv) the entries of H1 (E, u') are analytic in (E, Et) E D(r) and H1 (E, 6) = 0 for
E E O(r),
(v) at P, (t, e, µ') exists for (t, E, p) E R x D(r) and is given by
19
Pi(t,e,u) ={Bo(e) + B1(t,e,µ)) (In + Pi(t,E,ls)}
(VII.6.11)
+ P1 (t, i,#)) { Bo(E) + Hi (E, µ) }
for (t, E, µ) E IR x D(r-), where I, is the n x it identity matrix.

Remark VII-6-5. Equation (V1I.6.11) can be simplified as
(VII.6.12)
a P1(t,Ej) = Bo(E)P1(t,E,i) - Pi(t,E., )Bo(Ej
+ {B1(t,E,N)P:(t,E,Ji) - +B1(t,,rA-)} -Hl(E,u-)-
Proof of Lemma VII-6-4.
M
Given p = (pi,... ,p,,,), where the p, are non-negative integers, denote EIpjI
=1
and 4' um by Ipl and 91, respectively. Set
Pi (t, E, u') _ gP1,p(t, E), B, (t, e,

ipl>1 $pl?1
y
Ip1?1
Then, equation (VII.6.12) is equivalent to
8
(VII.6.13) p = B0(E)P1,p - P1,pBo(E) + Q1,p(t, E) -
where
(VII.6.14) Q1,p(t,E') = E (B1,ptP1,p2 - Pi,p2Hl,ps) + B1,p.

P1+p2=p, 0p1IJ{r21>_1)
Hence,
Pi,p(t,e-) = {c( + 1 exp[-sBo(0
(VII.6.15)
x 1 Q1,p(s, ) - H1,$,(e)) exP[sBo(E))ds} exp[-t B0(Z)1,
where C(e) and H1,p(e) are n x n matrices to be determined by the condition that
P1,p(t,e) is periodic in t of period w, i.e.,
exp[wBoQ-)1C(E) - C(i) exp[wBo(E)1
- exp[wBo(E)1 exp[-sBo(e)1H1,p(i) exp[sBo(i)1ds

(VII.6.16) J0
_ - exp[wBo(E)1 exp[-sBo(e)1Ql.p(s, e) exp[sBo(t1ds.
0
It is not difficult to see that condition (VII.6.16) determines the matrices C(ep)
and Hl,p(ej. Then, the matrix P1,p(t, E) is determined by (VII.6.15). The conver-
gence of power series P1 and H1 can be shown by using suitable majorant series. 0
In the same way as the proof of Theorem VII-6-3, the following theorem can be
proven.
Theorem VII-6-6. The transformation
(VII.6.17) it = {In + Pi (t,h(t),h'(t)))v"

changes differential equation (VII.6.10) to
dv = f
H(i(t)) + H1(h(t), h'(t))
dt l
(VII.6.18) - [In + P1(t, h(t), hh'(t)),'1 I hj(t) aP1(t, h(t), h'(t})

l
+ h (t) a_ (t, h{t), h'(t))J j v ,

where P1(t, E, µ) and H1(E, µ) are those two matrices given in Lemma VII-6-4.
Observe that under assumption (VII.6.3), we obtain
In + P1(t,h(t),h'(t))j_1 8Ej
f0+00
(VII.6.19)
+ hJ (t) a. (t, h(t), h'(t))J dt < +oo.
Observe also that the matrix H(h(t))+H1(h(t), h'(t)) does not contain any periodic
terms. Furthermore, H(6) + H1(6, 6) = H(0). It is clear that the derivative of
H(h(t)) is not necessarily absolutely integrable over to < t < +oo. Therefore, in
order to apply the argument of MI-5, the matrix H(h(t)) must be examined more
closely in each application. Details are left to the reader for further observation.
EXERCISES VII
VII-l. Find the Liapounoff's type number of each of the following four functions
f (t) at t = +00:
(1) exp [t2sin (1L1) sint(dt](3) inin(exp[3t],exp[5t]sin67rt]),
]' (2) exp Lfo
(4) the solution of the initial-value problem y"-y'-6y = eat, y(O) = 1, y'(0) = 4.
VII-2. Find a normal fundamental set of four linearly independent solutions of
the system dy = Ay" on the interval 0 < t < +oo, where
252 498 4134 698

A= -234 -465 -3885 -656
15 30 252 42
-10 -20 -166 -25
Hint. See Example IV-1-18.

VII-3. Let 4i(x) be a fundamental matrix solution of the system
log(l+x) e= expx2
= x3 1+ 1+ x sin x
exp(e') cosx arctanx
on the interval 0 < x < +oo. Find Liapunoff's type number of det 4i(x) at x = +oo.
EXERCISES VII 231
VII-4. Assuming that the entries of an n x n matrix A(t) are convergent power
series in t-1, calculate lim p log I (t) for a nontrivial solution ¢(t) of the differential
equation t = A(t)9 at t = +oo.

Hint. Set t = e".
VII-5. Let = xP-'A(x)i be a system of linear differential equations such that
dY
y E C" is an unknown quantity, p is a positive integer, the entries of an n x n matrix
A(x) are convergent power series in x-1, and lim A(x) is not nilpotent. Show that
ln(l y-(Teie)j)
there exists a solution g(x) of this system such that lim is a positive
r-.+oo rp
number for some real number 6.
Hint. Set x = re!e for a fixed 0 and t = rp. Then, the given system becomes
d# e+'0
= -A(re's)y'. Note that lim e'-A(re'B)
PO e'
_ -A(oo). The eigenvalues
P r-.+oo P P
A(oo) are ei_)Iof
ei>'
p ' where the a, are the eigenvalues of A(oo). Now, apply
,
P
Corollary VII-3-7.
VII-6. For each of two matrices A(t) given below, find a unitary matrix U(t)
analytic on (-oo, oo) for each matrix such that U-' (t)A(t)U(t) is diagonal or upper-
triangular.
r 1 t 0
(a) A(t) = Ot 0J, (b) A(t) = t 1 + 2t
11
0
i
L
0 0 1+t2
Hint. See [GH].
VII-7. Show that if a function p(t) is continuous on the interval 0 < t < +oo and
lim t-Pp(t) = I for some positive integer p, the differential equation d2
t +p(t)y =
t +oo
0 has two linearly independent solutions rlf(t) such that
77.k(t) = P(t)-114(1 +o(1))exp [±ij t )ds(,

JJ
77, (t)
= P(t)114(+1 +o(1))exp [f iJ t
to
d4] ,
as t +oo, where to is a sufficiently large positive number and o(1) denotes a

quantity which tends to zero as t - +oo.
Hint. Use an argument similar to Example VII-4-8.
m-1
VII-8. Let Q(x) = x' + E ahxh and P(x, e) _ ,Tn+l + Q(x), where in is a
h=o
positive integer, x is an independent variable, and am-1) are complex
parameters. Set A(x, f) = {(0) 11. Also, let Ro, po, and ao be arbitrary but
fixed positive numbers. Suppose that 0 < po < 2. Show that the system
d:V
= A(x,f)y, y = 17121
has two solutions y, (x, f) (j = 1, 2) satisfying the following conditions:

(a) the entries of y, (x, f) are holomorphic with respect to (x, f, ao,... , am_ 1) in
the domain
D={(x,f,ao,...,am_,): rEC, 0<Ifl<ao, Iargfl<Po,
laol + ... + Iam-I I < Ro},
(b) y', (x, f) (j = 1, 2) possess the asymptotic representations
111 +o(1)JP(x,f)
exp P(t,f)1/2dtl
L fro J L it - o(1)]P(x,f)1/4
11 +o(1)] P(x,E)-1/4
exp
l
- Jro P(t,E) 1/2dt ,
f -l + o(1)]P(x, f)'/4,
respectively as x - +oc on the positive real line in the x-plane, where xo
is a positive number depending on (Ra.po.ao) and o(1) denotes a quantity
which tends to 0 as x ---. +oo on the positive real line uniformly with respect
to (f, ao,... , am-I) for IEI < ao, I arg EI < po, and Iaol + ... + lam-1l < Ro.
Hint. Use a method similar to that for Exercise V1I-7. See, also, [Nful and [Si13,
Chapter 3].
VII-9. Let A1(t)...... (t) be n continuous functions of t on the interval Zo =
{tER:0<t<+oo}such that R.eja1(t)-aJ+1(t))>1(j=1,...,n-1) on
Also, let A(t) be an n x n matrix such that the entries are continuous in t on Zo
and that
sup(1 + p - t) 1 1 p I A(s) l ds - 0 as t -+ +oo.
p>t 1
Show that there exist a non-negative number to and an n x n matrix T(t) such that
(1) the derivative (t) exists and the entries of T(x) and . (x) are continuous
E
in t on the interval 2 = It : to < t < +oo},
(2) t limo T(t) = 0,
(3) transformation (I + T(x)) z changes the system
(diag[A1(t),A2(t),... +T(t))y
dt =
to
d
= diag(A1(t) + bl (t), A2(t) + b2(t), ... , An(t) +
dt
where y E C, i E C, n functions b1(t),... ,b,,(t) are complex-valued and
continuous in t on T, and lim b,(t) = 0 (j = 1, ... , n).
t +00
EXERCISES VII 233
VII-10. Show that if R(t) is a real-valued and continuous function on Zo = {t E It :

+00
0 < t < +oc} satisfying the condition J JR(t)j < +oo, the differential equation
0
d2y
+ (1 + R(t))y = 0 has a solution q5(t) such that lim (O (t) - sint) = 0. Also,
dt2 c-+OQ
show that 0(t) has infinitely many positive zeros an such that lim = 7r.
n+00 n
n
VII-11. Show that every solution of a differetial equation tz + R(t)y = 0 has
at most a finite number of zeros on the interval Zo = {t E Lea : 0 <_ t < +oo}
if R(t) is a real-valued and continuous function on Zo satisfying the condition
+a,
tIR(t)j < +oc.
L
+00
Remark. See (CL, Problem 28 on p. 1031. For the case when f tjR(t)j = +oo,
o
f+a
but J IR(t)l < +oo, see Example VI-1-i.
0
VII-12. Suppose that u(t) is a real-valued, continuous, and bounded function of t
r+00
on the interval 0 < t < +oo. Also, assume that J ju(t)jdt < +oc. Show that if
0
X is an eigenvalue of the eigenvalue problem
d'y f+00
7t2 + u(t)y = Ay, y(O) = 0, y(t)2dt < +oo,
J0
then 0 < A < sup Ja(i)l.
o<:<+c*
VII-13. Let A(t) be an n x it matrix whose entries are real-valued, continuous,
and periodic of period 1 on R. Show that there exist two n x n matrices P(t, c) and
B(c) such that
(a) the entries of P(t, c) and B(c) are power series in c which are uniformly con-
vergent for -oc < t < +oo and small JEJ,
(b) P(t, 0) = I, (-oo < t < +oo), where In is then x n identity matrix,
(c) the entries of P(t, c) is periodic in t of period 1,
(d) P(t, c) and B(E) satisfy the equation
eP
8t = cIA(t)P(t, e) - P(t, c)B(c)j.
VII-14. Apply Lemma VII-6-2 to the following system:

dy
dt
= [nL + (a + 8cos(2t))(K - L)Jg, g_
[Y2J
where n is a positive integer, the two quantities a and,3 are real parameters, and
f0
K= U]
and L = [ 01 0]
VII-15. Using Theorem VII-6-3, find the asymptotic behavior of solutions of the
differential equation
2 + {1 + h(t) sin(at)} rl = 0, h(t) = ln(2 + t)
as t -+ +oo, where a is a real parameter.

CHAPTER VIII
STABILITY
In the previous chapter, we explained the behavior of solutions of linear systems

as t -. +oo. In this chapter, we look into similar problems for nonlinear systems.
To start with, in §VIII-1, we introduce the concepts of stability and asymptotic
stability of a given particular solution as t +oo. We illustrate those concepts
with simple examples. Reducing the given solution to the trivial solution by a
simple transformation, we concentrate our explanation on the stability property
of the trivial solution. It is well known that the trivial solution is asymptotically
stable as t +oo if real parts of eigenvalues of the leading matrix of the given
system are all negative. This basic result is given as Theorem VIII-2-1 in §VIII-2.
The case when some of those real parts are not negative is treated in §VIII-3. In
particular, we discuss the stable and unstable manifolds. In §VIII-4, we look into
the structure of stable manifolds more closely for analytic differential equations.
First we change a given system by an analytic transformation to a simple standard
form. By virtue of such a simplification, we can construct the stable manifold in a
simple analytic form. This idea is applied to analytic systems in IR2 in §VIII-6. In
§§VIII-7-VIII-10, using the polar coordinates, we explain continuous perturbations
of linear systems in J2. In §VIII-5, we summarize some known facts concerning
linear systems with constant coefficients in JR2. The topics discussed in this chapter
are also found in [CL, pp. 371-388], [Har2, pp. 160-161, 220-227], and [SC, pp.
49-96]. The materials in §§VIII-4 and VIII-6 are also found in [Du], [Huk5], and
[Si2].
VIII-1. Basic definitions

Let us consider a system of differential equations
dy
(VIII.1.1) t=
d
where y E lR" and the R'-valued function At, M is continuous on a region
A(ro) = Zo x D (ro) = {(t, y1 E J"+I : 0 < t < +oo, 1171 < ro}.
Also, assume that a solution mo(t) of (VIII.1.1) is defined on the entire interval 10
and that (t, mo(t)) E A(ro) on Yo. The main topic in this chapter is the behavior
of solutions of the initial-value problem
(VIII.1.2)
dt = f (t, o'
as t +oo. To start, we introduce the concept of stability.
235
236 VIII. STABILITY
Definition VIII-1-1. The solution do(t) is said to be stable as t - +oo if, for
any given positive number c, there exists another positive number b(c) such that
whenever I&(0) - 7 < 5(c), every solution 4(t) of initial-value problem (VIII.1.2)
exists on the entire interval Zo and satisfies the condition
(VIII.1.3) I$o(t) - (t)I < E on 10.
Remark VIII-1-2. Suppose that the initial-value problem
(VIII.1.2.r) = At' y), y(r) = it

dt
has the unique solution y = (t, r, i) if (r, rl) E 0(ro). Then, qs(t, r, 771 is continuous
with respect to (t, r, '). Therefore, for any r E Zo and any given positive numbers
T and f, there exists a positive number p(r,T,e) such that whenever 100(r) - )1 <
p(r, T, c), the solution 0(t, r, 71) exists on the interval 0 < t < T and I0o(t) -
¢(t, r, Y-7)1 < E on the interval 0 < t < T (cf. §11-1). This implies that if the solution
0o(t) is stable as t +oo, then for any r E Zo and any positive number c, there
exists another positive number b(r, c) such that whenever loo (,r) - n7 < b(r, E), the
solution (t, r, rt) exists on the entire interval 10 and (t, r, J) satisfies the condition
IV'0(t) - (¢(t, r, T-D I < E on I.
We also introduce the concept of asymptotic stability.
Definition VIII-1-3. The solution do(t) is said to be asymptotically stable as
t-4 +00 if
(i) the solution do(t) is stable as t - +oo,
(ii) there exists a positive number r such that whenever I o(0) - i7 < r, every
solution fi(t) of initial-value problem (VIII. 1.2) satisfies the condition
(VIII.1.4) lim I&(t) - (t)I = 0.
Remark VIII-1-4. Set g= E+ ¢0(t). Then, system (VIII.1.1) is changed to
(VIII.1.5) dt = At' z + do(t)) - f'(t, do(t)).
Hence, the study of the solution o(t) of (VIII.1.1) is reduced to that of the trivial
solution z1(t) = 0 of (VIII.1.5). Thus, the solution .o(t) of (VIII.1.1) is stable
(respectively asymptotically stable) as t -+ +oo if and only if the trivial solution of
(VIII.1.5) is stable (respectively asymptotically stable) as t -- +co. In the following
sections, we shall study stability and asymptotic stability of the trivial solution.
The following three examples illustrate stability and asymptotic stability.
Example VIII-1-5. The first example is the system given by
(VIII.1.6) ddtl = -yi, dt = (yl - y2)y2

1. BASIC DEFINITIONS 237
To find the general solution of (VIII.1.6), solve the first-order equation
(VIII.1.7) dye = -y2 + .

dyi y1
The transformation u = b2 changes this equation to a first-order linear differential

du 2 vt a-2n
- = 2u - -. Thus, we obtain u(yl) = ezvi c - 2'I
equation7Y1 dq . Since
yl rl
u = yz > 0, the quantity c must be positive. It is evident that y1(t) = y e-t
satisfies the first equation of system (VIII.1.6) with the initial value yt(0) = y.
Therefore,
1
(VIII.1.8) yl(t) = -( e-t, 112(t) = ±
u(yi(t))
is a solution of (VIII.1.6). Observe that

y<yl(t)<0 if y<0 and 0<y1(t)<y if y>0
ry, (t)-2 e-2Yt(t)
for t > 0. Therefore, dt) < 0 for t > 0. Thus, y2(t)2 < C
J
J. 77
e2171
= e21',1e2'ry2(0)2 for t > 0. This proves that the trivial solution of system
c
(VIII.1.6) is stable as t +oc. Furthermore, since
fvt(t) a-2q fYi a-2n
lim J
t-+o0
-dt7 = lim
r) Y1 -0
di = -oo,
11
the trivial solution of system (VIII.1.6) is asymptotically stable as t -- +oo. This

result is shown also by Figure 1.
Y2
FIGURE 1.
Example VIII-1-6. The second example is a second-order differential equation
(VIII.1.9)
d-t2
+ g(t7) = 0,
238 VIII. STABILITY
where 9(17) is a real-valued and continuously differentiable function of g on the real

line R. Equation (VIII.1.9) is equivalent to the system
(VIII.1.10) dti = Y2, = -9(yi)

d
If g(a) = 0, then system (VIII.1.10) has a constant solution yt = a, Y2 = 0. Set
yi = zi + a and y2 = z2. Then, system (VIII.1.10) becomes
(VIII.1.11)
di = Az + #(z-),
iii
where
z = lz2J , A = j-9'(a) 0] [-[g(21 + a) -9 ()z1)]

The eigenvalues of A are ±(-g'(a))I/2. Note also that
z,-0 g(zl + a)zi- g'(a)zt = 0.

lim
Set
17
(VIII.1.12) G(i7) = / g(s)ds.

0
Then, G(n) takes a local minimum (respectively a local maximum) at n = a if

g(a) = 0 and g'(a) > 0 (respectively g'(a) < 0). Also, set
(VIII.1.13) H(yi , y2) = G(yi) + y2

2
Then,
dH(yi(t),y2(t)) dyl(t) (t) dy2(t)
dt
= 9 ( Y1 (t))
dt
+ Y2
dt
Therefore,
dH(yi(t),y2(t))
= 9(YI(t))y2(t) - y2(t)9(yI(t)) = 0
dt
if (yi(t),y2(t)) is a solution of (VIII.1.10). This means that

(t) + a) + z2(t)2 = (0) + a) + z2 0)2
(VIII.1.14) G(zi G(zi for all t
2 2
if z(t) is a solution of (VIII.1.11).

Case when g'(a) < 0: If g'(a) < 0, there exists a positive number PO such that
G(rl) < G(a) for 0 < 177 - al < Po Let t;(t) be the unique solution of (VIII.1.11)
satisfying the initial condition z1(0) = 0, z2(0) = Co > 0. Since t;2(t) = (1(t),
1. BASIC DEFINITIONS 239
we obtain ('(t) = 2(G(a) - G((1(t) + a)) + co >- Co as long as 0 < (1(t) < po.
Hence, there must be a positive number to depending on (o such that (I (to) = Pa
no matter how small to may be. This implies that the trivial solution of (VIII.1.11)
is not stable.
Case when g'(a) > 0: For a given positive number p, let V (p) be the connected
component containing 0 of the set {[; : G(( +a) < G(a)+p}. Then, V (pl) C V (P2)
if p1 < p2. Furthermore, if g(a) = 0, g(a) > 0, and if a positive number p is
sufficiently small, there exist two positive numbers e1(p) and e2(p) such that
(1) e2(P) < e1(P),
(2) Plmoe1(P) = 0,
(3) {( : ISO < C2 W} C 7) (P) C < el(p)} (cf. Figure 2).
S=o
G= G(S+a)
G=G(a)+p
G=G(a)
1 P)
FIGURE 2.
Observe also that G(a) < G(( + a) < G(a) + p for [; E V (p) if p is a sufficiently
small positive number.
For a given positive number e, choose another positive number p so that ei(p) < e
and p:5 e. Choose also the initial value z(0) so that Jz1(0)j < e2 (2) and z2(0)2 < p.
Then, G(z1(0)+a) < G(a)+ since z1(0) EP (2). Hence, (VIII.1.14) implies that
G(zi(t) + a) < G(a) + p for all
2 t. Observe that the set {z1(t) : all t} is connected
and zi(0) E D (2) C V (p). Therefore, zi(t) E D(p) for all t. Hence, Iz1(t) <
ei (p) < e. On the other hand, G(a) < G(z1(t) + a) < G(a) + p since z1(t) E D (p)
for all t. Hence, (VIII.1.14) implies that2(tz2< G(z1(0) +a) - G(a) + < p < e.
This proves that the trivial solution of system (VIII.1.11) is stable as t -+
2 +oo.
The analysis given in Example VIII-1-6 is an example of an application of the
Liapounoff functions to which we will return in Chapters IX and X.
The third example is the following result.
Theorem VIII-1-7. If f (x, y) and g(x, y) are real-valued continuously differen-
tiable functions such that
(i) f (0, 0) = 0 and g(0, 0) = 0,
(ii) (f (x, y), g(x, y)) 0 (0, 0) if (x, Y) # (0, 0),
(iii) 8x (x, y) + 49Y (x, y) = 0,
then, the trivial solution (x, y) = (0, 0) of the system
(S) dt = f(x,y), dt = g(x,y)

240 VIII. STABILITY
is not asymptotically stable as t -- +oo.

Proof.
A contradiction will be derived from the assumption that the trivial solution
x(t, ct, c2)
is asymptotically stable. Let (,t CI,t,C2)2) y(t,cl,c2)
be the unique solution of
system (S) satisfying the initial condition 0(0.cl, c2) =

[ci]. If the trivial solution
is asymptotically stable as t -. +oo, the trivial solution is also stable as t -. +oo.
Therefore, for every positive number c, there exists another positive number 8(e)
such that it (t,cl,c2)1 < e for 0 < t < +oo whenever max{jcj1,lc21} < b(e). Since
f and g are independent of t, O(t - r, cl, c2) is also a solution of (S) and satisfies
the initial condition (x(r), y(r)) = (Cl, c2). This implies that I¢(t, cl, c2)1 < e for
r < t < +oo if I¢(r,c1,c2)I < 8(e). Denote by 0(r) the disk {(c1,c2) E 1R2 :
Ic112 + 1c212 < r}. Also, for a fixed value r of t, let us denote by D(r, r) the
set {0r,C1,c2) : (cl,c2) E A(r)}. Then, the mapping (C1, C2) --+ (r,Cl,C2) is a
homeomorphism of 0(r) onto D(r, r) (cf. Exercise 11-4). Fix a sufficiently small
positive number ro. Since (0, 0) is asymptotically stable as t +00 and the disk
A(ro) is compact, there exists a positive number ro such that D(ro,ro) C 0 ro l
2 ).
This implies that the area of V(ro, ro) is definitely smaller than the area of A(ro).
Observe that
(VIII.1.15)
area of D(ro, ro) = det
tO(roc1c2) a$(ro, C1, C2) dc 1 de,.
fA(r0) 8c1 8c2
aj(ro,c1,c2)
It is known that the matrix 1L(t,cl,c2) = is the
L ac, 0Ce
unique solution of the initial-value problem
1of of
dX ax 8y X. X (0) = I2.
dt a9 09
ax Dy tr,y)=Oit.c,,cz)
where 12 is the 2 x 2 identity matrix (cf. Theorem 11-2-1). Therefore,

r
8f (x, y) + ag (x, y)1 I dtJ = 1
/'zJ
det '(t, cl, c2) = expLI (

o l ax Oly ft=,v>=ecl.c,,cz1
(cf. (4) of Remark IV-2-7). Now, it follows from (VIII.1.15) that the area of
D(ro, ro) is equal to the area of 0(ro). This is a contradiction. 0
2. A SUFFICIENT CONDITION FOR ASYMPTOTIC STABILITY 241
VIII-2. A sufficient condition for asymptotic stability

In this section, we prove a basic sufficient condition for asymptotic stability. Let
us consider a system of differential equations
(VIII.2.1) = Ag + g(t,yj
under the assumptions that
(i) A is a constant n x n matrix,
(ii) the entries of the 1R"-valued function §(t, y-) are continuous and satisfy the
estimate
(VIII.2.2) I9(t,y)l < k(t)lyy + colylt+w for (t, y)) E A(ro),
where, vo > O, co > O, ro > O, k(t) > O and bounded fort > 0, lim k(t) = 0,
t+x
and A(ro) = {(t, yj : 0 < t < +oo, Iy1 < ro).
The following theorem is the main result in this section.
Theorem VIII-2-1. If the real part of every eigenvalue of A is negative, the tnvial
solution g = 0 of (VIII. 2.1) is asymptotically stable as t -i +oo.
Proof.
Step 1. Let A, (j = 1, 2,... , n) be the eigenvalues of A. Then, RIA,I (j = 1, 2, ... ,
n) are Liapounoff's type numbers of the system d= Ay (cf. Example VII-2-8).
t
Therefore, choosing a positive number p so that 0 < p < -9RIA,I f o r j = 1, 2, ... , n,
we obtain
IeXpIAtll < Ke-"t on the interval To = {t : 0 < t < +oc}
for some positive constant K.
Step 2. Fixing a non-negative number T, change system (VIII.2.1) to the integral
equation
e(t-T)Ag(T)
g(t) = + fTte(t-')Ag(s,g(s))ds.
If Ig(t)I < b for some positive number b on an interval T < t < T1, then
<- Ke-"(t-T)Ig(T)I + Kff e-"(t-d) {k(s) + cabv0} Ig(s)Ids

ly(t)l
on the same interval T < t < Ti. Setting p(T) = sup k(s), change this inequality
a>T
to the form
e"(t-T)Ig(t)I e"('-T)Iy(s)Ids
<- Klg(T)I + K {p(T) + coa"o}
T
for T:5 t < Tt. Then,
e"(t-T)lg(t)I <- Klg(T)I exp[K(p(T) + cob&°)(t -T)I
and, hence,
(VIII.2.3) lg(t)I <- KI g(T)I exp I- (u - K(p(T) + cotVO))(t - T)J
for T:5 t < Tl (cf. Lemma 1-1-5).
242 VIII. STABILITY
Step 3. Fix a non-negative number T and two positive numbers 6 and 61 in such
a way that
p-K(p(T)+co6'o) > 0, 0<61 <6, K61 <6.
Assume that 19(T)J < 61. Then, inequality (VIII.2.3) holds for T < t < T1 as long
as Iy(t)I < bon the interval T < t < T1. This, in turn, implies that
Jg(t)J < K61 < 6 for T < t < T1.
This is true for all T1 not less than T. Hence, inequality (VIII.2.3) holds for t > T.
Step 4. If Jy(t)J < b < 1 for 0 < t < T, there exists a positive number is such
Id t)
that I < n19(t)j, for 0 < t < T. This implies that Jy-(t)J < Jy-(0)JeK' as long
as Jy(t)I < 6 T if Iy(0)J is small. This completes the
proof of Theorem VIII-2-1. O
Example VIII-2-2. For the following system of differential equations d = A#+
where
y= [y
Y
j A= r -0.4
-
-2
.2
1 (Y2 (
+ y2) I 11
1
the trivial solution y = 0 is asymptotically stable as t -' +oc. In f act, the char-
acteristic polynomial of the matrix A is PA(a) _ (A + 0.3)2 + 1.99. Therefore, the
real part of two eigenvalues are negative.
Remark VIII-2-3. The same conclusion as Theorem VIII-2-1 can be proven, even
if (VIII.2.2) is replaced by
(VIII.2.4) 19(t, y-)J < (k(t) + h(t)Jyj")Jy7 for (t, y) E A(ro),

where p is a positive number, and two functions h(t) and k(t) are continuous for
t > 0 such that k(t) > 0 for t > 0, lim k(t) = 0, h(t) > 0 fort > 0, and
Liapounoff's type number of h(t) at t = +oo is not positive (cf. [CL, Theorem 1.3
on pp. 318-319]).
Remark VIII-2-4. The converse of Theorem VIII-2-1 is not true, as clearly shown
in Example VIII-1-5. In that example, the eigenvalues of the matrix A are -1 and 0,
but the trivial solution is asymptotically stable as t -+ +oo. Also, in that example,
solutions starting in a neighborhood of 0 do not tend to 0 exponentially as t -+ +oo.
Remark VIII-2-5. Even if the matrix A is diagonalizable and its eigenvalues are
all purely imaginary, the trivial solution is not necessarily stable as t -+ +oo. In
such a case, we must frequently go through tedious analysis to decide if the trivial
solution is stable as t -+ +oo (cf. the case when g'(a) > 0 in Example VIII-1-6).
We shall return to such cases in IR2 later in §§VIII-6 and VIII-10 .
3. STABLE MANIFOLDS 243
Remark VIII-2-6. If the real part of an eigenvalue of A is positive, then the

trivial solution is not stable as it is claimed in the following theorem.
Theorem VIII-2-7. Assume that
(i) A is a constant n x n matrix,
(ii) the entries of y(t, yam) are continuous and satisfy the estimate
I9(t,y)l 5 e(t,yyly7 for (t,y-) E o(ro) _ {(t,y-) : 0 5 t < +oo, Iy-I <_ ro},
where ro > 0, e(t, y-) > 0 for (t, y) E i(ro), lim e(t, y) = 0,
and
t-1+19{o
(iii) the real part of an eigenvalue of the matrix A is positive.
dy
Then, the trivial solution of the system = Ay + §(t, y-) is not stable as t -. +oo.
dt
A proof of this theorem is given in (CL, Theorem 1.2, pp. 317-3181. We shall prove
this theorem for a particular case in the next section (cf. (vi) of Remark VIII-3-2).
An example of instability covered by this theorem is the case when g'(a) < 0 of
Example VIII-1-6. The converse of Theorem VIII-2-7 is not true. In fact, Figure 3
shows that the trivial solution of the system
dY2
I
= -yI, = (yi + yi)112
is not stable as t - +oo. Note that, in this case, eigenvalues of the matrix A are
-1 and 0.
Y2
Yt
FIGURE 3.
VIII-3. Stable manifolds

A stable manifold of the trivial solution is a set of points such that solutions
starting from them approach the trivial solution as t --. +oo. In order to illustrate
such a manifold, consider a system of the form
dx
(VIII-3-1) = AIi + LY = Alb + 92(40,
where Y E iR", 11 E R n, entries of R"-valued function gl and RI-valued function g2
are continuous in (1, y-) for max(IiI, Iyj) 5 po and satisfy the Lipschitz condition
19't(_,y)
- 9,(f,nil 5 L(P)max(Ii-.1, Iv-ffl) (1=1,2)
244 VIII. STABILITY
for max(jil, Iy1) < po and max(11 i, Ii1) < po, where po is a positive number and
u oL(p) = 0. Furthermore, assume that gf(0, 0) = 0 (j = 1, 2). Two matrices AI
and A2 are respectively constant n x n and m x m matrices satisfying the following
condition:
le(t-9)A' I < Kie-o1(t-a)
for t > s,
(VIII.3.2)
t Je(t-s)A21 < K2e-02(t-8) for t < s,
where K, and of (j = 1,2) are positive constants. Condition (VIII.3.2) implies
that the real parts of eigenvalues of AI are not greater than -01, whereas the real
parts of eigenvalues of A2 are not less than -02. Assume that
(VIII.3.3) a1 > 02.
Let us change (VIII.3.1) to the following system of integral equations:
I
x(t, c) = e1AI C + J e(1-a)A' gi (Y(s, c-), y(s, c))ds,
0
(VIII.3.4)
At, c ) = Jt to0 e(t 8)A2 2(x(s, c), y(s, c))ds.
The main result in this section is the following theorem.

Theorem VIII-3-1. Fix a positive number c so that a1 > 02 + E. Then, there
exists another positive number p(() such that if an arbitrary constant vector c' in
R' satisfies the condition KI I c1 < 2 , a solution (i(t, c), y"(t, cam)) of (VIII.3.1) can
be constructed so that
(VIII.3.5) 1(0,61=6 and max(Ii(t, c)I, I y(t, c)I) < p(e)e-(Ol -t)t
for 0 < t < +oo. Furthermore, this solution (i(t, cl, #(t, c)) is uniquely determined
by condition (VIII.3.5).
Proof
Observe that if
max(Ix(t,c)I,Iy(t,c)I) < pe-(°'-`)t
for 0 < t < +00,
then
t e(t-s)A'91(Y(s, c), y(s, c))ds1 < K1 L(P)P f t e-`(t-8)e-(" -`)sds

I fo 0
KiL(P)Pe-0,t (e" - I) < Ki L(P)Pe-(o, -c)t
and
t +00
e-02(t-8)e-(0 $ -')8ds
IJ e(t-a)A2 2(j(s, c), y(s, c))dsl < K2L(p)p f
+ 00 t
K2L(P)P e
al -o2-E
3. STABLE MANIFOLDS 245
K2L(p) < 1
This implies that if a positive number p is chosen so small that
01-o2-e - 2
and K1L(p) < and if the arbitrary vector c' in R" satisfies the condition
Kl I i < 2
, then, using successive approximations, a solution (i(t, cl , y(t, c)) of
2
(VIII.3.4) can be constructed so that
x(O, c-)= c and max(Ii(t, c1I, Iy(t, c)I) < pe-(c for 0 < t < +oo.
Details are left to the reader as an exercise.
Remark VIII-3-2.
(i) The positive number a is given to start with and the choice of p depends on
e. However, since this solution approaches the trivial solution, the constant e
may be eventually replaced by any smaller number, since the right-hand side
of (VIII.3.1) is independent of t. This implies that the curve (i(t, c), y'(t, c-))
is independent of a as t -. +oo. More precisely, if a solution (x"(t), y"(t)) of
(VIII.3.1) satisfies a condition
max(Ii(t)I, ly(t)I) < Ke-(°'-'0)t

for 0 < t < +00
with some positive constants K and eo such that of -o2-eo > 0, then for every
positive a smaller than co, there exists to > 0 such that (i(to + t), y(to + t)) _
(i(t, ), y(t, cam)), where c = ?(to).
(ii) The initial value of y(t, c-) is given by
0
00, cl = f e-sA292(i(s, c_), y(s, c_))ds.
+oo
(iii) If and exist and are continuous in a neighborhood of (6,6) and if

ay 09
8 (6, U = 0' and (0, 0) = 0', then i(t, c) and y"(t, cl are continuously
89
differentiable with respect to c.
(iv) If the real parts of all eigenvalues of the matrix Al of (VIII.3.1) are negative
and the real parts of all eigenvalues of the matrix A2 are positive, then the
stable manifold of the trivial solution of system (VIII.3.1) is given by S =
((6, 9(0, c-)) : 161 < p}, where p is a sufficiently small positive number.
(v) Consider a system
dy-
(VIII.3.6) Ay + 9(y
dt
in the following situation:
(a) A is a constant n x n matrix,
(b) the entries of §(y1 are continuous for Iy1 < po and satisfy the Lipschitz
condition 19(y-) - L(p)ly - r11 for ly-I < po and i31 < po,
where PO is a positive number and lim L(p) = 0,
p-0
246 VIII. STABILITY
(c) 9(0,6) = 6.
Suppose further that A has an eigenvalue with positive real part. Then, applying
Theorem VIII-3-1 to the system = -Ay' - g(y-), we can construct the stable
manifold U of the trivial solution. This means that if a solution fi(t) of (VIII.3.6)
starts from a point on U, then urn Q(t) = 6. This shows that the trivial solution of
t - +oo, and Theorem VIII-2-7 is proved for (VIII.3.6).
The set U is called the unstable manifold of the trivial solution of (VIII.3.6).
The materials in this section are also found in [CL, §§4 and 5 of Chapter 81 and
[Hart, Chapter IX; in particular Theorem 6.1 on p. 2421.
VIII-4. Analytic structure of stable manifolds

In order to look closely into the structure of the stable manifold of the trivial
solution of a system of analytic differential equations
dy
(VIII.4.1)
dt
= Ay + f (y),
let us construct a formal simplification of system (VIII.4.1). To do this, consider
system (VIII.4.1) under the following assumption.
Assumption VIII-4-1. The unknown quantity g is a vector in C" with entries
{y1, ... , y" }, A is a constant n x n matrix, and
(VIII.4.2) f (Y-) = E #vfa

jpI>2
is a formal power series with coefficients fp E C', where p = (pl,... , p") with
n
non-negative integers pl, ... , p", > Ph, and yam' = yl' yP,,-.
h=1
The following theorem is a basic result concerning formal simplifications of sys-
tem (VIII.4.1).
Theorem VIII-4-2. Under Assumption VIII-4-1, there exists a formal power se-
ries
(VIII.4.3) P'(u) = Pou" + EPP,,

IpI>2
in a vector u E C" with entries {ul,... ,u"} such that

(i) Po is an invertible constant n x n matrix and PP, E C",
(ii) the formal transformation y' = P(u") reduces system (VIII 4. 1) to
(VIII.4.4) j = Boii + g"(u)

4. ANALYTIC STRUCTURE OF STABLE MANIFOLDS 247
with a constant n x n matrix Bo and the formal power series g(ui) _ E upgp
Ipi>2
with coeficients gp in Cn such that
(iia) the matrix B0 is lower triangular with the diagonal entries Al.... , An,
and the entry bo(j, k) on the j-th row and k-th column of B0 is zero
whenever A, j4 Ak,
(iib) for p with IpI > 2, the j-th entry gp, of the vector gp is zero whenever
A1 n
(VIII.4.5) 1: PhAh
h=1
Proof
Observe that if y" = P(d), then
d9 = p p Pp
dt
Po + 1: [ L Pp P22 Pp ... un Pp Boi + F 4Lpgp
W>2 Ipl>2
and
Ay + f(y) = A Pod + F u-'pPp + E P(u)p fp.

IpI>2 IPI>_2
Furthermore,
poop P_p
Po + [Plup,5, P2upP-p ... Bo u'
ul u2 Un J
Ipl>2
n
= PoBou" + phAh u"PPp + uh
plp),A-
ul
lpl>2 h_1 1<h<35n
where 03,k is the entry of B0 on the 7-th row and k-th column.
Let us introduce a linear order pi -< p2 for p. = (p,',... , pin) (j = 1, 2) by the
relation
Plh = p2h for (h < he) and P1ho < P2ho
Now, calculating the coefficients of iip 1) on both sides of (VIII.4.1), we
obtain
(VIII.4.6) PoBo = APo

and
(PhAh)P + Po9p - APp

(VIII.4.7) h=1
=j (I5,' p -<p) + yp(Pp',ff : 10 <ItI)

248 VIII. STABILITY
for Jpj > 2. From (VIII.4.6), it is concluded that the diagonal entries A1, ... , A,, of
Bo are eigenvalues of A and that this allows us to set A = B0 and P0 = In, where
I,, is the n x n identity matrix. Then, (VIII.4.7) becomes
(VIII.4.8)
(PhAh)PP + 9p - Bo,6,
h=1
= ff(PP, p) + 9$'05P" 9P, : 10 < 1p1)
Solve (VIII.4.8) by solving equations of the form
(VIII.4.9) (PhAh - A7 PP,1 + 9P.J = F'p,J

h- J
successively, where PP,, and gp,, are the j-th entries of the vector P. and
respectively, and FP,, are known quantities. If >phAh - Aj 36 0, set 9,j = 0 and
h=1
solve (VIII.4.9). If >phAh - A? = 0, then set 9P,, = Fr,, and choose PPJ in any
h=1
way. This completes the proof of Theorem VIII.4.2.
Observation VIII-4-3. Assume that R(AE) < 0 for j = 1, ... , r and R(A3) > 0
for j 1,... , r. In this case, if uh = 0 for h I, ... , r, the a-th entry of the
vector Bou + g"(u) is zero for t ¢ I, ... , r. In fact, look at g"(u"). Then, the f-
n
th entry of the coefficient g"p is zero if Al 96 >phAh. Note that, if t > r and
h=1
n
At = > Ph Ah, then (Pr+ 1, . Pn) 0 (0, ... , 0). Hence, in such a case, uP = 0 if
h=1
(ur+1, ... , u,) = (0, ... , 0). Therefore, the 1-th entry of the vector Boti + 9"(uis
zero forI>rif(ur+1,...,u,,)=(0,....0).
Observation VIII-4-4. Under the same assumption on the Aj as in Observation
VIII-4-3, set (ur+1, ... , un) = (0,... , 0). Then, the system of differential equations
on (u1,... has the form
dud A3u3
= + QJ,huh
dt
(VIII.4.10)
... u p , ( i=1 , -...- - , r)-
+ 9(P,.
.P04ul1
A,=p,a,+ +p.a.,lai>2
r r
Observe that since A. - ph Ah # 0 if ph is sufficiently large, the right-hand
h=1 h=l
members of (VIII.4.10) are polynomials in (ul,... ,ur).
4. ANALYTIC STRUCTURE OF STABLE MANIFOLDS 249
Observation VIII-4-5. Assume that R[Ah+1J <_ R[Ah]. Then, (V111.4.10) can be
written in the form
du
dul
= Aiu1 and = A)u) + u_,-i) for j=2,... ,r.
dt
Hence, system (VIII.4.10) can be solved with an elementary method. To see the
structure of solutions of (VIII.4.10) more clearly, change (ul, ... , ur) by u2 _
ea'ivi (j = 1,... , r). This transformation changes (VIII.4.10) to
dv)
Q),hvh
dt
an
g(P,..
...vf (.i
A, =P, a,+...+prAr,Ipl>2
This, in turn, shows that the general solution of (VIII.4.10) has the form u) _
ea'ii) (t, cl, ... , cr) (j = 1.... , r), where (cl,... , c,.) are arbitrary constants and
0,(t,c1,....c,-) are polynomials in (t.cl,... ,G)
An analytic justification of the formal series 15(ii) is given by the following the-
orem.
Theorem VIII-4-6. In the case when the entries of the Z"-valued function f'(y-)
on the right-hand side of (VI11.4.1) are analytic in a neighborhood of 0, under the
same assumption on the A) as in Observation VIII-4-3, the power series P(d) is
convergent if (ur+i, ... , u") = (0,... , 0).
The proof of this theorem is straight forward but lengthy (cf. [Si2J). The key
fact is the inequality
A) - >PhAh
h=1 h=1
r
for some positive number a if E ph is large. In this case, a iajorant series for P
h=1
can be constructed.
The construction of such a majorant series is illustrated for a simple case of a
system
dy
dt
= Ay + f (y),
where A is a nonzero complex number and AM is a convergent power series in y

given by (VIII.4.2). According to Theorem VIII-4-2, in this case there exists a
dd
formal transfomation ff = d + Q(u) such that = Ad, where
dt
IpI?2
This implies that E AIpJi1PQg, = AQ(u) + f (u" + Set f (ii + Q(u)) _
IeI>2
250 VIII. STABILITY
ui"A,. Then,
lal>2
Ap
a= AOPI-1) for all p (IpJ > 2).
Set
1
' (y1= > IfP11F

IpI>2 1
Then, F(y-) is a majorant series for ft y-). Determine a power series V (u-) = uupii
IpI>2
by the equation v = ICI 0(u" + v'). Set F(u + V(U)) _ iBa. Then,
Ipl>2
B
vp = Ii for all p (JpJ > 2).
It can be shown easily that Y(ul) is a convergent majorant series of Q(ur). This
proves the convergence of Q(ii).
Putting P(u") and the general solution of (VIII.4.10) together, we obtain a par-
ticular solution y" P(#(t, cl) of (VIII.4.1), where
=
(t,cj = (e'\It7Pi(t,cl,... ,cr),... ,e*\1 'Wr(t,cl,... ,cr),0,... ,0).
This particular solution is depending on r arbitrary constants c = (cl,... , c,.).
Furthermore, this solution represents the stable manifold of the trivial solution of
(VIII.4.1)if W(Aj)>0for j#1,...,r.
Remark VIII-4.7. In the case when y, A, and AM are real, but A has some
eigenvalues which are not real, then P(yl must be constructed carefully so that the
a b
particular solution P(i(t, c)) is also real-valued. For example, if A = ,
b a
the eigenvalues of A are a ± ib. If IV = ['] is changed by ul = 1/i + iy2i and
= - iy, system (VIII.4.1) becomes
dul
= (a + ib)ul + 9P,,p2u 'uZ
OFF
Pl+P2>2
(VIII.4.11)
due
_ (a - ib)u2 + ,
dt
where gP,,- is the complex conjugate of g,,,p,. If a 54 0, using Theorem VIII-4-2,

simplify (VIII.4.11) to
(VIII.4.12) dtl = (a + ib)vl, = (a - ib)v2

ddt
S. TWO-DIMENSIONAL LINEAR SYSTEMS 251
PP,.P2'UP11t?221
VI +
p, +p2>_2
(VIII.4.13)
7P,.p2V2P1VP13.
V2 +
P,+p2>2
V 2 V2
Now, system (VIII.4.12), in turn, is changed back to - = At by wI =
VI - V2
and w2 = ,
where (w1, w2) are the entries of the vector tu. Observe that
2i
U l + u2
2
w1 + E
p,+p2>2
uI - u2
U'2 + g2.P,,pzu'P1 u"
2i
p, +p3 22
where ql,p,,p2 and g2,p,,p2 are real numbers. Similar arguments can be used in
general cases to construct real-valued solutions. (For complexification, see, for
example, [HirS, pp. 64-65].)
For classical works related with the materials in this section as well as more
general problems, see, for example, [Du).
VIII-5. Two-dimensional linear systems with constant coefficients

Throughout the rest of this chapter, we shall study the behavior of solutions of
nonlinear systems in 1R2. The R2-plane is called the phase plane and a solution
curve projected to the phase plane is called an orbit of the system of equations. A
diagram that shows the orbits in the phase plane is called a phase portrait of the
orbits of the system of equations. As a preparation, in this section, we summarize
the basic facts concerning linear systems with constant coefficients in R2.
Consider a linear system
dy =
(VIII.5.1) Ay,
dt
where E R2 and A is a real, constant, and invertible 2 x 2 matrix. Set p = trace(A)
and q = det(A), where q ¢ 0. Then, the characteristic equation of the matrix A is
1\2 - pA + q = 0. Hence, two eigenvalues of A are given by
AI =
2
+
4 q and A2 = 2 - 4
- q.
It is known that
P2
p = AI +A2, q = AIA2, and Al - A2 =2 - q.
4
252 VIII. STABILITY
Also, let t and ij be two eigenvectors of A associated with the eigenvalues Al and
A2, respectively, i.e., At = All;, l 0 0, and Au = A2ij, ij 0 0. Observe that, if
y(t) is a solution of (VIII.5.1), then cy(t +r) is also a solution of (VIII.5.1) for any
constants c and T. This fact is useful in order to find orbits of equation (VIII.5.1)
in the phase plane.
Case 1. Assume that two eigenvalues Al and A2 are real and distinct. In this
case, two eigenvectors { and ij are linearly independent and the general solution of
differential equation (VIII.5.1) is given by
c2e(1\2-,,,)t11= ea2t[cie(AI -a2)t (+ c2i17,

1!(t) = cl eAI t + c2ea'tti = e)lt (cl +
where cl and c2 are arbitrary constants and -oo < t < +oo.
2
la: In the case when Al > A2 > 0 (i.e., p > 0, q > 0 and 4 > q), the phase portrait
of orbits of (VIII.5.1) is shown by Figure 4. The arrow indicates the direction in
which t increases. The trivial solution 0 is unstable as t +oo. Note that as
t -oo , the solutions y(t) tends to 0 in one of the four directions of and
-ij. The point (0, 0) is called an unstable improper node.
2
lb: In the case when 0 > Al > A2 (i.e., p < 0, q > 0 and 4 > q), the phase
portrait of orbits of (VIII.5.1) is shown by Figure 5. The trivial solution 0 is stable
as t --i +oo. The point (0, 0) is called a stable improper node.
4
FIGURE 4. FIGURE 5.
1c: In the case when Al > 0 > A2 (i.e., q < 0), the phase portrait of orbits of
(VIII.5.1) is shown by Figure 6. The trivial solution 0 is unstable as Itl +oo.
Note that as t -+ -oo (or +oo), only two orbits of (VIII.5.1) tend to 0. The point
(0, 0) is called a saddle point.
2
Case 2. Assume that two eigenvalues Al and A2 are equal. Then, q = 4 and
Al=A2=29& 0.
2a: Assume furhter that A is diagonalizable; i.e., A = 212i where 12 is the 2 x 2

identity matrix. Then, every nonzero vector c is an eigenvector of A, and the general
solution of (VIII.5.1) is given by y"(t) = exp [t]e. In this case, the phase portrait
of orbits of (VIII.5.1) is shown by Figures 7-1 and 7-2. As t -+ +oo, the trivial
solution 6 is unstable (respectively stable) if p > 0 (respectively p < 0). Note that,
5. TWO-DIMENSIONAL LINEAR SYSTEMS 253
for every direction n, there exists an orbit which tends to 6 in the direction n" as t
tends to -oo (respectively +oo). The point (0, 0) is called an unstable (respectively
stable) proper node if p > 0 (respectively p < 0).
p>0 p<0
FIGURE 6. FIGURE 7-1. FIGURE 7-2.
2b: Assume that A is not diagonalizable; i.e., A = p 212 + N, where 12 is the 2 x 2
identity matrix and Nris a 2 x 2 nilpotent matrix. Note that N 0 and N2 = O.
Hence, exp[tA] = exp 12t] {I2 + tN}. Observe also that a nonzero vector c' is an
eigenvector of A if and only if NcE = 0. Since N(Nc) = 0, the vector NcE is
either 6 or an eigenvector of A. Hence, NE = ct(-) , where is the eigenvector
of A which was given at the beginning of this section and a(c) is a real-valued
linear homogeneous function of F. Observe also that at(c) = 0 if and only if c' is a
constant multiple of the eigenvector . The general solution of (VIII.5.1) is given
by y(t) = exp [t] {c + ta(c7}, where c" is an arbitrary constant vector. In this
case, the phase portrait of orbits of (VIII.5.1) is shown by Figures 8-1 and 8-2. The
trivial solution 6 is unstable (respectively stable) as t - +oo if p > 0 (respectively
p < 0). The point (0, 0) is called an unstable (respectively stable) improper node if
p > 0 (respectively p < 0).
p>0 p<0
FIGURE 8-1. FIGURE 8-2.
Case 3. If two eigenvalues \I and .12 are not real, then q > and
4
Al =a+ib, A2=a-ib, a= 2,
Note that b > 0. Set A = a12 + B. Then, B2 = -b212, since two eigenvalues of B
254 VIII. STABILITY
Smbbt)
are ib and -ib. Therefore, exp[tB] = (cos(bt))I2 + B.
3a: Assume that a = 0 (i.e., p = 0). Then, the general solution of (VIII.5.1) is
slnbbt)
given by '(t) = exp[tB]c = (cos(bt))c"+ Bc, which is periodic of period 2b
in t. The phase portrait of orbits of (VIII.5.1) is shown by Figures 9-1 and 9-2.
The trivial solution 0 is stable as It 4 +oo. The point (0, 0) is called a center. It
is important to notice that every orbit y(t) is invariant by the operator . In fact,
b
B j(t)
T
= cos(bt)
b
Bc-(sin(bt))c = (cos (bt + ))e+
sin bt +
b
r,
2 Be= 17 (t +
j
2b
)
In other words, dy(t) = by (t + Note that "" is 1 of the period

2r
.
dt 2b) 2b 4 b

3b: Assume that a 0 0 (i.e., p 0 0). Then, the general solution of (VIII.5.1) is
given by y(t) = exp[at]il(t), where u(t) is the general solution of dt = Bu. The
solution 0 is unstable (respectively stable) as t - +oo if p > 0 (respectively p < 0).
The orbits, as shown by Figures 10-1 and 10-2, go around the point (0, 0) infinitely
many times as y(t) -- 0. The point (0, 0) is called an unstable (respectively stable)
spiral point if p > 0 (respectively p < 0).
(0.0)
p>0 p<o
Let us summarize the results given above by using Figure 11:
(1) (0, 0) is an improper node.
(2) (0, 0) is a saddle point.
6. ANALYTIC SYSTEMS IN R2 255
(3) (0, 0) is a proper or improper node.

(4) (0, 0) is a center.
(5) (0, 0) is a spiral point.
Stability of the trivial solution 6 is summarized as follows:
(1) If q < 0, the trivial solution 0' is unstable as Itl -+ +oo.
(II) If q > 0 and p > 0, the trivial solution 0 is unstable as t +oo.
(III) If q > 0 and p < 0, the trivial solution 0 is stable as t -+ +00.
(IV) If p = 0 and q > 0, the trivial solution 0 is stable as It( - +oo.
(Cf. Figure 12).
q(P0) q(P=0)
P2
(5) (5) 4
(4) (IV)
(3) (3) (111) (Ill
-------- -------
VIII-6. Analytic systems in R2

In this section, we apply Theorems VIII-4-2 and VIII-4-6 to an analytic system
in R2. Consider a system
(VIII.6.1)
dt Ay" + E ypfp,
tp!>2
where y' E LR2 with the entries yj and y2, A is a real, invertible, and constant
2 x 2 matrix. p = (pi, p2) with two non-negative integers pi and p2i Ipl = pi + p2,
yam' = y'' y2 , the entries of vectors fp E R2 are real constants independent of t, and
the series on the right-hand side of (VIII.6.1) is uniformly convergent in a domain
A(po) = (y E R2 : fyj < po) for some positive number po. Let us look into the
structure of solutions of (VIII.6.1) in the following five cases.
Case 1. If the point (0, 0) is a stable proper node of the linear system = Ay, then
A = )J2i where A is a negative number and 12 is the 2 x 2 identity matrix. To apply
Theorem VIII-4-2 to this case, look at the equation A = pi.1 + p2A on non-negative
integers pi and p2 such that pl + p2 > 2. Since no such (pi, p2) exists, there exists
an R2-valued function P(u) whose entries are convergent power series in a vector
it E R2 with real coefficients such that 5 () = 12 and that the transformation
= P( u reduces system (VIII.6.1) to dt` = Ail. This, in turn, implies that the
256 VIII. STABILITY
point (0, 0) is also a stable proper node of (VIII.6.1) and that the general solution
of (VIII.6.1) is y = P(eJ1°cl, where c is an arbitrary constant vector in R2.
Case 2. If the point (0, 0) is a stable improper node of the linear system dt = Ay,
then we may assume that either (1) A = AI2+N, where 1A is a negative number and
N 4 0 is a 2 x 2 nilpotent matrix, or (2) A = I 1 a2 J , where Al and A2 are real
negative numbers such that Al > A. In case (1), the same conclusion is obtained
concerning the existence of an 12-valued function P(i7). Therefore, the point (0,0)
is a stable improper node of (VIII.6.1), and the general solution of (VIII.6.I) is
y" = P(eAt(12 + tN)c), where c is an arbitrary constant vector in 1R2. In case (2),
looking at the equations Al = p1 A 1 + p2A2 and A2 = p1 At + p2A2 on non-negative
integers pi and p2 such that pl + p2 ? 2, it is concluded that there exists an 1R2-
valued function P(ui) whose entries are convergent power series in a vector ii E R2
dP
with real coefficients such that (d) = 12 and that the transformation y" = P(u)
['Ju.&Ajul
reduces system (VIII.6.1) to dt = 2u21, where (u1, u2) are the entries of
1
i , M is a positive integer such that A2 = MA1, and -y is a real constant which must
be 0 if A2 54 MA1 for any positive integer M. Therefore, in case (2), the general
A,:
([(''
solution of (VIII.1.6) is Cty"7t=+ C2)e
PAlt ) , where cl and c2 are arbitrary
constants. This, in turn, implies that the point (0, 0) is a stable improper node of
(VIII.6.1).
dy
Case 3. If the point (0, 0) is a stable spiral point of the linear system = Ay",
at
then we may assume that A = I b ab 1, where a and b are real numbers such that
a < 0 and b # 0. The eigenvalues of A are At = a ± ib. This implies that there are
no non-negative integers pi and p2 satisfying the condition A = p1 A+ +P2A_ and
pl + > 2. Therefore, there exists an R 2-valued function P(g) whose entries are
convergent power series in a vector u" E R2 with real coefficients such that LP (0) _
8u
12, and the transformation y = P(u) reduces system (VIII.6.1) to dt = Au (cf.
Remark VIII.4.7). This, in turn, implies that the point (0, 0) is also a stable spiral
point of (VIII.6.1) and that the general solution of (VIII.6.1) is y = P(eAC1, where
c" is an arbitrary constant vector in 1R2.
Case 4. If the point (0, 0) is a saddle point of the linear system = Ay, the
eigenvalues At and A2 of A are real and At < 0 < A2. Construct two nontrivial
and real-valued convergent power series fi(x) and rG(x) in a variable x so that
¢(eA'tcl) (t > 0) and ,G(eA2tc2) (t < 0) are solutions of (VIII.6.1), where cl and c2
are arbitrary constants (cf. Exercise V-7). The solution y' = *(eAt tct) represents the
stable manifold of the trivial solution of (VIII.6.1), while the solution y" =1 (eA2tc2)
represents the unstable manifold of the trivial solution of (VIII.6.1). The point (0, 0)
is a saddle point of (VIII.6.1). In the next section, we shall explain the behavior of
solutions in a neighborhood of a saddle point in a more general case.
dy
Case 5. If the point (0, 0) is a center of the linear system - = Ay, both eigenvalues
at
01
of A are purely imaginary. Assume that they are ±i and A = [0
J.
Set
y" = 2
1i
1 v, where v =
[t1]. Then, the given system (VIII.6.1) is changed
L J 2
to
du Ig(vl,v2)l
(VIII.6.2)
dt - 9(27)
012, v1)
where
+00
g(tvl, v2) = ivl + 9P'9111 2, 012,V0 = -iv2 + p.9v2v1
p+9=2 p+9=2
Here, a denotes the complex conjugate of a complex number a. Let us apply

Theorem VIII-4-2 to system (VIII.6.2).
Observation VIII-6-1. First, setting Al = i and A2 = -i, look at Al = p1A1 +
p2A2 and A2 = Q1A1 +g2A2, where p1, p2, q1, and q2 are non-negative integers such
that p1 + P2 > 2 and ql + q2 > 2. Then, p1 - 1 = p2 and q2 - 1 = q1. This implies
that system (VIII.6.2) can be changed to
(VIII . 6 . 3)
du1
dt = i W ( ulu2 ) u1,
d = -iC
due
0 ( u1U2 ) U2,
+oo
where w(z) = 1 + E Wmxm, by a formal transformation
m=1
ul + h(ul,u2)
(T) v = f(17) = u2 + h(u2,u1)
Here,
h(ul,u2) _
+00
hp.qupU2, h(u2,u1) _
+0 _
hP,9u'lU2
p+q=2 p+9=2
In particular, h(ul, u2) can be construced so that
(VIII.6.4) the quantities hp+1,p are real for all positive integers p.
Observation VIII-6-2. We can show that if one of the Wm is not real, then y = 0
is a spiral point (cf. Exercises VIII-14). Hence, let us look into the case when the
+00
Wm are all real. Furthermore, if a formal power series a(t) = 1 + with
m=1
real coefficients am is chosen in a suitable way, the transformation
(VIII.6.5) u1 = a(B11j2)131, u2 = W10002
258 VIII. STABILITY
changes (VIII.6.3) to another system

d,31
= in(AiQ2)Qi, 2 _ -0#10002,
where f2(() is a polynomial in C with real coefficients which has one of the following
two forms:
1, Case I
Q(o
- 1 + coS"'°, Case II
where co is a nonzero real number and tno is a positive integer.
Observation VIII-6-3. Let us look into Case I. Assume that system (VIII.6.3)
is
dul due
(VIII.6.3.1) = iu1, -iu2.
dt dt =
Note that transformation (VIII.6.5) does not change (VIII.6.3.1). Using a trans-
formation of form (VIII.6.5), change h(ul, u2) so that
(VIII.6.6) 1,P = 0 for all positive integers p.
Since (u1,u2) = (ce`t,de-'t) is a solution of system (VIII.6.3.1) with a complex

arbitrary constant c, the formal series
cue-,t)
cell + h(eett,
(FS-1) v = f(-) =
ce-u + h(ce-tt cett)
is a formal solution of system (VIII.6.2) which depends on two real arbitrary con-
stants.
Let
Vt + H(t,
(S) u(t, , ) _
fe-'t
be the solution of system (VIII.6.2) satisfying the initial condition
(C) t'(0, f, 6 = III].
Note that
+«0
H(t, , ) _ HP,q(t)ef°, H(tto = E
D+9=2 P+v=2
are power series in { and which are convergent uniformly on any fixed bounded
interval on the real t line.
Using (VIII.6.6), we obtain

Zx
r 2* h(Ce-'t, cett)e'Ldt
I0
0
h(ceic, ee ")e-'tdt = 0,
0
= 0.
Fix c and c by the equations
c = £ +T7r- / 2x H(s, t:, )e-"ds and e = { + 2 J2w H(s, {, {)e'sds.

