Introduction To Functional Differential Equations, J Hale, Springer, 1993,458s PDF

Applied Mathematical Sciences
Volume 99
Editors
F. John J.E. Marsden L. Sirovich
Advisors
M. Ghil J .K. Hale J. Keller
K. Kirchgiissner B.J. Matkowsky
J.T. Stuart A. Weinstein
1. John: Partial Differential Equations, 4th ed.

2. Sirovich: Techniques of Asymptotic Analysis.
3. Hale: Theory of Functional Differential Equations, 2nd ed.
4. Percus: Combinatorial Methods.
5. von Mises/Friedrichs: Fluid Dynamics.
6. Freiberger/Grenonder: A Short Course in Computational Probability and Statistics.
7. Pipkin: Lectures on Viscoelasticity Theory.
8. Giacoglia: Perturbation Methods in Non-linear Systems.
9. Friedrichs: Spectral Theory of Operators in Hilbert Space.
10. Stroud: Numerical Quadrature and Solution of Ordinary Differential Equations.
11. Wolovich: Unear Multivariable Systems.
12. Berkovitz: Optimal Control Theory.
13. Bluman/Cole: Similarity Methods for Differential Equations.
14. Yoshizawa: Stability Theory and the Existence of Periodic Solution and Almost Periodic Solutions.
15. Braun: Differential Equations and Their Applications, 3rd ed.
16. Lefschetz: Applications of Algebraic Topology.
17. Collatz/Wenerling: Optimization Problems.
18. Grenander: Pattern Synthesis: Lectures in Pattern Theory, Vol. I.
19. Marsden/McCraclcen: Hopf Bifurcation and Its Applications.
20. Driver: Ordinary and Delay Differential Equations.
21. Courant/Friedrichs: Supersonic Flow and Shock Waves.
22. Rouche/Habets/Laloy: Stability Theory by Uapunov's Direct Method.
23. Lamperti: Stochastic Processes: A Survey of the Mathematical Theory.
24. Grenander: Pattern Analysis: Lectures in Pattern Theory, Vol. II.
25. Davies: Integral Transforms and Their Applications, 2nd ed.
26. Kushner/Clark: Stochastic Approximation Methods for Constrained and Unconstrained Systems.
27. de Boor: A Practical Guide to Splines.
28. Keilson: Markov Chain Models-Rarity and Exponentiality.
29. de Veubeke: A Course in Elasticity.
30. Shiatycki: Geometric Quantization and Quantum Mechanics.
31. Reid: Sturmian Theory for Ordinary Differential Equations.
32. Meis/Markowitz: Numerical Solution of Partial Differential Equations.
33. Grenander: Regular Structures: Lectures in Pattern Theory, Vol. III.
34. Kevorkian/Cole: Perturbation Methods in Applied Mathematics.
35. Carr: Applications of Centre Manifold Theory.
36. Bengtsson/Ghil/Kiillen: Dynamic Meteorology: Data Assimilation Methods.
37. Saperstone: Semidynamical Systems in Infinite Dimensional Spaces.
38. Lichtenberg/Lieberman: Regular and Chaotic Dynamics, 2nd ed.
39. Piccini/Stampacchia/Vidossich: Ordinary Differential Equations in R".
40. Naylor/Sell: Unear Operator Theory in Engineering and Science.
41. Sparrow: The Lorenz Equations: Bifurcations, Chaos, and Strange Attractors.
42. Guclcenheimer/Holmes: Nonlinear Oscillations, Dynamical Systems and Bifurcations of Vector Fields.
43. Oclcendon/Taylor: lnviscid Fluid Flows.
44. Pazy: Semigroups of Unear Operators and Applications to Partial Differential Equations.
45. Glashoff/Gustafson: Linear Operations and Approximation: An Introduction to the Theoretical Analysis
and Numerical Treatment of Semi-Infinite Programs.
46. Wilcox: Scattering Theory for Diffraction Gratings.
47. Hale et al: An Introduction to Infinite Dimensional Dynamical Systems-Geometric Theory.
48. Murray: Asymptotic Analysis.
49. Ladyzhenskaya: The Boundary-Value Problems of Mathematical Physics.
50. Wilcox: Sound Propagation in Stratified Fluids.
51. Golubitsky/Schaeffer: Bifurcation and Groups in Bifurcation Theory, Vol. I.
(continued following index)

Jack K. Hale Sjoerd M. Verduyn Lunel
Introduction to
Functional Differential
Equations
With 10 Illustrations
Springer Science+Business Media, LLC

Jack Hale Sjoerd M. Verduyn Lunel
School of Mathematics Vrije Universiteit
Georgia Institute of Amsterdam
Technology De Boelelaan 1081 a
Atlanta, GA 30332 1081 HV Amsterdam
USA The Netherlands
Editors
F. John J .E. Marsden L. Sirovich
Courant Institute of Department of Division of
Mathematical Sciences Mathematics Applied Mathematics
New York University University of California Brown University
New York, NY 10012 Berkeley, CA 94720 Providence, RI 02912
USA USA USA
Mathematics Subject Classification (1991): 34K20, 34A30, 39A10
Library of Congress Cataloging-in-Publication Data

Hale, Jack K.
Introduction to functional differential equations/Jack K. Hale,
Sjoerd M. Verduyn Lunel.
p. em. - (Applied mathematical sciences ; v.)
Includes bibliographical references.
ISBN 978-1-4612-8741-4 ISBN 978-1-4612-4342-7 (eBook)
DOI 10.1007/978-1-4612-4342-7
I. Functional differential equations.
I. Verduyn Lunel, S. M.
(Sjoerd M.) II. Title. Ill. Series: Applied mathematical sciences
(Springer-Verlag New York Inc.) ; v.
QAI.A647
[QA372]
510 s-dc20
[515'.35] 93-27729
Printed on acid-free paper.
© 1993 Springer Science+Business Media New York

Originally published by Springer-Verlag New York Inc. in 1993
Softcover reprint of the hardcover 1st edition 1993
All rights reserved. This work may not be translated or copied in whole or in part without the
written permission of the publisher (Springer Science+Business Media, LLC), 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 nnder-
stood by the Trade Marks and Merchandise Marks Act, may accordingly be used freely by
anyone.
Production managed by Natalie Johnson; manufacturing supervised by Jacqui Ashri.

Photocomposed copy produced using the authors' TeX files.
987654321
Preface
The present book builds upon an earlier work of J. Hale, "Theory of Func-
tional Differential Equations" published in 1977. We have tried to maintain
the spirit of that book and have retained approximately one-third of the
material intact. One major change was a complete new presentation of lin-
ear systems (Chapters 6~9) for retarded and neutral functional differential
equations. The theory of dissipative systems (Chapter 4) and global at-
tractors was completely revamped as well as the invariant manifold theory
(Chapter 10) near equilibrium points and periodic orbits. A more complete
theory of neutral equations is presented (see Chapters 1, 2, 3, 9, and 10).
Chapter 12 is completely new and contains a guide to active topics of re-
search. In the sections on supplementary remarks, we have included many
references to recent literature, but, of course, not nearly all, because the
subject is so extensive.
Jack K. Hale
Sjoerd M. Verduyn Lunel
Contents
Preface.................................................. .......... v
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1. Linear differential difference equations . . . . . . . . . . . . . . . . . . . . 11

1.1 Differential and difference equations. . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2 Retarded differential difference equations. . . . . . . . . . . . . . . . . . . . . . . 13
1.3 Exponential estimates of x( ¢,f) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4 The characteristic equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.5 The fundamental solution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.6 The variation-of-constants formula............................. 23
1. 7 Neutral differential difference equations . . . . . . . . . . . . . . . . . . . . . . . . 25
1.8 Supplementary remarks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2. Functional differential equations: Basic theory . . . . . . . . . . 38

2.1 Definition of a retarded equation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.2 Existence, uniqueness, and continuous dependence . . . . . . . . . . . . . 39
2.3 Continuation of solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.4 Differentiability of solutions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.5 Backward continuation........................................ 51
2.6 Caratheodory conditions....................................... 58
2.7 Definition of a neutral equation................................ 59
2.8 Fundamental properties of NFDE.............................. 61
3. Properties of the solution map.............................. 67

3.1 Finite- or infinite-dimensional problem?........................ 68
3.2 Equivalence classes of solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.3 Small solutions for linear equations............................ 74
3.4 Unique backward extensions................................... 86
3.5 Range in IRn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
viii Contents
3.6 Compactness and representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

3.7 The solution map for NFDE...................... ............. 91
4. Autonomous and periodic processes ....................... 100

4.1 Processes .............................. ........................ 100
4.2 Invariance .............................. ....................... 104
4.3 Discrete systems-Maximal invariant sets and attractors ....... 107
4.4 Fixed points of discrete dissipative processes . . . . . . . . . . . . . . . . . . . 115
4.5 Continuous systems-Maximal invariant sets and attractors .... 119
4.6 Stability and maximal invariant sets in processes ............... 121
4. 7 Convergent systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.8 Supplementary remarks .............................. .......... 127
5. Stability theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

5.1 Definitions .............................. ...................... 130
5.2 The method of Liapunov functionals ........................... 132
5.3 Liapunov functionals for autonomous systems .................. 143
5.4 Razumikhin theorems .............................. ............ 151
6. General linear systems .............................. ......... 167

6.1 Resolvents and exponential estimates .......................... 167
6.2 The variation-of-constants formula ............................. 173
6.3 The formal adjoint equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
6.4 Boundary-value problems .............................. ........ 180
6.5 Stability and boundedness .............................. ....... 184
7. Linear autonomous equations .............................. . 193

7.1 Strongly continuous semigroups .............................. .. 193
7.2 Spectrum of the generator-Decomposition of C ............... 197
7.3 Characteristic matrices and equivalence ........................ 200
7.4 The generalized eigenspace for RFDE .......................... 205
7.5 Decomposing C with the adjoint equation ...................... 208
7.6 Estimates on the complementary subspace ..................... 213
7. 7 An example .............................. ..................... 215
7.8 Spectral decomposition according to all eigenvalues ............ 218
7.9 The decomposition in the variation-of-constants formula ........ 225
7.10 Parameter dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
Contents IX
8. Periodic systems .................... .................... ...... 236

8.1 General theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
8.2 Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
8.3 An example: Integer delays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 7
8.4 Supplementary remarks .................... .................... 254
9. Equations of neutral type .................... ............... 255

9.1 General linear systems .................... .................... . 255
9.2 Linear autonomous equations .................... .............. 262
9.3 Exponential estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
9.4 Hyperbolic semigroups .................... .................... . 275
9.5 Variation-of-constan ts formula .................... ............. 279
9.6 Strongly stable D operators .................... ................ 284
9.7 Properties of equations with stableD operators ................ 290
9.8 Stability theory .................... .................... ........ 292
10. Near equilibrium and periodic orbits .................... . 302

10.1 Hyperbolic equilibrium points .................... .............. 302
10.2 Nonhyperbolic equilibrium points .................... .......... 312
10.3 Hyperbolic periodic orbits .................... ................. 317
10.4 Nondegenerate periodic orbits of RFDE .................... .... 324
11. Periodic solutions of autonomous equations ............. 331

11.1 Hopf bifurcation .................... .................... ....... 331
11.2 A periodicity theorem .................... .................... . 335
11.3 Range of the period .................... .................... ... 338
11.4 The equation x(t) = -ax(t- 1)[1 + x(t)] .................... .. 341
11.5 The equation x(t) = -ax(t- 1)[1- x 2 (t)] .................... . 347
11.6 The equation x(t) + f(x(t))x(t) + g(x(t- r)) = 0 .............. 348
12. Additional topics .................... .................... .... 364

12.1 Equations on manifolds-Definitio ns and examples ............. 364
12.2 Dimension of the global attractor .................... .......... 369
12.3 A-stability and Morse-Smale maps .................... ......... 372
12.4 Hyperbolicity is generic .................... .................... 376
12.5 One-to-oneness on the attractor .................... ............ 379
12.6 Morse decompositions .................... .................... . 382
12.7 Singularly perturbed systems .................... .............. 386
12.8 Averaging .................... .................... ............. 396
x Contents
12.9 Infinite delay .................... .................... .......... 401

Appendix: Stability of characteristic equations .............. 414

Bibliography .................... .................... .............. 419
Index .................... .................... .................... .. 444
List of Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 7
Introduction
In many applications, one assumes the system under consideration is

governed by a principle of causality; that is, the future state of the system
is independent of the past states and is determined solely by the present.
If it is also assumed that the system is governed by an equation involving
the state and rate of change of the state, then, generally, one is consider-
ing either ordinary or partial differential equations. However, under closer
scrutiny, it becomes apparent that the principle of causality is often only
a first approximation to the true situation and that a more realistic model
would include some of the past states of the system. Also, in some prob-
lems it is meaningless not to have dependence on the past. This has been
known for some time, but the theory for such systems has been extensively
developed only recently. In fact, until the time of Volterra [1] most of the
results obtained during the previous 200 years were concerned with spe-
cial properties for very special equations. There were some very interesting
developments concerning the closure of the set of exponential solutions of
linear equations and the expansion of solutions in terms of these special
solutions. On the other hand, there seemed to be little concern about a
qualitative theory in the same spirit as for ordinary differential equations.
In his research on predator-prey models and viscoelasticity, Volterra
[1, 2] formulated some rather general differential equations incorporating
the past states of the system. Also, because of the close connection be-
tween the equations and specific physical systems, Volterra attempted to
introduce a concept of energy function for these models. He then exploited
the behavior of this energy function to study the asymptotic behavior of
the system in the distant future. These beautiful papers were almost com-
pletely ignored by other workers in the field and therefore did not have
much immediate impact.
In the late thirties and early forties, Minorksy [1], in his study of ship
stabilization and automatic steering, pointed out very clearly the impor-
tance of the consideration of the delay in the feedback mechanism. The
great interest in control theory during these and later years has certainly
contributed significantly to the rapid development of the theory of differ-
ential equations with dependence on the past state.
2 Introduction
In the late forties and early fifties, a few books appeared which pre-
sented the current status of the subject and certainly greatly influenced
later developments. In his book, Mishkis [1] introduced a general class of
equations with delayed arguments and laid the foundation for a general
theory of linear systems. In their monograph at the Rand Corporation,
Bellman and Danskin [1] pointed out the diverse applications of equations
containing past information to other areas such as biology and economics.
They also presented a well organized theory of linear equations with con-
stant coefficients and the beginnings of stability theory. A more extensive
development of these ideas is in the book of Bellman and Cooke [1]. In
his book on stability theory, Krasovskii [1] presented the theory of Lia-
punov functionals emphasizing the important fact that some problems in
such systems are more meaningful and amenable to solution if one con-
siders the motion in a function space even though the state variable is a
finite-dimensional vector.
With such clear indications of the importance of these systems in the
applications and with the number of interesting mathematical problems
involved, it is not surprising that the subject has undergone a rapid de-
velopment in the last forty years. New applications also continue to arise
and require modifications of even the definition of the basic equations. We
list below a few types of equations that have been encountered merely to
give an idea of the diversity and give appropriate references for the specific
applications.
The simplest type of past dependence in a differential equation is that
in which the past dependence is through the state variable and not the
derivative of the state variable, the so-called retarded functional differential
equations or retarded differential difference equations. For a discussion of
the physical applications of the differential difference equation
. dx
x(t) = F(t, x(t), x(t- r)), x=-
dt
to control problems, see Minorsky [2, Ch. 21]. Lord Cherwell (see Wright
[1, 2]) has encountered the differential difference equation
x(t) = -cxx(t- 1)[1 + x(t)]
in his study of the distribution of primes. Variants of this equation have

also been used as models in the theory of growth of a single species (see
Cunningham [1]). Dunkel [1] suggested the more general equation
x(t) =-ex [1° 1

x(t +B) dry((:J)] [1 + x(t)]
for the growth of a single species.

In his study of predator-prey models, Volterra [1] had earlier investi-
gated the equations
Introduction 3
x(t) = [E1- ''!lY(t)- j_or F1(8)y(t + 8)d8] x(t)
y(t) = [ -El + 'Y2x(t) + j_or F2(8)x(t + B)d()] y(t)

where x and y are the number of prey and predators, respectively, and all
constants and functions are non-negative. For similar models, Wangersky
and Cunningham [1, 2], have also used the equations
x(t) = ax(t) [m ~x(t)] - bx(t)y(t)

y(t) = -(3y(t) + cx(t- r)y(t- r)
for predator-prey models.
In an attempt to explain the circummutation of plants (and especially
the sunflower), Israelson and Johnsson [1, 2] have used the equation
a(t) = -k 1 00
!(8) sina(t- ()-to) d()
as a model, where a is the angle the top of the plant makes with the vertical
(see also Klein [1]). For other applications, see Johnson and Karlsson [1].
Under suitable assumptions, the equation
N
x(t) = L Aix(t- Ti)
i=O
is a suitable model for describing the mixing of a dye from a central tank
as dyed water circulates through a number of pipes. An application to the
distribution in man of labeled albumin as it circulates from the blood stream
through the interstitial fluids and back to the blood stream is discussed by
Bailey and Reeve [1] (see also Bailey and Williams [1]). Boffi and Scozzafava
[1, 2] have also encountered this equation in transport problems.
In an attempt to describe the spread of measles in a metropolitan area,
London and Yorke [1] have encountered the equation
S(t) = -(3(t)S(t)[2'Y + S(t- 14)- S(t- 12)] + 'Y

where S(t) is the number of susceptible individuals at timet, 'Y is the rate
at which individuals enter the population, (3(t) is a function characteristic
of the population, and an individual exposed at time t is infectious in the
time interval [t- 14, t- 12].
In an analysis of gonorrhea, Cooke and Yorke [1] have studied the
equation
where I represents the number of infectious individuals and g is a non-

negative function vanishing outside a compact interval.
4 Introduction
A more general equation describing the spread of disease taking into

account age dependence was given by Cooke [1] and Hoppenstadt and Walt-
man [1]. For other equations that occur in the theory of epidemics, see
Waltman [1]. For other models in the biomedical sciences, see Banks [1].
Grossberg [1, 2] has encountered interesting differential equations in the
theory of learning.
The equation
i:(t) = -l~r a(t- u)g(x(u)) du

was encountered by Ergen [1] in the theory of a circulating fuel nuclear
reactor and has been extensively studied by Levin and Nohel [1]. In this
model, x is the neutron density. It is also a good model in one-dimensional
viscoelasticity in which x is the strain and a is the relaxation function.
Taking into account the transmission time in the triode oscillator,
Rubanik [1, p. 130] has encountered the van der Pol equation
x(t) + ai:(t)- f(x(t- r))i:(t- r) + x(t) = 0

with the delayed argument r. Taking into account the retarded connections
between oscillating systems, Starik [1] has encountered the system
y(t) + [w 2 + EAsin¢(t- rl)]y(t) = -E[hy(t) + '"YY3 (t- r 2 )]

J¢(t) = E(L(¢(t))- H(<j>(t)- CYiy 2 (t- r 3 ) cos<j>(t))
- CY2 sin ¢(t)- CY3 cos ¢(t)).
In the theory of optimal control, Krasovskii [2] has studied extensively

the system
i:(t) = P(t)x(t) + B(t)u(t)
y(t) = Q(t)x(t)
u(t) = /_or d[ry(t, B)]y(t +B)+ /_or d[JL(t, B)]u(t +B).

There are also a number of applications in which the delayed argument
occurs in the derivative of the state variable as well as in the independent
variable, the so-called neutral differential difference equations. Such prob-
lems are more difficult to motivate but often arise in the study of two or
more simple oscillatory systems with some interconnections between them.
For simplicity, it is usually assumed that the interaction of the components
of the coupled systems takes place immediately. In many cases, the time
for the interaction to take place is important even in determining the qual-
itative behavior of the system. It often occurs that the connection between
the coupled systems can be adequately described by a system of linear
hyperbolic partial differential equations with the motion of each individ-
ual system being described by a boundary condition. In some cases, the
Introduction 5
connection through the partial differential equations (considered as a con-

nection by a traveling wave) can be replaced by connections with delays.
Generally, the resulting ordinary differential equations involve delays in the
highest-order derivatives. A general discussion of when this process is valid
may be found in Rubanik [1] and Cooke and Krumme [1].
For example, Brayton [1] considered the lossless transmission line con-
nected as shown in Fig. 0.1, where g(v) is a nonlinear function of v and
gives the current in the indicated box in the direction shown. This problem
may be described by the following system of partial differential equations
0 <X< 1, t > 0,
with the boundary conditions
dv(1, t) .
E- v(O, t) - Ri(O, t) = 0, cl dt = z(l, t)- g(v(1, t)).
Fig. 0.1.
We now indicate how one can transform this problem into a differential
equation with delays. If s = (LC)- 112 and z = (L/C) 112 , then the general
solution of the partial differential equation is given by
v(x, t) = ¢(x- st) + '1/J(x + st)
i(x, t) = ~[¢(x-
z
st)- 'lj;(x + st)]
or
2¢(x- st) = v(x, t) + zi(x, t)
2'1/J(x + st) = v(x, t) - zi(x, t).
6 Introduction
This implies
2¢( -st) = v(1, t + ~) + zi(1, t + ~)
z s
2'1/J(st) = v(1, t- ~)- zi(1, t- ~).
z s
Using these expressions in the general solution and using the first boundary
condition at t- (1/s), one obtains
i(1, t)- Ki(1, t- ~) =a-~ v(1, t)- K v(1, t- ~)

s z z s
where K = (z-R)j(z+R), a= 2E/(z+R). Inserting the second boundary
condition and letting u(t) = v(1, t), we obtain the equation
u(t)- Ku(t- ~)
s
= f(u(t),u(t- ~))
s
where s = .,fLC,
1 K
Cd(u(t), u(t- r)) =a-- u(t)-- u(t- r)- g(u(t))
z z
+ Kg(u(t- r)),
all constants are positive and depend on the parameters in the original
equations. Also, if R > 0, then K < 1.
If generalized solutions of the original partial differential equation were
considered, the delay equation would require differentiating the difference
u(t)- Ku(t- (2/s)) rather than each term separately; that is, one would
consider the equation
d 2 2
dt [u(t)- Ku(t- -_;)] = f(u(t), u(t- -_;)).
The prescription for passing from a linear partial differential equation

with nonlinear boundary conditions to a delay equation is certainly not
unique and other transformations may be desirable in certain situations.
This fact is illustrated following the ideas of Lopes [1]. Let JKI < 1 (i.e.,
R > 0) and let p be any solution of the difference equation
2 z 1
p(t)- Kp(t--) = -b(t), b(t) = - E ( t - -).
s z+R s
If E is periodic, one can choose p periodic with the same period. Using the
first boundary condition at t- (1/s) and the general solution, one obtains
¢(1-st)= b(t)- K'lj;(st- 1).

If w(t) = 'lj;(1 + st) - p(t), then evaluation in the general solution gives
v(1, t) = w(t)- Kw(t- r)

Introduction 7
I K
i(1 , t) z w(t- r) + q
= --z w(t)--
where r = 2/ s, zq(t) = -p(t) -Kp(t-r) +b(t). Using the second boundary
condition one obtains the equation
d 1 K
C 1 -[w(t)- Kw(t- r)] = q- -w(t)- -w(t- r)
& z z
- g(w(t)- Kw(t- r)).
In his consideration of shunted transmission lines, Lopes [2] encountered

equations of the preceding type with two delays.
Sometimes boundary control of a linear hyperbolic equation can be
more effectively studied by investigating the corresponding control problem
for the transformed equations (see Banks and Kent [1]).
Another similar equation encountered by Rubanik [1] in his study of
vibrating masses attached to an elastic bar is
x(t) + wfx(t) = Eh (x(t), x(t), y(t), y(t)) + 'Ydi(t- r),

jj(t) + w~x(t) = Eh(x(t), x(t), y(t), y(t)) + 'Y2x(t- r).
In studying the collision problem in electrodynamics, Driver [1] con-
sidered systems of the type
x(t) = JI(t,x(t),x(g(t))) + h(t,x(t),x(g(t)))x(g(t))

where g(t) ::; t. In the same problem, one encounters delays g that also
depend on x.
El'sgol'tz [1, 2], Sabbagh [1], and Hughes [1] have considered the vari-
ational problem of minimizing
V(x) = 1 1
F(t, x(t), x(t- r), x(t), x(t- r)) dt
over some class of functions x. Generally, the Euler equations are of the
form
x(t) = f(t, x(t), x(t- r), x(t), x(t- r), x(t- r))
with some appropriate boundary conditions.
In the slowing down of neutrons in a nuclear reactor, the asymptotic
behavior as t --+ oo of the equation
x(t) = 1
t
t+l
k(s)x(s) ds
or
x(t) = k(t + 1)x(t + 1)- k(t)x(t)
seems to play an important role (see Slater and Wilf [1]). The state at time
t depends on the future state of the system. This can be considered as a
8 Introduction
special case of the retarded equation if we replace t by - T and investigate

solutions in the direction of decreasing T.
In his study of the dynamics of certain types of elastic materials,
Volterra [3] suggested that an appropriate model for the system would be
partial differential equations of the following type:
t
Ut(x,t)- .du(x,t) + Jo .L Ux;xi(x,T)¢i,j(t,T)dT = f(x,t)
n
•,J=l
Ut(x, t)- .du(x, t) + R(x, t)u(x, w(t)) = f(x, t)

where R(x, t) is a linear differential operator in x of first order and w(t) E
[0, t] for each t. Systems of this type and even more general ones have been
studied by Artola [1], Baiocchi [1, 2], Dafermos [1, 2], MacCamy [1, 2, 3],
and Slemrod [1].
There are many other systems for which the future behavior depends
on the past and yet there are no derivatives involved at all. One of these is
the difference equation
x(t) = g(t, x(t- 1), x(t- 2), ... , x(t- N)),
so important in problems in economics and models of heredity.

Finally, the Volterra integral equations
x(t) = f(t) +lot a(t, s, x(s)) ds
occur often in applied mathematics. For specific references and many more
examples and applications, see the Miscellaneous Exercises and Research
Problems in Bellman and Cooke [1] and the books of Miller [1], Corduneanu
[1], Halanay [1], and Grippenberg et al. [1].
The preceding examples have amply illustrated the importance and
frequency of occurrence of equations that depend on past history. The di-
versity of the different types of equations makes it seem at first glance
to be almost impossible to find a class of equations that contains all of
these and is still mathematically tractable and interesting. Of course, one
could write an equivalent integral equation for all differential equations and
then consider general operator equations to obtain existence, uniqueness,
etc. Some such general papers have appeared (see, in particular, Tychonov
[1] and Neustadt [1]) and include a few of these types. The difficulty in
this approach is to incorporate into the resulting functional equation all
of the distinct properties associated with the original differential equation.
One obtains a general existence theorem for a functional equation and it
becomes a major task to verify that one of the special equations satisfies
all of the hypotheses. But more importantly, some of the dynamics and
geometry of the original problem are lost.
Introduction 9
In this book, we emphasize the dynamics and the resulting flow in-
duced by the equations. Our objective is to obtain a theory for classes
of equations that begins to be as comprehensive as the available theory
for ordinary differential equations. We continually attempt to emphasize
the underlying ideas involved, hoping that further research in similar di-
rections will lead to extensions of more complicated problems that occur
in the applications. To accomplish our objective, we first discuss at great
length equations with no delays in the derivatives (the so-called retarded
functional differential equations). We then introduce a class of equations
with delays in the derivatives (neutral equations), which includes many but
not all of these special types. We hope in this way to isolate a class of
equations that is small enough to have a rich mathematical structure and
yet is large enough to include many interesting applications. As remarked
earlier, the experience and information gained by this approach are useful
in the discussion of other types of equations.
A brief description of the organization of the book follows. The first
chapter introduces the subject through linear differential difference equa-
tions of retarded and neutral type with constant coefficients. In this way,
the reader becomes familiar at an elementary level with the characteristic
equation, the fundamental solution, and the role of the fundamental so-
lution in determining precise exponential bounds on the solutions of the
homogeneous equation as well as the behavior of the solutions of nonhomo-
geneous equations.
For a rather general class of retarded and neutral functional differential
equations, Chapter 2 contains the basic theory of existence, uniqueness,
continuation, and continuous dependence on parameters and initial data.
Chapter 3 is fundamental for an understanding of some of the dif-
ferences between ordinary differential equations and functional differen-
tial equations. This chapter contains numerous examples depicting various
types of behavior of the solutions. A careful study of these examples will
develop the type of intuition that should allow the reader to avoid pitfalls
as well as make sensible conjectures of what to expect in specific problems.
In addition, in Chapter 3, we discuss exponentially small solutions for lin-
ear autonomous equations and give a very useful characterization of the
solution operator of nonlinear systems in terms of a contracting semigroup
and a completely continuous operator.
Chapter 4 contains an abstract theory of dissipative processes, concen-
trating particularly on maximal compact invariant sets and compact global
attractors. This theory leads to a procedure for comparing the flows of
infinite-dimensional systems. It also has applications to stability theory as
well as the existence of periodic solutions of periodic functional differential
equations.
Chapter 5 is an extensive study of the theory of stability. We empha-
size and contrast the method of Liapunov functionals and the method of
Razumikhin. Many examples are given illustrating the results.
10 Introduction
Chapters 6-8 deal with linear systems of retarded FDE. Chapter 6 con-
tains the general theory of time-varying systems, including the variation-of-
constants formula, formal adjoint equations, and boundary-value problems.
Some relationships are also given between the various types of stability for
linear systems. Chapter 7 contains the fundamental theory of linear au-
tonomous equations. It shows how the theory for functional differential
equations is related to the theory of linear ordinary differential equations
with constant coefficients, including a decomposition analogous to the Jor-
dan block decomposition for matrices. These results are fundamental in the
study of perturbed linear systems as well as the generic theory. Chapter 8
deals with the same questions as Chapter 7 except for periodic systems.
Chapter 9 is devoted to topics in neutral functional differential equa-
tions, especially the theory for linear systems analogous to Chapters 6-8
for retarded equations and the stability theory analogous to Chapter 5 for
retarded equations.
Chapter 10 is devoted to the discussion of stable, unstable, and center
manifolds near an equilibrium point for retarded and neutral functional
differential equations. For retarded equations, similar results are given for
periodic orbits of autonomous systems. Analogous but weaker results are
given for periodic orbits of neutral equations.
Chapter 11 deals with the existence of nonconstant periodic solutions
of autonomous equations. The Hop£ bifurcation is given as well as a general
method for determining periodic solutions. The latter method is probably
the most powerful one available at the present time. Various illustrations
are given.
In Chapter 12, we present selected directions in which the field of
functional differential equations has been going in recent years. The topics
in Sections 12.1-12.9 are self-explanatory and are presented in some detail
with few proofs.
Each chapter has a section called "Supplementary Remarks" that con-
tains many other directions with references in which the field of functional
differential equations is going.
The Appendix contains the classical procedure for determining when
the roots of a characteristic equation are in the left half-plane. The examples
needed in the text are discussed in detail.
There are now several books devoted to FDE and their applications
that complement the present text (for example, see Diekmann, van Gils,
Verduyn Lunel, and Walther [1], Driver [3], El'sgol'tz and Norkin [1], Gos-
plamy [1], Kolmanovskii and Nosov [1], and Kolmanovskii and Myshkis
[1]).
1
Linear differential difference equations
In this chapter, we discuss the simplest possible differential difference equa-

tions; namely, linear equations with constant coefficients. For these equa-
tions, a rather complete theory can be developed using very elementary
tools. The chapter serves as an introduction to the more general types of
equations that will be encountered in later chapters. It also is intended
to bring out the roles of the characteristic equation and the Laplace trans-
form and to emphasize some of the differences between retarded and neutral
equations. Since ordinary differential equations and difference equations are
special cases of the theory, we begin the discussion with the latter.
1.1 Differential and difference equations
Let lR = ( -oo, oo ), lRn be any real n-dimensional normed vector space. For
the scalar differential equation
(1.1) x=Ax,
where A is a constant, all solutions are given by (exp At )c, where c is an
arbitrary constant. In case A is ann x n matrix and x is an n-vector, the
same result is true, of course, with c an n-vector. Each column of exp At
has the form LPj(t) exp AA where each Pj(t) is ann-vector polynomial in
t and each Aj is an eigenvalue of the matrix A; that is, each Aj satisfies the
characteristic equation
(1.2) det (M- A)= 0.
The coefficients of the polynomials Pj are determined from the generalized

eigenvectors of the eigenvalue Aj. As a consequence of this representation
of the solutions of Equation (1.1), complete information of the solution is
obtained from the eigenvalues and eigenvectors of the matrix A. Note that
the characteristic equation (1.2) can be obtained by trying to find nontrivial
solutions of Equation (1.1) of the form (exp>.t)c.
Now consider the nonhomogeneous equation
12 1. Linear differential difference equations
(1.3) x =Ax+ f(t)

where f is a given continuous function from ffi to m,n. It is well known that
the solution of Equation (1.3) with x(O) = c is given by the variation-of-
constants formula
(1.4)
The derivation is easily obtained by making the transformation of variables

x(t) = (expAt)y(t) and observing that
iJ =e-At f(t).
Representation (1.4) will be needed later in this chapter to prove an exis-
tence theorem for a differential difference equation.
The theory for simple difference equations follows closely in spirit the
theory for Equation ( 1.1). Consider the difference equation
(1.5) x(t) = Ax(t- 1) + Bx(t- 2)

where A and B are constants. By introducing an additional variable y(t) =
x(t - 1), this scalar equation (1.5) is equivalent to the two-dimensional
equation
(1.6) z(t) = Cz(t- 1),
To obtain a solution of Equation (1.6) defined for all t 2: 0, one must

e
specify a 2-vector function ¢ on [-1, OJ. For any E [-1, OJ, the solution z
of Equation (1.6) is given by
(1.7) z(t) = ct-e¢(8), t = e, e+ 1, ... , e+ k, ....

Thus we see that the behavior of the solutions is determined by the eigen-
values and eigenvectors of the matrix C. The eigenvalues of Care the roots
of the characteristic equation
(1.8) p2 -Ap- B = 0.
Even though an initial function on [-r, 0] is needed to have Equation (1.6)
define a function on [0, oo ), the problem actually is finite dimensional since
the detailed structure of the solution is determined by the mapping C on the
plane. As for the differential equation (1.1), notice that the characteristic
equation (1.8) can be obtained by seeking nontrivial solutions of Equation
(1.5) of the form x(t) =pte, where cis a nonzero constant.
In this form, Equation (1.5) seems to be no more complicated than
Equation (1.1) since it is very similar to a linear map of the plane into itself.
However, the analysis of this equation is very sensitive to the numbers 1
and 2 on the right-hand side. If we consider the equation
1.2 Retarded differential difference equations 13
(1.9) x(t) = Ax(t- r) + Bx(t- s)

where r Is is irrational, s > r > 0, the problem is completely different.
In contrast to the case r = 1, s = 2, one cannot obtain any solution
of Equation (1.9) by specifying initial values only at x(-r),x(-s). The
problem is basically infinite-dimensional and one sees that a reasonable
initial-value problem for Equation (1.9) is to specify an initial function on
[-s, OJ and use Equation (1.9) to determine a solution for t 2 0.
In analogy to the preceding cases, one would expect the characteristic
equation for Equation (1.9) to be obtained by seeking nontrivial solutions
of Equation (1.9) of the form x(t) = pte, where c =f. 0 is constant. The
resulting equation is
(1.10)
This equation for r Is irrational has infinitely many solutions. Therefore, it

is not obvious that the solutions of Equation (1.9) can be obtained as linear
combinations of the characteristic functions. Even without discussing the
question of representation of solutions in series, it is not even obvious that
the asymptotic behavior (stability, etc.) of the solutions of Equation (1.9)
is determined by the solutions of the characteristic equation (1.10). Both
of these problems have a positive solution and, as we shall see, one method
of attack is through the Laplace transform.
A generalization of Equation (1.9) would be the equation
0
x(t) = ]_
00
d[tL(B)]x(t +B)
where fL is a function of bounded variation.
1.2 Retarded differential difference equations
The simplest linear retarded differential difference equation has the form
(2.1) x(t) = Ax(t) + Bx(t- r) + f(t)

where A, B, and r are constants with r > 0, f is a given continuous function
on IR, and x is scalar.
The first question is the following: what is the initial-value problem for
Equation (2.1)? More specifically, what is the minimum amount of initial
data that must be specified in order for Equation (2.1) to define a function
for t 2 0? A moment of reflection indicates that a function must be specified
on the entire interval [-r, 0]. In fact, let us prove
Theorem 2.1. If ¢is a given continuous function on [-r, 0], then there is a
unique function x( ¢,f) defined on [-r, oo) that coincides with ¢ on [-r, OJ
and satisfies Equation (2.1) fort;:::: 0. Of course, at t = 0, the derivative in
Equation (2.1) represents the right-hand derivative.
Proof. If xis a solution of Equation (2.1) which coincides with¢ on [-r, 0],
then the variation-of-constants formula (1.4) implies that x must satisfy
x(t) = ¢(t), t E [-r, 0],
(2.2)
x(t) = eAt¢(0) + 1t eA(t-s)[Bx(s- r) + f(s)J ds, t;:::: 0.
Also, if x satisfies Equations (2.2), then x must satisfy Equation (2.1).

It is only necessary to show that Equations (2.2) have a unique solution. But
this is trivial to demonstrate since we may explicitly calculate the solution
by the method of steps. In fact, on the interval 0::::; t::::; r, the function xis
given uniquely by
x(t) = eAt¢(0) + 1t eA(t-s)[B¢(s- r) + f(s)J ds.
Once x is known on [0, rJ and since it is continuous on this interval, the

second of Equations (2.2) can be used to obtain x on [r, 2rJ. The process
may be continued to prove the theorem. D
If f is not continuous but only locally integrable on IR, then the same
proof yields the existence of a unique solution x( ¢,f). Of course, by a solu-
tion, we mean a function that satisfies Equation (2.1) almost everywhere.
Theorem 2.2. If x( ¢,f) is the solution of Equation (2.1) defined by Theorem

2.1, then the following assertions are valid.
(i) x(¢, f)(t) has a continuous first derivative for all t > 0 and has a
continuous derivative at t = 0 if and only if ¢(8) has a derivative at
e
= 0 with
(2.3) ¢(0) = A¢(0) + B¢( -r) + f(O).
If f has derivatives of all orders, then x( ¢, f) becomes smoother with
increasing values oft.
(ii) If B -1- 0, then x(¢, f) can be extended as a solution of Equation (2.1)
on [-r - E, oo), 0 < E ::::; r, if and only if ¢ has a continuous first
derivative on [-E, OJ and Equation (2.3) is satisfied. Extension further
to the left requires more smoothness of¢ and f and additional boundary
conditions similar to Condition (2.3).
Proof. Part (i) is obvious from Equation (2.1). The necessity of Part (ii) is
also obvious. To prove the sufficiency, simply observe that the extension to
the left of -r can be accomplished by using the formula
1.3 Exponential estimates of x( ¢,f) 15
1 .
(2.4) x(t- r) = B [x(t)- Ax(t)- f(t)J
by the method of steps; that is, if ¢ satisfies the stated conditions, then
the right-hand side is known explicitly fort E [-E, OJ and, therefore, x(s)
is known for s E [-r- E, -rJ. If¢ has a first derivative on [-r, 0], then
Relation (2.4) defines the solution on [-2r, -rJ and, therefore the solution
is extended to [-2r, oo).
To extend the solution to the interval [- 2r - E, oo), 0 < E ::; r, re-
quires that the function x (s), s E [-r - E, -r], defined by Formula ( 2.4) be
continuously differentiable and satisfies
(2.5) x( -r) =Ax( -r) +Ex( -2r) + f( -r).

This requires the right-hand side of Formula (2.4) to be continuously differ-
entiable on [-E, 0], which imposes conditions on f and¢ as well as boundary
conditions at 0. It is sufficient for f to be differentiable on [-E, 0], ¢to have
two continuous derivatives on [-E, OJ and to satisfy some additional bound-
ary conditions obtainable from Formulas (2.4) and (2.5) and the relation
. 1 .. . .
(2.6) ¢(-r)= B [¢(0)-A¢(0)-f(O)].
This completes the proof of the theorem. D
Due to the smoothing property (i) of Theorem 2.2, many results from
ordinary differential equations are valid for retarded equations. The simi-
larities will become more apparent as further results are obtained.
1.3 Exponential estimates of x( ¢,f)
In this section, we derive an estimate on how the solution x( ¢,f) of Equa-

tion (2.1) depends on¢ and f. These estimates are basic to the application
of the Laplace transform and to obtaining an analogue of the variation-of-
constants formula.
To obtain these estimates, we need the following fundamental lemma.
Lemma 3.1. If u and a are real-valued continuous functions on [a, b], and
{3 :2': 0 is integrable on [a, bJ with
(3.1) u(t)::; a(t) + 1t {3(s)u(s) ds, a ::; t ::; b,
then
(3.2) u(t)::; a(t) + 1t {3(s)a(s) [exp it {3(T) dT] ds, a ::; t ::; b.
If, in addition, a is nondecreasing, then
(3.3) u(t) ~ a(t)exp(1t (3(s)ds), a~ t ~b.
Proof. Let R(t) = J: (3(s)u(s) ds. Then

dR
dt = (3u ~ (3a + (3R
and
d [R(s) exp(-
ds 18 f3)]
a ~ (3(s)a(s) exp(- 18f3).
a
Integrating from a to t, one obtains
and Inequality (3.1) yields Inequality (3.2). If a is nondecreasing, then

Inequality (3.2) yields
u(t) ~ a(t) [1 + 1t (3(s)(exp it f3) ds]
for t E [a, b]. A direct integration gives Inequality (3.3) and the lemma is
proved. D
Theorem 3.1. Suppose x(¢,!) is the solution of Equation (2.1) defined by

Theorem 2.1. Then there are positive constants a and b such that
(3.4) lx(¢, f)(t)l ~ aebt(l¢1 +lot lf(s)l ds), t2':0
where 1¢1 = sup_r:Se:So 1¢(8)1.

Proof. Since x = x(¢, f) satisfies Equation (2.1) fort 2': 0,
x(t) = ¢(0) + lot[Ax(s) + Bx(s- r) + f(s)] ds
fort 2': 0 and x(t) = ¢(t) fortE [-r, 0]. Therefore, fort 2': 0,
lx(t)l ~ 1¢1 +lot lf(s)l ds +lot IAIIx(s)l ds + j_tr IBIIx(s)l ds

~ (1 + IBir)l¢1 +lot lf(s)l ds +lot (IAI + IBI)Ix(s)l ds.
Applying Lemma 3.1, Formula (3.3), to this inequality, we obtain
1.4 The characteristic equation 17
lx(t)l ::; [(1 + IBir)l¢1 + lt lf(s)l ds] exp(IAI + IBI)t.

Since IBI ;::: 0, it follows that Inequality (3.4) is satisfied with a= 1 + IBir
and b = IAI + IBI. 0
Some immediate implications of Theorem 3.1 are the following. Since

Equation (2.1) is linear and solutions are uniquely defined by¢, it is obvious
that the solution x( ¢, 0) of the homogeneous equation
(3.5) x(t) = Ax(t) + Bx(t- r)

that coincides with¢ on [-r, OJ is linear in¢; that is, x(¢+'1', 0) = x(¢, 0) +
x( '1', 0) and x( a¢, 0) = ax(¢, 0) for any continuous functions ¢ and '1' on
[-r,OJ and any scalar a. For f = 0, Inequality (3.4) implies that x(¢,0)(t)
is continuous in¢ for all t; that is, x(·, O)(t) is a continuous linear functional
on the space of continuous functions on [-r, OJ. The Riesz representation
theorem then implies that the solution can be represented as a Stieltjes
integral.
By the remark after Theorem 2.1 and the proof of Theorem 3.1, Esti-
mate (3.4) is valid for locally integrable functions f. Using the same rea-
soning as earlier, the function x(O, ¢) is a solution of the nonhomogeneous
equation (2.1) with zero initial data. Estimate (3.4) shows that x(O, ·)(t) is a
continuous linear functional on the locally integrable functions. Therefore,
an integral representation of x(O, f)(t) is known to exist. Such a represen-
tation corresponds to the variation-of-constants formula.
This functional analytic approach will be exploited in detail in later
chapters for more general problems. However, for this simple equation, the
use of the Laplace transform is just as convenient and brings in some other
concepts that are important in the general theory.
1.4 The characteristic equation
The characteristic equation for a homogeneous linear differential difference

equation with constant coefficients is obtained from the equation by looking
for nontrivial solutions of the form e>.tc where c is constant. For example,
the scalar equation
(4.1) x(t) = Ax(t) + Bx(t- r)

has a nontrivial solution e>.tc if and only if
(4.2) h(.A) ~f .A- A- Be~>.r = 0.
We need some properties of the characteristic equation (4.2). In the

following, Re .A designates the real part of .A.
Lemma 4.1. If there is a sequence { Aj} of solutions of Equation ( 4.2) such

that f>..Jf ----+ oo as j ----+ oo, then Re AJ ----+ -oo as j ----+ oo. Thus, there is
a real number a such that all solutions of Equation (4.2) satisfy Re >.. < a
and there are only a finite number of solutions in any vertical strip in the
complex plane.
Proof. For any solution >.. of Equation (4.2),
Consequently, if f>..f ----+ oo, then exp( -rRe >..) ----+ oo, which implies the first
statement of the lemma. This also implies the existence of a as in the
lemma. Since h(>..) is an entire function, there can be only a finite number
of zeros of h(>..) in any compact set. These facts imply there are only a
finite number in any vertical strip in the complex plane and the lemma is
proved. D
Theorem 4.1. Suppose >.. is a root of multiplicity m of the characteristic

equation (4.2). Then each of the functions tk exp >..t, k = 0, 1, 2, ... , m- 1,
is a solution of Equation (4.1). Since Equation (4.1) is linear, any finite
sum of such solution is also a solution. Infinite sums are also solutions
under suitable conditions to ensure convergence.
Proof. If x(t) = tke>-.t, then
e->-.t[x(t)- Ax(t)- Bx(t- r)] = tk>.. + ktk- 1 - Atk- B(t- r)ke->-.r
= L (~) tk-j h(j) (>..)

k
j=O J
where the last expression is obtained by expanding (t-r)k by the binomial

theorem and making the observation that the coefficients are related to the
derivatives h(Jl(>..), h( 0 l(>..) = h(>..), of the function h(>..) as indicated.
If )... is a zero of h(>..) of multiplicity m, then h(>..) = kC 1 l (>..) = · · · =
h(m- 1)(>..) = 0. Consequently, x(t) = tke>-.t is a solution of Equation (4.1)
for k = 0, 1, ... , m- 1. This proves the first part of the theorem. The last
assertions are obvious and the theorem is proved. D
1.5 The fundamental solution
In the next section, we are going to apply the Laplace transform to obtain
the solution of the initial-value problem for the nonhomogeneous equation
(2.1). The characteristic equation (4.2) arises naturally as does the need for
a function whose Laplace transform is h- 1 (>..). It is actually possible to find
1.5 The fundamental solution 19
a solution of the homogeneous equation (4.1), called the fundamental solu-

tion of Equation (4.1), whose Laplace transform is h - l (.>,). This function
will now be defined without further motivation.
Let X (t) be the solution that satisfies Equation (4.1) for t ~ 0 and
satisfies the initial condition
(5.1) X(t)={O, t~O,

1, t- 0.
Our basic existence theorem does not apply to the initial data (5.1), but the
same proof as in Theorem 2.1 may be used to prove existence. Furthermore,
the exponential estimate in Theorem 3.1 is likewise valid. Also, it is clear
that X(t) is of bounded variation on any compact set.
Our next objective is to apply the Laplace transform to X(t). The
following two lemmas needed from the theory of the Laplace transform are
stated without proof.
Lemma 5.1. (Existence and convolution of Laplace transform). If f

[0, oo) ----+ IR is measurable and satisfies
(5.2) t E [0, oo),
for some constants a and b, then the Laplace transform .C(f) defined by
(5.3) £(!)(>..) = 1 00
e->-t f(t) dt
exists and is an analytic function of>.. for Re >.. > b. If the function f *g is
defined by f * g(t) = J;
f(t- s)g(s) ds, then .C(f *g)= .C(f).C(g).
The following notation is used
1 (c)
=lim-
T->oo 21Ti
1 lc+iT
c-iT
where cis a real number.
Lemma 5.2. (Inversion theorem). Suppose f : [0, oo) ----+ IR is a given func-
tion, b > 0 is a given constant such that f is of bounded variation on any
compact set, and t f----7 f( t) exp( -bt) is Lebesgue integrable on [0, oo). Then,
for any c > b,
(5.4) 1(c)
.C(f)(>..)e>-td>.. = { ~[f(t+) + f(t-)],
2f(O+),
t: 0,
t-0.
Theorem 5.1. The solution X(t) of Equation (4.1) with initial data (5.1) is
the fundamental solution; that is,
(5.5) .C(X)(A) = h- 1 (A).

Also, for any c > b,
(5.6) X(t) = { e>-.th- 1 (A) dA, t>O

J(c)
where b is the exponent associated with the bound on X(t) in Theorem 3.1,
IX(t)i :<::; aexp(bt), t
~ 0.
Proof. Since X(t) satisfies the exponential bounds in Theorem 3.1, .C(X)
exists and is analytic for Re A > b. Multiplying Equation (4.1) by e->-.t,
integrating from 0 to oo, and integrating the first term by parts, one obtains
Equation (5.5). Since X(t) is of bounded variation on compact sets and
continuous for t ~ 0, Relation (5.6) follows from the inversion formula
(5.4). The theorem is proved. D
Our next objective is to obtain very precise exponential bounds on

X(t) in terms of the maximum of the real parts of the solutions of the
characteristic equation.
Theorem 5.2. If a 0 = max{Re A: h(A) = 0}, then, for any a> ao, there is
a constant k = k(a) such that the fundamental solution X(t) satisfies the
inequality
(5.7) t ~ 0.
Proof. From Theorem 5.1, we know that
X(t) = { e>-.th- 1 (A) dA

J(c)
where cis some sufficiently large real number. We may take c > a. We want
to prove first that
(5.8) X(t) = { e>-.th- 1 (A)dA.

J(n)
To show this, consider the integration of the function e>-.th- 1 (A) around
the boundary of the box r in the complex plane with boundary L 1 M 1 L 2 M 2
in the direction indicated, where the segment £ 1 is the set {c + iT : - T :<::;
T :<::; T}, the segment £ 2 is the set {a + iT : - T :<::; T :<::; T}, the segment
M 1 is the set {a+ iT : a :<::; a :<::; c}, and the segment M 2 is the set
{a- iT, a:<::; a:<::; c} (see Fig. 1.1). Since h(A) has no zeros in this box,
it follows that the integral over the boundary is zero. Therefore, Relation
(5.8) will be verified if we show that
1.5 The fundamental solution 21
as T---+ oo.
Fig. 1.1.
Choose To so that
(1 + ;: ) 112 - ~(IAI + IBie-ar) ~~

for all T ~ T 0 . If T ~ T 0 and A E M 1; that is, A= a+ iT, a::::; a::::; c, and
T~ T0 , then
Therefore,
I
}Ml
r e>.th- 1(A) dAI ::::; ~ect(c-
T
a)---+ 0 as T---+ 00.
In the same way, the integral over M 2 ---+ 0 as T---+ oo. This proves Equation
(5.8).
Suppose T0 as above. If g(A) = h- 1(A)- (A- a 0 )-I, then
lg(A)I = I A+ ~e->.r- ao h-1(A)I ::::; T22 (IAI + IBie-ar +lao I)

-ao
for A= a+ iT, ITI ~To. Therefore,
1 (a)
lg(A)I dA < +oo and lj (a)
e>.tg(A) dAI ::::; K1eat, t > 0,
where K 1 is a constant. Furthermore, the integral f(a) e>.t(A- a 0 )- 1 dAis

known to exist and admit an estimate of the form
lj(a)
e>.t(A- ao)- 1 dAI ::::; K2eat, t > 0.
Consequently, using Equation (5.8), the proof of the theorem is complete

with the constant k given by K 1 + K 2 . D
The proof of Theorem 5.2 does not only provide a precise exponential
bound for the fundamental solution X(t), but also a method to expand X(t)
as a series of characteristic solutions Pj (t)e>.jt. Here Pj is a polynomial and
>.1 a root of the characteristic equation.
We proceed as follows. In the region {>. E <C : I>.- AI > IBie-Re>.r}
the characteristic equation has no roots. So, if we choose the box r in the
preceding proof such that
(5.9)
then, along the segments M 1 , M 2 , Estimate (5.9) holds. So
as T ----too.
In the same way, the integral over M 2 ----t 0 as T ----t oo. Since the zeros
of h(>.) have an asymptotic distribution, we can choose a sequence O:m,
m = 1, 2, ... , so that the line ReA = O:m does not contain a root of the
characteristic equation. Therefore, the Cauchy theorem of residues implies
(5.10)
where ).. 1 , ... , Akrn are the roots of the characteristic equation such that
Re >.1 > O:m· It is easy to see that
and that Pj (t )e>.j t is a solution of Equation (4.1) with initial function

¢1 (8) = P1 (B)e>.je, -r :::; e :::; 0, j = 1, 2, .... Since Equation (4.1) is
1
linear,
Yc,(t) = e>.th- 1 (>.) d)..
(a)
is a solution of Equation (4.1) as well. Further, in the same way as in the

proof of Theorem 5.2,
To analyze the series as a ____, -oo, it suffices to analyze the solution
Ya(t) = m->oo
lim Ya,. (t), t ~ 0.
1.6 The variation-of-constants formula 23
In this particular case, one can prove that Yo(t) = 0 for t > 0. In general,
Y0 (t) =/= 0, which is surprising since Yo is a solution that decays faster
than any exponential. Such solutions are called small solutions and will be
studied in Chapter 3. Series expansions will be studied in Chapter 7.
1.6 The variation-of-constants formula
In this section, we obtain a representation of the solution x(¢, f) of the

nonhomogeneous equation (2.1) in terms of the fundamental solution X(t)
of the previous section. In particular, it will be shown that the solution is
given by
(6.1) x(¢, f)(t) = x(¢, O)(t) +fat X(t- s)f(s) ds, t ~ 0.
The representation in this form will be referred to as the variation-of-

constants formula. This terminology is not a precise description but it is
easy to remember because of the similarity to ordinary differential equa-
tions.
It is also possible to represent the solution x(¢, 0) of the homogeneous
equation in terms of the fundamental solution X(t). The precise result is
contained in the following theorem.
Theorem 6.1. The solution x(¢, f) of the nonhomogeneous equation (2.1)

can be represented in the form (6.1). Furthermore, fort~ 0,
(6.2) x(¢, O)(t) = X(t)¢(0) + B /_or X(t- e- r)¢(0) de.
Proof. As remarked earlier, there are many ways to obtain this result. The
present proof is based on the Laplace transform and other proofs in more
general situations will be given later. To apply the Laplace transform, the
function f should be bounded by an exponential function. For any compact
interval [0, T], one can redefine f as a continuous function so that it is
zero outside the interval [0, T + E], E > 0. Then f will be bounded by an
exponential function. If we prove the theorem for this case, then the theorem
will be valid for 0 :::; t :::; T. Since T is arbitrary, the result will be proved.
Therefore, we may assume f is bounded by an exponential function.
From Theorem 3.1, it is valid to apply the Laplace transform to each
term in Equation (2.1). Let x = x(¢, f) be a solution of this equation.
Multiplying this equation by e->-.t, Re .>.. > c, sufficiently large, integrating
from 0 to oo, and integrating the first term by parts, we obtain
Using the inversion forniula in Lemma 5.2, the expression for x(t) is
(6.3)
x(t) = 1(c)
eAth- 1 (-X) [¢(0) +Be-Ar 1°-r
e-A9¢(0)d0
+loco e-At f(t) dt] d.X.
Evaluation of the last term in this expression is easy since it represents

merely the inverse Laplace transform of X from Theorem 5.1. Therefore,
the convolution theorem says this term is J~ X(t- s)f(s) ds. Since the first
two terms depend only on ¢, this proves that the variation-of-constants
formula (6.1) is valid.
To prove that the first two terms in Equation (6.3) coincide with Equa-
tion (6.2), we proceed as follows. The first term is X(t)¢(0) by the inversion
formula and Theorem 5.1.
If w: [-r, oo)---+ [0, 1] is defined by w(O) = 0, 0;::: 0, w(O) = 1 for 0 < 0
and the definition of¢ is extended to (-r,oo) by defining ¢(0) = ¢(0) for
0;::: 0, then
e-Ar 10-r
e-Af)¢(0) dO= roo e-As¢(-r + s)w(-r + s) ds
Jo
= .C(¢( -r + ·)w( -r + ·)).
Consequently, the second term in Equation (6.3) is B multiplied by the

inverse Laplace transform of the product of the Laplace transforms of X
and ¢( -r + ·)w( -r + ·). The convolution theorem, therefore, implies this
second term for t > 0 is equal to
Blot X(t- s)¢( -r + s)w( -r + s) ds = B for X(t- s)¢( -r + s) ds.
The latter relation follows from the definition of X and w. Letting s = r+O,
one obtains the complete formula (6.2) and the theorem is proved. D
Theorem 6.1 has many interesting implications. For example, it implies

the nontrivial result that the exponential behavior of the solutions of the
homogeneous equation (4.1) is determined by the characteristic equation
(4.2) as stated in the following result.
Theorem 6.2. Suppose a 0 = max{Re .X: h(A) = 0} and x(¢) is the solution
of the homogeneous equation (4.1), which coincides with ¢ on [-r, 0]. Then,
for any a> a 0 , there is a constant K = K(a) such that
(6.4) t;::: 0, 1¢1 = sup 1¢(0)1.

-r:-s;e:-s;o
In particular, if a 0 < 0, then one can choose ao <a < 0 to obtain the fact
that all solutions approach zero exponentially as t---+ oo.
1.7 Neutral differential difference equations 25
Proof. The proof is an obvious application of the exponential bounds for X

in Theorem 5.2 and Formula (6.2). D
Knowing that the asymptotic behavior of the solutions of the homo-

geneous differential difference equation is governed by the solutions of the
characteristic equation, it is possible to use the variation-of-constants for-
mula to determine the asymptotic behavior of perturbed nonlinear systems
of the form
(6.5) x(t) = Ax(t) + Bx(t- r) + f(x(t), x(t- r)).

In fact, if we assume existence for the initial-value problem, then the so-
lution x = x(¢) of Equation (6.5) with initial data ¢ on [-r, 0] is given
by
(6.6) x(t) = y(t) + 1t X(t- s)f(x(s), x(s- r)) ds
where y = y( ¢) is the solution of the homogeneous equation (4.1) and X is

the fundamental solution of that equation.
If f(x, y) satisfies f(O, 0) = 0, 8f(O, 0)/B(x, y) = 0 and o:o < 0 in
Theorem 6.2, then one can prove the analogue of the Poincan~-Liapunov
theorem on stability with respect to the first approximation. The method
of proof is similar to the proof of the corresponding result for ordinary
differential equations. The proof is omitted since a more general result will
be given later and our main purpose in this chapter is only to acquire
intuition.
Finally, we remark that other perturbation results can be obtained in
a similar manner.
1. 7 Neutral differential difference equations
In this section, we introduce another class of equations depending on past

and present values but that involve derivatives with delays as well as the
function itself. Such equations historically have been referred to as neutral
differential difference equations. The presentation will not be as detailed as
the one for the retarded equations of the previous section. We concentrate
only on those proofs that are significantly different from the ones for re-
tarded equations. We also point out some of the differences between neutral
equations and retarded equations.
The model nonhomogeneous equation is
(7.1) x(t)- Cx(t- r) = Ax(t) + Bx(t- r) + f(t).

where A, B, C, and rare constants with r > 0, C =/= 0 and f is a continuous
function on IR. The corresponding homogeneous equation is
(7.2) x(t)- Cx(t- r) = Ax(t) + Bx(t- r).

For this equation, it is a little more difficult to define the concept of
a solution and the appropriate space of initial data. In the retarded case,
the initial space did not play an important role in the initial-value problem
since the solution becomes continuously differentiable for t 2: 0 and, in
particular, it is continuous. For Equation (7.1) with C "!=- 0, we will see that
such smoothing of the solution does not occur with increasing t.
Let us define the initial-value problem as follows. Suppose¢ is a given
continuously differentiable function on [-r, OJ. A solution x = x(¢, f) of
Equation (7.1) through ¢ is a continuous function x defined on [-r, oo)
with x(t) = ¢(t) fortE [-r, 0], x continuously differentiable except at the
points kr, k = 0, 1, 2, ... and x satisfying Equation (7.1) except at these
points.
Using the same proof as in the proof of Theorem 2.1, one can show
there always exists a solution of Equation (7.1) through¢.
For C "!=- 0, Equation (7.1) implies that the solution can never have
more derivatives than the initial function ¢. Also, the solution x = x( ¢, f)
will in general have a discontinuous derivative at kr, k = 0, 1, 2, .... In fact,
suppose
¢(0) i=- C¢( -r) + A¢(0) + B¢( -r) + f(O).
Then the solution x(t) has a discontinuous derivative at t = 0. From Equa-
tion (7.1), this implies x(r) is discontinuous since C "!=- 0. The same reason-
ing applies for all values kr, k = 0, 1, 2, ....
On the other hand, if
(7.3) ¢(0) = c¢( -r) + A¢(0) + B¢( -r) + f(O)

then the solution x(t) has a continuous derivative at t = 0. Equation (7.1)
can again be used to show that x(t) has a continuous derivative for all
t 2: -r. Consequently, Relation (7.3) is necessary and sufficient for the
solution x to have a continuous derivative for all t 2: -r.
Notice also that C "!=- 0 implies that one can obtain a unique solution
of Equation (7.1) on ( -oo, 0], which coincides with¢ on [-r, 0]. In fact, if
y(t) = x(t- r), then y satisfies the differential equation
B A 1 1
i;(t) = - C y(t)- C x(t) + C x(t)- C f(t).
One may prove existence with the argument in the proof of Theorem 2.1
using the variation-of-constants formula in the backward direction. These
two results are summarized in the following theorem.
Theorem 7.1. If C "!=- 0 and ¢ is a continuously differentiable function on

[-r, 0], then there exists a unique function x : ( -oo, oo) ---> IR that coincides
with¢ on [-r, 0], is continuously differentiable and satisfies Equation (7.1)
except maybe at the points kr, k = 0, ±1, ±2, .... This solution x can have
no more derivatives than ¢ and is continuously differentiable if and only if

Relation (7.3) is satisfied.
For Equation (7.1), one can continue to develop the theory with the
preceding definition of a solution. However, if one wishes to consider many
delays and more general functional equations, the exceptional set where the
derivative is not required to exist becomes impossible to describe. There are
many ways to overcome this difficulty and we now discuss one.
Rewrite Equation (7.1) as
d
(7.4) dt [x(t)- Cx(t- r)] = Ax(t) + Bx(t- r) + f(t).
In this form, it is at least meaningful to consider the following initial-value

problem. Suppose¢ is a continuous function on [-r, OJ. A solution of Equa-
tion (7.4) through ¢ is a continuous function defined on [-r, oo), which
coincides with ¢ on [-r, 0] such that the difference x(t) - Cx(t- r) is
differentiable and satisfies Equation (7.4) fort~ 0.
Can one obtain such a solution? It is not difficult to prove that a
solution exists by letting x(t) = (expAt)y(t) and observing that y satisfies
the equation
d
dt [y(t)- e-ArCy(t- r)] = e-Ar(AC + B)y(t- r) +e-At f(t).
This latter equation may now always be integrated by the method of steps
in the direction of increasing t and C =f 0, in the direction of decreasing t.
These remarks are summarized in the following theorem.
Theorem 7.2. If¢ is a continuous function on [-r, 0], then there is a unique
solution of Equation (7.4) on [-r, oo) through ¢. If C =f 0, this solution
exists on ( -oo, oo) and is unique.
Equation (7.4) can be considered as a generalization of the retarded

equations (C = 0) as well as a generalization of difference equations [A =
B = f = 0, ¢(0) = C¢( -r)]. This is certainly sufficient motivation to
consider this class of equations in its own right. Also, we shall see that
many generalizations are easily accomplished. Therefore, in the remainder
of this section, we will consider only Equation (7.4) with initial data¢ that
is continuous. The corresponding homogeneous equation is
d
(7.5) dt [x(t)- Cx(t- r)] = Ax(t) + Bx(t- r).
As the theory evolves, one cannot keep from observing that retarded
equations are very similar to ordinary differential equations and parabolic
partial differential equations and that neutral equations resemble difference
equations and hyperbolic partial differential equations.
Theorem 7.3. Let x(</J, f) be the solution of Equation (7.4) given in Theorem
7.2. Then there are positive constants a and b such that
(7.6) lx(¢, f)(t)l::; aebt [I <PI+ lot lf(s)l ds], t 2: 0,
where
I<PI = sup I<P(B)I.
-r::;o::;o
Proof. Let a:= (1 + 2ICI) and fJ = IAI +lEI. If xis a solution of Equation
(7.4) for t 2: 0, then
x(t) = ¢(0)- C¢( -r) + Cx(t- r) +lot [Ax(s) + Bx(s- r)] ds

+lot f(s) ds.
If 1</JI = SUP-r<O<o 1</J(B)I and y(t) = SUP-r<O<O lx(t+B)I, and 0::; t::; r/2,
then - - - -
lx(t)l :S o:l¢1 +lot lf(s)l ds + fJ lot y(s) ds.

Since a: 2: 1 and x(t) = ¢(t) for t ::; 0, it follows that
y(t) :S o:l¢1 +lot lf(s)lds + fJ lot y(s) ds, < ~

0 -< t -2
An application of Lemma 3.1 yields
y(t) :S [o:l¢1 +lot lf(s)l ds] expfJt, 0 -< t -< ~

2
Also, the same argument implies the following estimate
y(t)::; [a:y(T) + 1t lf(s)l ds] expfJ(t- T), 0<

- T
r
- T + -.
- t <
< 2
Let 1 > fJ be such that a: exp(fJ -1 )r /2 < 1 and let us prove that
y(t)::; [o:l¢1 +lot lf(s)l ds] exp1t, O<t<-

- - 2
kr
where k is an integer. We know this relation is true for k = 1. Assume that

it is true for some k 2: 1. If r/2::; t::; (k + 1)r/2, then
r
y(t)::; [o:y(t--) +
2
it
t-r/2
lf(s)l ds] exp -fJr
2
and, therefore, the induction hypothesis implies

1. 7 Neutral differential difference equations 29
y(t) :::; [a{ al¢1 + 1 t-r/2

lf(s)l ds }e~'(t-r/ 2 ) +
1t
lf(s)l ds]ef3r/ 2
1
0 t-~2
t-r/2 1t
:::; [al¢1 + lf(s)l ds]e~'t + e,Br/ 2 lf(s)l ds.
0 t-r/2
Since t :2: r /2 and 'Y > {3, we have
y(t):::; [al¢1 + 1t lf(s)l ds] exp'"{t
and the theorem is proved. D
Our next objective is to discuss the characteristic equation for Equa-

tion (7.5). If Equation (7.5) has a solution e~'t, then >. must satisfy the
characteristic equation
(7.7) H(>.) ~f >.(1- Ce-Ar) -A- Be-Ar = 0.
Lemma 7.1. There is a real number a such that all solutions of Equation
(7.7) satisfy Re >.<a. If C =/:- 0, then all solutions of Equation (7.7) lie in
a vertical strip {[3 < Re >. < a} in the complex plane. If C =/:- 0 and there
is a sequence {>.j} of solutions such that 1>-il ~ oo as j ~ oo, then there
is a sequence { >.j} of zeros of
(7.8) 1- ce-Ar =0
such that Aj - >.j ~ 0 as j ~ oo. Also, one can show there always exists
such a sequence {>.j} when C =/:- 0.
Proof. For >. =/:- 0, Equation (7.7) is equivalent to the equation
eAr(l-~)-C-~ =0.
This relation implies the existence of a as stated for any C. If C =/:- 0,

then the same relation implies >. must lie in a vertical strip in the complex
plane. In this strip, Iexp(>.r)l is bounded. Therefore, if 1>-1 ~ oo, then
eAr - C ~ 0. This implies the statement concerning Aj and >.j. The proof
of the last statement is left as an exercise. D
One important observation is immediate from Lemma 7.1. For C =/:- 0,

the roots of Equation (7.8) are given by
(7.9) >.=InC+ 2k7ri k = 0, ±1, ±2, ....

r 4 '
If InC > 0, then Lemma 7.1 implies there are always infinitely many so-
lutions of Equation (7.5) which approach oo at an exponential rate and,
therefore, one can never have stability of Equation (7.5).
To exploit this further, suppose one attempts to consider Equation

(7.5) for r small as some type of perturbation of the ordinary differential
equation
d
(7.10) dt (1 - C)x(t) = (A+ B)x(t),
and suppose all solutions of this equation approach zero as t ----+ oo; that is,
(A+B)/(1-C) < 0. Lemma 7.1 and Formula (7.9) imply there are infinitely
many solutions of Equation (7.5) approaching infinity at an exponential rate
if ICI > 1 and there can be at most a finite number of values ..\ with Re
>. > 0 if ICI < 1. Furthermore, from Equation (7.7), it is easy to see that
for any compact set U in the complex plane and any neighborhood V of
the complex number (A+ B)/(1- C), there is an r 0 sufficiently small so
that Equation (7.7) has no solution in U - V and exactly one solution
in V for 0 < r ::; r 0 . Therefore, except for the root of Equation (7.7),
which approaches (A+ B)/(1- C) as r----+ 0, the moduli of the other roots
approach +oo as r----+ 0. Thus, for r sufficiently small, ICI < 1, Lemma 7.1,
and Formula (7.9) imply all roots of Equation (7.7) satisfy Re >. < -15 < 0
for some 15 > 0. If these roots govern the asymptotic behavior of solutions,
then at least asymptotic stability of the ordinary differential equation (7.10)
is preserved for Equation (7.5) if ICI < 1.
For Equation (7.5) with C = 0, that is, for the retarded equation,
Lemma 7.1 implies there are never more than a finite number of>. with
Re >. > 0. As in the argument for ICI < 1, one concludes that all roots of
the characteristic equation except the one close to A + B are in the left
half-plane for r sufficiently small.
In order to put more emphasis on these important remarks, we sum-
marize them in the following corollary.
Corollary 7.1. Suppose the supremum of the real parts of the roots of Equa-
tion (7. 7) govern the asymptotic behavior of the solutions of Equation (7.5).
If there is a 15 > 0 such that every solution of Equation (7. 7) satisfies
x(t)exp/5t----+ 0 as t----+ oo, then it is necessary that ICI < 1. If ICI < 1
and every solution of the ordinary differential equation (7.10) approaches
zero, then there are 15 > 0 and r 0 > 0 such that every solution satisfies
x(t)exp/5t----+ 0 as t----+ oo for 0::; r < ro.
To define the fundamental solution X(t) of Equation (7.5), one pro-

ceeds in a manner similar to the one used for the retarded case. The initial
data for X(t) will be given by Condition (5.1), which is discontinuous at
zero. From our experiences so far with neutral equations, it is to be ex-
pected that this discontinuity will persist at multiples of r for the solution,
if it exists. Therefore, we define a solution X of Equation (7.5) with initial
data (5.1) as one for which X(t) - CX(t- r) is continuous and satisfies
Equation (7.5) for t ::::: 0 except at the points kr, k = 0, 1, 2, .... With this
definition, and in the same manner as before, it is now possible to prove

that a unique solution X(t) of Equation (7.5) exists that satisfies initial
data (5.1).
The function X(t) will actually have a continuous first derivative on
each interval (kr, (k + 1)r), k = 0, 1, 2, ... , the right- and left-hand limits
of X(t) exist at each of the points kr, k = 0, 1, 2, .... Therefore, X(t) is of
bounded variation on each compact interval and satisfies
(7.11) X(t)- CX(t- r) = AX(t) + BX(t- r)

for t -1- kr, k = 0, 1, 2, .... These assertions are proved easily from the fact
that X (t) satisfies the integral equation
X(t) = 1 + CX(t- r) +fat [AX(s) + BX(s- r)] ds, t ~ 0.
As in the proof of Theorem 7.3, one shows that X(t) satisfies the
inequality
t E IR.
Since X(t) has an exponential bound one can define the Laplace transform
of X. It is easy to verify that .C(X) = H- 1 (>,).
Theorem 7.4. The solution X(t) of Equation (7.11) with initial data X(t) =
0, t < 0, X(O) = 1, is the fundamental solution of Equation (7.4); that is,
.C(X) = H- 1 (>. ).
The same proof as in the proof of Theorem 6.1 is easily modified

to show that the solution x(¢, f) of Equation (7.1) is obtained from the
variation-of-constants formula,
(7.12) x(¢, J)(t) = x(¢, O)(t) +fat X(t- s)f(s) ds.

It is also possible to express the solution x( ¢, 0) in terms of the funda-
mental solution X. The formula is more complicated than the one for the
retarded equation and will now be derived. Following the same proof as in
I:
the retarded case, one obtains
x(¢, O)(t) = X(t)[¢(0)- C¢( -r)] +B X(t- ()- r)¢(()) d()
+ C¢( -r + t)w( -r + t) + C [or X(t- ()- r)cp(()) d()

where w(()) = 1 for () < 0 and w(()) = 0 for () ~ 0. If t ~ r, the term
involving w is not present. If t < r, then this term is C¢( -r + t), which
is precisely the value of the Stieltjes integral - J~~r d[X(t- ()- r)]¢(()).
Therefore, if we make use of the Stieltjes integral, the relation for x(¢,0)(t)
can be written as
(7.13)
x(¢, O)(t) = X(t)[¢(0)- C¢( -r)] +B I: X(t- 0- r)¢(0) dO
- C [or d[X(t- 0- r)]¢(0).

The results are summarized in the following theorem.
Theorem 7.5. If X(t) is the fundamental solution of Equation (7.5), then

the solution x(¢, f) of Equation (7.1) is obtained from the variation-of-
constants formula (7.12) with x(¢,0) as given in Formula (7.13).
Of course, the usefulness of the representation formulas (7.12) and

(7.13) depends on knowing that the asymptotic behavior ofthe fundamen-
tal solution X(t) is determined by the solutions of the characteristic equa-
tion. For the retarded equation, these estimates were obtained in a rather
straightforward manner knowing that the Laplace transform of X wash(..\).
For the neutral Equation (7.5), we know that .C(X) = H- 1 (..\) and so
it should be possible to obtain similar estimates.
The analysis is not a trivial modification of the proof of Theorem 5.2.
Theorem 7.6. If ao = sup{Re ..\ : H(..\) = 0}, then, for any a > ao, there
is a constant k = k(a) such that the fundamental solution X of Equation
(7.5) satisfies the inequality
Var[t-r,t] X ~ keat, t ~ 0,
where Var[t-r,t] X denotes the total variation of X on [t - r, t].
Proof. Since a > a 0 , it follows from Lemma 7.1 that ar > ln ICI· Con-
sequently there is an interval I containing a such that 1 - ce->.r is uni-
formly bounded away from zero in the strip S = {..\ E C : Re ..\ E I}.
In a manner similar to the proof of Theorem 5.2, one can show that
X(t) = f(a) e>.t H- 1 (..\) d..\. To estimate the value of this integral, observe
that
1 1 A+ Be->.r
H(..\) = ..$1 - Ce->.r) + ..\(1- Ce-M)H(..$.
On the line Re ..\=a, the integral
1.(a)
A+ Be->.r
e>. t
--,-----,----,---:-::-=--c:-:-
..$1 - Ce->.r)H(..$
is absolutely convergent since the kernel is like ..\ - 2 . Therefore, this part
of the inverse Laplace transform of H- 1 (..\) admits an estimate of the type
specified in the theorem.
1. 7 Neutral differential difference equations 33
The term involving >.- 1 (1- Ce->.r)- 1 is more difficult. In the stripS,
the function (1- Ce->.r)- 1 is analytic and, if).= f3 + iw, (3 E I, then this
function is periodic of period 21r / r. Therefore, this function possesses an
absolutely convergent Fourier series
u- L
00
ce->.r)-1 = hke>.kr, ). E S,
k=-oo
L
00
lhk ie,Bkr < oo, f3 E I.

k=-oo
Consequently, if t > 0, t + kr -/=- 0, then
1(a)
1
>.(1- Ce->-r)
e>.td\
/\
= ~
~
k=-oo
h
k
1 (a)
e>-(t+kr) \ -1d\
/\ /\
= ~
~
a(t+kr)>O
h
k·
Finally,
L L L
00
hkl -::;eat lhkieakr-::; eat lhkleakr,

a(t+kr)>O a(t+kr)>O k=-oo
and we have shown that X (t) has an estimate of the type stated in the
theorem.
To assert that the variation of X has an estimate of the desired type, we
use the difference differential equation (7.11) to show that X(t) is bounded
by a constant times exp at except at the points kr, k = 0, 1, 2, .... If y(t) =
X(t), then y(t) -Cy(t-r) = p(t), where p(t) = AX(t) +BX(t-r) satisfies
an inequality of the form lp(t)l -::; (constant)expo:t. If y(t) = z(t)expo:t,
q(t) = p(t) exp( -o:t), then
z(t)- C'z(t- r) = q(t),

where IC'I = IC exp( -o:r)l < 1 and lq(t)l is bounded. It is now very easy
to show that this equation has all solutions bounded fort 2': 0. To compute
the variation of X(t), we also must consider the jump in X(t) at kr. We
observe that
and, thus the jumps are bounded by a constant times eat. This proves the
theorem. 0
An interesting corollary on the asymptotic behavior of the solutions of

Equation (7.5) can now be stated.
Corollary 7.2. If ao = sup{Re .A: .A(1- Ce->.r) =A+ Be->.r} and x(¢) is
the solution of Equation (7.5), which coincides with¢ on [-r, 0], then, for
any a> a 0 , there is a constant K = K(a) such that
t :2:: 0, 1¢1 = sup 1¢(0)1.
-r::;o::;o
In particular, if ao < 0, then all solutions of Equation (7.5) approach zero
exponentially.
Proof. This is an immediate consequence of the representation formula
(7.13) and Theorem 7.6. D
Perturbation results similar to the ones for retarded equations can also
be stated.
1.8 Supplementary remarks
The presentation in this chapter was certainly influenced by the discussion

in the first five chapters of Bellman and Cooke [1]. The treatment of the
neutral equations uses also the papers by Hale and Meyer [1] and Henry
[1].
In this chapter, we have discussed only scalar equations with one delay.
It is essentially a matter of notation to generalize the theory to the matrix
case and to include any finite number of delays. Of course, any discussion of
specific properties of the characteristic equation will be much more difficult
since this equation will be of the form
N
(8.1) ao(.A) +L aj(.X)e->.ri = 0
j=l
where the aj(.A) are polynomials of degree S:: nand a 0 (.A) is a polynomial
of degree n.
The theory of this chapter will be generalized even more in subsequent
pages with the emphasis being on qualitative and geometric properties of
the solutions. An application of this theory to a concrete example will often
require very detailed knowledge of the solutions of equations of the form
of Equation (8.1). Therefore, it is absolutely essential that one be aware of
existing methods for analyzing Equation (8.1). The theory of this equation
will not be developed in this book, but there is an excellent presentation in
Chapters 12 and 13 of Bellman and Cooke [1]. In the Appendix, we present
some of the available theory on determining conditions for the zeros of
Equation (8.1) to have negative real parts.
We briefly mentioned the very interesting question of the expansion
of solutions of linear differential difference equations as an infinite series of
the form
1.8 Supplementary remarks 35
LPJ(t)e>-jt
j
where the Aj are the roots of the characteristic function and the PJ are
polynomials. Some results on this question are contained in Chapter 7. Ad-
ditional results can be found in Bellman and Cooke [1], Banks and Manitius
[1], Gromova and Zverkin [1], Verduyn Lunel [2,3,4], and Zverkin [3].
As remarked earlier, our objective in later chapters of this book is to de-
velop a comprehensive qualitative theory for general functional differential
equations of retarded and neutral type. Before entering into this theory, it
is instructive to discuss in an intuitive manner the types of function spaces
that should be considered as appropriate spaces of initial functions.
Consider the scalar retarded equation
(8.2) i:(t) = f(t, x(t), x(t- r))
where r > 0 and f : IR3 --t IR is continuous. For a given a E IR and
¢ : [-r, 0] --t IR one would certainly want ¢to satisfy enough smoothness
conditions to ensure that finding a solution of Equation (8.2) for t ;:::: a
satisfying x( a + B) = ¢(B), -r :::; B :::; 0, would be equivalent to finding a
solution of the integral
(8.3)
x(t) = ¢(0) + lt f(s, x(s), x(s- r)) ds, t;:::: a,
x(a +B)= ¢(B), -r:::; B:::; 0.

There is very little difficulty finding spaces of initial functions that have
this property, and the space C of continuous functions certainly will be
sufficient. Even if some space other than continuous functions is used for
initial data, then the solution lies in C for t ;:::: a + r. Therefore, for the
fundamental theory, the space of initial data does not play a role that is
too significant. However, in the applications, it is sometimes convenient
to take initial functions with fewer or more restrictions. Even though our
subsequent discussion will center around the initial space of continuous
functions, we will occasionally amplify on this last remark.
Consider the neutral equation
(8.4) i:(t) = f(t,x(t),x(t- r),i:(t- r))
where r > 0 and f : IR4 --t IR is continuous. Since a specification of x at
t and x, i; at t - r uniquely determines i:(t), it is natural to specify the
following initial-value problem. Suppose a is a given real number and ¢
is a given function on [a - r, a] that is continuous together with its first
derivative. A solution of Equation (8.4) through (a,¢) is a function x defined
on an interval [a- r, a+ A), A> 0, which coincides with¢ on [a- r, a] has
a continuous first derivative except at the points a+kr for all k = 0, 1, 2, ...
for which a+ kr belongs to [a, a+ A). A theory of Equation (8.4) along
this line is developed in Bellman and Cooke [1].
One shortcoming of this definition of the initial-value problem is that it

cannot be generalized to the situation in which there is general dependence
of x(t) on values of x(s) for s:::; t. Even in Equation (8.4), and even more so
when r depends on t, there are great difficulties in discussing the dependence
of solutions on the initial data (a,¢). Other objections arise if one tries to
develop a geometric theory for Equation (8.4) in the same spirit as for
ordinary differential equations. There have been many papers devoted to
the formulation of the initial-value problem and the reader may consult
Turdy Sem. Teor. Diff. Urav. Otkl. Argumentom, Vols. 1-8 (1962-1973)
and the survey articles of Zverkin, Kamenskii, Norkin, and El'sgol'tz [1],
Kamenskii, Norkin, and El'sogl'tz [1], and Mishkis and El'sgol'tz [1].
A very significant contribution to this question was made by Driver
[2], who gave a formulation that has also been generalized by Melvin [1, 2].
We illustrate the ideas for Equation (8.4). Suppose 1> is a given absolutely
continuous function on [a-r, a]. A solution of Equation (8.4) through (a, 1>)
is an absolutely continuous function defined on an interval
[a - r, a+ A), A>O,
coinciding with 1> on [a- r, a] and satisfying Equation (8.4) almost every-
where on [a, a+ A). Of course, in order for this initial-value problem to
make sense, the function f must satisfy the following property: If x is any
given absolutely continuous function on [a - r, a + A) and if
F(t) = f(t,x(t),x(t- r),x(t- r)), a:::; t <a+ A,
then the function F must be locally integrable on [a, a+ A). A satisfactory
function f is
(8.5) f(t, x, y, z) = g(t, x, y)z + h(t, x, y),
which is linear in z (the term that corresponds to x(t- r) in Equation
(8.4)). Iff satisfies Equation (8.5) and the initial-value problem is defined as
earlier, then a theory of existence, uniqueness, and continuous dependence
on the initial data is developed in Driver [2].
One can also consider the case where the initial value 1> is continuous
and has a pth-power integrable first derivative with 1 :::; p :::; +oo, includ-
ing +oo. A solution of Equation (8.4) is required to lie in the same class.
This theory has been developed by Melvin [1, 2]. For further discussion of
topologies for neutral equations, see Driver [5].
For certain classes of neutral equations that occur frequently in the
application, it is also possible to obtain a well-posed initial-value problem
with the initial space being only the continuous functions. We have already
encountered this situation for linear equations in Section 1.7. Other non-
linear equations of a special form can also be discussed in a similar setting.
For example, consider the equation
(8.6) x(t) = g(t, x(t- r))x(t- r) + h(t, x(t), x(t- r))
where g and h are continuous functions of their arguments and g( t, x) has

a continuous first derivative in t. If
(8.7) G(t,x) = 1x g(t,s)ds,
then Equation (8.6) can be written as

d
(8.8) dt [x(t)- G(t,x(t- r))] = H(t,x(t),x(t- r))
where H(t, x, y) = h(t, x, y) - 8G(t, y)j8t. It is now possible to pose the

following initial-value problem for Equation (8.8). Suppose cjJ is a given
continuous n-vector function on [a - r, a]. A solution of Equation (8.8)
through (a,¢) is a continuous function x defined on [a- r, a+ A), A> 0,
coinciding with cjJ on [a- r, a] such that the function x(t)- G(t, x(t- r)),
not x(t), is continuously differentiable on [a, a+ A) and satisfies Equation
(8.8) on [a, a + A). A theory in this direction was initiated in Hale and
Meyer [1] and Cruz and Hale [1]. It will receive more attention later.
The neutral equations introduced later will be modeled after Equation
(8.8). The main reason for taking this approach is that retarded equa-
tions and difference equations will be included without imposing too many
smoothness conditions on the initial data. Furthermore, this class is math-
ematically simpler than the general equation of neutral type and yet it is
sufficiently general to include many applications.
2
Functional differential equations:
Basic theory
In this chapter, we introduce a general class of functional differential equa-

tions that generalize the differential difference equations of Chapter 1. The
basic theory of existence, uniqueness, continuation, and continuous depen-
dence for retarded equations will be developed in the first five sections. In
the last two sections, we introduce a fairly general class of neutral differen-
tial equations for which one can extend the basic theory.
2.1 Definition of a retarded equation
Suppose r 2: 0 is a given real number, lR = (-oo,oo), lRn is ann-

dimensionallinear vector space over the reals with norm I · I, C([a, b], lRn)
is the Banach space of continuous functions mapping the interval [a, b]
into lRn with the topology of uniform convergence. If [a, b] = [-r, OJ we
let C = C([-r, 0], lRn) and designate the norm of an element ¢ in C by
1¢1 = SUP-r<O<o 1¢(0)1. Even though single bars are used for norms in dif-
ferent spaces, no confusion should arise. If
a E lR, A 2:0, and x E C([-a- r, a+ A], lRn),
then for any t E [a, a+ A], we let Xt E C be defined by Xt(B) = x(t +B),
-r::::; B::::; 0. If Dis a subset of lR x C, f: D ---7lRn is a given function and
"·" represents the right-hand derivative, we say that the relation
(1.1) x(t) = J(t, xt)
is a retarded functional differential equation on D and will denote this equa-

tion by RFDE. If we wish to emphasize that the equation is defined by j,
we write the RFDE(f). A function x is said to be a solution of Equa-
tion (1.1) on [a- r, a+ A) if there are a E lR and A > 0 such that
x E C([a- r, a+ A), lRn), (t, Xt) E D and x(t) satisfies Equation (1.1) for
t E [a, a+ A). For given a E lR, ¢ E C, we say x(a, ¢,f) is a solution of
Equation (1.1) with initial value¢ at a or simply a solution through (a,¢)
2.2 Existence, uniqueness, and continuous dependence 39
if there is an A > 0 such that x(a, ¢,f) is a solution of Equation (1.1) on

[a- r, a+ A) and Xa(a, ¢,f)=¢.
Equation (1.1) is a very general type of equation and includes ordinary
differential equations (r = 0)
x(t) = F(x(t)),
differential difference equations

x(t) = f(t,x(t),x(t- T1(t)), ... ,x(t- Tp(t)))
with 0 :::; Tj ( t) :::; r, j = 1, 2, ... , p, as well as the integra-differential equa-
j_:
tion
x(t) = g(t,O,x(t+O))dO.
Much more general equations are also included in Equation (1.1).

We say Equation (1.1) is linear if f(t, ¢) = L(t)¢ + h(t), where L(t)
is linear; linear homogeneous if h = 0 and linear nonhomogeneous if h =I 0.
We say Equation (1.1) is autonomous if f(t, ¢) = g(¢) where g does not
depend on t. The proof of the following lemma is obvious, using Lemma 2.1
of the next section.
Lemma 1.1. If a E lR, ¢ E C are given, and f(t, ¢) is continuous, then

finding a solution of Equation ( 1.1) through (a, ¢) is equivalent to solving
the integral equation
Xa = c/J
(1.2)
x(t) = ¢(0) + 1t f(s, X8 ) ds, t ?:. a.
2.2 Existence, uniqueness, and continuous dependence
In this section, we give a basic existence theorem for the initial-value prob-
lem of Equation (1.1) assuming that f is continuous. Also, a rather general
result on continuous dependence will be given as well as a simple result on
uniqueness.
The ideas in this section are very simple, but the notation naturally will
involve some complications. To prove the existence of the solution through
a point (a,¢) E lR x C, we consider an a > 0 and all functions x on
[a- r, a+ a] that are continuous and coincide with ¢ on [a- r, a]; that
is, Xa = ¢. The values of these functions on [a, a+ a] are restricted to the
class of x such that lx(t)- ¢(0)1 :::; (3 fortE [a, a+ a]. The usual mapping
40 2. Functional differential equations: Basic theory
T obtained from the corresponding integral equation is defined and it is

then shown that o: and (3 can be so chosen that T maps this class into itself
and is completely continuous. Thus, Schauder's fixed-point theorem implies
existence.
Continuous dependence is slightly more difficult because the mapping
T depends on parameters and one must discuss the dependence of the fixed
points of T on these parameters. If one is attempting to prove continuous
dependence at a value .A0 of the parameter, then the usual procedure is to
assume a unique fixed point at .A0 and then prove that the "compactness"
property ofT is uniform with respect to compact sets containing .Ao. This
is the pattern followed here.
Our first observation is the following lemma.
Lemma 2.1. If x E C([a- r, a+ o:), lRn) then Xt is a continuous function

oft for t in [a, a + o:).
Proof. Since x is continuous on [a- r, a+ o:), it is uniformly continuous and
thus for any E > 0 there is a 8 > 0 such that lx(t)- x(r)l < E if It- rl < 8.
Consequently, fort in [a, a+ o:), It- rl < 8, we have
lx(t+B) -x(r+B)I < E
for all B in [-r, OJ. This proves the lemma. 0
To bring out the ideas in the proof of existence as well as the results
of subsequent sections, it is convenient to introduce some notation and to
prove a few technical lemmas.
For any (a,¢) E lR x C, let'¢ E C([a- r, oo), lRn) be defined by
(2.1) </Ja = </J, ;{>(t +a) = ¢(0), t :2:: 0.
Suppose x is a solution of Equation (1.1) through (a,¢). If x(t +a)

;f>(t +a)+ y(t), t :2:: -r, then Lemma 1.1 implies y satisfies
Yo= 0
(2.2)
y(t) = 1t f(a + s, 'i>a+s + Ys) ds,
Conversely, if y is a solution of this equation, then one obtains a solu-

tion x of Equation (1.1) by this transformation. Therefore, finding a solution
of (1.1) is equivalent to finding an o: > 0 and a function y E C([-r, o:), lRn)
such that Equation (2.2) is satisfied for 0 ~ t ~ o:.
If V is a subset of lR x C, then C(V, lRn) is the class of all functions
f: V----+ lRn that are continuous and C 0 (V, lRn) ~ C(V, lRn) is the subset of
bounded continuous functions from V to lRn. The space C 0 (V, lRn) becomes
a Banach space with the norm
(2.3) lflv = sup lf(t, ¢)1.

(t,¢)EV
For any real a and /3 define
I a = [0, a], B13 = {7/J E C : 11/JI :::; /3},

(2.4)
A(a,/3) = {y E C([-r,a],ffin): Yo= 0, Yt E B13, t E Ia}·
Lemma 2.2. Suppose fl ~ ffi x C is open, W ~ fl is compact and f 0 E

C(fl, ffin) is given. Then there exists a neighborhood V ~ fl of W such
that f 0 E C 0 (V, ffin), there exists a neighborhood U ~ C 0 (V, ffin) of f 0 and
positive constants M, a, and f3 such that
(2.5) If(a, ¢)1 < M for (a,¢>) E V and fEU.
Also, for any (a 0 , ¢>0 ) E W, we have (a 0 + t, Yt + '¢uo+to) E V fort E I a

andy E A(a,/3).
Proof. Since W is compact and f 0 is continuous, there is a constant M such
that lf0 (a 0 ,¢0 )1 < M for (a 0 ,¢>0 ) E W. Furthermore, for the same reason,
there are positive constants a, jj, and E such that
lf 0 (a 0 + t, ¢>0 + 1/1)1 < M - E for (a 0 , ¢>0 ) E Wand (t, 1/1) E Ia x B13.
If V = {(a 0 + t,¢> 0 + ¢) : (a 0 ,¢0 ) E W, (t,¢) E Ia x B13}, then

jO E C 0 (V, ffin) and there is a neighborhood U ~ C 0 (V; ffin) of f 0 such
that Inequality (2.5) is satisfied.
To prove the last assertion of the lemma, suppose 0 < /3 < jj and choose
a so that a< a and l'¢uo+to -¢0 1< jj-/3 for all (a 0 , ¢>0 ) E W, t E ! 01 • Since
W is compact, this last choice is possible. Therefore, IYt + '¢uo+to - ¢ 0 1 <
f3+fj-f3 = jj for yEA( a, /3). From the manner in which V was constructed,
the proof of the lemma is complete. D
The next lemma will be used to apply fixed-point theorems for exis-
tence and continuous dependence of solutions of Equation (1.1).
Lemma 2.3. Suppose fl ~ ffixC is open, W ~ fl is compact, f 0 E C(fl, ffin)

is given, and the neighborhoods U and V and constants M, a, and /3 are
the ones obtained from Lemma 2.2. If
T: W XU X A( a, /3)-> C([-r, a], ffin)
T(a,¢>,f,y)(t) = 0, t E [-r,O],
T(a, ¢>, f, y)(t) = 1t f(a + s, '¢u+s + Ys) ds,

then Tis continuous and there is a compact set Kin C([-r,a],ffin) such
that
T : W x U x A( a, ,6) ---+ K.
Furthermore, if M a ~ ,6, then
T: W xU x A(a,,B)---+ A(a,,B).
Proof. It is clear that T: W xU x A( a, ,6)---+ C([-r, a], IRn). Also, Relation

(2.5) implies
IT(a,cf;,J,y)(t) -T(a,cf;,J,y)(r)l ~ Mlt-rl

IT(a,cf;,J,y)(t)l ~ Ma
for all t, T E Ia. If
K = {g E C([-r,a],IRn): lg(t)- g(r)l ~Mit- rl, lg(t)l ~ Ma},

then K is compact, T: W xU x A(a,,B) ---+ K. If Ma ~ ,6, then K ~
A( a, ,6).
It remains only to show that Tis continuous. Suppose (ak, ¢k, Jk, yk) E
W xU x A(a,,B) and (ak,¢k,Jk,yk)---+ (a 0 ,¢0 ,j0 ,y0 ) E W xU x A(a,,B)
as k---+ oo. Since T(ak, ¢k, Jk, yk) E K and K is compact, there is a subse-
quence we designate with the same symbol and a "'( E K such that
T(ak,cf;k,fk,yk)-+"'f ask-+oo.
Since
(2.6) fk(ak + s, ¢a-k+sk + y:) ---+ f 0 (a 0 + s, ¢a-o+ 8o + y~)

for all s E Ia and all of these functions are uniformly bounded by Lemma
2.2, the Lebesgue dominated convergence theorem implies
for all t E Ia. This implies the limit of any convergent subsequence is inde-
pendent of the subsequence. But since every subsequence has a convergent
subsequence, this obviously implies the sequence itself converges. Therefore,
T is continuous and the lemma is proved. D
Even with the proof given earlier, the operator T is continuous under
weaker conditions than the ones stated in Lemma 2.3. In fact, only Ex-
pression (2.6) was used. This implies the convergence of Jk to f 0 was not
required in the uniform way as specified by the norm I · Iv. This will affect
the generality of our later result on continuous dependence. Our purpose
is to give only the basic results that are essential for the development to
follow in spite of the temptation to develop a more comprehensive theory

of existence, etc.
To prove our basic existence theorem, we need the Schauder fixed- point
theorem. If U is a subset of a Banach space X and T : U -'> X, then T is
said to be completely continuous if T is continuous and for any bounded set
B ~ U, the closure of TB is compact. We now state the Schauder theorem
without proof.
Lemma 2.4. (Schauder fixed-point theorem). If U is a closed bounded convex

subset of a Banach space X and T : U -'> U is completely continuous, then
T has a fixed point in U.
Theorem 2.1. (Existence). Suppose f2 is an open subset in lR x C and f 0 E

C(fl,lRn). If(a,¢) E fl, then there is a solution of the RFDE(J 0 ) passing
through (a,¢). More generally, if W ~ f2 is compact and f 0 E C (f2, lRn) is
given, then there is a neighborhood V ~ f2 of W such that f 0 E C 0 (V, lRn),
there is a neighborhood U ~ C 0 (V, JRn) of f 0 and an a > 0 such that
for any (a,¢) E W, f E U, there is a solution x(a, ¢,f) of the RFDE(J)
through (a,¢) that exists on [a - r, a+ a].
Proof. For the first part, take W ={(a,¢)}, a single point. Lemma 2.3 and
Schauder's fixed-point theorem imply that T( a,¢, f 0 , ·) has a fixed point
in A(a,,8) since A(a,,8) is a closed bounded convex set of C([-r,a],lRn).
This yields a solution of the RFDE(J 0 ) by Formula (2.2) with f replaced
by f 0 . To obtain the last statement of the theorem, simply apply the same
reasoning but use the general form of Lemma 2.3. D
Theorem 2.2. (Continuous dependence). Suppose f2 ~ lR x C is open,

(a 0 , ¢ 0 ) E fl, f 0 E C(fl, lRn), and x 0 is a solution of the RFDE(J 0 ) through
(a 0 , ¢ 0 ) which exists and is unique on [a 0 -r, b]. Let W 0 ~ f2 be the compact
set defined by
W 0 = {(t, x~): t E [a 0 , b]}
and let V 0 be a neighborhood of W 0 on which f 0 is bounded. If (ak, ¢k, fk),
k = 1, 2, ... satisfies ak -'> a 0 , ¢k -'> ¢ 0 , and lfk - f 0 Iva -'> 0 as k -'> oo,
then there is a k 0 such that the RFDE(Jk) for k ~ k 0 is such that each
solution xk = xk (ak, q;k, fk) through (ak, ¢k) exists on [ak - r, b] and xk -'>
x 0 uniformly on [a 0 - r, ,8]. Since all xk may not be defined on [a 0 - r, b],
by xk -'> x 0 uniformly on [a 0 - r, b], we mean that for any E > 0, there is a
k1(E) such that xk(t), k ~ k1(E), is defined on [a 0 - r + E, b], and xk-'> x 0
uniformly on [a 0 - r + E, b].
Proof. The proof is an easy application of Lemma 2.3. In fact, the set
W 0 U {(ak, ¢k): k = 1, 2, ... } is compact. By taking
for ko sufficiently large and restricting o: and f3 so that the resulting neigh-
borhood V of Lemma 2.2 belongs to V 0 , one may now apply Lemma 2.3
and Theorem 2.1 in the following manner.
From Theorem 2.1, each of the solutions xk = xk((Tk, ¢k, fk) through
((Tk, ¢k) exists on [(Tk -r, (Tk +o:] where o: is independent of k. Furthermore,
Lemma 2.3 asserts that the yk(t) = xk((Tk + t) -'J}((Tk + t) belong to a
compact set K of C([-r, o:), lRn). Therefore, there is a subsequence labeled
the same way such that yk converges uniformly to some function y* on
[-r, o:]. Since yk = T((Tk, ¢k, Jk, yk) and Tis continuous by Lemma 2.3, this
implies y* = T( (To, ¢ 0 , f 0 , y0 ) = y0 • Since every subsequence of the sequence
{yk} has a convergent subsequence that must converge to y 0 , it follows that
the entire sequence converges to y 0 . Translating these remarks back into xk
gives the result stated in the theorem for the interval [(T 0 - r, (To+ o:]. The
proof is completed by successively stepping intervals of length o:, which we
know is possible by Theorem 2.1. D
Theorem 2.3. Suppose fl is an open set in lRx C, f : [l ___. lRn is continuous,_

and f(t, ¢) is Lipschitizian in¢ in each compact set in fl. If ((T, ¢) E fl,
then there is a unique solution of Equation (1.1) through ((T, ¢).
Proof. Define Ia and B13 as in Equation (2.4) and suppose x and y are
solutions of Equation (1.1) on [(T- r, (T + o:] with x,. = ¢ = y,.. Then
x(t)- y(t) =it [f(s,x.)- f(s, Ys)] ds, t ~ (T,
x,.- y,. = 0.
If k is the Lipschitz constant of f(t, ¢) in any compact set containing the
trajectories {(t,xt)}, {(t,yt)}, t E Ia, then choose 0: so that ka < 1. Then,
fort E fc,,
lx(t)- y(t)l :=::; 1t klxs- Yslds :=::; ka sup lxs- Ysl

u u:s;s:s;t
and this implies x(t) = y(t) for t E ! 0 . One completes the proof of the
theorem by successively stepping intervals of length 0:. D
2.3 Continuation of solutions
Suppose f in Equation (1.1) is continuous. If x is a solution of Equation

(1.1) on an interval [(T, a), a > (T, we say x is a continuation of x if there is
a b > a such that x is defined on [(T- r, b), coincides with x on [(T- r, a),
and x satisfies Equation (1.1) on [(T, b). A solution xis noncontinuable if no
such continuation exists; that is, the interval [(T, a) is the maximal interval
2.3 Continuation of solutions 45
of existence of the solution x. The existence of a noncontinuable solution

follows from Zorn's lemma. Also, the maximal interval of existence must be
open.
Theorem 3.1. Suppose n is an open set in lR x C and f E C(n, lRn). If x

is a noncontinuable solution of Equation (1.1) on [O"- r, b), then, for any
compact set W inn, there is a tw SUCh that (t, Xt) ¢: W for tw :::; t <b.
Proof. The case b = oo is trivial so we suppose b finite. Consider first
the case r = 0 (an ordinary equation). Since W is compact, Theorem 2.1
implies there is an a > 0 such that the equation has a solution through any
(c, y) E W that exists at least on [c, c + a]. Now suppose the assertion of
the theorem is false; that is, there is a sequence (tk, x(tk)) E W, y E lRn,
(b, y) E W such that tk -+ b-, x(tk) -+ y ask-+ oo. Using the fact that f is
bounded in a neighborhood of (b, y), the function xis uniformly continuous
on [O", b) and x(t) -+ y as t -+ b-. There is obviously an extension of x to
the interval [O", b +a]. Since b +a > b, this is a contradiction. The proof for
the case r = 0 is complete.
If the conclusion of the theorem is not true for r > 0, then there are a
sequence of real numbers tk-+ b- ask-+ oo and a 'ljJ E C such that
as k -+ oo. Thus, for any f. > 0,

lim sup lxtk(0)-'1/J(O)I=O.
k->oo IIE[-r,-E]
Since Xt(O) = x(t + 0), -r :::; 0 :::; 0, and r > 0, this implies x(b + 0) =
'1/J(O), -r :::; 0 < 0. Hence limt-;b- x(t) exists and x can be extended to a
continuous function [O"- r, b] by defining x(b) = '1/J(O). Since (b, xb) E n,
one can find a solution of Equation (1.1) through this point to the right of
b. This contradicts the noncontinuability hypothesis on x and proves the
theorem. D
Corollary 3.1. Suppose n is an open set in lR x C and f E C(n,JRn). If

x is a noncontinuable solution of Equation (1.1) on [O"- r, b) and W is the
closure of the set {(t, Xt) : O":::; t < b} in lR x C, then W compact implies
there is a sequence {tk} of real numbers, tk -+ b- as k -+ oo such that
(tk,Xtk) tends to an ask-+ 00. Ifr > 0, then there is a '1/J E C such that
(b, '1/J) E an and (t, Xt)-+ (b, '1/J) as t-+ b-.
Proof. Theorem 3.1 implies that W does not belong to n and proves the
first part of the corollary. If r > 0, then the same argument as in the proof
of Theorem 3.1 implies limt_,b- x(t) exists and, thus, x can be extended as
a continuous function on [O"-r, b]. Clearly, (b, Xb) E an and (t, Xt) -+ (b, Xb)
as t-+ b-. D
Theorem 3.2. Suppose n is an open set in IR X c, f : n ~ IRn is completely

continuous; that is, f is continuous and takes closed bounded sets of n into
bounded sets of IRn, and x is a noncontinuable solution of Equation (1.1)
on [a- r, b). Then, for any closed bounded set u in IR X c, u inn, there
is a tu such that (t, Xt) ¢. U for tu ::; t < b.
Proof. The case r = 0 is contained in Theorem 3.1. Therefore, we suppose
r > 0 and it is no restriction to take b finite. Suppose the conclusion of the
theorem is not true. Then there is a sequence of real numbers tk ~ b- such
that (tk,Xtk) E U for all k. Since r > 0, this implies that x(t), a-r::; t < b
is bounded. Consequently, there is a constant M such that lf(r,¢)1::; M
for (r,¢) in the closure of {(t,xt): a::; t < b}. The integral equation for
the solutions of Equation (1.1) imply
fx(t+r)-x(t)i=l t l t+r
f(s,xs)dsi:SMr
for all t, t + T < b. Thus, x is uniformly continuous on [a- r, b). This

implies {(t, Xt) : a::; t < b} belongs to a compact set inn. This contradicts
Theorem 3.1 and proves the theorem. D
Theorem 3.2 gives conditions under which the trajectory (t, x) in IR x C

of a noncontinuable solution of [a, b) approaches the boundary of Jl as
t ~ b-. The approach to the boundary of Jl was described by saying that
the trajectory must leave and remain outside every closed bounded set in
n. If the condition that f is completely continuous is not imposed, then
it is conceivable that the trajectory {(t, Xt) : a ::; t < b} itself is a closed
bounded subset of Jl; that is, the curve (t, x(t)) oscillates so badly as a
subset of IR x IRn that there are no limit points of (t, Xt) in IR x C as
t ~b-. An explicit example illustrating this fact will now be given.
Let L1(t) = t 2 and select two sequences {ak}, {bk} of negative numbers,
a1 < a2 < · · · , b1 < b2 < · · · , ak ~ 0, bk ~ 0 as k ~ oo such that
For example, choose bk = -2-k, k = 1, 2, ....

Let 'ljJ(t) be an arbitrary continuous differentiable function satisfying
'ljJ(t) = { +1, fort ~n ( -oo, a1], [b2~ a2k+l], k = 1, 2, ... ,

-1, fort m [b2k-b a2k], k- 1, 2, ...
'1/J'(t) # 0, t E (ak, bk), k = 1, 2, ....