0
0
Then,
= c + -(c, c) and t; = e + ?(c, c),
where =(c, e) and c(e, c) are convergent power series in (c, e).
Now, we can prove that two formal solutions
+h(ce",Ce-'t)
ce`t
v'(t,c+F(c,e),e+ =2(c' i!)) and
ee-tt + h(&-'t, cent ) j
of system (VIII.6.2) are identical. The first of these two is convergent; hence, the
second is also convergent. This finishes Case I.
Observation VIII-6-4. Let us look into Case II. Assume that system (VIII.6.3)
has the form
(VIII.6.3.2) dtl = iul(l + co(u1u2)'0),
dt2 = -iu2(1 +
co(uiu2)"'°)
Note that we have (VIII.6.4).

Since (ul,u2) = (ce't(1+co(ce)'"°) ee-u(1+c0(ct)'"0)) is a solution of system
(VIII.6.3.2) with a complex arbitrary constant c, the formal series
ce't(I+co(ce)'"0) + h(ceit(l+co(ct)"'O), ee-'t(1+c0(ct)_0))
(FS-2) v = f (u') = h(ce-tt(1+co(ce)'"° ), celt(1+co(ct)"O) )
ee-'t(1+c0(ce)'"0) +
is a formal solution of system (VIII.6.2) which depends on two real arbitrary con-
stants.
Again, as in Observation VIII-6-3, let (S) be the solution of system (VIII.6.2)
satisfying the initial condition (C). Set
(VIII.6.7) t; = ?(c, cJ = c + h(c, c) and = =(c, c) = e + h(e, c).

Then,
h(ceit(1+co(ce)'"°),ee-'t(l+co(ct)"'0)
I ce1t(1+co(ct)"°)
+
ee-tt(1+c0(ct)'"0) + h(ee-it(t+co(ct)'"O),ceut(l+co(et)"O))
260 VIII. STABILITY
as formal power series in (c, e). This series has a formal period
27r
T(cc) = with respect to t, i.e.,
1 + co(ce)m0
E(c, c)
(VIII.6.8) v(T(cc), =[c, e], E[c, c])
E(e, C)
Solving (VIII.6.7) with respect to (c, e), we obtain two power series in
c= and e = e( , £).
Set t) = T(c(., t)e(, £)). Then, (VIII.6.8) becomes
(VIII.6.9)
Using (VIII.6.9) as equations for P({,£), it can be shown that the formal power
series is convergent. This implies that
1/7+o
C(S,S)C(CS) [(CO) 1/ J
is convergent. (Here, some details are left to the reader as an exercise.)

On the other hand,
T(ct)
{{ese H(s,e-.(1+c0(c4),"O)ds
+
0
T(cz)
1Cesa(1+co(ct)'"°) h(ceps(1+Co(cr)m0),ce-1e(1+co(Ce)m°))J
+
L
+00
x e-fs(1+G)(cz)m°)ds = c 1 + Ehp+l,p(cz)P
P=1
and
`T(cC) +oo
J/ {e + H(s, , ) } e'8(1+co(cc)"° )ds
1 + t he+1,P(cz)p
0
P=1
since the quantity hP+1,P are real (cf. (VIII.6.4)). This proves that C(Cr/ ) is conver-
C(l, 7)
gent. Hence, c(l;, ) and 4) are convergent. Thus, the proof of the convergence
of A g) in Case II is completed.
Thus, it was proved that if all of the coefficients w,,, on the right-hand side of
system (VIII.6.3) are real, the point (0,0) is a center of system (VIII.6.1). The
general solution of (VIII.6.1) can be constructed by using various transformations
of (VII.6.1) which bring the system to either (VIII.6.3.1) or (VIII.6.3.2). Periods
of solutions in t are independent of each solution in Case I, but depend on each
solution in Case H.
7. PERTURBATIONS OF AN IMPROPER NODE AND A SADDLE PT. 261
Remark VIII-6-5. The argument given above does not apply to the case when
the right-hand side of (VIII.6.1) is of C(O°) but not analytic. A counterexample is
given by
2 _
r2 = Y1 + y2+
dt [h(1) h(r) y'
where h(r2) = e-1jr2 sin I I. Using the polar coordinates (r,O), this system is
r
changed to
dt = rh(r2) and
d
and periodic solutions are given by r2 = 1
MR
with any integers nt. This is an
example of centers in the sense of Bendixson (cf. [Ben2, p. 26] and §VIII-10).
For more information concerning analytic systems in JR2, see [Huk5].
VIII-7. Perturbations of an improper node and a saddle point

In the previous section, the structure of solutions of an analytic system in JR2
was showen by means of the method with power series. Hereafter, in this chapter,
we consider a system
dy _
(VIII.7.1) Ay + §(y-),
dt
under the assumption that
(i) A is a real constant 2 x 2 matrix,
(ii) the entries of the R2-valued function §(y) are continuous in g E JR2 near 0' and
satisfies the condition
(VIII.7.2) lim 9(y) = 0.

y-o Iy1
In this section, it is also assumed that the initial-value problem
(VIII.7.3) df = Ay + §(y-), g(0) = g

has one and only one solution as long as g belongs to a sufficiently small
neighborhood of 0.
Remark VIII-7-1. In the case, when the entries of an 1R2-valued function f (Z-)
are continuously differentiable with respect to the entries z1 and z2 of i in a neigh-
borhood of i = 0 and f (0) = 0, write the differential equation dt = f(E) in form
(VIII.7.1) by setting
aft eft
of (o) - (o) (o)
A 19x2
= 0z' and g(y)=f(y)-Ay.
i9 aft
Lof2 ( )
8x1 az2
262 VIII. STABILITY
Using thepolar coordinates (r, 0) in the g-plane, write the vector g in the form
y=r [c?}. Then, system (VIII.7.1) is written in terms of (r, B) as follows:
d=r[a, i cos2(B) + a22 sin2(9) + (a12 + a21) sin(g) cos(9)]
+ 91(-1 COs(0) + 92() sin(g),

(VIII.7.4)
r dte =r[-a12 sin2(0) + a21 COS2(0) + (a22 - all) sin(e) cos(0)]
- 91(y-) sin(g) + 92 M cos(e),
where A =
all a12 J and g"(y) = [iW)l Here, use was made of the formula
a21 22 92 M
dr = r de
dy1
cos (9) + dye sin(6), = -dy1 sin(g) + dye cos(9).
dt dt dt dt dt dt
In this section, we consider the case when the two eigenvalues Ai and A2 of the
matrix A are real, distinct, and at least one of them is negative, i.e., A2 > Al
and Ai < 0. Throughout this section, assume that all = Al, a22 = A2, and
a12 = a21 = 0. Then, in terms of the polar coordinates (r, 0), system (VIII.7.1) can
be writen in the form
(VIII.7.5)
1
r [Ai(cos(8))2 + A2(sin (e))2 + (e) + g2(g)sin (e))]
dt =
d9
= (A2 - Ai )sin (0) cos (0) + r (e) + 92(y-)cos (e))
Observation VIII-7-2. The point (0,0) is not a proper node of (VIII.7.1). In

fact, if (r(t), 9(t)) is a solution of (VIII.7.5) such that r(t) 0 and e(t) --+ w as
t - +oo or -oo, then d8 --* (A2 - AI) sin w cos w as t -+ +oo or -oo. Hence,
7r,
sin w cos w = 0. This implies that w = -7r, - 0, or
2
Observation VIII-7-3. For any given positive number e > 0, there exists another
positive number p(e) > 0 such that
>0 for - 7r + e < 0 < - 2 - e,

de <0 for -2 + e << 0 < - e,
dt >0 for e<e< -
I<0 for -+e<0<ir-e,
IT
2
if 0 < r < p(e) (cf. Figure 13).
Observation VIII-7-4. If 0 > A2 > Al, there exists a positive number ro > 0
such that dt < Air for 0 < r < ro.
Observation VIII-7-5. In the case when 1\2 > 0 > AI, find a real number wo
such that 0 < wo < 2 and tanwo = -1. Then, for any given positive number
2
E > 0, there exists another positive number p(E) > 0 such that
>0 for (wo - 7r) + E < 0 < -"lo - e,

dr <0 for - wo + f < 0 < wo - E,
dt >0 for wo + E <0< (7r - wo) - E,
<0 for (7r - wo) + f < 8 < (7r + wo) E,
if 0 < r < p(E) (cf. Figure 14).

Observation VIII-7-6. If two positive numbers f and p are sufficiently small,
Observations VIII-7-2 and VIII-7-3 show that the behavior of orbits of (VIII.7.1)
in the sectorial region S = {(r cos 0, r sin 0) : $6 < E, 0 < r < p} looks like Figure
15.
Denote by fl the set of all real numbers w such that
(i) IwI 5 E,
(ii) the orbit (r(t), 8(t)) of (VIII.7.1) such that r(0) = p and 0(0) = w leaves the
sectorial region S through the boundary 0 = -E.
Set a = sup(w : w E 0). Then, ja$ < E. Furthermore, the orbit (r(t), B(t)) of
(VIII.7.1) such that r(0) = p and 0(0) = a satisfies the condition lim r(t) = 0
and Iim 6(t) = 0. This can be shown by using the continuity of orbits with +00 respect
c-.+oo
to initial data (cf. Figures 16-1 and 16-2). Here, use was made of the uniqueness
of solutions of initial-value problem (VIII.7.3).

264 VIII. STABILITY
Observation VIII-7-7. In the case when 0 > A2 > A1, the point (0, 0) is a stable
improper node of (VIII.7.1) as t -+ +oo (cf. Figure 17).
Observation VIII-7-8. In the case when A2 > 0 > A1, the point (0,0) is a saddle
point of (VIII.7.1) (cf. Figure 18).

Observation VIII-7-9. Figures 17 and 18 indicate that there are many orbits
tend to the point (0,0) in the direction 0 = 0 as t +oc. This is the general
situation, as the following example shows.
Fixing two positive numbers c and v, define two regions Do and DI in the y"-plane
as follows (cf. Figure 19):
'D0 = (!7: Iy2I Al, and
A3
(VIII.7.6) 51 >1+ v.
Let µ(y) be a real-valued function such that (a) u is bounded on tit, (b) p is continu-
ouslv differentiable in y" on e - {(0.0)}. (c)U(y) = A3 on Do. and
(d) µ(y') = A2 on DI. Set 9(y) = 0 Then, 9(y) = 0 if y" E Dl.
(l4(y) - A2) Y21-
J
Furthermore, 0 < yl. 11121 5 (1 + ()y1+" if y"" D1, and
I9'(y)I = I (µ(y) - A2)y2I 5 KIy2I < K(1 + E)yi"" 5 K(1 + e)Iyj1+
where K is a suitable positive constant.

Consider the system
(VIII.7.7)
ddll
= -\01- dt = AM + (µ(y) - A21y2
If y"(0) E Do, then

db2 = (.!) 2
Hence, y2 = y1 ' . Now, condition (VIII.7.6)
dy1 ,\I 1/i
implies that g(t) E Do for t > 0, lim r(t) = 0, and lim 9(t) = 0. (cf. Figure 20).
t-+oo t-+oo
.V
(I +E)Yj
(I +e)Yi
Observation VIII-7-10. If we assume some condition on smoothness of §(y-),
there is only one orbit of (VIII.7. 1) that tends to the point (0, 0) in the direction
9 = 0 as t - +oc.
Theorem VIII-7-11. Assume that all = A1, a22 = A2, and a12 = a2I = 0, and
that
(1) A2 > \1 and Al < 0,
(2) g(j) is continuous in a neighborhood of
(3)y-.o
lim " = 0,
lye
(4) a9(y exists and is continuous in a neighborhood of 6.

2
Then, there exists a unique orbit y(t) _ [01(1)1 such that

02(t)
01(t) > 0 for t > 0,
b2(t)
urn 01(t) = 0,
i t-.+oo
lim 02(t) = 0, lim
t--+co 01(t)
= 0.
Proof.
The existence of such orbit is seen in previous observations. To show the unique-
ness of such orbit, rewrite (VIII.7.1) in the form
I
(VIII.7.8) dye = 1\2Y2 + 92(Y) _ All yz + a y192(y - (Aty1)291(Y)

dy1 AIyI+91(91 \Y1 1+ 1 91()J)
AIYI
y2
Next, introduce a new unknown u by y2 = y1u 1or u = yI. Then, (VIII.7.8)
becomes
1 JI2u
1y192(YIU) - \I-91(y1u)
( VIII.7.9 ) y1-
du + u = (L2) u+ 1
dy1
1 + Iy1gI(y1u)
'\
266 VIII. STABILITY
[:1].
where u = Observe that (a) 161 = 1 if Jul < 1, (b) -L (d) = U g(0, y2)
-0 Y2
= d, and (c) - 1 < 0.
i
Write system (VIII.7.9) in the form
l
- 1 I u + G(yi.u).
Q 1 A2u
92(yiu) - a 91(yl U
G(yi, u) = -
1
1 + -9i(y1iZ)
aiyi
ac
It is easy to show that lim 5:u (y1, u) = 0 uniformly on Jul < 1. Hence, there exists
v --o
a non-negative valued function K(yj) such that
!G(yi, u) - G(y1, v)! < K(yi )lu - vl
whenever Jul < 1, !v! < 1 and !y1! is sufficiently small. Furthermore, slim K(yt)
=0.
Let u = 01(yi) and u = ip2(yi) be two solutions of (VIII.7.9) such that
(1) ti i (yi) and .'2(y1) are defined for 0 < yi < r) for some sufficiently small
positive number 'I,
(2) lim i'1(yi) = 0 and lim i2(yl) = 0.
vl o+ v,
Set '(yi) = 01 (y1) - s (yi ). By virtue of uniqueness, assume without any loss
of generality that y(yi) > 0 for 0 < yi < ri. Then, for a sufficiently small positive r?,
d
y1 -)tt'(yi) for 0 < y1 < i , where 7 = 2 1 I < 0. This implies that
J
d
[y1 "V,(yi)j < 0 for 0 < yi < rj. Hence, 0 < y1 lim yi 'W(yi) = 0 for
dyi Y:- 0
0 < y1 < 17. Therefore, ;1(yi) = 0 for 0 < y1 < r).
The materials of this section are also found in (CL, §§5 and 6 of Chapter 151.
VIII-8. Perturbations of a proper node

In this section, we consider a system of the form
(VIII.8.1) = All + where y" E R2,
dt
under the assumptions that
(i) A is a negative number,
(ii) the entries of the R2-valued function g(y} are continuous in a neighborhood
of 0,
(iii) 1§(y -)l < K!y"l1+i in a neighborhood of 6, where K is a nor-negative number
and v is a positive number.
The main result is the following theorem.
8. PERTURBATIONS OF A PROPER NODE 267
Theorem VIII-8-1. Under the assumption given above, the point (0, 0) is a stable
proper node of system (VIII.8.1) as t --, +oo.
Before we prove this theorem, we illustrate the general situation by an example.

Example VIII-8-2. Look at the system
(VIII.8.2) dt = -y + 9(yj, 9(y _ (yi + y2) [

y
where y E R2 with the entries yl and y2 It is known that the point (0,0) is a
stable proper node of the linear system dt = -y. To find the general solution of
(VIII.8.2), change (VIII.8.2) to
(VIII.8.3)
by the transformation y' = e-'Z. Next, introduce a new independent variable s by

s = 2e-2t. Then, system (VIII.8.3) becomes
(VIII.8.4)
ds
-g(z)
Observe that
If a solution y(t) of system (VIII.8.2) satisfies an initial condition
171
(VIII.8.5) 9(o) = n = ?12 '
the corresponding solution z(s) = et y'(t) of system (VIII.8.4) is the solution of the
initial-value problem ds = r(7)2 f zz2
l
= rj, where
z" C 2 /
(VIII.8.6) r(i) = (n,)2 + (m)2.
Hence,
zl(s) = r(ir)sin(r(i)2s + 6(rl)), z2(s) = r(i)7)cos(r(1-7)2s + 0(r1)),

where
2
z, (0) = r(rl sin 111 = r(i) sin( 0(rl

{ Z2(0) = r(n7) cos in = r( n oos ( 2
268 VIII. STABILITY
(cf. Figure 21).

The general solution of (VIII.8.2) is
2
yI(t) = e-tr(7f)sin (L(7 a-2t
y2(0 = e-tr(771 cos (!2e_2t

where y(0) and r(fl) are given by (VIII.8.5) and (VIII.8.6), respectively. Every orbit
of (VIII.8.2) goes around the point (0, 0) only a finite number of times. Figure 22
shows that the point (0, 0) is a stable proper node as t -+ +oo.

Remark VIII-8-3. If condition (iii) of the assumption given above is relaxed, the
point (0, 0) may become a stable spiral point (cf. Exercise VIII-10).
Proof of Theorem VIII-8-1.
We prove Theorem VIII-8-1 in two steps. To start with, using the polar coordi-
nates, write system (VIII.8.1) in the form
dr
= Ar + gI (y) cos(9) + g2(y) sin(g),
(VIII.8.7) J dt
dO
r t = -g1(y) sin(g) + g2(y COs(9)
(VIII.7.4)). By virtue of the assumption given above, there exists a positive

(cf.
number ro such that
(VIII.8.8) 2Ar < Ar + gi (y-)cos(9) + g2(y) sin(g) < 2r < 0

for0<r<ro.
Step 1. We show first that the point (0, 0) is a stable node. To do this, let
(r(t), 9(t)) be a solution of (VIII.8.7) satisfying the initial condition
(VIII.8.9) r(0) = p, 9(0) = w,

where p is a positive number and w is a real number. This solution exists for
t > 0. Furthermore, estimate (VIII.8.8) implies that the function r(t) satisfies
8. PERTURBATIONS OF A PROPER NODE 269
the following two conditions: (a) r(t) is strictly decreasing as t -+ +oo and (b)
lim r(t) = 0. Therefore, we obtain (c) 0 < r(t) <_ p for t >_ 0.
t-. +00
Since r(t) is strictly decreasing for t _> 0, the variable t can be regarded as a
function of r, i.e., t = t(r), where t(p) = 0 and lim t(r) = +oo. Use r as the
independent variable. Then, (VIII.8.7) becomes
-gi (y-)sin(g) + g2(Y)cas(e)

( V1II . 8 . 10 ) d9 F (r, 0 ) _
Tr r[Ar + g1(y) cos(0) + g2(y) sin(0)J
Let 0 = 4D(r) = 0(t(r)) (0 < r < ro) be the solution of (VIII.8.10) satisfying the
initial condition (b(p) = w. Then,
fi(r) = w + r F(s, I(s))ds (0 < r < ro).

JP
Estimate (VIII.8.8) implies that
I - gt (y sin(0) + 92(Y) cos(0) I F(r, 0) < (I g1(y)I + Ig2(y1I)

2Ar2 ar2
for 0 < r < ro. Now, by virtue of condition (iii), the improper integral
to
r-.0 JP F(s, 4(s))ds exists. Observe that 0(t) = 4(r(t)) = w +
F(s, $(s))ds = lim r
J P
r(t)
F(s, (s))ds and that lim r(t)
t--+oo
= 0. Therefore, lim 0(t)
t-'+oo
=w+
P
0
Thus, we conclude that the point (0,0) is a stable node of
JP
(VIII.8.1) as t +oc.
Step 2. In this step, we prove that the point (0,0) is a proper node. To do
this, for a given real number c, find a solution (r(t), 0(t)) of (VIII.8.7) such that
rim 0(t) = c and t -+oo
t-+oo
lim r(t) = 0. Selecting a sequence {pm : m = 1, 2, ... } of
positive numbers such that 0 < p,,, < ro and lim pm = 0, define a sequence of
functions {cm(r) : in = 1, 2.... } by the initial-value problems
d0
= F(r, 0), 0(Pm) = c (m = 1, 2.... ),
WT
respectively. Those functions are defined for 0 < r < ro, and
c+/r F(s, 4m(s))ds (m = 1, 2, ... ).

P
Set
- +«; 4)m(r) = c + jP1.o

Cm = lim F(s, Om(s))ds.
270 VIII. STABILITY
Then, 4b,,,(0) = c,,,, lira c,,, = c, and

m=+oo
(r) = cm + J r F(s, 4,,,(s))ds (m = 1,2.... )

0
for 0 < r < ro. Since the sequence {4D m(r) : m = 1, 2.... } is bounded and
equicontinuous on the interval 0 < r < ro, assume without any loss of generality
that lim fi(r) exists uniformly on the interval 0 < r < ro (otherwise
M +00
choose a subsequence). It can be shown easily that
fi(r) = c + fF(s,(s))ds (0 r ro).
Therefore, 7(r) is a solution of (VIII.8.10) such that rim 1(r) = c.

Define r(t) by the initial-value problem
dr
dt
= )r + 92(17(r))sm(4'(r)), r(0) = ro,
where g(r) = r [c?1]. The function r(t) is defined for t > 0, is decreasing,
and tends to 0 as t -4 +oo. Set 0(t) = $(r(t)). Then, it is easily shown that
(r(t), B(t)) is a solution of (VIII.8.7) such that limo 0(t) = 1i mt(r) = c.
t +0
0
The materials of this section are also found in (CL, §3 of Chapter 151.
VIII-9. Perturbation of a spiral point

In this section, we show that (0,0) is a stable spiral point of (VIII.7.1) as t -, +oo
if (0, 0) is a stable spiral point of the linear system = Ay". The main result is
the following theorem.
Theorem VIII-9-1. If
(i) two eigenvalues of A are not real and their real parts are negative,
(ii) the entries of the L2-valued function ff(y-) is continuous in a neighborhood of
the point (0,0),
(iii) lira MI = 6,
g-alyl
then the point (0, 0) is a stable spiral point of (VIII. 7.1) as t +oo.
Proof.
Assume that
(VIII.9.1) all = a22 = a < 0 and a12 = -a21 = b 0 0.

10. PERTURBATIONS OF A CENTER 271
Then, from system (VIII.7.4), it follows that

dr
ar + 91(yjcos(9) + 92(ylsin(g),
dt
rt = -br - 91(y-) sin(g) + g2(Y)cos(0).
1at
Therefore, a positive number ro can be fouind so that r(t) < r(0) exp J for t > 0
if 0 < r(0) < ro. This, in turn, implies that lim r(t) = 0 and that dt < - if
t -+oo
2
b > 0, while de > - 2 if b < 0, whenever 0 < r(0) < ro. Therefore, lim00(t) = -oo
if b > 0 and lim 0(t) = +oo if b < 0, whenever 0 < r(0) < ro. Thus, we conclude
t too -
that (0, 0) is a stable spiral point of system (VIII.7.1) as t -+ +oo. 0
The materials of this section are also found in fCL, §3 of Chapter 151-
VIII-10. Perturbation of a center

We still consider a system (VIII.7.1) under the assumption that the entries of the
RI-valued function §(y) is continuous in a neighborhood of 0 and satisfies condition
(VIII.7.2). In this section, we show that if the 2 x 2 matrix A has two purely
imaginary eigenvalues Al = ib and A2 = -ib, where b is a nonzero real number,
then the point (0,0) is either a center or a spiral point of (VIAssume
without any loss of generality that the matrix A has the form b and system
[°b J
(VIII.7.4) becomes
dr _
(0) ,
(VIII . 10 . 1) dt = 91Mcos (0) + 92(y-)sin
dB
dt
= -b + r 1
(B) + 92(y')cos (B)).
Observe that system (VI I I.10.1) can be written in the form

dr = 91(y')cos (0) + 92(y-)sin (0)
(VIII.10.2)
d8
-b + r (-91(yjsin (0) + g2(y-')cos (0))
If a positive number p is sufficiently small, the solution r(0, p) of (VIII.10.2) satis-

fying the initial condition r(0, p) = p exists for 0 < 0 < 27r.
Case 1. If there exists a sequence (pm : rn = 1, 2, ...) of positive numbers such
that
lim p,,, =0 and r(2xr, Pm) = Pm (m. = 1, 2, ... ),
m-.+00
then the point (0,0) is a center of (VIII.7.1) in the sense of Bendixson (Bent, p.
261 (cf. Figure 23).
272 VIII. STABILITY
Case 2. Assume that b < 0. If there exists a positive number po such that
r(2ir, p) > p (respectively < p)
whenever 0 < p < po, then the point (0, 0) is an unstable (respectively stable)
spiral point as t - +oo (cf. Figures 24-1 and 24-2).

Example VIII-10-1. The point (0,0) is a center of the system
d
dt [y2J [ yuj + yi +y2 I yy1J
dr
since = 0.
d8
Example VIII-10-2. The point (0, 0) is an unstable spiral point of the system
d = [ Y2 + ` y 2I + Y22 yI
dt l Y2 l -Yi J Y2
1
as t - -oc, since dr
dt
= r2, which has the solution r = 1
ro - t
r(0) = and
ro
r +oo when t ro. (Note: t < ro.)
Example VIII-10-3. For the system
dt = - y, d y = x + x2 - xy + cry2,
the point (0, 0) is (a) a stable spiral point if a < -1, (b) a center if a = -1, and
(c) an unstable spiral point if a > -1.
Proof.
Use the polar coordinates (r, 0) for (x, y) to write the given system in the form
(1+rcos8(cos20-cos0sin8+asin28)) = r2sin8(cos28-cos0sin8+asin20).
10. PERTURBATIONS OF A CENTER 273
Setting
too
r(9, c) _ E r,,, (9)c"', where r(0, c) = c,
m=1
and comparing r(2r, c) and c for sufficiently small positive c, it can be shown that
(1 + Or
r1(t) = 1, r2(2r) = 0, r3(2r) = 4
Thus, r(27r, c) < c if a < -1. Therefore, (a) follows. Similarly, r(2r, c) > c if
a > -1. Therefore, (c) follows.
In the case when a x2 - xy - y2 = (x - +2 ) (x - (1-2 ) . Set
W = - (1 Then. a = (1 Therefore, changing x and y by u =
X - (1 + ,/)y
2
and v = x - (1 -
2
)y, the given system is changed to
du dv
Wv(1 + u), = ---u(1 + v).
dt
Hence, on any solution curve, the function

t2W2v2
W2(v - ln(1 + v)) + (u - ln(1 + v)) = u2 +
is independent of t. This implies that in the neighborhood of (0, 0), orbits are closed
curves. This shows that (0, 0) is a center. 0
Example VIII-10-4. The point (0,0) is a center of the system
dy = bAy' + 9(y), y= y'J

dt y2
if (1) b is a nonzero real number, (2) A = 101 1J, (3) the entries of the R2-
"
valued function g"(yj is continuous and continuously differentiable in a neighborhood
of 0, (4) lim = 6, and (5) there exists a function M(y) such that M is positive
g-6 1Y1
valued, continuous, and continuously differentiable in a neighborhood of 0 and that
8(Mf1)(y) + 8(M 2)(y..) = 0 in a neighborhood of 0, where f, (y-) and f2 (y-) are
the entries of the vector MY + §(y).
Proof
The point (0, 0) is either a center or a spiral point. Look at the system
dt =bAy + 9(y) and

7t- = M(y-
274 VIII. STABILITY
Using s as the independent variable, change the given system to ds = M(y)(bAy+

g"(y)). Upon applying Theorem VIII-1-7, we conclude that (0, 0) is not asymptoti-
cally stable as t - ±oo. Hence, (0, 0) is a center.
The materials of this section are also found in [CL, §4 of Chapter 15] and [Sai,
§21 of Chapter 3, pp. 89-1001.
EXERCISES VIII
VIII-I. For each of the following five matrices A, find a phase portrait of orbits
d:i
of the system = AY.
5 1
I 14 1]' [ 11 31 ' [ 15 11]' [1 5
VIII-2. Find all nontrivial solutions (x, y) = (d(t), P (t)). if any, of the system
-xy2, _ -x4y(1 + y)
dt =
satisfying the condition lim (th(t),tb(t)) = (0,0).
c-.+00
VIII-3. Let f (x, y) and g(x, y) be continuously differentiable functions of (x, y)
such that
(a) f (0, 0) = 0 and g(0, 0) = 0,
(b) (f(x,y),g(x,y)) # (0,0) if (x,y) 0 (0,0),
(c) f and g are homogeneous of degree m in (x, y), i.e., f (rx, ry) = rm f (x, y) and
g(rx, ry) = rmg(x, y). Let (r, 8) be polar coordinates of (x, y), i.e., x = r cos 8
and y = r sin 0. Set F(8) = f (cos 0, sin 8) and G(O) = g(cos 8, sin 8).
(I) Show that
dr
= rm(F(8) cos 8 + G(8) sin 8),
(S) dt
d8
dt
= rm-1(-F(8)sin8 + G(O)cosO).
(II) Using system (S), discuss the stability property of the trivial solution of each
of the following three systems:
(i) = x2 - y2
dt
= 2xy;
4ii
= x3 (x2 + y2) - 2x(x2 + y2)2, = -y3(x2 + y2);
(iii)
d = x4 - 6x2y2 + y4, dy = 4x3y - 4xy3.
EXERCISES VIII 275
VIII-4. Let J be the 2n x 2n matrix defined by (IV.5.2) and let H be a real constant
2n x 2n symmetric matrix. Show that the trivial solution of the Hamiltonian system
a = JHy is not asymptotically stable as t +oo.
V11I-5. Show that the trivial solution of the system
dy'
d = Ay" + 9(t, y-)
is asymptotically stable as t +oo if the following conditions are satisfied:

(i) A is a real constant n x it matrix,
(ii) the real part of every eigenvalue of A is negative,
(iii) the entries of the IRI-valued function g(t, y-) are continuous in the region
0(ro) = To x D(ro) = {(t, yl : 0 < t < +oo, Iyj < ro) for some ro,
(iv) g(t, yj satisfies the estimate
19(t, y) 1 _< eo(t, y')Iyj for (t, y) E o(ro),
where Eo(t, y) is continuous, and positive, andfun Eo(t, y) = 0 in L(ro).

+Ii7 -.o
VIII-6. Show that the point (0, 0) is a stable improper node of the system
dy = Ay + Ng + 9(y), y = Iyil,
where
(1) A is a negative number,
(2) N is a real constant nilpotent 2 x 2 matrix and N # 0,
(3) the 1R2-valued function y'(yj is continuous in a neighborhood of
(4) 19(y-) 11-- cly-1I+" for some positive numbers c and v in a neighborhood of
CA
Hint. Suppose N = 0 2 . If we set yj = r cos 0 and y2 = r sin 6, we obtain
0 0
d = r IA _ EA sin2 cos 0 + 9i (y-)cos 9 + 92(y-)sin O

,
where 9"(y'i = {9i]. Hence, if r(0) is small, r(t) is bounded by r(0) and
92(V
lim r(t) = 0 for t >- 0.
t+0 This implies that if 1#(0)I is small, g(t) is bounded
and lim y(t) = 0'. Look at
t-+o
y(t) = eAtetN {(o) + -sN9(y($))d$]

276 VIII. STABILITY
If we set y(t) = e' e'Nu, we obtain
d(t) = y-(0) + r e-aae-aN9(g(s))ds.

0
Choose e > 0 so that 1 - e > 1 Then, condition (4) implies that
+ +00
e-aae_a'
6(+ = 9(0) +
0
exists.
Now,
(1) if d(+oo) = 0, then d(t) = 0 identically, since, in this case,
d(t) = f+a
e-X$e-ajV9(y(s))ds;
(2) if t7(+oo) = 0 0, then t line-a`y"(t) = i"(+00);
(3) if iZ(+oc) = RZJ with 2 # 0, then limo ( e _)

02
L
Hence, the point (0, 0) is a stable improper node of the given system.
VIII-7. Determine whether the point (0,0) is a center or a spiral point of the
system
= -y, dt = 2x + r3 - x2(2 - x)y.

dt
VIII-8. Show that the point (0, 0) is the center of the system
dt = y
+ xy3 - y7, _ -x + xy2 - yg.
Hint. This system does not change even if (t, y) is replaced by (-t, -y).
VIII-9. For the system
3y + 5x(x2 + y2), dty = 27x + 5y(x2 + y2),

dt =
find an approximation for the orbit(s) approaching (0,0) as t - +oo.
VIII-IO. Show that the point (0, 0) is a stable spiral point of the system
d [y,] [yj + 1 [ y2
dt y2 --
- yIn yi + y2 yi
as t-++oc.
EXERCISES VIII 277
Hint. If we set yl = r cos 0 and y2 = r sin 0, the given system becomes

dr _ dO 1
dt -r' dt = In r'
VIII-11. Suppose that a solution ¢(t) of a system
dy _
AJ + §(y-)
dt
satisfies the condition
lim
t-r 4(t) = 0
for a positive number r. Show that if
(1) A is an n x n constant matrix,
(2) the n-dimensional vector W(if) is continuous in a neighborhood of 0,
(3) lien 9(Y) = 0,
V-6 !y1
then ¢(t) = 6 for all values oft.
(Note that the uniqueness of solutions of initial-value problems is not assumed.)
VIII-12. Assume that
(i) AI.... .. 1n are complex numbers that are in the interior of a half-plane in the
complex A-plane whose boundary contains A = 0,
(ii) there are no relations \, = plat + P2 1\2 + + pnan for j = 1, ... , n and
non-negative integers p,,... , pn such that pi + - - + pn >_ 2,
(iii) f(y' = yam' ft, is a convergent power series in ff E Un with coefficients
jp1>2
fpEC'.
Show that there exists a convergent power series ((u) _ ul'Qp with coefficients
Inl>2 _
dg
Qp E Cn such that the transformation if = u" + Q(ur) changes the system - _
dt
A#+ f (y-) to
j = Au, where A = diag[.1t, A2,4, ... , A ].
VIII-13. Show that the trivial solution of the system
dx dr
- -st(x +X22) + x2ezl+ 2, dt
-22(22 + 22) _ Xlez'+ 2
is aymptotically stable as t +oo. Sketch a phase portrait of the orbits of this
system.
VIII-14. In Observation VIII-6-2, it is stated that if one of the is not real,
then y" is a spiral point. Verify this statement.
VIII-15. Discuss the stability of the trivial solution of the system
j1 = x21 722 = xi(I - x1)

as t - +00.
278 VIII. STABILITY
VIII-16. Find the general solution of the system
= 223 + 2j + X122 + 22,
= 3x4 + + X122 t 2122 + 21X3 + x223,
explicitly.
CHAPTER IX
AUTONOMOUS SYSTEMS
In this chapter, we explain the behavior of solutions of an autonomous system

dg
dt
= f (y). We look at solution curves in the y-space rather than the (t, y-)-space.
Such curves are called orbits of the given system. In general, an orbit does not
tend to a limit point as t -+ +oo. However, a bounded orbit accumulates to a set
as t -4 +oo. Such a set is called a limit-invariant set. In §IX-1, we explain the
basic properties of limit-invariant sets. In §IX-2, using the Liapounoff functions,
we explain how to locate limit-invariant sets. The main tool is a theorem due to
J. LaSalle and S. Lefschetz [LaL, Chapter 2, §13, pp. 56-71] (cf. Theorem IX-2-1).
The topic of §IX-4 is the Poincare-Bendixson theorem which characterizes limit-
sets in the plane (cf. Theorem IX-4-1). In §IX-3, orbital stability and orbitally
asymptotic stability are explained. In §IX-5, we explain how to use the indices of
the Jordan curves to the study of autonomous systems in the plane. Most of topics
discussed in this chapter are also in [CL, Chapters 13 and 16], [Har2, Chapter 7],
and [SC, Chapter 4, pp. 159-171].
IX-1. Limit-invariant sets

In this chapter, we explain behavior of solutions of a system of differential equa-
tions of the form
dg
(IX.1.1)
fW
where y" E Ill;" and the entries of the RI-valued function f (y) are continuous in the
entire y-space II8". We also assume that every initial-value problem
dy
(IX.1.2)
dt
= Ay), y(0) = if
has a unique solution y = p(t, y7). System (IX.1.1) is called an autonomous sys-
tem since the right-hand side f (y) does not depend on the independent vari-
able t.
Observe that p(t + r, r) is also a solution of (IX.1.1) for every real number T.
Furthermore, At + r, rl) = p(r, il) at t = 0. Hence, uniqueness of the solution of
initial-value problem (IX.1.2) implies that p(t +r, rt) = p(t, p(-r, rt)) whenever both
sides are defined. For each il, let T(rl) be the maximal t-interval on which the
solution p(t, r) is defined. Set C(rt) = {y" = p(t, rl) : t E Z(rl)}. The curve C(rl) is
called the orbit passing through the point il. Two orbits C(iji) and C(rj2) do not
intersect unless they are identical as a curve. In fact,
(IX.1.3) C(ijl) = C(i@ if and only if 7)2 E C(il, ).
279
280 IX. AUTONOMOUS SYSTEMS
If f (n) = 0, the point ij is called a stationary point. If , is a stationary point, the

orbit consists of a point, i.e., C(17) = Generalizing property (IX.1.3) of
orbits, we introduce the concept of invariant sets which play a central role in the
study of autonomous system (IX.1.1).
Definition IX-1-1. A set M C R is said to be invariant if i E M implies
C(n) C M.
For example, every orbit is an invariant set.
Remark IX-1-2. If M1 and M2 are invariant sets, then M1 UM2 and M1 f1M2
are also invariant. For a given set f2 C 1R", let Ma (A E A, an index set) be all
invariant subsets of S2, then U M>, is the largest invariant subset of Q.
AEA
Hereafter, assume that every solution p(t, nJ) is defined for t > 0, i.e., (t : t >
0) C Z(n). In general, lim Pit, nom) may not exist. However, if p(t, n) is bounded for
t > 0, the orbit C(171 accumulates to a set as t -+ +oo. This set is very important
in the study of behavior of C(n) as t -+ +oo.
Definition IX-1-3. Let C+(i7-7) denote the set of all y" E II8" such that
lim p(tk, n ) = y f o r some increasing sequence {tk : k = 1, 2, ... } of real numbers
k-++oo
such that lim tk = +oo. The set C+(n) is called the limit-invariant set for the
k-+oo
initial point r7.
The basic properties of are given in the following theorem.

Theorem IX-1-4. If p(t, n) is bounded fort > 0, then C+(n) is nonempty, bounded,
closed, connected, and invariant.
Proof.
(1) C+ (1-71 is nonempty: In fact, let {sk : k = 1,2,... } be an increasing sequence
of real numbers such that lim sk = +oo. Since p(t, n) is bounded for t > 0,
k-.+oo
there exists a subsequence {tk : k = 1, 2.... } such that lim tk = +oo and that
lim p(tk, exists. This limit belongs to f_+ (17).
k-.+oo
(2) The boundedness of C+(n7) follows from the boundedness of p(t,i7) immedi-
ately.
(3) is closed: To prove this, suppose that lim yk = y for 9k E L+(1-7).
k-+oo
It must be shown that y" E L+(771. Since g k E it follows that ilk =
lira p(tk,t, r 7 ) for some {tk,t : I = 1, 2, ... } such that lim tk,l = +00. Choose tk
t-.+oo t-+oo
so that lim tk,t,, = +oo and 19k - P(tk,t,,,'Th !5 ! Then, since ly - P(tk,t4, n)I -
k-.+oo
+ ly - ykl, we obtain lim oP(tk,l,,, n) = y E C+(1_7)
k k
(4) C+(777) is connected: Otherwise, there must be two nonempty, bounded, and
closed sets Sl and S2 in R" such that
2. LIAPOUNOFF'S DIRECT METHOD 281
(1) S1 n S2 = 0,
(2) S1 U S2 L+(rf)
Set d = distance(S{{l, S2) = min{lyl - y21 yl E S1, y2 E S2}. Note that d >
1. :
0. Set also So = g: distance(y, SI) = Then, So is not empty, bounded,
and closed. Furthermore, So n C+(i) =20. Choose two points y"I E SI and
y ' 2 E S 2 and t w o sequences {tk : k = 1, 2, ... } and {sk : k = 1,2,...) of
lim tk = +oo, lim sk = +oo,
real numbers so that tk < Sk (k = 1, 2, ... ), k-+oo k-+oo
limp P(tk, y"1, and k lim G p1sk, r7) = 92. Assume that r), SO <
k
2 and distance(p(sk, r), SI) > Then, there exists a Tk for each k such that
2.
tk < Tk < sk and p-(Tk, rl) E So. Choose a subsequence (at : e = 1,2.... ) of
{rk : k = 1, 2, ... } so that lim at = +oo and lim p(at, r) = # exist. Then,
Y E So n C+ (Y-)) = 0. This is a contradiction.
(5) C+ (1-7) is invariant: It must be shown that if ff E C+ (r), then p1t, y) E L+(1-7)
f o r all t E 11(y). In fact, there exists a sequence {tk : k = 1, 2, ... } of real numbers
such that Iii o tk = +oo and k lim o P-14, r) = 17. From (IX.1.1) and the continuity
k
of P(t,y) of y", it follows that for each fixed t.
li tk, n) = UM P(t, Pptk, r1)) = Pit, y) E L+(7). El
k +co k
The materials of this section are also found in [CL, Chapter 16, §1, pp. 389-3911
and [Har2, Chapter VII, §1, pp. 144-1461.
IX-2. Liapounoff's direct method

In order to find the behavior of pit, r)) as t +oo, it is important to locate
L+(rl). As a matter of fact, if p(t, r)) is bounded for t _> 0 and if a set M contains
,C+ (1-7), then p(t. r)) tends to M as t -+ +oo, i.e., lim inf(1p(t, r)) - yj : y E M) = 0.
t +a:
Otherwise, there must be a positive number co and a sequence {tk . k = 1, 2.... } of
real numbers such that lim tk = +00, limo P140 7) exists, and ipjtk, >)) -y"1 eo
k k
for all y E Jul. Hence, lim -1) V M. This is a contradiction. Keeping this
k-.too
fact in mind, let us prove the following theorem (cf. [LaL, Chapter 2, §13, pp.
56-71]).
Theorem IX-2-1. Let V(y) be a real-valued, continuous, and continuously differ-
entiable function for Jyl < ro, where 0 < ro < +oo. Set
fDt = J#: V (y-) < e}, St = {y" E DI : V-(y) J(y) = 0}.
Mt = the largest invariant set in St,
F' a`, yl fl(
where Vi(y) ' f(y) _ ayJ fJ(y), y , and f( y) _ . Suppose
J-1 yn fn (y)
that there exist a mat number t and a positive number r such that 0 < r < ro,
Dt C {y : Iy1 < r}, and Vg(y-) f (y) < 0 on De. Then, L+ (1-7) C Me for all n E Vt.
Proof.
Set u(t) = V(p(t,rl")). Then,
dot)
= V9(r(t,r7))' dtp(t,i) =
du(t)
Therefore, < 0 as long as p(t, t) E Vt. This implies that u(t) < V(1-7) < e
dt
for r) E Vt as long as p(t, n-) E Vt. Hence, p(t, rl E Dt for t > 0 if ij E Vt.
Consequently, Ip(t, rlI < r for t > 0. It is known that C+ (171 is nonempty, bounded,
closed, connected, and invariant (cf. Theorem IX-1-4). Furthermore, C Vt.
d t)
Let po be the minimum of V(rl') for Iy7 < r. Then, po < u(t) and < 0 if
it E 1)t. Hence, lim u(t) = uo exists and no > po. This implies that V(y) = uo
for all y E C+(t)). Since C+(71 is invariant, V(p(t, y)) = uo for all y E L+(171 and
t > 0. Therefore, VV(p(t, y)) &I t, y)) = 0 for all y E C+ (q) and t > 0. Setting
t = 0, we obtain Vc(y) f (y) = 0 if y E C+(rl, i.e., C St if i E Vt. Hence,
C+() C Mt for if E Vt.
The following theorem is useful in many situations and it can be proved in a way
similar to the proof of Theorem IX-2-1.
Theorem IX-2-2. Let V (y) be a real-valued and continuously differentiable func-
tion for ally E lR' such that VV(y) f (y) < 0 for all y" E R". Assume also that the
orbit p(t, rte) is bounded fort > 0. Set S = 1g: Vi(y) f(y) = 0) and let M be the
largest invariant set in S. Then, L+ (7-1) C M.
The proof of this theorem is left to the reader as an exercise.
In order to use Theorem IX-2-2, the boundedness of p(t, must be shown in
advance. To do this, the following theorem is useful.
Theorem IX-2-3. If V (y) is a real-valued and continuously differentiable function
for all y" E lR^ such that VV(y) f(y) < 0 for Iyi > ro, where ro is a positive number,
and that lim V (y) = +oo, then all solutions p(t, r)7) are bounded for t > 0.
1Q1+00
Proof.
It suffices to consider the case when p(to,rl > ro for some to > 0. There are two
possibilities:
(1) IP-(t,r1)I > ro fort > to,
(2) IP(t,n')I > ro for to < t < t, and Ip1ti,17)1 = ro for some tl > to.
Case (1). In this case, dt V (p( t, r)1) = Vj(p(t, rl) f (p(t, 4-7)) < 0 fort > 0. Therefore,
V (r(t, V (r(to, t))) for t > to. Hence, p(t, tt") is bounded for t > 0.
Case (2). In this case, it can be shown that
V(p(t, 71)) <_ max IV(P(to,' )), max(V(y) : 1yi <_ ro)1
3. ORBITAL STABILITY 283
in a way similar to Case (1). The details are left to the reader as an exercise. 0
IX-3. Orbital stability

In this section, we introduce a concept of stability (respectively asymptotic sta-
bility) in a sense different from the stability (respectively the asymptotic stability)
of Chapter VIII (cf. Definitions VIII-1-1 and VIII-1-3). We denote again by C(rf
the orbit passing through fl. Also, assume that every solution p(t, if) is defined for
t>0.
Definition IX-3-1. An orbit C(sjo) is said to be orbitally stable as t -+ +00 if,
for any given positive number e, there exists another positive number b(e) such that
distanee(p"(t, rl"), C(rlo)) < c for t > 0 whenever 14 - rlo1 < b(e).
This definition is independent of the choice of the point ijo on the orbit C(ilo). If
C(i-lo) = C(#,) and if the condition of this definition is satisfied in terms of ilo, then
the same condition is satisfied also in terms of , with a different choice of b(e).
The following example illustrates these new concepts.
Example IX-3-2. The orbit of the system
[y2] = yl +y2
d. [ - 2]
passing through the point ij = I ] is given by
yj = r(f[)cos{r(i))t + 9(t)}, y2 = r(ij)sin{r(if)t + 9(f))),
where it = [c?]. It is easy to show that if r(n) - r(q) = b > 0 and

9(ili) = 9(ij2), then distance(p(t, f2), C(#,)) = b. Hence, the orbit C(ij) is orbitally
stable as t -+ +oo for every i ) . However, i h ) - p (b ff) I = r(il') + r(')2 )
since 1r(ffi)t+9(ili)1-1r(il2)t+9(il2)j = 6t (cf. Figure 1). Therefore, every nontrivial
solution p"(t, i)) is not stable as t +oc in the sense of Definition VIII-1-1.
Definition IX-3-3. An orbit C(fo) is said to be orbitally asymptotically stable
as t - +00 if
(i) C(ilo) is orbitally stable as t +oo,
(ii) there exists a positive number 6o such that lim C(i )) = 0
whenever {il - iloj < 60.
This definition is independent of the choice of the point 4 on the orbit. However,
the choice of 6o depends on rjo.
Consider a system of differential equations
(IX.3.1) = f (y,
where the entries of the R"-valued function f is continuously differentiable on the

entire y-space W1. Assume also that the solution At, rjo) is periodic in t of period
1 (i.e., #(t + 1, ito) = p1t, i"Jo) for -oo < t < +oo) and that f'(p1t, ijo)) 54 0. The
system of linear differential equations
di7 Of
(IX.3.2) =
is called the first variation of system (IX.3.1) with respect to the solution pit, rlo).
The coefficients matrix of (IX.3.2) is periodic in t of period 1.
Let pi, p2, ... , p be the multipliers of (IX.3.2) (cf. Definition IV-4-5). Since
2 p-(t,
(pit, o)) d p1t, ijo), linear system (IX.3.2) has a nontrivial periodic
solution fjt, im). This implies that one of the multipliers must be 1. Set pl = 1.
The following theorem gives a basic sufficient condition for orbitally asymptotic
stability.
Theorem IX-3-4. If n - 1 multipliers p2,... , p, satisfy the condition ipiI < 1
(j = 2,... , n), then the periodic orbit C(rjo) is orbitally asymptotically stable as
t-4+oo.
Prof.
We prove this theorem in five steps. It suffices to find an (n - 1)-dimensional
manifold M in a neighborhood of the point so that
(1) M is transversal to the orbit C(i) at ilo; i.e., the tangent of C(yo) is not in
the tangent space of M at ilo,
(2) there exist two positive numbers K and r such that
Iu(t,n - p1t,i )I <- KIn - iole-°:
for i EM and t>0

(cf. Figure 2).
FIGURE 1. FIGURE 2.
Step 1. There exists an invertible n x n matrix P(t) whose entries are real-valued,
continuously differentiable, and periodic in t of period 1 (or 2) such that the trans-
formation
(1X.3.3) 6 = P(t)v"
changes (IX.3.2) to
(IX.3.4)
dt =
AV, A = [ 0 B,
where B is a constant (n - 1) x (n - 1) matrix (cf. Theorems IV-4-1 and IV-4-3).
Since the absolute values of n -1 multipliers are less than 1, there exist two positive
numbers Co and a such that
(IX.3.5) Iexp[tB] I < Coe-tot for t > 0.
A fundamental matrix solution of (IX.3.2) is given by
0
exp[tA] = [1
1
1/(t) = P(t) exp[tA],
o exp[tBj J
This implies that the first column vector of P(t) is a periodic solution of (IX.3.2)
and all the other columns of W(t) are not periodic. Hence, assume that
(IX.3.6) P(t) _ [d AtA), Q(t),

where Q(t) is an n x (ii - 1) matrix. Note that det[P(t)] 56 0.
Step 2. Change (IX.3.1) by
(IX.3.7) y = z' + At, rlo)
to
Tt 8 ay(p1t,
(IX.3.8) no))z + h(t, z
dt =
where
h(t,z-) = f(z + pit,t)o)) - I(plt,no)) - y(pit,qo))z
It is easy to show that h(t, 6) = 0 and az (t, 6) = 0.

Step 3. Change (IX.3.8) by the transformation i = P(t)w". Since A =
P(t)-1 18 (p-(t, jo))P(t) - d (,t) , it follows that
L
(IX.3.9) dt = Aw + #(t, tip),

where g(t, w) = P(t)-'h(t, P(t)w). Note that g(t, 6) = 0 and 8w (t, 0') = 0. This
implies that for any given positive number r, there exists another positive number
K(r) such that
(IX.3.10) 19(t, w1) - g"(t, w2)I < K(r)[w1 - t62I for t>0
whenever kw"II < r and I w2I < r and that
(IX.3.11) lim K(r) = 0.
r-.0
Set ti = ti and w2 = 0. Then, (IX.3.10) becomes
(IX.3.12) jg"(t,w)I < K(r)Iti'j for t > 0.
Step 4. Let us write (1X.3.9) in the form

dv _
(IX.3.13) = 9i(t,19), By + g2(t, w),
dt dt
where v and g2 are (n - 1)-dimensional vectors w" _ V and g = I _ J . System

J 111
(IX.3.13) can be changed to the system of integral equations

t
u(t, C))ds,
(IX.3.14) J00
+
v(t, ) = p ( tB1 + f exp [(t - s)BI g2(s, w(s, ))ds,

0
where 1; is an (n -1)-dimensional arbitrary constant vector. In this step, a solution

w(t, t) of (IX.3.14) will be constructed in such a way that
ko 1 ie-ot
(IX.3.15)
for
(IX.3.16) t>0 1t1 < 60, and

where ko and bo are suitable positive numbers.
First fix three positive numbers ko, ro, and 6o so that Ico > 2, kobo < ro, and
K (r)
< 2 for 0 < r < ro. Then, from (IX.3.5), (1X.3.10), (IX.3.11), (IX.3.14), and
a
(IX.3.15), it follows that
+
t9i (s,+u(s,&ds < K(ko[ e-O°ds
(I) t JL0 itt
If1)
K(a
and
exp (tBJ ( + jexE(t -

e-2otlrj + f eo"ds
t
S
0
Se of
I1+ K( (1 + ) < ko1 1e-OL
Next, set lIJt ({) = sup (eot 1 (t, 4)1 : t > 0) if the entries of an R"-valued func-
tion t(i(t, £) are continuous for (IX.3.16) and l}} < kolle-ot for (1X.3.16).
Then,
I
J+9i(s,t,ii(s,())ds - ft 9i(s,tG2(s,&ds l
+oo
< K(ko at
< 2IItLI - 1211(&-ot
and
f exp[(t-s)B]92(s,))ds
t
- f+ exp((t-s)B](s,(IV)
<
Q
<- 2[[i1 - &IOe-oc
for (IX.3.16) if the entries of R'-valued functions 1(t, )) and (t, t; are continuous
for (1X.3.16) and that J ,(t,£)I < (j = 1,2) for (IX.3.16).
ko1&-OC
Let us define successive approximations as follows:
15m(t, =
L gm(t,
where %ao(t, = 0 and
f= g1(s,(s,))ds,
ao
exp[tB) + exp[(t-s)B]92(s, ' m

J0t
Then, it can be shown without any difficulties that lim >im(t, exists
m --+oc
uniformly for (1X.3.16) and the limit %i(t, { is a solution of integral equations
(1X.3.14) satisfying condition (IX.3.15) for (IX.3.16). Note that i(O,t;)
0
where a(t) = f gl (s, z/i(s, })ds. This implies that
00
Step 5. Set
00,6 = Plt,vio)
Then, is a solution of system (IX.3.1) such that
(IX.3.18) At, f) - pjt, i o)[ 5 Ko{f [e-°` for (IX.3.16),
where Ko is a positive constant. Define an (n - 1)-dimensional manifold M by
M = (g=4(0'6: 01<<601-
Using (IX.3.6), (IX.3.17), (IX.3.18), and (1X.3.11), we obtain
(0, f) = P(0)W (0, a + i?o = dP (0, 40) + Q(O) + rlo,
det dp(0, 'b), Q(0)] -76 0.

Thus, the manifold M is transversal to the orbit C(fjo) at rp. Note that (t, t
At, ¢(0, 6).
Moreover, by continuity of the solutions of (IX.3.1) with respect to the initial
conditions, there exist two positive numbers 61 and ro such that for i satisfying the
condition
11 -401
there is a real number r(q such that 1r(q-)l < ro and p(r(g71,'7) E M. Therefore,
(IX.3.18) implies that
1pit+r(17),q - P-(t,il0){ <Koboe-°` for t > 0.