Let H be the set of points (t, x) such that fxl < 1-t. These inequalities
are equivalent to
X +t < 1 if X > 0,
-x + t < 1 if x < 0.
2.3 Continuation of solutions 47
Thus His the wedge in Fig. 2.1. We now define a function h(t,x) on H.
On the graph of the curve '¢(t), let
h(t- L1(t),'¢(t- L1(t))) = '¢'(t),
where the prime denotes the derivative, -oo < t < 0. The function h is
continuous on the graph of'¢. For any tin (ak, bk), k ~ 2 (i.e., a point of
increase or decrease on the graph), t- L1(t) E [bk-1, ak]· Fort in ( -oo, b1],
t - L1(t) E ( -oo, a 1]. Therefore, h = 0 for any t in ( -oo, b1], (ak, bk),
k ~ 2, and, in particular, h = 0 on all points of increase or decrease of the
graph of the curve '¢(t). Now continue the function h(t, x) in any manner
whatsoever as long as it remains continuous and is equal to zero in the
square P : ltl + lxl :::; 1.
'I' (t)
X+t =1
x-t = 1
Fig. 2.1.
Now consider the equation
(3.1) x(t) = h(t- L1(t),x(t- L1(t))), t < 0 and L1(t) = t 2 .
Choose a < a 1 and let r = a- min{(t- t 2 ) : a :::; t :::; 0}. We consider the
initial-value problem starting at a. The function x(t) = '¢(t) is a solution
of this equation fort < 0 and is a noncontinuable solution on [a- r, 0).
If the right-hand side of Equation (3.1) is denoted by f(t,xt), t E IR,
Xt E C ( [-r, 0], IR), then f (t, ¢>) does not map closed bounded sets of f? =
IR x C([-r,O],IR) into bounded sets. In fact, the set {(t,'¢t): t ~ 0} is a
bounded set and it is closed since there are no sequences tk ----+ 0 such that
'¢tk converges.
This strange behavior is one of the ways in which the noncompactness of

the unit ball in C([-r,O],IRn), r > 0 can influence the solution of RFDE.
The particular property exhibited in this example is actually a general
phenomenon. In fact, one can prove the following result.
Theorem 3.3. Suppose x : [-r, a) ___, IRn, r > 0, a finite, is an arbitrary

bounded continuously differentiable function satisfying the property that x(t)
does not approach a limit as t ___, a-. Then there exists a continuous func-
tion f : C ___, IRn such that x is a noncontinuable solution of the RFDE(f)
on [-r, a).
Proof. Suppose x as stated and let A = {xt : t E [0, a)}. By hypothesis,
the set A is closed in C. Let g : A ___, IRn be defined by g(xt) = x(t),
t E [0, a). The function g is continuous on A. Therefore, by a generalization
of the Tietze extension theorem due to Dugundji [1], there is a continuous
extension f of g to all of C. It is clear that x is a noncontinuable solution
of the RFDE(f) and the proof is complete. D
2.4 Differentiability of solutions

In Theorem 2.2, sufficient conditions were given to ensure that the solution
x(a, ¢,f) on a RFDE(f) depends continuously on (a,¢, f). In this section,
some results are given on the differentiability with respect to (a,¢, f).
If [l is an open set in IR x C, let CP(fl, IRn), p 2: 0, designate the
space of functions taking [l into IRn that have bounded continuous deriva-
tives up through order p with respect to ¢ on D. The space CP(fl, IRn)
becomes a Banach space if the norm is chosen as the supremum norm over
all derivatives up through order p. The norm will be designated by I · lv·
For our main theorem, we make use of some basic results on the depen-
dence of fixed points of contraction mappings on parameters. The proofs
are omitted.
Definition 4.1. Suppose U is a subset of a Banach space X and T: U ___,X.

The mapping T is said to be a contraction on U if there is a .\, 0 ::; ,\ < 1,
such that
ITx- Tyl ::; .\lx- Yl for all x, y E U.
If V is also a subset of a Banach space Y and T : U x V ___, X, then T is
said to be a uniform contraction if there is a 0 ::; ,\ < 1, such that
IT(x, v)- T(y, v)l ::; .\lx- Yl for all x, y E U and v E V.
Lemma 4.1. (Contraction mapping principle). If U is a closed subset of a

Banach space X and T : U ___, U is a contraction, then T has a unique fixed
point in U.
2.4 Differentiability of solutions 49
Lemma 4.2. If U is a closed subset of a Banach space X, V is a subset

of a Banach spaceY, T: U x V----+ U is a uniform contraction, and T is
continuous, then the unique fixed point x(v) ofT(·, v) in U is continuous in
v. Furthermore, if U, V are the closures of sets U 0 , V 0 and T(x, v) has con-
tinuous first derivatives in x, v, then x(v) has a continuous first derivative
with respect to v. The same conclusion holds for higher derivatives.
Theorem 4.1. Iff E CP(f.?, IRn), p ~ 1, then the solution x(a, </>, f)(t) of
the RFDE(f) through (a,¢) is unique and continuously differentiable with
respect to (¢,f) for t in any compact set in the domain of definition of
x( a,¢, f). Furthermore, for each t ~ a, the derivative of x with respect to ¢,
Dq,x(a, </>, f)(t) is a linear operator from C to IRn, Dq,x(a, ¢,!)(a) =I, the
identity, and Dq,x(a, ¢, f)'lj;(t) for each 'lj; inC satisfies the linear variational
equation
(4.1) y(t) = Dq,f(t,xt(a,<J>,f))yt.

Also, for each t ~a, Dtx(a,<J>,f)(t) is a linear operator from CP(f.?,IRn)
into IRn, Dtx(a, ¢,!)(a)= 0, and Dtx(a, ¢, f)g(t) for each g E CP(f.?, IRn)
satisfies the nonhomogeneous linear variation equation
(4.2) i(t) = Dq,f(t, Xt(a, </>, f))zt + g(t, Xt(a, ¢, !)).
Proof. Since p ~ 1, it follows from Theorem 2.3 that the solution x

x(a, ¢,f) of the RFDE(f) through (a,¢) is unique. Let the maximal interval
of existence of x be [a- r, a+ w) and fix b < w.
Our first objective is to show that x(a, ¢, f)(t) is continuously differ-
entiable with respect to¢ on [a- r, a+ b]. There is an open neighborhood
U of¢ such that x(a, 'lj;, f)(t), 'lj; E U, is defined for t E [a- r, a+ b]. If
W = {(t,xt) : t E [a, a+ b]}, then W is compact. Using the notation of
Section 2.2, we can determine M, a, /3, U, and Vas in Lemma 2.2. Choose
a so that M a ~ j3 and ka < 1, where k is a bound of the derivative of
f with respect to¢ on f.?. If x(a + t) ='¢(a+ t) + y(t), t E Io" then y is
a fixed point of the operator T(a, ¢,f) of Lemma 2.3. On the other hand,
the restriction on a, (3 implies that T( a,¢, f) takes A( a, /3) into itself for
each a, j3 and is a contraction. Furthermore, the contraction constant is
independent of (a,¢,/) E V x U. Since the mapping T(a, </>,f) is easily
shown to be continuously differentiable in f.?, it follows from Lemma 4.2
that the fixed point y = y(a, ¢,f) is continuously differentiable in f.?. The
same proof shows that x(a, ¢, f)(t) is continuously differentiable in f for
t E [a, a+ a]. Using the fact that the basic interval [a- r, a+ b] is compact,
one completes the proof of the differentiability.
Knowing that x(a, ¢,f) is continuously differentiable with respect to
¢, f, one can use the interval equation for x to easily obtain Formulas (4.1)
and (4.2). D
The next question of interest is the differentiability of x with respect

to u under the same hypotheses on f as in Theorem 4.1. In general, this
derivative will not exist unless the function f (t, ¢) satisfies some additional
smoothness properties in t. To see this, consider the following example,
(4.3) x(t) = a(t)x(t- 1)

where a is a continuous function. If x( u, ¢) is the solution of Equation (4.3)
through (u, ¢) and u < t 0, then
x(u + h, cp)(t) = ¢(0) + 1t a(s)x(u + h, cp)(s- 1) ds

a+h
= ¢(0) + 1t a(s)cp(s- u- h- 1) ds,
a+h
x(u, cp)(t) = ¢(0) + 1t a(s)x(u, cp)(s- 1) ds
= ¢(0) + 1t a(s)cp(s- u- 1) ds.

Therefore,
lim x(u + h, cp)(t)- x(u, cp)(t) = -a(t)cp(t- u- 1)

h--->0 h
l . 1ta(s+h)-a(s) 'i"s-u-1
+1m "'( )ds
h--->0 <T h
and, without some additional smoothness of a, the expression on the right-
hand side will not exist. If a is such that its derivative a is integrable, then
ax( u, ¢) (t) 1au exists and satisfies
ax(~:)(t) = -a(t)cp(t- u- 1) + 1t a(s)cp(s- u- 1) ds.
It would be interesting to discuss the general problem of when
ax(u, ¢,f)/au
exists. To carry out this investigation the following formula for t > u + h,
h > 0, is useful:
x(u + h, ¢, f)(t)- x(u, ¢, f)(t) = x(u + h, ¢, f)(t)
- x(u + h, Xa+h(u, ¢,f), f) (t).
For the ordinary differential equation (r = 0), this last formula shows im-
mediately that
ax(u, ¢, f)(t) = - ax(u, ¢, f)(t) f(u, ¢).

au a¢
2.5 Backward continuation 51
Finally, we remark that the pth-derivatives of x(a, ¢,f) with respect

to ¢, f exist under the hypotheses of Theorem 4.1, but the proof of this
fact will not be given.
2.5 Backward continuation
In ordinary differential equations with a continuous vector field, one can

prove the existence of a solution through a point (a, xo) defined on an
interval [a - a, a+ a], a > 0; that is, the solution exists to the right and
left of the initial t-value. For an RFDE, this is not necessarily the case. We
will state some sufficient conditions for existence to the left of the initial
t-value.
Definition 5.1. Suppose D <,;;; IR x C is open and f E C(D, IRn). We say

a function x E C([a- r- a,a],IRn), a> 0, is a solution of Equation
(1.1) on [a- r- a, a] through (a,¢) for (a,¢) ED if Xa =¢and for any
a 1 E [a- a, a], (a 1 , xaJ E D and x is a solution of Equation (1.1) oa
[a 1 - r, a] through (a 1 , Xa 1 ). We sometimes refer to such a solution as the
backward continuation of a solution through (a,¢).
Definition 5.1 is very natural and says only that a function defined on
[a- r- a, a] is a solution of Equation (1.1) on this interval if it has the
property that it will satisfy the equation in the forward direction of t no
matter where the initial time is chosen.
General results on backward continuation are very difficult to prove
although the ideas are relatively simple. To motivate the definitions to
follow, let us consider a simple example
(5.1) i:(t) = a(t)x(t- 1).
If a = 0 and ¢ is a given function in C and there exists a backward continu-

ation of a solution through (0, ¢),then it is necessary that¢ be continuously
differentiable on a small interval ( -E, 0] and ¢(0) = a(O)¢( -1). Conversely,
if this condition is satisfied and a(t) -j. 0 fortE ( -E, OJ, then one can define
1 .
x(t- 1) = a(t) ¢(t), t E ( -E, OJ
and x will be a solution of Equation (5.1) on ( -r - E, OJ with x 0 = ¢.

Therefore, there is a backward continuation through (0, ¢ ).
We wish to generalize this idea to a general RFDE(f). The important
remark in the example was a(t) -j. 0 fortE ( -E, 0]; that is, the evolution of
the system x(t) actually used the information specified at x(t- 1). In the
general case, the manner in which f(t,¢) varies with ¢(-r) is determined
by a "coefficient" that depends on t and ¢. The precise specification of this

coefficient leads to the definition of "atomic at -r", given later.
A careful examination of the technical details shows that the proof of
existence is also a natural generalization of the preceding remarks.
For the statement of the main result, we need some definitions. For
Banach spaces X and Y, £(X, Y) is the Banach space of bounded linear
mappings from X toY with the operator topology. If L E £(C,lRn), then
the Riesz representation theorem implies there is an n x n matrix function
'fJ on [-r, 0] of bounded variation such that
For any such 'TJ, we always understand that we have extended the definition
to IR so that 'TJ(O) = 'TJ(-r) for 0:::; -r, 'fJ(O) = 'fJ(O) for 0 ~ 0.
Definition 5.2. Let A be an open subset of a metric space. We say L : A ~

£(C,IRn), has smoothness on the measure if for any f3 E IR, there is a
scalar function ')'(.A, s) continuous for .A E A, s E IR, ')'(.A, 0) = 0, such that
if L(.A)¢ = f~r d'fJ[(.A, 0)]¢(0), A E A, 0 < s, then
(5.2) I lim
h--+0+
1(3+s + 1{3-h
f3+h {3-s d['fJ(.A,0)]¢(0)1:::; 'Y(.A,s)l¢1.
If f3 E IR and the matrix A(.A;f3,L) = 'fJ(.A,f3+) -'T](.A,/3-) is nonsingular at
.A= .A0 , we say L(.A) is atomic at f3 at .Ao. If A(.A: /3, L) is nonsingular on a
set K s;:; A, we say L(.A) is atomic at f3 on K.
Lemma 5.1. If L E C(A, £(0, IRn)), then L has smoothness on the measure.
Proof. Let elements of A be denoted by t. Iff is a function of bounded
variation on an interval I, let Var1 f denote the total variation off on I.
If 'fJ = ('TJij) is a matrix of bounded variation on [-r, 0], let
n
II'TJII = 1~\a.fn L Var[-r,O] 'f/ij

- - j=1
n N
= 1~ffnL sup L I'TJij(rk)- 'f/ij(Tk-1)1
- - j=1 P[-r,O] k=1
where the supremum is taken over all partitions of [-r, 0].

If L(t)¢ = f~r d['fJ(t, 0)]¢(0), then there is a constant k > 0 such that
kii'TJ(t,·)ll:::; IIL(t)ll:::; II'TJ(t,·)ll· Since L E C(A,£(C,IRn)), for any tEA
and E > 0, there is aD > 0 such that II'TJ(t, ·)- 'TJ(r, ·)II < E if It- rl < 8.
This means for any [a, b] s;:; [-r, 0] and any i, j = 1, 2, ... , n,
(5.3) if It- rl < 8.

For a fixed i and 0 < s ~ r, let

n
1-li(t, s) = L:var[,e+,,e+s]u[,e-s,,e-] 'T/ij(t, ·).
j=l
From Inequality (5.3), we have /-Li(t, s) is continuous in t uniformly with

respect to s. Also, 1-li(t, s) is nondecreasing ins, is uniformly bounded ins
and 1-li(t, s)---+ 0 ass---+ 0.
In 1R?, consider the set {(s,y) : y = 1-li(t,s), s E (O,oo)}, and its
closed convex hull ri(t). Let 'Yi(t, s) = sup{y: (s, y) E ri(t)}. Then 'Yi(t, s)
is continuous in t uniformly with respect to s. Also, for each fixed t, it is
continuous in s with 'Yi(t, s) ---+ 0 as s ---+ 0. If we define 'Yi(t, 0) = 0, then
'Yi(t, s) is jointly continuous in t, s. If 'Y(t, s) = max1::;i::;n 'Yi(t, s), then 'Y
satisfies the conditions of the lemma and the proof is complete. D
In the following our interest lies in the case where A = il ~ lli x C; that
is, L E C(il,.C(C,llin)). If D: Q---+ llin has a continuous first derivative
with respect to ¢, then Lemma 5.1 implies D¢ has smoothness on the
measure. This remark justifies the following definition.
Definition 5.3. Suppose Q ~ lli x Cis open with elements (t, ¢).A function
D : Q ---+ llin (not necessarily linear) is said to be atomic at (3 on Q if
D is continuous together with its first and second Frechet derivatives with
respect to¢; and D¢, the derivative with respect to¢, is atomic at (3 on fl.
If D( t, ¢) is linear in ¢ and continuous in (t, ¢) E lli x C,
D(t)¢ =[or d[ry(t, 0)]¢(0)

then A(t, ¢, (3) = A(t, (3) is independent of¢ and
A(t, (3) = ry(t, /3+) - ry(t, (3- ).
Thus, D(t) is atomic at (3 on lli x C if det A(t, (3) # 0 for all t E lli. In
particular, if (3 # 0, (3 E [-r, 0], D(t)¢ = ¢(0) + B(t)¢((3), then A(t, (3) =
B(t) and D(t) is atomic at (3 on lli x C if det B(t) # 0 for all t E lli.
For the following results, the smoothness conditions on the function D
in the definition of atomic at (3 are more severe than necessary. What is
needed for a given f is the existence of functions L( t,¢), g( t, ¢, '1/J), t:( t, ¢, s),
continuous such that t:(t, ¢, 0) = 0, Relation (5.2) is satisfied, and
(5.4) f(t, ¢ + '1/J) = f(t, ¢) + L(t, ¢)'1/J + g(t, ¢, '1/J)
(5.5) lg(t, ¢, '1/J)- g(t, ¢,~)I ~ t:(t, ¢, s)I'I/J- ~I, 1'1/JI, 1~1 ~ s
for all (t,¢) E fl, (t,¢+'1/J) E fl, (t,¢+~) E fl, s 2::0. If A(t,cp,(J;L) is
the matrix defined in Definition 5.2, then for any (t, ¢) E fl, there is an
s0 = s0 (t,¢>) such that A(t,¢>,(3;L) and E are related by
(5.6) ldetA(t,¢,(J;L)I-e(t,¢,'1j;)> O fori'I/JI:Sso.
If ldetA(t,¢,(J;L)I 2:: a> 0 for all (t,¢) Efland e(t,¢,s)::::; Eo(s) for all
(t, ¢) E fl, then Relation (5.6) holds uniformly on fl. We will make use of
this remark in the examples.
We can now state
Theorem 5.1. (Backward continuation). If fl is an open set in 1R x C,

f : fl ----t 1Rn is atomic at -r on fl, (lT, ¢>) E fl, and there is an a, 0 < a < r,
such that /p(B) is continuous for() E [-a, 0], ¢(0) = f(lT, ¢), then there is
an a > 0 and a unique solution x of the RFDE(J) on [lT- r- a, lT] through
[lT, ¢).
Proof. A function x is a solution of the RFDE(J) on [lT- r- a, lT] through
(lT,cp) if and only if x" = ¢>, (t,xt) E fl, t E [lT- a,O"], and
(5.7) f(t, Xt) = i:(t) = ¢(t- lT), t E [lT- a,O"]
since 0 < a < r.

For any a > 0, let J;: [-r- a, 0] ____, 1Rn be defined by '¢;(t) = ¢(t),
t E [-r, 0], '¢;(t) = ¢>( -r), t E [-r - a, -r]. If x(lT + t) = '¢;(t) + z(t),
t E [-r- a,O], then xis a solution of Equation (5.7) if and only if z
satisfies z 0 = 0 and
(5.8) f(lT + t, J;t + zt) = ¢(t), t E [-a, 0].
If f(t,¢>+'1/J) = f(t,¢>) +L(t,¢>)'1/J+g(t,¢>,'1/J), where L(t,¢) = f¢(t,¢),

then the hypotheses on f imply that g(t,¢,'1/J) is continuous in (t,¢,'1/J),
g(t,¢>,0) = 0 and, for any (t,¢) E fl there is a (3 = (J(t,¢) 2:: 0 and a
function e(t, ¢>, (3) continuous in (t, ¢>, (3) such that e(t, ¢>, 0) = 0 and
lg(t, ¢>, '1/J)- g(t, ¢>,~)I ::::; E(t, ¢>, (3)1'1/J- ~1, 1~1 :S (3 and 1'1/JI :S (3.
If we make use of this in Equation (5.8), then Equation (5.7) is equivalent
to z0 = 0 and
t E [-a, 0].
Let A(t, ¢) = A(t, ¢>, -r; L) be the matrix associated with L(t, ¢)given
in Definition 5.2. Then Equation (5.7) is equivalent to zo = 0 and
z(t- r) = A- 1 (0" + t, J;t) [/_:+ d[ry(t, J;t, B)]zt(B)- f(lT + t, J;t)

(5.9)
- g(lT + t, J;t, Zt) + /p(t)], t E [-a, 0].
For any (3 > 0, let Bf3 = {'¢ E C: 1'1/JI ~ (3}. For any v, 0 < v < 1/4, there
are o: > 0, (3 > 0, such that (a + t, ¢ + '1/J) E Q,
IA- 1 (a+t,¢+'1/J)IE(a+t,¢+'1/J,(3) <v

IA- 1 (a+t,¢+'1/J)I'Y(a+t,¢+'1/J,o:) <v
fortE [-a, OJ, '1/J E Bf3, and')' the function in Relation (5.2).
Choose a and (3 so that these relations are satisfied. For any non-
negative real a and /3, let B(a, /3) be the set defined by
B(a, /3) = {( E C([-r- a, 0], lRn) : ( 0 = 0, (t E Bjj, t E [-a, 0]}.
For any 0 < /3 < (3, there is an a, 0 < a < a, so that l¢t - ¢1 < (3 - /3.
Further restrict a so that
If( a+ t, ¢t)- f(a, ¢)1 ~ v/3

l¢(o)- ¢(t)1 ~ v/3
fortE [-a, 0].
For any ( E B(a, /3), define the transformation
T: B(a,/3)-+ C([-r- a,O],lRn)
by the relation (T()o = 0 and
(T()(t- r) = A- 1 (a + t, ~) [/_:+ do[Tl(t, ¢t, O)](t(O)

(5.10) - f(a + t, ~) + f(a, ¢)- g(a + t, ¢t, (t)
+¢(t)-¢(o)], tE[-a,o].
By hypothesis ¢(0) = f(a, ¢) and therefore the fixed points ( on T in

B(a,/3) yield the solution x of Equation (5.7) on [a-r-a,a] with x(a+t) =
¢(t) + ((t), t E [-a, o].
We now show that Tis a contraction on B(a, /3). It is clear from Equa-
tion (5.10) and the restrictions on a and /3 that
I(T()(t- r)l ~ v/3 + v/3 + v/3 + v/3 < /3,

1
I(T()(t- r)- (T~)(t- r)l ~ vl(t- ~tl + vl(t- ~tl < 21(t- ~tl
for all t E [-a, OJ,(,~ E B(a, /3). Therefore, T: B(a, /3)-+ B(a, /3) and is a
contraction. Thus, there is a unique fixed point in B(a, /3). This proves the
theorem. D
Suppose the conditions of Theorem 2.1 are satisfied. For any (a,¢) E il,
there exists a tu,</> and a function x that is a noncontinuable solution of
the RFDE(J) on [a- r,ta,¢) through (a,¢). Let Da ~ C be defined by

Da = {¢ E C : (a,¢) E D}. If we only speak of noncontinuable solutions,
then we can define a map T(t, a) : Da ----+ C defined for each ¢ E f2 8 , and
t E [a,ta,¢) by T(t,a)¢ = Xt(a,¢). This map will be called the solution
map of the RFDE(J). We shall discuss many properties of this mapping in
the following section and content ourselves with the following observation
at the present time.
Corollary 5.1. Suppose f2 is an open set in IR x C, f : f2----+ IRn is contin-

uous, and T(t, a) : Da ----+ C, t E [a, ta,¢) is defined above. Iff is atomic at
-r on D, then T(t, a) is one-to-one.
Proof. If the assertion is not true, then there are '1/J i= ¢ in C and a h > a
such that Xt 1 (a,¢) = xda, '1/J), Xt(a, ¢) i= Xt(a, '1/J) for a :S t < t1. If
x = x(a,¢) andy= x(a,'l/J), then x(t) = f(t,xt) and y(t) = f(t,yt) for
a ::; t ::; t 1. Since f is atomic at -r on f2 there is an a = a(tl) > 0 such
that there is a unique solution of the RFDE(J) on [h - r - a, t 1] through
(tl,Xt 1 ) = (h,Yt 1 ). Thus (t,xt) = (t,yt) for t1- a :S t :S h. This is a
contradiction and proves the corollary. D
Let us now consider some examples. Consider the linear system
(5.11) x(t) = j_: d[77(t, B)]x(t +B) ~f L(t)xt

where 77(t, -r+)- 77(t, -r) = A(t) is continuous and
if_~r+s d[77(t, 8)]'1/J(B)- A(t)'l/J( -r)l :S 'Y(t, s) -r~~~r+s 1'1/J(B)I

for a continuous scalar function 'Y(t, s), t E IR, s 2: 0, 'Y(t, 0) = 0. If
detA(t) i= 0 for all t, then L(t) is atomic at -ron IR x C and the map
T(t, a) defined by the solutions of Equation (5.11) is one-to-one.
As another example, consider the equation
(5.12) i;(t) = L(t)Xt + N(t, Xt) ~f F(t, Xt)

where L is the same function as in Equation (5.11), N(t, ¢) is continuous
for (t, ¢) E IR x C, and the Frechet derivative Nq,(t, ¢) of N with respect
to¢ is continuous and INq,(t,¢)1 :S tL(I¢1) for (t,¢) E IR x C, where IL
is a continuous function with tL(O) = 0. If ldetA(t)l 2: a> 0 fortE IR,
then F(t, ¢) is atomic at -r on IR x U where U is a sufficiently small
neighborhood of the origin in C. Here, we are using the relations (5.4)-
(5.6) after the Definition 5.3 of atomic at -r. Consequently, the solution
map T(t,¢) defined by the solutions of Equation (5.12) is one-to-one on its
domain of definition.
As a final example, consider the equation
(5.13) x(t) = -ax(t- 1)[1 + x(t)], a> 0.

For this case, f(t, ¢)=-a¢( -1)[1 + ¢(0)] and
J¢(t, ¢)'¢=-a'¢( -1)[1 + ¢(0)]- a¢( -1)'¢(0)

A(t, ¢) = -a[1 + ¢(0)].
As long as ¢(0)-:/:- -1, the function f(t, ¢)is atomic at-rand the mapping
T(t, CJ) will be one-to-one as long as the solution x(CJ, ¢)(t)-:/:- -1. But, from
Equation (5.13), we have
x(t) = -1 + [1 + ¢(0)]exp[-1t ax(s -1)ds],
Therefore, any solution with ¢(0) -:f. -1 will always have x(t) -:/:- -1 and
T(t, CJ) defined by T(t, CJ)¢ = Xt(CJ, ¢) is one-to-one on the sets
{¢EC:¢(0)>-1}, and {¢EC:¢(0)<-1}.
On the set
C-1 ~f{¢EC:¢(0)=-1}
the map T(t, CJ) is not one-to-one and, in fact, T(t, CJ)¢ is the constant
function one fort::::: CJ + 1 and¢ E C-1·
Let us state more precisely the type of general question that should be
discussed in connection with the results in this section and consider only
the autonomous equation
(5.14) x(t) = f(xt)·
Suppose r is an open set inC, f E C 1(r,IRn) with the usual norm, lflr =
SUP,pa[lf('¢)1 + IDq,('¢)1]. A residual set in C 1(F, IRn) is a set that is the
countable intersection of open dense sets.
Question. For a given Jl, does there exist a residual setS~ C 1(Jl, IRn)
such that the solution map Tt(t) : Jl---+ C of Equation (5.14) is one-to-one
from Jl to its range?
If we restrict ourselves to the subclass of linear RFDE consisting only
of
(5.15)
then the answer is yes. In fact, for any 8 > 0 there is an n x n matrix B
such that IBI < 8 and
L¢+B(-r)
is atomic at -ron C. Furthermore, if Lis atomic at -ron C, then every
continuous linear I: C---+ IRn sufficiently close to Lin £(C, IRn) is atomic at
-r. Therefore, we can takeS to be the set of continuous linear L: C---+ IRn
such that L is atomic at -r. The set S is actually open and dense in
.C(C, JRn).
In later sections, the importance of this question to the qualitative
theory will be discussed. Also, some further specific results for both the
autonomous and nonautonomous equations will be given.
2.6 Caratheodory conditions
In Section 2.1, we defined a functional differential equation for continuous

f : 1Rx C __., 1Rn. On the other hand, it was then shown that the initial-value
problem was equivalent to
=¢
1t
Xa
(6.1)
x(t) = ¢(0) + f(s, x 8 ) ds,
Equation (6.1) is certainly meaningful for a more general class of functions

f if it is not required that x(t) have a continuous first derivative fort > CJ.
We give in this section the appropriate generalization to functional dif-
ferential equations of the well-known Caratheodory conditions of ordinary
differential equations.
Suppose [lis an open subset of 1R X C. A function f : [l --t 1Rn is said
to satisfy the Caratheodory condition on [l if f(t, ¢) is measurable in t for
each fixed¢, continuous in¢ for each fixed t and for any fixed (t, ¢) E fl,
there is a neighborhood V (t, ¢) and a Lebesgue integrable function m such
that
(6.2) \f(s,'lji)\:::; m(s), (s,'lji) E V(t,¢).
If f : [l __., 1Rn is continuous, it is easy to see that f satisfies the

Caratheodory condition on fl. Therefore, a theory for Equation (6.1) in
this more general setting will include the previous theory.
If f satisfies the Caratheodory condition on fl, (CJ, ¢) E fl, we say
a function x = x( CJ, ¢>, f) is a solution of Equation (6.1) through (CJ, ¢) if
there is an A > 0 such that x E C([CJ- r, CJ +A], 1Rn), Xa = ¢ and x(t)
is absolutely continuous on [CJ, CJ +A] and satisfies x(t) = f(t, Xt) almost
everywhere on [CJ, CJ +A].
Using essentially the same arguments, one can extend all of the previ-
ous results to the case where f satisfies a Caratheodory condition on fl. The
most difficult result is the analogue of Theorem 2.2 on continuous depen-
dence. To obtain a result on continuous dependence, it is sufficient to require
that all fk satisfy the Caratheodory condition on fl, fk(s,'ljJ) __., f 0 (s,¢)
as k __., oo, 'ljJ __., ¢ for almost all s and satisfy the following condition: For
any compact set Win fl, there is an open neighborhood V(W) of Wand
2. 7 Definition of a neutral equation 59
a Lebesgue integrable function M such that the sequence of functions fk,

k = 0, 1, 2, ... , satisfies
lfk(s, 'lj;)l::::; M(s), (s,'lj;)EV(W), k=0,1,2, ....

We remark in passing that more general existence theorems are easily
given if the function f(t, ¢) depends on¢ in some special way. In particular,
if for any E ~ 0 we let ¢< denote the restriction of¢ to the interval [-r, E]
and
f(t,¢) = F(t,¢(0),¢'),
then the basic existence theorem can be proved by the process of stepping
forward a step of size less than E (if E > 0) under very weak conditions on
the dependence of F(t, x, 'lj;) on 'lj;.
2. 7 Definition of a neutral equation
In this section, we define a neutral functional differential equation (NFDE)

and give some examples.
Definition 7.1. Suppose fl ~ IR x C is open, f : fl --+ IRn, D : fl --+ IRn

are given continuous functions with D atomic at zero (see Definition 5.3 of
Section 2.5). The relation
(7.1)
is called the neutral functional differential equation NFDE(D, f). The func-
tion D will be called the difference operator for the NFDE(D, f).
Definition 7.2. For a given NFDE(D, f), a function xis said to be a solution
of the NFDE(D, f) if there are (J E IR, A> 0, such that
(t, Xt) E fl, t E [(]", (J +A),
D(t, Xt) is continuously differentiable and satisfies Equation (7.1) on [(]", (J +
A). For a given (J E IR, ¢ E C, and((]",¢) E fl, we say x((J,¢,D,f) is
a solution of Equation (7.1) with initial value ¢ at (J or simply a solution
through ((]" 1 ¢) if there is an A > 0 such that x( (J, ¢, D, f) is a solution of
Equation (7.1) on [(J"- r, (J +A) and Xa((J, ¢, D, f) = ¢.
Definition 7.3. If D(t, ¢) = D 0 (t)¢- g(t), f(t, ¢) = L(t)¢ + h(t) where

D 0 (t) and L(t) are linear in¢, the NFDE(D, f) is called linear. It is linear
homogeneous if g = =
0, h 0 and linear nonhomogeneous if either g =/'. 0 or
h =/'. 0. An NFDE(D, f) is called autonomous if D(t, ¢) and f(t, ¢) do not
depend on t.
Let us now consider some examples of NFDE.

Example 7.1. If D¢ = ¢(0) for all¢, then Dis atomic at 0. Therefore, for any
continuous f: fl--+ lRn, the pair (D,f) defines an NFDE. Consequently,
RFDE are NFDE.
Example 7.2. IfT > 0, B is an nxn constant matrix, D(¢) = ¢(0)-B¢( -T),
and f: fl--+ lRn is continuous, then the pair (D, f) defines an NFDE; that
is, the equation
d
(7.2) dt[x(t)- Bx(t- T)] = f(t,xt)
is an NFDE.
Example 7.3. If T > 0, xis a scalar, D¢ = ¢(0)- ¢ 2 ( -T) and f: fl--+ lR is

continuous, then the pair (D, f) defines an NFDE; that is, the equation
(7.3)
is an NFDE.
Example 7 .4. Suppose T j, 0 < T j ::; T, j = 1, 2, ... , N, are given real numbers
and g : lRN --+ lRn is a given continuous function. If
D(¢) = ¢(0)- g(¢( -TI), ... , ¢(-TN))

and 0 is the zero function on lR x C, then the pair (D, 0) defines an NFDE
on lR x C; that is, the equation
d
(7.4) dt [x(t)- g(x(t- TI), ... , x(t- TN))] = 0
is an NFDE. Obviously, the solutions of Equation (7.4) are given by
(7.5) x(t) = g(x(t- T1), ... , x(t- TN))+ c,

where cis a constant given by c = ¢(0) -g(¢( -T 1 ), ... , ¢(-TN)). Therefore,
if one defines the subset U of C by
U = {c/J E C : ¢(0) = g( c/J( -Tl), ... , ¢(-TN))}

then any solution of the NFDE (7.4) with initial data in U is a solution of
the difference equation
(7.6) x(t) = g(x(t- T1), ... , x(t- TN)).

Conversely, any solution of Equation (7.6) is a solution of Equation (7.4).
Consequently, NFDE defined as in Definition 7.1 include difference equa-
tions.
2.8 Fundamental properties of NFDE 61
In the definition of a solution of Equation (7.1), it is only required that

D(t, Xt) be differentiable in t. The function x(t) may not be differentiable.
The same situation was encountered in Chapter 1. If the function x(t) is
continuously differentiable on [a-- r, a-+ A) and the function D( t, ¢) is also
differentiable in t, ¢ then the function x satisfies the equation
(7.7) Dq,(t, Xt)Xt = f(t, Xt) - Dt(t, Xt)

for a- :"::: t < a- + A. The symbol Xt designates the function in C given by
±t(e) = x(t +e), -r :"::: e :"::: o.
As a result of this discussion, we see that smooth solutions of Equation
(7.1) satisfy Equation (7.7), which is linear in the derivative of x. This is a
basic limitation of the development of the theory using the Definition 1.1
for an NFDE. For the special cases, Equations (7.2) and (7.3), Equation
(7.7) is given by
(7.8) x(t)- Bx(t- r) = f(t,xt),
(7.9) x(t)- 2x(t- r)x(t- r) = j(t, Xt)·

Example (7.9) also shows that our theory will not include any general theory
of NFDE based only on the assumption that the derivative of x enters in
a linear fashion. In fact, the terms involving x that multiply x must occur
with the same delays.
2.8 Fundamental properties of NFDE
In this section, we consider the questions of existence, uniqueness, and

continuous dependence of solutions of NFDE.
Theorem 8.1. (Existence). If [l is an open set in IR x C and (a-, ¢) E fl,

then there exists a solution of the NFDE(D,f) through (a-,¢).
Proof. We give an outline of the proof and leave the details of the compu-
tations to the reader. If the derivative Dq,(t,¢) of D(t,¢) with respect to¢
is represented as
Dq,(t,¢)ij; = A(t,¢)ij;(O)- /_: d[tL(t,B,¢)]if;(B)
then the definition of atomic at 0 implies det A( t, ¢) -=1- 0, A( t, ¢) is continu-

ous in t, ¢,and the linear operator Dq,(t, ¢)has smoothness on the measure
(see Definition 5.2 and Lemma 5.1 of Section 2.5).
Let J: [-r, oo) -+ IRn be defined by Jo
= ¢, 'J(t) = ¢(0), t ~ 0.
If xis c:_solution of Equation (1.1) through (a-,¢) on [a-- r,a- +a] and
Xt+a- = c/Jt + Zt, then z satisfies the equation
(8.1) D((J' + t, i>t + Zt) = D((J', </J) +fat f((J' + s, i>s + Z8 ) ds
for 0:::; t:::; a and z0 = 0. Using the preceding notation for Dq,(t,¢), we
have
(8.2) z=Sz+Uz
where the operators Sand U are defined by (Sz)(t) = 0 and (Uz)(t) = 0
for -r :::; t :::; 0, and
A(t + (]', ;f>t)(Sz)(t) =lor- de[JL((]' + t, B, ;f>t)]z(t +B)
+ [D((J',</J)-D((J'+t,;f>t)]- [D((J'+t,;f>t+zt)
- D((J' + t, ;f>t)- Dq,((J' + t, ;f>t)zt]
A(t + (]', ;f>t)(Uz)(t) =fat f((J' + s, i>s + Zs) ds
for 0:::; t:::; a.

Any z E C([-r, a], 1Rn) satisfying the operator equation (8.2) is a
solution of Equation (8.1) and, therefore, Xt+u = i>t + Zt is a solution of
Equation (7.1). Therefore, the existence of a solution of Equation (7.1)
through ((]', <P) is equivalent to determining an a > 0 such that Equation
(8.2) has a solution in C([-r,a],lRn).
Let A(a,,B) = {( E C([-r,a],1Rn) : (o = 0, l(tl :::; ,8, t E [O,a]}. It
is now a relatively simple matter to show that one can choose a and ,8 so
that the mapS+ U: A( a, ,8) --+A( a, ,8). Furthermore, Sis a contraction
on A( a, ,8) and U is completely continuous on A( a, ,8). Therefore, S + U
is an a-contraction and Darbo's theorem (see Theorem 6.3 of Section 4.6)
implies there is a fixed point of S + U in A( a, ,8). This completes the proof
of the theorem. D
Continuous dependence is more difficult. Suppose A is a subset of a

Banach space and [2 is an open set in 1R x C. If D : [2 x A --+ 1Rn is a
given function, we have different possibilities for the manner in which the
property of being atomic at 0 depends on the parameter A. One can assume
that D(t, ¢,A) is atomic at each (t, ¢) E [2 for each A E A. One can also
assume the stronger hypothesis that D(t, ¢,A) is atomic at zero at each
(t, ¢) E [2 uniformly with respect to A E A, with uniformly meaning that
all estimates on determinants, measures, etc., in the definition of atomic at
0 are uniform with respect to A E A. Each of these hypotheses on D will
require corresponding hypotheses on f : [2 x A --+ 1Rn in order to obtain
continuous dependence of the solution of the NFDE(D, f).
A useful tool in obtaining results on continuous dependence is the
following lemma on continuous dependence of fixed points of condensing
mappings.
2.8 Fundamental properties of NFDE 63
Lemma 8.1. Suppose r is a closed, bounded, convex set of a Banach space,

A is a subset of another Banach space, and T : r X A ----> r is a given
mapping satisfying the following hypotheses:
(hi) T(·, A) is continuous for each A E A and there exists a Ao E A such
that T(x, A) is continuous at (x, Ao) for each x E r.
(h 2 ) For every T' ~ r, a(r') > 0, there is an open neighborhood
B = B(r') of Ao such that for any precompact set A' ~ An B,
we have a(T(T', A')) < a(r'), where a is the Kuratowskii measure
of noncompactness.
(h3) The equation
(8.3) x = T(x, A)
for A= Ao has a unique solution x(Ao) in A.
Then the solutions x(A) in r of Equation (8.3) are continuous at A at
A= Ao.
Proof. If T' ~ r and a(T') = 0, define B(r') = r. Suppose {Ak} c
An B(T) is a sequence converging to Ao ask----> oo, and let x(Ak) E r be
a solution of Equation (8.3) for A= Ak. If r' = {x(Ak)}, choose k so large
that A' = {Ak} ~ An B(r'). Since A' is precompact and Hypothesis (h 2)
is satisfied,
if a(T') > 0. Since this is impossible, a(T') = 0 and T' is precompact.

Since r is closed, there is a subsequence {vk} of {Ak} and z E T such that
x(vk)----> z ask----> oo. Hypothesis (hi) implies z = T(z, Ao) and Hypothesis
(h 3 ) implies z = x(A 0 ). Since every convergent subsequence of the x(Ak)
must converge to the same limit, it follows that x(Ak) ----> x(Ao) as k ----> oo.
Since the sequence { Ak} was an arbitrary sequence converging to Ao, the
proof of the lemma is complete. D
Corollary 8.1. Suppose r and A are as in Lemma 8.1, T : r X A ----> r

satisfies Hypotheses (hi) and (h3 ), T = S + U, and, for each compact set
A' ~A,
(h4) 8(·, A) is a contraction on r uniformly with respect to A E A'.
(h 5 ) U(r, A') is precompact.
Then the solutions x(A) of Equation (8.3) are continuous at A0 .
Proof. To verify that Hypothesis (h 2 ) in Lemma 8.1 is satisfied, simply
observe that for every r' ~ r, a(T(r',A')) = a(S(T',A')):::; ka(T') for
some k E [0, 1). D
Corollary 8.2. Suppose r and A are as in Lemma 8.1, T : r X A ----> r

satisfies Hypothesis (h3 ), T = S+U where U satisfies Hypothesis (h 5 ), and
(h 6 ) S(·, .A) is a contraction for), E A and U(·, .A) is continuous for each
), E A',
(h 7 ) S(x, .A) is continuous at .Ao uniformly for x E r.
Then the solutions x(.A) of Equation (8.3) are continuous at >.0 .
Proof. Hypothesis (hl) is obviously satisfied. To prove Hypothesis (h 2 ) is
satisfied, let B13 = {A: I.A- .Aol < ,8} and, for any E > 0, choose ,B(E) such
that
IS(x, .A)- S(x, .Ao)l < E if .A E B/3(<), x E r.
This choice is possible by Hypothesis (h 7 ). Let the contraction constant for
S(·, .Ao) be ko < 1. If r' <;;; r, then a((S(·, .A)- S(·, .Ao))T') < E if), E Bf3(e)·
If we choose E so that k 0 + E < 1, then
a(T(-, .A)T') = a(S(·, >.)r')::; (ko + E)a(T') < a(T')

for all T' <;;; r. Therefore, Hypothesis (h 2 ) is satisfied and the conclusion of
the corollary follows from Lemma 8.1. 0
We are now in a position to prove a basic result on continuous depen-

dence.
Theorem 8.2. (Continuous dependence). Suppose fl <;;; IR x C is open, A is

a subset of a Banach space, D : fl x A ____, IRn, f : fl x A ____, IRn satisfy the
following hypotheses:
(i) D(t, cp, .A) is atomic at zero for each (t, cp) E fl uniformly with respect
to >..
(ii) D(t,cp,.A) and f(t,cp,.A) are continuous in (t,cp) E fl for each), E A
and continuous also at (t,cfJ,.Ao) for (t,cp) E fl.
(iii) The NFDE(D(·,.Ao),f(·,.Ao)) has a unique solution through (u,cp) E fl
that exists on an interval [u- r, b].
Then there is a neighborhood N(u, cp, .Ao) of (u, cp, >. 0 ) such that for any
(u', cp', .A') E N(u, cp, .Ao), all solutions x(u', cp', .A') of the NFDE(D(·, A'),
f(·,.A')) through (u',cp') exist on [u'- r,b] and Xt(u',cp',A') is continuous
at (t,u',cp',.Ao) fortE [u,u+A], (u',cp',.A') E N(u,cp,.A 0 ).
Proof. For each ), E A, one can define the operators S(z, .A) and U(z, >.)
as in Equation (8.2) for the solutions of the NFDE(D(·, .A), f(·, .A)) through
(u', cp'). It is left as an exercise to verify that the hypotheses of Corollary 8.1
are satisfied for the set r chosen as an appropriate set A( a, ,B) used in the
proof of Theorem 8.1. Using the compactness of the set {(t, Xt(u, cp, >. 0 )) :
t E [u, b]}, one can successively step intervals of length a to complete the
proof of the theorem. 0
Almost exactly as in the proof of Theorem 2.3 of Section 2.2, one

obtains the following result on the uniqueness of solutions.
Theorem 8.3. (Uniqueness). If Q ~ lR x C is open and f : Q ----. JRn is

Lipschitzian in ¢ on compact sets of Q, then, for any (u, ¢) E Q, there
exists a unique solution of the NFDE(D, f) through (u, ¢).
The problems concerning the behavior of the solutions of an NFDE as

one approaches the maximal interval of existence is not as well understood
as the ones for RFDE. The proofs of results are also much more techni-
cal. The following results are stated without proof with references given in
Section 2.9.
Theorem 8.4. (Continuation). Suppose Q ~ lR x C is open, (D, f) de-

fines an NFDE on Q and W ~ Q is closed and bounded and there is a
8-neighborhood ofW in D. Iff maps W into a bounded set in lRn, D(t, ¢)
and Dq,(t, ¢) are uniformly continuous on W, D is uniformly atomic at 0
on W, and x is a noncontinuable solution of the NFDE(D, f) on [u- r, b),
then there is at' E [u, b) such that (t', Xt') r:J. W.
If D( t, ¢) is linear in ¢, we obtain the same conclusion with weaker

hypotheses.
Theorem 8.5. (Continuation). Suppose (D, f) defines an NFDE on an open

set Q ~ lR x C and D(t)¢ = D(t, ¢) is linear in¢. If x is a noncontinuable
solution of the NFDE(D, f) on [u- r, b), D(t) is defined on [u, b], and W
is a closed bounded set in Q for which f(W) is bounded, then there is a
t' E [u, b) such that (t', Xt') r:J. W.
Except for technical details, the methods used in proving the existence and
continuous dependence theorems in Section 2.2 are natural generalizations
of methods from ordinary differential equations. One can investigate these
questions with much weaker conditions on the function f as well as different
types of hereditary dependence than the one implied by Xt (see Neustadt
[1], Tychonov [1], Jones [1], Cruz and Hale [1], Imaz and Vorel [1]). Coffman
and Schaffer [1] have investigated the weakest possible conditions on linear
operators L(t) that will permit the development of a general theory. The
paper of Stokes [1] contains some results on differentiability of solutions.
By applying differential inequalities, one can obtain very general results
on uniqueness of solutions (see Laksmikantham and Leela [1]).
The problem of continuation of solutions has features not encountered
in ordinary differential equations as Example (3.1), due to Mishkis [2],
demonstrates. Theorem 3.3 is due to Yorke [1] and the extension theorem
used in the proof is due to Dugundji [1].
The first general results on backward continuation are due to Hastings

[1]. The presentation in the text follows Hale [1] with the improvements
made possible by Definition 5.2 and Lemma 5.1 introduced by Hale and
Oliva [1]. Lillo [1] discusses the backward continuation of solutions of lin-
ear equations that are not necessarily atomic at -r. A similar situation is
encountered in Hale and Oliva [1] in their discussion of the "size" of the
set of linear RDFE for which the solution operator is one-to-one. For some
other results on the inverse of the solution operator, see Kamenskii [1].
In the supplementary remarks at the end of Chapter 1, we gave a rather
lengthy discussion of and references for the different ways in which one can
define a neutral functional differential equation. The approach taken in this
chapter corresponds to one of the simplest definitions. The definition of a
NFDE in Section 2.7 is due to Hale [1] (see also Hale [19]). A more general
definition using only Lipschitz continuous functions D as well as a more
general dependence on the past history is contained in Cruz and Hale [1].
The existence Theorem 8.1 is due to Hale [1]. The general result on
continuous dependence on parameters of fixed points of condensing maps
(Lemma 8.1) is due to Hale [12] (see also Artstein [1]), and generalized
earlier results of Melvin [3] and Cruz and Hale [3]. Corollary 8.1 is due to
Melvin [3] and Corollary 8.2 is due to Cruz and Hale [1]. The continuation
Theorem 8.4 may be found in Hale [1] and the continuation Theorem 8.5
is due to Lopes.
As remarked in Section 2.8 the solution operator for a NFDE(D, f) is
a homeomorphism if D is atomic at zero and -r. If f is also C 1 , then it
can be shown that the solution operator is differentiable with respect to
the initial data and so it is a diffeomorphism. In this case, as in ordinary
equations, the equation defines a group rather than a semigroup.
We have shown how the theory of a-contractions played an important
role in the existence theory of Section 2.8. We will see in the next chapter
that this same concept plays an important role in the representation formula
for the solution operator. Nussbaum [8] also has interesting applications of
a-contractions to existence. Other applications of a-contractions to NFDE
as well as ordinary differential equations are contained in Hale [16]. For the
applications of coincidence degree to NFDE, see Hale and Mawhin [1] and
Hetzer [1].
3
Properties of the sol uti on map
In the study of retarded functional differential equations, the space of ini-

tial functions is preassigned, but the space in which one considers the
trajectories is not. To be more specific, if x(a, ¢,f) is a solution of an
RFDE(f) through (a,¢), should the solution map be considered as the
map x(a, ·, f)(t) : C-+ IRn or the map T(t, a) : C-+ C defined in Section
2.5 as T(t, a)¢= Xt(a, ¢,f)?
By the consideration of simple examples, one can see that x(a, ·)(t) has
some rather undesirable properties. In fact, the scalar equation
(1) ±(t) = -x(t- ~)

2
has a unique solution through each (a,¢) E IR x C, but it also has the
solutions x(t) = sint and x(t) = cost. These latter solutions plotted in
(x, t) space intersect an infinite number of times on any interval [a, oo) and
yet are not identical on any interval.
If we use the map T(t, a) for this example, then Corollary 5.1 of Sec-
tion 2.5 implies T(t, a) is one-to-one on all of C. In the general situation,
uniqueness implies if there is aT> a such that T(r,a)¢ = T(r,a)'!f;, then
T(t, a)¢= T(t, a)'!f; for all t 2:': a.
For autonomous equations, it is more natural to consider the orbits
of solutions rather than the trajectories, that is, the path traced out by
the solution in the phase space X rather than the graph of the solution
IR x X. If the phase space for Equation (1) is chosen as IR and the orbits
as Ut>ox(O,¢)(t), then the orbits for the solutions x(t) = sint and x(t) =
cost coincide and are equal to the interval [-1, 1]. That the orbits coincide
is expected because sin(t + (7r/2)) = cost. Equation (1) is autonomous
and therefore, a solution shifted in phase is still a solution. The difficulty
encountered by choosing the phase space IR is that the orbit of one solution
may completely contain the orbit of another solution and not be related in
any way to a phase shift. The orbit of the solution x = 0 is contained in
the orbit of cost.
On the other hand, if the phase space is chosen as C = C([-7r/2, OJ, IR),
then the orbit of the solution sint of Equation (1) is the set
68 3. Properties of the solution map
r = {¢: ¢(0) = sin(t + 0), - 21r :S 0 :S 0, fortE [0, oo) },

of points in C. The set r, as before, is also the orbit of the solution cost.
Furthermore, because of uniqueness of solutions and one-to-oneness of the
mapping T(t, cr), any solution x of Equation (1) for which there is aT with
Xr E r must be a phase shift on sin t. Therefore, r is determined by phase
shifts of a solution. Finally, r is a closed curve in C that is intuitively
satisfying since sin t is periodic.
This simple example suggests the geometric theory for Equation (1)
will probably be richer if the map T(t, cr) is used. However, in some situa-
tions, it is very advantageous to know that T(t, cr)¢ is determined by taking
a restriction over an interval of a function in lRn.
It is the purpose of this chapter to discuss a few of the good and
bad properties of the solution map TJ(t, cr) of an RFDE(f) defined by
Tt(t, cr)¢ = Xt(cr, ¢,f). We will assume, unless otherwise explicitly stated,
that f is continuous and there is a unique solution of the RFDE(f) through
(cr, ¢) so that Tt (t, cr )¢is continuous in (t, cr, ¢,f) by Theorem 2.2 of Section
2.2.
If the RFDE(f) is autonomous, let Tt(t, 0) = Tt(t), t 2: 0. It follows
that
TJ(O) =I
(2) TJ(t)TJ(T) = Tt(t + T), t, T 2: 0,
Tt(t)¢ is continuous in (t, ¢,f);
that is, {T(t)}, t ;::: 0 is a strongly continuous semigroup of transformations
on a subset of C. Of course, it is understood here that t and Tare allowed
to range over an interval that may depend on ¢ E C. Since our interest
in this chapter does not concern the dependence of Tt(t, cr) on f, we write
simply T(t, cr).
3.1 Finite- or infinite-dimensional problem?
From the special way in which an RFDE is defined, it is necessary to discuss

whether the problem is actually infinite-dimensional or finite-dimensional.
More specifically, can the noncompactness of the unit ball in C have any
adverse effects on the solutions considered either in lRn or C? The purpose
of this section is to consider this question in some detail.
Theorem 3.2 of Section 2.3 on the continuation of solutions states that
a noncontinuable solution of an RFDE(f) must leave every closed bounded
set W in the domain of definition il of the equation, provided f is com-
pletely continuous on il. A continuous function on il need not be completely
continuous on il if r > 0, that is, if C is infinite-dimensional. Thus, it is
natural to ask if this latter condition is necessary. The answer is yes and
we state this as
3.1 Finite- or infinite-dimensional problem? 69
Property 1.1. The continuation theorem is not valid iff is not a completely
continuous map.
Proof. Equation (3.1) and Theorem 3.3 of Section 2.3 demonstrate this
result. D
We say that a mapping from one metric space to another metric space
is bounded if it takes closed bounded sets into bounded sets. The map is
locally bounded if it takes some neighborhood of each point into a bounded
set.
Property 1.2. T(t, a) is locally bounded fort::;:: a.
Proof. Since T(t,a)¢ is assumed to be continuous in (t,a,¢), it follows that
for any t 2: a,¢ E C for which (a,¢) E D and T(t, a)¢ is defined, there is a
neighborhood V(t, a,¢) of¢ inC such that T(t, a)V(t, a,¢) is bounded. D
Property 1.3. T(t, a) may not be a bounded map.

Proof. Suppose r = i, C = C([-r, 0], 1R) and consider the equation
(1.1) x(t) = f(t,xt) ~f x 2 (t)- {
0
fx(s)f ds.
Jmin(t-r,O)
It is clear that f takes closed bounded sets into bounded sets and is locally
Lipschitzian. If B = {¢ E C: f¢1 :=:; 1} and x(b), bE B, is the solution
of Equation (1.1), then x(b) is always::;:: -1. Also, forb =I= 0, x(b)(O) ::; 1,
x(b)(t) < x 2 (t) for all t implies x(t) < y(t) for all 0 < t < 1, where iJ = y 2 (t),
y(O) = 1. Therefore, x(b)(t) exists on [0, 1),
x(b)(t) < y(t) = (1- t)- 1
and, in particular, x(b)(r) < (1- r)- 1 for all bE B. For t ::;:: r, x(b)(t) =
x 2 (b)(t) and the fact that x(b)(r) < (1 - r)- 1 implies x(b)(t) exists for
-r ::; t ::; 1.
If we show that for any E > 0, there is a b E B such that
x(b)(r) > (1- r)- 1 - E,
then the set x(B)(1) is not bounded since the solution iJ = y 2 (t) through
(r, (1- r)- 1 ) is unbounded at t = 1 and x(t) = x 2 (t) fort ::;:: r. Therefore,
suppose E > 0 is given, C = [1- rf- 1 .
Choose b E B so that b(O) = 1, f~r fb(t)f dt < 2CE and let y(t) =
y(t, 0, 1), y(O, 0, 1) = 1, be the solution of y(t) = y 2 (t) and x(t) = x(b)(t).
If 'l/;(t) = y(t)- x(t) for 0 < t < r, then 'l/;(t) ::;:: 0 and -J;(t) ::; 2C'lj;(t) + 2CE.
Since '1/J(O) = 0, one thus obtains '1/J(r) :=:;E. This shows that
x(r) = y(r)- 'lj;(r) = (1- r)- 1 - 'lj;(r)

~ (1- r)- 1 - f
and proves the general assertion made earlier. 0
As another illustration of the infinite dimensionality of an RFDE, con-

sider the control problem
(1.2) x(t) = Ax(t- r) + Bu
where A and Bare constant matrices, r > 0, x E IRn, u E JRP, lui ~ 1, and
u = u(t) is a locally integrable function. Suppose ¢ E C, t ~ 0 are given
and let x(¢, u) designate the solution of Equation (1.2) with xo(¢, u) = ¢.
Suppose ,,
A(t, ¢) = {'lj; E C: there is a locally integrable u, lui ~ 1,

with Xt(¢,u) = 'lj;}.
The set A(t, ¢) is the set attainable at timet along solutions of Equation
(1.2) using the controls u and starting at t = 0 with ¢. For ordinary differ-
ential equations (r = 0), it is known that every element of the attainable
set at timet can also be reached by using only the bang-bang controls in
Equation (1.2); that is, by only using control function u(r) with iu(r)l = 1
for 0 ~ r ~ t.
Property 1.4. Bang-bang controls are not always possible for RFDE.
Proof. The followipg counterexample demonstrates this property. Suppose
¢ = 0 and consider
(1.3) x(t) = x(t- 1) + u(t), lui ~ 1.
Then
x(O, u)(t) = 1t u(s) ds for 0 ~t~1
and A(1, 0) contains zero since the control u(t) = 0, 0 ~ t ~ 1, gives
x1 (0, u) = 0. On the other hand, there is no way to reach zero with a bang-
bang control. 0
3.2 Equivalence classes of solutions
In the previous chapter, Equation (5.13) of Section 2.5, we gave an exam-

ple in which the mapping T(t, a) was not one-to-one. To reemphasize this
remark and to discuss some more geometry of the solutions, we state this
result explicitly and give another example.
3.2 Equivalence classes of solutions 71
Property 2.1. The map T(t, a) may not be one-to-one.
Proof. Consider the equation
(2.1) x(t) = -x(t- r)[1- x 2 (t)].
Equation (2.1) has the solution x( t) = 1 for all t in ( -oo, oo). Furthermore,
if r = 1, a = 0, and ¢ E C, then there is a unique solution x(O, ¢) of
Equation (2.1) through (0, ¢) that depends continuously on ¢. If -1 ::;
¢(0) ::; 1, these solutions are actually defined on [-1, oo). On the other
hand, if¢ E C, ¢(0) = 1, then x(O, ¢)(t) = 1 for all t :2: 0. Therefore, for all
such initial values, Xt(O, ¢), t :2: 1, is the constant function 1. A translation
of a subspace of C of codimension one is mapped into a point by T(t, 0) for
all t :2: 1. D
'
----~------------ t
'
'
'
'
'
'L------------
Fig. 3.1.
The function x(t) = -1 is also a solution of Equation (2.1) and for

any¢ E C, ¢(0) = -1, the solution x(O, ¢)(t) is -1 fort :2: 0. Therefore,
Xt (0, ¢) is the constant function -1 for t :2: 1.
For this example, it is interesting to try to depict the trajectories in
IR x C. For any constant a let Ca = {¢ E C: ¢(0) =a}. The set Ca is the
translate of a subspace of C of codimension 1 (a hyperplane) and IR x C
can be represented schematically as in the accompanying diagram. We have
put on this diagram the sets IR X cl and IR X c_l as well as the constant
functions 1 and -1 and representative trajectories in these planes. Notice
that solutions are trapped between these planes if the initial values ¢ satisfy
-1 ~ ¢(0) ~ 1. Also notice that any solution that oscillates about zero must
have a trajectory that crosses the set 1R x Co.
That the solution map T(t, a) may not be one-to-one is an annoying
feature of the theory of RFDE. Sufficient conditions for one-to-oneness were
given in Corollary 5.1 of Section 2.5. We also posed the general question of
whether or not there is a residual set in the set of all RFDE for which the
map T( t, ¢) is one-to-one. Even if the answer to this question is affirmative,
it does not necessarily take care of applications. It may be that the form of
the equation is fixed and one is not allowed to change it too much. In this
situation, a better understanding of one-to-oneness is needed.
One way to begin to understand why the map T(t, a) is not one-to-one
is to define and study equivalence classes of initial data in the following
manner. Suppose [2 = 1R x C and all solutions x(a, ¢)of the RFDE(f) are
defined on [a-r, oo). We say (a,¢) E 1Rx Cis equivalent to (a, 1/J) E 1Rx C,
(a,¢)"' (a,'l/J), if there is aT~ a such that x 7 (a,¢) = x 7 (a,'ljJ); that is
(a,¢) is equivalent to (a,'ljJ) if the trajectories through (a,¢) and (a,'I/J)
have a point in common. It is easy to see that "' is an equivalence relation
and the space Cis decomposed into equivalence classes {Va} for each fixed
a. If T(t, a) is one-to-one, then each equivalence class consists of a single
point; namely, the initial value (a,¢). For each equivalence class Va, choose
a representation element ¢ 17' 0 and let
(2.2) W(a) = U ¢17,<>.
From the point of view of the qualitative theory of functional differen-

tial equations, the set W (a) is very interesting since it is a maximal set on
which the map T(t, a) is one-to-one. However, it seems to be very difficult
to say much about the properties of W(a). In fact, without some more pre-
cise description of the manner in which ¢ 17'a. is chosen from Va., one cannot
hope to discuss such topological properties of W(a) as connectedness. For
example, consider the scalar equation
x(t) = o
considered as a functional differential equation on C. If Ca = { f E C :
¢(0) = a}, then¢ E Ca implies Xt(a,¢) is the constant function a for
t ~ a+ r. Therefore, for each a, the equivalence classes Va. are the sets Ca.,
-oo < a < oo. An arbitrary choice of ¢ 17'a. leads to a very uninteresting
set W (a). On the other hand, W (a) consisting of all constant functions is
certainly the set that is of interest for the equation.
In a general situation, we know nothing about the "appropriate" choice
of ¢ 17 'a.. On the other hand, in the example discussed in some detail in Prop-
erty 2.1, the equivalence classes are also very easy to determine. In fact,
for any ¢ E C, ¢ ~ C 1 , ¢ ~ C_ 11 it follows from Corollary 5.1 of Section
2.5 that the equivalence class corresponding to¢ consists only of the single
3.2 Equivalence classes of solutions 73
element ¢. On the other hand, if ¢ E cl' then the equivalence class corre-
sponding to¢ is C 1 . Similarly, C_ 1 is the equivalence class corresponding
to ¢ E C_ 1 . A good choice for W(O) in Equation (2.1) in this case would
be C\ { (C1 \ {1}) u (C_ 1 \ { -1})} where 1 and -1 are constant functions.
We say that an equivalence class Va is determined in a finite time if
there exists T > 0 such that for any¢, 'ljJ EVa, Xa+t(a, ¢) = Xu+t(a, '1/J) for
t ~ T. We now give some rather surprising results about this concept.
Property 2.2. The equivalence classes may not be determined in finite time.
(2.3) x(t) = !3[lxtl- x(t)].
For any ;3 > 0, we show the equivalence classes for Equation (2.3) are in
one-to-one correspondence with the constant functions. Also, we show there
is a ;3 > 0 such that the equivalence classes are not determined in finite
time.
For a given ¢ in C = C([-1, 0], lR), there is a unique solution x =
x(¢, !3)(t) of this equation through (0, ¢) that is continuous in (¢, ;3, t).
If ¢(0) ~ 0,¢ =1- 0, then x(¢,!3)(t) is a positive constant fort~ 1. In
fact, since ±(t) ~ 0, it follows that lxtl = x(t) for t ~ 1 and uniqueness
implies x(t) is a constant ~ ¢(0) for t ~ 1. Also, if ¢(0) = 0, then ¢ =1-
0 implies ±(0) > 0 and x(t) > 0 for t ~ 1. Therefore, for any positive
constant function, the corresponding equivalence class contains more than
one element. Also, the preceding argument and the autonomous nature of
the equation show that the equivalence class corresponding to the constant
function zero contains only zero.
If ¢(0) < 0, then it is clear that x(¢,!3)(t) approaches a constant as
t ~ oo. If x(¢,;3)(t), ¢(0) < 0, has a zero z(¢,/3), it must be simple, and
therefore, z(¢, ;3) is continuous in ¢, ;3. For any ;3 > 0, there is a ¢ E C,
¢(0) < 0, such that z(¢,;3) exists. In fact, let¢ E C, ¢(0) = -1, ¢(0) = -"(,
'Y > 1, -1 ::::; 0 ::::; -~ and let ¢(0) be a monotone increasing function for
-~ ::::; 0::::; 0. As long as x(t) ::::; 0 and 0::::; t::::; ~'we have lxtl = 'Y and
±(t) = ;3["1- x(t)] ~ /3"(.
Therefore, x(t) ~ ;3"(t- 1 if x(t) ::::; 0 and 0 ::::; t ::::; ~· For /3"(/2 > 1, it
follows that x must have a zero z( ¢, /3) < ~.
The closed subset C_ 1 = {¢ E C : ¢(0) = -1} can be written as
C-1 = C_lo UC_ln where C_lo = {¢ E C-1: z(¢,;3) exists} and C_ln =
{¢ E C-1: z(¢,/3) does not exist}. Since z(¢,;3) is continuous, the set C_ 1o
is open and, therefore, c_ln is closed. If c_ln is not empty, then there is
a sequence </Yj E C_lo, </Yj ~ </J E C_ln as f ~ oo and z(</Jj,/3) ~ oo.
There is a /3o > 0 such that C_ln is not empty. In fact, choose /3o > 0
less than or equal to that value ;3 for which the equation >..+;3 = -!3e->. has
a real root >.. 0 of multiplicity two. For this ;30 , the equation>..+ ;3 = -!3e->.
has two real negative roots. If -,\ 0 is one of these roots, then x(t) = -e->-.ot
is a solution of the Equation (2.3) with initial value ¢ 0 (8) = -e->-.oe, -1:::;
B :S: 0, ¢o E C-1· Therefore, C_1n is not empty. It follows that
8(f3o) ~f sup{z(¢, fJo) : ¢ E C_1o} = oo.

Since the original equation is positive homogeneous of degree 1 in x, it
follows that, for any positive constants a and t 0 , there exists ¢ E C, such
that x(¢,(30 )(t) =a, t ~ t 0 , and x(¢,(30 )(t) <a for 0:::; t < t 0 . This proves
the assertion. 0
3.3 Small solutions for linear equations
A small solution x is a solution such that
(3.1) lim ektx(t)

t--->00
= 0 for all k E IR.
In this section we study the existence of small solutions of linear au-
F
tonomous RFDE(L)
(3.2) {±:: : d["(O)]x(t + 0)
The zero solution is always a small solution, and the question becomes
whether there are initial conditions ¢ -/=- 0 such that the solution x( ·, ¢)
to System (3.2) is a small solution. Such solutions will be called nontrivial
small solutions. It is easy to see that nontrivial small solutions can exist.
For example, consider the system
±1(t) = x2(t- 1)
(3.3)
x2(t) = x1(t).
Any initial condition ¢ = (¢ 1, ¢ 2)T with ¢ 1(0) = 0 and ¢ 2 = 0 yields a
small solution.
In this section we present necessary and sufficient conditions for the
existence of nontrivial small solutions. The proof is based on a fine analysis
of the Laplace transform of System (3.2) and for this we need the notion of
the exponential type of an entire function.
An entire function h : <C --+ <C is of order 1 if and only if
.
l1msup log log M(r)
l = 1
r-->oo og r
where
3.3 Small solutions for linear equations 75
M(r) = max {lh(rei(l)l}·

0~(1~211'
An entire function of order 1 is of exponential type if and only if
.
l1msup log M(r) = E(h)
r-too r
where 0 :::; E(h) < oo. In that case, E(h) is called the exponential type of
h. A vector-valued function h = (h 1 , ... , hn) : <C ---+ <Cn will be called an
entire function of exponential type if and only if the components hj of h
are entire functions of order 1 that are of exponential type. In this case, the
exponential type will be defined by
E(h) = max E(hj)·

l~J~n
The following lemma is a special case of the Paley-Wiener theorem,

theorem 6.9.1 of Boas [1].
Lemma 3.1. Let h : <C ---+ <C be an entire function. If h is uniformly bounded
in the closed right half plane Re z 2 0, then h is of exponential type T and
L 2 -integrable along the imaginary axis if and only if
where ¢ E L 2[0, T] and ¢ does not vanish a. e. in any neighborhood ofT.
Next we collect some consequences of this lemma.
Lemma 3.2. For j = 1, 2, let O:j be a function of bounded variation on the

interval [0, aj], normalized so that O:j is continuous from the left on (0, aj)
and O:j (0) = 0. If o: 1 and o: 2 are not constant in neighborhoods of a 1 and
a 2 , respectively, then
(3.4)
and
Proof. Using partial integration,
laj e-zt do:j(t) = Z laj e-zt(o:j(t)- O:j(aj)) dt, j = 1, 2
and Relation (3.4) follows from Lemma 3.1. From the convolution property,
we find
where a3 is defined to be the convolution of a 1 and a 2
Therefore, Relation (3.5) follows from Lemma 3.1 as well. D
In the sequel we call entire functions of the form lj(z) = J;

e-ztda(t),
where T 2: 0 and a is a function of bounded variation, finite Laplace-Stieltjes
transforms. From Lemma 3.2, we find the following useful representation
for the determinant of the characteristic matrix Ll(z) for Equation (3.2)
n
detLl(z) = zn- Llj(z)zn-j
j=l
where, for j = 1, ... , n, the coefficient lj is a finite Laplace-Stieltjes trans-

form of exponential type at most jr. So det Ll(z) is an entire function of
exponential type with
E(detLl(z)) :S nr.
Let adj Ll(z) denote the matrix of cofactors of Ll(z). Since the cofactors
Cij(z) are (n- 1) by (n- 1)-subdeterminants of Ll(z), it follows that
n-1
Cij(z) = L lj(z)zn-j
j=l
where, for j = 1, ... , n- 1, the coefficient lj is a finite Laplace-Stieltjes

transform of exponential type at most jr. So the exponential type of the
cofactors is less than or equal to (n- 1)r. After these preliminary results,
we return to Equation (3.2). The following result is the first ingredient in
the characterization of the small solutions of Equation (3.2).
Theorem 3.1. Suppose that x( ·; ¢>) is a small solution of Equation (3.2),

then x(t; ¢>) = 0 fort 2: nr- E(det Ll(>.)).
Proof. The proof is based on the observation that the Laplace transform
of a small solution is an entire function which is £ 2-integrable along the
imaginary axis. First we compute the Laplace transform of the derivative
of the solution of Equation (3.2):
.C(x)(z) = 1oo e-zsx(s) ds

= 1 00
e-zs lor d['T](B)]x(s+B)ds
00
=lor d['T](B)]1 e-zsx(s +B) ds
=lor ez 0 d['T](B)l[ -1°r e-zT ¢(7) d7 + 1 : e-zsx(s) ds]
=lor ez&d['T](B)]ler e-zT ¢(7) d7 +lor ez&d['T](B)] x(z).
Therefore, the Laplace transform of the solution of Equation (3.2) becomes
So, x is explicitly given by the equation

x(z) = Ll(z)- 1 [e-zr ¢(0) + z 1: e-z(r+s)¢(s) ds
10
(3.6)
-lor d['T](B)]e-z(r+s-B)¢(s) ds].
Since x decays faster than any exponential, the Laplace transform of x

converges for every z in <C. Further, the Plancherel theorem implies that xis
L 2 -integrable along the imaginary axis. Sox is an entire function that is L 2 -
integrable along the imaginary axis. From the explicit representation (3.6)
for x, x.
we next compute the exponential type of From this representation
of the right-hand side of Equation (3.6), it follows that the exponential type
of xis at most nr- E(det Ll(z)). Thus, the Paley-Wiener theorem (Lemma
3.1) implies the result. 0
In contrast to the general nonlinear case, the following result is true

for linear systems.
Corollary 3.1. For linear autonomous systems of RFDE, the equivalence

classes are determined in a finite time.
Proof. If ¢ and 7/J belong to the same equivalence class of a linear sys-
tem, then the solution x corresponding to ¢ - 7/J is a solution and must
be identically zero after some finite time. Therefore, this solution ap-
proaches zero faster than any exponential and, thus, is a small solution
of (3.2). Theorem 3.1 implies that any small solution is identically zero for
t;::: nr- E(detL\(z)). Therefore, each equivalence class is determined in a
finite time and the assertion is proved. D
Observe that in order to study the nontrivial small solutions of Equa-

tion (3.2), it suffices to analyze the null space of the solution operator.
Define the ascent a of a semigroup T(t) by the value
(3.7) a=inf {t: forall€>0:N(T(t)) =N(T(t+€)) }.
We shall prove an explicit expression for the ascent of the solution operator
for Equation (3.2) solely in terms of the characteristic matrix L\(z). For
this, we have to introduce the following two numbers: Define € and a by
(3.8) E(det L\(z)) = nr- €, ~a,.x E(Cij(z)) = (n- 1)r- a.

l::;z,J::;n
Theorem 3.2. The ascent a of the solution operator T(t) associated with
Equation (3.2) is finite and is given by
(3.9) a= €-a.
Before we sketch a proof of Theorem 3.2, we would like to state the

following important consequence.
Theorem 3.3. The solution operator T(t) associated with Equation (3.2) is
one-to-one if and only ifE(detL\(z)) = nr.
The following example will illustrate the results.
Example 3.1. Consider the following system of differential difference equa-

tions
±1(t) = -x2(t) + x3(t -1)
(3.10) X2(t) = X1(t- 1) + X3(t- ~)
X3(t) = -X3(t).
+
The characteristic matrix is given by
<l(z)~ ( 1
Z -e 2
-e_-L)
0 z+1
with determinant
det Ll(z) = (z + 1)(z2 + e-z).

So, E = 2. The cofactor
-e-z I
-e -lz
2
has exponential type 2, and hence a= 0. Therefore, from Theorem 3.2, the
ascent of the System (3.10) equals two.
Sketch of proof of Theorem 3.2. For t 2:: 0, the solution of Equation

(3.2) satisfies the integral equation
x(t) =lot ry(s- t)x(s) ds + F¢(t)

where
F¢(t) = ¢(0) +lot j_sr d[ry(B)]¢(s +B) ds.
Let F be the Banach space of continuous functions on the half-line [0, oo)
that are constant on [r, oo), provided with the supremum norm. The map-
ping F, considered as F : C ---+ F is not onto and it is not easy to charac-
terize the range of F. Therefore, we first consider the auxiliary problem to
analyze the small solutions of the integral equation
(3.11) y(t) =lot ry(s- t)y(s) ds + f(t)

where f belongs to F. The Laplace transform of y satisfies the equation
(3.12)
Suppose that the Laplace transform of y is a finite Laplace transform given

by J0°e-zty(t) dt. This implies that the right-hand side of (3.12) is an entire
function. The exponential type of the right-hand side of (3.12) is easily
computed to be less than or equal toE- a. So
E(lo 0
e-zty(t) dt) :::; E-a.
Therefore, if y is a solution of (3.11) that is identically zero after finite

time, then y(t) = 0 for t ;:::: E-a. On the other hand, in Verduyn Lunel
[1], we constructed a function f in F such that the right-hand side of
(3.12) is a finite Laplace transform of exponential type precisely E - a.
So for this function f, we have a solution y of Equation (3.11) such that
y(t) =f 0 fort in a neighborhood of E-a. To complete the proof, we claim
a one-to-one correspondence between solutions of Equation (3.2) and the

integral equation (3.11). For fin F, the solution y( ·;f) of Equation (3.11)
is continuous on the interval [0, r] and since f is constant on [r, oo), y( ·;f)
£:
satisfies the differential equation
y(t) = d7](e)y(t +e) on [r, oo).
So, if we define ¢(e) = y(r +e), -r ::; e ::; 0, then the solution x( ·; ¢) of
Equation (3.2) if given by x( ·; ¢) = y(r + · ). On the other hand, given an
initial condition¢ inC, there exists a solution to Equation (3.11) such that
y = ¢(r + ·) on the interval [-r, OJ; just define fin F by
(3.13) f = ¢(r + . ) - lt [~s d7](e)¢(r + s +e) ds.
This proves the correspondence between the solutions of the integral equa-
tion (3.11) and Equation (3.2) and completes the proof of Theorem 3.2. D
Proof of Theorem 3.3. By Theorem 3.2, it suffices to prove that for all E > 0,
a < E. Suppose a = E. We shall calculate E ( det adj Ll( z)) in two different
ways. Since a = E, we have
E(detadj Ll(z))::; n((n -1)r- E)

=(n-1)(nr-E)-E.
On the other hand, by Lemma 3.2, we have
E( det adj Ll(z)) = E( (det Ll(z) t~ 1 )

= (n- 1)(nr- E).
This implies that
(n- 1)(nr- E) ::; (n- 1)(nr- E)- E,
which is a contradiction if E > 0. D

In Chapter 1, Section 5, we have seen that the existence of nontriv-
ial small solutions is closely related to the question whether a solution of
Equation (3.2) has a convergent series expansion in characteristic solutions
cj(t)e>-.jt, where Pj is a polynomial and Aj a root of the characteristic
equation
(3.14) det Ll(z) = 0.
The procedure to obtain a series expansion was through the Laplace trans-
form. From Representation (3.6) for the Laplace transform of a solution of
Equation (3.2) one finds for t > 0,
1 l'"Y+wi
(3.15) x(t;¢) = lim - . eztH(z,¢)dz
W-HXl 21!"~ ')'-Wi
where
H(z, ¢) = Ll(z)-- 1 [e-zr ¢(0) + z /_or e-z(r+s)¢(s) ds

(3.16)
-/_or 10 d[ry(B)]e-z(r+s-O)¢(s) ds].
By Cauchy's residue theorem, we deduce that
m
x(t;¢) = L~~seztH(z,¢)+xam(t)
j=1 J
(3.17) m
= LPi(t)e,\Jt + Xam (t)
j=1
where the summation is over the roots of the characteristic equation (3.14)
with real part larger than am. For the remainder, we have the following
integral representation:
Xa(t) =
1
lim - .
w->oo 21!"~
1 a+wi
a-wi
ezt Ll(z)- 1 [e-zr ¢(0) + z
Jo
-r
e-z(r+s)¢(s) ds
-/_or 10 d[ry(B)]e-z(r+s-O)¢(s) ds] dz.

In general, one cannot expect that the series expansion converges to the
solution. In particular, small solutions have an entire Laplace transform
and hence a series expansion in which all terms are zero. However, if the
series expansion converges, we can decompose the solution as a convergent
series of characteristic solutions and a small solution. At this point, we
encounter two problems. Firstly, when does the series expansion actually
converge to the solution, and secondly, can we identify the set of initial
values that correspond to a solution with a convergent series expansion.
The first question can be answered similarly to the presentation in Chapter
1, Section 1. The second question boils down to the analysis of the closure
of the span of the system of eigenvectors and generalized eigenvectors of
an unbounded operator. We return to these questions in Chapter 7. In the
remaining part of this section, we present the main results and illustrate
them with several examples.
To get a better feeling for these questions, let us consider, in detail,
the equation
x(t) = Ax(t- 1), t ~ 0,
(3.18)
The determinant of the characteristic matrix is a polynomial det Ll( z) = z 3.

So the series expansion in characteristic solutions is always finite. As before,
let x(z) = f 000 e-ztx(t- 1) dt. An easy computation yields the following
representation for the Laplace transform of a solution
x1(z) = e-z [¢1(0)

z
+ jo-1
e-zs¢1(s) ds + x2(z)]
(3.19) x2(z) = e-z [¢2(0) + jo e-zs¢2(s) ds + x3(z)]

z -1
X3(z) = e-z [¢3(0) + jo e-zs¢3(s) ds].

z -1
System (3.19) contains all of the information. First we compute the char-
acteristic solutions at >.. = 0 by computing the residue at z = 0. This yields
c3(t) = c, c2(t) = b + ct and c1(t) =a+ bt + ct 2, where a,b, and care
arbitrary constants. Define M to be the subspace spanned by the initial
conditions corresponding to characteristic solutions
M = {¢ = (¢1, ¢2, ¢3) E C: ¢3(()) = c, ¢2(()) = b + c(),

(3.20) ()2
¢1(()) =a+ b() + c 2 }·
So M is a subspace of dimension three that is invariant under the solution
operator; that is, T(t)M <:;;; M for all t 2: 0. In fact, if¢ EM, then x3(t) = c
for -1 :::; t:::; 0, and so x 2(t) = b + f~ cds fort 2: 0. Since x 2(()) = b + ce,
-1 :::; () :::; 0, it follows that x 2(t) = b + ct for t 2: -1. In the same way, one
shows that x 1(t) =a+ bt + (ct 2/2) fort 2: -1. This proves Xt(¢) EM for
all t 2: 0 if¢ E M.
If xis a solution of Equation (3.18) with initial value¢ that approaches
zero faster than any exponential, then ¢3(0) = 0 and x3(t) = 0 fort 2: 0.
Therefore,
x2(t) = ¢2(0) +lot ¢3(s- 1) ds, 0:::; t:::; 1
x2(t) = ¢2(0) + 11
¢3(s- 1) ds, t 2: 1,
and one must have ¢ 2(0) = - J~ 1 ¢ 3(()) d(). In the same way, one must have
(3.21)
Therefore, if x is a small solution of Equation (3.18), then the initial value

1> of x must belong to the set S defined by
0
S = { 1> = (¢1, 1>2, ¢3) E C: ¢3(0) = 0, 1>2(0) = - /_ 1 ¢3(8) dB,
(3.22)
1>1(0) satisfies Equation (3.21) }·
The closed subspace S has codimension three. By Theorem 3.2 S

N(T(3) ), which can be verified directly by choosing 1> 1 arbitrary, 1> 1(0) =
-3/7r,¢2 (B) = 2/7r, -1 :=::; B :=::; 0, and ¢ 3 (8) = sin1rB, -1 :=::; B :=::; 0. Observe
that M n S = { 0} and we can decompose C as the direct sum of two closed
invariant subspaces
(3.23)
For any initial condition 1> not belonging to M, it follows from System
(3.18) that the solution x( ·; 1>) has a component that is a small solution.
What is the form of the original equation (3.18) on the invariant sub-
space M? Let us first observe that one can extend the definition of T(t) for
negative values oft on Min such a way that T(t)M ~ U, t E (-oo,oo).
To accomplish this, one simply defines the polynomial functions (x 1 , x 2 , x 3 )
as earlier and verifies that Equation (3.18) is satisfied. Choose as a basis
forM the functions E M is given as in Expression (3.20), then one observes
that y(O) = (a, b, c). Also, a direct calculation shows that y satisfies the
ordinary differential equation
(3.25) iJ = Ay
where A is the same matrix as in Equation (3.18). Since A is in Jordan
canonical form, it is certainly natural to say that the canonical form for
Equation (3.18) is the equation itself.
The small solutions can also be found from System (3.19). Recall that
the Laplace transform of a small solution is an entire function. Therefore, an
initial value 1> corresponds to a small solution if and only if the right-hand
side of System (3.19) is entire.
This decomposition could be computed directly without invoking the
Laplace transform, but we have chosen to illustrate the use of the Laplace
transform since this approach also can be used to analyze more complicated
examples when the characteristic equation does not reduce to a polynomial.
In that case, the space M becomes infinite dimensional and the question
becomes: Does Decomposition (3.23) still hold with M replaced by M the

closure of M.
First we introduce some more notation. Set H(z, ¢) ~f J000 e-ztx(t; ¢)dt.
Then
H(z ¢)- P(z, ¢)
' - det Ll(z)
where
P(z, ¢) = adj Ll(z) [e-zr ¢(0) + z lor e-z(r+s)¢(s) ds

(3.26)
-£: 10 d[TJ(B)]e-z(r+s-e)¢(s) ds].
From the representation for the Laplace transform of the solution x( ·; ¢)

of Equation (3.2), we find the following characterization of the null space
of the solution operator.
Corollary 3.2. N(T(o:)) = {¢ E C: Ll(z)- 1 H(z,¢) is entire}.
In general, convergence results for the series expansion in characteristic

solutions are delicate. However, it turns out that M, the closure of the set
of initial values such that the solution has a convergent series expansion
can be characterized.
Theorem 3.4. M = {¢ E C: E(P(z,¢)) :<:; E(detLl(z))}.
An important consequence of these characterizations is the fact that
M nN(T(o:)) = {0}.
In fact, ¢ E M n N( T(o:)) implies that the Laplace transform x(z)

J000 e-ztx(t- r) dt of the solution x = x( ·; ¢) is an entire function of zero
exponential type, which is L 2 -integrable along the imaginary axis. There-
fore, the Paley-Wiener theorem (Lemma 3.1) implies that x(t- r) = 0 for
t :2: 0. Hence ¢ = 0. Another easy consequence of the theorem is that M is
invariant under the solution operator.
Furthermore, we have the following "almost" decomposition.
Theorem 3.5. C = M ffi N( T(o:)).
So the subspace M represents, in a certain sense, the minimal amount

of initial data needed to specify a solution of Equation (3.2). If more initial
data are specified, then the extra part belongs to M for t :2: o:.
The proofs of these results are involved and will be discussed in Section
7.8. Here we would like to illustrate the results with an example.
Example 3.2. Consider the following system

X1(t) = X2(t- 1)
(3.27)
X2(t) = X1(t).
The characteristic matrix is given by
L\(z) = [ ~1 -:-z] '
So the ascent a of the solution map equals one and it is easy to see that
N(T(1)) = {(<h,¢2) E c: ¢1(0) = 0, ¢2 = 0}.
Next, we use Theorem 3.4 to compute M. Using Equation (3.27), the
Laplace transform of the solution satisfies the system of equations
ezzx1(z) = cf>1(0) + x2(z) + z [ 0

1 e-ztcf>1(t) dt,
ezzx2(z) = cf>2(0) + ezx1(z)- [~ e-ztcf>1(t) dt + z [ 0

1 e-ztcf>2(t) dt.
Hence
L\(z) [~1(z)] = [ft(z,cp)]
x2(z) h(z, ¢)
where
ft (z, cf>) = cf>1 (0) + z [~ e-ztcf>1 (t) dt
h(z, cp) = e-z¢2(0)- e-z [ 01 e-ztcf>1(t) dt + ze-z [ 01 e-ztcf>2(t) dt.
Therefore,
[ x1(z)]- 2 1 [zft(z,cf>)+e-zh(z,cp)]
X2(z) - z +e-Z ft(z,cp)+zh(z,cp) •
The exponential type of both ft(z, ¢)and h(z, ¢)is at most one. By The-
orem 3.4
M = {¢ E C: E(h(z,cp)) = 0}
which implies that ¢2(t) = cf>2(0) + J~ cf>1(s) ds. So
M = {4> E C: cf>2(t) = cf>2(0) +lot cf>1(s) ds }.
Note that indeed N( T(1)) n M = {0}, but that N( T(1)) EB M is only

dense in C.
3.4 Unique backward extensions
In this section, we wish to discuss backward extensions of solutions of an

RFDE that exists on ( -oo, OJ. The question of interest is: When is this
backward extension unique?
Property 4.1. Backward extension on ( -oo, OJ of linear autonomous RFDE
is unique.
Proof. Suppose x andy are backward extensions on ( -oo, OJ of¢ E C for
a linear autonomous RFDE. These solutions have unique extensions for
t 2 0. If z = x - y, then z is a solution of the equation on ( -oo, oo) with
z 0 = 0. Fix 17 < 0 and consider the solution z on [17 - nr - r, oo). This is
a solution that approaches zero faster than any exponential. Consequently,
Theorem 3.1 implies that the solution Zt through Zcr-nr is equal to zero for
t 2 17 - nr + nr = 17. Therefore, Zcr = 0. Since 17 is arbitrary, this implies
z(t) = 0 for t :::; 0 and therefore, the backward extension is unique. 0
Property 4.2. There may exist two distinct backward extensions on ( -oo, OJ
for an autonomous RFDE.
Proof. Let r = 1, f(s) = 0, 0 :::; s :::; 1, f(s) = -3(s 113 - 1) 2 , s > 1, and
consider the equation
x(t) = f(lxtl).
The function x = 0 is a solution of this equation on ( -oo, oo). Also, the
function x(t) = -t 3 , t < 0, = 0, t 2 0 is also a solution. In fact, since x:::; 1
for t 2 -1, it is clear that x satisfies the equation for t 2 0. Since x is
monotone decreasing fort:::; 0, lxtl = x(t -1) = -(t -1) 3 and x(t) = -3t 2 •
It is easy to verify that -3t 2 = f((1- t) 3 ) fort< 0. The proof is complete.
0
If f : C --> IRn has infinitely many bounded derivatives and x is a

bounded solution of the RFDE(f) on ( -oo, 0], then x will have infinitely
many derivatives. Even in this case, it is possible to give examples for which
bounded backward extensions on ( -oo, OJ are not necessarily unique. When
can one make a positive statement? One can prove:
Property 4.3. Iff : C --> IR n is analytic, then bounded backward extensions
on ( -oo, OJ are unique.
Proof. If one could show that any bounded backward extension x on ( -oo, OJ
is an analytic function, then the uniqueness would be immediate. That this
is true is a consequence of the following result, which is stated without
~~ 0
Theorem 4.1. If x is a bounded solution on ( -oo, OJ of an analytic RFDE(f),

f : C --> IRn, then x is analytic in a region containing ( -oo, OJ.
3.5 Range in IRn 87
3.5 Range in lRn
For ordinary differential equations, the solution map defines a homeomor-

phism and so the image of a ball under this map contains a ball. For an
RFDE, the phenomenon may not be true even if we consider the map
x (a, ·) (t) : C --t ffi.n. More precisely, we have
Property 5.1. For nonautonomous linear equations x(a, C)(t) may be zero
dimensional for some t >a. In fact, one may have x(a, C)(t) = 0 for all t
greater than or equal to some T.
(5.1) ±(t) = b(t)x(t- 3;)
where, for an arbitrary g E C([-oo,O],rn.n), the function b is defined by

g(t) t ~ 0,
{ 0 O<t<
- -
3 11"
2'
(5.2) b(t) =
-cost 3 11"
2 -
<t<
-
311"
'
1 t:::: 31!".
For a = 0 and an arbitrary ¢ E C([-37r /2, 0], ffi.), Equation (5.1) has the
solution
¢(0) 0~t ~ 3; '
(5.3) {
x(O,cp)(t)= (-sint)¢(0) t;:::~~
Thus, x(O,cp)(t) = 0 fort= k1r, k = 2,3, ... and all¢ E C([-37r/2,0],ffi.).
This proves the first statement of Property 5.1.
To prove the more striking conclusion in the second part, we consider
the equation
(5.4) ±(t) = -a(t)x(t- 1)

where
a(t) = { 2sin2 (1rt) t E [2n,2n+ 1]
0 t E (2n- 1, 2n)
for each integer n. For any a Ern.,¢ E C, we show T(t, a)¢= 0, t;::: a+ 4.
In fact, if N is the smallest odd integer such that N ;::: a, then x(t) = x(N),
t E [N, N + 1] and
±(t) = -a(t)x(N), tE [N+1,N+2].
1
Thus,
N+2
x(N + 2) = x(N) [1- 2 sin 2 (1rs) ds] = 0
N+l
and x(t) = 0 fort E [N + 2,N + 3] and x(t) = 0 fort 2:: N + 2. This

completes the proof. D
For autonomous linear systems, one has the following surprising result.
Property 5.2. For n-dimensional autonomous linear differential difference
equations, x(a, C)(t) may have dimension< n for some t.
Proof. The following example is an autonomous linear equation of dimen-
sion 3 for which x(O, C)(t) has dimension 2 for t 2:: some t 0 . Consider the
solutions of
±(t) = 2y(t)
(5.5) y(t) = -z(t) + x(t- 1)
i(t) = 2y(t- 1)
defined on [-1, oo). Fort 2:: 1, y(t) = 0, x(t) = 0 and thus y(t) is at most a
first-degree polynomial in t and x(t) is at most a second-degree polynomial
in t. Therefore
y(t- 1) - y(t) = -y(t)
x(t)- x(t -1) = ~[±(t) + x(t- 1)]
and
(x, y, z)(t) · (1, -2, -1) = x(t)- 2y(t)- z(t)
= x(t- 1) + x(t- 1) + x(t) - 2y(t) - z(t)
2
= x(t- 1) + y(t- 1) + y(t)- 2y(t) - z(t)
= x(t- 1) - y(t) - z(t) = 0.
Consequently, (x, y, z) (t) is orthogonal to (1, -2, -1) and must lie in a plane
for t 2:: 1.
With one delay as in the preceding example, it is actually possible to
prove that dimension 3 is the lowest dimension for which this phenomenon
can happen. We do not give the proof. For more than one delay, it can even
happen in dimension 2. Suppose x = (x1, x 2 ) is a 2-vector and consider the
equation
(5.6) x(t) = (1 -1) x(t)+.(-4_ 3)

0 0 4 4 x(t-ln2)+ (4 -2)
8 _4 x(t-ln4).
If x(t) = cexp>.t, c = (c 1 ,c2 ) is a solution, then>. must satisfy the char-

acteristic equation >. 2 - >. = 0 and the corresponding exponential solutions
are x(t) = (0, 1) and x(t) = (0, 1)et.
One could prove that x(O, C)(t) is one-dimensional for t 2:: some to
by showing directly that this set is orthogonal to some particular nonzero
vector. However, we choose to give a different proof using Theorem 3.1
3.6 Compactness and representation 89
and some elementary results on linear systems that will be developed later.
Even though this is not the logical way to proceed, the reader should find
the proof interesting.
Consider the linear operator T(t) = T(O, t) : C -----+ C defined by the
solution of Equation (5.6). Later we show T(t) is completely continuous for
t;::: ln4 and that the only elements in the spectrum ofT(t) are 1, et and 0.
Furthermore, 1 and et are simple eigenvalues with corresponding eigenvec-
tors (0,1) and (0, 1)et. Consequently, any solution with zero projection on
the span of these two eigenvectors must be associated with the zero-point of
the spectrum of T(t) and thus approach zero faster than any exponential.
Finally, Theorem 3.1 implies the corresponding solution is identically zero
fort ;::: /3, j3 = (n- 1)r- r. This proves the result. 0
3.6 Compactness and representation
The previous sections have been devoted mainly to peculiar properties that
some RFDE may have. In this section, we state some positive results that
are valid for all equations and will be extremely useful in later chapters.
Lemma 6.1. For any t;::: a+ r, the map T(t, a) is locally completely contin-
uous; that is, for any t;::: a+ r, ¢ E C, there is a neighborhood V(t, a,¢)
of¢ such that T(t, a)V(t, a,¢) is in a compact set of C.
Proof. Since f is continuous and T(r,a)'lj; is continuous in (r,a,'lj;), for any
t ;::: a, there is a neighborhood V (t, a, ¢) of ¢ and a constant M such that
\T(r,a)V(t,a,¢)1::; M, if(r,T(r,a)V(t,a,¢))1::; M for a::; T::; t. Thus
\±(a, V (t, a, ¢)) (T) I ::; M for a ::; T ::; t. This implies the family of functions
{ Xt (a, 'ljJ) : 'ljJ E V (t, a,¢)} is precompact for t ;::: a + r and proves Lemma
6.1. 0
For simplicity in the statement of the global properties of T(t, a), let
us assumeD= IR x C and T(t, a) : C-----+ Cis defined for all t;::: a and, of
course, T(t,a)¢ is continuous in (t,a,¢). The following definitions are also
convenient.
Definition 6.1. Suppose a mapping A(t, a) : C-----+ Cis defined for all t;::: a.
The mapping A(t, a), t ;::: a, is said to be conditionally completely contin-
uous if A(t, a)¢ is continuous in (t, a,¢) and, for any bounded set B ~ C,
there is a compact set B* ~ C such that A( r, a)¢ E B for a ::; T ::; t implies
A(t, a)¢ E B*.
Definition 6.2. If X, Y are metric spaces, then a mapping A : X -----+ Y is

bounded if A takes bounded sets of X into bounded sets of Y. If A(A.) :
X --+ Y depends on a parameter in a metric space A, then A(>.) is said to

be bounded uniformly on compact sets of A if, for any compact set A 0 <;;: A
and any bounded set X 0 <;;: Y, there is a bounded set Y 0 <;;: Y such that
A(>.)Xo <;;: Yo for all >. E Ao.
The following lemma is obvious from the definitions.
Lemma 6.2. If A(t, u) : C --+ C is defined fort ~ u and A(t, u) is a map

bounded uniformly on compact sets of [u, oo), then A( t, u) conditionally
completely continuous implies A( t, u), t ~ u completely continuous.
Define the operator S(t) : C--+ C, t ~ 0, by the relation
(6.1) S(t)¢(B) = { ¢(t +B)- ¢(0) t + () < 0

0 t + () ~ 0, -r ::; () ::; 0.
The operator S(t) is a bounded linear operator on C for each t ~ 0 and

satisfies
S(t + T) = S(t)S(T) fort~ 0, T ~ 0
(6.2)
S(t)=O fort~r.
The last relations are obvious consequences of the definition. Therefore, for
any a> 0, there is aKa such that
(6.3)
In the statement of the next result, we use the concept of an a-contraction,

which is described in Definition 3.6 of Chapter 4.
Theorem 6.1. For a given RFDE(f) for which f : lR x C --+ lRn is a bounded
continuous map, the solution map T(t, u), t ~ u, can be written as
(6.4) T(t, u) = S(t- u) + U(t, u)
where S(t), t ~ 0, is defined in (6.1), satisfies (6.3), and U(t, u) : C --+

C, t ~ u, is conditionally completely continuous. Thus, T(t, u) is an a-
contraction for t > u and is conditionally completely continuous for t ~
u+r.
Proof. With S(t), t ~ 0, defined as in Equation (6.1), it follows that U(t, u)¢
is defined as
¢(0) t + () < u,
{
U(t, u)¢(O) = ¢(0) + J;+IJ f(s, T(s, u)¢) ds t + () ~ u, -r::; ()::; 0.
For any bounded set B <;;: C, the fact that S(t) is a bounded linear operator
implies there is a bounded set B 1 <;;: C such that T(s, u)¢ E B1 for u::; s::; t
provided that U(s, u)¢ E B for u::; s::; t. Since f is a bounded continuous
3.7 The solution map for NFDE 91
map from IR x C to IRn, there is an M depending only on B, u, and t such

that lf(s, T(s, u)¢)1 ::; M for u ::; s ::; t if U(s, u)¢ E B for u ::; s ::; t.
Therefore, under these suppositions,
if3 If( s, T( s, u )¢)Ids ::; M({J -a)
for u ::; a ::; {3 ::; t. Let
B* = {¢ E C: ¢ E B, 1¢(0)- ¢(()1::; MIO- (1, 0,( E [-r,O]}.

These computations show that U(t, u)¢ E B* since U(t, u)¢(0) is constant
for t + 0 ::; u. The proof of the theorem is completed by observing that
S(t) = 0 fort 2: r. D
The following corollaries are immediate.
Corollary 6.1. Iff : IR x C ---+ IRn is a bounded continuous map and

T(t,u): C---+ Cis a map bounded uniformly on compact sets of [u,oo),
then
T(t, u) = S(t- u) + U(t, u), t 2: u,
where S(t), t 2: 0 is defined in Equation (6.1) and U(t, u), t 2: u, is com-
pletely continuous.
Corollary 6.2. Iff : IR x C ---+ IRn is a bounded continuous map and

T( t, u) : C ---+ C is a map bounded uniformly on compact sets of [u, oo),
then T(t, u), t 2: u + r, is completely continuous.
Corollary 6.3. For r > 0, the solution map T(t, u) fort > u can never be
a homeomorphism iff : IR x C ---+ IRn is a bounded continuous map and
T(t, u), t 2: u, is a bounded map.
Proof. This follows immediately from the preceding theorem and the fact
that the unit ball in C([-r,O],IRn) is not compact for any r > 0. D
3. 7 The solution map for NFDE
For a NFDE( D, f) on S? and any (u, ¢) E S?, there is a ta,¢ and a non-
continuable solution x through (u,¢) defined on [u- r,ta,q,). If we speak
of only noncontinuable solutions, then we can define the solution map
T(t, u): Da---+ C where Da = {1P E C: (u, ¢)ED} and T(t, u) = Xt(u, ¢).
For RFDE, the solution map T(t, u) smoothes the initial data as t
increases. Also, if a solution of an RFDE can be extended in the negative
direction through a point (u, ¢), then¢ must satisfy some differentiability
conditions. For a large class of NFDE, neither of these properties holds.

This large class is the one for which the function D uses the information
about the function x at t- r as well as t. In particular, we can state the
following result.
Theorem 7.1. If the NFDE(D, f) on an open set [l satisfies the property that
D(t,¢) is atomic at -r for each (a,¢) ED, then there are t 1(a,¢) <a<
t 2(a,¢) and a noncontinuable solution x of the NFDE( D, f) on h (a, ¢) -
r < t2(a, ¢);that is, T(t, ¢): Da--+ Cis defined for t1(a, ¢) < t < t2(a, ¢).
Furthermore, if f(t, a) has a continuous first derivative in¢, then
T(t, a) : Da--+ C
is a homeomorphism for each t 1(a, ¢) < t < t 2(a,¢).

Proof. The same ideas used in the proof of Theorem 8.1 of Chapter 2 are
easily adapted to prove the existence of a solution through (a,¢) in the
negative direction. The extra smoothness condition on f implies uniqueness
and continuous dependence of the solution. The proof is complete. D
A very simple example illustrating the last result is the equation

d
dt [x(t)- Ax(t- r)] = f(t, xt)
where A is an n x n constant matrix with det A =/:- 0. Iff (t, ¢) is continuously

differentiable in ¢, then the solution operator of this equation defines a
homeomorphism.
These methods can be used to prove results on NFDE concerning the
differentiability with respect to parameters, but we do not dwell on this
question. For the case of Theorem 7.1, the implication will be that the
solution operator is a diffeomorphism.
The fact that solutions of NFDE can have backward extension for every
(a,¢) E [l allows one to prove the following result for linear systems, which
contrasts drastically with RFDE (see Section 3.3).
Theorem 7.2. Suppose T E IR is given, [l = (T, oo) x C, (D, f) defines

an NFDE on D, D(t, ¢) is linear in ¢, f(t, ¢) is linear in¢, there is a
positive constant k such that ID(t, ¢)1:::; kl¢1, lf(t, ¢)1:::; kl¢1 for (t, ¢)ED,
D(t,¢) is atomic at 0 and -r uniformly with respect to (t,¢) ED. Then
for any (a, ¢) E [l, there is a unique solution of the NFDE( D, f) through
(a, ¢) defined on (T, oo) and, if a solution x approaches zero faster than any
exponential as t--+ oo, then x(t) = 0 for all t E (T, oo).
Proof. Existence and uniqueness follow by applying these results. Following
the same reasoning as in the proof of Theorem 7.3 of Section 1.7, there are
positive constants a and b such that for any a E ( T, oo), any solution x
satisfies
3. 7 The solution map for NFDE 93
t E (T, oo).
Now, suppose xu -+ 0 faster than any exponential as a-+ oo and there is
atE (T, oo) such that lxtl > 0. For a 2: t, we have
lxtlexpbt:::; alxulexpba.
This gives a contradiction and proves the theorem. D
Our next objective is to obtain a representation of the solution operator

of an NFDE in a form similar to one in Theorem 6.1 for RFDE. We restrict
our attention to the case where the equation is autonomous in order to
avoid excessive notation.
As preparatory material, we need some notation associated to the ho-
mogeneous "difference" equation
(7.1) t 2: 0,
where D : C -+ 1Rn is linear, bounded, and atomic at 0. Without loss in
generality, we may assumeD¢= ¢(0)- f~r d[~-t(0)]¢(0), where Var[-s,o]l-£-+
0 ass-+ 0.
If
(7.2) Cv = { ¢ E C : D¢ = 0}
then Equation (7.1) defines a semigroup of linear operators Tv(t), t 2: 0,
on Cv, where Tv(t)'ljJ = Yt('l/J), t 2: 0, 'ljJ E Cv, and Yt('l/J) is the solution of
Equation (7.1) through (0, '1/J).
Lemma 7.1. If D: C-+ 1Rn is linear, bounded and atomic at zero, then there
are n functions ¢1, ... , ¢n such that DiP= I, where iP = (¢1, ... , ¢n)·
Proof. For any s E [0, r], let 'ljJ E C([-r, O],R) be defined by
0
'ljJ(O) = 0, -r :::; 0 :::; -s, '1/J(O) = 1 + -, -s < 0 :::; 0.
I:
s
Then
D('!jJI) =I- d[~-t(0)](1 + ~).
If s > 0 is sufficiently small, then det D('ljJI) =f. 0 and the columns of
the matrix D('ljJI) form a basis for 1Rn. If we let iP = 'ljJI(D('ljJI))-1, then
DiP = I and the lemma is proved. D
Let f : C -+ 1Rn be a continuous function and let Tv,t(t) be the

semigroup defined by the NFDE (D,f),
d
(7.3) dt Dxt = f(xt), t 2: 0.
With this notation, we can state the following important representation

theorem for the semigroup associated to Equation (7.3).
= I, be defined by Lemma 7.1 and let IJr = I- D.

Theorem 7.3. Let , D
Iff is completely continuous and ifTD,J(t) is a bounded map for each t > 0,
then
TD,t(t) = TD(t)!Jr + U(t)
where U(t) : C-+ C, t ;?: 0 is completely continuous.
To prove this result, we need the following auxiliary result on the non-
homogeneous "difference" equation
(7.4) Dyt = h(t), t;::: 0,

with hE C([O, oo), IRn).
Lemma 7.2. If D : C -+ IRn is bounded, linear, and atomic at zero and

y('lj;, h) is the solution of Equation (7.6) satisfying Yo('l/J, h)= 'lj;, then there
are positive constants a and b such that fort ;?: 0,
iy('lj;, h)(t)i S beat[ 1'1/JI + sup lh(s)!].

o::;s::;t
Proof. The proof is technical, but can be supplied by modifying the argu-
ment used in the proof of Theorem 7.3 of Section 1.7. See also Theorem 3.4
of Chapter 9. D
Proof of Theorem 7.3. If U(t)¢ = TD,J(t)<P- TD(t)!Jr¢, then U(t)¢ satisfies

the equation
DU(t)¢ = D¢ +lot f(TD,t(s)</>) ds.
Lemma 7.1 and the boundedness assumption on TD,t(t) imply that U(t)
takes bounded sets into bounded sets. Also,
D[U(t + T)</>- U(t)¢] = t i t+r

f(TD,t(s)</>) ds.
Now suppose that -
 D<PI can be made arbitrarily small by taking T small. Since f is completely
continuous and no<s<t TD,t(s)B is bounded for each t, it follows from
Lemma 7.2 that for -any fixed a> 0, t E [0, a], the expression IU(t + T)</>-
U(t)<Pi can be made arbitrarily small by taking T small. This shows that
U(t) is completely continuous. D
We now assume that the kernel J-L in the definition of D has no singular
part; that is,
1:
3. 7 The solution map for NFDE 95
D¢ =Do¢- A(0)¢(0) dO
(7.5)
2: Ak¢( -rk)
00
Do(¢) = ¢(0) -
k=l
where
t; IAkl +
00 0
0 < rk $ r, [r IA(O)I dO< oo.
Theorem 7.4. Suppose that D satisfies (7.5), let Po, DoPo = I be defined
by Lemma 7.1, and let llio =I- PoDo. Then
TD(t) = TD 0 (t)llio + U(t)

where U(t) : C---. C is completely continuous fort 2:: 0.
Proof. If U(t)¢ = TD(t)¢- TD 0 (t)llio¢, then U(t)¢ is linear and continuous
1:
in ¢ and satisfies the equation
DoU(t)¢ = A(O)(TD(t)¢)(0) dO.
For any r E IR, t 2:: 0, we have
Do[U(t + r)¢- U(t)¢]
=[or A(O)(TD(t + r)¢)(0) dO- 1: A(O)(TD(t)¢)(0) dO
= 1-r+r A(O- r) (TD(t)¢) (0) dO- [or A(O) (TD(t)¢) (0) dO
= ([~r+r + 1r )(A(O- r) - A(O) )(TD(t)¢) (0) dO
=: h(t,r,¢).
J:
Now we proceed as in the proof of Theorem 7.3 using the fact that
IA(s- r)- A(s)i ds---> 0 as r---> 0. D
Corollary 7.1. Suppose D as in (7.5), let P, DP = I, Po, DoPo = I, be

defined by Lemma 7.1, and let lli = I- PD : C ---> CD, llio = I- PoDo :
C ---> CDo. Iff : C ---> IRn is completely continuous and if TD,J(t) is a
bounded map for each t > 0, then
(7.6)
where U(t) : C---. C, t 2:: 0, is completely continuous.
Let av 0 be the order of the semigroup Tv 0 (t); that is,

(7.7) avo= in£{ a E lR: there is a k such that 11Tv (t)ll::; keat, t
0 ~ 0 }.
Lemma 7.3. If av0 is defined as in (7.7), then, for any a> av 0 , there is an
equivalent norm I· Ia inC such that
(7.8)
Proof. For any a > av 0 , there is a constant k such that

(7.9)
For any '1/J E C Do, we define
From (7.9), we have 11/JI ::; 11/Jia ::; ki'I/JI and so l·la is an equivalent norm on
Cv 0 and can be extended to an equivalent norm on C. Also,
ITvo(t)'l/Jia =sup 1Tv0 (t + s)'l/Jie-as
s~O
=eat sup 1Tv0 (t + s)'l/Jie-a(t+s)

s~O
::; eat l¢la,

and the lemma is proved. D
Corollary 7.2. If D, f satisfy the hypotheses in Corollary 7.1 and av 0 < 0,

then there is an equivalent norm in C such that in this new norm, the semi-
group Tv,J(t) defined by (7.4) is the sum of a contraction and a completely
continuous map for each t > 0.
Proof. If we choose a so that av 0 < a < 0, then this is an immediate
consequence of Lemma 7.3 and Equation (7.5). D
In Chapter 9, we will show that the number av 0 is determined by the

characteristic equation for the difference equation
00
(7.10) y(t)- L Aky(t- rk) = 0.

k=l
In fact,
00
(7.11) av 0 = sup{Re>.: det [I- LAke-.>.rk] = 0}.

k=l
Krasnovskii [1] was the first to emphasize the importance of considering

the state of a system defined by a functional differential equation as the
element Xt(a, ¢)of C. He made the observation that the converse theorems
of Liapunov on stability could not be proved by using a scalar function
V(t, x) that depends only on (t, x) in IRx IRn. In fact, if uniform asymptotic
stability of the solution x = 0 of
x(t) = J(xt)
implies the existence of a positive definite function V(x) such that
(oVfox)f < o,
then the solution x = 0 of
x(t) = kf(xt)
would be uniformly asymptotically stable for any positive k since
(oVfox)(kf) < o.
On the other hand, the linear equation
x(t) = -kx(t- 1)
has all roots of the characteristic equation A. = -ke->. with negative real
parts if k < 1r /2 and some with positive real parts if k > 1r /2. Therefore,
the converse theorems of Liapunov must make explicit use of the infinite
dimensionality of the problem.
Example (1.1) is due to Hannsgen. The example in Property 1.4 is due
to Charrier [1]. The same property was observed by Banks and Kent [1].
Example 2.3 and the example in Property 4.2 are due to Hausrath.
The paper of Henry [2] contains Theorem 3.1 and other interesting
implications of this result. Theorem 3.2 and 3.3 are from Verduyn Lunel
[1]. The results with their proofs carry over verbatim to linear autonomous
neutral functional differential equations NFDE( D, L). Related results can
be found in Bartosiewicz [1] and Kappel [4]. Theorem 3.1 is false for time-
dependent delay equations. The following example is due to Oliva
(8.1) x(t) = -2te 1 - 2 tx(t- 1)
which has the solution x(t) = e-t 2 on [-1, oo). From the result of Cooke
and Verduyn Lunel [1], it follows that the sign change in the coefficient of
Equation (8.1) is necessary in order to have nontrivial small solutions. Also
for linear periodic RFDE, Theorem 3.1 is false as follows from the following
example due to Stokes [2]
(8.2) x(t) = sin(27rt)x(t- 1).
In Section 3 of Chapter 8 we shall further analyze the existence of nontrivial

small solutions for periodic RFDE.
Theorem 4.1 is due to Nussbaum [1]. Example (5.1) is due to Zverkin
[1], and Example (5.4) to Winston and Yorke [1].
The special systems considered in Property 5.2 belong to the general
category of pointwise degenerate systems. Consider a differential difference
equation
m
(8.3) x(t) = Ax(t) + LBjx(t- jr)
j=l
where A, Bj, j 2: 1, are n x n constant matrices. Equation (8.3) is said to be

pointwise degenerate at h with respect to the p x n, p < n, matrix Q if for all
continuous initial functions given on [-mr, 0], the corresponding solution
x(t) satisfies Qx(tl) = 0. Otherwise, the equation is pointwise complete.
If a system is pointwise degenerate at t 1, then it is pointwise degenerate
for all t 2: t 1. Therefore, the solutions of Equation (8.3) will lie in the
subspace MQ of IRn defined by MQ = {x E IRn : Qx = 0} fort 2: t1. Weiss
[1] was the first to conjecture that there might exist pointwise degenerate
systems. With Example (5.5), Popov [1] showed the existence of pointwise
degenerate systems and gave some interesting properties of these systems.
With one delay, one cannot obtain a pointwise degenerate system for n ~ 2.
Zverkin [2] was the first to show by means of Example (6.6) that one could
have a pointwise degenerate system for n = 2.
A number of authors have been concerned with necessary and sufficient
conditions for System (8.3) to be pointwise degenerate and the construction
of pointwise degenerate systems (see, for example, Asner [1], Asner and
Halanay [1, 2], Charrier [1], Kappel [1, 2, 3], Popov [1, 2], Zmood and
McClamroch [1], and Zverkin [2]).
Let us reinterpret Example (5.5) of Popov [1] in terms of a control
problem. Let q E IRn be the vector (1, -2, -1). For System (5.5), it was
shown that qx(t) = 0 for all t 2: 1. Therefore, the ordinary differential
equation
w=Aw,
can be steered to the plane Mq = {x E IR3 : qx = 0} in time~ 1 and it will

remain in this plane for t 2: 1 provided that the feedback control is chosen
to be
Bw(t -1),
Therefore, the time optimal control problem is meaningful for this situation.
This question was discussed in more detail in Popov [2]. Asner and Halanay
[3, 4] considered similar questions by allowing the delayed feedback control
to involve the derivative x of x as well as x itself.
Popov [1] also observed in this first paper on degenerate systems that
it is impossible to solve this problem by using delayed feedback of the form
bcw(t- r) where b is a column vector and c is a row vector. Asner and
Halanay [4] showed the result of Popov remained valid even with several
delays.
Definition 6.1 and the representation of the solution operator in The-
orem 6.1 are due to Hale and Lopes [1].
Theorem 7.2 is due to Hale [1] and generalizes an earlier result of
Wright [3]. The representation theorem 7.3 first occurred in Henry [1]. The
importance of the difference operator Do has been recognized by Cruz and
Hale [4].
4
Autonomous and periodic processes
In this chapter, we discuss some fundamental properties of autonomous and

periodic RFDE. Because many of these properties also hold for abstract
processes, we will develop the theory in an abstract setting and point out
special features that apply to RFDE. The abstract setting also will be
sufficiently general to apply to the neutral functional differential equations
in Chapter 9.
In Sections 4.1 and 4.2, we define and develop properties of w-limit
sets and invariant sets of discrete and continuous processes. Section 4.3
is devoted to giving conditions on a discrete dynamical system to ensure
the existence of maximal compact invariant sets and global attractors. In
Section 4.4, we give sufficient conditions for the existence of fixed points.
Sections 4.5 and 4.6 deal with the same topics for continuous w-periodic
processes. Section 4. 7 deals with convergent systems.
4.1 Processes
In this section, we give the definition of processes, demonstrate how pro-

cesses arise from RFDE, and give a few elementary properties of orbits of
processes.
Definition 1.1. Suppose X is a Banach space, JR+ = [0, oo), u : 1R x X x

JR+ --+X is a given mapping and define U(a, t) :X--+ X for a E 1R, t E JR+
by U(a, t)x = u(a, x, t). A process on X is a mapping u: 1R x X x JR+ --+X
satisfying the following properties
(i) u is continuous.
(ii) U(a, 0) =I, the identity.
(iii) U(a + s, t)U(a, s) = U(a, s + t).
A process u is said to be an w-periodic process if there is an w > 0 such

that U(a + w, t) = U(a, t) for all a E 1R, t E JR+. A process is said to
4.1 Processes 101
be an autonomous process or a (continuous) dynamical system if U(O', t)

is independent of O'j that is, the family of transformations T(t) = U(O, t)
t;::: 0, is a C 0 -semigroup; T(t)x is continuous for (t,x) E lR+ X X,
T(O) =I, T(t + r) = T(t)T(r), t, T E JR+.
If S: X--+ X is a continuous map, the family {Sk, k 2: 0} of iterates of S
is called a discrete dynamical system.
If u is an w-periodic process and S = U(O,w), then the discrete
dynamical system {Sk : k 2: 0} coincides with the family of mappings
{U(O,kw) : k 2: 0}. In fact, S = U(O,w), 8 2 = U(O,w)U(O,w) =
U(w,w)U(O,w) = U(O, 2w), etc. This particular discrete dynamical system
associated with the process u will be referred to as the dynamical system
generated by the period map of the w-periodic process.
In a process, u(O', x, t) can be considered as the state of a system at
time 0' + t if initially the state at time 0' was x.
Let us now discuss how processes arise from RFDE. Suppose f: lR x
C--+ lRn is completely continuous and let x(O', ¢)denote the solution of the
RFDE(f)
(1.1) ±(t) = f(t, xt)
through (0', ¢) and suppose x is uniquely defined for t ;::: 0' - r. From
Theorem 2.2 of Section 2.2, x(O', cp)(t) is continuous in 0', ¢, t for 0' E lR,
¢ E C, and t 2: 0'. If u(O',cp,r) = Xa+r(O',cjJ) for (O',cp,r) E lR X X x lR+,
then u is a process on C. Furthermore, if U is defined as before, then
U(O', r) = T(O' + r, 0') where T(t, 0') is the solution operator for Equation
(1.1), T(t, 0')¢ = Xt(O', ¢).In the following, we refer to the process obtained
from Equation (1.1) in this way as the process generated by the RFDE (!).
If there is an w > 0 such that f(O'+w, ¢) = f(O', ¢)for all (0', ¢) E lRxC,
then the process generated by the RFDE(f) is an w-periodic process. If
f (0', ¢) is independent of 0', then the process generated by the RFDE(f) is
a dynamical system. For the neutral equations of Chapter 2, one obtains a
process in the same manner.
Definition 1.2. Suppose u is a process on X. The trajectory r+ (0', x) through

(0', x) E lR x X is the set in lR x X defined by
r+(O',x) = {(O'+t,U(O',t)x): t E lR+}.
The orbit '/'+ (0', x) through (0', x) is the set in X defined by
'1'+(0', x) = {U(O', t)x: t E lR+}.
If H is a subset of X, then r+(O', H) = UxEH r+(O', x), '1'+(0', H) =
UxEH'I'+(O',x). If {Tk : k 2: 0} is a discrete dynamical system, the or-
bit 'l'+(x) through x EX is defined as 'l'+(x) = {Tkx: k 2: 0} and the orbit
'/'+(H) through a set H ~X is 'I'+(H) = UxEH 'l'+(x).
102 4. Autonomous and periodic processes
If u is an w-periodic process on X, then the trajectory T+ (O" + kw, x)

for any integer k E ffi. is a translate along the reals by a distance kw of
the trajectory T+(O",x). The orbits satisfy 'Y+(O"+kw,x) = 'Y+(O",x) for any
integer k E ffi.. If u is a dynamical system on X, then T+ (O" + s, x) is a
translate by s of 7+ (O", x) for any s E ffi. and ')'+ (O", x) = 'Y+ (0, x) for all
O" E ffi.. In this latter case, we will simply write 'Y+(x) for orbits.
A point c E X is saiq to be an equilibrium or critical point of a process
u if there is an O" E ffi. such that 'Y+(O", c) = { c}; that is, U(O", t)c = c for
tErn.+. If there is a O" E ffi., p > 0, x EX such that U(O", t + p)x = U(O", t)x
for all t E ffi. +, then the trajectory T+ (O", x) is said to be p-periodic.
The proof of the following lemma is obvious.
Lemma 1.1. For a continuous dynamical system, a trajectory is p-periodic

if and only if the corresponding orbit is a closed curve. For an w-periodic
process u, a trajectory through (O", x) is kw-periodic for some positive integer
kif and only ifTkx = x where T = U(O",w).
Lemma 1.2. Let {T(t) : t 2:: 0} be a dynamical system on X. If there are

sequences {xn} ~ X, {wn} ~ (O,oo) satisfying T(wn)Xn = Xn, Wn - t 0
as n -+ oo and some subsequence of {xn} converges to xo, then xo is an
equilibrium point.
Proof. Changing the notation if necessary, we may assume Xn converges to
Xo. Let kn(t) be the integer defined by kn(t)wn :$ t < (kn(t) + l)wn. Then
T(kn(t)wn)Xn = Xn and
IT(t)xo- xol :$ IT(t)xo- T(kn(t)wn)xol
+ IT(kn(t)wn)Xo- T(kn(t)wn)xnl + lxn - xol·
Since kn(t)wn -+ t, Xn -+ Xo as n-+ oo, the right-hand side of this expres-
sion approaches zero as n -+ oo. Therefore, T(t)xo = xo for all t 2:: 0 and
the lemma is proved. 0
Lemmas 1.1 and 1.2 give an effective way of changing the problem of
finding kw-periodic trajectories of w-periodic processes and critical points
of dynamical systems into a problem involving fixed points of mappings.
We will make explicit use of this fact in the following pages.
Definition 1.3. Suppose u is a process on X. A point y E X is said to be in

thew-limit set w(O", x) of an orbit 'Y+(O", x) if there is a sequence tn -+ oo
as n-+ oo such that U(O", tn)x-+ y as n-+ oo, or equivalently, if
(1.2) w(O",X) = nu
t2:0 7" 2:t
U(O",T)x.
A point y E X is said to be in the a-limit set a( O", x) of an orbit ')'- (O", x) =

Ut::s;o U(O", t)x if U(O", t)x is defined fort:$ 0 and there is a sequence
4.1 Processes 103
tn ---. -oo as n---. oo such that U(a, tn)x---. y as n---. oo, or, equivalently, if
(1.3) a(a,x) = nU
t:::;or:::;t
U(a,r)x.
For any subset H ~X, Relations (1.2), (1.3) with x replaced by H can also
be used to define w(a, H), a( a, H), the w- and a-limit sets of a set H. If
the process u is a dynamical system, the dependence of all sets on a will be
deleted. If {Tk : k ~ 0} is a discrete dynamical system of X and H ~ X,
then the w-limit set of H is defined as
(1.4) w(H) = nU
j~On~j
TnH,
and the a-limit set of H is defined as
(1.5) a(H) = nU
j::=;on:::;j
TnH.
Lemma 1.3. Suppose u is a process on X. If 'Y+(a,x) is the precompact,

then w( a, x) exists, is nonempty, compact, connected, and
dist(U(a,t)x,w(a,x))---. 0 as t ---. oo.
If 'Y- (a, x) is precompact, then the same is true of a( a, x) with t ---. -oo.
The same conclusion is valid for any connected subset H ~ X for which
'Y+(a,H) is precompact. If {Tk: k ~ 0} is a discrete dynamical system and
'Y+(H), H ~ X, is precompact, then w(H) is nonempty and compact and
yk H ---. w(H) as k ---. oo.
Proof. The nonemptiness and compactness of w( a, x) follow from Expression
(1.2). For the remainder of this part of the proof, let T(t) = U(a, t) and let
w(x) = w(a,x). If dist(T(t)x,w(x)) -ft 0 as t---. oo, then there is ~tn E > 0
and a sequence tk---. oo ask---. oo such that dist(T(tk)x,w(x)) > E fork=
1, 2, .... Since T(tk)x belongs to a compact set, there must be a convergent
subsequence. This limit belongs to w(x), which is a contradiction. Therefore,
dist(T(t)x,w(x))---. 0 as t---. oo.
If w(x) is not connected, then w(x) would be the union of two disjoint
compact sets that are a distance 8 apart. This obviously contradicts the
fact that T(t)x---. w(x) as t---. oo and so w(x) is connected. The assertions
concerning a( a, x) are verified in the same way.
The reader can supply the details for a set H ~ X and a process u as
well as the assertions concerning a discrete dynamical system to complete
the proof of the lemma. D
Lemma 1.4. For w-periodic processes generated by w-periodic RFDE, every

bounded orbit is precompact. The same conclusion is true for any H ~ X
for which 'Y+ (a, H) is bounded.
Proof. If IU(O", t)¢1 = lxt+u(O", ¢)1 ~ m for t ~ 0, then f is com-

pletely continuous and w-periodic implies there is a constant N such that
li:(O", ¢)(t + O")l ~ N, t ~ 0. Ascoli's theorem implies the result. A similar
argument applies to arbitrary H <:;; X for which --y+ (O", H) is bounded. D
4.2 Invariance
In this section, we discuss invariant sets for autonomous processes, w-

periodic processes, and discrete dynamical systems.
Definition 2.1. If u is a process on X, then an integral of the process on JR

is a continuous function y : JR ---+ X such that for any O" E JR, T+ (O", y( O")) =
{ (O" +t, y( O"+t)) : t ~ 0}. An integral y is an integral through (O", x) E JR x X
if y(O") = x. An integral set on JR is a set M in JR x X such that for any
(O",x) EM, there is an integral yon JR through (O",x) and (s,y(s)) EM
for s E JR. For any O" E JR, let Mu = {x EX: (O",x) EM}.
For a discrete dynamical system, an integral is defined in the same

manner. In fact, if 7l. denotes the integers in JR, an integral y of a discrete
dynamical system {Tk : k ~ 0} is a function y : 7l. ---+ X such that for any
integer k 0 E 71.., Tky(ko) = y(k + ko) for all integers k ~ 0. An integral set
M of a discrete dynamical system is a subset of 7l. x X such that for any
(k 0 ,x) EM, there is an integral yon 7l. through (k 0 ,x) and (k,y(k)) EM
for all k E 71...
Definition 2.2. If {T(t), t ~ 0} is a dynamical system on X, then a set

Q <:;; X is said to be an invariant set if T(t)Q = Q for t ~ 0. This is
equivalent to saying that, through every point x E Q, there is an integral
y through (O,x) such that (y(s) E Q, s E JR. If {Tk: k ~ 0} is a discrete
dynamical system on X, a set Q <:;; X is said to be invariant if TQ = Q.
If u is an w-periodic process in JR x X, an integral set M <:;; JR x X is an
invariant set of u if, for each O" E JR, the set Mu = {x E X : (O", x) E M}
is an invariant set of the discrete dynamical system {Uk(O", w) : k ~ 0} and
Mu+w = Mu for all 0" E JR.
The concept of invariance for discrete dynamical systems in Definition

2.2 is actually equivalent to the following: Q is invariant if and only if it
is possible to extend the definition of Tk on Q to negative integers k and
TkQ <:;; Q for k E ( -oo, oo ). In fact, if T satisfies the last condition stated
then TQ <:;; Q, r- 1 Q <:;; Q, which implies TQ = Q. If TQ = Q, then, in
particular, Q <:;; TQ. Therefore, we can define r- 1 on Q and r-n = (T- 1 )n.
Our next objective is to give sufficient conditions for the w-limit set of
an orbit of a process to be invariant.
4.2 Invariance 105
Lemma 2.1. If u is a dynamical system on X, and an orbit 1+ (x) is pre-

compact, then w(x) is nonempty, compact, connected, and invariant. The
same conclusion is true for any connected set H <:;;; X for which 1+ (H) is
precompact. Finally, T(t)H ~ w(H) as t ~ x.
Pmof. We only need to prove invariance since the other assertions are con-
tained in Lemma 1.3. To show invariance, let U(O, t) = T(t), t ~ 0, suppose
y E w(x), tk ~ oo and T(tk)x ~ y as k ~ oo. For any integer N ~ 0,
there is an integer k0 (N) such that T(tk +t)x is defined for -N :S: t < +oo
if k ~ k 0 (N). Since l+(x) is precompact, one can, therefore, find a subse-
quence {tk,N} of the {tk} and a continuous function G: [-N, N] ~ w(x)
such that T(tk,N + t)x ~ G(t) as k ~ oo uniformly for t E [-N, N].
Using the diagonalization procedure, we can find a subsequence we label
again as {tk} and a continuous function G: (-oo,oo) ~ w(x) such that
T(t + tk)x ~ G(t) ask~ oo uniformly on compact sets of ( -oo, oo).
From the definition of G, we have G(O) = y and G(t + T) = T(t)G(T)
for all t ~ 0, T E ( -oo, oo). Consequently, G : 1R ~ X is an integral
through (0, y) and w(x) is invariant. The lemma will be proved when the
reader supplies the details for sets H <:;;; X. D
Suppose now the autonomous process {T(t) : t ~ 0} on Cis generated

by an RFDE(f) with f completely continuous and 1+(¢) is a bounded
orbit. Lemma 1.4 implies the orbit 1+(¢) is precompact and Lemma 2.1
implies w(¢) is nonempty, compact, connected, and invariant. Invariance in
Definition 2.2 means that a function P : 1R x M ~ M can be defined so that
P(rr,¢) EM for all (rr,¢) E lRxM, T(t)P(rr,¢) = P(t+rr,¢) for all rr E lR,
t ~ 0. This relation can be used to extend the definition of T(t) on M to
negative t. How do we know that the extension of the definition of T(t) on
M to negative t can be made in such a way that T(t)¢ = gt, t E (oo, oo),
for some function g : ( -oo, oo) ~ 1Rn that is a solution of the RFDE (1.1)
on ( -oo, oo )? To see that this is indeed the case, it is necessary to return to
the proof of Lemma 2.1 and use the fact that T(t) is now generated from an
RFDE. Using the same type of argument as in the proof of Lemma 2.1, it
is possible to construct a continuous function g : 1R ~ 1Rn such that g0 = ¢
and for any rr E lR, T(t)gu = gt+u, t ~ 0; that is, the function G(t) = gt,
t E lR, is an integral through (0, ¢ ). Let Xt = T(t)¢. For t in any compact
set [- N, N], there is a K (N) such that
lg(t +h)- g(t)- hf(gt)l :S: lg(t +h)- x(t + h + tk)l

+ lx(t + h + tk)- x(t + tk)- hf(xt+tk)l
+ lx(t + tk)- g(t)l + hlf(xt+tk)- f(gt)l
for all k ~ K(N). Choose k(h) in such a way that k(h) ~ oo ash~ 0 and
lg(t)- x(t + tk(hJ)I = o(lhl) ash~ 0 for all tin [-N, N]. The right-hand
side of this inequality is now o(lhl) as h ~ 0, which proves g(t) = f(gt)·
Therefore, g is a solution of Equation (1.1) on (-oo,oo) and gt E w(¢),
tE ( -oo, oo). If H ~X is an arbitrary set for which '"Y+(H) is precompact,

one can prove the same result to obtain the following corollary.
Corollary 2.1. If an autonomous RFDE(f) generates a dynamical system

and'"'!+(¢) is a bounded orbit then w(¢) is nonempty, compact, connected,
and invariant, where invariance means for any 7/J E w(¢), there is a solution
x of the RFDE(f) on (-oo, oo) with x 0 = 7/J and Xt E w(¢), -oo < t < oo.
If H ~ C is connected and '"Y+ (H) is bounded, the same conclusion is true.
Lemma 2.2. If {Tk : k 2 0} is a discrete dynamical system on X, H ~

X, and '"Y+(H) is precompact, then w(H) is nonempty, compact, invariant,
and Tk H ---+ w(H) as k ---+ oo. Also, H compact, '"Y+(H) precompact, and
w(H) ~ H implies w(H) = nn;:-o Tn H.
Proof The fact that w(H) is nonempty and compact and that Tk H---+ w(H)
is contained in Lemma 1.3. The continuity ofT implies Tw(H) ~ w(H).
If y E w (H), then there is a sequence {x j} ~ H and a sequence {nj}
of integers, nj ---t 00 as j ---t oo, such that rnj Xj ---t y as j ---t 00. By the
precompactness hypothesis, we can select a subsequence that we again label
as nj such that rnrlxj ---t z. Then z E w(H) and Tz = limj-.oo rnj Xj = y
by the continuity ofT. Therefore, w(H) ~ Tw(H), which shows w(H) =
Tw(H). The first part of the lemma is proved.
Suppose now His compact, '"Y+(H) is precompact, and w(H) ~ H. If
J(H) = nn>o Tn H, then clearly J(H) is compact and J(H) ~ w(H). To
prove the converse, suppose y E w(H) and rnj Xj ---t y as j ---t 00 where
nj ---+ oo as j ---+ oo and each Xj E H. Since '"Y+(H) is precompact, for any
integer i, we can find a subsequence of the sequence rnj -i, which we label
as before and a Yi E w(H) ~ H such that Tnj- 1 Xj ---+ Yi as j ---+ oo. Thus
TiY; = y for all integers i. Therefore y E J(H). This proves w(H) ~ J(H)
and the proof of the lemma is complete. D
Lemma 2.3. If u is an w -periodic process on X and an orbit '"Y+ (CJ, x) is

precompact, then there is an invariant integral set M(CY, x) E IR x X such
that if M 8 (CY, x) = {x: (s, x) E M(CY, x)}, then dist(U(CY, t)x, Ma+t(CY, x))---+
0 as t ---+ oo. The same conclusion is true for any set H ~ X such that
'"Y+ (CJ, H) is precompact.
Proof. If T = U (CJ, w), then {Tk : k 2 0} is a discrete dynamical system with
'"Y+(x) = {Tkx: k 2 0} precompact. Lemma 2.2 implies thew-limit set Q
is an invariant set of this discrete system. For a given integer k E ( -oo, oo)
and t E [O,w), define the compact set Ma-kw+t = U(CY,t)Q. Then Ms is
defined for all s E IR and Ms+w = M 8 • Consider the set M = M(CY, x) in
IR x X such that for each integer k E ( -oo, oo), t E [O,w), {(CY- kw +t, x):
x E Ma-kw+t} = M. The set M is an invariant set for the w-periodic
process u. In fact, since U(CY- kw + t, w) = U(CY + t, w) and U(CY, w)Q = Q,
we have
4.3 Discrete systems-Maximal invariant sets 107
U(u- kw + t,w)M~-kw+t = U(u + t,w)U(u, t)Q = U(u,w + t)Q

= U(u+w,t)U(u,w)Q
= U(u, t)Q = M~-kw+t·
To prove dist(U(u, s)x, M~+s) ----+ 0 ass----+ oo, observe first that Lemma
1.3 implies U(u, kw)x----+ Q ask----+ oo. If s = kw+t, then M~+s = U(u, t)Q
and
dist(U(u, kw + t)x, U(u, t)Q) = dist(U(u + kw, t)U(u, kw)x, U(u, t)Q)
= dist(U(u, t)U(u, kw)x, U(u, t)Q)----+ 0
as k ----+ oo for each t > 0. This proves the first part of the lemma. The reader
may easily supply the details for the case of a set H <;;; X to complete the
proof of the lemma. D
One can combine the ideas in the proofs of Lemmas 1.4, 2.3, and Corol-
lary 2.1 to obtain the following result.
Corollary 2.2. If an w-periodic RFDE(f) generates an w-periodic process

u and 'Y+ (u, ¢) is a bounded orbit, then there is an invariant integral set
M(u,¢) <;;; IR x C of the RFDE(f), M 8 (u,¢) = {~ E C: (s,~) E M(u,¢)}
is compact, and dist(xt(u, ¢), M~+t(u, ¢))----+ 0 as t----+ oo. That is, there is
a set of functions Y (u, ¢) = {y : IR ----+ IRn} such that each y is a solution
of the RFDE(f) on ( -oo, oo), the set M(u, ¢) = {(s, Ys) : s E IR, y E
Y( u, ¢)} is invariant for the w-periodic process generated by the RFDE(f)
and dist(xt(u, ¢), M~+t(u, ¢))----+ 0 as t----+ oo.
4.3 Discrete systems-Maximal invariant sets and

attractors
In this section, we discuss discrete dynamical systems {Tk : k ~ 1} on

X and give conditions that ensure there is a maximal compact invariant
set that is a stable global attractor. For any set H <;;; X and E > 0, let
B(H,E) = {x EX: dist(H,x) < E}.
Definition 3.1. For a given continuous function T : X ----+ X, we say a set

K <;;;X attracts a setH<;;; X if, for any E > 0, there is an integer N(H, E)
such that Tn(H) <;;; B(K, E) for n ~ N(H, E). We say K attracts points of
X if K attracts each point of X. We say K attracts compact sets of X if K
attracts each compact set of X. We say K attracts neighborhoods of points
(compact sets) of X if, for each point x E X (each compact set H <;;; X),
there is a neighborhood Ox of x (H0 of H) such that K attracts Ox(H0 ).
We say K attracts bounded sets of X if K attracts each bounded set of
X. We say that K is a global attractor if K is invariant and attracts each

bounded set of X.
Definition 3.2. A set M ~ X is said to be stable (with respect to the

dynamical system {Tk : k 2:: 0}) if, for any E > 0, there is a 8 > 0 such
that x E B(M, 8) implies Tnx E B(M, E) for n 2:: 0. The set M is said to
be asymptotically stable if it is stable and there is an Eo > 0 such that M
attracts points of B(M, Eo). The set M is said to be uniformly asymptotically
stable if it is asymptotically stable and, for any TJ > 0, there is an integer
no(TJ, Eo) such that Tnx E B(M, TJ) for n 2:: no(TJ, Eo) and x E B(M, Eo).
Definition 3.3. If B is a bounded set in a Banach space X, the K uratowski

measure a( B) of noncompactness of B is defined as
a(B) = inf{d: B has a finite cover of diameter< d}.
Some properties of a(B) that will be needed in the following are now
stated without proof.
(i) a(B) = 0 if and only if B is compact.
(ii) a(A U B) = max(a(A), a(B)).
(iii) a( coB)= a(B), where coB is the closed convex hull of B.
Lemma 3.1. If T : X --+ X and there is a compact set K ~ X that attracts

compact sets of X, then for any compact set H ~ X, 'Y+ (H) is precompact.
Proof. Let a(B) be the Kuratowski measure of noncompactness of a
bounded set B given in Definition 3.3. For any compact setH~ X, the set
A ~r U TnH = 'Y+(H)
n:::O:O
is bounded since K attracts compact sets. Since T1 H is compact for any

j, we have a(A) = a(Un:::O:j TnH) for any j. But, for any E > 0, we have
Un:::O:j rn H ~ B(K, E) for j 2:: N(H, E). Consequently, a( A) :::; 2E for every
E > 0. This implies a(A) = 0 and so A is precompact. D
Our next objective is to define a maximal compact invariant set for a

discrete dynamical system satisfying the conditions of Lemma 3.1.
Suppose T is continuous and there is a compact set K ~ X that
attracts compact sets of X. Lemma 3.1 implies 'Y+(K) is precompact and
Lemma 2.2 implies w(K) exists. For any E > 0, there is an integer no(K, E)
such that Tn K ~ B(K, E) for n 2:: n 0 (K, E). Thus w(K) ~ B(K, E) for every
E > 0. Therefore, w(K) ~ K. It follows from Lemma 2.2 that
(3.1) J~f nrnK=w(K)

n:::O:O
is nonempty, compact, and invariant. We claim that J is independent of the

set K, which attracts compact sets of X. In fact, if we designate J by J(K)
and K 1 is any other compact set that attracts compact sets of X, then
there is an integer n 0 (K, K 1, E) such that yn J(K) ~ B(K1, E), yn J(KI) ~
B(K,E), for n 2': n 0 (K,K 1 ,E). Since J(K) and J(K1 ) are invariant, this
implies J(K) ~ K1, J(K1) ~ K and J(K) ~ yn K1, J(K1) ~ yn K for all
n 2': 0. Therefore, J(K) = J(K1).
Theorem 3.1. Suppose T: X---> X and there is a compact set K ~X that

attracts compact sets of X and let J be defined by Expression (3.1). Then
the following conclusions hold:
(i) J = w(K) is independent of K, is a nonempty compact invariant set,
and is maximal with respect to this property.
(ii) J is connected.
(iii) J is stable.
(iv) For any compact setH C X, J attracts H.
Proof. The remarks preceding the statement of the theorem imply that
J = w(K) is a nonempty compact invariant set that is independent of K.
To prove that J is maximal, suppose that H is any compact invariant
set. Since K attracts H, for any E > 0, there is an integer n 1 ( H, E) such
that yn H C B(K, E) for n 2': n 1(H, E). From the invariance of H it follows
that H = yn H for all n. So we can choose a sequence Ej ---> 0 as j ---> oo to
obtain H C K. Likewise, H C yn K for all n 2': 0 and soH C J.
Now suppose that x EX is arbitrary. Then l+(x) is precompact from
Lemma 3.1. Thus, w(x) is nonempty, compact, and invariant and w(x) c J.
Since Tkx---> w(x) ask---> oo, it follows that J attracts points of X.
We now prove that J is connected. If co J is the closed convex hull
of J, then co J is compact and connected and J attracts co J. If J is not
connected, then there are open sets U, V with U n J "/=- 0, V n J "/=- 0. By
the continuity of T, the set yn (co J) is connected for each n 2': 0. Since
J c rn(co J) for all n 2': 0, we have unrn(co J) -1=- 0, vnrn(co J) -1=- 0 for
all n 2': 0. Since yn (co J) is connected, there is an Xn E yn (co J) \ (U U V).
Since J attracts { Xn, n 2': 0 }, this set must be compact and we may assume
that Xn---> x E J. Clearly, x t/:. U U V, which is a contradiction.
To prove that J is stable, suppose that this is not the case. Then for
some E > 0, which may be chosen as small as desired, there are sequences
of integers nj and Yj E X such that nj ---> oo, Yj ---> J as j ---> oo, Tnyj E
B(J, E), 0::; n::; nj, and ynj+lyj is not in B(J, E). Furthermore, J compact
implies that we may assume without loss of generality that Yj ---> y E J as
j ---> oo. The set H = { y, yj; j 2': 1} is compact. Lemma 3.1 implies
that I+(H) is precompact. Therefore, Lemma 2.2 implies that w(H) is
nonempty, compact, and invariant. The first part of the theorem being
proved implies that w(H) C J. Also, I+(H) precompact implies that with
no loss in generality we may assume that TnjYi __, z as j __, oo. Then
z E w(H) C J. But the choice of nj, Yi implies that Tz ¢:. B(J, E) and,
therefore, Tz ¢:. J. Since J is invariant, this is a contradiction and the proof
of assertion (iii) is complete.
To prove (iv), suppose that E > 0 is given and 8 is the number associ-
ated with E in the definition of stability. For any x E X, there is an integer
N(x) such that Tnx c B(J, 8) for n ~ N(x). Since Tis continuous, there
is a neighborhood Ox of x such that TN(x)Ox C B(J, 8). As a consequence,
TiOx C B(J, E) for j ~ N(x). If His an arbitrary compact set in X, then
a finite covering argument proves (iv). The proof of the theorem is com-
plete. D
From Theorem 3.1, we see that the existence of a maximal compact

invariant set that is stable and attracts compact sets of X is ensured under
very weak conditions on the map T. From the discussion in Section 1 deal-
ing with w-periodic RFDE, a discrete dynamical system is obtained from
the solution map T(t, cr) by defining T = T(w, 0). Furthermore, Lemma 1.1
implies that kw-periodic solutions of the RFDE are obtained from the fixed
points of Tk. As a consequence, it is important to know under what con-
ditions such fixed points exist and, even more importantly, to know when
k = 1; that is, when does T itself have a fixed point?
Intuitively, one might suspect that the conditions of Theorem 3.1 would
imply that T has a fixed point since J is stable and attracts co~pact sets
of X. This problem, however, as well as the problem of whether even some
iterate of T has a fixed point, is unsolved. The ~ext section is devoted to
the specification of further conditions on T, which will ensure that T or
some iterate of T has a fixed point.
However, before developing the theory of fixed points, it is convenient
from the point of view of later applications to specify conditions on T,
which will imply a type of stability that is stronger than the conclusion in
Theorem 3.1(iv). More specifically, we want the same conclusion for any
bounded set H c X. Some additional definitions are required.
Definition 3.4. A continuous map T : X __, X is said to be point (compact)

(local) (locally compact) (bounded) dissipative if there is a bounded set B C
X such that B attracts points (compact sets) (neighborhoods of points)
(neighborhoods of compact sets) (bounded sets) of X.
Obviously, bounded dissipative implies locally compact dissipative im-

plies local dissipative implies compact dissipative implies point dissipative.
Definition 3.5. A continuous map T : X __, X is said to be asymptotically

smooth if, for any bounded set B C X for which T B c B, there is a compact
set B* C X such that B* attracts B.
It is not difficult to show that this definition is equivalent to the fol-

lowing one. A continuous map T : X --+ X is asymptotically smooth if,
for any bounded set B C X, there is a compact set B* C X such that for
any E > 0, there is an integer no(B, E) with the property that rnx E B for
n ~ 0 implies that rnx E B(B*, E) for n ~ n 0 (B, E).
The concept of asymptotically smooth is a result of a detailed investi-
gation of the relationship between asymptotic stability and uniform asymp-
totic stability of compact invariant sets in a Banach space X. In fact, it
is an interesting exercise to prove the following fact: If T is asymptotically
smooth and J is a compact invariant set ofT, then J is asymptotically
stable if and only if it is uniformly asymptotically stable. As we will see, the
importance of the concept is related to the fact that uniform asymptotically
stable sets enjoy some type of persistence under perturbations of the map
T, whereas this is not the case when the set is only asymptotically stable.
Lemma 3.2. If T : X --+ X is local dissipative and asymptotically smooth,

then there is a compact set K C X that attracts neighborhoods of compact
sets of X.
Proof. If B 1 is the bounded set in the definition of local dissipative and
8 > 0, let B = B(B1, 8) and let B* be the corresponding; compact set in
the definition of asymptotically smooth. If H is an arbitrary compact set
and xis an arbitrary point in H, then there is a neighborhood Ox of x and
an integer N(x) such that rnox c B for n ~ N(x). Selecting from the
covering {Ox : x E H} of H a finite subcovering, we see that there is an
integer N(H) and an open neighborhood H 1 of H such that rn H 1 c B
for n ~ N(H). Since His compact and Tis continuous, one can choose a
neighborhood H 0 of H, H 0 c H 1 , such that rnH0 is bounded for 0:::; n:::;
N(H). Therefore, { rn H 0 : n ~ 0} is bounded since H 0 c H 1 . Since Tis
asymptotically smooth, for any E > 0, this implies that rn Ho c B(B*' E)
for n ~ N(H) + n 0 (B, E). If we let K = B*, the proof of the lemma is
complete. D
Theorem 3.2. If T : X --+ X is local dissipative and asymptotically smooth,

then the set J in Expression (3.1) defined by K in Lemma 3.2 satisfies all
of the conclusions of Theorem 3.1 and
(v) For any compact setH C X, there is a neighborhood H 1 of H such
that ')'+ ( H 1) is bounded and J attracts H 1; in particular, J is uniformly
asymptotically stable.
(vi) If B C X is an arbitrary bounded set and 'Y+(B) is bounded, then J
attracts B; in particular, if 'Y+(B) is bounded for every bounded set
B C X, then J is a compact global attractor.
Proof. Lemma 3.2 implies that T satisfies the hypotheses of Theorem 3.1
and, therefore, the first part of the theorem is proved. To prove Property
(iv), suppose that B c X and /'+(B) C U for some bounded set U c X.

Let U* be the corresponding compact set in the definition of asymptotically
smooth. The definition of asymptotically smooth implies that for any E > 0,
there is an integer no(U, E) such that rn B c B(U*' E) for n 2 no(U, E).
Since J attracts a neighborhood of the compact set U*, we may assume that
E > 0 is fixed and so small that J attracts B(U*, E). For any 7) > 0, there is
an integer mo(U*, E) such that TB(U*, E) c B(J, rJ) form 2 m 0 (U*, TJ). If
we let N = n 0 (U, E)+ m 0 (U*, rJ), then
rno(U, E)+m B c TmB(U*' E) c B(J, TJ) form 2 mo(U*' TJ)'

or Tn B c B(J, TJ) for n 2 N. This completes the proof of the theorem. D
Corollary 3.1. If T : X ---+ X is asymptotically smooth, then the following

are equivalent:
(i) There is a compact set K that attracts compact sets of X.
(ii) There is a compact set K that attracts neighborhoods of compact sets
of X.
It is possible to obtain a result on the existence of a compact global

attractor, which is easier to use in the applications than Theorem 3.2. We
need the following lemma.
Lemma 3.3. Let T be asymptotically smooth and point dissipative. If the

orbit of any compact set is bounded, then T is locally compact dissipative.
If the orbit of any bounded set is bounded, then T is bounded dissipative.
Proof. Let B be a nonempty closed bounded set that attracts points of X
and let U = { x E B: l'+(x) c B }. Then U attracts points of X and l'+(u)
is bounded. Since T!'+(U) c I'+(U) and Tis asymptotically smooth, there
is a compact set K C Cl U such that K attracts U and attracts points of
X. The set K also attracts itself and so I'+(K) is precompact. If J = w(K),
then J is a compact invariant set that attracts points of X.
Now suppose that the orbit of any compact set is bounded. We first
show that there is a neighborhood V of J such that l'+(v) is bounded. If
this is not the case, then there is a sequence Xj E X and a sequence of
integers kj ---+ oo and a y E J such that limj_, 00 Xj = y and ITki Xj I ---+ oo as
j---+ oo. But then Cl { Xj, j 2 1} is a compact set with I'+(Cl { Xj, j 2 1})
unbounded. This is a contradiction. Thus, there is a neighborhood V of
J such that /'+ (V) is bounded. Since J attracts points of X and T is
continuous, for any x EX, there is a neighborhood Ox of x and an integer
n 0 such that TnOx c I'+(V) for n 2 no; that is, I'+(V) attracts Ox. For
any compact set H in X, one can find a finite covering of H and thus an
open set H 1 containing H such that I'+(V) attracts H1. Thus, Tis locally
compact dissipative.
Now suppose that orbits of bounded sets are bounded. Then orbits of
compact sets are bounded and T is locally compact dissipative from the
first part of the lemma. We now apply Theorem 3.2 to complete the proof.
0
As a consequence of Theorem 3.2 and Lemma 3.3, we have the following

result.
Theorem 3.3. If T is asymptotically smooth, point dissipative, and orbits

of bounded sets are bounded, then there exists a connected compact global
attractor.
We now give some sufficient conditions for the hypotheses of this the-
orem to hold. To do this, some additional definitions are required.
Definition 3.6. Suppose that a is the Kuratowski measure of noncompact-

ness of a set in X and that T : X -+ X is a continuous map. The map T is
conditionally condensing if a(TB) < a(B) for any bounded set B C X for
which a(B) > 0 and T B is bounded. The map Tis said to be a conditionally
a-contracting if there is a constant k, 0::::; k < 1, such that a(T B) ::::; ka(B)
for any bounded set B c X for which T B is bounded. If T takes bounded
sets into bounded sets, then a conditionally a-contracting (resp. condens-
ing) map is an a-contracting (resp. condensing) map. The map Tis said to
be conditionally completely continuous if T B is precompact for any bounded
set B in X for which T B is bounded. The discrete dynamical system gen-
erated by Tis said to be conditionally completely continuous fork ~ n 0 , if,
for any bounded set B in X, there is a compact set B* in X such that for
any integer N ~ n 0 and any x EX with Tn X E B, 0 ::::; n::::; N, it follows
that Tnx E B* for no::::; n::::; N.
Note that, if Tis conditionally completely continuous, then Tis a con-

ditional a-contraction, and this implies that T is conditionally condensing.
It is also obvious that, if T takes bounded sets into bounded sets and
rna is completely continuous for some n 0 , then the discrete dynamical
system generated by Tis conditionally completely continuous fork~ n 0 •
One of the most important mappings T, which is an a-contraction
and is not necessarily completely continuous, is T = S + U, where U is
a completely continuous mapping and the mapping S is a bounded linear
map on X with IIBII < 1. If Sis only a bounded linear map with modulus of
the spectrum < 1, then the space X can be renormed so that the property
is satisfied (see Lemma 4.4).
For RFDE, it was shown in Theorem 6.1 of Section 3.6 that the solution
map T(t, a-)= S(t-o-)+U(t, a-), where U(t, a-) is conditionally completely
continuous for all t ~ a- and the bounded linear map S (t) = 0 for t ~ r.
Therefore, T(t, a-) is conditionally completely continuous for t ~ a-+ r.
Also, for any given w > 0, we can renorm the space X so that the map
r = r(a + w, a) is a conditional a-contraction.
Lemma 3.4. If r : X ----+ X is point dissipative and the dynamical system

generated by r is conditionally completely continuous fork ~ n 0 , then there
is a compact set K in X such that for any compact set H in X, there is
an open neighborhood H 0 of H and an integer N(H) such that !+(H0 ) is
bounded and rn H 0 c K for n ~ N(H). In particular, there is a compact
set K that attracts compact sets of X.
Proof. Since r is point dissipative, for any E > 0, x E X, there is an
integer N(x, E) such that rnx E B(B, E) for n ~ N(x, E), where B is the
bounded set in Definition 3.4. We will fix E and therefore suppress the
explicit dependence of constants and sets on E. Let B* be the compact set
of Definition 3.6 corresponding to B(B, E). By the continuity of r, there is
an open neighborhood Ox of X such that rnox c B(B, E) for N(x) ::; n::;
N(x)+n 0 . Therefore, rn(x)Ox c B*, where n(x) = N(x)+n 0 . Suppose that
His an arbitrary compact set of X. The neighborhoods Ox of x E H form
a covering of H. Selecting from this cover a finite subcover {Ox, (H)} and
letting N(H) =maxi n(xi), we have rn(x;)Ox, (H) C B* and 0 ::; n(xi) ::;
N(H). Let Ho = Ui Ox, (H) and let K be the compact set consisting of
the union of B*, rB*, ... , rN(B*)B*. We claim that rnB* C K for n ~
N(B*). In fact, if x E B* and n ~ N(B*), then there is a least integer
j, 0 ::; j ::; n, such that rn-j X E B* and rn-kx tJ. B* for 0 ::; k < j. But
rn-Jx E Ox,(B*) for some i, rn(x;)rn-Jx c B*, 0 ::; n(xi) < N(B*).
This implies that 0 ::; j ::; N(B*) and so rnx E K for n ~ N(B*).
Let N(H)' = N(H) + N(B*) and suppose that x E H 0 , n ~ N(H)'.
Then x E Ox, (H) for some i, 0 ::; n(xi) ::; N(H) and rn(x;)Ox, (H) =
rn-n(x,)rn(x,)Ox,(H) c K for n ~ N(H)'. The fact that U1:::: 0 r1Ho is
bounded is obvious, and the lemma is proved. D
If r is a mapping satisfying the conditions of Lemma 3.4, then r

satisfies the hypotheses of Theorem 3.1. Therefore, there is a compact set J
that is stable and attracts neighborhoods of compact sets. However, more
is implied from Lemma 3.4 if r takes bounded sets into bounded sets. In
fact, we can prove the following result.
Corollary 3.2. If r : X ___, X takes bounded sets into bounded sets, is point
dissipative, and rna is completely continuous for some n 0 , then for any
bounded set B C X, there is a bounded set U c X, a compact set K C X,
and an integer N(B) such that rJ B C U, j ~ 0 and rJ B C K, j ~ N(B).
In addition, K is a connected compact global attractor.
Proof. If B is bounded, then the hypotheses on r imply that Cl rna B = H
is compact. Therefore, Lemma 3.4 implies that rn H c K for n ~ N(H),
where K is a compact set. Thus, rn B C K for n ~ N(H) + n 0 . Since r
4.4 Fixed points of discrete dissipative processes 115
takes bounded sets into bounded sets, the existence of the bounded set U
is clear and the corollary is proved. D
Lemma 3.5. A conditionally condensing map is asymptotically smooth.

Proof. Let B be a closed bounded set in X such that B C X, TB C B. We
want to show first that every sequence { Tki Xj } with kj ----+ oo as j ----+ oo has
compact closure. Let C = {{ Tki Xj} : { Xj} C B, { ki} integers, 0 ~ ki ----+
oo} and 'TJ = sup{ a( h) : h E C }. Since h E C C B, we have 'T} ~ a( B).
We claim that there is an h* E C such that a(h*) = ry. Let {he} C C be
such that a(he)----+ 'TJ and define he= {Tkixi: Tkixi E he, ki > t'}. Then
a(he) =a( he). If h* = Ue he ordered in any way, then h* E C and a(h*) ?:
a(he). Therefore, a(h*) = ry. If h* = {Tk;-lxj: Tk;xj E h* }, then h* E C
and a(h*) ~ TJ. Since T(h*) = h* and a(h*) = ry, we have a(h*) ?: a(h*).
Since Tis conditionally condensing, it follows that a(h*) = 0. Thus, h* and
h* are precompact and 'TJ = 0. Thus every sequence inC is precompact.
This argument shows that w(B) exists and is given by w(B) =
nn>o Tn Band thus w(B) is invariant. Since Tis condensing, the set w(B)
is invariant. Finally, w(B) attracts B. In fact, if this is not the case, the•,
there is a sequence {Tkixi} and an E > 0 such that Tkixi ¢:_ Ne(w(B))
for all j ?: 1. However, the set { Tk; xi } is precompact by the previous
argument and contains a subsequence that converges to a point z outside
Ne(w(B)). This obviously is a contradiction and completes the proof of the
lemma. D
4.4 Fixed points of discrete dissipative processes
In this section, we give sufficient conditions on a continuous map T : X ----+

X, which will ensure that Tor some iterate ofT has a fixed point. These
fixed points are sometimes called periodic points ofT.
To motivate the approach to be taken, let us examine some of the
implications of Theorem 3.1. If J is the maximal compact invariant set in
Theorem 3.1, and K =co J, the closed convex hull of J, then K is compact
and there is a convex neighborhood B of K such that 'Y+(B) is bounded and
J (and, therefore, K) attracts B, i.e., there exist convex subsets K ~ B ~ S
of X with K compact, S closed and bounded, and B open in S such that
'Y+(B) ~Sand K attracts B. For nested sets of this type and even weaker
properties of attraction, one can prove the following interesting result.
Lemma 4.1. Suppose K B ~ S are convex subsets of X with K compact,

~
S closed and bounded, and B open in S. If T : S ----+ X is continuous,
'Y+(B) ~ S, and K attracts points of B, then there is a closed, bounded,
convex subset A of S such that
A= co [U Ti(B n A)], AnK-=f.0.

j::C:l
Proof. Let F be the set of convex, closed, bounded subsets L of S such that
Ti(BnL) ~ L for j;::: 1 and LnK -=f. 0. The family F is not empty because
S E F. If L E F, let
L1 =co [Uri(BnL)].
j::C:l
Since K attracts points of B, there is a sequence Xn E L 1 such that Xn ____, K

as n ____, oo. Therefore, ( Cl Ll) n K -=f. 0. But L 1 is closed and so L 1 n K -=f. 0.
Also, L 1 is convex and L 1 ~ S. Since L E F, we have L :2 L 1, and thus
the definition of L 1 implies L 1 :2 Ti ( B n L) :2 Ti ( B n Ll) for j 2:: 1. Thus,
L 1 E F and a minimal element A of F will satisfy the condition of the
lemma.
To prove such a minimal element exists, let (La)aEJ be a totally or-
dered family of sets in F. The set L = naEI La is closed, convex, and
contained inS. Also, Ti(B n L) ~ Ti(B n La) ~La for any a E I and
j ;::: 1. Thus, Ti (B n L) ~ L for j ;::: 1. If J is any finite subset of I, then by
the same reasoning as for the set L 1, we have K n (naEJLa) =f. 0. From
compactness of K, it follows that K n (naEI La) =f. 0. Thus, L E F and
Zorn's lemma yields the conclusion of the lemma. D
To proceed further, we need the following asymptotic fixed-point the-

orem of Horn [1], which is stated without proof.
Theorem 4.1. Let S 0 ~ S 1 ~ S 2 be convex subsets of a Banach space X

with S 0 , S 2 compact and S 1 open in S 2 . LetT : S 2 ____, X be a continuous
mapping such that for some integer m > 0, Ti S 1 ~ S 2 for 0 :::; j :::; m - 1
and TJ S1 ~ S 0 form :::; j :::; 2m- 1. Then T has a fixed point.
Our basic lemma is now derived from this result and Lemma 4.1.
Lemma 4.2. Suppose K ~ B ~ S are convex subsets of X with K compact,

S closed and bounded, and B open in S. If T : S ____, X is continuous,
'Y+(B) ~ S, K attracts compact sets of B, and the set A of Lemma 4.1, is
compact, then T has a fixed point.
Proof. Without loss in generality, we may take B = l3(K, E) n S for some
E > 0 since K is compact and convex. Let So = An Cl B(K, E/2), sl =
An B(K, E), and s2 = A. Then So ~ sl ~ s2 are convex subsets of X
with S 0 ,S2 compact and S 1 open in S 2 . Also, TJS1 = TJ(AnB(K,E)) ~A,
j 2:: 1, 'Y+(S1) ~A. Therefore. TJ S 1 ~ S 2 for j 2:: 1. Also, the fact that K
attracts compact sets of B implies there is an integer n 1 = n 1(K, E) such
that TJ sl ~ B(K, E/2) for j;::: nl. If nl ;::: 1, then TJ sl ~A, for j ;::: nl and
4.4 Fixed points of discrete dissipative processes 117
so Ti 8 1 ~ 8 0 for j ~ n 1 . We may now apply Theorem 4.1 to the mapping

T to complete the proof of the lemma. D
Lemma 4.2 gives a good motivation for determining conditions on the

mapping T, which will ensure that the set A is compact. The following
lemma deals with this problem.
Lemma 4.3. If T is conditionally condensing, then the set A in Lemma 4.1

is compact.
Proof. If A = UJ_.> 1 Ti (B n A) then A = T(B n A) U T(A) and a(A) =
a(A) =max [a(T(B n A)), a(T(A))]. Since a(T(A)) < a(A) if a(A) > 0,
it follows that a(A) = a(A) = a(T(B n A)). Thus, if a(B n A) > 0,
then a(A) = a(A) < a(B n A) ::; a(A) and this is a contradiction. Thus,
a(B n A) = 0. This implies a(T(B n A)) = 0 and thus a(A) = 0. But this
implies A is compact and the lemma is proved. D
Lemmas 4.2 and 4.3 give sufficient conditions for the existence of fixed
points of the map T. We now state a result using the theory of Section 4.3
and these lemmas.
Theorem 4.2. If T : X ----t X, there is a compact set that attracts compact

sets of X and T is conditionally condensing, then T has a fixed point.
Proof. This is a direct application of Theorem 3.1 and Lemmas 4.2 and
4.3. D
Our next objective is to give other conditions on the map T that imply
the hypotheses of Theorem 4.2 and are easier to verify in applications. These
conditions involve the concept of dissipativeness introduced in Definition
3.4.
Corollary 4.1. If T : X ----t X is continuous and point dissipative and {Tk :

k ~ 0} is conditionally completely continuous fork ~ n 0 , then Tno has a
fixed point.
Proof. Lemma 3.3 implies the existence of a compact set that attracts com-
pact sets. Since Tno is conditionally condensing, Theorem 4.2 may be ap-
plied to Tno to yield the assertion. D
Theorem 4.3. If T : X ----t X is conditionally condensing and point dissipa-

tive and {Tk : k ~ 0} is conditionally completely continuous for k ~ n 0 ,
then T has a connected compact global attractor and T has a fixed point.
Proof. Corollary 3.2 implies the existence of the global attractor. Therefore,
Theorem 4.2 implies the existence of a fixed point ofT. D
Theorem 4.4. IfT is conditionally condensing and compact dissipative, then

T has a fixed point.
Pmof. This is a consequence of Lemma 3.5 and Theorems 3.2 and 4.2. 0
Lemma 4.4. If S : X --+ X is a bounded linear operator with spectrum

contained in the open unit ball, then there is an equivalent norm, I · 11 , in
X such that ISI1 < 1.
Proof. Define lxl 1= 'Ej:;:o ISJxl. The assumption on the spectrum implies
there is an r, 0::::; r < 1, such that ISnl < rn if n is sufficiently large. Thus,
there is a constant K 2: 1 such that lxl : : ; lxl 1 : : ; Klxl for all x EX. Also,
for x =1- 0,
1Sxl1 = 1- [1 + ISxl + IS2xl + .. ·r1 = 1- [lx11J -1 < 1- 2_,

lxl1 lxl lxl lxl - K
This proves the lemma. 0
Theorem 4.5. IfT: X--+ X is continuous, T = S+U where Sis linear with
spectrum contained in the open unit ball and U is conditionally completely
continuous, then
(i) T compact dissipative implies T has a fixed point.
(ii) T point dissipative implies Y has a fixed point if sno is completely
continuous for some integer no.
Proof. Changing the norm in the space does not affect the existence of fixed
points. Lemma 4.4 implies T is conditionally condensing in an appropriate
norm. Theorem 4.4 implies Part (a). In Part (b), rno = sno +V and one can
show that V is conditionally completely continuous. Theorem 4.3 implies
the conclusion of Part (b) and the theorem is proved. 0
Theorem 3.1 can also serve as a general motivation for the so-called
asymptotic fixed-point theorems, one of which is stated in Theorem 4.1. To
see this, we first observe the following easy consequence of Theorem 3.1.
Lemma 4.5. If T : X --+ X is continuous and there is a compact set that at-
tracts compact sets of X, then there exists an integer m and convex bounded
sets S 0 ~ S1 ~ S2 with So, S2 closed and S1 open, such that !'+(SI) ~ S 2
and l'+(rmsl) ~Sa.
Pmof. Let J be the maximal compact invariant set of Theorem 3.1 and
let K be the closed convex hull of J. Then K is compact and convex.
Theorem 3.1(iii) guarantees the existence of a neighborhood K 1 of K such
that 'Y+(KI) is bounded and J attracts K1. Since J ~ K, it follows that
K attracts K1 . Let S 2 be a closed bounded convex set containing 'Y+(KI).
Choose E > 0 such that Cl B(K, E) ~ K 1 and let S 0 = Cl B(K, E/2),
4.5 Continuous systems-Maximal invariant sets 119
sl = B(K, E). Since K attracts Kl, there is an m = no(K, E) such that

Tn K 1 ~ S 0 for n 2: m. Therefore, TnS 1 ~ S 0 for n 2: m. Also, the
definition of S1 and S2 implies 1'+(SI) ~ S2 and the lemma is proved. D
Lemma 4.5leads in a natural way to the following fixed-point problem:

Suppose there exists an integer m, bounded convex subsets So ~ S 1 ~ S 2
of a Banach space X such that So and S 2 are closed and S 1 is open, and a
continuous map T: S2 ---> X such that 1'+(SI) ~ S2 and 1'+(TmSI) ~ S 0 .
What additional conditions on T imply the existence of a fixed point ofT?
The famous asymptotic fixed-point theorem of Browder states that T
completely continuous is sufficient to imply the existence of a fixed point
of T if there are sets So ~ S 1 ~ S2 as asserted in the. preceding problem.
Using the techniques we have been employing in this chapter, it is possible
to prove the following interesting extension of this result.
Theorem 4.6. Suppose So ~ S1 ~ S2 are convex bounded subsets of a Ba-

nach space X, S 0 and S2 are closed, and S1 is open in S2, and suppose
T : S2 ---> X is condensing in the following sense: if D and T( D) are
contained in S2 and a(D) > 0, then a(T(D)) < a(D). If 1'+(SI) ~ S2
and, for any compact set H ~ S 1 , there is a number N(H) such that
1'+ (TN(H) H) ~ S 0 , then T has a fixed point.
Proof. Repeating the same argument as in Lemma 3.4, one can show there
is a compact set J!~ S 0 that attracts compact sets of S1 . Since S 0 is
convex, K = co K ~ S 0 . Let B ~ S 1 be an open convex neighborhood
of K. The conditions of Lemma 4.1 are therefore satisfied and Lemma 4.3
implies that A is compact. Thus, the conditions of Lemma 4.2 are satisfied
~T~a~d~~- D
4.5 Continuous systems-Maximal invariant sets and

at tractors
In this section, we consider the same topics as in Section 4.3 for contin-
uous dynamical systems or equivalently C0 -semigroups of transformations
T(t), t 2: 0, on a Banach space X; that is, T(t)x, t 2: 0, x EX, is continuous
in t, x and satisfies the properties T(O) =I, T(t + s) = T(t)T(s), t, s 2:0.
For any set B c X, the positive orbit 1'+(B) is defined as 1'+(B)
Utzo T(t)B and thew-limit set w(B) is defined as
w(B) = nU
szOtzs
T(t)B.
A negative orbit through a point x is a function ¢ : ( -oo, 0] ---> X such that

¢(0) = x and, for any s ::=; 0, we have T(t)¢(s) = ¢(t + s) for 0 ::=; t ::=; -s.
In an obvious way, we can define the a-limit set of a negative orbit. In the
following, we do not need the concept of a-limit set of a set.
We say that a set K C X attracts a set H c X if, for any E > 0, there
is a t 0 (H, E) such that T(t)H C B(K, E) fort~ 0. We say that T(t), t ~ 0,
is point dissipative if there is a bounded set K C X such that K attracts
points of X. We say that T(t), t ~ 0, is bounded dissipative if there is a
bounded set K C X such that K attracts bounded sets of X. We say that
K is a global attractor if K is invariant and attracts bounded sets of X.
A set M c X is said to be stable if, for any E > 0, there is a 8 > 0 such
that x E B(M, 8) implies that T(t)x E B(M, E) fort ~ 0. A set M is said
to be asymptotically stable if it is stable and there is an Eo > 0 such that M
attracts the set B(M, Eo). A set M is said to be uniformly asymptotically
stable if it is asymptotically stable and, for any TJ > 0, there is a t 0 (ry, Eo) > 0
such that T(t)x E B(M, ry) fort~ to(TJ, Eo) and x E B(M, Eo).
The semigroup T(t), t ~ 0, is said to be asymptotically smooth if, for
any bounded set B C X such that T(t)B c X, there is a compact set K
such that K attracts B. The semigroup T(t), t ~ 0, is said to be an a-
contraction if, for each bounded set B C X, the set {T(s)B, 0:::; s:::; t} is
bounded and there is a continuous function k : [0, oo) ---> [0, oo) such that
k(t) ---> 0 as t---> oo, and, for each t > 0 and for each bounded set B c X,
we have a(T(t)B) :::; k(t)a(B), where a denotes the Kuratowski measure
of noncompactness. As in Section 3, we can show that a-contracting semi-
groups are asymptotically smooth.
All of the results of Section 3 are valid for continuous semigroups with
the proofs being essentially the same. However, we are going to state one
of the most important ones, which will be useful for our later discussion.
An equilibrium point of T(t), t ~ 0, is a point x EX such that T(t)x = x
for t ~ 0.
Theorem 5.1. If T(t), t ~ 0, is asymptotically smooth, point dissipative and

positive orbits of bounded sets are bounded, then there exists a connected
compact global attractor and there is an equilibrium point ofT(t).
Theorem 5.2. If there is a t 1 ~ 0 such that T(t) is completely continuous

fort ~ t 1 and T(t), t ~ 0, is point dissipative, then there exists a connected
compact global attractor A and there is an equilibrium point ofT(t).
The only statement that is not verified by the same method as in the
previous section is the assertion about the existence of an equilibrium point.
This is proved in the following way. Choose a sequence of positive numbers
wn ---> oo as n ---> oo and consider the discrete dynamical system defined by
the map Tn = T(wn)· From the previous results on discrete maps, there is
a fixed point Xn of Tn. It is clear that Xn E A, the global attractor in the
preceding theorems. Since A is compact, we can extract a subsequence and
apply Lemma 4.2.
4.6 Stability and maximal invariant sets in processes 121
4.6 Stability and maximal invariant sets in processes
In this section, we generalize the results of Section 4.3 to w-periodic pro-

cesses.
Definition 6.1. For a given process u on X and a given a E JR, we say that
a set M c JR x X attracts a setH c X at a if, for any E > 0, there is a
to(E, H, a) such that (a+t, U(a, t)H) C B(M, E) fort 2:: to(E, H, a). If M
attracts a setH at a for each a E JR, we simply say that M attracts H. We
say that M attracts points of X (at a) if M attracts each point of X (at
a). We say that M attracts neighborhoods of points of X (at a) if, for any
x EX, there is a neighborhood Ox of X such that M attracts Ox (at a).
We say that M attracts bounded sets if M attracts each bounded set of X.
We say that M is a global attractor if M is invariant and attracts bounded
sets of X. When we use the term uniform attracts, we shall mean that the
number t 0 can be chosen independently of a. Similar definitions are given
for M attracting other types of sets of X.
Definition 6.2. For a given process u on X and a given a E JR, we say that a
set M C JRxX is stable at a if, for any E > 0, there is a c5(E, a) > 0 such that
(a, x) E B(M, c5(E, a)) implies that (a+ t, U(a, t)x) E B(M, E) fort 2:: 0.
The set M is said to be stable if it is stable at a for all a E JR. The set M
is unstable if it is not stable. The set M is uniformly stable if it is stable
and the number c5 in the definition of stability is independent of a. The set
M is said to be asymptotically stable at a if it is stable at a and there is an
Eo(a) such that (a+ t, U(a, t)) --+ M as t--+ oo for (a, x) E B(M, Eo(a)).
The set M is said to be uniformly asymptotically stable if it is uniformly
stable and there is an Eo > 0 such that for any "7 > 0, there is a t 0 (ry, Eo)
having the property that (a+ t, U(a, t)) E B(M, ry) fort 2:: to(ry, Eo) and
all x such that (a, x) E B(M, Eo).
If a process is generated by an ordinary differential equation in JRn,

then M C JR x JRn stable at a fixed a E JR implies M is stable at every
a E JR; that is, M is stable. For general processes, this is no longer true. In
fact, it is not even true for RFDE. To see this, consider the process generated
by the linear RFDE discussed in Equation (5.1) of Section 3.5. For a= 0,
the solution is given by Equation (5.3) and, therefore, the solution x = 0
(that is, the set M = JR x { 0 }) is stable. On the other hand, for any a > 37r,
the solution x(a, ¢) must satisfy
x(t) = x(t- 3;).
For any>. satisfying>.= exp(-37r.A/2), the function aexp>.t is a solution

for any a E JR. But this latter equation has a solution >.0 > 0 and so the
solution x = 0 is unstable at a > 37r.
If u is a process and the set M = IR x K with K compact is stable at u,

then M is stable at ( < u. This is an immediate consequence of continuity
of U(u, t)x in (u, t, x). For an w-periodic process, U(u + kw, t) = U(u, t)
for all integers k. Thus a set M = IR x K with K compact is stable at u if
and only if it is stable.
It is difficult to determine when stability at u is equivalent to stability.
From a practical point of view, it does not seem to be significant to consider
systems that are stable at u and not stable. As a consequence, this weaker
concept will not be discussed in more detail.
Now suppose that u is an w-periodic process, K C X is compact, and
M c IR x K attracts compact sets of X. For any u E IR, consider the
discrete dynamical system {Uk(u, w), k 2 0}. For this discrete system, K
attracts compact sets of X and we can define, as in Section 4.3,
(6.1) J" = n
n2':0
Un(u, w)K, u E IR.
It was proved in Section 4.3 that J" is independent of K. If .:J c IR x X is

defined by
(6.2) .:J= U(u, Jcr),

crEIR
then .:J is an invariant set for the process u and
Jcr ~f{x EX: (u, x) E .:J} = J"

is compact. Also, Ucr J" is compact in X.
Theorem 6.1. Suppose that u is an w-periodic process on X, there is a

compact set K C X such that IR x K attracts compact sets of X, and let
.:J c IR x X be defined by (6.2). Then the following conclusions hold:
(i) .:J is connected.
(ii) .:J is independent of K, is a nonempty invariant set with :fer compact,
and .:J is maximal with respect to this property.
(iii) .:J is stable.
(iv) For any compact setH C X, .:J attracts H.
Proof. The proofs of Properties (i) and (ii) are the same as in the proof of
Theorem 3.1.
To prove Property (iii), we first observe that Theorem 3.1 implies that
the set J" in Equation (6.1) is a stable global attractor for the discrete
dynamical system defined by U(u, w). Therefore, for any E > 0, there is a
8(E, u) > 0 such that x E B(Jcr, E) for k 2 0. Since U(u, t) is continuous
in t, we may further restrict 8 to be assured that U(u, t)x E B(Jcr+t, E) for
0 ::; t ::; w if x E B(Jcr, 8). Since :fer = J" and .:J is invariant, it follows
4.6 Stability and maximal invariant sets in processes 123
that :Tis stable at a for every a E JR. Since U(a + kw, t) = U(a, t) for all
integers k, the set :Tis stable at a+kw with the constant 8 independent of k.
Therefore, it is only necessary to vary a E [0, w]. For 0 ~a~ wand t 2: 0,
U(a, t + w- a)= U(a, t)U(a, w- a). If 8(E, a) is the stability constant at
a, then the continuity of u and the compactness of JO' imply that there is a
0 < 81 (E) ~ 8(E, a), 0 ~a~ w, such that U(a, w-a)B(JO', 81) C B(Jw, 81)
for 0 ~ a ~ w. This shows that uniform stability and Part (iii) of the
theorem is proved.
The proof of Part (iv) may be supplied in a similar fashion using the
results of Theorem 3.1(iv) to complete the proof of the theorem. D
Definition 6.3. A process u on X is said to be point (compact) ( loca[) (locally

compact) (bounded:) dissipative if there is a bounded set B C X such that
IR x B attracts points (compact sets) (neighborhoods of points) (neighbor-
hoods of compact sets) (bounded sets) of X.
Definition 6.4. A process u on X is said to be asymptotically smooth if, for

any bounded set B C X, there is a compact set B* C X such that for any
E > 0, there is a t 0 = to(E, B) such that for any a E IR, U(a, t)x E B for
t 2: 0 implies that U(a, t)x E B* fort 2: t 0 •
Following arguments similar to the proof of Theorem 3.2 and Theorem

6.1, one can demonstrate the following result.
Theorem 6.2. Suppose that u is an w-periodic process on X that is local

dissipative and asymptotically smooth. Then there exists a set :T c IR x X
that satisfies all of the conclusions of Theorem 6.1 and
(v) For any compact setH C X, there is a neighborhood H 1 of H such that
'Y+(Hl) is bounded and :T attracts H 1 ; in particular, :T is uniformly
(vi) If B C X is bounded and 'Y+(a, B) is bounded, then :T attracts B.
In particular, if 'Y+(a, B) is bounded for every bounded set B c X,
a E IR, then :T is a compact global attractor.
Corollary 6.1. If u is an w-periodic process, then the following are equivalent

statements:
(i) There is a compact set K c X such that IR x K attracts compact sets
of X.
(ii) There is a compact set K C X such that IR x K attracts neighborhoods
of compact sets of X.
We also have the analogue of Theorem 3.3.

Theorem 6.3. If U (O", t) is asymptotically smooth, point dissipative and pos-

itive orbits of bounded sets are bounded, then there exists a connected global
attractor. Finally there is an w-periodic trajectory.
One also can prove the following generalization of Lemma 3.4 and
Corollary 3.2.
Theorem 6.4. Suppose that u is an w-periodic process on X that is point

dissipative and there is a t 0 > 0 such that for every O" E IR, U(O", t) is
conditionally completely continuous for t 2': to. Then there is a compact set
K C X such that for any compact setH C X, there is an open neighborhood
H 0 of H and a h > 0 such that for any O" E IR, ')'+(O", Ho) is bounded and
U(O", t)Ho C K fort 2': t1. In particular, there is a compact set K C X
such that M = IR x K attracts compact sets of X. If, in addition, U(O", t)
takes bounded sets into bounded sets, then, for any bounded set B c X,
the set 'Y+(O", B) is bounded and there is a h > 0 such that U(O", t)B c K
for t 2': t 1 . In particular, there is a compact, connected, global attractor.
Finally, there is an w-periodic trajectory.
Using Corollary 6.1 of Section 3.6 and Theorem 6.4, we have the fol-
lowing result for RFDE.
Theorem 6.5. If an w-periodic RFDE(f) generates an w-periodic process u

on C, U(O", t) is a bounded map for each O", t and u is point dissipative,
then there is a compact, connected, global attractor. Also, there is an w-
periodic solution of the RFDE(f). If the RFDE(f) is autonomous and the
same conditions are satisfied, then there is an equilibrium solution of the
RFDE(f).
For an RFDE(f), Theorem 6.5 gives the existence of a periodic solution

under the weak hypothesis of point dissipative. Some results are available
for special types of equations even with weaker hypotheses. For example, for
linear equations, one can prove the existence of a periodic solution assuming
only the existence of a bounded solution. To do this, we need the following
result of Darbo, which we state without proof.
Theorem 6.6. If r is a closed, bounded, convex subset of a Banach space

and T : r ---+ r is a-condensing, then T has a fixed point in r.
Theorem 6.7. For an w-periodic linear nonhomogeneous RFDE, the exis-

tence of a solution that is bounded for t 2': 0 implies the existence of an
w-periodic solution.
Proof. We need some results from linear systems in Chapter 6, but they are
very elementary and will be repeated here. For a linear RFDE
4. 7 Convergent systems 125
i:(t) = j(t, Xt) + h(t),

where f(t, ¢) is linear in¢, the solution x(cr, ¢,h) through (cr, ¢) can be
written as x(cr, ¢,h) = x(cr, ¢, 0) + x(cr, 0, h). Therefore, the period map
taking ¢into Xa+w(cr, ¢, h) is an affine transformation T : C----> C, T¢ =
L¢ + '1/J, where ¢ E C, 'ljJ E C and L : C----> Cis a bounded linear map.
Corollary 6.1 of Section 3.6 implies that L = S + U, where U is completely
continuous and the spectrum of S contains only the zero element. Therefore,
Lemma 4.4 implies that there is an equivalent norm in C such that S is
a contraction with contraction constant k. For any bounded set B in C,
a(LB) = a(SB) ::; ka(B). Therefore, L is an a-contraction. Since T is
only a translation of L, it follows that T is an a-contraction.
Suppose that x: [-r, oo) ____, IRn is a bounded solution of the RFDE.
If ¢ = x 0 , this implies that { Tk¢ : k ~ 1} is a bounded set in C. Let
D = co { ¢, T¢, T 2 ¢, ... }. If 'T} E D, then 'T} = LiEJ aiTi¢, where J is a
finite subset of the positive integers, ai ~ 0, LiEJ ai = 1. Furthermore,
iEJ iEJ
iEJ iEJ
and TD S: D. Since Tis continuous, T(Cl D) S: Cl D. Since Cl D is a

closed bounded convex set of C and T is an a-contraction, there is a fixed
point ofT in D by Theorem 6.6 and, thus, an w-periodic solution of the
RFDE. D
4. 7 Convergent systems
In this section, we investigate the simplest possible processes, namely, ones

for which all solutions are asymptotically stable and at least one solution is
bounded. We first treat discrete dynamical systems in detail and only state
the results for w-periodic processes.
Lemma 7.1. Suppose T : X ----> X is continuous and there is a compact set

that attracts compact sets of X. If the set J in Theorem 3.1 consists of a
single point, then there is a bounded orbit of T on ( -oo, oo), every orbit is
stable and attracts neighborhoods of points of X.
Proof. Let J = {x 0 }. Since J is compact and invariant, rnx 0 is defined for
-oo < n < oo, rnxo = Xo for all n and 'Y = {Tnxo : -oo < n < 00} = J
is bounded. By Theorem 3.1, 'Y is stable and so, for any E > 0, there is a
8 = 8(E) > 0 such that lx- xol < 8 implies ITnx- xol < E for n ~ 0 since
'Y = J. Also, Theorem 3.1 implies 'Y attracts neighborhoods of points of X.
Therefore, for any x EX, there is an n 0 (x) and a neighborhood Ox of x such

that ITno(x)y- xol < 8 for y E Ox. Stability of"'( implies ITny- rnxl < 2E
for n 2:: no(x), y E Ox. Continuity ofT therefore implies we can choose Ox
so that ITny- Tnxl < 2E for 0::; n::; n 0 (x). This implies the orbit "Y+(x)
is stable. The same type of argument also gives the fact that each orbit
attracts neighborhoods of points of X and the proof is complete. 0
Lemma 7.2. If T : X --+ X is continuous, if there is a bounded orbit ofT

on ( -oo, oo) and every trajectory is uniformly asymptotically stable, then
T is local dissipative.
Proof. Let {Tnx 0 : -oo < n < oo} = "'( be a bounded orbit of T. We first
show that
(7.1) lim ITnx- Tnxol =0
n---+oo
for all x E X. Let G = { x E X : Equation (7.1) is satisfied}. The set

G is not empty since x 0 E G. The set G is open because of the uniform
asymptotic stability. Also 8G is empty. Otherwise x* E 8G would not be
stable. Therefore G = X. This proves that T is point dissipative. The
hypothesis of uniform asymptotic stability implies the local dissipativeness
of T and the lemma is proved. 0
Definition 7.1. A continuous map T : X --+ X is said to be convergent if

there is a unique fixed point ofT that is stable and attracts neighborhoods
of points of X.
Theorem 7.1. If T is conditionally condensing, then T is convergent if and

only if there is a bounded orbit of T on ( -oo, oo) and every orbit is uni-
formly asymptotically stable.
Proof. IfT is conditionally condensing and convergent, then Tis condensing
and local dissipative. Therefore, Lemma 3.4 implies there is a compact set
that attracts compact sets of X. The set J in Theorem 3.1 therefore exists
and must contain only a single point. Lemma 7.1 implies the result.
Conversely, if there is a bounded orbit of T on ( -oo, oo) and every
orbit is uniformly asymptotically stable, then Lemma 7.2 implies Tis local
dissipative. We use Lemma 3.4 again and conclude T is convergent directly
from Theorem 3.1. The proof is complete. 0
It is possible to weaken the definition of convergent T and require only

that the unique fixed point is stable and a global attractor. An analogue
of Theorem 7.1 is then obtained by requiring that T be conditionally con-
densing and some iterate of T be conditionally completely continuous. One
must repeat arguments similar to the ones used in proving Theorem 3.1
and Lemma 3.1.
The analogue of these results for w-periodic processes are the following.
Definition 7.2. An w-periodic process u is said to be convergent if

(i) there is a unique w-periodic orbit;
(ii) the trajectory of this orbit is uniformly stable and uniformly attracts
neighborhoods of points of ffi. x X.
Theorem 7.2. If the w-periodic process u is conditionally condensing, then u

is convergent if and only if there is a bounded orbit on ( -oo, oo) and every
trajectory is uniformly asymptotically stable.
As for discrete processes, these hypotheses can be modified to allow

only asymptotic stability.
The theory of dissipative processes had its origin in a fundamental paper

of Levinson [1] dealing with periodic differential equations in the plane. By
considering the period map T, he formulated very clearly the basic prob-
lems. First, one should characterize the smallest set J that contains all
of the information about the limits of trajectories and then discuss T re-
stricted to J. And Levinson [1] formulated the concept of point dissipative
and proved that point dissipative implies the existence of a maximal com-
pact invariant set. From this fact, he was able to prove that some iterate
ofT must have a fixed point. In their remarkable work on the van der Pol
equation,
x + k(x 2 -1)± + x = bk)..cos>-.t,
with k large and b < 2/3, Cartwright and Littlewood (see Littlewood [1]
and Pliss [1] for references) in the late 1940s and 1950s discovered a large
number of periodic solutions of very high periods. These results indicated
the possible complicated structure of the set J even for rather innocent-
looking nonlinear equations. Levinson [2] also noted this same phenomenon.
Over the years, a tremendous literature on this subject has accumu-
lated and one may consult LaSalle [1], Pliss [1], Reissig, Sansone, and Conti
[1], and Yoshizawa [1] for references. Continuing in the spirit of Levinson for
finite dimensions, Pliss [1] showed that the maximal compact invariant set
is globally asymptotically stable. For the special case of RFDE for which
the period w is greater than the delay, Jones [2] and Yoshizawa [1] showed
the existence of w-periodic solutions by using the asymptotic fixed-point
theorem of Browder [1]. For a discrete point dissipative dynamical system
T on an arbitrary Banach space X, the existence of fixed points ofT were
proved by Horn [1] and Gerstein and Krasnoselskii [1] when Tis completely
continuous. Billotti and LaSalle [1] proved the same result when Tis com-
pletely continuous, but also characterized the maximal compact invariant
set and proved it is globally asymptotically stable.
Gerstein [1] considered the case when T is point dissipative and is

condensing on balls in X and showed the existence of a maximal compact
invariant set, but concluded nothing about stability of this set or the ex-
istence of fixed points ofT. Hale, LaSalle, and Slemrod [1] and Hale and
Lopes [1] proved many of the same results that are contained in Sections
4.3~4.6 although in a different order and sometimes less generally. Nuss-
baum [2] has also obtained fixed-point theorems similar to Theorem 4.3.
Theorem 4.1 is due to Horn [1]. Theorem 4.6 forT completely continuous
is due to Browder [1]. Theorem 6.4 is due to Chow [1] with the proof in the
text due to Chow and Hale [1]. The latter paper also contains an interesting
class of mappings, called strongly limit compact, which were motivated by
Lemma 4.1.
In IRn, Jones [3] and Jones and Yorke [1] have proved results on ex-
istence of periodic solutions using the concept of a set S c IRn being
compactly constrained by a dynamical system.
For dynamical systems on locally compact spaces, Theorem 6.1 was
given in a more general form by Bhatia and Hajek [1]. More specifically,
they assume that there is a compact set M such that w(x) n M -=/= 0 for
each x EX. In particular, this weak hypothesis in IRn implies that there is
a compact set that attracts compact sets. This latter property was basic to
our entire investigation in this chapter. It is not known how to weaken the
hypotheses on Theorem 6.1 for Banach spaces that are not locally compact.
Section 4. 7 is a generalization of the corresponding results of Pliss and
Krasovskii (see Pliss [1]) in finite-dimensional space. For a detailed dis-
cussion of the relationship between stability and the existence of periodic
and almost periodic solutions of ordinary and functional differential equa-
tions, see Burton [3], Yoshizawa [2], Hino, Murakami, and Naito [1] and the
references therein.
The measure of noncompactness of a set was introduced by Kuratowski
[1]. The fixed-point theorem stated in Theorem 6.6 is due to Darbo [1].
For an excellent survey of measures of noncompactness and fixed-point
theorems, see Sadovskii [1].
These remarks appeared in the original version of the book of Hale
[22]. At that time, he was unaware of the very important work of Ladyzen-
skaya [1] in which she showed the existence of a compact global attractor for
the Navier-Stokes equation on a bounded domain in two space dimensions.
During the last fifteen years, there has appeared a tremendous amount of
literature on global attractors, especially for partial differential equations.
The topics center around the existence of attractors, the finiteness of the
Hausdorff dimension of the attractor and estimates of this dimension and
the structure of the flow on the attractor. It would take us too far afield
to attempt to make a survey of these developments. The interested reader
should consult the recent books of Babin and Vishik [1], Hale [23], Ladyzen-
skaya [2], and Temam [1] for a guide to some of the literature.
The processes described in Section 4.1 can be given a much more
dynamical interpretation by using the concept of skew product flows. To

see this, let W be the set of processes endowed with some metric and, for
some u E W, define the translation a(T), T E R, of the process as
(8.1) (a(T)u)(s, x, t) = u(T + s, x, t).

Then a(O) =I, a(t + T) = a(t)a(T) for all t, T E R, and a(T), T E R, is a
semigroup. Also, let a: X x R+ x W- X be defined by
(8.2) a(x, t, u) = u(O, x, t).

With a and a as in (8.1) and (8.2), define
(8.3) 7r(t) : X x W- X x W, ?T(t)(x, u) = (a(x, t, u), a(t)u).

One now verifies that 7r(O) =I, ?T(t+T) = 7r(t)7r(T); that is, 7r(t), t?: 0,
is a semigroup of transformations. The nonautonomous process has been
converted into an autonomous process by enlarging the system! If the map
?T(t)(x, u) were continuous in t, x, u, then 7r(t), t ?: 0, would be a C 0-
semigroup of transformations and all of the previous theory could be ap-
plied. This approach has led to many interesting results, especially in the
case where the process arises from a system of evolutionary equations for
which the vector field is almost periodic in time (see Miller and Sell [1],
Sacker and Sell [1], Sell [1, 2], Hale [23], and the references therein).
For the spectral theory of the linearization of such processes with ap-
plications to FDE and PDE, see Magalhaes [6, 8]. His results apply at least
to periodic systems and asymptotically autonomous systems.
5
Stability theory
In this chapter, we discuss methods for determining stability and ultimate

boundedness of solutions of RFDE. The method of Liapunov functionals is
discussed as well as the method based on the use of functions on IRn in the
spirit of Razumikhin.
5.1 Definitions
Suppose f : IR x C----+ IRn is continuous and consider the RFDE(f)
(1.1) x(t) = f(t,xt)·
The function f will be supposed to be completely continuous and to satisfy

enough additional smoothness conditions to ensure the solution x(u, rjJ )(t)
through (u, rjJ) is continuous in (u, ¢, t) in the domain of definition of the
function. The following theory for f : (a, 00) X n ----+ IRn' where n is an
open set in C is valid and it is only for notational purposes that the domain
of definition off is taken to be IR x C.
The definition of stability of the solution x = 0 was given in Section
4.5 via the process u generated by the RFDE(f). It is convenient to restate
these definitions directly in terms of the solution of Equation (1.1).
Definition 1.1. Suppose f(t, 0) = 0 for all t E IR. The solution x = 0 of

Equation (1.1) is said to be stable if for any u E IR, E > 0, there is a
8 = 8(E,u) such that rjJ E !3(0,8) implies Xt(u,¢) E l3(0,E) fort;:::: u. The
solution x = 0 of Equation (1.1) is said to be asymptotically stable if it
is stable and there is a b0 = b0 ( u) > 0 such that rjJ E !3( 0, b0 ) implies
x(u,rjJ)(t)----+ 0 as t----+ oo. The solution x = 0 is said to be uniformly stable
if the number 8 in the definition is independent of u. The solution x = 0
of Equation ( 1.1) is uniformly asymptotically stable if it is uniformly stable
and there is b0 > 0 such that for every 'T/ > 0, there is a t 0 (ry) such that
rjJ E l3(0,b0 ) implies Xt(u,¢) E B(O,ry) fort;:::: u+to('TI) for every u E IR.
5.1 Definitions 131
If y(t) is any solution of Equation (1.1), then y is said to be stable if

the solution z = 0 of the equation
z(t) = f(t, Zt + Yt) - f(t, Yt)
is stable. The other concepts are defined in a similar manner.

For some RFDE, there is no distinction between stability and uniform
stability. In fact, we now prove
Lemma 1.1. If there is an w > 0 such that f(t + w, ¢) = f(t, ¢) for all
(t, ¢) E lR. x C, then the solution x = 0 is stable (asymptotically stable) if
and only if it is uniformly stable (uniformly asymptotically stable).
Proof. Suppose the solution x = 0 is stable. Since Xt+CT+kw(a + kw, ¢) =
Xt+"'(a, ¢) for all t ~ 0, a E JR., it is only necessary to show that the
number 8(E, a) in the definition of stability can be chosen independent of
a E [0, w] in order to prove x = 0 is uniformly stable. For 0 ~ a ~ w
and t ~ 0, Xt+w(a, ¢) = Xt+w(w, xw(a, ¢)). Therefore, the continuity of
x(a, ¢)(t) in (a,¢, t) implies there is a 81 > 0, 81 ~ 8(E, a), 0 ~ a ~ w,
such that xw(a,¢) E B(0,8(E,w)), if¢ E B(0,81) for 0 ~a~ w. Since
Xt(w, ¢) E B(O, E) fort~ w if¢ E B(O, 8(E, w)), this proves uniform stability.
The proof of uniform asymptotic stability is more difficult. If x = 0
is asymptotically stable, it is stable and therefore, uniformly stable. In
addition, if¢ E B(O, b0 ), then Xt(a, ¢)-+ 0 as t-+ oo, that is, the set M =
lR. x {0} attracts points of the ball B(O, b0 ). The representation theorem,
Theorem 6.1 of Section 3.6, for the solution operator T( t, a) implies there is
a t 0 > 0, independent of a such that T( a+ t, a) is conditionally completely
continuous fort~ t 0 . The complete proof of the uniform approach to zero
is now supplied in the same way as for the proof of Theorem 5.3 of Section
4.5. D
Definition 1.2. A solution x(a, ¢) of an RFDE(f) is bounded if there is a

{3(a, ¢) such that lx(a, ¢)(t)1 < {3(a, ¢) for t ~ a- r. The solutions are
uniformly bounded if for any a > 0, there is a {3 = {3(a) > 0 such that for
all a E JR., ¢ E C, and 1¢1 ~ a, we have lx(a, ¢)(t)l ~ {3(a) for all t ~ a.
The solutions are ultimately bounded if there is a constant {3 such that for
any (a,¢) E lR. x C, there is a constant to(a, ¢) such that lx(a, ¢)(t)1 < {3
fort~ a+to(a,¢). The solutions are uniformly ultimately bounded if there
is a {3 > 0 such that for any a > 0, there is a constant t 0 (a) > 0 such that
lx(a, ¢)(t)l ~ {3 fort~ a+ to(a) for all a E JR., ¢ E C, 1¢1 ~a.
This definition of ultimate boundedness coincides with the concept of

point dissipative for the process generated by the RFDE(f). As a conse-
quence of Theorem 5.2 of Section 4.5, we can state the following result.
132 5. Stability theory
Lemma 1.2. If a periodic RFDE(f) is such that the solution map T(t, 0') :
C ---> C is defined for all t 2:: O", T(t, 0')¢ is continuous in (t, O", ¢) and
T(t, 0') takes bounded sets into bounded sets, then ultimate boundedness is
equivalent to uniform ultimate boundedness.
5.2 The method of Liapunov functionals
In this section, we give sufficient conditions for the stability and instability
of the solution x = 0 of Equation (1.1) that generalize the second method
of Liapunov for ordinary differential equations. The results are illustrated
by examples.
If V : 1R x C ---> 1R is continuous and x( O", <P) is the solution of Equation
(1.1) through (0', ¢), we define
. 1
V(t, ¢)=lim sup -h [V(t + h, Xt+h(t, ¢))- V(t, ¢)].
h-+O+
The function V(t, <P) is the upper right-hand derivate of V(t, <P) along the
solution of Equation (1.1). If we wish to emphasize the dependence on
Equation (1.1), we write i(u)(t, ¢).
Theorem 2.1. Suppose f: 1R x C---> 1Rn takes 1R x (bounded sets of C) into

bounded sets oflRn, and u,v,w: JR+---> JR+ are continuous nondecreasing
functions, u( s) and v( s) are positive for s > 0, and u(O) = v(O) = 0. If
there is a continuous function V : 1R x C ---> 1R such that
u(l¢(0)1)::; V(t, ¢)::; v(l¢1)

V(t, ¢)::; -w(l¢(0)1)
then the solution x = 0 of Equation (1.1) is uniformly stable. Ifu(s)---> oo as
s---> oo, the solutions of Equation (1.1) are uniformly bounded. Ifw(s) > 0
for s > 0, then the solution x = 0 is uniformly asymptotically stable.
Proof. There is no loss of generality in assuming r > 0. For any E > 0, there
is a 8 = 8(E), 0 < 8 < E, such that v(8) < u(E). If )(t)l)::; V(t,xt(O",¢))::; V(O",¢)::; v(t5) < u(E), t 2:: (]',

Therefore, lx(O", ¢)(t)1 < E, t 2:: 0'. Since I<PI < 8 < E, this proves uniform
stability.
If u( s) ---> oo as s ---> oo and a > 0 is any given constant, there is a
(3 > 0 such that u((J) = v( a). Consequently, if I<PI ::; a, the solution x( 0', <P)
through (0', ¢) satisfies lx(O", ¢)(t)l ::; (3 for all t 2:: 0'. This obviously implies
uniform boundedness.
5.2 The method of Liapunov functionals 133
To prove the assertion of the theorem concerning uniform asymptotic

stability, for E = 1, choose 80 = 8(1) as the constant for uniform stability.
For any E > 0, we wish to show there is a t 0 (80 , E) > 0 such that any
solution x(u, ¢) of Equation (1.1) with 1¢1 < 8o satisfies lxt(u, ¢)1 < E for
t ~ u+t0 (80 , E). Let 8 = 8(E) be the constant for uniform stability. Suppose
that a solution x = x(u, ¢), 1¢1 < 80 satisfies lxtl ~ 8, t ~ u. Since each
interval of length r contains an s such that lx(s)l ~ 8, there exists a sequence
{tk} such that
u + (2k- 1)r :::; tk :::; u + 2kr, k = 1,2, ... '
and
lx(tk)l ~ 8.
By the assumption of j, there exists a constant L such that lx(t)l < L for
all t ~ u. Therefore, on the intervals tk- (8/2L) :::; t :::; tk + (8/2L), we
have lx(t)l > 8/2. Therefore,
8 8
tk - - <t <
2L- -
tk + -.
2L
By taking a large L, if necessary, we can assume that these intervals do not
overlap, and hence
8 8
V(tk>Xtk)- V(u,¢):::; -w("2)L (k -1).
Let K(8o,L) be the smallest integer~ v(80 )/((8/L)w(8/2)). If k > 1 +

K(8o, L), then
which is a contradiction. Therefore, at some T = T( u, ¢) such that u <

T:::;u + 2rK(8o, L) we have lxrl < 8 and lxtl < E fort~ u + 2rK(80 , L).
This proves the uniform asymptotic stability. The proof of the theorem is
therefore complete. D
Let us consider a method of constructing a particular functional satis-

fying the conditions of Theorem 2.1 for the equation
(2.1) x(t) = Ax(t) + Bx(t- r), r > 0,
where A and B are constant matrices. Suppose the eigenvalues of A have
negative real parts and choose the symmetric matrix C such that C > 0
and ATC + CA = -D < 0, where AT is the transpose of A. If E is a
positive definite matrix and
(2.2) V(¢) = ¢(0fC¢(0) +I: ¢(BfE¢(B) dB

then vl¢(0)1 2 ~ V(¢) ~ Kl¢1 2 for some positive v, K. Furthermore,
V(¢) = -¢(0)TD¢(0) + 2¢(0fCB¢(-r) + ¢(0)T E¢(0)

(2.3)
- ¢( -r)T E¢(-r).
If we consider the right-hand side of Equation (2.3) as a quadratic form
in ¢( 0), ¢( -r) and impose conditions on the matrices A and B to ensure
there exist matrices C and E such that this quadratic form is negative
definite in ¢(0), ¢( -r), then Theorem 2.1 will imply uniform asymptotic
stability of the solution x = 0. More specifically, we try to determine A, B,
C, and E so that the symmetric matrix
(2.4) [ D-E (CBf]

-CB E
is positive where - D = AT C + CA. We already know that D and E are
positive. Consequently, Matrix (2.4) for B = 0 is positive if D-E > 0.
As a result of this, this matrix is positive for B sufficiently small and the
solution x = 0 is asymptotically stable. This latter fact is also an immediate
consequence of the results in Chapter 1, where it was shown that asymptotic
stability prevailed if the roots of the characteristic equation
(2.5) det [I- A- Be-,\r] = 0
satisfy ReA< 0.
One can actually use this method to obtain explicit estimates on B for
which Matrix (2.4) is positive. To be more specific, suppose E < D and
Then
V(¢) ~ -AI¢(oW + 2IICBIII¢(O)II¢(-r)l- JLI¢(-rW.
If AJL-IICBII 2 > 0, then V(¢) ~ -k(I¢(O)I 2 +1¢(-r)l 2 ), r > 0, for a suitable
positive constant k and Theorem 2.1 implies uniform asymptotic stability.
These estimates are certainly not optimal and the best estimates using
this procedure are not easy to obtain. However, there is one important
qualitative remark that can be made about proving stability by insisting
that Matrix (2.4) is positive. If Matrix (2.4) is positive, then the solution
x = 0 of Equatio:q (2.1) is asymptotically stable for every value of the delay
r.
For the scalar equation
(2.6) ±(t) = -ax(t)- bx(t- r),

if
(2.7) JL > 0,
then
The corresponding Matrix (2.4) is positive if and only if
(2.8) a> f.1 > 0,

The choice of f.1 that will allow lbl as large as possible is f.1 = a/2. For
this choice of f.l, we see that the solution x = 0 of Equation (2.6) is asymp-
totically stable if lbl <a. The exact region of asymptotic stability is shown
in Figure 5.1 and is obtained by applying the theory in the appendix to the
equation>.= -a- bexp(->.r) (see Theorem A.5). The upper boundary of
the region of stability is given parametrically by the equation
7r
a= -bcos(r, bsin(r = (, 0 < ( < -.
r
The region lbl < a is precisely the region for which there is asymptotic
stability for all r > 0 and an estimate of the rate of decay of zero that is
independent of the delay r.
(0, n/(2r))
(-1 /r, 1/r)

-------------------- ------------------------
'
'
'
'
'
'
'
'
'
''
Fig. 5.1.
Even if a and b in Equation (2.6) depend on t, we obtain very in-

teresting information about stability using the same type of functional V
in Equation (2.7). In fact, suppose a(t) and b(t) are bounded continuous
functions on ffi with a(t) ?: 8 > 0 for all t and consider the equation
(2.9) i:(t) = -a(t)x(t)- b(t)x(t- r).

Using the functional V in Equation (2.7) with f-l = 8/2, one sees that V(¢)
is the same expression as before and the solution x = 0 of Equation (2.9) is
uniformly asymptotically stable if Inequalities (2.8) are satisfied; that is, if
(2a(t) -8)8 > b2 (t) uniformly in t. In particular, the conditions are satisfied
if there is a constant B, 0 :S (} < 1 such that [b(t)[ :S (}8 for all t E IR.
If r = r(t) is continuously differentiable and bounded, then using the
same argument and the same V as in Equation (2. 7) one observes that
uniform asymptotic stability prevails if a(t) 2: 8 > f-l > 0 and
(2a(t)- f.-l)(1- r(t))f.-l > b2 (t)

uniformly in t.
For the autonomous matrix Equation (2.1), the exact region of stability
as an explicit function of A, B, and r is not known and probably will never
be known. The reason is simple to understand because the characteristic
equation (2.5) is so complicated. It is therefore, worthwhile to obtain meth-
ods for determining approximations to the region of stability. One possible
approach is to make use of Theorem 2.1. If this approach is taken, then the
functional V(¢) must be more complicated than the one in Equation (2.2)
since the corresponding stability region is independent of the delay r.
For Equation (2.6), we can give a very simple result on a possible form
for the Liapunov functional.
Suppose a,(3: [-r,O]--> IR, "(: [-r,O] x [-r,O]--> IR are continuously
differentiable functions and consider the functional V on C defined by the
quadratic form
V(¢) = ¢ 2 (0) + 2¢(0) [or a( B)¢( B) d(} +[or (3(B)¢ 2(B) d(}
(2.10)
+[Or [or ¢(~)¢(ry)"f(~, ry) d~dry.
A quadratic form H on C is said to be positive if H (¢) > 0 for ¢ =f. 0.
If such an H is positive, we write H > 0. To obtain the stability region
for Equation (2.6) by means of a functional of the form given in Equation
(2.10), we make use of the following result.
Lemma 2.1. If there exist a, (3, "(, and H > 0 such that the derivative of V
in Equation (2.10) along the solutions of Equation (2.6) satisfies
V(¢) = -H(¢)
for all¢ E C, then no root of the characteristic equation
(2.11) A= -a- be->.r
lies on the imaginary axis. Furthermore, V > 0 if and only if no roots of

Equation (2.11) have positive real parts.
Proof. Suppose there exists an H as specified and there is also a purely

imaginary root of Equation (2.11). Then there is a periodic solution x of
Equation (2.6) such that Xt -/=- 0 for all t. Thus, V(xt) is a periodic function
and V(xt) = -H(xt) < 0 for all t. This is obviously a contradiction.
To prove the second statement, suppose V > 0. If some roots of char-
acteristic equation (2.11) have positive real parts then there is a solution x
of Equation (2.6) for which V(xt) is unbounded. But this contradicts the
fact that 0 < V(xt) :::; V(x 0 ) for t 2': 0. Conversely, suppose no roots of
Equation (2.11) have positive real parts and Vis not positive. If there is a
nonzero ¢ such that V (¢) :::; 0, then the solution x through ¢ is such that
Xt -/=- 0 for 0 :::; t < E for some E sufficiently small. Therefore, V (Xt) < 0 for
0 < t < E. Thus, there is a 'ljJ -/=- 0 such that V ('ljJ) < 0. The instability The-
orem 3.3 of Section 5.3 gives a contradiction and the lemma is proved. D
The problem remains to determine a, (3, "(, and H. If we simply pro-

ceed to calculate V along the solutions of Equation (2.6), we obtain a rather
complicated set of equations for which it is difficult to recognize the nat-
ural choices for these functions. A more reasonable approach is to try to
determine the analogue of the converse theorem of Liapunov for ordinary
differential equations. Let us be more specific. For a matrix autonomous or-
dinary differential equation (Equation (2.1) with B = 0), if V(x) = xTCx
is a given quadratic form, then the derivative along the solutions is given
by V(x) = xT(ATC + CA)x = -xTWx. If all solutions approach zero as
t ----> oo, then it is the classical theorem of Liapunov states that, for any
positive definite matrix W, there is a positive definite matrix C such that
AT C + C A = - W. Conversely, if the latter equation has a solution for a
positive definite matrix W, then the solutions of the differential equation
approach zero. We now formulate a similar result for RFDE.
Consider the general linear autonomous system
(2.12) x(t) = Lxt = j_or d[ry(B)] x(t +B),

where 17 is an n x n matrix function of bounded variation. Exponential
solutions of (2.12) are determined from the characteristic equation
(2.13)
If all of the solutions of (2.13) satisfy Re .A. < 0, then we will see in Chapter
7 that all solutions of (2.12) approach zero as t----> oo.
Theorem 2.2. If all solutions of (2.13) satisfy Re .A. < 0, then for any positive
definite n x n matrix W, there is a quadratic functional V : C ----> lR such
that for¢ E C, the derivative V(¢) along the solutions of (2.12) satisfies
(2.14) V(¢) = -rp(O?W¢(0)
and there is a constant k > 0 and for any a > 0 a constant ka. > 0 such
that for I<PI :S a,
(2.15)
We remark that even though V(¢) is a quadratic form in¢, the lower
bound on V(¢) involves 1¢(0)1 3 and the lower bound also depends on the
upper bound of 1¢1. For an equation (2.12) for which there exists a function
V satisfying the conditions of Theorem 2.2, we can apply Theorem 2.1
to conclude that the origin is uniformly asymptotically stable. Since the
system (2.12) is linear, this implies that all of the solutions approach zero.
We now define the method for constructing the function V. For any
given n x n matrix W, define the matrix Y(s) = Yw(s), -r :::; s :::; r, by
the relation
(2.16)
and the function V (¢) by
V(¢) =¢(0fY(0)¢(0) + 2¢(0f lor 1° Y( -u +e) d77(u) ¢(e) de
(2.17) + [or1° ds¢(sf d77T(h)

[I: 1°Y( -s + h- u+ e)d7](u) ¢(e) de].
It can then be shown that Y ( s) is continuously differentiable for s -=/= 0 and

satisfies the properties:
Y(O) is symmetric and Y(s) = YT(-s), s 2:0,

(2.18)
Y(s)= I:d[7]T(e)]Y(s+e), o::;s::;r,
where Y = dYjds. Also,
r
0
V(¢) =¢(of [[ Y( -e) d[77(e)] +[or d[77T(e)] Y(e)]¢(0)
(2.19)
=¢(Of [Y(O) + yr (0)] ¢(0)
= -¢(ofw ¢(0).
If the matrix W is positive definite and the solutions of (2.13) have neg-
ative real parts, then the function V satisfies the properties stated. The
proofs of these facts are far from trivial and the reader may consult the
supplementary remarks for references to the details.
We now give more details on the construction of the functional V for

the special equation (2.1). Relation (2.18) implies that
This implies that
since
Y(s) = ATY(s)- BT:YT(r- s)
= ATY(s)- BTYT(r- s)A- BTYT( -s)B
= ATY(s)- BTY(s- r)A- BTY(s)B
= ATY(s)- Y(s)A + ATY(s)A- BTY(s)B.
From (2.19), we see that not only mustY satisfy the second-order ordinary
differential equation, but it must also satisfy the initial condition
(2.21) :Y(o) + :YT(o) = -w

If the eigenvalues of the equation
(2.22) det [AI- A- Be->.r] = 0
have negative real parts, then Theorem 2.2 asserts that for any positive
definite matrix W, the function V defined in (2.17) withY satisfying (2.20),
(2.21), and (2.18) satisfies the conditions (2.14) and (2.15).
Even though it will in general be impossible to give the explicit formula
for the functional V, we can use the fact that we know that it must exist
and be of the preceding form to obtain approximately the region of stability.
In fact, we can make intelligent guesses for the function Y and then verify
that the conditions of Theorem 2.1 are satisfied.
It is possible to give the explicit form of V for the scalar equation (2.6)
with r = 1; that is,
(2.23) x(t) = -ax(t) - bx(t- 1).
If we choose W = 1, then the equations (2.20) and (2.21) for this case are
. 1
(2.24) Y(O) = - -
2
and (2.18) implies that
(2.25) Y(s) = -~Y(s)- bY(s- 1) = -aY(s)- bY(1- s).
The functional V is given by

0
V(¢) =Y(0)¢ 2 (0)- 2b¢(0) /_ 1 Y(O + 1)¢(0) dO
(2.26)
+ b2 /_01 ¢(u)[/_01 Y(O- u)¢(0) dO] du.
After some rather lengthy computations, it can be shown that if (3 +
a+ be-!3 =f- 0 (which is always the case if the solutions of the characteristic
equation have negative real parts) and
1 1 +a
Y(s)=- 2 s+~, O:::;s:::;1, a=b=f-0
andY( -s) = Y(s), 0:::; s:::; 1, then

bef3 ef3s ((3 + a )e-f3s
Y(s) = - 2(3((3 +a+ be-f3) + 2(3((3 +a+ be-f3)' 0 :::; s:::; 1
and (3 2 = a 2 - b2 =f- 0.
Our next objective is to give sufficient conditions for the instability of
the solution x = 0 of an RFDE(f).
Theorem 2.3. Consider an RFDE(f) and designate the solution through

(cr, ¢) by x(cr, ¢). Suppose V(¢) is a continuous bounded scalar function on
C. If there exist a 'Y > 0 and an open set U in C such that
(i) V(¢) > 0 on U, V(¢) = 0 on the boundary of U,
(ii) 0 belongs to the closure ofUnB(O,"f),
(iii) V(¢):::; u(l¢(0)1) on U n B(O,"f),
(iv) v_(<f>);::: w(l¢(0)1) on [O,oo) X UnB(O,"(),
. 1
V_(<f>) = liminf -h[V(xt+h(t, ¢))- V(¢)]
h--->0+
where u(s), w(s) are continuous, increasing, and positive for s > 0, then the
solution x = 0 of the RFDE(f) is unstable. More specifically, each solution
Xt(cr, ¢) with initial function¢ in U n B(O, 'Y) at cr must reach the boundary
of B(O, 'Y) in a finite time.
Proof. Suppose¢ E U n B(O,"f), cr E JR. Then V(¢) > 0. By Hypothesis
(iii), 1¢(0)1 ;::: u- 1 (V(¢)) and Hypotheses (iii) and (iv) imply Xt = Xt(cr, ¢)
satisfies lx(t)l ;::: u- 1 (V(xt)) ;::: u- 1 (V(¢)) as long as Xt E U n B(O,"f).
Consequently,
V_(xt);::: w(lx(t)l);::: w(u- 1 (V(¢))) > 0 if xt E u n B(O, "f).

If we let rJ = w(u- 1 (V(¢))), then this implies
V(xt) ;::: V(¢) + ry(t- cr)

as long as Xt E U n B(O, f'). Hypotheses (i) and (iv) imply that Xt cannot
leave UnB(O,'Y) by crossing the boundary of U. Since V(¢) is bounded on
u n B(O, 'Y) this implies there must be a h such that Xtl E aB(O, f'). This
proves the last assertion of the theorem. But Hypothesis (ii) implies that in
each neighborhood of the origin of C, there are 4> in U n B(O, f'). Instability
follows and the proof of the theorem is complete. 0
As an example, consider Equation (2.6) with a + b < 0. We wish to

prove by use of Liapunov functions that the solution x = 0 of Equation
(2.6) is unstable, for some r < r 0 (a, b). Even though better results were
obtained before, this Liapunov function may be used for nonlinear and
nonautonomous equations.
If F is any given continuously differentiable function and
x2(t)
V(xt) = - - -
2 2 t-r
lit
F(t- u)[x(u)- x(t)] 2 du,
then it easily seen that
V_(xt) = V(xt) =-(a+ b)x2 (t)- b[x(t- r)- x(t)]x(t)

1
+ 2F(r)[x(t- r)- x(t)] 2
1 rt
- 2 lt-r F(t- u)[x(u)- x(t)] 2 du
+ i~r F(t- u)[x(u)- x(t)]

x [-(a+ b)x(t)- b{x(t- r)- x(t)}] du.
If the expression for V is written as an integral from [t - r, t], then the

integrand will be a positive definite quadratic form in x(t), [x(t - r) -
x(t)], [x(u)- x(t)] if the following inequalities are satisfied
a+b < 0,
Ll1 ~f- (a;b) F(r)- ~ > 0,

-.1 1 ( -~
F(O))- (a+ b) 2 F 2 (0)F(r) > 0, 0 ::; () ::; r.
r2 2 8r
If a+ b < 0, then one can determine an r 0 (a, b) and a continuously dif-
ferentiable positive function F( 0), 0 ::; () ::; r < r 0 (a, b) such that these
inequalities are satisfied. Consequently, there exists a positive number q
such that
V(¢) ::; ¢2~0)'
for all ¢ in C. If
then U satisfies Hypotheses (i) and (ii) of Theorem 2.2 and the solution
x = 0 of Equation (2.6) is unstable if a+ b < 0 and r < r0 (a, b).
Notice that the same conclusions for this example are valid if a and b
are functions oft provided that a and bare bounded and a(t)+b(t) < 8 < 0
for all t.
(2.27) x(t) = a(t)x 3 (t) + b(t)x 3 (t- r)

where a(t) and b(t) are arbitrary continuous bounded functions with a(t) 2':
L:
8 > 0, lb(t)l < q8, 0 < q < 1. For
V(¢) = ¢4~o) - ~ ¢6(e) de,
we have V(¢) ::=:; ¢ 4(0)/4 and
V_(¢) = V(¢) = [a(t)- ~]¢ 6 (0) + b(t)¢3 (0)¢3 ( -r) + ~ ¢ 6( -r).