Thus, C(ilo) is orbitally asymptotically stable as t --+ +oc.
Remark IX-3-5. If the entries of an R'-valued function f (t,1/) is continuously
differentiable on the entire f.-space lRn and system (IX.3.1) has a periodic solution
p ( t , i l o ) of period 1 such that f (p(t, i )) j4 0, there exist n - 1 vectors 4j(t) C
Rn (j = 2, ... , n) such that
(i) the entries of vectors 9'J (t) (j = 2,... , n) are continuously differentiable and
periodic in t of period 1,
(ii) n vectors dpig0), q"2(t), ..., form an orthogonal system.
Proof.
If n = 2, it is easy to find 42(t). For n > 3, there is the recipe.
Note first that
dpdtilo} - -
f(p1t,ilo))00. Set g1(t)=
f(At, 40))
f (pit, no)) f (pit, ilo))
where u" b denotes the usual dot product. Since fl(t) is smooth, there exists a
constant vector i71 such that nl ill = 1 and that ili + qi (t) A 0 for 0 < t < 1.
Choosing an orthonormal set {ill.... , in I, where n, are constant vectors, define
Q'n(t) = T)n - +ai(t) (rJi + ji(t))

on(t) (1 = 2, ... , n),
where ah(t) _'h - ji(t). 0

For this construction, see, for example, (U).
Example IX-3-6. The orthogonal system of Remark IX-3-5 is useful to study
solutions of (IX.3.1) in a neighborhood of a periodic solution. To illustrate this
application, let us look at a system
dy
dt
=f
where f is an 1R3-valued function whose entries are continuously differentiable on
the entire. y-space R3. Assume that (SI) has a periodic solution p(t, irjo) of period
I such that f(plt,r'b)) 54 0. In this case, there exist two vectors j2(t) E IR3 and
4fi(t) E L3 such that
(i) the entries of vectors 4"4(t) (j = 2,3) are continuously differentiable and peri-
odic in t of period 1,
(ii) three vectors d
Q , 4'2(t), and 43(t) form an orthogonal system.
dp(ttrlo)
Note that = f (plt, no)).
For any fixed real non-negative number r, the set
P(T) = {l7-,m) + u142(r) + u2Q3(T) (ul,u2) E R2}
is the plane which is perpendicular to the orbit at piT,, o). Letting t, u1, and
u2 be three functions of r to be determined, set
(IX.3.19) f(t) = p( r, no) + u1Q2(r) + u243(T)
From (IX.3.19) and the given system (S1), we derive
dge(r)
dt {/,-/
f(f(t))dT = f(P(r>io)) +
du1
42(r) +
due _
+ u1 + u2 dQ3(r)
dT dT 93(r) d-, dT
y'
This yields
(IX.3.20)
df11 dt ( 42(T) 1
((t) d93(r)1
W- 2(T) f(y(t))dr - 2(t)' dr l u1 - dT J u2,
due dt
((t) dQ2(T)\
dr u1 - 93(t) dg3(r))
dr
and
= [f(PIT'770)) f(PIT,, )) + (fcvlr,no)) u1

(IX.3.21)
d9d(T)) uz]
+ (f(p1r,TIo)) x [f(Ar, *lo}).N(t))}
f( -',
where a" b denotes the usual dot product and we assumed that q"1(r} = 1 (j =
2, 3).
Note that
AM)) = RAT, 7lo))+ul (PIT,r, ))42(r)+ u2 f(YlT,r/o))93(T)+O(ItII +Iu2I)

89 09
Hence, from(IX.3.20) and (IX.3.21), we derive
dt
= 1 + 00U11 + Iu21)
77.
and
(IX.3.22)
dul
42(r) ' l ul + [i(T). (r))1 U2
dr
d9drr))
- (o) . _d_r) Jut - (0) - u2 + O(lul l + 1U21),
93(T) ' (FFIr, 1I0))42(r) 1 ul + [i(r). r, 90))93(r) 11 U2

d7- W
(i(t). ddrr)) ul - (t(t). d (r) J U2 + O(lulI + 1U21)
Let Q(t) be the 3 x 3 matrix whose column vectors are (at, qo)), fi(t), and
,fi(t), i.e., Q(t) = [ f (p1t, rlo)) 6(t) q"3(t)! . The transformation w = Q(t)v changes
the linear system
dtr,
(S2) _ L(PIt,rl *6
to
dv' 0 /31(t) 32(t)

= 0 atl(t) a12(t) 16,
0 a21(t) a22(t)
Using these notations, write (IX.3.22) in the form
dul
d7-
= all(r)ul + a12(r)u2 + 91(T,ul,u2),
(IX.3.23)
1 dug
= a21(r)u1 + a22(r)u2 + 92(T,u1,u2),
dr
where
191(r,u1,u2) - K(r)(lul-V11+1u2-V21) (i = 1,2)

with K(r) > 0 and limK(r) = 0 for lull + lull < r and Ivll + 1v21 < r. Observe
r-O
that the two multipliers of the linear system
j
du' = [all(t) a12(t)
a21(t) a22(t)
U
are also multipliers of system (S2). Therefore, using system (IX.3.23), Theorem
IX-3-4 can be proven. In general, we obtain more precise information concerning
the behavior of solutions in a neighborhood of a periodic solution in this way.
The materials of this section are also found in 1CL, Chapter 13, §2, pp. 321-3271.
4. THE POINCARE-BENDIXSON THEOREM 291
IX-4. The Poincare-Bendixson theorem

In this section, we explain the structure of C+(i-) on the plane. Consider an
R2-valued function f (y) of y E R2 such that the entries of f (y7) are continuously
differentiable on the entire g-plane R2. Denote again by pit, ill the unique solution
of the initial-value problem dg = f (y-), y(0) = i. The main result of this section
is the following theorem due to H. Poincare [Poll] and 1. Bendixson [Ben2].
Theorem IX-4-1. Suppose that the solution pit, qo) is bounded for t > 0 and
that C+(ilo) contains only a finite number of stationary points. Then, there are the
following three possibilities:
(i) C+(i-Io) is a periodic orbit,
(ii) C+(i o) consists of a stationary point,
(iii) C+(i)) consists of a finite number of stationary points and a set of orbits each
of which tends to one of these stationary points as I ti tends to +oo.
To prove this theorem, we need some preparation.
Definition IX-4-2. A finite closed segment t of a straight line in R2 is called a
transversal with respect to f if f (y1 0 0 at every point on t and if the vector f (y)
is not parallel to t at every point on t (cf. Figure 3).
Observation IX-4-3. For every transversal t and every point i , the set tnC+(
contains at most one point (cf. Figures 4-1 and 4-2).
I
PU,

The following lemma is the main part of the proof of Theorem IX-4-1.
Lemma IX-4-4. If pit, qo) is bounded fort > 0 and if there exists a point i7 E
C+(7) such that C+ (11) contains a nonstationary point, then C+(i ) is a periodic
orbit.
Proof.
We prove this lemma in three steps. Since C+(qo) is invariant and rj1 E C+(7 ),
it follows that p1t,i7l) E C+(jo) for all t. Furthermore, C+(ijt) C C+(io) since
C+(ilo) is closed.
Step 1. Let rj be a nonstationary point on C+(t7 ). Also, let t be a transversal with

respect to f that passes through I. Then, t n C+(10) = {77} since rj E G+(ijl) C
C+(no).
Step 2. Since tj E C+(r)l ), there exists a sequence {tk : k = 1,2.... } of real

numbers such that m+oo lira tk = +oo and p1tk, ill) E t (k = 1, 2, ... ). Note that
p(tk, ill) E C+(ilo). Therefore, p(tk, m) = #(k = 1,2.... ). This implies that
there exist two distinct real numbers rl and T2 such that it = p(rl, ill) = (r2, ill)
and hence p(t, rl = p(t, p(ri , ili )) = p1t, p(r2i ill )). This, in turn, implies that
p(t + T1, ill) = p(t + r2i ill) for t > 0. Therefore, the orbit C(ill) is periodic in t of
period Irl - r21. Furthermore, C(ill) C C+(rjo). Note that there is no stationary
point on C(qj ).
Step 3. Since C+(ilo) is connected, it follows that distance(C(ijl), C+(ilo)-C(ni))

= 0. Hence, if C(ijl) # L+ (10), there exists a sequence { k E C+(i o) : k = 1, 2.... }
such that tk C(rj,) (k = 1, 2, ...) and lim £k = { E C(ill ). Assume that
k-.+oo
there exists a transversal t such that E e and lk E t (k = 1, 2.... ). Note that
l; E C(ill) C C+(ilo). Then, k = t c C(ill) (k = 1, 2, ... ). This is a contradiction.
Thus, it is concluded that C+(rlo) = C(iji).
Now, we complete the proof of Theorem IX-4-1 as follows.
Proof of Theorem IX-4-1-
If C+ (, ) does not contain any stationary points, then (i) follows (cf. Lemma
IX-4-4). If C+(no) consists of stationary points only, we obtain (ii), since C+(ilo) is
connected. If f-+ (Q contains stationary and nonstationary points, then C+(7-) for
any point it E C+(ilo) does not contain nonstationary points (cf. Lemma IX-4-4).
This is true also for t < 0. Hence, (iii) follows.
Observation IX-4-5. In cases (i) and (iii), the set 1R2 - C+(ilo) is not connected.
Furthermore, if an orbit C(i)) is contained in C+(i o), two sides of the curve p1t, tl")
belong to two different connected components of R2 - C+(t ). In fact, if we consider
a simple Jordan curve C which intersects with the orbit C(tl at n transversally, then
the curve p(t,,o) intersects with C in a neighborhood of two distinct points on C
infinitely many times (cf. Figure 5).
Theorem IX-4-6. If C(ijo) n C+(ilo) 0 0, then C+(ijo) = C(rjo) and either no is
a stationary point or the orbit C(ilo) is periodic.
Proof
In this case, C(ilo) C C+(ilo). If C+ (8o) contains nonstationary points, the orbit
C(i o) consists of nonstationary points. Choose a transversal a at ilo. Then, it can
be shown that C(ilo) is periodic in a way similar to the proof of Lemma IX-4-4,
since r1o E C+(tlo).
Observation IX-4-7. If C(ilo) n L+(no) = 0, then lim distance(plt,ilo) and

t-+oo
,C+(Q) = 0. It follows that if C+(ilo) consists of a stationary point then
lim p(t, irjo)
t+oo If C+(ijo) is a periodic orbit, then C+(rjo) is called a limit
cycle (cf. Figures 6-1 and 6-2).
5. INDICES OF JORDAN CURVES 293

The materials of this section are also found in [CL, Chapter 16, §§1 and 2, pp.
391-3981 and (Har2, Chapter VII, §§4 and 5, pp. 151-1581. In [Hart[, the Poincare-
Bendixson Theorem was proved without the uniqueness of solutions of initial-value
problems.
IX-5. Indices of Jordan curves

In this section, we explain applications of index of a Jordan curve in the plane to
the study of solutions of an autonomous system in 1R2. Consider again an R'-valued
function f (y-) of y E 7,k2 whose entries are continuously differentiable on the entire
y-plane R. Denote also by pit, J) the unique solution of the initial-value problem
dg
dt
= f (y1, yi(0)
To begin with, let us introduce the concept of indices of Jordan curves. Let C
be a Jordan curve y = il(s) (0 < s < 1) with the counterclockwise orientation (cf.
Figure 7). Assume that &(s)) # 0 for 0 < s < 1. Set u(s) = f ('l(s)) (0 <
If(s))1
s < 1). Then, il(s) is the unit vector in the direction of f'(il(s)). There exists a
real-valued continuous function 0(s) defined on the interval 0 < s < I such that
a (s)
(s) _ Isin0(s)I'
Definition IX-5-1. The index of the Jordan curve C with respect to the vector
0(1) - 8(0) .
field f (y) is given by I f{C) =
2r,
This definition is independent of the choice of a parameterization rj(s) of C and the
function 0(s). Let us denote by D the domain bounded by C (cf. Figure 7).
Observation IX-5-2. If the domain D is divided into two domains DI and D2
by a simple curve £, then 1 f{C) = l f{",) + If-(8D2) if f (rl) 0 on £ U C, where
8D? (j = 1, 2) denote the boundaries of domains DI and D2, respectively, as the
portions of I f{8Di) and I j{d Dt) along £ canceled each other (cf. Figure 8).
Observation IX-5-3. If All) A 0 on C U D, then I f{C) = 0. To show this, divide
V into sufficiently small subdomains, use the fact that the vector field 19(s) has no
change on a small subdomain, and apply Observation IX-5-2.
Observation IX-5-4. Assume that a point d E V is a stationary point, i.e.,

f(d) = 0. Assume also that f(,) 96 0 on C U V except at a. Then, If(C) depends
only on aa. Therefore, we define the index of an isolated stationary point d with
respect to f by Ilea) = I f(OV), where V is a neighborhood of as such that there are
no stationary points in V other than a.
Observation IX-5-5. If D contains only a finite number of stationary points
N
dl, d2, ... , dN, then I j<C) = FI J(a"k). Proof of this result is similar to that of
k=1
Observation IX-5-3.
Observation IX-5-6. If f (y) # 0 on C and if f(,7(s)) is tangent to C at every
point i(s) (0 < s < 1) of C, then I1{C) = 1 (cf. (CL, Chapter 16, §4, Theorem
4.31).
Proof.
It is evident that the index of C with respect to f and the index of C with respect
to the tangent vector to C are the same, i.e.,
If4C) = hn)(C)-
Assume without any loss of generalities that
# ( 0 )=# ( I ) ) = 0 and i (s) > 0 for 0 < s < 1,
where rig (s) (j = 1, 2) are the entries of the vector r"(s) (cf. Figure 9).
- __11
n2=0
(0.0)
FIGURE 7. FIGURE 8. FIGURE 9.
For 0<r<s<1,set
if T = s,
V(T,s) = if T < s (mod 1),
1 (T, s) _ (0, 1).
Then, v(T, s) is continuous in (T, s) for 0 <_ r < s < 1. Hence, studying how 16(0, s)
changes from s = 0 to s = 1 and how i6(T,1) changes from T = 0 to T = 1, we
obtain Ij.(.)(C) = 1. 13
Observation IX-5-7. If C is a periodic orbit of the autonomous system
di
(IX.5.1) f(+1+
dt
then 1 f-(C) = 1.
Observation IX-5-8. If C is a periodic orbit of autonomous system (IX.5.1), then

the domain D contains at least a stationary point (cf. Observations IX-5-3 and
IX-5-7).
Observation IX-5-9. Suppose that f(it(s)) is tangent to C only at a finite number

of points ij(rl ), #(7-2),. .. ,1t(TN) on C. Assume also that at each point it(rk), either
(1) p(t,it(rk)) V V for ItI < 5k, or
(2) p(t, it(rk)) E D for 0 < ItI < bk,
for some sufficiently small positive numbers 51, ... , SN. Let us call rj(rk) an exterior
(respectively interior) contact point in case (1) (respectively (2)) (cf. Figure 10).
Let us denote by E (respectively H) the total number of interior (respectively
exterior) contact points among N points it(-r1), ij(r2), ... , #(T-N). Then E+H = N
and
(IX.5.2) 1 f-(C) =
E-H+2
2
Sketch of proof.
Let rj(r,) and fj(rk) be two consecutive contact points such that T. < rk. Then,
(i) if both of these two points are exterior contact points, then the tangent to C
changes 7r in angle more than the vector field f does from the point #(T.) to
the point fj(rk) (cf. Figure 11-1),
(ii) if both of these two points are interior contact points, then the vector field f
changes it in angle more than the tangent to C does from the point #(T,) to
the point it(rk) (cf. Figure 11-2),
(iii) if these two points are an exterior contact point and an interior contact point,
then the tangent to C and the vector field f change in the same amount in
angle (cf. Figures 12-1 and 12-2).
Since the total amount of change of the tangent to C in angle is 27r (cf. Observation
IX-5-6), we arrive at formula (IX.5.2). 0
q(

Observation IX-5-10. Let d be an isolated stationary point. Assume that a

neighborhood V of d is divided into a finite number of sectorial regions S1, S2,... , SN
by a finite number of orbits C(rji ), C(il3),... , C(iv) in such a way that
(a) rli, Tt2, - , nN E 8V,
(Q) either p1r,rjk) E V fort > 0 and tends to 99 as t -+ +oo, or p5r,ilk) E V for
t < 0 and tends to d as t - -cc (cf. Figures 13-1 and 13-2).
FIGURE 12-1. FIGURE 12-2. FIGURE 13-1. FIGURE 13-2.

Let us assume also that each sectorial region Sk satisfies one of the following
four conditions:
(I) if it E Sk and {il - d] is sufficiently small, then C(r)) C Sk and pit, q-) tends to
d as Itl +oo (cf. Figure 14),
(II) if i) E Sk and Jil - dl is sufficiently small, then pit, i)) E Sk only on a finite
t-interval ak < t < Yk (cf. Figure 15),

(111) if it E Sk and ail - dl is sufficiently small, then p1t,i) tends to dd in Sk as
t +oc, but plt, q") does not tend to dd in Sk as t -+ -oo (cf. Figure 16-1),
(IV) if ij E Sk and Jrj - al is sufficiently small, then p1t,71) tends to ad in Sk as
t -. -oc, but plt, q+) does not tend to dd in Sk as t - +oo (cf. Figure 16-2).
The sectorial region Sk is said to be elliptic (respectively hyperbolic) in case (I)
(respectively (II)). In cases (III) and (IV), the sectorial region Sk is said to be
parabolic. Let us denote by E (respectively H) the total number of elliptic sectorial
regions (respectively hyperbolic sectorial regions) among St, ... , SN. Then, it is
known that
(IX.5.3)
E--H+2
2
This result is similar to formula (1X.5.2) of Observation IX-5-9. For a proof in

detail, see (Har2, Chapter VII, §9, pp. 166-172).
Observation IX-5-11. Suppose that the vector field f (y e) depends on a param-

eter e E 0 continuously, where 0 is a connected set. In this case, if f (if(s), e) # 0
at every point if(s) on the Jordan curve C for all e E A, then the index I f(.,()(C) is
a constant independent of e. In fact, in this case, the integer I f( E)(C) depends on
e continuously, and hence it is a constant.
Example IX-5-12. If an isolated stationary point d is a node, a spiral point, or
a center, then E = H = 0. Hence It{a") = I.
Example IX-5-13. If an isolated stationary point d is a saddle point, then E = 0
and H = 4. Hence -1 (cf. Figure 17).
FIGURE 16-1. FIGURE 16-2. FIGURE 17.

Example IX-5-14. More general cases of isolated stationary points a are shown
by Figures 18-1, 18-2, and 18-3. In fact,
1E = 0, H = 2 and hence II-(a") = 0 in the case of Figure 18-1,
E = 2, H = 0 and hence It-(d) = 2 in the case of Figure 18-2,
E = 1, H = 1 and hence l1-(d) = 1 in the case of Figure 18-3.
Example IX-5-15. Let p(z) be a polynomial in a complex variable z with complex

coefficients. Set y' = t(zj and y2 = !3(zJ (i.e., z = yl + iy2) and regard the
dz
differential equation = p(z) as a system of two differential equations
(IX.5.4) _ 9(p(yl + iy2)], dtY2 = (p(yl + iy2)]

ddt'
on the y"-plane. It is easy to see that, if zo = 171 +ir12 is a zero of p(z) of multiplicity
m, then if#?) = m, where AM _ R(p(yl + iy2)] and i) = [112, 111
. For example,
%P(YI + iy2)]
if p(z) = iz2, system (IX.5.4) becomes
dtl -2y1Y2, ddt2

= y1 - y2
The point 0 is an isolated stationary point and II<O) = 2.
If we define a vector field f (y, e) by f (W, e) = [!1], then, 1- (0) = 2 for

e > 0 (cf. Observation IX-5-11).
Example IX-5-16. Assume that f ( = fl (pl'Y2) satisfies the condition f (arty)

f2(y142)
= aP f (rte, where A is a real variable and p is an integer. Set y = r os0 I. Then,
the autonomous system = f(it) can be written in the form
dr
r dO rPF2(0),
dt = rPFI (0), dt
where
F1(0) = f 1(cos 8, sin o) cos 0 + f2(cos 0, sin 0) sin 0,
Sl F2(0) = -f1(cos0,sin6)sin0 + f2(cos0,sin0)cos0
(cf. Exercise VIII-3). If a ray to is defined by F2(0) = 0, then to is an orbit of the
system dff = f (r))). Thus, the entire y"-plane can be divided into sectorial regions
by those orbits QB determined by equation F2(0) = 0. For example, in the case of
system (IX.5.5), we obtain Fl (0) = - sin 0 and F2(0) = cos 0. Observe that
>0 for -r<0<0, >o for -2 <0<2,

dr
=0 for 0=-r and 0, and =0 for 0 = - and
2
.
2'
dt dt
<0 for 0<0<r <0 for 2<0< 3 1r
Hence, E = 2 and H = 0 (cf. Figure 19). Therefore, I f-{0) = 2 (cf. Example

IX-5-15).
FIGURE 18-1. FIGURE 18-2. FIGURE 18-3. FIGURE 19.

The materials of this section are also found in [CL, Chapter 16, §§4 and 5, pp.
398-402] and [Har2, Chapter VII, §§2 and 3, pp. 146-151, and §§8 and 9, pp.
161-1741. For a geometric and topological treatment of indices, see (Mil.
EXERCISES IX
IX-1. For each of the following three systems, using the given function V (x, y),
show that all orbits are bounded as t -. +oo .
EXERCISES IX 299
Y2;
(1) _ -xy2 - 4y, _ -yx2 + 3x, V(x,y) = 3x2 +4
Y2;
(2) = y, = -x3 - y, V(x,y) = x4 +2
dy
= y, = - (xs - 3x4 + 2x3 + 120x2 - 23x + 5) - (1 + x2)y,
V(x,y) = + J (ss - 3x4 + 2x3 + 120x2 - 23s + 5)ds.

2 0
IX-2. Show that every solution y(t) and its derivative L(t)
(t) of the differential
equation j2 + dt + y3 =0 tend to 0 as t -i +oo.

IX-3. Consider an autonomous system
yj __ fj(11j,112)
(S1) Wt [y2] [f2(yi,y2) '
where f j and f2 are continuously differentiable on the entire (yj, y2)-plane. Assume
that
(1)
fl(yl,y2)>0 for y2>0 and -oo<yl<+oo,
fi(yi,y2) < 0 for y2<0 and - oo < yl < +oo,
(2) fj(yi,1) > 1 and fi(yi, -1) < -1 for -oo < yl < +00,
(3) f2(yi,1) = 0 and f2(yi, -1) = 0 for -oo < y1 < +oo,
(4)
2 - 1)
f2(11j,y2) <_ YI(112 for jy2j < 1 and 0 < yi < +oo,
1
f2(Y1,Y2) ? Y1 (Y22 - 1)
for jy21 < 1 and - oo < yi < 0,
Find G+((n1,n2)) for (771,772) such that In2t < 1.

Hint. There are two possibilities:
Case 1. The solution (yi(t),y2(t)) of (Si) that satisfies the initial condition
(C) yi(0) = 771 and 112(0) = 172
is bounded as t -+ +oo.
Case 2. The solution (y1(t),y2(t)) of (SI) that satisfies the initial condition (C) is
unbounded as t +oo.
In Case 1, G+((nj, n2)) is either {(0, 0)} or a periodic orbit. In Case 2, G+((nj, 92))
is{(x,±1):-oo<x<+oo}.
Examples.
(a) Every orbit in Iy21 < 1 of the system with fl(yl,y2) = y2 and f2(yi,y2) _
yl(y2 - 1) is periodic.
(b) The stationary point (0,0) is a stable spiral point if fl(yl,y2) = y2(2 -
sin(yly2)) and f2(yj,y2) = 2yi(y2 - 1).
(c) The stationary point (0, 0) is an unstable spiral point if ff(yi, y2) = y2(2 +
sin(yiy2)) and f2(yi,y2) = 2y, (y22 1). -
Verify these statements by using the function V(yl, y2) = yi - ln(1 - y2).
IX-4. Let us consider a system
(52) dt - W= [y2 ] , f(U _ [f2(Y-)

where f, (y-) and f2(y-) are continuously differentiable with respect toy in a domain
Do C R2. Assume that
(i) system (S2) has a periodic orbit p-(t, ik) of period 1 which is contained in the
domain Do,
(ii) RA t' v) 6,
(iii) the integral (1 ( (5t, o))) dt is negative.

i (pjt, qo)) +
,9y 0
Show that the periodic orbit p(t, jo) is orbitally asymptotically stable as t -' +00.
Hint. This is called Poincnre's criterion. Look at
dzV 49f
dt =
Then, I det Y(1)I < 1 for the fundamental matrix solution Y(t) of this system such
that Y(0) = 12 (cf. (4) of Remark IV-2-7). Hence, an eigenvalue p of Y(1) must
satisfy the condition dpi < 1. (The other eigenvalue of Y(1) is 1.) Now, use Theorem
IX-3-4.
IX-5. Assume that two functions f (x) and g(x) are continuously differentiable in
drt
x E R. Assume also that the differential equation dt2 + f (x) + g(x) = 0 has
a nontrivial periodic solution x(t) of period I such that f f (x(t))dt > 0. Show
0
that (x(t), x'(t)) is an orbitally asymptotically stable orbit as t +00 in the (x, x')
phase plane.
Hint. Use Exercise IX-4-
IX-6. Show that there exists a nontrivial periodic orbit of the system
dI7 W)
=f (y), y= Ly2J, Lf2h (Y-)
where
(1) the entries of the 1R2-valued function f (y-) is continuously differentiable on the
entire y-plane,
EXERCISES IX 301
(2) f(6) and f(y) ifil96 ,

(3) of (O) = i, eft (0) = 8, 2Z-2 (0) = -2 and (p)
= 1,
ay1 OW ay1
liM +oo(ft(Y1,Y2) + yr) and 1Y11+liM +oo(f2(yt,N2) + yz) exist.
(4) Ivll+l
IX-7. Given that
Y1 - y2 y2
find the total number E of elliptic sectorial regions, the total number H of hyper-
bolic sectorial regions, and the total number of parabolic sectorial regions in the
neighborhood of the isolated stationary point 0 of the system = f (y, E) in the
dt
following two cases: (i) E = 2 and (ii) E _ 2. Also, find I y(0) for e 96 0.
IX-8. Let us consider a system
(S3) dt = f (y),
where the entries of the l 3-valued functionf is continuously differentiable on the
entire y-space R3. Assume that (S3) has a periodic orbit pit,ip) of period 1 such
that f (r'(t, o)) 0 0. Assume also that the first variation of system (S3) with respect
to the solution p(t,' o), i.e.,
d9 Of
(r(t,io))u,
dt
has three multipliers p1 = 1, p2i and p3 satisfying the condition: 1P21 < 1 and jp3j >
1, respectively. Construct the general orbits pit, r) of (53) such that distance(At,17),
C(s'7o)) tends to0ast-+too.
drt
IX-9. Show that the differential equation d-t2 + (x + 3)(x + 2) + x(x + 1) = 0
does not have nontrivial periodic solutions.
Hint. Two stationary points are a node (0,0) and a saddle (-1,0). Furthermore,
setting f(x1,x2) = x2 , we obtain divf(x1,x2) _
-(xt + 3)(x1 + 2)x2 - xt(xt + 1) J
I
-(x1 + 3)(x1 + 2) < 0 if x1 > -1. Also, use index in §IX-5.
IX-10. For the system
= f(x,y) = x(1 - x2 - y2) - 3y, = g(x,y) = y(1 - x2 - y2) + 3x,

dt
(1) find and classify all critical points,

(2) find
= f (x, y) + 9(x, y)
dt
.-
5
for the function V (x, y) = x2 + y2,

dV = 01,
(3) find the set S = {(x): dt
(4) examine if S is an invariant set,
(5) find the phase-portrait of orbits.
IX-11. For the system
dx y2)2
f(x,y,x) = - x(1 - x2 _ + 3xz + y,
dt =
9(x,y,z) = - y(1 - x2 - y2)2 + 3yz - x,
dt =
dz
h(x,y,z) = - z - 3(x2 + y2),
dt =
(1) find all critical points and determine if they are asymptotically stable,
(2) find
dV 8V + h(x, y, x) 8V
dt
= f (x, y, x) 8x
8V + g(x, y, z) 8y 8z
for the function V(x, y, z) = x2 + y21+ z2,
dV
(3) find thesetS=((x,y,z): =0},
(4) find the maximal invariant set M in S,
(5) find the phase portrait of orbits.
IX-12. Find the a phase portrait of orbits of the system
dt = y + (1 - x2)x2(4 - x2).
dt = y,
IX-13. Let f (x, y) and g(x, y) be real-valued, continuous, and continuously dif-
ferentiable functions of two real variables (x, y) in an open, connected, and sim-
ply connected set D in the (x, y)-plane such that 8f (x, y) + Lf (x, y) j4 0 for all
dxt
(x, y) E D. Show that the system = f (x, y), d = g(x, y) does not have any
nontrivial periodic orbit that is contained entirely in D.
IX-14. Let p(z) be a polynomial in a complex variable z and deg p(z) > 0. Set
x = 3t(z] and y = Q [z]. Verify the following statements.
(i) In the neighborhood of each of stationary points of system
(A) d = R[p(z)], dt = `3`[6 (x)],
there are no hyperbolic sectors. Also, system (A) does not have any isolated
nontrivial periodic orbit.
(ii) In the neighborhood of each of stationary points of system
(B)
dt
= 3R[p(z)], L = -`3'[p(x)],
there are no elliptic sectors. Also, system (B) does not have any nontrivial
periodic orbits.
EXERCISES IX 303
IX-15. Find the phase portrait of orbits of the system
d = x2 - y2 -3x+2, d _ -2xy + 3y.
Hint. See Exercise IX-14 with p(z) = (z - 1)(z - 2).

dx dy
IX-16. Find explicitly a two-dimensional system dt = f (x, y),
d t= 9(x, y) so
that it has exactly five stationary points and all of them are centers.
IX-17. Consider a system
dy
(S4) dt = f (y1,
where the entries of the R'-valued function f are analytic with respect toy in a
domain Do C R'. Assume that
(i) system (S2) has a periodic orbit pit, ijo) of period 1 which is contained in the
domain Do,
(ii) AP-Tt,no)) 0
(iii) for any open subset V of Do which contains the periodic orbits p(t, i"p), there
exists an open subset U of V which also contains p'(t, o) such that for any
point i in U, the orbits p'(t, i of (S4) is contained in V and periodic in t.
Show that if U is sufficiently small, for any point i in U, we can fix a positive period
of p1t, y) so that is bounded and analytic with respect to i in any simply
connected bounded open subset of U.
Hint. Apply the following observation.
Observation. Let Do be a connected, simply connected, open, and bounded set in
1Rk and let T2 (j = 1, 2, ...) be analytic mappings of Do to l . Suppose that, for
any pointy E Do, there exists a j such that T3 [yl = y, where j may depends on Y.
Then, there exists a jo such that Tjo [yj = y for all y E Do.
Proof.
Set
E j _ {y E D o : T j [ y 1 = Y - ) . j = 1, 2, ... .
Then,
(1) E. is closed in Do,
+00
(2) Do = U E,,
3=1
(3) Do is of the second category in the sense of Baire.
Hence, for some jo, the set Ego contains a nonempty open subset (cf. Baire's
Theorem). Since Tea is analytic, we obtain Tao [ 7 = y" for all y" E Do. O
For Baire's Theorem, see, for example, [Bar, pp. 91-921.
CHAPTER X
THE SECOND-ORDER DIFFERENTIAL
EQUATION dt2 + h(x) + g(x) = 0
In this chapter, we explain the basic results concerning the behavior of solutions
of a system
112
[-h(yi)112 - 9(111)
as t -- +oo. In §X-2, using results given in §IX-2, we show the boundedness of
solutions and apply these results to the van der Pol equation
(E) x + E(x2 - 1)
d +x=0
(cf. Example X-2-5). The boundedness of solutions and the instability of the unique
stationary point imply that the van der Pol equation has a nontrivial periodic
solution. This is a consequence of the Poincar&$endixson Theorem (cf. Theorem
IX-4-1). In §X-3, we prove the uniqueness of periodic orbits in such a way that it
can be applied to equation (E). In §X-4, we show that the absolute value of one
of the two multipliers of the unique periodic solution of (E) is less than 1. The
argument in §X-4 gives another proof of the uniqueness of periodic orbit of (E). In
§X-5, we explain how to approximate the unique periodic solution of (E) in the case
when a is positive and small. This is a typical problem of regular perturbations. In
§X-6, we explain how to locate the unique periodic solution of (E) geometrically as
e - +oo. In §X-8, we explain how to find an approximation of the periodic solution
of (E) analytically as a +oc. This is a typical problem of singular perturbations.
Concerning singular perturbations, we also explain a basic result due to M. Nagumo
[Na6] in §X-7. In §X-1, we look at a boundary-value problem
y = F (t, y, d i ) , y(a) = o, y(b) = 8.
Using the Kneser Theorems (cf. Theorems 1II-2-4 and III-2-5), we show the exis-
tence of solutions for this problem in the case when F(t, y, u) is bounded on the
entire (y, u)-space. Also, we explain a basic theorem due to M. Nagumo [Na4] (cf.
Theorem X-1-3) which we can use in more general situations including singular
perturbation problems (cf. [How]).
For more singular perturbation problems, see, for example, [Levi2], [LeL], (FL],
[HabL], [Si5], [How], [Wasl], and [O'M].
304
1. TWO-POINT BOUNDARY-VALUE PROBLEMS 305
X-1. Two-point boundary-value problems

In this section, first as an application of Theorems 111-2-4 and 111-2-5 (cf. [Kn]),
we prove the following theorem concerning a boundary-value problem
(X.1.1) d2 = F t, y, dt) , y(a) = a, y(b) = Q
Theorem X-1-1. If the function F(t, yt, y2) is continuous and bounded on a region
11 = {(t, Y1, y2) : a < t < b, lyt I < +oo, Iy2I < +oo}, then problem (X.1.1) has a
solution (or solutions).
Proof.
For any positive number K, the set Aa = {(a, or, y2) : Iy21 < K} is a compact
and connected subset of ft We shall show that A0 satisfies Assumptions 1 and 2
of §111-2 for every positive number K. In fact, writing the second-order equation
(X.1.1) as a system
dyt _ dye
(X.1.2) - Y2, = F(t, yt, y2),
dt dt
we derive
y1(t) = y1(a) + J y2(s)ds,
a
y2(t) = y2(a) + JF(s,yi(s),y2(s))ds.

`
Hence, if (a, y1(a), Y2 (a)) E A0, we obtain
Jy2(t)I < K + M(b - a),

ly1(t)I < lal + [K + M(b - a)](b - a),
where I F(t, yi, Y2)1 < Al on Q. Therefore, Ao satisfies Assumptions 1 and 2 of
§111-2. Thus, Theorem 111-2-5 implies that SS is also compact and connected for
every c on the interval Te = {t : a < t < b}.
We shall prove that if K > 0 is sufficiently large, the set Sb contains two points
(771,772) and (t;t, (2) such that
(X.1.3) n, < 0 < (1.
In fact, by using the Taylor series at t = a, write yt (b) in the form
yt(b) = a + y2(a)(b - a) + (c) (b - a)2,

2 dt
where c is a certain point in the interval 10. Since I ddt2 (c) M, the quantity
yI(b)I can be made as large as we wish by choosing Iy2(a)I sufficiently large. Thus,
there are two points (nt, n2) and ((1, (2) in Sb such that (X.1.3) is satisfied.
Since the set Sb is compact and connected, there must be a point (Q, () in the
set Sb. This implies the existence of a solution of problem (X.1.1). 0
306 X. THE SECOND-ORDER DIFFERENTIAL EQUATION
Example X-1-2. Theorem X-1-1 applies to the following two problems:
(X.1.4) dt2 + sin y = 0, y(a) = a, y(b)
and
(X.1.5)
+ p-+-, =
0, y(a) = a, y(b) = Q
dt2
However, Theorem X-1-1 does not apply to
(X.1.6) d 22 + y = 0, y(a) = a, y(b) = Q.
For more general cases, the following theorem due to M. Nagumo (Na4] is useful.
Theorem X-1-3. Assume that
09 f
(i) a real-valued function f (t, x, y) and its derivatives az and are continuous
in a region V = {(t, x, y) : (t, x) E A, -oo < y < +oo}, where 0 is a bounded
and closed set in the (t, x) -space;
(ii) in the region D, the function f satisfies the condition
(I) 1f(t,x,y)1 : 0(IyI),
where 0(u) is a positive-valued function on the interval 0 < u < +oo such
that
+°° udu
(II) +00;
(iii) two real-valued functions wi (t) and w2(t) are twice continuously differentiable
on an interval a < t f 1 t, wi (t), d
2 dt t) )
(IV)
d2dt d
2 (t) <f t, W2 (t), dt t)) , for a < t <b;
(iv) two real numbers A and B satisfy the condition
(V) wi(a) < A < w2(a), and wi(b) < B < w2(b).
Then, the boundary-value problem
d 22 = f I t, x, ') , x(a) = A, x(b) = B,

has a solution x(t) such that (t, x(t)) E Ao for a < t < b, i.e.,
wi(t) < x(t) < w2(t), for a<t< b.
Proof.
The main tools are the following two lemmas.
1. TWO-POINT BOUNDARY-VALUE PROBLEMS 307
Lemma X-1-4. Let x(t, to,rl) be the solution of the initial-value problem
d2x d .)
= f (t, x, , x(to) x'(to) = n,
where a < to < b, (to, l;) E Ao. Then, for any given positive number M, there exists
a positive number a(M) such that l x'(t, to, l;, il) j < a(M) if
(X.1.7) jr71 < M and (T, x(r, to, e, rl)) E Do for to < r < t or t:5 r < to.
Proof.
Letting L be a positive number such that
(X.1.8) w2(t) - w1(t) < L for a < t < b,

choose a(M) > 0 for any given positive number M in such a way that a(M) > 111
and
a(M) udu
(X.1.9) > L.
+! q5( u)
Suppose that there exist ri and r2 such that to < Ti < r2 < t and that
x'(r1) = M < x'(r) < x'(r2) = a(M) for r1 < r < T2,
where x(T) = x(r, to, t ,17). Then, since x'(r) > 0 for ri < r < r2, it follows that
xO(X'x
x/ (7-) for T1 < -r < -r2.
(T ))) <
Hence,
a(M) u/du
f 0u) -
This contradicts the choice of L by (X.1.9). Therefore, Lenuna X-1-4 is true for
to < r < t. We can treat the case t < r < to similarly, since if we change t by -t,
the differential equation x" = f (t, x, x') becomes x" = f (-t, x, -x'). 0
Lemma X-1-5. Set
7 = rim {a(ba)I wi(a)+e, -w2' (a)-eI,
where e is an arbitrarily fixed positive number. Let also x(t, c) be the solution to the
d2x
= f (t, x, , x(a) = A, i (a) =
)
Then,
(1) two curves x = x(t, c) and x = w2(t) meet for some t on the interval a < t 'Y;
(2) two curves x = x(t, c) and x = wl (t) meet for some ton the interval a < t if (r,x(t,c)) E Ao for
a < r < t. The proof of part (2) is similar.
Proof of Theorem X-1-3.
Now, let us complete the proof of Theorem X-1-3. The main point is that when
two curves x = x(t,c) and x = w2(t) or two curves x = x(t,c) and x = w, (t) meet,
they cut through each other. So look at Figure 1.
(b, B)
(a. A)
t=a r=b
FIGURE 1.
Example X-1-6. Theorem X-1-3 applies to the boundary-value problem
d2x _
(X.1.10) x(0) = A, x(1) = B
dt2 Ax'
if A is a positive number. In fact, assume that 0(u) is a suitable positive constant. If

wi(t) = sinh(ft) -a and w2(t) = sinh(ft)+$ with two positive numbers a and
,3 such that -a < A < j3 and sinh(/) - a < B < sinh(VrA_) +,Q, all requirements
of Theorem X-1-3 are satisfied.
If A is negative, Theorem X-1-3 does not apply to problem (X.1.10). Details are
2. APPLICATIONS OF THE LIAPOUNOFF FUNCTIONS 309
X-2. Applications of the Liapounoff functions

In this section, using the results of §IX-2, we explain the behavior of orbits of a
system
d [yj __ Y2
(X.2.1)
dt y[-h(yi)y2 - 9(yi)
as t
l)y2- 9(y1)
+oo. Set y = I yl
J
and f (yl = [_h(yi
h(x), g(x), and dd(x) are continuous with respect to x on the entire real line R.
JLet us assume that
Also, we denote by p(t,r) the solution of (X.2.1) satisfying the initial condition
y(0) = n
Fi rst set V (y) = + G(yl ), where G(x) = fox g(s)ds. Then,
2Y2
= [9(yi ), Y21, 8y'

- f (y-) = -h(yi) y22.
09
Set also, S = { g : R Y) = 0 y. Then, U E S if and only if either h(yl) = 0 or

l JJJ
Y2 = 0.
Observation X-2-1. Denote by M the set of all stationary points of system
(X.2.1), i.e., M = {9: g(yl) = 0, y2 = 01. Then, M is the largest invariant set in
S if the following three conditions are satisfied:
(1) h(x) > 0 for -oo < x < +oc,
(2) h(x) has only isolated zeros on the entire real line IR,
(3) g(x) has only isolated zeros on the entire real line R.
The proof of this result is left to the reader as an exercise (cf. Figure 2, where
0, 0 and 0).
By using Theorem IX-2-2, we conclude that lim p(t, y) = 17 E M if conditions
t-+00
(1), (2), and (3) are satisfied and if the solution p(t,g) is bounded for t _> 0. Note
that G+(rt) is a connected subset of M.
In Observation X-2-1, the boundedness of the solution p(t, i) for t _> 0 was
assumed. In the following three observations, we explore the boundedness of all
solutions of (X.2.1). Set
G(x) = o
Jo
g(s)ds and H(x) = f
o
X
h(s)ds.
Observation X-2-2. Every solution p(t, q) of (X.2.1) is bounded for t > 0 if

(i) h(x) > 0 for -oc < x < +oo and (ii) lim G(x) = +oo.
1x1 +00
This is a simple consequence of Theorem IX-2-3. In fact, lim V (Y-) = +oo.
191-.+00
Observation X-2-3. Every solution pit, of (X.2.1) is bounded for t > 0, if

(i) h(x) > 0 for -oo < x < +oc, (ii) urn IH(x)l = +oo, and (iii) xg(x) > 0
I=t-+o0
for -oc < x < +oo.
Proof
Change system (X.2.1) to
(X.2.2)
d [zl] Z2 - H(z1)
dt 22 -g(z1)
Y1 = Z1, Y2 = z2 - H(zi).
Denote by qqt,() the solution of (X.2.2) such that z(0) _ . Set V, (z-) = zz2 +
G(z1). Then,
L9 V,
ay - [g(zl), z21, ,P = -g(z1)H(zl)
8z"
Note that g(x)H(x) > 0 for -oo < x < +oo and that
dt V1(ggt,C))
- z(g-(t,C)) F(glt,S)) S 0 for t > 0.
Hence, setting q"(t, S) = z1(t, and V, = c > 0, we obtain
2[z2{t,S)]2 < c for t>0,

since G(x) > 0 for -oo < x < +oo. Therefore, Figure 3 clearly shows that q1 t, C)
is bounded for t > 0. This implies that all solutions of (X.2.2) are also bounded for
t>0.
-
Y2=0 z2=0
{s2, 0) (43,0) (41.0)
di-1 z1=0
FtcuRE 2. F1cuRE 3.
2. APPLICATIONS OF THE LIAPOUNOFF FUNCTIONS 311
Observation X-2-4. Every solution p(t, >)1 of (X.2.2) is bounded for t > 0 if
(i) = lim
+00
H(x) = +oo, (ii) = --Cc
lim H(x) = -oo, (iii) g(x) > a for x > as > 0,
and (iv) g(x) < -a for x < -ao < 0, where a and ao are some positive numbers.
Proof.
In Observation X-2-3, the Liapounof function
Vi(z) = G(zi) = 2[y2 + H(yl)]2 + G(yi)

was used. Now, let us modify Vl to a form
V2(yl = 2[y2 + H(yi) - k(yi)J2 + G(yl)

Then,
OV2
09
_ [[y2 + H(yi) - k(yi )J {h(yi) - dyyt )1 + 9(yi ), 1
Y2 + H(1h) - k(yi) I
and J
(Y ')
J
az {y22
dy, + [H(yi) - k(yi)J y2}
- 9(yi) [H(yi) - k(yl)J
Using (i) and (ii), three positive numbers M, a, and c can be chosen so that
fa [H(x) - cJ > M for x > a > ao,
-a [H(x) + c] > M for x < -a < -ao.
Also choose a function k(x) so that
{c for x > 2a,
(1) k(x) =
-C for x!5 -2a,
dk(x)
(II} Jk(x)J < c and >0 for - on < x - +oo ,
dx
and
dk(x) >
m > 0 for JxJ < a
dx
for some positive number m (cf. Figure 4).
k
k=c
k=-c
x=-2a x=-a x=O x=a x=2a
FIGURE 4.
If a positive number b is chosen sufficiently large,
OV2 .f 0 for lyil ? 2a, or 1y21 > b.
In fact,
- a IH(yl) - cI < -M < 0 for yl > a,

(A) -g(yl) [H(bi) - k(yl)) a IH(y1) + cj < -M < 0 for yl < -a
and
(B) Y2 + IH(yj) - k(yi )11/2 ? 1 for 1yi 15 2a, 1y21 > b
if b > 0 is sufficiently large. Therefore,

av2
(i) ' f = -g(yi) IH(yi) - k(yi)1 < 0 for Iyi I >- 2a, 11/21 < +00,
(u) f< -- (y1 <0 for a < 1yi I G 2a, 1y21 ? b,

09
and
OV2 f < - m {y22 + IH(yi) - k(yi)1 y2}

(iii)
g(yt)IH(yi) - k(yi)I < 0 forlyil < a, 1y21 ? b
if b > 0 is sufficiently large.

Since lim G(x) = +oo, Theorem IX-2-3 implies that every solution of (X.2.2)
1XI +00
is bounded.
Example X-2-5. For the van der Pol equation
2 + e(x2 - 1) dt + x
= 0,
where a is a positive number, h(x) = e(x2 - 1) and g(r) = x. Hence,
G(x) = 2x2 and H(x) = e (3x3 - x)

Therefore, conditions (i), (ii), (iii), and (iv) of Observation X-2-4 are satisfied. This
implies that every solution of the van der Pol equation
d yl y2
X.2.3
( ) dt y2 -E(Y - 1)y2 - yi,
is bounded for t > 0. System (X.2.3) has only one stationary point 6. It is easy
to see that 0 is an unstable stationary point as t -+ +oo. Therefore, using the
Poincaare-Bendixson Theorem (cf. Theorem IX-4-1), we conclude that there exists
at least one limit cycle. In §X-3, it will be shown that system (X.2.3) has exactly
one periodic solution.
3. EXISTENCE AND UNIQUENESS OF PERIODIC ORBITS 313
Example X-2-6. For a given positive number a, the system

[yi _ y2
(X.2.4)
d Y2 J [ -aye - sin (yl )
satisfies three conditions (1), (2), and (3) of Observation X-2-1. But, (X.2.4) does
not satisfy conditions of Observations X-2-2, X-2-3, and X-2-4. Therefore, in order
to prove the boundedness of solutions of (X.2.4), we must use some other methods.
-aye < 0, where f (y = [ij.

In fact, using the Liapounoff function V (y) = -yZ - cos (yl ), we obtain
t
lim
2Since V (y-)
+oo
a
ev. f =
WY=
exists for every solution p-(t, ii) of (X.2.4). Now, observe that
(1) We must have a < 1. Otherwise we would have y2(t,,)2 > 2(a+cos(yi(t,rte))
for t > 0. This implies that dt V (pl t, r7-')) < -2a(o - 1) < 0. This contradicts
(X.2.5).
(2) If -1 < a < 1, the solution p(t, 7) must stay in one of connected components
of the set {y" : V(y) < a + e < 1} for large positive t. Those connected
components are bounded sets.
(3) In case a = 1, we can show the boundedness of p(t, t) by investigating the
behavior of solutions of (X.2.4) on the boundary of the set {g: V(y-) < 1).
X-3. Existence and uniqueness of periodic orbits

In this section, we prove the following theorem (cf. (CL, p. 402, Problem 51).
Theorem X-3-1. Assume that
dg
(i) two real-valued functions h(x) and g(x), and (x) are continuous for -oo <
X < +00,
(ii) g(-x) = -g(x) and h(-x) = h(x) for -oo < x < +oo,
(iii) g(x) > 0 for x > 0,
(iv) h(0) < 0,
(v) H(x) = f h(s)ds has only one positive zero at x = a,
0
(vi) h(x) > 0 for x > a,
(vii) H(x) tends to +oo as x -- +oo.
Then, the system
d [Yyj
(X.3.1)
dt [-h(yi)y22- 9(yi)
has exactly one nontrivial periodic orbit and all the other orbits (except for the
stationary point 0) tend asymptotically to this periodic orbit as t +oo.
Proof.
Change system (X.3.1) to

d zi)
(X.3.2)
dt Lz2J Lz2-
Y2J Lz2 -zH(zi),

Setting
V(z) = + G(z1),
where 2
G(x) = z g(s)ds and z=
J0 [z2]
look at the way in which the function V(z) changes along an orbit of (X.3.2). For
example,
dt V
(zI = 9(zl)[z2 - H(zi)] - z29(z1) = -g(z1)H(z1)
along an orbit of (X.3.2). Hence,
dz2 g(zl) dzi z2 - H(zi)
dzi z2 - H(zl)' dz2 9(z1)
dV 9(z1)H(z1) dV _
H(zi)
dz1 z2 - H(z1)' dz2
along an orbit of (X.3.2).

z1(t' a)be the orbit of (X.3.2) such that z(0, a)
Observation 1. Let z(t, a) = IZ24,a)]
[0]. Then, V (i(0, a)) = 2 a2. There exists exactly one positive number ao such
h
that
[
z1(ao, ao) = 0]
and z2 (t, ao) >0 for 0:5t < oo
for some positive number co, where a is the unique positive zero of H(x) given in
condition (v) (cf. Figure 5).
Observation 2. Since 0 when z2 = H(z1), there exist two positive numbers

r(a) and Q(a)such that
0
z(r(a), a) and 0 < zi (t, a) :5a for 0< t < r(a)
-Q(a)
if 0 < a < ao. Also, since H(x) < 0 for 0 < x < a, we obtain
dtV(z'(t,a)) = -g(zi(t,a))H(zi(t,a)) > 0 for 0 < t < r(a)

except for a = ao and t = oo. Therefore,
(A) V(z"(r(a),a)) - V(z(0,a)) > 0 for 0 < a < a0 (cf. Figure 6).
(0.ao)
zI = 0
FIGURE 5. FIGURE 6.
Observation 3. If ao < a, there exists a positive number r0(a) such that
J z1(ro(a), a) = a, 0 < z1(t, a) < a for 0 < t < ro(a),

10 < z2(t,a) - H(zl(t,a)) for 0 < t < ro(a)
(cf. Figure 7). In particular ro(ao) = Co (cf. Observation 1).

If the variable t is restricted to the interval 0 < t < 7-0 (a), the quantity z2(t, a)
can be regarded as a function of z1(t,a), i.e.,
z2(t,a) = Z(zl(t,a),a),
where Z(x, a) is a continuous function of (x, a) for 0 < x < a and a > a0, and
continuously differentiable for 0 < x < a and a > ao except for x = a and a = a0.
Furthermore, Z(x, al) < Z(x, a2) for 0:5 x:5 a if a0 < a1 < 02 (cf. Figure 7).
Set 2(x, a) = I Z(z, a) J . Then,
g(x)H(x)
,
dx
d
V(Z(x,a)) - Z(x,a) - H(x) > 0 for 0<x<a
and, hence,
3jV(Z(x,a1)) > ±V(i(x,a2)) for 0<x<a
if 00 < 01 < 02. Thus, we obtain
(I) V(zro(al),al)) - V(z(O,al)) > V(E(ro(a2),a2)) - V(E(O,a2)) > 0

for go<a1<a2.
Observation 4. If a0 < a, there exists a positive number ri(a) such that ri(a) >
ro(a), zl (r1(a), a) = a, and zi(t, a) >a for ro(a) <t <7-1 (a) (cf. Figure 8).
(a, z2(rp(a), a)) (a. z2(r0(a), a))
(0. a) O,a
(O.a0) I
z2=H(zt) (
(O.ao)
) z2=H(z0
z2=0 =0
zi =0 zi=a Z, =0
FIGURE 7. FIGURE 8.
Note that z2(ri(a),a) < 0 and that H(zi(t,a)) > 0 for ro(a) < t < 7-1(a) since
zi (t, a) > a on this t-interval. Regarding zi (t, a) as a function of z2(t, a) for
z2(r1(a), a) < z2 < z2(ro(a), a), we obtain
(II) 0 > V(zi(ri(ai),ai))-V(zlro(ai),ai)) > V(z(ri(a2),a2))-V(z(ro(a2),a2))

for ao < al < a2 in a way similar to Observation 3 (cf. Figure 8).
Observation 5. If ao < a, there exists a positive number r(a) such that r(a) >
71(a), z1(r(a), a) = 0, and 0 < zi (t, a) < a for r1(a) < t < r(a) (cf. Figure 9).
Note that z2(t, a) < H(zi (t, a)) < 0 for r2 (a) < t < r(a). Again, regarding Z2 (t, a)
as a function of zi (t, a) in the same way as in Observation 3, we can derive
(III) V(z(r(ai),al))-V(z-'(ri(ai),ai)) > V(i(r(a2),a2))-V4flri(a2),a2)) > 0

if a0 < aI < a2 (cf. Figure 9).
Observation 6. Thus, by adding (I), (II), and (III), we obtain
(B) V(z(r(aI),at)) - V(z(0,a1)) > V(zr(a2),a2)) - V(z0,a2)) > 0

if a0 < al < a2. This implies that the function G(a) defined by
g(a) = V((r(a),a)) - V(z(0,0 = Zz2(r(a),a)2 - 2a2

is strictly decreasing for a > ao as a -+ +oo. Also, 9(a) > 0 for 0 < a 5 a0 (cf.
(A)).
Observation 7. Since
lim
dV =0 uniformly for 0 < z1 < a,
jz21-+oo dzi
z1lim00
a = +oo uniformly for - oo < z2 < +oo,
it follows that
lim (V(i(ro(a),a)) - V(z(0,0J = 0,
a+oo
lim (V(z(rl(a),a)) - V(z(ro(a),a))J = -oo,
a+oo
lim JV(z(r(a),a)) - V(i(ri(a),a))J = 0.
a-.+00
Therefore,
(C) lim Cg(a) = -oo.

a-'+00
Thus, we conclude that g has exactly one positive zero a+, i.e.,
> 0, 0<a<a+,
(X.3.3) 9(a) Zz2(r(a), a)2 - 2a2
= 0, a = a+,
< 0, a>a+

From (X.3.3) and symmetric properties (ii) of the functions h(x) and g(x), we
conclude that a(t, a+) is the only periodic orbit, and all the other orbits tends to
z"(t, a+) asymptotically since
Iz2(r(a), a)I > a if 0<0<a+,

Iz2(r(a),a)I < Of if a> a+.
Thus, we complete the proof of Theorem X-3-1. 0
Remark X-3-2. Condition (iv) of Theorem X-3-1 can be replaced by the following
condition:
(iv') there exists a positive number 6 such that H(x) < 0 for 0 < x < S.
X-4. Multipliers of the periodic orbit of the van der Pol equation
In Example X-2-5, we looked at the van der Pol equation
(X.4.1) LX + e(x2 - 1) d + x = 0,
where a is a positive number. Using Observation X-2-4, it was shown that every
orbit of the system
(X.4.2) d i
[y112 / 112
] dt
[ -E(y1 1)112 - 111 -
is bounded for t > 0. It was also remarked that 0 is the only stationary point of
system (X.4.2) and that the stationary point 0 is not stable as t - +oo. In fact, the
linear part of the right-hand side of (X.4.2) at y = 0 is Ay, where A = [ 01 and
[i]. Since trace[A]=f and det (A) = 1, the stationary point is an unstable
Y21-
node for e > 2, while 0 is an unstable spiral point for 0 < e < 2. Therefore, using
the Poincare-Bendixson Theorem (cf. Theorem IX-4-1), we conclude that there
exists at least one limit cycle. Now, Theorem X-3-1 implies that system (X.4.2)
has exactly one limit cycle and all the other orbits except for the stationary point 6
approach this limit cycle asymptotically as t - +oo. In fact, since h(x) = e(x2 -1),
3
H(x) = e I - x] , and g(x) = x, the seven conditions (i) - (vii) of Theorem X-3-1
3
are satisfied. In particular, the positive zero of H(x) is a = J > 1.
In this section, we prove the following theorem concerning the multipliers of the
unique periodic solution x = x(t, e) of the van der Pol equation (X.4.1).
Theorem X-4-1. The multipliers of the periodic solution x(t, e) of (X.4.1) at 1
and p such that )pI < 1.
Proof.
z
xz
d , then dt = -e(x2 - 1)y2. This implies
11
If we! set v = , wwhere 11 =
2
that eJ (x(t)2 - 1)dt = - / . Now
Now look at Figure 11.
FIGURE 11.
5. THE VAN DER POL EQUATION FOR A SMALL PARAMETER 319
On each of two curves Cl and C2, let us denote y as a function of v by yl (v) and y2(v)
respectively. Note that x2 > 1 on C1, but x2 < 1 on C2. Hence, yl(v)2 < y2(v)2.
This implies that
fT
e (x(t)2 - 1)dt > 0,
where T is a period. Therefore, we can complete the proof by using the following
lemma.
Lemma X-4-2. Assume that two functions f (x) and g(x) are continuously differ-
entiable in R. Assume also that the differential equation
d2
2 + f(x) dt + g(x) = 0
r1
has a nontrivial periodic solution x(t) of period 1 such that f (x(t))dt > 0. Then,
J0
the multipliers of the periodic solution x(t) are 1 and p such that jpI < 1.
Remark X-4-3. If system (X.4.2) has more than one periodic orbit, then at least
one of them must be orbitally unstable. Therefore, the proof of Theorem X-4-1 is
another proof of the uniqueness of periodic orbit of (X.4.2).
X-5. The van der Pol equation for a small e > 0

In this section, we explain a method to locate the unique periodic orbit of dif-
ferential equation (X.4.1) (or system (X.4.2)) for a small e > 0.
Set e = 0. Then, system (X.4.2) becomes
(X.5.1)
dt[y2J LyYi
Every orbit of (X.5.1) is a circle and of period 2n in t. We expect that the periodic
orbit of (X.4.2) must be approximated by one of those circles as e -+ 0. The main
problem is to find the radius of the circle which approximates the periodic orbit of
(X.4.2) for a sufficiently small e > 0.
Since the periodic orbit and its period are functions of e, we normalize the
independent variable t by the change of independent variable
(X.5.2) t = (I +eW)T
so that the period of the periodic orbit becomes 27r for every e > 0. Transformation
(X.5.2) changes differential equation (X.4.1) to
x722
+ e(1+eW)(x2-1)d + (1+&,d)2x = 0
that can be written in the form
d 2X
(X.5.3) + x = e(1 + ew)(1 - x2) - e(2W + ew2)x.
22
Observation X-5-1. In the case when a real-valued function f (r) is continuous

and periodic of period 27r in r, the general solution of the linear nonhomogeneous
d2x
(X.5.4) -T' + x = f(r)
d
is given by
/'r+2n
(X.5.5) x(r) = K cos(r + m) + a- J sf(s)sin(r - s)ds
r
as it is shown with a straight forward calculation, where K and 0 are arbitrary
constants. This solution is periodic in r of period 2ir if and only if
2w 2w
(X.5.6) f (s) sin(s)ds = 0 and f (s) cos(s)ds = 0.
10 J0
Observation X-5-2. In the case when condition (X.5.6) is not satisfied, set
27r 2, f (S)
(X.5.7) S[ f ] = 1 fo f(s) sin(s)ds and C[J] = 1 cos(s)ds.
i 7t J
o
Then,
v(T) = Kcos(r + 0)
(X.5.8) jr+21r 1
+ 2a s f (s) - S[ f ] sin(s) - C[ f ] cos(s)Jsin(r - s)ds
is the general solution of the differential equation
dr2 + x = f(r) - S[f]sin(r) -
Furthermore, v(r) is periodic of period 27r.

Observation X-5-3. Set
f ( X, ,W, E) _ (I + (w)(I - x2) - (2W + EW2)x.
Letting K be a parameter and a function v(r, K, w, f) be periodic in r of period 27r,

set
1 2w
S(K,w,) _
j
f (v(sKw,_(s,K,WE)WJf
2w
I C(K,, E) = -
7r
f (1T cos(s)ds.
5. THE VAN DER POL EQUATION FOR A SMALL PARAMETER 321
Now, let us consider an integral equation

+2 a
v(r, K, w, E) = K cos(t) + s I f lV,
(X.5.9)
2n fr dr, w, EL
- S(K, w, c) sin(s) - C(K, w, f) tos(s)] sin(T - s)ds.
Solutions of (X.5.9) satisfy the differential equation

d 2V Va
(X.5.10) d-,r2 + V = E [f( V, , w, E I - S(K, w, c) sin(r) - C(K, w, E) cos(r)
as long as v is periodic in r of period 2ir. This integral equation can be solved by

using successive approximations in such a way that the solution v(r, K, W, E) is a
convergent power series in c:
(X.5.11) v(r,K,w,E) _ EmVm(T,fi,w) = Rcos(T)+E vl(r,K,w)+ ,
m=0
where vm(r, K, w) are polynomials in (K, w) with coefficients periodic in r of period

2n. For any given positive numbers Ko and wo, there exists another positive number
Eo(Ko,wo) such that series (X.5.11) is uniformly convergent for
I K1 < K0, IwI <- wo, IEI <- Eo(Ko,wo), -00<T<+00.

This implies that S(K, w, c) and C(K, w, E) are also convergent power series in c:
00 0C
(X.5.12) S(K,w,E) = E EmS,,,(K,w) and C(K,w,E) _ F, E"'Cm(K,w),
m=0 m=0
where Sm(K,w) and C,,,(K,w) are polynomials in (K,w) with constant coefficients.
These two power series also converge uniformly for
I K1 <- Ko, IwI < wo, IEI Eo(Ko,wo)
Observation X-5-4. Inserting series (X.5.11) and (X.5.12) into system (X.5.10),
we obtain
d2vi
+ vi = (1 - K2cos2(r))(-Ksin(r)) - 2wKcos(r)
dr2
- So(K,w) sin(r) - Co(K,w) cos(r)
= K3 cos2(r) sin(r) - K sin(r) - 2wK cos(r)
- So(K,w) sin(1T) - Co(K,w) cos(r)
1
= 4 K sin(3r) + - K - So(K, w)J sin(r)
4
- [2wK + Co(K, w)1 cos(r).
Since v1 is periodic in r of period 2,r, we must have

So(K,w) = I K3 - K and Co(K,w) _ -2wK.
This implies that S(2, 0, 0) = 0 , C(2, 0, 0) = 0, and
(2,0,0) x(2,0,0)
8K _ 2 0
--8.
200 8C200 -10
41
Therefore, the system of equations S(K, w, e) = 0, C(K, w, e) = 0 has a solution

K(e) = 2 + 0(e), w(e) = 0(e). The functions K(E) and w(e) are power series in e
which converge if Iej is sufficiently small.
Observation X-5-5. Set
x(t, e) = v (T+ f1w(E) I K(E), w(E), e/f
Then, x(t, e) is a periodic solution of (X.5.3) and
x(t, e) = K(e) cos t + 0(e) as a --+ 0.

(1+Ew(e)
From the fact that
y2(t, E) _ (t, f) = - 1 K(f) t
sin C 1 + ew(e)) + 0(e) as e -+ 0,
we obtain the following conclusion.

Conclusion X-5-6. The unique periodic orbit of (X.4.2) tends to the circle yl +
Y22 = 4 ase-.0+.
X-6. The van der Pol equation for a large parameter

In this section, we consider the van der Pol equation (X.4.1) for a large e. Let
us write (X.4.1) in the form
_ (`1
(X.6.1)
dt [z21
-z1
3
by setting x=z1 and =z2-ef 3 - zl).
21
Set {z2] - I Eu' and t = er. Then, system (X.6.1) becomes
I
d [$2 Il
(X.6.2)
dr
1
where a = -
E
6. THE VAN DER POL EQUATION FOR A LARGE PARAMETER 323
Observation X-6-1. Note that, along an orbit of (X.6.2),

dud
(X.6.3) Q2w1
dwl 01
3 - wtl
Therefore, if 3 > 0 is sufficiently small, the slope dw, of any orbit is small at a
point (wl, w2) far away from the curve C : w2 = 3 - w!. This implies that every
orbit moves toward the curve C almost horizontally (cf. Figure 12).
Observation X-6-2. From Observation X-6-1 a rough picture of orbits of (X.6.2)
is obtained (cf. Figure 13).

Actually, defining the curve Co by Figure 14, we prove the following theorem.
Theorem X-6-3. For a sufficiently small positive number J3, the unique periodic
orbit of (X.6.2) is located in an open set V(/3) such that the closure of V(f3) contains
the curve Co and shrinks to CD as 0 -+ 0+.
Proof of Theorem X-6-8. In eight steps, we construct an open set V(13) so that
(1) the closure of V(O) contains the curve Co,
(2) the closure of V(J3) shrinks to the curve Co as O 0+,
(3) if an orbit of (X.6.2) enters in the open set V(!3) at r = ro, then the orbit
stays in V(/3) for r > ro.
Step 1. Fixing a number a(3) > 2, we use the line segment
C1(a) = j wi, a(3)3 - a((3) I : 0 < wi <_ a(/3)}

as a part of the boundary 8V(i3) of V(/3) (cf. Figure 15).

Note that the slope of an orbit given by (X.6.3) is negative on C1 (O) and decreasing
to -oo as wt increases to a(f3). Also, is 0 on C1(f3).
dw2
Step 2. Let us consider a curve
CO) = j (wi, ws) : w2 = 3' - wl - Op(wi, 0), 1 < wl < a(fl)} ,
where
(i) µ(w1,3) is continuous for 1 < wl < a(0) and continuously differentiable for
1 < wl < a($),
(ii) µ(w1, 0) > 0 for 1 < w1 < a(0), and (iii) µ(a(,3), 0) = 0.
On the curve C2(f3),
3
= -0µ(w1,13)
\ 31 - wl
and, hence,
_ /32w1 Owl
W2 _ f wl _ wl /f l µ(w1, 0)
3
On the other hand, the slope of the curve C2($) is
dw2 _ w1
2
- 1 - ,8 dEc(w1,f)
dwl - dwl
This implies that if µ is fixed by the initial-value problem
(X.6.4)
dwl
1, µ(a(0),,8) = 0,
we obtain
/32w1 dw2
for 1 < wl < a($),
1 dwl
W2 - -
C 3 w'/
where dw2 denotes the slope of C2(0). The unique solution to problem (X.6.4) is
i
given by
a(/3)2 - wl.
Thus, we choose the curve
C2(O) _ {(wt, wz) : u,2 = - wl - O a(0)2 - wi, 1 < w1 < a(0) }

3
as a part of the boundary OV(#) of V(f3) (cf. Figure 16). Note that on the curve
C2(,8),
w2 = -3 - Q a(f3)2 - 1 at w1 = 1.
6. THE VAN DER POL EQUATION FOR A LARGE PARAMETER 325
Step 3. We choose the curve
C3(/3) = {(wIw2): W2 = -3 - Q a(N)2 - w1, 0 < wl < 11
as a part of the boundary OV(/3) of V(/3) (cf. Figure 17).

C
C1(p)

On the curve C3(/3),
3 - 31 - wl l - /3 a(/3)2 - wi
1w3 2
W2
31 wl)
and, hence,
/32w1 0'w1
(,w3
31 wl/
/ a(Q)2 - w1 + 3
+
(31 - wl
dw2
On the other hand, the slope of the curve C3(/3) is This
dw, 1 a(Q)2 - wl
implies that
02W, <dm2
for 0<w1<1,
w2-
(u, -wI ) dw,
3
where denotes the slope of C3(0) (cf. Figure 17). Note that on the curve
awl
C3(3), W2 = - 3 - Qa(Q) at wI = 0.
Step 4. Now, let us fix the positive number a(/3) > 2 by the equation
3
2
(X.6.5) + 3a(3) = a(3) - a(13).
Denote by C+(/3) the curve consisting of Cl(/3), C2(13), and C3()3), i.e., C+(Q) _
C1(Q) UC2(/3) UC3($). Let C_(0) be the symmetric image of C+(Q) with respect to
the point (0, 0). Then, C+(3) U C_ (/3) is a closed curve if a(,Q) satisfies condition
(X.6.5). We use the closed curve C+(/3) UC_(/3) as a part of the boundary 8V(/3)
of V(/3) (cf. Figure 18).
b(3) 3
Step 5. Fixing a positive number b($) < 2 so that 3 = - b(j3) +,A(0), we
choose the curve
b(3)3-
r!(,3) = {twiw2) : w2 = b(j3) +,6b($)2 - w;, 0 < w, < b(13)}
as a part of the boundary 8V(j3) of V($) (cf. Figure 19). Note that, on r,(j3),
2
w2=3 at w, = 0.

On the curve r, (,B),
/32w1 02w1
b(3) 3
w2 Q b(B)2 - wl + - b(8) - (3' w,)
Q wl
>-
b(3) - w,
On the other hand, the slope of the curve 171(3) is dullw2
ap)2'-
wi
Step 6. We use the two curves
r2(0) = {(wiw2): w2
3
3
-wi, 1 <w, <b(S)
ll
r3($) = {(wltw2): U2 -2,
3
0 < wl <
JJ
as parts of the boundary 8V(j3) of V(8) (cf. Figure 20). Note that on r3(Q), the
slope of an orbit given by (X.6.3) is positive and dd is negative.
Step 7. Denote by r+(j3) the curve consisting of r,(8), r2(8), and r3(8), i.e.,
r+w) = r1(Q) u r2(a) u r3(8).
Let r_ (j3) be the symmetric image of r+(j3) with respect to the point (0, 0). Then,
r+(p) U r_(8) is a closed curve. We use the closed curve r+(j3) U r_ (f3) as a part
of the boundary OV(0) of V(8) (cf. Figure 21).
(iii) the function Y(x) is a solution of the differential equation
F (x, Y, I = 0,
d'Yx(x)
and M for a positive constant M on the interval 0 < x < t.
Let y(x) be any solution of the differential equation
AL2 + f (x,y,LY) = 0 (A>0)

satisfying the conditions y(O) = Y(0) and ly'(0) - Y'(0)l l< p. Then,
ly(x) - Y(x)I < {h, + a(L + k)}exp[ Lx]

on the interval 0 < x < if c and A are sufficiently small.
Proof.
Step 1. Setting
y=u+Y(x) and -__v+dY),

change the equation
ad.z + f (x, y, d-T ) = 0 (A > 0)
to the systern
du dv
(X.7.1)
dx = v, F(X, u, v, a),
where
F (x, Y(x), v + d ) I + F(x, u, v,.)l < (e + AM) + Kf ul,

as long as (x, u, v) is in the regi/on l
Do = {(x, u, v) : 0 < x < t, Jul <- a(x), lv) < pe-`= + b(x)}.
Also,
F (riY(x)v + d ()) > Lv for v > 0,
F x, Y(x). v + d ()) < Lv for v < 0,
in Do. Hence, in Do,
(X.7.2))
F(x, u, v, A) < -Lv + Klul + (e + AM) for v > 0,
F(x, u, v, A) > -Lv - Klul - (E + AM) for v 0.
8. A SINGULAR PERTURBATION PROBLEM 329
Step 2. Suppose that two functions wl (x, .A) and w2(x, A) satisfy the following
conditions:
(X.7.3)
10 < wl(x,.A) < a(x), 0 < w2(x,A) < pe-' + b(x),
wl(0,A) > 0, w2(0,.1) > p,
wi (x, A) > W&, A), Aw2(x, A) > -Lw2(x, A) + Kwl (x, A) + (E + AM)
on the interval 0 < x < e. Then, as long as (x, u, v) is in the region
Dl = {(x,u,v): 0 <_ x < e, IuI < wi(x,A), IvI < w2(x,A)},

it holds that
IvI < wi (x, A),
.F(x, u, w2(x, A), A) < Aw2(x, )1),
F(x. u, -w2(x, A), .A) > -aw2(x, I\).
Look at the right-hand side of (X.7.1) on the boundary of V. Then, it can be
easily seen that if a solution of (X.7.1) starts from Dl, it will stay in Dl on the
interval 0 < x < e.
Step 3. Show that two functions
Jwi(x,A) = x A)l + 62(1+1),

J0
pe-Lx/A + Alex:/L
w2(x A) =
ApA e + AM + K62(e + 1)
where bl = L2 + L , satisfy the requirements (X.7.3) if f, A,
and a positive constant 62 are sufficiently small. Observe that two roots of AX2 +
LX -K=0 are -L -(o and (o = L +O(A).
Step 4. Note that
wl (x, A) = (L (1 - e-Lx/A) + bK I (eKx/L - 1) + 82(x + 1).

/
To complete the proof, look at Jul < wl (x, A) as 62 -4 0. 0
d
The inequality
I VI v< wr
2( a, )as 62 0, yields the following estimate of
dy
dx
dY(x) < pe-Lx/A +
dx
fE
L
+
L
(h + M1 ] ex'/L.
L J
Note that
lim a-Lx/A =0 if x > 0.
X-8. A singular perturbation problem

In this section, we look at behavior of solutions of the van der Pol equation
(X.4.1) as a --+ +oo more closely. Setting t = Er and A = c2, let us change (X.4.1)
to
d2x
J1dr2 + (x -
2 1)dx
dr
+x= 0.
Set A = 0. Then, (X.8.1) becomes
(X.8.2) (x2 - 1) d + x = 0.
Solving (X.8.2) with an initial value x(O) = xo < -1, we obtain x = 0(r), where
2 )2 - In I0(r) I = -r + 2 - In(-xo).
Observe that
d0(r)
dr
0(r)
0(7)2-1 >0 if d(r) < -1.
2
Note also that setting ro = 2 - In(-xo) -- 2 > 0, we obtain ,(ro) = -1 and
45(7-)2 - 1 > 0 for 0:5 r < ro. The graph of 0(r) is given in Figure 22.
FIGURE 22.
(i) Behavior for 0 < r < 7-o -bo, where do > 0: Let us denote by x(r, A) the solution
to the initial-value problem
d2z
(X.8.3)
a dr2 +
(x - 1) dx
2
dr- +
x = 0, x(O, A) = xo, 40,A) = 7-7,
where the prime is d7- and q is a fixed constant. Using Theorem X-7-1 (Na6J, we
derive
Ix(r, A) - O(T)I 5 AK,
I Ix'(r, A) - 4'(r)l 5 I7-7 - 0'(0)le-P'1x
+ AK,
for 0 < T < To - 60, where K and a are suitable positive constants.
(11) Behavior for lx+1l< o: First note that lim p0'(T) _ +oo. Set ¢'(ro-2bo) _
1Pi > 0. Then, there exists T1(A) for sufficiently small A > 0 such that
{ O<ri(A)<ro_So, X
llm T1(A) _ 7.0 - 280,
x(T1(A),A) _ O(Tp - 26p), x'(T, A) > for r1(A) < T < TO - 60.
2P1
1A)
Set p = Then,
(X.8.4) ALP = p2(x2 - 1) + p3x,
regarding p = p(x, A) as a function of x and A. Note that
0 < p(x, A) < 2p' for ¢(ro -26o) < x < x(ro-b0, A) < -1, 0 < A < AO,
where A0 is a sufficiently small positive number. It is important to notice that if
we make 60 small and if we make A0 also small accordingly, we can make p1 small.
Set
(X.8.5) o = -1 - ¢(T0 - 26o) and M0 = max Ix2 - 11.

I1+xI<a
Then,
(X.8.6) o > 0, lien o = 0, and lim Mp = 0.

0-+0+
Furthermore, 0 < p1 < Mo since P1 = 1 - ((TO - 25p)2 Therefore,

0(TO - 260)
0 < p(z, A) < 2Mo for ¢(r0-25o) < x < x(ro - 50, A), 0 < A < A0.
Now, we shall show that
(X.8.7) 0 < p(x, A) < 2Mo for Ii + xt < o, 0 < A < A0.
If (X.8.7) is not true, there must exist a such that
p({, A) = 2M0, 0 < p(x, A) < 2Mo for -1-a:5 x<
Then, this implies that
5 A)2(Mo + A)2(1 + g) < 0.

This is impossible.
(iii) Behavior for -1 + or < x < 2 - al: To begin with, it should be remarked that
8
J(2_1)d
[..-_]2 i
= -2- ( -3+1J = 0,
(X.8.8)
1)d. = -x)-3<0, for -1<x<2.
C3
Fix a positive number al so that
(X.8.9) j(2 - 1)4 < - y for -1+a< x < 2-al,

where y is a sufficiently small positive constant. Note that ry --+ 0 as do 0+ and
al -0+-
x,,\-
Integrating (X.8.4), it follows that
p(z ) = P( I, A) - J (C2 - 1) - f 1 agd
(I)P( x<0. Hence, if we choose \0>0

so that
A
y < 4K
7
for 0 < A < A0, where K= f
0
2
w e obtain 0 < p(x, A) < 9K for -1+a <x<0, 0<A <A0.

a
(II) p(x A) > y -K max A) for 0 < x < 2 - al. Suppose that
0< for 0< < x, 0 < A < A0.

2K
2
Then, 0 < p(x, A) < < L. Thus, we proved that
2K
(X.8.10) 0 < p(x,A) < 2K for -1+a < x < 2 - al, 0 < A < A0.
(iv) Behavior for 2 - al < x < 2 + a2: First look at
p(2 - at, A) 2&0), A) -

1) + p(, )ld
Notice that p(2 - al, A) is independent of 5o. Then, it can be shown that
A
0+ as a, -+ 0+ and A 0+.
p(2 - a,, A)
For a fixed positive number a2, assume that
0 < p(x, A) < 1 for 2 - al < x < 2 + a2i 0 < A < Ao.
Then,
2+02
g2 - 1) + P( A)fl > Q.
P(2 + 2, A) p(2 -- o , A) 12-,,
Hence,
2+0a
g2 - 1) + < p(2 - Aal, A) for 2 - al <x < 2 + a2.
J -o,
This is a contradiction if al and A are small. Therefore, if Ao > 0 is sufficiently
small, there exists an x(A) such that
2 - al < x(A) < 2 + a2 and p(x(A),A) = 1 for 0 < A < Ao.
It can be shown that ao*

Urn x(A) = 2.
Setting r(x(A)) = r(A), it can be shown that liM +r(A) = ro. Now again, apply
Theorem X-7-1 (Na6) to the initial-value problem
AT2 + (x2 - 1)d + x = 0, x(r(A)) = x(A), x'(r(A)) = 1.
Figure 23 shows the general behavior of x(r, A) for small positive A.
TO
FIGURE 23.
EXERCISES X
X-1. (1) Show that if F(t,yl,y2) is continuous and bounded on

Il = {(t, Y1, Y2) a < t < b, [yl[ < +oo, 1y2[ < +oo},
the boundary-value problem
F (t, y, &) !LY (a) = a, y(b)
dt2 =
()
has a solution.
(2) Does the boundary-value problem
dtY
dt2
= F t, y, dY
dt
dy(a)
dt = a'
dY (b)
dt
_3
have a solution?
(3) Show that the boundary-value problem
d2x
= tx + x3, x(-2) = A, x(3) = B
dt2
has a solution for any real numbers A and B.
Hint. A counterexample for (2) is
d2 = 0, (n) = 0, (b) = 1.
For (3), apply Theorem X-1-3 [Na4j with w,(t) = a and w2(t) where -a
and 3 are sufficiently large positive constants.
X-2. Find the global phase portrait of each of the following two differential equa-
tions:
d2x
(i) + x2(1 - x)2 dt + (x - 1)2(x + 1)x = 0;
2
d2x dx
dx2
+
dt
- 1
l + x2
= 0.
X-3. Suppose that

(i) f (x, y, t) is continuously differentiable everywhere in the (x, y, t)-space (i.e.,
in p3),
(ii) g(x) is continuously differentiable in -oo < x < +oo,
(iii) e(t) is continuous on 0 < t < +oo,
(iv) f (x, y, t) 0 for x2 + y2 > r2 and t > 0, where r is a positive number r,
r>
(v) G(x) = J g(t)dC --' +oo as jx[ +oo,
0
(vi) 1 je(r)Idr is bounded for 0 < t < +oo.

0`
Show that every solution (x(t), y(t)) of the system
dx _ dy =
-f(x,y,t)Y - g(x) + e(t)
dt - Y ' df
is well defined and bounded for 0 < t < +oo.
EXERCISES X 335
Hint. Set V (x, y) = 2 + G(x). Then, d = - f (x, y, t)yz + e(t)y. There exists a
positive number ro > r such that V(x, y) > 4 for x2 + y2 > ro. Hence, if an orbit
(x(t), y(t)) satisfies conditions that x(t)2 + y(t)2 > ro for to < t < ti, we obtain
d V(x(t),y(t)) < Ie(t)Iiy(t)I <2(e(t)j V(x(t),y(t)) for to <t <t,.
This yields
V(x(t),y(t)) < V(x(to),y(to)) + ft., (e(r)!dr for to <t < t,.
X-4. Show that the differential equation

d yi _ ill
( E.1 ) dt [y2 ] [ -E(yi - 1)y2 - i/i - Eyi
has exactly one periodic orbit, where a is a positive parameter-
X-5. Find an approximation for the unique periodic orbit of (E.1) for small e > 0
and for large e.
X-6. Considering the system of two differential equations
d yi y2
(E.2)
dt [ y2] -
[ -e(yi - 1)y2 + yi - byl
where a and 6 are positive numbers,
(a) show that every orbit is bounded for t > 0,
(b) determine whether each stationary point is stable or unstable, examining every
possibility,
(c) using the function
! y,
-y2 + E(-yi + 6YNY2 + J (-s + bs3J[1 + E2(si)]ds
0
and Theorem IX-2-2, show that if 0 < 6< 3, every orbit of system (E.2) tends to
one of the stationary points as t -' +oo,
(d) discuss the uniqueness of periodic orbits.
X-7. Show that the differential equation
d2x
dz
+ f (x)
d+ 9(x) = 0
has at most one nontrivial periodic solution if fix) and g(x) are continuously dif-
ferentiable in !R and satisfy the following conditions
<0 for 1x3 < 1,
(i)
f (x) { > 0 for IzI > 1,
(ii) 9(x)
<0 for -2<x<0,
> 0 for (x + 2)x > 0,
(iii) J g(x)dx = 0.
Hint. See [Sat].

The main ideas are as follows:
(1) If there are more than one nontrivial periodic orbits, then at least one of them
should not be orbitally asymptotically stable.
(2) For a nontrivial periodic orbit (x(t), x'(t)) of period T > 0, the inequality
fT
f (x(t))dt > 0 implies that this orbit is orbitally asymptotically stable
o
(cf. Exercise IX-5).
(3) Set
a(t) = V(x(t), x'(t)) =

( x'(t))2
2
+ f x(t)
Then,
da(t)
_ -f(x(t))(x'(t))2 .
dt
Therefore, we obtain
(I) J f(x( t) )dt = - Jr (x

dA
' )
2
Now, investigate the quantity (x'(t))2 as a function of A along the periodic orbit
(x(t), x'(t)). For a fixed value of A, compare (x'(t))2 for different values of x(t),
using the assumptions on f and g.
Step 1. First the following remarks are very important.
(a) If we set y = x', the given differential equation is reduced to the system
(S) dt = y, dt = -f (x)y -
A careful observation shows that the index of the critical point (0, 0) is 1,
while the index of the critical point (-2,0) is -1. There are only two critical
points of (S).
(c) The periodic orbit and any line {x = a constant) intersect each other at most
twice.
(d) The critical point (-2,0) should not be contained in the domain bounded by
the periodic orbit. To show this, use the fact that the index of any periodic
orbit is 1.
These facts imply that the periodic orbit should be confined in the half-plane x >
-2.
2 r:
Step 2. A level curve of V (x, y) = 2 + is a closed curve if the curve is
J0
totally confined in the strip jxj < 1. In particular, (1, 0) and (-1, 0) are on the same
level curve V (x, y) = / 1 _ 1 g(C)d{. Since A(t) is increasing as long as
0 Io
jx(t)I < 1, the periodic orbit cannot intersect the level curve V(x,y) = / ,
0
Therefore, the periodic orbit must intersect the lines x = 1 (as well as the line
x = -1) twice. Denote these four points by (1,s1(A)), (1,-rt(B)), (-1,-q(C)),
and (-1, 77(D)), where r)(A), n(B), q(C), and q(D) are positive numbers.
EXERCISES X 337
Step 3. Set
AA = V(1, rl(A)), AB = V (1, -n(B)),

Ac = V(-1, -n(C)), AD = V (-1, rl(D))
Then,
AA > AB, AC > AB, AC > AD, AA > AD-
Set
Z(A) = { A : AA > A > max(AB, AD) },
Z(B) = { A : AB < A < min(AA, )'C) },
Z(C) = { A : Ac > A > max(AB, AD) },
Z(D) = { A: AD < A < min(AA, Ac) }.
Note that the interval lo = JA: min(AA, Ac) > A > max(AB, AD)} is contained in
Z(A) n Z(B) n Z(C) n I(D).
Now,
(a) on the arc between (1, rl(A)) and (1, -rl(B)) of the periodic orbit, regard y as
a function of A and denote it by yi (A), where AB < A 5 AA,
(Q) on the arc between (1, -rl(B)) and (-1, -r1(C)) of the periodic orbit, regard
y as a function of A and denote it by y2 (A), where AB < A < AC,
(ry) on the are between (-1, -rl(C)) and (-1, rl(D)) of the periodic orbit, regard
y as a function of A and denote it by yi (A), where AD < A < AC,
(6) on the arc between (-1, rl(D)) and (1, rl(A)) of the periodic orbit, regard y as
a function of A and denote it by yz (A), where AD :5 A < AA.
Then, it can be shown that
yi (A)2 < y2 (A)2 on I(A),
yi (A)2 < y2 (A)2 on Z(B),
yi (A)2 < y2 (A)2 on I(C),
yi (A)2 < y2 (A)2 on Z(D).
Step 4. Now, fixing a AO E Z(A) n Z(B) n 1(C) n T(D), evaluate the integral (I)
as follows:
rT {JA8 dA rc dA rAD dA AA dA 1
J
f(s(t))dt = -
w yi (A)2 + J y2 (A)2 +'Ac yl (A)2 + IA a y2 (A)2 T ,
where
dA dA r" dA
(A)2,
J\A yi (A)2 as yi (A)2 lao yi
rac dA _ ra° dA
AC
dA
+
Jib
y2(A)2,
Joe y2 (A)2 = Jaa y2 (A)2

I" dA ra° dA rAc dA
yi (A)2 = - JaD yi (A)2 - ao yi
(A)2,
(A)2.
ao ys (A)2 aD y2 (A) 2 Jao y2
Thus, we conclude that

T f (x(t))dt > 0.
J
This implies that every periodic orbit is orbitally asymptotically stable. Hence,
there exists at most one periodic orbit.
X-8. For a nonzero real number c and a real-valued, continuous, and periodic
function f (t) of period T which is defined on -oo < t < +00, find a unique
solution 0(t, E) of the differential equation
E
dt = -y + f (t)
which is periodic of period T. Also, find the uniform limit of 0(t, c) on the interval
-oo<t<+ooasE-+0.
X-9. Assuming that X(t) satisfies the condition
(X(t)2 - X(t) = 0 and IX(t)I > 1
on an interval 0 < t < t, where a is a positive constant, and that x(t, A) is the
unique solution to the initial-value problem
22 + (x2 -1) dt + x = 0, x(O, A) = X (O), x'(O, A) = X'(O),
find lim d2x(2' as a function of X (t) for 0 < t < Q.

a-.o+ dt
X-10. Assuming that y(t, E) is the solution to the initial-value problem
d2 Q
t2Y
dt2
+ 2ydy
dt
= 0, y(0) = a, y'(0) =
where e is a positive parameter, or is a real number, and ,8 is a positive number,
find lim y(t, e) for any fixed t > 0.
f 0+
Hint. Look at E dt = c - y2. Then, c = Q + a2 > a2. Hence, y2 increases to c

very quickly. Then, use Theorem X-7-1 [Na6], or find the solution explicitly (cf.
Exercise I-1).
X-11. Find, if any, solution(s) y(t, c) of the boundary-value problem
4
dt2 + 2ydty = 0, y(0) = A, y(1) = B,
in the following six cases: (1) 0 IAA, (4) B = -A > 0,
(5) CBI < -A, and (6) B < 0 < A, assuming that e is a positive parameter. Also,
find lim y(t, e) for 0 < t < 1 in each of the six cases.
e 0+
EXERCISES X 339
Hint. Use explicit solutions together with the Nagumo Theorems (Theorems X-
1-3 and X-7-1) on boundary-value and singular perturbation problems. See also
Exercises X-10 and X-12, and [How].
X-12. Let f (x, y, t, e) be a real-valued funcion of four real variables (x, y, t, e).
Assume that
(i) 0(t) is a real-valued, continuous, twice-continuously differentiable function
on the interval Zo = {t : 0 < t < 1) and satisfies the conditions 0 =
f (¢(t), 4'(t), t, 0) and 4(1) = B on Zo, where B is a given real number,
(ii) the function f (x, y, t, e) and its partial derivatives with respect to (x, y) are
continuous in (x, y, t, e) on a region R = {(x, y, t, e) : Ix - ¢(t)I < rl, IyI <
+00, t E Zo, 0 < E < r2}, where r1 and r2 are positive nmbers,
(iii) If (4(t), 4'(t), t, e) I < Ke for t E Zo, where K is a positive number,
(iv) there exists a positive number µ such that Of (x, y, t, e) < -p on R,
(v) there is a positive-valued and continuous function '(s) defined on the interval
J+00
0 < s < +oo such that +oo and that If (x, y, t, e)I < +'(IyI) on
R,
(vi) A is a given real number.
Then, there exists a positive number co such that for each positive f not greater
than co, there exists a solution x(t, E) of the boundary-value problem
d2x
f d#2 = f (X, d , t, E) , x(0, e) = A, x(1, c) = B
such that
Ix(t,E) - di(t)I < IA - h(0)led=te-ft/` +C2e on Zo,
x'(t, e) - 4'(t) C3e + co for 0 < c5e <_ t < 1,

e
where c1 i c2, c3, c4, and cs are positive numbers.

Hint. Cf. Theorem X-1-3. See also [How].
X-13. Let a(t), b(t), and f (t) be real-valued, continuous, and continuously differ-
entiable functions of t which are defined on the interval 0 < t < 1. Also, let 4o(t)
be a real-valued solution of the differential equation a(t) dt + b(t)z = f (t).
For a positive number e, denote by y(t, e) the unique solution of the initial-value
problem
d3y dy
+ d-t2 + a(t) + b(t)y = f (t),
3
y(0) ='7o(E), y'(0) = 171 W, y"(0) = 12(E)-
Show that ly(t, e) -00(t)I+ Iy'(t, e) -40 (t)I +Iy"(t, e) -40 (t) I tends to zero uniformly
on the interval 0:5 t < 1 if 4o(0)I + Iti1(E) - 40(0)I + Irn(E) - 0011(0)1 --+ 0 as
e-0+.
X-14. Let Y E lR", y E Rm, t E R, and e E R. Also, let ft f, y, t, e) and g(x", y, t, e)

be respectively R'-valued and 1R'"-valued functions of (x, y, t, e). Assume that
(i) an R"-valued function fi(t) and an R'-valued function t%i(t) are continuous
and continuously differentiable on an interval Zo = {t : a < t < b} and satisfy
the system of equations
dx"
= f (Y, i, t, 0), 0 = g(x, y, t, 0) on Zo,
dt
(ii) f (Y, y, t, e), g(x', y, t, e), and their partial derivatives with respect to (x, y-) are
continuous on a region R = {(:e, y, t, e) : I x - m(t) I < rl, Iy - j(t) I < r2, t E
lo, 0 < e < r3 }, where rl, r2, and r3 are positive numbers,
(iii) all eigenvalues of the matrix (1(t), ti(t), t, 0) are less than a negative num-
ber M on .70.
Show that the initial-value problem
d:5
T = f (: , y, t, E), a Y
dt =
§(Y, 9, t, f), x(a) = ((e), y(a) = #(f)
has a unique solution (2(t, e), y(t, f)) for t E Zo and 0 < e < r4 if e + IC(e) -
(a)I + Ir(e) - t'(a)I is sufficiently small on 0 < e < r4i where r4 is a positive
number. Also, show that Ii(t, e) - ¢(t)I + I y(t, e) - tj,(t)I - 0 uniformly on Zo as
e + IC(E) - (a)I + I#(E) - (a)I - 0.
Hint. See [LeL].
X-15. Let x E lR", y E R"', t E R, and e E R. Also, let t, e) and g(i, y, t, e)
be respectively R"-valued and R'-valued functions of (i, y", t, e) which are periodic
in t of period T. Assume that
(i) an R"-valued function fi(t) and an R"'-valued function t%'(t) are continuous,
continuously differentiable, and periodic of period T on the interval Zo = it :
-00 < t < +oo} and satisfy the system of equations d = f (:F, y", t, 0), 0 =
g"(x, y, t, 0) on Za,
(ii) f(i, y", t, e), g(i, y", t, e), and their partial derivatives with respect to (x, y') are
continuous on a region R = {(i, y, t, e) : Ix - m(t)I <- ri, I y - 7G(t)I < r2, t E
Zo, IEI < r3}, where rl, r2, and r3 are positive numbers,
(iii) there exists a real m x m matrix P(t) such that
(iiia) the entries of P(t) and P(t)-1 are real-valued, continuous, continuously
differentiable, and periodic of period T[Bat
on Zo,
(iiib) P(t)8b(m(t),1(t),t,0)P(t) = B2(t)], where Bi(t) is a real
m1 x ml matrix with eigenvaules havingnegative real parts on 10i while B2(t)
is a real (m - ml) x (m - ml) matrix with eigenvalues having positive real
parts on Zo,
(iv) the system
EXERCISES X 341
d,F=
.iF
8iq(t), (t),t,0)
- Of (fi(t), G(t), t, 0) [ w(t), +G(t), t, ON(* it), t, 0) z

0)J
does not have nontrivial periodic solutions of period T.
Show that the system
di dy"
dt - f(s,y,t,f), adt
=9'(2,y,t,e)
has a unique periodic solution (a(t, e), y(t, e)) of period T on To if 0 < Iej is suffi-
ciently small. Also, show that jT(t, e) - (t)I + lc(t, e) - iJJ(t)j 0 uniformly on To
Hint. See [FL]

CHAPTER XI
ASYMPTOTIC EXPANSIONS
In §§V-1 and V-2, we defined formal solutions of a system of analytic dif-

ferential equations. Formal solutions are not necessarily convergent. For ex-
ample, as we mentioned it in Remark V-1-4, the divergent formal power series
00
dy
f= (-1)m (m! )x'"+1 is a formal solution of x2 + y - x = 0. This equation
m=0 r_
has an actual solution f (x) = el/=J t-1e-hIldt for x > 0. Integrating by parts, we
0
N
obtain f = (-1)m(m!)xm+1 + (- 1)N+1((N + 1)!)e1/x tNe-l'tdt. Since 0 <
M=0 fox
el/=10,7 tNe-l/tdt = xN+2 - (N + 2)el/= r stN+1e-l/tdt < xN+2, we conclude that

0
N
f (X) - (_l)m(m!)xm+l < ((N + 1)!)xN+2 for x > 0. This is an example of
M=0
an asymptotic representation of an actual solution by means of a formal solution.
In this chapter, we explain the asymptotic expansions of functions in the sense of
Poincare and in the sense of the Gevrey asymptotics. In the Poincare asymptotics,
flat functions are characterized by the condition lim E f) = 0 for all positive inte-
xm
gers m, whereas in the Gevrey asymptotic, flat functions are characterized by the
condition If (x) I exp(cIxI-k) < M as x -i 0, where c, k, and M are some positive
numbers. Generally speaking, the Poincare asymptotics is too general for the study
of ordinary differential equations. A motivation of the Gevrey asymptotics is also
given by the Maillet Theorem (cf. Theorem V-1-5). In §XI-1, we summarize the
basic properties of asymptotic expansions of functions in the sense of Poincare. The
Gevrey asymptotics is explained in §§XI-2-XI-5.
For more information concerning the Poincare asymptotics, see, for example,
[Wasl]. The Gevrey asymptotics was originally introduced in [Wat] and further
developed in [Nell. To understand the materials concerning the Gevrey asymptotics
of this chapter, [Ram 1], (Ram 21, [Ram3], [Si17, Appendices], [Si18], and (Si19] are
helpful.
XI-1. Asymptotic expansions in the sense of Poincare

In this section, we explain the asymptotic expansions in the sense of Poincare.
Let x = a be a point on the extended complex x-plane. Consider a formal power
series
00
(XI.1.1) P(x) = E c,(x - a)m.
M=0
342
1. ASYMPTOTIC EXPANSIONS IN THE SENSE OF POINCAR$ 343
Let D be a sector in the x-plane with vertex at x = a and Do be a neighborhood

of x = a in D. Assume that f (x) is defined and continuous in Do.
Definition XI-1-1. The formal series (XI.1.1) is said to bean asymptotic (series)
expansion of f (x) as x -k a in V if for every non-negative integer N, there exists
a constant KN such that
N
(XI.1.2)
f(x)->2cm(x-a)m
< KN Lx - aIN+r, N = 0, 1, 2,- ..

M-0
for all x in Do.
Such an asymptotic relation is denoted by
(XI.1.3) f(x) p(x) as x a in D.
This definition of an asymptotic expansion of a function was originally given by H.

Poincare [Poi2J.
Before we explain some basic properties of asymptotic expansions, it is worth-
while to make the following remarks.
Remark XI-1-2. The vertex x = a can be x = oo. In that case, the asymptotic
on
series is in the form C nx-"'.
m=0
Remark XI-1-3. Assume that f (x, t) is a function defined and continuous in (x, t)
for x in D and t a parameter in a domain H in the t-plane. A formal power series
N
F c,,,(t)(x-a)'", where c,(t) is a function oft, is said to bean asymptotic (series)
M=0
expansion of f (x, t) as x 0 in V if for every non-negative integer N, there exists
a function KN(t), independent of x, such that
N
f(x, t) - Cm(t)(x - a)m <K1v(t)Ix-a1N+1, N=0,1,2,...
M=0
for all x in Do. If, moreover, KN(t) are independent of t, then the asymptotic
expansion is said to be uniform with respect to t.
Remark XI-1-4. If f (x) is holomorphic (i.e., analytic and single-valued) in a
neighborhood of x = a, by virtue of the Taylor's Remainder Theorem, f (x) admits
its Taylor's series expansion as its asymptotic expansion in any sector with vertex
at x = a.
Theorem XI-1-5. For a continuous function f (x), there is at most one asymptotic
expansion as x -+ a in a sector with vertex at x = a.
Proof
Assume that there are two asymptotic expansions of f (x) at x = a
00
(XI.1.4) f (x) ^_- c,,(x - a)14 as x -a in D
M=0
344 XI. ASYMPTOTIC EXPANSIONS
and
f(x)^, 7m(x-a)' as x -i a in D.
M=0
Then, for every non-negative integer N, there exist two constants KN and LN such
that
N
N+i ,
f (x) - > cm(x - a)m < KN Ix - al N = 0, 1, 2, .. .
m=0
N
.f (x) - ,1.(x - a)m <LNIx-aIN+1 N=0,1,2,...
m=0
for all x in a neighborhood Do of x = a in D. For N = 0, we have Ieo - 7ol <

(Ko + Lo)lx - al for x in Do. Let x a in D; then, we obtain co = 70
Now, assume that Ck = 7k for k = 0, 1,... , N - 1. Then, from (XI.1.5) and
(XI.1.6), it follows that ICN -7NIIx-aIN < (KN+LN)Is-alN+l for x in Do. Let
x -+ a in D. Then, cN = 7N. Thus, cm = 7m is true for every non-negative integer
m. O
Theorem XI-1-6. Assume that f (x) and g(x) are two continuous functions such
that
(XI.1.7)
00
f (x) > cm(x - a)' and g(x) ^-' E 7m(x - a)"' as x --+ a in D.
M=0 m=0
f (x) ± g(x) 00 (cm ± 7.)(x - a)m as x -+ a in D

M=0
f(x)g(x) 1 Ch7k) (x - a)m as x -+a in D.

m=0 h+k=m
Proof.
N
Fix a non-negative integer N and put f (x) _ > cm(x - a)'+Ei(x)(x-a)N+i
M=0
N
and g(x) _ > 'ym(x - a)m + E2(x)(x - a)N+1 Then, there exists two constants
m=0
KN and LN such that
(XI.1.10) IE1(x)I <_ KN, IE2(x)I <_ LN
for x in Do. Therefore,
N
(XI.1.11) If (x) ± 9(x)] - E I(Cm ± 1'.](x - a)m < (KN + LN)I x - aIN+1
M=0
for x in Do. Thus, (XI.1.8) holds. Also, it is easy to see that there exists a positive
constant k such that
f(x)9(x) - Ch7k) (x - a)m < KIx - aIN+1

(
m=0 h+k=m
for x in Do. Thus, (XI.1.9) holds.

Theorem XI-1-7. Suppose that f (x) is a function holomorphic in 0 < Ix - al < 8
and admits an asymptotic expansion (XL1.4) with V = (0 < Ix - at < b}. Then,
the asymptotic series actually converges to f (x) in 0 < Ix - at < S.
Proof.
ra
Since lim f (x) = co, x = a is a removable singularity. Thus, f (x) is holomorphic
in Ix - al < b. Therefore, the asymptotic series agrees with the Taylor's series (cf.
Remark XI-1-4 and Theorem XI-1-5). Thus, the asymptotic series converges in
Ix - aI < 8.
For a vector z E cm with entries (z1,z2,... zm) and p = (p1,p2,... Pm) with
non-negative integers p, (j = 1, 2,... , m), we define I6oI = P1 + p2 + + p,,,
z' = zr z2 ... zm , and Iz7 = max{Iz1I, IZ21, ... , Ixml}.
Theorem XI-1-8. Suppose that F(x, z) is a function with power series expansion
00
F(x, ) = F.(x)zY, which converges uniformly forx E D, I zl < bo, where Fp(x)
Ip1=o
00
is continuous in D and admits an asymptotic expansion Fp(x) c 1: Fpk(x - a)k
k=0
as x -, a E D. Define the formal power series in (x,
00
(XI.1.12) 4i(x, 2) = E 4ip(x)iV,
lpI=0
where
0
(XI.1.13) 4ip(x) _ E Fpk(x - a)k.
k=0
If f (x) is a continuous Cm-valued function with entrywise asymptotic expansions

00
f (x) ^-- > f (x - a)' __ p"(x) as x -' a in D,

1=0
then 45(x, pjx) - fo) defines a formal power series of (x - a) and
(XI.1.14) F(x, f (x) - fo) = 4'(x, p(x) - fo) as x -+ a in D.
Proof.
Choose p so small that I f (x) - fol < 6o for Ix - al < p. Fix a positive integer N
and put
N 00
F(x,f(x) - fo) = E F,(x)(f(x) - fo)p + > Fp(x)(f(x) - fo)p.

Ip1=0 Ipl=N+1
Then, by virtue of Theorem X1-1-6, we obtain

N N
/.-y o)'.
F,,(x)(f(x) - f0)p
Ip1=0 1p1=0
00
E Fp(x)(f(x) - fo)p = O(Ix - alN+1).

Ipi=N+1
Set
N(x) =,D(x,Plx) - fo) = >2 Hm(x - a).

M=0
Then,
N N
(XI.1.17) op(x)Q3(x) - f0)p = Hm(x - a)m +O((x - a)N+1)
1p1=0 m=0
and, hence,
N N
(XI.1.18) E Fp(x)(f(x) - fo)p = E Hm(x - a)m +O(Ix - aIN+1).
Ip1=0 m=o
Since this is true for all positive integers N, the theorem follows immediately. 0
Theorem XI-1-9. Suppose that f (x) is a continuous function with asymptotic ex-
pansion (XI.1.4) and co Let p(x) denote the formal power series
34 0.
(X7.1.1). Then, Pox) defines a formal power series in (x - a) and f(x) ox) as
x -+a in D.
Proof.
Consider F(z) = I . Then, the theorem follows from Theorem X1-1-8. 0

co+z
Theorem XI-1-10. Suppose that f (x) is a continuous function with asymptotic

expansion (XL 1.4) as x -' a in D. Let p(x) denote the formal power series (XI. 1.1).
Ten,
1
f(t)dt m+1cm(x-a)m+1 as x --* a in D,
J
where the path of integration is the line segment joining a and x, i.e.,
as x -a in D.
f aaZ f(t)dt f a=P(t)dt
Proof
N
Set pN(x) _ E c, (x - a)'. From (XI.1.4) and (XI.1.5), it follows that
m-0
If (t) -PN(t)Idtl 5 PN(t)l ds

J 0x_al If(C)
Js-al
KN sN+1 = KN Ix - al N+2
J0 N+2
for every non-negative integer N, where s = It - al. O

Theorem XI-1-11. Suppose that f (x) is a holomorphic function with asymptotic
expansion (X1.1.4). Let p(x) denote the formal power series (XI.1.1). Let D be a
proper subsector of D with vertex x = a. Then, I ax as x -- a in D, where
dp
= E mr,,,(x -
00 a)m-1 .
m=1
Proof.
For a non-negative integer N, put f (x) = pN(x) + EN(x). Then, there exists a
positive constant KN such that
(XI.1.19) EN(x)I 5 KNIx - alN+'
for x in a neighborhood of x = a in D (cf. Definition XI-1-1). Let the radial

boundaries of D be arg(x - a) = 6t and those of D be arg(x - a) = 0± (9_ <
0_ < 0+ < 9+). Let 0 = min{6+ - 0+, 0_ - 9_ }. Let x be a point of D. Consider
a circle I' _ {t1 It - xI = 2Ix - al sin 9 }. Then, r and its interior are contained
in D. Using Cauchy's integral formula, we obtain
df (x)
dx
= 1 r f
2rri Jr (t - x)2
__ 1 r PN (t)
2rri Jr (t x)2 -
I
+ 2rri
Ir (t -
EN (C)
x)2
dt.
(
ei4'
2Ix-al sin 0.Then, r_(rll>I= e'4'2Ix-aIsin0,0<0<27r
l 11
and (XI.1.19) implies

1 1 EN () I fEN()drl
tai r (x)2 172
2x
_ 1 1
EN(e)e"'dcb = 0 (Ix - aIN) .
7
Thus, the theorem is proved.

Theorem XI-1-12. Suppose that {fn(x)In = 1,2,3,...) is a sequence of contin-
uous functions such that they admit unifonn asymptotic expansions
00
(XI.1.20) fn(x) ^-- > a)' (n = 1, 2, 3, ...) as x --+ a in D.

M=O
Assume that {f(x)In = 1,2,3.... } converges uniformly to a function f (x) in a

subsector Do of D with vertex x = a. Then,
lira Cnm = Cm
00
f (x) ^- E cm(x - a)' as x -+ a in Do.

M=0
Proof
The assumption implies that for each non-negative integer N, there exists a
positive constant KN, independent of n, such that
N
(XI.1.23) a)'1 :5
(n = 1,2,3,... )
I
M=0
for x in a neighborhood of a in D. Furthermore, for each pair of positive integers

(j, k), there exists a positive constant bjk such that
(XI.1.24) If,(x) - fk(x)I S bjk
for x in Do, where
(XI.1.25) bjk-+0 as j, k -+ oo.
Put
N
(XI.1.26) pjN(x) =E cjm(x - a)n' (j = 1, 2, 3, ... ).
M=0
1. ASYMPTOTIC EXPANSIONS IN THE SENSE OF POINCARE 349
Then,
(XI.1.27) Ifj(x) -P)N(x)I 5 KNIx-alN+1 (j = 1,2,3,...).
Thus,
(XI.1.28) IP) N(x)-PkN(x)I 5 5)k+2KNIx-alN+l (j,k=1,2,3,...).

In particular,
(XI.1.29) Ic)o - ckol < b)k+2Kolx-al (j,k= 1,2,3,...)

for N = 0 and x in Do. Therefore, (XI.1.25) and (XI.1.29) imply that lim c)o = CO
exists.
Assume that lim c),,, = c,,, exist for m < N. Then, from (XI.1.28), it follows
1 00
that
N-1
E (C),,, - Ckm)(x - a)m + (CjN - CkN)(x - a)N
M=0
< S)k + 2KNIx - alN+1 (j,k = 1,2,3,...)
for x in Do. Thus,
IC)N - CkNI Ix - aIN < e)k + 61k + 2KNIx - alN+1 k=1 ,2 , 3 . . . .
where e)k 0 as j, k -. oo. Hence,
Ic)N -CkNl :5-

11
aIN +2KNIx-al (j,k= 1,2,3,...).
Iz
Therefore, lim c)N = cN. Consequently, lim c),, = c,,, for all m.
) --00 l-00
Furthermore, since (XI.1.27) holds independently of j, we obtain
N
f(x) - > cm(x - a)m < KNlx - alN+1
M=0
for x in Do. Thus, (XI.1.22) holds.

The following basic theorem is due to E. Borel and J. F. Ritt.
Theorem XI-1-13 ((Bor( and [Ril). For a given formal power series in (x - a)
(XI.1.30) P(x) = > c,,,(x - a)m

m=0
and a sector with vertex x = a
(XI.1.31) D={xIO<Ix -al <ry, (31<arg(x-a)<p2},

them exists a function f (x) which is continuous in D and holomorphic in the inte-
rior of D, and
00
(XI.1.32) f (x) ^_- E c,,,(x - a)' as x - a in D.

M=0
Proof.
Without loss of generality, assume that the sector D contains the ray arg(x -
a) = 0. Construct functions a,,, (x) for m = 1, 2.... such that f (x) = co +
00
E ca,,,(x)(x - a)m is continuous in D, holomorphic in the interior of D, and
m=1
satisfies (XI.1.32). Let bm and 0 be positive numbers, where bm are to be specified
later, whereas 0 is chosen to satisfy 0 < 0 < 1 and
(XI.1.33) 01 arg(x - a)I < 2 for x E D.
Consider
r 1
(XI.1.34) am(x) = 1 - exp l - 2m1,m(x - a)0 J , m = 1, 2, ... .
Then, am(x) are holomorphic in a `sector containing D. Also, from (XI.1.33) and
the fact that
I1-esl=l jetdt <Izi for z<0,
it follows that
bm
(XI . 1 . 35) lam(x)I <
- 2mymlx - aIe' m= 1 , 2 ,...,
in D. Put
(XI.1.36)
bm = I Icml-' if c,,, 0 0,
0 if Cm=0.
Then,
00 00
Ix - alm-e
Icmam(x) (x - a)mI < E 2m.ym for x E D.
M=1 m=1
Hence,
Oo
f (x) = co + E c,nam(x) (x - a)m

m=1
converges uniformly in D. Consequently, f (x) is continuous in D and holomorphic

in the interior of D.
1. ASYMPTOTIC EXPANSIONS IN THE SENSE OF POINCARE 351
To show that f (x) satisfies (XI.1.32), let N be a positive integer and observe
that
N+1 N+1
AX) -0D - E cm(x - a)m =EC exp ax - a)m
m=1 m=1
2--y-(Z - 7)01(
00
+ (x - a) N+1 i cmam(x)(x -a) m-N-1]
.
Lm=N+2
Since (XI.1.33) implies that
r = 0. 1, 2,... , N,
iexp [ 2'n7m(x - a)e ] Ix - aj-"I,
are bounded for x E D, there exists a positive constant H1 such that
c-exp _
N r bm
,,=1 <Hljx-aIN+1 for xED.
m-1
I 2mrym(x - a)e] (x - a)m
Also, (XI.I.35) implies that there exists a positive constant H2 such that
00 -aim-N-1
a)m-N-1 1 [00 Ix
E cmam(x) (x - (27)N+1Ix - aj0 m=N+2 (2'r)m-N-1
m=N+2
for X E D.
- i (2y)N+2
+2
I
{x - aI - H2
2y
Thus, there exists a positive constant KN such that

N
f(x)- c,n(x-a)m <KNIx-aIN+1 for xED.
Therefore, (XI.1.32) is satisfied. 0

Similarly to Theorem XI-1-13, we can prove the following theorem.
Theorem XI-1-14. For a given formal series p(x, t) _ 00 cm(t)(x - a)m in pour
m
ers of (x - a and a sector D given by (XI.1.31), where q. (t) (m = 0, 1, ...) are
holomorphic and are bounded in a domain S2, there exists a function f (z, t)
which is holomorphic with respect to (x, t) in D x S2 and satisfies the conditions
(XI.1.37) A x , t) c,,,(t)(x - a)m as x -+ a in D

M=0
and
00
( X,
(XI.1.38) as x -. a in V
dt m=o dt
uniformly fort in Il.
Proof.
I dcd't(t)
As are bounded in Sl (m = 0,1, ... ), choose
{ I ax 1cm(t)I] +I if c -(t) 0,
6m = t r: I d d (t) J }-1
0 L ` ifC,,,(t)-0.
Set
00
f (x, t) = co(t) + Cm(t)am(x)(x - a)m,

m=1
where a,,, (x) are given in (XI.1.34). Then, it can be shown that f (x, t) is holo-
morphic with respect to (x, t) in D x 1 and satisfies (XI.1.37) and (XI.1.38) in a
manner similar to the proof of Theorem XI-1-13. O
Examples XI-1-15. The followings are two examples of functions admitting as-
ymptotic expansions.
1. Let r(z) denote the Gamma function and Log z denote the principal branch of
the complex natural logarithm of z. If 0 = arg z satisfies IOI < 7r, a real number w
satisfies Iw I < 2 , and 10 + w l < 2 , then
rN m-1
Log[r(z)] _ (z - Logz - x +
2m(2m
M=1
BN+1 -(2N+1)
+ eN(z)
2(N+1)(2N+1)z
where IeN(z) < ( ( )I and the B,,, are the Bernoulli numbers
cos 8 + w ))2N+1I sin(2w
(see, for example, [01, p. 294]).
2. The exponential integral function Ei(z) = ettdt has the following form:
J zo0
N Z
Ei(z) = er J(k - 1)!z-k + N!100 ett-N-ldt.

k=1
From this observation, it follows that
N
IzI-N-1
e-ZEi(x) - E (k - 1)!z-kI
< {N! + (N + 1)!(7r + (N + 1)-1))
k=1
32
for I arg(-z)I < 7r. The asymptotic expansion is valid for I arg(-z)I < (see, for
example, (Wasl, p. 31[).
2. GEVREY ASYMPTOTICS 353
XI-2. Gevrey asymptotics

The Gevrey asymptotics is based on the following two definitions.
Definition XI-2-1. Let s be a non-negative number. A formal power series p =
+oo
E anx' E C[[x]) is said to be of Gevrey order s if there exist two non-negative
m=0
numbers C and A such that
(XI.2.1) [am[ < C(m!)'Am for m = 0,1,2,... .

We denote by C([x]], the set of all power series of Gevrey order s.
Definition XI-2-2. A function 0 of x is said to admit an asymptotic expansion
of Gevrey orders (s > 0) as x -+ 0 on a sector
D(r, a, b) = {x : 0 < fix[ < r, a < arg(x) < b},
where a and b are two real numbers such that a a,,,x"' E G[[x]], such that an in-
m
equality
N-1
(XI.2.2) E amxm < Kv,a,A(N!)'(Bv,a,3)8[x[N
0(x) - M=0
holds on D(p, a, 0) for every positive integer N and every (p, a, 03) satisfying
the inequalities 0 < p < r and a < a <,3 < b, when KP,Q,o and Bp,,,,o are
non-negative numbers determined by (p, a, /3).
We denote by A, (r, a, b) the set of all functions admitting asymptotic expansions
of Gevrey order s as x -' 0 on the sector D(r, a, b). We also set J(4) = p for
0 E A. (r, a, b). In §XI-3, we shall explain the basic properties of ar[[x]] A. (r, a, b),
and the map J : A. (r, a, b) -+ G[[x]],.
In the example given at the beginning of this chapter, the formal solution p =
00
E (-1)m(m!)x"+1 of X2!Ly +y-x = 0 is of Gevrey order 1 and the solution f (x) =
M=0
eL/z Jxt-1e-lltdt admits an asymptotic expansion of Gevrey order 1. Furthermore,
JJa
J(f) = p. Also, the Maillet Theorem (cf. Theorem V-1-5) states that any formal
solutions of an algebraic differential equation belong to C[[x]], for some s, where s
depends on each solution.
It is well known that if a complex-valued function 0(x) is holomorphic and
bounded on a domain 0 < [xj < r, where r is a positive number, then 0 is repre-
sented by a convergent power series in x. The Gevrey asymptotic expansions arise
354 XI. ASYMPTOTICS EXPANSIONS
in a similar but more general situation. To explain such a situation, let us consider
N sectors
St={x: 0<Ixl <r, at<arg(x)<b1} (1=1,2,...,N)

N
which satisfy the condition USt = (x: 0 < IxI < r}. The set {S11S2,... SN} is
t=1
called a covering at x = 0. Also, a covering {S1, S2, ... , SN } at x = 0 is said to be
good if
(i) at < at+l (t = 1, 2, ... , N), where aN+l = al + 27r,
(ii) bt - at < x (1 = 1, 2,... , N),
( ) S t n St+1 0 (1 = 1, 2, ... , N) and Stn Sk = 0 otherwise if 1,,E k, where
SN+1 = S1.
The following theorem is the most basic fact in the Gevrey asymptotics.
Theorem XI-2-3. Assume that a covering {Si, S2, ... , S1 %r) at x = 0 is good and
that N functions 01(x), ¢2 (x), ... -ON (x) satisfy the conditions
(1) 4t(x) is holomorphic on St,
(2) dt(z) is bounded on St,
(3) it holds that
(XI.2.3) I¢e(x) - 01+1(x)I < 'Yexp [- on St n St+1,
where -y > 0, A > 0, and k > 0 are suitable numbers independent of 1.

Set
(X1.2.4)
+00
Then, there ex,sts a formal power series p = E a ..xm E C[[x)]. such that ¢t E
M=0
A. (r, at, bt) and J(41) = p for each 1.
There are various situations in which Gevrey asymptotic expansions arise. To
illustrate such a situation, let ¢(x) be a convergent power series in x with coefficients
in C. For two positive numbers r and k, set
k rr
0(t)e-('1x)}tk- 1 dt.
x J0
This integral is called an incomplete Leroy transform ofd of order k.
+oo
Theorem XI-2-4. For every ¢ = cx"j E C{x}, it holds that
m=0
R if
4,k(0) E Al/k P,-2k'2k
2. GEVREY ASYMPTOTICS 355
and
+00
J(tr,k(4)) _ r 1 1 + k Cm2'",
M=0
where k and p are any positive numbers and r is any positive number smaller than
the radius of convergence of 0.
Proof.
The following proof is suggested by B. L. J. Braaksma. For an arbitrary small
positive number e, let {Sl (e), S2(e), ... , SN(c)} be a good covering at x = 0, where
St(E) _ {x : I argx - dtI < 2k - e, 0 < jxj < r}

with real numbers dt such that -ir < d 1 < . - - < dN ir, and set
J OtW _ (t = 1, 2, ... , N).
I. 4t(x) = tr,k(0t)(xe-d`)
In particular, choose dt = 0 for some t. Then, {Sj(e),S2(e),... ,SN(e)) and
{01, 02, . . , ON} satisfy conditions (1), (2), and (3) of Theorem XI-2-3. In fact, (1)
-
and (2) are evident. To prove (3), note that

k / re'd' 0(a)e-(o/s) o
r
'da,
Qe(x) = J
zk 0
where the path of integration is the line segment connecting 0 to re{d'. Therefore,
of (x) - bt+1(x) = k
X j t
0(a)e-(ol=)'ak-1d,
where the path ryt of integration is the circular arc connecting re'd'+1 to re'd'.
Statement (3) follows, since
/
e-(°/=)` exp - t fx cos[k(arg a - arg x)J
and kj arga-argxj < 2 -ke for a E ?Y and x E St(e)f1St+1(E) Since a is arbitrary,

it follows that
61 E Allk I P, de - r,-, dt + (t = 1, 2, ... , N).

2k
Furthermore, J(01) can be computed easily (cf. Exercise XI-13).
+00
Observe that a power series p = 1: a,nxm E C[[xjj belongs to C[[xJ), if and only
m=0
+ao
if d(x) = belongs to C{x}. Therefore, we obtain an important
I'(l
sm)xm
+
corollary of Theorem XI-2-4.
Corollary XI-2-5. For any p E G[[x]], and any real number d, there exists a
function ¢(x) E A. d - 2 , d + ) such that J(O) = p.
(r,
2
This corollary corresponds to the Borel-Ritt Theorem (cf. Theorem XI-1-13) of
the Poincar6 asymptotics. Also, this corollary implies that the map J
A8 (r, d - s7r , d + 2) -+ c[[x]], is onto.
2
Theorem XI-2-3 is a corollary of the following lemma.
Lemma XI-2-6. Assume that a covering {St : e = 1, 2,... , N} at x = 0 is good
and that N functions 61(x), 62(x), ... , 6v (x) satisfy the following conditions:
(i) 6t is holomorphic on St n St+1,
(ii) I6t(x)I 7exp[-Alxl-k] on St n St+1, where -y > 0, A > 0, and k > 0 are
suitable numbers independent of e.
Define s by (XI. 2-4). Then, there exist N functions ), (x), t.b2(x), ... ,'IPN(x) and a
+oo
formal power series p = > E C[[x]], such that
m=o
(a) 4,t E A. (r, at, bt) and J(4't) = p, where St = {x : 0 < I x[ < r, at < arg(x) <
bt} (e = 1, 2, ... , N),
(b) 6t(x) _ t(x) - i/'e+l(x) onSt n St+1.
Let us prove Theorem XI-2-3 by using Lemma XI-2-6.
Proof.
Set
5t(x) = 4t(x) - 41+1(x) (e = 1, 2, ... , N).
Then, there exist N functions 4111(x), 02(z), ... , ON (x) satisfying conditions (a) and
(b) of Lemma XI-2-6. In particular, (b) implies that
Oe(x) - 4,t+1(x) = t t(x) - 4Gt+1(x) (e = 1, 2,... , N)
on St n St+1. This, in turn, implies that
01(x) - 4't(x) = 4,1+1(x) - 4Gt+1(x) (e = 1, 2,... , N)
on St n St+1. Define a function 0 by
4,(x) = 4t(x) - 4Vt(x) on St (e = 1, 2, ... , N).
Then, 0 is holomorphic and bounded for 0 < IxI < r. Therefore, 0 is represented
by a convergent power series. Since .01 = -01 + 0, Theorem XI-2-3 follows immedi-
ately. 0
We shall prove Lemma XI-2-6 in §XI-5.
Because the Gevrey asymptotics of functions containing parameters will be used
later, we state the following two definitions.
3. FLAT FUNCTIONS IN THE GEVREY ASYMPTOTICS 357
00
Definition XI-2-7. Lets be a positive number. A formal power series am(u-')em

m=0
is said to be of Gevrey order s uniformly on a domain V in the u-space if there
exist two non-negative numbers Co and C, such that
(XI.2.5) Iam(i )I < Co(m!)sCr
for u E D and rn = 0, 1, 2, ....
Set V = D(60, a0, Qo) = {e : 0 < Ie] < 5o, ao < arg e < (30} and W = D(b, a, Q).
Definition XI-2-8. Let s be a positive number. A function f (u e) is said to admit
00
an asymptotic expansion am(u")em of Gevrey order s as c 0 in V uniformly
m=0
on D if
(i) > am(i )em is of Gevrey order s uniformly on D,
m=0
(ii) for each W such that a0 < a < 0 < X30 and 0 am('tL)cm < 1)!]8L
M=0
fori ED,eEIV andN=1, 2, ....
Theorem XI-2-3, Theorem XI-2-4, Corollary XI-2-5, and Lemma XI-2-6 can be
extended in a natural way so that we can use them for functions containing param-
eters. We leave such details to the reader as an exercise.
The materials of this section are also found in [Rain 11, [Ram 21, [Si17, Appen-
dices], and [Si18).
XI-3. Flat functions in the Gevrey asymptotics

In the next section, we shall show that C[[x)), and A. (r, a, b) are differential
algebras over C and the map J : A, (r, a, b) C[[x)), is a homomorphism of
differential algebras over C. In §XI-2, it was shown that the map J is onto if
b - a < sir (cf. Corollary XI-2-5). In this section, we explain the basic results
concerning the nullspace of J. To begin with, we introduce the following definition.
Definition XI-3-1. A function f (ii, e) is said to be flat of Gevrey order s as e - 0
in a sector
V = D(r,a,b)={e: 0<je]<r, a<arg(e)<b)
uniformly on a domain D in the u'-space if f (ii, e) admits an asymptotic expansion
+00
p = E am (u")em of Gevrey orders as f 0 in V uniformly on D and the expansion
m=0
off is 0, i.e., all the coefficients a, (ii) of p are equal to zero.
The following theorem characterizes flat functions of Gevrey order s.
Theorem XI-3-2. A function f (u", c) is flat of Gevrey order s as e -- 0 in a

sector D(r, a, b) uniformly on a domain D in the ii-space if and only if for each
W = D(p, a, Q) such that 0 < p < r and a < c< < ft < b, there exist non-negative
numbers KW and Aw such that
(XI.3.1) If(u,e)I<KtyexpI-ikI for (i,e)EDxW,

where
I.
(XI.3.2) k =
s
Proof.
Suppose that for a fixed W, there exist two non-negative numbers C and A such
that
I f ( u " , e)j < C(m!)"Am jelm for (11,f) E D x W and m=0,1,2,....
Then, by virtue of Stirling's formula
r`k) 2m
(XI-3.3)
m! = 2nm mme'm [i
l
+0(01,
If(u,e)I < Com"'Ag'jem for (u, e) E D x W and m = 0, 1, 2, .. .