This last expression is a positive definite quadratic form in ¢ 3 (0), ¢ 3 ( -r).
If
then the same argument as in the previous example shows that x = 0 is an

unstable solution of Equation (2.27).
If a(t) ::=:; -8 < 0, lb(t)l < q8, one can choose
V(¢) = ¢4(0) + ~
4 2
jo
-r
¢6(e) de
and use Theorem 2.1 to prove the zero solution is uniformly asymptotically
stable.
Notice that a more refined argument using the same V functionals may
be employed to show that the zero solution of
is stable or unstable according to whether a < 0 or > 0, provided only that

lb(t) I is bounded on IR. One simply must operate in a sufficiently small
neighborhood of the origin.
5.3 Liapunov functionals for autonomous systems 143
5.3 Liapunov functionals for autonomous systems
Consider the autonomous equation
(3.1) x(t) = f(xt)
where f: C---> IRis completely continuous and solutions of Equation (3.1)

depend continuously on the initial data. We denote by x( ¢) the solution
through (0, ¢).
If V : C ---> IR is a continuous function, we define the derivative of V
along the solution of Equation (3.1) as in Section 5.2; namely
. . 1
V(¢) = V(3.1)(¢) =lim sup -h [V(xh(¢))- V(¢)].
h-+O+
Definition 3.1. We say V : C ---> IR is a Liapunov function on a set G in C

relative to Equation (3.1) if Vis continuous on G (or Cl G), the closure of
G, and V ~ 0 on G. Let
s = {¢ E G: V(¢) = 0}
M =largest set in S that is invariant with respect to Equation (3.1).
Theorem 3.1. If V is a Liapunov function on G and Xt(¢) is a bounded

solution of Equation (3.1) that remains in G, then Xt(¢) tends to M as
t - t 00.
Theorem 3.2. IfV is a Liapunov function on Uz = {¢ E C: V(¢) < l} and

there is a constant K = K(l) such that¢ in Uz implies 1¢(0)1 < K, then
any solution Xt(¢) of Equation (3.1) with¢ in Uz tends toM as t---> oo.
Proof of Theorem 3.1. If lxt(¢)1 < K, Xt(¢) E G, t ;::: 0, then {xt(¢)}

belongs to a compact set of C and has a nonempty w-limit set, w(')'+(¢)).
Thus, V ( Xt (¢)) is nonincreasing, bounded below, and must approach a limit
cast---> oo. Since Vis continuous on Cl G, V('l/J) = c for 7/J in w('Y+(¢)).
Since w('Y+(¢)) is invariant V('ljJ) = 0 on w('Y+(¢)). Since every solution
approaches its w-limit set, this proves Theorem 3.1. 0
Proof of Theorem 3.2. If ¢ is in Uz and V ~ 0 on Uz, then Xt(¢) E Uz,
t;::: 0. Also lx(¢)(t)1 ~ Kif t;::: 0, which implies Xt(¢) bounded. Now use
Theorem 3.1. 0
Corollary 3.1. Suppose V C ---> IR is continuous and there exist non-

negative functions a(r) and b(r) such that a(r)---> oo as r---> oo
a(l¢(0)1) ~ V(¢), V(¢) ~ -b(l¢(0)1).

Then the solution x = 0 of Equation (3.1) is stable and every solution

is bounded. If, in addition, b(r) is positive definite, then every solution
appmaches zem as t ---> oo.
Pmof. Stability is immediate. The solutions are bounded since a( r) ---> oo
and lx(t)l bounded fort 2: 0 implies Xt bounded fort 2: 0. If b is positive
definite, the conditions of Theorem 3.2 are satisfied for any l. Furthermore,
S = {¢: ¢(0) = 0}, M = {0}. D
Theorem 3.3. (Instability). Suppose zem belongs to the closure of an open

set U inC and N is an open neighborhood of zem in C. Assume that
(i) V is a Liapunov function on G = N n U.
(ii) M n G is either the empty set or zem.
(iii) V(¢) < rJ on G when¢ -=f. 0.
(iv) V(O) = rJ and V(¢) = rJ when¢ E 8G n N.
If N 0 is a bounded neighborhood of zem pmpe'rly contained in N, then ¢ -=f. 0
in G n No implies there exists a T such that XT (¢) E aN0.
Pmof. If¢ E GnNo, ¢-=f. 0, then V(xt(¢)) s
V(¢) < rJ for all t;:::: 0 as long
as Xt(¢) remains in N 0 n G. If Xt(¢) remains in the bounded set N 0 n G
for all t ;:::: 0, then the w-limit set w(l+(¢)) <:;; N 0 n G. Also, w(r+(¢))
is invariant. Hypothesis (ii) implies w(r+(¢)) = {0}. On the other hand,
V(O) = TJ· This is a contradiction. Therefore, there is a T > 0 such that
xT(¢) E 8(N0 n G). Condition (iv) implies xT(¢) E 8No and the theorem is
proved. D
As a first example, we reconsider Equation (2.27) in the autonomous

case; that is,
(3.2)
where a and b are constants, a -=f. 0. If
V(¢) = - ¢4(o)
2a
+ jo
-r
¢6(e) de,
then
V(¢) = - [¢6(0) + 2b ¢3(0)¢3( -r) + ¢6( -r)].
a
Consequently, V is a Liapunov function on C if lbl S lal. If a < 0, then
V(¢);:::: ¢ 4 (0)/(2lal) and Corollary 3.1 implies the origin is stable and every
solution is bounded.
If a< 0, lbl < Ia I, then Sin Definition 3.1 is {¢ E C: ¢(0) = ¢( -r) =
0}. Obviously, the set M = {0} and Corollary 3.1 implies the solution x = 0
is globally asymptotically stable.
If a < 0, b = a, then S = {¢ : ¢(0) = -¢( -r)}. Therefore, M must

be the set of initial values of solutions satisfying x(t) = -x(t- r) for all
t E JR; that is, x = 0 and x(t) = c, a constant. But x(t) = -x(t- r) implies
c = 0 and again the origin is globally asymptotically stable.
If a< 0, b =-a, then S = {¢ E C: ¢(0) = ¢(-r)}. Using the same
type of reasoning, one concludes that
M = {constant functions on [-r, 0]}.
To obtain information about w(¢) for ¢ E C, suppose V(xt(¢)) -----t c as

t _____, oo. Then w(¢) ~v- 1 (c) n M and this latter set consists of a finite
number of constant functions since V (o:) is a polynomial of sixth degree in
the indeterminate o:. Since w(¢) is connected, this implies w(¢) is a single
point and each solution of Equation (3.2) approaches a constant.
If a > 0 and lbl < a (orb= a), then the set G = {¢ : V(¢) < 0}
is nonempty and positively invariant. As before the set M = {0}. Conse-
quently, Theorem 3.3 implies instability of the solution x = 0 and, in fact,
every solution starting in G is unbounded.
-1:
As a more sophisticated example, suppose n = 1 and
!(¢) = a( -e)g(¢(8)) de
where
(3.3) G(x) ~flax g(s) ds _____, oo as lxl _____, oo
(3.4) a(r) = 0, a(t);::: 0, a( t) ::; o, ii(t) ;::: 0, 0 ::; t ::; r,
are continuous. We consider the special case of Equation (3.1) given by
(3.5) x(t) = -lr 0

a( -B)g(x(t +B)) dB= -l~r a(t- u)g(x(u)) du.
Any solution of Equation (3.5) satisfies
(3.6) x(t) = a(O)g(x(t)) = -l~r a(t- u)g(x(u)) du
I:
or
1:
x(t) + a(O)g(x(t)) = -a(r) g(x(t +e)) de
(3.7)
+ 0
ii(-B)(i g(x(t+u))du)de.
Equation (3.6) is the model of a special type of circulating fuel nuclear

reactor where x represents the neutron density. It can also serve as a one-
dimensional model in viscoelasticity where x is the strain and a is the
relaxation function.
If we define V : C ----> 1R by the relation
V(¢) = G(¢(0))-- 110

2 -r
a(-B)[ 10 e
g(q)(s))ds] 2 dB,
then the derivative of V along solutions of Equation (3.5) is given by
V(q))
1
=2 a(r)[
10
-r g(q)(B))dB] 2 -
110
2 -r
ro
a(-B)[le g(q)(s))ds] 2 dB.
Since the hypotheses on a imply that V(q)) :::; 0, it follows from Corollary
3.1 that all solutions are bounded.
Let us now apply Theorem 3.1 to this equation. If for any s E [0, r],
we let
(3.8) Hs(¢) = [~ g(q)(B)) dB

then the set S of Theorem 3.1 is
S = {¢ E C: Hr(¢) = 0 if a(r) i= 0, Hs(¢) = 0 if a(s) i= 0}.

From Equation (3.7), the largest invariant set M of Equation (3.5) inS is
contained in the set in C generated by bounded solutions x of the ordinary
differential equation
(3.9) x + a(O)g(x) = 0
for which
Hr(Xt) = 0, t E (-oo,oo) if a(r) i= 0
Hs(Xt) = 0, tE(-oo,oo) ifa(s)i-0.
If a(r) i= 0, x satisfying Equation (3.9) is bounded and Hr(xt) = 0

for t E (-oo,oo), then x(t) = x(t- r) for all t. Therefore, x(t) =
kt +(a periodic function of period r) and boundedness of x implies x(t) =
x(t- r) for all t. If there is an So, a( so) i= 0, then there is an interval Iso
containing so such that a(s) i= 0 for s E Iso· If X satisfies Equation (3.9), is
bounded, and Hs(Xt) = 0 for -()() < t < oo, s E Iso, then x(t) is periodic
of period s for every s E ! 80 • Therefore, x(t) is constant and boundedness
of x implies x is constant.
Theorem 3.4. If System (3.5) satisfies Conditions (3.3) and (3.4) and g has
isolated zeros, then
(i) If there is an s such that a(s) > 0, then, for any¢ E C, thew-limit set
w(¢) of the orbit through¢ is an equilibrium point of Equation (3.5) ;
that is, a zero of g.
(ii) If a(s) = 0, a :f=. 0 (that is, a linear), then for any¢ E C, thew-limit
set w( ¢) of the orbit through ¢ is a single periodic orbit of period r
generated by a solution of Equation (3.9).
Proof. (i) The remarks preceding the theorem imply w(¢) contains only
equilibrium points and, thus, only zeros of g. Since w(¢) is connected and
the zeros of g are isolated, we have the result in Part (i).
(ii) Suppose a(s) = 0 and choose a(s) = (r- s)jr. If xis a solution of
Equation (3.9) of period r, then
-1t r-(t-u) g(x(u))du=lt r-(t-u) x(u)du

t-r r t-r r
= x(u) r- (t- u) It
- i t ~ x(u) du
r t-r t-r r
= x(t)
that is, x is a solution of Equation (3.5). From the remarks preceding the
theorem this implies M consists of the periodic solutions of Equation (3.9)
of period r.
We first prove that if w (¢) contains an equilibrium point c of Equation
(3.5), then w(¢) = c. We know that w(¢) is a closed connected set and
must be the union of r-periodic orbits of Equation (3.9). If cis not a local
minimum of G(x), then the nature of the orbits of Equation (3.9) in the
(x,±)-plane implies there can be nor-periodic orbits in w(¢) except c. If c
is a local minimum of G (x), then
V(¢)- G(c) = G(¢(0))- G(c) 1 10 [10 g(¢(s)) ds] 2 dB > 0

+-
2r -r ()
for¢ :f=. c in a neighborhood of c. Since V(xt(¢)) =constant for '1/J E w(¢),

it follows that w( ¢) = c.
Therefore, assume w( ¢) contains no constant solutions of Equation
(3.9). Since the solutions of Equation (3.9) must lie on the curves
j;2
2 + G(x) =constant,
it follows that any periodic orbit must be symmetrical with respect to the
x-axis. Let u(t, a) be a nonconstant periodic solution of Equation (3.9) of
least period p with u(O, a) = a and u(O, a) = 0. Then there is an integer
m such that mp = r. If there is an interval of periodic orbits in w(¢), then
pis independent of a in this interval. In fact, p = p(a) is continuous and,

therefore, m = m(a) is continuous. But m is an integer and, therefore, must
be independent of a. Also
10_)Jo{0
V(ut(a)) = V(uo(a))
1
= G(a) + 2r g(u(s))ds] 2 d0
= G(a) + - 10
1
2r it?(O, a) dO
+ - 10 u
-r
= G(a) 1 2 (0,a)d0
2mp -mp
= G(a) + 2 10
1 i£2 (0, a) dO
p -p
= G(a) + -11p/2 i£2 (0, a) dO

p 0
= G(a) +-21p/2 [G(a)- G(u(O, a))] dO

p 0
J21')'(e>)
= G(a) +- [G(a)- G(r)p1 2 dr
p "'
where /'(a) = u(a,p/2). On the other hand, the derivative of this latter
expression with respect to a is not zero. Therefore, V (Ut (a)) is not constant
for a in an interval. This implies w( ¢) is a single orbit and proves the
theorem. 0
It is also possible to analyze the stability and instability properties

of the limiting equilibrium point. If c is an equilibrium point in Equation
(3.5), let {3 = g'(c). If x = y+c in Equation (3.5), then the linear variational
equation for y is given by
(3.10) y(t) =-[or a(-O)g'(c)y(t+O)dO.

If g'(c) > 0, then we can use the preceding theorem to conclude that every
solution of Equation (3.10) approaches zero as t---. oo if either
(i) ii(s) ¢-0, or
(ii) if a(s) = (r- s)/r and
m27r
(g'(c))l/2 =f r
for all integers m.
If g'(c) < 0, then using the negative of the function V(¢) and instability
Theorem 3.3, one sees that the solution x = 0 of Equation (3.10) is unstable.
Later, we will see that these properties of the linear equation also hold true
for the nonlinear equation.
Consider now the system
(3.11) Ax(t) + Bx(t) =faT F(B)x(t- B) dB

where A, B, and F are symmetric n x n matrices and F is continuously
differentiable. Let
M = B -faT F(B) dB
and write Equation (3.11) as
x(t) = y(t)
(3.12)
Ay(t) = -Mx(t) +faT F(B)[x(t- B)- x(t)] dB.
Theorem 3.5.
(i) If A> 0, M > 0, F(B) ~ 0, F(B) :::; 0, and there is a B0 in [0, r] such
that F(Bo) < 0, then every solution of Equation (3.12) appmaches zem
as t ---t oo.
=
(ii) If A > 0, M > 0, P 0, and F > 0, then all solutions of Equation
(3.12) are bounded and thew-limit set of any solution must be generated
by periodic solutions of period r of the ordinary system
(3.13) x = y, Ay = -Bx.
(iii) If A> 0, M < 0, F(B) ~ 0, 0:::; B:::; r, F(B) :::; 0, 0:::; B:::; r, and there
is a B0 in [0, r] such that F(Bo) < 0, then the solution x = 0, y = 0 of
Equation (3.12) is unstable.
Pmof. Let ¢, '1/J be the initial values for x, y in Equation (3.12) and define
V(¢;,'1/J) = ~ cj;(OfM¢(0) + ~ '1/J(Of A'lj;(O)

+~ r [¢(-B)- ¢(0)f F(B)[¢( -B)- ¢(0)] dB.
2 lo
A few simple calculations yield
. 1 T
V(¢, '1/J) = - 2 [¢( -r)- ¢(0)] F(r)[¢( -r)- ¢(0)]
(3.14)
+ 21 lo r [¢(-B)- ¢(0)]T F(B)[¢( -B)- ¢(0)] dB:::; 0
if Condition (i), (ii), or (iii) is satisfied. If either Condition (i) or (iii) is
satisfied, then there is an interval Ie 0 containing B0 such that F(B) < 0
forB in leo· From Equation (3.14), V(¢, '1/J) = 0 implies¢( -B)- ¢(0) for
0 E le 0 • For a solution x, y to belong to the largest invariant set where

V(¢, '1/J) = 0, we must, therefore, have x(t- 0) = x(t) for all tin ( -oo, oo),
0 in le 0 • Therefore x(t) = c, a constant. From Equation (3.12), this implies
y = 0 and thus M c = 0. But c = 0 if M > 0 or < 0. Therefore, the largest
invariant set in the set where V( ¢, 'ljJ) = 0 is {(0, 0)}, the origin. If Condition
(i) is satisfied, then V satisfies Theorem 3.1 and Corollary 3.1 and every
solution of Equation (3.12) approaches (0, 0) as t--+ oo. If Condition (iii) is
satisfied, then V satisfies Theorem 3.3 and (0, 0) is unstable.
If Condition (ii) is satisfied, then F(O) = 0 and V(¢, '1/J) = 0 implies
¢( -r) = ¢( 0). Thus the largest invariant set in the set where V(¢, 'ljJ) = 0
consists of the r-periodic solutions of Equation (3.12). On the other hand, if
x(t), y(t) is an r-periodic solution of Equation (3.12) with F(O) =constant,
then the fact that M > 0 implies the integrals of x, y over the interval [0, r]
are zero. Therefore, x(t), y(t) must satisfy Equation (3.13). This proves the
theorem. 0
As a final example, consider the system of equations

x(t) = y(t)
(3.15)
y(t) = -~r y(t)- ~r sinx(t) +~
r
1°
-r
[cosx(t + O)]y(t + 0) dO
where a, b and r are positive constants. If t 2: r, solutions of this equation

satisfy
(3.16) x(t) + ~r x(t) + ~r sinx(t- r) = 0
which is a special case of the equation mentioned in the introduction for

the circummutation of plants.
If
then
V(xt, Yt) = j_)- 2ar y (t)

2 + bcosx(t + O)y(t)y(t + 0)
- 2ar y2(t + 0)] dO
~ 1°[-
-r
.!:_ y 2 (t)
2r
+ b sgn(y(t)y(t + 0) )~(t)y(t + 0)
- .!:_ y 2 (t + 0)] dO.
2r
If r < a/b, then the quadratic form in the integral is negative definite in
y(t), y(t + 0). For any b > 0, let
5.4 Razumikhin theorems 151
Then V(¢1, ¢2) is a Liapunov function on Ub and, for any (¢1, ¢2) E Ub,
the corresponding solution x(¢ 1, ¢ 2), y(¢ 1, ¢ 2) of System (3.15) satisfies
lx(¢1, ¢2)(t)1 < n/2 and IY 2(¢1, ¢2)(t)1 < 2bjr. We may therefore apply
Theorem 3.2. We know V = 0 if and only if ¢ 2(0) = 0. Therefore, any
solution that remains in M for all t E ( -oo, oo) must satisfy y(t) = 0 for
all t. But this implies x(t) = 0 for all tor x(t) =constant for all t. These
constants must be the zeros in sinx. Therefore, we have proved that r < ajb
implies every solution of Equation (3.16) approaches one of the constants
kn, k = 0, ±1, ±2, ... ' if the initial values are in ub.
5.4 Razumikhin theorems
In the previous section, sufficient conditions for stability of an RFDE were

given in terms of the rate of change of functionals along solutions. The use
of functionals is a natural generalization of the direct method of Liapunov
for ordinary differential equations. On the other hand, functions are much
simpler to use and it is natural to explore the possibility of using the rate of
change of a function on IRn to determine sufficient conditions for stability.
Results in this direction are generally referred to as theorems of Razumikhin
type.
If v : IRn --> IRn is a given positive definite continuously differentiable
function, then the derivative of v along an RFDE(f) is given by
(4.1) . ( ( )) - av(x(t)) !( )
V X t - ax Xt •
In order for v to be nonpositive for all initial data, one would be forced
to impose very severe restrictions on the function f (¢). In fact, the point
¢(0) must play a dominant role and, therefore, the results will apply only
to equations that are very similar to ordinary differential equations.
A few moments of reflection in the proper direction indicate that it
is unnecessary to require that Equation (4.1) be nonpositive for all initial
data in order to have stability. In fact, if a solution of the RFDE(f) begins
in a ball and is to leave this ball at some timet, then lxtl = lx(t)l; that is,
lx(t + Bl :S: lx(t)l for all B E [-r, 0]. Consequently, one need only consider
initial data satisfying this latter property. This is the basic idea exploited
in this section.
If V : IR x IRn --> IR is a continuous function, then V(t, ¢(0)), the
derivative of V along the solutions of an RFDE(f) is defined to be
. 1
V(t, ¢(0)) =lim sup -h [V(t + h, x(t, ¢)(t +h))- V(t, ¢(0))]
h---+0+
where x(t, ¢) is the solution of the RFDE(f) through (t, ¢).

Theorem 4.1. Suppose f: lR x C ___, lRn takes lR x (bounded sets of C) into

bounded sets of lRn and consider the RFDE(f). Suppose u, v, w : lR+ ___,
lR+ are continuous, nondecreasing functions, u(s), v(s) positive for s > 0,
u(O) = v(O) = 0, v strictly increasing. If there is a continuous function
V : lR x lRn ___, lR such that
(4.2) u(lxl) :S: V(t, x) :S: v(lxl),
and
(4.3) V(t, ¢(0)) :::; -w(l¢(0)1) if V(t + e, ¢(e)) :::; V(t, ¢(0)),
fore E [-r, OJ, then the solution x = 0 of the RFDE(f) is uniformly stable.
Proof. If
V(t, ¢) = sup V(t + e, ¢(e))
-r:'OO:'OO
for t E lR, ¢ E C, then there is a e0 in [-r, OJ such that V(t, ¢) = V(t +

eo, ¢(eo)) and either e0 = 0 or e0 < 0 and V(t + e, ¢(e))< V(t + e0 , ¢(e0 ))
if eo < e :::; 0. If eo < 0, then for h > 0 sufficiently small
V(t + h, Xt+h(t, ¢)) = V(t, ¢)

. .
e
and V = 0. If 0 = 0, then V :::; 0 by the Condition (4.3). Also, Relation
(4.2) implies that u(l¢(0)1) :S: V(t, ¢) :::; v(l¢1) fortE lR, ¢ E C. Theorem
2.1 implies the uniform stability of the solution x = 0 of the RFDE(f) and
the theorem is proved. D
Theorem 4.2. Suppose all of the conditions of Theorem 4.1 are satisfied
and in addition w(s) > 0 if s > 0. If there is a continuous nondecreasing
function p( s) > s for s > 0 such that Condition (4.3) is strengthened to
(4.4) V(t,¢(0)):::; -w(l¢(0)1) ifV(t+e,¢(e)) <p(V(t,¢(0))
fore E [-r, OJ, then the solution x = 0 of the RFDE(f) is uniformly asymp-
totically stable. If u( s) ___, oo as s ___, oo, then the solution x = 0 is also a
global attractor for the RFDE(f).
Proof. Theorem 4.1 implies uniform stability. To complete the proof of the
theorem, suppose b > 0, H > 0 are such that v(b) = u(H). Such numbers
always exist by our hypotheses on u and v. In fact, since v(O) = 0 and
0 < u(s) :::; v(s) for s > 0, one can preassign Hand then determine a b
such that the desired relation is satisfied. If u( s) ___, oo as s ___, oo, then one
can fix b arbitrarily and determine H such that v(b) = u(H). This remark
and the reasoning that follows will prove the uniform asymptotic stability
of x = 0 and the fact that x = 0 is a global attractor.
If v(b) = u(H), the same argument as in the proof of Theorem 4.1
shows that 1¢1 :::; b implies lxt(to, ¢)1 :S: H, V(t, x(to, ¢)(t)) < v(b) for
- r. Suppose 0 < 'Tf :::::; H is arbitrary. We need to show there is a

t ;:::: t 0
number t = t(ry, 8) such that for any t 0 2: 0 and 1¢1 :::::; 8, the solution x(to, ¢)
of the RFDE(f) satisfies lxt(t 0 ,¢)1 :::::; ry, t 2: t 0 + t + r. This will be true
if we show that V(t,x(t 0 ,¢)(t)):::::; u(ry), fort 2: t 0 +t. In the remainder of
this proof, we let x(t) = x(t 0 , ¢)(t).
From the properties of the function p( s), there is a number a > 0
such that p( s) - s > a for u( 'Tf) :::::; s :::::; v( 8). Let N be the first positive
integer such that u(ry) +Na 2: v(8), and let 'Y = infv-l(u('l)):::;s:::;Hw(s) and
T = Nv(8)h.
We now show that V(t, x(t)) : : :; u(ry) for all t 2: t 0 + T. First, we show
that V(t, x(t)) :::::; u(ry) + (N -1)a fort 2: to+ (v(8)h). If u(ry) + (N -1)a <
V(t,x(t)) for to : : :; t <to+ (v(8)h), then, since V(t,x(t)) :::::; v(8) for all
t 2: t 0 - r, it follows that
p(V(t, x(t))) > V(t, x(t)) +a 2: u(ry) + Na 2: v(8) 2: V(t + 0, x(t + 0)),
v(8)
to :::::; t :::::; to +- , 0 E [-r, OJ.
'Y
Hypothesis (4.4) implies
V(t, x(t)):::::; -w(lx(t)l):::::; -"(

for to :::::; t : : :; to + (v (8) h). Consequently,
V(t, x(t)):::::; V(to, x(to))- 'Y(t- to):::::; v(8)- 'Y(t- to)

on this same interval. The positive property (4.2) of V implies that
V(t, x(t)) : : :; u(ry) + (N - 1)a at t 1 = t 0 + v(8)h. But this implies
V(t,x(t)) :::::; u(ry) + (N- 1)a for all t 2: t 0 + (v(8)h), since V(t,x(t))
is negative by Condition (4.4) when V(t,x(t)) = u(ry) + (N -1)a.
Now let tj = jv(8)h, j = 1, 2, ... , N, t 0 = 0, and assume that for
some integer k 2: 1, in the interval tk- 1 - r:::::; t- t 0 :::::; tk, we have
u(ry) + (N- k)a:::::; V(t, x(t)) : : :; u(ry) + (N- k + 1)a.
By the same type of reasoning, we have
V(t, x(t)):::::; -"(, tk-1:::::; t- to:::::; tk

and
V(t, x(t)):::::; V(to + tk-1, x(to + tk-1))- 'Y(t- to- tk-d
: : :; v(8)- 'Y(t- to- tk-1):::::; 0
if t- to- tk-1 2: v(8)h. Consequently,
+ tk, x(to + tk)) :::::; u(ry) + (N- k)a,

V(to
and, finally, V(t, x(t)) :::::; u(ry) + (N- k)a for all t 2: t 0 + tk. This completes
the induction and we have V(t, x(t)) :::::; u(ry) for all t 2: t 0 + Nv(8)h. This
proves Theorem 4.2. 0
As a first example, consider the equation
(4.5) i:(t) = -a(t)x(t)- b(t)x(t- r 0 (t))
where a, b, and r0 are bounded continuous functions on IR with lb(t) I :::; a(t),
0:::; r 0 (t) :::; r, for all t E IR. If V(x) = x 2 /2, then
V(x(t)) = -a(t)x 2 (t)- b(t)x(t)x(t- r 0 (t))

:::; -a(t)x 2 (t) + lb(t)llx(t)llx(t- ro(t))l
:::; -[a(t) -lb(t)l]x 2 (t):::; 0
if lx(t)l 2: lx(t-ro(t))l. Since V(x) = x 2 /2, we have shown that V(x(t)):::; 0

if V(x(t)) 2: V(x(t- r 0 (t))). Theorem 4.1 implies the solution x = 0 of
Equation (4.5) is uniformly stable.
If, in addition, a(t) 2: 8 > 0, and there is a k, 0 :::; k < 1 such that
lb(t)l :::; k8, then the solution x = 0 of Equation (4.5) is uniformly asymp-
totically stable. In fact, choose p( s) = q2 s for some constant q > 1. If
V(x) = x 2 /2 as before, then
V(x(t)):::; -(1- qk) 8x 2 (t)
if p(V(x(t))) > V(x(t- ro(t))). Since k < 1, there is a q > 1 such that
1 - qk > 0 and Theorem 4.2 implies the uniform asymptotic stability of
the solution x = 0. This is an improvement over the results obtained with
functionals for Equation (2.9) since the delay can be an arbitrary bounded
continuous function.
If we use the same V (x), then a similar argument shows that the zero
solution of
n
i:(t) = -a(t)x(t)- 2:= bj(t)x(t- rj(t))
j=l
is uniformly asymptotically stable for all bounded continuous functions

a, bj, rj if a(t) 2: 8 > 0, L-7=
1 lbj(t)l < k8, 0 < k < 1, 0 :::; rj(t) :::; r
for all t E IR.
As an example of a nonlinear problem, consider the first-order equation
(4.6) i:(t) = f(x(t- "'f(t)), t), 0 :::; "'((t) :::; r, f(O, t) =0

where "Y(t) is a continuous function oft and f(x, t) is a continuous function
of x, t for t 2: 0, -oo < x < oo, has a continuous partial derivative such
that iof(x, t)joxi < L, t 2: 0, -oo < x < oo. For t 2: 2r we can rewrite
Equation (4.6) as
i:(t) f(x(t), t)- [f(x(t), t)- f(x(t- "Y(t)), t))
-it
=
(4.7)
= f(x(t), t) ~~ (x(B), t)f(x(B- "'!(B)), B) dB.
t-"((t) ux
For V(¢) = ¢ 2 (0), we have
V(xt) = 2x(t)f(x(t), t) - 21t x(t/1

8 (x(O), t)f(x(O- !'(0)), 0) dO
t-1(t) X
:::; 2x(t)f(x(t), t) + 2L2 1t lx(t)x(O- 1'(0))1 dO

t-'Y(t)
:::; 2x(t)f(x(t), t) + 2L 2'Y(t)lx(t)llxt-1 (t) I, t ~ 2r.
Consequently, if q > 1 is fixed and we consider the set of all x(t) such that
0:::; ~:::; 2r,
then
V(xt):::; 2[!(:~~, t) + L 2'Y(t)q]x 2 (t) < -2J.Lx2 (t)
+ L 2 'Y(t)q < -J.L < 0. For J.L > 0, t ~ 0, x E ( -oo, oo), and
if (f(x, t)jx)
p(s) = q2 s, Theorem 4.2 implies the origin is uniformly asymptotically
stable and a global attractor.
Let us now consider the linear equation
(4.8) ±(t) = Ax(t) + Bx(t- T)

where A and Bare matrices (A constant, B may be a bounded continuous
matrix function) and T = T(t), 0:::; T(t) :::; r, is continuous. lfV(x) = xT Dx,
where D is positive definite, then
V = x(tf (AT D + DA)x(t) + 2x(t)T DBx(t- T).
If there are constants q > 1, 17 > 0 such that
V(x(~)) < qV(x(t)), t- r:::; ~:::; t implies V:::; -rylx(tW,
then Theorem 4.2 implies the solution x = 0 of Equation (4.8) is uniformly

The difficulty in obtaining results along this line arises from attempting
to estimate V for the restricted class of initial curves satisfying V(x(~)) <
qV(x(t)), t- r :::; ~ :::; t. Furthermore, there are numerous directions in
which one may proceed. In particular, one may wish to obtain stability
conditions that are independent of T or conditions that depend on T. In the
first case, one must obviously have the zero solution of
±(t) = (A+ B)x(t)

asymptotically stable. By an appropriate change of coordinates, one can
take V (x) = xT x and be assured that V along the trajectories of this
ordinary differential equation is a negative definite function. In these new
coordinates, V along solutions of Equation (4.8) is
V= x(tf[(A +B)+ (A+ B)T]x(t) + 2x(tf Bx(t- r)- x(t)T (B + BT)x(t)
and one can estimate V along curves satisfying V(x(~)) < qV(x(t)) (or
equivalently, lx(~)l < qlx(t)l), q > 1, t- T ~ ~ ~ t in the following way.
Since (A+ B)+ (AT+ BT) is negative definite, there is a A> 0 such that
V~ + 2qiBIIx(tW + IB + BTIIx(tW
-Aix(t)l 2
=-[A- 2qiBI-IB + BTIJix(tW.
Consequently, if 2qiBI + IB +BTl < A, then the solution x = 0 of Equation

(4.8) is uniformly asymptotically stable. Razumikhin [1] has carried out
this type of procedure for a second-order system
x(t) + ax(t) + bx(t) + cx(t- r) = 0

to obtain estimates on a, b, and c ensuring asymptotic stability independent
ofT.
To obtain estimates that depend on the delay function T, one may
proceed in the following manner. For simplicity in notation, let us consider
the initial time to be zero and let x(t), t ~ 0, be the solution of Equation
(4.8) through (0, ¢). Since x(t) is continuously differentiable fort ~ 0 one
can write
x(t- r) = x(t) - /_: x(t + 0) dO
= x(t)- /_: [Ax(t + 0) + Bx(t- T + 0)] dO
fort~ T. If we return to Equation (4.8) using this expression for x(t- r),
l:
we obtain the equation
(4.9) x(t) =(A+ B)x(t)- B [Ax(t + 0) + Bx(t- T + 0)] dO
for arbitrary continuous initial data on [-2r, 0]. If the zero solution of Equa-
tion (4.9) is asymptotically stable, then the zero solution of Equation (4.8)
is asymptotically stable since Equation (4.8) is a special case of Equation
(4.9) with continuous initial data'¢ on [-2r,O] given by '¢(s) arbitrary
for s E [-2r, -r- r(O)], '¢(s) = cp(s + r(O)), -r- r(O) ~ s ~ -r(O),
and '¢(s) = x(t + s), -r(O) ~ s ~ 0, where x is the solution of
x(t) = Ax(t) + Bx(t- r) through (0, ¢).
As an example, consider the equation
x(t) = -bx(t- r), r > 0,
and the auxiliary problem on [-2r, 0] given by

±(t) = -bx(t)- b2 1t-r x(s) ds.

t-2r
If V(x) = x 2 /2, then, for any constant q > 1,
V = -bx 2 (t)- b2 1t-r x(t)x(s) ds

t-2r
:s; -b(1- qbr)x 2 (t)
if V(x(~)) < q2 V(x(t)), t- 2r :s; ~ :s; t. Consequently, if br < 1, then there

is a q > 1 such that qbr < 1 and Theorem 4.2 implies asymptotic stability.
There are many ways to extend the ideas of the last computation to
obtain more precise information about stability for the RFDE(f). For any
integer k 2:0, one can artificially interpret this equation on C([-r,O],IRn)
as an equation on C([-(k + 1)r, 0], IRn). Of course, arbitrary initial condi-
tions in this larger space need not be considered from the point of view of
stability in the original equation. One should restrict consideration to initial
data that are related in some way to the original equation. An obvious way
is the following. Let¢> E 0([-r,O],IRn) and let x(¢>) be the solution of the
RFDE(f) through¢> at initial time zero. For initial data 'ljJ E 0[-(k+ 1)r, OJ
for the artificial problem, take '1/J(s) = x(<f>)(kr + s), -(k + 1)r :s; s :s; 0.
Such considerations will take into account integration of the equation over
k delay intervals.
One can obviously apply analogues of Theorems 4.1 and 4.2 to obtain
information about the equation. Rather than go into detail, let us consider
specifically what could be done to obtain stability of solutions.
Suppose there is a continuous function V : IRn --+ IR such that
(4.10) u(lxl) :s; V(x) :s; v(lxl)

where u and v satisfy the same properties as in Theorem 4.1. If x(¢>) is the
solution of the RFDE(f) through ¢> E 0([0, r], IRn), let
(4.11) V(¢) = sup V(x(¢)(s)).

-r~s~kr
:s; 0 along the solutions of the ~FDE(f), then the solution x = 0 is

If V (¢>)
stable. Therefore, we need only make V(¢>) :s; 0. Using the same argument
as in the proof of Theorem 4.1, this condition will be satisfied if one can
show that the set
(4.12) P = {1> E 0: V(¢>) = V(x(<f>)(kr)) > 0; V(x(<f>)(kr)) > 0}
is empty.
Let us apply the previous remarks to the scalar equation
(4.13) x(t) = -bx(t- r), b>O

with k = 2 and V(x) = x 2 /2. If qi in Expression (4.12) is not empt:y, then

there is an E > 0 and a¢ E C such that lx(¢)(2r)l = E, V(¢) = V(x(¢)(2r)),
and
1 d
2 dt x 2 (¢)(2r) = -bx(¢)(2r)x(¢)(r) > 0.
Without loss of generality, we may assume x(¢)(2r) = E and the inequality
then implies x(¢)(r) < 0. Consequently, we will obtain a contradiction if
we show that
P ~f sup{x(¢)(2r) : ¢ E C, lx(¢)(t)l S E,
(4.14)
-r s t s 2r, x(¢)(r) < 0} <E.
The variational problem (4.14) imposes a restriction of the magnitude of

the parameter b. Rather than integrate the equation for the solution up
to t = 2r, we can take advantage of the equation to obtain a variational
problem that only involves integration over an interval of length r. In fact
if 1¢1 < E, then the equation implies 1±(¢)(t)1 s bE for 0 s t s rand then
P S sup{x(¢)(r): lxt(¢)1 S E, 1±(¢)(t)1 S bE, 0 S t S r, ¢(0) < 0}.
Let W:L be the subset of C consisting of those functions that are absolutely
continuous and have an essentially bounded derivative on [-r, OJ and define
P(y) = sup{x(~)(r): ~ E W:L, 1~1 S E, 1~(0)1 S bE a.e. in 0, ~(0) = y}.

It is clear that P s sup{P(y) : -E < y < 0}. This latter problem involves
integration only over [0, r]. For any y E lR, -E < y < 0, there is a function
'such that x(,, r) = P(y). In fact one can take' as the function
-E 0 E [-r, -(E + y)jbE],

{
'(O) = y+bE0 0 E [-(E+y)/bE,O].
For this' and br > 1,
- 1 2
x(~)(r) = y + bEr- 2E (E + y)
1
P s sup{P(y): -E < y < 0} = bEr- 2 = E(br- 2).
- E
For br < 1, a similar computation gives P s E(1- br). Therefore, if 0 < b <
3/2r, P < E and we have stability of the solution x = 0. This is a significant
improvement over the estimates obtained before.
Our next result is concerned with uniform ultimate boundedness.
Theorem 4.3. Suppose f : 1R x C ---. JRn takes 1R x (bounded sets of C) into

bounded sets ofJRn and consider the RFDE(f). Suppose u, v, w: JR+ ---. JR+
are continuous nonincreasing functions, u(s) ---. oo as s ---. oo. If there
is a continuous V : 1R x JRn ---. JR, a continuous nondecreasing function
p: JR+ ---.JR+, p(s) > s for s > 0, and a constant H 2::0 such that
(4.15) u(lxl)-::; V(t, x)-::; v(lxl), t E JR, X E JRn
and
(4.16) V(t, ¢)-::; -w(l¢(0)1)

if
1¢(0) I 2:: H, V(t + B, ¢(B)) < p(V(t, ¢(0)), BE [-r, o],
then the solutions of the RFDE(f) are uniformly ultimately bounded.
The proof of this result will not be given since it is essentially a repe-
tition of the arguments used in the proof of Theorems 4.1 and 4.2.
In the applications one often needs a generalization of this result. More
specifically, one may be able to verify Inequality (4.16) only for some coor-
dinate, say ¢ 1 , of the function ¢; that is, one may be able to verify that
(4.17) V(t, ¢)-::; -w(l¢1(0)1)

if
V(t + B, ¢1 (B), Xz, ... 'Xn) < p(V(t, ¢1 (0), Xz, ... 'Xn))
for
BE [-r, OJ, Xj E JR.
In this case, one can prove that the first coordinate of the solutions of
the RFDE(f) is uniformly ultimately bounded. The proof of this result is
essentially the same as before.
One can also prove results by replacing l¢ 1(0) I 2:: H by ¢1(0) 2:: H with
the conclusion being that there is an a > 0 such that the first coordinate
satisfies x 1(t) -::; a for all t 2:: a.
As a first example of the application of Theorem 4.3, consider the
second-order system
(4.18)
±(t) = y(t)
y(t) = -tP(t, y(t))- f(x(t)) + p(t) + 1: g(x(t + B))y(t +B) dB.
If g(x) = df(x)/dx, then System (4.18) includes the second-order scalar

equation
(4.19) x(t) + tP(t, ±(t)) + f(x(t- r)) = p(t).

We make the following assumptions on System (4.18):

(i) lJ> : IR? ---+ IR is continuous, P takes IR x (bounded sets of IR) into
bounded sets and there are constants a> 0, H > 0, such that
l/>(t,y) 0
-->a>
y
for IYI2 H,
(ii) f: IR---+ IRs continuous and f(x) sgn x---+ oo as lxl ---+ oo,
(iii) p: IR---+ IRis continuous and bounded by k,
(iv) g: IR---+ IRis continuous and lg(x)l :S L for all x E IR,
(v) Lr <a.
Of course, it is always assumed that a uniqueness result holds for the
solutions of System (4.18).
Under the hypotheses, we will show that the solutions of System (4.18)
are uniformly ultimately bounded. If, in addition, there is a w > 0 such that
l/>(t+w, y) = l/>(t, y), p(t+w) = p(t) for all t E IR, then Equation (4.18) has
an w-periodic solution. This latter remark is a consequence of the uniform
ultimate boundedness and Theorem 6.2 of Section 4.6.
If V(x, y) = F(x) + y2/2, F(x) = J;
f(s) ds and q > 1, qLr <a, then
V(x(t), y(t)) = -y(t)ll>(t, y(t)) + y(t)p(t)

+ y(t) [or g(x(t + ()))y(t + ()) d()
s; -(a- qLr)y2(t) + ly(t)lk
if ly(t)l 2 Hand ly(t+())l :S qly(t)l. By choosing H1 2 H appropriately, one
obtains a positive constant fJ such that V(x(t), y(t)) :S -{ly 2 (t) for ly(t)l 2
H1 and ly(t +())I :S qly(t)l. Therefore, the remarks following Theorem 4.3
imply the y coordinate of the solutions is uniformly ultimately bounded by
a constant c.
If IYI :S c and V1(x,y) = V(x,y) +y, then there is a constant k1 such
that
V1 (x(t), y(t)) =
+ 1:
V(x(t), y(t))- P(t, y(t))- f(x(t))
:S - f(x(t)) + k1.
g(x(t + e))y(t + ()) d()
+ p(t)
One can choose a constant b > 0 such that
V1(x(t),y(t)) < -1 if x(t) 2 b.
Consequently, there is an a > 0 such that the x coordinate of the solu-

tions satisfies x(t) :S a, in the strip IYI :S c. Using the function V2(x, y) =
V(x,y)- y one obtains the x coordinate that satisfies -a :S x(t) in the
strip IYI ~ c. This clearly implies the uniform ultimate boundedness of the
solutions of System (4.18).
It is natural to ask if this result is valid without any restriction on the
delay. Consider the linear equation
x(t) + ax(t) + b2 x(t- 2r) = 0.

The characteristic equation is
For a = 0, it is easy to verify that 2rb > 1r implies there are at least two
roots with positive real parts. Therefore, for a > 0 sufficiently small, there
will be unbounded solutions of the linear equation and one does not have
uniform ultimate boundedness.
x(t) = y(t)
(4.20)
y(t) = -ay(t- r)- f(x(t)) + p(t)
with a> 0, f, p satisfying (ii) and (iii). If V(x, y) = F(x) + y2 /2, F(x) =
J; f(s) ds, q > 1, then
V(x(t), y(t)) ~ -ay(t)y(t- r) + y(t)p(t)

~ -aqy2 (t) + iy(t)ik
~ -p?y(t)
if iy(t)i 2:: Hand iy(t-r)i ~ qly(t)i. Therefore, we obtain uniform ultimate

boundedness of the y-coordinate of the solutions of (4.20) without any re-
striction on the delay. Using arguments on the x-coordinate of the solutions
as in the previous example, one obtains uniform ultimate boundedness of
the solutions of (4.20) only under the hypotheses a> 0 and (ii), (iii).
The proof of Lemma 1.1 had its origin in the work on dissipative processes
of Hale, LaSalle, and Slemrod [1] and Hale and Lopes [1]. The result was
independently discovered by Ize [1]. Lemma 1.2 is implicitly contained in
Billotti and LaSalle [1] and was independently discovered by Pavel [1] (see
also Yoshizawa [2]).
Krasovskii [1, p. 151 ff.] proved the asymptotic stability in Theorem 2.1.
The proof in the test is due to Yoshizawa [1]. In Theorem 2.1 (and Theorem
4.1), we have required that f takes IR x (bounded sets of C) into bounded
sets of IRn. Burton [1, 2] has shown that it suffices to require that f is a
completely continuous map. However, Makay [1] has shown that the condi-
tions cannot be weakened if one only assumes that the estimates on V and
V are required to be satisfied along the solutions of the differential equation
(this is the only requirement in the proofs).
Lemma 2.1 is due to Repin [1] and Datko [1]. The method of con-
structing the Liapunov functional in Theorem 2.2 is due to Huang [1]. For
the special case of (2.1), Infante and Castelan [1] earlier had proved that
a quadratic functional exists as in Theorem 2.2 by approximating the dif-
ference differential equation by a system of ordinary differential equations,
using the Liapunov theorem for this approximate equation and then taking
a limit. Mansurov [1] has also considered difference approximations to ob-
tain stability. A special case of the instability Theorem 2.2 was proved by
Shimanov [1]. The material in Section 5.3 on the stability of autonomous
systems is based on Hale [2], taking into account the improvements by
LaSalle [2] and Onuchic [1]. Much more general results for compact and
uniform processes have been given by Dafermos [3].
Example (3.2) is due to LaSalle. Example (3.5) and a special case of
Theorem 3.4 was originally given by Levin and Nohel [1] by different meth-
ods. Under the assumption (i) in Theorem 3.4, it follows from Theorem 5.2
of Chapter 4 that there is a compact connected global attractor. Further-
more, the system is gradient-like with hyperbolic equilibrium points and
the attractor is the union of the unstable manifolds (see Chapter 10 for
the definition) of the equilibrium points (see Hale [23], for example). The
unstable points correspond to the zeros of g for which the derivative at the
point is negative and the dimension of the unstable manifold is one. Hale
and Rybakowski [1] have discussed the types of orbit connections between
equilibrium points and, surprisingly, it is shown that these connections do
not always preserve the natural order of the real numbers. By using the
recent results on convergence of Hale and Raugel [1], we remark that the
same conclusion as in part (i) of Theorem 3.4 (that is, convergence to a
single equilibrium point) can be shown to be valid without the hypothesis
that the zeros of g are isolated.
Onuchi [1] has an interesting instability theory for Equation (3.5). The-
orem 3.5 was first proved by Hale [2] and motivated by Volterra [2].
The Liapunov functional for Equations (3.15) is due to A. Somolinos.
Equation (3.16) often is referred to as the sunflower equation because of
its origin in the circummutation of plants. It is shown in the text that,
for r < ajb, the system is gradient-like. If we consider the flow defined
by this equation on the space X = C([-r, OJ, S 1 x IR), then there is a
compact global attractor from Theorem 5.2 of Chapter 4. Since the equi-
librium points are hyperbolic, it follows that the attractor is the union of
the unstable manifolds, and it is easy to check that these have dimension 1.
Since there is a Liapunov functional, only two equilibrium points, and the
attractor is connected, it follows that the attractor is homeomorphic to S 1 .
For more details and further properties, see Lizano [1]. For some interesting
stability problems in car following, see Harband [1].

Theorems of the type given in Theorems 4.1 and 4.2 originated with
Razumikhin [1, 2] with versions also being stated in the book of Krasovskii
[1, p. 157 ff.]. The proof of uniform asymptotic stability was first given by
Driver [2]. The global nature of this theorem was independently discovered
by Seifert [1]. Example (4.6) is due to Krasovskii [1, p. 174], the reduction of
the stability problem for Equation (4.13) to an optimization problem is due
to Barnea [1]. Other interesting examples with several delays are in Barnea
[1], Bailey and Williams [1], and Noonburg [1]. An instability result similar
to Theorem 4.4 is also contained in Barnea [1]. Theorem 4.3 on ultimate
boundedness was first stated by Lopes [1, 2, 3] for more general neutral
functional differential equations. Example (4.18) was first considered by
Yoshizawa [1, p. 208] with a more restrictive hypotheses on the delay.
As indicated in the main text, it is possible to extend Razumikhin
Theorems 4.1 and 4.2 by taking into consideration the value of the solu-
tion over a few lag intervals. An even more general extension has been
given by J. Kato [1] based on comparison principles (see earlier versions in
Laksmikantham and Leela [1]). We now summarize some of the results of
Kato [1].
Consider a scalar ordinary differential equation
u(t) = U(t, u(t)),
where U : JR? _____, IR is continuous and denote by r(t, s, o:) the maximal
solution of this equation through s, o: fort 2 sand by l(t, s, o:) the maximal
solution through s, o: for t ::; s.
Theorem 5.1. If v : [0' - r, oo) _____, IR is a continuous function whose upper

right-hand derivative v satisfies
v(t)::; U(t, v(t)) fort 2 0', if v(s)::; l(s, t, v(t)), s E [t- r, t],
then
v(t) ::; r(t, 0', o:) fort 2 0'
where o: is chosen so that v(s)::; l(s,O',o:), s E [0'- r,O'].
Theorem 5.1 can be shown to include Theorem 4.2 by noting an ap-

propriate function U(t, u) is given by the function
. 1 2 2
mm{ 3r u,p(3' u)- 3' u,w(u)}.
Kato gives applications of this more general result to the theory of

asymptotic stability, stability with respect to delays, and the following in-
teresting result of Yorke [2].
Theorem 5.2. Suppose f : lR x C --+ lR is continuous and satisfies

(5.1) -aM(¢)::; f(t,¢)::; aM(-¢)
for some constant a~ 0, where M(¢) = max{O, SUP-r<B<O ¢(0)}. Then the
following conclusions hold: - -
(i) If ar ::; ~, the zero solution of the RFDE(f) is uniformly stable.
(ii) If ar < ~' then the zero solution of the RFDE(f) is uniformly asymp-
totically stable if for any sequence tm --+ oo and any sequence {<Pm}
converging to a nonzero constant function, the sequence f(tm, <Pm) does
not converge to zero.
A special case of the last result is the linear equation

(5.2) x(t) = -bx(t- r(t))
where b ~ 0 and r(t) is a continuous function satisfying 0 ::; r(t) ::; r. If
br ::; ~, then the solution x = 0 is uniformly stable and if br < ~, the
solution x = 0 is uniformly asymptotically stable. Notice the constant ~ is
the same as the one obtained in the analysis of Equation (4.13) by using
the knowledge of the solution over [0, 2r]. The original proof of Yorke also
uses knowledge of the actual solution, but over the larger interval [0, 3r].
The upper bound ~ for br has been previously shown to be the best
possible for Equation (5.2) and it is, therefore, quite remarkable that the
same upper bound can be obtained for the nonlinear problem as in Theorem
5.2. In fact, it was shown by Mishkis [1] and Lillo [2] that for br = ~, there
are Equations (5.2) that have periodic solutions and if br > ~, there are
equations with unbounded solutions. The paper of Lillo [2] also contains
further remarkable properties on the asymptotic behavior of general linear
scalar systems that satisfy the conditions of Theorem 5.2. For other results
along this line, see Halanay and Yorke [1].
Generalizations of Theorem 5.2 to n-dimensions have been given by
Grossman [1] (see also Kato [1]).
Grimmer and Seifert [1] have applied Razumikhin arguments to discuss
the asymptotic behavior of the solutions of integrodifferential equations of
the type
x(t) = Ax(t) +lot B(t, s){x(s) + g1(x)(s)}ds + g2(x)(t) + f(t)

where gj(x)(s) are certain higher-order functionals of x, depending on val-
ues of x( 0) for 0 ::; t.
Seifert [2, 3] has also used arguments in the spirit of Razumikhin to
prove the existence of a solution of an RFDE(f) that has the range of f in
a closed convex set in lRn.
It is certainly worthwhile to discuss the advantages and disadvantages
of Liapunov functionals and Razumikhin arguments. As we have seen by
examples, it appears to be easier generally to use the Razumikhin theorems

for stability of the trivial solution than to construct Liapunov functionals.
On the other hand, if the limit of a solution is more complicated than a
point, it is not clear how one can take advantage of arguments in the spirit
of Razumikhin. This becomes very apparent if we consider autonomous
equations. For example, if we consider a scalar equation that has a stable
nontrivial periodic solution, then the trace of this solution in IR will be an
interval that could contain a constant solution. How can one possibly detect
the asymptotic behavior of a solution by only observing what happens
in IR? The Razumikhin arguments use mainly IR whereas the Liapunov
functionals take advantage of C. This is very apparent in the example of
Levin and Nohel (see Equation (3.5)).
In Section 5.2, there is the obvious omission of a theorem on ultimate
boundedness using Liapunov functionals. A result in this direction is con-
tained in Yoshizawa [1, p. 206]. The reason for not stating this particular
result in the text is that the proof imposes a restriction on the size of the
delay, in addition to the usual conditions on Liapunov functionals. To ob-
tain a result that depends only on the rate of change of certain functionals
seems to put severe restrictions on the form of the RFDE. On the other
hand, Theorem 4.3 on ultimate boundedness does not have such additional
restrictions on the delays.
Another area of investigation that uses ideas in the same spirit as
the ones in the Razumikhin theorems is the method of guiding functions
introduced by Krasnoselskii et al. around 1958. The following definition is
due to Mawhin [1, 2].
Definition 5.1. A continuously differential function V : IRn -t IRis a guiding

function for an RFDE(f) if there is a p > 0 such that
(5.3) av~(o)) f(t,¢) > o
for every ¢ E C satisfying 1¢(0)1 2: p and IV(¢(0))1 > IV(¢(0))1, for 0 E

[-r,OJ.
Krasnoselskii was interested in determining periodic solutions of peri-

odic ordinary and differential difference equations. His original definition re-
quired that Inequality (5.3) be satisfied for every¢ E C such that 1¢(0) I 2: p.
This imposed very severe restrictions on the functional f since it required
that ¢(0) be the governing factor for the asymptotic behavior of f. In ad-
dition, the proof of the results on the existence of periodic solutions by
Krasnoselskii et al. was very complicated and the extension of the proof to
general RFDE was not clear. Mawhin [1] found a different proof that was
much simpler. With this simpler proof, it became clear that the guiding
function could be as general as given in Definition 5.1.
The topological method of Wazewski using ingress and egress points

on the boundary of an open set is one of the most important tools in the
study of the asymptotic behavior of the solutions of ordinary differential
equations. By insisting that the direction of the flow in IRn at time t is
determined in a very strong way by the value of the solution x at the same
time t, one can give an extension of Wazewski's principle (see Onuchic [2]
and Mikolajska [1]).
The general principle ofWazewski has been given by Rybakowski [1, 2,
3] together with its connection to a homotopy index and the Conley index.
Due to lack of space, many areas of the theory of stability of RFDE
have not been mentioned. For example, no converse theorems have been
given concerning the existence of Liapunov functionals for stable systems.
These results are certainly important and, as remarked earlier, were the
original motivation for Krasovskii [1] to begin the vigorous development
of the theory of RFDE in the state space C as opposed to IRn. If one
understands well the method of obtaining converse theorems for ordinary
differential equations, it is not too difficult to extend the results of RFDE.
The basic ideas are contained in the books of Halanay [1], Krasovskii [1],
and Yoshizawa [1] and the paper of Hale [3]. The converse theorems have
applications to the implications of stability; for example, stability under
constantly acting disturbances (see Corduneanu [2]), the behavior of the
solutions of perturbations of stable systems (see Laksmikantham and Leela
[1] and Onuchic [3]), and the absolute stability of systems of the form
x(t) = Ax(t) + Bx(t- 7) + Q(iJ(t))

(5.4)
IJ =ex
(see Halanay [1], and Somolinos [1]). The book of Razvan [1] contains an
excellent bibliography on Equation (5.4). Halanay [2] also has used the
converse theorems to discuss Equation (5.4) with nonhomogeneous almost-
periodic forcing functions. The book of Martynyuk [1] is devoted entirely
to the stability theory of functional differential equations.
For a more complete discussion of the recent developments in stability
theory and the general theory of RFDE, see Burton [2, 3] Grippenberg,
Londen, and Staffans [1], and Hino, Murakami, and Naito [1].
6
General linear systems
This chapter is devoted to the development of linear RFDE, including the

variation-of-constants formula and the formal adjoint of a linear system.
For an arbitrary two-point boundary-value problem, it is then shown
that there is a two-point boundary-value problem for the formal adjoint
equation that fulfills the conditions of the usual Fredholm alternative. Also,
relations between various types of stability for linear systems are given.
The chapter is completed with a discussion about perturbed linear
systems.
6.1 Resolvents and exponential estimates
For (a,¢) Em. x C, consider the linear system
:i:(t) = L(t)xt + h(t), t ~a,

(1.1)
x,. = 4>
where hE .Cioc([a,oo),m.n), the space of functions from [a,oo) into m.n

that are Lebesgue integrable on each compact set of [a, oo). Also, assume
that there is ann x n matrix function ry(t, 0), measurable in (t, 0) Em. x m.,
normalized so that
ry(t, 0) = 0 for 0 ~ 0, ry(t, 0) = ry(t, -r) for 0:::; -r,

ry(t, 0) is continuous from the left in 0 on ( -r, 0) and has bounded variation
in 0 on [-r,OJ for each t. Further, there is an mE .Cioc((-oo,oo),m.) such
that
(1.2) Var[-r,O] ry(t, ·) ::=; m(t)

and the linear mapping L(t) : C---+ m.n is given by
(1.3) L(t)¢ =I: d[ry(t, 0)]¢(0)
168 6. General linear systems
for all t E ( -oo, oo) and 4> E C. Obviously, the norm of L(t) satisfies
JL(t)4>l ~ m(t)J4>J.
Theorem 1.1. Suppose these conditions on rJ and J.L are satisfied. For any
given u E lR, 4> E C([-r,O],lRn), and hE .Cioc([u,oo),lRn), there exists a
unique function x(u, 4>) defined and continuous on [u- r, oo) that satisfies
System (1.1) on [u,oo).
Proof. Condition (1.2) implies the Caratheodory conditions are satisfied.
Therefore, we have local existence from Chapter 2. Local uniqueness follows
as in Chapter 1 since L(t) is Lipschitzian. To prove global existence, let x
be a noncontinuable solution of System (1.1) on [u- r, b). Integration of
System (1.1) yields
Jx(t)J ~ 14>(0)1 + lt m(s)lxsl ds + lt lh(s)l ds
for all values oft E [u, b). Thus,
lxtl ~ 14>1 + lt m(s)lxsl ds + lt lh(s)l ds
fortE [u, b). The inequality in Lemma 3.1 of Section 1.3 implies
(1.4)
fortE [u, b). But this relation implies l±(t)l is bounded by a function in .Cioc.
Following the same proof as in Theorem 3.2 of Section 2.3 for equations with
continuous right-hand sides, one shows b = oo and the theorem is proved.
D
The most common type of linear systems with finite lag known to be
useful in the applications has the form
+ j_r A(t, B)x(t +B) dB+ h(t)

N 0
(1.5) ±(t) = ( ; Akx(t- Wk)
where 0 ~ w1 < w2 < · · · < WN ~ r and A(t, B) is integrable in B for each t

and there is a function a E .Cioc ( (-oo, oo), lR) such that
1/_: A(t,B)4>(B)dBI ~ a(t)14>1

for all t E lR and 4> E C.
To derive a representation for the solution, it is useful to rewrite the
equation. First we split off the part that explicitly depends on the initial
data
6.1 Resolvents and exponential estimates 169
±(t) =it d[ry(t, e- t)]x(e) + I:-t de[ry(t, e)]¢(t- a+ e)+ h(t)
= -ry(t, a- t)x(a)- it ry(t, e- t)±(e) de
+ I:-t de[ry(t, e)]¢(t- a+ e)+ h(t).
This equation is a Volterra equation (of the second kind)
(1.6) y(t) =it k(t, s)y(s) ds + g(t), t?: a,
where y(t) = i:(t), k(t, s) = ry(t, s- t) and
g(t) = -ry(t, a- t)¢(0) + I:-t de[ry(t, e)]¢(t- a+ e)+ h(t).
Let J be an interval. A measurable function k : J x J __, IRnxn is called

a Volterra kernel of type £1 on J if k(t, s) = 0 for s > t and llkll1 < oo,
where
II kill d~f sup r r ik(t, s)f(s)l ds dt = ess sup Jr ik(t, s)l dt.
lfh 9 } J }J sEJ J
The kernel k(t, s) = ry(t, s- t) is a kernel of type £ 1 on [a, oo).

We call a kernel of type £ 1 if J = IR and of type .Ciac if for every
interval [a, b] C IR, the kernel is of type £ 1 on [a, b].
If k is a kernel of type £ 1 on [a, oo), then
(1.7) f.._.... it k(t, s)f(s) ds
maps £1 [a, oo) into itself and
A kernel R(t, s), t?: s, of type £ 1 is called a Volterra resolvent of kif
R(t,s) = k(t,s) -1t R(t,a)k(a,s)da

(1.8)
= k(t, s)- 1t k(t, a)R(a, s) da.
A simple contraction argument shows that if k is a kernel of type £ 1 with

llkll1 < 1, then k has a resolvent of type £ 1. Further forgE £1, Equation
(1.6) has a unique solution in £ 1. This solution is given by the variation-
of-constants formula
(1.9) y(t) = g(t)- [t R(t, s)g(s) ds.
Lemma 1.1. If the hypotheses on TJ are satisfied, then the kernel k(t, s) =
TJ(t, s- t), t?: s, has a resolvent of type c~oc.
Proof. If we define
R(t,s) = R(t,s)e1 (t-s), ij(t,s) = TJ(t,s)e- 18 , k(t,s) = ij(t,s),
then R( t, s) satisfies the equation
R(t,s)=if(t,s-t)-1tR(t,a)if(a,s-a)da, t?:s.
If we choose "Y such that
sup
sE[O',oo)
1"'
00
[TJ(a,s-a)[e- 1 (s-a)da < 1,
then llkll 1 < 1 and R(t, s) is a resolvent of type C 1 on [a, oo ). Consequently,

R(t, s) = R(t, s)e-1(t-s)
is a resolvent of type £ 1 on [a, T] for the kernel TJ(t, s-t), t?: s. This proves
the lemma. D
Using the resolvent equation, we can give a representation theorem for

the solutions of System (1.1).
Theorem 1.2. If the hypotheses on TJ are satisfied, then for any given a,
¢ E C, and h E C~oc([a, oo ), IRn), there exists a unique solution x( ·;a,¢)
defined and continuous on [a- r, oo) that satisfies System (1.1) on [a, oo).
Furthermore, this solution is given by
(1.10) x(t; a,¢) = X(t, a)¢(0) + [t X(t, a) da[F(a, a;¢, h)], t ?: a,
where
(1.11) X(t, a)= I- [t R(s, a) ds
and F( ·,a;· ,h): C ~ c~oc([a,oo),IRn) is defined by
(1.12) F(t, a;¢, h) = ¢(0) + [t j_~s do[TJ(s, B)]¢(s +B) ds + [t h(a) da.
Proof. From Representation (1.9), we find that the derivative of any solution
of System (1.1) has the form
6.1 Resolvents and exponential estimates 171
(1.13) x(t) = g(t) - i t R(t, o:)g(o:) do:.
From Representation (1.13), the resolvent equation (1.8), and Fubini's the-
orem, we deduce that
x(t) = ¢(0) +it x(s) ds
= ¢(0) +it g(s) ds - i t is R(s, o:)g(o:) dads
= ¢(0)- it 'fJ(s, a- s)¢(0) ds +it is R(s, o:)'f)(o:, a- o:)¢(0) do: ds
+it da[F(o:, a;¢, h)] - i t is R(s, a) da[F(o:, a;¢, h)] ds
= ¢(0) - i t R(s, a)¢(0) ds +it da[F(o:, a;¢, h)]
- i t is R(s, o:) da[F(o:, a;¢, h)] ds
= X(t,a)¢(0) +it da[F(o:,a;¢,h)]
- i t it R(s,o:)dsda[F(o:,a;¢,h)]
= X(t, a)¢(0) +it X(t, o:) da[F(o:, a;¢, h)].
Thus, any solution of System (1.1) has the representation (1.10) and the
theorem follows from Lemma 1.1. D
The matrix solution X(t, s ), t 2:: s, has a natural interpretation for the
homogeneous differential equation
(1.14) x(t) = L(t)xt.
With respect to the initial data
forB= 0,
Xo(B) = {I,
0, for -r:::; B < 0,
Equation (1.6) becomes the resolvent equation
(1.15) R(t, s) = 'f)(t, s- t) -1t 'f)(t, o:- t)R(o:, s) do:, t :=:: s.
Therefore the matrix solution X(t, s) = I- J:

R(o:, s) do:, t 2:: s, can be
regarded as the solution of System (1.14) with respect to the discontinuous
initial data X 0 .
172 6o General linear systems
In the sequel, we define X (t, s) = 0 for t < s and call X (t, s) the
fundamental matrix solution of System (1.14)0
Corollary 1.1. The fundamental matrix solution X(t, s), t ;::: s, of System
( 1.14) satisfies the following estimates
(1.16) IX(t, s)l ~ exp[1t m(a) da]

(1.17) Vaqs,t] X( 0, s) ~ exp[1t m(a) da] -1, t;::: So
Proof. Since X(t,s), t;::: s, satisfies the integrated equation, the same esti-
mate as in Theorem 1.1 gives (1.16)0
The estimation
IR(t, s)l ~ m(t) + 1t IR(t, a)lm(a) da
and the inequality in Lemma 301 of Chapter 1 yield the a priori bound
IR(t, s)l ~ m(t) exp[1t m(a) da]o

Since X(t, s) = I - J: R(T, s) dT, we find
Var[s,t] X( 0, s) ~ 1t IR(T, 0')1 dT
~ 1t 7
m(T)exp[1 m(a)da] dT = exp[1t m(a)da]-1.
This proves estimate (1.17)0 D
As an illustration, we consider Equation (1.5)0 It is easy to verify that

Representation (1.10) becomes
x(t) = X(t, 0')¢(0) +~

N 1"a-wk X(t, a+ wk)Ak(a + Wk)¢(a- 0') da
(1.18) + 1a-w
a 1a+w
a X(t,s)A(s,a-s)ds<f>(a-O')da
+ lt X(t, s)h(s) dso
6.2 The variation-of-constants formula
In this section, it is our aim to present an abstract version of Equation

(1.10) that holds in the state space C. It will become clear that such a
formula has great advantages over Equation (1.10).
A two-parameter family T(t, IT), t ~ O", of bounded linear operators on
a real Banach space B is called a (forward) evolutionary system on B if
(i) T(IT,O")=I;
(ii) T(t, s)T(s, IT)= T(t, IT), t ~ s ~ 0".
From the existence and uniqueness for solutions of System (1.1), it follows
that translation along the solution defines an evolutionary system on C:
(2.1) T(t, iT)¢= Xt(.; 0", ¢),

From the variation-of-constants formula given by Equation (1.10), we find
that the solution of System (1.1) is given by
(2.2) Xt(B; IT,¢, h) = T(t, iT)cp(B) + 1u

t+&
X(t+B, a)h(a) da, -r ::::: () ::::: 0,
where it is understood that the integral is considered as a family of Eu-

clidean space integrals parameterized by B. It our objective to give an in-
terpretation of Equation (2.2) as a Banach space integral.
In order to do so, we have to introduce a little bit of vector-valued
integration. Let S : .C 1 ([a, b], IRn) . . . ., B be a continuous linear operator,
and let E be the IT-field of Lebesgue measurable subsets of [a, b]. Define
F: E......., B, E r---+ S(XE),
where XE denotes the characteristic function of E. One can easily verify that
whenever E 1 and E 2 are disjoint members of E then F(E1 UE2 ) = F(EI) +
F(E 2 ). Further, from the fact that there is a >..(E) > 0 (the Lebesgue
measure of E) such that
(2.3) IIF(E)II ::::: >..(E)IISII,
one verifies that F(U~ 1 Ej) = 2::~ 1 F(Ej) in norm for all sequences of
pairwise disjoint members of E such that U~ 1 Ej E E. A function F with
this property is called a countable additive vector measure. The variation of
F is the extended nonnegative function
(2.4) IFI : E......., [0, oo], IFI(E) =sup

7r
L IIF(A)II,
AE1r
where the supremum is taken over all partitions 1r of E into a finite number
of pairwise disjoint members of E. From (2.3), it follows that IFI([a, b]) :::;
(b- a)IISII- We call Fa vector measure of bounded variation.
It is easy to define the integral of an integrable scalar function with

respect to the vector measure F. First, define SF on the space of simple
functions
n n
(2.5) SF(L:CjXEJ) = "L.cjF(Ej)·
j=l j=l
One can show that (2.5) defines a linear operator and
n
ISF(f)l =I L CjF(Ej)l s; IFI([a, b])lfll·

j=l
So if the space of simple functions is given the .C 1 norm, then SF has a
continuous linear extension to .C 1 ([a, b], IR). Therefore, one can define
(2.6) r
J[a,b]
1 dF~f sF(!).
It is convenient to write the integral in (2.6) as a Stieltjes integral. This

can be done as follows: If we define K : [a, b] --+ T3 by K(t) = F([a, t]), then
K is of strong bounded variation over [a, b]; that is, the strong variation
function
N
(2.7) IKI(t) = sup
P(a,t)
L
j=l
IK(aj)- K(aj-1)1,
where P(a, t) denotes a partition a= a 0 < a 1 < · · · < aN =t of [a, t], is

bounded. We write the integral in (2.6) as follows
(2.8) { f dF = {b d[K(a)]f(a).
}[a,~ Ja
Iff is continuous, then the integral in (2.8) can be understood as a vector-
valued Riemann-Stieltjes integral.
After these preparations, we return to Equation (2.2). The following
lemma is the key to a variation-of-constants formula in the space C.
Lemma 2.1. Fix t 2 O" and define S(t) : .C 1 ([O", t], IRn) --+ C by
(2.9) S(t)h(B) = 1a
t+O
X(t + (}, a)h(a) da
where X(t, s) denotes the fundamental matrix solution to System (1.1). The
linear operator S(t) is completely continuous and can be represented by a
vector-valued integral
(2.10) S(t)h = lt d[K(t, a)]h(a),

where the kernel K(t, ·) : [a, t] ----> C is given by
(2.11) K(t, s)(O) = 1 8

X(t + 0, o:) do:.
Proof. First we show that S(t) is a completely continuous operator. Set

B1 ={hE .C1([a,t],IRn) : lhl 1 :::; 1}. By the Arzela-Ascoli theorem, we
have only to check that S(t)B 1 is uniformly bounded and equicontinuous.
Fix t > 0. From Corollary 1.1, we have
IS(t)h(O)I :::; exp[ " 1 t+O

m(o:) do:] lhl1·
Since for > 0, there is a 8 > 0 such that for 101 - 02 1 < 8, we have
E
IX(t + 81, o:)- X(t + 82, o:)l <E. Therefore, if lhl1 :::; 1 and I01- 82l < 8
IS(t)h(Ol)- S(t)h(02)I :::; (t- a)E.
This shows that S(t) is completely continuous. Note that X(t, s) = 0 fort<
s. Therefore, Representation (2.10) follows immediately from the definition
K(t, s) = S(t)X[u, 8 J· D
Corollary 2.1. The solution Xt = Xt( ·;a,¢, h) of System (1.1) inC satisfies
the following abstract variation-of-constants formula
(2.12) Xt = T(t, a)¢+ 1t d[K(t, o:)]h(o:)
where the kernel K(t, s) is given by (2.11). In the special case that L(t) =L
is independent oft in System (1.1), we have
T(t, s) = T(t- s, 0) ~f T(t- s), X(t, s) = X(t- s, 0) ~r X(t- s),
and
(2.13) Xt = T(t- a)¢+ 1t d[K(t, o:)]h(o:)
where the kernel K(t, ·) : [a, t] ----> C is given by
K(t, s)(O) = 18
X(t + 0- o:) do:.
6.3 The formal adjoint equation
In this section, we consider the general linear system
(3.1) i:(t) = =I:

L(t)xt d[ry(t, B)]x(t +B), t 2: rJ,
with Xa = ¢, ¢ E C, and where T) satisfies the conditions from Section 1.

The purpose of this section is to introduce the formal adjoint equation for
Equation (3.1). A two-parameter family V(s, t), s::; t, of bounded linear
operators on B is called a backward evolutionary system if
(i) V(t, t) =I;
(ii) V(s, rJ)V(rJ, t) = V(s, t), s ::; (J ::; t.
Let B* denote the dual space of Band let T(t,s), t 2: s, be a forward
evolutionary system on B. For every (t, s ), t 2: s, we can define the adjoint
operator T(t, s )* on B*. If we set T* (s, t) = T(t, s )*, then
(¢,T*(s,r)T*(r,t)¢*) = (T(t,r)T(r,s)¢,¢*)
= ( T(t, s)¢, ¢*)
= ( ¢, T*(s, t)¢* ).
So T*(s, t), s::; t, is a backward evolutionary system on B* and is called
the adjoint system of T(t, s), t 2: s. Note that, in particular, we have that
(3.2) ( T(s, rJ)¢, T*(s, t)¢*), (J ::; s ::; t,
is independent of s. This property plays an important role in the Fredholm

alternative, when dealing with boundary-value problems. In order to use
(3.2) we have to compute the adjoint system of T(t, s), t 2: s.
Let B 0 denote the Banach space of row-valued functions 'lj; : ( -oo, OJ ---+
IRn* that are constant on ( -oo, -r], of bounded variation on [-r, 0], contin-
uous from the left on ( -r, 0) and vanishing at zero with norm Var [-r,OJ ?j;.
Together with the pairing
(3.3) (f, ¢) =I: df(B)¢(8), f E Bo, ¢ E C,
the space B 0 becomes a representation for the dual space of C.

So the adjoint system of T(t, s ), t 2: s, is a backward evolutionary sys-
tem on B 0 , and to compute T*(s, t), s::; t, we use the pairing (3.3). Before
we do so, we would like to have some more information about T*(s, t), s::; t.
An obvious question would be: Is T*(s,t),s::; t, the evolutionary system
for a differential equation?
Define the formal adjoint equation by
(3.4) y(s) +it y(T) ry(T, s- T) dT =constant, s ::; t - r,

6.3 The formal adjoint equation 177
where y is a solution that vanishes on [t, oo), satisfies Equation (3.4) on

(-oo, t- r], and such that y(t + 0) = '1/J(O), -r ::; (J ::; 0 for 'ljJ in Bo. In
general, we cannot differentiate the integral equation to obtain a differential
equation, since the solution on the interval [t- r, t] is only in B 0 .
Many applications have the form given in Equation (1.5). In this case,
the formal adjoint equation can be written as
(3.5)
dy(s)
----;{8 = - ~
~
y(s + wk)Ak(s + wk) -
jo y(s-
-w ~)A(s- ~' ~) d~
with Yt = '1/J, 'ljJ E Bo and s ::; t.

To clarify the relation between the adjoint evolutionary system T*(s, t)
and the formal adjoint equation (3.4), it turns out to be useful to study the
adjoint equation as a forced Volterra equation. If we define g to be
0, for (J = 0,
(3.6) g(O) = { '1/J(O) + J~ '1/J(r)ry(t + T, s- r) dr, for -r::; (J < 0,
g(-r), for (J::; -r,
then g belongs to B 0 and the solution y of Equation (3.4) satisfies the

Volterra equation
(3.7) y(s) + 1t y(r) ry(r, s- r) dr = l(s), s ::; t,
where l(s) = g(s -t). The form of this equation resembles Equation (1.6).
Recall that we introduced the formal resolvent equation associated with
ry(t,s- t) by
R(t, s) = -ry(t, s- t) + 1t R(t, a)ry(a, s-a) da
or equivalently,
(3.8) R(t, s) = -ry(t, s- t) + 1t ry(t, a- t)R(a, s) da.
In Lemma 1.1, we showed the existence of a unique solution to this equation.

The proof of the next theorem follows the same lines as the proof of Theorem
1.1, using the (equivalent) resolvent equation (3.8).
Theorem 3.1. If the hypotheses on 'f/ from Section 1 are satisfied, then for
any given tin ffi and g in B 0 , there exists a unique solution y( ·; t, g) defined
and locally of bounded variation on (-oo, t] that satisfies the formal adjoint
equation
(3.9) y(s) + 1t y(r) ry(r, s- r) dr = l(s), s ::; t,

where gt ( s) = g( s - t). Furthermore, a representation of the solution is

given by
y(s; t, g)= l(s)- 1t l(a)R(a, s) da

(3.10)
= -1t d[g(a- t)]X(a, s), s:::; t.
Translation along a solution of Equation (3.9) induces a two-parameter

family of bounded operators V(s, t), s:::; ton B 0 .
'''----- -:-- ~--

'' ;:-----' ''''' ~
'~_;: ',
:: g :: l: ',, ', ___ _
I
I
: t
: g
I
'
cr-r t-r
Fig. 6.1.
In fact, the solution y restricted to ( -oo, a] satisfies a Volterra integral

equation given by
s:::; a,
where g belongs to B 0 and is given by
fore= 0;
g(B) = {~a(B)+ f0°ya(T)'TJ(T+a,B-T)dT, for -r::::: e < 0;
g( -r), fore::::: -r
(see Fig. 6.1). Define V(a, t)g = g, so that V(a, t) maps the forcing function
for the solution on ( -oo, t] onto the forcing function for the solution on
( -oo, a]. From the uniqueness property for the Volterra equation, it is easy
to see that V(s, t), s:::; t, defines a backward evolutionary system on B 0 .
6.3 The formal adjoint equation 179
We shall see shortly that V(s, t), s $ t, is the adjoint system of

T(t, s), t ~ s. First we derive a representation for V(a, t), from its defi-
nition. The integral equation yields
g(a + 0- t) = y(a + 0) + 1t y(r)ry(r, a+ 0- r) dr

u+f.l
(3.11)
= Yu(O) + it-u Yu(T)'fJ(T +a, 0- r) dr.
Therefore
g(O) = g(O- (t- a)) -lt-u Yu(T)rJ(T +a, 0- r) dr
and, using Representation (3.10), V(a, t) : B 0 ---+ B 0 is given by
V(a, t)g(O) = g(O- (t- a))- i t y(r)ry(r, 0- r) dr
(3.12) = g(O- (t- a))
+it lt d[g( -t + a)]X(a, r)ry(r, 0- r) dr.
By definition, V(a, t)g(O) = 0, and using Equation (3.11),
(3.13) V(a, t)g(O-) ~f lim V(a, t)(O) = y(a).

f.l--+0
Theorem 3.2. Let T(t, s), t ~ s be the evolutionary system associated with
System (3.1) on C. If V(s, t), s $ t denotes the backward evolutionary sys-
tem for the adjoint equation defined by (3.12), then V(s, t), s $ t is the
adjoint system ofT(t, s), t ~ s, that is,
T(t, s)* = V(s, t).
Proof. Compute
( ¢, V(s, t)g) = /_: ¢(0) d[V(s, t)g(O)]
= -cp(O)y(s)
+/_or ¢(0) d[g(O- (t- s))]
+/_or ¢(0) 1t lt d[g( -t + a)]X(a, r) dry(r, 0- r) dr.
For the first term, we use the representation for y given in Theorem 3.1,
rjJ(O)y(s) = -¢(0) 1t d[g(a- t)]X(a, s)
= -1:t X(a + t, s)¢(0) d[g(a)].
Since ¢(0) d[g] is a scalar-valued measure, it commutes with X and
(3.14) cp(O)y(s) =-I: X(a + t, s)¢(0) d[g(a)]
where we used that X(t,s) = 0, t:::; s. The second integral becomes
(3.15) lor ¢(8) d[g(8- (t- s))] =I: ¢(a+ (t- s)) d[g(a)].
For the third term, integrating first with respect to 8 and using that ¢(0) d[g]
is a scalar-valued measure, yields
lor cp(8) 1t 1t d[g(-t+a)]X(a+T)dry(T,8 -T)dT
1t 1:t X(a+t,T) [~T de[ry(T,8)]¢(T+8)d[g(a)]dT
[orlo+t X(a+t,T) [~ 7 de['17(T,8)]rjJ(T+8)dTd[g(a)]

where in the last identity, we used Fubini's theorem to reverse the order of
integration, and the convention that X(t, s) = 0, t:::; s. Altogether, we find
( ¢, V(s, t)g) =lor X(a + t, s)¢(0) d[g(a)] +lor ¢(a+ (t- s)) d[g(a)]
+ 1:1a+t X(a+t,T) [~ 7 de[ry(T,8)]¢(T+8)dTd[g(a)]

= ( T(t, s)¢, g).
D
6.4 Boundary-value problems
In this section, we discuss two-point boundary-value problems for the non-

homogeneous equation (1.1) and obtain results of the "Fredholm alterna-
tive" type. The notation of the previous section will be employed without
explanation.
6.4 Boundary-value problems 181
Suppose V is a real Banach space, cr < T are given real numbers,

M, N : C ---+ V are linear operators with domain dense in C, and 'Y E V is
fixed. The problem is to find a solution x of
(4.1) x(t) = L(t)xt + h(t)

subject to the boundary condition
(4.2) Mxa + Nxr = 'Y·
Let V* be the dual space of V, and M* and N* the adjoint operators of

M and N respectively. The fundamental result is
Theorem 4.1. In order that Equations (4.1) and (4.2) have a solution, it is
necessary that
(4.3) 17
y(a)h(a)da = -(8,"/)v
for all 8 E V* and solutions y of the system of adjoint equations
(4.4) {
y(s) + 1
7
y(a)rJ(a, s-a) da = N*8, S:::; T,
y(s) + la y(a)rJ(a, s-a) da = -M*8, s :::; cr.
IfR( M + NT(r,cr)) is closed, then the preceding condition is sufficient.
Proof. Define S(r,cr): .C 1 ([cr,r];IRn)---+ C by
S(r,cr)h(O)= a 1 r+O
X(r+O,a)h(a)da.
The solution to Equation (4.1) is given by
X7 =T(r,cr)¢+S(r,cr)h
and Relation (4.2) is equivalent to
NS(r, cr)h- 'Y E R( M + NT(r, cr) ).

Therefore, it is necessary and, under the closure hypothesis, sufficient that
NS(r, cr)h- 'Y E Cl R( M + NT(r, cr)) = N( M* + T*(cr, r)N* )1.;

that is, for any 8 E V* such that M* 8 + T* (cr, T )N* 8 = 0, we must have
( 8, NS(r, cr)h- 'Y )v = 0, or
(8,"f)v = (8,NS(T,a)h)v
= (N*8,S(T,a)h)v
(4.5) = £:1r+IJ X(T+0,a)h(a)dad[N*8(0)]
= 1 7
h(a) 1~ 7 d[N*8(0)]X(T + 0, a) da.
In the last equality, we used Fubini's theorem to change the order of inte-
gration. From Representation (3.10) for the solution of Equation (4.4), we
derive that
1~ 7 d[N*8(0)]X(T + 0, a)= lr d 0 [N*8(0- T)]X(O, a)
= -y(a),
where we used that X(t, s) = 0 for t :::; s. Substitution in Equation (4.5)

yields Equation (4.3) and proves the theorem. D
As another boundary-value problem, suppose P, Q : V --> C are linear

operators with domain dense in V and that p and q are fixed elements of
C. The problem is to find avE V and a solution of Equation (4.1) on [a, T]
such that
(4.6) x(T = Pv + p, Xr = Qv + q.
Let P* and Q* be the adjoints of P and Q, respectively.
Theorem 4.2. In order for Equations (4.1) and (4.6) to have a solution, it
is necessary that
(4.7) 1y(~)h(~)d~
7
= (T*(a,T)g,p)- (g,q)
for every solution y of the formal adjoint problem
(4.8) y(s)+ lr y(a)ry(a,s-a)da=g, s:::;T
with g in B 0 such that
(4.9) P*T*(a, T)g = Q*g.

If R( Q - T( T, a )P) is closed in C, this condition is both necessary and
sufficient.
Proof. Proceeding as in the proof of Theorem 4.1, Relation (4.6) is equiva-
lent to
T(T, a)p- q + S(T, a)h E R( Q- T(T, a)P).
6.4 Boundary-value problems 183
Therefore, it is necessary and, under the closure hypothesis, sufficient that
T(T, a)p- q + S(T, a)h EN( Q*- P*T*(a, T) ).l;
that is; for any g E B 0 such that Q* g - P*T* (a, T )g = 0,
0 = (g, T(T,a)p- q + S(T,a)h)

(4.10)
= (T*(a,T)g,p)- (g,q) + (g,S(T,a)h).
The same argument as in the proof of Theorem 4.1 shows
(g,S(T,a)h) = -1r y(~)h(~)d~,

where y is the solution of (4.8). Substitution into (4.10) yields the conditions
(4.7)-(4.9) and this proves the theorem. 0
Our next objective is to consider Equation (4.1) with hE Pw; that is,
his continuous and h(t + w) = h(t) for all t :2: 0. Let if" : Pw ~ Pw be a
continuous projection ofPw onto the periodic solutions of the homogeneous
equation
(4.11) x(t) = L(t)xt
of period w. For example, one can define if" in the following manner. Let
U = (</>1, ... , </>d) be a basis for thew-periodic solutions of Equation (4.11)
and define
We can now state the following result.
Corollary 4.1. Suppose h(t) and L(t) in Equation (4.1) are periodic in t
of period w > 0. The necessary and sufficient condition that there exist
w-periodic solutions of Equation (4.1) is
(4.13) 1w y(a)h(a) da = 0
for all w-periodic solutions y of the formal adjoint problem (3. 7). Further-
more, there is a continuous projection J : Pw ~ Pw such that the set of
h in Pw satisfying Equation (4.13) is (I- J)Pw and there is a continuous
linear operator K : (I- J)Pw ~ (I- ir)Pw such that Kh is a solution of
Equation (4.1) for each hE (I- J)Pw·
Proof. Take a = 0, T = w, V = C, "! = 0, and M = -N = I in Theorem
4.1. Since the solution operator is linear, Corollary 6.1 of Section 3.6 implies
that T(T, a)= U(T, a)+ S(T- a), where U(T, a) is completely continuous
and the spectral radius of S(t- T) is zero. Therefore, R( I- T(w, 0)) is

closed. The corollary now follows from Theorem 4.1.
Let V = col ('lj; 1 , ... , '1/Jd) be a basis for the w-periodic solutions of
Equation (3.7). If we define J: Pw--+ Pw
(4.14)
then Jh = 0 if and only if Condition (4.13) is satisfied.

If Condition (4.13) is satisfied, then we have seen there is an w-periodic
solution x(h) of Equation (4.1). If Kf =(I -ii")x(h), then all assertions of
the theorem are true and the proof is complete. 0
6.5 Stability and boundedness
In this section we consider the homogeneous linear equation (4.1); that is,
(5.1) x(t) = L(t)xt
where L satisfies the conditions of Section 6.1. Also, X(t, s) denotes the
fundamental matrix solution of Equation (5.1) given in Theorem 1.1, and
T(t, a) denotes the solution operator of Equation (5.1).
Lemma 5.1. The following statements are equivalent:

(i) The solutions of Equation (5.1) are bounded.
(ii) The solution x = 0 of Equation (5.1) is stable.
(iii) There is a constant c(a) for each a E lR such that IT(t,a)l ::::; c(a),
t?. a.
(iv) There is a constant C(s) for each s E lR such that IX(t, s)l ::::; C(s),
t?. s.
Proof. If each solution of Equation (5.1) is bounded, then, for any (a,¢) E
lR x C, there is a constant c( a, ¢) such that the bounded linear operator
T(t, a) satisfies
IT(t, a)¢1::::; c(a, ¢) for t?. a.
The principle of uniform boundedness implies that there is a constant c(a)
such that IT(t, a)l ::::; c(a) and so (i) implies (iii).
Since X(t, s) satisfies Equation (1.11), it follows from Lemma 1.1 that
(5.2) t ?. s.
6.5 Stability and boundedness 185
Consequently, IXt(-, s)l ~ -y(s) = exp J:+r m(u) du, s ~ t ~ s + r. Since

X(t, s) satisfies Equation (5.1) fort 2: s+r, it follows that (iii) implies that
IXt( ·, s)l ~ c(s + r)'Y(s) fort 2: sand so (iii) implies (iv).
Theorem 1.2 shows that (iv) implies (ii). Obviously (ii) implies (i) and
the proof is complete. 0
One could have avoided the principle of uniform boundedness by prov-

ing that (i) implies (iv) implies (iii) implies (ii) implies (i). However, the
application of the principle of uniform boundedness can be used to show the
equivalence of properties similar to (i) and (iii) for general linear processes.
Lemma 5.2. If there is a constant m 1 such that
(5.3) l t+r
t m(u) du ~ m1 for t E lR,
then the following statements are equivalent:

(i) The solutions of Equation (5.1) are uniformly bounded.
(ii) The solution x = 0 of Equation (5.1) is uniformly stable.
(iii) There is a constant c such that for all a E lR, IT(t, a)l ~ c, t 2: a.
(iv) There is a constant C such that for all s E lR, IX(t, s)l ~ C, t 2: s.
Proof. If the solutions of Equation (5.1) are uniformly bounded, then there
is a c > 0 such that for all a E lR,
for t 2: a, ¢ E C, 1¢1 ~ 1.
Therefore, IT(t, a)l ~ c, t 2: 0 and (i) implies (iii). Hypothesis (5.3) and
Inequality (5.2) imply that for all s E lR, IXt( ·, s)l ~ -y, 'Y = exp m1, for
s ~ t ~ s + r. As in the proof of the previous lemma, for all s E lR, this
implies that IXt( ·, s)l ~ c-y fort 2: s if (iii) is satisfied. Thus, (iii) implies
(iv). Using Theorem 1.1, one easily sees that (iv) implies (ii). It is obvious
that (ii) implies (i) and the lemma is proved. 0
Note that statements (i)-(iii) of Lemma 5.2 are equivalent without

Hypothesis (5.3).
Lemma 5.3. If Inequality (5.3) is satisfied, then the following statements are
equivalent:
(i) The solution x = 0 of Equation (5.1) is uniformly asymptotically stable.
(ii) The solution x = 0 is exponentially asymptotically stable; that is, there
are constants c > 0, a > 0 such that for all a E lR,
t 2: a.
(iii) There are constants C > 0, a> 0 (a is the same as in (ii)) such that
for all s E lR,
IX(t, s)i :s; ce-o:(t-s)' t ?_ s.
Proof. If (i) is satisfied, then, for any ry > 0, there is a T = T(ry) > 0 such
that for all u E lR, 1¢1 :s; 1,
lxt(u, ¢)1 < 'T/ for all t ?_ u + T.

Consequently, IT(t, u)l < ry fort?_ u + T. Choose ry < 1 and let
a= -T- 1 logry, c = c0 exp aT
where c0 is the constant in Lemma 5.2(iii) guaranteed by the uniform sta-

bility of the solution x = 0 of Equation (5.1). For any t ?_ u, there is an
integer n ?_ 0 such that nT :s; t - u < (n + 1)T. Thus,
iT(t,u)i :s; iT(t,u+nT)iiT(u+m,u)i

:s; coiT(u + nT, u)i
:s; c0 ryiT(u + (n- 1)T, u) I
:s; Co'T/n = co exp( -anT)
= cexp[-a(n + 1)T] :s; cexp[-a(t- u)].
This proves (ii).

The proofs that (ii) implies (iii) and (iii) implies (i) are supplied in the
same manner as in the previous two lemmas. D
Note that Properties (i) and (ii) of Lemma 5.3 are equivalent without
Hypothesis (5.3).
The theory of resolvents used in Section 6.1 is basic classical material and
one can consult Grippenberg et al. [1] and Miller [1,2]. Another approach
is to integrate Equation (1.1); see Section 9.1 for details. For the theory
of vector measures, we refer to Diestel and Uhl [1]. The theory for vector-
valued Riemann-Stieltjes integrals can be found in Hille and Phillips [1].
The formal adjoint equation has been used in functional differential
equations since 1920. For a complete list of references on its origin and
evolution, see Zverkin [3] and Hale [22].
The idea to study a functional differential equation as a Volterra in-
tegral equation has been used by Diekmann [1,2,3], Delfour and Manitius
[1,2], Staffans [1,2] and Verduyn Lunel [1,2,3].
The interpretation of the adjoint evolutionary system as a backward

evolutionary system on a Volterra integral equation was given by Diekmann
[2] for autonomous equations. The approach in Section 6.3 has the advan-
tage that it eliminates the use of the "true adjoint" (see Henry [3] and Hale
[22]).
Now that we have a good version of the variation-of-constants formula
(see Equation (2.12)), we can subject the linear equation to many types
of perturbations and use techniques very similar to the ones in ordinary
differential equations to obtain properties of solutions. We indicate briefly
a few of the results here and suggest that the reader consult the references
for more details.
6.6.1 Boundary-value problems
We keep the same notation as in Section 6.4. Also, we make the following
hypothesis:
(6.1) R(M +NT(r,a)) = VE, N(M +NT(r,a)) = Cu

where U : C -+ C and E : V -+ V are continuous projections with ranges
Cu and VE, respectively. In particular, this hypothesis implies that these
sets are closed. Let JC : VE -+ C1-u be a bounded right inverse of M +
NT(r, a). If')' E R(M +NT( r, a)), then the function T(t, a)IC')', a ::;: t::;: r,
is a solution of the linear equation
(6.2) ±(t) = L(t)Xt
and the boundary conditions (4.2).
Now suppose that f: IR x C-+ IRn is a given function that is continu-
ous in (t, ¢)and continuously differentiable in¢, and consider the solutions
of the RFDE
(6.3) ±(t) = L(t)Xt + f(t, Xt),
which satisfy the boundary conditions (4.2). If we let Xt(a, ¢)be the solution
to Equation (6.3) with xu(a, ¢)=¢and define
(6.4) W(t,a,¢) = 1t d[K(t,s)]f(s,x 8 (a,¢)),
then Xt(a, ¢) is a solution of the boundary-value problem if and only if

(6.5) [M + NT(r, a)]¢= 'Y- NW(r, a,¢).
Using the projection operators U and E, we observe that Equation
(6.5) is equivalent to the system of equations
(6.6a) v =ICE[!'- NW(r, a, u + v)],

(6.6b) (I- E)[-y- NW(r, u, u + v)] = 0,
where¢= u + v, u E Cu = .N(M + NT(r,u)), v E C1-U·

We are in a position now to apply all of the existing methods for the
solution of Equation (6.6)-fixed point theorems, the method of Liapunov-
Schmidt, or more generally the method of alternative problems (see Cesari
[1,2], Hale [7,21] for extensive references) and degree theory (see Mawhin
[3], for example).
One of the most elementary, but at the same time very important,
observations is that we can easily apply the method of Liapunov-Schmidt
to Equations (6.6) if the function f and its derivative Dq,f with respect to
¢are small.
Theorem 6.1. If (6.1) is satisfied, then there are constants E > 0, 81 >
0, 82 > 0 such that for any function f satisfying lf(t, ¢)1 < E, IDq,f(t, ¢)1 <
E fortE [u, r], 1¢1 < 81, there is a unique function v*(u, ¢) satisfying (6.6a),
having norm< 82, continuously differentiable in u, f and v*(u, 0) =KE-y.
Furthermore, the boundary-value problem (6.3), (4.2) has a solution x(u, ¢)
with norm < 81 if and only if¢ = u + v* (u, f), where u is a solution of the
bifurcation equation
(6.7) G(u, ¢) ~f(I- E)['Y- NW(r, u, u + v*(u, ¢))] = 0.
The proof of this result is an elementary application of the implicit

function theorem.
A trivial consequence of Theorem 6.1 arises when E =I; that is, the
range of M + NT( r, u) is the whole space.
Corollary 6.1. If the conditions of Theorem 6.1 are satisfied and E = I,

then, for any u E Cu, there is a solution x( u, ¢) of the boundary-value
problem (6.3), (4.2) with ¢ = u + v*(u, f) provided that 1¢1 < 82. If, in
addition, .N(M + NT(r,u)) = {0} (that is, U = 0), then this solution is
unique.
For results related to Corollary 6.1, see Fennell and Waltman [1] and
Mosjagin [1].
The most interesting situation is when E =/:- I and U =/:- 0. In this
case, the solution of the original boundary-value problem is reduced to the
discussion of the solutions of the bifurcation equation (6.7).
Let us translate these results to the case where the boundary conditions
(4.2) correspond to periodicity conditions (see Perella [2], Shimanov [5]).
More specifically, suppose there is a positive constant w such that L(t+w) =
L(t), f(t+w, ¢) = f(t, ¢)for all (t, ¢)and let us determine periodic solutions
of Equation (6.3) of period w. In our previous notation, this is the same
as taking u = 0, r = w, V = C, M = -N = I, 'Y = 0. The space Cu is
the space of w-periodic solutions of the linear equation (6.2) and the space
CI-E is the space of w-periodic solutions of the formal adjoint equation
(3.7). The spaces Cu and CI-E have the same dimension d and d < oo. If
we choose a basis iP for Cu and a basis lJ! for CI-E, and let u = iPa, where
a E IRd, then there is ann-vector function B(a,f) such that
(6.8) G(iPa, f) = lJ! B(a, f),
and the bifurcation equation (6.7) is equivalent to the equation
(6.9) B(a, f)= 0.
Therefore, the problem of the existence of an w-periodic solution of Equa-
tion (6.3) is reduced to the discussion of solutions of equation (6.9) in IRd.
Let us state another result of a more global nature.
Theorem 6.2. If (6.I) is satisfied with E = I and lf(t, ¢)1/1¢1 --+ 0 as

1<1>1 --+ oo uniformly fort E [a, r], then there exists at least one solution of
the boundary-value problem (6.3), (4.2).
We only give an outline of the proof. Since E = I, the problem is

to find a solution of Equation (6.6a); that is, a solution of the equation
v + Su(v) = IC"(, where Su(v) = JCNW(r,a,u + v). It is possible to show
that the function Su(-) is completely continuous and IBu(v)l/lvl --+ 0 as
lvl --+ oo. Now, we apply a fixed point theorem of Granas [I].
For a complete proof of Theorem 6.2, see Hale [22]. Theorem 6.2 is
a generalization (since it is not assumed that T - a 2: r) of a result of
Waltman and Wong [I].
Related results on boundary-value problems are contained in Kwapisz
[I] and Nosov [3,4]. For other results on the existence of periodic solutions,
see Burton [3], Perella [3], Ziegler [I]. For periodic solutions of parabolic
equations with delays, see Biroli [I], and Comincioli [I]. Boundary-value
problems for special equations is an entire subject in itself and has an
extensive literature. See Norkin [I], Nosov [I], Grimm and Schmitt [I,2],
Medzitov [I], Sentebova [I], Kobyakov [I], Kovac and Savcenko [I] and de
Nevers and Schmitt [I]. For general functional boundary conditions, see
Henry [3], Myjak [I], and Kwapisz [2].
6.6.2 Exponential dichotomies, bounded and almost periodic

perturbations
In Corollary 6.I and Theorem 6.2, we have seen that if the range of the
solution operator of a homogeneous linear boundary-value problem is the
whole space, then the existence of a solution of a perturbed problem can be
reduced to the discussion of the existence of a fixed point of some operator.
It is possible to extend this idea to more general situations; for example, to
the existence of almost periodic solutions.
We say that the linear system (6.2) admits an exponential dichotomy

on JR (see Henry [5], Hale [28]) if the solution operator T(t, s) satisfies
the following property: there exist positive constants K, a, and projection
operators P(s): C---> C, s E JR such that if Q(s) =I- P(s), then
(i) T(t, s)P(s) = P(t)T(t, s), t ~ s.
(ii) The restriction T(t,s)R(Q(s)), t ~ s, is an isomorphism of R(Q(s))
onto R(Q(t)) and we define T(s, t) as the inverse mapping.
(iii) IT(t, s )P( s) I :::; K e-a(t-s)' t ~ s.
(iv) IT(t,s)Q(s)l:::; Ke-a(t-s), s ~ t.
The operator P(t) is called the projection operator of the dichotomy.
The solution of System (6.2) with initial value at time s in the range of
P(s) (resp. Q(s)) approach zero exponentially as t---> oo (resp. t---> -oo).
Let BC(JR, Y) be the bounded continuous functions from JR to a real
Banach space Y. Suppose that L E BC(JR, .C(C, JRn)) and that System
(6.2) admits an exponential dichotomy. For any h E BC(JR, JRn), it follows
that the equation
(6.10) x(t) = L(t)xt + h(t)

has a unique solution Kh E BC(JR, C), and it is given by
(6.11) Kh(t) =[too d[P(s)K(t, s)]h(s) -1 00

d[Q(s)K(t, s)]h(s).
If L(·), hare in AP (AP is the class of almost periodic functions), then it

also can be shown that Kh is AP and the frequency module is the union
of the frequency modules of L( ·) and h.
Iff(·,¢) E BC(JR, JRn), then the equation
(6.12) x(t) = L(t)xt + f(t, xt)

has a solution t ~---> Xt E BC(JR, C) if and only if
(6.13) Xt = Kf(t, Xt), t E JR.
Using the contraction mapping principle, we can obtain very easily an ex-
tension of Corollary 6.1 for the existence of solutions of Equation (6.12)
that are bounded on JR. If the perturbation function!(·,¢) is in AP uni-
formly with respect to¢ in bounded sets, then the bounded solution is AP.
Special cases of such perturbation results have been given by Halanay [1]
and Konovalov [1].
Results on A'P-solutions without assumptions of smallness of the per-
turbed vector field are very difficult to obtain. For some results and refer-
ences, see Fink [1] for ODE, Yoshizawa [2] for ODE and RFDE and Hino,
Murakami, and Naito [1] for RFDE.
If L E BC(JR,.C(C,JRn)) and, for each hE BC(JR,JRn), there is at

least one solution of the Equation (6.10) in BC(JR, JRn), then the homo-
geneous equation System (6.2) must admit an exponential dichotomy (see
Burd and Kolesov [1], and Pecelli [1]). If there is an exponential dichotomy
for (6.2), then there is the solution Kh E BC(JR, C) if h E BC(JR, JRn).
This suggests the following more general concepts. If X, Yare real Banach
spaces, we say that (X, Y) is admissible for Equation (6.10) if, for every
h E X, there is at least one solution of Equation (6.10) in Y. Coffman
and Schaffer [2], Schaffer [1], Pecelli [1] and Corduneanu [1] have discussed
admissible pairs for RFDE.
6.6.3 Asymptotic behavior
The variation-of-constants formula in the function space C also is very

useful for the study of the asymptotic behavior of solutions of a perturbed
system
(6.14) x(t) = LXt + M(t)xt

where L C --+ C is a continuous linear operator, M(t) : C --+ C is a
continuous linear operator that is "small" at oo. Since the operator L is
independent of t, we would expect to be able to relate the solutions of
Equation (6.14) to the autonomous equation
(6.15) x(t) = Lxt.
More specifically, if we have an eigenvalue!" of Equation (6.15), under what

conditions on M(t) will there be a solution of Equation (6.14) that has an
asymptotic behavior similar to ell-t? Using the variation-of-constants for-
mula in C, Hale [6,22] has obtained very precise results, which are more
general than the ones obtained by Bellman and Cooke [2,3] using the
variation-of-constants formula in JRn. Kato [2] has given similar results.
For other results on asymptotic behavior, see Kato [3] and Onuchic [3]. For
a differential difference equation,
(6.16) x(t) = Ax(t) + Bx(t- r(t)) + Cx(t- s(t)),

it also is of interest to determine the asymptotic behavior of the solutions
if r(t), s(t) are "close" to constants for t large. In this case, the space C
is too large in general. Cooke [2] has shown that the theory of Chapter
7 and the variation-of-constants formula in the function space W 1 •00 can
be effectively used to discuss this more general situation (see also Bellman
and Cooke [4]). The space W 1 •00 or W 1 ·P for some p, 1 < p < oo, has
also been used for the existence of periodic solutions (see Ginzburg [2],
Ruiz-Clayessen [1], and Stephan [1,2]).
Perturbation problems that involve a small parameter in the highest

derivative have received much attention and probably should be reexam-
ined in the light of the preceding remarks. For some of the literature, see
the nine volumes of Trudy Sem. Teorii Differentialnije Uravneniyija c Otk-
lonyayushimcya Argumentom, published by Univ. Patrisa Lumumba and
the papers of Cooke [3], Cooke and Meyer [1], Habets [1], and Magalhaes
[1,2,3,4]. A special singular perturbation problem is considered in some de-
tail in Chapter 12.
7
Linear autonomous equations
A linear autonomous RFDE has the form
(1) x(t) = Lxt

where L is a continuous linear mapping from C into IRn. This hypothesis
implies that there exists ann X n matrix ry(B), -r ~ e ~ 0, whose elements
are of bounded variation, normalized so that 17 is continuous from the left
on ( -r, 0) and ry(O) = 0, such that,
(2) L¢ = j_or d[ry(B)]¢(0),

The goal is to understand the geometric behavior of the solutions of Equa-
tion (1) when they are interpreted in C.
7.1 Strongly continuous semigroups
Let B be a real Banach space. A strongly continuous semigroup of linear

operators, in short a C0 -semigroup, is a one-parameter family T(t) : B----> B,
t ?: 0, of bounded linear operators that satisfy the properties
(i) T(O) =I;
(ii) T(t1 + t2) = T(tl)T(t2),
(iii) limno IIT(t)¢- ¢11 = 0,
To every C0 -semigroup T(t), we can associate an infinitesimal generator
A: V(A) ----> B defined by
A¢= lffti ~ [T(t)¢- ¢], ¢ E V(A),
that is, for all ¢ E B for which the limit exists in the norm topology of B.
The following lemma is standard.
Lemma 1.1. If T(t) is a Co-semigroup on B, then

194 7. Linear autonomous equations
(i) for every¢ in B, t ~-----' T(t)¢ is a continuous mapping fmm IR+ into B;
(ii) A is a closed densely defined operator;
(iii) for every¢ E V(A), t ~-----' T(t)¢ satisfies the differential equation
:t T(t)¢ = AT(t)¢ = T(t)A¢.
Next, we investigate the abstract properties of the solution operator

of Equation (1). Let ¢be a given function in C. If x( ·; ¢) is the unique
solution of Equation (1) with initial function ¢ at zero, then the solution
operator T(t) : C---+ Cis defined by the relation
(1.1) Xt(-; ¢) = T(t)¢.