For a given e E W, choose a non-negative integer m so that
k
me < < (m+1)e.
(;i)
Then,
1
I f(u'
e) 5 Comm/k (me) = Coe-"'/k = Co exp I ke {e - (m + 1)e}]
Hence, I f (u", e) I < K exp[-A ej-k] for every (6,e) ED x W, where K = Coexp[s]
and A = . The converse is evident. 0
e 0
3. FLAT FUNCTIONS IN THE GEVREY ASYMPTOTICS 359
Remark XI-3-3. It is known that a function f (11, e) is identically zero on D x V

if
(i) f is holomorphic in a on V for each fixed ii in D,
(ii) f is flat of Gevrey order s as e --+ 0 in V uniformly on D,
This implies that the homomorphism J :A, (r, a, b) - C[[xBJ, is one-to-one if
b - a > sa. Remark XI-3-3 is a consequence of the following lemma.
Lemma XI-3-4 ([Wat]). Let V = D A - T , ), when p is a positive number.

2k 2k
If f (x) is holomorphic in V and If (x)1 < Cexp[-Blxl-kJ for x E V for some
positive numbers k, B, and C, then f = 0 identically in V.
Proof
Fora>k,set
B
F(x,a) = f(x) P cos(k ) x_k,2a
Then, for each fixed value of a, F is holomorphic in V and
IF(x,a)l < Cexp - B C1 - coss((k9) (k7r)) IxE-k I for x E V,
where 0 = argx. In particular,
IF(x,a)l < C for argx = ±7r, x E V,
IF(x,a)1 < Cexp IB

() - 1 r-k
( cos-E
1
for Ixj = r <pand IargxI <
Therefore, by virtue of the Phragmen-Lindelof Theorem (cf. Lemma XI-3-5), we

conclude that
IF(x,a)I < Cexp [B (COS I for I arg x1 < , x E V.

cos(U')
In particular,
If (x)I < C exp [BI i)p-k - COS ) s k] for argx = 0, x E V.
Letting a -+ k, we derive f (x) = 0 if argx = 0 and x < p. This completes the proof
of Lemma 111-3-5.
Since use was made of the Phragmdn-Lindelof Theorem, we shall state and prove
the theorem precisely (cf. [Ne2, pp. 43-44J).
Lemma XI-3-5 (Phragmen-Lindelof). Let W be a closed sector in C given by

W = {x : a < arg x < ,Q, 0 < IxI < p(arg x) },
where p(w) is a positive-valued and continuous function on the interval a < w < Q.
Assume that
(1) Q - a < A , where k is a positive number,
(2) f is continuous on W,
0 0
(3) f is holomorphic in W, where W is the interior of W,
(4) I f(x)I < Kexp[AIxI-k] for x E W, where A and K are positive numbers,
(5) If (x) I < M for x on the boundary of W except for x = 0 , where M is a
positive number.
Then,
(XI.3.4) If (x)I < M for x E W.
Proof.
a
Choose I > k so that Q - a < < 7r. Set g(x) = f (x) exp[-e(e-'sx)-r], where
P 1.
0= a Q
and e> 0. Then,
2
I9(x)I = If(x)Iexp[-ecos(t(argx - 0))Ixl-`]
< Kexp[-IxI-r(ecos(f(argx - 0)) - AIxIt'k)]
for x E W. Note that eI arg x - 01 < f (3- a)J < 2 for x E W. Therefore,
2
lim g(x) = 0 and I9(x)I 5 If(x)I for x E W. Since g is holomorphic in W, it
Izl-u
follows that Ig(x)I < M for x E W. Therefore,
If (x)I < 11f lexp(e(e-sex)-t)I for xEW.
Letting e - 0, we derive (XI.3.4).
We can also prove the following theorem without any complication.
Theorem XI-3-6. If f is holomorphic in e on D(r, a, b) for each fixed u" in V and
f is flat of Gevrey orders as e - 0 in D(r, a, b) uniformly on a domain D in the
u"-space, then
a, b)
j uniform'
f (u, a)do and of (a, e) are also flat of Cevney orders as e -' 0 in
ly on D.
Of
The proof of this theorem is left to the reader as an exercise.

For more information see [R.aml], [Ram2] and [Si17, Appendices].
XI-4. Basic properties of Gevrey asymptotic expansions

In this section, we summarize the basic properties of the Gevrey asymptotic
expansions. First we prove the following theorem.
4. BASIC PROPERTIES OF GEVREY ASYMPTOTIC EXPANSIONS 361
Theorem XI-4-1. Assume that

(i) a C-valued function F(ul,... , un, e) is holomorphic in (ul,... , un, e) in a do-
main D(rl,... ,rn) x D(r,a,b), where D(rl,... ,rn) = {(u1,... ,un) : Iu3f <
r,, j = 1, ... n},
(ii) flu l, ... , un, e) admits an asymptotic expansion p(ul, ... , tin, E) _
+oo
Ean,(u1 i ... , un)E' of Gevrey order s as a -' 0 in D(r, a, b) uniformly
mc0
on D(rl,... , r.), where the coefficients a,n(u1 i ... , un) are holomorphic in
D(ri, . . . , rn),
(iii) n C-valued functions 61 (v", e), .... On(v, e) are holomorphic in a domain D x
D(r, a, b), where D is a domain in the 6-space,
(iv) for each j, the function 03 (v, e) admits an asymptotic expansion p, (v, e) _
+oo
E b3,,n(v')e'n of Gevrey order si as a -+ 0 in D(r, a, b) uniformly on V, where
m=1
the coefficients bj,,n(v) are holomorphic in D.
Regarding the coefficients a,n(u1,... , un) as power series in ( u1 , , u.), define a
formal power series in a by
P(16, c) = p(p1(i', e).... Pi-(Of"

M=0
where the coefficients P,n(v") are holomorphic in D. Then, for any (a, 0) satisfying
the condition a < a < Q < b, there exists a positive number p(a, 3) such that
0 < p(a, (3) < r and that the function F(Oi (6, e), ... , thin (6,t), c) is holomorphic in
D x D(p(a, 0), a, Q) and admits the formal power series P(U, e) as its asymptotic
expansion of Gevrey order max(s, si , ... , sn) as e 0 in V(p(a, 03), a, 03) uniformly
on the domain D.
Remark XI-4-2. As a consequence of this theorem. we conclude that under the
same assumptions as in Theorem XI-4-1, the formal power series P(v, e) in a is of
Gevrey order max(s, s1, ... , sn) uniformly on D.
Proof of Theorem XI-4-1.
Fixing a pair (a, (3) satisfying the condition a < a < 0 < b, choose a suitable
good covering {S1, ... , SN } ate = 0 and N(n + 1) functions Ft(u1, ... , un, e),
¢1,e(vU, e), ... , QSn,t(U, e) (e = 1,... , N) such that
(1) for each e, the function Ft(ui,... ,un,e) is holomorphic in the domain
D(r1 i ... , r,,) x St and admits the formal power series p(u1,... , Unit) as
its asymptotic expansion of Gevrey order s as e -' 0 in St uniformly on
D(r1,... , rn),
(2) for each (j, e), the function 6,,e(17, e) is holomorphic in the domain V x St and
admits the formal power series pj (u", e) as its asymptotic expansion of Gevrey
order sJ as e -+ 0 in St uniformly on V.
(3) there exist two positive numbers K and A such that
a
,un,E) - Ft+1(ul,... Un,()! < Kexp - Iefk
on D(rl,... , rn) x (St n St+l ), where k = 1,

(4) for each j = 1, ... , n, it holds that
101,1 (V,e) -o,,e+1(V,e)I Kexp - kJEI

on D x (St n St+1),
1
where k, = ,
Si
(5) $t = D(p, a, E3), Ft. = F, and O,,& n) for some fo.
We can accomplish this by virtue of Corollary XI-2-5 and Theorem XI-3-2.
Observe that
Fe+1(&,e+1(v,e),...
Fl(41,t(V, E), ... , ¢n,[(ve E), E) - (v, E), ... , 41n,,+1(v, e), e) I
tl
+ JFt(¢1,1+1(Lt,E), +dn,[+1(v,e),E) - F1+1(dl.[+1(6,e), ,1+1(V,e), e)1
n
< LKexp [-I E Ik
+ Kexp [-(
IC`J'
-1 E
for (6, e) E V x (St n S[+1), where L, K, and A are suitable positive numbers.
Therefore, using Theorem XI-2-3, we complete the proof of Theorem XI-4-1. 0
As a corollary of Theorem X1-4-1, the following result is obtained without any
complication.
Theorem XI-4-3. C([x]], and A. (r, a, b) are commutative differential algebra
over C, t.e.,
(i) f + g E C((x]], (respectively A. (r, a, b)) if f and g E C((x]], (respectively
A, (r, a, b)),
(ii) f9 E C([xJ], (respectively A. (r, a, b)) if f and g E C((xi], (respectively
A. (r, a, b)),
(iii) cf E C[(x]], (respectively A. (r, a, b)) if c E C and f E C((x]J, (respectively
A. (r, a, b)),
(iv) 1 E C((xJ], (respectively A, (r, a. b)) if f E C((x)J, (respectively A. (r, a, b)),
(v) / Z f dx E C((x]], (respectively A, (r, a, b)) if f E C((xf, (respectively

A. (r, a, b))
Furthermore, the map J : A. (r, a, b) --t C((xi], is a homomorphism of differential
algebra over C.
Remark XI-4-4. Using Theorem XI-3-6, we can prove (iv) and (v) of Theorem
XI-4-3 in a way similar to the proof of Theorem X1-4-1. Also, it can be shown
1
that if f (x) E A, (r, a, b) with J(f) = >+"Qa,,,x' and ao 54 0, we obtain E
m=o
f
5. PROOF OF LEMMA XI-2-6 363
A. (p(a, 8), a, 0) and J = J{f) , where a < a < /3 < b and p(a, /3) is a
\I/
suitable positive number depending on (a,fl).
The materials in this section are also found in [Ram1], [Ram2], and [Si17, Ap-
pendices].
XI-5. Proof of Lemma XI-2-6

In order to prove Lemma XI-2-6, choose N line segments C1, C2, ... , CN so
that Ct E St n St+1 (t = 1, 2,... , N), i.e., Ct : z = teie1 (0 < t < r), where
for each t, Bt is a fixed number satisfying the condition at+1 < 8t < bt. These N
line segments divide the open punctured disk V = {x : 0 < IzI < r} into N open
sectors S1, S2, ... , SN, where
St = {x : 9t_ 1 < arg(x) < 01, 0 < txI < r}.

Set
N
Zt(x)
lr
= 1
21rt
h=1
for x E St, t = 1, 2,... , N. The functions +Gt can be continued analytically onto
St by deforming Ct_1 and Ct without moving any of their endpoints. In doing this,
we do not change other line segments Ch (h i4 t - 1, t). Thus, the function tyt is
holomorphic on St, t = 1, 2,... , N, respectively.
Now, assuming that x E St n St+1, compute t/t(x) - tPt+1(x). To do this, write
'Pt and tLt+1 in the following forms respectively:
tt't(x) _ {2 =) / ft (2 =) Ln2
E C"
h#lJ
Ot+1 {x) = + 2Ri -x

27ri e", t - x c,
where the paths Ct and Ct+1 of integration are obtained by deforming C, without
moving either of its endpoints so that
(1) Ct C St n St+l and Ct+l c St n St+1,
(2) -Ct + Ct+1 is a simple closed curve whose interior contains x.
Thus, (b) follows, i.e.,
0&) - 01+1(x) = tai fe, + e«, at(x) on St nSt,.1.
Let us derive asymptotic properties of ib1. To do this, fix an t and a closed sector
W contained in St. Let Ct (respectively Ct_1) be a path obtained by deforming Ct
(respectively CL_1) without moving the endpoints so that W is contained in the
interior of the simple closed curve Ct_ 1 + 7t - C1, where 1't is a circular arc
joining two points reist-' and rest. For x E W, it holds that
ot (x) =
{-1)
- 27ri f sew
x
(-1)
+ 2?7 f bt-1(1)
x
d{ (-1)
+ 27ri f x
It may be assumed that the path Ct is given as a union of a line segment L1 and a
curve rt defined by
Lt : x = te'" (0 < t < r1 < r),

rt : x = Ftt(T) (0 <T < 1),
where we is a constant such that 01 < we < bt, and
pt(0) = r1e'", pt(i) = re'B', r1 < litt(r)I < r (0 < 7- < 1).
It can be also assumed that there exists a positive number a < 1 such that JxI < ar1
for x E W. Then, 1 fr, c - x
8t(-)
dC is represented by a convergent power series in x
21ri
whose radius of convergence is not less than r1.
Let us estimate the integral 1J 27ri

bt()
L, c - x
Since
mM+1
1
x,
x
1 1
x Q)m +
we obtain
1 M
1 6t{f) xM+lEM+1(x),
[Ymx"'
27ri L-x m=0
where
__ 1 r 5() (m > 0) and EAi+I(x) = 1f 61W <.

am
21ri JL, Cm+1 - 27ri £M+1({ - x)
Now, estimate the a,,, as follows:
= 1 I fr' ebt(Teu.")eudJ I /r'

lQm
J
+oo
27r f Tm+1 dT
7 l
7
27r f r_m-1e_ar-
dr = 2k-,r l A-Ilk! m r(\ k 1
-I- I \A-1lk\m
( \_7r
27)1-k (v') -m/k (m!)l",
where the Stirling's formula (XI.3.3) was used as m -' +co.

EXERCISES XI 365
To estimate EN+i(x), note that there exists a positive number 0 depending on

W such that
it - xJ > Jtj sin(0) for t E L1, x E W.
Hence,
1 r ^ye-ar ' .ya-(N+1)/k N+ 1

<
JEN+i(x)J
2a Jo
rN+2sin(0)dr
2kirsin(0) r( k
7/N+llity+l)/k
< k sin(O) ` )eke )
We can estimate
-11 and
-1 (h94 Q,t-1)
2ri c,_, f - x 2iri Jc - x
in a similar manner. Thus, the proof of Lemma XI-2-6 is completed.
The material of this section is also found in [Si18].
EXERCISES XI
XI-1. Let S be an open sector D(ro, a, b), i.e.,
S = {x : 0 < JxI < ro, a < argx < b},

where a and b are real numbers and ro is a positive number. Denote by
(a) A(S) the set of all functions which are holomorphic in S and admit asymptotic
expansions in powers of x as x -. 0 in every open subsector
W = {x : 0 < JxJ < r < ro, a < a < argx < Q < b}
of S,
(b) J[f] the asymptotic expansion for f E A(S),
(c) Ao(S) the set of all those functions f (x) in A(S) such that f (x) ^-- 0 as x 0
in every W (i.e. AO(S) = If E A(S) : J[f] = 0}).
Show that C[Jxll. A(S), and Ao(S) are differential algebras over the field C and
that the map J : A(S) C[[x] is a homomorphism of differential algebras over
the field C.
XI-2. Using the same notations as in Exercise XI-1 and considering a linear dif-
N
ferential operator P = E ak(x)Dk with coefficients ak(x) which are holomorphic
k-o
in a disk JxJ < r containing S, where D = d , show that
(i) P defines three homomorphisms

P1 = P : Ao(S) - Ao(S), P2 = P : A(S) -- A(S), P3 = P : C[[x]] C[[x]]
of vector spaces over the field C,

(ii) two vector spaces A(S)/Ao(S) and C[[x]] are isomorphic,
(iii) dimensions of three vector spaces Nullspace ofP1i Nullspace ofP2, and
Nullspace ofP3 over C are not greater than the order of P,
(iv) two vector spaces {Nullspace of P2}/{Nullspace of P1} and J(Nullspace of P2]
are isomorphic,
(v) there exists a homomorphism T : Nullspace of P3 .A0(S)/P1 [A0(S)] of vec-
tor spaces over C such that Nullspace of T = J[Nullspace of P2] and
T[Nullspace of P3] = {P2[A(S)] nAo(S)}/P1[Ao(S)]
Hint.
(ii) J is onto (cf. Theorem X1-1-13) and .Ao(S) is Nullspace of J.
(iii) Use Theorem IV-2-1 for Nullspace of P1 and Nullspace of P2; use an idea
similar to the proof of Theorem V-I-3 for Nullspace of P3.
(iv) Nullspace of the homomorphism J : Nullspace of P2 -+ J[Nulspace of P2] is
Nullspace of P1. Also, this homomorphism is onto.
(v) Assume that P3[¢] = 0 in C{[xJI. There exists a b E A(S) such that J[4] = 4
(cf. Theorem XI-1-13). Such b is not unique. Actually, ¢ + also satisfies
the condition J[4, + '] = 4, if and only if iP E A0(S). So define T by T[4,] =
P2[4J (mod P1[Ao(S)]). Note that J[P2[4,]] = P3[4'] = 0. This implies that
P2[O] E Ao(S).
X1-3. For each of the following three power series, find s such that f E C[[x]],.
cc
00 r(1+ 2)
(a) f(x) _ (mx)m,
(b) f (x) =m_or
E 1+ m
,n=o 3
00 2m
(c) f (x) _ Exm 11 (5+ V5-h).
m=o h=m
XI-4. For two real numbers s > 0 and A > 0 and for a formal power series
+00 +00
f = 1: a,nxm E C[[x]1, set I I f 11 &,A = E ( Denote by E(s, A) the set of
M=0
all f E C[[x]] satisfying the condition IIf II..A < +oc. Show that
7=0
(i) E(s, A) is a Banach space with the norm IIf II.,A,
(ii) IIf9II.,A <_ IIf II..A I19II..A if f E E(s, A) and g E E(s, A),
(iii) if B > A, then E(s, A) C E(s, B), and II f II,,a < IIf II,,A for f E E(s, A),
(iv) if B > A, then Il f II.,e - If (0)I < [IIf II.,A - If (0)I] for f E .6(s, A), where
f (0) = a0 if f = E.m>o a. xm, B
(v) for f E E(s, A) and an integer q such that sq > 1, it holds that
df
11 5 -LIIfII.,A
eA
EXERCISES XI 367
Comment. The formal power series f E C([x]] is of Gevrey order s if and only if
f E E(s, A) for some positive number A.
XI-5. Show that f (e) = 0 identically in a sector S if f is holomorphic and flat of
Gevrey order 0 as e --+ 0 in S.
Hint. I f (e)I < KA' IeIr' fore E Sand N = 0,1,2,..., where K and A are positive
numbers independent of N.
XI-6. Show that C[[x]]o = C{x}.
Hint. f = 00E E C[[x]]o if and only if Ic,.. I < KAt If I' (m = 0, 1, 2.... ),
=G
where K andmA are positive numbers independent of m.
X1-7. Consider a system of differential equations
(S)
xy+t dy = f0 + bg +
Ip!>_o
where q is a positive integer, fo and fp E C[(x]]n, 4> is an n x n matrix whose entries

are formal power series in x, p = (pl, ... , p,>) with n non-negative integers pt, ... ,
Pn, IpI = Pt + - - - + p,,, and yam' = yP' - - - VP-. Assume that sq > 1 and that
(i) fo and f. E E(s, A)n, and the entries of 0 belongs to E(s, A) for some A > 0,
(ii) the power series F, IIfpIIs,A9P is a convergent power series in y", where
Ipl>2
IIY1Ia,A = max{IIy,II5,A :1 < j < n},
(iii) fo(0) = 0,
(iv) (3(0) is invertible.
Show that there exists a unique power series E C[[x]]n satisfying system (S) and
the condition ¢(0) = 0. Furthermore, E (C[[x]],)n.
Hint. Write system (S) in a form y" = go + x9+1 1P + E y'gp and use the
ipl>2
dx
Banach fixed-point theorem in terms of the norm II IIs.A-
XI-8. Let D(r) = {z : IzI < r}, S(r, p) = {z : 0 < IzI < r, I arg zI < p}, and a power
+w
series f (z, e) = E fn(e)zn is convergent uniformly in a domain D(ro) x S(r, p),
n=o
where ro, r, and p are positive numbers. Assume that the coefficients fn(e) of the
series f are holomorphic in S(r, p) and admit asymptotic expansions in powers of
e as e -' 0 in S(r, p). Suppose also that limo fo(e) = 0 and limo f, (c) 0 0. Show
C- 1-
that there exist a positive number rt and a function O(e) such that (1) O(e) is
holomorphic in S(rt, p) and admits an asymptotic expansion in powers of a as
e -+ 0 in S(rt, p), (2) l ¢(e) = 0, and (3) f (c(e), e) = 0 identically in S(rt, p).
X1-9. Consider a formal power series
(P) F(x, y, _ #,
y fa,,n2,
ip,l+IpI>o
yi z1
"
where %' _ , fn:,as E C[[x]] , pi = (pll,... ,pln), and Pz =
yn xm
(p21, ... , p2",) with non-negative integers P1k and p2k. Denote by Fg(x, y z) the
Jacobian matrix of ! with respect to W. Assume that power series (P) satisfies the
conditions (1) f;(0,6,6) = 0, (2) Fy(0, 0, 0') is invertible, (3) fn, a, E E(s, A)" for
some s > 0 and A > 0, and (4) the power series i[ fn, a, Ij, Aye' is a
ia,I+?P2I?0
convergent power series in y and F. Show that there exists a unique power series
¢(x, z) 92 in (x, z) such that (i) drr,, E C[[x]]", (ii) b(0,0) = 0, and (iii)
lack>o
F(x, (x, z), z) = 0. Also, show that (a, E E(s, B)" for some B > 0 and the power
series E II( a' i" is a convergent power series in z.
IP3i>o
Hint. This is an implicit function theorem. Write f in the form f = fo + fiy +

F'gv' z"P2 fn,,r,, , where fo and fn,,n, are in E(s, A)", $ E M"(E(s, A)), and >'
is the sum over all (pi i p) such that either jell = 0 and [g-,j = I or Jial j +
jp2j > 2. Here, M.(R) denotes the set of all n x it matrices with entries in a ring
R. Assumption (1) implies that fo(0) = 0. Assumption (2) implies that -t(0) is
invertible. Therefore, I ' exists in M"(C[[x]],), and hence II4 'jI, A < +oo for
some A > 0, where III-' IIB,A = sup{jj4V' f Ij, A : f E E(s, A)n, Ij f [j,.A = 11. Since
IIfIIe,B 5 11A .,A for f E (C[[x]],)" if B > A, assume without any loss of generality
that A = A. Then, t' IF = 90 + y" + zP2§p., p,, where g"o E E(s, A)",
9a,,a, E E(s, A)", 90(0) = 0, and 'II9n,,a2I1a,A9" i is a convergent power series
ul
in y" and F. Consider the equation 0 = u"+ y"+ F,'#P- z g"n, ,n, , u" = . . Then,
un
there exists a unique power series 5(x, u", _ lip ' z'na"n, such that
fa,l+lpiI>I
dn,,a, E E(s, A)" and 0 = u'+ a'+ identically. The unique power
series satisfying conditions (i), (ii), and (iii) is given by
(S) (x, z_) = 5(x, 90, zfl

Now, consider the equation
&'$121 Is,A
(R) y'=u".+z-V Gn,,p, where Ga,,i7 = ER".

9P% ,P2 Il .w
EXERCISES XI 369
Equation (R) has a unique solution y = E 9P'z-r'pp,,p, such that Qp,,p, E

1p1I+1P21>>-1
]R" , the entries of pP,,p, are non-negative, and the series is a convergent power
series in u and z1 . Use this series as a majorant to show that defined by (S)
satisfies conditions (iv) and (v).
XI-10. Assume that a covering (S,,$2,... , SN} at x = 0 is good and that N
functions 01(x), 02(x),... , ON (x) satisfy the conditions:
(1) 0t(x) is holomorphic in St,
(2) of(x)^_-Oasx-.0 in St,
(3) 10e(x) - Ot+1(x)l < -yexp[-AIxI-kJ on St n St+1i where -y > 0, A > 0 and
k > 0 are suitable numbers independent of e.
Show that there exists a positive number H such that
[4t(x)I < Hexp[-AIxI-kJ in St.
Hint. See [Si15J.

XI-11. Assume that a covering {S1, S2, ... , SN } at x = 0 is good and that N
functions 01(x), 02(x),... , y` N (x) satisfy the conditions
(1) 4,(x) is holomorphic on St,
(2) 4c(x) is bounded on St,
(3) we have
IOt(x) - 01+1(x)l < Knlxl' (n = 1,2,...) on St n St+1,
where K are positive numbers.

Show that there exists a formal power series p = > a,,,xm E d[[xJJ such that for
m=0
each t, we obtain ¢t E A(S1) and J(¢1) = p, where the notations A(S) and J are
defined in Exercise XI-1.
XI-12. Assume that a covering {S1, S2.... , SN } at z = 0 is good and that n x n
matrices 4i1(x),t2(x),... ,4N(x) satisfy the following conditions: (1) the entries
of tt(x) belong to A(S1 n St+1) and (2) J(tt) = I,. Show that there exists a
formal power series Q = I,, + E xmQm having constants n x n matrices Qm as
m=1
coefficients, and n x n matrices P1(x), P2(x), ... , PN(x) such that
(i) for each e, the entries of P1(x) and P1(x)-1 belong to A(S1),
(ii) J(P1) = Q (t = 1, 2,... , N),
(iii) It(x) = PI(x)-1Pt+1(x)(e= 1,2,... ,N),
where the notation A(S) and J are defined in Exercise XI-1. Also, show that if
the entries of (t(x) belong to A.(St), respectively, then the entries of P,(x) and
P, 1(x) also belong to A,(Se), respectively.
Hint. See [Si17, Theorem 6.4.1 on p. 150, its proof on pp. 152-161, and §A.2.4 on
pp. 207-2081.
XI-13. Prove the formula

00
M=0
which is given in Therorem XI-2-4.
Hint. Assume that j argx1 < 2k - b, where b is a small positive number. Then,
k
k rT tme-(t/x)ktk-Idt = xm (r'/) v"'/ke °do
J0 0
m) Jr°O
r (1 + xm - xm Om/ke-`do.
T J (t/z)k
Hence,
N
1,,k(6) _ r (i + m ) Cmxm - E cmxm Om/ke °d0
k m-0 (t/k)k
M=0
/r +00
+ k J cmtm e-(t/x)ktk-ldt.
xk 0 m=N+1
XI-14. Let a be a positive number larger than 1. Also, let
SJ={x: a, <argx<bj, 0<jxj <p) (1=1,2,... v)
be a good covering of the sector S = {x : j argxj < 'r,0 < jxj < p}, where a1 =
,b= , and p is a positive number. Assume that v functions &1(x), 4(x),
satisfy the following conditions:

(i) Oj(x) is holomorphic in Sj and continuous on the closed sector
19, ={x:aj5argx<bj,O<lxj<p},
(ii) 103(x)1 < Aexp[clxl-1] in S, for some positive numbers A and c,
(iii) 10,+1(x) - Oj(x)j < Mo in S, fl S,t1 for some positive number M0,
(iv) 1O1(x)j < M0 on the line segment argx = a1, 0 < lxj < p,
(v) < M0 on the line segment argx = b,,, 0 < lxj < p.
Show that there exists a positive number M such that
IOj(x)1 < M in SS (j=1,2....,v).

Hint. This is a generalization of the Phragmen-Lindelof Theorem (Lemma XI-3-5).
See ]Lin].
EXERCISES XI 371
XI-15. Let k be a positive number. A formal power series f E C[[x]] is said to be

k-summable in a direction arg x = 0 if
\
f E I m a g {e J : A 1/k (Ap> e
IT
- 2k - E, 0+
7r
2k
+ EJ C{[x]]1/k
for some positive numbers po and E. Show that if a formal power series f E C[[x]]
is k-summable in a direction arg x = 0, there exists one and only one function
F E Al/k (po, 9 - 2k - e, 0 + 2 + EJ such that J[F] = f.
Hint. Use Remark XI-3-3. For more informations concerning summability, see, for
example, [Ba13], (Ram2), [Ram3], [Si17, Appendices], and [Si19].
r e_t-s
XI-16. Consider the integral f (x) = f - x dl;, where -1 is a line segment 0 <
x < 1 in the sectorial domain D 1, - 4 27r + 7r !. Using the argument given in

\
§XI-5, show that /
(i) f(x) admits an `àsymptotic expansion J[f] E C[[x])112 as x -. 0 in
DI 1,-427x+ 4),
(ii) J[f] is 2-summable in the direction argx = 0 if 0 < 0 < 27r.
CHAPTER XII
ASYMPTOTIC SOLUTIONS IN A PARAMETER
In this chapter, we explain asymptotic solutions of a system of differential equa-

tions E°fy = f (x, y, f) as a -i 0. In §§XII-1, XII-2, and XII-3, existence of such as-
ymptotic solutions in the sense of Poincar6 is proved in detill. In §XII-4, this result
is used to prove a block-diagonalization theorem of a linear system E° d = A(x, e)y.
In §XII-5, we explain similar results in the Gevrey asymptotics. In §XII-6, we ex-
plain how much we can simplify a linear system by means of a linear transformation
with a coefficient matrix whose entries are convergent power series in the param-
eter. This result is given in [Hs1] and similar to a theorem due to G. D. Birkhoff
[Bi] concerning singularity with respect to the independent variable x (cf. Theorem
XII-6-1 and (Si17, Chapter III]). The materials in §§XII-1- XII-5 are also found in
[Wasl], [Si3], [Si7], and [Si22].
XII-1. An existence theorem

In §§XII-1, XII-2, and XII-3, we consider a system of differential equations
dv
(XII.1.1)
E° V - = f3 (x, V 1 , V2, ... , Vn, E) (j = 1, 2, ... n),
where a is a positive integer and f J (x, V1, v2, ... , vn, f) are holomorphic with respect
to complex variables (x, v1, v2, ... , vn, E) in a domain
(XII.1.2) Ix[ < do, 0< IEI < Po, I arg EI < ao, Iv. I < 'ro (j = 1, 2, ... , n),
6o, po, ao, and 'Yo being positive constants. Set
f) () n
(XII.1.3) f j (x, V, = fJO(x, a,h(x, e)Vh + 1 fi, (x, E)iJ ,
h=1 Icl?2
where i I E Cn with the entries (V1, V2, ... , vn).

We look at (XII.1.1) under the following three assumptions.
Assumption I. Each function f j (x, v, c) (j = 1, 2,... , n) has an asymptotic ex-
pansion
(XII.1.4) fi (x, v, E) "'00E f v)E"

L=o
in the sense of Poincare as c 0 in the sector

(XII.1.5) 0 < IEI < po, I arg EI < ao,
372
1. AN EXISTENCE THEOREM 373
where coefficients f j,, (x, v) are holomorphic in the domain
(XII.1.6) W < 6o, IV1 < 7o.
Furthermore, we assume that
(XII.1.7) fjo(x,) = 0 (j = 1, 2, ... , n) for 1x1 < do.
Observation XII-1-1. Under Assumption I, fjo(x, e), a jh(x, e), and f jp(x, e) ad-
mit asymptotic expansions
00 00
(XI1.1.8) fjo(x, () E f o (x)e", fjp(x, e) ' E fJp.(x)f"

=1 v=0
and
(XII.1.9) aj h(x, e) '- > (j, h = 1, 2, ... , n)

--o
as e - 0 in sector (XII.1.5) with coefficients holomorphic in the domain
(XII.1.10) 1x(< 60-
Let A(x, e) be the n x n matrix whose (j, k)-entry is a.,&, e), respectively (i.e.,
A(x, e) = (ajk(x, e)). Then, A(x, c) admits an asymptotic expansion
(XII.1.11) A(x, e) ^_- E e"A,,(x)

=o
as a -+ 0 in sector (XII.1.5), where the entries of coefficient matrices

(ajk,,(x)) are holomorphic in domain (XII.1.10). The following second assumption
is technical and we do not lose any generality with it.
Assumption II. The matrix Ao(0) has the following S-N decomposition
Ao(0) = diag [µh,02,... ,p1 + N,
where Al, µ2f ... , An are eigenvalues of Ao(0) and .,V is a lower-triangular nilpotent
matrix.
Note that [Ni can be made as small as we wish (cf. Lemma VII-3-3).
The following third assumption plays a key roll.
Assumption III. The matrix Ao(0) is invertible, i.e.,
t<J#0 (j = 1, 2,. .. , n).
In §§XII-2 and XII-3, we shall prove the following theorem.

374 XII. ASYMPTOTIC SOLUTIONS IN A PARAMETER
Theorem XII-1-2. Under Assumptions I, II, and III, system (X11.1.1) has a
solution
(XII.1.12) vJ = pj (x, c) (j = 1,2,... ,n)

such that
(r) pi(x,e) are holomorphie in a domain
(XII.1.13) jxl < b, 0 < lei < p, l arg el < a,
where b, p, and a are suitable positive constants such that 0 < 6 < bo, 0 <
p<po, and 0<a<ao,
(ii) p, (x, e) admit asymptotic expansions
00
(XII.1.14) p3(x,e) - (j = 1,2,... n)
as e - 0 in the sector
(XII.1.15) 0 < Iej < p, urge) < a,
where coefficients of (XII.1.14) are holomorphic with respect to x in

the domain {x : lxi < 6}.
XII-2. Basic estimates

In order to prove Theorem XII-1-2, let us change system (XII-1-1) to a system
of integral equations.
Observation XII-2-1. Expansion (XII.1.14) of the solution p3(x,e)
00
(XII.2.1) vJ = Epi,(x)e" (j = 1,2,... n)
1-1
must be a formal solution of system (XII.1.1). The existence of such a formal

solution (XII.2.1) of system (XII.1.1) follows immediately from Assumptions I and
III. The proof of this fact is left to the reader as an exercise.
Observation XII-2-2. For each j = 1, 2.... , n, using Theorem XI-1-14, let us
construct a function g1(x,e) such that
(i) q, is holomorphic in a domain
(XI1.2.2) lxi < 6', 0 < jej < p', l arg el < a',
where 0<6'<6o,0<p'<po,and 0<a'<ao,

2. BASIC ESTIMATES 375
(ii) qj and !j admit asymptotic expansions

d-z
(XII . 2 . 3) gj (x ,E ) > p,v (x) "

=i
E an d d9, ,E)
_ E dp,, x) "
V=1
E
as f -. 0 in the sector
(XII.2.4) 0 < IEI < p', I argel < o .
Consider the change of variables
v, = u, + q,(x,E) (j = 1,2,...,n).
Denote (ql, q2, ... , qn) and (U I, U2,. .. , un) by q" and u, respectively. Then, ii sat-
isfies the system of differential equations
dd
(XII.2.5) E° xj = gj(x,19, c) (j = 1,2,... ,n),
where
E° dq,(x,E)
93 (x,u,E) = fj(x,U + q' f) - (j = 1, 2,. .. , n).
dx
Set
n 00
g)(x,u,E) = g2o(x,E) + bjk(x,E)uk + E bJp(x,E)u-P

(XII.2.6) k=1 IpI?2
(j = 1, 2, ... , n).
In particular,
(XII.2.7) g,o(x,E) = fi(x,9,E) -E°dq)'E) ^ 0 (j = 1,2,... n)

and
b,k(x, E) - a,k(x, c) = O(E) k = 1, 2,... , n)
as c - 0 in sector (XII.2.4). Therefore,
(XII.2.8) bik(x, E) = ajko(x) + O(e) (j, k = 1, 2,... , n)
asE-+0insector (XII.2.4).
Set
(XII.2.9) g, (x, ",E) = p, u, +R.(x,u,E) (j = 1,2,... ,n).

Then, for sufficiently small positive numbers 6, p, and -y, there exists a positive
constant c, independent of j, such that for every positive integer N, the estimates
(XII.2.10) IR7(x,u,E)I CIUI+BNIEIN (j=1,2,...,n)
and
(XII.2.11) eIu - u I (j = 1, 2,... , n)

hold whenever (x, u", f) and (x, u~', e) are in the domain
(XII.2.12) lxI <8, 0<lEH<p, largfl <a', lull <-y (j=1,2,... n).
Here, BN is a positive constant depending on N and 1191.
From (XII.2.5) and (XII.2.9), it follows that
(XII.2.13) E° uj = µ,u, + R.(x,u,E) (j = 1,2,... ,n).
Change system (XII.2.13) to the system of integral equations
I
u, = exp of( (t - x) R, (t, u, E)dt
(XII.2.14) E
where the paths of integration must be chosen carefully so that uniformly convergent
successive approximations can be defined.
Hereafter in this section, we explain how to choose paths of integration on the
right-hand side of (XII.2.14).
Observation XII-2-3. Set
(XII.2.15) w,=argµ, (j=1,2,...,n)
and suppose that
(XII.2.16) - 27r < wj < 27r (j = 1, 2,... , n).
Then, there exists a positive number a less than a such that
21
-2rr <w, - e < 7r and w, - e -2ir ,n).
Without loss of generality, suppose that n real numbers w, are divided into the
following two groups:
(I) <w,-e<2ir (j =1,2,...,m')
and
(II) -27r<w,-6<-27r (3=m'+1,...,n).

Choose two positive numbers a and /3 sufficiently small such that
-2ir+aa+3 <wl -6 < 7 r - (Qa+Q) (.7 = m'),

3 1
2"+oa+13<wj -A<-27r-(aa+p) (j=m'+1,...,n).
Set
x(1) = be-i°, x(2) = ix(1) tan [3, x(3) = -x(1), and x(4) _ -x(2),
where b is a sufficiently small positive number. Then, a rhombus is defined by its
four vertices x( j) (j = 1,2,3,4) (cf. Figure 1). Note that the angle at x(1) is 20.
FIGURE 1.
Denote the interior of this rhombus by D(5). It is noteworthy that the domain
D(b) contains a small open neighborhood of x = 0 and is contained in the domain
{x : IxI < b}.
The basic estimates for the proof of Theorem XII-1-2 are given by the following
lemma.
Lemma XII-2-4. For each j = 1.... , n, consider the function
(XII.2.17) U3(x,E) = =exp - µi(t_ dt

J z, [
L J
where the path of integration is the straight line 77! and

( x(1) (j = 1, 2,... , m'),
xj=51 x(3) (j = m' + 1, ... , m).
Then, there exists a positive number c such that
(x - xl)
(XII.2.18) I Ul (x, E)1 :5 c1cl, lexp '1l
L E° J
for
(XII.2.19) x E D(8), I argel <a, IEI > 0.
Proof
We prove this lemma for the cases j = 1, 2,... , m'. The cases j = m' + 1,... , m
can be treated in a similar manner. Set f = It - x(l)I and 0 = arg(t - x(')) for
t E D(b) and set w = arg e. Then,
(XII.2.20)
11
I U j (x, e) I s Ix- (l, l exp [-19'J' £ cos(w, + 0 - aw)] d { (j = 1, 2, ... m').
From the definition of D(8), it follows that 7r - e - f3 < 0 < 7r - e +,6 for x E D(8).
Thus, if a is in the sector I arg eI < a, the inequalities
1 3
2a<wj +7r-(aa+(3) <wj +0-aw<wj + ir - 0 + (aa + 0) < 2a
hold f o r x E D(b) and j = 1, 2, ... , m'. Therefore, there exists a positive constant
c' such that
(XII.2.21) -cos(wj +0-aw)>c (j=1,2,...,m')
for (x, e) in (XII.2.19). By virtue of (XII.2.20) and (XII.2.21), we obtain
IUj(x,E)I <
µ,( xj) (j = 1,2,... ,m')
leXp [
for (x, e) in (XII.2.19). Thus, Lemma XII-2-4 is proved. D

In the next section, we shall consider system (XII-2-14) of integral equations
assuming that x E 1)(6) and the paths of integration are chosen in the same way
as in Lemma XII-2-4.
XII-3. Proof of Theorem XII-1-2

Let us construct a solution u, _ 0, (x, e) of (XII.2.14) so that
(XII.3.1) Oj(x,e) 0 (j = 1,2,... n)
as e-'0in (XII.1.15).
Now, by virtue of Assumption II, three positive quantities 6, p, and (Nj can be
chosen so small that c in (Xl.2.10) and (XII.2.11) satisfies the condition
(XII.3.2) cc < 1,
where c is the constant given in (XII.2.18).
Define successive approximations of a solution of (XII.2.14) in the following way:
u°) (x, e) 0,
I i (Z 1
(XII.3.3) Uh(xe) _ exp [__i.(t - x)} Ri (t, u"("-1)(t, e), e)dt
(j = 1,2,... ,n; h=1,2,...)1
3. PROOF OF THEOREM XII-1-2 379
where x E D(b) and the integration is taken over the straight line . For a given
positive integer N, it will be shown that
(i) for each j, the sequence I (h) (x, E) : h = 0,1, ... } is well defined,
(ii) for the given integer N, there exist positive constants KN and PN (0 < PN
p) such that
(XII.3.4) Iu(h)(x,E)I<KNIEIN (j=1.2,...,n; h=0,1,...),

uniformly for (x, e) in the domain
(XII.3.5) x E D(b), 0 < IEl < pN, I argel < a',
(iii) the sequence {u(') (x, f) : h = 0,1, ...) converges uniformly to (x, E) _
(01(x, E), ¢2(x, E), ... , 0n (x, E)) in (XII.3.5), where iZ (x, f) is the C"-valued
function with the entries (u(h) (x, E), ... , u,(th) (x, E)).
The limit function (x, e) is independent of N since the successive approximations
are independent of N.
If (i), (ii), and (iii) are proved, it follows that
(XII.3.6) 0l(x,E) = llim u(jh)(x,E) ^-0 (j = 1,2,... ,n)
as e - 0 in the sector S = E : 0 < lei < sup(pN), I argel < a' }, and the functions
inll N
(x, c) are holomorphic the domain D(b) x S and satisfy (XII.2.14). Setting
pf(x,E) = j(x,E) +g2(x,E)(j = 1,... ,n), we obtain a solution v. = p., (x,E)(j =
1,2,... , n) of (XII.1.1) which satisfies all of the requirements of Theorem XII-1-2.
Thus, the proof of Theorem XII-1-2 will be completed.
To show (i) and (ii), choose two constants KN and pN so that KN >
PNNKN < ry. This is possible since condition (XII.3.2) is satisfied. Now, assuming i BNOC and
that (i) and (ii) are true for ugh-1)(x,E) in (XII.3.5), let us prove that
also satisfy conditions (i) and (ii). First from (XII.3.3), it follows that
..(h)i_ _' _ I _-._ I uJ I_ _ ,l

(XII.3.7)
x j:exp { (t -x.,)]Ri(tu(h-1)E)dt (J = 1,2,....n).
Then, from Lemma XII-2-4, (XII.2.10), (XII.3.4), and the inductive assumption,
we conclude that
E)I < C{CKN + BN}IEI N < KNIEIN
for (x, e) in (XII.3.5). Thus, (i) and (ii) are true for ujh) (x, E) (j = 1, 2, ... , n).
To show (iii), from (XII.2.11), (XII.3.3), and Lemma XII-2-4, we obtain

Iu-(h+l)(x,E)-u-(h)(x,E)I :5c8 sup Iiz(h)(t,E)-tL(h-1)(t,E)I
tED(6)
in (XII.3.5) for h = 1 , 2, 3, .... Hence, the sequence (u -1 h ) (x, E) : h = 0,1, 2, ... }

is convergent to (x, f) uniformly in (XII.3.5). Furthermore, uj = Oj (x, e) (j =
1, 2,... , n) satisfy system (XII.2.14). Thus, the proof of Theorem XII-1-2 is com-
pleted. 0
XII-4. A block-diagonalization theorem

Consider a system of linear differential equations
(XII.4.1) E° fy = A(x, E)y,
where a is a positive integer, y E C", and A(x, c) is an n x n matrix. The entries

of A(x, E) are holomorphic with respect to a complex variable x and a complex
parameter c in a domain
(XII.4.2) lxi < ao, 0 < IEI < po, I argEl < no,
where bo, po, and ao are positive numbers. Assume that the matrix A(x, e) admits
a uniform asymptotic expansion in the sense of Poincar6,
00
(XII.4.3) A(x, e) E E"A, (x),
V=o
in domain (XII.4.2) as c --# 0 in the sector

(XII.4.4) 0< IEI < po, I
ao,
where the entries of coefficients are holomorphic with respect to x in the
domain
(XII.4.5) Ixl < do.
Suppose that Ao(O) has a distinct eigenvalues A1, A2i ... , At with multiplicities
n1, n2i ... , nt, respectively (n1 + n2 + - - + ne = n). Without loss of general-
ity, assume that Ao(0) is in a block-diagonal form
(XII.4.6) Ao(0) = diag [Al, A21 ... , At] ,
where Aj are n, x n, matrices in the form

(XII.4.7) Aj =A,I",+N, (j=1,2,...,t).
Here, I,, is the nj x n, identity matrix and Yj is an n, x nj lower-triangular
nilpotent matrix. The main result of this section is the following theorem.
4. A BLOCK-DIAGONALIZATION THEOREM 381
Theorem XII-4-1 ([Si7]). Under assumptions (X11.4.8) and (X11.4.6), then

exists an n x n matrix P(x, e) such that
(i) the entries of P(x, e) are holomorphic with respect to (x, e) in a domain
(XII.4.8) Ix j < 6, 0< kkI < p, l argeI < a,
where 6, p, and a are positive numbers such that 0 < 5 < 60, 0 < p < po and
0<or <ao,
(ii) P(x, e) admits a uniform asymptotic expansion
00
(XII.4.9) P(x, e) etPP(x) (P0(0) = I,,)

V=0
in domain (XII.4.8) as e - 0 in the sector
(XII.4.10) 0< IkI < p, argel < a,
where the entries of coefficients P,(x) are holomorphic with respect to x in

the domain
(XII.4.11) fix, < 6,
(iii) the transformation
(XII.4.12) y" = P(x, e)zi
reduces system (X11.4.1) to a system
(XII.4.13) e° d = B(x, e)z
with the coefficient matrix B(x, e) in a block-diagonal form
(XII.4.14) B(x, e) = diag[Bj (x, e), B2(x, e), ... , B,(x, e)],
where B , (x, e) is an n, x n, matrix (j = 1, 2, ... , e),
(iv) the matrix B, (x, e) admits a uniform asymptotic expansion
(XII.4.15) Bj (x, e) ^- 00
1:
e in sector (XII.4.10), where the entries of coef-

ficients are holomorphic with respect to x in domain (X11.4.11).
Remark XII-4-2.
(a) Set
(XII.4.16) diag [B1v(x), B2 ,(x), ... , Bl,(x)] -
Then, the coefficient matrix B(x, e) of system (XII.4.13) admits a uniform
0
asymptotic expansion B(x, e)
'=a
(b) When a fundamental matrix solution Z(x, e) of (XII.4.13) is known, a funda-
mental matrix solution Y(x, e) of (XII.4.1) is given by Y(x, e) = P(x, e)Z(x, e).
(c) In the case when the matrix Ao(O) has n distinct eigenvalues, by Theorem
XII-4-1, we can diagonalize system (XII.4.1).
(d) In the case when eigenvalues of the matrix Ao(O) are not completely distinct,
the point x = 0 is, in general, a so-called transition point. In order to study
behavior of solutions in the neighborhood of a transition point, we need a
much deeper analysis of solutions of system (XII.4.1). For these informations,
see, for example, [Wash], [Was2], and [Si12].
Proof of Theorem XII-4-1.
We prove this theorem in two steps. The proof is similar to that of Theorem
VII-3-1.
Step 1. We show that there exist a positive number 6 (< oo) and an n x n matrix
Po(x) such that
(i) the entries of Po(x) and Po(x)-1 are holomorphic in the domain {x : JxI < b}
and Po(0) = In,
(ii) the matrix Co(x) = Po(x)-1Ao(x)Po(x) is in a block-diagonal form
(XIL4.17) Co(x) = diag [Col)(x),Co2)(x),... ,C(c)(x)]
where Ca) (x) is an n, x nj matrix such that
Co)(0) = Aj=a,ln,+Arj (J=1,2,...,e).
In fact, two matrices Po(x) and Co(x) must be determined by the equation
(XII.4.18) Ao(x)Po(x) - Po(x)Co(x) = 0.
Set
4(11)(x)
'4(12)(x) 4(13)(x) ... 40(2)
Ao(x)
Ao2 1) (x 4 ) (x) AA(x)
o2) ... Aou) (x)
Aat1)(x) Ao r)(x)
Aotz}(x) Ao)(x)
nd
Poll)(x) Po12)(x) Po's)(x)
Pol() (x)
Pa21)(x)
P Po2)(x) Po2'(x)
Po(x) = o )(x)
,
Poti)(x) Poti)
P ' (x) Po )(x) (x) 1
k)
where A4(jk) (x) and PO (x) are nj x nk matrices. Furthermore, set PO l) (x) _
In, (j = 1,2,... , t). Then,
(XIL4.19) Co ) (x) = Ao(jj) (x) + 1` 4h) (x) p(hj)(x) (j = 1, 2,... , f)

h#3
and
Ao(jj)(x)po
k)(x) - po k)(x)Cc(k)(x)
(XII.4.20)
+ +AD(jk)(x) =
0
(j i4 k)
h#j,k
from ( XII.4.18). Combining (XII.4.19) we obtain

41) (x)p0 k) (X) p(1k) (x)("Okk)(x) 4h)
- + (x)pOhk)(x))
hjAk
(XII.4.21) vh)(x) p(hk)(x) +
0 "(j # k).
h#j,k
Upon applying the implicit function theorem to (XII.4.21), matrices POk)(x) (j L

k) can be constructed. Then, Co(x) is given by (XII.4.19) and (XII.4.17).
Step 2. Now, assume without loss of generality that Ao(x) is in a block-diagonal
form
(XII.4.22) Ao(x) = diag [4)(x),4)(x),... ,Afl[)(x)] .

To prove Theorem XII-4-1, it suffices to solve the differential equation
dP(x, e)
(XII.4.23) Ea = A(x, E)P(x, E) - P(x, E)B(x, E),
where
A(x, E) = Ao(x) + EA(x, E), P(x, E) = In + EP(x, E) and

(XII.4.24)
B(x, E) = Ao(x) + EB(x, e).
Set
A(11)(x,E) A(12)(x,E) ... A(11)(x,E)
A(21)(,, ) A(22) (x, E) ... A(21) ('T' E)
A(x, c) _
A([I) (x, E) A(M) (x, E) ... A(t[) (x, E)
P(11)(x,E) P(12)(x,E) ... P(I[)(x'E)
P(21) (x, E) P(22) (x, E) ... P(2[) (x, E)
P([1)(x E) ... PIP (x,

p(t2)(x, E)
and
B(x, e) = diag [b(') (X, E), b(2) (X, E), . . . , B(') (x, E)] ,
where A(jk) (x) and P(Jk) (x) are nj x nk matrices and BJ (x, c) is an nj x nJ matrix
(j, k = 1,... , e). Furthermore, set
(XII.4.25) P(JJ)(x,E) = 0 (j = 1,2,...

Then, equation (XII.4.23) becomes
l
B(.)(x, f) = A(Jj)(x,E) + Ey: A(Jh)(x,E)P(hJ)(x,e)
(XII.4.26) h=1
(j = 1,2,...,t)
and
(XII.4.27)
EodP(jk)(x,E)
= A(J)(x)P(Jk)(x,E) - P(Jk)(x,E)A(k)(x) + A(jk)(x,E)
+EE A(jh)(x f)P(hk)(x,E) - P(jk)(x,f)P(k)(x,E) (i k).

1hjAk
Combining (XII.4.26) and (XII.4.27), we obtain
Eo dP(J c) (x, E)
4 J)
x P(Jk) x E - P(Jk) x, e A(k) x + A(Jk) x, E
I
+E L.:A(Jh)(x, E)P(hk)(x, E) - E p(Jk)(x, E)A(kk) (x, E)
h=1
I
- (2P(Jk) (x, c) A(kh) (x, E)P(hk) (x, E) (j0k).
h=1
Replacing P(jk)(x,E) by an nJ x nk matrix X(Jk), we derive a system of nonlinear

differential equations
E
o dX (2k) = A (J) ( x) X (jk) - X (jk) A(k)x+(X)A(Jk) x( E),
dx
I
(XII.4.28) + E E A(Jh)(x, C)X(hk) - EX (Jk)A(kk)(x, f)
h=1
I
- E2X(jk) FA(kh)(x,e)X(hk) (j0k)

h=1
Consider the entries of X (jk) (j, k = 1, 2, ... , e; j # k) altogether to form

a system of nonlinear differential equations. Upon applying Theorem XII-1-2, we
5. GEVREY ASYMPTOTIC SOLUTIONS IN A PARAMETER 385
construct an analytic solution of this nonlinear system in a domain (XII.4.8) admit-

ting an asymptotic expansion in e as described in Theorem XII-4-1. This completes
the proof of Theorem Xll-4-1. 0
XII-5. Gevrey asymptotic solutions in a parameter

In this section, using Theorems XI-2-3 and the proof of Theorem XII-1-2, we
construct Gevrey asymptotic solutions of a system of differential equations
(XII.5.1) = Ax, y, E),

Ea dx
where x is a complex variable, y E Cn, E is a complex parameter, a is a positive

integer, and Ax, if, e) is a Ca-valued function of (x, y', f). Define three domains by
A(So) = {x E C : IxI < bo}.

(XII.5.2)
S(ro, no) = {E E , arg E < ao, 0 < H < ro}.
Also, define two matrices A(x, E) and Ao(x) by
(XII.5.3) A(x, e)
89
Oyn (x, 0, f)
8y1(x,E)
and
(XII.5.4) Ao(x) = limo A(x,e),
respectively.
We first prove the following lemma.
Lemma XII-5-1. Assume that
(i) Ax, y", e) is holomorphic with respect to (x, f, c) in a domain A(bo) x Q(po) x
S(ro, no), where 6o, po, ro, and no are positive numbers,
(ii) f (x, y", c) is bounded on A(60) x i2(po) x S(ro, no),
(iii) the matrix Ao(x) defined by (XII. 5.4) exists as e - 0 in S(ro, no) uniformly
in x E A(6o) and Ao(0) is invertible,
(iv) f (x, 0, e) is flat of Gevrey order -r as e -+ 0 in S(ro, no) uniformly in x E
A(60), where r is a non-negative number.
Then, there exist three positive numbers b, r, and or such that system (XIL5.1) has
a solution O(x, e) which is holomorphic in (x, t) E A(6) x S(r, a) and that (x, c)
is flat of Gevrey order -r as e -' 0 in S(r,a) uniformly in x E L1(6).
Proof.
Note first that
(XII.5.5) I exp1- k1I = expJ-cJfJ-'cos(k(arge))]
for any positive numbers c and k. Note next that assumptions (i) and (iv) imply tha
in the case when r = 0, Ax, 0, e) is identically equal to 0 for (t, f) E A(b) x S(r, a).
Hence, in this case, 0 is a solution of (XII.5.1). In the case when r > 0, it holds
that
(XII.5.6) If(x,o,e)I < h exp[-2cIEl-k1
for some positive numbers K and c if (x, e) E A(b) x S(r, a) for sufficiently small
positive numbers 6, r, and a and if
k
(XII.5.7)
(cf. Theorem X1-3-2). Therefore, (XII.5.5) and (XII.5.6) imply that

I exp[ce-k] I If (x, O, E)I = if (x, o, e)I expf clkl-k cos(k(arg e))]
(XII.S.8)
<K exp(-c{E1,'k]
for (x, e) E A(6) x S(r, a). Note also that
(XII.5.9) cos(k(arge)) > cos(ka) > 0 for f E S(r,a) if ka < 2.
Let us change l/ in (XII.5.1) by
(XII.5.10) Ii = exp(-ce-'16.
Then, (XII.5.1) is reduced to
(XII.5.11) el" = exp[m-kJf(x,eXP[-ce-k]u,e).