Lemma 1.2. The solution operator T(t), t 2: 0, defined by Relation (1.1), is
a C0 -semigmup with infinitesimal generator
V(A) = {¢ E C: ~: E C, ~: (0) =/_or d[ry(B)]¢(8) },

(1.2)
A¢= d¢
de·
Furthermore, T(t) is completely continuous fort 2: r; that is, T(t), t 2: r,
is continuous and maps bounded sets into relatively compact sets.
Pmof. From the uniqueness of solutions of Equation (1), it is obvious that
T(t) is a linear transformation that satisfies the semigroup property. By
definition, T(O) =I. Since L : C---+ IRn is continuous and linear, it follows
that there is a constant 'Y such that IL¢1 :::; 'YI¢1 for all ¢ in C. From the
definition of T(t), we have, for any fixed t 2: 0 and -r:::; 8:::; 0,
¢(t +e), t + e :::; o,
{
(1. 3) T(t)¢(B) = ¢(0) + f~+e LT(s)¢ds, t + 8 > 0.
It follows that IT( t)¢1 :::; 1¢1 + 'Y J~ IT( s )¢1 ds. The inequality in Lemma 3.1
of Section 1.3 then implies that
(1.4) t 2: 0, </J E C
and thus, T(t) is bounded. From (1.3) it is readily seen that T(t) is strongly
continuous at zero. So T(t) is a C0 -semigroup. Let R > 0. If S = {¢ E C:
I<PI < R}, then for any 'ljJ in T(t)S, t 2: r, Relation (1.4) implies 11'1/JII :::; e"'~t R,
and Equation (1) implies I~ I :::; "(e"'~t R. Since these functions are uniformly
bounded with a uniform Lipschitz constant, T(t)S, t 2: r, belongs to a
compact subset of C. To finish the proof, we compute the infinitesimal
generator. From Lemma 1.1 and the strong continuity of T(t), it follows
that, for every¢ in V(A), we have
7.1 Strongly continuous semigroups 195
A¢= d¢
d()
and V(A) <;;; {¢ E C I ~: E C}.
On the other hand, if ¢ is continuously differentiable, the limit
lim~t [T(t)¢(8)- ¢(8)]

tlO
exists for () > 0. For () = 0, we find
d¢ (0) =lim-
d() tlO t o
11t
LT(s) ds = LT(O)¢ = L¢.
This proves the lemma. D
We remark that more information on the properties of the semigroup

T(t) fort E [0, r] is contained in Theorem 6.1 of Section 3.6.
From the adjoint theory developed in Section 6.3, we know that the
adjoint semigroup T*(s) = T(s)* corresponds to the Volterra equation
(1.5) y(s) + 1° y(T)ry(s- T) dT = g(s), s::::: 0,
with g E E 0 . Recall that Eo denotes the Banach space of row-valued func-

tions 'ljJ : ( -oo, OJ ____, m,n* that are constant on ( -oo, -r], of bounded vari-
ation on [-r, 0], continuous from the left on ( -r, 0), and vanishing at zero
with norm Var [-r,o] '1/J.
The adjoint semigroup T* (s) has the following explicit representation
g(B- s)
(1.6) T*(s)g(()) = { - f~s I: d[g(a)]X(a- T)ry(()- T) dT if()< 0;
0 if()= 0.
Here X denotes the fundamental matrix solution to System (1). So it is clear
that T*(t) is not a Co-continuous semigroup. Later, in the decomposition
theory, the spectral properties of both A and its adjoint A* have to be
studied. The adjoint A* of A is defined by f E V(A*) if and only if g E Eo
exists such that
(!,A¢)= (g,¢)
for all¢ E V(A) and in that case A* f =g.
Lemma 1.3. The adjoint operator A* : V (A*) ____, Eo is given by
V(A*) = {f E Eo: ~ E Eo},

(1.7)
A* f(B) = f(O- )ry(B)- ~(B), -r:::; ():::; 0.
Proof. It is obvious that given the action of A*, the domain of definition of
A* cannot be larger than the subspace { f E B 0 : df /dB E B 0 }. So assume
that g E B 0 with g( -r) = 0, and
f(B) = f(O-)- leo g(s) ds, for B < 0.
From Relation (1.2)
(!,A¢) =£:A¢( B) df(B)
= L<f>f(O-) + £: ~: (B)g(B) dB
=[or dry(B)</>(B)f(O-)- £:¢(B) dg(B)
=£:</>(B) d[f(O- )ry(B)]- £:</>(B) dg(B)
= (A* J, </> ).
This proves the lemma. 0
In this chapter, we shall also study the semigroup associated with the
transposed equation. Define C' = C([O, r], IR.n*). For each s E [0, oo) let y 8
designate the element inC' defined by y 8 (~) = y( -s + ~), 0::; ~::; r. The
transpose of System (1.1) is defined to be
iJ(s) =[or y(s- B) d[ry(B)], s ::; 0,

(1.8)
'!f;EC'.
Let y be a solution of Equation (1.8) on an interval ( -oo, r]. Following the

ideas in this section, we can associate a C0 -semigroup TT (s) with Equation
(1.8), the transposed semigroup, defined by
(1.9)
where y is the solution of Equation (1.8). In precisely the same way as we
proved Lemma 1.2, one can prove the following result.
Lemma 1.4. The solution operator TT(s), s :2: 0, defined by Relation (1.9)
£:
is a Co -semigroup with infinitesimal generator
D(AT) = {'!f; E C': ~~ E C', ~~(0) = - '!f;(-B)d[ry(B)l},

(1.10)
A T.f, = - d'lj;
'f' d~.
7.2 Spectrum of the generator-Decomposition of C 197
Furthermore TT (s) is completely continuous (compact) for s ::::: r.
By integration, Equation (1.8) can be written as a Volterra integral

equation
s:::; 0,
-1°I:
where pT : C' ___. B 0 is given by
(1.11) (FT '1/J)(s) = '1/J(O) '1/J(a- ()) d[ry((J)] da.
Since pT cannot be an invertible mapping from C' onto B 0 , the transposed

equation (1.8) is not the same as the formal adjoint equation (1.5), but
the equations are closely related. From the definitions, it is not difficult to
verify that
(1.12)
7.2 Spectrum ofthe generator-Decomposition ofC
It is our objective to determine the nature of a(T(t)) and a(A) for the
solution operator arising from Equation (1). To introduce the spectra, we
have to work with complex Banach spaces. Let Bee be a complexified real
Banach space, and let Bee: 'D(Bee) ___.Bee be a complexified linear operator.
By this we mean that there exist a real Banach space B and an operator
B : 'D(B) ___. B such that Bee = B EB iB and Bee(bl + ib2) = Bb1 + iBb2 for
b1 , b2 E 'D (B). For Bee, we can define the complex conjugate, denoted by an
overbar, and given by b1 + ib2 = b1 - ib 2 • Whenever there is no confusion,
we shall write, by abuse of notation, B for Bee and B for Bee.
The resolvent set p(B) of B is the set of values in the complex plane
for which the operator >..I - B has a bounded inverse with domain dense
in B. This inverse will be denoted by R(>.., B), >.. E p(B), and is called
the resolvent of B. The complement of p(B) in the complex plane is called
the spectrum of B and is denoted by a(B). The spectrum of an operator
may consist of three different types of points, namely, the residual spectrum
Ra(B), the continuous spectrum Ca(B), and the point spectrum Pa(B).
The residual spectrum consists of those >..in a( B) for which R(>.., B) exists
but 'D ( R( >..,B)) is not dense in B. The continuous spectrum consists of
those >.. in a(B) for which >..I- B has an unbounded inverse with dense
domain. The point spectrum consists of those>.. in a( B) for which >..I- B
does not have an inverse.
For A E Pa(B), the generalized eigenspace of A will be denoted by

M>.(B) and is defined to be the smallest subspace of B containing all el-
ements of B that belong toN( (AI- B)k ), k = 1, 2, .... The dimension
of M>.(B) is called the algebraic multiplicity of A and the smallest integer
k such that N( (AI- B)k) = N( (AI- B)k+l) is called the ascent of A.
Points of the point spectrum of A are called eigenvalues. Isolated points of
the point spectrum of A with finite-dimensional generalized eigenspace are
called eigenvalues of finite type or normal eigenvalues.
Lemma 2.1. If A is defined by Equation (1.2), then a(A) = Pa(A) and A is
I:
in a(A) if and only if A satisfies the characteristic equation
(2.1) det Ll(A) = 0, Ll(A) =AI- e>.o d17(0).
For any A in a(A), the generalized eigenspace M>.(A) is finite dimensional

and there is an integer k such that M>.(A) = N( (AI- A)k) and we have
the direct sum decomposition:
C = N( (AI- A)k) EB n( (AI- A)k).
Proof. Let 'ljJ = R(A, A)¢. From Relation (1.2), it follows that (AI -A)'ljJ = ¢
if and only if 'ljJ satisfies the differential equation
with boundary condition A'ljJ(O) - L'ljJ = 0. Define
'ljJ(O) = e>. 0 '1jJ(O) + foo e>.(O-s)¢(s)ds, .,--r ~ () ~ 0.
Then 'ljJ satisfies the differential equation and the boundary condition be-
+I: +()
comes
Ll(A)'l/J(O) =c d[rJ(O)Jfo-O e->.s¢(s ds.
If det Ll(A) # 0 this equation can be solved. So
{A E <C : det Ll(A) # 0} c p(A).

To prove the reverse inclusion, choose A E <C such that det Ll(A) = 0 and
define
for - h ~ t ~ 0,
where ¢ 0 # 0 is an element of the null space of Ll(A). Then
A¢=¢= A¢.
Therefore, we conclude that A E Pa(A). This proves the first part of the
lemma.
7.2 Spectrum of the generator-Decomposition of C 199
The characteristic function det L1 is an entire function and therefore

has zeros of finite order. So the representation for the resolvent implies that
the resolvent R(z, A) has a pole of order k, k ~ ko at >.o if >.o is a zero of
det L1 of order k0 . Since A is a closed operator, it follows from Hille and
Phillips [1, p.306] that M>. (A) is finite dimensional and has the properties
stated in the lemma. D
From Lemma 2.1, we know that >. in O"(A) implies that M>. is finite
dimensional and M;.(A) = N( (>.I -A)k) for some integer k. The subspace
M;.(A) satisfies AM;.(A) <;;;: M;.(A) since A commutes with R(>.,A). Let
M;.(A) have dimension d, let ¢~, ... , ¢2 be a basis for M;.(A) and let
P;. = {¢~, ... ,¢2}- Since AM;.(A) <;;;: M;.(A), there is ad x d constant
matrix B;. such that AtP;. = P;.B;..
Lemma 2.2. The only eigenvalue of B;. is >..

Proof. For any d-vector a, (>.I- A)ktJ>;.a = 0 and so P;.(>.I- B;.)ka = 0
for all d-vectors a. Therefore, (>.I- B;.)ka = 0 for all d-vectors a. But this
implies (>.I- B;.)k = 0 and the result follows. D
From the definition of A in Expression (1.2), the relation AtP;. = P;.B;.

implies that
-r ~ (} ~ 0.
From Lemma 1.1(iii), one also obtains
fort :2: 0, which together with the expression for P;., implies that
-r ~ (} ~ 0.
This relation permits one to define T(t) on M;.(A) for all values oft in
(-oo, oo). Therefore, on the generalized eigenspace of an eigenvalue of
Equation (1), that is, an element of O"(A), the differential equation has
the same structure as an ordinary differential equation.
From Lemma 1.1, we also know that T(t)A¢ = AT(t)¢ for all ¢ in
D(A). This implies that R( (>.I- A)k) is also invariant under T(t). By a
repeated application of the preceding process we obtain
Theorem 2.1. Suppose A is a finite set {>. 1 , ... , Ap} of eigenvalues of Equa-
tion {1) and let tP A = {P>. 1 , . • • , P;.p}, BA = diag(B;. 1 , ••• , B;.p), where P>.j
is a basis for the generalized eigenspace of Aj and B;.j is the matrix defined
by AtP;.j = P;.j B;.j, j = 1, 2, ... , p. Then the only eigenvalue of B;.j is Aj
and, for any vector a of the same dimension as tP A, the solution T( t )tP Aa
with initial value tP A a at t = 0 may be defined on (-oo, oo) by the relation
T(t)<l>Aa = <l>AeBAta,
(2.2)
<f>A(B) = <f>A(O)eBAO, -r:::; B:::; 0.
Furthermore, there exists a subspace QA of C such that T(t)QA ~ QA for
all t ~ 0 and
c = PA ffiQA,
where PA = {¢ E C I¢= <l>Aa, for some vector a}.
Theorem 2.1 gives a very clear picture of the behavior of the solutions
of Equation (1). In fact, on generalized eigenspaces, Equation (1) behaves
essentially as an ordinary differential equation and the decomposition of C
into two subspaces invariant under A and T(t) tells us that we can separate
out the behavior on the eigenspaces from the other type of behavior. The
decomposition of C allows one to introduce a direct sum decomposition
of C, which plays the same role as the Jordan canonical form in ordinary
differential equations. As we know in ordinary differential equations, this is
very important for studying systems that are close to linear.
The decomposition of C will be complete provided that we can explic-
itly characterize the projection operator defined by this decomposition. We
shall also need bounds for T(t) on the complementary subspace QA in order
to apply the results to perturbed linear systems. In the next two sections,
we shall address these questions for a more general class of generators than
given by Relation (1.2). The reason for this extension becomes clear when
we study neutral and periodic functional differential equations.
7.3 Characteristic matrices and equivalence
In the last section we saw that the spectrum of the infinitesimal generator A
defined by (1.2) is precisely given by the roots of the characteristic equation
det L1(z) = 0, L1(z) = zi- j_or e>.O dry(B).
We call L1(z) the characteristic matrix for A. If >.o is an eigenvalue of A, then

>.0 is a characteristic value of L1(z); that is, the matrix L1(>.o) is singular.
In this section, we shall prove that the null space of L1(>.o) describes the
complete geometric structure of the generalized eigenspace M>.(A).
A naive approach to prove such a theorem would be to try to find an
equivalence
L1(z) = F(z)(zi- A)E(z), z E <C,
where E(z) and F(z) are invertible operators that depend analytically on
the parameter z. But, of course, this would imply that C is finite dimen-
sional, which is a contradiction. So it is obvious that we have to associate
with L1(z) a simple operator acting on an infinite-dimensional space.
7.3 Characteristic matrices and equivalence 201
Set C = IRn x C endowed with the product norm topology
Define the embedding j : C-+ C by¢ r--+ (¢(0), ¢).~The solution operator
T(t) : C-+ C induces a solution operator on jC C C by
jT(t)¢ = T(t)j¢.
It is not _possible to extend the solution operator on jC to a solution oper-
ator on C since Xt( ·; c, ¢) has a discontinuity at -t, 0 :::; t :::; r. Indeed, the
operator A : V (A) -+ C
A\ ~ d¢
V(A; = {(c, ¢) E C: dB E C, c = ¢(0)},
(3.1)
~ d¢
A(c,¢) = (L¢, dB)
is a closed unbounded operator, but the domain of A is not dense in C, and

hence, A is not the generator of a semiffroup. The closure of the domain is
precisely given by jC and the part of A in jC is given by Relation (1.2).
~ For the complexified operators, it Jollows that the s_pectrum of A and
A are the same and jM>.(A) = M>.(A), where jC-+ C denotes the em-
bedding¢ r--+ (¢(0), ¢).So the spectral analysis of A and A are one and the
same.
We shall prove that the operator A is equivalent to
where I denotes the identity on C. This result solely depends on the struc-
ture of the operator A. For this reason we present the results in this section
for a more general class of operators that includes the infinitesimal gener-
ators associated with neutral functional differential equations.
There is a general scheme to construct characteristic matrices for a
rather general class of unbounded operators. For this purpose, we need
auxiliary operators D, L, and M.
The operator M : V ( M) -+ B is a closed linear operator acting in a
complex Banach space B and M is assumed to satisfy the following two
conditions:
(Hl) N := N ( M) is finite dimensional and N =1- { 0};
(H2) The operator M has a restriction M 0 : V(Mo) -+ B such that
(i) V(M) = N ffi V(Mo),
(ii) fl := p(Mo) =/- 0.
Apart from M we need two bounded linear operators
where n = dimN. One may think about M as a maximal operator and

about D and L as generalized boundary-value operators.
With D, L, and Mas earlier, we associate two operators A: V(A) --+ B
and A: v(.A) --+ B, where fj =en X B. The definitions are as follows
V(A) = { ¢ E C: ¢ E V(M), DM¢ = L¢}, A¢=M¢

(3.2) V(.A) = {(c,¢) E C: ¢ E V(M), c = D¢},
A(c, ¢) = (L¢, M¢).
The operators A and A are well-defined closed linear operators and are
closely related. In fact, if j : B--+ f3 denotes the embedding¢~--+ (D¢, ¢),
then j A = Aj. We shall refer to A and A, respectively, as the first and
second operator associated with D, L, and M.
Next, we define the candidate for the characteristic matrix function.
Let l : en --+ N be some isomorphism, and set
(3.3) Ll(z) = -(zD- L)M0 (zi- Mo)- 1 l, zED.
Here M 0 is the operator appearing in Hypothesis (H2) on M and Dis as

defined in (H2).
Let B(A) denote the domain of A provided with the graph norm
def ~
II·IIA= II·II+IIA ·II·

Since A is closed, B(A) becomes a Banach space and A : B(A) --+ f3 a
bounded operator.
We then have the following theorem.
Theorem 3.1. Suppose that A: V(.A) --+ f3 is the second operator associated
with D, L, and M. Then the matrix function Ll(z) defined in (3.3) is a
characteristic matrix for A and the equivalence is given by
(3.4) ( Ll~z) I~) = F(z)(zi- A)E(z), zED,
where E(z) : f3--+ B(A) and F(z) : f3--+ f3 are bijective mappings that de-
pend analytically on z in D. Furthermore, these operators have the following
representation
E( ) (c) = ( -DMo(zi- Mo)- 1 lc + D(zi- M 0 )- 1 ¢)

z ¢ -Mo(zi- Mo)- 1 lc + (zi- Mo)- 1 ¢ '
(3.5)
F(z) (;) = ( c- zD(zi- Mo)- 1: + L(zi- M 0 )- 1 ¢) .
7.3 Characteristic matrices and equivalence 203
Given the formulas for E(z) and F(z), the theorem above is easy to
verify directly. With Equivalence (3.4), the problem to determine the struc-
ture of the generalized eigenspace M.>.(A), ).. E O"(A), has been reduced to
the structure of the null space of ..:1(.>..) when det ..:1(.>..) = 0.
The following corollary holds for analytic matrix functions and there-
fore, using (3.4), for A and A.
Corollary 3.1. Let A satisfy the assumptions in Theorem 3.1. If det ..::::l(z) =/:.
0, then
(i) The set O"(A) n [l consists of eigenvalues and
O"(A) n [l = {z E [l: det..::::l(z) = 0};
(ii) For .Ao E O"(A) n D,

dimM.>. 0 (A) = m,
where m = m(.Ao, ..:1), the order of .>..a as a zero of det ..:1;
(iii) For .>..a E O"(A) n D, the ascent of .>. 0 equals k; that is,
where k = k(.>..o, ..:1), the order of .>..a as a pole of ..:1- 1 .
To proceed further, we must analyze the null space of ..:1(.>..). Let B1 and
B2 be complex Banach spaces and let K(z) : B1 ---+ B2 be an operator-valued
function that depends analytically on z in D. For example,
K(z): <Dn---+ <Dn, K(z) = L1(z),
or
K(z): v(A) ---+ c, K(z) = z- A
with [l = <D and A defined by (3.1). A point ).. 0 is called a characteristic
value of K(z) if there exists a vector x 0 E B 1 , x 0 -f=. 0, such that,
(3.6) K(.>..o)xo = 0.
An ordered set (xo, x 1 , ... , xk-d of vectors in B 1 is called a Jordan chain

for K(.>..o) if x 0 -f=. 0 and
(3.7) K(z)[xo + (z- .Ao)xl + · · · + (z- .Ao)k-lxk-1] = O((z- .Ao)k).

The number k is called the length of the chain and the maximal length of
the chain starting with x 0 is called the mnk of x 0 . The function
k-1
L(z- .Ao) 1xl
l=O
in (3.7) is called a root function of K corresponding to .>. 0 . We remark that

if K(z) = z- A and if (xo, x1, ... , Xk-d is a Jordan chain for z- A at .>. 0 ,
then Xj EN( (.>.. 0 ! - A)H 1 ), j = 0, 1, ... , k- 1.
Let .1(z) : (Cn---> (Cn be defined by Equation (2.1). If .>. 0 is an isolated
characteristic value of .1(z), then the Jordan chains for -1(.>.. 0 ) have finite
rank and we can organize the chains according to the procedure described
by Gohberg and Sigal [1]. Choose an eigenvector, say x 1,0 , with maximal
rank, say r1. Next, choose a Jordan chain (x 1,0 , ... ,x1,r 1 -1) of length r 1
and let N 1 be the complement inN( -1(.>.. 0 )) of the subspace spanned by
x 1,o. In N 1 we choose an eigenvector x 2,0 of maximal rank, say r 2, and
let (x2,o, ... , x2,r 2 -1) be a corresponding Jordan chain of length r 2. We
continue as follows: let N2 be the complement in N 1 of the subspace spanned
by x2,o and replace N1 by N2 in the described procedure.
In this way, we obtain a basis {x1,o, ... ,xp,o} of N(-1(.>.. 0 )) and a
corresponding canonical system of Jordan chains
(3.8)
for .1(.Ao).
Lemma 3.1. Let A satisfy the assumptions in Theorem 3.1. If det .1(.>..0 ) = 0,
then there is a one-to-one correspondence between the Jordan chains of z- A
and L1(z) at .>. 0 .
Proof. Since relation (3.4) implies that the null spaces N( .1(.>.. 0 )) and
N( .>. 0 -A) are isomorphic, it suffices to show that there is a one-to-one
correspondence between the Jordan chains of length k, k;:::: 1, of z- A and
-1 at .Ao.
Let (x 0 , ... , Xk-d be a Jordan chain for z- A at .>. 0 of length k. The
equivalence relation (3.4) implies that
k-1
(3.9) L1(z)E(z)- 1 L:)z- .>..o) 1xt = O((z- A.o)k)
l=O
for lz- .Aol---> 0. If I:7~~(z- .>..o) 1Yt denotes the Taylor expansion of order
k around z = .>. 0 for the holomorphic function
k-1
E(z)- 1 l:(z- .>..o) 1xt,
l=O
then
k-1
L1(z) l:(z- A.o) 1Yt = O((z- A.o)k) for Iz - .Ao I ---> 0
l=O
and (y 0 , ... , Yk-d is a Jordan chain for -1 at .>. 0 of length k. So we proved

that a Jordan chain for z- A at .>. 0 of length k induces a Jordan chain for
7.4 The generalized eigenspace for RFDE 205
L1 at >. 0 of length k. Since the roles of z- A and L1 can be interchanged,

the proof is complete. 0
7.4 The generalized eigenspace for RFDE
As a first application, we apply the result from Section 3 to retarded func-

tional differential equations.
Lemma 4.1. Let M: V(M) --+ C be the operator defined by
d¢
V(M) = {¢ E C: d() E C},
Then M satisfies hypotheses (H1) and (H2) in Section 7.3 with M 0

V(Mo) --+ C defined by
V(Mo) = { ¢ E V(M) : ¢(0) = 0}, Mo¢ = M¢

and fl = <C. Furthermore, the infinitesimal generator A defined by (1.2)
and the infinitesimal generator A defined by (3.1) are the first and second
operator associated with D, L, and M, where D¢ = ¢(0).
Proof. Clearly, the kernel of M consists of the constant functions. It follows
that N = N( M) has dimension n. We have V(M) =NEB V(Mo), and
for each z E <C the operator z - M 0 is invertible and the resolvent of M 0 is
given by
(4.1) ((zi- Mo)- 1 ¢)(B) =leo e<O-a)z¢(rr) drr, -r:::; ():::; 0.
Thus M satisfies (H1) and (H2) with [l =<C. Furthermore, if we set D¢ =

¢(0), then A is given by
V(A) = {¢ E C: ¢ E V(M), DM¢ = L¢}, A¢=M¢.
So A is the first operator associated with D, L, and M. Finally, it is clear

that A defined by (3.1) is the second operator associated with D, L, and
M. o
Theorem 4.1. The matrix function
(4.2)
is a characteristic matrix for the infinitesimal generator A defined by Re-

lation (3.1). The equivalence is given by
(4.3) ( L1bz) ~) = F(z)(zi- A)E(z), zE<C
with E(z) : C--+ V(A)
(4.4) E(z)(c, ¢)(0) = (c, e9 zc +leo e<O-u)z¢(a)da)
and F(z) : C--+ C

(4.5) F(z)(c, ¢) = (c + L(zi- Mo)- 1 ¢, ¢).
Proof. Let M and M 0 be as in the statement of Lemma 4.1. Define l : IRn --+
N,N=N(M), by
(lc)(O) = c, -r ::; ()::; 0.
From Theorem 3.1, we know that
L1(z) = -(zD- L)Mo(zi- M0 )- 1 l, z E <C,
is a characteristic matrix for A. To verify the concrete representation (4.2)

we use the resolvent formula (4.1) for Mo and calculate
-(zD- L)M0 (zi- M 0 )- 1 lc(O) = (lc- z(zi- M0 )- 1 lc)(O)

=c+z 1 9
e<O-u)zcda
= c- e(O-u)zcl~
= e8 zc, -r::; () ::; 0.
Finally, the concrete representations for E(z) and F(z) are verified in a
similar way. 0
From the equivalence relation (4.3), we have the following representa-

tion for the resolvent of A.
Corollary 4.1. Let A: V(A) --+ C be the generator defined by Relation (1.2).
The resolvent of A has the following representation
for -r ::; () ::; 0, where

7.4 The generalized eigenspace for RFDE 207
Since the spectral analysis of A and A are one and the same, we have
actually characterized N( (>.I- A)k) in a manner that is convenient for
computations.
Theorem 4.2. The spectrum of the infinitesimal generator A defined by (1.2)

consists of eigenvalues of finite type only,
(4.8) a( A) = {A : det ..1(>.) = 0}.

For>. E a( A), the algebraic multiplicity of the eigenvalue>. equals the order
of>. as a zero of det ..1, the ascent of>. equals the order of>. as a pole of ..1- 1 .
Furthermore, a canonical basis of eigenvectors and generalized eigenvectors
for A at >. may be obtained in the following way: If
{(')'i,O, · · ·, 'T'i,k;-1) : i = 1, ... ,p}

is a canonical system of Jordan chains for ..1 at>., then
i = 1, ... ,p,
where
()l
I: 'T'i,v-llf
v
(4.9) <Pi,v (o) = e).()
l=O
yields a canonical basis for A at >..

Proof. In order to apply Corollary 3.1, we first show that det ..1 ¢. 0. From
1:
the representation (4.2) for ..1(z), it follows that it suffices to prove
lz- 1 eztdry(t)l ~0
as Re z ~ oo. But this is obvious since 17 is of bounded variation.
Next we prove the representation for the canonical basis for A at >..
Let ('Yo, ... , 'T'k-d be a Jordan chain for ..1 at >.of length k. Definer(>.) =
'Yo + 1'1 (z - >.) + · · · + 'T'k-1 (z - >.)k- 1 . Since
E(z)(r(>.),o) = (r(>.),ez· r(>.)),
we derive from the equivalence that
for lz - >.1 ~ 0.
From Lemma 3.1, it follows that there is a one-to-one correspondence be-
tween the Jordan chains of ..1(z) at >.and the Jordan chains of z- A at >..
So we have to expand ez · up to order k in a neighborhood of >.. Since
ok-1
ezo = e>- 0 [1 + O(z- >.) + · · · + (k _ 1)! (z- >.)k- 1 + O((z- >.)k)],
the Jordan chain for (zi- A) at A becomes { ¢ 0, ... , </Jk-1}, where
and it becomes clear that this procedure yields a canonical basis for A at
A from a canonical system of Jordan chains for .1(z) at A. D
From Theorem 4.2, we see, in particular, that there is a one-to-one

correspondence between the Jordan chains of .1(z) at A oflength k and the
solutions of the equation
The Jordan chains of length 1 are precisely the vectors in the null space
of .1(A) and already in Lemma 2.1 we saw that if b E N( .1(A)), then
e>-. 0 b satisfies (AI- A)e>-. 0 b = 0. The higher-order Jordan chains can also be
characterized by vectors in the null space of a certain matrix. Define
.1(j)(z) = dJ.Ll(z), j = 0, 1, ...

dzJ
and the matrices Ak of dimension ( kn) x (kn)
k = 1,2, ....
Then ('-ro, ... , /'k-d is a Jordan chain of length k if and only if
Akbo ... /'k-1f = 0.
7.5 Decomposing C with the adjoint equation
In the previous section, in which we proved Theorem 4.2, we have seen that
we can characterize the generalized eigenspace of A corresponding to an
eigenvalue A. In this section, it is our goal to compute the corresponding
projection onto this generalized eigenspace. For this, we use that
k = 1,2, ....
Therefore, we start this section with the spectral analysis of the adjoint
operator A* : V(A*) --'> B 0 . Let Eo = 1Rm x Bo be the dual space of C,
with the pairing
7.5 Decomposing C with the adjoint equation 209
(5.1) ((a, f), (c, ¢)) = ac +I: d[f(O)] ¢(0)
where (a, f) E B0 , (c, ¢) E C. As we observed in the previous section, the

operator A is similar to the part of A in jC. This implies that the operator
A* : V (A*) ---. B 0 is similar to the part of the operator A* : Eo ---. C (A)* to
j* E 0 • Here C(A) denotes the domain of A provided with the graph norm.
In particular, the spectral analysis of A* and A* are one and the same.
Together with the properties of the adjoint operation-<. the equivalence
(4.3) for A implies the following equivalence relation for A*:
(5.2) (a f) ( Llbz) ~) = E(z)*(zi- A*)F(z)*(a, f), (a, f) E Eo
where E(z)* : C(A)* ---. Eo and F(z)* : Eo ---. Eo are bijective mappings
that depend analytically on z, z E <C.
In particular, we find that the Jordan chains of length k of LlT at z = ..\
are in one-to-one correspondence with the solutions of
(5.3) (..\I- A*)k'l/J = 0.

More precisely, we have an explicit representation for the eigenfunctions
and generalized eigenfunctions of A*. This is the contents of Theorem 5.1.
First we compute F(z)* : Eo ---.Eo.
Lemma 5.1. The mapping F(z)* : Eo ---.Eo is bijective, depends analytically
I:
on z, z E <C, and is given by F(z)*(a,f) =(a, g), where
(5.4) g(O) =leo d[ary(T)]e-(o--r)z da + f(O).
Proof. Since F(z) : C ---. C is bijective and depends analytically on z,

z E <C, it follows that F(z)* : Eo ---.Eo has the same properties. It remains
to compute F(z)*. From the definition
( F(z)*(a, f), (c, ¢)) = ((a, f), F(z)(c, ¢))

(5.5)
= a[c + L(zi- Mo)- 1 ¢] +[or ¢(0) d[f(B)].
Using Fubini's theorem to reverse the order of integration, we have the

following identity
1°
I:
aL(zi- Mo)- 1 ¢ = ad[ry(B)] { 0 e-(o--O)zrp(a) da
-r Jo
=[or rp(a) ad[ry(B)]e-(o--O)z da,
Substitution of this identity into (5.5) yields
(5.6) (F(z)*(a,f),(c,¢)) =ac+ [or cj;(B)d[g(B)],
where g is given by (5.4). This proves the lemma. 0
Theorem 5.1. The spectrum of the infinitesimal generator A* defined by

(3.1) consists of eigenvalues of finite type only,
(5.7) O"(A*) ={A: det Ll(A) = 0}.
For ).. E O"(A*), the algebraic multiplicity of the eigenvalue A equals the
order of A as a zero of det Ll, the ascent of A equals the order of A as a
pole of Ll- 1 . Furthermore, a canonical basis of eigenvectors and generalized
eigenvectors for A* at ).. may be obtained in the following way: If
{(f3[o,
'
· · ·, f3[k' ' -1): i = 1, · · · ,p}
is a canonical system of Jordan chains for LlT at A E <C, then
Xi,O, · · ·, Xi,k,-1, i=1, ... ,p,
where Xi,v = FT '1/Ji,v with
C) - ->.~ ~ (3 ( -~)l
'1/Ji,v (" - e ~ i,v-l_l_!-'
l=O
yields a canonical basis for A* at A. Here FT : C[O, r] -+ B 0 is the mapping

from an initial condition into a forcing function of the integrated equation
introduced in Section 7.1
(5.8)
Proof. The first part of the proof is the same as the proof of Theorem
f3L
4.2. Let ((3{;, ... , 1 ) be a Jordan chain for LlT at ).. of length k and let
a(z)T = (3{; + (3f(z- A)+···+ f3L
1 (z- A)k- 1 denote the corresponding
root function. From the equivalence relation (5.2), we have F(z)*(a(z),O)
is a root function for A*. Therefore,
is a root function for z- A* at >-.. From Lemma 3.1, it follows that there is
a one-to-one correspondence between the Jordan chains of Ll(zf at A and
the Jordan chains of z- A* at A. So we have to expand
7.5 Decomposing C with the adjoint equation 211
up to order k in a neighborhood of A. A similar expansion, as in the proof

of Theorem 4.2, shows that
is a root function for z- A*. Since A* : V (A*) ---+ B 0 is similar to the part of
A* in j* B0 , it remains to compute the adjoint of j: C---+ C, j¢ = (¢(0), ¢).
An easy computation shows that j* : B0 ---+ B 0 is given by
'*( !)(e) {o fore=o,

J a, = a+ f for -r ::::: e < 0.
So the corresponding Jordan chain for z- A* at A becomes {xo, ... , xk},

where
and it becomes clear that this procedure yields a canonical basis for A* at
A from the canonical system of Jordan chains for Ll(z)T at A. This proves
the theorem. D
From the preceding theorem, we conclude that the (generalized) eigen-

functions of A* are precisely given by the images under the operator pT
of the (generalized) eigenfunctions of the transposed generator AT. This
is to be expected from the adjoint theory in Section 6.3. In particular,
T*(t)FT = FTTT(t), and hence
In general, the mapping pT is not one-to-one (see Section 3.3), but, on the
generalized eigenspace of AT, it is. Hence, if'¢ is a (generalized) eigenfunc-
tion of AT, then FT'¢ is a (generalized) eigenfunction of A*. This moti-
vates the introduction of the following bilinear form (see also Hale [22]).
For '¢ E C' and ¢ E C define
('¢,¢)~f(FT'¢,¢) = j_or d[FT'¢(e)]¢(e)

(5.9) = '¢(O)¢(o)- j_: Je '¢(e- T) d[ry(T)]¢(e) de
= '¢(0)¢(0)- j_or foe '¢(e- T) d[ry(T)]¢(e) de

between C and C'. With respect to this bilinear form, the transposed op-
erator AT satisfies
and we have proved the following lemma.
Lemma 5.2. For .A in a( A), let lJi>. =col (7/! 1 , ... , 7/!p) and<~>>.= (cfJI, ... , r/Jp)
be bases for M>.(AT) and M>.(A), respectively, and let (lJi>., <~>>.) = (7/!i, cpj),
i,j = 1,2, ... ,p. Then (lJi>-.,<1>>.) is nonsingular and thus may be taken as
the identity. The decomposition of C given by Lemma 2.1 may be written
explicitly as
¢ = ¢P>- + ¢Q>-, ¢P>- in P>., ¢Q>. in Q>.,
P>. = M>.(A) = {¢ E C: ¢ = <l>>.b for some p-vector b},
Q>. = {¢ E C: (lJi>.,¢) = 0},
cpp>- = lf>>.b, b = (lJi>., cp), cpQ>. = cp- cpp>-.
It is also interesting to note that (lJi>., <~>>.) = I, and ATlJi>. = B~lJi>. and
A. = <l>>.B>. implies B~ = B>.. In fact,
(lJi>., A.) = (lJi>., <l>>.B>.) = (lJi>., <l>>.)B>. = B>.

= (ATlJi>.,<l>>.) = (B~lJi>-.,<1>>.) = BHlJi>-.,<1>>.) = B~.
We have already defined the generalized eigenspace of a characteristic

value of Equation (1) as the set M>.(A). If A= {>. 1 , ... , .Ap} is a finite set of
characteristic value of Equation (1), we let P = PA be the linear extension
of the M>.j (A), Aj E A, and refer to this as the generalized eigenspace of
Equation (1) associated with A. In a similar manner, we can define pT =PI
to be the generalized eigenspace of the transposed equation (1.8) associated
with A. If and lJi are bases for P and PI, respectively, (lJi, ) = I, the
identity, then
C=PAEBQA
(5.10) PA ={ E C: = <Pb for some vector b}
QA = {<!> E C: (lJi,¢) = 0}
and, therefore, for any¢ inC,
cp = cppA + cpQA
(5.11)
cppA = lf>(lJi, ¢) •
When this particular decomposition of C is used, we shall briefly express
this by saying that C is decomposed by A.
7.6 Estimates on the complementary subspace 213
7.6 Estimates on the complementary subspace
If C is decomposed by A, we know, from Theorem 2.1, that there is a

constant matrix B = B A whose eigenvalues coincide with A such that
For the application of the theory of linear systems, we need to have an

estimate for the solutions on the complementary subspace QA. Such an
estimate requires detailed knowledge of the spectrum of T(t). In particular,
we need to know the spectral radius of the semigroup T(t) restricted to QA.
A first step in this direction is answered by the following result:
Lemma 6.1. If the semigroup T( t) is strongly continuous on [0, oo) with

infinitesimal generator A, then Pa(T(t)) = etPu(A) plus possibly {0}. More
specifically, if f.L = f.L(t) -1- 0 is in Pa(T(t)) for some fixed t, then there is
a point >. in Pa(A) such that e>.t = f.L· Furthermore, if An consists of all
distinct points in Pa(A) such that e>-nt = f.L, then N( (f.LI- T(t))k) is the
closed linear extension of the linearly independent manifolds N( (f.LI -A)k).
Proof. Lemma 4.1 is a special case of Theorem 16.7.2, p. 467 of Hille and
Phillips [1] for k = 1. The reader may complete the proof for arbitrary
k. D
The spectral radius p of an operator T mapping a Banach space into

itself is the smallest disk centered at the origin of the complex plane that
contains a(T).
We also need the following
Lemma 6.2. If T(t) is a strongly continuous semigroup of operators of a

Banach space B into itself, if for some r > 0, the spectral radius p =
PT(r) -1- 0 and if f3 is defined by f3r = log p, then for any "( > 0, there is a
constant K('Y) 2: 1 such that
IIT(t)¢11 :::; K('Y)eCB+-y)tll¢11, for all t 2: 0, ¢> E B.
Proof. Since T(t) is strongly continuous, it is certainly bounded for each t

and, in particular, T(r) is bounded. It then follows that
P = ef3r = lim IITn(r)lll/n.

n---+oo
Therefore, for any 'Y > 0,

e--yr = lim e-((3+-y)r IITn(r) 111/n
n---+oo
and there is a number N such that

where e- 1 r +En ::::; L < 1 for all n : : ": N. Therefore,
as n __, oo.
Since T(t) is strongly continuous, there is a constant B such that IIT(t)ll ::::;
B for 0::::; t::::; r. Define K(r) for any 'Y to be
K(r) = Beli3+lir maxe-(i3+l)nri1Tn(r)ll·
n2:0
IfO::::; t::::; r, then, for any¢ in B,
IIT(t)¢11 S: IIT(t)ll · 11¢11 S: Bll¢11 S: K(r)e(i3+l)tll¢11·
If t : : ": r, then there is an integer n such that nr::::; t < (n + 1)r and, for all
¢in B,
IIT(t)¢11 = IIT(t- nr)T(nr)¢11 S: BIITn(r)ll · 11¢11

= [Be-(!3+/)(t-nr) e-(!3+/)nr IITn(r) II] e(i3+l)t 11¢11
::::; K(r)e(i3+1)tll¢11·
This completes the proof of the lemma. 0
If we now turn to our original problem posed before the statement of

Lemma 6.1, we obtain the following information. Since T(t) is completely
continuous fort::::": r, it follows that for any 11 in a(T(r)), 11 -=1- 0 is an element
of Pa(T(r)) and that the only possible accumulation point in a(T(r)) is
zero. Furthermore, if 11 -=1- 0 is in Pa(T(r)), then N( (f.tl- T(r))k) is of
finite dimension for every k and M"(T(r)) is finite dimensional. These
are well-known properties of completely continuous operators that can be
found in Taylor [1, pp. 180-182]. Lemmas 6.1 and 2.1 imply there are only
a finite number of >. in a( A) such that Re >. > f3 for any real number /3.
Consequently, if A= A(/3) = {>. E a( A) : Re >. > /3} and Cis decomposed
by A, then there are constants 'Y > 0 and K = K(r) > 0, such that
We summarize these results in
Theorem 6.1. For any real number /3, let A = A(/3) = {>. E a(A) : Re >. >
/3}. If C is decomposed by A as C = PA EB QA, then there exist positive
constants 'Y and K = K ('Y) such that
IIT(t)¢P11 II S: K e(i3+l)t II¢P11 II, t::::; 0,

(6.1)
IIT(t)¢Q 11 II S: Ke(i3+l)tii¢Q 11 II, t::::: 0.
7. 7 An example 215
The first of Inequalities (6.1) follows from Theorem 2.1 since we know
that T(t) can be defined on PA for -oo < t < oo and the eigenvalues of the
corresponding matrix BA associated with PA coincides with the set A.
An important corollary of Theorem 6.2 concerning exponential asymp-
totic stability is
Corollary 6.1. If all of the roots of the chamcteristic equation (2.1) of Equa-
tion (1) have negative real parts, then there exist positive constants K and
8 such that
t 2:: 0,
for all¢ in C.
Proof. The proof is obvious since, by choosing (3 < 0 in Theorem 6.1 suffi-
ciently close to 0, the set A is empty. D
Theorem 6.1 can also be obtained without such sophisticated theory.

In fact, the same approach as in Chapter 1 applies (see Chapter 9).
However, Theorems 2.1 and 6.1 give a very clear picture of the behavior
of the solutions of an autonomous linear RFDE in C. In particular, one can
choose (3 = 0 in Theorem 6.1, separate out the eigenvalues with real parts
2:: 0, and then be assured that the finite-dimensional subspace PA attracts
all solutions of the differential equation exponentially. Furthermore, the flow
on PA is equivalent to an ordinary differential equation. In fact, if iP is a
basis for PA and T(t)¢A = iPy(t), then y(t) = (exp Bt)b, where b = (lP, ¢A).
7.7 An example
Consider the scalar equation
(7.1) x(t) 10
= - 27r x(t- 1) = _1 d[ry(B)]x(t +B)
where
0 = -1,
ry(O) = { ~ -1 < 0:::; 0,
and the transposed system
(7.2) y(s) = ~y(s + 1).
The bilinear form is
(7.3) ('lj;, ¢) = 1/J(0)¢(0)- 2

7r 10
_1 'lj;(r + 1)¢(r) dr,
and the operators A, AT are given by
d¢ . 7r
V(A) = {¢ E C: d() E C, ¢(0) = -2¢(-1)},
V(AT) = {¢ E C': dd~ E C', cjJ(O) = -~¢(1)},

2
s
Moreover, ¢ is in N ()...]-A) if and only if ¢(B) = e;..e b, -r :s; () :s; 0, where
b is a constant and >. satisfies the characteristic equation
(7.4)
Also,~ belongs to N(>.I- AT) if and only if ~(7) = e->-.rc, 0 :s; 7 :s; r,
where cis a constant and >. satisfies Equation (7.4).
It is easy to prove (using the Appendix) that Equation (7.4) has two
simple roots ±i~ and the remaining roots have negative real parts.
If A= {i~, -i~}, then it is obvious that
(7.5) -1 :s; () :s; 0,

is a basis for the generalized eigenspace P = PA of Equation (7.1) associated
with A and that
1JF = col (~f, ~f), 0 :S:: 7 :S:: 1,
is a basis for the generalized eigenspace pT = PJ

of Equation (7.2) asso-
ciated with A. We wish to decompose C by A. In addition, we have seen
that the transformations are simpler if (wT, cf>) = (~J, ¢k), j, k = 1, 2, is
the identity matrix, and (~' ¢) is defined in Equation (7.3). However, if we
compute this matrix, we see that it is not the identity. Therefore, we define
a new basis l[t for by PJ
w = (wr, cf>)-1wr
and then (W, cf>) = I. The explicit expression for the basis l[t is
l[t = col(~1,~2),
(7.6) ~ 1 (7) = 2f.L[sin(~7) + ~ cos(~7)],

7r . 7r 7r 1
~ 2 (7) = 2{-l[- 2sm( 2 7) +cos( 2 7)],
If we now decompose C by A and let Q = QA for simplicity in notation,

then any ¢ in C can be written as
cp = cpp +cpQ,
(7. 7) b1 = f.L7r¢(0)- fL1r Jo

-1
7r 7r 7r
[cos( -s)-- sin( -s)]¢(s) ds,
2 2 2
b2 = 2f.l¢(0) + fL1r Jo
-1
7r
2
7r
2
7r
[sin( -s) +-cos( -s)]¢(s) ds.
2
7.7 An example 217
The explicit expressions for b1 and b2 are obtained by simply substituting

the expressions for 1[1 into Equation (7.3).
From Theorem 6.1, we know that there are positive constants K and
'Y such that
(7.8) t ~ 0.
Consequently, the subspace P of C is asymptotically stable. More specifi-

cally, with A and iP defined as earlier, we have
(7.9) Ail!= iPB, B= (

0 _1[)
2
~ 0
and therefore, T(t)iP = iPe 8 t. Since ¢Q = ¢- ¢P, ¢P = iPb, b = (w,¢), it
follows from estimate (7.8) that
exponentially as t-+ oo for every¢ E C, where b = (w, ¢)is given explicitly

in Equation (7.7). That is, any solution of Equation (7.1) approaches a
periodic function oft given by b1 sin(?rt/2) + b2 cos(?rt/2) where b1 and b2
satisfy Equations (7.7).
Fig. 7.1.
In the (x, t)-space, it is very difficult to visualize this picture, but in C,

everything is very clear. In the subspace P, T(t)¢ = iPe 8 tb, the elements
¢1 and ¢2 of iP serve as a coordinate system in P and for any initial value
Pb in P, we have T(t + 4)Pb = T(t)Pb since exp[B(t + 4)] = expBt, and

in particular, T(4)Pb = Pb; that is, the trajectories in C on Pare closed
curves. We can, therefore, symbolically represent the trajectories in C as in
Figure 7 .1. The pictorial representation of P by a (¢I, t/!2 )-plane is precise,
but it should always be kept in mind that Q is an infinite-dimensional space.
7.8 Spectral decomposition according to all

eigenvalues
If Cis decomposed by A0 = A(,60 ) = {.A E u(A) : Re.A > ,60 }, then we

know that there is a constant matrix B 0 whose eigenvalues coincide with
A 0 such that
T(t)q)Po = PoeBota, q)Po = Poa,
liT(t)q)Qo II : : ; M eflot, t ~ 0.
The projection Po= PA 0 is given by
(8.1)
where Po is a basis of eigenvectors and generalized eigenvectors of A at A0

and lf/0 is a basis of eigenvectors and generalized eigenvectors of AT at A 0
such that (lf/o, Po) =I.
Suppose we decrease ,6 and cross another pair of eigenvalues of A, say
AI = A(,BI) = A0 U {Ak, .Xk}. It is our objective to compute the new projec-
tion PI = PA, from P0 . First of all, observe that if .AI =f. >. 2 , AI =f. 3. 2 and
if ¢I EM.>., (A), 'l/;2 E M.>. 2 (AT), then ('~h, ¢1) = 0. So eigenvectors of A
and AT corresponding to different pairs of eigenvalues are "perpendicular"
with respect to the bilinear form ( ·, · ). Therefore, adding eigenvalues to A
does not affect the matrix representation for the projection P0 and
where the eigenvectors and generalized eigenvectors are normalized accord-

ing to
(8.2)
Next, it is our aim to analyze the projection Pm as ,Bm decreases to -oo.
In order to do so, we need an abstract expression for the coefficients
so that we can analyze the decay rate as Re >. tends to -oo. Since we have
explicit expressions for the eigenfunctions and generalized eigenfunctions of
A and AT, we can compute ('1/J.>., ¢)explicitly.
7.8 Spectral decomposition according to all eigenvalues 219
Lemma 8.1. For >. E a( A), let 'ljJ~, ... , '1/Jm and </h, ... , ¢m be bases for
MA(AT) and MA(A), respectively, such that ('1/Jj, cpj) = 1, for j = 1, ... , m.
The projection PA onto MA(A) is given by
m
PA¢ = "f)'I/Jj, cp)cpj
j=l
and can be written explicitly as
In particular, if>. is a simple eigenvalue, then
Proof. To avoid technical difficulties, we restrict ourselves to the case that

>. is a simple eigenvalue. The general case follows by integration by parts.
Since >.is simple, we know from Section 7.4 that
1/JI(~) = dle-A 8 , 0:::; S:::; r, d1Ll().) = 0,

¢I(O) = c1eA8 , -r:::; 0:::; 0, Ll(>.)c1 = 0.
Further, from Ll(>.)T d[ = 0 it follows that
f E <Cn.
So
(1/J1,¢) = /_: d[FT'l/J1(0)]cp(O)
= d1 [¢(0) + /_: /_8r d[ry(T)]e-A(B-r)cp(O) dO]
= JT~~Ll(z)- 1 [¢(0) + j_:j_8rd[ry(T)]e-A(B-r)cp(O)dO].

From the normalization (1/J~, ¢ 1) = 1
(1/J~,¢1) = JT~~Ll(z)- 1 [cl + j_orj_8rd[ry(T)]eA 7

CldO]
= JT :Z Ll(>.)-l [I+ /_or 10 eAr dO] C1

=!Tel= 1.
Thus
and this completes the proof of the lemma. D
As a corollary to the lemma, the spectral projection PA corresponding

to a set A = {>. 1 , ... , Am} of eigenvalues of A can be written as
(8.5) (PA¢>)(8) = f Res P(z,¢>)( 8)

z=>-. 1 det Ll(z)
j=l
where
Observe that the residues correspond precisely to the residues in Section

3.3. In particular, the spectral projection PA projects ¢> onto the initial
value of a characteristic solution corresponding to A.
To analyze the behavior of the residues, we invoke Cauchy's theorem
and rewrite the sum of residues as a contour integral. Let rA be a simple
closed smooth curve. If rA encloses A, but no other eigenvalues of A, then
PA¢>=-1-j P(z,¢>) dz
27ri rA det Ll(z) ·
Note that from Representation (4.6) for the resolvent of A, it follows that
PA as defined is the Riesz projection
PA¢> = -21 . j R(z, A)¢>dz

7rZ TA
onto the generalized eigenspace of A corresponding to A.

Define the generalized eigenspace of A to be the linear subspace M(A)
generated by all M>-.(A), i.e.,
(8.7) M(A) = EB M>-.(A).

>-.Ea(A)
The system of eigenfunctions and generalized eigenfunctions is called com-

plete if M(A) is dense inC, i.e., M(A) =C. The space M(A) is precisely
the space of initial functions such that the corresponding solution has a
finite expansion in characteristic solutions. In fact, the solution through
¢> E M(A) exists for all time and the semigroup T(t) extends to a flow
on M(A). In general, however, the space M(A) can be too small to be
interesting, can even be finite dimensional (see the example in Section 3.3).
Further, if it is infinite dimensional it is difficult to characterize and not
closed. Therefore, we turn to the closure of M(A). In Section 3.3, we claimed
that this is an invariant subspace on which the semigroup is one-to-one and

which contains all information about the equation.
First we analyze the range of the semigroup T(t) using duality. From
the results in this chapter, we know that the adjoint semigroup T* (t)
B 0 ---+ B 0 corresponds to the Volterra integral equation
y(s) + 1°
y(T)ry(s- T) dT = g(s), s::; 0,
with g E B 0 . Now note that the numbers E and a introduced in Section 3.3
are invariant under transpose; that is,
Therefore, an argument similar to the proof of Theorem 3.2 of Section 3.3

implies that the ascent of T* (t) equals E - a as well. Since
R(T(t)) =N(T*(t)).l,
we conclude that the closure of the range of T(t) becomes independent

of t for t :::0: E - a. (The range itself becomes smaller with increasing t
due to the fact that the solutions become more smooth.) By definition,
M(A) c R( T(t)) fort :::0: 0. Hence
M(A) ~ R( T(t) ).
Theorem 8.1. M(A) = R( T(E- a)).
Sketch of the proof. The orthogonal complement of M(A) is given by
M(A).i = n N( (>..I-
,\EPD"(A)
A*)m>- ).
Therefore, from the Neumann expansion for the resolvent, we have that for
--.l
any 1/J E M(A) , the function
(8.8) Zf--->R(z,A*)'l/J
is an entire function. But the resolvent is the Laplace transform of the
semigroup, so that t f---> T*(t)'ljJ(O) is a solution of the adjoint equation
which decays faster than any exponential. So an argument similar to the
proof of Theorem 3.1 of Section 3.3 implies that t f---> T*(t)'ljJ(O) is identically
zero after finite time and 1/J E N ( T* (E - a) ) . The opposite inequality is
clear and we have shown that
This proves the theorem. 0

A simple observation yields the following important corollary.
Corollary 8.1. The system of eigenfunctions and generalized eigenfunctions

of the generator of Equation (1) is complete; that is, M(A) = C, if and
only if
E(det.1(z)) = nr.
Or, equivalently, if and only if there are no small solutions to System (1).
Let rN be a simple closed smooth curve in <C and let AN be the corre-
sponding set of eigenvalues of A enclosed by rN. Assume that M(A) =C.
To analyze whether the series expansion limN-.oo PAN¢ converges to¢, we
must estimate the resolvent of A outside small circles centered around the
eigenvalues of A. From the representation
R(z, A)¢= P(z, 1>)

det .1(z)
where P(z, 1>) is given by Formula (8.6), it follows that it suffices to have
good estimates for
.1(z)_ 1 = adj .1(z).
det .1(z)
As a consequence of general properties of entire functions, the zeros of
det .1(z) cannot accumulate. So there exists a sequence of simple closed
r
smooth curves N such that
(i) There is a complex function aN : [0, 21!"] -+ <C that is differentiable
on [0, 21r] such that rN = {z = aN(t), t E [0, 21!"]}, for t E [0, 21r]
laN(t)l-+ oo as N-+ oo and la~(t)l :::; laN(t)l;
(ii) rN encloses AN but no other eigenvalues of u(A);
(iii) there exists an E > 0 such that dist (rN, u(A)) > E for all N = 0, 1, ....
In order to control the behavior of l.1- 1 (z)l as lzl -+ oo, we make the
following assumption on the kernel ry:
(J) The entries 'f/ij of 'f/ have an atom before they become constant, i.e.,
there is a tij with 'f/ij(t) = 'f/ij(tij+) fort 2 tij and 'f/ij(tij-) -=1-
'f/ij(tij+ ).
For example 'f/ can be a step function.
Theorem 8.2. Suppose that 'f/ satisfies (J) and M (A) = C. Then for 1> E
V(A 3 )
where AN denotes the set of eigenvalues enclosed by the contour rN.
Sketch of the proof. From Cauchy's theorem, it suffices to show that

¢ = hm
. -1.
N ->oo 27rz rN
1 R(z,A)¢dz.
Using the Newton polygon, one can show that if 'f} satisfies (J) and the
exponential type of det Ll(z) is equal to nr, then for every¢ E V(A) there
exists a positive constant M such that
(8.9) IR(z, A)¢1 ::::; M for zEFN, N=0,1, ....
For¢ E V(A 3 ), it follows that

1 1 1 2
R(z,A)¢ = -¢+ 2 A¢+ 2 R(z,A)A ¢.
z z z
1
Clearly,
N
lim - 1
--->00 27ri
1TN
¢dz
- =¢
Z '
l i m1-
N ->oo21ri rN
dz =0
A¢-
z2 '
and
where we used the properties of the contours rN and Estimate (8.9).

Therefore
and this proves the theorem. 0

Pointwise expansion for the solution is possible under weaker conditions.
Theorem 8.3. Suppose that 'f} satisfies (J) and M(A) =C. Then for¢ E C
the solution x( ·; ¢) to System (1) has the following expansion
= .
hm -21 .
N->oo7rZ
1 TN
ezt(R(z, A)¢)(0) dz, t > 0.
Sketch of the proof It is easy to see that for A E a( A), the solution of System
(1) corresponding to initial condition P>.¢ is given by Xt ( ·; P>.¢) = e>.t P>.¢·
In particular,
x(t;P>.¢) = e>.t(P>.¢)(0) = (P>.e>.t¢)(0)
= ~ { ezt (R(z, A)¢) (0) dz

2m lr;..
where r>.. is a small circle surrounding >.. Since the spectrum of A is con-
tained in a left half plane Re z < 'Y, we can modify the contours N by r
replacing the arc in Re z ~ 'Y by the corresponding line segment Re z = 'Y.
From the Laplace inversion formula (in ffin), it follows that
1 1'Y+ik
x(t;¢) = lim -2 . ezt(R(z,A)¢)(0)dz for t > 0.
k--+oo7rZ '}'-ik
Let eN denote the arc of the contour rN contained in the half plane Re z <
'Y. In order to prove that
x(t;¢) = lim -21 . { ezt(R(z,A)¢)(0)dz, t > 0,

N --+oo7rZ } TN
it suffices to prove
(8.10) lim -21

N--+oo 7rZ
·1 CN
ezt(R(z,A)¢)(0) dz =0 for t > 0.
Since we evaluate the resolvent of A at 0, we can improve the estimate for

R(z,A) on CN. Indeed, using the Newton polygon, it follows that
j (R(z, A)¢)(0)j :::; Mmin{ le~;z1, 1 },

The limit in Equation (8.10) now follows from a standard result, see for
example, the proof of Lemma 4.2 of Bellman and Cooke [1]. D
Without the assumption (J) the results do not hold. However, using
similar techniques one can prove that if M (A) = C, then for ¢ E V (A 00 )
¢ = N--+oo
lim PAN¢.
Since A is the generator of a C0 -semigroup, the domain V (A 00 ) is still dense,

but too small to imply a convergent series expansion for the solution for
t > 0.
We end this section with some remarks about what to do ifE(det L1(z))
is less than nr; that is, there are small solutions and M(A) =f C.
Define the following subspace of C:
(8.11) £ = {¢ E c I E(P(z,¢)(0)):::; E(det..d(z)), -r:::; 0:::; o}.

Note that, £ =f C, if and only if the ascent of T(t) is strictly positive.
Furthermore, we have the following lemma.
Lemma 8.2. The restriction of the solution operator T(t) to£ is one-to-one.
Proof. Suppose that there exists a¢ =f 0 inC such that T(a)¢ = 0 for some
a> 0. Then
7.9 The decomposition in the variation-of-constants formula 225
(8.12) R(z, A)¢= 1 00

e-ztT(t)¢dt = 1rr e-ztT(t)¢dt.
On the other hand,
P(z, ¢)
(8.13) R(z, A)¢= det L1(z) ·
A combination of Equations (8.12) and (8.13) implies that E(P(z, ¢)) >
E(det L1(z)). So¢ tf. £. 0
The following two theorems are stated without proof.
Theorem 8.4. The subspace £ defined by (8.11) is a closed subspace of C

that is invariant under the solution operator T(t) defined by (1.2).
Theorem 8.5. Let A be the infinitesimal generator given by Relation (1.2).

If£ is the subspace defined by (8.11), then
M(A) = R(T(E- a))=£.

Furthermore, the space C can be decomposed as
(8.14) C = M(A) rJJN(T(E- a))
7.9 The decomposition in the variation-of-constants

formula
Consider the equation
(9.1) t 2:: a,
and the inhomogeneous system
(9.2) x(t) = Lxt + f(t)

where f E £1oc ([a, oo ), lRn). From our general results on linear systems,
the solution of Equation (9.2) with initial value ¢ at a is
(9.3) Xt = T(t- a)¢+ i t d[K(t, s)]f(s)
where T(t) is the solution semigroup of Equation (9.1),

(9.4) K(t, s)(B) = [s X(t + e- a) da,
and X(·) denotes the fundamental matrix solution to System (9.1). We

now wish to obtain a decomposition of this integral relation according to
the results on the decomposition of the homogeneous equation (9.1).
Suppose that A = {>. 1 , ... , Ap} where each Aj belongs to a(A) and
that C is decomposed by A as C = P EB Q, tJ) is a basis for the general-
ized eigenspace P of A, lf/ is a basis for the generalized eigenspace of the
transposed equation associated with A, (lf/, tP) =I, where ( ·, ·) is given by
Equation (5.9) and the matrix B is defined by the relations
AtP = tPB,
If the decomposition of any element ¢ of C is written as ¢ = ¢P + cpQ,
¢P in P, cpQ in Q, then ¢P = tP(lf/, ¢). Suppose that x is the solution of
Equation (9.2) with initial value ¢ at a, Xt = xf + x~, and let us compute
xf directly from the preceding definition. To do this, we observe that y is
a solution of the transposed equation on ( -oo, oo) and x is a solution of
Equation (9.2) for t :::=: 0, then
(9.5) (y\xt)= [ty(s)f(s)ds+(y 17 ,xl7)
for all t:::: 0. Each row of the matrix e-Btw, W(T) = e-BTlf/(0), 0::::; e::::; r,
is a solution of the transposed equation on ( -oo, oo), and therefore,
and
xf = tP(lf/, Xt)
= [t tPeB(t-s)w(O)f( s) ds + tPeB(t-<7) (lf/, ¢)
= [t T(t- s)tPlf/(O)f(s) + T(t- cr)tP(lf/, ¢)
= [tT(t-s)Xf:f(s)+T(t-a)cpP,
where Xf: = tPlf/(0). Note that, although X 0 tj. C, Xf: = tPlf/(0) can be
considered as the projection of X 0 onto P.
This formula also shows another interesting property: Namely,
(lf/, xt) = [t eB(t-s)w(O)f(s) ds + eB(t-<7)(w, ¢).
Therefore, if we let y(t) = (lf/, Xt) then

y(t) = eB(t-a)y(a') + l t eB(t-s)l[!(O)f(s) ds.
With xf given as earlier, define K(t,s)Q = K(t,s) -.P(w,K(t,s)). From

the variation-of-constan ts formula (9.3), we find
x~ = Xt- xf = T(t- a)¢Q + l t d[K(t, s)Q]f(s).

If A = {.A E a( A) : Re .A > 1' }, then the estimate on the complementary
space Q (see Section 7.6) implies the exponential estimate for T(t- a)¢Q.
To estimate the kernel K(t,s)Q, we need a lemma.
Lemma 9.1. Let tP is a basis for the generalized eigenspace P of A, l[! is

a basis for the generalized eigenspace of the transposed equation associated
with A, (w,.P) =I, and A= {.A E a(A): Re.A > 1'}· If K(t,s) denotes the
kernel in the variation-of-constan ts formula (9.3) and K(t, s)Q = K(t, s)-
tf>(l[!, K (t, s)), then for any {3 > 0, there are constants M 1 , M 2 > 0 such that
IK(t,s)QI:::; M 1 eb+f3lt, t?_s

(9.6)
Var[a,t] K(t, · )Q :::; Mzeb+f3)t.
Proof. Without loss of generality assume that a = 0. First consider the

kernel K(t, s) fort-s?_ r. In this case we can write
K(t, s)(O) = ~~~~s X(r + {3 +B) d{J.

So fort-s?_ r, the columns of X(r+f3+ ·)belong to C and X(r+f3+ ·) =
T({J)Xr. Therefore, we can compute the projection
(9.7) .P(w,K(t,s)) = ~~~~s T({J).P(w,Xr)d{J.

Since Xr -.P(w,Xr) belongs to Q, the exponential estimates in (6.1) imply
On the other hand, if t - s :::; r, we write
(9.9) K(t, s) = t-r T({J)Xr d{J + ~s-t X_a da

Jo -r
and K(t, s) can be estimated by a constant independently oft. The projec-

tion K(t, s)Q has the following representation
(9.10)
ftt~;-s T(/3) ( Xr - <P(Iff, Xr)) d/3 if t- s ~ r;
K(t,s)Q = { J;-rT(f3)(Xr -<P(tP",Xr))d/3
+ f~r X_o: da- <P(tP, f~r X_o: da) if t- s < r.
So the exponential estimates (9.6) now follow immediately from Estimate

(9.8). 0
These results are summarized in
Theorem 9.1. Let <P be a basis for the generalized eigenspace P of A, Iff is
a basis for the generalized eigenspace of the transposed equation associated
with A, (Iff, <P) = I. If C is decomposed by A as P EB Q, then the solution
x(a, ¢) of Equation (9.2) satisfies •
xf = T(t- a)¢P + 1t T(t- s)Xt' f(s) ds

(9.11)
x~ = T(t- a)¢Q + 1t d[K(t, s)Qlf(s), t ~a,
where Xf: = <PtP(O) and K(t, s)Q = K(t, s)-<P(tP, K(t, s)) is given by (9.10).
Also, if xf = <Py(t) then
(9.12) y(t) = By(t) + tP(O)f(t).
Furthermore, if A= {A E a(A) : Re>. > 1'}, then for any {3 > 0, there is
an M > 0 such that
(9.13a) IT(t)¢QI:::; Meh+i3)ti¢QI,
(9.13b) IK(t, s)QI:::; Meh+i3)t
(9.13c) Var[O",tjK(t, ·):::; Meh+i3)t.
As an example, consider
• 7C'
(9.14) x(t) = - 2 x(t- 1) + f(t)
and choose A= {+i~, -i~}. If we let P = P(A), Q = Q(A), ¢P = <P(tP, ¢)
for any¢ E C, where <Pis defined in Equations (7.5), Iff in Equations (7.6),
¢Q = ¢- ¢P, and Xt = xf + x~ = <Py(t) + x~, then
y(t) = By(t) + tP(O)f(t)

(9.15)
x~ = T(t- a)¢Q + 1t d[K(t, s)Q]f(s)
where
lli(O) =col (7rJL, 2JL),
1
(9.16)
If we let the components of y be Yt and y2, then the first of Equations

(9.15) is given explicitly as
If we let
7l' 7l'
Zt = 2Y1 +y2, Yl = JL( 2Z1 + Z2)
(9.17) 7l'
Z2 = Yl- -y2,
2
Y2 = JL(Zt - ~z2)
then
(9.18)
and
(9.19) ..
Zl + (7!')2
2 Z1 = -7!' f ·
It is interesting to see this second-order equation (9.19) for a special type of

forcing term, f = f(x(t), x(t- 1)). From the definition of <P and Relations
(9.17) and (9.18), we have
Xt(O)- x~(O) = <P(O)y(t) = Y2(t) = JL(zt(t) + it(t))

and
2 7!'2
Xt( -1)- x~(-1) = <P(-1)y(t) = -yt(t) = --JL( -z1(t)- it(t)).
7l' 4
Consequently, if we neglect the terms x~ (0) and x~ ( -1), the second-order
equation (9.18) becomes
In the applications to nonlinear oscillations, this equation plays a funda-

mental role.
7.10 Parameter dependence
In this section, we present some simple results on the dependence of simple

eigenvalues on a parameter in the equation.
For any a E IR, suppose L(a) : C-> IRn is a continuous linear operator
that is continuous, together with its first derivative in a. Suppose >. 0 is a
characteristic root of the RFDE[L(O)]. It is important in applications to
know the dependence on a of the characteristic roots of the RFDE[L(a)],
which reduce to >. 0 for a = 0. We do not discuss the general problem, but
prove only the following simple application of the implicit function theorem.
Lemma 10.1. Suppose that the family of bounded linear operators { L( a) :

a E IR} from C to IRn is continuous, together with the first derivative in
a. If Ao is a simple characteristic root of the RFDE[L(O)], then there is
an a 0 > 0 and a simple characteristic root >.(a) of the RFDE[L(a)] that
is continuous, together with its first derivative for fa[ < a 0 , >.(0) = >.0 .
Furthermore, if C is decomposed by >.(a) as C = Pa EB Qa and 1l"a is the
corresponding projection with range Pa, then 1l"a is continuous together with
its first derivative.
(10.1) .::\(>.; a)b = 0,

where L\(z; a) is the characteristic equation L\(z; a)= zi- L(a)ez · corre-
sponding to the RFDE[L(a)].
Since we have assumed >. 0 is a simple eigenvalue of the RFDE[L(O)],
the null space N( L\(>.0 ; 0)) has dimension one and <Cn = N( L\(>. 0 ; 0)) EB
R( L\(>.o; 0)). Let p, I- p be projection operators defined by this decompo-
sition. The existence of a smooth eigenvalue follows easily from the implicit
function theorem. We give another proof, which will produce the eigenvalue
as well. Rewrite Equation (10.1) as
(10.2) L\(>.o; O)b = [L\(>.o; 0) - .::\(>.; a)]b.
Fix a complex eigenvector bo of L\(>.0 ; 0). If b = b0 + d, dE (I- p)<Cn, then

(10.2) is equivalent to
(10.3a) L\(>.o; O)d =(I- p)[L\(>.o; 0)- .::\(>.; a)](bo +d)

(10.3b) p[L\(>.o; 0)- .::\(>.; a)](bo +d)= 0.
Since .1(>.0 ;0) is an isomorphism on (I -p)<Cn, the implicit function theo-

rem implies that there is a 8 > 0 and a unique solution d = d*(a,>.,b 0 )
of Equation (10.3a) for [>.->.of < 8, fa[ < 8, [bo + d*(a, >., bo)[ "/= 0,
and the function d*(a,>.,b 0 ) is continuously differentiable in a,>., b0 and
d*(O, >., b0 ) = 0. Also, the linearity of (10.3a) implies that d*(a, >., bo) =
7.10 Parameter dependence 231
D*(a, >.)bo where D*(a, >.) is ann x n-matrix. Therefore, if there is a so-
lution b of L1(>.; a)b = 0, with b = bo for a = 0, >. = >.o, then b must be
defined by b = [bo + D*(a, >.)bo] for lal < 8, 1>-- >-ol < 8 and satisfy
(10.4) f(a, >.o) ~ p[..d(>.o; 0)- L1(>.; a)][I + D*(a, >.)]bo = 0.

Therefore, the existence of an eigenvalue near >.o for a near zero is equivalent
to the existence of a solution of Equation (10.4). Since an·J~,>.o) satisfies
the equation
..d(O >. )aD*(O,>.o) =-(I_ )aL1(>.0 ;0)

, o a>. P a>. '
it follows that
aj(O, >.) b _ a..d(O, >.) b
a>. o -P a>. o-
By choosing basis vectors so that L1(0, >. 0 ) = diag(O, B), where B is an
(n-1) x (n-1) nonsingular matrix, one can choose b0 as col(1, 0, ... , 0) and
the projection p can be identified with p = (1, 0, ... , 0). Then p..d(>.; O)bo =
a(>.) where >.is the term in the left-hand corner of the matrix L1(>.; 0). An
easy computation now shows that
a..d(>.o; 0) ]b = (d B)_ 1 aL1(>.o; 0)

p[ a>. 0 et a>. .
Thus, aj~/o) -:J 0. Since f(O, >. 0 ) = 0, the implicit function theorem implies
there is a 8 > 0 (which can be taken as the same 8 as before) and a
continuously differentiable function >.(a), >.(0) = >.o, such that f(a, >.(a))=
0 for lal < 8. The corresponding eigenvector is then b(a) = bo +D*(a, >.)bo.
The same remarks apply to the construction of left eigenvectors c( a) of
L1(>.(a); a). The remainder of the proof of the lemma follows directly from
the equivalence in Section 7.4. 0
Later, we need an explicit formula for the derivative of the eigenvalue

>.(a) with respect to a.
Lemma 10.2. Suppose that the conditions of Lemma 10.1 are satisfied and
let <Pa(O) = b(a)exp(>.o(a)O), -r ~ (} ~ 0, 1/Jo(s) = a(a)exp(->.o(a)s),
0 ~ s ~ r, are bases for N( A(a)->.0 (a) ), N( AT(a)->.0 (a) ), respectively,
for the simple eigenvalue >.(a) of the RFDE[L(a)]. If (1/Ja, <Pa) = 1, then
A'(a) = -a(a)L'(a)e>.(a)b(a)
for all a E IR, where prime denotes differentiation with respect to a.
Proof. From the definition of a(a), b(a), we have
a(a)L1(>.(a); a)= 0, L1(>.(a); a)b(a) = 0
(10.5)
a(a)..d(>.(a); a)b(a) = 0
for all o: E IR. Let us first observe that
(10.6)
(see the proof of Lemma 8.1) Differentiating the latter expression in (10.5)
with respect too: and using Equation (10.6), we have
>.'(o:) = -a(o:) ad(~~); o:) b(o:)
for all o: E IR. From the definition of L1(>.(o:); o:), one observes that this
latter expression is the same as the one in the statement of the lemma.
Thus, the lemma is proved. D
Lemma 10.3. Suppose that the conditions of Lemma 10.1 are satisfied and
A(o:) = {A1(o:), ... , Ap(o:)}, >-2j-1(o:) = X2j(o:), j = 1.2, ... , k, Aj(o:) real
j = 2k+ 1, ... ,p, is a set of simple eigenvalues of the RFDE[L(o:)]. Let iP 0 ,_,
1/10 , (iPa, Wa) =I, be real bases for MA(a) 1 MA(a)' respectively, A(o:)iPa =
iPaB(o:). Then
for all o: E IR.

Proof. If A2j_ 1(o:) = J.tj(o:) + iaj(o:), J.tj(o:), aj(o:) real, j = 1, 2, ... , k,
then there is no loss in generality in assuming that B(o:) is in real canon-
ical form. Then there is a further change of basis taking i.Pa into iPaM
where M is a matrix independent of o: such that the corresponding B(o:)
is diagonal. Therefore, we may assume B(o:) is diagonal. In this case,
1/1 = col( 'l/J1a, ... , '1/!pa), iP"' = (¢I a, ... , c/Jpa) satisfy the relations
'1/!ja(s) = aj(o:)e->.i(a)s, 0:::; s:::; r,

c/Jja(O) = bj(O)e>-i(a)O, -r:::; ():::; 0
for j = 1, 2, ... ,p. From Lemma 10.2 one observes that

'( ) Ak(a)b ( )
aj (o: )L o: e k o: =
{ 0,
->.j(o:),
j
j
-1 k,
= k.
Since Wa(O) = col(a1(o:), ... ,ap(o:)), i.Pa(O) (b1(o:), ... ,bp(o:)), one ob-
tains the result stated in the lemma. D
For basic theory on semigroups of transformations, see Hille and Phillips [1],
Yoshida [1], or Pazy [1]. The idea of treating a linear autonomous functional
differential equation as a semigroup of transformations on C originated

with Krasovskii [1] in his studies on stability. Shimanov [2] continued this
investigation and actually used the decomposition in the space C in a very
special case. The general theory was first given by Hale [4] and Shimanov
[3].
An abstract perturbation theory for delay equations is a nontrivial
problem. The reason is the following. If one writes System (1) as an abstract
evolutionary system in the Banach space C
it(t) =Au, u(O) = uo E C,
then A is given by Relation (1.2). Observe that the action of A is inde-

pendent of Equation (1), the equation only enters in the definition of the
domain of A. This means, in particular, that perturbing Equation (1) is the
same as perturbing the domain of an infinitesimal generator. This makes
perturbations hard to handle.
An important idea to solve this problem is the notion of the part of an
operator. If M: v(M) ____, z is an unbounded operator and X c Z, then
M, the part of Min X, is defined to be
v(M) = {¢ E v(M): M¢ EX}, M¢ = M¢.
It is an immediate consequence of the general theory that if M : V ( M) ----> Z

is the generator of a semigroup T(t), then the part of Min X C Z is a
generator of a semigroup T(t) if and only if T(t)X ~X and, in that case,
T(t) = T(t)ix.
Now the first question becomes: Can one find a space Z ::J X and an
infinitesimal generator A such that A, defined by Relation (1.2), is the part
of A in C.
In the text, we have seen that for Z = IRn x C, A(c, ¢) = (L¢, d¢jde),
the generator A is indeed the part of A in C. However, A is not the generator
of a semigroup on Z = IRn x C. In fact, one can prove that A is the generator
of a once integrated semigroup (see Thieme [1]).
Another choice would be Z = IRn x L 00 ([-r,O],IRn), where the gen-
erator A is still the part of A in C and in this case, A is the generator
of a semigroup on Z = IRn X L 00 ([-r,O],IRn). Unfortunately, T(t) is not
a strongly continuous semigroup, but only a weak* continuous semigroup.
This is where duality enters the problem. A general method is the (sun
star) 8*-framework developed by Clement et al. [1].
The idea is the following. Start with a C0 -semigroup T 0 (t) on a Ba-
nach space X and consider the adjoint semigroup T 0(t) on the dual space
X*. In general, the adjoint semigroup is only weak* continuous and not
strongly continuous (see Section 7.1). Therefore, one can restrict to the
largest T0(t)-invariant subspace, denoted by X 0 , of X* on which T 0(t) is
strongly continuous. It is a general result due to Hille and Phillips [1] that
X 0 = V (A*), the norm closure of the domain of A* in X*. From the re-
marks made earlier, the restriction T~(t) of T 0(t) to X 0 is generated by
the part of A* in X0.
Now one can repeat the procedure and find a space X0* and a weak*
continuous semigroup T 0 *(t). Since X0 is a subspace of the dual space of
X, X is embedded in X 0 *. Furthermore, from the construction, it follows
that T 0 *(t)[x = T(t). So the largest invariant subspace x00 on which
T0*(t) is strongly continuous contains X. When X 3:! x00 one calls X
sun-reflexive with respect to T0 (t).
Let X= C and To(t) be the C0 -semigroup corresponding to
u(t) = Aou(t), u(O) = uo EX,

where
d¢
V(Ao) = {¢ E C: ¢(0) = 0}, Ao¢ = d(}'
that is, A 0 is the generator corresponding to x(t) = 0 considered as a delay
equation. It can be shown that C with respect to T0 (t) is sun-reflexive. Sys-
tem (1) on X0* = 1Rn x £ 00 ([-r, OJ, 1Rn) corresponds to the evolutionary
system
u(t) = A~*u + Bu,
where B : C __, X0*, ¢ ~---+ (L¢, 0) and
V(A~*) = {(c, ¢) E Z: ~: E L 00 , ¢(0) = 0}, A~*(c, ¢) = (0, ~:).

On X 0 *, we have a good perturbation theory and the core of the theory is
the following abstract variation-of-constants formula for T(t), the perturbed
semigroup on C,
T(t) = To(t) +lot T~*(t- s)BT(s) ds

where the integral has to be interpreted as a weak* integral. Thus, even
though the map B takes C into the larger space X0*, the convolution
integral is in C and one obtains solutions in C.
So although the initial investment is large, the result yields a powerful
abstract variation-of-constants formula for a larger class of perturbations
than those bounded from C into C. The method not only applies to func-
tional differential equations, but also to different types of problems, for ex-
ample, structured population dynamics. See Diekmann, van Gils, Verduyn
Lunel, and Walther [1].
In this chapter, we have chosen an approach that minimizes the use of
functional analysis. Our variation-of-constants formula, which is an abstract
integral equation in the state space C, allows us to handle perturbations
of System (1). The only disadvantage might be that we do not have a
perturbation theory at the generator level.
The text in Section 7.3 follows Kaashoek and Verduyn Lunel [1]. We
have included the abstract result since it also can be applied to neutral
and periodic functional differential equations as will be illustrated in later
chapters. Theorem 4.2 is called a "folk theorem in functional differential
equations" and was first proved, using a different approach, by Levinger
[1]. See also Kappel and Wimmer [1]
The applications of the decomposition theory of this chapter are nu-
merous. Some applications to perturbed linear systems and behavior near
constant and periodic solutions of nonlinear autonomous equations will be
given in later chapters.
The theory of existence of small solutions has important applications
in the theory of control for linear functional differential equations. See the
papers by Delfour and Manitius [1,2], and Manitius [1].
Early work on the closure of M(A) and, in particular, the convergence
of the infinite series representation of a solution in terms of eigenfunctions
is contained in Bellman and Cooke [1], Pitt [1] and Banks and Manitius [1].
Estimates for ,1-l ( z) using the Newton polygon are given in Bellman
and Cooke [1], Banks and Manitius [1] and Verduyn Lunel [2,3]. The results
presented in Section 7.8 hold for neutral equations as well and can be found
in Verduyn Lunel [4,5,6]. Convergence in norm rather than pointwise re-
quires delicate estimates and Theorem 8.2 is not optimal. Under assumption
(J) the spectral projections PAk¢ converge pointwise to¢ when summation
in the sense of Cesaro is used. See Verduyn Lunel [7] for the details.
8
Periodic systems
The purpose of this chapter is to develop the theory of a linear periodic

RFDE that is analogous to the Floquet theory for ordinary differential
equations. It is shown by example that a complete Floquet theory does not
exist. However, it is possible to define characteristic multipliers and exploit
the compactness of the solution operator to show that a Floquet representa-
tion exists on the generalized eigenspace of a characteristic multiplier. The
decomposition of the space C by a characteristic multiplier is also applied
to the variation-of-constants formula. The case of periodic delay equations
with integer lags is considered in detail.
8.1 General theory
Suppose L : ffi ---+ .C(C, rn.n) satisfies the conditions of Section 6.1 and
suppose there is an w > 0 such that L(t + w) = L(t) for all t. In this
section, we consider the system
(1.1) x(t) = L(t)xt
and the extent to which there is a Floquet theory.
For any 8 E ffi, <P E C, there is a solution x = x(8, ¢) of Equation
(1.1) defined on [8, oo) and Xt(8, ¢) is continuous in t, 8, and ¢. As usual,
let T(t, 8)</J = Xt(8, ¢)for all t;::: 8 and <P E C. The operator T(t, 8) always
satisfies T(t, 8)T(8, r) = T(t, r) for all t;::: 8;::: T and periodicity of Equation
(1.1) implies T(t + w, 8) = T(t, 8)T(8 + w, 8) for all t;::: 8. Let U : C---+ C
be defined by
U 0, there is an integer m > 0 such that mw ;::: r. Therefore;
um = T(mw, 0) is completely continuous. The polynomial spectral theorem
and the theory of completely continuous operators imply the spectrum of
a(U) of U is at most countable, a compact set of the complex plane with
the only possible accumulation point being zero. Also, if 1-L =1- 0 is in a(U),
then 1-L is in the point spectrum Pa(U) of U; that is, there is a . for which J.L = e.X.w is called
a characteristic exponent of Equation (1.1).
Lemma 1.1. IL = e.X.w is a characteristic multiplier of Equation (1.1) if and

only if there is a¢::/=- 0 inC such that T(t+w, 0)¢ = J.LT(t, 0)¢ for all t 2:': 0.
Proof. If J.L E Pa(U), then there is a ¢ ::/=- 0 in C such that U ¢ = J.L¢.
Periodicity of Equation (1.1) implies T(t + w, 0)¢ = J.LT(t, 0)¢ for all t. The
converse is trivial. D
Since um is completely continuous, for any characteristic multiplier 1L

of Equation (1.1), there are two closed subspaces E~-' and Q~-' of C such that
the following properties hold:
(i) E~-' is finite dimensional;
(ii) E~-' EB Q~-' = C;
(iii) UE~-' <;;; E~-', UQ~-' <;;; Q~-';
(iv) a(UIE~-') = {J.L}, a(UIQ~-') = a(U) \ {J.L}.
The dimension of E~-' is called the multiplicity of the multiplier IL·
Let ¢1, ... , c/Jd~" be a basis forE~-', P = {¢1, ... , c/Jd~"}. Since UE~-' <;;; E~-',
there is a dl-' x d~-' matrix M = M~-' such that UP = PM and Property
(iv) implies that the only eigenvalue of M is J.L ::/=- 0. Therefore, there is a
d~-' x d~-' matrix B = B~-' such that B = w- 1 log M. Define the vector P(t)
with elements inC by P(t) = T(t, O)<Pe-Bt. Then, for t 2:': 0,
P(t + w) = T(t + w, O)<Pe-B(t+w) = T(t, O)T(w, 0)<Pe-Bwe-Bt

= T(t,0)U<Pe-Bwe-Bt = T(t,0)<PMe-Bwe-Bt = P(t);
that is, P(t) is periodic of period w. Thus, T(t, O)<P = P(t)e 8 t, t 2:': 0. Extend
the definition of P(t) fort in ( -oo, oo) in the following way. If t < 0, there
is an integer k such that t + kw 2:': 0 and let P( t) = P( t + kw). The function
Xt(O, <P) = T(t, O)<P = P(t)e 8 t is well defined for -oo < t < oo and it is
easily seen that each column of this matrix is a solution of Equation (1.1)
on (-oo,oo). We therefore have
Lemma 1.2. If J.L is a characteristic multiplier of Equation (1.1) and P is a

basis forE~-' of dimension d~-', there is ad~-' x d~-' matrix B, a(e 8 w) = {J.L}
and an nx d~-'-matrixfunction P(t) with each column inC, P(t+w) = P(t),
t E ( -oo, oo) such that if¢ = Pb, then Xt(¢) is defined fort E ( -oo, oo)
and
t E ( -oo, oo).
Therefore, in particular, J.L = e.x.w is a characteristic multiplier of Equation

(1.1) if and only if there is a nonzero solution of Equation (1.1) of the form
238 8. Periodic systems
x(t) = p(t)e)..t
where p(t + w) = p(t).
Since Xt(O, ¢)(B) = x(O, ¢)(t +B) = Xt+O(O, ¢)(0), -r :=:; B :=:; 0, and
¢EEl-', it follows that P(t)(B) = P(t + B)(O)eB£1, -r :=:; B :=:; 0. Therefore, if
we let P(t +B) = P(t + B)(O), then if>( B) = P(B)eBIJ and
x(O, ¢)(t) = P(t)eBtb, t E ( -oo, oo), ¢ = if>b.
Therefore, the solutions of Equation (1.1) with initial value in El-' are of the
Floquet type; namely, if J.L = e)..w, the solutions are of the form e)..t times a
polynomial in t with coefficients periodic in t of period w.
We also need the following remark: If T(t, O)if>b = 0 for any t, then
b = 0. In fact, if there is at such that T(t, O)if>b = 0 and mw ~ t, m ~ 1,
then ·
Since the eigenvalue of Mm is J.Lm # 0, the result follows immediately.

We have defined the characteristic multipliers of Equation (1.1) in
terms of the period map U of Equation (1.1) stftrting with the initial time
0. To justify the terminology, it is necessary to show that the multipliers
do not depend on the starting time. We prove much more than this.
For any s in IR, let U(s) = T(s + w, s). As before, for any J.L =F 0,
J.L E a(U(s)), there exist two closed subspaces El-'(s) and Ql-'(s) of C such
that Properties (i)-(iv) hold with the appropriate change in notation. Let
if>(s) be a basis for El-'(s), U(s)if>(s) = if>(s)M(s), a(M(s)) = {J.L}. As for
the cases= 0, one can define T(t, s)if>(s) for all t E ( -oo, oo). For any real
number r
U(r)T(r, s)if>(s) = T(r + w, r)T(r, s)if>(s)
= T(r + w, s)if>(s)
= T(r, s)T(s + w, s)if>(s)
= T(r, s)if>(s)M(s).
If M = J.Ll - N, then N is nilpotent and
[J.Ll- U(r)]T(r, s)if>(s) = T(r, s)if>(s)N.
Since T(t, s)if>(s)b = 0 implies b = 0, it follows that J.L is in a(U(r)) and the
dimension of El-'(r) is at least as large as the dimension of El-'(s). Since one
can reverse the role of s and r, we obtain the following
Lemma 1.3. The characteristic multipliers of Equation (1.1) are independent

of the starting time and if if>( s) is a basis for El-' (s), then T( t, s )if>( s) is a
basis for El-'(t) for any t in JR.
8.1 General theory 239
Lemma 1.3 shows in particular that the sets E~-'(s) and E~-'(t) are dif-
feomorphic for all s and t. Similarly the sets Q~-'(s) and Q~-'(t) are home-
omorphic. In fact, let 1ft : C ~ Q~-'(t) and I- 7rt : C ~ E~-'(t) be pro-
jections defined by the decomposition C = E~-'(t) EB Q/1-(t). Since Equation
(1.1) is periodic in t, the mapping 7rt is uniformly continuous in t. Thus,
there is a {j > 0 such that for any s E lR, lo:l < o, the linear mapping
1rs+aiQ~"(s) : Q~-'(s) ~ Q~-'(s + o:) is an isomorphism. Therefore, each Q~-'(s)
is homeomorphic to each Q~-' (t).
There is more information in Lemma 1.3. In fact, let tP be a basis for
E~-'(0). Then cf?(B) = P(B)eBO, -r ~ (} ~ 0, where P(B + w) = P(B) and B
is a constant matrix. Lemma 1.3 implies tPt, tPt(B) = P(t + B)eB(t+O), -r ~
(} ~ 0, is a basis for E/1-(t). Therefore, the mapping h(t) : E~-'(0) ~ E/1-(t)
defined by h(t)¢ = tPtb, where ¢ = cf?b is differentiable in t. This implies
that the set UtEIR(t, E~-'(t)) ~ lR x Cis diffeomorphic to lR x E~-'(0) through
the mapping (t,¢) ~ (t,h(t)¢).
It is not known if the remarks in the preceding paragraph hold for
Q~-' (t). Our homeomorphism taking Q~-' (t) into Q~-' (s) was constructed
through the projections 7rt· In general, these mappings seem to be only
continuous in t. This mapping is differentiable if the function L(t)¢ is con-
tinuously differentiable in t. We now prove this fact.
Let L(t) be as in Equation (1.1) and assume that the derivative
8L(t)¢/at is continuous. For any o: E lR, consider the equation
(1.2) x(t) = L(t + o:)xt.
If J.L is a characteristic multiplier of Equation (1.1), then J.L is a characteristic

multiplier of Equation (1.2). Furthermore, if Xt(a, ¢)is a solution of Equa-
tion (1.2) through (a,¢), and t + o: = s, x(a, ¢)(t) = z(s), then z satisfies
Equation (1.1) and Zu+a = ¢. Therefore, if T(t, a, o:), T(a, a, o:) =I, is the
solution operator of Equation (1.2), then T(t, a, o:) = T(t+o:, a+o:, 0). Now
take a= 0, o: = s. Then the decomposition C = E~-'(s) EB Q~-'(s) according
to the multiplier J.L is obtained from the mapping
U(s) ~f T(w, 0, s) = T(w + s, s, 0).
Consequently, if 8L(t, ¢)jot is continuous, then 8U(s)j8s is continuous

from Equation (1.2) and the mapping 7r8 will be differentiable in s.
Let
(1.3) Q/1- = u(t, Q/1-(t)),

tEIR
(1.4) £~-' = U (t, E~-'(t)).
tEIR
These results are summarized in the following lemma.

Lemma 1.4. For System (1.1), the sets E 1_.(t) and Ep,(s) are diffeomorphic
and the sets Q J.t (t) and Q p, ( s) are homeomorphic for all t, s E IR. The
set Ep, in Expression (1.4) is diffeomorphic to IR x Ep,(O) and the set QJ.t
is homeomorphic to IR x Qp,(O). If fJL(t, ¢)/at is also continuous, then
QJ.t is diffeomorphic to IR x Qp,(O). These diffeomorphisms are defined by
(a,¢) ~ (a,g(a)¢), a E IR, ¢ E Ep,(a), (a,'lj;) ~ (a,h(a)'lj;), a E IR,
'ljJ E Q J.t (a), where g and h are continuously differentiable.
The next result relates the Floquet representation in each eigenspace

to the solution Xt( ¢, s) for an arbitrary¢ E C. Suppose O"(U) = {0} U {flm}
where {flm} is either finite or countable and each flm =f. 0. For fln in O"(U),
let Pn(s): C......, EJ.tn(s), I - Pn(s): C......, QJ.tn(s), be projections induced
by EJ.tn(s) and QJ.tn(s) defined earlier and satisfying Properties (i)-(iv).
Theorem 1.1. Suppose ¢ is a given element of C and the notation is as

described earlier. If a is an arbitrary real number, then there are constants
(3 = (3(a) > 0, M = M(a) > 0 such that
lxt(s,¢)- L Xt(s,Pn(s)¢)1 :S: Me(a-,B)(t-s)l¢1, t :2: s.

IJ.tnl~expaw
Proof. For any integer k, one has
C = Ep, 1 (s) EB · · · EB EJ.tk(s) EB Fk(s)

for any set /ll, ... , /lk of nonzero elements of O"(U(s)), where each Ep, 1 (s)
and Fk(s) are invariant under U(s) and
O"(U(s)iFk(s)) = O"(U(s)) \ {fll, ... , Ilk}·

Order the fln so that lfl1l :2: lfl2l :2: · · · > 0 and, for any a, let k = k( a) be
the integer satisfying iflk I :2: exp aw and iflk+ll < exp aw. If
k
Rk(s)¢=¢- LPp,1 (s)¢,
j=l
then Rk(s)¢ E Fk(s), Xs+mw(s, Rk(s)¢) E Fk(s) for all m = 0, 1, ... and
¢ E C. If e1 w = iflk+ 11, then the spectral radius of
U1 ~r U(s)iFk(s)
is e'w. Therefore, limn-->oo 1Ufl 1 /n = e1 w. The proof is completed exactly as
we did for the case when Equation (1.1) was independent oft (see Section
7.6). 0
Corollary 1.1. The solution x = 0 of Equation (1.1) is uniformly asymp-

totically stable if and only if all characteristic multipliers of Equation (1.1)
have moduli less than 1.
8.2 Decomposition 241
Proof. The "if" part follows from Theorem 1.1 with a= 0. The "only if"
part is a consequence of the Floquet representation associated with any
characteristic multiplier. 0
In a similar manner, one obtains
Corollary 1.2. The solution x = 0 of Equation (1.1) is uniformly stable if

and only if all characteristic multipliers of Equation (1.1) have moduli :::::: 1
and further, if p, is a multiplier with lf.J-1 = 1, then all solutions of Equation
(1.1) with initial value in Ell- are bounded.
Corollary 1.3. Suppose that Lm PI-Lm ( s )¢ converges, where the sum is taken
over all projections associated with the nonzero eigenvalues. If R¢ = ¢ -
Lmpi-Lm(s)¢, then, for any real number k,
that is, the solution corresponding to the initial condition R¢ is a small

solution of Equation (1.1).
In Section 8.3 we shall determine the characteristic multipliers and

analyze the existence of small solutions in particular cases. In contrast to
autonomous linear equations, small solutions need not be identically zero
after some finite time as we shall see in Section 8.3.
First we present the decomposition theory.
8.2 Decomposition
In this section, we consider the linear periodic system (1.1) and the problem
of the decomposition of the space C using the generalized eigenspaces of the
characteristic multipliers of System (1.1). The results are easy consequences
of the adjoint theory of Section 6.3.
For a given cr, if U(cr) = T(cr + w, cr), then U*(cr) = T*(cr, cr + w) is the
period map corresponding to the periodic Volterra integral equation
(2.1) y(s)+ i
8
u+w
y(r)ry(r,s-r)dr=gu+r(s),
with g E B 0 . Here U*(cr)g is the forcing function corresponding to the

solution y restricted to ( -oo, cr].
We have already seen in Section 8.1 that, for some integer m, EI-L(cr) =
N( (p,I- U(cr))m ), QI-L(cr) = R.( (p,I- U(cr))m) and
(2.2) C = N( (p,I- U(cr))m) E9 R.( (p,I- U(cr))m ).
Furthermore,
(2.3) Bo = N( (f.lJ- U*(cr))m) EB R( (f.lJ- U*(cr))m)

with d = dimN( (f.lJ- U(cr))m) = dimN( (f.lJ- U*(cr))m) < oo.
Let lJf;(cr) = col(1/Ji, ... ,1/J~) and <P'"'(cr) = (¢i, ... ,¢d,) be bases for
N( (f.lJ- U*(cr))m) and N( (f.lJ- U(cr))m ), respectively. Then
(2.4) R( (f.lJ- U(cr))m) = { ¢ E C: (IJf~(cr), ¢) = 0}
(2.5) R( (f.lJ- U*(cr))m) = { 1/! E Bo: ( 1/J, <P/l>(cr)) = 0}.
Relation (2.4) implies that the d x d-matrix ( l]/; (
cr ), <P'"' (cr)) is nonsingular
and without loss of generality can be chosen as the identity. Therefore,
decompositions (2.2) and (2.3) can be written as
(2.6) C={¢EC:¢=<P'"'(IJf~(cr),¢)}EB{¢EC:(IJf~(cr),¢)=0}
(2.7) Bo = { ¢ E C: 1/J = lJf~( 1/J, <P'"'(cr))} EB {1/J E Bo: ( 1/J, <P'"'(cr)) = 0}.
Relations (2.6) and (2.7) are sufficient for the applications, but some re-
marks are in order to clarify the relationship between this decomposition
and the one given in Section 7.5 for autonomous equations. Define the
transpose of System (1.1) to be
(2.8)
iJ(s) = j_or y(s- e) d[ry(s, e)], s:::; t,
y(t + ~) = 1/J(~), 0 :::; ~ :::; r, 1/! E C'

In the autonomous case, the adjoint and the transposed equation are related
through the mapping pT : C' -+ B 0 (see Section 7.1). If one defines the
time-dependent analog of pT to be
then FT(t): C'-+ B 0, but we no longer have that
Therefore, we cannot introduce the generalization of the bilinear form de-

fined in (5.9) of Chapter 7. But, there is another way to relate the adjoint
equation and the transposed equation when we restrict the forcing func-
tion in Equation (2.1) to be continuous (which is no restriction when we
compute the eigenfunctions of U* (cr)).
Define 81{! to be (81/!)(e) = 1/!(r +e) for -r < e < 0 and define
I- .fl(t) : Bo-+ Bo,
g(s) f--t l(s) -it l(a)R(a, s) da, s:::; t,

where R(t, s) satisfies the formal resolvent equation (3.8) of Chapter 6.

From Chapter 6, Theorem 3.1, it follows that (I- D(t))g = y( · ;t,g) with
y( · ;t,g) the solution of the adjoint equation (2.1). Therefore, I- D(t) is
invertible and it is easy to verify that
(2.9)
for continuous g belonging to B 0 . Relation (2.9) clarifies the connection

between T*(s, t) and TT(s, t), when restricted to continuous functions, and
can be used to find a representation for the eigenfunctions and generalized
eigenfunctions of U* (a) from ur (a). This is sufficient for the application
and there is no need to introduce a special bilinear form.
To this end, we wish to obtain the same type of decomposition as
earlier in the variation-of-constants formula for
(2.10) x(t) = L(t)xt + j(t),

where L(t) is the same function as in Section 8.1 and f is locally integrable
on (-oo,oo). The variation-of-constants formula for Equation (2.10) is
(2.11) Xt = T(t, a)xa +it d[K(t, s)]f(s), t 2': a,
where the kernel K(t, ·) : [a, t] ---+Cis given by
K(t, s)(B) =is X(t + e, o:) do:
and X(t, s) is the fundamental matrix function to the homogeneous equa-

tion.
For any characteristic multiplier f.1. =f. 0 of Equation ( 1.1), let
for any t 2': a. To find the integral equation for the components of Xt, let
(2.12)
The first object K(t, s)E"'(t) has a simple meaning. Recall that each column
of X(t, s) belongs to C fort 2': s +rand
X(t, s) = T(t, s + r)Xs+r( ·, s), t 2': s + r.

Therefore, each column of Xs+r ( · , s) can be decomposed into its compo-
nents according to the decomposition E!l(s+r) EBQ!l(s+r). Since T(s+r, s)
is a homeomorphism on E !l (s + r), this allows us to define in a unique man-
ner ann x n matrix x:"'(t) whose columns are in E!l(t) so that
This justifies
Theorem 2.1. If x is a solution of Equation (2.10) fort ~ u and f-L -/=- 0

is a characteristic multiplier of Equation (1.1) that decomposes C for any
s E (-oo,oo) as C = E.,(s) E9 Q.,(s) with E.,(s) and Q.,(s) as in Section
8.1, then Xt satisfies the integral equations
x;'"(t) = T(t, u)x~"(t) + 1t T(t, s)X~"(t) f(s) ds,

(2.13)
x~"(t) = T(t, u)x~"(t) + 1t ds[K(t, s)Q"(tl]f(s), t ~ u,
where K(t, s)Q"(t) = (s)( tJr*(s), K(t, s) ).
Proof. Suppose (O) = .,(O) and tJr;(o) are bases chosen as stated be-
fore Equations (2.6) and (2.7) with ( tJr;(o), (O)) = I, the identity. Let
(t) = T(t, O)(O) and tJr*(t) = T*(t, O)tJr*(O) with tJr*(t) the matrix solu-
tion of the adjoint equation on ( -oo, t]. Lemma 1.3 and the corresponding
generalization for the formal adjoint equation imply that ( t) is a basis for
the solutions of Equation (1.1) on [t- r, t] corresponding to the multiplier
f-L and tJr*(t) is a basis for forcing functions of Equation (2.1) on (-oo,t]
corresponding to the multiplier f-L· From Chapter 6, Relation (3.2),
( T*(t, O)tJr*(O), T(t, O)(O)) =constant.
Thus (tJr*(t),(t)) =I for all t E (-oo,oo). Furthermore,
xf"(t) = (t)(tJr*(t),xt)
= (t)( tJr*(u), Xa) + (t) 1t tJr(s)*(O)f(s) ds
= T(t, u)(u)( tJr*(u), Xa) + 1t T(t, s)(s)tJr(s)*(O)f(s) ds
= T(t, u)x~"(a) + 1t T(t, s)X~"(t) f(s) ds
since X~"(s) = (s)( tJr*(s), X 0 ) = (s)tJr*(s)(O). Using the fact that

x~"(t) = Xt- xf"(t), one completes the proof of the theorem. D
The first of equations (2.13) is equivalent to an ordinary differential

equation. In fact, if(t) = T(t)(O) = P(t)eBt, U = <J>eBw, and
(2.14)
then Equations (2.13) are equivalent to the system

y(t) =By+ tP*(t)(O)f(t)

(2.15)
x~~'(t) = T(t, u)x~~>(t) + lt ds[K(t, s)Q~>(t)Jf(s), t :2: u.
System (2.15) is now in a form to permit the discussion of problems con-

cerning the perturbation of Equation (1.1) in a manner very similar to that
when Equation (1.1) was autonomous. We do not devote any time to a
detailed discussion of these questions since they proceed in a manner that
is very analogous to ordinary differential equations.
Theorem 2.1 can obviously be employed to make further decomposi-
tions of the space C. In fact, for any 8 > 0, let A= A(8) be the characteristic
multipliers of Equation (1.1) satisfying IJLI > 8. As before, one can obtain
bases tPA_ (u) and <P A(u) for the formal adjoint equation (2.1) and Equation
( 1.1) corresponding to all multipliers in A such that ( tPA_ (u), <P A(u) ) = I
and
C = EA(u) EB QA(u)
(2.16) EA(u) = {cP E C: cP = <PJL( tP~(u), cP)}
QA(U) = {cP E C : ( W~ (0"), cP ) = 0} .
The corresponding Formulas (2.13) become
x~A(t) = T(t, u)x:A(t) + lt T(t, s)X~A(t) f(s) ds,

(2.17)
x~A(t) = T(t, u)x~A(t) + lt d [K(t, s)QA(t)]f(s), t :2: u.
8
It is clear that the spectral radius of T(t, u)IQA(u) is less than 8 = exp"(W
and the spectrum ofT(t,u)IEA(u) is equal to A. Therefore, for any {3 > 0,
there is an M > 0 such that
(2.18)
IT(t,u)¢1:::; Meh+.B)(t-a)l¢1, t :2: u, u E IR, ¢> E QA(u),
IK(t,s)QA(s)l:::; Meh+.B)(t-s), t :2: s, s E IR,
t :2: s, s E IR.
The results in this and the previous section have natural generalizations
to NFDE. Consider the linear periodic NFDE
(2.19)
where D(t)¢, L(t)¢ are continuous in t, ¢>,linear in¢, D(t + w)¢ = D(t)¢,
L(t + w)¢ = L(t)¢ for all t, ¢>and some fixed w > 0, and D(t)¢ is atomic
at zero. Let TD,L(t, u), t :2: u, be the solution operator of Equation (2.19).
A complex number p is called a characteristic multiplier of the NFDE
(2.19) if pis an eigenvalue of finite type ofTD,L(u+w, u), that is, an isolated
point of the spectrum of TD,L (a+ w, a) with finite-dimensional generalized

eigenspace. As in Section 1, one shows that the characteristic multipliers
are independent of a. By definition, for each characteristic multiplier p
and each a E JR., there is a decomposition of C as C = Pa EB Q8 , where
Pa = N(TD,L(a + w,a)- pi)k, Qa = R(TD,L(a + w,a)- pi)k, and Pa
has finite dimension, say d. If tPa is a basis for Pa, then there is a d x d
constant matrix Ba with the spectrum of eB"w equal to p and an n x d
matrix Ca(t), Ca(t + w) = Ca(t), such that if¢ EPa, ¢ = tPab, then
(2.20) TD,L(t +a, a)= Ca(t)eBut)b.
In this sense, there is a Floquet representation on the generalized eigenspace
Pa of the characteristic multiplier p.
The representation in Theorem 7.3 of Chapter 3 is valid for Equation
(2.19). In fact, one can determine a constant k such that for any a E JR.,
there is a tl>a = (¢r, ... ,¢~) such that D(a)tl>a =I, and lc/Jjl::; k, j =
1, 2, ... 'n. If wa =I- tJ>U D(a), then
(2.21) TD,L(t, a)= TD(t, a)l/Fa + U(t, a),

where U(t, a) is completely continuous fort:::: a and TD(t, a) is the solution
operator generated by the equation D(t)yt = 0.
Using the same type of reasoning as for the autonomous case, if eavw is
the spectral radius ofTD(a+w, a), one proves that any pin the spectrum of
TD(a+w, a) with IPI > eadw must be a normal eigenvalue and, thus, a char-
acteristic multiplier of Equation (2.19). Furthermore, there are only a finite
number of characteristic multipliers p satisfying IPI :::: eaw for any constant
a > aD. The space C therefore can be decomposed as C = Pa,a EB Qa,a,
where Pa,a and Qa,a are invariant under TD(a + w, a) and the spectrum of
TD(a+w,a)IPa,a consistsonlyofthemultipliers IPI with IPI:::: eaw, a> aD,
and the spectrum of TD (a + w, a) IQa,a lies inside the disk with center zero
and radius < e<a-<)w for some E > 0. Therefore, there is a constant K such
that
(2.22) ITD,L(t, a)IQa,al ::; Ke<a-<)(t-a)' t:::: a.

Also, there is a constant matrix Ba,a and an w-periodic matrix Ca,a(t) such
that
(2.23)
where tPa,a is a basis for Pa,a and the only eigenvalues of eBu,aw are those
characteristic multipliers p of Equation (2.19) with IPI:::: eaw.
As for RFDE, similar decompositions hold in the variation-of-constants
formula for the nonhomogeneous linear equation. The reader may supply
the details.
8.3 An example: Integer delays 247
8.3 An example: Integer delays
In this section, we analyze the following system of linear periodic delay

equations
m
(3.1) x(t) = LBj(t)x(t- jw), X8 = cp, t ~ s,
j=O
where Bj are continuous matrix-valued functions such that Bj(t + w) =

Bj(t).
Define D!(z) to be the fundamental matrix solution of the periodic
ordinary differential equation
m
(3.2) y(t) = L zj B3(t)y(t)
j=O
with D!(z) =I. Let U: C--+ C be defined by Ucp = T(w,O)cp. Then U is
t, /_~ n~(O)Bj(r)cp(r-
given by
(3.3) (Ucp) (0) = n~w(O)cp(O) + (j- 1)w) dr
for -w::; (}::; 0 and (Ucp)(O) = cp(O+w) for -mw::; (}::; -w.
Theorem 3.1. If det Bm has only isolated zeros, then the spectrum of the op-
erator U : C[-mw, OJ --+ C[-mw, OJ defined by (3.3) consists of eigenvalues
of finite type only,
(3.4) a(U) = {J.t : det Ll(J.t) = 0},

where Ll(z) = z- DQ'(z- 1 ). For J.t E a(U), the algebraic multiplicity of the
eigenvalue J.t equals the order of J.t as a zero of det .:1.
For a proof of the theorem, we are going to apply the abstract results
about characteristic matrices from Section 7.3. Before we can do so, we
have to make some preparations. Since um is compact and one-to-one, the
inverse of U is a well-defined unbounded closed operator on C. To avoid
technical complications we shall only prove the theorem form= 1.
For m = 1, the space C equals C([-w, OJ, lRn) and the mapping U :
C--+ Cis given by
(3.5) (Uc/J)(O) = n~w(O)cp(O) + /_~ n~B1(r)cp(r) dr.

We shall show that A = u- 1 is the first operator associated with the
operators D, L, and M, which we now define. Let M: V(M) --+ C be the
operator
with maximal domain V(M) = {'lP E C: M'lj; E C}. The kernel of M is

given by
N(M) = {'1/J E C: '1/;(8) = il~w(O)'I/;(-w)}.
So (Hl) of Section 7.3 is satisfied.
Next we define the restriction Mo : V(Mo) --> C
(3.7) V(Mo) = {'1/J E C: '1/J E V(M), '1/;(-w) = 0}, Mo'I/J=M'I/J.

To find the resolvent of M0 , put (z- M 0 )- 1 ¢ = '1/J. Then '1/J satisfies the
equation
,j; = (Bo + zB1)'1/J- B1¢.
The initial condition is '1/;( -w) = 0 and we can solve for '1/J. Consequently,
(3.8) (z- Mo)- 1 ¢(8) =- j_ow il~(z)B1(r)¢(r) dr.