Set
f (x+ FJ, e) = f (x, 0, e) + A(x, e)il + E yPfp(x, e).
JpI>2
Then,
exp[cE-k1f(x, eXp[-cE-k]u, e)
= eXPfce-kJf (x, 0, e) + A(x, e)U + explc(1 - lpl)E-k]ul fp(x, e)'
ip1>2
Using a method similar to the proof of Theorem XII-1-2, we construct a bounded

solution u = tlr(x, e) of (XII.5.11). Therefore, system (XII.5.1) has a solution of
the form y" _ ¢(x,e) = expl-ce-k]y,(x,e). This completes the proof of Lemma
XII.5.1.
The main result of this section is the following theorem, which was originally
conjectured by J: P. Ramis.
Theorem XII-5-2. Assume that

(i) f (x, y', e) is holomorphic in (x, y, e) on a domain 0(6o) x f2(po) x S(ro, ao),
where bo, po, ro, and ao are positive numbers,
(ii) f (x, y, e) admits an asymptotic expansion F(x, y, e) of Cevmy orders as a --+ 0
in S(ro, ao) uniformly in (x, y-) E A(bo) x f2(po), where s is a non-negative
number,
(iii) the matrix Ao(x) given by (XIL5.4) is invertible on 0(bo),
(iv) it holds that
(XII.5.12) lim f (x, 0, e) = 0 on 0(bo).
Then,
(1) (XIL5.1) has a unique formal solution
(X1.5.13) Plx, e) _ eep'e(x)

c-
with coefficients p"t(x) which are holomorphic on A(bo),
(2) there exist three positive numbers 6, r, and a such that (XII. 5.1) has an actual
solution ¢(x, e) which is holomorphic in (x, e) E A(5) x S(r, a) and that ¢(x, e)
admits the formal solution P(x, e) as its asymptotic expansion of Gevrey order
max { 1, s } as e - 0 in S(r, a) uniformly in x E L1(b).
a J11
Proof.
If a positive number & is sufficiently small, for every real number 6, there exists
a C"-valued function fe(x,y,e) such that
(a) fe(x, y', e) is holomorphic in (x, y", e) on a domain A(60) x f2(po) x Se(ro, &),
where
(XIL5.14) Se(ro,&) = {e: jarge - 61 < &, 0 < fej < ro},
(b) fe(x, y, e) admits an asymptotic expansion F(x, y e) of Gevrey order s as
e - 0 in Se(ro, &) uniformly in (x, y) E 0(bo) x f2(po), where s is the non-
negative number given in Theorem XII.5.2.
Such a function fe(x, y, e) exists if & is sufficiently small (cf. Corollary XI-2-5). In
particular, set
(XII.5.15) fo(x, b, e) = f (x, v, e)
Let y = & (x, e) be a solution of the system
(XII.5.16) e° = fe(x, y, e)
such that &(x, e) is holomorphic and bounded in (x, e) E i (b1) x Seri, al) for
suitable positive numbers 61, r1, and al. Using Theorem XII-1-2, it can be shown
that such a solution of (XII.5.16) exists.
Suppose that So,(ri,al) nSo,(ri,al) 0 0. Set

(XII.5.17) /i(x, E) = 4 (x, E) - 4 (x, E)
for (x, E) E A(5j) X {So,(r1,a1)nSo.(rl,a,)}. Then, obi satisfies the following linear
system
(XII.5.18) E°L = B(x, E)/+ 6(x, E),

where
B(x,E)W(x,E) = fet(x,iez(x,E),E),
(XII.5.19)
6(x,0 =
Note that
(XII.5.20) B(x, c) _ JlOfJ (x,tiet(x,E) + (1 -
and that, if s > 0, then

(XII.5.21) jb(x,E)I <pexp{-vlEl-k}
for (x,E) E 0(6) x {Se,(rl,a,)nS02(rl,a,)}, where
k
(XII.5.22)
and u and v are suitable positive numbers (cf. Theorem XI-3-2). From (XII.5.20),
it follows that
(XII.5.23) lim B(x, E) = Ao(x),
where the matrix Ao(x) is given by (XII.5.4). If s = 0, the power series f is

convergent in c (cf. Exercise XI-6). Hence, b(x, E) = 0.
If
(XII.5.24) Se(r2,a2) C Se,(ri,a'i)nSo,(ri,ai)

and if positive numbers 62, r2, and a2 are sufficiently small, applying Lemma XII-
5-1 to (XII.5.18), the functions j'(x, f) can be written in the form
(XII.5.25) i(x, E) = I (x, E)y'(E) + w(x, E),

where
1=0 if s=0,
..
(XII526) 1.-.f .)l
for (x, e) E A(S2) x Se (r2, a2), K and c are suitable positive numbers, t(x, e) is
a fundamental matrix solution of the homogeneous system
(XII.5.27) to dV = B(x, E)v,

and I(e) is a C"-valued function of a independent of x.
Assuming that Ao(0) has distinct eigenvalues A1,.12, ... , Al with multiplicities
nl, n2i ... , ne, respectively (n1 + n2 + + ne = n) and that AO(O) is in a block-
diagonal form (XII.4.6) with (XII.4.7), where N, (j = 1, 2,... , l) are n, x nj nilpo-
tent matrices, respectively, apply Theorem XII-4-1 to block-diagonalize (XII.5.27).
More precisely speaking, there exist three positive numbers 53, r3, 03, and an n x n
matrix P(x, e) which is holomorphic in (x, e) E A(53) x So (r3, a3) such that
(i) P(x, e) admits an asymptotic expansion (XII.4.9) in the sense of Poincart as
e -+ 0 in S(r3,a3) uniformly in x E A(53),
(XII.5.28) V = P(x, E)z
reduces (XII.5.27) to
di'
(XII.5.29) E°
T = E(z, E) L,
where E(x, e) is in the block-diagonal form
(XII.5.30) E(x, e) = diag [El (z, e), E2(z, c),. . . , Et(x, E)J ,
with an n1 x n, matrix E J (x, c) for each j = 1, 2, ... , t, and E(x, e) admits
an asymptotic expansion in the sense of Poincar6 as t -+ 0 in
SB(rs,a3) uniformly in x E A(63), where Eo(0) = Ao(0).
Let us construct a rhombus 7)(64) with vertices x{'), x(2), xi3i, and xt4i as it is
defined in §XII-2, in such a way that
(a) 7)(54) C a(53),
(b) it holds that
(Ai(x_r(I))\ < _i4> HIx--x""I for 3=1,...,m',
E° J IEI°
(XII.5.31)
i4 A,1 Ix - x3l for j=m'+1,...,1
(A,(z_z(3))\
E° < IEI°
on the domain V(54) x So(r4,Q4), where 64i r4, and a4 are suitable positive
numbers.
El
Write fin the form z" _ I where z"J E C"' (j = 1, 2, ... , f). Then, (XII.5.29)
ztJ
can be written in the form
(XII.5.32) E° L = (j = 1,2... ,f).

For each j, let 4ij(x, e) be a fundamental matrix solution of cv = Ej(x, e)z j.

Then, the general solution of (XII.5.29) is given by
(XII.5.33) %(x,E) _
where
xl for j = 1,... , m',
(XII.5.34)
x3 for j = m+ I,... , 1.
In this way, we can prove that t 5(x, e) - w'(x, e) is flat of Gevrey order 1 as e - O in
S e, (r1, a i) n Sez (ri , a 1) uniformly for x E A(S) if a positive number 6 is sufficiently
small (cf. [Si6; Proof of Lemma 1 on pp. 377-379]). Therefore, by using Theorem
X1-2-3, we can complete the proof of Theorem XII-5-2.
Materials of this section are also found in JSi22J.
XII-6. Analytic simplification in a parameter

For a system of linear differential equations of the form
(E) dx = xkAl ly,
where k is an integer, y E C', and A(t) is an n x n matrix whose entries are

holomorphic in a domain
(D) { t : it( < }, (t = x; Ro : positive constant),
G. D. Birkhoff proved the following theorem.

Theorem XII-6-1 ((Bil). There exists an n x n matrix P(t) such that the entries
of P(t) and P(t)-' are holomorphic in domain (D) and that the transformation
(T) y= PMg
reduces system (E) to a system of the form
(Eo)
= xkB(z}v,
where B(t) is an n x n matrix whose entries are polynomials in t. Moreover, the
matrix P(t) can be chosen so that x = 0 is, at worst, a regular singular point
of (E).
Observe that if we set t = x-3, (E) becomes
dg _t-k-2
{E1) = A(t)9.
dt
6. ANALYTIC SIMPLIFICATION IN A PARAMETER 391
If k < -2, (E1) has the coefficient holomorphic in (D). Hence, there is a funda-
mental matrix solution V(t) of (E1) whose entries are holomorphic in (D). Then,
the transformation v = V (1) i reduces (E) to the system f = 6. Therefore, the
main claim of Theorem XII-6-1 concerns the case when k > -1. In this theorem
of G. D. Birkhoff, the entries of the matrix B(t) are polynomials in t. However,
even though we can choose P(t) so that x = 0 is, at worst, a regular singular point
of (Eo), the degree of B(t) with respect to t may be very large. In order that the
degree of B(t) with respect to t is at most k + 1 so that x = 0 is a singularity
of the first kind of system (Eo), we must impose a certain condition on A(t) (cf.
Exercises XII-8, XII-9, and XII-10). For interesting discussions on this matter, see,
for example, [u2], [JLP], [Ball], and [Ba12j. A complete proof of Theorem XII-6-1
is found, for example, in [Si17, Chapter 3].
In this section, we prove a result similar to Theorem XII-6-1 for a system of
linear differential equations
(XII.6.1) f '!L = A(x,e)y,

dx
under the assumption that a is a positive integer, y" E C", and A(x, e) is an n x n
matrix whose entries are holomorphic with respect to complex variables (x, e) in a
domain
(XII.6.2) X E Do, [e[ < 60,
where Do is a domain in the x-plane containing x = 0, and 6o is a positive constant.

Let
00
(XII.6.3) A(x,e) = E ekAk(x)

k=0
be the expansion of A(x, e) in powers of e, where the entries of coefficients Ak(x) are
00
holomorphic in Do. We assume that the series >26o (Ak(x)I is convergent uniformly
k=0
in Do. The main result of this section is the following theorem, which was originally
proved in [Hsl].
Theorem XII-6-2 ([Hsl]). For each non-negative integer m, there is an n x n
matrix P(x, e) satisfying the following conditions:
(i) the entries of P(x, e) are holomorphic in (x, e) in a domain
(XII.6.4) x E D1i jeJ < 60,
where V1 is a subdomain of Do containing x = 0,

(ii) P(x, 0) = In for X E V1 and P(0, e) = In for je[ < 6o,
(iii) the system (XII.6.1) is reduced to a system of the form
m o-1
(XII.6.5) e°
dii
_{kA(X)
E + em+1 r` c'Bk(: W ) u
k=0 k--O
(XII.6.6) P(x, e)u,
where Bk(x) (k = 1, 2, ... , or - 1) are n x n matrices whose entries are holo-

morphic for x E D1.
Remark XII-6-3. In case when a = 1, (XII.6.5) becomes
dil
ed En CkAk(x)+em+IB0(x)I!
.
k=0
In particular, if a = I and m = 0, (XII.6.5) has the form
dil
e = {Ao(x) + eB0(x)} u.
It should be noticed that in Theorem XII-6-2, if we put m = 0, then the right-hand

side of (XII.6.5) is a polynomial in a of degree at most a without any restrictions
on A(x, e). We prove Theorem XII-6-2 by a direct method based on the theory of
ordinary differential equations in a Banach space. (See also [Si8j.)
Proof of Theorem XII-6-2.
The proof is given in three steps.
Step 1. Put
a m+o
(XII.6.7) P(x, e) = I + Em+1 [-` ekPk(x) and B(x, e) = E EkBk(x),
k=0 k--0
where
Bk(x) _ J A&. (x) (k = 0,1, ... , m),

(XII.6.8)
Bk-m-1(x) (k = m+ 1,m+2,... ,m+a).
From (XII.6.1), (XII.6.5), and (XII.6.6), it follows that the matrices P(x,e) and
B(x, e) must satisfy the equation
QdP
(XII.6.9) e = A(.x, e)P - PB(x, e).
From (XII.6.3), (XII.6.7), (XII.6.8), and (XII.6.9), we obtain

k
Am+1+k(x) - Bm+1+k(x)+ E{Ak-h(x)Ph(x) - Ph(x)Bk-h(x)}
(XII.6.10) h=O
(k=0,1,...,a-1)
6. ANALYTIC SIMPLIFICATION IN A PARAMETER 393
and
o+k
dPk(x)
=A m +1+ o +k(x) + Ad+k-h(x)Ph(x)
dx
h=o
(XII.6.11) o+k
E Ph(x)Bo+k-h(x) (k = 0,1,2,...),
h=k-m
where
(XIL6.12) Ph(x) = 0 if h < 0.
It should be noted that the formal power series P and B that satisfy the equation
(XII.6.9) are not convergent in general. In order to construct P as a convergent
power series in e, we must choose a suitable B. To do this, first solve equation
(XII.6.10) for B,,,+1+k(x) to derive
Bm+1+k(x) = Am+I+k(x)+Hm+1+k(x;P0,P1,... Pk)

(XII.6.13)
(k=0,1, ..,Q-1),
where H, are defined by
(XII.6.14)
H , = 0, (k = 0,1, ... , m),
k k
Hm+I+k(x; Po, PI, ... , Pk) = E{Ak-h(x)Ph - PhAk-h(x)) - E PhHk-h,
h=0 h=0
(k = 0,1,...,a - 1).
Denote by P an infinite-dimensional vector {Pk : k = 0, 1, 2,... }. Then, by substi-
tuting (XII.6.13) into (XII.6.11), we obtain
dPk(x)
(XII.6.15) = fk(x; P) (k = 0,1, 2, ... ),
dx
where
o+k o+k
fk(x; P) = Am+1+o+k(x) + E Ao+k-h(x)Ph - 1: PhA,,+k-h(x)
h=0 h=k-m
(XII.6.16)
a+k
- E PhHo+k-h(x;P) (k=0,1,2,...).
h=k-m
Denote by -F(x; P) the infinite-dimensional vector { fk(x; P) : k = 0,1, 2,... }.

Then, equation (XII.6.15) can be written in the form
dP = F(x; P).
(XII.6.17)
dx
Solve this differential equation in a suitable Banach space with the initial condition
P(0) = 0. Actually, we solve the integral equation
(XII.6.18) P(x) = f0F(;P())de.

This is equivalent to the system
(XII.6.19) Pk(x) = ffk(;P())de (k = 0,1, 2, ... ).
If P is determined, then the matrices Bk are determined by (XII.6.13) and (XI1.6.14).

Step 2. We still assume that A(x, c) is holomorphic in domain (XII.6.2). Denote
by B the set of all infinite-dimensional vectors P = {Pk : k = 0, 1, 2.... } such that
(i) Pk are n x n matrices of complex entries,
00
(ii) j:boilPkll < oo, where IlPkll is the sum of the absolute values of entries
k=0
of Pk.
For each P, define a norm 11P11 by
00
(XII.6.20) IIPII = Ebo IlPkll
k=0
Then, we can regard B as a Banach space over the field of complex numbers.
Now, we can establish the following lemma.
Lemma XII-6-4. Let F(x; P) be the infinite-dimensional vector whose entries
fk(x,P) are given by (XII.6.16). Then, for each positive numbers R, there exist
two positive numbers G(R) and K(R) such that
(XII.6.21) flF(x; P) II 5 G(R) for IIPII <- R
and
(XII.6.22) II.(x; P) - F(x; P)ll 5 K(R)IIP - PII for IIPII 5 R, IIPI{ 5 R.
In order to prove this lemma, consider a formal power series of e which is defined
by
(XII.6.23) .r(x,P,E) _ Ekfk(x;P)-

k
Then, using (XII.6.16), we obtain
(XII.6.24)
00
.F(x, P, e) _ >2 EkA,,,+1+o+k(x) + 1
\ k=0 EkAk(x)/
(e'Pk)
k=0 l k=0
a-1 k
C* M+V r
Ik(x; PA)
- Ek=OEk hE Ak-h Ph k=o Ekpk/ k o Ek [Ak(x) +
0-1 k
- k=0
> ek E Ph[Ak-h(x) + Hk-h(x;P)1
h=O
l
}.
Hence, Lemma XII-6-4 follows immediately.

EXERCISES XII 395
Step 3. We construct the matrix P(x, e), solving the integral equation (XII.6.18)
by the method of successive approximations similar to that given in Chapter I. By
virtue of Lemma XII-6-4, we can construct a solution P(x) in a subdomain Dl of
Do containing x = 0 in its interior. Since (XII.6.18) is equivalent to differential
equation (XII.6.15) with the initial condition P(0) = 0, the solution P(x) gives
the desired P(x, e). The matrix B(x, c) is given by (XII.6.13) and (XII.6.14). 0
EXERCISES XII
00
XII-1. Find a formal power series solution y = F, e'd,,,(x) of the system of

m=0
differential equations
CO
LY = Ay" + > Embm(x),

m=0
where y E C", A is an invertible constant n x n matrix, and 5,n (x) and bm(x) are
C"-valued functions whose entries are holomorphic in x in a neighborhood of x = 0.
XII-2. Using Theorem MI-4-1, diagonalize the system
dy = 0 1+x1 where
e dx 11-X ex y' [1.
XII-3. Using Theorem XII-4.1, find two linearly independent formal solutions of
each of the following two differential equations which do not involve any fractional
powers of e.
E2d2
Y
(1) 2 + y = eq(x)y, (2) a2 + y = eq(x)y,
where q(x) is holomorphic in x for small jxj.
Hint. If we set '62 = e, differential equation (2) has two linearly independent
solutions e4=/00(x, 0) and a-1/190(x, -/3). The two solutions
= Q reiz/°4(x, Q) - -Q),
do not involve any fractional powers of e.

XII-4. Let x be a complex independent variable, y" E C", z" E C, e be a complex
parameter, A(x, y, z, e) be an n x n matrix whose entries are holomorphic with
respect to (x, y, z, e) in a domain Do = {(x, y, X, e) : jxj < ro, jyi < Pi, Izl < p2, 0 <
jej < ao, j arg ej < flo}, f (x, y, z, e) be a C-valued function whose entries are
holomorphic with respect to (x, y, 1, e) in Do, and #(x, ,F, e) be a C"-valued function
whose entries are holomorphic with respect to (x, z, e) in the domain Uo = ((x, zl, e) :
jxj < ro, jzj < p2,0 < jej < ao, j argej < /3o}. Assume that the entries of the matrix
A(x, if, z, c), the functions f (x, y, z, e), and §(x, .F, e) admit asymptotic expansions as
f - 0 in the sector So = {E : 0 < IEI < ao, I arg EI < /30} uniformly in (x, y, z) in the
domain Ao = {(x, y, z) : IxI < ro, Iv'I < pl, I1 < p2} or (x, z-) in Vo = {(x, z) : I xI <
ro, Izl < p2}. The coefficients of those expansions are holomorphic with respect to
(x, y", I) in Do or (x, z-) in V0. Assume also that det A(0, 0, 0, 0) 0 0. Show that the
system
f dy = A(x, y, z, E)y + E9(x, Z, f), = Ef (x, g, Z, f)
has one and only one solution (y, z) = (¢(x, f), e)) satisfying the following
conditions:
(a) the entries of functions O(x, E) and '(x, f) are holomorphic with respect to
(x, e) in a domain {(x, f) I xI < r, 0 < lei < a, I arg EI < /3} for some positive
:
numbers r, a, and 0 such that r < ro, a < ao and /3 < /io,
(b) the entries of functions (x, c) and i/ (x, c) admit asymptotic expansions in
powers of f as c - 0 in the sector if : 0 < lei < a, I argel < (3} uniformly
in x in the disk (x : Ixi < r), where coefficients of these expansions are
holomorphic with respect to x in the disk {x: Ixi < r},
(c) 0(0,c) = O for f E {E : 0 < IEI < a, I arg EI < ,0}.
Hint. Use a method similar to that of §§X11-2 and XII-3.
r 1
at
XII-5. Find the following limits: (1) lim /
c-0 0 f (t) sin 1 dt and (2) lim
\E
j f (t) sine I I dt, where a is a nonzero real number and f (t) is continuous
and continuously differentiable on the interval 0 < t < 1.
XII-6. Discuss the behavior of real-valued solutions of the system
Edg = rEE -1 +Ex1 y

as e -40, where y"_ [y2J
XII-7. Discuss the behavior of real-valued solutions of the following two differential
d3 y
equations as f O+: (1) E2 + 2 + xty = O, (2) E3 1 (1 x)y = O.
XII-8. Assume that
(i) the entries of a C"-valued function Ax, y", e) are holomorphic with respect to
(x, y", e) in a domain A1(Eo) X II(po) X O2(ro), where b0, po, and ro are positive
numbers and
0l(6o)={xE(C:IxI <bo}, I(po)={iEC":Iv7<po},
A2(ro) = {E E C : IEI < ro},
(ii) the matrix Ao(x) =

Of -
(x, 0, 0) is invertible on 0(bo),
(iii) f(x, 0, 0) = 0 on A(bo),
(iv) a is a positive integer.
EXERCISES XII 397
Denote by S(r, a, 0) the sector (EEC : 0 < IEI < r, I arg e - 01 < a). Show that
y
(1) the system (S) c'±dx = f (x, y, E) has a unique formal solution p(x, E) _
+00
with coefficients p"l(x), which are holomorphic in A(50),
c=1
(2) for any real number 0, there exist three positive numbers b, r, and a such that
(S) has an actual solution ¢(x, e), which is holomorphic in (x, E) E 0(b) x
S(r, a, 0), and that ¢(x, c) has the formal solution p(x, e) as its asymptotic
expansion of Gevrey order 1 as c -. 0 in S(r, a, 0) uniformly in x E 0(b)
or
Hint. This is a special case of Theorem XII-5-2.
+M
XII-9. Assume that p(x,E) _ E Emp',,,(x) is a formal solution of a system (S)
M=1
dy
ca = f (x, 17, f), where a is a positive integer, y" E C, Ax, y", e) is a C"-valued
dx
function whose entries are holomorphic in a neighborhood of (x, y, f) = (0, 0, 0),
and the entries of pm(x) (m = 0.1, ...) are holomorphic in a disk Ixi < ro, where
ro is a positive ember. Assume also that p(x, E) E {G[[E]]3}" uniformly for Ixi < ro
and that 0 < s < 1. Show that if S is an open sector in the E-plane with vertex at
e = 0 and whose opening is smaller than sir, there exists two positive numbers ri
and r2 and a solution $(x, e) of (S) such that the entries of q,(x, e) are holomorphic
in (x, E) for Ixt < r1, IEI < r2,( E S, and that (x, E) admits the formal solution
p(x, c) as the asymptotic expansion of Gevrey order s as e -+ 0 in S n {I-El < r2}
uniformly for IxI < r1.
Note. No additional conditions on the linear part of Ax, y", c) are assumed.
+00
XJI-10. Assume that p(x, c) = F E"`p""m(x) is a formal solution of a system
m--O
e = f (z, y, E), where y E C", f (x, y, e) is a C"-valued function whose entries are
holomorphic in a neighborhood of (x, y, c) = (0, 6, 0), p".. E C[(x]]" (m = 0,1, ... ),
p"o(0) = 6, and p(O, e) is convergent. Show that p(x, e) is convegent for ]x) < r and
IEI < p for some positive numbers r and p.
Hint. Regard p(ET, e) as the solution of the initial-value problem y = f (.ET, y, E),
d
00) = P(0. E.
XII-11. Let t and e be complex variables, y" E G", and the entries of a G"-
valued function f(t, y, e) be holomorphic with respect to (t, y. E) in a domain 1)o =
E) : ICI < do, -oo < tt < oo, Jyj < po, 0 < IEI < ro, I arg E) < ao). As-
00
some that f(t, y, E) E E'f,,,(t, y-) as e -+ 0 in the sector So = {E : 0 < IEI <
m=0
ro, I arg EI < ao } uniformly with respect to (t, y-) in the domain Do = {(t, y) : J 3'tI <
do, -oo < Iftt < +oc, Iyl < po}, where the coefficients fm(t,y-) are holomorphic in
Ao. Assume also that f(t, y, e) and the coefficients fm(t, y) are periodic in t of
dy
period 1. Show that the differential equation = 27riy + e[y + e f (t, y, e)] has a
dt
periodic solution U = e) of period 1 such that
(a) the entries of (t, e) are holomorphic with respect to (t, e) in a domain D =
{ (t, e) : I `3'tJ < d, -oo < Rt < +oo, 0 < JeJ < r, I arg el < a} for some (d, r, a)
such that0<d<do,0<r<r0,and0<a<ao,
00
(b) j(t, e) ^ - - E em&m(t) as a -+ 0 in the sector S = If : 0 < Jej < r, J arg el <
M=0
a}, where coefficients pm(t) are holomorphic and periodic of period 1 in the
domain D' = It : J:33tI < d, -oo < Rt < +oo}.
Hint.
Step 1. Construct the solution +li(t, c", e) to the initial-value problem
(IP) dt = 27riy+ e[y+ ef (t, y, e)], 9(0) = c.

Substep 1. First, construct a formal solution
+00
+G(t, , e) _ E em'+Gm(t, l = e2x1cd + ete2attC + O(e2).
M=0
The coefficients 1& (t, a) are holomorphic in a domain

S2o = {(t, cl : 13'tI < do, -oo < Rt < +oo, Icl < ryo),
where -yo is a positive number. To prove this, set y = e2i't(c + CZ-) to change (IP)
to
(IP') d = 6 + e[z + e-2i't f (t, e2"`t(c + ez_), e)J, z(0) = 0.

The function f (t, e2x't (c" + ez-),e) admits an asymptotic expansion in powers of e
(cf. Proof of Theorem XI-1-8) whose coefficients are polynomials in the entries of
the vector z. Coefficients ci can be calculated successively.
Substep 2. Show that iP(t, c", e) V;(t, c, e) as e -. 0 in a sector S = {e : 0 < je1 <
r, I argel < a} uniformly in a domain 1Y = {(t,c) : J)tj < d, IRtI < T, Icl < ry},
where T is an arbitrary positive number and -y > 0 depends on T. In this argument,
the Gronwall's inequality (Lemma 1-1-5) is useful.
Step 2. Solve the equation 1 (1, cE, e) = E. This equation has the form
(E) c" = eh(c, e),
where h(c", e) admits an asymptotic expansion as a -' 0 in the sector S uniformly
for Icd < y. To solve this, we can use successive approximations as follows:
4W = 0, Ch(e) = e46y_1(e),e).
Then, the approximations 64(e) converge to a limit c(e). Using Theorem XI-1-12,
we can conclude that ee) admits an asymptotic expansion in powers of e.
EXERCISES XII 399
Step 3. The particular solution ¢(t, e) = j(t, cue), e) is the periodic solution satis-
fying all of the requirements.
XII-12. Consider a system of differential equations
dy
(s) = xka(x,e)y,
dx
where x is a complex independent variable, a is a complex parameter, k is a non-

negative integer, and y is an unknown element in a Banach algebra U over the field
C of complex numbers with a unit element I . Assume that
+00
(a) a(x, e) _ E x-ma,.,(e),
m=o
where am(e) E U and these quantities are holomorphic and bounded with respect
toe in a sector S = {e : 0 < je' < bo, I argel < bl}, and the series (a) is convergent
in norm for lx[ > Ra uniformly for e in S. Also, assume that a(x, e) admits an
asymptotic expansion in powers of a uniformly for x in jxl > Ro as e -. 0 in S.
Show that if a positive integer N is sufficiently large, there exist elements p(x, e),
bo(e ), ... , bM (e) of U such that
(i) p(x, e) is holomorphic and bounded in S and large [x[,
(ii) p(x, e) admits an asymptotic expansion in powers of e uniformly for large [xi
a positive integer and the quantities bo(e), ..., bkf(e) are holomorphic
in S and admit asymptotic expansions in powers of e as a -+ 0 in S,
(iv) m>k+l,
(v) the transformation y = [I + x-(N+1)p(x, e)Ju changes (s) to
N M
du E x-mam(e) + x- (N+1)
(S')
dx
xk
E x-mb,,,(e)I U.
Comment and Hint. See [Sill; in particular Theorem 2 on p. 157]. The point
x = 0 is not necessarily a regular singular point of (s') as in Theorem XII-6-1.
Step 1. The main idea is, assuming that N > k, to solve equations of the following
forms:
(1)
(N + 1) x -(k+l)Q(x) - x -k d((x) + a (x )Q( x ) - Q( x ) a ( x ) - B(x )
+ x-(N+1) [a(x)Q(x) - Q(x)B(x)] = F(x)
and
(N + 1)x-(k+l)Q(x) - x-kdQ(x) + a(x)Q( x ) - Q(x ) a(x ) - B( x )

(II)
+ x-(N+1) [a(x)Q(x) - Q(x)$(x) - 'Y(x)B(x)) = F(x),
where
(1) Q(x) is an unknown quantity which should be a convergent power series
in x-1,
(2) B(x) is an unknown quantity which should be a polynomial in x-1 of
degree k,
(3) a(x) is a given polynomial in x-1 of degree N,
(4) a(x) is a given convergent power series in x-1,
(5) 0 is a given polynomial in x-1 of degree k,
(6) -r(x) is a given convergent power series in x-1,
(7) F(x) is a given convergent power series in x-1.
To solve these equations, first express B in terms of Q. In fact, setting
too k N
Q(x) _ E x-'Q,,,. B(x) _ F, x-mBm,
a(x) _ E X -0m,
m=0 m=0 m=0
we obtain
M
Bm = 1:fam-hQh - Qham-hl, m = 0,1,..., k.
h=1
Then, we can write (I) and (II) in the form
1
(III) QM
= (m>0),
where 9,,, is either quadratic or linear in Q (n > 0).
Step 2. If N > 0 is sufficiently large, we can solve (III) by defining a norm for a
+oo +oo
convergent power series P(x) = E x-'P,,, by 11P11 = E p'11 P,,, 11, where p is a
m0 m=0
sufficiently small positive number.
Step 3. Using Steps 1 and 2, we can construct a formal power series q(x, e) _
+00
E e'ge(x) such that the coefficients qt(x) are holomorphic and bounded in a disk
t=0
A = {x : Ix] > R > 0} and that the formal transformation y = [I+x-(N+1)4(x, e)Ju
changes (s) to (s').
Step 4. Find a function q(x, e) such that q is holomorphic and bounded in 0 x S
and that q(x, e) q(x, e) as e --+ 0 in S uniformly for x E A. Then, transformation
Y= [I + x-(N+1)q(x, e)Iv changes (s) to
(s")
k +oo
dv
dx
= xk
F x-mam(E) + x- (N+1) E x-mbm(E) + E x-mbm(0) 1 u,
M= m=0 m=k+1
where
bm(E) ^_- b,1 (E) as E 0 in S
and
x-mam(E) ^- 0
m=k+1
as 0 in S uniformly for x E A.
EXERCISES XII 401
Step 5. Using again Steps 1 and 2, find a function r(x, e) such that
(a) r(x, e) is holomorphic and bounded for jxj > R' > 0 and e E S,
(b) r(x, e) = 0 as a - 0 in S uniformly for jx[ > R',
(c) for a sufficiently large M' > 0, the transformation v = [I +x-(h1'+1)r(x,e)jw
changes (s") to
N M
dw = xk ` x-mam(f) + x -(N+1) E ? bm(e) w,
Lm=o
M-
whereb,,,(e)-0ase-0 inSform>k+1.
XII-13. Let A(t) be a 2 x 2 matrix whose entries are holomorphic in a disk It :
it[ < po} such that the matrix
(M) P ()' {kA(!)P(1) -P (-X1)1
is not triangular for any 2 x 2 matrix P(t) such that the entries of P(t) and P(t)-1
are holomorphic in the disk D = It : jtj < po}. Show that there exists such a 2 x 2
)
matrix P(x) for which matrix (M) has the form xk B Gl with a 2 x 2 matrix B(t)
wh ose entries are polynomials in t and whose degree in t is at most k + 1.
Hint. See [JLP]. This result can be generalized to the case when A(t) is a 3 x 3
matrix (cf. [Ball) and [Bal2J).
XII-14. Let P(t) be a 2 x 2 matrix such that the entries of P(t) and P(t)-1
are
holomorphic in a disk V = it : jtj < po} and that the transformation y" = P (!) X
u"
changes the system
dil [ l +x-1 0 1
dx x-3 y =
[ i12,
1-x'1Jy'
I.
to
dil u1
ds
Blx/u, u= 112
with a 2x 2 matrix B(t) whose entries are polynomials in t. Show that the degree
of the polynomial B(t) is not less than 2. Also, show that there exists P(t) such
that the degree of the polynomial B(t) is equal to 2.
Hint. To prove that there exists P(t) such that the degree of the polynomial B(t)
-17+ at
is equal to 2, apply Theorem V-5-1 to the system t d'r = 1 +tbt ] v with
.ii L t
suitable constants a, Q, y, and b. To show that the degree of the polynomial B(t)
is not less than 2, assuming that the degree of the polynomial B(t) is less than 2,
derive a contradiction from the following fact:
0
The transformation y = 0 u changes the system
dy rl+x'1 0 1
dx l x-3 xtJ',
to
dil 1-x-11
dx = [lx-I u, u = IuaJ.
XII-15. Let A(t) be a 2 x 2 triangular matrix whose entries are holomorphic in a

disk it : Its < po}. Show that there exists a 2 x 2 matrix Q(x) such that
(i) the entries of Q(x) and Q(x)-I are holomorphic for jxi > o and meromorphic
at x = oo,
(ii) the transformation y" = Q(x)t changes the system dx = xkA y to _
\ x/
xkB 1 f u" with a 2 x 2 matrix B(t) whose entries are polynomials in t and
whose degree in t is at most k + 1.
Hint. See IJLP]. This result can be generalized to the case when A(t) is a 3 x 3
matrix (cf. [Ball] and [Bal2j).
CHAPTER XIII
SINGULARITIES OF THE SECOND KIND
In this chapter, we explain the structure of asymptotic solutions of a system of

differential equations at a singular point of the second kind. In §§XIII-1, XIII-2, and
XIII-3, a basic existence theorem of asymptotic solutions in the sense of Poincar6
is proved in detail. In §XII-4, this result is used to prove a block-diagonalization
theorem of a linear system. The materials in §§XIII-I-XIII-4 are also found in
[Si7J. The main topic of §XIII-5 is the equivalence between a system of linear
differential equations and an n-th-order linear differential equation. The equivalence
is based on the existence of a cyclic vector for a linear differential operator. The
existence of cyclic vectors was originally proved in [De1J. In §XIII-6, we explain a
basic theorem concerning the structure of solutions of a linear system at a singular
point of the second kind. This theorem was proved independently in [Huk4] and
(Tull. In §XIII-7, the Newton polygon of a linear differential operator is defined.
This polygon is useful when we calculate formal solutions of an n-tb-order linear
differential equation (cf. [Stj). In §XIII-8, we explain asymptotic solutions in the
Gevrey asymptotics. To understand materials in §XIII-8, the expository paper
[Ram3J is very helpful. In §§XIII-I-XIII-4, the singularity is at x = oo, but from
§XIII-5 through §XIII-8, the singularity is at x = 0. Any singularity at x = oo is
changed to a singularity of the same kind at = 0 by the transformation x = I
XIII-1. An existence theorem

In §§XIII-1, XUI-2, and XIII-3, we consider a system of differential equations
(XIII.1.1) = x'f,(x,L'i,v2,... ,un) (.7 = 1,2,... n),
where r is a non-negative integer and ff (x, v1, t,2,... , vn) are holomorphic with
respect to complex variables (x, v1, v2, .... vn) in a domain
(XIII.1.2) JxJ>No, JargxI<ao, Jv21<6o (j=1,2,...,n),

No, oo, and 6a being positive constants. Set
n
(XIII.1.3) f2 (x, ffo(x) + E ain(x)va + f, ,(x)v +
h=1 Ip1?2
where v" E C" with the entries (vl, V2.... , vn).

We look at (XIII.1.1) under the following three assumptions.
403
404 XIII. SINGULARITIES OF THE SECOND KIND
Assumption I. Each function f j (x, V-) (j = 1, 2,... , n) admits a uniform asymp-

totic expansion
k
(XIII.1.4) fi(x,v1 =00>fjk(la
k=0
in the sense of Poincare as x - oo in the sector

(XIII.1.5) Jx j > No, J argxj < cto,
where coefficients fjk(v-) are holomorphic in the domain
(XIII.1.6) jvl < bo.
Furthermore. we assume that
(XIII.1.7) fio(d) = 0 (j = 1, 2,... , n).
Observation XIII-1-1. Under Assumption I, fo(x), ajh(x), and fl,(x) admit

asymptotic expansions
oe 00
f}0(x) f70kx-k,
fjp(x) ti 1: fjpkx-k,
k=1 k=0
a,h(x) ti00E
ajhkx-
(j, h = 1, 2,... , n)
k=0
as x oo in sector (XIII.1.5) with coefficients in C. Let A(x) be then x it matrix

whose (j, k)-entry is ajk(x), respectively (i.e., A(x) = (a,h(x)). Then, A(x) admits
an asymptotic expansion
00
(XIII.1.10) A(x) '=- E x-kAk
k=0
as x - oo in sector (XIII.1.5), where Ak = (ajhk). The following assumption is

technical and we do not lose any generality with it.
Assumption II. The matrix A0 has the following S-N decomposition:
(XIII.1.11) Ao = diag [µ1,µ2, ... 44n] + N,

where µ1,µ,2, ... , µ are eigenvalues of Aa and N is a lower-triangular nilpotent
matrix.
Note that WI can be made as small as we wish (cf. Lemma VII-3-3).
The following assumption plays a key roll.
1. AN EXISTENCE THEOREM 405
Assumption III. The matrix A0 is invertible, i.e.,
(XIII.1.12) p,54 0 (j=1,2,...,n).

Set
(XIII.1.13) w = arg x, w , = arg p3 1 ,2 , .n ) .

and denote by D, (N,'y,q) the domain defined by
(XIII.1.14)
D3 (N, y, q)
l
x: Ixi > N, jwj < ao,
(2q_)r+Y<wj+(r+1)w< (2q+)1r_v}
q is an integer, N is a sufficiently large positive constant, and -y is a sufficiently
small positive constant. For each j, there exists at least one integer q, such that
the real half-line defined by x > N is contained in the interior of V., (N, y, qj). Set
,w
n
(XIIL1.15) 7)(N,7) = n V, (N,y,qj).
=1
In §§XIII-2 and XIII-3, we shall prove the following theorem.

Theorem XIII-1-2. If N-1 and y are sufficiently small positive numbers, then
under Assumptions 1, II, and III, system (XIII. 1.1) has a solution
(XHI.1.16) vj = pj (x) U = 1,2,... , m),

such that
(i) p,(x) are holomorphic in D(N,-y),
(ii) p, (x) admit asymptotic expansions
00
(XIII.1.17) Pi (z) >2 P,kx-k (j = 1, 2, ... , m)

k=1
as x - oo in 7)(N,7), where p,k E C.

To illustrate Theorem XIII-1-2, we prove the following corollary.
Corollary XIII-1-3. Let A(x) be an n x n matrix whose entries are holomorphic
and bounded in a domain Ao = {x : lxi > Ro} and let f (x) be a C'-valued function
whose entries are holomorphic and bounded in the domain Do. Also, let A1,A2,
... An be eigenvalues of A(oo). Assume that A(oo) is invertible. Assume also
that none of the quantities Aje1ka (j = 1, 2,... , n) are real and negative for a real
number 9 and a positive integer k. Then, there exist a positive number e and a
solution y"= J(x) of the system d-.yi = xk-1A(x)y+x-1 f(x) such that the entries of
O(x) are holomorphic and admit asymptotic expansions in powers of x-1 as x -+ 00

in the sector S = {x : IxI > Ro, I argx - 01 < Zk + e}.
Proof.
The main claim of this corollary is that the asymptotic expansion of the solution
is valid in a sector I argx-8 < 2k +e whose opening is greater than k So, we look
at the sector D(N, y) of Theorem XIII-1-2. In the given case, ao = +oo, r = k -1,
and µ, = Aj (j = 1,... , n). The assumptions given in the corollary imply that
w, + k6 -A (2p + 1)7r, where wj = arg Aj (j = 1, ... , n),
for any integer p. Therefore, for each 3, there exists an integer q j such that
either - it < w, - 2qjr, + k8 < 0 or 0 < w, - 2gjir + k8 < rr.

Therefore, either
- 3r. < wj - 2q, 7r + k8 - 2 < w) - 2q,rr + kO + 2

<2
or
<w) - 2 q , i r + k 8 - 2 < wj - 2gjrr + k8 + 2 < 32

2 T.
2nk
This proves that a sector S ={ x : x > &,, I argx - 81 < + e is in D(N, y)
for a sufficiently large Ro > 0 and a sufficiently small e > 0 if we use xe-i8 as the
independent variable instead of x.
XIII-2. Basic estimates

In order to prove Theorem XIII-1-2, let us change system (XIII-1-1) to a system
of integral equations.
Observation XIII-2-1. Expansion (XIII.I.17) of the solution pj (x)
00
(XIII.2.1) uJ =
Epjkx_k
(j = 1,2,... n)
k=1
must be a formal solution of system (XIII.1.1). The existence of such a formal

solution (XIII.2.1) of system (XIII.1.1) follows immediately from Assumptions I
and III. The proof of this fact is left to the reader as an exercise.
Observation XIII-2-2. For each j = 1,2,... , n, using Theorem XI-1-14, let us

construct a function z3(x) such that
(i) z j (x) is holomorphic in a sector
(XIII.2.2) In l > No, I argxl < ao,
where No is a positive number not smaller than No,

(ii) zj (x) and
dz) admit asymptotic expansions
(XIII.2.3) z,(x) L.f,kx-k and

dz) E(-k)P,kx-k-1
k=1 k=1
as x - oo in sector (XIII.2.2), respectively.

Consider the change of variables
(XIII.2.4) v, = u, +z,(x) (j = 1,2,... ,n).

Denote (zl, z2, ... , z,a) and (u1, U2,. ,- - by i and u, respectively. Then, u' sat-
isfies the system of differential equations
du,
(XIII.2.5) = x'g,(x,U-) (.7 = 1,2,... ,n),
where
dx
Set
m
(XIII.2.6) g,(x, u") = go(x) + >2 b,k(x)uk +00E b,p(x)iI (j = 1, 2,... , n).
k=1 jp1>2
In particular,
(XIII.2.7) 92o(x) = f, (x, a) - x-' d 0 (j = 1, 2,... , n)

dxx)
and
b3k(x)-ajk(x)=O(IxI-') (j,k=1,2,...,n)
as x -+ oo in sector (XIII.2.2). Thus,
()III.2.8) b,k(x) = a,k(oc) +o(IxI-1) (j, k = 1,2,... , n)
as x -' oo in sector (XIII.2.2).

Observation XIII-2-3. Set

(XIII.2.9) g1(x,u") = u, uj +R1(x,u') (j = 1,2,... ,n).
Then, by virtue of Assumption III, (XIII.2.7), and (XIII.2.8), for sufficiently small
positive numbers No 1 and 5, there exists a positive constant c, independent of j,
such that for any positive integer h, the estimates
(XIII.2.10) lR3(x,u)I ( clui +bhlxl-(n+1)

(j = 1,2,... n)
and
(XIII.2.11) IR,(x,u) - R, (x,1U') I < clu"- u l (j = 1,2,... ,n)

hold whenever (x, u") and (x, u"') are in the domain
(XIII.2.12) lxl > No, Iargxl <oo, luil <6 (j =1,2,... ,n).
Here, by, is a positive constant depending on h. Furthermore, by virtue of Assump-
tion II, it can be assumed without loss of generality that the constant c satisfies
the condition
(XIII.2.13) c< r+1

H
,
where H is a positive constant to be specified later.

From (XIII.2.5) and (XIII.2.9), it follows that
d
(XIII.2.14) ds = X'" [µuj + R, (x, u)]
Change system (XII/I.2.14) to a system of integral equations
(XIII.2.15) u, = J
t trR, (t, u) exp Ir - 1(tr+l _ xr+1)1 dt (j = 1, 2,... , n),
J
where the paths of integration L3 start from x. The paths of integration must
be chosen carefully so that uniformly convergent successive approximations can
be defined in such a way that the limit is a solution uf(x) of (XIII.2.15) which
satisfies the conditions (i) u , (x) (j = 1, 2, ... , n) are holomorphic in D(N,'y) and
(ii) u., (x) ^ _ - 0 (j = 1, 2, ... , n) as x - oo in D(N, ry) for suitable positive numbers
N and ry.
Hereafter in this section, we explain how to choose paths of integration on the
right-hand side of (XIII.2.15).
Observation XIII-2-4. Since for each j, the domain Dj (N, y, q,) contains the
real half-line defined by x > N in its interior, their common part D(N, ry) is given
by the inequalities
(XIII.2.16) l xI > N, -t < argx < e',

where t and f are positive constants. It is noteworthy that if x E D(N, y), then x
satisfies the inequalities
(XIII.2.17)
argxI <ao, - 2r+y<wj +(r+1)argx< 2r-y
(j=1,2,...,n),
provided that w1 = arg u, are chosen suitably. Therefore,
r <w2 - (r+1)f <wj +(r+1)argx

2
(XIII.2.18)
3
<w, +(r+1)P'< 2r-y
Moreover, since D(N,1) is the common part of Dj (N, y, qj), the equalities
(XIII.2.19) -Zr+y =w3 -(r+ 1)f

and
3
(XIII.2.20) r-y=Wh+(r+1)f
hold for some j and h. This implies that t and f depend on y. Thus, the quantity
y can be chosen so that
(XIII.2.21) Wa - (r+ 1)f 36 f2r

and
(XIII.2.22) wj +(r+1)f 54± r

for all j.
Observation XIII-2-5. Define the paths of integration in each of the following
two cases.
Case 1. Consider first the case when
(XIII.2.23) (r+1)(f+f) <Tr.

In this case, the set of indices J = {j : j = 1, 2, ... , n} is divided into four groups:
(XIII.2.24) G1 = jj : 2r <W3 -(r+1)f <w, +(r+1)f' < 2r
(XIII.2.25) G2 {j: -2tr<w1-(r+1)f<wj +(r+1)t<-2r},

For every point E D(N), the lines Lj4 in D(N) (except possibly for their
starting points) are defined by
I s = fo + t expji arg( - £o)], 0 < t < I - toI for j E G1 u G2,
(XIII.2.36) s = t + texp[i(r + 1)t'], 0 < t < oo for j E G3,
s=C+texp[-i(r+1)t], 0<t<oo for jEG4.
Note that from (XIII.2.28) and (XIII.2.29), it follows that
Ir+ <w (j E G1),
2 2 2 2
J-2r+ 2 <wj 2 (jEG2)

Case 2. For the case when
(XIII.2.37) (r + 1)(t + t') > r,
the set of indices J is divided into three groups:
-2r+ry' <wj -(r+1)t <-r

(XIII.2.38.1)
-2-.+y <wj+(r+l)C<2r-ry
62= j: -2r+ry'<c3-(r+1)t<Zr
(XIII.2.38.2)
2r+-y' <w) +(r+1)t' < r-ry'},
-2r +'y' <W3 - (r + 1)t < -r - -y',

(XIII.2.38.3)
2r+
where -y' is a sufficiently small positive constant such that r - 2^y' > 0. Note that
(r + 1) (t + t') > r - y'. These inequalities imply that 2 r - 2 ry' > 0 and
-(r + 1)t < -(r + 1)t - 27' + 17r < (r + 1)t' + 27' - 2r < (r + 1)t'.
In D(N), let
1 = Nr+1 exp i{ (r + 1)t' +

ifl
12'Y'
- Zr}J ,
C2Nr+lexp i -(r+1)t-+2r}
1,Y ,
For every t E D(N), the paths of integration in t(N) are defined in the following
manner:
(a) For j E G1, from (XIII.2.37) and (XIII.2.38.1), it follows that -27r + 27' <
w, -(r+1)t-21+rand w, -(r+l)t-2ry'+r<w1+(r+1)f'<2r-'y'and,
hence,
-2r+2ry'<wj -(r+1)P-2ry'+r<w, +(r+1)e<2r-ry'.

The lines L14 in D(N) are defined by
(XIII.2.39) s = + texp(i(O({)) (0 < t < cc),
where 0({) is a real-valued and continuous function off such that
-(r+1)(- 2y'+r (r+1)t"
in D(N). This implies that - 2 + 2 < wJ + 8(t;) < 2 - 2 . Precisely speaking,

the function is defined in the following way. Note that (XIII.2.37) implies
-(r+1)t- 12 7 + 2
<(r+ 1)e-2lr<(r+1)(- ir+
2
lry`.
2
Let
6 =Nr+1expli{(r+1)f -2r}J.
Then, o is on arc (A-3). Let t;' be a point on arc (A-3) such that argt2 < argt' <
arg {o. Then, the tangent T4, of the circle j = Nr+1 at the point t:' is given
by s = £' + t < +oo), where 0(l;') is continuous in {' and
-(r + 1)t - 2 + r < (r + 1)e'. Moreover, the part of T. in D(N) is given
by
(XIII.2.40) s + t exp(iQ(t:')) (0 < t < +oo).

If a point f in D(N) is on such a tangent (XIII.2.40), set 9(e) = Note that
the tangent of the circle I{1 = Nr+' at to is parallel to the line (A-1). If a point
is in a domain between these two lines, then set O(ff) = (r + 1)Q'. For all points
other than those given above, set 9(e) = -(r + 1)t - try' + r.
(b) For j E G2, (XIII.2.37) and (XIII.2.38.2) imply that
-2r+y'<wj -(r+1)Q<w,+(r+1)f +2'?'-r<2r-Zry'

The lines LSE is defined by (XIII.2.39) with 9(4) such that
-(r+1)e<B(t) <(r+1)P'+ I r
in )(N). This implies that - 2 + 2 < wj + O(l:) < 2 - 2 The function O(C) is
defined in a way similar to the previous case.
(c) For j E C3, from (XIII.2.38.3), it follows that
I
-17r+
2 2
< wj - (r+ 1)e- 'y'+1r <
2
17r
2
- 32
2
lry'-7r< 2
7- 1>'
Let
o = N''+1 exp [i{(r+1)t'+.i'_ 32rr1l .
In the case when (r + 1)(t + e') >_ 2a - ry', the inequality
(XIII.2.41) arglo ? argf2,

holds, whereas in the case when (r + 1)(t + e') < 2a - ', it holds that
(XIII.2.42) arg o < arg t;2
In the case when (XIII.2.41) holds, the lines Lj( are defined by (XIII.2.39) with
-(r+1)e- 2y'+a (r+1)e'+ 2ry'-7r
in D(N). This implies that - 2 + 2 < wj + B({) < 2 - 2 . Since to is on arc

(A-3), O(l;) is defined as given above. In the case when (XIII.2.42) holds, the lines
Ljf are defined by (XIII.2.39) with
(r+1)e'+ try'-7r <9({) -(r+l)e- Zry'+7r
32 Y.
in D(N). This implies that - 2 + < wj + 9({) < 2 - 2 In this case, o is
not on arc (A-3), but 8(e) is still defined in a way similar to the previous cases.
For all of the cases considered above, we prove the following lemma.
Lemma XIII-2-6. There exists a positive constant H independent of h such that
the inequalities
I -
(XIII.2.43)
Isi-11 I exp[
r+ 1
a11 Ids) <Hjl:I-h l exp 1r l ]
(j = 1,2,... n)
hold in a domain D(Nh) for any positive number h, where Nh is a sufficiently large
positive constant depending only on h.
Proof
First let j E G1 U G2. Then,
exp r+l{to+texP(iarg(t-to))}JI
exp [ r+1ji - I [-
- I exp [ r + 1 to] I I exp I - r +tl exp(i arg(t - Wd
exp
L r "J 1 to] I
exp I +I1 cos(wj + g(t - to))]
for E D(N). By virtue of (XIII.2.28), (XIII.2.29), and (XIII.2.35), it holds that
27r+2Y'<w, (9EG1),
-rr2 + try' < W, + arg( - W) < -Zrr - 27 (j E G2)-
1
Therefore, - cos(w, + arg(C - co)) > sin (-') for E D(Nh) and j E G1 U G2 .
Moreover, (XIII.2.34) implies
arg( - o) - 2(r + 1)(e+ e') + 2ry1
and (XIII.2.28) and (XIII.2.29) imply (r + 1)(t + t') < n - ry'. Hence,
(XIII.2.44) I arg( - {o) - ir

- 27
Observe that idol = Nn+' implies 1812 = M2 + t2 + 2cMt, where M = Nh+1 and
c= -cos(rr-0) with 0= (cf. Figure 4).
Let b = sin ('). Then,
(XIII.2.45) 0<b<6 <1

(cf. (XIII.2.44)). Set
Y(r) = (M2 + 7-2 + 2aMr)h/2 exp(M,r) or dt,

(M2 + t2 + 26Mt)h/2
where M, = iµj I cos(Wj +arg( - co)) . Then,

r
(XIII.2.46) dr = { Mj+ h M2 + T2 IY + 1
+ 2f Mr
'r + am
and Af, < - rµ+l < 0. From (XIII.2.45), it follows that
M2 + r + 2c"Mr
i 2hµJIb21),
Mb (r > 0). If M satisfies an inequality M > then Y+
2(1
T- < r + 1)
2(r + 1)
1. Since Y = 0 for r = 0, we obtain Y(r) < (r > 0). This implies that
Ipj lb
exp 1 112r+
i
1)
x ISI-h exp I exp{i arg(e - o)}J I Idsl < 2(lu
JG,c l - r +tl Ib
Therefore, (XIII.2.43) follows. In this case, H = 2(rub 1), where y = min{p,).

In other cases (i.e., G3, G4, G1, G2, and G3), the lines L, are given by (XIII.2.39),
where 9(£) satisfies - 2 it + 27' < wj + 9(e) < 27r - in D(Nh). Therefore,
(XIII.2.47) cos(w, + 9(e)) > sin try' I =b
for E D(N),).
(i) Consider the case when 10(x) - a given point { E D(Nh). Then,
Isle =112 + t2 - cos(sr - IB(S) - argil) ? IC12.
f
Therefore,
exp exp [ -
r±1 ( +
I31-h
Idyl <
1
J,< L
r+1
p1 S 0
I dt
=141-h exp _ Iexp (wj + dt

+tl
< exp l - Nj, 1I(r1}
1 ` r+1J lµ)Ib
This implies (XIII.2.43).
(ii) Consider the case when 27r < 19(x) - a given point t E TD(Nh).
Since the lines Lg are in D(Nh), the distance D from the origin to L,t is not less
than M = Then, IS12 can be written in the form 1x12 = D2+o2 (cf. Figure
Nh+i.
5).
FIGURE 4. FIGURE 5.
3. PROOF OF THEOREM XIII-1-2 417
Consider a function Y(r) = (D2 + r2)h/2exp(Mjr)

100 xp + M, ) da, where
M2 = 1'U2I òs(+ + OW). Note that b > 0. Hence, for r > 0, (XIII.2.47) implies
that Y(r) < M . On the other hand, Y(r) satisfies the differential equation dY =
{
M+h
D2 + r2 }
Y -1. Since
r
D2+72
< 1
2D
, it follows that
dY
dr
>
lit, Ib
2(r + 1)
Y-1
h(r + 1)
if M > . Here, a use was made of the inequality M -< D. Since Y(0) <
1i2 1b
1 < (r + 1)
we obtain Y(r) <
2(r i 1) (r < 0). Therefore, (XIII.2.43) follows.
M2 1A2 ib 1µ, b
2(r + 1)
This completes the proof of Lemma XIII-2-6. In this case, again, H = µb ,
where µ =min{µ2 }. 0
2
We fix the paths of integration on the right-hand sides of (XIII.2.15) as explained

above. The constant H in (XIII.2.13) is chosen to be equal to the constant H given
in Lemma XIII-2-6. Note that H is independent of h.
XIII-3. Proof of Theorem XIII-1-2

Let us construct a solution u2 = 02 (x) of (XIII.2.15) so that
(XIII.3.1) O2 (x) 0 (j = 1,2,... n)
as x oo in (XIII.1.15).
Observation XIII-3-1. As in the previous section, denote by Do(Nh) the domain
in the x-plane which corresponds to the domain D(Nh) in the -pllane. The domain
Do(Nh) depends on h. Set o i = Nh+1 exp {i(r + 1)(e + and denote by
xoh) the corresponding point in Do(Nh ). Also, setting h' = (h = 1,2,3.... ),

r+1
consider the domain Do(Nh,). As mentioned in the previous section, the constant
H in (XIII.2.13) is chosen to be equal to the constant H in Lemma XIII-2-6.
H
Choose a positive constant 8 so that r 1 { cb + Nh2 <_ S. Furthermore, by
111
virtue of (XIII.2.13), Nh, can be chosen so large that Nh, > No and b < d (cf.
(XIII.2.12)). Fix Nh, in this way. Then, by a method similar to that in §XII-3 for
each j (j = 1, 2, ... , n), successive approximations can be defined to construct a
solution u2 = O,(x) of (XIII.2.15) which is holomorphic in Do(Nh-) and 101(x)l <
(j = 1,2,... ,n), where = xT+'. In this way, the existence of a bounded
solution u2 = 02 (x) (j = 1, 2, ... , n) of system (XIII.2.15) is proved.
Observation XIII-3-2. In order to prove that Oj (x) 0 (j = 1, 2,... , n) as

x -+ oo in Do(Nh'), use the fact that, for each positive integer k, the functions
4j (x) (j = 1, 2, ... , n) also satisfy the integral equations
.(xr+1 _ (xok))r+1)I
Oj(x(k))exp
1r+1
+ f tTRj(t, i (t)) exp [r
-JAI (tr+l
- xr+1) I dt
Ljc J
(XIII.3.2) uj(x) = (j E G1 U G2),
_xr+1)ldt
I trR1(t,, (t))exp [r+i(tr+1
JJ
UVG1UG2)
in the domain Do(Nk). Here, u(x) denotes the Cn-valued function whose entries are
(u1(x),u2(x),... ,un(x)). Note that we can assume Do(Nk) C Do(Nh') without
loss of generality.
Upon applying successive approximations similar to those of §XII-3 together
with Lemma XIII-2-6 to (XIII.3.2), it can be proved that (XIII.3.2) has a solution
u,(x) = 0., (x) (j = 1,2,... ,n) such that IiP, (x) I < CkIxI-k U = 1,2,... ,n) in
Do(Nk), where Ck is a suitable positive number. Also, using Lemma XIII-2-6 and
(XIII.2.13), it can be shown that ¢j (x) = ikj (x) (j = 1,2,... , n) in VO(Nk), since
1(tr+1 _ xr+1)l dt,

Oj (x) - j (x) = j tr [RR (t, fi(t)) - Rj(t,,(t))I exp Ir +
J
where j = 1, 2,... , n, and ¢(x) and tG(x) denote en-valued functions whose entries
are (01(x), 02(x), ... , On(x)) and (ti1(x), ,11 (x)), respectively. Thus, we
conclude that 0, (x) 0 (j = 1, 2,... , n) as x -' oo in Do(Nh,).
Now, let us return to Observation XIII-2-2. If we set pj(x) = Oj(x) + zj(x), the
solution v , = pj(x) (j = 1, 2, ... , n) of (XIII.1.1) satisfies all of the requirements of
Theorem XIII-1-2. 0
Remark XIII-3-3. Let Ax, y', µ, e) be a Cn-valued function of a variable (x, e)
E C x Cn x Cn x C such that
(a) the entries of f are bounded and holomorphic in a domain
Do = { (x, b, µ, e) : IxI > Ro, Iu-1 < do, lµl < p o, IeI < Eo, I arg EI < po},
where R0, 60, µo, co, and po are some positive constants,
(b) f admits a uniform asymptotic expansion in Do
+0
r/x,
s/
a/, e)
ti Ehfh(x, . p) as e - 0,
h=0
where coefficients fh (x, y, i) are single-valued, bounded, and holomorphic in

the domain
Ao = {(x, 9, #) : IxI > Ro, I vi < 6o, Iµl < po},
3. PROOF OF THEOREM XIII-1-2 419
(c) if f (x, y, µ', e) = fo (x, #,c) + A(x, µ, e)y + O(I y12), then
fo(oo,µ',e) = 0 and detA(oo,0,0) 0 0.

Let )q, A2, ... , )1n be n eigenvalues of A(oo, 0', 0) and set wf = arg A, (j _
1, 2, ... , n). Note that the w, are not unique. Set also S(pl, p2) _ {x : Pi <
argx < p2}. Let a and r be two non-negative integers. Assume that the sector
S(pl, p2) is contained in the sector
32
32 - min{wj} + apo < (r+ 1)argx < - max{wj} - apo
f o r a suitable choice of {wj : j = 1, 2, ... , n}.

Under these assumptions, we can prove the following theorem in a way similar
to the proof of Theorem XIII-1-2.
Theorem XIII-3-4. The system e° fy = x"f (x, c) has two solutions

¢(x, 91, e)and i (x, µ, e) such that
(i) these two solutions are bounded and holomorphic in the domain
So = {(x, 9, e) : Ix1 > N, x E S(p1, p2), 1141 < ul, 0 < IEI < e1,1 arg el < p0},
if p1 and E1 are sufficiently small,

(ii) the solution ¢(x, µ, e) admits an uniform asymptotic expansion
+oo
F, x-k&(µ, e) as x -. oo
k=1
in So, where coefficients &(µ,e) are bounded, holomorphic, and admit uni-
form asymptotic expansions in powers of a as c -+ 0 in the domain {(µ, e)
Iµ1 < {31, 0 < lei < el, I argel < po},
(iii) the solution 1'(x, N, e) admits an uniform asymptotic expansion:
+oo
(x, , e) ' ehWh(x, µ) as f -4 0
h=0
in So, where coefficients >Gh(x,#) are bounded, holomorphic, and admit uni-
form asymptotic expansions in powers of x'1 as x - oo in the domain
{(x,µ') : Ixf > N, x E S(p1,p2), Jill < p1}.
FLrther more, there exist functions tc(x, µ, e) for t = 0,1, ... such that
(a) these functions ¢'c satisfy conditions (i) and (is) given above,
(b)
h=0
A complete proof of Theorem XIII-3-4 is found in [Si7J and (Si101.
Remark XIII-3-5. In the proof of Theorem XHI-1-2, we used the assumption that
the matrix Ao on the right-hand side of (XIII.1.10) is invertible (cf. Assumption
III of §XIII-1). Without such an assumption, we can prove the following theorem.
Theorem XIII-3-6. Let F(x, yo, yr, ... , yn) be a nonzero polynomial in yo, yl, ,
00
yn whose coefficients are convergent power series in x'1, and let p(x) = an,x-°`
M--0
E C[(x-1)) be a formal solution/ of the differential equation
(XIII.3.3) F(x,y,Ly,..., fn)=0.