So Q= p(M0 ) = e and (H2) of Section 7.3 is satisfied. Finally, set C=
en XC and
L :C --> en, ¢ ~ ¢( -w),
Then the operator A = u- 1 can be represented by

(3.9) V(A) = {'1/J E C: '1/J E V(M), DM'Ij; = L'I/J}, A'lj;=M'I/J.
So the operator A defined by (3.9) is the first operator associated with D, L,
and M. The second operator A: v(A) --> C associated with D, L, and M
is given by
V (A) = {(c, '1/J) E C : '1/J E V (M), c = D'lj;},

(3.10)
A(c, '1/;) = (L'Ij;, M'lj;).
Theorem 3.2. The matrix function Ll : e --> .C( en) defined by

(3.11) Ll(z) = zil0(z)- I
is a characteristic matrix for A. The equivalence relation is given by
(3.12) ( Ll~z) ~) = F(z)(z- A)E(z),
where E: e--> .C(C, C(A)) is given by
E(z)(c,'lj;) = (il~w(z)c- j_: il~(z)B1(r)¢(r)dr,x),

X( B)= fl~w(z)c -l fl~(z)B1(T)</J(T)

0
w dT,
and F : <C ---+ £(C) is given by
Proof. From Theorem 3.1 of Chapter 7, it follows that the characteristic

matrix for A is given by
Ll(z) = (zM- L)(j- z(z- M0 )- 1 j)
where j : <Cn---+ N( D), c ~--+ fl~w(O)c. Using Representation (3.8) for the
resolvent of M 0 , we conclude
So j - z(z- Mo)- 1 j = fl~w(z) and

Ll(z) = z[l~w(z)- I.
The concrete representations for E and F are verified in a similar way. 0
Proof of Theorem 3.1 Since u- 1 =A, an eigenvalue f.L of U corresponds to

an eigenvalue f.L- 1 of A and the spectral data of U and A are the same. This
implies that the spectral data of U and A are the same. Hence the theorem
follows from Corollary 3.1 of Section 7.3. 0
In the special case that the matrices Bj, j = 0, ... , m, are diagonal,
we find a characteristic equation for the exponents:
where
j=O, ... ,m
denotes the average of Bj. So the characteristic matrix for the exponents
(since they are determined only up to multiples of 21rijw) maybe taken to
be
m
A= LBje-iwA,
j=O
which is just the characteristic matrix for the autonomous equation

m
y(t) = L Bjx(t- jw).

j=O
Theorem 3.2 leads to a situation where we can apply the results from
Section 7.8. To simplify the arguments, we restrict our attention to the
scalar case.
Consider a scalar linear periodic delay equation
m
(3.13) x(t) = L. bj(t)x(t- jw), t:::: S, X8 = </>,
j=O
where b1 , j = 0, ... , m, are continuous periodic functions with period w. As

before, we will assume that the solution map T(t, s) : C ____,. C for (3.13) is
one-to-one, i.e., the zeros of bm are isolated. We have the following results.
Theorem 3.3. Suppose that the zems of bm are isolated. Then the system of
Floquet solutions is complete if and only if bm has no sign change.
As in the autonomous case there exists a relation between the com-

pleteness of the system of Floquet solutions and the existence of small
solutions.
Theorem 3.4. Suppose that the zeros of bm are isolated. Then (3.18) has
small solutions if and only if bm has a sign change.
Define the following subspaces of C:
S= {1> E C: T(t,s)</> is a small solution of (3.18)}

M = EB Mp,(U).
p,EPa(U)
Theorem 3.5. The system of Floquet solutions is complete if and only if

there are no small solutions. Furthermore,
C = M EBS.
Example 3.1. The equation

1
x(t) = (2 + sin(27rt))x(t- 1)
has small solutions, although there are infinitely many independent Floquet
solutions
with .A- - ~e->-
2 .
In the remaining part of this section we shall sketch the proof of these
results for m = 1.
The proof consists of several steps.
(i) Recall that in the scalar case
Furthermore
(3.14)
From (3.13), we deduce that the resolvent of A has the following represen-
tation
R(z,A)¢= P(z,¢)
det Ll(z)
where
P(z,¢)(e) = st~w(z)[c+z [~ st~(z)h(T)cjJ(T)dT]

(3.15) (}
- det Ll(z) lw st~(z)b 1 (T)cp(T) dT.

(ii) Similar arguments as used in Section 8 of Chapter 7 show that
there exists a sequence of simple closed smooth curves rN such that
There is a complex function aN : [0, 2n] --+ <C that is differentiable
on [0, 2n] such that rN = {z = aN(t), t E [0, 2n]}, for t E [0, 2n]
iaN(t)i--+ oo as N--+ oo and ia~(t)l ~ iaN(t)i;
the contour rN encloses AN but no other eigenvalues of a(A);
there exists an E > 0 such that dist (FN, a(A)) > E for all N = 0, 1, ....
(iii) Define
(3.16) E= { ¢ E C: E(P(z, ¢)) ~ lwb1l}.
Then for every¢ E En D(A) there exists a positive constant M such that
(3.17) IR(z, A)¢1 ~ M for z ErN, N = 0, 1, ....
The same argument as used in the proof of Theorem 8.2 of Chapter 7 yields
that for¢ E En D(A 3 )
(3.18)
where AN denotes the set of eigenvalues enclosed by the contour rN.

(iv) Next we show M = E. Theorem 2.3 of Verduyn Lunel [4] yields
that Eisa closed subset of C. Since for¢ EM, the function z ~ R(z, A)¢
is rational, we conclude that M C E. On the other hand, from (3.18) we
conclude that
Thus
(v) Define
s= n N( p)..) = {¢: z f--t R(z,A)¢ is entire}.

>-.Ea(A)
Suppose¢ EM n Sand¢# 0. Using the invariance of M and S
Therefore, it follows from (3.18) that
since P11 and R(>., A) commute. However ¢ E S and hence P11 ¢ = 0 for
every >. E a(A). This proves M n S = {0}. To prove the density of the
direct sum, we use duality. From the Neumann series for the resolvent
S= n
>-.Ea(A)
N(P>-.)·
So the density of M tt! Sis equivalent toM* n S* = {0}, where M* is the

linear subspace generated by R( P; ), ).. E a(A*), and
S* = n
>-.Ea(A*)
N(P;).
Since the characteristic matrix for A* is Ll(z) and the hypotheses are in-
variant under duality, a similar argument yields M* nS* = {0}. This shows
C =M tt!S.
(vi) It suffices to compute the exponential type of z r--t P(z, ¢).Since
z.fl~w(z) lOw .fl~(z)bl (T)cP(T) dT- z.fl~w(z) l~ fl~(z)b1 (T)cP( T) dT

= z.fl~w(z)[l~ .fl~(z)b1 (T)c/J( T) dT- zfl2(z) l~ .fl~(z)b1 (T)cP( T) dT]
= z.fl~w(z) foO .fl~(z)bl(T)cP(T)dT,
we find
P(z, ¢) = n~w(z)c + zil~w(z) io il~(z)b1(r)¢(r) dr

(3.19)
+ /_~ n~(z)b1(r)¢(r) dr.
First suppose that b1 does not change sign, say b1 (s) 2: 0. The entire func-
tion
is of order 1, has maximal type in the left half plane, and is bounded in the
right half plane. Using the Paley-Wiener theorem we find
E(J:w ef: bo(s)ds+z J: bl(s)dsb1(r)¢(r) dr) = -~a;'~olo b1(s)ds

:::; wb1.
Furthermore,
E(zil~w(z) io il~(z)b1(r)¢(r)dr) =/_ow b1(s)ds+ 6~~0 1° b1(s)ds

=wb1.
Similarly, for b1 ( s) :::; 0, the functions have exponential growth in the right
half plane and are bounded in the left half plane, and the same exponential
estimates (with b1 replaced by -b1 ) hold. In both cases, it follows that
M = C. On the other hand, if b1 does change sign, the maximum of the
function
on [0,0] is larger than wb1. Since z t-t zil~w(z)f~ il~(z)b 1 (r)¢(r)dr can
not be canceled by any other term in (3.19) the type of z t-t P(z, ¢)(0) is
larger than wb 1 • Hence C =1- M. Since the eigenfunctions and generalized
eigenfunctions of A correspond to the Floquet solutions of (3.1), we obtain
Theorem 3.3.
(vii) To prove the remaining results, it suffices to prove that the sub-
space
S = { ¢ : z t-t R(z, A)¢ is entire }
corresponds to the space of initial conditions S that yield small solutions
of (3.1). It is clear from the exponential estimates in Section 1 that x(t) =
T(t, s)¢ is a small solution of (3.1) if and only if
for all f.L E a(U)
or, equivalently, z t-t z(I- zU)- 1 ¢ is an entire function. But U = A- 1 and

hence z t-t -zA(z- A)- 1 must be entire. Hence x(t) = T(t, s)¢ is a small
solution of (3.1) if and only if
Pvp = o for all .A E <1(A).
This completes the proof of Theorems 3.4 and 3.5.
The paper of Stokes [2] contained the first general discussion of the Floquet
theory for periodic functional differential equations. Theorem 1.1 is due to
Stokes [2]. Shimanov [4] was the first to state the decomposition theorem
for periodic systems for the special case when the function rJ(t, ·) has no
singular part. The extension to the general case was given by Henry [3].
The case of periodic differential difference equations with integer lags
has received considerable attention, with the main effort devoted to the
expansion of solutions in terms of a series of Floquet solutions. The presen-
tation in Section 8.3 follows Verduyn Lunel [5] and extends results obtained
by Hahn [1], Zverkin [3,5,6], and Lillo [3,4,5]. Recently general scalar differ-
ential delay equations with one delay were studied by Huang and Mallet-
Paret [1,2] under small divisor conditions.
9
Equations of neutral type
In this chapter we discuss a particular class of neutral equations for which

a qualitative theory is available and for which one can reproduce a theory
similar to the one for RFDE.
9.1 General linear systems
For (a,¢) E lR x C, consider the linear system

d
(1.1)
dt [D(t)xt) = L(t)xt + h(t), t 2: a,
x.,. = ¢
where h E .Cioc and for all t E lR, D(t) : C -t lRn and L(t) : C -t lRn are
= I: =I:
given by
(1.2) D(t)¢ ¢(0)- d[JL(t, 0)]¢(0), L(t)¢ d[1J(t, 0)]¢(0).
The following assumptions on the kernels JL and 1J will be maintained

throughout this chapter.
The kernel17: lR x lR-----" lRnxn, (t, 0) f-+ 17(t, 0), is measurable in (t, 0),
normalized so that
1](t, 0) =0 for 0 2: 0, 17(t, 0) = 17(t, -r) for 0:::; -r,

1J( t, 0) is continuous from the left in 0 on ( -r, 0) and has bounded variation
in 0 on [-r, OJ for each t. Further, there is an m E .C~oc ( ( -oo, oo), lR) such
that
Var[-r,O] 17(t, ·):::; m(t).
The kernel JL : lR x lR -----" lRnxn, (t, 0) f-+ JL(t, 0), is measurable in (t, 0),
normalized so that
JL(t, 0) = 0 for 0 2: 0, JL(t, 0) = 17(t, -r) for 0:::; -r,

256 9. Equations of neutral type
p,(t, B) is continuous from the left in Bon ( -r, 0) and has bounded variation
in B on [-r, 0] uniformly in t, and such that t ~ D(t)¢ is continuous for
each ¢. Further, p, is uniformly nonatomic at zero, i.e., for every E > 0,
there exists a 8 > 0 such that
(1.3) Var[-6,0] p,(t, ·) < E, for all t E JR.
Following the proofs of the corresponding results in Sections 6.1 and 1. 7

one shows that there exists a unique global solution.
Theorem 1.1. Suppose the conditions on 7J and p, are satisfied. For any given
a E lR, ¢ E C([-r,O],lRn), and hE .Cioc([a,oo),lRn), there exists a unique
function x( ·;a,¢) defined and continuous on [a-r, oo) that satisfies System
(1.1) on [a, oo).
To derive a representation for the solution, it turns out to be useful to

integrate System (1.1). If we split off the part that explicitly depends on
the initial data, we find
x(t) = i~t d[p,(t, B)]x(t +B)+ i t i~.,. d[ry(r, B)]x(r +B) dr

(1.4) + f(t)
=it d[p,(t, s- t) + 1t ry(r, s- r) dr]x(s) + f(t),
where we used the Fubini theorem to reverse the order of integration, and
f(t) = D(a)¢ + i:-t d[p,(t, B)]¢(B + t- a)

(1.5)
+it i:-r d[ry(r, B]¢(B + T - a) dr +it h(s) ds.
Equation (1.4) is a Stieltjes-Volterra equation
(1.6) x(t)- i t d[k(t, s)]x(s) = f(t)
with
(1. 7) k(t, s) = p,(t, s- t) + 1t ry(r, s- r) dr.

A function k: lR x lR+ ~ lRnxn, (t,s) ~ k(t,s), is called a Stieltjes-
Volterra kernel of type B 00 , if k is a bounded function that is measurable in
t for each fixed s, vanishes for s > t, and is of bounded variation, continuous
from the right ins on (0, t) for each fixed t. In addition, the total variation
of k in its second argument is bounded uniformly in t.
9.1 General linear systems 257
From the assumptions on ry and J-L, it is clear that k defined by (1.7) is

a Stieltjes-Volterra kernel of type B 00 •
Every function ( of bounded variation on [0, T], normalized so that
((0) = 0 and (is left continuous on (0, T) represents a Borel measure on IR
with no mass outside [0, T]. This measure will be denoted by d(. For any
d(-measurable function f: [0, T] --+ IR
(1.8) lot d[((s)]f(t- s)

denotes the convolution off with respect to the measured(. Iff is contin-
uous, then (1.8) is just a Riemann-Stieltjes integral. More generally, (1.8)
is defined for any f E Lp[O, T] and is called a Lebesgue-Stieltjes integral.
If k is a kernel of type B 00 , then
g(t) = (Cf)(t) =lot d[k(t,s)]f(s),

maps the functions of bounded variation into itself and
Vaqo,T] g:::; llkllr Var[o,r] J,

where
llkllr = sup Var[o,T] k(t, · ).
tE[O,T]
A kernel p is called a Stieltjes-Volterra resolvent of type B 00 corre-

sponding to a function k if
p(t, s) = -k(t, s) + lt d[p(t, a)]k(a, s)
= -k(t, s) + lt d[k(t, a)]p(a, s), t ?_ s.
A simple contraction argument shows that if k is a kernel of type B 00 with

llkllr < 1, then k has a Stieltjes-Volterra resolvent of type B 00 , and iff
is a function of bounded variation on [0, T], Equation (1.6) has a unique
solution of bounded variation given by the variation-of-constants formula
(1.9) x(t) = f(t)- 1t d[p(t, a)]!( a).
Here the integrals are understood as Lebesgue-Stieltjes integrals. To apply

this result to Equation (1.4), we use a scaling argument similar to the one
used in Section 6.1. If we define
p(t,s) = p(t,s)e-'Y(t-s), k(t,s) = k(t,s)e-'Y(t-s),
then p( t, s) satisfies the equation

p(t, s) = -k(t, s) + 1t d[p(t, a)]k(a, s)
and it suffices to prove that we can choose "' > 0 such that
llkllr = sup Var[o,T] k(t, ·) < 1.

[O,T]
From Definition (1. 7) for k and the assumptions on 'fJ and J.L, we find
llkllr :::; sup Var_r::;s-t<c5 J.L(t, s- t)e-"'(t-s)

tE[O,T]
+ sup Var_o::;s-t::;o J.L(t, s- t)e-"'(t-s)
tE[O,T]
+ M1(t- s)e-"'(t-s)
:::; Moe-"~
6
+ M1
18 + M2€,
where f. can be made arbitrarily small according to the uniformly nonatomic

condition on J.L. So if we choose f.> 0 such that M 2 E < 1/3 and"' so large
that
M
Moe-"~ 6 + "'; < 2/3,
then we have llkllr < 1. This proves the following theorem:
Theorem 1.2. Suppose that the assumptions on 'TJ, J.L, and h are satisfied. If
x = x(a, ¢,h) is the solution of System (1.1) on [s, oo) such that Xu = ¢,
then fort 2: a,
(1.10) x(t) = X(t, a)f(a) +it X(t, a) df(a),
where
1:
f(t) = D(a)¢ + j_or d[J.L(t, 0 +a- t)]¢(0)
(1.11)
+it de['TJ(T, 0 +a- T)]¢(0) dT +it h(s) ds
and X(t, s) =I +p(t, s), t 2: s, is the matrix solution to the integral equation
(1.12) X(t, s) =I+ 1t d[X(t, a)]J.L(a, s-a) -1t X(t, a)ry(a, s-a) da.
Proof. Since Representation (1.9) is well defined for continuous f and Sys-
tem (1.1) has a unique solution, it follows that any solution of System (1.1)
can be represented by (1.9). Therefore, we have
x(t) = f(t) - i t d[p(t, a)]f(a)
= f(t) + p(t, CJ)f(CJ) +it p(t, a) df(a)
= X(t, CJ)f(CJ) +it X(t, a) df(a),
where we define X (t, s) = I + p( t, s), t ;:::: s. To complete the proof of the

theorem, it remains to show that X(t, s) satisfies Equation (1.12). From the
resolvent equation, it follows that
I+ p(t, s) =I+ i t d[k(t, a)](I + p(a, s)),
integration by parts, using Definition (1.7) fork, proves the theorem. D
Corollary 1.1. Suppose that the assumptions on TJ, f-1, and h are satisfied. If
x(CJ,c/>,h) is the solution of System (1.1) on [s,oo) such that Xa =¢,then
there are positive constants C and '/ such that
(1.13)
Proof. From the assumptions on TJ and f-1, it follows that we can choose '/
such that
k(t,s) = k(t,s)e-"!(t-s)
satisfies llkll supt::O:O Var[o,oo)k(t, ·) < 1. Therefore k has a Stieltjes-

Volterra resolvent of type B 00 on [0, oo ). This proves that
(1.14) IX(t, s)i =II+ p(t, s)i :::; Ce'Y(t-s)
Estimate (1.13) now follows from Representation (1.10). D
As we have seen in the retarded case, the matrix solution X(t, s),
t ;:::: s, has a natural interpretation for System (1.1). In fact, it is easy to
see that X(t, s), t;:::: s, is the solution to System (1.1) corresponding to the
discontinuous initial data X(e, s) = X 0 (0), -r:::; e:::; 0, where
X (O) = {I, fore= 0,

0 0, for -r :::; e < 0.
Note that the fundamental solution X(t,s) is of bounded variation on

compact intervals, but not necessarily absolutely continuous for t > s, in
contrast to the retarded case, where the fundamental solution X(t, s) 1s
absolutely continuous for t > s.
From the existence and uniqueness for solutions of Equation (1.1) and
the continuity assumptions, it follows that translation along the solution
defines a forward evolutionary system on C:
(1.15) T(t, a)¢= Xt(.; a,¢), t ?_a.
From the variation-of-constants formula, given by Equation (1.10), we find

that the solution of System (1.1) is given by
(1.16) Xt(8; a,¢, h)= T(t, a)¢(8) + 1

u
t+O
X(t + 8, a)h(a) da
where it is understood that the integral is considered as a family of Eu-
clidean space integrals parameterized by 8, -r :=:; 8 :=:; 0.
From the results of Section 6.2, it follows that we can write Equation
(1.16) as an abstract integral equation in the Banach space C
(1.17) Xt(-; a,¢, h) = T(t- a)¢+ 1t d[K(t, a)]h(a),
where
(1.18) K(t, s)(8) = 1s X(t + 8, a) da

and X(t, s) is the solution to Equation (1.12).
One can discuss the formal adjoint in the same manner as in Section
6.3. Following similar arguments, one can show that the adjoint evolution-
ary system T*(s, t) can be expressed in terms of a backward evolutionary
system on forcing functions for the Stieltjes-Volterra equation
(1.19) y(s) + ~t y(T) d[k(T, s)] = gt(s), s :=:; t,
where gt(s) = g(s- t) belongs to B 0 and k is given by (1.7).

From the abstract theory, it follows that Equation (1.19) has a unique
solution of bounded variation on compact intervals, given by the variation-
of-constants formula
(1.20)
y(s; t, g) = l(s) + lt d[p(t, a)Jl(a)
= -~t d[g(a- t)]X(a, s), s :=:; t.
Translation along a solution of Equation (1.19) induces a two-parameter

family of bounded operators V(s, t), s :=:; t on B 0 . In fact, the solution y
restricted to ( -oo, a] satisfies a Volterra integral equation given by
s :=:; a,
where g belongs to B 0 and is given by

for 0 = 0;
g(O) = { ~o-(0) + J~ Yo-( r)dk(r +a, 0- r), for -r ~ 0 < 0;
for 0 ~ -r.
g( -r),
Define V(a, t)g = g, so that V(a, t) maps the forcing function for the solu-
tion on ( -oo, t] onto the forcing function for the solution on ( -oo, a]. From
the uniqueness property for the Volterra integral equation, it is easy to see
that V(s, t), s ~ t, defines a backward evolutionary system on Bo.
Next we derive a representation for V(s, t), s ~ t, from its definition.
The integral equation yields
g(a+O-t)=y(a+0)+1t y(r)dk(r,a+O-r)
o-+9
(1.21)
=y00 (0)+ it-o- Yo-(r)dk(r+a,O-r).
Therefore
g(O) = g(O- (t- a)) -fot-o- Yo-(r)dk(r +a, 0- r)
and, using Representation (1.20), V(a, t) : B 0 -t B 0 is given by
V(a, t)g(O) = g(O- (t- a))+ i t y(r)dk(r, 0- r) dr

(1.22) = g(O- (t- a))
+it l t d[g( -t + a)]X(a, r)dk(r, 0- r).
By definition V(a, t)g(O) = 0 and, using Equation (1.21),
V(a, t)g(O-) = y(a).
Similar to the retarded case, one can prove the following result.
Theorem 1.3. Let T(t, s), t 2: s be the evolutionary system associated with
System (1.1) on C. IJV(s, t), s ~ t denotes the backward evolutionary sys-
tem for the adjoint equation defined by (1.21), then V(s, t), s ~ t is the
adjoint system ofT(t,s), t 2: s, that is,
T(t, s)* = V(s, t).
To end this section, we illustrate how the theory can be applied to dis-
cuss two-point boundary-value problems for the nonhomogeneous equation
(1.1). The theory parallels the theory for RFDE given in Section 6.4.
Suppose V is a Banach space, a < T are given real numbers, M, N :

C--+ V are linear operators with domain dense in C, and 'Y E Vis fixed.
The problem is to find a solution x of
d
(1.23) dt [D(t)xt] = L(t)xt + h(t)
subject to the boundary condition

(1.24)
Let V* be the dual space of V and M* and N* the adjoint operators of M
and N respectively.
Theorem 1.4. In order that Equations (1.23) and (1.24) have a solution, it
is necessary that
(1.25) ir y(a)h(a) da = -( 8, "( )v
for all 8 E V* and solutions y of the system of adjoint equations
y(s) + 1 7
y(a)d[k(a,s)] = N*8, S ::; T,
+ 1a
(1.26) {
y(s) y(a)d[k(a, s)] = -M*8, s ::; a,
where k is given by (1.7). IfR( M +NT(r, a)) is closed, then the condition
is sufficient.
9.2 Linear autonomous equations
In this section, we consider the theory of a linear autonomous NFDE(D, L)

where D and L are continuous linear functions from C into lRn. The theory
completely parallels the theory for RFDE given in Chapter 7 and, in fact,
the proofs in Chapter 7 were given in such a way that many would carry
over almost verbatim for this situation.
Consider the homogeneous linear autonomous NFDE( D, L)
(2.1)
where D, L : C --+ lRn are continuous and linear and D is atomic at zero.
Following Representation (1.2), we write
D¢ = ¢(0) - /_: d(J.L(0)]¢(0),

9.2 Linear autonomous equations 263
where Var[s,o] JL --+ 0 as s --+ 0. Let x( ¢) denote the solution of Equation

(2.1) with x 0 (¢) = ¢. By making slight modifications to the proofs of
Lemma 1.2 of Section 7.1 and Lemma 2.1 of Section 7.2 and of the proof
of Theorem 2.1 of Section 7.2, one obtains the following result.
Lemma 2.1. The solution operator T(t), t?: 0, defined by
T(t)cp ~f Xt(c/J),
is a Co -semigroup with infinitesimal generator

d¢
(2.2) A¢= dB.
If A is defined by Equation (2.2), then a(A) = Pa(A) and>. is in a(A) if

and only if >. satisfies the characteristic equation
(2.3) det Ll(>.) = 0, Ll(>.) = >.D(e>..· I)- L(e>..· I).
For any>. in a(A), the generalized eigenspace M>..(A) is finite dimensional

and there is an integer k such that M>..(A) = N( (>.I- A)k) and we have
the direct sum decomposition:
(2.4) C= N( (>.I- A)k) EEl R( (>.I- A)k ).
Suppose A is a finite set {>.1, ... , >.p} of eigenvalues of Equation (2.1)

and let PA = {<hp···,if>>..p}, BA = diag(B>..p···,B>..p), where P>..i is a
basis for the generalized eigenspace of Aj and B>..i is the matrix defined by
Ail>>..i = P>..i B>..i, j = 1, 2, ... , p. Then the only eigenvalue of B>..i is Aj and,
for any vector a of the same dimension as if> A, the solution T(t) Aa at t = 0 may be defined on (-oo, oo) by the relation
T(t) AeBAta,

(2.5)
PA(B) = PA(O)eBAII, -r ~ () ~ 0.
Furthermore, there exists a subspace QA of C such that T(t)QA ~ QA for

all t?: 0 and
c = PA EElQA,
where PA = {¢ E C I¢= il>Aa, for some vector a}.
The spectral analysis of unbounded operators developed in Section

7.3 extends to NFDE. Set 8 = lRn x C endowed with the product norm
topology
(2.6)
and define the operator .A: v(.A) --+ 8

A\ ~ d¢
V(A; = {(c, ¢) E C: d() E C, c = D¢},
(2.7)
~ d¢
A(c,¢) = (L¢, dB).
This is a closed unbounded operator, but the domain of A is not dense in

C, and hence, A is not the generator of a semigroup. The closure of the
domain is precisely given by jC and the part of A in jC is given by Relation
(2.2).
Since v(A) c jC and A is the part of A in jC, it follows that the
spectrum of A and A are the same and JM>.(A) = M>.(A).
The structure of A is similar to the one studied in Section 7.3 and an
application of Theorem 3.1 of Section 7.3 yields the following result.
Theorem 2.1. The matrix function .:1 : <C----> <Cnxn
is a characteristic matrix for A and the equivalence is given by
(2.9) ( .:10(z) o)
I = F(z)(zi- A)E(z),
~
z E <C,
where the explicit formulas forE : <C ----> £( C, C(A)) and F : <C ----> £( C, C)
are given in Theorem 3.1 of Section 7.3.
Corollary 2.1. The spectrum of the operator A: V(A) ----> C defined by (2.2)
consists of eigenvalues of finite type only,
(2.10) a-(A) = {A : det .d(A) = 0}.

For A E a-( A), the algebraic multiplicity of the eigenvalue A equals the order
of A as a zero of det .:1, the partial multiplicities of the eigenvalue A are equal
to the zero-multiplicities of A as a characteristic value of .:1, and the largest
partial multiplicity (ascent) of A equals the order of A as a pole of .::1- 1 .
Furthermore, a canonical basis of eigenvectors and generalized eigenvectors
for A at A may be obtained in the following way: If
{(!'i,o, ... ,')'i,ki-d Ii = 1, ... ,p}

is a canonical system of Jordan chains for .:1 at A ED, then
Xi,O, · · ·, Xi,ki-1' i = 1, ... ,p,
where
v ()l
Xi,v (B) = e>.IJ L 'Yi,v-l[f'
l=O
9.2 Linear autonomous equations 265
yields a canonical basis for A at >..

From the equivalence relation (2.9), we have the following representa-
tion for the resolvent of A.
Corollary 2.2. Let A: V(A) ---+ C be the generator defined by Relation (2.2).
The resolvent of A has the following representation
where
From the adjoint theory developed in Section 9.1, we know that the
adjoint semigroup T*(s) = T(s)* corresponds to the Volterra equation
(2.13) y(s) + 1°y(r) d[k(r- s)] = g(s), s:::; 0,
with g E Bo and k given by
(2.14) k(t) = k(t, 0) = JL( -t) + 1t rt( -r) dr.
Recall that B 0 denotes the Banach space of row-valued functions '1/J

(-oo, OJ ---+ JRn* that are constant on ( -oo, -r], of bounded variation on
[-r,O], continuous from the left on (O,r), and vanishing at zero with norm
Var [-r,o] '1/J. The generator of T*(t) equals A* and can be easily computed.
Lemma 2.2. The adjoint operator A* : V(A*) ---+ Bo is given by

d
(2.15a) V(A*) = {f E Bo : d() [f(O- )JL + f] E Bo}
(2.15b) A* f(O) = f(O- )ry(B)- :() [f(O- )JL(B) + f(B)].
Proof. It is obvious that given the action of A*, the domain of definition
of A* cannot be larger than the subspace defined in the right-hand side of
(2.15a). So assume that g E B 0 with g(-r) = 0, and
f(O- )JL(B) + f(O) = f(O-)- leo g(s) ds, for () < 0.
From Relation (2.2)

(f, A¢) = j_or A¢( e) dj(e)

= j_: dry(e)¢(e)f(o-) + j_: ~:(e) d[f(O- )M(e) + f(e)]
= j_: ¢(e)d[f(o-)ry(e)]- j_: ¢(e)dg(e) = (A*J,¢).
As in the retarded case, the adjoint equation (2.13) is closely related

to the transposed equation. Define C' = C([O, r], IRn*). For each s E [0, oo)
let ys designate the element inC' defined by y 8 (~) = y( -s + ~), 0::::; ~::::; r.
The transpose of System (2.1) is defined to be
(2.16) ddS [y(s)- ! 0

-r
y(s- e) d[JJ(e)l] = !-r
0
y(s- e) d[ry(e)], s::::; 0,
y0 = '1/J, 'ljJ E C'.

Let y be a solution of Equation (2.16) on an interval ( -oo, r]. We associate
a C0 -semigroup TT(s) with Equation (2.16), defined by translation along
the solution,
(2.17)
where y is the solution of Equation (2.16). By making slight modifications
to the proof of Lemma 1.2 of Section 7.1, one can prove the following result.
Lemma 2.3. The solution operator TT(s), s 2 0, defined by Relation (2.17)

is a C0 -semigroup with infinitesimal generator
(2.18)
Integration of Equation (2.16) yields a Volterra equation
y(s) + 1° y(T) d[k(T- s)] = pT '1/J(s), s::::; 0,
where pT : C' __, B 0 is given by
(Fr '1/J)(s) = 1°
'1/J(O) + '1/J(s- e) dJJ(e)
(2.19)
-1°j_: '1/J(a- e)d[77 (e)] do:
902 Linear autonomous equations 267
and k is given by (2014)0 From the definitions, it is not difficult to verify

that
(2020)
One can give an explicit characterization of Decomposition (204) via

the formal adjoint A* Together with the properties of the adjoint operation,
the equivalence (209) for A implies the following equivalence relation for A*:
0
(2021) (a f) ( li(~)T ~) = E(z)*(zi- A*)F(z)*(af),
where E(z)* : C(A)* --+Eo and F(z)* : Eo --+ E0 are bijective mappings
that depend analytically on z, z E ~0
In particular, we find that the Jordan chains of length k of Li(z)T are
in one-to-one correspondence with the solutions of
The proof of the following result follows immediately from the correspond-
ing results in the RFDE case; see Lemma 501 and Theorem 501 of Section
7050
Theorem 2.2. The spectrum of the infinitesimal generator A* defined by

(1. 7) consists of eigenvalues of finite type only,
(2022) a(A*) ={A: det Li(A) = O}o
For ,\ E a( A*), the algebraic multiplicity of the eigenvalue ,\ equals the

order of ,\ as a zem of det li, the ascent of ,\ equals the order of ,\ as a
pole of ,1- 1 Furthermore, a canonical basis of eigenvectors and generalized
0
eigenvectors for A* at ,\ may be obtained in the following way: If
{(,6i,O, o o o ,,6i,ki-d: i = 1, o o o ,p}
is a canonical system of Jordan chains for LiT at ,\ E ~~ then
{Xi,O o o o, Xi,ki-1 : i = 1, o o o ,p}

where Xi,v = FT '1/Ji,v with
,/, (c) _ -A~~ r:; ( -~)l

o/i,v <, - e ~ fJi,v-l_l_! -'
l=O
yields a canonical basis for A* at ,\0 Here FT : C[O, r] --+ B 0 is the mapping
defined by (2019) 0
From the theorem, we conclude that the (generalized) eigenfunctions

of A* are precisely given by the images under the operator FT of the (gen-
eralized) eigenfunctions of the transposed generator AT. Furthermore, from
(2.20)
In general, the mapping FT is not one-to-one (see Section 3.3), but on the
generalized eigenspace of AT, it is. Hence, if'¢ is a (generalized) eigenfunc-
tion of AT, then FT '¢ is a (generalized) eigenfunction of A*. This motivates
the introduction of the following bilinear form. For '¢ E C' and ¢ E C define
('¢,¢)~f(FT'¢,¢) =I: d(FT'lj;(B)]¢(B)

= '¢(0)¢(0) + 1: 0
d[i 'lj;(B- a) d[~t(a)l]¢(B)
(2.23) +[or 1°'lj;(B- r) d[ry(r)]¢(B) dB
= '¢(0)¢(0)- [or d[1° 'lj;(B- a)d[~t(a)l]¢(B)
-1: 1a. 'lj;(B- r) d[ry(r)]¢(B) dB
between C and C'. With respect to this bilinear form, the transposed op-
erator AT satisfies
Lemma 2.3. For,\ in u(A), let l]t;. =col ('¢1, ... , 'l/Jp) and iP;. = (¢1, ... , ¢p)
be bases for M;.(AT) and M;.(A), respectively, and let (l]t;., iP;.) = ('l/Ji, ¢j),
i,j = 1,2, ... ,p. Then (l]t;.,iP;.) is nonsingular and thus may be taken as
the identity. The decomposition of C given by Lemma 2.1 may be written
explicitly as
¢ = ¢P>- + ¢Q>-, ¢P>- in P;., ¢Q>. in Q;.,

P;. = M;.(A) = {¢ E C: ¢ = iP;.b for some p-vector b}
Q;. = {¢ E C: (!P;.,¢) = 0},
qyP>. = P;.b, b = (l]t;., ¢), qyQ>. = ¢- qyP>-.
If A is a finite set of eigenvalues and C is decomposed as in Lemma

2.2, we say Cis decomposed by A. In Chapter 1, we have seen that the set
A = {,\ E u(A) : Re ,\ 2: ,B} is not necessarily finite. So it is not clear that
we can decompose C according to this infinite set of eigenvalues. We will
return to this question in Section 9.4.
9.3 Exponential estimates 269
Suppose C can be decomposed by A as C = PA EB QA. Can we estimate

T(t) on the complementary space QA? In the next section, we shall address
this question.
9.3 Exponential estimates
In Corollary 1.1, we proved that the solutions to System (2.1) are exponen-
tially bounded. It is our aim to prove precise exponential estimates for the
solution.
In Section 9.1, we have seen that we can rewrite the System (2.1) with
initial condition x 0 = ¢ as a Volterra-Stieltjes equation
(3.1) x(t) -lot d[k(t- s)]x(s) = f(t)

where
(3.2) k(t) = J.t(-t) +lot ry(-r)dr,

(3.3) f(t) = ¢(0) +lot d[tL(B)]¢(t +B)+ lot j_~a d[ry(B)]¢(a +B) da.
From Theorem 1.2, we find that the solution of Equation (3.1) has a simple
representation
(3.4) x(t) = X(t)f(O) +lot X(t- a) df(a)

where X(t) is the matrix solution to
(3.5) X(t) =I+ lot d[k(t- a)]X(a).

To estimate the solution of Equation (3.1) we are going to apply the
Laplace-Stieltjes transform to X (t).
First we present some results from the theory of vector-valued Laplace
transforms needed in the sequel. Let B be a complex Banach space, cJ> a
function on [0, oo) to B, and let cJ> be of strong bounded variation over
every compact interval. If there exist constants C and a such that
cf>*(t) = Var[o,t) cJ> :=:; Ceat, t ~ 0,

then the Laplace-Stieltjes transform of a is defined by
(3.6) f(z) = looo e-zt dcf>(t).

For Re z > a, the integral is absolutely convergent and exists as a Bochner

integral. We let CJ"a(<P) denote the abscissa of absolute convergence of (3.6);
that is,
CJ"a(<P) = inf{CJ" E lR I 1 00
e-atd<P*(t) converges}.
If g: lR+-+ B is bounded on bounded intervals and <P(t) = J~ g(s) ds, then

 is of strong bounded variation and
The Laplace-Stieltjes transform of is given by
(3.7) .C(g)(z) = 100

e-ztg(t) dt
and is called the Laplace transform of g.

Since Vaqo,t) X(·) satisfies the exponential bound (1.14), there exists
an a E lR such that .C(X) exists and is analytic for Re z > a. To compute
the Laplace-Stieltjes transform of X, we use the following lemma.
Lemma 3.1. (Convolution of Laplace-Stieltjes transform.) Let 7] be ann x

n matrix-valued function of bounded variation and let a be an n-vector
function of bounded variation. If the Laplace-Stieltjes integral
f(z) = 1 00
e-zt da(t), g(z) = laoo e-zt dry(t),
are absolutely convergent for Re z > era, then
(3.8) g(z)f(z) = 1 00
e-zt d((t) with ((t) =fat d[ry(B)]a(t- B).
So from Equation (3.5), we find for Rez >a
1
00
e-ztd[X(t)] =1 00
e-ztd[k(t)]1
00
e-ztd[X(t)]
= -Ll(z)
1
z
10
00
e-ztd[X(t)]
and a simple computation yields
(3.9) .C(X)(z) = Ll(z)- 1
1:
where Ll(z) is given by Formula (2.8). Set Ll(z) = zL1 0 (z)- J~,.. ezt d[ry(t)]
where
Llo(z) =I- ezt d[t~(t)].
In order to control the behavior of IL1(z) I as lzl --+ oo, we make the
following assumption on the kernel f.L:
(J) The entries /-Lij of f.L have an atom before they become constant, i.e.,
there is a tij with /-Lij (t) = f.Lij (tij +) for t ;:::: tij and /-Lij (tij-) =/=
f.Lij(tij+ ).
For example, f.L can be a step function. In that case
00
L1 0 (z) =I- L_:>-zrj Aj

j=l
and det .:1 0 is an almost-periodic function. The jump condition (J) is much
more general and implies that det .:1 0 is asymptotically almost-periodic.
Define
<C, 1 , ,2 = {z E <C l1'1 < Rez < 1'2} and <C, = {z E <C I Rez > ')'}.
Lemma 3.2. If f.L satisfies (J), then the zeros of det .:1 0 are located in a finite
strip <CI'om and there exist positive constants E, m and M, such that
outside circles of radius E centered around the zeros of det .:1 0 . Here T is the
exponential type of det .do.
Proof. From Lemma 3.1, it follows that there is a function f.L* such that
(3.10)
If f.L satisfies (J), then f.L* has jumps at both endpoints 0 and T. An appli-
cation of Verduyn Lunel [2], Theorem 4.6, yields the lemma. 0
Theorem 3.2. If av,L = sup{Rez: det Ll(z) = 0}, then, for any a> av,L,
there is a constant C = C(a) such that the fundamental solution X of
System ( 2.1) satisfies the exponential estimates
(3.11)
Before we prove the theorem we need some preparation. Let S denote

the Schwartz space, that is, the space of C 00 -functions such that
sup lxkq~(q)(x)l :::; mkq k,q = 0, 1, ....

xElR
These pseudonorms can be used to make S a complete linear topological

vector space. Tempered distributions are representations of continuous lin-
ear functionals on S.
If f : lR ----* lR is a bounded function, then P f-t (!, P) denotes the

tempered distribution defined by f on S. For PES, the Fourier transform
:FP(s) = -21
7f
!-00
00
eistP(t) dt
exists, and is an injective continuous mapping from S onto S. Its inverse is

also continuous, and is given by
By duality, one defines the Fourier transform :Ff of a tempered distribution

f onS
(3.12) (:Ff,P) = (f,:FP), PES.
Note that
I:FPI1 = 2~ IF- 1PI1
where I · 11 is the £ 1-norm. The following lemma will be useful in the
estimations.
Lemma 3.3. Let f : lR ----* lR be a bounded function. The (generalized)

Fourier transform of f belongs to Coo if and only if
(3.13) PES.
Proof. From (3.13) one finds
I(:Ff,:F- 1P)I = l(r,P)I ~ KI:FPI1 ~ !IIF- 1PII1> PES.
Since :F[S] = S and S is dense in £1, we conclude that :Ff E .Coo. On the
other hand, if :Ff belongs to Coo, then
Proof of Theorem 3.2. In any right half plane Re z > "Y

(3.14) det Ll(z) = zndet Ll 0 (z) + O(zn- 1) as lzl ----* oo in Re z > "Y·
Let o: > an,L· From the Rouche theorem, it follows that det Llo(z) i- 0
for Re z > aD ,L. Therefore, there exists a strip o: - E < Re z < o: + E
such that detLl 0 (z) i- 0 on this strip. From Lemma 3.2, we derive that
ldet Ll 0 (z) I 2: C for z E o:- E < Re z < o: + E From Representation (3.11)
for det Ll 0 (z) and the fact that there exists a constant C > 0 such that
ldet Llo (a + iv) I 2: C for v E lR, it follows that there exists a function (
such that ((t) = 0 fort< 0, e"' · ( is of bounded variation on lR, and
(detLlo(a+iv))- 1 = 1 00
e(e>+iv)td((t)
(see for example Hille and Phillips [1], pp. 144-150). The cofactors of Ll(z)
are polynomials of degree n - 1 with entire coefficients that are bounded in
(;-w,w· Therefore, we can expand
(3.15) Ll(a+iv)- 1 = r 00 e(e>+iv)td[((t)l(~+0( 1. ))

Jo a+w (a+w) 2
as lvl--t oo. Let x E lRn andy E lRM and set f(v;x,y)) = yLl(a+iv)- 1x.
From the Laplace inversion formula and a simple contour integration, we
find
So
le-"'tX(t)xl = sup le"'tyX(t)xl = sup IF!Ioo
IYI9 IYI9
and it remains to prove that Ff belongs to £ 00 • From (3.15) it follows that
it suffices to analyze the Fourier transform of
(3.16) roo e-(e>+iv)t d((t)y~x.

Jo a+w
In order to use Lemma 3.3, we have to estimate
l1
I
oo
-oo o
oo
d((t)y
A e-(e>+iv)t
1 .
a+ zv
x(v) dvl
=I roo d((t)
Jo J±oo
t
e-C<CTF(a)yA1xdal
: :; 11 00
e-"'t d((t)IFI1IYIIA1IIxll
::::; IAIIIxiiiiYIIIIFII1,
where we have used, in order to reverse the order of integration, that E S
and e-"' · ( is of bounded variation on JR. So it follows from Lemma 3.3 that
the Fourier transform of (3.16) is bounded. This shows that X(t) satisfies
the estimate in Formula (3.13).
In order to estimate the variation of X(t) we write Equation (3.5) as
follows
(3.17) X(t) =I+ 1t d[X(t- a)]J.l( -a) -1t X(t- a)ry( -a) da.
If X*(t) = Var[o,t)X = f~ ldX(T)I denotes the variation of X, then it follows

from Equation (3.17) that X* satisfies the inequality
X*(t)::::; fat X*(t- B) diJ.L(B)I + Ceo.t.

The assertion now follows from Lemma 3.1 of Chapter 1. D
Corollary 3.1. IfaD,L = sup{Rez: detLl(z) = 0}, then, for any a> aD,L,
there is a constant c = c(a) such that the solution x( ·; ¢) of System (2.1)
with xo = ¢ satisfies the exponential estimates
In particular, if aD,L < 0, then all solutions of System (2.1) approach zero
exponentially.
Proof. This is an immediate consequence of Representation (3.4). D
We end this section with an application to homogeneous and inhomo-

geneous difference equations
(3.18) t2::0
and
(3.19) Dyt = h(t),

where hE C([O,oo),IRn). If
(3.20) CD = {¢ E c : D¢ = 0}
then the theory ofthe previous section implies that Equation (3.18) defines
a Co-semigroup TD(t) :CD-+ CD.
As before, we can rewrite Equation (3.19) as a Volterra-Stieltjes
equation
(3.21) x(t) -fat d[J.L(t- s)]x(s) = ¢(0) +fat d[J.L(B)]¢(t +B)+ h(t).
So the theory developed earlier can be applied, and we find the following
result.
Theorem 3.4. Let y('lj;, h) denote the solution of Equation (3.19) satisfying
Yo('I/J,h) = '1/J. If aD= sup{Rez: detLlo(z) = 0}, then for any a> aD,
there is a constant C = C(a) such that
ly('lj;, h)(t)l::::; C(a) [1'1/Jieo.t + sup lh(s)l]

O~s~t
fort 2:: 0.
9.4 Hyperbolic semigroups 275
Definition 3.1. SupposeD : C--+ IRn is linear, continuous, and atomic at 0.

The operator D is said to be stable if the zero solution of the homogeneous
difference equation (3.18) with y0 = '1/! E Cv is uniformly asymptotically
stable.
The following theorem is a simple consequence of these results.
Theorem 3.5. The following statements are equivalent:

(i) D is stable.
(ii) av < 0.
(iii) There are constants C > 0 and a > 0 such that for any h E
C([O, oo ), IRn), any solution y of the nonhomogeneous equation (3.19)
satisfies
(3.22) jy('lj!, h)(t)l :::; C(a) [1'1/Jie-at + sup jh(s)j].

O:Ss:St
(iv) If D¢ = ¢(0)- f~r d[J.L(B)]¢(8), Var[-s,O] --+ 0 ass--+ 0, and J.L satisfies
(J), then there is a 8 > 0 such that all solutions of the characteristic
equation
det.:1 0 (,\) = det [I- [or e>- d[J.L(B)]

0 = 0
satisfy Re ,\ :::; -8.
9.4 Hyperbolic semigroups
In Corollary 3.1, we have proved exponential estimates for the solution of

System (2.1). These estimates, in particular, imply that for the semigroup
associated with System (2.1), we have
IIT(t)!l :::; Meat
where a> av,L = sup{Rez: detLl(z) = 0}. In this section, we shall

further extend the results in Section 7.6 for neutral equations and give nec-
essary and sufficient conditions such that the space C can be decomposed
according to Ap = {z : Re z ?: ,6}.
Let B be a complex Banach space and A : D (A) --+ B be the generator
of a Co-semigroup (T(t))t?_o on B. We say that a Co-semigroup (T(t))t?_o
on a Banach space B is hyperbolic when the space B decomposes into B =
B_ EBB+ such that T(t)B± c B±,
T_(t): B_--+ B_, T_(t)x = T(t)x

extends to a C0 -group on [3_ over -oo < t < oo, and there are positive
constants K, a, (3 such that
(4.1) IIT-(t)ll:::; KetJtiiiixll, t:::; 0,

(4.2) IIT+(t)xll:::; Ke-"tii(I -II)xll, t ~ 0.
It is clear that a semigroup is hyperbolic if and only if there is an open

annulus containing the circle { z E <C : Iz I = 1} in the resolvent set of
T(t). In many applications, however, only the generator is explicitly known.
Therefore, one would like to have necessary and sufficient conditions on the
generator such that the semigroup is hyperbolic.
First, we need a more general version of the Laplace inversion formula.
Let .P be a locally (Bochner) integrable function on [0, oo) with values in
!3. Denote by
(4.3) f(z) = 1
00
e-zt.P(t) dt
the Laplace transform of .P. If we define aa(<P) to be the abscissa of absolute

convergence of the Laplace transform of .P, that is, aa(<P) is the infinimum
of the real numbers a such that
converges for Re z > a, then for c > a a ( <P)
fort< 0;
1 lc+ioo { 0,
(4.4) ~(C,1)- eztf(z)dz= ~.P(O+), fort= 0;
1f~ c-ioo ~[<P(t+)+<P(t-)], fort> 0
whenever the expressions in the right-hand side of (4.4) have a meaning.

Here, the Cesaro mean of integrals is given by
(C,1)- l
c+ioo
eztf(z)dz= lim
JN e<c+•v)tf(c+iv)(1--)dv.
. lvl
c-ioo N-+oo -N N
The Cesaro mean is a weaker notion than the principal value of the integrals,
but if the principal value exists, it equals the Cesaro mean.
For example, (4.4) holds if .Pis locally of bounded variation. To prove
(4.4), apply a continuous linear functional to both sides of (4.4), use the
scalar version of (4.4) from Widder [1], theorem II.9.2, and apply the Hahn-
Banach theorem.
In particular, if A: V(A) --) l3 is the generator of a Co-semigroup on
l3 and ¢(t) = T(t)x, <P(t) = T(t)x, then it is known that the order of T(t)
is the supremum of aa(T( · )x) over x E !3. Since the resolvent of A equals
the Laplace transform of the semigroup, we conclude that for c > w(A),
9.4 Hyperbolic semigroups 277
1 ~c+ioo
(4.5a) T(t)x = -.(C, 1)- eztR(z,A)xdz, t>O
2nz c-ioo
and
1 1 ~c+ioo
(4.5b) -x=-.(C,1)- R(z,A)xdz
2 2nz c-ioo
The idea is to use the inversion formula (4.5a) to analyze the semigroup
T(t) from properties of the resolvent only.
One has the following abstract characterization result (see the Supple-
mentary remarks for references).
Theorem 4.1. Let A : V(A) ---* B be the generator of a Ca-semigroup

(T( t) )t>o. The semigroup T( t) is hyperbolic if and only if there is an open
strip containing the imaginary axis on which the resolvent is uniformly
bounded and there exists an w > 0, such that
(i) for each ¢ E B and 0 < lhl < w, the integral
(C, 1)-1~:~00 R(z, A)¢dz exists;
(ii) for each ¢ E B, a E B* and 0 < Ihi < w, the function r( ·, h; ¢,a)
IR ---* <C defined by
(4.6) r = r(v, h; ¢,a) = (a, R(h + iv, A)¢)
satisfies
l(r,«P)I::; Kll¢1111aiiiiF«PIIl, for all 4> E S
where S denotes the Schwartz space.
The first condition implies that one has the following integral repre-
sentation for the semigroup
1 1p+ioo
(4.7) T(t)x= -.(C,1)- eztR(z,A)xdz, t > 0.
2nz p-ioo
Therefore, it is easy to see that
e-ptT(t)x 1
= -(C,1)- joo . tR(p+iv,A)¢dv
e211
2n _ 00
= Fr(t).
So e-PtT(t)x equals the generalized Fourier transform of r where r is given
by Equation (4.6). From the second condition and Lemma 3.4, one derives
that Fr E L 00 •
It is not difficult to apply the theorem to the C0 -semigroup associated
with System (2.1).
Theorem 4.2. If f..L satisfies (J), and det L1 0 has no zeros in a strip -6 <
Rez < 6, then the C0 -semigroup TD,L(t) associated with System (2.1) is
hyperbolic if and only if det L1 has no zeros on the imaginary axis.
The proof of the theorem will be an application of Theorem 4.1. The argu-
ments are similar to those given in the proof of Theorem 3.2 and are left to
the reader.
We end this section with some simple corollaries of the exponential
estimates in Section 9.3.
Theorem 4.3. If aD,L denotes the order ofTD,L(t), the semigroup associated
with System (2.1), then
aD,L = sup{Rez: detLl(z) = 0}.
Theorem 4.3 gives a way to determine the asymptotic behavior of the

semigroup TD,L(t). In particular, we have the following important result.
Corollary 4.1. If there is a 6 > 0 such that the zeros of det L1 are in the left
half plane Re z :S: 6 < 0, then there are positive constants K and o: such
that
t :2: 0,
that is the zero solution of System (2.1) is uniformly asymptotically stable.
In a similar manner, we can discuss the behavior of C0 -semigroup T D ( t)

associated with the homogeneous difference equation.
Theorem 4.4. If aD denotes the order of TD(t), the semigroup associated

with System (2.1), then
(4.8) aD= sup{Rez: det Llo(z) = 0}.
The semigroup TD(t) yields important information about the per-

turbed semigroup TD,L· Note that from the Rouche theorem, it follows
that in the half plane Re z >aD, the characteristic equation det Ll(z) = 0
can only have finitely many roots. So if we decompose C by A = {Re z > a},
a> aD, then A is finite and C can be decomposed as in Lemma 2.3.
An eigenvalue f..L of a bounded linear operator T on a Banach space is
called a normal eigenvalue if it is an isolated point of the spectrum of a(T)
of T and the corresponding generalized eigenspace is finite dimensional. A
normal point of T is either a normal eigenvalue or a point of the resolvent
set p(T) ofT. Let p(A) denote the set of normal points ofT. The essential
spectrum of T is the set <C \ p(T) in the complex plane.
9.5 Variation-of-constants formula 279
Theorem 4.5. The mdius re(D, L, t) of the essential spectrum of Tv,L(t)

satisfies
(4.9) t;:::o
where av is given in (4.8). If, in addition, D satisfies
D¢ = D 0 ¢ + /_: A(0)¢(0) dO
00
(4.10) Do¢= ¢(0)- LAk¢(-rk), 0 < Tk:::; r

k=l
then
(4.11) r
e
(D L t) <
' ' -
eanot
'
t:::: 0.
The estimates can easily be proved using the Rouche theorem, see
the remarks made before the theorem. A different proof for Estimate (4.9)
follows from the representation for Tv,L(t) in Theorem 7.3 of Section 3.7.
If D satisfies (4.10), then the representation for Tv(t) in Theorem 7.4 of
Section 3.7 yields estimate (4.11).
9.5 Variation-of-constants formula
The purpose of this section is to discuss the variation-of-constants formula

for linear inhomogeneous systems.
If G E C([O, oo), lRn), FE £ioc([O, oo), lRn), then the nonhomogeneous
linear NFDE(D - G, L +F) is
d
(5.1) dt [Dxt- G(t)] = Lxt + F(t), t:::: 0.
Following the proofs of the corresponding results in Section 6.1 and 1.7,
one shows there are constants a and b such that the solution x( ·; ¢, G, F)
of Equation (4.2) with xo ( · ; ¢, G, F = ¢) satisfies the estimate
fort;::: 0.
Following representation (1.2), we write
-I:
(5.3) D¢ = ¢(0) d[JL(0)]¢(0),
where Var[-s,O] 11 --> 0 as s --> 0. From the results in Section 9.1, it follows
that Equation (5.1) has a fundamental solution X(t), t 2: -r, i.e., X is of
bounded variation on compact sets, X(t) = 0 for -r ~ t < 0 and fort 2: 0
satisfies the equation
(5.4) X(t) =I+ lot d[X(t- a)]tL( -a) -lot X(t- a)ry( -a) da.
Theorem 5.1. Let T(t), t 2: 0, denote the solution operator defined by

the NFDE(D,L). If X is defined by Equation (4.4), then the solution
x = x(¢, G, F) of Equation (5.1) with xo(¢, G, F) = ¢ satisfies the
relation
x(t)- X(O)G(t) = T(t)¢(0)- X(t)G(O) +lot X(t- s)F(s) ds

(5.5)
-lot d8 [X(t- s)]G(s)
fort 2: 0.
Proof. From Theorem 1.2, it follows that
x(t) = X(t)¢(0) +lot X(t- a) df(a),
where
f(t) = ¢(0) - G(O) + G(t) +lot d[JL(O)]¢(t + 0)
+lot [~a d[ry(O)]¢(a + 0) da.
This yields
x(t) = T(t)¢(0) +lot X(t- s)F(s) ds +lot X(t, a) d[G(a)- G(O)].
Together with
lot X(t- a) d[G(a) - G(O)] = X(O)G(t) - X(t)G(O)
-lot d[X(t- a)]G(a).
This proves the theorem. D

If G = 0, this formula is the same as the one for RFDE. So with

G =0, results on perturbed linear systems will follow in a manner similar
to RFDE. If G ¢. 0, then new ideas must be employed.
First we use Lemma 2.1 of Chapter 6 to write Equation (5.5) as an
integral in the Banach space C
Xt- XoG(t) = T(t)¢- XtG(O) +lot d[K(t, s)]F(s)

(5.6)
d rt
-de lo d[K(t, s)]G(s)
where K(t, s)(e) = J; X(t+e-a) da. Note that if G ¢. 0, then the variation-
of-constants formula is no longer an integral equation.
Note that Xt - X 0 G(t) and T(t)¢ - XtG(O) do not belong to C. In
order to estimate the solution, we would like to make a transformation
of variables. In order to do so, we have to introduce the product space
C = JRn x C with the embedding into Coo given by (c, ¢) = X 0 c + ¢. On
C, one defines the following solution operator T(t) : C-+ Coo
T(t)(c, ¢) = Xtc + T(t)¢.
Define the unbounded operator
If we set Zt = Xt- X 0 G(t), then the variation-of-constants formula becomes
(5.7) Zt = T(t)zo +lot d[K(t, s)]F(s)- M lot d[K(t, s)]G(s).
Definition 5.1. If G: C-+ JRn is continuous, we say that G(¢) is indepen-

dent of ¢(0) if there is an E E [-r, 0) such that G(¢) depends only on the
e
values c/J( e) of the function c/J for E [-r, E].
We are now in a position to prove the following result.
Theorem 5.2. Suppose D, L : C -+ JRn are linear and continuous and

the zero solution of the NFDE( D, L) is uniformly asymptotically stable. If
F, G : C-+ JRn are continuous together with their first derivatives Fq,, Gq,
and F(O) = G(O) = 0, Fq,(O) = Gq,(O) = 0 and G(¢) is independent of ¢(0),
then the zero solution of the equation
is exponentially asymptotically stable.

Proof. Let j : C ---+ £ 00 be the embedding j (c, ¢) = Xoc + ¢. Consider the

map
l: c---> c, l¢ = (-G(¢), ¢).
Since G(¢) does not depend on ¢(0), the mapping lhasa continuous inverse,
and since h is a homeomorphism in a fixed neighborhood of¢ = 0 E C,
(c, 'lj;) = (0, 0) E C. Furthermore, there are constants k1 > 0 and k2 > 0 such
that in this neighborhood, (c,'lj;) = l(¢) implies 1'1/JI:::; k1I<PI and 1¢1:::; k2l'l/JI-
For Equation (5.1), the variation-of-constants formula is given by (5.5).
Seth: C---+ L 00 , h = j o l. If Zt = h(xt) and 'ljJ ish(¢), then Zt satisfies the
equation
(5.8) Zt = T(t)'lj; + lt d[K(t, s)]F(h- 1 (z 8 )) ds- M lt d[K(t, s)]G(s).
Choose 8 so that o:- Kk 2 8 > 0. Applying the estimates from Theorem

3.2 and the hypotheses on F and G, there is an E(8) > 0 such that z in
Equation (5.8) satisfies
lztl < Ke-ati'1/JI + lt Ke-a(t-s)8k2lzsl ds
as long as lzsl < E(8). Applying the inequality in Lemma 3.1 of Section 1.3
to lztleat, we obtain
:21xtl :::; lztl:::; Ke-(a-Kk26)tl'l/JI :::; Kkle-(a-Kk26)tl¢1
as long as lzsl < E(8). Since this clearly can be assured for all t ~ 0 if 1¢1 is
sufficiently small, we obtain the result stated in the theorem. D
As one sees from this proof, the transformation h reduces the discussion
to an argument very similar to the one for ordinary differential equations.
One could easily generalize the results to more general perturbations G(t, ¢)
and F(t,¢) of the NFDE(D,L) where D and L can even depend on t.
Next we discuss the decomposition in the variation-of-constants for-
mula for the inhomogeneous linear equation (5.1). Observe that we can
also write Equation (5.6) as follows
(5.9) Xt=T(t)¢+ ltd[K(t,s)]F(s)- :e~atd[K(t,s)](G(s)-G(O)).

Suppose A= {.X 1, ... , Ap}, >.1 E O"(A), and suppose C is decomposed
by A as C = P EB Q. Let ¢ E C be written as ¢ = ¢P + ¢Q, ¢P E P,
¢Q E Q. Let X6 = <P!lf(O), where P is a basis for P and !lf is a basis
for pT, (!lf,<P) =I. Then K(t,s)P = J;T(t- o:)X6 do: and K(t,s)Q =
K(t, s)- <P(!lf, K(t, s)). Exactly as in Section 7.9, one obtains the following
result on the decomposition in the variation-of-constants formula.
Theorem 5.3. If A is a finite set of elements of a( A) and C is decomposed

by A as C = P EB Q, and Xt = xf + x~, xf E P, x~ E Q, then x satisfies
the equations
(5.10)
xf = T(t)¢P +lot T(t- s)Xt F(s)
-:e lot T(t- s)X6(G(s)- G(O))

x~ = T(t)¢ +lot d[K(t, s)Q]F(s)- :()lot d[K(t, s)Ql(G(s)- G(O))
fort~ 0.
It is now tempting to let xf = Py(t), where cf_j is a basis of P and

try to determine an ordinary differential equation for y as in Section 7.9.
However, the function y need not be differentiable in t. In fact, if r = 0;
that is, our equation is an ordinary differential equation,
d
dt [x(t) - G(t)] = Bx(t),
where B is ann x n matrix, then x(t)- G(t) is differentiable but x(t) is not
differentiable unless G(t) is differentiable. For an ordinary equation, one
could let z = x- G to obtain
d
dt [z(t)] = Bz(t) + BG(t)
and z(t) is differentiable in t. This is the same type of transformation that
was made earlier in this section. Using this transformation again, we find
from the variation-of-constants formula (5.7), that
z[ = T(t)zf: +lot T(t- s)Xt F(s)- M lot T(t- s)X6G(s)

(5.11)
z~ = T(t)z~ +lot d[K(t, s)Q]F(s) + M lot d[K(t, s)Q]G(s)
for t ~ 0. If Zt = Py(t), where cf_j is a basis for P and T(t)P = Pexp Bt,
t E ( -oo, oo) then y satisfies the equation
Py(t) = PeBty(O) +lot PeB(t-s) F(s) ds -lot d[PeB(t-s)IJi(O)]G(s)
= PeBty(O) +lot PeB(t-s) F(s) ds +lot PeB(t-s) BIJi(O)G(s)
for t ~ 0. Therefore, y(t) is continuously differentiable and-=satisfies the

ordinary differential equation -
(5.12) y(t) = By(t) + Blf/(O)G(t) + lf/(O)F(t).

For G = 0, this is the same equation as obtained in Section 7.9.
These relations can be generalized to the case of linear periodic equa-
tions, but the details are not given.
of xrIn the sequel, we must also have an estimate of the exponential growth
and K(t,s)Q. In particular, we need an exponential estimate for
X- X P, but this follows immediately from the hyperbolicity property of the
semigroup TD,L(t) and the proof of Theorem 3.2. We state the fundamental
inequalities.
Lemma 5.1. Let TD,L(t) be the semigroup generated by System (2.1) with
infinitesimal generator AD,L· If A= {A E a(AD,L) :Re-A 2 a} and C is
decomposed by A as C = P E9 Q, then there are positive constants M and E
such that for ¢ E C
ITD,L(t)¢pl :S Me(a-e)ti¢PI, Var(t,O] IT(t)X6'1 :S Me(a-e)t, t :S 0

IK(t, · )QI :S Me(a-e)t, Var[o,t) K(t, · )Q :S e<a-e)t, t 2 0.
Similar results can be obtained for the periodic case if we use the
decomposition theory from Chapter 8.
9.6 Strongly stable D operators
In Definition 3.1 of Section 9.3, we introduced the concept of a stable D

operator. In this section, we discuss how variations in parameters in D affect
the property of being stable. We shall only be concerned with difference
operators D(r, A) of the form
N
(6.1) D(r, A)¢ = ¢(0) - L Ak¢( -rk)
k=l
where each Ak is an n x n constant matrix, each rk is a constant, A =

(A1, ... ,Ak), and r = (r1, ... ,rk)·
In a problem, both the Ak and the rk are generally parameters that are
not known exactly. It is important, therefore, to know the effect that small
changes in the parameters will have on the stability of the zero solution of
the equation
(6.2)
From Theorem 3.5, we know that D is stable if and only if there is a

6 > 0 such that
9.6 Strongly stable D operators 285
N
(6.3) det Ll(r, A)(-\) = 0, Ll(r, A)(-\)= I - L Ake->-rk,
k=l
implies Re ,\ ::::; -8. It is not too difficult to show the following fact: for any
0 < 81 < 8, there is an E > 0 such that all zeros of det Ll(r, B)(-\) = 0
satisfy Re ,\::::; -81 if IAk- Bkl < E, k = 1, 2, ... , N.
The situation for variations in the rk is much more complicated. For
example, consider the equation
y(t) = ay(t- 1) + by(t- 2), t;:::: 0

whose characteristic equation is
1 - ae->- - be- 2 >- = 0.
The roots of this equation are given by
p=
a± va 2 +4b
p =eA.
2
For a= b = -1/2, IPI 2 = 1/2. Therefore, if 28 = -ln(~), then ReA=
-8 < 0. For a given integer n > 0, consider the equation
1 1 1
y(t) = -2y(t- 1 + 2n + 3) - 2y(t- 2).
It is easy to check that y(t) = sin(n+ (3/2))7rt is a solution of this equation

and it does not approach zero as t --> oo. Therefore, by taking n large, the
perturbation 1/(2n + 3) in the first delay can be made as small as desired
and, for each such perturbation, the equation has a solution that does not
approach zero. This happens even though the perturbed equation had all
roots of the unperturbed equation with real parts bounded away from zero.
In terms of the notation of the semigroups Tv(t), this example shows
that avis not continuous when one makes changes in the delays. It becomes
important, therefore, to characterize those D operators of the form (6.1)
for which one can preserve stability when small perturbations are made in
the delays. This section is devoted to this characterization.
Definition 6.1. Let (IR+)N be the cross product of IR+ by itself N times.
The operator D(r, A) is said to be stable locally in the delays if there is an
open neighborhood I(r) ~ (IR+)N of r such that D(s, A) is stable for each
s E I(r).
Definition 6.2. The operator D(r, A) is stable globally in the delays if it is

stable for each r E (IR+)N. In this case, we also say D(·,A) is strongly
stable.
The main result of this section is the following theorem.

Theorem 6.1. The following statements are equivalent:

(i) For some fixed r E (JR+)N, r = (rb ... , TN) with rk > 0 rationally
independent, D(r, A) is stable.
(ii) If "!(B) is the spectral radius of a matrix B, then 'Yo(A) < 1 where
N
(6.4) 'Yo(A) ~f sup{ 'Y(L Akei6k) : Ok E [0, 21r], k = 1, 2, ... , N}.
k=l
(iii) D(r, A) is stable locally in the delays.

(iv) D(r, A) is stable globally in the delays.
Proof. The scheme for the proof is (i) => (ii) => (iii) => (i) and (iii) {::} (iv).
(i) => (ii). Suppose r satisfies the hypotheses stated in (i), D(r, A) is
stable, and Statement (ii) does not hold. Since the Ok vary over a compact
set, there are 'Yo ~ 1 and ILk E [0, 21r], k = 1, 2, ... , N, such that 'Yo =
'Y(L:k Ak exp(iJ.Lk)). Let f(a) = I:;k Ak exp(iJ.Lk -ark) for a E JR. Since
rk > 0, "!(f(a))---+ 0 as a---+ oo. Since "!(f(O)) ='Yo~ 1, there are ao ~ 0
and Oo E [0, 21r], such that 'YU(ao)) = 1 and
(6.5) 0 = det [I- e-iOo f(a)J = det [I- L Ake-uorkei(JLk-Oolj.
k
Since the rk are rationally independent, it follows from Kronecker's theorem

that there is a sequence {tn} of real numbers such that
lim (tnr1, ... , tnrN) = (Oo- {L1, ... , Oo- {LN) (mod 27r).
n-HXl
We may also assume that {tn} is such that the sequence {exp(a0 + itn)}
converges to some (o = e>-o, Re >.o = ao, as n ---+ oo. Therefore, using
Equation (6.5), we have
O = det [I_ L Ake-(uo+itn)rkei(p.k-Oo+tnrk)J
k
and so
0 = det [I- L Ake->.orkj.
k
This contradicts the fact that D(r, A) is stable since Re >.0 = a 0 ~ 0. This
completes the proof that (i) implies (ii).
(ii) => (iii). Assume Equation (6.4) is satisfied and D(r, A) is not stable
locally in the delays. Then there exists a sequence {si} ~ (JR+)N and a
sequence {>.j} of complex numbers such that lr - si I < 1!J, Re Aj ~ -1 j j
and detLl(sJ,A)(>.j) = 0 for j = 1,2, ... Suppose there is a subsequence
of the {si}, {>.j}, which we label the same way, such that Re Aj ---+ 0 as
j ---+ oo. If Aj = a-3 + i(Jj, then the (3j satisfy the equation
N
det [I - L A{eif3jrk] = 0
k=l
where A{ = Ak exp( -Ajs{ + i(Jjrk) for all k and Aj = (A{, ... , A~,) --+ A
as j --+ oo. Since the spectral radius 'Y(B) of a matrix B is continuous in
B, this contradicts Equation (6.4).
Therefore, we may assume each element of the sequence { Aj} satisfies
Re Aj ~ t5 > 0 for some constant t5. Also, we may assume the sj are rational
since this can be accomplished by a small change in A. To simplify notation,
let the generic element of the sequences {Aj}, {sJ} be ..\, s = ( s 1 , ... , s N).
For any real x, consider the solutions z = z(x) of the equation
N
det [xi- L Ak exp( -skz)] = 0.

k=l
Since the Sk are rational, this is a polynomial in f.L = exp( -z/q) for some
integer q and the solutions J-L(x) of this polynomial equation can be chosen
as a continuous function of x. Also, ReJ-L(x)--+ -oo as x--+ oo. For x = 1,
we have a root..\ with Re ..\ > 0. Therefore, there is an x* E ffi, Jx* J> ..\ 0 (A)
in Equation (6.4) and a f.L* = exp iO*, 0* E ffi such that
N
det [x* I- L Ak exp iO*] = 0.
k=l
This contradicts Statement (ii) and completes the proof of the assertion
that (ii) implies (iii).
The fact that (iii) implies (i) is obvious. Since (iv) implies (i) and (i)
is equivalent to (iii), (iv) implies (iii). Since (ii) and (iii) are equivalent and
condition (ii) is independent of the delays, it follows that (iii) implies (iv).
The proof of the theorem is complete. D
An immediate consequence of Theorem 6.1 is the following result for

the scalar case.
Corollary 6.1. If each Ak is a scalar, then D(r, A) is stable locally in the

delays if and only if ~~=l JakJ < 1.
As another example, consider the real scalar difference equation
(6.6) x(t) = ax(t- r) + bx(t- s) + cx(t- r - s)

where a, b, and c are constants and r > 0, s > 0. The characteristic equation
is
(6.7)
Corollary 6.1 does not apply to this equation since the three delays r, s,
and r + s, cannot be varied independently.
To apply the theory of this section, we transform Equation (6.6) to
an equivalent matrix equation (equivalent meaning that the characteristic
equation is the same as Equation (6. 7)) involving only the two delays r and
s. The particular matrix equation is not important since Theorem 6.1 is a
statement only about the solutions of the characteristic equation (6.7).
If we let
x( t) ]
[ x(t-
y(t) = r) '
then Equation (6.6) is equivalent to the system
(6.8) y(t) = Ay(t- r) + By(t- s).
Theorem 6.1 implies that System (6.8) will be stable locally in the
delays if and only if
It is clear that we can take 81 = 0 and obtain this supremum. Therefore, it

is necessary to discuss the second-order equation
for all 8 E [0, 27r].

If we let z = (1 + >.)/(1- >.),then z inside the unit circle is equivalent
to ,\in the left half-plane. If we make this transformation and multiply the
resulting expression by (1 - ,\) 2 , then ,\ must satisfy the equation
f(,\) ~f [1 +a+ (b- c)ei 0 ]>. 2

(6.10)
+ 2(1 + ceie),\ + 1 -a- (b + c)eie = 0.
One may apply the Routh-Hurwitz criteria to the polynomial in Equa-
tion (6.10). After some rather lengthy but straightforward calculations, one
obtains the following necessary and sufficient conditions for the solutions
of Equation (6.10) to have real parts negative and bounded away from zero
uniformly in 8:
1 +a> lb+ ci
(6.11)
1-a>lb-cl.
This region in the a, b, c space is larger than the region Ia I + lbl + lei < 1
obtained in Corollary 6.1 for the equation
x(t) = ax(t- r 1 ) + bx(t- r2) + cx(t- r3)

X=O
I II II II II II I
\I\ I I I\ I I I\ I
\1 II II II II II
/\ /\ /\ /\ ,'\ /\ c(v) I
I I I I I I I I I I I
I II II II II 1/ I
Fig. 9.1.
where the parameters r 1 , r 2 , and r 3 could vary independently and not

through the relation r 3 = r1 + r 2.
As an illustration of where this latter example occurs in the applica-
tions, consider the lossless transmission line shown in Figure 9.1 with a non-
linear capacitor connected at x = 0; that is, the shunted transmission line.
Using the derivative outlined in the introduction, letting L and C be the
mutual inductance and capacitance in the line z = (L/C) 112 , a= (LC)- 1 12 ,
r1 = 2h/a, r2 = 2lz/a, f.l = (R1- z)j(R1 + z), and q = (R- z)/(R + z),
and supposing c( v) > 0, h is the inverse of the function J0v c( s )ds, one can
show that the function x(t) = J0v(t) c(s)ds, where vis the voltage at x = l2
satisfies the equation
d 2 2qu
(6.12) - Dxt = -- g(x(t)) + - g(x(t- r1- r2)) + e(t)
dt z z
where e(t) has the same properties as E(t) and
(6.13) D¢ = ¢(0)- q¢( -r1)- f.l¢( -r2) + qf.l¢( -r1- r2).

If f.l -=f. 0, then the operator contains three delays that are not inde-
pendent. The parameters r 1 and r 2 are physical parameters and we should
have stability of D subject to small changes in these parameters. For a= q,
b = f.l, and c = -ab, Inequalities (6.11) are equivalent to lql < 1, lf.ll < 1
and these conditions are always satisfied if R -=f. 0 and R 1 -=f. 0. Thus, the
operator D is stable locally in the delays. For this particular example, it is
also easy to observe this fact directly from the characteristic equation since
it factors into two simple factors.
9. 7 Properties of equations with stable D operators
In this section, we consider the special class of NFDE(D, f) for which D

is stable. Also, we assume f : [} ~ 1Rn, [} ~ C is open, that is, f is
independent of t. This latter hypothesis simplifies the notation, but some
of the results hold when f depends on t. The reader may supply the details
for this case or consult the references in Section 9.9.
For an autonomous NFDE(D, f), the concepts of positive orbit 'Y+(¢),
w-limit set w(¢), a-limit set a(¢), and invariant set, are defined as for
RFDE.
Theorem 7.1. If[}~ C is open, D is stable and f: [} ~ 1Rn is completely

continuous, then a positive orbit "(+(¢) of the NFDE(D, f) is relatively
compact if and only if 'Y+ (¢) is bounded. If 'Y+ (¢) is bounded, then w( ¢) is
a nonempty, compact, connected, invariant set.
Proof. The latter part of this theorem is proved in essentially the same
manner as for the RFDE and is thus omitted.
It remains to prove the first part. Obviously, 'Y+ (¢) relatively compact
implies 'Y+ (¢) is bounded. Conversely, if 'Y+ (¢) is bounded, there is a con-
stant M such that lf('Y+(¢))1 :SM. Also, for any T;::: 0, t;::: 0, the solution
Xt through ¢ satisfies
D(xt+r - Xt) = i.
t
t+r
f(xs)ds.
From Part (iii) of Theorem 3.4 and, in particular, Relation (3.22), D stable
implies
Since X 7 is continuous in r, this implies x(¢)(t) is uniformly continuous on

[-r,oo). Since x(¢)(t) is bounded, this implies that 'Y+(¢) is precompact
and the theorem is proved. D
For an RFDE(f), any solution defined on an interval of length (k+ l)r,

k;::: 1, say [a- r, a+ kr] must have k derivatives on [a+ (k -l)r, a+ kr] if
f E ck-l. The solution operator for an RFDE(f) continues to smooth the
data as time increases. If a solution of an RFDE(f) is defined on ( -oo, oo),
then obviously this solution is Ck+ 1 iff is Ck. For an NFDE(D, f), the
solution does not smooth on finite intervals. However, if the operator D is
stable, the solution operator does tend to smooth on infinite intervals in
the following sense.
Theorem 7.2. Suppose [} ~ C is open. For any NFDE(D, f) with D

stable and f having continuous, bounded derivatives on [} through order
k ;::: 0, any bounded solution on ( -oo, a] must be Ck+ 1 . Also, the w-limit
9. 7 Properties of equations with stable D operators 291
set of a bounded orbit consists of uniformly bounded, equicontinuous Ck+l

functions.
Iff is analytic, we can also obtain the following result (see the refer-
ences in Section 9.9 for a proof).
Theorem 7.3. Suppose Q ~ C is open. For any NFDE(D, f) with D stable

and f analytic in D, any bounded solution on ( -oo, a] must be analytic.
Also, thew-limit of any bounded orbit is analytic.
As a consequence of Theorems 7.2 and 7.3, we have the following im-

portant remark.
Corollary 7.1. If Q ~ C is open, D is stable, and f : Q---? lRn has continu-

ous, bounded derivatives of order k 2: 0 on Q (resp. analytic in D), then any
periodic solution of the NFDE(D, f) has continuous, bounded derivatives of
order k + 1 (resp. analytic).
Corollary 7.1 is very important in the applications. For example, con-

sider the autonomous equation
(7.1)
where D is stable and suppose there exists a nonconstant w-periodic so-

lution p of this equation. Then p is continuously differentiable. If f also
has a continuous derivative then d2 Dptf dt 2 exists and is equal to dDptf dt.
Therefore,
(7.2)
has the nontrivial w-periodic solution p. This remark is fundamental to the

discussion of the behavior of solutions of Equation (7.1) near a periodic
orbit (see Chapter 10).
To prove this theorem, we make use of the following lemma.
Lemma 7.1. If D is stable, h : ( -oo, a] ---? lRn is uniformly continuous and

bounded, and if x is a bounded solution of
Dxt = h(t)
on ( -oo, a], then x is bounded and uniformly continuous on ( -oo, a]. If h
is also bounded and uniformly continuous on ( -oo, a], then x(t) exists and
is uniformly continuous and bounded on ( -oo, a].
Proof. Suppose 7 2: 0. Since D(xt+r - Xt) = h(t + 7) - h(t) for all t E
( -oo, a - 7], Relation (3.22) implies
!xt+r- xd S be-a(t-u)!xu+r- Xu!+ b sup !h(u + r)- h(u)!

u$u$t
Therefore, letting u ---+ -oo, we have
!xt+r- Xt! S b sup !h(u + r)- h(u)j.

-oo<u$t
Thus, xis uniformly continuous on ( -oo, a]. Also, using the same argument,
for any T > 0, s > 0,
Xt+r - Xt - Xt+s - Xt I :::; b sup Ih(U + T) - h(U) - --'-----'----'--'-

h(U + S) - h( S) I
I-'------
r s uE{-oo,t] T S
= b sup I [ 1 [h(u + vr)- h(u + vs)]dvl.

uE{ -oo,t] Jo
Since his uniformly continuous, this implies [x(t+r) -x(t)]!r approaches a
continuous limit x(t) as r---+ 0. Obviously, D(xt) = h(t) and, as before, xis
bounded and uniformly continuous on ( -oo, a]. This proves the lemma. 0
Proof of Theorem 7.2. Suppose k = 0. If h(t) = Dxt, then x bounded and

continuous on (-oo,a] implies his bounded and continuous on (-oo,a].
Also, dDxt/dt = f(xt) implies h is bounded and continuous on ( -oo, a].
Thus, his uniformly continuous and Lemma 6.1 implies xis bounded and
uniformly continuous. Thus f(xt) = h(t) is bounded and uniformly contin-
uous on ( -oo, a]. Lemma 6.1 implies x exists, is bounded and uniformly
continuous on ( -oo, a]. This proves the theorem for k = 0. The proof for
the case k > 0 is left for the reader. 0
For the NFDE(D, f) with D stable, we know that TD,J(t) is an a-
contracting semigroup (see Theorem 7.3 of Section 3.7 and Theorem 4.5).
As a consequence of the results in Chapter 4, if {TD,J(t)B, t ~ 0} is
bounded for B bounded and TD,J(t) is point dissipative, then there exists
a compact global attractor AD,!· The functions in AD,! must satisfy the
regularity properties stated in Theorem 7.2 and 7.3. In particular, iff is
analytic, then TD,J(t) is one-to-one on the attractor AD,!· Since AD,f is
compact, this implies that TD,J(t) is a group on AD,f·
9.8 Stability theory
In this section, we indicate the changes that are sufficient to adapt the
methods of Liapunov and Razumikhin for RFDE to an NFDE(D, f) with
a stable D operator. The functional D¢ = ¢(0) played a very important
role in all of the results of Chapter 5 on RFDE. For an arbitrary stable D
9.8 Stability theory 293
operator, it turns out that one can obtain similar results by letting D¢ play
the role of ¢(0).
If V : 1R x C __, IRn is continuous and x(u, ¢) is the solution of an
NFDE(D,f) through (u,¢) we define
. 1
V(u, ¢)=lim sup -h [V(t + h,xt+h(u, ¢))- V(u, ¢)].
h->O+
Theorem 8.1. SupposeD is stable, f: 1R x C __,JRn, f: 1R x (bounded sets

of C) into bounded sets of IRn and suppose u( s), v( s), and w( s) are con-
tinuous, nonnegative, and nondecreasing with u(s),v(s) > 0 for s =f. 0, and
u(O) = v(O) = 0. If there is a continuous function V : 1R x C __, IRn such
that
u(ID¢1):::; V(t, ¢):::; v(l¢1)
V(t, ¢) :::; -w(ID¢1)
then the solution x = 0 of the NFDE( D, f) is uniformly stable. If u( s) __, oo
as s __, oo, the solutions of the NFDE(D, f) are uniformly bounded. If
w(s) > 0 for s > 0, then the solution x = 0 of the NFDE(D, f) is uniformly
asymptotically stable. The same conclusion holds if the upper bound on
V(t, ¢) is given by -w(l¢(0)1).
Proof. The proof is omitted since the basic ideas are contained in the proof
of Theorem 2.1 of Section 5.2 and that proof can be appropriately modified
if one uses Property (3.22) of stable D operators. 0
For autonomous equations, one can also generalize the results of Sec-
tion 5.3.
Definition 8.1. We say V: C __, IRn is a Liapunov functional on a set Gin

C for an autonomous NFDE(D, f) if V is continuous on G, the closure of
G, and V :::; 0 on G. Let
S={¢EG:V(¢)=0}
M = largest set in S that is invariant with respect to the
NFDE(D,f).
Theorem 8.2. If D is stable and V is a Liapunov functional on G for the

autonomous NFDE(D, f) and"!+(¢) <:;;; G and either Xt(¢) or Dxt(¢) is
bounded fort 2: 0, then Xt(¢) __, M as t __, oo. If
G = Uz ~f {¢ E C: V(¢) < l}
and there is a constant K = K(l) such that¢ in Uz implies either 1¢(0)1 < K
or ID¢1 < K, then any solution Xt(¢) of the NFDE(D, f) with ¢ E Uz
approaches M as t __, oo.
Proof. The proof is essentially the same as the proofs of Theorems 3.1 and
3.2 of Section 5.3, taking into account Inequality (3.22) and Theorem 7.1.
D
As a first example, consider the scalar equation

d
(8.1) dt [x(t)- cx(t- r)] + ax(t) = 0
where lei < 1 and a> 0. The operator D¢ = ¢(0)- c¢( -r) is stable. If
V(¢) = (D¢) 2 + ac2 /_: ¢ 2(0) d(),
then
V(¢) = -a(D¢) 2 - a(l- c2 )¢2(0):::; -a(D¢) 2.
Theorem 7.1 implies uniform asymptotic stability.
Since the solution operator for this equation is an a-contraction and
therefore has the radius of the essential spectrum less than one for t ;::: 0,
the approach to zero is exponential. Therefore, this proof, using Liapunov
functions, implies there is a 8 such that all solutions of the equation .\(1 -
ce-r>.) +a= 0 have Re.\:::; -8 < 0.
As remarked in the introduction, the following equation arises in the
theory of transmission lines:
d
(8.2) dt Dxt = -ax(t)- qax(t- r)- g(Dxt)
where D¢ = ¢(0)- q¢( -r), JqJ < 1, a > 0. The operator D is stable. Let
us use Theorem 7.1 to prove the following result.
Theorem 8.3. The zero solution of Equation (7.2) is uniformly asymptoti-

cally stable and every solution approaches zero if JqJ < 1, g(O) = 0, and
m ~f inf g(x) > -a 1 -JqJ.

X 1 + JqJ
Proof. Let 'Y be such that m > 'Y > -a(1 -JqJ)/(1 + Jqi). If (3 = JqJa,
V(¢) = ~(D¢) 2 + (3 j_: ¢ 2 (0) dB.
Then
V(¢) = (D¢)[-a¢(0)- qa¢( -r)- g(D¢)] + (3[¢2(0)- ¢ 2 ( -r)J
= -"((D¢) 2 - (D¢) 2 [gcg:) -"(] + (D¢)[-a¢(0) -qa¢(-r)]
+ (3¢2(0)- (3¢2( -r)
:::; -"((D¢) 2 + (D¢)[-a¢(0)- qa¢(-r)] + (3¢ 2(0)- (3¢ 2( -r).
9.8 Stability theory 295
Expanding the expression on the right-hand side of the last inequality, we

have
It is easy to check that the right-hand side of this expression is a negative

definite quadratic form in ¢(0), ¢( -r). Therefore, there is a positive con-
stant k such that V(¢) ~ -k¢ 2 (0). Theorem 7.1 implies the stated result
on stability. D
In some problems, it is difficult to construct Liapunov functions and

one needs an analogue of the Razumikhin-type theorems given in Section
5.4. Before stating such a result for neutral equations, some additional
notation is needed.
Suppose D is a stable operator, IIDII = K. Let 0 ~ u(s) ~ v(s),
s ;::: 0, be continuous, nondecreasing functions, u(s) --+ oo as s --+ oo, and
suppose there is a continuous function o:(ry), 17 ;::: 0, satisfying v(Kry) ~
u(o:(ry)). Let !3(17) > b(ry + o:(ry)) be a continuous function where b > 0 is
defined in Inequality (3.22). Finally, let F : [0, oo) --+ m,+ be a continuous
nondecreasing function such that F(v(Kry)) > v({3(ry)) for 17 > 0. Under
these conditions, we can state the following result.
Theorem 8.4. Suppose the preceding notation, D is stable, f : ffi x C--+ m,n
is continuous, and takes ffi x (bounded sets of C) into bounded sets of m,n
and consider the NFDE(D, f). If there is a continuous, positive function
w(s), s ;::: 0 and a continuous function V : ffin --+ ffi such that u(jxi) ~
V(x) ~ v(ixi) for all x E ffin and
V(D¢) ~ -w(ID¢1)
for all functions¢ satisfying F(V(D¢)) ;::: V(¢(0)), -r ~ fJ ~ 0, then the
solution x = 0 of the NFDE(D, f) is uniformly asymptotically stable and
all solutions approach zero at t --+ oo.
Proof. The proof follows along the lines of the proof of Theorem 4.2 of
Section 5.4 using properties of stableD operators given in Inequality (3.22).
The reader may consult the references in Section 9.9 for the complete proof.
D
Let us apply Theorem 8.4 to the shunted transmission line, Equation

(6.12), with JL = 0 and E = 0; that is, the equation
d
dt Dxt = -g(x(t))
(8.3)
D¢ = ¢(0)-q¢( -r), jqj < 1,
and g continuous. Choose V(x) = x 2 , u(s) = v(s) = s 2 , and F(s) = N 2 s,
where
N > (1-lql)-1, o:(77) = (1 + lql)77, (3( ) = 1 + lql

77 1 - lql 77·
One can show that these functions satisfy the requirements of Theorem 8.4.
Also,
(8.4) V(D¢) = -2(D¢)g(¢(0)).
We need to impose conditions that will imply V(D¢)::; -w(ID¢1) for

all¢ satisfying F(V(D¢)) ::; V(¢(8)), -r ::; 8 ::; 0. Suppose lql < 1/2. If
¢satisfies this latter inequality, then, in particular, 1¢( -r)l ::; NID¢1 and,
thus,
¢(0) = 1 + q ¢( -r) > 1 _ lqll¢( -r)l > 1 _I IN> 1- 2lql > 0

D¢ D¢ - ID(¢)1 - q 1 -lql .
If lg(x)l ----+ oo as lxl ----+ oo and xg(x) > 0 for x -1- 0, define E = 1-lqiN and
w1(s) = { ming(u), ES::; u::; s(1-lql)-1, s 2:0,

-maxg(u), s(1-lql)- 1 ::; u::; Es, s < 0.
If ¢(0) 2: ED¢ and D¢ 2: 0, then g(¢(0)) 2: w 1 (D¢) and V(D¢) ::;

-D¢w 1 (D¢). If ¢(0) 2: ED¢ and D¢::; 0, then g(¢(0)) ::; -w 1(D¢) and
V(D¢)::; D¢w1(D¢). If w(s) = mins;:>:o(sw 1(s),sw 1(-s)), then w(s) > 0
for s > 0 and V(D¢) ::; -w(ID¢1) for all ¢satisfying ¢(0) 2: ED¢ and,
therefore, for all ¢ satisfying F(V(D¢)) 2: V(¢(8)), -r 2: 8 ::; 0. These
computations have proved the following result:
Theorem 8.5. If lql < 1/2, xg(x) > 0 for x -1- 0, lg(x)l ----+ oo as lxl ----+ oo,
then the solution x = 0 of Equation (8.3) is uniformly asymptotically stable
and every solution approaches zero as t----+ oo.
Early work on the properties of NFDE (1.1) are due to Bellman and Cooke
[1], Cruz and Hale [3], Hale [9,15], Hale and Meyer [1] and Henry [1,4].
The approach in this chapter is new; it uses the theory of resolvents
and makes the approach very similar to the retarded case. The approach
generalizes earlier work and allows the kernel p to have a singular part as
well, although we still need a condition on p. (See also Kappel and Zhang
[1].) The Fredholm alternative for periodic systems has been given by Hale
[17] and Nosov [2].
The abstract theorem on hyperbolic semigroups is from Kaashoek and
Verduyn Lunel [2]. If B is a Hilbert space, then conditions (i) and (ii) in
Theorem 4.1 are automatically satisfied. For details and a proof of Theorem
4.2 see Kaashoek and Verduyn Lunel [2]. Theorem 4.2 is more general than
the original result by Henry [1]. See also Greiner-Schwarz [1]. The results
in Section 4 can also be formulated as a spectral mapping theorem. If
D¢ = ¢(0) - L:%"= 1 Ak¢(rk), then
etcr(Av) ~ O"(Tn(t)) ~ etA+ilRU{o}
where A = {ReA : det [I - De·>. I]}. (See Greiner-Schwarz [1] and Henry
[1].) In Henry [7], this result has been refined and it has been shown that
etcr(Av) \ {0} = O"(Tn(t)) \ {0}

for almost all t ~ 0.
Neves, Ribeiro and Lopez [1] study mixed initial-value problems for
hyperbolic partial differential equations in one space dimension that gen-
eralize the NFDE (2.1) studied in this chapter. They obtain similar results
and are able to characterize the growth bound (see also Kaashoek and Ver-
duyn Lunel [2]).
Lemma 3.1 and Theorems 3.4 and 4.4 on difference equations are valid
for equations D(t, Yt) = h(t). One must, of course, impose conditions that
are uniform in t in order to obtain the estimates. For example, the analogue
of Theorem 4.4 would assume a> an, where
an= inf{a: there is a K such that IIT(t,O")II ~ Kea(t-cr),t 2': O" 2': 0}
where T(t,O") is the solution operator of D(t,yt) = 0. There are more in-
teresting problems in the theory of linear systems. For example, precise
conditions for convergence of the spectral projections P>.¢, A E O"(A), to
the state ¢ when A is the generator associated with System (2.1). The
topology on the state space becomes important, and for precise results we
refer to Verduyn Lunel [7].
Some results on perturbed linear systems are also contained in Bellman
and Cooke [1] and Nosov [3, 4], but they generally involve more hypotheses
than would be required by using ideas similar to the ones in Section 9.5.
Stable D operators in connection with NFDE were introduced in Cruz
and Hale [2]. Cruz and Hale proved the equivalence of (i) and (ii) in Theorem
3.5 and gave more information on the asymptotic behavior of solutions of
nonhomogeneous difference equations (3.19). This paper also considers the
nonautonomous D operators. The equivalence of (i) and (iv) in Theorem
3.5 is due to Henry [2].
The method of proof of Theorem 5.2 is due to Hale and Martinez-
Amores [1] (see also Hale and Ize [1]).
Henry [4], Melvin [4], and Moreno [1] were the first to observe that
small changes in the delays could drastically change the stability properties
of a simple difference equation, and Corollary 6.1 is an immediate conse-
quence of their work. Hale [13] showed that stability locally in the delays
implies stability globally in the delays. Theorem 6.1 is contained in the

thesis of Silkowski [1]. For the Routh-Hurwitz criteria used in the example
after Corollary 6.1, see Coppel [1] or Gantmacher [1]. One also can discuss
the analogue of Theorem 6.1 for the preservation with respect to delays
of hyperbolicity of the origin. Necessary and sufficient conditions for this
property and many examples are given in Avellar and Hale [1].
If the delays in (6.1) are not allowed to vary independently, then it is
possible to obtain results with an appropriate modification of Theorem 6.1.
In such situations, the region of strong stability in specific examples will be
much larger. To formulate the results, let r = (r 1 , ... , rM) E (IR+)M, rk =
(rkl, ... '/kM ), /kj integers, /k =!= 0, rk. r = L~Irkjrj, k = 1, 2, ... 'N,
and consider the difference equation
N
(9.1) D(r, /, A)yt ~f y(t)- L Aky(t -!k · r) = 0,

k=l
where each Ak is ann x n constant matrix. Theorem 6.1 remains valid with
Bk replaced by rk · 8 with 8 E IRM.
It also is of interest to consider stability globally in the delays for
NFDE of the form
d N N
(9.2) dt [x(t)- L
Akx(t -!k · r)] = B 0 x(t) +L Bkx(t -!k · r),
k=l k=l
where the notation is as earlier and each Bk is an n x n constant matrix.

For the retarded case (all matrices Ak = 0), such problems have been
considered by Koval and Carkov [1], Rep in [2], Cooke and Ferreira [1],
Hale, Infante, and Tsen [1]. For NFDE, Zivotovskii [1] has considered the
scalar equation with independent delays and Datko [3], Hale, Infante and
Tsen have considered the general equation (9.2). For Sj E <C, j = i, ... M,
if we define
P(>., /, s 1 , ... , SM, A, B)
N N
= det [>.-(I- L Aksikt · s:z;M)- 1(L Bksikt · s:z;M)],
k=l k=l
then Hale, Infante, and Tsen [1] prove the following result.
Theorem 9.1. The NFDE (9.2) is stable globally in the delays if and only if
the following conditions hold:
(i) Equation (9.1) is stable globally in the delays,
(ii) P(iy,/,s 1 , ... ,sM,A,B) =/= 0, for ally E IR,y =/= 0, isJI = 1,j
1, ... ,M,
(iii) Recr[(I- I:;~=l Ak)- 1 I:;~=l Bk] < 0, where cr denotes spectrum.
If we specialize this result to the scalar equation with independent

delays
d N N
(9.3) dt [x(t)- L akx(t- rk)] = box(t) + L bkx(t- rk),
k=l k=l
then Equation (9.3) is stable globally in the delays if and only if
N N
(9.4) L lak I < 1, bo < 0, L lbkl:::; lbol·
k=l k=l
For an equation with dependent delays,
d
L akx(t- 'Yk · r)] = +L
N N
(9.5) dt [x(t)- b0 x(t) bkx(t- 'Yk · r),
k=l k=l
the stability criteria are somewhat more complicated to state, but are de-
termined from the properties of the solutions of the difference equations
N N
(9.6) y(t)- L aky(t- 'Yk · r)] = 0, boz(t) +L bkz(t- '/'k · r) = 0.

k=l k=l
In fact, if we define
N
L bk sin 'Yk · B,
N
(9.7) a(B, b)= bo +L bk cos'Yk · B, (3(0, b) =
k=l k=l
then Equation (9.6) is stable globally in the delays if and only if the fol-
lowing conditions hold:
N
2::::#0,
(9.8) k=l k=O
either a(B,a) # 0 or a(B,a) = 0, (3(B,a) = 0.
Other results and examples can be found in Hale, Infante, and Tsen [1].
It also is of interest to investigate NFDE when the delays depend on
time. Of course, it will be necessary to understand well the corresponding
difference equation when the delays depend on time. Very little information
of a general nature is known, and the following observations are made to
illustrate some of the difficulties.
Suppose that 'Yk ·r = rk, k = 1, ... , M, and that (9.1) is stable globally
in the delays. Will the zero solution of the difference equation,
N
(9.9) y(t)- L Aky(t- 'Yk · r(t)) = 0,
k=l
be asymptotically stable if the function r(t) is continuous and bounded

on 1R? If y is a scalar, this is true as a consequence of the inequalities
l::~=l IAkl < 1. If y is a vector, an example has been given by Sigueira
Marconato and Avellar [1] with r(t) periodic for which a solution of Equa-
tion (9.9) is unbounded on [0, oo). In fact, the system
2 2
Yl(t) = 3Y1(t- r1(t)) + 3Y2(t- r2(t))
2 2
Y2(t) = -3y1(t- r2(t))- 3Y2(t- r1(t))
is stable globally in the delays if the delays are constant in time and there
is an unbounded solution if r1 (t), r2(t) are continuous periodic functions of
period 3 such that r 1(0) = 1,r1(1) = 3,r1(2) = 3, and r 2(0) = 2,r2(1) =
1,r2(2) = 1, and t - r 1 (t) ~ -2 fort~ 0, j = 1,2. For continuous initial
data 'ljJ: [-2,0]---+ JR2 with 'lj;(-1) = (1,0), 'lj;(-2) = (0, 1), it is possible to
show that the solution y(t) through (0,'1/J) is unbounded on [O,oo).
Theorem 7.1 is due to Cruz and Hale [2]. A special case of Theorem
7.2 for k = 0 was proved by Hale [14] and the result as stated is due to
Lopes [5]. Theorem 7.3 is due to Hale and Scheurle [1]. For more results on
attractors for NFDE with stable D operators and the existence of periodic
solutions of systems periodic in time, see Hale [23].
Theorems 8.1 and 8.2 are due to Cruz and Hale [2] (see also Chary
[1] and Minsk [1,2]). Example (8.2) is due to Lopes [1]. As remarked in
the introduction, one can derive neutral equations from the transmission
line problem in different ways. Slemrod [2] used the other form of Problem
(8.2) and employed Liapunov functionals to obtain sufficient conditions for
stability. Theorem 8.4 is due to Lopes [6]. Example 8.3 is due to Lopes [2].
Lopes [1,2,6] has also generalized these results to obtain uniform ultimate
boundedness of solutions of nonautonomous equations. If the equations are
also w-periodic, he has applied the results of Chapter 4 to obtain the exis-
tence of w-periodic solutions.
Infante and Slemrod [1] have used Liapunov functionals and Theo-
rem 8.1 to obtain sufficient conditions on the coefficients in linear au-
tonomous neutral differential difference equations, which will ensure the
uniform asymptotic stability of the zero solution. For symmetric systems,
Brayton and Willoughby [1] obtained sufficient conditions for the stability
of such systems directly from the characteristic equation.
Liapunov functionals also have been used for NFDE where the deriva-
tive occurs explicitly and the space of initial data involves the function and
its derivative. The reader may consult the volumes of Trudy Sem. Teorii
Diff. Urav Otkl. Argumenton from the People's Friendship University in
Moscow for references.
The relationship between the different types of stability of linear sys-
tems of NFDE is not as simple as for RFDE (for a general discussion, see
Hale [18]). Even for autonomous equations, many surprising results occur.
It is easy to show that stability implies uniform stability. On the other

hand, one can have asymptotic stability and not have uniform asymptotic
stability. This follows because uniform asymptotic stability is equivalent
to exponential asymptotic stability. On the other hand, one can have all
eigenvalues AJ of the linear system in the left half-plane with Re AJ ---+ 0
as j ---+ oo and all solutions approaching zero as t ---+ oo. Such a system
cannot be uniformly asymptotically stable. One can also have all eigen-
values ).. with Re).. < 0 and have some solutions unbounded (see Brumley
[1]). An even more striking example was given by Gromova and Zverkin [1].
They gave an example of a linear neutral differential difference equation
with all eigenvalues simple and on the imaginary axis and yet the equation
has unbounded solutions. If the operator D is stable, one obtains the same
relationship between the concepts of stability for linear autonomous and
periodic equations as for RFDE.
Ize [2] and Ize and de Molfetta [1] have given very general results
of the asymptotic behavior of linear equations that are nonautonomous
perturbations of autonomous systems.
For the invariance principle using Razumikhin-type Liapunov functions
on IRn, see Haddock, Krisztin, Jerj'ecki, and Wu [1].
In certain problems concerning control systems containing gas, steam,
or water pipes (see Kobyakov [1] and Solodovnikov [1]), one encounters lin-
ear hyperbolic partial differential equations with mixed initial and deriva-
tive boundary conditions. The same is true in loss-less transmission lines
(see Brayton [2]). Using the process described in the introduction, these
problems are equivalent to a system of equations of the following form:
x(t)= Ax(t) + By(t- r) + f(x(t), y(t), y(t- r))
y(t)- Ex(t)- Jy(t- r)- g(x(t), y(t), y(t- r)) = 0
where A, B, E and J are matrices, f and g are given functions and r > 0.
Brayton [2] discussed the linear version of this equation by means of the
Laplace transform. Razvan [1, 2] treated this equation as a special case of
a NFDE by letting y(t) = i(t). For further remarks, see Zverkin [4]. Hale
and Martinez-Amores [1] discussed the equation by writing the equation as
x(t) = Ax(t) + By(t- r) + f
d
dt [y(t)- Ex(t)- Jy(t- r)- g] = 0
and applying the results of this chapter to the set of initial data x(O) = a,
y0 (B) = ¢(8), -r ::; () ::; 0, restricted to the set ¢(0)- Ea- J¢( -r) - g =
0. The resulting theory thus follows in a very natural manner from the
known results on NFDE. Equations of this type occur also in certain models
describing lazer optics (see Chow and Huang [1] for references).
Datko [2] has discussed linear autonomous NFDE in a Banach space
setting and has thus obtained some interesting generalizations of the Hille-
Yoshida theorem.
10
Near equilibrium and periodic orbits
In this section, we consider autonomous FDE of retarded or neutral type

and discuss the behavior of the solutions near equilibrium points and peri-
odic orbits. We concentrate particularly on the existence of stable, unstable,
center-stable, and center-unstable manifolds.
10.1 Hyperbolic equilibrium points
Let f2 be a neighborhood of zero inC and let Cb'(fl,IRn) c CP(Jl,IRn)

be the subset of functions from [2 into IRn that have bounded continuous
derivatives up through order p with respect to¢ E fl. The space Cb(fl, IRn)
becomes a Banach space if the norm is chosen as the supremum norm over
all derivatives up through order p. The norm will be designated by I · lp·
Throughout this chapter, we shall assume that D E .C (C, IRn) is stable and
F E Cl ([2, IRn). Consider the equation
(1.1)
If F(O) = 0, then 0 is an equilibrium point and the linearization about 0 is
d
(1.2) -Dxt = Lxt,
dt
where L E .C( C, IRn), L'ljJ = Dq,F(O)'l/J. We say that 0 is a hyperbolic equi-
librium point of (1.1) if the roots of the characteristic equation
(1.3) det Ll(>-) = 0, Ll(>-) = D(e>.· I)- L(e>.· I),
have nonzero real parts.

If 0 is a hyperbolic equilibrium point of (1.1) and A denotes the set of
roots of (1.3) with positive real part, then the space C can be decomposed
by A as
c = u ffi s.
10.1 Hyperbolic equilibrium points 303
The decomposition of Cas U ffiS defines two projection operators 1ru : C - t

U, 1ruU = U, 7r8: C - t S, 1r3S = S, 7r8 =I -7ru. If eli is a basis for U and lft
is a basis for ur' (lft, eli) = 1, then the projection is given by 7ru¢ = eli(lft, ¢).
Let¢ E C be written as¢= ¢U + ¢ 8 , ¢U E U, ¢ 8 E S. Let K(t, s)(B) =
J; X (t + (} - a) da where X ( · ) denotes the fundamental matrix solution to
the linear equation (1.2). If T(t) is the semigroup generated by the linear
equation (1.2), then U and S are invariant under T(t) and T(t) is defined
on U for all t E ffi. Define X[/= elilft(O). Then K(t, s)u = J;T(t-a)X[/ da
and K(t, s) 8 = 1r3K(t, s) = K(t, s)- eli(lft, K(t, s)).
From Sections 7.9 and 9.5, it follows that there are positive constants
M, a such that for¢ E C
IT(t)¢ul :S Me"tl¢ul, Var(t,O] IT(t)X[/1::; Me"t, t::; 0,

(1.4)
IT(t)¢8 1::; Me-"tl¢8 1, Var[o,t) K(t, · )8 ::; Me-"t, t ~ 0.
Relations (1.4) and the fact that D and L are linear imply that the origin
of System (1.2) is a saddle point with the orbits inC behaving as shown in
Figure 10.1.
Fig.lO.l.
The set U is uniquely characterized as the set of initial values of those

solutions of Equation (1.2) that exist and remain bounded for t ::; 0. Re-
lations (1.4) imply that these solutions approach zero exponentially as
t - t -oo. The set S is characterized as the set of initial values of those
solutions of Equation (1.2) that exist and remain bounded fort ~ 0. These
solutions approach zero exponentially as t -too.
It is natural to ask if the solutions of Equation (1.1) have the same
qualitative behavior near x = 0 as the solutions of Equation (1.2). Of
course, the meaning of qualitative behavior must be defined very carefully.
It is tempting to say that Equation (1.1) has the same qualitative behavior
as Equation (1.2) near x = 0 if the orbits of Equation (1.1) can be mapped
homeomorphically onto the orbits of (1.2). The following example suggests
that such a definition is too strong.
304 10. Near equilibrium and periodic orbits
Consider the scalar retarded equation
y(t) = -y(t- 1)[1 + y(t)].

The constant function y(t) = -1, t E IR, is an equilibrium point. If we let
x(t) = y(t) + 1, then x satisfies the equation
x(t) = x(t)- x(t)x(t- 1),

which is a special case of (1.1) with
D¢ = ¢(0), F(¢) = ¢(0)- ¢(0)¢( -1), r = 1.
The linear equation (1.2) is x(t) = x(t), which is a special case of (1.4) with
D¢ = Lcf> = ¢(0). This ordinary differential equation must be considered
as a FDE in C. It is easy to verify that the sets U and S are given by
U = { ¢: ¢(0) = e 0 ¢(0), 0 E [-1, 0]}

S = { cf>: ¢(0) = '¢(0)- e 0 '¢(0), 0 E [-1, OJ, 'lj; E C}.
If¢ E S, then T(t)¢ = 0 for t 2: 1. Therefore, each orbit in C must lie on

U for t ;::: 1. The semigroup for the linear equation is not one-to-one. For
the particular perturbation -¢(0)¢( -1), the semigroup generated by the
nonlinear equation also is not one-to-one. In fact, any initial function¢ with
¢(0) = 0 has the property that the solution x(t, ¢) = 0 for all t 2: 0. On the
other hand, if we were to consider an arbitrary higher-order perturbation
of the linear equation, it is not unreasonable to expect that the semigroup
generated by the nonlinear equation is one-to-one. As a consequence, we
cannot expect the orbits of the two systems to be homeomorphic.
On the other hand, we can show that some of the important properties
of the trajectories are preserved. More specifically, we show that the set of
initial values of those solutions of Equation (1.1) that exist and remain in
a 8-neighborhood of x = 0 for t :::; 0 is diffeomorphic to a neighborhood
in U of zero and these solutions approach zero exponentially as t --> -oo.
The same result is proved fort 2: 0 and S. Any other solution must leave a
neighborhood of zero with increasing t and, if it exists fort :::; -r, it must
also leave a neighborhood of zero with decreasing t.
Let us now be more precise. Let x(t, ¢) be the solution of (1.1) with
initial value ¢ at t = 0. We define the stable set and unstable set of the
equilibrium point 0 of (1.1) as, respectively,
W 8 (0) = {cf> E C: Xt(-,cf>) -t 0 as t -too}

(1.5)
Wu(O) = { cf> E C: Xt( ·, cf>) --> 0 as t--> -oo }.
For a given neighborhood V of 0, we also can define the local stable and
local unstable sets
W1~c(O) ~f W 8 (0, V) = { ¢ E W 8 (0) : Xt( ·, ¢) E V fort 2 0}

(1.6)
W1~c(O) ~f Wu(O, V) = { ¢ E Wu(O) : Xt( ·, ¢) E V fort :S: 0 }.
We say that wu(o, V) is a Lipschitz graph (resp. Ck-graph) over nuC if

there is a neighborhood V of 0 in nuC and a Lipschitz continuous function
(resp. Ck-function) g such that
wu(o, V) = { '¢' E c: '¢' = g(¢), ¢ E V}.

The set wu(o, V) is said to be tangent to nuC at 0 if lns'¢11/lnu'¢11 --t 0 as
'¢' --t 0 in wu(o, V). Similar definitions hold for W 8 (0, V).
A basic result on stable and unstable sets of equilibrium points is the
following.
Theorem 1.1. If 0 is a hyperbolic equilibrium point of Equation (1.1), F E

ct(.n,IRn), and Dis stable, then there is a neighborhood V ofO inC such
that wu(o, V) (resp. W 8 (0, V)) is a Ck-graph over nuC (resp. nsC) that
is tangent to nuC (resp. nsC) at 0.
To prove this result, we let f(¢) = F(¢)- L¢, and rewrite Equation
(1.1) as
(1. 7)
The function f has the property that f(O) = 0, Dq,f(O) = 0. We need the
following result.
Lemma 1.1. If x( t, ¢) is a solution of (1. 7) that is defined and bounded for

t :=;; 0, then Xt ( ·, ¢) is a solution of the following integral equation in C:
y(t) = T(t)¢u+lot T(t- T)Xff f(y(T)) dT

(1.8)
+[too d[K(t, T) Jf(y(T)). 8
If x(t, ¢) is a solution of (1.7) that is defined and bounded fort 2 0, then

Xt( ·, ¢) is a solution of the following integral equation inC:
y(t) = T(t)¢ 8 +lot d[K(t,T) ]f(y(T)) 8
(1.9)
-1 00
T(t- T)Xff f(y(T)) dT.
Conversely, if x(t, ¢) is a solution of (1.8) (resp. (1.9)) that is defined and

bounded fort:=;; 0 (resp. t 2 0), then Xt( ·, ¢) is a solution of (1.7).
Proof. If y(t) ~f Xt( ·, ¢), t::; 0, is a bounded solution of (1.7), then for any
t E ( -oo, 0],
wsy(t) = T(t- f)wsy(f) + 1t d[wsK(t, T)]f(y(T)).
If we let t---+ -oo and use Relations (1.4), we deduce that
Since
wuy(t) = T(t)cpu +fat T(t- T)X!/ f(y(T)) dT,
we see that y(t) must satisfy (1.8). The prooffor the case when cjJ E W 8 (0)
is similar and therefore omitted.
The converse statement is proved by direct computation. D
We now begin the proof of Theorem 1.1, considering first the case
where the function is assumed only to be Lipschitz continuous. In this
case, we prove that the stable and unstable sets are Lipschitz graphs. More
specifically, suppose that there is a continuous function TJ : [0, oo) ---+ [0, oo ),
TJ(O) = 0, and let us consider those functions fin (1.7) that satisfy
f(O) = 0
( 1.10)
If(¢)- !(1/J)I::; TJ(u)l¢- 1/JI if 1¢1, 11/JI ::; u.
With the constants K, K 1 , a as defined earlier, choose 15 > 0 so that
8KK1 TJ(15) < a, 8K 2 KrTJ(I5) < a. For cjJ E wuC with 1¢1 ::; I5/2K, define
S( ¢, 15) as the set of continuous functions y : ( -oo, 0] ---> C such that IYI =
sup_oo<t<O IY( t) I ::; 15 and wuy(O) = ¢. The set S( ¢, 15) is a closed bounded
subset ofthe Banach space of all continuous functions taking ( -oo, OJ into
C with the uniform topology. For any y E S (¢, 15), define
(1.11)
(Ty)(t) = T(t)cp + lt T(t- T)X!/ j(y(T)) dT
+ j_too d[K(t, T) 8 ]f(y(T)).

From the estimates (1.4), we see that
ITy(t)l::; Ke"'tl¢1 + KKlTJ(I5) 1° ea(t-T)IY(T)I dT
+ KKlTJ(I5) j_too e-a(t-T)IY(T)I dT
::; l5( ~ + 2K; TJ( 15)) ::; ~l5.

Thus, T: S(¢,15)--+ S(¢,15). In the same way, we observe that Tis a con-
traction mapping with contraction constant 1/4, independent of¢. There-
fore, T has a unique fixed-point y* ( ·, ¢) E S( ¢, 15). This fixed point satisfies
(1.8) and thus is a solution of (1. 7) from Lemma 1.1. Also, y* ( · , ¢) is con-
tinuous in ¢.
For¢ E nuC with 1¢1 ::; 15j2K, we now consider the setS(¢, 15, a) of
continuous functions y : ( -oo, OJ --+ C such that
IYia = sup e-at/ 2 ly(t)1::; 15

-oo<t~O
and nuy(O) =¢.It is clear that S(¢,/5,a) C S(¢,15) since a> 0. From
(1.11), we deduce that
e-at/ 21Ty(t)1 ::; ~15 + KK~1J(I5) (2 + ~)IYia::; 15.

Therefore, T maps S(¢, 15, a) into itself. Since T has a unique fixed point
in S(¢,/5) and S(¢,/5,a) C S(¢,15), we infer that y*( · ,¢) E S(¢,/5,a) and
ly*(t, ¢)1::; 15eat/ 2 fort::; 0.
We now obtain better estimates on y* ( · , ¢). The function y* ( · , c;)
satisfies the integral inequality
ly*(t, ¢)1::; Keatl¢1 + KKI1J(I5) 1°ea(t-T)IY~(T, ¢)1 dT
+ KK11](8) [too e-a(t-T)Iy*(T, ¢)1 dT.
Since ly*( ·, ¢)Ia is bounded, we have
e-at/21y*(t, ¢)1::; Keat/21¢1 + KKI1J(I5) 10 ea(t-T)/2e-aT/21Y(T)I dT
+ KKI1J(I5) [too e-3a(t-T)/2e-aT/21Y(T)I dT
::; Keat/ 21¢1 + KK~1J(I5) (2 + ~)ly*( ·, ¢)Ia

1
::; Kl¢1 + 2ly*( · '¢)Ia·
From this inequality, we obtain IY* ( · , ¢)I ::; 2Kinu¢1, which implies that
(1.12)
This estimate shows that y*(t,O) = 0 and that y*(O,¢) E wu(O).

The same type of computations show that
(1.13) ly*(t, ¢)- y*(t, ?j;)l::; 2Keat/ 21¢- 'lj;l, t::; 0.
In particular, y* ( ·, ¢) is Lipschitzian in ¢.
We next observe that

0
ly*(O,¢) -y*(0,1/J)I2: 1¢-1/JI- [ 00
KK17](8)early*(T,¢) -y*(T,1/!)IdT
2: 1¢-1/!1[1- 4 K 2 ~ 77 ( 8 )]2: ~1¢-1/JI.

Thus, the mapping¢,_. y*(O, ¢)is one-to-one with a continuous inverse.
These estimates, together with the fact that x(t, ¢), t:::; 0, ¢ E wu(O)
must satisfy (1.8), imply that there exists a neighborhood V of 0 inC such
that the conclusions of Theorem 1.1 are valid for wu(o, V) if we replace
the word Ck-graph by Lipschitz graph.
We can repeat the same type of argument to Equation (1.9) to obtain
that ws(o, V) is a Lipschitz graph over 1rsC.
To show that these local sets are C 1-graphs, we make use of the fol-
lowing result, which is of independent interest.
Lemma 1.2. Let X, Y be Banach spaces, Q C X an open set, and g : Q ____, Y

be locally Lipschitzian. Then g is continuously differentiable if and only if
for each¢ E Q,
(1.14) lg(1j; +h)- g(1j;)- g(¢ +h)+ g(¢)1 = o(lhlx)

as (1/!, h)____,(¢, 0).
Proof. It is easy to see that (1.14) holds if g is a C 1-function. Without loss
in generality, we take Q to be a ball and g to be Lipschitzian in Q. If the
derivative g' of g exists at each point of Q and (1.14) is satisfied, then g' is
continuous. Thus, it is enough to prove that g' exists at each point of Q.
Case 1. Let us first suppose that X = Y = lR.. Since g is absolutely
continuous, it is differentiable almost everywhere. For any ¢ E Q, E > 0,
there is a 8 > 0 such that
lg(1/J +h)- g(1j;)- g(c/J +h)+ g(¢)1 :::; Elhl if 11/J- ¢1 + lhl < 8.
There is a 1/J* E ( ¢ - 8, ¢ + 8) such that g' (1/J*) exists. Thus, for h "/= 0
sufficiently small,
Ig(¢ + h~- g(¢) - g'(1j;*)l :::; 2E,
0<{1 . . f}g(¢+h)-g(¢)
l"1mm <4
_ 1msup- h _ E.
h--->0 h--->0
Since E is arbitrary, this implies that g' (¢) exists.

Case 2. Now suppose that X = lR. and Y is a Banach space with Y*
being the dual space. If 77 E Y*, then Case 1 implies that 7]g is a C 1-function
and I(7Jg)'(1/J)I :::; 1771 Lip g. If D(1j;) : 7] ,_. (7Jg)'(1/!), then D(1j;) E Y** is
continuous in 1j; from (1.14).
For ¢> E Q, TJ E Y*, ITJI :'S 1, we have
TJ(g(¢> +h)- g(¢>) - D(c/>)TJ) =.!. {Hh (D('ij;)- D(¢>))TJ-+ 0

h h }q,
ash-+ 0 uniformly for ITJI :'S 1. LetT : Y-+ Y** be the canonical inclusion.
Then T[g(¢> +h)- g(¢>)]/h-+ D(¢>) ash-+ 0. Since Tis an isometry, this
implies that [g(¢ +h)- g(c/>)]/h-+ a limit in Y as h -+ 0; that is, g'(c/>)
exists.
Case 3. Finally, let X, Y be arbitrary Banach spaces. From Case 2, for any
'ljJ E X, h EX, the map t f--+ g('l/J + th) taking 1R toY is C 1 if tis small.
Thus, the Gateaux derivative
d ("'' h)= dg('l/J + th) I exists.

g '~-'' dt t=O
Condition (1.14) implies that dg('lj;, h) -+ dg(¢>, h) in Y as '1/J -+ ¢> E Q,

uniformly for lhl :'S 1. This implies that h f--+ dg('lj;, h) is linear and con-
tinuous. Thus, dg( '1/J, h) is the Frechet derivative at h of g and the proof is
complete. D
We now use Lemma 1.2 to prove that wu(o, V) is a C 1 -graph over

TiuC. Let y* ( ·, ¢>) be the fixed point inS(¢>, 8) of the map Tin (1.11) that
defines wu(o, V). We show that y*( ·,¢>)is C 1 in¢. Define
z*('lj;, ¢>, h)(t) = y*(t, '1/J +h)- y*(t, '1/J)- y*(t, ¢>+h)+ y*(t, ¢>).
From Lemma 1.2, it is sufficient to show that
(1.15) lim sup lhll z*('lj;, ¢>, h)(t) = 0

( ,P,h)--+( ¢,0)
uniformly for -oo < t :'S 0.

From the definition of y*(t, ¢>)and the fact that f is a C 1 -function, we
have, for t :'S 0,
z*('l/J, ¢>, h)(t) =fat T(t- T)X(j Dq,f(y*(T, c/>))z*('lj;, ¢>, h)(T) dT
+[too d[K(t, T) 5 ]Dq,f(y*(T, cp))z*('lj;, cp, h)(T)
+fat T(t- T)X(j[Dq,f(y*(T,'ij;))
- Dq,f(y*(T, ¢))] (y*(T, '1/J +h)- y*(T, '1/J)) dT
+[too d[K(t, T) l[Dq,f(y*(T, '1/J))

5
- Dq,f(y*(T, ¢))] (y*(T, '1/J +h)- y*(T, '1/J))

+ o(h) as lhl-+ 0.
Using the estimates (1.4), (1.10), and (1.13), we have
lz*('lj!,¢,h)(t)1:::; KK1ry(8) 1° ea(t-T)Iz*('lj!,¢,h)(T)IdT
+ K K177( 8) /_too e-a(t-T) lz* (7/!, ¢,h) (T) ldT

8K 2 K 12
+ lhleat/ 2 sup ID¢f(y*(T, 7/!))- D¢f(y*(T, ¢))1
0! T:S::O
+ o(h) as lhl ~ 0.
As in the proof of the existence of wu(o, V), one can show that there
is a positive constant K 1 such that
for t :::; 0. Since the function y* ( · , 7/J) is continuous in 7/J and f is a 0 1 -

function, we have (1.15) and we have proved Theorem 1.1 for wu(o, V) for
k = 1. The same type of argument applies for W 8 (0, V).
The proof that Theorem 1.1 is valid for arbitrary k requires an in-
duction argument going from functions that are ck- 1 with the (k- 1)st-
derivative Lipschitz continuous to functions that are Ck. This will not be
given (see the references in the supplementary remarks).
In Definition (1.5), we defined the global stable and unstable sets for
the equilibrium point 0 and Theorem 1.1 asserts that these sets locally near
0 are Ck-manifolds. If we let T(t) be the semigroup generated by Equation
(1.7) (that is, T(t)¢ = Xt( ·,¢)),then we have the following relations:
u
t~O
(1.16)
wu(o) = T(t)Wu(o, V).
t~O
It is natural to investigate whether or not W 8 (0) and wu(o) are Ck-

manifolds. Without further hypotheses on the vector field f, this is not
the case. Rather surprisingly, this is not true in general even for the finite-
dimensional set wu(o). In fact, consider the retarded delay equation
( 1.17) x(t) = a(x(t))x(t) + (3(x(t))x(t- 1),

where X E IR, a(x) and (3(x) are 0 00 -functions defined as
2e-1 e2 • 1 1< .
(a(x), (3(x)) = { ( e-1 ' - e-1 ), ~f x - 1,
(1, 0), 1f lxl 2:: 2;
and
a(x) + j3(x)e- 1 = 1 when 1:::; lxl :::; 2.
The origin 0 is an equilibrium point of (1.17). Equation (1.17) is linear in a
neighborhood of 0 and has .X 1 = 1 and .X 2 = 2 as the positive characteristic
values. Thus, there is a neighborhood V of 0 such that dim wu(o, V) = 2.
Let x(t) = Eet be a solution of (1.17) initiating in wu(o, V). There is a
t 0 such that inL1::;e::;o lx(to + 0)1 > 2, and, in a neighborhood of Xt 0 , the
Equation (1.17) becomes ±(t) = x(t). If¢ is in a small neighborhood of Xt 0
and the solution through 1/J is defined for negative t, then '¢( 0) = 'T/et+O,
where 'T/ is close to E. Therefore, the unstable set in this neighborhood of
Xt 0 is a smooth manifold but of dimension 1. The local two-dimensional
unstable manifold collapses into a one-dimension manifold as we follow the
manifold along the solutions.
If it is assumed that the map T(t) is one-to-one together with D¢T(t),
then it is possible to show that both W 8 (0) and wu(o) are embedded
submanifolds of C.
We present now an interesting result concerning the manner in which
solutions leave the stable manifold. Suppose that the hypotheses of Theorem
1.1 are satisfied, cp = (¢1, ¢2, ... ,¢d) is a basis for U, and let 1ru¢ = c}jb,
b = b(¢) E IRd. The mapping b: C ~ lRd is continuous and linear and we
take the norm of b to be the Euclidean norm. For any continuous function
V : C ~ IR, we let
V_(¢) = liminf ~[V(xt(¢))- V(¢)],

t--+0+ t
where x(¢) is the solution of Equation (1.1) through¢.
Lemma 1.3. Under the assumptions of Theorem 1.1 and the preceding no-
tation, there is a positive definite quadratic form V(¢) = bT Eb on lRd with
the property that for any constant p > 0, there is a 80 > 0 such that for any
8, 0 < 8:::; 8o, V_(¢) > 0 ifV(¢) ~ p 282, ¢ E C, 1¢1:::; 8.
Proof. Let 1rUXt = c]jy(t), where Xt = Xt(¢) is the solution of Equation (1.1)
with Xo = ¢. From Theorem 9.1 of Section 7.9, there is ad x n constant
matrix C and a d x d constant matrix B with the spectrum of B equal to
the roots of the characteristic equation (1.3) with positive real parts such
that
y(t) = By(t) + Cf(xt),
where f(¢) = F(¢)- L¢. Suppose that Eisa d x d positive definite matrix
satisfying BT E + EB = I and define V(¢) = bT Eb, where 1ru¢ = cpb. If
g(¢) = Cf(¢), then
V_(¢) = bTb+ 2gTEb.
Let /3 2 = min{ bT Eb : lbl = 1} and 'Y = max{ bT Eb : lbl = 1 }. Suppose
that 'T/: [0, oo) ~IRis a continuous nondecreasing function, 'f/(0) = 0, such
that lg(¢)1 :S: 7](8)1¢1 for I<PI :S: 8 and choose 8o > 0 so that 4')'IEI7J(8o) < p8.
Then as long as 1¢1:::; 8, 0 < 8:::; 80 , and V(¢):::: p2 82 , we have
This last lemma holds even if some eigenvalues of Equation (1.3) are
on the imaginary axis. It is easily checked that the same proof is valid.
Remark 1.1. It is possible to prove parameterized versions of all of these

results; that is, we can consider the equation
where ,\ is a parameter in a Banach space, F( ¢, 0) = F( ¢ ), and prove

analogues of Theorem 1.1 and the lemmas for I.XI small. We also can consider
nonautonomous equations
provided that we suppose that the linear variational equation near the origin
has an exponential dichotomy. See the references for the definition of this
term.
10.2 Nonhyperbolic equilibrium points
If 0 is a nonhyperbolic equilibrium point of (1.1), then the space C can be

decomposed as
where U is finite dimensional and corresponds to the span of the general-

ized eigenspaces of the roots of (1.3) with positive real parts, N is finite
dimensional and corresponds to the span of the generalized eigenspaces of
the roots of (1.3) with zero real parts. This decomposition of C defines three
projection operators 1ru: C----> U, 1ruU = U, 1rN: C----> N, IrNN = N, Irs:
C----> S, 1rsS = S, 1ru + 7rN +Irs =I. If is a basis for U and tJt is a basis
for ur, (tJt,<J>) = 1, then the projection Jru is given by Jru¢ = <J>(tJt,¢).
Similarly, the projection 7rN can be given by 7rN¢ = o(tito, ¢), where
<!> 0 is a basis for N and tJt0 is a basis for NT, ( tJto, o) = 1. Let <P E C
10.2 Nonhyperbolic equilibrium points 313
be written as ¢ = cpu + ¢N + ¢ 8 , cpu E U, ¢N E N, ¢ 8 E S. Let

K(t, s)(O) = J; X(t+O-a) da where X(·) denotes the fundamental matrix
solution to the linear equation (1.2).
If T(t) is the semigroup generated by the linear equation (1.2), then
U, N, and S are invariant under T(t), T(t) is defined on U, N for all t E JR.
Define X[/ = <Ptli(O), Xf = <P 0 tli0 (0). Then K(t, s)u = J;
T(t- a)X[/ da,
K(t, s)N = J; T(t - a)Xf da and K(t, s) = 1rsK(t, s) = K(t, s) -
8
<P(tli, K(t, s))- <Po(tlio, K(t, s)).
From Sections 7.9 and 9.5, it follows that there are positive constants
M, a, and, for any E > 0, a positive constant M€, such that for¢ E C
IT(t)¢ul:::; Me"'tl¢ul, Var(t,o] T(t)X[/ :S Me"'t, t :S 0,
(2.1) IT(t)¢NI :S M€e€ltii¢NI, Var(-t,t) T(t)Xf:::; Me€1t1, t E IR,
IT(t)¢81:::; Me-atl¢81, Var[o,t) K(t, · )8 :::; Me-at, t ~ 0.
As in the previous section, the set U is uniquely characterized as the
set of initial values of those solutions of Equation (1.2) that exist, remain
bounded fort :::; 0, and approach zero exponentially as t ~ -oo. The set
S is characterized as the set of initial values of those solutions of Equation
(1.2) that exist, remain bounded fort~ 0, and approach zero exponentially
as t ~ oo. The set N is characterized as the set of initial values of those
solutions of Equation (1.2) that exist for all t E IR and increase less rapidly
than any fixed exponential function.
Definition 2.1. For a given neighborhood V of 0 E C, the local strongly

stable set (or manifold) W1~~(0)~fW 88 (0, V) of the equilibrium point 0 of
Equation (1.1) is the collection of points ¢ E C with the property that
the solution Xt ( · , ¢) E V for t ~ 0 and approaches zero exponentially
as t ~ oo. In the same way, we define the local strongly unstable set (or
manifold) W1~~(0) ~ wsu(o, V).
Definition 2.2. For a given neighborhood V of 0 E C, a local center manifold

W1~c(O) ~f wc(o, V) of the equilibrium point 0 of Equation (1.1) is a C 1 -
submanifold that is a graph over VnN inC, tangent toN at 0, and locally
invariant under the flow defined by Equation (1.1). In other words,
W1~c(O) n V =N E C: '1/J = ¢ + h(¢), ¢EN n V}

where h : N ~ U EB S is a C 1 -mapping with h(O) = 0, Dq,h(O) = 0.
Moreover, every orbit that begins on W1~c(O) remains in this set as long as
it stays in V.
Definition 2.3. For a given neighborhood V of 0 E C, a local center-stable

manifold Wj~~ (0) ~f wcs (0, V) of the equilibrium point 0 of ( 1.1) is a set in
C such that W1~~(0)nV is a C 1-submanifold that is a graph over (NEBS)nV,
tangent to NEBS at the origin and is locally invariant under the flow. In
other words,
where h : N EB S -; U is a C 1 -mapping with h(O) = 0, Dcph(O) = 0.

Moreover, every orbit that begins on W1~~(0) remains in this set as long as it
stays in V. Furthermore, any orbit that stays in V for all t 2: 0 must belong
to W1~~(0). In the same way, we define the local center-unstable manifold
W1~~(0) ~r weu(o, V) of the equilibrium point 0 of (1.1) by replacing t 2: 0
by t:::; 0, the set NEBS by NEB U, and the set U by S.
The basic result on the existence of the invariant manifolds is the

following.
Theorem 2.1. IfF in (1.1) is a Ck-function, then there is a neighborhood

V of 0 E C such that each of the sets W1~~(0), W1~~(0), W1~c(O), W1~~(0),
and W1~~(0) exists and is a Ck-submanifold of C. The manifolds W1~~(0)
and W1~~(0) are uniquely defined, whereas the manifolds W1~c(O), W1~~(0),
and W1~~ (0) are not. Furthermore, every invariant set of (1.1) that remains
in V must belong to W1~c(O).
We do not give all of the details of the proof of Theorem 2.1, but simply
outline the major steps. To be specific, we concentrate on the local center
manifold. With the decomposition C = U EB NEBS, we can write
(2.2)
Let <l>e be a basis for N and tJie be a corresponding basis for the solutions
of the adjoint equation with (<Pc, tJie) =I, where ( ·, ·) is the usual bilinear
form defined by Relation (2.23) of Chapter 9. We know that T(t)<Pe =
<[> eeBct, where the eigenvalues of Be have zero real parts and correspond to
the solutions of (1.3) that lie on the imaginary axis. If we let xf ( · , cj;) =
<l>ey(t), then the solution Xt = <l>ey(t) + xfEllS with Xo = cj; is a solution of
the system
Y =Bey+ tJie(O)f(<l>eY + xfEll 8 )

(2.3)
xfEllS = T(t)cj;UEllS +lot d[K(t, T)UEllS]f(<Py(T) + x~Ells),
and conversely.
It is convenient to modify the function f in the direction of N so that
we can consider the arbitrary elements of N rather than those elements
of N that are in a small neighborhood of 0. Let X : JRdN - ; [0, 1] be a
C 00 -function with x(y) = 1 for IYI :::; 1, x(y) = 0 for IYI 2: 2. For any TJ > 0,
define
10.2 Nonhyperbolic equilibrium points 315
(2.4)
Let c,., = { ¢ E C : ¢ = PeYX(YTJ-1) + ¢UFBS, y E JRdN' [¢UFBS[ ::; TJ }. For

any 'I > 0, there exists a positive number TJ such that
(2.5)
Let
S(b, Ll) = { h: JRdN --+ u EEls: [h[ ::; b, [h(y)- h(Y)[ ::; Ll[y- ill}.
For any h E S(b, Ll), define y(t, y0 , h), y(O, Yo, h) = Yo, to be the unique
solution of the equation
(2.6) iJ =BeY+ lfFe(O)l(y, h(y))

and define the operator Ton S(b, Ll) by the formula
(2.7)
(Th) = -1 00
d[K( -T)Ujj(y(T, y0 , h), h(y(T, Yo, h))) dT
0
+[ 00
d[K(-T) 8 j](y(T,yo,h),h(y(T,yo,h)))dT.
After several estimates, it is now possible to show that for appropriate

constants b, L1 (that is, for a sufficiently small TJ), the operator Tis a uniform
contraction on S(b, Ll). Therefore, there is a unique fixed point he E S(b, Ll).
By applying the same type of reasoning as in the proof of Theorem
1.1, we deduce that the function he is a Ck-function.
The local center manifold W1~e(O) is then given by
The flow on W1~c(O) is obtained from the solutions of the ordinary differ-
ential equation
(2.8) iJ =BeY+ lfFc(O)f(PcY + hc(Pey)) ~f Bey+ Ye(Y)

with the initial data Yo= (lfFe, ¢).In fact, the solution Xt( ·,¢)of (1.7) with
cP E W1~e(O) is given by Xt( ·, ¢) = Pey(t, Yo)+ he(Pcy(t, Yo)), where y(t, Yo)
is the solution of (2.8) with Yo = (IJFe, ¢).
In the same way, we obtain the other invariant manifolds in the state-
ment of Theorem 2.1. Also, the solutions of (1.1) on W1~~(0) are de-
scribed by the solutions of an ordinary differential equation. In fact, let
dim U EEl N = dutf!N, Pcu be a basis for N EEl U and lfFeu be a corresponding
basis for the solutions of the adjoint equation with (Peu, lfFcu) = I, where
( ·, ·)is the usual bilinear form. We know that T(t)Peu = Peue 8 cut, where
the eigenvalues of Bcu have nonnegative real parts. Then there exists a
function hcu : IRduE!lN --+ S such that
(2.9) W~~~(O) = {¢ E C: c/> = .PcuZ + hcu(z), izi :S 17 }.

If cj> E W1~~(0), then Xt( ·, ¢) = .Pcuz(t, zo) + hcu(z(t, zo)), where z(t, zo) is
the solution of the ordinary differential equation
(2.10)
with the initial data zo = (lftcu, ¢).
The local center-unstable manifold has a certain type of stability prop-
erty that is sometimes referred to as asymptotic phase. Any solution off
the center-unstable manifold decays exponentially toward a solution on the
center-unstable manifold as long as it remains in a neighborhood of the
origin. The precise description of this property is the following theorem.
Theorem 2.2. Suppose that the hypotheses of Theorem 1.1 are satisfied. Then
there exists a neighborhood V of 0 E C, positive constants K 1, a1, and a
Ck-function H : V - t U EB N such that if(/> = H(¢), then the solution
Xt( ·, (/> + ¢ 8 ) of (1.7) satisfies the property that
(2.11) U(f)N (-, ¢- + '¢) = H (.Pcuz(t, zq,)

Xt + Xt8 (-, ¢- + '¢) ) ,
where z(t, zq,) is the solution of (2.10) with zq, = (lftcu, ¢). In addition,
t:::: 0
lxf(f)N ( ·, (/> + '¢)- .Pcuz(t, zq,)l :S K1e-a. 1 t,
(2.12)
lxf(-, (/> + '¢)- hcu(z(t, zq,))l :S K1e-a. t, t:::: 0
1
as long as the solution remains in V.
We now outline the proof of Theorem 2.2. The first step is to introduce
a coordinate system in which the center-unstable manifold replaces U EB N
as coordinate. If Xt is a solution of (1. 7), let
(2.13) Xt = .Pcuz(t) + hcu(z(t)) + yf,

A few computations will show that
Z = BcuZ + lftcu(O)f(.PcuZ + hcu(z) + yf)

~ BcuZ + Z(z, yf),
1t
(2.14)
yf = T(t)(¢ 8 - hcu(zq,)) + d[K(t, r) 8 ]F(z(r), y~)
where
(2.15)
We now consider a class offunctions L(z, '¢) = z + M(z, '¢) E lR.duE!)N,
z E '¢ E S, with M(z, 0) = 0. Our objective is to determine the
JR.duE!lN,
function M(z, '¢) so that if¢ E Cis given, zq, = (lftcu, ¢) and
10.3 Hyperbolic periodic orbits 317
{iJ = ([>cu(Zcf> + M(zcf>,'ljJ)),

then the solution (z(t, Z¢>H), yf (¢ + '1/J)) of (2.14) satisfies the relation
s -
z(t, Z¢>H) = L(z(t, Zcf>), Yt (¢ + '1/J))
(2.16) ~f z(t, Zcf>) + M(z(t, Zcf>), yf ({iJ + '1/J)),
s - s -
Yt (¢ + '1/J) = Xt (-, ¢ + '1/J)- hcu(z(t, Zcf>)) ___. 0 as t ___. oo.
Recall that z(t, Zcf>) is the solution on the center manifold in our coordinate
system. This implies that the function w(t) = M(z(t, Zcf>), Yt(¢ + '1/J)) with
w(O) = M(zcf>, '1/J) is a solution of the equation
where yf (¢ + 'ljJ) is the solution of the equation
(2.18) yf = T(t)'ljJ +lot d[K(t, r) 8 ]F(L(z(r, Zcf>), y~), y~) dr.

This suggests that we attempt to obtain the function M(z, '1/J) as a
fixed point of the operator
(2.19)
(TM)(zcf>, '1/J) = 1 00
e8 cur [z(L(z(r, Zcf>), y~(¢ + '1/J)), y~(¢ + '1/J))
- Z(z(r, Zcf>), 0)] dr,
where yf(¢ + '1/J) is the solution of (2.18).

As in the proof of the center manifold theorem, let S (8, Ll) be the class
of Lipschitz continuous functions M : V x V ---. IRUEilN, where V (resp.
V) is a neighborhood of 0 E IRUE!lN (resp. S), such that JM(¢,'1/J)I :::; 8,
M(¢,0) = 0 and M has Lipschitz constant< Ll. It is possible to show that
the neighborhoods V, V and constants 8, Ll can be chosen so that Tis a
uniform contraction on S(8, Ll) and thus has a unique fixed point M* in
S(8, Ll). In the proof, we use the fact that Jyf(¢+'1/J)I:::; K 1 e-a 1 t, t 2: 0, for
some positive constants K1, a 1 . This completes the proof of the theorem.
As in Remark 1.1, we can have parameterized versions of Theorem 2.1
and dependence of the vector field on t.
10.3 Hyperbolic periodic orbits
In this section, we study the neighborhood of a period orbit "( of an FDE

of either retarded or neutral type. Since a periodic orbit is a C 1 -manifold,
we can define a transversal E at a point p E 'Y and, therefore, obtain a
Poincare map on E. Since the solutions T(t)¢ of RFDE become smoother
in t as t increases, we can be sure that the Poincare map is Ck if the

vector field is Ck. We merely take the period of the periodic orbit large
enough. We may then develop the manifold theory near a fixed point for
the Poincare map in a manner analogous to the theory near equilibrium
points in Sections 10.1 and 10.2. This artifice will not work for NFDE since
the solution operator does not smooth with time. We present an approach
that gives partial information in this more general situation.
We will need a special case of the following result that is of indepen-
dent interest and stated without proof (see the supplementary remarks for
references).
Theorem 3.1. IfF is ck (resp. analytic) and X : ( -oo, 0] -+ lRn is a

solution of Equation (1.1) on ( -oo, 0] that is bounded, then x E Ck+ 1 (resp.
analytic).
We say that 'Y is an w-periodic orbit of (1.1) of minimal period w > 0

if 'Y = { p( t), t E lR}, where p( t) is a periodic solution of ( 1.1) of minimal
period w > 0.
Corollary 3.1. IfF is Ck (resp. analytic) and 'Y = {p(t), t E lR} is an

w-periodic orbit, then p E Ck+ 1 (resp. analytic) and 'Y is a Ck+ 1 -manifold
( resp. analytic manifold).
If 'Y = { p(t), t E lR} is an w-periodic orbit, then the linear variational

equation about p(t) is
(3.1) L(t) = D¢F(p(t)).
Equation (3.1) is a linear equation with coefficients periodic in t of

period w. Therefore, we can define the Floquet multipliers as in Chapter
8. From Corollary 3.1, if F E Ck, k ;::: 1, then p E Ck+ 1 and thus, p(t)
is a nontrivial periodic solution of (3.1). As a consequence, 1 is always a
characteristic multiplier of (3.1).
Definition 3.1. The Floquet multiplier of a periodic orbit 'Y are the Floquet
multipliers of the linear variational equation (3.1) except that 1 is not a
multiplier of 'Y if 1 is a simple multiplier of (3.1).
Definition 3.2. A periodic orbit 'Y is hyperbolic if each Floquet multiplier of

'Y has modulus different from 1. The index i('Y) of a hyperbolic orbit 'Y is
the number (counting multiplicity) of Floquet multipliers with moduli > 1.
Let x(t, ¢) be the solution of (1.1) with initial value ¢ at t = 0. We

define the stable set and unstable set of the periodic orbit 'Y of (1.1) as,
respectively,
W 8 ('y) = {¢ E c: XtC '¢) -Fy as t - 00 },

(3.2)
wu('y) = { ¢ E c: XtC '¢)- 'Y as t--oo}.
For a given neighborhood V of 'Y, we also can define the local stable and
local unstable sets
W1~c('y) ~f W 8 ('y, V) = { </> E W 8 ('y): Xt( ·, ¢) E V fort 2:: 0}

(3.3)
Wi~c('y) ~f Wu('y, V) = { </> E Wu('y) : Xt(-, ¢) E V fort:::; 0 }.
Theorem 3.2. IfF is a Ck-function, k 2:: 1, D¢ = ¢(0) (that is, (1.1)

is an RFDE), and 'Y is a periodic orbit of the RFDE (1.1), then there is a
neighborhood V of 'Y such that W 8 ( 'Y, V) and wu ('Y, V) are Ck -submanifolds
of C with dim wu('y, V) = i('y) + 1 and codim W 8 ('y, V) = i('y).
Proof. Let 'Y = {p(t), t E lR} have minimal period wand let T(t)¢ be the
solution of (1.1) with initial data¢ at t = 0. Fix a E [O,w) and let Ea. be a
codimension one transversal to 'Y at Po.· We can choose an E-neighborhood
N('y, E) of 'Y and the transversals Ea. in such a way that, for any¢ E N('y, E),
there are a unique a E [O,w) and <Po. E Ea. nN('y, E) such that¢= Po. +<Po.·
The orbit 'Y and the set of transversals form a coordinate system near 'Y.
Since the solutions of RFDE smooth in t as t increases, it follows
that T(t)¢ is a Ck-function oft,¢ if t > kr. Choose an integer j so that
jw > kr. Let D(rro.) C Ea. be a neighborhood of Po. in Ea. with the property
that, for each ¢ E D(rr0 ), there is a unique t = t(¢) near jw such that
T(t(¢))¢ E E 0 • If we define 7r0 : D(rro.)- Ea. by 7r0 (¢) = T(t(¢))¢, then
1r0 is a Ck-function and 7ra.(Po.) =Po.·
We introduce the stable and unstable sets of the fixed point Po. of 7r 0 :
W1~c(Po.) ~f W 8 (Po., V) = { ¢ E C: rr':¢ E V, m 2:: 0, rr':¢- Po.

as m-oo}
(3.4)
Wi~c(Po.) ~f Wu(Po., V) = { ¢ E C : rr':¢ E V, m :::; 0, rr':¢- Po.
as m-oo}.
It is possible to show that the elements of the point spectrum of the

linear mapping D¢7ra.(Pa.) are the Floquet multipliers of 'Y· Since 'Y is hyper-
bolic, no element of the spectrum of D¢7ro.(Po.) has modulus one. Now, we
can use the same type of arguments as in Section 1 for stable and unstable
manifolds near equilibrium points to conclude that there is a neighborhood
V of Po. such that W 8 (p 0 , V), wu(p 0 , V) are Ck-submanifolds of C. In the
previous proof, every infinite integral is replaced by an infinite sum. Since
'Y is compact, we also can choose V independent of a. Also, we extend the
definition of W 8 (p 0 , V) to all a E lR by setting W 8 (p 0 , V) = W 8 (Po.+w, V),
wu(Po., V) = Wu(Po.+w, V).
For any (3 E [O,w) and any neighborhood V of p13, define
WJ,loeb) = W b) n (Ep n V),

8
W$,loeb) = wub) n (Ep n V).
We claim that W$,loeb) is a Ck-submanifold of dimension ib). In

fact, for any integer j such that (j- l)w > kr, it is obvious that there is a
neighborhood V of p 0 such that
Also, since wu(p0 , V) = wu(PJw, V) for all j, there is a neighborhood V

of Pf3 such that
that is, the inverse map of T(jw + {3) : wu(po, V) __, W$,loeb) is
T(jw)- 1 T(jw- {3), which is Ck.
If we extend the definition of W$,loe b) so that it is periodic in {3, then
the argument shows that w,::,loe ('Y) = T(jw - {3) W$,loe ('Y). If we denote the
tangent space of a submanifold M of C by T M, then
(3.5) TW,::,loeb) = D¢T(jw- {3)TW$,loeb).

Using the fact that Ep is transversal to"( at Pf3 (that is, Ep E9 [Pp] = C)
and the implicit function theorem, we can find an E > 0 such that the set
u8E((3-E,(3+E) w~loeb) is a Ck-submanifold modeled on [Pp] E9 TW$,loeb).
Since 'Y is compact, this shows that Wub) satisfies the properties stated
in Theorem 3.2.
We need to show the same properties for the stable manifold of 'Y· For
any integer j 2 1, we have WJ,loeb) = T(jw- {J)W~,loeb). Choose j so
that (j-l)w > kr. Let Y C C be the set of¢ E C such that Dc/>T(jw-{3)¢ E
TW~ loeb). Since TW~ loeb) is a closed subspace of C, it follows that Y is
aclos'edsubspaceofC. Also, [Pp]EBTWJ,loeb)EBY = Cfrom (3.5). We now
use the implicit function theorem to obtain that WJ,loeb) is a Ck-graph
over Y and is thus a Ck-submanifold of C.
The stable set of "( in a neighborhood of {3 is defined by
w;:foeb) = {P8 + ¢: T(jw- 8)¢ E W~,loeb), 8 E ({3- E, {3 +E)}
for some E > 0. Using the fact that we have [Pp] E9 TWJ,loeb) E9 Y = C,
we can use the implicit function theorem to show that w;:foeb) is a Ck-
submanifold modeled on lPa] E9 Y = lPa] E9 TWJ,loeb). Since"( is compact,
we can obtain the conclusion in Theorem 3.2 for the stable manifold of 'Y·
This completes the proof of the theorem. D
Let us now turn to the analogue of Theorem 3.2 for NFDE. In this
case, we have been unable to find a way to use the Poincare map
c/J 1-+ 7r01 (c/J) = T(t(cjJ))cjJ.
If t(¢) is not a constant, then 1r01.(¢) is not differentiable. As a consequence,

we proceed in a different manner. We introduce the synchronized stable and
synchronized unstable sets of a point P/3 on the periodic orbit "(:
W~('Y) = { <f> E C: Xt(-, ¢)- Pt+/3--+ 0 as t--+ oo },

(3.6)
W»('Y) = { c/J E C: Xt( ·, ¢)- Pt+/3--+ 0 as t--+ -oo }.
For a given neighborhood V of"(, we also can define the local synchronized
stable and local synchronized unstable sets of a point P/3 on the orbit "(:
W~,locb) ~f W$('Y, V) = { 1> E W~('Y) : Xt( ·, ¢) E V fort :2:: 0 },

(3.7)
W»,locb) ~f W~('Y, V) = { ¢ E W~('Y) : Xt( ·, ¢) E V fort~ 0 }.
We define the local synchronized stable and local synchronized unstable sets
of 'Y as
Wk,c('Y) = Ws('Y, V) = U W~,locb),

/3EIR
(3.8)
W1~c ('Y) = wu ('Y, V) = U W~,1oc ('Y) ·
/3EIR
Theorem 3.3. IfF is a Ck-function, k :2:: 2, and 'Y is a periodic orbit of ( 1.1),
then there is a neighborhood V of 'Y such that the synchronized stable man-
ifold W 8 ('Y, V) is a ck- 1 -submanifold of C and the synchronized unstable
manifold wu ('Y, V) is a Ck -submanifold of C with dim ('Y, V) = i('Y) + 1 wu
and codim W 8 ('Y, V) = i('Y).
The proof of Theorem 3.3 involves several results of independent in-

terest.
Lemma 3.1. IfF is a Ck-function, k :2:: 1, and 'Y is a periodic orbit of (1.1),
then there is a neighborhood V of 'Y such that W~('Y, V) and W~('Y, V) are
Ck-submanifolds of C with dim W»("f, V) = i('Y) and codim W~('Y, V) =
i('Y)+l.
Proof. To simplify the notation, we first take f3 = 0. If x(t) = p(t) + z(t) in
(1.1), then z satisfies the equation
(3.9)
(3.10) L(t, '1/J) = D<PF(pt)'l/J, G(t, ¢) = F(Pt + ¢)- F(pt)- L(t, ¢).
The variation-of-constants formula for Equation (3.9) is
(3.11) Zt = T(t, a)¢+ lot d[K(t, T)]G(T, Z7 ).
For each fixed T E IR., the results in Chapter 8 imply that the space C
can be decomposed as
C = U(T) EB N(T) EB S(T)

where U (T) is finite dimensional and corresponds to the span of the gen-
eralized eigenspaces of the Floquet multipliers of (3.1) with moduli > 1,
N(T) is one dimensional and is the span [p 7 ] of p7 • This decomposition of
C defines three projection operators
1l'U(r) : C -t U(T), 1l'U(r)U(T) = U(T),
1l'N(r) : C -t N(T), 1l'N(r)N(T) = N(T),
1l'8(r) : C -t S(T), 1l'8(r)S(T) = S(T)
where 1l'U(r) + 1l'N(r) + 1l'8(r) =I. If T(t, a) is the solution operator of the
linear equation (3.9), then
T(t, a)U(a) = U(t), T(t, a)N(a) = N(t), T(t, a)S(a) = S(t).
This relation holds on S( ·) for all t ~a and it holds on U( · ), N( ·) for all

t E IR..
Also, there are positive constants M, a such that for¢ E C,
IT(t,a)1l'u(a)¢1:::; Me"'(t-a)l¢1, t:::; a,

Var(t,o] 1l'u(a)K(t, a) :::; M e<>(t-a), t :::; a,
(3.12)
IT(t,a)11'8(a)¢1:::; Me-<>(t-a)l¢1, t ~a,
Var[o,t) 1l'8(a)K(t,a):::; Me-<>(t-a), t ~a.
With this notation, consider the integral equation
Zt = T(t,0)¢ 8 +lot d[11'8(r)K(t,T)]G(T,Z 7 )
(3.13)
-1 00
d[(11'u(r)K(t,r) + 1l'N(r))K(t, T)jG(T, Zr)
fort~ 0, ¢8 E 1!'8 (o)C arbitrary, and the integral equation
Zt = T(t, 0)</Ju +lot d[1l'u(r)K(t, T)]G(T, Z 7 ) dT

(3.14)
+[ 00
d[(11'N(r) + 1l'8(r))K(t, T)jG(T, Z 7 )
for t :::; 0, <Pu E 1l'u(o)C arbitrary. Any solution of these integral equations
will be a solution of Equation (3.9).
The proof of Lemma 3.1 will be complete if we make use of the following
result, which is verified using the same type of arguments as in the proof
of Theorem 1.1.
Lemma 3.2. If F E Ck, k 2': 1, and 1 is a hyperbolic periodic orbit of

(1.1), then there are positive constants 8, v, such that for any cp 8 E n 8 (o)C,
lcp 8 1 < v, there is a unique solution z;(cp 8 ) of Equation (3.13) with
lz;(cp 8 )1 :::; tJe-at/ 2 fort 2': 0. Furthermore, if H 8 (cp 8 ) = z0(cp 8 ), then the
set { H 8 (cp 8 ), lcp 8 1< v} is a Ck-graph over the set { cp 8 : lcp8 1< v} through
the projection operator 1r8(o)· Also, Dq,H 8 (0) =I.
For any cpu E nu(o)C, lcpul < v, there is a unique solution z;*(cpu)
of Equation (3.14) with lz;*(cpu)l :::; 8eatf 2 for t :::; 0. Furthermore, if
Hu(cpu) = zo*(cpu), then the set { Hu(cpu), lcpul < v} is a Ck-graph over
the set {cpu : Icpu I < v} through the projection operator nu(o). Also,
Dq,Hu(o) =I.
Proof of Theorem 3.3. If we fix j3 E IR and let x(t) = p(t +,B)+ z(t) in (1.1),
then z( t) satisfies the equation
d
(3.15) dt Dzt = L(t + /3, Zt) + G(t + ,8, Zt),
where L and G are given in (3.10). Proceeding exactly as in the proof of
Lemma 3.1, we obtain Ck-functions
H 8 (cpg,,B), z*(cpg,/3), and Hu(cp~,/3), z**(cp~,/3),
defined for cpg E 1r8(fJ)C, lcpgl < v, cp~ E 7ru(f3)C, lcp~l < v. Also,
Dq,H 8 (0, ,B) = I, Dq,Hu(o, ,B) = I. Since 1 is compact, the constants 8, v
can be chosen to be independent of ,8. It is clear from the preceding con-
struction that
WJ,locb) = {pp + Hs(cpg, /3), lcpgl < V },
W~,locb) = {pp + Hu(cp~, /3), lcp~l < v }.
It also is clear that
(3.16) Wl~c(!) = U WJ,locb), Wi:,c(!) = U W~,loc(!).

(3EIR (3EIR
It remains to show that the sets W1~c (!) and Wk,c (!) are submanifolds
satisfying the properties stated in Theorem 3.3. We discuss only the stable
set since the unstable set is treated in a similar way.
We claim first that, if v > 0 is sufficiently small and 'T] E W1~c(!),
then there are unique /3 E [0, w), cpg E 1r 8 ((3) C, Icpg I < v, such that 'T] =
pp + H 8 (cpg). Define the function
f(ry, {3, ¢%) ~ 'TJ- Pf3- Hs(¢%, {3)
for (ry, {3, <P%) E C x ill. x 1fs(f3)C· It is clear that f(Pf3, 0, 0) = 0 and the
derivative Df3f with respect to {3, <P% evaluated at (h, '1/J) E ill. x 1fs(f3)C
satisfies the relation Df3f(Pf3, 0, O)(h, '1/J) = -·Pf3h- '1/J. Therefore, the map
Df3f(Pf3, 0, 0) is an isomorphism on ill. x 1fS(f3)C· The implicit function theo-
rem implies that there are a positive number v and Ck-functions {J(ry), <P%(ry)
with I'TJI < v, lfJ(ry)l < v, 1¢%(ry)l < v, such that f(ry,{3(ry),¢%(ry)) = 0. Since
"( is compact, the claim is proved.
Our next assertion is that W1~c('Y) is a ck- 1-submanifold of C. Since
'Y is compact, it is sufficient to show that, for any {30 < w, the set
S(f3o) = U W$,Jocb)
(3E[O,f3o]
is ck- 1-diffeomorphic to [0, f3o] X Bo(v), where Bo(v) = { <Pg E 7fs(oP :

1</Jgl < v }. From Lemma 1.4 in Section 8.1, we know that there is a ck- 1_
isomorphism If3 : 7fs(o)C-+ 1fS(f3)C. For any 'TJ E W1~c('Y), there are unique
{J(ry) E [O,w), ¢~ E 7fS(f3("1)) such that 'TJ = Pf3("1) + ¢%(.,)· The mapping
ry f--.+ (f3(ry),Tfi<~)¢%(.,)) is a ck- 1 -diffeomorphism.
In the same way, we conclude that Wk,c('Y) is a Ck-submanifold of C.
We obtain Ck rather that ck- 1 because the isomorphism from Lemma 1.4
of Section 8.1 is Ck. This completes the proof of Theorem 3.3. D
At this time, we have been unable to show the following natural con-
jecture.
Conjecture. W1~c ('Y) = W1~c ('Y) and W1~c ('Y) = Wk,c ('Y).
It is clear that ~~c('Y) ::J Wfoc('Y). However, the proof that W1~c('Y) c
W1~c('Y) is not so easy. It is not obvious that a solution of (1.1) that ap-
proaches 'Y must satisfy the integral equation (3.13) or, more precisely, the
one corresponding to (3.15).
10.4 Nondegenerate periodic orbits ofRFDE
Suppose that (1.1) has a periodic orbit 'Y of period w. We say that 'Y is
nondegenerate if 1 is not a Floquet multiplier of 'Y· We remark that this
is equivalent to saying that the linear variational equation (3.1) has 1 as a
simple characteristic multiplier. The concept of nondegeneracy depends on
w. More precisely, if Wm is the minimal period of"(, then it is possible for
'Y to be nondegenerate and have other Floquet multipliers that are roots of
unity. If this were the case, then the orbit would not be nondegenerate if
we choose w = jwm for j an appropriate integer.
10.4 Nondegenerate periodic orbits of RFDE 325
If A is a Banach space and F: C X A---+ lRn, (¢,>..) ~---+ F(¢,>..), is

continuous in (¢, >..) and continuously differentiable in ¢, F(¢, 0) = F(¢),
we consider the RFDE
(4.1)
as a perturbation of the RFDE
(4.2) ±(t) = F(xt)·
Our objective is to prove the following result.
Theorem ~.1. If 'Y is a nondegenerate periodic of (4.2) of period w and the

function F satisfies the hypotheses, then there exist a positive constant v
and a neighborhood V of"( such that for each>.. E A, 0 :-: :; 1>..1 < v, Equation
(4.1) has a nondegenerate periodic orbit"(>-. E V of period W>-., "(>-. and W>-.
depend continuously on>.., 'Yo = "(, w0 = w, and "f>-. is the only periodic orbit
in V whose period approaches w as >..---+ 0.
Remark 4.1. Under further smoothness properties on the function F, it is

possible to obtain that the orbit and period in Theorem 4.1 are smooth in
>... More precisely, if F is ck' then these functions are Ck One can even
0
take k = oo, but it is not known if we can extend the result to analyticity.
Remark 4.2. If w > r, then we can use the Poincare map introduced in the
proof of Theorem 3.2 together with the implicit function theorem to obtain
the conclusion in Theorem 4.1. On the other hand, if w :-:::; r, we cannot
proceed in this way since this map is not differentiable. If we take some
multiple jw so that jw > r, then, as remarked earlier, the orbit 'Y may not
be nondegenerate with respect to the period jw.
Remark 4.2 suggests that the standard implicit function theorem may
not be appropriate to prove Theorem 4.1. We will make use of the following
version of the parametric implicit function theorem, which involves only the
differentiability with respect to the parameter along the fixed-point set. The
proof may be supplied by the reader as an application of the contraction
mapping principle.
Lemma 4.1. Let E be an open subset of a Banach spaceY, F be a closed

subset of a Banach space X, int F =I= 0, where int F denotes the interior of
F. Assume that T: F x E---+ F satisfies the following hypotheses:
(i) T(x, ·) : E ---+ F is continuous.
(ii) T( ·, y) : F---+ F is continuous and, for each y E E, has a unique fixed
point x(y) that depends continuously on y.
(iii) If x(E) = F 1 C F, then T(x, y) is continuously differentiable in y for
(x,y) E F 1 x E.
(iv) There is an open set F 2 C X such that F C F 2 and the deriva-

tive DxT(x, y) of T(x, y) with respect to x is continuous and satisfies
IDxT(x, y)l ~ 8 < 1 for all (x, y) E F2 x E.
Then the fixed point x(y) E F, y E E, of T(x, y) is continuously differen-
tiable in y.
Proof of Theorem 4.1. For any real number /3 > -1, if t = (1 + /3)T and
x(t) = y(T), then x(t + 0) = y(T + 0/(1 + /3)), -r ~ 0 ~ 0. If we define
Yr,(3(0) = y(T + 0/(1 + /3)), -r ~ 0 ~ 0, then Equation (4.1) becomes
(4.3) y(T) = (1 + /3)F(Yr,f3, A).

If there is a periodic solution of Equation (4.1) of period w, then there is a
periodic solution of Equation (4.3) of period (1 + f3)w, and conversely. Let
"( = {p(t) : 0 ~ t < w }, where pis a periodic solution of (4.3) of period w.
If y(T) = p(T) + z(r) in (4.3), then
(4.4) i(r) = L(r)zr,o + H(r, z, A, /3),

where
(4.5)
and
(4.6) H(r, z, A, /3) = -F(Pr,o)- L(r)zr,O + (1 + /3)F(pr,(3 + Zr,(3, T).

We remark that H(r,O,O,O) = 0 and the derivative DzH(r,z,A,/3) exists
and satisfies DzH(r,O,O,O) = 0.
Let C(2r) = C((-2r, 0], JRn). For l/31 ~ 1/2, we can interpret Equation
(4.4) as an RFDE on C(2r). The linear equation
(4.7) i(T) = L(T)Zr,O
has one as a simple characteristic multiplier corresponding to the periodic
solution p. Therefore, we may decompose C(2r) relative to the multiplier
1 as C(2r) = E EB K, where E = E 1 (0) and K = K 1 (0) (see Chapter
8). Suppose that P : C(2r) ---+ C(2r) is the projection induced by this
decomposition and P takes C(2r) onto K.
Let no be the set of continuous w-periodic functions in IRn with lzlo =
SUPt lz(t)l for z E no. Corollary 4.1 of Section 6.4 implies that the non-
homogeneous linear equation
i(r) = L(T)Zr,O + h(T), hE no,
has a solution in no if and only if
law q(T)h(T) dT = 0
10.4 Nondegenerate periodic orbits of RFDE 327
where q(T) is a nonzero w-periodic solution of the equation adjoint to (4.7).

We may assume that fow q(T)qT(T) dT = 1. For any h E Do, let F(h)
J:;' q( T) h(T) dT. Then r : Do ---> IR is a continuous linear mapping.
For any h E D0 , the equation
z(T) = L(T)Zr,O + h(T)- r(h)q'(T)

has a solution in D0 and it has a unique solution whose (I- F)-projection
is zero. If we designate this solution by Kh, the K is a continuous linear
operator taking Do into Do.
For any ..\, (3, and u E D 0 , consider the equation
(4.8) R(u, ..\, (3) ~r u- K[H( ·, u, ..\, (3)- F(H( ·, u, ..\, (J))qT] = 0.
Using the fact that H(T,O,O,O) = 0, DzH(T,O,O,O) = 0, we deduce from
the implicit function theorem that there are positive constants v 0 , (30 , Do,
such that Equation (4.8) has a unique solution u* (..\, (3) E D 0 , Ju* (..\, (3) I ::::;
Do, J>.J ::::; vo, lf31 ::::; f3o, u*(..\, (3) is continuous in ..\, (3, u*(O, 0) = 0 and
u* (..\, (3) satisfies the equation
(4.9) z(T) = L(T)Zr,O + H(T, z, ..\,(3)- B(..\,(J)qT(T)
where we have put
(4.10) B(..\,(3) =low q(T)H(T,u*(..\,(3),..\,(J)dT.
In particular, u* (..\, (3)( T) is continuously differentiable in T.

If u is a C 1 -function in .f20 , then the function H( ·, u, ..\, (3) is contin-
uously differentiable in (3. Thus, Lemma 4.2 implies that u* (..\, (3) has a
continuous first derivative with respect to (3. This implies that the function
B(..\, (3) is continuously differentiable in (3. To compute this derivative, we
observe that the function au*(..\, (3)/8(3 is a solution of the equation
. 8B(..\, (3) T
(4.11) v(T)=L(T)Vr,o+Ll(T,v,..\,(3)- o(J q (T)
where
L1 (T, v, ..\, (3) = (1 + (3)Dq,F(Pr,{3 + u~,{3' ..\)vr,{3 - Dq,F(Pr,o)Vr,O
+ F(Pr,{3 + u~,{3, ..\)
- 1 ~ (3Dq,F(Pr,{3 + u;,{3, ..\)(( · )Pr,{3 + ( · )u;,{3)
where (( · )Pr,f3)(B) = Bj;(T+B/(1+(3)) and (( · )u;,f3)(B) = Bu*(T+B/(1+(3)).
From (4.6) and (4.9), we see that u~,{3---> 0 as ..\,(3---> 0. From this fact and
the relationj;(t) = F(pt), it follows that L 1 (t,v,O,O) = J(t,p), where
(4.12) J(t,j;) = j;(t)- L(t)( · )Pt·

Since Equation (4.11) has thew-periodic solution 8u*(>.,(3)j8(3, we must

have
and so
(4.13)
8B(O,O)
8(3 = lo
r q(T)J(T,p) dT
where J is given in (4.12).

We show now that Expression (4.13) is not zero. If Expression (4.13)
is zero, then there is a nontrivial w-periodic solution z of the equation
z(t) = L(t)zt + J(t,p).
If x(t) = z(t) - tp(t), then a few elementary computations imply that x
is a solution of (4.7) with z and p being w-periodic functions. Since p also
is a solution of (4. 7), we deduce that the characteristic multiplier one is
not simple. This is a contradiction to the assumption that the orbit "( is
elementary and proves that the Expression (4.13) is not zero.
Since B(O, 0) = 0 and 8B(O, 0)/8(3 =1- 0, the implicit function theo-
rem implies the existence of positive constants (31 ~ (30 , v1 ~ v 0 , and a
continuous function (3(>.), 1>.1 ~ v1, 1(3(>.)1 ~ (31, so that (3(0) = 0 and
B()-.,(3(>.)) = 0. Since u*(>.,(3) is a solution of Equation (4.9), it follows
that u*(>., (3(>.)) is an w-periodic solution of Equation (4.4). This proves
the existence of a periodic solution y*(>.) of (4.3) and thus a solution x*(>.)
of Equation (4.1) of period w(>.) = (1 + (3(>.))- 1 w, which is continuous in)..
for 0 ~ 1>.1 ~ v1, x*(O) = p. The linear variational equation associated with
this periodic solution x* (>.) is a continuous function of ).. and, therefore,
the multiplier one will have a generalized eigenspace of dimension one for
o ~ I>.I ~ v2 ~ v1 .
Since each periodic orbit of (4.1) near"! with a period close tow can be
obtained in this manner, we have completed the proof of Theorem 4.1. D
Corollary 4.1. lf"f is a nondegenerate periodic orbit of the RFDE (4.2), then
there are a neighborhood V of"( and a positive number 8 > 0 such that (4.2)
contains no periodic orbits in V \ {"(} of period w satisfying lw- wl < 8. In
particular, V \ {"!} contains no w-periodic orbits.
Remark 4.3. It should be possible to extend this result to NFDE with a

stable D operator. It will be necessary to use the method of Liapunov-
Schmidt to obtain the analogue of (4.9) and then to show that the periodic
functions that satisfy that relation are C 1 -functions of t. The method of
Hale and Scheurle [1] could perhaps be used to prove this fact.
For RFDE, the original formulation of Theorem 1.1 for Lipschitz pertur-
bations of a linear vector field is due to Hale and Perella [1]. Lemma 1.2 is
due to Henry [6]. In ordinary differential equations, we have the classical
result of Hartman and Grohman that the flow near a hyperbolic equilib-
rium point is topologically equivalent to its linearization. Sternberg [2,3]
has given the appropriate extension of this theorem for the situation where
the RFDE has a global attractor. The topological equivalence is relative to
the solutions on the attractor.
Chafee [1, 2] discussed the existence of center manifolds of an equi-
librium point of RFDE in connection with the bifurcation of periodic or-
bits from an equilibrium point for equations containing a small parameter.
Kurzweil [1, 2] and Fodcuk [1, 2], have also considered this general problem.
Diekmann and van Gils [1] also have proved the existence of center man-
ifolds for RFDE by using the sun-star theory of dual semigroups and the
variation-of-constants formula. Ruiz-Claeyssen [1] has discussed the case in
which the initial data are chosen from the space Wf' of absolutely contin-
uous functions with essentially bounded derivatives in order to study the
effects of delays on the behavior near equilibrium. Lima [1] and Hale [24]
have considered the same problem in the space C and more general fading
memory spaces by making use of the fixed-point theorem in Lemma 4.1.
The center manifold theorem allows one to prove the following result:
For any FDE of the form (1.1) with p eigenvalues of Equation (1.2) on the
imaginary axis, there is an ordinary differential equation u = h( u), u E IRP,
defined in a neighborhood of zero in IRP such that h(O) = 0 and the stability
properties of the solution u = 0 of this equation are the same as the stability
properties of the solution x = 0 of Equation (1.1)
(see Hale [8]). Special cases of this general result were used earlier by
Prokopev and Shimanov [1], Hale [9] and Hausrath [1] to obtain sufficient
conditions for the stability of the solution of Equation (1.1) when no roots
of (1.2) have positive real parts. In these latter papers, it was necessary to
extend some of the classical transformation theory of Liapunov to equations
in infinite-dimensional spaces. The remarks of Hausrath [1] were later used
by Henry [5] and Carr [1], for parabolic equations as well as ODE. Further
extension of this transformation theory was used by Chow and Mallet-Paret
[1] to understand the the Hopf bifurcation for RFDE. See also the remarks
in Section 12.10.
Theorem 3.1 is due to Nussbaum [1] for RFDE and to Hale and
Scheurle [1] for NFDE. Hale [10] was the first to formulate a version of
Theorem 3.2. The proof in the text is based on Hale and Lin [2]. The proof
of Theorem 3.3 is based on ideas from Hale [10].
There are several approaches that could possibly be used to verify
the conjecture in Section 10.3 that the local stable and unstable sets of a
periodic orbit coincide with the synchronized local and unstable sets of the
orbit. The functions H 8 and Hu can be used as part of a coordinate system
in a neighborhood of f. In fact, using the implicit function theorem and the
compactness of{, we can deduce that there are positive constants p,, v > 0
such that for any rJ E C with dist (ry, C) < p,, there are unique (3 E [0, w),
¢~ E 1rS(f3)C, ¢~ E 7ru(f3)C, such that 1¢~1 < v, 1¢~1 < v, and
rJ = Pf3 + Hs(¢~, (3) + Hu(¢~, (3).

It is reasonable to expect that one could prove the conjecture with this
coordinate system and appropriate modifications of the arguments used by
Hale and Raugel [2] in their study of convergence to equilibrium of gradient-
like systems.
Another approach would be to develop a more extensive theory of
invariant manifolds (center stable, center unstable, and the theory of folia-
tions near the periodic orbit) or even more generally near an invariant set.
There seem to be some difficulties in using the approach in Hirsch, Pugh,
and Shub [1] for maps adapted for flows. If we attempt to use something
similar to the approach in the text by considering the variation of solutions
from a particular solution on the invariant manifold, then we encounter
nonautonomous equations for which we can construct the synchronized
manifolds. The same difficulty as in the conjecture occurs. On the other
hand, if the nonautonomous equation can be considered as a skew product
flow, then it is feasible that invariant manifolds in the skew product flow
can be used to determine the invariant manifolds in the original space. If
this is the case, then the conjecture could be proved. For finite-dimensional
problems, Chow and Yi are investigating this problem by this approach.
Success in the finite dimensional case will certainly lead to some results in
infinite dimensions.
Theorem 4.1 was given by Hale [10] and generalizes an earlier result of
Halanay [1]. Stokes [3] has given a different proof of this result based on a
general theory of when approximate periodic solutions of an equation imply
the existence of an exact periodic solution. His procedure also is effective
numerically.
Stokes [1,4,5] has studied the orbital stability of periodic orbits in a
general setting. Stokes [6] has introduced a local coordinate system around
a periodic orbit 1 of a RFDE and has done this in such a way as to be able
to determine the many properties of the solutions in a neighborhood of f.
This result has implications for nonautonomous perturbations of the vector
field. The coordinate system is not a natural generalization of the finite
dimensional case in the sense that the "angle" coordinate appears with
retarded terms. At the present time, we are investigating the possibility of
obtaining a better coordinate system that will permit an easier transcription
of results from the finite-dimensional setting to RFDE and NFDE.
11
Periodic solutions of autonomous
equations
The purpose of this chapter is to give a procedure for determining periodic

solutions of some classes of autonomous RFDE.
For equations that are close to linear, the analysis in Chapter 10 can be
effectively applied using the period and amplitude as undetermined param-
eters chosen in such a way as to satisfy the bifurcation equations. In some
applications, it is desirable to determine other parameters so that these
equations are satisfied. Such a situation arises from the case of a Hopf
bifurcation from a constant solution to a nonconstant periodic solution as
some parameter varies. This is discussed in Section 11.1
Sections 11.2 and 11.3 are devoted to general fixed-point theorems that
apply to equations that are not necessarily close to linear. Three types of
equations are used as illustrations of these general results.
11.1 Hopf bifurcation
In this section, we discuss one of the simplest ways in which nonconstant

periodic solutions of autonomous equations can arise--the so-called Hopf
bifurcation. More specifically, we consider a one-parameter family of RFDE
of the form
(1.1) x(t) = F(a,xt)

where F( a,¢) has continuous first and second derivatives in a, ¢ for a E lR,
¢ E C, and F(a, 0) = 0 for all a. Define L: lR x C----+ lRn by
(1.2) L(a)'lj; = D¢F(a, 0)'1/;
where D¢F(a, 0) is the derivative of F(a, ¢)with respect to¢ at¢= 0 and
define
(1.3) f(a, ¢) = F(a, ¢)- L(a)¢.

Additional hypotheses also will be imposed.
332 11. Periodic solutions of autonomous equations
(H1) The linear RFDE(L(O)) has a simple purely imaginary characteristic

root .A0 = iva =f. 0 and all characteristic roots Aj =f. .A 0 , 5. 0 , satisfy
Aj =f. m.A 0 for any integer m.
Since L(a) is continuously differentiable in a, Lemma 10.1 of Section 7.10
implies that there is an a 0 > 0 and a simple characteristic root .A( a) of the
linear RFDE(L(a)) that has a continuous derivative .A'(a) in a for lal < a 0 .
We suppose
(H2) Re A' (0) =f. 0.
We will show that Hypotheses (H1) and (H2) imply there are noncon-
stant periodic solutions of Equation (1.1) for a small that have period close
to 27r / v0 • Before stating the result precisely, we introduce some notation
that will be needed in the proof. The additional notation will also make the
statement of the result more specific.
By taking ao sufficiently small, we may assume Im .A( a) =f. 0 for lal <
ao and obtain a function ¢a. E C that is continuously differentiable in a
and is a basis for the solutions of the RFDE(L(a)) corresponding to .A( a).
The functions
(Re¢a.,Im¢a) ~f Pa
form a corresponding basis for the characteristic roots .A( a), A( a). Similarly,
we obtain a basis lfra. for the adjoint equation with (lfra., Pa.) = I. If we
decompose C by (.A( a), A( a)) as C = Pa. EB Qa., then Pa. is a basis for Pa..
We know that
(1.4)
and the eigenvalues of the 2 x 2 matrix B(a) are .A(a) and A( a). By a change
of coordinates and perhaps redefining the parameter a, we may assume that
B(a) = voBo + aB1(a)

(1.5)
Bo = [ ~1 ~] , B1(a) = [ _ 11(a) "~\a)]
where 'Y(a) is continuously differentiable on 0:::; lal < ao.
We can now state the Hopf bifurcation theorem; we refer to the con-
clusions stated in this theorem as a Hopf bifurcation.
Theorem 1.1. Suppose F( a,¢) has continuous first and second derivatives
with respect to a,¢, F(a, 0) = 0 for all a, and Hypotheses {Hl) and {H2) are
satisfied. Then there are constants ao > 0, ao > 0, 8o > 0, functions a( a) E
IR, w(a) E IR, and an w(a)-periodic function x*(a), with all functions being
continuously differentiable in a for lal < ao, such that x*(a) is a solution
of Equation (1.1) with
(1.6) x0(a)Q" = z0(a),

11.1 Hopf bifurcation 333
where y*(a) =(a, Of+ o(lal), z0(a) = o(lal) as lal ----+ 0. Furthermore, for
lal < ao, lw- (27r/vo)l < 8o, every w-periodic solution of Equation (1.1)
with lxt I < 8o must be of this type except for a translation in phase.
Proof. We prove this result by applying a classical procedure in ordinary
differential equations. Let /3 E [-1, 1], Wo = 271"/ZJo, t = (1 + /3)T, x(t +e)=
u( T + e I (1 + /3))' -r ::::: e ::::: 0, and define Ur,(3 as an element of space
C([-r, 0], IRn) given by Ur,(3(e) = u(T + 8/(1 + /3)), -r:::::
e::::: 0. Equation
(1.1) is then equivalent to the equation
(1.7)
If this equation has an w0 -periodic solution, then Equation (1.1) has a

(1 + /3)w 0 -periodic solution, and conversely.
Let us rewrite Equation (1. 7) as
du(T)
(1. 8 ) ~ = L(O)ur + N(/3, a, Un U ,(3)
7
N(/3, a, Un U7 ,(3) = (1 + /3)L(a)ur,(3- L(O)u7 + (1 + /3)f(a, U 7 ,(3)

This means that we are going to consider Equation (1.8) as a perturbation
of the autonomous linear equation
(1.9)
du(T) _ L( )
dT - 0 U7 .
We know that the columns of U(T) ~f Po(O) exp(B(O)T), T E IR, form a

basis for the w 0 -periodic solutions of Equation (1.9) and the rows of V( T) ~f
exp(- B(O)T )Po (0), T E IR form a basis for the w 0 -periodic solutions of the
formal adjoint equation of Equation (1.9).
We may now apply Corollary 4.1 of Section 6.4 to obtain necessary
and sufficient conditions for the existence of w0 -periodic solutions of Equa-
tion (1.8). In fact, a direct application of that corollary shows that every
w 0 -periodic solution of Equation (1.8), except a translation in phase, is a
solution of the equations
(1.10a) u(T) = U(T)(a, Of+ K(I- J)N(/3, a, u., u.,f3)
(1.10b ), JN(/3,a,u.,u.,(3) = 0,
and conversely. The operators K, J are defined in Corollary 4.1 of Section
6.4.
One can now apply the implicit function theorem to solve Equation
(1.10a) for u = u* (a, /3, a) for a, /3, a in a sufficiently small neighborhood of
zero, u*(a, 0, 0)- U(·)(a, O)T = o(lal) as lal ----+ 0. The function u*(a, /3, a)
is continuously differentiable in a, a from the implicit function theorem.
Since u*(a, /3, a)(t) satisfies Equation (1.10a), it also satisfies a differential
integral equation and is, therefore, continuously differentiable in t. From

Lemma 4.1 of Section 10.4, it follows that u* (a, /3, a) is also continuously
differentiable in f3.
Therefore, all w 0 -periodic solutions of Equation (1.8) are obtained by
finding the solutions a, /3, a of the bifurcation equations
(1.11) JN(j3,a,u~(a,j3,a),u~,f3(a,j3,a)) = 0.
Using the definition of J in Equation (4.12) of Section 6.4, Equation

(1.11) is equivalent to the equation
G(a,f3,a) = 0
where
(1.12) G(a, /3, a) ~f Jo

ra e-B(O)sl[r (0)N(j3, a, u:(a, /3, a), u:,f3(a, a, /3)) ds.
0
From the preceding discussion, it follows that it remains to solve

the equation G(a,f3,a) = 0. Since G(0,/3,a) = 0 for all f3 and a, let
H(a,j3,a) = G(a,/3,a)/a. Noting properties of u*(a,f3,a) and the defi-
nition of G(a,j3,a) in Expression (1.12), one easily observes that
where e 1 = (1, O)T. One may now apply Lemma 10.3 of Section 7.10 directly
to obtain
oH(O,O,a) _ [ 1 ]
oa - Wo -")1(0) .
Furthermore,
Writing this as two separate integrals, changing s into s / ( 1 + /3) in the first
integral, and noting that
dU(s/(1 + /3))
d(s/( 1 + /3)) = L(O)Us/(Hf3),f3,
one sees that
H(O, /3, 0) =
ra e-B(O)slJro(O)cl>o(O)e
f3 Jo 8 (0)s B(O)e1ds.
If x is a solution of Equation ( 1.9) and y is a solution of the adjoint equation,

then (yt, Xt) = constant for all t. Therefore,
11.2 A periodicity theorem 335
I= ( e-B(o)(s+)wo(O), ll>o(O)eB(o)(s+·))
-1: 1()
= e-B(O)sllio(O)IPo(O)eB(O)s
e-B(O)(sH-O)llio(O) d[ry(O)]IPo(O)eB(O)(sH) d~
for all s E JR. Integrating this from 0 to wo and using the fact that the
second integral is zero, one obtains
H(O, (3, 0) = f3woB(O)el.

Combining all of this information, one has
H(O, 0, 0) = 0,
aH [
a((3, a) (0, 0, 0) = wo _ 110
0
Therefore, the implicit function theorem implies the existence of f3(a) and
a( a) such that (3(0) = 0, a(O) = 0, and H(a, (3(a), a(O)) = 0 and the solu-
tion is unique in a neighborhood of zero. The fact that the corresponding
w-periodic function x((1 + (3)t) = u(T) satisfies Equation (1.1) and the
properties stated in the theorem are obvious. 0
11.2 A periodicity theorem
Our objective in this section is to give a general fixed-point theorem that

has been very useful in obtaining periodic solutions of autonomous func-
tional differential equations that are not necessarily perturbations of linear
systems as Section 11.1. Proofs are given only for those properties that
relate directly to the functional differential equations.
Definition 2.1. Suppose X is a Banach space, U is a subset of X, and x is

a given point in U. Given a map A: U\{x}-+ X, the point x E U is said
to be an ejective point of A if there is an open neighborhood G ~ X of x
such that for every y E GnU, y =/=- x, there is an integer m = m(y) such
that Amy tj. GnU.
For any M > 0, we let SM = {x EX: lxl = M}, and BM = {x EX:

lxl < M}. Then SM = aBM·
For references to the following two theorems, which are stated without
proof, see Section 11.7.
Theorem 2.1. If K is a closed, bounded, convex, infinite-dimensional set

in X, A : K\ {x 0 } ---+ K is completely continuous, and x 0 E K is an ejec-
tive point of A, then there is a fixed point of A in K\{x 0 }. If K is finite
dimensional and x 0 is an extreme point of K, then the same conclusion
holds.
Theorem 2.2. If K is a closed convex set in X, A : K\ {0} ---+ K is completely

continuous, 0 E K is an ejective point of A, and there is an M > 0 such
that Ax = AX, X E K n SM implies ).. < 1, then A has a fixed point in
K n B M \ { 0} if either K is infinite dimensional or 0 is an extreme point of
K.
Remark 2.1. Theorems 2.1 and 2.2 remain valid for mappings A that are
a-contractions. This generalization should play a role in studying periodic
solutions of equations with a period smaller than twice the delay, equations
with infinite delays, and certain neutral functional differential equations.
This area has not been exploited very much at the present time.
In the application of Theorems 2.1 and 2.2 to retarded functional dif-

ferential equations, the mapping A is usually similar to the "transversal"
map of Poincare in ordinary differential equations. In fact, one obtains a set
K c::; C such that every solution x(¢), ¢ E K, of the RFDE(f), f: C---+ IRn,
returns to Kin some timeT(¢) > 0; that is, xT(¢)(¢) E Kif¢ E K. The
mapping A : K ---+ K is then defined by A¢ = xT( q,) ( ¢). If A were completely
continuous and K were closed, convex, and bounded, then there would be a
¢ E K such that A¢=¢, and thus, a periodic solution of the RFDE(f). It
looks as if the problem is over, but it is not. We wish to obtain nonconstant
periodic solutions of the RFDE(f), and, if there is a constant a E IRn such
that the constant function a E C defined by a( B) = a, -r ::; B ::; 0, satisfies
a E K, f(a) = 0, then the only fixed point of A in K could be a. If K
contains no such constant functions, then the problem of the existence of
nonconstant periodic solutions is solved. Unfortunately, in the applications,
the construction of such a K is very difficult and, often, the set K contains
only one constant solution x 0 , and it is an extreme point. The theorems
assert that if x 0 is ejective, then there is a nonconstant periodic solution of
the RFDE(f).
From the preceding discussion, it is clear that an efficient method is
needed for determining when a constant solution of an RFDE(f) is ejective
relative to some set K and the mapping A defined earlier. Such a result
will now be given.
Suppose L : C ---+ IRn is linear and continuous, f : C ---+ IRn is com-
pletely continuous together with a continuous derivative f' and f(O) = 0,
f' (0) = 0. Consider the equations
(2.1)
11.2 A periodicity theorem 337
(2.2) y(t) = LYt·
For any characteristic root A of Equation (2.2), there is a decomposition

of Cas C = P>. EB Q>.., where P>. and Q>. are invariant under the solution
operator TL(t) of Equation (2.2), TL(t)¢ = Yt(¢), ¢ E C. Let the projection
operators defined by the decomposition of C be 1!">., I - 1!">. with the range
of 1!">. equal to P>..
Theorem 2.3. Suppose the following conditions are fulfilled:

(i) There is characteristic root A of Equation (2.2) satisfying ReA > 0.
(ii) There is a closed convex set K ~ C, 0 E K, and 8 > 0, such that
v = v(8) ~f inf{ln>.¢1 : ¢ E K, 1¢1 = 8} > 0.
(iii) There is a completely continuous function T: K\{0}-+ [a, oo), 0 :<::;a

such that the map defined by
(2.3) ¢ E K\{0}
takes K\{0} into K and is completely continuous.
Then 0 is an ejective point of A.
Proof. If P>. = (iP 1 >., ... , iP)d>.) is a basis for P>. and 7r>.P = iPb, then b =
b( ¢) E lRd is a continuous d- vector linear functional on lRd and we take the
norm of b to be the Euclidean norm. For any continuous V : C -+ lR, let
V(¢) = liminf ~ [V(xt(¢))- V(¢)].

t--+O+ t
Exactly as in the proof of Lemma 1.3 of Section 10.1, one can prove the
following
Lemma 2.1. There is a positive definition quadratic form V(¢) = bT Bb

with the property that for any p > 0, there is a 80 > 0 such that for any 8,
0 < 8 < 8o, V(¢) > 0 if V (¢) ?:_ p 2 82 , ¢ E fh.
With the function V(¢) in Lemma 2.1, observe that Condition (ii)
implies that
v2 ~f inf{V(¢/1¢1) : ¢ E K, 1¢1 =/= 0} > 0
since there are a> 0, f3 > 0, such that ln>.¢1 2 :<: ; aiW :<: ; af3- 2 V(¢).
Fix p < v and choose 8o as in Lemma 2.1. Lemma 2.1 implies V(xt(¢))
is increasing in t as long as V(xt(¢)) ?:_ p 2 82 and lxt(¢)1 :<: ; 8, 0 < 8 < 80 . If
then G(8) is an open set, 0 E G.

Since T is completely continuous, there is a k > 0 such that T( ¢) ::; k

for all ¢ E G(8o) n K. Choose 0 < 81 < 80 sufficiently small so that the
solution x(¢) through ¢ E Cl G(81) satisfies Xt(<P) E G(8o) for 0 ::; t ::; k.
For any <P E G(8o) n K, 1¢1 = E, then the fact that p < v implies
Consequently, Lemma 2.1 implies V(¢) is increasing along the solutions of

Equation (2.1) for 1¢1 < 8o, ¢ E K\{0}. Therefore, for any¢ E G(8 1 ) n K,
¢ -1- 0, there are h > t2 > 0 such that Xt(<P) E G(8I), 0 ::; t < t 2,
Xt(<P) E G(8o)\G(81), t2 ::; t < t1 and xd¢) E 8G(8o). Let no(¢) be
the least integer such that Ak (¢) E { Xt( ¢ ), 0 ::; t ::; t 2 }, k = 0, 1, ... , n 0(¢ ).
Since t 1-t2 2: k by the choice of 81, it follows that Ano(</>)+ 1(¢) tJ_ G(81)nK.
This proves 0 is an ejective point of A. The proof of the theorem is complete.
D
We are now in a position to state a useful tool for obtaining periodic

solutions of an RFDE by combining Theorems 2.1, 2.2, and 2.3.
Theorem 2.4. Suppose K is a closed convex set in C, 0 E K, such that

Conditions (i)-(iii) of Theorem 2.3 are satisfied. If either of the conditions
(iv') K is bounded and infinite-dimensional,
(iv") K is bounded, finite-dimensional, and 0 is an extreme point of K,
(iv"') there is an M > 0 such that Ax= AX, x E K n SM implies A < 1;
hold, then there is a nonzero periodic solution of Equation (2.1) with initial
value in K\{0}.
11.3 Range of the period
In this section, we return to the discussion of the one-parameter family of

RFDE considered in Section 11.1. More specifically, suppose the hypotheses
of Theorem 1.1 are satisfied so that a Hopf bifurcation occurs. We will
show how to determine some information about the range of the periods of
Equation (1.1) as a is varied by combining the Hopf bifurcation with some
results on fixed points of mappings that depend on a parameter a.
Definition 3.1. Suppose X is a Banach space, x 0 E X is given, U <:;;; X is an

open neighborhood of x 0 , and A: U - t X is a continuous map. The point
x 0 is said to be a uniform attractive point for A if there is neighborhood Uo
of x 0 , Uo <:;;; U, such that for any neighborhood V of xo, there is an integer
m(V) such that AJ(U0 ) <:;;; U for j 2: 0, AJ(Uo) <:;;; V, j 2: m(V).
11.3 Range of the period 339
For references to the following theorem, which is stated without proof,

see Section 11.7.
Theorem 3.1. Suppose K is a closed convex set of a Banach space X, 0 is

an extreme point of K, {0} =f. K, and J =(a, oo), -oo <::::a< oo. Suppose
A : K x J -+ K satisfies the following hypotheses:
(i) A(O,a) = 0 for a E J.
(ii) A is completely continuous.
(iii) There is an a 0 E J such that for any compact interval Jo C J, ao ~ Jo,
there is an E = E(J0) > 0 such that A(x, a) =f. x for a E Jo, 0 < lxl <::::E.
(iv) There is an open interval I 0 about a 0 such that 0 is a uniform attractive
point of A(·, a) for a E Io, a< ao, and 0 is an ejective point of A(-, a)
for a E Io, a > a 0 , or conversely.
(v) There is an a 1 E J such that A(x, a) = x for x =f. 0 implies a;::: a1.
If the hypotheses are satisfied and S <;;; K x J is defined by
S = Cl{(x,a) E K x J: x =f. O,A(x,a) = x},

then S 0 , the maximal closed connected component of S that contains (0, a 0 )
is unbounded.
We now discuss the way in which one can apply Theorem 3.1 to the
RFDE in Section 11.1. For System (1.1), suppose the following conditions
are satisfied:
(3.1) There is a closed convex set K <;;; C and a < 0 such that 0 is an
extreme point of K, {0} =f. K, and, for any ¢ E K, ¢ =f. 0 and
a E (a, oo), there is a completely continuous function T( ¢,a) > 0
such that the solution x( ¢,a) of Equation (1.1) through ¢ satisfies
X 7 (¢,a)(¢,a) E K.
(3.2) If A( a)¢= X 7 (¢,aJ(¢, a),¢ E K, ¢=f. 0, A(a)O = 0, a E (a, oo), then

A( a) : K -+ C is completely continuous.
Hypotheses (3.1) and (3.2) imply Hypotheses (i) and (ii) of Theorem 3.1 are
satisfied. For the linear RFDE(L(O)) with L(a) defined in Equation (1.2),
suppose
(3.3) the linear RFDE(L(O)) has a simple purely imaginary characteristic
root Ao = iva =f. 0 and all characteristic roots Aj =f. Ao, 5.o satisfy
Re>-1 < 0.
With Hypothesis (3.3), Lemma 10.1 of Section 7.10 implies there is an
a 2 > 0 and a simple characteristic root >.(a) of the linear RFDE(L(a)),
which has a continuous first derivative X(a) in a for lal < a 2 • Suppose
(3.4) Re >.' (0) =f. 0.

Let 1f>.(a) be the projection operator defined by the decomposition of C as

C = P.\(a) EB Q>.(a) with the range of 1f>.(a) equal to P.\(a), the generalized
eigenspace of .A(a). Suppose
(3.5) for any compact interval J 0 <;;; (a,oo), there is aD> 0 such that
v ~f inf{l1f>.(a)cPI : ¢ E K, 1¢1 = D, a E Jo} > 0.
If Hypotheses (3.3), (3.4), and (3.5) are satisfied, then Theorem 1.1 of
Section 10.1 and Theorem 2.3 imply that Condition (iv) of Theorem 3.1
is satisfied. Furthermore, the proof of Theorem 2.3 and the uniformity of
Hypothesis (3.5) in a imply that Condition (iii) is satisfied.
If Conditions (3.4) and (3.5) are satisfied, then Theorem 1.1 implies
there are b0 > 0, a2 > 0, Do > 0 and functions a(b) E (a, oo), w(b) E IR,
¢*(b) E C, lbl < bo, a(O) = 0, w(O) = 21flvo, ¢*(0) = 0, such that each
function is continuous and continuously differentiable in b, the solution
x*(¢*(b),a(b)) of Equation (1.1) for a= a(b) through ¢*(b) is an w(b)-
periodic solution. Furthermore, any periodic solution of Equation (1.1) for
lal < ao, lxl < Do, must be of the preceding form. We can now prove the
following.
Theorem 3.2. Suppose System (1.1) satisfies Hypotheses (3.1)-(3.5) and

(3.6) there exists a 1 E (a,oo) such that A(¢,a) =¢for¢-:/= 0 implies
a 2:: a1,
then S 0 is unbounded where S 0 is defined in Theorem 3.1 with a 0 = 0.
Also, if
(3.7) (¢*(b), a(b)) E S for lbl < bo
then the period w(¢, a) of each periodic solution x(¢, a) of Equation (1.1),
with initial value¢ satisfying(¢, a) E So, is a continuous function of a.
Proof. With Conditions (3.1)-(3.6) satisfied, we have already shown that
all conditions of Theorem 3.1 are satisfied. Therefore, S 0 is unbounded. If
Condition (3. 7) is satisfied, then in a sufficiently small neighborhood U of
(0, 0) E S 0 , the only periodic solutions of Equation (1.1) with initial value
¢ satisfying (¢,a) E U must have ¢ = ¢*(b), a = a(b). Therefore, the
period w(¢, a) is continuous at¢= 0, a= 0. For other values of(¢, a) the
continuity follows because we have assumed T(¢, a) is continuous. D
Theorem 3.2 shows that the range of the period won So is an interval
containing 27r I v 0 . If one knows that there exist (¢, a) E So for each a E
[O,oo) and w00 = limsupw(¢,a) as a---+ oo, then w(So) :2 [27rlvo,woo) if
Woo > 27f Iva and w(So) :2 (woo, 27f Iva] if Woo < 27f Iva.
11.4 The equation ±(t) = -ax(t- 1)[1 + x(t)] 341
11.4 The equation x(t) = -ax(t- 1)[1 + x(t)]

In this section, we apply the results of the previous sections to the scalar
equation
(4.1) i:(t) = -ax(t- 1)[1 + x(t)]
where a > 0. Let us first show that a Hopf bifurcation occurs at a = 1r /2.
To do this, we will apply Theorem 1.1 and, therefore, we need information
about the behavior of the zeros of the characteristic polynomial
(4.2) .Xe>- =-a
of the linear part of Equation (4.1),
(4.3) y(t) = -ay(t- 1).
This information is contained in the following lemma.
Lemma 4.1. If 0 < a < 1r /2, every solution of Equation (4.2) has a negative
real part. If a > e-1, there is a root .X(a) = f'(a) + ia(a) of Equation
(4.2), which is continuous together with its first derivative in a and satisfies
0 < a(a) < 1r, a(w/2) = w/2, 'Y(w/2) = 0, 'Y'(w/2) > 0, and f'(a) > 0 for
a> w/2.
Proof. The assertion concerning the real parts for 0 < a < 1r /2 is in The-
orem A.5 of the Appendix. To prove the remaining part of the lemma,
let p(J.t) = -J.teiL. Then p'(J.t) = -(1 + J.t)e!L and, therefore, p'(J.t) > 0,
-oo < J.t < -1, p'(-1) = 0, p'(J.t) < 0, -1 < J.t < oo. Consequently, p(J.t)
has a maximum at J.t = -1, p( -1) = e- 1 . Therefore, Equation (4.2) has no
real roots for a > e- 1 . If a > e-I, .X = /' + ia, J.t = -!', satisfies Equation
(4.2), then J.t- ia = aexp(J.t- ia) and
J.t = aeiL cosw, a= aeiL sin a
or
ae-"" cot"" def
J.t =a cot a, a= . = f(a).
sma
Let us consider f(a) for 0 <a< 1r. It is clear that f(a) > 0.
f'(a) 1 2 2 (1-acota) 2 +a 2
f(a) =-;;- cot a+ a cosec a= a > 0.
Furthermore, f(a) ----t oo as a ----t 1r and f(a) ----t e- 1 as a----t 0. Therefore,
there is exactly one value of a say a 0 = a 0 (a), 0 < a 0 (a) < 1r, for which
f(ao(a)) =a if a> e- 1 . Let 'Yo(a) = -a0 (a)cota0 (a). The functions
a 0 (a) and 'Yo(a) are clearly differentiable in a.
Also, f(w/2) = w/2, 'Yo(w/2) = 0, and 'Yo(a) > 0 if a> w/2. From the
equation .X( a) exp.X(a) =-a, one observes that 'Y'(w/2) > 0. The lemma
is proved. D
Using Lemma 4.1 and Theorem 1.1, we may now state

Theorem 4.1. Equation (4.1) has a Hopf bifurcation at a= 1rj2.
Our next objective is to show that Equation (4.1) has a nonzero peri-
odic solution for every a > 1r /2.
Let x( ¢,a) be the solution of Equation (4.1) through ¢. It is clear that
x(cjJ,a)(t) > -1, t 2': 0, if ¢(0) > -1. Also, it is clear that there is no t 0 > 0
such that x(¢, a)(t) = 0 fort 2': t 0 unless¢= 0. We say the zeros of x(¢, a)
are bounded if x(¢, a)(t) has only a finite number of positive zeros.
Lemma 4.2.
(i) If ¢(0) > -1 and the zeros of x(¢, a) are bounded, then x(¢, a)(t)---+ 0
as t ---+ oo .
(ii) If ¢(0) > -1, then x(¢, a)(t) is bounded. Furthermore, if the zeros of
x( ¢, a) are unbounded, then any maximum of x( ¢,a) (t), t > 0, is less
thane"- 1.
(iii) If ¢(0) > -1 and a > 1, then the zeros of x( ¢,a) are unbounded.
(iv) If cjJ(B) > 0, -1 -1, ¢(8) < 0, -1 0 such that
x(t) ~f x(cjJ,a)(t)
is of constant sign fort 2': h -1. Since x(t) > -1 for all t 2': 0, x(t)x(t-1) <
0, t 2': h. Therefore, x(t) is bounded and approaches a limit monotonically.
This implies x(t) is bounded and therefore x(t)---+ 0 as t---+ oo. This implies
x(t)---+ 0 or -1, but -1 is obviously excluded.
(ii) x satisfies
(4.4) 1 + x(t) = [1 + x(t0 )] exp [-a lt- x(~)d~J

1
to-1
for any t 2': t 0 2': 0. If the zeros of x(t) are bounded, then Part (i) implies x
is bounded. If there is a sequence of nonoverlapping intervals h of [0, oo)
such that x is zero at the endpoints of each h and has constant sign on h,
then there is a tk such that x(tk) = 0. Thus, x(tk- 1) = 0. Consequently,
Equation (4.4) implies for to = tk - 1, t = tk.
since x(t) > -1, t 2': 0. Finally, x(tk) :S: e"- 1 for all tk. This proves Part
(ii)'
(iii) If the zeros of x are bounded, then Part (i) implies x( t) ---+ 0 as
t ---+ oo and, thus, the existence of a t 0 > 0 such that a(1 + x(t)) > 1 for
t 2': t 0 and x (t) has constant sign for t 2': t 0 . Thus
x(t)x(t- 1) = -o:x 2 (t- 1)(1 + x(t)) < -x 2 (t- 1) < 0, t::::: to+ 1,
and x(t)-+ 0 monotonically as t-+ oo. If xis positive on [t0 , oo), then
x(t 0 + 3)- x(t0 + 2) = 1 to+3
t 0 +2
x(t) dt <
1to+2
t 0 +1
x(t) dt < -x(to + 2)
and x(t 0 + 3) < 0. This is a contradiction. If x(t) is negative on [t 0 ,oo),

then a similar contradiction is obtained.
(iv) Suppose x(t 0 ) = 0 and x(t) > 0, t 0 -1 < t < t 0 . For t 0 < t < t 0 + 1,
x(t) < 0. Similarly, if x(t) < 0 for t 0 - 1 < t < t 0 and x(t 0 ) = 0, then
x(t) > 0, t 0 < t < t 0 + 1. Thus, the assertions of (iv) are obvious and the
lemma is proved. 0
Let K be the class of all functions ¢> E C such that ¢>(B) ::::: 0, -1 <
() :::; 0, ¢>( -1) = 0, ¢> nondecreasing. Then K is a cone; that is, K is a closed
convex set in C with the following properties: if ¢> E K, then >..¢> E K for
all>..> 0 and if¢> E K, ¢> i= 0, then-¢> tJ_ K. If a:> 1, ¢> E K, ¢> i= 0, let
z(¢>, a:)= min{t: x(¢>, o:)(t) = 0, x(¢>, o:)(t) > 0}.

This minimum exists from Lemma 4.2, Parts (iii) and (iv). Also z(¢>, a:) >
2. Furthermore, Lemma 4.2, Part (iv) implies x(¢>, o:)(t) is positive and
nondecreasing on (z( ¢>,a:), z( ¢>,a:)+ 1). Consequently, if r( ¢>,a:) = z( ¢>,a:)+ 1,
then the mapping
A(o:)O=O
A( a:)¢>= X7 (¢,o:)(¢>, a:),
is a mapping of the cone K into itself. Since x(¢>, o:)(r(¢>, a:) - 1) > 0,
continuity of x(¢>,o:)(t) in t,¢>,o: implies that r(¢>,o:) is continuous in
K\ {0} X (1, oo ).
Lemma 4.3. The map T: (K\{0}) x (1, oo) -+ (0, oo) defined by r(¢>, a:) =
z( ¢>,a:) + 1 is completely continuous.
Proof. First of all, we claim a solution x = x(¢>, a:), ¢> E K, cannot take a
time longer than 2 to become negative because, if x(1) = rJ > 0, we have
1 + x(2) = (1 + ry) exp [-a: fo 1

x(s) ds] :::; (1 + ry)e-o:ry
and so x(2) :::; (1 + ry)e-o:7J- 1, and this quantity is negative because the
function h(ry) = (1 +ry)e-o:'7 -1 satisfies h(O) = 0 and h'(ry) =(-a:- o:ry+
1)e-o:'7 < 0 for rJ > 0.
For any bounded set B ~ K and any 'ljJ E B, a: E (1, oo), let t 0 (¢>, a:):::;
3 denote the point where the solution x = x(¢>, a:) has a minimum. Since
to(¢>, a:)::::: 1, the set H(o:) = Cl U¢EB Xto(</>,o:)(¢>, a:) is compact and
H(a) <;;; K 1 ~ N E C: -1 < '¢(0) ~ 0,-1 ~ 0 ~ 0, '¢ nonincreasing}.

For any '¢ E K1, define the continuous function
71: [K1\{0}] X (1,oo)----+ (O,oo)
by the relation 71('¢, a) = min{t > 0 : x('¢, a)(t) = 0}. If we prove

71(H(a)\{O}, a) is bounded for each a E (1, oo), then 7(B\{O}, a) is
bounded for each a. Since H(a) is compact, it is therefore only neces-
sary to prove that 71 is bounded on a neighborhood of zero in K1. This is
easy to verify in the following manner. If '¢ E K 1\ {0} and 71 ('¢, a) > 1,
then x('¢,a)(1) = (3 < 0 and
1 + x('¢, a)(2) = (1 + (3) exp [-a 1 1

x(s, '¢,a) ds] ~ (1 + (3)e-af3.
So
x('¢, a)(2) ~ h((J) ~f (1 + (3)e-af3- 1.
If (3 is small and negative, then this function is positive since h(O) = 0 and
h'(O) = 1- a< 0. This shows that 71 is bounded.
Since 7(¢,a) is continuous for (¢,a) E [K\{0}] x (1,oo), and IRis
locally compact, one obtains the conclusion stated in the lemma. 0
From Parts (ii) and (iv) of Lemma 4.2, it follows that IA(a)¢1 ~ e" -1
for each cjJ E K and A(a) takes any bounded set B in K\{0} into the set
{¢ E C: 1¢1 < e" -1}. Since 7: [K\{0}] x (1,oo)----+ (O,oo) is completely
continuous and 7(¢, a) > 1, it follows that the closure of A(a)B is compact.
Also, if cPk E K\{0}, cPk----+ 0, then we may assume 7(¢k,a)----+ 7o(a) ask----+
oo. Continuity of solutions x(¢, a)(t) of Equation (4.1) with respect tot, cjJ
implies Xr(c/>k,a) (¢k, a) ----+ X 70 (a)(O, a) = 0. Therefore, A(a) is continuous at
0 and A(a) is completely continuous.
Let >.(a) be the root of Equation (4.2) given in Lemma 4.1, let C be
decomposed as C = P>.(a) E9 Q>.(a) in the usual manner, and let 7r>.(a) be
the usual projection on P>.(a)·
Lemma 4.4. If Jo is a compact set of (1, oo) then
J.L = inf{ln>.(a)¢1, cjJ E K, 1¢1 = 1, a E Jo} > 0.
Proof. Let).= >.(a) be the solution of Equation (4.2) given by Lemma 4.1,
¢(0) = e>.e /(1 + >.), -1 ~ 0 ~ 0, 'lf;(s) = e->.s, 0 ~ s ~ 1, iP = (¢,¢),
lJ! =('¢,if;). The formal adjoint of Equation (4.3) is
i(t) = az(t + 1)
and the bilinear form is
It is easily seen that (lft, P) is the identity. Therefore, for any¢ E C, 7r>.¢ =
P(lft, ¢). To show the conclusion of the lemma is true, it is therefore sufficient
to show that
inf{l(lft, ¢)1, ¢ E K, 1¢1 = 1, a E Jo} > 0.
Since ('¢, ¢), (i[J, ¢) are complex conjugate, it is sufficient to look at ('¢, ¢).
If¢ E K, 1¢1 = 1, then ¢(0) = 1 and
('¢,¢,a) ~f ('¢, ¢) = R(¢) + iJ(¢)

0
R(¢) = 1- o: /_ 1 ¢(B)e-"'(B+l) cosa(B + 1) d(}
I(¢) =a /_ 1 ¢(B)e-"1(0+ 1 ) sina(B + 1)dB

0
Since J0 is compact, there is an E > 0 such that E <a= a( a) < 1r- E for
a E Jo.
Now suppose there exist sequences ¢n E K, ¢n(O) = 1, an E J 0 such
that ('¢, ¢n, O:n) --+ 0 as n--+ oo. We may assume O:n --+ {3 as n--+ oo since
Jo is compact. Since I(¢n) --+ 0, we have ¢n(B) --+ 0, -1 ::; (} ::; 0. Thus,
R(¢n) --+ 1 as n--+ oo. This is a contradiction to the fact that R(¢n) --+ 0
as n --+ oo and the lemma is proved. D
If we now take M > e"' - 1, then the lemmas imply that all of the
conditions of Theorems 2.2 and 2.3 are satisfied. Therefore, we have proved
the following result.
Theorem 4.2. If o: > 1rj2, Equation (4.1) has a nonzero periodic solution.
The lemmas also contain more information. Let

S = Cl{(¢, o:) E K x (1, oo): A( a:)¢=¢,¢ =f 0}
(4.5) 8 0 =maximal closed connected component of S that
contains (0,7r/2).
Since the root A( a:)= I'( a:)+ ia(o:) of Equation (4.2) given by Lemma 4.1
satisfies a( 1r /2) = 1r /2 and the solution x* obtained by the Hop£ bifurcation
Theorem 1.1 can be chosen (by a phase shift) to satisfy x*(t) = cos7r(t/2)
for a = 1r /2, it is clear that the initial value of Hop£ bifurcating solution
belongs to K for o: close to 1r /2. With this remark and the lemmas we have
shown that all conditions of Theorem 3.2 are satisfied except Hypothesis
(3.6). The next lemma shows this condition is also satisfied.
Lemma 4.5. There is an a 1 > 1 such that for any a E (1, ar), the only
solution of A( a)¢=¢ inK is¢= 0.
Proof. Suppose¢ E K, A( a)¢=¢, and let z1(¢, a)= z1 and z2(¢, a)= z2
be the first and second zeros of the periodic solution x( ¢, a) of Equation
(4.1). From Equation (4.4) and the fact that 1¢1 = 1¢(0)1 = lx(z2 + 1)1,
x(~) 2': x(z 1 + 1) for~ E [z2- 1, z2], we have
-x(z 1 + 1) = 1- exp [-a 1z 1
Zl-1
x(~) d~J ::=; 1- e-al¢1 ~f 'Y(I¢1)
1¢1 = x(z2 + 1) = exp[-a 1:~ 1 x(~) d~] -1
:::; e-ax(zl +1) - 1 :::; e<>f'(l¢1) - 1.
If f(/3) = exp( +a'Y(/3)) - 1 - (3, then there is a unique solution (3( a) of

f(/3) = 0 for each a > 1 and (3(a) ----* 0 as a ----+ 1. Furthermore, f(/3) < 0
if (3 > f3(a). Therefore, if there is a¢ such that A(a)¢ = ¢, ¢ E K, then
I<PI :::; (3( ¢ ). Since the zero solution of Equation (4.1) is asymptotically stable
for 0 < a < 1r /2, there are 8 > 0 and E > 0, such that the only periodic
solution of Equation (4.1) with a E (1-E, 1+E), 1¢1 < 8, is¢= 0. Therefore,
if a 1 is chosen such that (3(al) < 8, the lemma is proved. D
The following result is now an immediate consequence of Theorem 3.2

and the fact that fixed points of A(a) satisfy 1¢1 < e<>- 1.
Theorem 4.3. If So is defined as in Expression (4.5), then S 0 is unbounded

and for any a 2 > 1, there is an a> a 2 and¢ E K such that(¢, a) E S 0 .
As described in Section 11.3, Theorem 3.2 together with Theorem 1.1

can now be used to obtain an estimate of the range of the period w(¢, a)
of periodic solutions of Equation (4.1) with initial values ¢ in K. Since we
know that S0 contains an element for each a E [1rj2, oo), and w(O, K/2) = 4
for a= 1rj2, we need only estimate a= limsupw(¢,a) as a----+ oo.
The following is stated without proof and is obtained from Theorem
3.2 and estimates showing that W 00 = limsup{w(¢, a) : (¢,a) E S 0 , a ----+
oo} = oo (see Section 11.7 for references).
Theorem 4.4. For any p > 4, there is a periodic solution of Equation (4.1)
of period p.
11.5 The equation x(t) = -ax(t- 1)[1- x 2 (t)] 347
11.5 The equation x(t) = -ax(t- 1)[1- x 2 (t)]

In this section, we consider the equation
(5.1) x(t) = -ax(t- 1)[1- x 2 (t)]

where a > 0. Our purpose is only to show the modifications of the previous
section necessary to obtain a nonconstant periodic solution of Equation
(5.1).
Since the linear part of this equation is the same as in the previous
section, the following result holds.
Theorem 5.1. Equation (5.1) has a Hopf bifurcation at a= rr/2.
If x(¢, a) is the solution of Equation (5.1), then -1 < x(¢, a)(t) < 1
fort 2:0 if -1 < ¢(0) < 1. Also, there is a t 0 > 0 such that x(¢,a)(t) = 0
fort 2: t 0 only if¢= 0. The analogue of Lemma 4.2 is
Lemma 5.1.
(i) If -1 < ¢(0) < 1 and the zeros of x(¢, a)(t) are bounded, then
x(¢, a)(t)----+ 0 as t----+ oo.
(ii) If -1 < ¢(0) < 1, then -1 :::; x(¢, a)(t) :::; 1, t 2: 0, and if the zeros
of x( ¢,a) are unbounded, then any maximum [or minimum] of x( t),
t > 0, is less than (e 2 "' -1)/(e2"'+ 1) [greater than -(e2 "' -1)/(e 2"'+ 1)].
(iii) If -1 < ¢(0) < 1 and a> 1, then the zeros of x(¢, a) are unbounded.
(iv) If 1 > ¢(0) > 0, -1 < 0 < 0 [or if ¢(0) > -1, ¢(0) < 0, -1 < 0 < 0],
then the zeros (if any) ofx(¢,a)(t) are simple and the distance from
a zero of x(¢, a)(t) to the next maximum or minimum is 2: 1.
Proof. The proof of Parts (ii), (iii) and (iv) are the same as the proof in
Lemma 4.2 except for obvious modifications. To prove Part (ii), observe
that x = x(¢, a) satisfies
ln 1 + x(t) -ln 1 + x(to) = -2a

1 - x(t) 1- x(to)
1t-l x(~) d~
to-1
for any t 2: t 0 2: 1. Using the same argument as in Lemma 4.2. Part (iii),
for this equation, one proves Part (ii). D
Let
K={¢EC:¢(-1)=0, O:S¢(0)<1, -1:::;0:::;0,
¢ nondecreasing}.
Then K is a truncated cone; that is, K is the intersection of a cone with a
ball with center zero. Define the operator A : K ----+ K as in Section 11.4.
As before, one shows that r and A are completely continuous mappings.