\arg
Then, for any given direction x = 0, there exist two positive numbers 6 and a
and a function 0(x) such that
(i) ¢(x) is holomorphic in the sector S = {x E C:0< [xj <6, (arg x - 0[ < a},
(ii) 0(x) admits the formal solution p(x) as its asymptotic expansion as x o0
in S,
(iii) 0(x) satisfies differential equation (XIII.3.3) in S.
The main idea of the proof is similar to the proof of Theorem XIII-1-2. However,
we need the thorough knowledge of the structure of solutions of a linear system
xd = A(x)y that we shall explain in §XIII-6. Also, it is more difficult to define
the paths of integration (cf. [I] and (RS11). A complete proof of Theorem XIII-3-6
is found in [RS1). See, also, §XIII-8.
XIII-4. A block-diagonalization theorem

Consider a system of linear differential equations
dg
(XIIi.4.1) = x' A(x}y,
dx
where r is a positive integer, y" E C', and A(x) is an n x n matrix. The entries of
A(x) are holomorphic with respect to a complex variable x in a sector
(XIII.4.2) jxl > No, f argx[ < ao,
where No and ao are positive numbers. Assume that the matrix A(x) admits an
asymptotic expansion in the sense of Poincare
00
(X1II.4.3) A(x) E x'"A

&'=o
as x - oo in sector (XIII.4.2), where coefficients A are constant n x n matrices.

Suppose also that Ao = A(oo) has a distinct eigenvalues A1, A2, ... , J1t with mul-
tiplicities n1, n2, ... , nt, respectively (nl + n2 + + nt = n). Without loss of
generality, assume that Ao is in a block-diagonal form:
()UII.4.4) AO = diag [Al, A2, ... , At) ,
where Aj are of x n, matrices in the form
(XIII.4.5) A, =A,In, +Nj (j=1,2,...,1).

Here, A , is a lower-triangular nilpotent n, x n, matrix. The main result of this
section is the following theorem (cf. (Si7J).
Theorem XIII-4-1. Under the assumption (XIIL4.3) and (XIII.4.4), there exists
an n x n matrix P(x) whose entries are holomorphic in a sector
(XIII.4.6) IxI > N1, I argxj < al,
where Ni 1 and a1 are sufficiently small positive numbers, such that

(i) P(x) admits an asymptotic expansion
00
(XIII.4.7) P(x) Ex-"p- (Pp = In),

"=0
as x -' oo in sector (X111.4.6), where coefficients P" are constant n x n

matrices,
(XIII.4.8) y = P(x)zi
reduces system (X111.4.1) to
di ,
(XIII.4.9) = x B(x)i,
where B(x) is in a block-diagonal form
(XIII.4.10) B(x) = diag [BI (x), B2(x),... , Bt(x)J .
Hen, B,(x) an n, x n. matrices and admit asymptotic expanswns

00
(XIII.4.11) B, (x) ^-- E xBj"

"=o
as x oo in (XIII.4.6), where coefficients B,, are constant n) x n) matrices.

Proof.
From (XIII.4.1), (XIII.4.8), and (XIII.4.9), we derive the equation
dP
(XIII.4.12) = xr(A(x)P - PB(x)]
that determines the matrices P(x) and B(x). Set
(XIII.4.13) A(x) = Ao + E(x), B(x) = A0 + F(x), P(x) = In + Q(x).

Then, E(x) = O(x-1), F(x) = O(x'1), and Q(x) = O(x-1). Furthermore,

(XIII.4.12) becomes
(XIII.4.14) dQ = x''[AoQ - QAo + E - F + EQ - QF].

dx
Write each of three matrices E(x), F(x), and Q(x) in a block-matrix form according
to that of Ao in (XIII.4.4), i.e.,
(XIII.4.15)
F(x) = diag]Fl,F2,... ,Ft],
E11 E12 ... Ell Q11 Q12 ... Qlt
E21 E22 ... E2t
E(x) _ Q(x) =
Ell E2 Eu Qtt
Q11 Q2t .
where EJk and QJk are nJ X nk matrices and F, are nJ x nJ matrices. Set
(XIII.4.16) QJJ = 0 (j = 1,2,... ,t).
From (XIII.4.4), (XIII.4.14), (XIII.4.15), and (XIII.4.16), it follows that
(XIII.4.17) F,=EJJ+FEJhQh, (j=1,2,...,t)
h#J
and
dQjk
(XIII.4.18) = x' [43Qik - QJkAk + EJk + EJhQhk - QJkAk] (j j4 k).
h*k
Substituting (XIII.4.17) into (XIrII.4.18), a system of nonlinear differential equations
dQJk
= x' LAJQJk - QJkAk + EJhQhk
dx
{ h*k
(XIII.4.19)
- QJk (Ekk + EkhQhk) + EJkJ (j # k)
h*k
is obtained. Since it is assumed that .11i ... , Al are distinct eigenvalues of Ao and
that A0 is in the block-diagonal form (XIII.4.5), upon applying Theorem XIII-1-2
to (XIII.4.19) we can construct a desired holomorphic solution Qjk(x) of (XIII.4.19)
00
which admit an asymptotic expansion Qjk(x) > x "Qjk" (j, k = 1, 2,... , 3; j
"=1
k), where Qjk" are constant nj by nk matrices. Defining Fj by (XIII.4.17) and then
B(x) by (XIII.4.13), the proof of Theorem XIII-4-1 is completed. 0
Theorem XIII-4-1 concerns the behavior of solutions of system (XIII.4.1) near
x = oo. Since it is useful to give a similar result concerning behavior of solutions
near x = 0, we consider, hereafter in this section, a system of differential equations
(XIII.4.20) xa+1 dY = A(x)y,

where d is a positive integer and the entries of n x n matrix A(x) are holomorphic
in a neighborhood of x = 0. Also, assume that A(0) is in a block-diagonal form
()III.4.21) A(0) = diag[A11,,, +N2,... +JVt],
where a1, ... , .1t are distinct eigenvalues of A(0) with multiplicities n1, n2, ... , nt,
respectively (n1 + n2 + + nt = n), and, for each j, .W is a lower-triangular and
nilpotent nj x nj matrix.
Comparing the present situation with that of Theorem XIII-4-1, we notice the
following two differences:
(a) singularity is at x = 0 in the present situation, while singularity is at x = 00
in Theorem XIII-4-1,
(b) The power series expansion of A(x) is convergent in the present situation,
while A(x) in Theorem XIII-4-1 admits only an asymptotic expansion in a
sector containing the direction arg x = 0.
We can change any singularity at x = 0 to a singularity at x = oo by changing
1.
the independent variable x by Also, any direction argx = 8 can be changed
x
to the direction arg x = 0 by rotating the independent variable x. Furthermore,
the asymptotic expansion P of Po(x) and the expansion b of Be are formal power
series satisfying the equation xd+i dP = AP - PB. This implies that two matrices
P and b are independent of P. Hence, using Corollary XIII-1-3, the following result
is obtained.
Theorem XIII-4-2. Let A(x) be an n x n matrix whose entries are holomorphic
in a neighborhood of x = 0. Also, let d be a positive integer. Assume that the matrix
A(0) is in block-diagonal form (XIII.4.21), where al, ... , at are distinct eigenvalues
of A(0) with multiplicities n1, n2, ... , nt, respectively (n1 + n2 +, + nt = n), and
for each j, JVj is a lower-triangular and nilpotent n, x nj matrix. Fix a real number
0 so that (a, - Xk)e-utO V EP for j 96 k. Then, there exists two positive numbers be
and ee and an n x n matrix Pe(x) such that
(a) the entries of Po(x) are holomorphic and admit asymptotic expansions in pow-
ers of x as x - 0 in the sector Se = Ix: 0 < IxI < be, I arg x - 01 < + to },
00
(b) if >2 xmP,,, is the asymptotic expansion of Pe(x), then this expansion is
m=0
independent of 8 and Po = I,,,
(c) the transformation PB(x)u" changes system (XIIL4.20) to a system
(XIII.4.22) xd+1 = Be(x)u,

where the matrix Be(x) is in a block-diagonal form
(XIII.4.23) Be(x) = diag (B1e(x), B2e(x), ... , Bte(x)]
For each,, Bje is an n, xn2 matrix which admits also an asymptotic expansion
in x as x--i0 in So.
The main claims of this theorem are
(i) the asymptotic expansion of Pe(x) is independent of 0,

(ii) the opening of the sector So is larger than k.
Proof of Theorem XIII-4-2 is left to the reader as an exercises.

Remark XIII-4-3. Using Theorem XIII-3-4, we can generalize Theorem XIII-
4-1 to the system e°dx = xr'A(x, µ, e)y", where r and o are non-negative integers,
y" E C^, and A(x, µ', e) is an n x n matrix with the entries that are bounded and
holomorphic with respect to the variable (x, µ, e) E C x C'° x C in a domain
Do = {(x, µ, e) : 1x1 > N, Iµ1 < µo, 0 < 1e1 < eo, 0 < 1 arg e1 < po}. Also, assume
00
that A admits a uniform asymptotic expansion A(x, µ, e) - > ehAh(x, µ) in Do
h=0
as a -+ 0, where coefficients Ah (X, µ) are bounded and holomorphic in the domain
{(x,µ) : 1x1 > N, 1µ1 < µo}. A complete proof of this result is found in [Si7] and
[Hs2].
XIII-5. Cyclic vectors (A lemma of P. Deligne)

In the study of singularities, a single n-th-order differential equation is, in many
cases, easier to treat than a system of differential equations. In this section, we
explain equivalence between a system of linear differential equations and a single
n-th-order linear differential equation.
Let us denote by 1C the field of fractions of the ring C[(x]] of formal power series
in x, i.e.,
K=
I 9:
p E C[Ix]J, 9 E C[Ix]j, 9 34 0}
Also, denote by V the set of all row vectors (cl(x), c2(x), ... , c,,(x)), where the
entries are in the field K. The set V is an n-dimensional vector space over the
field 1C.
Define a linear differential operator C : V - V by G[vl = by + 7l(x) (v' E V ),
where b = x d
and S2(x) is an n x n matrix whose entries are in the field 1C. We
first prove the following lemma.
Lemma XIII-5-1 (P. Deligne [Del]). There exists an element iio E V such that
{v"o, Gv"o, G2v"o, ... , G"-lvo} is a basis for V as a vector space over IC.
Proof.
For each nonzero element v of V, denote by µ(v'' the largest integer t such that
{ii, Cii, C2v, ... ,.CeV) is linearly independent over K. In two steps, we shall derive
a contradiction from the assumption that max{µ(v') : v' E V} < n - 1.
5. CYCLIC VECTORS 425
Step 1. First, we introduce a criterion for linear dependence of a set of elements of

V. Consider a set {i11, ... vm } C V, where m is a positive integer not greater than
n = dim, V. Let 6j = (c31,c12,... ,Cin) (j = 1,2,... ,m). Set .7 = {(jl,... ,jm)
1 < it < j2 < . < jm < n}, and introduce a linear order .7 -' { 1, 2,... , (M-)) in
the set J. Let us now define a map
(.)
Vm={(v1,V2,...,Vm): 61 EV (j=l,...,m)} -+ Id.)
(VI, ... , Vm) --+ V1 A v2 A ... A vm, where
1,AV2A ... A vm
f cI31 Cj32 ... clim
det : : : (jl,... ,jm) E .7

C-i 1 Cmj2 C-3-
It is easy to verify the following properties of V1 A v"2 A A v,,,:
Vk_1 A 1m_1 A Um
(1) = Vk_1 A Vm_1 A v"m
+ v'k_1 A Vm_1 A Vm,
V1 A ... A Vk_1 A (a Vk) A vk+I A ... A Vm_1 A Vm

(2)
= 6k_1 A irk A vm),
for all aEIC,
VIA.-.A irk A ...A...AVjA ... A Vm
(3)
= -(11A...A iY A A Vm),
(4) a sety{ {6j, 62,... , Vm } E V is linearly dependent if and only if v'1 AV2 A A 17m =
6 in K('^).
Step 2. Fix an element Vo of V such that µ(6o) = max{p(V) : v" E V} < n-1. Since
p(VO) < n-1, another element w of V can be chosen so that {VO, Cv"o, CZi3o, ... , CnOv',
w"} is linearly independent, where no = max{µ(v") : v E V}. Set v' = vo+Ax"'O E V,
where A E C and m is an integer. Then, Cwt = Cw"o + C'(Ax-ti) = Crvo +
Axm(C + m)tw. Since {V, Cv", ... , Cn0+1V} is linearly dependent, it follows that
v A DU A A CnOv' A C"O+I V = 6 for all A E C and all integers in. Note that
v A,06 A A C"OiYA C"°+16 is a polynomial in A. Since this polynomial is iden-
tically zero, each coefficients must be zero. For example, the constant term of this
polynomial is i6o A CVO A ... A 00+I V"o. This is zero since {iio, Ciio..... G"0+Iuo}
is linearly dependent. Compute the coefficient of the linear term in A of the poly-
nomial. Then,
m)n0,+1ti
w A C60 A ... A Cna+1v'o + v'Y A ... A 00 Vo A (C +
no
+ 1:60A...ACi-IVoA (C+m)lw A CJ+'A A...ACnb+Ii = 0
J=1
identically for all integers m. The left-hand side of this identity is a polynomial in
in of degree no + I. Hence, each coefficient of this polynomial must be zero. In
particular, computing the coefficient of mn0+1, we obtain vOAGvOA . A GnbtloAw' =
0. This is a contradiction, since {vo, Gvo, ... , L"Ovo, w} is linearly independent.
This completes the proof of Lemma XIII-5-1.
Definition XIII-5-2. An element vo E V is called a cyclic vector of G if {6o, Lu0,
L2v'o, ... , Ln-1%) is a basis for V as a vector space over X.
Observation XI II-5-3. Let uo be a cyclic vector of Land let P(x) be the n x n ma-
Uo
Gvo
trix whose row vectors are {vo,G'o,... ,L' 'io}, i.e., P(x) = . Then,
Gn-lvo
nv"o
L21 yo
o
= and, hence, setting A(x) = L[P(x)]P(x)-1 = bP(x)P(x)-l +
P(x)f2(x)P(x)-1, we obtain
0 1 0 0 .. 0
0 0 1 0 ... 0
A
0 0 0 0 ... 1
ap al a2 a3 ... an-1
with the entries a, E C. Thus, we proved the following theorem, which is the main
result of this section.
Theorem XIII-5-4. The system of differential equations
y1
(XIII.5.1) by' = fl(x)y", where g _
Y.
becomes
(XIII.5.2) bu" = A(x)u,
if y is changed by u = P(x)y". System (XIll.5.2) is equivalent to the n-th-order

n-1
(XHI.5.3) bnq - arbrq = 0, where q = yl.
1=o
5. CYCLIC VECTORS 427
Example XIII-5-5.
(1) Let us consider the system
y1
(a) by" = 0, where y=
yn
The transformation
(T) u = diag[l, X'... , xn-11y
changes system (a) to
(E) bu = diag[0,1,... ,n - 1]u.
Further, the transformation
1
2
-- 22 U
(r)
2n-1
changes (E) to the form

0 1 0 0 0
0 0 1 0 0
w,
0 0 0 0 1
ao a1 a2 a3 ...
an-1
where ao, a,,.. . , an-1 are integers. Hence, the transformation

1 1 1 ... 1
0 1 2 .. n- 1
w" = 0 1 22 ... (n - 1)2 diag [1, x,... ,x°-1
y
0 1
2n-1 ... (n - 1)n-1
changes (a) to (E'). This implies that vo = (1, x, x2'. .. xn-1) is a cyclic vector in
this case.
(II) Next, consider the system
(b) by = Ay,
where A is a constant diagonal n x n matrix. Choose a transformation similar to
(T) of (a) to change system (b) to
(E") oiZ = A'u
so that A' is a diagonal matrix with n distinct diagonal entries. Then, a transfor-
mation similar to (T') can be found so that (E") is changed to (E') with suitable
constants ao, al, ... , an_ I .
XIII-6. The Hukuhara-Turrittin theorem

In this section, we explain a theorem due to M. Hukuhara and H. L. TLrrittin that
clearly shows the structure of solutions of a system of linear differential equations
of the form
(XIII.6.1)
where the entries of the n x n matrix A are in K (cf. §XIII-5). In order to state
this theorem, we must introduce a field extension f- of K. To define L, we first set
+a
E a.nxm1" : a.., E C and M E Z ,
M=M I
+a
where 7L is the set of all integers. For any element a = F a,nxn'/° of K,,, we
M=M
+00
define x ji by x da = F \ v) amx'nl '. Then, K, is a differential field. The field
M=M
+oo
L is given by L = U K which is also a differential field containing K as a subfield.
L=1
Furthermore, L is algebraically closed. The Hukuhara-T rrittin theorem is given
as follows.
Theorem XIII-6-1 ({Huk4j and [Tu1j). There exists a transformation
(XIII.6.2) y' = UI
such that
(i) the entries of the matrix U are in L and det U ;j-1 0,
(ii) transformation (XIII.6.2) changes system (X111.6.1) to
(XIII.6.3) xd = Bz',
where B is an n x n matrix in the Jordan canonical form
B=diag[Bl,B2,...,Bpj, Bj=diag[BJi,Bj2i...,B,,n,],
(XIII.6.4)
Bjk =AI In,,. + Jn,,,.
6. THE HUKUHARA-TURRITTIN THEOREM 429
Here, I,,,, is the njk x njk identity matrix, J.,,, is an n 1k x n3k nilpotent
matrix of the form
0 1 0
0 0 1
(XIII.6.5) Jn",
0 0 0
0 0 0
and the Aj are polynomials in xfor some positive integer s, i.e.,

d,
(XIII.6.6) Aj = > A x-'18 where all E C (j = 1, 2, ... , p)
r=O
and
(XII1.6.7)
A,d, 0 if dJ > 0 and A., - A, are not integers if i 0 j.
Proof.
Without loss of generality, assume that the matrix A of system (XIII.6.1) has
the form
0 1 0 0 0
0 0 1 0 ... 0
A
0 0 0 0 ... 1
00 a1 a2 a3 "' an-1
an- 1
where ak E 1C (k = 0, 1, ... , n - 1) (cf. Theorem XIII-5-4). Set E= A - n In.
Then,
(XIIi.6.8) trace [E] = 0.
Consider the system
(XIII.6.9) X E = E.
Case 1. If there exists an n x n matrix S with the entries in X such that det S 0 0
dii
and the transformation w = Sii changes (XIII.6.9) to x = A(x)u", where the
entries of A are in C[[x]], then there exists another n x n matrix S with the entries
in X such that det S 96 0 and the transformation
(XIII.6.10) w = Suu
dig
changes (XIII.6.9) to x = Aoii, where the entries of the matrix A0 are in C.
F irthermore, any two distinct eigenvalues of A0 do not differ by an integer (cf.
Theorem V-5-4). Hence, in this case, system (XIII.6.1) is changed by transforma-
tion (XIII.6.10) to
dil
xaj = [an 1 In+Ao]u.
This proved Theorem XIII-6-1 in this case.
Case 2. Assume that there is no n x n matrix S with entries in X such that

det.S 0 and the transformation w = Su' changes (XM.6.9) to xdu = A(x)u,
where the entries of A are in C([x]]. Since
1 0 0 0
- ln an_1 1 0 0
E =
0 0 0 - ln an_1 1
ap al a2 a3 an-2 an-1 -- -an-1,
a cyclic vector can be found for system (XIII.6.9) by using the matrix W defined
by
w11 win
with
writ Wnn
[wll ... Win] = (10 ... 01,
[w,,1
... wJn] = V1-1[1 0 01,
where V([c1 cn]) = x[c1 [c1 - - cn]E. The matrix W is lower-
triangular and the diagonal entries are {1, ... ,1}, i.e.,
1 0 0 0
1 0 0
(XIII.6.11) W=
1
If (XIII.6.9) is changed by the transformation v" = Wtv", then
(XIU.6.12) X!LV _ x + WEJ W-Y.
It follows from (XIII.6.8) and (XIII.6.11) that
( XIII . 6 . 13 ) t racel x + WE]W- l = 0.
Also,
0 1 0 0 0
0 0 1 0 0
x + WE] W-1 = V[ W]W-1
11
0 0 0 0 l
22 A
W i31 /32 An-1
where 1k E )C (k = 0, 1,... , n - 1). In particular, from (XIII.6.13), it follows that

On-1 = 0. Under our assumption, not all Qt are in C[[x]]. Set 9 = (t: at V C[[x]]}
+00
and set Also, /3t = x-u" E Qtmxm
(t E 3), where, for each t, the quantity µt is a
m=0
+oo
positive integer, 1: /3tmxm E C[[xJ] and #to 54 0. Set
m=0
k=max(n µt t :IE91. J
Then, ut < k(n - t) for every t E J and µt = k(n - t) for some t E J. This implies
that
(I) Of = L.
Qt,mxm
(t=0,1,... n- 1)
m>-k(n-t)
and
(11) Qt = x-k(n-t) (Ct + xqt)

for some t such that k(n - t) is a positive integer, ct is a nonzero number in C,
qt E C[[x]], and Qt,m E C. We may assume without any loss of generality that
k = h for some positive integers h and q.
9
Let us change system (XIII.6.12) by the transformation
v" = diag [1, x-k, ... , x-(n-1)kI u".
Then,
(XIII.6.14)
where
0 1 0 0
0 0 1 0
(XIII.6.15) F= + kxkdiag [0,1, ... , n - 1]
0 0 0 0
70 71 72 73
and
7t = xk(n-t)o, =:m>-k(n-t)
$,,mxm+k(n-t) E C[ [x'1911 (0 < t < n - 1).
In particular, 7n-1 = 0.
+oo
Setting F = E xm/9Fm, where the entries of F,,, are in C, we obtain
m=o
0 1 0 0 0 01
Fo=
CO
C1 C2 C3 ... Cn_2 0
where the constants co, cl ... , cn_2 are not all zero. This implies that the matrix
F0 must have at least two distinct eigenvalues. Hence, there exists an n x n matrix
T such that
(1) T = XM19Tm, where the entries of the matrices Tm are in C and To is
m=o
invertible,
(XIII.6.16) y = Ti
changes system (XIII.6.14) to a system xd = x-kGxi with a matrix G in

a block-diagonal form G = [ 0 0 , where Gl and G2 are respectively
G2 J
n1 x nl and n2 x n2 matrices with entries in C[[x1/91and that nl + n2 = n
and n., > 0 (j = 1, 2) (cf. §XIII-5). Therefore, the proof of Theorem XIII-6-1
can be completed recursively on n. 0
Observation XIII-6-2. In order to find a fundamental matrix solution of (XIII.6.1),
let us construct a fundamental matrix solution of (XIII.6.3) in the following way:
Step 1. For each (j, k), set
4'1k = xA,0 eXp[AJ (x)] exp[(log x)Jf, ],
where
if dj = 0,
-t/s if d1>0.
Step 2. For each j, set
Dj = diag ['j1, xAJ0 exp[A,(x)] exp[(logx)Jj],
where J. = diag [J,,,,, J,,72 , ... J,mJ ] .

Step 3. Set
A = ding [Al(x)In A2(x)I,,,, ... ,

C = diag [,\101., + J1, 1\20I., + J2, ... ,
Then,
(XIII.6.17) 4) = diag [451, 4i2, ... , 4ip] = xC exp[A]
is a fundamental matrix solution of (XIII.6.3), where
xC = diag [xA1OxJ, xl\2OxJ2' ... 'X aPOxJP] , xJ, = exp[(logx)JjJ.
The matrix
(XIII.6.18) U4i = Uxc exp[A]
is a formal fundamental matrix solution of system (XIII.6.1), where U is the matrix

of transformation (XIII.6.2) of Theorem XIII-6-1. The two matrices exp[A] and xC
commute.
Observation XIII-6-3. Theorem XIII-6-1 is given totally in terms of formal
power series. However, even if the matrix A(x) of system (XIII.6.1) is given ana-
lytically, the entries of U of transformation (XIII.6.2) are, in general, formal power
series in xl1", since the entries of the matrix T(x) of transformation (XIII.6.16)
are formal power series in general. Transformation (XIII.6.16) changes system
(XIII.6.14) to a block-diagonal form. Therefore, in a situation to which Theo-
rem XIII-4-1 applies, transformation (XIII.6.2) can be justified analytically. The
following theorem gives such a result.
Theorem XIII-6-4. Assume that the entries of an n x it matrix A(x) are holo-
morphic in a sector So = {x E C 0 < lxJ < ro, I argxi < ao} and admit asymptotic
expansions in powers of x as x 0 in So, where ro and ao are positive numbers.
Assume also that d is a positive integer and y" E C'. Let S be a subsector of So
whose opening is sufficiently small. Then, Theorem XIII-6-1 applies to the system
(XIII.6.19) xd+1dy = A(x)f

dx
with transformation (XIII.6.2) such that the entries of the matrix U of (XIII.6.2)
are holomorphic in S and each of them is in a form x,00(x) where p is a rational
number and O(x) admits an asymptotic expansion in powers of xl"° as x 0 in S,
where s is a positive integer.
Observation XIII-6-5. In the case when the entries of the matrix A(x) on the
right-hand side of (XIII.6.19) are in C[[x]l and A(0) has n distinct eigenvalues, the
matrix A(x) also has n distinct eigenvalues A1(x), A2(x), ... , which are in
C((xJJ. Furthermore, the corresponding eigenvectors p"1(x), p"2(x), ... ,15n(x) can be
constructed in such a way that their entries are in Chill and that p""1(0),7"2(0), ... ,
p",,(0) are n eigenvectors of A(O). Denote by P(x) the n x n matrix whose column
vectors are p1(x), p2(x), ... , p,+(x). Then, detP(O) 36 0 and P(x)-lA(x)P(x) =
diag[A1(x),A2(x),... ,An(x)]. This implies that the transformation y = P(x)ii
changes system (XIII.6.19) to
(XIII.6.20) xd+1 dx {diagiAi(x)A2(x).... , An(x)] - xd+1P(x)-1 d ( )

It is easy to construct another n x n matrix Q(x) so that (a) the entries of Q(x)
are in C[(x]], (b) Q(0) = 4,, and (c) the transformation i = Q(x)v changes system
(XIII.6.20) to
(XIII.6.21) xd+1 dv = diag [ft1(x), µ2(x), ... , µn(x)] v,

dx
where µ1(x), µ2(x), ... , i (x) are polynomials in x of degree at most d such that
A, (x) = it) (x) + O(xd+1) ( j = 1, 2, ... , n). Therefore, in this case, the entries of
the matrix U of transformation (XIII.6.2) are in 1C.
Observation XIII-6-6. Assume that the entries of A(x) of (XIII.6.19) are in
C([x]]. Assume also that A(O) is invertible. Then, upon applying Theorem XIII-6-1
d1
to system (XIII.6.19), we obtain = d for all j. Using this fact, we can prove the
s
following theorem.
Theorem XIII-6-7. Let Q;i1(x) and Qi2(x) be two solutions of a system
(XIII.6.22) xd+1 dy = A(x)yf + x f (x),

dx
where d is a positive integer, the entries of the n x n matrix A(x) and the C"-valued
function 1 *(x) are holomorphic in a neighborhood of x = 0, and A(O) is invertible.
Assume that for each j = 1, 2, the solution ¢,(x) admits an asymptotic expansion
in powers of x as x - 0 in a sector S3 = {x E C : lxl < ro, aj < arg x < b3 },
where ro is a positive number, while aJ and bi are neat numbers. Suppose also that
S1 n S2 0. Then, there exist positive numbers K and A and a closed subsector S =
{x : lxl < R,a < argx < b} of Sl nS2 such that K exp[-Alxl -d]
in S.
Proof.
Since the matrix A(O) is invertible, the asymptotic expansions of1(x) and ¢2(x)
are identical. Set 1 (x) = ¢1(x) -$2(x). Then, the C"-valued function >G(x) satisfies
system (XIII.6.19) in S, n$2 and ii(x) ^_- 6 as x 0 in S1nS2. By virtue of Theorem
XIII-6-4, a constant vector 66 E C" can be found so that r%i(x) = U4D(x)6, where
lb(x) is given by (XIII.6.17). Now, using Observation XIII-6-6, we can complete
the proof of Theorem XIII-6-7.
Observation XIII-6-8. The matrix A = diag [AI In,, A2In3, ... , API,y] on the
right-hand side of (XIII.6.18) is unique in the following sense. Assume that an-
other formal fundamental matrix solution Uxcexp[A] of system (XIII.6.1) is con-
structed with three matrices U, C, and A similar to U, C, and A. Since the
matrices Uxc exp[A] and UxO exp[A] are two formal fundamental matrices of sys-
tem (XHI.6.1), there exists a constant n x n matrix r E GL(n, C) such that
Uxc exp[A] = U5C exp[A]I' (cf. Remark IV-2-7(1)). Hence, exp[A]T exp[-A] =
x-CU-IUxC. Using the fact that r is invertible, it can be easily shown that A = A
if the diagonal entries of A are arranged suitably. For more information concerning
the uniqueness of the Jordan form (XIII.6.3) and transformation (XIII.6.2), see, for
example, [BJL], [Ju], and [Leve].
Observation XIII-6-9. The quantities A.,(x) are polynomials in x1l'. Set w =
2a[!]
exp and x 1/a = wx 1/a Then, i = x. Therefore, if z 1/e in Ur Cexp[A] is
replaced by zI/', then another formal fundamental matrix of (XIII.6.1) is obtained.
This implies that the two sets {Aj (i) : j = 1, 2,... , p} and (A.,(x) : j = 1, 2,... , p)
are identical by virtue of Observation XIII-6-8.
Observation XIII-6-10. A power series p(x) in x1/' can be written in a form
a-1
Ax) = Exh1'gh(x), where ql(x) E C[[x]] (j = 0, 1, ... , s - 1). Using this fact and
h=0
Observation XIII-6-9, we can derive the following result from Theorem XIII-6-1.
Theorem XIII-6-11. There exist an integer q and an n x n matnx T(x) whose
entries are in C[[x]] such that
(a) det T (x) 96 0 as a formal power series in x,
dil
(b) the transformation y" = T(x)t changes system (XII1.6.1) to x = E(x)iZ
with an it x it matrix E(x) such that entries of x9E(x) are polynomials in x.
The main issue here is to construct, starting from Theorem XIII-6-1, a formal
transformation whose matrix does not involve any fractional powers of x in such a
way that the given system is reduced to another system with a matrix as simple as
possible. A proof of Theorem XIII-6-11 is found in [BJL].
Changing the independent variable x by x-1, we can apply Theorem XIII-6-1 to
singularities at x = oo. The following example illustrates such a case.
Example XIII-6-12. A system of the form
P(x) 0]b, y = lyd,

where P(x) = xm+ ahxin is a positive odd integer, and the ah are complex
h=1
numbers, has a formal fundamental matrix solution of the form
x F(x)
1 0 f1 1 e 0
0 x-1/2 ] l I -111 0 eE(t,a) I
where
m 1/2 +00
+ E ak
xk + E bk(a)
xk
i
k=1 k=1
2 (a)x(m+2-2h)/2,
E(x, a) = (m 2) x(m+2)/2 + (.m + 2 - 2h) bh
+ 1<h<(m+2)/2
+00
and F(x) = xq> x-h Fr, with an integer q and 2x2 constant matrices Fh such that
h--O
+00
det h=O x-h Fh ,-6 0 as a formal power series in x-l. For details of construction,
see [HsSj and [Sil3j.
XIII-7. An n-th-order linear differential equation at a singular point of

the second kind
Let us look at the formal fundamental matrix solution (XIII.6.18) of system
(XIII.6.1). First, notice that if we set k = max { as :j = 1,2,... pthen k is the
JJJ
order of singularity of system (XIII.6.1) at x = 0 (cf. Definition V-7-8). Let

(XIII.7.1) 0<k1 <k2 <... <k9_1 <kq
be the distinct values among p non-negative rational numbers
set th =
d1 d2
s s
, ... ,
s
Also, .
n, (h = 1, 2,... , q). It is easy to see that lh > 0 (h = 1, 2,... , q)
d /s=kh
q P
and Elh = En, = n.
h=1 7=1
Observation XIII-7-1. System (XIII.6.1) has th linearly independent formal so-
lutions of the form
(XIII.7.2) 'ih,v(x) = x'1h - exp[Qh,v(x)Wh,,,(x) (v = 1, ... ,;h = 1, ... , q),
where -yh,,, E C, Qh,,,(x) is either equal to 0 or a polynomial in x-1/' of the form
Qh.v(x) = I2h,vx-kh(I +O(x1/')) (11h,v E C and I2h,, # 0),
and the entries of lh,,,(x) are polynomials in logx with coefficients in
Define q + 1 points (Xh, Yh) (h = 0,1, ... , q) recursively by
(Xo,Y0) (010),
(Xh, Yh)== (Xh-1 + eh, Yh-t + kheh) (h = 1, 2,... , q).
Let us denote by N the polygon whose vertices are q + 1 points (Xh, Yh) (h =
0,1, ... , q). The polygon N has q distinct slopes kh given in (XIII.7.1).
7. AN N-TH-ORDER LINEAR DIFFERENTIAL EQUATION 437
Definition XIII-7-2. The polygon N is called the Newton polygon of system

(XIII.6.1) at x = 0.
Observation XIII-7-3. In §XIII-5, it was shown that system (XIII.6.1) is equiv-
alent to an n-th-order linear differential equation
n-1
(XIII.7.3) anb" r) + E atbtrl = 0,
t=o
where b = x , at E C((xj), and an # 0 (cf. Theorem XIII 5 4). For the differential
operator
n-1
(XIII.7.4) C(>,J = anb"r) + j:atdtrt.
t=o
the Newton polygon N(C) is defined in the following way.
If a power series a = E c,,x'n E C((xj] is not 0, we set v(a) = min{m : cn

m=0
0}. If a = 0, set v(0) = +oo. For operator (XIII.7.4), consider n + 1 points
(Q, v(at)) (t = 0, 1, ... , n) on an (X, Y)-plane. Set
Pt = ((X, Y) : 0 < X < e, Y > v(at)}, and P = UP,.

t=o
Definition XIII-7-4. The boundary curve C of the smallest convex set containing
P is called the Newton polygon of the operator C at x = 0.
Denote by N(C) the Newton polygon of C at x = 0.
Definition XIII-7-5. Two Newton polygons are said to be identical if the two
polygons become the same by moving one or the other upward in the direction of
the Y-axis.
Now, we prove the following theorem.
Theorem XIII-7-6. If system (XIII.6.1) and differential equation (X111.7.3) are
equivalent in the sense of Theorem X111-5-4. then the two Newton polygons N and
N(C) are identical.
Proof.
The proof of this theorem will be completed if the following three statements are
verified:
(a) If./V(C) has only one nonvertical side with slope k, then differential equation
(XIII.7.3) is equivalent to a system xk+1 = A(x)g with a matrix A(x)
whose entries are power series in x11", and A(0) is invertible if k > 0, where
s is a positive integer such that sk is an integer.
(b) If (XIII.7.1) gives slopes of all nonvertical sides of N(L), then the operator L
is factored in the following way:
(XIII.7.5)
where, for each h, N(Lh) has only one nonvertical side with slope kh.
(c) If L is factored as in (b) and each differential equation Lh[r)h] = 0 is equiv-
alent to a system xdih = Ah(x)uh, then the differential equation (XIII.7.3) is
equivalent to x !L = diag [A, (x), A2(x), ... , Aq(x)] y'.

Statement (a) can be proved by an idea similar to the argument which is used to
reduce system (XIII.6.12) to (XIII.6.14) in Case 2 of the proof of Theorem XIII-6-1.
A proof of Statement (b) is found in [Mall], [Si16] and [Si17, Appendix 1]. Look
at the system
Lq]u] = V, L1... L9_1[v] = 0.
Then, Statement (c) can be verified recursively on q without any complication. The
proof of Theorem XIII-7-6 in detail is left to the reader as an exercise.
Combining Observation XIII-7-1 and Theorem XIII-7-6, we obtain the following
theorem.
Theorem XIII-7-7. If the distinct slopes of the nonvertical sides of N(L) are
given by (XIII.7.1), then the n-th-order differential equation (XIIL7.S) has, at x =
0, n linearly independent formal solutions of the form
(XIII.7.6) rlh,v(x) = xtin.,. exp[Qh.v(x)]Oh,v(x)
where
(i) if
{(X, Yh-I + kh(X - Xh_1)) : Xh-1 C X < Xh}
is the nonvertical side of N(L) of slope kh, then
It, = Xh - Xh-1 (h = 1, 2,... , 9),
(ii) ryh,v E C, Qh,v(x) is either equal to 0 or a polynomial in x- 1/11 of the form
Qh,v(x) = lAh,vx-k"(1 + O(xi'°)) (I<h,v E C and uh,v 54 0),
and the quantities Oh,v(x) are polynomials in log x with coefficients in C[[x'/']]
Here, s is a positive integer such that skh (h = 1,... , q) are integers, and PI + P2 +
+ Pq = n.
A complete proof of Theorem XIII-7-7 is found in [St].
We can construct formal solutions (XIII.7.6), using an effective method with the
Newton polygon N(L). The following example illustrates such a method.
7. AN N-TH-ORDER LINEAR DIFFERENTIAL EQUATION 439
Example XIII-7-8. Consider the differential operator C = x63 - x62 - 5 - 1 or

the third-order differential equation C[q] = 0. The Newton polygon N(C) is given
by Figure 6. In this case, q = 2, kl = 0, and k2 = 1.
1 3
FIGURE 6.
+00
(i) For k1 = 0, set i _ E c,,,xX+m (co 0). Then,
M=0
+00
C[q] _ { (A + m)3 - (,\ + m)2} c,,,xa+m+1
M=0
t o0
(A + m + 1)C,,,xa+m = 0.
m=o
Hence,
((A + 1)co = 0,
1
(A+m+1)Cm = {(A+m-1)3-(A+m-1)2}C,,,_1 for m > 1.
Therefore,
2)2(m -3)
A = -1 and c,,, _ (m - c,,,_1 for m > 1.
m -
Thus,
A = -1, cl = -2co, c, ,, = 0 for m>2.
This implies that q = CO (x-1 - 2) is a solution of C[r7] = 0.
(ii) For k2 = 2, set n = exp[Ax-112](.
Then,
x_1 2)
b"['1] = exp[Ax-112] \a - [S] (n =0,1,2.... ).
Therefore, C[n] = 0 is equivalent to
3 )2_
ft Cb - Zx-1/2) - x I tS - -\X-112 (b - 2x-1/2) - 1 0.
Since
2
Ax-1/2
2x-1/21 - \x-112d + + 42x-1,
J
3
2x-1/2 =d3 - tax-1/252 + (x_1/2 4,\2x-I
/ + / d
- 18,\x-1/2
+
3,\2X-1 + 8A3x-3/21
,
it follows that
/ 2_
XId- 2x-1/2/3 - x I d - 2x-1/2 (d - x-1/2) - 1
= x63 - (xhI2 + x 62
+ (.x2 - 1 + 4Ax1/2) d
2 4
+1 - g31 x-1/2 - 1 + g-) + x1/2.

2
The Newton polygon of this operator has only one nonvertical side of slope 2 for
arbitrary A (cf. Figure 7-1). However, if A is determined by the equation
A \3
2 8 0,
then the Newton polygon has a horizontal side (cf. Figure 7-2).

Hence, we can find two formal solutions of equation £[77] = 0:
exp [2x-1 12] xv, 11 + x1 /2f, J,

exp [ - 2x-1/2]xl2[1 + x1/2f2J,
where p1 and p2 are constants and f1 and f2 are formal power series in C[[x1/2JJ.
8. GEVREY PROPERTY OF ASYMPTOTIC SOLUTIONS 441
XIII-8. Gevrey property of asymptotic solutions at an irregular singular

point
In this section we prove a result which is more precise than Theorem V-1-5
([Mai]). In §XIII-3, we stated an existence theorem of asymptotic solutions for a
given formal solution of an algebraic differential equation (cf. Theorem XIII-3-6).
If differential equation (XIII.3.3) has a formal solution, we can transform (XIII.3.3)
to the form
(XIII.8.1) f_[y] = x"'G(x,y,by,... b"-1y) (b = x;),

where
n
(XIII.8.2) G = F_ ah(x)bh,
h=o
and G(x, yo, yl, ... , y.-I) is a convergent power series in (x, yo, yl ... y,,-,) (cf.
(SS31 and [Mal2j). Here, it can be assumed without any loss of generality that
(i) ah (h = 0,1, ... , n) are convergent power series in x and a # 0,
00
(ii) differential equation (XIII.8.1) has a formal solution p(x) = E a,,,x"' E
m=0
CI[x]I,
(iii) M is an integer such that for any differential operator At of order not greater
than n, the two Newton polygons N(C - xMIC) and N(C) are identical (cf.
Definitions XIII-7-4 and XILI-7-5).
Using Theorem XIII-3-6, we can find
(a) a good covering {S1, $2, ... , SN } at x = 0,
(b) N solutions 01(x), 02(x), ... , ON (x) of (XIII.8.1) in S1, S2,... , SN, respec-
tively such that of are holomorphic and admit the formal solution p(x) as
their asymptotic expansions as x -+ 0 in St, respectively.
Set ui = 01 - 0t+1 on Se n Sjt+1. Then, ue are flat in the sense of Poincare
in sectors Se n S(+1, respectively, where Sv+1 = S1. Furthermore, if we define
differential operators IQ by
Kt E -(x,...
h
bh (tot + (1 - t)-Ol+l I.... )dt bh,
O<h<n-1 LJO 1
then
(C - x"'K t)[ut] = 0 on se n St+l (f = 1, 2, .... N).
If the Newton polygon N(,C) has only one nonvertical side of slope 0, then x = 0
is a regular singular point of C[iI] = 0. Therefore, in this case the formal solution
p(x) is convergent (cf. Theorem V-2-7). Let us assume that./V(C) has at least one
side of positive slope. In such a case, let
(XIII.8.3) 0 < kl < k2 < . . . < k9 < +00

be all of the positive slopes of the Newton polygon N(,C). Then, since N(1 C -xMJCe)
and N(G) are identical, we must have
Ixl-k]
lut(x)l : -yeXp(-A on Sr n Se+1
for a non-negative number -y and a positive number A, where k E {k1, k2 ... , ky}
(cf. Theorem XIII-7-7). Now, by virtue of Theorem XI-2-3, we obtain the following
theorem.
Theorem XIII-8-1. Under the assumptions given above, the formal solution p(x)
is a formal power series of Gevrey order 1 and, for each t, the solution 0e admits
p(x) as its asymptotic expansion of Gevrey order k as x -. 0 in S1, where k E
{k1ik2... ,kq}.
This theorem was originally prove in (Raml] for a linear system. For nonlinear
cases, see, for example, ISi17, §A.2.4, pp. 207-211J.
Remark XIII-8-2. In Exercise XI-14, we gave the definition of a k-summable
power series. As stated in Exercise XI-14, if a formal power series f E C((x]]
is k-summable in a direction argx = 0, there exists one and only one function
F E Al /k (Po, 0 - 2k
-e,6+ 2k + e) such that J(F] = f , where Po and a are
positive unmbers. This function F is called the sum of f in the direction arg x = 0.
If we use the idea of Corollary XIII-1-3 and Theorem XIII-6-7, we can prove the
following theorem concerning the k-summability of a formal solution of a nonlinear
system
(XIII.8.4) xk+1 dy = A(x)y" + xb(x, yj.

dx
Theorem XIII-8-3. Under the assumptions

(i) k is a positive integer,
(ii) A(x) is an n x n matrix whose entries are holomorphic in a neighborhood of
x = 0 and b(x, y-) is a C"-valued function whose entries are holomorphic in a
neighborhood of (x, yl = (0,
(iii) A(0) is invertible,
00
system (XIII.8.4) has one and only one formal solution y = p(x) xmpm and
m=1
p(x) is k-summable in any direction arg x = 0 except a finite number of values of 6.
Furthermore, the sum of p(x) in the direction arg x = 0 is a solution of (XIII.8.4).
To prove this theorem, it suffices to choose the good covering {S1, S2, ... , SN }
at x = 0 in the proof of Theorem XIII-8-1 so that opening of each of sectors
{S1, S2, ... , SN } is larger than k . We can prove a more general theorem.
EXERCISES XIII 443
Theorem XIII.8.4. Let a linear differential operator L = 6 - A(x) be given,

where 6 = , and A(x) is an n x n matrix whose entries are meromorphic in a
neighborhood of x = 0. Also, let (X111.8.3) be all the positive slopes of the Newton
polygon N(C) of the operator C. Assume that
(1) k1 ?
51
(2) C[ f] is meromorphic at x = 0 for a f (x) E C[[xiln.
Then, there exist a finite number of directions arg x = 0 (1= 1,2,... , p) such that
i f 0 34 Ot f o r I = 1 , 2, ... , p, there exist q formal power series j, (v = 1, 2, ... , q)
satisfying the following conditions:
(a) for each v, the power series f is in the direction argx = 0,
(b)f=fi+fz+...+fq.
A complete proof of this theorem is found in [BBRS]. In this case, f is said to

be {k1, k2 ... , kq }-multisummable in the direction arg x = 0. We can also prove
multisummability of formal solutions of a nonlinear system. For those informations,
see, for example, [Ram3], [Br), [RS21, and [Bal3j.
EXERCISES XIII
XIII-1. Using Observation XIII-6-5, diagonalize the following system:
dg _ x+ 5 x+ 8 where `yil
dx 3 -x+ 1 y l yz J
for large )x[.

XIII-2. Show that there is no rational function f (x) in x such that
(a"f)(r) + rzf(x) = x
Hint. One method is to show that the given equation has a unique power series
solution in X-1 which is divergent at x = oo. Another method is to observe that
any solution of this equation has no singularity in ]x] < +oo except possibly at
x = 0. Furthermore, if p(x) is a rational solution, then some inspection shows that
p(x) does not have any pole at x = 0. This implies that p(x) must be a polynomial.
But, we can easily see that this equation does not have any polynomial solution
(cf. {Si20J).
XIII-3. Find a cyclic vector for the differential operator £[ 3 = x;ii + 'A, with a
constant n x n matrix A of the form A = I At 0 1, where for each j = 1, 2, the
quantity Aj is an of x n, matrix of the form

0 1 0 0 ... 0
0 0 1 0 ... 0
A?=
0 0 0 0 ... 1
CYO, a 1,) a2,j C13,j ... ant -1j

with complex constants ah,j.
XIII-4. Find the Newton polygon and a complete set of linearly independent for-
mal solutions for each of the following three differential equations:
(a) x(52y + 45y - y = 0, (b) x2(52y + xby - y = 0, (c) xb2y + x6y - y = 0,
d
where b = x .
XIII-5. Find the Newton polygon of the following system:

1 1 0
x3 d = 0 x x
dx
10 0 x2
Hint. For the given system, a cyclic vector is (1, 0, 0). This implies that the trans-
formation
1 0 0
(T) X-2 x-2 0
X-4-2x-2 x-4 + x-3 - 2x-2 x-
changes the given system to
(E) (5 - 1)(x5 - 1)(x2(5- 1)u1 = 0, where b = x .
Note that transformation (T) is equivalent to u1 = yl, u2 = dyl, u3 = 52y1, and

the given system can be written in the form
1/2 = (x25 - 1/3 = (xb - 1)1/2, (5 - 1)1/3 = 0.
The Newton polygon of (E) has three sides with slopes 0, 1, and 2, respectively.
On the other hand, The standard form of the given system in the sense of
Theorem XIII-6-1 is
a(x) 0 0
3 du
X - = 0 b(x) 0 u,
dx
0 0 x2
where
a(x) = 1 + 0(x), b(x) = x + 0(x2).
Hence, this also shows that the Newton polygon of the given system has three sides
with slopes 0, 1 and 2, respectively.
EXERCISES XIII 445
XIII-6. Let A(x) be an n x n matrix whose entries are holomorphic and bounded
in a domain Ao = {x : 3x! < ro} and let j (x) be a Cn-valued function whose
entries are holomorphic and bounded in the domain Ao. Also, let Al, A2, ... , An
be eigenvalues of A(O). Assume that det A(0) # 0. Assume further that two real
numbers 01 and 82 satisfy the following conditions:
(1) 01 < 02,
(2) none of the quantities Aye-'ke (j = 1, 2, ... , n) are real and positive for a
positive integer k if 01 < 8 < 02,
(3) Ape-ike' and Aqe-'k93
are real and positive for some p and q.
Show that there exist one and only one solution f = fi(x) of the system x'`+1 ds =
A(x)f+ x f (x) such that the entries of fi(x) are holomorphic and admit asymptotic
expansions in powers of x as x - 0 in the sector S = {x : 0 < !xj < r0, 81 - <
2k
argx<82+2k
Hint. If we use Corollary XIII-1-3 at x = 0, it can be shown that for each 8 in the
interval B1 - < B < B2 + 7, a solution ¢'(x; 8) is found so that the entries of
2k
O(x; 8) are holomorphic and admit asymptotic expansions in powers of x as x - 0
'r
in the sectorial domain Sa = x : 0 < jxj < ro, I arg x - 81 < 2k + ea }, where Ee is
a sufficiently small positive constant depending on 8. If 10 - 8'l is sufficiently small,
then ¢(x; 0) = d(x; 8).
XIII-7.
(a) Show that j (x) _ (-1)m(m!)xr+L is a formal solution of
m=o
2 dy
(E) dx+y-x=0
and that f is not convergent except at x = 0.

(b) For a given direction 8, find a solution fe(x) of (E) such that fe(x) ^- j (x) as
x 0 in the direction arg x = 8.
(c) Calculate fe, (x) - fe, (x) for two given directions 81 and 02.
Hint. For Part (b), use the following steps:
Step 1. Apply Theorem XIII-1-2 to the given differential equation. To do this, we
must change x = 0 to t = oo by x = 1. Then, the differential equation becomes
du 1
= y - t . In this case, n = 1, r = 0, and the eigenvalue is ,u = 1. Set arg u = 0.
Then, the domain D(N, ry) (cf. (XI 11. 1. 15)) is
D(N,7) = {t:iti > N, I argt - 2gir! < 32

- f}
where q is an integer, N is a sufficiently large positive number, and -y is a sufficiently

small positive number. In terms of x, the domain D(N, y) becomes
S(N, -t) = {x:O < IxI < N, I arg x - 2pirI <

2 - 'Y}
where p = -q. Since there is no singular point of the given differential equation in
the domain 0 = {x: 0 < Ixi < oo}, we can conclude that, for each fixed integer p,
there exists a solution ¢p(x) of the given differential equation such that
(i) op(x) is analytic (but not single-valued) in !2,
(ii) -ip(x) j (x) as x 0 in the domain
37r
Dp = {x: 0 < xl < oc, argx - 2pl <
_ 2
Step 2. It is easy to see that f (x) = el/=J 2t-1e-1l`dt (x > 0) satisfies the given
0
differential equation and has the asymptotic property f (x) = j (x) as x 0. Since
¢p(x) - f (x) is a solution of the homogeneous differential equation x2dy + y = 0,
it follows s that ¢p(x) = f (x) + cell=, where c is a constant. From this, op(x) _
elI1 t-le-ll`dt follows for argx = 2pir.
Step 3. Using an argument similar to that of Step 2,

rt
4p(x) if 10 - 2p7ri <
21
fe(x) =
Op(x) + cge,/, if 2 < 0 - 2pir < 2 ,
is obtained, where co is an arbitrary constant.

Step 4. Note that
t-le-1'`dt = 27ri,
IC
where C is a counterclockwise oriented circle with the center at x = 0. Hence, using
analytic continuation of f (x), we obtain
fi(x) = ¢p(x) + 2pirie' for argx = 2pir.
XIII-8. Show that the following differential equation has a nontrivial convergent
power series solution:
3
2x
d
+ dX+
2
y = Q.
EXERCISES XIII 447
Hint.
Step 1: The given differential equation has three linearly independent formal so-
lutions
+oo +- +oo
1(x) = ell= 1+ amxm 2(x) = 1+ L. bmxm, and o3(x) = x+ amxm.
M=1 m=2 m=3
In fact, Ol can be found through some calculation with the Newton polygon. The
other two can be found by solving the equations
f ao + 2a2 = 0, al + 603 = 0,
(l a,, + (m + 1)m(m - 1)0,,,+1 + (m + 2)(m + 1)am+2 = 0 for m > 2
+00
for a formal solution E a,,,x"`.
m=o
Step 2: The given differential equation has three linearly independent actual solu-
3ir
tions such that e-'/=O1(x) ^ as x 0 in the sector j argx + 7r1 <
T,
and 02(x) fi(x) and 03(x) ^_- 3(x) as x -. 0 in the sector I argxI < 32
Step 3: In the direction arg x = -7r,
02(2e2x,) 03(xe2ai)
02(x) - = C201(x) and 03(x) - = C301 (-T)
for some constants c2 and C3. Then, c3 (x) - c2.3(x) is a convergent power series
solution of the given differential equation.
Remark:
(1) See [HIJ.
(2) This result was originally proved for a more general case in [Per1].
(3) There is another proof based on Exercises V-18 and V-19 (cf. [HSWI).
XIII-9. Consider a system of differential equations
du
xp+1
dx = F(x, y, u),
(E)
where p is a positive integer, x is a complex independent variable, y is a complex

parameter, it and F are n-dimensional vectors (i.e., E C'), and entries of F(x, y, u")
are holomorphic with respect to (x, y, u-) in a neighborhood of (x, y, u) = (0, 0, 0).
Assume that there exists a formal solution of system (E)
00
u = V, (X, y) _ yh+Gh(x),
h=0
where coefficients zlih(x) are R'-valued functions whose entries are holomorphic in
a neighborhood of x = 0. Assume also that O0(0) = 0 and det [(oo)]
o9p
54 0.
Show that i (x, y) is convergent in a neighborhood of (x, y) = (0, 0).

Hint. See ]Si14] and [Si21].

XIII-10. Consider a Pfaffian system
aaxu
= F(x, y, u,
(S)
J xl
y4+t ala = G(x, y, u-),
where p and q are positive integers, x and y are two complex independent variables,
iZ, F, and d are n-dimensional vectors (i.e., E C"), the entries of F and G are
holomorphic with respect to (x, y, 11) in a neighborhood of (x, y, u) _ (0, 0, and
system (S) is completely integrable, i.e., F and d satisfy the condition
yq+I
5F (x, y, v') + (x, y, u)G(x, y, u = xp+l (x, y, u) + 5 (x, y, U-)F(x, y, u).
041
Assume that F(0, 0, 0) = 0, d(0, 0, 0) = 0, det [(oo)J 36 0, and
det (0, 0, 0)J j4 0. Show that system (S) has one and only one solution
L
11 = (x, y) such that i (0,O) = 0 and that entries of y) are holomorphic

with respect to (x, y) in a neighborhood of (x, y) = (0, 0).
Hint. This is an application of Exercise XIII-9.
Step 1. Construct a formal power series solution
+00
(FS) 11 = '1G(x, y) = E yhrGh(x)
h=o
of the system
yq+1
ay = G (x y , ,
in such a way that coefficients t/ih(x) are holomorphic in a neighborhood of x = 0.

Step 2. For t _ l%i(x, y), note that
yl+9ay (i+ P F(x,y,11))

ax -
= yl+qxl+p 0211 y1+qaF - OF'yl+ga11
8i-ax ay su ay
= yi+qxl+p 8211 aF OF -
0;-ax Y
1+q
ay - &a
G(x,y,ur)
= xl+pyl+q 8211 x1+paG

- auF(x,y,u)
5y-ax TX
= x1+p ad ac 011
- x
+paG - ad
ax + au ax ax su F'(x, y, u)
(xl+P
ax - F(x,y,i)
811
EXERCISES XIII 449
Using this result, it can be shown that (FS) also satisfies the system
(E) xP+i a = Ax, y, u)
Step 3. Upon applying Exercise XIII-9 to (E), the convergence of t%i(x, y) is proved.
XIII-11. Complete the proof of Theorem XIII-7-6 by verifying rigorously state-
ments (a), (b), and (c) in the proof.
XIII-12. Show that the series
(FS)
E
+00 (3)h xexp [_x3/2}
y = P(x) = x-1/4 1 + >( -1)h
[ 54hh!r(h +) 3
d2-xZ
is a formal solution of the differential equaton -xy = 0, where r is the Gamma-
function.
d2-xZ
XIII-13. Show that the differential equation - xy= 0 has a unique solution
O(x) such that (1) b(x) is entire in x and (2) ¢(x)exp [3x3/21 admits the formal
series p(x) exp [x3I2J as its asymptotic exansion as x -oc in the sector I arg xj <
r, where p(x) is givenby (FS).
Remark. Ai(x) = 2(--X) is called the Airy function (cf. [AS, p. 446), [Wasl, pp.
v/Fr
124-1261, and [01, pp. 392-394]).
XIII-14. Using the same notations in Exercises XIII-12 and XIII-13, show that if
w =xpel(2ril
3 J then ¢(w'ix) and Q(wx) are two solutions of equation (S). Also,
(i) derive asymptotic expansions of ¢(w-lx) and m(wx),
(ii) show that {Q(x),y5(w-lx)}, {0(w-1x),O(wx)}, and {d(wx),Q(x)} are three
fundamental sets of solutions of (S),
(iii) show that if we set m(x) = c1O(w-1x)+c2y5(wx), then c2 = -w and
[ct O
1
3
w 11J
is equal to the 2 x 2 identity matrix,
(iv) using (iii), show that cl = -w-1.
XIII-15. Show that if O(x, A) is an eigenfunction of the eigenvalue problem
(EP) +00
d2 y
da l La
then
( i) Q(x) is entire in x and Q(x) exp is a polynomial,
ITJ
(ii) all negative odd integers are eigenvalues of (EP) and there is no other eigen-
value,
(iii) for every non-negative integer n, Hn(x) = (-1)ne=2 Un (e-y2) is a polyno-
r 2
mial, and -0n(x) = H.(x)exp I - 2 is an eigenfunction of (EP) for the
J
eigenvalue -(2n + 1), L
r+00 +00
(iv) On(x)/m(x)dx = 0 if n 9k m, and / On(x)2dx = 2nn!y'.
J 0o J o00
Remark. The polynomials Hn(x) are called the Hermit polynomials (cf. [AS, p.
775] and [01, p. 49]).
XIII-16. Construct Green's function of the boundary-value problem
- x2y = f (x),
j o0
y(x)2dx < +00. Show also that
(i)
j +00 r+00
00 J 00
G(x,)2dxd< +oo,
(ii) if f (x) is real-valued, f (x), f'(x), and f"(x) are continuous,

+
+00
f (x)2dx <
o0
+00 00
+00, and f{f"(x) - x2 f (x)}2dx < +oo, then the series Y 2nn!
00
(f, On)
xOn(x) converges to f(x) uniformly on the interval -oo < x < +oo, where
+00
On (x) are defined in Exercise XIII-15, and (f, g) _
Hint. See §VI-4.

j 00
f (x)g(x)dx.
n-1 h
XIII-17. Consider a differential operator C[y] _ dxn + 1: an-h(x) h , where
h=0
+00
aj(x) = ajmx-'n
E C[[x1]. Also, assume that aj,_m1 # 0 if a,(x) is not
m=-mj
equal to zero identically, while mj = -oo if aj(x) is zero identically. In the (X, Y)-
plane, consider the points Pj = (j, mj) (j = 1,... , n). Construct a convex polygon
II whose vertices are (0, 0), pi, p2, ... , p, such that each pk is one of those points
Pj, and that all other points Pj are situated below the polygon. Set po = (0, 0)
and Pk = (ak, Pk) (k = 1) ... , s), where an = n. Denote by pk the slope of the
segment k k+1 (k = 0, ... , s - 1). Then, Po > Pl > P2 > ... > p,-,. Assume
that pk>-1 (k<vo)and pk<-1 (k>vo). Show that
(i) the differential equation £[y] = 0 has n - at,,, linearly independent formal so-
Me
lutions of the form ye(x) = xe" E Req(x)(logx)q (e = a,,. + 1,... , n), where
q=o
Rrq(x) E C[[x-1]], b E C, and the Mt are non-negative integers,
(ii) if k < vo, then the differential equation £[y] = 0 has ak+1 - ak linearly
me
independent formal solutions of the form yr = e^<(x)x6, E Rrq(x)(log x)q (e =
q-0
EXERCISES XIII 451
1 + ok,... , nk+1), where At(x) = Arxl+1 b + terms of lower degree, At E C,

A, 96 0, bt E C, M, are non-negative integers, and Req E C[[x-lIP]], p being a
positive integer.
Hint. Compare the polygon in this problem with the Newton polygon at x=0
defined in §XIII-7.
XIII-18. Consider a differential operator C[yl = x9+1 dx + 1(x)y', where q is a

00
positive integer and f2(x) _ > xlf2, with Hi being n x n constant matrices. As in
t=o
§V-4, the operator C can be represented by the matrix
H1
f12
and J_ _ [Oq]
01
where H =
J,, +A I...
Here, 0, is the nr x oo zero matrix and I,,,, is the oc x oo identity matrix. Let
Qo=So+No and A=S+N
be the S-N decomposition of 0o and that of A in the sense of §V-3. Also, let
Po, P1, P2, ... , Pm, ... be n x n constants matrices and let AO be an n x n diagonal
matrix such that det Po 0 0 and that
A 1 0 0 0 0
0 A2 0 0 0
(1) SoPo = PoAo, Ao = 0 0 A3 0 0
0 0 0 0 An
and
(2) SP=PAo,
where
P1
P2
P0
P=
P.
Note that At are eigenvalues of So. Hence, A, are eigenvalues of 00. Show that
(a) the matrix S represents a multiplication operation: y -+ o(x)y for an n x n
matrix o(x) = So + O(x) whose entries are formal power series in x,
(b) the matrix N represents a differential operator Lo[yM = x4+1 fy + v(x)y for
an n x n matrix v(x) = No +O(x) whose entries are formal power series in x,
(c) C[f, = £o[yl + o(x)y and Go[o(x)yj = a(x)Go[yl,
+00
(d) if we set P(x) _ F xmP,,,, then P(x)-1o(x)P(x) = A0,
m=0
(e) if we set /C[uZ = P(x)-'C[P(x)uZ and Ko[u] = P(x)-1Go[P(x)ul, then K[u"j =
Ko[ul + Aou,
(f) if we set
Vii
du
Ko[i ] = xq+1 + vo(x)u, vo(x) = .
vn1
then
vjk(x) = 0 if .\.7 # Ak
(cf. [HKS]).
REFERENCES
[AS] M. Abramowitz and I. A. Stegun, Handbooks of Mathematical Functions, Dover,

New York, 1965.
[Ar] C. Arzelh, Sulle funzioni di line, Mem. R. Accad. Bologna (5), 5(1895), 225-244.
[As] G. Ascoli, Le curve limiti di una varieth data di curve, Mem. R. Accad. Lincei
(3), 18(1883/4), 521-586.
(AM] M. F. Atiyah and I. M. MacDonald, Introduction to Commutative Algebra,
Addison-Wesley, Reading, MA, 1969.
[Ball] W. Balser, Meromorphic transformation to Birkhoff standard form in dimension
three, J. Fac. Sci. Univ. Tokyo, Sect. IA, Math. 36(1989), 233-246.
[Bal2] W. Balser, Analytic transformation to Birkhoff standard form in dimension three,
Funk. Ekv., 33(1990), 59-67.
(Ba13] W. Balser, From Divergent Power Series to Analytic Functions, Lecture Notes
in Math. No. 1582, Springer-Verlag, 1994.
[BBRS] W. Balser, B. L. J. Braaksma, J.-P. Ramis, and Y. Sibuya, Multisuinmability of
formal power series solutions of linear ordinary differential equations, Asymptotic
Anal., 5(1991), 27-45.
[BJL] W. Balser, W. B. Jurkat, and D. A. Lutz, A general theory of invariants for
meromorphic differential equations, Funk. Ekv., 22(1979), 197-221.
(Bar] R. G. Bartle, The Elements of Real Analysts, John Wiley, New York, 1964.
[Bell] R. Bellman, On the asymptotic behavior of solutions of u" - (1 + f (t))u = 0,
Ann. Mat. Pure Appl., 31(1950), 83-91.
[Be12] R. Bellman, Stability Theory of Differential Equations, McGraw-Hill, New York,
1953.
[Be13] R. Bellman, Introduction to Matrix Analysis, SIAM, Philadelphia, 1995.
[Benl] I. Bendixson, Demonstration de 1'existence de l'integrale d'une equation aux
derivees partielles lineaire, Bull. Soc. Math. France, 24(1896), 220-225.
[Ben2] I. Bendixson, Sur les courbes definies par des equations differentielles, Acta
Math., 24(1901), 1-88.
[Bi] G. D. Birkhoff, Equivalent singular points of ordinary differential equations,
Math. Ann., 74(1913), 134-139; also in Collected Mathematical Papers, Vol. 1,
American Mathematical Society, Providence, RI, 1950, pp. 252-257.
[Bor] E. Borel, Leyons sur les serves divergentes, Gauthier-Villars, Paris, 1928.
[Bou] N. Bourbaki, Algebre II, Hermann, Paris, 1952.
[Br] B. L. J. Braaksma, Multisummability of formal power series solutions of non-
linear meromorphic differential equations, Ann. Inst. Fourier, Grenoble, 42-
43(1992), 517-540.
[Ca] C. Caratheodory, Vorlesungen fiber reelle F'unktionen, Teubner, Leipzig, 1927;
reprinted, Chelsea, New York, 1948.
[Ce] L. Cesari, Asymptotic Behavior and Stability Problems in Ordinary Differential
Equations. Academic Press, New York, 1963.
453
454 REFERENCES
(CK) K. Chiba and T. Kimura, On the asymptotic behavior of solutions of a system

of linear ordinary differential equations, Comment. Math. Univ. St. Pauli,
18(1970), 61-80.
[Cod] E. A. Coddington, An Introduction to Ordinary Differential Equations, Prentice-
Hall, Englewood Cliffs, NJ, 1961.
[CL) E. A. Coddington and N. Levinson, Theory of Ordinary Differential Equations,
McGraw-Hill, New York, 1955.
(Copl) W. A. Coppel, Stability and Asymptotic Behavior of Differential Equations,
Heath, Boston, 1965.
[Cop2] W. A. Coppel, Disconjugacy, Lecture Notes in Math. No. 220, Springer-Verlag,
New York, 1971.
[Cu) C. G. Cullen, Linear Algebra and Differential Equations, 2nd ed., PWS-Kent,
Boston, 1991.
(Dell P. Deligne, Equations differentielles d points singuliers reguliers, Lecture Notes
in Math. No. 163, Springer-Verlag, New York, 1970.
[Dev] A. Devinatz, The asymptotic nature of the solutions of certain systems of differ-
ential equations, Pacific J. Math., 15(1965), 75-83.
[DK] A. Devinatz and J. Kaplan, Asymptotic estimates for solutions of linear systems
of ordinary differential equations having multiple characteristic roots, Indiana
Univ. Math. J., 22(1972), 355-366.
(Du) H. Dulac, Points singuliers des equations differentielles, M6mor. Sci. Math.,
Fasc. 61, Gauthier-Villars, Paris, 1934.
(Dw) B. Dwork, Bessel functions as p-adic functions of the argument, Duke Math. J.,
41(1974) 711-738.
(El) M. S. P. Eastham, The asymptotic solution of linear differential systems, Math-
ematika, 32(1985), 131-138.
[E2) M. S. P. Eastham, The Asymptotic Solution of Linear Differential Systems, Ox-
ford Science Publications, Oxford, 1989.
[FL) L. Flatto and N. Levinson, Periodic solutions of singularly perturbed systems,
J. Rat. Mech. Anal., 4(1955), 943-950.
[Fl) G. Floquet, Sur les equations differentielles lineaires h coefficients periodiques,
Ann. Ecole Norm., Ser. 2, 12(1883), 47-89.
[Fu] L. Fuchs, Zur Theorie der linearen Differentialgleichungen mit veriinderlichen
Koeffizienten, J. Math., 66(1866), 121-160; 68(1868) 354-385.
[GelL) I. M. Gel'fand and B. M. Levitan, On the determination of a differential equation
from its spectral function, AMS 73nnsl., Ser. 2.1, Vol. 1, 1955, 253-304.
[GerL) R. Gerard and A. H. M. Levelt, Sur les connexion h singularites regulie res dans
le cas de plusieurs variables, Funk. Ekv., 19 (1976), 149-173.
(GerT] R. Gerard and H. Tahara, Singular nonlinear partial differential equations, As-
pects of Mathematics, Vol. E 28, Vieweg & Sohn Verlagsgesellschaft mbH, Braun-
schweig/Wiesbaden, 1996.
[Gi] H. Gingold, Almost diagonal systems in asymptotic integration, Proc. Edin.
Math. Soc., 28(1985), 143-158.
[GH] H. Gingold and P. F. Hsieh, Globally analytic triangularization of a matrix
function, Lin. Alg. Appl., 169(1992), 75-101.
(GHS] H. Gingold, P. F. Hsieh and Y. Sibuya, Globally analytic simplification and the
Levinson theorem, J. Math. Anal. Appl., 128 (1994), 269-286.
REFERENCES 455
[GrJ T. H. Gronwall, Note on the derivatives with respect to a parameter of the

solutions of a system of differential equations, Ann. Math. (2), 20(1919), 292-
296.
[Haa] A. Haar, Zur characteristikentheorie, Acta Sci. Math Szeged, 4(1928), 103-114.
[HabL] S. Haber and N. Levinson, A boundary value problem for a singularly perturbed
differential equation, Proc. Am. Math. Soc., 6(1955), 866-872.
(HarL11 W. A. Harris, Jr. and D. A. Lutz, On the asymptotic integration of linear
differential system, J. Math. Anal. Appl., 48(1974), 1-16.
[HarL2J W. A. Harris, Jr. and D. A. Lutz, Asymptotic integration of adiabatic oscillators,
J. Math. Anal. Appl., 51(1975), 76-93.
[HarL31 W. A. Harris, Jr. and D. A. Lutz, A unified theory of asymptotic integration,
J. Math. Anal. Appl., 57(1977), 571-586.
(HaSI] W. A. Harris, Jr. and Y. Sibuya, The reciprocals of solutions of linear ordinary
differential equations, Adv. Math., 58(1985), 119-132.
[HaS21 W. A. Harris, Jr. and Y. Sibuya, The n-th roots of solutions of linear ordinary
differential equations, Proc. Am. Math. Soc., 97(1986), 207-211.
(HaS31 W. A. Harris, Jr. and Y. Sibuya, Asymptotic behaviors of solutions of a system of
linear ordinary differential equations as t - oo, in Delay Differential Equations
and Dynamical Systems, Lecture Notes in Math. No. 1475, S. Busenberg and
M. Martelli eds., Springer-Verlag, New York, 1991, pp. 210-217,
[HSWJ W. A. Harris, Jr., Y. Sibuya and L. Weinberg, Holomorphic solutions of linear
differential systems at singular points, Arch. Rat. Mech. Anal., 35(1969), 245-
248.
[Harl] P. Hartman, Unrestricted solution fields of almost separable differential equa-
tions, 7hans. Am. Math. Soc., 63(1948), 560-580.
[Har2) P. Hartman, Ordinary Differential Equations, John Wiley, New York, 1964.
[HW1J P. Hartman and A. Wintner, On nonoscillatory linear differential equations, Am.
J. Math., 75(1953), 717-730.
(HW21 P. Hartman and A. Wintner, Asymptotic integration of linear ordinary differen-
tial equations, Am. J. Math., 77(1955), 45-87.
[Hi1P] E. Hille and R. S. Phillips, Functional Analysis and Semi-Groups, Colloquium
Pblications XXXI, American Mathematical Society, Providence, RI, 1957.
(HirSJ M. W. Hirsch and S. Smale, Differential Equations, Dynamical Systems and
Linear Algebra, Academic Press, New York, 1974.
[How] F. A. Howes, Effective characterization of the asymptotic behavior of solutions of
singularly perturbed boundary value problems, SIAM J. Appl. Math., 30(1976),
296-306.
[Hsl] P. F. Hsieh, On an analytic simplification of a system of linear ordinary dif-
ferential equations containing a parameter, Proc. Am. Math. Soc. 10(1968),
1201-1206.
(Hs2J P. F. Hsieh, Regular perturbation for a turning point problem, Funk. Ekv.,
12(1969), 155-179.
[HKS] P. F. Hsieh, M. Kohno and Y. Sibuya, Construction of a fundamental matrix
solution at a singular point of the first kind by means of SN decomposition of
matrices, Lin. Alg. Appl., 239(1996), 29-76.
456 REFERENCES
[HsS] P. F. Hsieh and Y. Sibuya, On the asymptotic integration of second order lin-
ear ordinary differential equations with polynomial coefficients, J. Math. Anal.
Appl., 16(1966), 84-103.
[HX1] P. F. Hsieh and F. Xie, Asymptotic diagonalization of a linear ordinary differ-
ential system, Kumamoto J. Math., 7(1994), 27-50.
[HX2] P. F. Hsieh and F. Xie, Asymptotic diagonalization of a system of linear ordinary
differential equations, Dynamics of Continu. Discre. Impula. Syst., 2(1996), 51-
74.
[HX3] P. F. Hsieh and F. Xie, On asymptotic diagonalization of linear ordinary differ-
ential equations, Dynamics of Continu. Discre. Impuls. Syst., 4(1998), 351-377.
(Huki] M. Hukuhara, Sur les systemes des equation differentielles ordinaires, Proc. Imp.
Acad. Japan, 4(1928), 448-449.
(Huk2] M. Hukuhara, Sur 1'ensemble des courbes integrals d'un systeme d'equations
differentielles ordinaires, Proc. Imp. Acad. Tokyo, 6(1930), 360-362; also [Huk8,
11-131.
[Huk3] M. Hukuhara, Sur les points singuliers des equations differentielles lineaires,
Domaine reel, J. Fac. Sci. Hokkaido Imp. Univ., 2(1934), 13-88; also (Huk8,
52-1301.
[Huk4) M. Hukuhara, Sur les points singuliers des equations differentielles lineaires, II,
J. Fac. Sci. Hokkaido Imp. Univ., 5(1937), 123-166; also [Huk8, 131-175].
(Huk5] M. Hukuhara, Sur un systeme de deux equations differentielles ordinaires non
lineaires h coefficients reels, Fac. Sci. Univ. Tokyo, 6(1951), 295-317.
(Huk6] M. Hukuhara, Sur la theorie des equations differentielles ordinaires, J. Fac. Sci.
Univ. Tokyo, Sec. 1, 7(1958), 483-510.
[Huk7] M. Hukuhara, Theoremes fondamentaux de la theorie des equations differentielles
ordinaires dans l'espace vectoriel topologique, J. Fac. Sci. Univ. Tokyo, Sec. I,
8(1959), 111-138.
(Huk8] M. Hukuhara, Selected Papers, Kyoritsu Shuppan Co., Ltd., Tokyo, 1997.
[HI] M. Hukuhara and M. Iwano, Etude de la convergence des solutions formelles
d'un systeme differentielles ordinaires lineaires, Funk. Ekv., 2(1959), 1-15; also
(Huk8, 423-437).
[HN1] M. Hukuhara and M. Nagumo, On a condition of stability for a differential
equation, Proc. Imp. Acad. Tokyo, 6(1930), 131-132; also (Na7, 136-1371.
[HN2] M. Hukuhara and M. Nagumo, Sur la stability des integrals d'un system
d'equations differentielles, Proc. Imp. Acad. Tokyo, 6(1930), 357-359; also
(Na7, 138-1401.
(HN3] M. Hukuhara and M. Nagumo, Un theoreme relatif h. 1'ensemble des courbes
intygrales d'un systeme differentielles ordinaires, Proc. Phys. -Math. Soc. Japan,
(3) 12(1930), 233-239; also [Na7, 141-147].
[Hum] J. E. Humphreys, Introduction to Lie Algebras and Representation Theory,
Springer-Verlag, New York, 1972.
[I] M. Iwano, Bounded solutions and stable domains of nonlinear ordinary differ-
ential equations, Analytic Theorey of Differential Equations, Lecture Notes in
Math. No. 183, P. F. Hsieh and A. W. J. Stoddart, eds., 59-127, Springer-
Verlag, New York, 1971.
REFERENCES 457
[IKSY] K. Iwasaki, H. Kimura, S. Shimomura and M. Yoshida, From Gauss to Painleve,

Aspects of Mathematics, Vol. E16, Vieweg & Sohn Verlagsgesellschaft mbH,
Braunschweig/Wiesbaden, 1991.
[Jai N. Jacobson, Basic Algebra I, Freeman and Co., New York, 1985.
[Jul W. B. Jurkat, Meromorphe Differentialgleichungen, Lecture Notes in Math., No.
637, Springer-Verlag, New York, 1978.
(JL) W. B. Jurkat and D. A. Lutz, On the order of solutions of analytic linear differ-
ential equations, Proc. London Math. Soc., 22(1971) 465-482.
[JLP] W. B. Jurkat, D. A. Lutz and A. Peyerimhoff, Birkhoff invariants and effective
calculations for meromorphic linear differential equations, Part 1, J. Math. Anal.
Appl., 53(1976), 438-470; Part II, Houston J. Math., 2(1976), 207-238.
[KS] J. Kato and Y. Sibuya, Catastrophic deformation of a flow and non-existence of
almost periodic solutions, J. Fac. Sci. Univ. Tokyo, 24(1977), 267-279.
[Ka] T. Kato, Perturbation Theory for Linear Operators, Springer-Verlag, New York,
1980.
[Kn] H. Kneser, Uber die Losungen eines Systems gewohnlicher Differentialgleichun-
gen das der Lipschitzschen Bedingung nicht geniigt, S.-B. Preuss. Akad. Wiss.
Phys.-Math. Kl. (1923), 171-174.
(Ko] H. Komatsu, On the index of ordinary differential operators, J. Fac. Sci. Univ.
Tokyo, Sec. IA, 18(1971), 379-398.
[LaL] J. LaSalle and S. Lefschetz, Stability by Liapunov's Direct Method with Applica-
tions, Academic Press, New York, 1961.
(LM1] E. B. Lee and L. Markus, Optimal Control for Nonlinear Processes, Arch. Rat.
Mech. Anal., 8(1961), 36-58.
[LM2] E. B. Lee and L. Markus, Foundations of Optimal Control Theory, John Wiley,
New York, 1967.
[Let] F. Lettenmeyer, Uber die an einer Unbestimmtheitsstelle regularen Losungen
eines Systems homogener linearen Differentialgleichungen, S.-B. Bayer. Akad.
Wiss. Miinchen Math.-nat. Abt., (1926), 287-307.
[LeL] J. J. Levin and N. Levinson, Singular perturbations of non-linear systems of
differential equations and an associated boundary layer equation, J. Rat. Mech.
Anal., 3(1954), 247-270.
[Levil] N. Levinson, The asymptotic nature of solutions of linear differential equations,
Duke Math. J., 15(1948), 111-126.
(Levi2] N. Levinson, Perturbations of discontinuous solutions of nonlinear systems of
differential equations, Acta Math., 82(1950), 71-105.
[Level A. H. M. Levelt, Jordan decomposition for a class of singular differential opera-
tors, Ark. Mat., 13(1975), 1-27.
[Lia] A. M. Liapounoff, Probleme general de la stabilite du mouvement, Ann. Fac.
Sci. Univ. Toulouse, 9(1909), 203-475; also Ann. Math. Studies, 17, Princeton,
1947.
[Lin] C.-H. Lin, Phragmen-Lindelof theorem in a cohomological form, Proc. Am.
Math. Soc., 89(1983), 589-597.
[Lindl] E. Lindelof, Sur l'application des methodes d'approximations successives a 1'etude
des integrales reeles des equations differentielles, J. Math., Ser. 4, 10(1894), 117-
128.
458 REFERENCES
(Lind2J E. Lindelof, Demonstration de quelques theorems sur les equations diffLrentielles,

J. Math., Ser. 5, 6(1900), 423-441.
[Lip] R. Lipschitz, Sur la posibilitd d'integrer completement un systeme done
d'6quations differentielles, Bull. Sci. Math. A8tron., 10(1876), 149-159.
[MW] W. Magnus and S. Winkler, Hill's Equation, Interscience, New York, 1966.
[Mahl K. Mahler, Lectures on Tl'nnscendental Numbers, Lecture Notes in Math. No.
546, Springer-Verlag, New York, 1976.
[Mail E. Maillet, Sur les series divergentes et les equations diffarentielles, Ann. Ecole
Norm., 20(1903), 487-517.
[Mall) B. Malgrange, Sur la reduction formelle des equations differentielles a singu-
larites irregulieres, Preprint, Grenoble, 1979.
[Ma12J B. Malgrange, Sur le theorem de Maillet, Asymptotic Anal., 2(1989), 1-4.
[Marl L. Markus, Exponentials in algebraic matrix groups, Adv. Math., 11(1973), 351-
367.
[Mil J. W. Milnor, Topology from the Differentiable Viewpoint, University of Virginia,
Charlottsville, 1976.
[Mol J. Moser, The order of a singularity in Fuchs' theory, Math. Zeitschr., 72(1960),
379-398.
[Mul F. E. Mullin, On the regular perturbation of the subdominant solution to second
order linear ordinary differential equations with polynomial coefficients, Funk.
Ekv., 11(1968), 1-38.
(Nall M. Nagumo, Eine hinreichende Bedingung fiir die Unitat der Losung von Differ-
entialgleichingen erster Ordung, Japan J. Math. 3, (1926), 107-112; also [Na7,
1-6].
[Na2] M. Nagumo, Eine hinreichende Bedingung fiir die Unitat der Losung von
gewohnlichen Differentialgleichungen n-ter Ordnuug, Japan. J. Math., 4(1927),
307-309; also [Na7, 33-35].
[Na3] M. Nagumo, Uber das Verfahren der sukzessive Approximationen zur Integration
gewohnlicher Differentialgleichung and die Eindeutigkeit ihrer Integrale, Japan.
J. Math., 7(1930) 143-160; also [Na7, 157-174].
[Na4J M. Nagumo, Uber die Differentialgleichung y" = f (x, y, y'), Proc. Phys. - Math.
Soc. Japan, 19(1937), 861-866; also [Na7, 222-227].
\
(Na5] M. Nagumo, Uber die Ungleichung ay > f (x, y, u, 8x 1, Japan. J. Math.,
15(1938), 51-56; also [Na7, 228-233]. \ ///
[Na6] M. Nagumo, Uber das Verhalten der Integrale von Ay" + f (x, y, y', A) = 0 fiir
A -+ 0, Proc. Phys.-Math. Soc. Japan, 21(1939), 529-534; also [Na7, 234-239].
[Na7) M. Nagumo, Mitio Nagumo Collected Papers, Springer-Verlag, New York, 1993.
[Nell F. Nevanlinna, Zur Theorie der Asymptotischen Potenzreihen, Ann. Acad. Sci.
Fennice Sarja, Ser. A, Nid. Tom. 12(1919), 1-81.
[Ne2J R. Nevanlinna, Analytic Functions, Springer-Verlag, New York, 1970.
[NU] A. F. Nikiforov and V. B. Uvarov, Special Functions of Mathematical Physics,
Birkhauser, Boston, 1988.
[O1] F. W. J. Olver, Asymptotics and Special Functions, Academic Press, New York,
1974.
REFERENCES 459
(O'M] R. E. O'Malley, Introduction to Singular Perturbations, Academic Press, New

York, 1974.
(Os] W. F. Osgood, Bewise der Existenz einer Ldsung einer Differentialgleichung
dy/dx = f (x, y) ohne Heinzunahme der Cauchy-Lipschizschen Bedingung,
Monat. Math.-Phys., 9(1898), 331-345.
[Pa] B. P. Palka, An Introduction to Complex Function Theory, Springer-Verlag, New
York, 1990.
[Peal] G. Peano, Demonstration de l'integrabilite des equations differentielles ordi-
naires, Math. Ann., 37(1890), 182-228.
[Peat] G. Peano, Generalith sulle equazioni differenziali ordinare, Atti R. Accad. Sci.
71ino, 33(1897), 9-18.
(Perl) O. Perron, Uber diejenigen Integrale linearer Differentialgleichungen, welche sich
an einer Unbestimmtheitsstelle bestimmt verhalten, Math. Ann., 70(1911), 1-32.
(Per2] O. Perron, Ein neue existenzbeweis fur die Integrale der Differentialgleichung
y' = f (x, y), Math. Ann., 76(1915), 471-484.
[Per3] O. Perron, Uber ein vermeintliches Stabilitatskriterium, Gott. Nachr. (1930),
128-129.
[Pi] E. Picard, Memoire sur la theorie des equations aux derivees partielles et la
methode des approximations successives, J. Math., ser 4, 6(1890), 145-210.
(Poil] H. Poincare, Sur les courbes definies par les equations differentielles, Oeuvres de
Henri Poincare, Vol. 1, Gauthier-Villars, Paris, 1928.
(Poi2] H. Poincare, Les methodes nouvelles de la mecanique celeste, Vol. II,
Gauthier-Villars, Paris, 1893.
[Pop] J. Popken, Uber arithmetische Eigenschaften analytischer FLnktionen, Disserta-
tion, Groningen, 1935.
[Rab] A. L. Rabenstein, Elementary Differential Equations with Linear Algebra, 4th
ed., Harcourt Brace Javanovich, New York, 1991.
(Rams] J.-P. Ramis, Devissage Gevrey, Asterisque, 59/60(1978), 173-204.
[Ram2] J.-P. Ramis, Les series k-sommables et leurs applications, Complex Analysis,
Microlocal Calculus and Relativistic Quantum Theory, Lecture Notes in Physics
No. 126, Springer-Verlag, New York, 1980, pp. 178-199.
(Ram3] J.-P. Ramis, Series divergentes et theories asymptotiques, Soc. Math. France,
Panoramas Syntheses, 121(1993), 1-74.
[RS1] J.-P. Ramis and Y. Sibuya, Hukuhara domains and fundamental existence and
uniqueness theorems for asymptotic solutions of Gevrey type, Asymptotic Anal.,
2(1989), 39-94.
[RS2] J.-P. Ramis and Y. Sibuya, A new proof of multisummability of formal solutions
of nonlinear meromorphic differential equations, Ann. Inst. Founer, Grenoble,
44(1994), 811-848.
(Ri] J. F. Pitt, On the derivatives of a function at a point, Ann. Math., 18(1916),
18-23.
[SC] G. Sansone and R. Conti, Nonlinear Differential Equations, MacMillan, New
York, 1961.
[Sail T. Saito, Introduction to Analytical Mechanics, Shibundo, Tokyo, 1964 (in Japan-
ese).
460 REFERENCES
[Sat] Y. Sato, Periodic solutions of the equation x" + f (x)x' + g(x) = 0 without the
hypothesis xg(x) > 0 for [xI > 0, Comment. Math. Univ. St. Pauli, 20(1971),
127-145.
(Sc] L. Schwartz, Theorie des Distributions, 3rd ed., Hermann, Paris, 1966.
[Sil] Y. Sibuya, Sur un systcme d'6quations differentielles ordinaires lnnaires h coef-
ficients periodiques et contenants des parami tres, J. Fac. Sci. Univ. Tokyo, I,
7(1954), 229-241.
[Si2] Y. Sibuya, Nonlinear ordinary differential equations with periodic coefficients,
Punk. Ekv., 1(1958), 77-132.
[Si3] Y. Sibuya, Sur reduction analytique d'un systeme d'i quations differentielles or-
dinaires lnnaires contenant un parametre, J. Fac. Sci. Univ. Tokyo, 7(1958),
527-540.
[Si4] Y. Sibuya, Notes on real matrices and linear dynamical systems with periodic
coefficients, J. Math. Anal. Appl., 1(1960), 363-372.
[Si5] Y. Sibuya, On perturbations of discontinuous solutions of ordinary differential
equations, Natural Sci. Rep. Ochanomizu Univ., 11(1960), 1-18.
[Si6] Y. Sibuya, On nonlinear ordinary differential equations containing a parameter,
J. Math. Mech., 9(1960), 369-398.
[Si7] Y. Sibuya, Simplification of a system of linear ordinary differential equations
about a singular point, Funk. Ekv., 4(1962), 29-56.
[Si8] Y. Sibuya, On the convergence of formal solutions of systems of linear ordinary
differential equations containing a parameter, MRC Tech. Sum. Report, No.
511, 1964.
(Si9] Y. Sibuya, A block-diagonalization theorem for systems of linear ordinary differ-
ential equations and its applications, J. SIAM Anal. Math., 14(1966), 468--475.
[Si10] Y. Sibuya, Perturbation of linear ordinary differential equations at irregular sin-
gular points, Funk. Ekv., 11(1968), 235-246.
[Sill] Y. Sibuya, Perturbation at an irregular singular point, Japan-U. S. Seminar on
Ordinary Differential and Functional Equations, Lecture Notes in Math. No.
243, M. Urabe, ed., Springer-Verlag, New York, 1971, pp. 148-168.
(Si12] Y. Sibuya, Uniform Simplification in a Full Neighborhood of a Thansition Point,
Memoirs of American Mathematical Society, No. 149, Providence, RI, 1974.
(Sil3] Y. Sibuya, Global Theory of a Second Order Linear Ordinary Differential Equa-
tion with a Polynomial Coefficient, Math. Studies 18, North-Holland, Amster-
dam, 1975.
[Si14] Y. Sibuya, Convergence of formal power series solutions of a system of nonlinear
differential equations at an irregular singular point, Lecture Notes in Math. No.
810, Springer-Verlag, New York, 1980, pp. 135-142.
[Si15] Y. Sibuya, A theorem concerning uniform simplification at a transition point
and the problem of resonance, SIAM J. Math. Anal., 12 (1981), 653-668.
[Si16] Y. Sibuya, Differential modules and theorem of Hukuhara-Turrittin, Proc. of
International Conf. on Nonlinear Phenomena in Math. Sci., Academic Press,
New York,1982, pp. 927-937.
[Si17] Y. Sibuya, Linear Differential Equations in the Complex Domain: Problems of
Analytic Continuation, Translations of Math. Monographs 82, American Math-
ematical Society, Providence, RI, 1990.
REFERENCES 461
[Si18] Y. Sibuya, Gevrey asymptotics, Ordinary and Delay Differential Equations, J.

Wiener and J. K. Hale, ed., Pitman Research Notes in Math. Series 272, Pitman,
London, 1991, pp. 207-217.
[Si19] Y. Sibuya, The Gevrey asypmtotics and exact asymptotics, Trends and Develop-
ments in Ordinary Differential Equations, Y. Alavi and P. F. Hsieh, eds., World
Scientific, Singapore, 1993, pp. 325-335.
(Si20] Y. Sibuya, A remark on Bessel functions, Differential Equations, Dynamical
Systems and Control Science, W. N. Everitt and B. Lee, eds., Lecture Notes in
Pure and Applied Math. No. 152, Marcel Dekker, New York, 1994, pp. 301-306.
[Si21] Y. Sibuya, Convergence of formal solutions of meromorphic differential equations
containing parameters, Funk. Ekv., 37(1994), 395-400.
[Si22] Y. Sibuya, The Gevrey asymptotics in case of singular perturbations, Preprint,
Univ. of Minnesota, 1997.
[SS1] Y. Sibuya and S. Sperber, Some new results on power series solutions of algebraic
differential equations, Proc. of Advanced Seminar on Singular Perturbations and
Asymptotics, May 1980, MRC, Univ. of Wisconsin-Madison, Academic Press,
New York, 1980, pp. 379-404.
[SS2] Y. Sibuya and S. Sperber, Arithmetic properties of power series solutions of
algebraic differential equations, Ann. Math., 113(1981), 111-157.
(SS3] Y. Sibuya and S. Sperber, Power Series Solutions of Algebraic Differential Equa-
tions, Publication I.R.M.A. et Departement de Mathematique, Univ. Louis Pas-
teur, Strasbourg, France, 1982.
[St] W. Sternberg, Uber die asymptotische Integration von Differentialgleichungen,
Math. Ann., 81(1920), 119-186.
[Sz] G. Szego, Orthogonal Polynomials, Colloquium Publications XXIII, American
Mathematical Society, Providence, RI, 1939.
[TD] S. Tanaka and E. Date, KdV Equation: An Introduction to Nonlinear Mathe-
matical Physics, Kinokuniya, Tokyo, 1979 (in Japanese).
[1111] H. L. Turrittin, Convergent solutions of ordinary linear homogeneous differen-
tial equations in the neighborhood of an irregular singular point, Acta Math.,
93(1955), 27-66.
[Tu2] H. L. Turrittin, Reduction of ordinary differential equations to the Birkhoff
canonical form, Trans. Am. Math. Soc., 107(1963), 485-507.
[U] M. Urabe, Geometric study of nonlinear autonomous oscillations, Funk. Ekv.,
1(1958), 1-83.
[Wasl] W. Wasow, Asymptotic Expansions for Ordinary Differential Equations, Dover,
New York, 1987. (Also John Wiley, New York, 1965.)
[Was2] W. Wasow, Linear Turning Point Theory, Applied Math. Sci. 54, Springer-
Verlag, New York, 1985.
[Wat] G. N. Watson, A theory of asymptotic series, Phil. Trans. Roy. Soc. of London
(Ser. A), 211(1911), 279-313.
[We] H. Weyl, The Classical Groups, Princeton University Press, Princeton, NJ, 1946.
[WhW] E. T. Whittaker and G. N. Watson, A Course of Modern Analysis, Cambridge
University Press, Cambridge, 1927.
[Z] A. Zygmund, Trigonometric Series, Vol. 1, 2nd ed., Cambridge University Press,
Cambridge, 1959.
INDEX
,A.(r,a,b), 353 Asymptotic solution, 374

Abel's formula, 80 Asymptotic stability, 236, 300
Abramowitz, M., 141, 449, 450 a sufficient condition for, 241
Adjoint equation, 79 Atiyah, M. F., 109
Airy function, 449 Autonomous system, 279
Algebraic differential equation, 109, 353
Analytic differential equation, 20 Balser, W., 371, 391, 401, 435, 443
Analytic simplification of a system, Baire's Theorem, 303
at a singular point, 421 Banach space, 25, 26, 103, 366, 394
with a parameter, 391 differential equation in, 25, 394
Approximate solutions, c-, 10 Bartle, R. G., 303
Approximations, successive, 3, 20, 21, Basis of the image space, 73
378, 418 Bellman, R., 71, 144, 148, 197
Arzelh, C., 9 Bendixson, 1., 27, 279, 291, 304
Arzelb-Ascoli Lemma, 9, 13, 30 Bendixson center, 261
Ascoli, G., 9 Bessel function, 125
Associated homogeneous equation, 97, 99 Bessel inequality, 162
Asymptotics, Gevrey, 353 Birkhoff, G. D., 372, 390
Poincar6, 342 Block-diagonalization theorem, 190, 204,
Asymptotic behavior of solutions, 197 380, 422, 423
of a linear second-order Block solution, 107
equation, 226 Blowup of solutions, local, 17
of linear systems, 209, 213, 219 Borel, E., 349, 356
Asymptotic (series) expansion, 343, 353 Borel-Ritt Theorem, 349, 356
definition, 343 Boundary-value problem, 144
differentiation, 347 of the second-order equation, 305
integration, 347 Sturm-Liouville, 148
inverse of, 346 Bounded set of functions, 9
of a holomorphic function, 345 Bourbaki, N., 69
of Ei(z), 352 Braaksma, B. L. J., 359, 355, 443
of Gevrey order s, 353 Branch point, 44
of Log(o(z)), 352
of uniformly convergent sequences, 348 C, field of complex numbers, 20
product, 344 C", set of all n-column vectors in C, 20
sum, 344 C[xJ, set of all polynomials, 109
Taylor's series as, 345 C{x}, set of all convergent power
uniform with respect to a parameter, series, 109
343, 357, 361 C[[xJJ, set of all formal power series, 109
of Gevrey order s, 357 C{x}", set of all convergent power
uniqueness, 343 series with coefficients in C", 110
Asymptotic reduction of a linear C[[xfJ", set of all formal power series
system with singularity, 405 with coefficients in C", 110
of a singularly perturbed linear C[[xJJ set of all power series of Gevrey
system, 381 order s, 353
462
INDEX 463
C-algebra, 70 in S-N decomposition, 74, 119, 122

Cayley-Hamilton Theorem, 71, 220 Diagonalization theorem, Levinson's, 213
Canonical transformation of a Differentiability with respect to initial
Hamiltonian system, 92, 94 values, 35
Cauchy, A. L., 1 Differential operator C, 120, 151, 169,
Cauchy-Euler differential equation, 140 424, 437, 443
Caratheodory, C., 15 S-N decomposition, 122
Center, 255, 257, 261, 272 normal form, 121
perturbation of, 271 Dirac delta function, 8
Cesari, L., 197 Distribution, of L. Schwartz, 8
Characteristic exponents of an of eigenvalues, 159
equation, 89, 203 Dulac, H., 235, 251
Characteristic polynomial of a matrix, Dwork, B., 8
71, 72
of a function of a matrix f(A), 84, 85 E(s, A), 366
partial fraction decomposition of the E(s, A)", 367
inverse of, 72, 76 Eastham, M. S. P., 197, 217
Chevalley, Jordan-, decomposition, 69 Eigenfunction, 162
Chiba, K., 197 Eigenvalues of a matrix, 70, 71
Classification of singularity of a Eigenvalue-problem, 153, 154, 156
homogeneous linear system, 132 of a boundary-value problem, 153, 186
Coddington, E. A.,14, 28, 36, 58, 65, 69, Eigenvectors of a matrix, 70
108, 137, 138, 144, 148, 153, 162, Elliptic sectorial domain, 296, 302
191, 197, 225, 233, 246, 266 274, Equicontinuity, 9
281, 290, 293, 298, 313 equicontinuous set, 47, 55, 63
Commutative algebra, 109 Euler, Cauchy-, differential equation, 140
Commutative differential algebra, 109, Existence, 1, 12
362 of solution of an initial-value problem,
Commutative differential module, 110 1, 3, 12, 21
Comparison theorem, 58, 144 with Lipschitz condition, 3
Complexification, 252 without Lipschitz condition, 12
Contact point, exterior (interior), 295 of solution of a boundary-value
Conti, R., 15, 235, 279 problem, 148
Continuity of a solution, 17, 29 of S-N decomposition of a matrix, 74
with respect to the initial point and Existence theorem for a nonlinear
initial condition, 29, 31 system at a singular point, 405
with respect to a parameter, 31 Existence theorem for a singularly
Convergence of formal solution, 113, 118 perturbed nonlinear system, 420
Coppel, W. A., 144, 148, 197 cxp[Al, exponential of a matrix A, 80
Covering, good, 354 exp[tAl with S-N decomposition, 82
Cullen, C. G., 71 Extension of a solution, 16
Cyclic vector, 424, 426 of a nonhomogeneous equation, 17
Date, E., 144 Finite-zone, potentials, 189

Deligne, P., 403, 424 First variation of an equation, 284
Denseness of diagonalizable matrices, 70 Flat of Gevrey order s, 342, 367, 385
Dependence on data, 28, Flatto, L., 341
Devinatz, A., 197 Floquet, G., Theorem, 69, 80, 225
Diagonalizable matrix, 70 Formal power series, 109, 343, 349
464 INDEX
Formal power series (cont.) Hamiltonian system with periodic

as an asymptotic series, 350 coefficients, 90
of Gevrey order s, 353, 355 Harr inequality, 27
with a parameter as an asymptotic Harr's uniqueness theorem, 27
series, 351 Harris, W. A., Jr., 142, 143, 197, 433, 447
Formal solution, 109, 342 Hartman, P., 14, 27, 28, 39, 58, 69, 144,
convergence, 115, 118 148, 153, 162, 197, 246, 281, 293,
Taylor's series as, 345 298
uniqueness, 115 Heaviside function, 8
Fourier series, 165, 166, 167 Hermit polynomials, 450
Fuchs, L., 137 Higher order scalar equations, 98
Fundamental matrix solution, 78, 105, Hille, E., 103
125, 122 Hirsch, M. W., 69, 251
of a system with constant coefficient, Holomorphic function asymptotic to a
81 formal series, 349, 356
with periodic coefficient, 87, 88 Homogeneous systems, 78
of an equation at the singularity of Homomorphism of differential algebras,
the first kind, 125 359, 362, 365
of an equation at the singularity of Howes, F. W., 304, 339
the second kind, 136, 432, 433 Hsieh, P. F., 78, 104, 108, 130, 197, 217,
Fundamental theorem of existence and 231, 372, 424, 436, 452
uniqueness, 3, 12 Hukuhara, M, 25, 26, 50, 108, 131, 197,
without Lipschitz condition, 8 212, 218, 261, 403, 428, 447
Fundamental set of linearly Hukuhara-Nagumo condition, 218
independent solutions, 78 for bounded solution of a second
normal, 202 order equation, 212
for bounded solution of a system, 218
Gel'fand, 1. M., 181 Hukuhara-Turrittin theorem, 428
Gel'fand-Levitan integral equation, 181 Humphreys, J. E., 69
General solution, 7, 105 Hyperbolic sectorial domain, 296, 302
Gerard, R., 108, 112, 125 Hypergeometric series, 138
Gevrey asymptotics, 353
Gevrey order 8, 356, 367 Improper node, 252, 254
asymptotic expansion of, 353 perturbation of, 261
flat of, 342, 367, 385 stable, 252, 253, 256
formal power series, 354 unstable, 252, 253
uniformly on a domain, 357, 361 Independent solutions, 78
Gevrey property of asymptotic Index of isolated stationary point, 293
solutions at an irregular singularity, Index of a Jordan curve, 293
441 Indicial polynomial, 112
solutions in a parameter, 385 Infinite-dimensional matrix, 118
Gingold, H., 104, 197, 217, 231 Infinite-dimensional vector, 120, 393
Global properties of solutions, 15 Infinite product of analytic function, 196
GL(n, C), general linear group, 69 Initial-value problem, complex
Good covering, 354 variable, 21
Green's function, 150, 154 partial derivative of a solution as a
Gronwall, T. If., Lemma, 3, 398 solution of, 33
Initial-value problem, real variable, 1, 28,
Haber, S., 304 32, 39
INDEX 465
nonhomogeneous equation, 96 Levinson, N, 14, 28, 36. 65, 69, 108, 137,
nonlinear, 1, 16, 32, 39 138, 144, 153, 162, 191, 197, 225,
second order equation, 324, 333 233, 246, 266, 274, 281, 290, 293,
Inner product, 151 298, 304, 340, 341
Instability, 243 Levinson's diagonalization theorem, 21:3
Instability region, 189 Levitan, B. M., 181
Integral inequality, 377, 418 Levitan, Gel'fand-, integral equation, 181
Invariant set, 280, 299 Liapounoff, A., 197
Irregular singular point, 134, 441 Liapounoff function, 239, 309, 311
Iwano, M., 447 Liapounoff's direct method, 281
Iwasaki, 1., 138 Liapounoff's type number, 198
calculation of, 203
J, map of A,(r,a,b) to C[[x]] 353 multiplicities, 202
one-to-one, 359 of a function, 198
onto, 356 of a system at t = oc, 201, 204
Jacobson, N., 69 of a solution, 199
Jordan-Chevalley decomposition, 69 properties, 198
Jordan canonical form, 421, 426 Lie algebra, 90
Jordan curve, index of, 293 Limit cycle, 292
Jost solution, 168, 180, 194 Limit invariant set, 280
Jurkat, W., 136, 391, 401, 402, 435 Lin, C.-H., 370
Lindelbf, E., 1, 28, 359
K, the field of fractions of C[[x[], 424 Lipschitz condition, 3, 43, 64
K[u'J, differential operator, 123 constant, 3
Kaplan, J., 197 sufficient condition for, 5
Kato, J., 44, 66 existence without, 8
Kato, T., 103 Local blowup of solutions, 17
Kimura, II., 138 Logarithm of a matrix, 86
Kimura, T., 197 log[ 1 + ltf ], 86, 88
Kneser, H., theorem, 41, 47, 305 log[8(w)[, 88
Kohno, M., 78, 108, 130, 452 log[8(w)2], 88, 94
Komatsu, H., 8 log[s], 86, 88
log[S2], 88
C, differential operator, 120, 151, 169, Lutz, D. A., 136, 197, 391, 401, 402, 435
424, 437, 443
normal form, 121 M,,(C), set of all n x n matrices
calculation of, 130 in C, 69
C+, limit-invariant set, 280 MacDonald, 1. C., 109
Laplace transform, inverse, 103 Magnus, W., 196
LaSalle, J., 279, 281 Mahler, K., 112
Lebesgue-integrable function, 15 Maillet, E., 112, 353, 441
Lee, E. B., 68, 106 Malgrange, B., 438
Lefschetz, S., 279, 281 Manifold, 45
Legendre, equation, 141 differential equation on, 45
polynomials, 141, 192 stable, 243
Leroy transform, incomplete, 354 unstable, 246
Lettenmeyer, F., 142 Markus, L., 68, 69, 96, 106
Levelt, A. H. M., 108, 125, 435 Matrix, 69
Levin, J. J., 340 functions of a matrix f (A), 84
466 INDEX
Matrix (cont.) Nonlinear initial-value problem, 17, 21

diagonalizable, 70 28, 32, 41
in S-N decomposition, 74 Nonstationary point, 291, 292
infinite-dimensional, 118 Nonuniqueness of solution of an
nilpotent, 71 initial-value problem, 41
in S-N decomposition, 74 Norm, (in C[[x]], 366
norm of a matrix A, 11A11, 69 in C[[x]l", 367
semisimple, 70 in CI[x]j 366
symplectic, 93, 95 of a continuous function, 162
upper-triangular, 70 in E(s, A)", 367
Maximal interval, 16 of a matrix A, 1PAIl, 69
Maximal solution, 52, 54, 56, 66 of a vector, 1, 345
Minimal solution, 52, 54, 66 of an infinite-dimensional vector, 394
Milnor, J. W., 298 Normal form of a differential
Moser, J., 135 operator, 121
Mullen, F. E., 232 Normal fundamental set, 202
Multiplicity of an eigenvalue of a Null space of a homomorphism, 366
matrix, 72, 74
Multiplicity of Liapounoff's type Olver, F. W. J., 138, 141, 449, 450
number, 201 O'Malley, R. E., 304
of a system of equations with constant Orbit, 251, 279
coefficients, 202 periodic, 304, 318, 319
of a system of equations with periodic of van der Pol equation, 313, 318
coefficients, 202 Orbitally asymptotical stability, 283
Multipliers, of an autonomous Orbital stability, 283
system, 284 Orthogonal sequence, 166
of a system with periodic Osgood, W. F., 63
coefficients, 89, 183, 203 Osgood condition for uniqueness of
solution, 63, 65
of periodic orbits, 318, 319
Palka, B. P., 196
N(1), Newton polygon of G, 437, 439
Parabolic sectorial domain, 296
Nagumo, M., 27, 50, 197, 212, 218, 304, Parseval inequality, 166
306, 327, 330, 333, 334, 342
Partial differential equation, 39
Nagumo condition for uniqueness of Peano, G., 1, 28
solution, 64, 65
Periodic coefficients, system with, 87
Nevanlinna, F., 342
Periodic orbit, 291, 297, 318, 319
Nevanlinna, R., 359 of van der Pol equation, 313, 318
Newton polygon, 437, 439 Periodic potentials, 183
Nikiforov, A. F., 141 Periodic solution, 184, 288
Nilpotent matrix, 71 of van der Pol equation, 313, 318
in S-N decomposition, 74 Perron, 0., 57, 212, 443
Node, 253, 254 Perturbation, of initial-value problem, 43
improper, 253, 255 of a center, 271
proper, 253, 256 of a proper node, 266
Nonhomogeneous equation, 17 of a saddle point, 263
Nonhomogeneous initial-value problem, of a spiral point, 270
96 of an improper node, 263
Nonlinear equation, 18, 372, 384 Peyerimhoff, A., 391, 401, 402
INDEX 467
Pfaffian system, 448 Schwartz, L., 8

Phase plane, 251 Second-order equation, boundary-value
Phase portrait of orbits, 251, 302 problem, 304, 308
Phillips, R. S., 103 Sectorial region, 296
Phragmen-Lindelof theorem, 359, 370 elliptic, 296, 302
Picard, E., 1 hyperbolic, 296, 302
Poincarc, H., 279, 291, 304 parabolic, 296
Poincarc asymptotics, 342 Self- adjoi nt ness, 151, 155
Poincarc-Bendixson Theorem, 291, 293 Semisimple matrix, 70
Poincarc's criterion, 300 Shearing transformation, 209
Popken, J., 112 Shimomura, S., 138
Potentials, 175, 178 Sibuya, Y., 44, 66, 69, 78, 96, 108, 112,
finite-zone, 189, 130, 131, 136, 142, 143, 197, 217,
reflectionless, 175, 177, 178, 181 227, 232, 304, 342, 360, 363, 365,
periodic, 183 369, 371, 372, 381, 382, 390, 391,
Projection, P,(A), 72 392, 399, 419, 420, 424, 436, 438,
properties, 73 447, 448, 452
Proper node, 253, 254 Singular solution, 42
perturbation of, 266 Singular perturbation of van der Pol
stable, 253, 255, 267 equations, 330
unstable, 253 Singularity of a linear homogeneous
system, 132, 134
R, real line, I of the first kind, 113, 132, 136
R", set of all n-column vectors in R, 1 of the second kind, 132, 134, 136, 403
Rabenstein, A. L., 36, 70, 101 n-th-order linear equation, 436
Ramis, J.-P., 342, 360, 363, 371, 386, 420, regular, 131, 133, 134, 394
443 irregular, of order k, 134, 441
Reflection coefficient, 175 Smale, S., 69, 251
Reflectionless potentials, 177, 178, 181 Solution, asymptotic, 374
construction, 178 asymptotically stable, 236, 241
Regular singular point, 131, 133, 134, 394 independent, 78
Ritt, J. F., 349, 356 fundamental matrix, 78, 105, 125, 222
periodic, 184, 288
S,,, set of all diagonalizable matrices, 70 of van der Pol equation, 313, 318
S-N decomposition, 74, 75 singular, 42
existence, 74 trivial, 236, 243
of a differential operator, 121 uniqueness, 1, 3, 21
of a function of a matrix f(A), 84 Solution curves, 42, 49
of a matrix for a periodic equation, 88 Spiral point, 254, 255
of infinite order, 118, 120 perturbation of, 270
of a real matrix, 75 stable, 254, 256, 270, 272
uniqueness, 75 unstable, 254, 272
Saddle point, 252, 255, 256 Sperber, S., 112
perturbation of, 261 Stability, 235
Sansone, G., 15, 235, 279 Stability region, 189
Saito, T., 274 Stationary point, 280, 291, 312
Sato, Y., 336 isolated, 294
Scalar equations, higher order, 98 Stable manifolds, 243
Scattering data, 172, 175 analytic structure, 246
468 INDEX
Stegun, 1. A., 141, 449, 450 of solution of an initial-value

Sternberg, W., 403, 438 problem, 1, 3, 21, 79
Stirling formula, 358, 364 Osgood condition, 63, 65
Sturm-Liouville problem, 148 sufficient conditions for, 61, 63
Subalgebra, 109 Unstable manifold, 246
Successive approximations, 3, 20, 21, 378, Upper-triangular matrix, 70
418 Urabe, M., 288
Sufficient condition, for asymptotic Uvarov, V. Q., 141
stability, 241
for uniqueness, 61, 63 V image of P, (A), 73
for unstability, 243 direct sum for C", 73
Summability of a formal power series, van der Pol equation, 312
371, 442, 443 for a large parameter, 322
Symplectic group Sp(2n, R), 92, 93 for a small parameter, 319
matrix of order 2n, 92, 93 107 singular perturbation of, 330
Variation of parameters, formula of, 101
Tahara, H., 112 Vector, infinite-dimensional, 120, 393
Tanaka, S., 144 norm of, 393
Taylor's series, 345 Vector space of solutions, 78
Transform, incomplete Leroy, 354
Transformation, shearing, 209 Wasow, W., 108, 304, 342, 372, 382, 449
Transversal, to orbit, 291, 292
Watson, G. N., 141, 342, 359
with respect to a vector, 291
Weinberg, L., 142, 447
Trivial solution, 236, 243
Weyl, H., 69
Turrittin, H. L., 391, 403, 428
Turrittin, Hukuhara-, theorem, 428 Whittaker, E. T., 141
Two dimensional system with constant Winkler, S., 196
coefficients, 251 Wintner, A., 197
Wronskian, 145, 149, 168, 172
Unbounded operators, 103
Uniformly asymptotic series, 343, 357, Xie, F., 197, 217
361 x"C{x}", 110
Uniqueness, Lipschitz condition, 1, 3, 64 x"C((Fx)l", 110
Nagumo condition, 64, 65
of S-N decomposition of a matrix, 75 Yoshida, M., 138
of asymptotic expansion, 343
of formal solution, 115 Zeros, of eigenfunctions, 185
of solution of a boundary-value of solutions, 144
problem, 150 Zygmund, A., 167
468
Universitext (continued )
Moise: Introductory Problems Course in Analysis and Topology

Morris: Introduction to Game Theory
Polster: A Geometrical Picture Book
Porter/Woods: Extensions and Absolutes of Hausdorff Spaces
Ramsay/Rlchtmyer: Introduction to Hyperbolic Geometry
Reisel: Elementary Theory of Metric Spaces
Rickart: Natural Function Algebras
Rotman: Galois Theory
Rubel/Colliander: Entire and Meromorphic Functions
Sagan: Space-Filling Curses
Samelson: Notes on Lie Algebras
Schif: Normal Families
Shapiro: Composition Operators and Classical Function Theory
Simonnet: Measures and Prohabilit
Smith: Power Senes From a Computational Point of View
Smorytiski: Self-Reference and Modal Logic
Stillwell: Geometry of Surfaces
Stroock: An Introduction to the Theory of Large Deviations
Sunder: An Invitation to von Neumann Algebras
Tondeur: Foliations on Riemannian Manifolds
Wong: Wcyl Transforms
Zhang: Matrix Theory Basic Results and Techniques
Zong: Strange Phenomena in Convex and Discrete Geometry
Zong: Sphere Packing-,
Universitext
The authors' aim is to provide the reader with the very basic knowledge
necessary to begin research on differential equations with professional
ability. The selection of topics should provide the reader with methods
and results that are applicable in a variety of different fields. The text is
suitable for a one-year graduate course, as well as a reference book for
research mathematicians. The book is divided into four parts. The first
covers fundamental existence, uniqueness, and smoothness with respect
to data, as well as nonuniqueness. The second part describes the basic
results concerning linear differential equations, and the third deals with
nonlinear equations. In the last part, the authors write about the basic
results concerning power series solutions.
Each chapter begins with a brief discussion of its contents and history.
The book has 114 illustrations and 206 exercises. Hints and comments
for many problems are given.
ISBN 0-387-98699-S
ISBN 0-387-98699-5 III
www.springer-ny.com 9 78038709869990>

Basic Theory of Ordinary Differential Equations Universitext PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Basic Theory of Ordinary Differential Equations Universitext PDF

Uploaded by

Copyright:

Available Formats

BASIC THEORY

Editors (North America): S. Axler, F.W. Gehring, and K.A. Ribet

Aksoy/Khamsl: Nonstandard Methods in Fixed Point Theory

(cnnrinued after index)

Basic Theory of Ordinary

With 114 Illustrations

Mathematics Subject Classification (1991): 34-01

Library of Congress Cataloging-in-Publication Data

Printed on acid-free paper.

r 1999 Spnnger-Verlag New York, Inc.

ISBN 0-387-98699-5 Springer-Verlag New York Berlin Heidelberg SPIN 10707353

This graduate level textbook is developed from courses in ordinary differential

Chapter I. Fundamental Theorems of Ordinary Differential Equations 1

I-1. Existence and uniqueness with the Lipschitz condition 1

Chapter II. Dependence on Data 28

Chapter III. Nonuniqueness 41

Chapter IV. General Theory of Linear Systems 69

Chapter V. Singularities of the First Kind 108

Chapter VI. Boundary-Value Problems of Linear Differential

Chapter VII. Asymptotic Behavior of Solutions of Linear Systems

Chapter VIII. Stability

Chapter IX. Autonomous Systems

Chapter X. The Second-Order Differential Equation

Chapter XI. Asymptotic Expansions 342

Chapter XII. Asymptotic Expansions in a Parameter 372

Chapter XIII. Singularities of the Second Kind 403

I-1. Existence and uniqueness with the Lipschitz condition

R = R((to,6),a,b) = {(t,yl: It - tot <a, ly - e< b},

(1.1.2) fi(t) = co + / t f(s, 0(s))ds for It -tot < fi.

(1.1.4) 1 (t)-ZbI<-MQ<Ma<Ma<MM=b for It - tol<f.

In particular, 0) - cot < b. This contradicts assumption (ii).

Remark 1-1-3. Using Lemma 1-1-2, it is concluded that if y = j(t) is a solution

(1.1.5) 0 < g(t) < K + L g(s)ds on to < t < ti,

exp[-L(t - to)]v(t) < {1 - exp(-L(t - to)]).

We use mathematical induction in Steps 1 and 2. The statement above is evident

Thus, (t.0k+1(t)) E R for It - tol < a.

14k+I (t) - Qk(t)l =J o

Step 4. The sequence {k(t) : k = 0, 1, 2.... } converges to

do(t) + F, I(&(t) - k-1(t))

uniformly on it - tol < a as k -+ +oo.

Step 5. fi(t) given by (I.1.10) is a solution of problem (P) on It - toI < a.

fi(t) = 6o + f f(s, (s))ds.

Step 6. fi(t) is the unique solution of problem (P) on It - tol < a.

on It - tol < &. Hence,

At) - lL(t)I = f' {f(s,.(s)) - f(s,iG(S))} dsl <- L I fo Id(s) - i (s)Idsl

on it - tol < a. Now, by applying Lemma 1-1-5 with K = 0 to the inequality

interval, D is a region, and Lo is a constant, and if V is a convex open set, then

9(O) = At, 92), 9(1) = f(t,91)-

where yh,1, Yh,2, ... , Yh.n are entries of yh (h = 1,2). Hence,

If(t,91)-f(t,y2)1 =II @(O)dO <nLolii-u21-

This is true for all t E Z, yl E D, and y2 E D. Thus, the lemma is proved.

has one and only one solution.

R = {(t,y): It-11 < a, ly-51 < b}

a = min ( a, ) =min f, 2 min(v ,10) = f.

R = {(t, y) : It - 41 < a, ly - 31 < b}.

as a formal Fourier series

(R) imam + am-1 + am+1 + am-2 + am+2 = 0 (m E Z),

1-2. Existence without the Lipschitz condition

Definition 1-2-1. A set.F of functions f (t) defined on an interval I is said to be

Step 2. The sequence { fh : h = 1, 2.... } of Step 1 converges uniformly on the

(1.2.1) Ifh(t) - ft(t)I < E

Then, (1.2.1) is satisfied whenever h > N(e) and I > N(E). 0

Step 1. There exists a function P(t, yj such that