Lemma 4.4 is true forK.
From Lemma 5.1, Part (ii),
IA¢1 -< (e2a- 1) ~f /3 < 1

(e2a + 1)
for all¢ E K. If we choose M > /3, then the conditions of Theorem 2.2 and
2.3 are satisfied. Therefore, we have proved
Theorem 5.2. If a> 7f/2, then Equation (5.1) has a nonconstant periodic
solution.
One can use the same type of arguments as in Section 11.4 to prove
the analogue of Lemma 4.5 and obtain
Theorem 5.3. So is unbounded and for any a 2 E (1, oo), there are a > a2
and¢ E K such that(¢, a) E So.
Some comments are made in Section 11.7, concerning the range of the
period of the periodic solutions of Equation (5.1).
11.6 The equation x(t) + f(x(t))x(t) + g(x(t- r)) = 0

In this section, we discuss the existence of nonconstant periodic solutions
of the equation
(6.1) x(t) + f(x(t))x(t) + g(x(t- r)) = 0

where r ;:::: 0, f is continuous, and g is continuous together with its first
derivative, f(O) = -k, g'(O) = 1.
As in the previous sections, we first discuss the Hopf bifurcation. To
do this we must consider the characteristic equation for the linear part of
Equation (6.1); namely, the equation
(6.2)
Lemma 6.1. Let ao (r), 0 < ao (r) < 7f /2r, be the unique solution of a 2 =
cosar, and let ko(r) = [ao(r)]- 1 sinao(r)r. If k < -ko(r), then all roots of
Equation (6.2) have negative real parts. There is an E > 0 and a root .A(k)
of Equation (6.2) that is continuous together with its first derivative in k
fork E ( -ko(r)- E, -ko(r) +E), .A( -ko(r)) = iao(r), Re A'( -ko(r)) > 0.
Finally, for each k > -k0 (r), there are precisely two roots ).. of Equation
(6.2) with Re).. > 0 and -1rjr < Im).. < 1rjr.
11.6 The equation x(t) + f(x(t))x(t) + g(x(t- r)) = 0 349
Proof. Suppose,\= p,+ia is a solution of Equation (6.2). Separating the real

and imaginary parts of this equation, one obtains the following equations
for p,,a:
p, 2 - a 2 - kp, + e-J.Lr cos ar = 0

(6.3)
2ap, - ka- e-J.Lr sin ar = 0.
It is shown in Theorem A.6 of the Appendix that all roots of Equation (6.3)
have negative real parts if k < -k0 (r), where k0 (r) is specified as in the
lemma. If k = -k0 (r), then J.L = 0 and a= ao(r) are solutions of Equations
(6.3). Also, the implicit function theorem implies there is an E > 0 and
a unique solution p,(k), a(k), p,(-ko(r)) = 0, a(-ko(r)) = ao(r), fork E
( -k0 (r)- E, -k0 (r) +E), which is continuous and continuously differentiable
in k. Also, it is easy to compute the derivatives of these functions at k =
-k0 (r) and observe that p,'(-k0 (r)) > 0. Thus, all of the lemma is proved
except the last assertion.
Let Uab = {A E C : a :::; Re A :::; b, lim .\I < 1r jr }. The application
of the implicit function theorem and the fact that all roots of Equations
(6.2) have negative real parts fork < -k0 (r) show there is a k1 > -k0 (r)
such that Equation (6.2) with k E [-k 0 (r), k1] has exactly two roots in the
region Uo,=· Now suppose k ?: k1. Our first observation is that there is no
solution of Equation (6.2) with either,\ = J.L + (in/r), J.L ?: 0 or A = ia,
0:::; a:::; njr. The last assertion is obvious from Equations (6.3) since we
would have a= a 0 (r) and k = -k0 (r), which is a contradiction to the fact
that k?: k1 > -k0 (r). If A= J.L + (injr), then Equations (6.3) imply that
J.L = k/2 and
0 = p, 2 - (~) 2 - kp,- exp( -rp,) = -k

2
- (~) 2 - exp( -rp,) < 0
r 4 r
which is a contradiction.
For any fixed k2 E IR, there is a real number b(k2 ) > k1 such that there
are no solutions of Equation (6.2) with Re.\?: b(k2 ) for any k E [k1,k 2 ].
Therefore, for any k E [k1, k2 ], there are no solutions of Equation (6.2) on
the boundary of Uo,b(k 2 )· Let ft(A) = .\2 - [(1- t)k1 + tk].\ + exp( -r.\),
0 :::; t :::; 1, k E [k1, k2 ]. Since (1- t)k1 + tk E [k1, k2 ] for t E [0, 1], and
k E [k1, k2 ], it follows from Rouche's theorem that the number of zeros of
ft(A) in Uo,b(k 2 ) is constant fortE [0, 1]. In particular, the number of zeros
of fi(.\) (Equation (6.2) for k1) in Uo,b(k 2 ) is the same as the number of zeros
of fo(.\) (Equation (6.2) for kl) in Uo,b(k 2 )· Since we have already observed
this latter number is precisely two, the proof of the lemma is complete. 0
Using Lemma 6.1 and Theorem 1.1, we may now state
Theorem 6.1. Equation (6.1) has a Hopf bifurcations at k = -k0 (r), where
ko(r) is defined in Lemma (6.1).
Under some additional hypotheses on f and g in Equation (6.1), we

will prove that Equation (6.1) has a nonconstant periodic solution for every
k > -ko(r).
The additional hypotheses on f and g are the following:
(6.4a) F(x) =fox f(s)ds is odd in x.
(6.4b) F(x) -too as jxj -too and there is a f3 > 0 such that F(x) > 0 and
is monotone increasing for x > f3.
(6.4c) g'(x) > 0, xg(x) > 0, x =f 0, g(x) = -g( -x), g'(O) = 1.
(6.4d) g(F- 1 (x))/x- 0, F- 1 (x)/x- 0 as x - oo.
In this notation, Equation (6.1) is equivalent to the system of equations
±(t) = y(t)- F(x(t)),

(6.5)
y(t) = -g(x(t- r)).
Let Co= C([-r,O],IR) x IR and designate elements in C 0 by 'ljJ =(¢,a),

¢> E C([-r, 0], IR), a E JR. For any 'ljJ E C0 • Equation (6.5) has a unique
solution z('lj;), z = (x, y), through 'ljJ at zero. In the following, the symbol
Zt E Co shall designate a solution of Equation (6.5) and Zt = (xt, y(t)).
Also, z(t) will designate z(t) = (x(t), y(t)), and 'lj;(O) = (¢(0), a) if 'ljJ E C0 .
Let
K = {'1/J =(¢,a) E Co: 0 ~a< oo,O = ¢>(-r) ~ ¢(0), -r ~ () ~ 0,

¢(B) nondecreasing in B}.
Lemma 6.2. If Hypotheses (6.4) are satisfied, then the following assertions
hold:
(i) There is a continuous r 1 : K\{0}- (r,oo) such that
(ii) There is a continuous r2: K\{0}- (r,oo) such that Z72 (..p)('t/J) E K.
(iii) For any 'ljJ E K\{0}, the solution z('lj;) of Equations (6.5) is oscillatory;
that is, both x('lj;)(t) and y('lj;)(t) have infinitely many zeros.
Proof. If 'ljJ E K and z('lj;) is a solution of Equations (6.5), then -z('lj;) =

z( -'1/J) is also a solution. Therefore Statement (i) implies Statement (ii).
Since Statements (i) and (ii) imply Statement (iii), it is only necessary to
prove Statement (i).
If 'ljJ E K and z = z('lj;) is the solution of Equations (6.5) through '1/J,
we analyze the curve in the (x, y)-plane traced out by z(t), t 2::: 0. Let
r = {(x,y) E IR2 : y = F(x),x E IR}.

'Jf(O) r
-~
------------- -------- -------- -------------X
'
'
'
'
'
'
'
Fig. 11.1.
The accompanying Figure 11.1 will be helpful in understanding the follow-

ing argument.
Suppose 7f(O) is above r. As long as z(t) is above r, x(t) = y(t) -
F(x(t)) > 0 and y(t) = -g(x(t- r))::::; 0. If z(t) does not intersect r,
then
x(t) 2': x(r) > 0 fort 2': rand y(t) = -g(x(t-r)) = g( -x(t-r)) ::::; -x(r) <
0 from Hypothesis (6.4c). This is clearly impossible and so there is a first
time t1 such that z(h) E r. Also x(tl) = 0 and y(tl) < 0.
As long as z(t) is below rand x(t) 2': 0, we have x(t) < 0 and y(t)::::; 0.
r
Since any crossing of must occur with a vertical slope, it follows that z(t)
cannot cross r for t > h and x(t) 2': 0. Therefore, there are 8 > 0 and
E > 0, such that x(t) < -8 < 0 fort> h + E and x(t) 2': 0. Thus, there is a
first value t2 > t 1 such that x(t 2 ) = 0 and x(t 2 ) < -8.
An argument similar to this implies there is a first t 4 > t 2 such that
z(t4) E r and x(t) < 0 for t2 < t ::::; t4. But, this clearly implies y(t4) > 0
and, thus, x(t4- r) < 0. If t3 = t2 + r, then Zt 3 E K. If 71(7f) = t3, then
7 1 : K\{0} ---+ (r, oo) is continuous and Zr,('i/J) E K. If 7f(O) is between r
and the x-axis, the same argument may be applied to complete the proof
of the lemma. 0
For any 7f E K\ {0}, let 7 1(7f) be the number given by Lemma 6.3, Part
(i), and define A : K\{0} ---+ K by A1f = -zr,(,p)(1f). If 7f -:f. 0, A7P = 7f,
then the symmetry in Equations (6.5) implies that
and '1/J corresponds to a nontrivial periodic solution of Equations (6.5)

of period 2r1 ('1/J).
Lemma 6.3. The map A : K\{0} ---> K defined by A'l/; = -z71 (..p)(W) is
completely continuous.
Proof. Let B = {'1/J E K\{0} : 1'1/J(O)I :::; Mo, a :::; ao}. Since r1('1/;) > r
and A'l/; = -Z71 (..pJ('I/;) for '1/J E K\{0}, we will have shown that A(B) has
compact closure if A(B) is bounded. Let e(a) be the largest positive root
of a+ F(e) = 0 and let M > max(Mo, e(a)). Then 0:::; x(t) < M for
Also, 0:::; -y(t) :::; g(M) for 0:::; t:::; t 2('1/;) + r. Now suppose there is arE
(0, t2('1/;)) such that Fm- y(r) = M/r where Fm = inf{F(x) : 0:::; x:::; M}.
Then ±(t):::; -M/r as long as r:::; t:::; r+r and x(t) ~ 0. Since x(O):::; M,
it follows that r :::; t2('1/;) :::; r + r. Since -y(t) = g(x(t- r)) :::; g(M) for
0:::; t:::; t 2('1/;) + r, it follows that -y(t):::; -y(r) + g(M)(t- r) and so
-y(t) :::; 2rg(M) + -Mr - Fm

def
= Ym
for 0 :::; t :::; t2('1/;) +r. Of course, if -y(t) > (-M/r) + Fm for 0 :::; t :::; t2('1/;),
one obtains the better estimate -y(t) :::; rg(M) + (M/r)- Fm.
If (3 = inf{s : F(s) = 0}, then (3 :::; 0. If x(t) > (3 for t2('1/;) :::; t :::;
t 2('1/;) + r, then lx(t)l :::; max(M, lf31) and we are done. If there is a r E
[t2('1/;), t 2('1/;) + r) such that x('l/;) = (3, then ±(t) ~ y(t) ~ -ym for r:::; t:::;
t 2('1/;) + r. Therefore x(t) ~ -ymr + (3 and
lx(t)l :::; max(M, lf31 + IYmlr) ~f Xm·
Thus, on [0, t 2('1/;)+1], both x(t) and y(t) are uniformly bounded for '1/J E B.
Thus, A(B) is bounded and the lemma is proved. D
Lemma 6.4. There is a constant M > 0 such that if A'l/; = J.L'I/J, '1/J E K\{0},
IJ.LI = M, then J.L < 1.
Proof. We use the notation in the proof of Lemma 6.2. If '1/J E K\{0},
A'l/; = J.L'I/J, and '1/;(0) is below r, then (A'I/;)(0) is above r since F is an odd
function.
Suppose now that '1/;(0) is above r. For a given a > 0, let e = e(a)
be the largest positive solution of a - F( e) = 0. If '1/J = (¢,a), then
the fact that '1/;(0) is above r implies that 1'1/JI = a and ¢(0) :::; a. In
the proof of Lemma 6.3, we obtained the following estimates on Zr1 (,p) =
(Xr1 (..p),y(rl(W)),rl(W) = t2('1/;) +r:
1Yh(7/J))I S: 2rg(~(a)) +~(a)+ IFml

= [2rg(F-l(a)) + p-l(a) + IFml]a
a a
lxT1 (w)l S: max(p-l(a), ~+ IYh(7/J))Ir)a.
a a a
Hypotheses (6.4d) imply there are a0 > 0 and k < 1 such that IY(TI('lj;))l ::;
ka, for a ;:::: a0 and 7/J E K, 'lj;(O) above r. Therefore, if A'lj; = f.L'lj; for 7/J E K,
7/J(O) above r and a 2': ao, then f.L < 1. This proves the lemma. D
Our next objective is to show that 0 is an ejective point of A. To do

this, we will make use of the decomposition of the space C as C = U E9 S
where U is the subspace of dimension 2 of C spanned by the initial values
of the two eigenvalues with positive real parts given in Lemma 6.1. Let 1ru
be the associated projection onto U.
Lemma 6.5. If ko(r) is as in Lemma 6.1, then inf¢E8B(l)nK l1ru¢l > 0 if

f(O) < ko(r).
Proof. For k = - f(O) > -k0 (r), the linear part of Equations (6.5) is
x(t) = y(t) + kx(t)
y(t) = -x(t- r)
and the formal adjoint equations are
u(t) = -ku(t) + v(t + r)

v(t) = -u(t)
and the associated bilinear form is
((, 7/J) = b¢(0) + ~(O)a- j_or ~(() + r)¢(8) d()
where ( = (b,~), bE lR, ~ E C([O,r],lR) and 7j; =(¢,a),¢ E C, a E JR.

If the eigenvalues in Lemma 6.1 are distinct, the appropriate basis for
the solutions of the adjoint equation defining the projection 1r is
(6.6) bJ = (->. J' -1) '

j = 1,2.
Let us consider first the case where >. = '/ + ia-, 0 < a- < 1r /r. If
b =a- i(J where a and (3 are real two vectors, then the real and imaginary
parts of ((, 7/J) are given by
Re ((, 7/J) = a'lj;(O) + /_: e--y(~-r) cos a-(~+ r)¢(~) d~,

-Im((,'lj;) =(37/J(O)+ /_: e-'Y(~-r)sina-(~+r)¢(~)d~
where;J1jl(O) = ai.AI ¢(0) 2 0. Ifthereisasequence1/Jn = (¢n,an) E 8B(1)n

K such that 1ru1/!n --+ 0 as n --+ oo, then Im ( (, 1/!n) --+ 0, Re ((, 1/!n) --+ 0
as n --+ oo. But this implies <Pn --+ 0 as n --+ oo and an --+ 0 as n --+ oo.
Therefore, 1/!n --+ 0 as n --+ oo, which is a contradiction.
If now )q > .A 2 > 0 are real roots of Equation (6.2) and (j are defined
in Equations (6.6), then
((1,1/l) = b11/J(O) + j_or e->.l(t;+rl¢(~)d~
= ->.1¢(0)- a+ j_or e->"(t;+rl¢(~) dC

((2, 1/l) = b21/J(O) + j_: e->.2 (t;+rl¢(~) d~
= ->.2¢(0)- a+ j_: e->. 2 (t;+rl¢(~) d~.
Suppose there is a sequence 1/!n = ( ¢n, an) E 8B(1)nK such that ((1, 1/Jn) --+
0, ((2, 1/!n) --+ 0 as n --+ oo. Without loss in generality we may assume
<Pn(O)--+ ¢ 0 , an--+ ao as n--+ oo. Since .A1 > .A 2, the assertion implies
Since k > .A 1 > .A 2, this is a contradiction unless ¢ 0 = 0. If ¢ 0 = 0, then

the assertion implies
as n--+ oo.
Since .A1 > .A2, this implies <Pn --+ 0 as n--+ oo. Then ((1, 1/ln) --+ 0 implies
an --+ 0 = ao and 1/Jn --+ 0, which is a contradiction.
If the two eigenvalues of Lemma 6.1 are complex, then one can show
that the map 7 1 in Lemma 6.2 is completely continuous. Therefore Lemma
6.5 and Theorem 2.3 imply 0 is an ejective point of A. If these eigenvalues
are real, then 7 1 is unbounded in a neighborhood of¢ = 0. However, one
can use Lemma 6.2 to show there are E > 0, 8 > 0 such that IA¢1 > 8 for
¢ E K\{0}, 1¢1 <E. Thus, 0 is an ejective point of A. The proofs of these
facts are left to the reader.
The case where ( 1(t) = bexp(.A 1t), (2(t) = (bt + c)exp(.A1t), .A1 > 0,
remains to be considered. This case is left for the reader. With this remark,
the proof of the lemma is complete. D
Using these results and Theorems 2.2 and 2.3, we have
Theorem 6.2. IfF and g satisfy Conditions (6.4a)-(6.4d) and f(O) < ko(r),
where k 0 (r) > 0 is given in Lemma 6.1, then Equation (6.1) has a noncon-
stant periodic solution.
A special case of this theorem is f(x) = k(x 2 - 1), k > 0; that is,
the van der Pol equation with a retardation. It is interesting to look at the
latter equation in more detail. If x = uj v'k, then u satisfies the equation
u(t) + (u 2 (t)- k)it(t) + u(t- r) = 0.
This equation satisfies the conditions of Theorem 6.2 for every k > 0.
Therefore, there is a periodic solution u*(k) with Ju*(k)J ;::: c > 0 for 0 :::;
k:::; 1. Thus, the solution x*(k) = u*(k)/v'k of
x(t) + k(x 2 (t)- 1)±(t) + x(t- r) = 0
approaches oo as k -. 0.
The conditions of Theorem 6.2 are also satisfied by f(x) = ax 2 + b,
a> 0, b < k0 (r), g(x) = x, and, in particular, for f(x) = x 2 , g(x) = x.
Hop£ [1] was the first to state Theorem 1.1 for ordinary differential equa-
tions. Many generalizations to infinite-dimensional systems have been given
(see Marsden and MacCracken [1] for references). To the author's knowl-
edge, the first statement similar to Theorem 1.1 for RFDE was given by
Chow and Mallet-Paret in a course at Brown University in 1974 with a
proof different from the one in the text. The proof in the text is easily
adapted to certain types of partial differential equations. It is of interest
to determine the stability and amplitude of the bifurcating periodic orbit.
Efficient procedures for doing this using a method of averaging have been
given by Chow and Mallet-Paret [1]. For Equation (4.1), Gumowski [1] has
computed the period and amplitude of the solution using the method of
undetermined parameters (see, also Strygin [1]). Ginzburg [1] has also con-
sidered the bifurcation problem. Other results giving more properties of the
orbits are contained in Chafee [2]. The global existence of the Hop£ bifurca-
tion as a function of the initial data, >., and the period has been discussed
by Chow and Mallet-Paret [2] and Nussbaum [9].
It is also possible that a bifurcation from a constant occurs as one varies
the delays in an equation. In the case of a single delay as in Equation (4.1)
of Section 10.4, one can reduce the discussion of Equation (4.2) of Section
10.4 and apply the Hop£ bifurcation Theorem 1.1. The case of several delays
has been discussed in Hale [24].
As remarked in Section 10.5, much research has been devoted to the
behavior of solutions of differential difference equations with a small delay.
Perell6 [1] has also discussed equations with a small delay and proved the
following interesting "continuity" result. Suppose one has shown there is
an invariant torus for an ordinary differential equation and a cross section
of this torus is mapped into itself by the flow induced by the differential
equation. Then the Brouwer fixed-point theorem implies there is a periodic

solution. Perell6 [1] has shown the differential difference equation obtained
from this ordinary equation by the introduction of small delays must also
have a periodic solution.
If it were possible to construct a closed bounded convex set K in C
that is free of equilibrium points and has the property that for each ¢ E K,
the trajectory through¢ of an RFDE returns to K, then the usual fixed-
point theorems could be applied to obtain the existence of nonconstant
periodic solutions. However, this construction is very difficult in an infinite-
dimensional space.
In a fundamental paper, Jones [4] introduced the idea of finding a cone
that maps into itself under the flow rather than a cross section of a torus.
The cone is easier to construct than a torus but some other complications
arise because most problems have zero as an equilibrium point. Therefore,
one is interested in finding nonzero fixed points of a mapping of a cone into
itself knowing that zero is a fixed point. This problem has been the moti-
vation for a number of interesting fixed-point theorems for cone mappings
(see Krasnoselskii [2] and Grafton [1]) as well as mappings of a convex set
into itself with an ejective fixed point as an extreme point of the convex
set (Browder [2]). Using the ideas of Browder [2], Nussbaum [3] proved
Theorem 2.1. Theorem 2.2 is also due to Nussbaum [3] and combines in
an interesting way ejectivity as in Theorem 2.1 and the concept of eigen-
value as in the work of Krasnoselskii [2] and Grafton's theorem (see the
statement in Hale [11] and Lopes [4]). Theorem 2.3 is essentially due to
Chow and Hale [2]. Smith [1] has also developed the ideas of Krasnoselskii
to obtain fixed-point theorems and periodic solutions.
Theorem 3.1 is due to Nussbaum [5] and was proved with the intention
of making applications to the range of the period. The usefulness of the Hop£
bifurcation theorem to assist in this effort was pointed out by Chow and
Hale [2].
Lemmas 4.1 and 4.2 are due to Wright [2]. Theorems 4.2 and 5.2 are due
to Jones [4, 5] with the proof following the ideas in Grafton [1]. Theorems
4.3, 4.4, and 5.3 are due to Nussbaum [5]. If we let 1 + x(t) = eu(t) in
Equation (4.1), then u satisfies the equation
(7.1) u(t) = -oJ(u(t- 1))

where f(u) = eu- 1. In a series of papers, Nussbaum [3,4,5,6,7] has dis-
cussed Equation (7.1) under very general hypotheses on f and has given
sophisticated theorems on the existence of periodic solutions and the range
of the periods of these solutions as a function of a. For the equation
(7.2) ±(t) = -ax(t- 1)[a- x(t)][b + x(t)],
the results of Nussbaum imply that the range of the period of the periodic
solutions of Equations (7.2) contains the interval4 < p < 2 +(a/b)+ (b/a).
If a= b = 1, that is, Equation (5.1), then no information is obtained about

the range of the period. For Equation (5.1), Jones [6] has shown there is
a periodic solution of period 4 for every a > 0. Kaplan and Yorke [1]
have given more general conditions on f in Equation (7.1) to have periodic
solutions of period 4.
A special case of Theorem 6.2 was first given by Grafton [1]. The proof
in the text is in the same spirit as Grafton [1] but with many corrections and
simplifications taken from Nussbaum [3]. One can also relax the conditions
on f and g in Equation (6.1) and prove an analogue of Theorem 6.2 (see
Grafton [2] and Nussbaum [3]). Nussbaum [7] has also given some results
on the range of the period of periodic solutions of Equation (6.1).
Periodic solutions of many other types of equations are known to exist.
Among these are equations of the form
(7.3) x(t) =-a[£: x(t + tJ) dj3(tJ)]l1 + x(t)]
(7.4)
where j3 is a nondecreasing function, (3(0) = 0, and (3(1) = 1. Some remarks

and numerical results for j3 a step function with two jumps is contained in
Jones [6]. Chow and Hale [2] discussed the case of Equation (7.3) where
(3( tJ) is absolutely continuous on [-1, OJ and !3( tJ) = 1 on [- ~, 0] and showed
there is a nonconstant periodic solution for a > a 0 > 0. Walther [1] has
eliminated the hypothesis that jJ(tJ) be absolutely continuous. The proof in
Walther [1] could possibly be simplified by using the ideas in Chow and
Hale [2].
In the analysis of more general Equations (7.3) and (7.4), it is necessary
to have detailed information about the behavior of the eigenvalues of the
linear equation as the function j3 varies in some class of functions. Some
interesting results on this general question are contained in Walther [2]. In
this same paper, Walther also gives sufficient conditions for all roots of the
linear equation to lie in the left half-plane for every a. In these cases, any
nonconstant periodic solution that arises by a variation in a cannot be due
to a Hop£ bifurcation.
Some results on the uniqueness and stability of periodic solutions of
special cases of Equation (7.3) also have been obtained (see Jones [6], Ka-
plan and Yorke [2], and Walther [3]).
Chow [2] considers the equation
x(t) = -ax(t) + e-x(t-r)'
0 < a < e- 1 , and shows there is a nonconstant periodic solution for r ;?: r 0
sufficiently large. The proof could be simplified by using the results in this
chapter.
Numerical studies have been made by Grafton [3] on the equations
x(t) + a[x 2 (t)- 1]±(t) + Ax(t- r) + Bx(t- s) = 0

x(t) + a[xn(t)- 1]±(t) + Ax(t- r) + Bx 3 (t- r) = 0
x(t) + a[x 2 (t- r) -1]±(t- r) + x(t) = 0
for various values of the constants a, A, B, r, s, and n. A variety of types
of oscillations occur that have been mathematically explained. Nussbaum
[7] has shown the existence of a periodic solution of the first equation for
a< 0, r > 0, s = 0, A= 1, and -1 < B :::; 0. In [7], Nussbaum has also
discussed some other cases of equations of this type.
Considerable attention has been devoted to the following equation,
which can be considered as a generalization of Equation (7.1):
(7.5) ±(t) = -~ 1 -1+6

_1 f(x(t + 8)) dB,
where a and 8 are positive numbers, 8 :::; !,
and the function f is C 1 ,
f'(O) = 1, xf(x) > 0 for x "I 0 and f(x) 2:: -B > -oo for all x. The
following result has been proved by Walther [4] (see also Nussbaum [10]):
Theorem 7.1. There is a constant a 0 > 0 such that for a > a 0 , there are
real numbers z 1 > 1-8, z 2 - z 1 > 1-8, and a periodic solution x 0 (t) = x(t)
of (7.5) of period z2 such that x(O) = 0, x(t) > 0 on (0, z1), x(t) < 0 on
(zr, z2).
The constant a 0 is the value of a at which the zero solution of the

linear variational equation becomes unstable. The method of proof uses
cone maps and the ejectivity properties of the origin.
Alt [1] also has considered the existence of periodic !>Olutions for more
general equations with one time delay or distributed time qelays.
Another interesting class of equations are those for which the delay is
determined by a threshold condition. More specifically, Alt [2] has consid-
ered periodic solutions of the equation
(7.6) x(t) =- f(x(t), x(t- a(xt))),
where the function a(¢) is determined by a threshold condition
(7.7) 1° -u(<f>)
k(¢(s)) ds = ko,
where k : lR--> JR+ is continuous, k;:::: c > 0 and k(O) = ko. The function
f is assumed to satisfy f(x, y) ;:::: 0 (resp. :::; 0) for x 2:: 0, y 2:: 0 (resp. x :::;
0, y:::; 0), f(O, y) < 0 for y < 0, f(x, x) "I 0 for x "I 0, f(O, x)) 2:: -B > -oo
for x E ( -oo, OJ. Assuming also that the linear variational equation has
an eigenvalue with positive real part, Alt [2] shows that there is a periodic
solution of (7.6) and (7.7). The method of proof is to show first that there
is set K c C that is homeomorphic to a convex bounded set and such that
for any¢ E K, there is a r(¢) such that the solution through¢ at 0 is inK
at r( ¢). The set K has 0 on the boundary of K and it must be shown that
it is ejective. Since o-( ¢) is not differentiable in ¢, the obvious linearization
cannot be used for the ejectivity. Alt [2] transforms the equation to a neutral
functional differential equation for which the method using projections onto
the unstable manifold can be applied.
Equations of the form (7.6) and (7.7) often serve as models in biological
problems. In the manner in which it is presented, the delay depends on
the state of the system. Smith [2] has shown how to eliminate the state-
dependent delay for certain types of threshold-type problems.
For a functional differential equation x(t) = f(xt) with delay interval
of length r > 0 and an equilibrium point 0, a solution x(t) is said to be
slowly oscillating about the point 0 if it has a sequence of zeros approaching
oo and the distances between the zeros of x(t) are > r. Most of the results
on the existence of periodic solutions of delay differential equations are for
slowly oscillating ones. In many situations, the slowly oscillating periodic
solutions are the ones that enjoy stability properties and thus are the ones
that will be observed in the system. To substantiate this remark, let us
restrict attention to Equation (7.1) and always assume that f' (0) = 1,
xf(x) > 0 for x f- 0, f(x) is bounded below, and a> ~- In this situation,
we know that (7.1) has a slowly oscillating periodic solution. We denote by
Sf the set of initial data for the slowly oscillating solutions of (7.1). Walther
[5] has shown that all of the slowly oscillating solutions of (7.1) eventually
enter a solid torus; more specifically, there are constants a > 0, and r > 0
such that for any solution x(t) of (7.1) with Xt E Sf of (7.1), there is a to
such that Xt E { ¢ E Sf :a:::; 11¢11 :::; r} fort 2: to. If we assume in addition
that f'(x) > 0 for all x, then Walther [6] has shown that there is an a> 0
such that ¢ is in the closure of Sf if the solution x(t, ¢) through ¢ at 0
satisfies limsupt--+oo lx(t, ¢)1 >a. He shows also that there is an a' > 0 such
that the set { ¢ E C : limsupt--+oo lx(t, ¢)1 2: a'} is dense in C. Since the
set Sf is open, we obtain openness and denseness of the slowly oscillating
solutions.
For the bifurcation of slowly oscillating periodic solutions of (7 .1) under
some symmetry conditions on J, see Walther [7]. Under the assumption that
f is odd and some restrictive monotonicity conditions, Chow and Walther
[1] have proved the hyperbolicity of the slowly oscillating periodic solution
(of period 4). When the function f is not odd, there may be several slowly
oscillating periodic solutions (see Saupe [1,2]). If a is large, then there is
such a solution with a very long period. Xie [1] has shown that this solution
is linearly stable. He accomplishes this by obtaining a good expression for
the eigenvalues of the Poincare map.
Another very interesting class of equations has the form
(7.8) x(t) = -f.Lx(t) + f.Lf(x(t- 1)),
where f.L > 0 is a constant and the function f has negative feedback; that
is, f(O) = 0, f'(O) < 0 and xf(x) < 0 for x "/c 0. Also, suppose that there
is an interval I such that f(I) C I. Under the additional assumption that
the origin of (7.8) is linearly unstable, Hadeler and Tomiuk [1] have shown
that there is a slowly oscillating periodic solution. Some earlier results were
obtained by Pesin [1].
The linearization of (7.8) about the origin has a two-dimensional invari-
ant manifold W 0 consisting of the span of the eigenfunctions of an eigenvalue
with positive real part and the solutions corresponding to these eigenfunc-
tions are slowly oscillating. For the complete equation near the origin, there
is a two-dimensional local invariant manifold Wloc(O) that is tangent to W 0
at 0. The solutions on this manifold are slowly oscillating. Let W(O) be the
global extension of Wloc(O) by following the solutions of (7.8). Under the
assumption that f is monotone and bounded either from above or below,
Walther [8] has shown that the boundary of W(O) is a periodic orbit that
attracts all orbits of (7.8) with initial data in W(O) \ { 0 }.
As remarked earlier, there may be more than one slowly oscillating
periodic solution of (7.8). Let'/ be a hyperbolic unstable periodic orbit of a
slowly oscillating periodic solution and let W, be the unstable manifold of
this orbit. Walther [9] has shown that W, is a smooth annulus-like graph of
dimension two whose boundary consists either of two other slowly oscillating
periodic orbits or one slowly oscillating periodic orbit and the equilibrium
solution 0.
A further generalization of (7.8) that is important in physiology and
laser optics is
d d
(7.9) (Em dt + 1) ···(Eo dt + 1)x(t) = f(x(t- 1)),
where each Ej > 0, j = 1, ... , m and f satisfies the same conditions as in

(7.8). Assuming that the origin is unstable and m = 1 (that is, a second-
order equation), an der Heiden [1] has proved the existence of a slowly
oscillating periodic solution. For any m, Hale and Ivanov [1] have proved
the same result, provided that the constants EJ, 0 ~ j ~ m, are sufficiently
small.
For a recent survey on oscillations in scalar delay differential equations,
see Ivanov and Sharkovsky [1].
Except for second-order differential equations, there are few results on
the existence of periodic solutions of systems of delay differential equations.
Taboas [1] has considered the system
x1(t) = -x1(t) +aF1(x1(t -1),x2(t -1)),

x2(t) = -x2(t) + aF1(x1(t -1),x2(t -1))
where a> 0 is a constant, F = (F1 , F 2 ) is a bounded C 3 -map with 8Fl/8x2

and 8F218x 1 different from 0 at the origin and F satisfies a negative feedback
condition:
X2F1(x1,x2) > 0, X2 f. 0, X1F2(x1,x2) < 0, X1 f. 0.

Under these hypotheses, Taboas [1] shows that there is an ao > 0 such that,
for any a > a 0 , there is a periodic solution of period > 4. The method of
proof is to show that the negative feedback condition causes the solutions of
the equation to rotate in the (x 1 , x 2 )-plane and then he uses modifications
of the method of cone maps and ejective fixed points for the cone
K = { (¢1, ¢ 2) E C: e 0 cj>j(B) nondecreasing- 1:::;: (}:::;: 0, j = 1, 2,

¢1(-1) = 0, ¢2(-1) 2: 0}
after renormalization so that 8 F1 (0) I 8x2 = -8F2 (0) I 8x1.
In the proof of ejectivity, he found the following interesting modifica-
tion of Theorem 2.3.
Theorem 7.2. Suppose that the following conditions are fulfilled:

(i) There is an eigenvalue A of Equation (2.2) satisfying ReA > 0.
(ii) There exists a closed convex subset K of C, 0 E K, and a continuous
function r: K \ {0}--+ [a, oo), a> 0, such that the map
Acj> = Xr(¢)(¢), cj> E K \ {0}, AO = 0,

is completely continuous and AK C K.
(iii) inf{ 17f>.Xtl : Xt = Xt(cf>), cf> E K, 1¢1 = 8, 0:::;: t:::;: r(cf>)} > 0.
(iv) Given G c C open, 0 E G, there is a neighborhood V of 0 such that
Xt(cf>) E G if cj> E V n K, cj> f. 0 and 0:::;: t:::;: r(cf>).
Then 0 is an ejectivity fixed point of A.
Leung [1] has proved the existence of a periodic solution of period > 2r
for the prey predator model
±(t) = x(t)[a- bx(t)- cy(t)]

y(t) = ay(t)[x(t- r)- ,6]
provided that all constants are positive, a > a- 1 b,6(b + a) and r >
[2a(a- b,6)]- 1 .
Recently, there have been several problems in ecology and physiology
for which the delay depends on the state of the system and it cannot be
eliminated by any change of variables. In this case, no general qualitative
theory is available. The primary difficulty arises from the fact that the
solutions of the equation are not differentiable with respect to the delay
function. Therefore, the theory given in the text is not applicable. In spite
of this fact, there have appeared recently some interesting results dealing
with special types of solutions of equations with state-dependent delays and,
in particular, the slowly oscillating periodic solutions. Consider Equation
(7.8), with the delay r = r(x(t)) depending on the present state of the
system. Under the same assumptions (negative feedback and the existence
of a positively invariant interval containing the origin), it has been shown
that there is a slowly oscillating periodic solution if f-l is large (see Mallet-
Paret and Nussbaum [5], Kuang and Smith [1,2]). The existence of the
periodic solutions is proved by using an appropriate extension of the method
of cone maps mentioned before. The papers of Kuang and Smith [1,2] apply
also to threshold-type problems. Mallet-Paret and Nussbaum [5,6] have
discussed detailed properties of the slowly oscillating periodic solution as
E----> 0. In particular, for the equation
ei:(t) = -f-lx(t)- kf-lx(t- r(x(t)), r(x) = 1 +ex,

they have shown that the solution approaches a saw-tooth function of period
k+1 given by the straight line segment y = c- 1 x on the interval-1 ::; x ::; k.
This is in marked contrast with what occurs with a constant delay (see
Chapter 12 for a discussion of this case).
For equations with several delays or RFDE with distributed delays
with the distribution function having several large peaks, the problem of
existence of periodic solutions is much more difficult than the same problem
for the equations that have been discussed earlier.
Even the local Hopf bifurcation theorem is difficult. If the bifurcation
parameters are taken to be the delay parameters, there is a difficulty due to
the fact that the solutions of an RFDE are not differentiable with respect to
the delays. However, using the fact that the periodic solutions of an RFDE
will be as smooth as the vector field, it is possible to use a variant of the
usual implicit function theorem to show that the Hopf bifurcation theorem
remains valid with the delays being the parameters (see, for example, Hale
[24]). For the types of differentiability properties of solutions of RFDE with
respect to delays, see Hale and Ladeira [1,2].
A greater difficulty arises in the understanding of the manner in which
the roots of the characteristic equation of a linear system with several delays
depend on parameters. For the linear equation,
(7.10) x(t) + ax(t)bx(t- r) + cx(t- T) = 0,

where a, b, c, r 2::: 0, and T 2::: 0, are constants, there have been many
attempts to understand the region in the parameter space for which the
solutions approach zero (for example, Bellman and Cooke [1], Hale [24],
Mahaffey [1], Marriott, Vallee, and Delisle [1], Nussbaum [11], Ragazzo and
Perez et al. [1], Ruiz-Claeyssen [1], Stech [1]). The boundary of this region
is of primary interest since it represents the points at which the origin can
undergo a bifurcation from stability to instability. The complete description
of this boundary is not available at this time. However, Hale and Huang
[1] have given a good description in the case when a, b, and c are fixed and
the delays r, T are considered as the parameters. To be more precise, the
stable region is defined as a maximal connected set D c [0, oo) x [0, oo) that
contains the origin (0, 0) such that for each (r, T) ED, the zero solution of
(7.10) is asymptotically stable. Hale and Huang [1] show the following:
Theorem 7.3. If a stable region is unbounded, then its boundary will ap-
proach a straight line parallel to the r-axis or T-axis as r + T----> oo.
As a consequence of this result, we see that, if a half line in the first

quadrant of the (r, T )-plane contains an unstable point, then the intersection
of this line and the boundary of the stable region contains at most finitely
many points and eventually leaves the stable region.
More detailed properties of the boundary of a stable region are dis-
cussed in Hale and Huang [1]. For a point on the boundary, if the char-
acteristic equation has only one pair of imaginary roots, then there is a
Hopf bifurcation to a periodic orbit as the delays cross the boundary. The
determination of whether it is supercritical or subcritical is very difficult
since it requires considerable computation (see Stech [2], Franke and Stech
[1] for how to use MACSYMA to do these computations and more compli-
cated ones). It is possible to have multiple eigenvalues on the boundary of
a stable region. In this case, very complicated oscillatory phenomena can
arise as the boundary is crossed-even chaos. The special situation where
r = 3T has been analyzed in some detail by Stech [1], where he showed that
three periodic solutions could bifurcate from the origin. Nussbaum [11] had
previously observed that, in this case for a special type of nonlinear equa-
tion, there were two periodic orbits that existed globally with respect to the
delays larger than some value. He conjectured that there should be another
periodic solution as well, which Stech confirmed near the bifurcation point
on the boundary of the stable region.
In modeling of physical systems, the manner in which the equations
depend on the past history is taken to be as simple as possible due to the
severe difficulties in the analysis of RFDE. In many situations, authors have
chosen the memory functions to be of the type that repeated differentiations
will reduce the problem to an ordinary differential equation of high order.
This leads to a simpler system, but it leaves open an important question.
If some property is discovered from the ordinary differential equation, does
it remain valid for the original problem if we choose a memory function
that is close in some sense to the one that was analyzed, but for which
no reduction to an ordinary differential equation is possible? Hines [1] has
many results in this direction, including the preservation of stability and
the upper and lower semicontinuity of attractors. Farkas and Stepan [1] has
similar results for the preservation of stability.
12
Additional topics
In the previous chapters, we have touched only the surface of the theory
of functional differential equations. In recent years, the subject has been
investigated extensively and now there are several topics that can be clas-
sified as a field in itself. In this chapter, we give an introduction to some
of these areas, describe the main results, and occasionally give indications
of proofs. We cover generic theory, equations with negative feedback and
Morse decompositions, slowly oscillating periodic solutions, singularly per-
turbed delay equations, averaging, and abstract phase spaces associated
with equations with infinite delay. In the supplementary remarks, we give
references for the detailed proofs and indicate other areas of functional
differential equations that are currently being investigated.
12.1 Equations on manifolds-Definitions and examples
In this section, we begin with a few examples that will serve as moti-
vation for the consideration of functional differential equations on finite-
dimensional manifolds.
Example 1.1. For any constant c, the scalar equation

(1.1) B(t) = csin(B(t- 1))
can be considered as an RFDE on the circle 8 1 = { y E 1R? : yf + y~ = 1 }
by considering B as an angle variable only determined up to a multiple of
211".
Example 1.2. If b, c are constants, then we write the second-order RFDE

(1.2) iJ(t) + bB(t) = csin(B(t- 1))
as a system of first-order RFDE
:.h(t) = x2(t)
(1.3)
12.1 Equations on manifolds 365
By considering x 1 as an angle variable only determined up to a multiple

of 21f, Equation (1.3) is an RFDE on the cylinder 8 1 x ffi. We remark
that we can take the space of initial data for the solution x of (1.2) as
C([ -1, 0], 8 1 ) X ffi.
Example 1.3. Let 8 2 = { (x, y, z) E ffi3 : x2 + y 2 + z2 = 1} and consider

the following system of RFDE:
x(t) = -x(t- 1)y(t) - z(t)

(1.4) y(t) = x(t- 1)x(t) - z(t)
z(t) = x(t) + y(t).
If (x(t), y(t), z(t)) is a solution of Equation (1.4), it is easy to see that
x(t)x(t) + y(t)y(t) + z(t)z(t) = o

for all t ?: 0. As a consequence, fort ?: 0, x 2 (t) + y 2 (t) + z 2 (t) = a 2 , a
constant. Thus, if an initial condition ¢ = (¢ 1 , ¢ 2 , ¢ 3 ) satisfies ¢(B) E 8 2
for all() E [-1,0], we conclude that the solution (x,y,z)(t;¢) E 8 2 for all
t?: 0.
With this remark, we can define an RFDE on 8 2 by the map
where F(¢) is the tangent vector to 8 2 at the point ¢(0) defined by
We now formalize the notions in these examples to obtain an RFDE

on an n-dimensional manifold. Roughly speaking, an RFDE on a man-
ifold M is a function F mapping each continuous path ¢ lying on M,
¢ E C([-r, 0], M), into a vector F(¢) tangent toM at the point ¢(0) EM.
Let M be a separable coo finite n-dimensional manifold, I the interval
[-r, 0], r?: 0, and C(I, M) the totality of continuous maps¢ of I into M.
Let TM be the tangent bundle of M and TM: TM ~ M its C 00 -canonical
projection. Assume that there is given on M a complete Riemannian struc-
ture (it exists because M is separable) with DM the associated complete
metric. This metric on M induces an admissible metric on C(I, M) by
8(¢, {fi) = sup8M(¢(B), {fi(B)).

OEI
The space C(I, M) is separable (since M is complete and separable) and

is a C 00 -manifold modeled on a separable Banach space. If M is imbedded
as a closed submanifold of a Euclidean space V, then C(I, M) is a closed
C 00 -submanifold of the Banach space C(I, V).
366 12. Additional topics
If p: C(J,M)---+ M is the evaluation map, p(¢) = ¢(0), then pis c=,

and for each a EM, p- 1 (a) is a closed submanifold of C(I, M) of codimen-
sion n = dimM. A retarded functional differential equation (RFDE) on M
is a continuous function F : C(J, M) ---+ TM such that TM oF= p. If we
want to emphasize the function F defining the RFDE, we write RFDE(F).
A solution of RFDE(F) is defined in the obvious way, namely, as a
continuous function x : [-r, a) ---+ M, a > 0, such that ±(t) exists and is
continuous for t E [0, a) and (x(t), x(t)) = F(xt) for t E [0, a). Locally, if
F(¢) = (¢(0), f(¢)) for an appropriate function f, then this is equivalent
to x(t) = f(xt)·
We remark that it is just as easy to define a neutral functional differ-
ential (NFDE) on M. We simply choose a map D : C(J, M) ---+ M such
that D is atomic at zero and ask that TM o F = D.
The basic theory of existence, uniqueness, and continuous dependence
on initial data for general RFDE on manifolds is the same as the theory
when M = IRn.
Example 1.4. Any Ck-vector field on M defines a Ck-RFDE on M. In fact,

if X: M---+ TM is a Ck-vector field on M, it is easy to see that F =X o p
is a Ck-RFDE o:p. M.
Example 1.5. To show that the equation considered in Example 1.2 is an

RFDE according to the definition, we need the concept of a second-order
RFDE on M. Let P : C (I, T M) ---+ T M x T M be a continuous function
that locally has the representation
F(¢,'1/J) = ((¢(0), '1/J(O)), ('1/J(O), f(¢, '1/J))).

The solutions (x(t),y(t)) of the RFDE(F) on TM satisfy the equations
(1.5) ±(t) = y(t), y(t) = f(xt, Yt)
where x(t) E M. If it is possible to perform the differentiations, then we
obtain the second-order equation
If we now return to Example (1.2), we see that the formulation requires

that we consider initial data in the space C(J, 8 1 ) x C(I, IR). However,
this does not affect the dynamics since the solution will be in the space
C(I, 8 1 ) x IR after one unit of time.
Example 1.6. (Delay differential equations on M). Let g : M x M ---+ T M

be such that (TM o g)(x,y) = x and let d: C(J,M)---+ M x M be defined
by d(¢) = (¢(0), ¢( -r)). The function F = god is an RFDE on M that
can be written locally as
x(t) = g(x(t),x(t- r)),
12.1 Equations on manifolds 367
where g(x,y) = (x,g(x,y)).
For notational purposes, it is convenient to let
where BCk denotes that the function and derivatives up through order k are
bounded. An FE Xk corresponds to a Ck-RFDE. As for an RFDE on IRn,
we let Tp(t) denote the solution operator for the RFDE(F) and introduce
the concept of dissipativeness, globally defined solutions, attractors, etc.
For an RFDE(F), we denote the set of globally defined solutions by A(F).
We know that if the equation is point dissipative, then A(F) is compact,
invariant and is the global attractor. Another important concept is the
following. An element 'lj; E A(F) is called a nonwandering point ofF if for
any neighborhood U of 'lj; in A( F) and any T > 0, there exist t = t(U, T) >
T and;(; E U such that Tp(t);f; E U. The set of all nonwandering points ofF
is called the nonwandering set and is denoted by D(F). The following result
is easy to prove but very important for the generic theory to be discussed
in a later section.
Lemma 1.1. IfF is a C 1 -RFDE on M that is point dissipative, then D(F)

is closed and, moreover, if Tp(t) is one-to-one on A(F), then D(F) is
invariant.
If M is a compact manifold without boundary, then A(F) is always

the global attractor for a C 1 -RFDE on M. Furthermore, it is not difficult
to prove the following result.
Theorem 1.1. If M is a compact manifold and FE X 1 , then A( F) is upper

semicontinuous in F; that is, for any neighborhood U of A( F) in M, there
is a neighborhood V ofF in X 1 such that A( G) C U if G E V.
The following result also is true.
Theorem 1.2. If M is a compact manifold and FE X 1 , then dim A( F) 2:

dim M and the restriction of p to A( F) is onto; that is, through any x EM,
there is at least one global solution. Furthermore, if M is without boundary
and A(F) is a compact manifold without boundary, then A(F) is homeo-
morphic to M.
If X is a C 1-vector field on M and M is a compact manifold without

boundary, then A( F) = M. A stronger version of the last part of Theorem
1.2 is
Theorem 1.3. If M is a compact manifold without boundary and X is a

C 1 -vector field on M, F =X o p, then there is a neighborhood U ofF in
X 1 such that for each G E U, the attractor A( G) is diffeomorphic to M,

A(G) --> A(F) in the Hausdorff sense as G --> F, and the restriction of
Ta(t) to A( G) is a one-parameter family of diffeomorphisms.
Example 1.1. (Revisited). In Equation (1.1), if c = 0, we have the ordinary

differential equation ±(t) = 0 on 8 1 . Theorem 1.3 implies that the attractor
for (1.1) is diffeomorphic to 8 1 if lei is sufficiently small. The same remark
also is true if we consider the equation
x(t) = sinx(t) + cg(x(t -1)),

where g(x + 27!') = g(x) is a given C 1 -function and lei is sufficiently small.
The following result also can be useful.
Theorem 1.4. Let M be a compact manifold without boundary and let F E

X 2 . If there is a constant k > 0 such that IID¢TF(c/>)ll:::; k, IID~TF(c/>)11:::; k
for all t 2: 0 and all cf> E C(J, M), then the attractor A(F) is a connected,
compact C 1 -manifold (without boundary) that is homeomorphic toM.
Furthermore, the restriction ofTp(t) to A(F) is a one-parameter fam-
ily of diffeomorphisms.
Finally, there is a neighborhood U ofF in X 2 such that for any G E U,
the attractor A( G) is a manifold that is diffeomorphic to A( F).
Example 1.7. Consider the scalar equation
(1.6) x(t) = b(t) sin[x(t) - x(t- 1)],
where b E C 1 (lR, lR) is such that b(O) = 0 and Idb/ dxl :::; a < 1. As in
Example 1.1, this equation defines an RFDE on 8 1 . Any constant function
is a global solution of (1.6). Conversely, any global solution is a constant
function. To see this, one considers the map
B: z(t) f--+ it
t-1
b(u) sin[z(u)] du
acting in the Banach space of all bounded continuous functions z(t) with
the sup norm. It is easy to see that B is a contraction map and that z(t) = 0
are its fixed points. On the other hand, if x(t) is a global solution of (1.6),
then [x(t)- x(t- 1)] is bounded and
x(t)- x(t- 1) =it t-1

b(u) sin[x(u)- x(u- 1)] du,
which shows that x(t) - x(t - 1) = 0. The equation gives ±(t) = 0 and
x(t) = constant. In this case, the attractor A(F) is a circle. Let us indicate
how to show that for any cf> E C, there is a unique constant function a(¢)
12.2 Dimension of the global attractor 369
such that T1 (t)---+ a(¢) as t---+ oo. For any constant function c, the linear
variational equation is
y(t) = b'(O)[y(t)- y(t- 1)].
The corresponding characteristic equation is A = b' (0) - b' (O)e-.\. From
the Appendix and the fact that lb'(O)I < 1, we know that all solutions
of this equation have negative real parts except for A = 0. Furthermore,
A = 0 is simple. This is enough to prove the assertion (for details, as well
as a reference to another proof, see the references in the supplementary
remarks). The map a is a C 1 -retraction, a 2 =a, aoT1 (t) = TJ(t) oa. The
image a (C) of a is the at tractor A( F) and, for any constant function c, the
set a- 1 (c) is a submanifold of C of codimension 1.
Example 1.8. Consider the scalar equation

• 7f 7f
(1. 7) B(t) = 2 (1- cosB(t)) + 2 (1- cosB(t -1))
on S 1 . The only equilibrium point on S 1 is () = 0. On the other hand,
the function B(t) = 1rt is a periodic solution of (1.7). The attractor A(F)
must contain the point 0 as well as this periodic orbit. As a consequence,
A(F) cannot be homeomorphic to S 1 . If A(F) is a manifold, it must have
dimension > 1.
12.2 Dimension of the global attractor
In this section, we present some results on the size of the global attractor
A( F) of an RFDE(F) in terms of limit capacity and Hausdorff dimension.
The principal results are applicable not only to RFDE but to many other
types of evolutionary equations. As we will see, the theory applies to the
situation where the solution operator T(t) can be written as T(t) = S(t) +
U(t), where U(t) is compact and S(t) is a linear operator whose norm
approaches zero as t ---+ oo.
Let K be a topological space. We say that K is finite-dimensional if
there exists an integer n such that for every open covering A of K, there
exists another open covering A' of K refining A such that every point of K
belongs to at most n + 1 sets of A'. In this case, the dimension of K, dim
K, is defined as the minimum n satisfying this property. Then dim IRn = n
and, if K is a compact finite-dimensional space, it is homeomorphic to a
subset of IRn with n = 2dimK + 1. If K is a metric space, its Hausdorff
dimension, dim H ( K) is defined as follows: for any a > 0, E > 0, let
M~(K) = inf 2:>r

where the inf is taken over all coverings BE; (xi), i = 1, 2 ... , of K with Ei <
E for all i, where BE; (xi)= { x: d(x, xi) < Ei}. Let J.L 0 (K) = limE--->O J.L~(K).
The function J.La is called the Hausdorff measure of dimension a. For a = n
and K a subset of IR.n with lxl = sup Ixi I, f..Ln is the Lebesgue outer measure.
It is not difficult to show that if J.L 0 (K) < oo for some a, then J.Lf3(K) = 0
if f3 > a. Thus,
inf{ a: J.L 0 (K) = 0} =sup{ a: J.L 0 (K) = oo },
and we define the Hausdorff dimension of K as
It is known that dim(K)::::; dimH(K), and these numbers are equal when
K is a submanifold of a Banach space. For general K, there is little that
can be said relating these numbers.
To define another measure of the size of a metric space K, let n(E, K)
be the minimum number of open balls ofradius E needed to cover K. Define
the limit capacity of K, c( K), by
. logn(E, K)
c(K) = hm sup 1 ( / ) .
E--->0 Og 1 E
In other words, c(K) is the minimum real number such that for every a> 0,
there is a 8 > 0 such that
1 ) c(K)+u
n(E, K) ::::; ( - if 0 < E < 8.
E
It is not difficult to show that dim H(K) ::::; c(K).
Theorem 2.1. Suppose that X is a Banach space, T : X --+ X is an a-

contraction, point dissipative and orbits of bounded sets are bounded. If T
is a 0 1 -map and DxT = S + U, where U is compact and the norm of S is
less than 1, then the global attractor A ofT has the following properties:
(i) c(A) < oo;
(ii) If d = 2c(A) + 1 and S is any linear subspace of X with dimS > d,
then there is a residual set II of the space of all continuous projections
P of X onto S (taken with the uniform operator topology) such that
PIA is one-to-one for every P E II.
The first part of this theorem says that the limit capacity c(A) of A
is finite, which in turn implies that the Hausdorff dimension is finite. The
second part of the theorem says that the attractor can be "flattened" in a
residual set of directions onto a finite-dimensional subspace of dimension
> 2c(A) + 1.
12.2 Dimension of the global attractor 371
The proof of Theorem 2.1 is not given. However, it is worthwhile to

point out that in proving that c(A) is finite, explicit estimates are obtained
(better estimates often can be obtained in a specific problem). The esti-
mates for the limit capacity of a compact attractor for a map T can be
obtained by an application of general results for the capacity of compact
subsets of a Banach space E with the property that T(K) =:> K for some
C 1-map T: U---+ E, U =:> K, whose derivative can be decomposed as a sum
of compact map and a contraction-a special case of an a-contraction. To
describe the nature of these estimates, we need some notation. For A > 0,
the subspace of J:(E) consisting of all maps L = L 1 + L2 with L1 compact
and IIL 2 II < A is denoted by L;..(E). Given a map L E L;..(E), we define
Ls =LIS and
v;.. (L) =min{ dimS : S is a linear subspace of E and IlLs II < A}.
It is easy to prove that v;.. ( L) is finite for L E J: ;..; 2 (E). The basic result for
the estimate of c(K) is contained in
Theorem 2.2. Let E be a Banach space, U c E an open set, T : U ---+ E

a C 1 -map and K c U a compact set such that T(K) =:> K. If the Prechet
derivative DxT E £ 1;4(£) for all x E K, then
c(K) < log(8[2(A(1 +a)+ k 2 )/Aa] 6 ),

- log[1/2A(1 +a)]
where k = supxEK IIDxTII, 0 < A < 1/2, 0 < a < (1/2A) - 1, 8

supxEK v;..(DxT 2 ). If DxT E £1(K) for all x E K, then c(K) < oo.
We may apply Theorem 2.1 directly to RFDE. In fact, if Tp(t) is the

solution map of RFDE(F), then the map T~f Tp(r) satisfies the conditions
of Theorem 2.1 since Tp(r) is compact. We summarize this remark in
Theorem 2.3. Suppose F E X 1 and A(F) is the set of globally defined and
bounded solutions of RFDE(F). For any (3 > 0, there is a positive number
d!3 suchthatc(A!3(F)) :S;df3, whereA!3(F) =A(F)n{¢E C 0 (I,M): 1¢1 :S;
(3}. If A(F) is the global attractor (which happens if the RFDE is point
dissipative), then c( A( F)) < oo. The same remarks apply to the Hausdorff
dimension.
Of course, there is a transcription of Theorem 2.2 to RFDE and we

leave this for the reader to do.
The finite dimensionality of the sets A{3(F) implies the finite dimen-
sionality of the period module of any almost periodic solution of RFDE(F),
generalizing the same result for ordinary differential equations. Any almost
periodic function x(t) has a Fourier expansion x(t) "' Eanexp ( -iAnt),
where Elanl 2 < oo. The period module of x(t) is the vector space M
spanned by the set { An } over the rationals. Saying the period module
is finite-dimensional is equivalent to saying that the function x is quasi-
periodic. As a consequence of Theorem 2.3, we have the following result.
Theorem 2.4. For a C 1 -RFDE(F) on a manifold M, every almost periodic

solution is quasi-periodic. If RFDE(F) has a global attractor, then there is
an integer N such that the period module of any almost periodic solution
has dimension ::; N.
The conclusion that A(F) has finite dimension depends upon the fact
that the RFDE is at least C 1 . In fact, let QL be the set of functions "Y :
JRn __, JRn with global Lipschitz constant L. For each "Y E QL, each solution
of the ordinary differential equation x = "Y(x) is defined for all t E JR.
Theorem 2.5. For each L > 0, there is a continuous RFDE(F), depending

only on L, such that for every "Y E QL, every solution of the ODE x =
"Y(x) is also a solution of the RFDE(F). In particular, A(F) has infinite
dimension.
12.3 A-stability and Morse-Smale maps
The primary objective in the qualitative theory of discrete dynamical sys-

tems T on a Banach space X is to study the manner in which the flow
changes when T changes. Due to the infinite dimensionality of the space
and the fact that T may not be one-to-one, a comparison of all orbits of two
different maps is very difficult and is likely to lead to severe restrictions on
the maps that are to be considered. For this reason, we restrict our discus-
sion to mappings that have a global attractor and then make comparisons
of orbits on the attractors.
Let cr(X,X), r ~ 1, be the space of cr-maps from X to X. Let
KCr(X, X) be the subset of cr(x, X) with the property that
(i) T E KCr(x, X) implies that T has a global attractor A(T).
(ii) A(T) is upper semicontinuous on KCr(X, X).
For T, S E K cr (X, X), we say that T is equivalent to S, T ,...., S, ifthere is
a homeomorphism h: A(T) __, A(S) such that hT = Sh on A(T). We say
that Tis A-stable if there is a neighborhood V ofT in cr(X, X) such that
T"' S for every S c V n KCr(X, X).
A fixed point x of a map T E cr (X, X) is hyperbolic if the spectrum
of DxT(x) does not intersect the unit circle in <C with center zero. For any
hyperbolic fixed point ofT, let
12.3 A-stability and Morse-Smale maps 373
W 8 (x, T) = { y EX: Tny-+ X as n-+ 00 },

Wu(x, T) = { y E X : r-ny is defined for n ~ 0
and T-ny-+ X as n-+ 00 }.
The sets W 8 (x, T), wu(x, T) are called, respectively, the stable and unsta-
ble sets ofT. If xis a hyperbolic fixed point ofT, then there is a neighbor-
hood V of x such that
W1~c(x,T) = W1~c(x,T, V)~{y E W 8 (x,T): Tny E V,n ~ 0},

W1~c(x, T) = W1~c(x, T, V) ~f { y E Wu(x, T): r-ny E V, n ~ 0}
are cr-manifolds. These sets are called the local stable and unstable mani-
folds. If the maps T and DT are one-to-one on X, then W 8 (x, T), Wu(x, T)
are immersed in X.
A point x E X is a periodic point of period p ofT if TPx = x, T1 x-=/= x,
j = 1, 2, ... , p - 1. A periodic point x of period p is hyperbolic if the
spectrum of DTP(x) does not intersect the unit circle in <C with center
zero.
A map T E KCr(X,X) is Marse-Smale if
(i) T, DxT are one-to-one on A(T).
(ii) fl(T), the nonwandering set ofT, is finite and consists of the periodic
points Per (T) ofT.
(iii) All periodic points are hyperbolic with finite-dimensional unstable
manifolds.
(iv) W 8 (x, T) is transversal to wu(x, T) for all x, y E Per (T).
The following result is basic in stability theory.
Theorem 3.1. IfT E KCr(X,X) is Marse-Smale, then Tis A-stable.
For a Marse-Smale map T E KCr(X,X), the global attractor A(T)

can be written as
A(T) = U wu(x, T).
xEPer(T)
When we apply these results to an RFDE(F), we fix a t 0 > 0 (for exam-

ple, we could take t 0 = r) and then define T =
TF(t 0 ). If Tis Morse-Smale,
then Theorem 3.1 asserts the topological equivalence on the attractors of
the discrete flow defined by the iterates of T to that defined by a map S
close toT. It does not say very much about the continuous flows defined by
two RFDE. If the continuous flows are gradient (that is, the a- and w-limit
sets of any bounded orbit belong to the set of equilibrium points), then the
consideration of the map captures the essential properties of the flow. On
the other hand, if there are periodic orbits for the continuous flow, then
the discrete map will not capture this behavior. If we assume that there
are only a finite number of equilibrium points and periodic orbits for the
continuous flow, each of which is hyperbolic, all stable and unstable mani-
folds intersect transversally, and the nonwandering set consists only of the
equilibrium points and periodic orbits, then we would like to call the contin-
uous flow a Marse-Smale flow. It is reasonable to expect that a continuous
version of Theorem 3.1 is true, but no one has written the proof.
We now present an example of a nontrivial Morse-Smale gradient sys-
tem for which we can give a considerable amount of information about the
flow on the attractor.
Suppose bE C 2 ([-1,0],lR),b(-1) = O,b(O) > O,b'(O) ~ 0, b"(B) ~ 0
for(} E (-1,0] and
(3.1) b"(Bo) > 0 for some Bo E [-1, 0].
Let g E C 1 (lR, lR) be such that
(3.2) G(x) ~f 1x g(s) ds ____, oo as lxl___, oo
and consider the equation
0
(3.3) x(t) = - /_ 1 b(O)g(x(t +B)) dO.
Let Tb,g(t)</> ~f T(t)¢ be the solution of (3.3) with initial value </> at
t = 0. We say that Ab,g is the minimal global attractor for (3.3) if Ab,g is
invariant and attracts any bounded set of C. We have the following result.
Theorem 3.2. If (3.1), (3.2) are satisfied, then there exists a locally com-
pact minimal global attractor for System (3.3). Furthermore, System (3.3)
defines a gradient flow with the w-limit set of any orbit being a single equi-
librium point. If the zero set E of g is bounded, then Ab,g is compact. If, in
addition, each element of E is hyperbolic, then dim wu(xo) = 1 for each
xo E E and
(3.4) Ab,g = U wu(x 0 ).

xoEE
In the proof of Theorem 3.4 of Chapter 5, we used a Liapunov func-

tional to show that positive orbits of bounded sets are bounded and that
the w-limit set of any orbit belongs to the set of equilibrium points, that
is, the zeros of g. Therefore, if we define Ab,g = Ur> 0 w(B(O,r)), where
B(O, r) is the ball in C of center 0 and radius r, then it is clear that Ab,g
is the minimal global attractor. The existence of the Liapunov functional
V shows also that the a-limit set of any orbit in Ab,g belongs to the set of
equilibrium points. If the zero set E of g is bounded, then System (3.3) is
12.3 A-stability and Morse-Smale maps 375
point dissipative and there is a compact global attractor. This set is Ab,g·
If xo is an equilibrium point (that is, g(x 0) = 0), we consider the linear
variational equation about xo:
(3.5) y(t) =- j_ b(O)g'(x )y(t + 0) d(}.

0
1
0
The characteristic equation is
(3.6)
It is possible to show that the equilibrium point x 0 is hyperbolic if g'(xo) =f 0

and x 0 is asymptotically stable if g'(x 0 ) > 0 and unstable if g'(xo) < 0.
Furthermore, if g'(xo) < 0, it is possible to show that dim wu(xo) = 1.
Therefore, if all of the equilibrium points are hyperbolic, we have (3.4). Since
thew-limit set of an orbit is connected, the only situation in which this limit
can contain more than one equilibrium point Xo is when g'(xo) = 0. In this
case, the only characteristic value is A = 0, and it is simple. It follows from
general convergence theorems for gradient systems that the w-limit set is
a singleton (see the supplementary remarks for references). This completes
the proof of the theorem.
Let us now suppose that the zero set E of g is finite and each zero of g
is simple. In this case, the compact global attractor Ab,g is represented as
in Equation (3.4) and is one-dimensional. If g'(x 0 ) < 0, then there are two
orbits in C leaving the point x 0 ; that is, there are two distinct solutions
¢(t), 'lj;(t) of System (3.3), defined for t ::::; 0 that approach x 0 as t-+ -oo.
The problem is to determine the limits of ¢(t), 'lj;(t) as t -+ oo. We know
that these limits must be equilibrium points of System (3.3). It is natural
to suspect that one of these limits will be larger than x 0 and one will be
less than xo (that is, the ordering of the real numbers is preserved on the
attractor). If E = { x1, x2, X3} with x1 < x2 < X3, this is obviously true
since Ab,g is connected.
It is a surprising fact that this is not in general true if the set E consists
of five zeros of g. To describe this situation, it is convenient to systematize
the notation. Let us use the symbol j[k, l] to mean that the unstable point
Xj is connected to Xk, X! by an orbit. If g has only the five simple zeros,
x1 < x2 < X3 < X4 < xs, then x2, X4 are unstable with one-dimensional
unstable manifolds and x1, x2, x 3 are asymptotically stable. The flow on
Ab,g is determined by the manner in which the points x2, X4 are connected
by orbits to the other equilibrium points. The main result is the following.
Theorem 3.3. For a given function b, one can realize each of the following
flows on the attractor .A.b,g by choosing an appropriate function g with five
simple zeros:
(i) 2[1, 3], 4[3, 5],
(ii) 2[1, 4], 4[3, 5],
(iii) 2[1, 5], 4[3, 5],
(iv) 2[1,3],4[2,5],
(v) 2[1, 3], 4[1, 5].
The only situation in which the flow on .A.b,g preserves the natural order
of the reals is case (i) (it is also Morse-Smale). Case (ii) has a nontransverse
intersection between wu(x2) and W 8 (x4) and case (iv) has a nontransverse
intersection between Ws(x2) and wu(x 4). In cases (iii) and (v), there is
transversal intersection of the stable and unstable manifolds (thus, Marse-
Smale) and the natural order of the reals is not preserved by the flow on
the at tractor.
12.4 Hyperbolicity is generic
The aim of the generic theory of differential equations is to study qualitative

properties of solutions that are typical in the sense that they hold for all
equations defined by functions of a residual set of the function space being
considered. More precisely, if X is a complete metric space, then a property
P on the elements x E X is said to be generic if there is a residual set Y c X
such that each element of Y has property P. We recall that a residual set
is either an open dense set or, more generally, a countable intersection of
such sets.
For any Banach space E and any coo n-dimensional manifold M with a
complete Riemannian metric, we let Ck(E, M) denote the space offunctions
from E to M that are continuous together with derivatives up through
order k. The metric on Ck(E, M) is the one induced by the Riemannian
metric on M, taking into account the differences between functions and
their derivatives up through order k.
For ordinary differential equations, the most basic result in the generic
theory is the Kupka-Smale theorem. To be specific, for the ordinary differ-
ential equation
(4.1) x= f(x),
where f E Ck(IRn,IRn), let P be the property that all critical points (equi-
librium points) and periodic orbits are hyperbolic and the stable and un-
stable manifolds intersect transversally. The Kupka-Smale theorem asserts
12.4 Hyperbolicity is generic 377
that this property P is generic. This theorem shows that the hyperbolicity
requirement in a Morse-Smale system is generic.
The complete proof of the Kupka-Smale theorem for retarded FDE is
not available at the present time. However, the following result is true.
Theorem 4.1. The setoff E Ck(C,ffin), for which all critical points and
all periodic orbits of the RFDE,
(4.2) x(t) = f(xt),
are hyperbolic, is residual.
We give an indication of the proof, emphasizing only those parts that

are distinctly different from the one for ordinary differential equations. We
recall that an equilibrium point is nondegenerate if zero is not an eigenvalue
of the corresponding linear variational equation. A periodic orbit defined
by a periodic function p(t) is called nondegenerate if the linear variational
equation for p(t) has 1 as a simple characteristic multiplier. For a fixed
compact set K c m,n and a fixed T > 0, we define the following subsets of
Ck(C,ffin):
Q0 (K) ={!:all critical points inK of (4.2) are nondegenerate },

Q1 (K) ={!:all critical points inK of (4.2) are hyperbolic},
93 ; 2 (T,K) = {! E Q1(K): all periodic orbits of (4.2) inK and having
a period in (0, T] are nondegenerate}
Q2(T,K) = {! E Q1(K): all periodic orbits of (4.2) inK and having
a period in (0, T] are hyperbolic}.
For each T, each of the sets Q0 (K), Q1(K), Q3 ; 2 (T,K), Q2(T,K) is open
as a consequence of the results in Chapter 10. If for each T, we show that
Q2(T, K) is dense, then it follows that Q2(K) = n'N= 1Q2(N, K) is residual,
and this will show that Theorem 4.1 holds at least for those solutions lying
in K. By taking a sequence Krn of compact sets whose union is ffin, and
intersecting the corresponding residual sets Q2(Krn), the theorem is proved.
To show that Q0 (K) and Q1 (K) are dense, first take any f E X 1 and
make two perturbations as follows:
(i) By Sard's theorem, there is an jl near f for which all critical points
inK are isolated and nondegenerate. Thus, j1 E Q0 (K).
(ii) By perturbing locally near each critical point, one may obtain P near
f 1 such that each critical point is hyperbolic. Thus, P E Q1(K).
To show that Q3 ; 2(T, K) and Q2 (T, K) are dense for any T, we begin
with some f E Q1(K). One may easily find a lower bound To> 0 for periods
of periodic solutions of Equation (4.2) in K and so, trivially, f E Q2(T, K)
for T < T 0 . An induction argument on the period T is then used. This

induction also involves two steps.
(a) An argument involving Sard's theorem can be used to prove that
(h(T,K) is dense in Q3 ; 2 (3T/2,K).
(b) A local perturbation around each nondegenerate periodic orbit of
some f E Q3 ; 2 (T, K) yields a hyperbolic orbit, thereby showing that
Q3 ; 2(T, K) is dense in Q2(T, K).
The most difficult step is (a), since we must be concerned with all of
the periodic solutions with period in (T, 3T /2]. These may not be nonde-
generate. Let x(t) be a periodic solution with least period T E (T,3T/2]
and consider the map
P: (¢, t,g)-+ Xt(r/J,g)- r/J,
where Xt (¢,g) E C is the solution of
(4.3) x(t) = g(Xt), Xo = r/J.

Clearly, zeros of <P correspond to the initial data for periodic solutions of
Equation (4.3). Therefore, <P(x 0 , T, f) = 0. What is of particular interest are
the periodic solutions in a neighborhood of (x 0, T) for gin a neighborhood
of f. To study the zeros of <P, the Fnkhet derivative of <P must be analyzed
and, in particular, the derivative r = Dg<P(xo, T, f) with respect to the
variable g E X 1 . This leads to the study of the inhomogeneous variational
equation
(4.4)
about the known periodic solution x(t) since r : X 1 -+ c is given by

(Fh)(t) = Yt· For simplicity, suppose that T > r. In order to use Sard's
theorem, we must verify that the range of the linear map r is dense inC:
that is, for any 'lj; E C and any E > 0, there is an h E X 1 such that the
solution of (4.4) satisfies IYr - 'l/JI < E. We may always choose such an h of
the form
-r -2r
(4.5) h(¢) = H(¢(0), ¢( N ), ¢( N ), ... , ¢( -r))
for some N, because of the following result.
Lemma 4.1. Let x(t) be a periodic solution of (4.2) of least period T > 0.
Then, for sufficiently large N, the map
r 2r
'Y(t) = (x(t),x(t- N),x(t- N), ... ,x(t- r))
is a one-to-one regular (that is, i'(t) =/: 0 for all t) mapping of the reals mod
T into lRn(N+I).
12.5 One-to-oneness on the attractor 379
We remark that it is possible to prove a generalization of Theorem 4.1

by using the Ck-Whitney topology on the functions from C to lR.n.
It is interesting to restrict the class of functions X 1 . For example, sup-
pose that the systems under investigation are differential difference equa-
tions of the form
(4.6) x(t) = F(x(t),x(t -1)).

To obtain a generic theorem about this restricted class of equations is more
difficult since there is less freedom to construct perturbations. For instance,
the function h in (4.5) cannot be used. Nevertheless, Theorem 4.1 remains
true for these equations. Instead of using Lemma 4.1 to produce h, one first
approximates Equation (4.6) with an analytic F and then uses the following
Lemma 4.2. If x(t) is a periodic solution of Equation (4.6) of least period

r > 0 and F is analytic, then the map
'Y(t) = (x(t),x(t -1))
is one-to-one and regular except at a finite number oft values in the reals
modr.
One may consider an even more restrictive class of equations of the

form
x(t) = F(x(t- 1)).
For this class of equations, Theorem 4.1 is not known.
It would be very interesting to obtain the complete Kupka-Smale theo-
rem for any of these situations, that is, assuming that there is a residual set
of f such that the conclusions of Theorem 4.1 hold and the stable and un-
stable manifolds of critical points and periodic orbits intersect transversally.
Some ideas distinctly different from the ones used for ordinary differential
equations seem to be required.
12.5 One-to-oneness on the attractor
The following result is almost obvious.
Proposition 5.1. Let F E Xk, k ;::: 1. If A is a compact invariant set of

RFDE(F) and Tp(t) is one-to-one on A, then Tp(t) is a continuous group
of operators on A.
In particular, if A( F) is the global attractor for Tp(t) and Tp(t) is one-

to-one on A(F), then Tp(t) is a group on A(F). This is certainly sufficient
reason to study the following question: When is Tp(t) one-to-one? As we
have noted in Chapter 3, this need not be true on all of C. Therefore, it is

natural to ask if perhaps the property of one-to-oneness is generic. At this
time, this question has not been answered completely, but there are some
results, which are discussed in this section.
Theorem 5.1. For an analytic RFDE(F), any globally defined bounded so-
lution x(t), t E lR, is analytic. In particular, Tp(t) is one-to-one on the set
A(F) of globally defined bounded solutions.
For a class of linear nonautonomous RFDE, analyticity is not required.

Let .C = .C(C, lRn) be the Banach space of all continuous linear mappings L:
C-+ lRn with the usual norm and let C 1 (1R, .C) be the space of continuously
differentiable mappings from lR to .C with the uniform C 1 topology on
compact sets of JR. It is possible to prove the following result.
Theorem 5.2. The set of L E C 1 (lR, .C) such that the corresponding solu-
tion operator of RFDE(L) is one-to-one on lR is dense. For any compact
set K C lR, the set of L E C 1 (1R, .C) for which the solution operator of
RFDE(L) is one-to-one on K is not open.
It is not known if the set of L in Theorem 5.2 is residual. For a more

restricted class, better results are obtained. Let V(lR) C C 1 (lR, .C) be the
set of L such that there is an integer N such that the corresponding measure
ry(t, e) of L has at most N discontinuities in e for all t E lR and, in addition,
ry(t, e) is a step function in e. For any compact set K c lR, the set V(K) is
defined in a similar way.
Theorem 5.3. The set of L in V(JR) for which the solution operator of
RFDE(L) is one-to-one on lR is residual. Also, for any compact set K c lR,
the corresponding set is open in V(K).
In the case where we restrict the solution map Tp(t) to the global
attractor, the one-to-oneness can be related to other interesting concepts.
If E, X are Banach spaces and B(e) C X, e E E, is a family of subsets, then
we say that B(e) is lower semicontinuous at eo if dist (B(eo), B(e))-+ 0 as
e-+ e0 . We recall that the distance between two subsets B, Cis defined by
dist (B, C) = supxEB infyEC llx- Yll·
Theorem 5.4. If M is a compact manifold and F0 E X 1 is given, then the

following conclusions hold:
(i) There is a compact set K :J A(F) such that F0 can be uniformly ap-
proximated on K by a C 1 -RFDE(F) for which Tp(t) is one-to-one on
A( F).
(ii) If F0 is A-stable, then Tp0 (t) is one-to-one on A(Fo).
(iii) If F0 is given and A( F) is lower semicontinuous at Fa, then Tp0 (t) is
one-to-one on A(Fo).
12.5 One-to-oneness on the attractor 381
For delay differential equations, this result can be modified in the fol-
lowing way.
Theorem 5.5. If M is a compact manifold and Fo E X 1 is a given de-

lay differential equation with a finite number of delays, then the following
statements hold:
(i) Fo can be uniformly approximated by a C 1 -delay differential equation
F with the same number of delays and Tp(t) is one-to-one on A(F).
(ii) If F0 is A-stable, then Tp0 (t) is one-to-one on A(Fo).
(iii) If Fo is given and A( F) is lower semicontinuous at Fo, then Tp0 (t) is

one-to-one on A(Fo).
The proof of these theorems use some auxiliary results of independent

interest.
Proposition 5.2. IfF, G E XI, F"' G and Tp(t) is one-to-one on A(F),

then Tc(t) is one-to-one on A( G).
This proposition yields immediately assertion (b) in the theorems.
Proposition 5.3. Suppose that Fi, FE Xl, i 2: 1, and that Fi -+ F as

i-+ oo. If A( F) is lower semjcontinuous in F and Tpi (t) is one-to-one on
A(Fi), then Tp(t) is one-to-one on A(F).
These propositions, together with the Stone-Weierstrass approximation

theorem, are the main ingredients of the proof of Theorems 5.4 and 5.5.
We remark that the conclusions of Theorem 5.4 remain valid for RFDE
on noncompact manifolds. In this case, we restrict consideration to the class
of RFDE that has a global attractor that is upper semicontinuous.
An RFDE is said to be gradient-like if all positive orbits are bounded,
the w-limit set of each positive orbit and the a-limit set of each bounded
negative orbit belongs to the set E of equilibrium points. If E is bounded,
then there is a global attractor. It is possible to prove the following result.
Theorem 5.6. If an RFDE(F0 ) on a compact manifold M is gradient-like

and all equilibrium points are hyperbolic, then A( F) is lower semicontinuous
at F0 . Thus, A(F) is Hausdorff continuous in F from Theorem 5.5.
Theorem 5.6 is intuitively obvious for the following reason. If the stable
and unstable manifolds were transversal, then the RFDE(F) would be A-
stable and, in particular, A( F) is Hausdorff continuous in F. If the stable
and unstable manifolds are not transversal along an orbit "YF whose a-limit
set is c/J- and whose w-limit set is¢+, then, under a small perturbation G of
F, and for any neighborhood U of "(F, there is an orbit "(G = UtEIR Ta(t)¢
of the RFDE(G) such that T 0 (t)¢ remains in U for as long as we like. This
certainly suggests that A( F) should be lower semicontinuous. Supplying a
precise proof requires considerably more effort.
12.6 Morse decompositions
Suppose that T 0 (t), t 2:: 0, is a C 0 -semigroup on a Banach space X for

which there is a compact global attractor A. A Morse decomposition of the
attractor A is a finite ordered collection A 1 < A 2 < · · · < AM of disjoint
compact invariant subsets of A (called Morse sets) such that for any ¢ E A,
there are positive integers Nand K, N 2:: K, such that a(¢) C AN and
w(¢) C AK and N = K implies that¢ E AN. In the case N = K, we have
T(t)¢ E AN fortE JR.
The Morse sets, together with the connecting orbits
for N > K, give the global attractor A.

There are two obvious Morse decompositions, namely, the set A itself
and the empty set. Neither of these decompositions is interesting. A Morse
decomposition becomes important when it gives some additional informa-
tion about the flow defined by the semigroup restricted to the attractor. In
this section, we consider in some detail a Morse decomposition for a special
class of differential difference equations. We only give some ideas of the
proofs; the reader may consult the references in the section on supplemen-
tary remarks for details.
(6.1) i:(t) = -(3x(t)- g(x(t- 1)),
where (3 2:: 0 and the following hypotheses are satisfied:

(HI) g E C 00 (IR, IR) and has negative feedback; that is, xg(x) > 0 for
x =/=- 0 and g'(O) > 0,
(H 2 ) there is a constant k such that g(x) 2:: -k for all x,
(H 3) the zero solution of Equation (6.1) is hyperbolic.
The results to be stated hold in more general situations. The right-
hand side of Equation (6.1) can be replaced by f(x(t),x(t- 1)) with a
modified definition of negative feedback. Hypothesis (H 2 ) can be replaced
by the assumption that the semigroup associated with Equation (6.1) is
point dissipative. We need only the existence of a compact global attractor.
We will prove that (H 1 ) and (H 2 ) imply this property. Hypothesis (H3) is
not necessary, but to eliminate it requires more complicated definitions.
12.6 Morse decompositions 383
In Chapter 11, and especially in the supplementary remarks of that

chapter, we mentioned several examples that have the form (6.1) and have
negative feedback.
Lemma 6.1. Under hypotheses (Hl) and (H 2 ), the semigroup T(t) generated
by Equation (6.1) is a bounded map and is point dissipative. Thus, there is
a compact global attractor A.
Proof. By integrating Equation (6.1) over [0, 1], one deduces that T(1) is a
bounded map. Thus, T(t) is a bounded map for any t > 0.
To show that T(t) is point dissipative, we observe that
for all t :::: 0. Therefore,
x(t):::; x(O)e-,6t + k(1- e-,Bt)/(3

and limsupt--> 00 x(t):::; 2k/(3. Since x(t) is bounded above, it follows that
-g(x(t- 1)) is bounded below by a constant K 1 . Therefore, arguing as
earlier, one obtains that lim inft-+oo x(t) :::: -2Kl/ (3. This shows that T(t)
is point dissipative and Corollary 4.3.2 completes the proof of the lemma.
D
We now describe an interesting Morse decomposition of the flow on

the attractor A defined by the semigroup T(t) corresponding to (6.1). To
do this, it is convenient to think of the flow on A in the following way. For
any ¢ E A, we know that T(t)¢ E A for all t E ffi. Since (T(t)¢)(8) =
(T(t+B)¢)(0) fore E [-1,0], the orbit T(t)¢, t E ffi can be identified with
the function x(t, ¢) = (T(t)¢)(0), t E ffi. With this observation, we make
the following definition. For any ¢ E A, ¢ "1- 0, let u :::: t be the first zero
of x(t, ¢) in [t, oo), if it exists. We define V(T(t)¢) as the number of zeros
of x(t,¢) (counting multiplicity) in the half-open interval (u -1,u]. If u
does not exist, we define V(T(t)¢) = 1. Thus, V(T(t)¢) is either a positive
integer or oo. We will refer to V(T(t)¢) either as the Liapunov function for
(6.1) or as the zero number of x(·, ¢).
Theorem 6.1.
(i) V(T(t)¢) is nonincreasing in t for each¢ E A,¢ "1- 0.
(ii) V(T(t)¢) is an odd integer for each¢ E A, ¢ "1- 0.
(iii) There is a constant K such that V(T(t)¢) :::; K for all t E ffi, ¢ E A,
¢ -I- 0.
We give only some of the intuitive ideas of why Theorem 6.1 is true.
The complete proof is very difficult and the interested reader should consult
the references. Let us first suppose that the zeros of x(t, ¢) are simple. Let
O'o < 0' 1 be consecutive zeros with x(t, ¢) > 0 in between. Then x(O'o, ¢) > 0
and x(0' 1 , ¢) < 0. From the negative feedback condition (HI), it follows that
x(O'o- 1, ¢) < 0 and x(0' 1 - 1, ¢) > 0. Thus, x(t, ¢) = 0 at some point in
(0' 0 -1, 0' 1 -1); that is, x(t, ¢)can have no more zeros in (0' 1 -1, 0' 1 ] than it
does in (O'o -1, 0'0 ]. This shows that V(T(t)¢) is nonincreasing in t. Again, if
we assume that the zeros of x(t, ¢) are simple, then x(O', ¢) = 0, x(O', ¢) > 0
(resp. x(O',¢) < 0) imply that x(0'-1,¢) < 0 (resp. x(0'-1,¢) > 0), which
in turn implies that the number of zeros of x(t, ¢)in (0'- 1, 0'] is odd.
If the zeros of x(t, ¢)are not simple, one first proves that V(x(·, ¢) < oo
for ¢ E A,¢ -1- 0. This requires several technical estimates. Also, if ¢ E
A,¢ -1- 0 has a zero of order exactly k at t = 0', then it is easy to see that
t = 0' -1 is a zero of exactly order k -1 and Dk- 1 x(O' -1, ¢)Dkx(O', ¢) < 0.
The proofs of (i) and (ii) are completed by noting sign changes near the
zeros.
The proof of property (iii) in the theorem is technical and difficult.
The difficulties arise in the determination of the behavior of the flow on A
near the origin. We can recognize the problem even in the case where the
origin is hyperbolic; that is, the solutions of the characteristic equation for
the linear variational equation around zero of (6.1),
(6.2) .A+ (3 + g'(O)e->- = 0,
have nonzero real parts. If ¢ E A, ¢ -1- 0 and the solution x( t, ¢) stays

in a small neighborhood of the origin for t ::; -T, then it must lie on the
unstable manifold wu of zero. Since this set is finite-dimensional, the solu-
tion will approach the origin as t ____, -oo along an eigenspace of the linear
equation and therefore should have the same type of oscillatory properties
as the eigenfunctions. Therefore, we should have an integer N* such that
V(x(t, ¢)::; N* for all¢ E wu \ {0}. If x(t, ¢)remains in a small neighbor-
hood of the origin fort ;::: T, then it must lie on the stable manifold ws of 0.
Again, one would expect that this solution would approach the origin along
one of the eigenspaces of the linear equation. If this were the case, then
the oscillatory properties would be the same as those on the eigenspace.
However, since wu is infinite-dimensional, this is far from obvious and it is
conceivable that there is a solution that approaches zero faster than any ex-
ponential (the so-called small solutions mentioned in Chapter 3). It is true
that no such solutions exist, but the proof is very difficult. Even knowing
this does not prove part (iii) of Theorem 6.1. If one takes into account the
fact that the orbits of (6.1) that are of interest lie on the compact attractor
A, then one can prove the following result.
12.6 Morse decompositions 385
Theorem 6.2. If the origin is hyperbolic and dim wu = N*, then N* is

even and there is a neighborhood U C A of the origin such that
¢E ws \ {0} implies V(x(t, ¢)) > N* for all t E IR,
¢ E wu \ {0} implies V(x(t, ¢)) < N* for all t E IR.

In particular, ws n wu = { 0 }, and so there is no orbit homoclinic to the
ongm.
Using Theorem 6.2, one can show part (iii) of Theorem 6.1 in the case
when the origin is hyperbolic. If the origin is not hyperbolic, more care is
needed.
With these results, we are now in a position to define a Morse decom-
position of the attractor A. It is tempting to consider, for each odd integer
N, the following sets as part of a Morse decomposition:
{ ¢ E A, ¢ =/= 0 : V (x (t, ¢)) = N for all t E IR } .
However, this will not work because the function V is not defined at the
origin and these sets in general are not closed. In fact, several of them may
contain the point 0 in their closure. The definition must be refined to keep
the orbits away from the origin.
We assume that the origin is hyperbolic (the definition can be given
without this hypotheses but is more complicated) and let N* =dim wu(o).
For any odd integer N, define
AN={¢ E A,¢=/= 0: V(x(t, ¢)) = N fortE IR and 0 FJ_ a(¢) Uw(¢) }.
Let AN* = { 0 }. With this definition, the sets AN for N odd are compact
and do not contain the origin. We remark that AN = 0 for large N by (iii)
of Theorem 6.1. It is now possible to prove
Theorem 6.3. If the origin is hyperbolic, the sets AN, N E { N*, 1, 3, 5, ... },
form a Morse decomposition of A with the ordering AN < AK if and only
if K < N.
Further properties also are known about the Morse sets AN. In particu-
lar, for Nan odd integer, if¢ E AN, then the zeros of x(t, ¢)are simple. This
allows one to prove that each AN for N < N* is not empty and contains
a periodic orbit xN(t) with least period T satisfying 2/N < T < 2/(N- 1)
and XN(t) has exactly two zeros in [0, T).
The proof of this last fact uses a special type of Poincare map. Consider
the map e : AN ----+ S 1 from the Morse set AN to the unit circle S 1 in the
plane with center 0, induced by the map
¢ E AN~ (x(O, ¢), x(O, ¢)) E IR 2 \ { 0 }.

From the properties mentioned earlier, the image of the orbit winds around
the circle infinitely often as t ---> ±oo. In particular, it has a transversal
cross section, namely, the half-line x = 0, ± > 0 in lR? \ { 0} and has
a corresponding Poincare map. It is this map that is used to prove the
existence of the periodic solutions mentioned before.
In order to obtain more information about the structure of the flow on
the attractor A, it is first necessary to understand the existence of connect-
ing orbits Cff for various Nand K. Using the Conley index and the theory
of connection matrices for isolated invariant sets, it has been recently shown
that CJI #- 0 for all N > K (see the supplementary remarks for references).
As we have seen, the existence of the Morse decomposition and of the
connecting orbits gives a much better picture of the flow on the attractor A.
This does not mean that the flow is simple. In fact, numerical studies suggest
that, in many cases, the flow in the set AN may have a very complicated
structure involving multiple periodic orbits arising from period-doubling
bifurcations and even chaotic dynamics (see the supplementary remarks for
references).
12.7 Singularly perturbed systems
(7.1) E±(t) = -x(t) + f(x(t- 1)),
where E > 0 is a parameter and f E C 1 (lR,lR).

If we formally take the limit of Equation (7.1) as E---> 0, then we obtain
a difference equation
(7.2) x(t) = f(x(t -1)),
which can be considered as a discrete dynamical system defined by the map
(7.3) x f---t f(x).
It is an interesting problem to determine how the dynamics of Equation

(7.1) mirror the dynamics of the difference equation (7.2) or the discrete
dynamical system (7.3) when E is small. In this section, we investigate some
of the known similarities and dissimilarities.
For any interval I C lR (closed or open), let X1~C([-1,0],I). We
let T, (t) denote the semigroup on C. The following result is very easy to
prove.
Proposition 7.1. (Positive invariance). If I is an interval such that f(I) C I,

then T,(t)XI c X1 fort 2 0.
12.7 Singularly perturbed systems 387
If x 0 is a fixed point of f, then the constant function Xo E C is an

equilibrium point of Equation (7.1), and conversely. If x 0 is an attracting
fixed point off, we say that an interval J is the maximal interval of attrac-
tion of x 0 if x 0 E J, f(J) C J, fn(x) -+ x 0 as n-+ oo for each x E J and
there is no interval J' => J with this property. We remark that the maximal
interval of attraction is open. It is possible to prove the following property.
Proposition 7.2. (Stability). If xo is an attracting fixed point off with max-

imal interval of attraction J, then the equilibrium solution Xo of Equation
(7.1) is asymptotically stable and, for each 'ljJ E XJ and every E > 0, we
have
lim TE(t)'lj;
t-HXl
= xo.
Under the assumption that f has negative feedback, we have seen in

the previous section that there is a Morse decomposition of the attractor.
Also, in the supplementary remarks to Chapter 11, we asserted that for
each E > 0, there is a slowly oscillating periodic solution if the origin is
unstable and if there is an interval I such that f(I) C I. It is natural to
discuss the limit of this solution as E -+ 0. To state a precise result, we say
that a point (a, b) is a period-two point off if a -=1- band f(a) = b, f(b) =a.
We say that a function w(t), t E IRis a square wave if there are constants
a -=1- b such that w(t) =a fort E (2n, 2n + 1), w(t) = b for t E (2n + 1, 2n)
for all integers n.
For the statement of the next result, we recall (see Section 11.7) that
a solution of (7.1) is slowly oscillating (about zero) if it has a sequence of
zeros approaching infinity and the distance between zeros is > 1.
Theorem 7.1. Suppose that there is an interval I such that 0 E I, f(I) C I,

f has negative feedback on I, and f'(O) < -1. Then there exists an Eo > 0
such that for 0 < E < Eo, there is a slowly oscillating periodic solution x€ of
Equation (7.1) that is continuous in E.
Furthermore, if f'(x) < 0 for x E I and (a, b) is a period-two point of
f in I that is asymptotically stable, then xE(t) has exactly one maximum
and one minimum over a period and approaches a square wave uniformly
on all compact sets ofiR \ { n = 0, ±1, ±2, ... } with the values (a, b) of the
square wave corresponding to the period-two point of the map f.
It might be expected that the conclusion of Theorem 7.1 would remain

true without the severe restriction that f is monotone on the interval I.
However, this is typically the exception. If f is not monotone, it can be
shown that the function xE(t) approaches a square wave uniformly on all
compact sets of IR \ { n = 0, ± 1, ±2, ... } with the values (a, b) of the square
wave corresponding to a period-two point of the map f. However, at the
points of transition near the integers, the function x€ (t) begins to oscillate
with the number of oscillations increasing to infinity as E --+ 0. The ampli-

tudes of the oscillations around the point a (resp. b) are bounded but do
not approach zero as E --+ 0. Thus, the limiting process exhibits a Gibbs'
type of phenomenon at the integers. This fact is easily observed numerically
but very difficult to prove. However, the underlying reason for the Gibbs
phenomenon has a very simple dynamical and geometric interpretation. It
is possible to write down some equations that serve to determine the tran-
sition curves that allow the solutions to pass from point a to point b. These
equations are essentially the same as Equations (7.10). The problem is to
determine the constant r so that there is a solution of these equations that
has a (resp. b) as its a-limit set (resp. w-limit set). If the function f is
monotone, then the dominant eigenvalue near a (resp. b) is real and the
transition curve should be monotone (and, thus, the conclusion of the theo-
rem). On the other hand, iff is not monotone, then the dominant eigenvalue
is complex and the transition curve will oscillate. See the bibliography for
details.
If the mapping f were to be a function of a parameter >., the period-
two points often arise through a period-doubling bifurcation from a fixed
point. Therefore, it is of interest to understand the implications of a period-
doubling bifurcation of the map on the dynamics of the flow of Equation
(7.1) forE near zero. We now describe this situation more precisely.
ForE> 0 and small and f E Ck(JR x IR), k 2: 3, we consider periodic
solutions of the equation
(7.4) ei;(t) = -x(t) + f(x(t- 1), >.)
under the assumption that the point >. = 0 corresponds to a generic period-
doubling point for the map x f-+ f(x, 0). More specifically, we assume that
(7.5) f(x, >.) = -(1 + >.)x + ax 2 + bx 3 + o(x3 ) as x --+ 0,
where a, bare constants such that (3 = a2 +b ¥- 0. Under this assumption on

f, for each small value of >. for which >.(3 > 0, there are nonzero constants
dv,, d2>., d1>. ¥- d2>., such that f(dl>., >.) = d2>., f(d2>., >.) = d1>. and so
j2(d 1>., >.) = d1>.. Furthermore, d1>., d2>. --+ 0 as >. --+ 0. The points dl>., d2>.
are periodic points of period two of the map f(·, >.). If (3 > 0, we say that
the bifurcation is supercritical (the fixed point 0 of the map f ( ·, 0) is stable)
and if (3 < 0, we say that the bifurcation is subcritical (the fixed point 0 of
the map f(·, 0) is unstable).
We are interested in how the period-doubling bifurcation of the map
is reflected into the bifurcation from the origin of periodic solutions of
Equation (7.4) of period approximately 2. The principal result is
Theorem 7.2. Suppose that f(x, >.) satisfies (7.5). Then there is a a neigh-
borhood U of (0, 0) in the (>.,E) plane and a sectorial region S in U such
that if(>., E) E U, then there is a periodic solution X>.,€ of Equation (7.4)
with period 2T(A, E) = 2 + 2E + 0(\EI(I>-1 +lei)) as (>.,E) ___, (0, 0) if and

only if ( >., E) E S. Furthermore, this solution is unique. If, in addition,
f(x, >.) =- f( -x, >.), then X>-.,E(t + T(A, E))= -X>-.,E(t).
Of course, the sectorS must belong to the set E > 0 in the (>.,E) plane.
If the period-two doubling bifurcation of the map is supercritical, then the
sector S c {(>.,E) : E > 0, >. > 0} and, for >. = >. 0 > 0, fixed, the set
{E: (E,Ao) E S} is an interval >. 0 x (0,E 0 (>.o)). At the point (>.o,Eo(>-o)),
there is a Hopf bifurcation and the periodic solution approaches a square
x
wave as E ___, 0; that is, the periodic solution \a ,E (t) has the property that
X>-.a,E(t) ___, d 1 >-. (respectively, d 2 >-.) as E ___, 0 uniformly on compact sets of
(0, 1) (respectively, (1, 2)). Part of this result is contained in Theorem 7.1.
In the supercritical case, the sector S is completely different and the
periodic orbits have a different structure as E ___, 0. The sector S contains
points (E, >.) with ,.\ both negative and positive. More precisely, for >. =
>. 0 > 0, fixed, the set {E: (E, >. 0 ) E S} is an interval Ao x (Eo(>-o), /3o(>-o)).
At the point (>. 0 , Eo(>- 0 )), there is a Hopf bifurcation. For >. = Ao < 0,
fixed, the set { E : (E, >. 0 ) E S} is an interval >. 0 x (0, o: 0 (>.o). As E ___, 0,
the unique periodic solution becomes pulse-like in the following sense: the
periodic solution X>-. 0 ,E(t) has the property that X>-. 0 ,E(t) ___, 0 as E ___, 0
uniformly on compact sets of (0, 1) U (1, 2). The magnitude of the pulse
exceeds max{ld 1 >-.l, ldz>-.1 }. The part of the period doubling in the map that
is reflected in the pulse-like solution is that the jumps in the solution occur
near the integers and are opposite in direction.
Let us briefly outline the proof since it makes use of so much of the
local theory that we have developed in the previous chapters. The linear
variational equation around the equilibrium solution 0 of Equation (7.4) is
(7.6) ey(t) = -y(t)- (1 + >.)y(t- 1).

By analyzing the characteristic equation
(7.7) EfJ + 1 + (1 + >.)e-~" = 0,
it is possible to see that if >. ::::; 0, then the origin is asymptotically stable
for all E > 0. On the other hand, if>. > 0, then there is an Eo(>.) > 0 such
that for E > Eo(>.), the origin is asymptotically stable, and, for 0 < E <
Eo(>.), the origin is unstable with a pair of complex solutions of (7.7) with
positive real part. For E = Eo(>.), there are two purely imaginary solutions
of (7.7). Furthermore, if the complex roots near E =Eo(>.) are denoted by
tJ(A, E), Ji(>., E), then 8RetJ(A, Eo(>.))/8E > 0. Therefore, there is a Hopf
bifurcation in Equation (7.4) at the origin at the point(>., Eo(>.)). It can be
shown also that there is a unique periodic orbit bifurcating from the origin
under the assumption that ,6 # 0 and the period is approximately 2.
The basic problem now is to determine the region near the origin in the
parameter space (>., E) for the existence of this bifurcating periodic orbit
and to determine the behavior of this orbit as E --+ 0. To accomplish this,

we introduce some scalings. We suppose that Equation (7.4) has a periodic
solution x(t) with period 2 + 2rE and let
(7.8) w1 (t)= x( -Ert), w2(t) = x( -Ert + 1 + Er ).
Since x(t) has period 2 + 2rE, we see that
w2(t) = x( -Er(t + 1)- 1)
(7.9)
w2(t- 1) = x( -Ert- 1).
If we use (7.8) and (7.9) in (7.4), we deduce that
w1(t) = rw1(t)- rj(w2(t- 1), .X)

(7.10)
w2(t) = rw2(t)- r f(wl (t- 1), .X).
This equation now is independent of E. We now look for periodic solutions
of System (7.10) in a neighborhood of the origin regarding it as a two-
parameter bifurcation problem with (.X, r) as parameters.
Some caution must be exercised at this point. Every periodic solution
x(t) of Equation (7.4) of period 2 + 2rE leads to a periodic solution of Sys-
tem (7.10) through the transformation (7.9). In addition, the corresponding
solution of System (7.10) must encircle the origin. The following converse
also is true: any periodic solution of System (7.10) that encircles the origin
and has period w > 2 corresponds to a periodic solution of Equation (7.4)
of period 2 + 2rE if and only if E satisfies the equation r(w- 2)E = 2.
The next step is to determine the approximate value of the constant r
in the period 2 + 2rE. The appropriate approximate value of r is obtained
by considering the linear variational equation around the zero solution of
System (7.10) for .X= 0,
w1(t) = rw1(t) + rw2(t -1)
(7.11)
w2(t) = rw2(t) + rw1 (t- 1).
The eigenvalues of System (7.11) are the roots of the characteristic equation,
(7.12)
where
(7.13) L¢ = ¢(0) + [~ ~] ¢(-1).

The left-hand side of Equation (7.12) always has 1-L = 0 as a zero. It is a
simple zero if r =/= 1 and a double zero if r = 1. Bifurcation from a simple
zero can never lead to any periodic orbits. Therefore, we are forced to take
r = 1 in the first approximation. For r = 1, the remaining eigenvalues of
(7.11) have negative real parts. If we let r = 1 + h, w = (wb w2), where h
is a small parameter, then (7.10) can be written as
(7.14)
where
¢2( -1) + f(¢2( -1), .X)]
(7.15) F.x,h(¢) = (1 +h) [ ¢1(-1) + f(¢1( -1), .X) '
and Wt(O) = w(t + 0) for -1 ::=; (} ::=; 0.

We now consider Equation (7.14) as a perturbation of the linear equa-
tion
(7.16) v(t) = Lvt.
Of course, we will consider Equation (7.14) with initial data in the space
C = C([-1,0],1R?). Since the characteristic eqlfation for the linear part
of (7.14) for (.X, h) = (0, 0) has a zero as a root of multiplicity two, we
know that the small periodic orbits of Equation (7.14) will lie on a two-
dimensional center manifold that is tangent to the subspace generated by
generalized eigenvectors ( 1, ( 2 associated with the eigenvalue zero of (7.16).
Therefore, the first problem is to determine the approximate vector field
on the center manifold. If we let Wt = z1(1 + z2(2 + Wt, where Wt lies in
the natural linear space complementary to the span of ( 1, ( 2, then we show
that the approximate flow on the center manifold is given by the system of
ordinary differential equations (there are several nontrivial computations
here)
i1 = 2hz1 + 2.X( 32 z1 + z2)- 2j3( 32 z1 + z2) 3 -a22 3

(3z1 + z2)z1
(7.17)
Z2 = -Z1.
For (h, .X) = (0, 0), we can use the theory of normal forms to make a
nonlinear change of variables to obtain the equation
Z1 = (2h + .X~)z1 + 2.Xz2- 2/3z~- 4f3z1z~

(7.18)
up through terms of order (h + .X) 2Izl + lzl 4 .

To analyze the periodic solutions of System (7.18), it is now convenient
to rescale variables
(7.19)
to obtain the new equations
(7.20)
U1 = -~t(28 + (sgn.X)~)u1- 2(sgn,X)u2 + 2j3u~ + 4/3/-LU1U~
U2 = U1.
Equation (7.20) is equivalent to the second-order scalar equation

(7.21) W+ p,(26 + (sgnA)~)W + 2((sgnA)W- ,6W 3 ) - 2p,,6W 2 W = o,
where we have put W = u 2 / v'2. For p, = 0, this is a conservative system
(7.22) W+ 2((sgnA)W- ,6W 3 ) = 0.
The bifurcation diagram in (A, h) space for the periodic orbits of Sys-
tem (7.18) (or the (p,, h) space for (7.20)) are well known. In spite of this
fact, there is still more work to do. In fact, we have remarked before that
not all periodic orbits of System (7.18) are valid candidates for periodic or-
bits of Equation (7.4). We must seek those periodic orbits of System (7.18)
that encircle the origin and have period > 2.
In the case ,6 > 0, there are periodic orbits of Equation (7.21) only
if A > 0. Each periodic orbit of Equation (7.21) encircles the origin and
has period > 2 and, therefore, corresponds to a periodic orbit of Equation
(7.4). The periodic orbits that approach a square wave correspond to pe-
riodic orbits of System (7.20) that approach the heteroclinic cycle of the
conservative system Equation (7.21).
If ,6 < 0, the periodic orbits of System (7.18) that are candidates for
periodic orbits of Equation (7.4) encircle the origin and are outside the
figure eight (the homoclinic orbits) for the conservative system (7.22). The
analysis in this case is very complicated and involves very delicate estimates
of Abelian integrals. The pulse-like solution of Equation (7.4) corresponds
to the periodic orbits of System (7.20) for A < 0 that approach the figure
eight for the conservative system (7.22).
It is possible to extend Theorem 7.2 to the matrix case. Let us describe
the setup and results. Suppose that E > 0 is a real parameter, A is an n x n
nonsingular real constant matrix, f E Ck(JRm x lR, lRm), k 2: 4, f(O, A) = 0
for all A, and consider the vector equation
(7.23) d:(t) + Ax(t) = Aj(x(t- 1), A).
As for the scalar case, we impose conditions on f so that the m-

dimensional map x ---> f(x, A) undergoes a generic period-doubling bifurca-
tion at the point (x, A) = (0, 0) and then investigate under what conditions
Equation (7.23) possesses a periodic solution of period approximately 2 for
E small and discuss the limiting behavior of this solution as E ---> 0. If a(L)
denotes the spectrum of an n x n matrix L, then our first hypothesis is
a(fx(O, 0)) n S 1 = { -1}
where S 1 is the unit circle in the complex plane with center at the origin.
We also suppose that a(fx(O, A)) contains the point -(1 +A) for A small.
With these hypotheses, we can make a change of coordinates in Equation
(7.23) and write, without loss of generality,
f(x,>-.) = (fi(x,>-.),h(x,>-.))T E IR x IRm-l

h(x, >-.) = -(1 + >-.)x1 + c1xi + x1C2x2 + c3xf
(7.24) + O(lxzl 2 + lxl 4 + i>-.llxl 2 )
h(x, >-.) = Gox2 + xiG1 + x1G2x2
+ O(l>-.llx2llx2l 2 + lxl 3 + i>-.llxl 2),
where X= (xl,x2) E IR X IRm-I, cl,C3 E IR, c2 E IRlx(m-1), Go,G2 E
IR(m-l)x(m-1)' Gl E IR(m-l)xl, and
<7(Go) n 5 1 = 0.
If we now apply the method of Liapunov-Schmidt for the existence of

period-two points of the map f(x, >-.) near (0, 0), then the generic condi-
tion for existence of such points is that
= C 2(Im-1 -Go )-1 G1

Ro clef + c12 + c3 "I- 0,
where Im-1 is the identity matrix in IRm-l_
To be able to say that these period-two points of the map are carried
over into periodic orbits of (7.23) of period 2 + 2n for (E, >-.) small, we need
some additional conditions that relate the matrix A -l to the operator G0 .
To motivate the hypotheses, we introduce in the matrix case the coordinates
and scaling (7.8), (7.9) to obtain
w1(t) = rAw1(t)- rAj(w2(t -1), >-.)

(7.25)
w2(t) = r Aw2(t) - r Af( w1 (t- 1), >-.).
As for the scalar case, we regard (7.25) as a two-parameter bifurcation

problem with (>-., r) as parameters.
To determine the approximate value of the constant r in the period
2 + 2n, we consider the linear variational equation around the zero solution
of (7.25) for >-. = 0,
w1(t) = rAw 1(t)- rADw2(t- 1)

(7.26)
w2(t) = rAw 2(t)- rADw 1(t- 1).
The eigenvalues of (7.26) are the roots of the characteristic equation,
(7.27)
where L is a continuous linear map from C([ -1, 0], IR 2m) into IR 2m,
L¢ = [ ~ ~] ¢(0)- [A~ A~] ¢(-1)

D = fx(O,O).
The left-hand side of Equation (7.26) always has p, = 0 as a zero. We

determine r so that Equation (7.27) has p, = 0 as a double zero. We make
the following hypotheses:
0 "1- ro E IR,
w here A 21 E IR (m-1)x1 , A 12 E IR1x(m-1) , A 22 E IR(m-1)x(m-1) ,
We suppose also that
(H5) det [iwim- roA(Im ± De-iw)] "1- 0 for w E IR \ { 0 }.

The justification of hypotheses (H 3 ), (H4 ), and (H 5) is contained in
Lemma 7.1. Let .6.(p,, r) be as in Equation (7.27). Under the hypotheses

(H 3 ), (H 4 ), the point p, = 0 is a double zero of .6.(·, r) if and only if r = ro.
With the additional hypothesis (H 5 ), no other zeros of .6.(p,, r) lie on the
imaginary axis.
The main result for the matrix case that generalizes Theorem 7.2 is
Theorem 7.3. Suppose that (HI)-(H5) are satisfied. Then there is a neigh-
borhood U of (0, 0) in the (.>..,E)-plane and a sectorial region S in U such
that if (.>..,E) E U, then there exists a periodic solution X>.,e of Equation
(7.23) with period 2r(.A, E) = 2 + 2roE + O(IEI(I.AI +lEI)) as (.>..,E) ---+ (0, 0)
if and only if (.>..,E) E S. Furthermore, this solution is unique. The set
as= {(.>..,E) E s: f.> 0, .AR1 > 0} corresponds to a Hopf bifurcation. Let
S>. = { E: (.>..,E) E S }. If RoR1 > 0 (respectively, RoR1 < 0) and .AR1 > 0
(respectively, .AR1 < 0), then X>.,e approaches a square wave (respectively, a
pulse-like wave) as E---+ 0.
The basic idea for the proof of this result is the same as for the scalar
case-treat (7.25) as a perturbation of (7.26), obtain the vector field on a
center manifold, and relate the periodic solutions on the center manifold
to periodic solutions of (7.23). The essential new ideas for the vector case
is Lemma 7.1 and it is a nontrivial task (although only computational) to
obtain the vector field on the center manifold.
In problems of transmission of light through a ring cavity (see the
references), the following model has been proposed:
d d
(7.28) (Em-+ 1) · · · (E1-d + 1)y(t) = g(y(t- 1), A),
dt t
where each E· > 0 is a small parameter. In the supplementary remarks of

Chapter 11, ~e have sufficient conditions for the existence of a slowly oscil-
lating periodic solution of (7.28). These conditions are satisfied if the Ej are
sufficiently small and the map y----> g(y, .X.) undergoes a generic supercritical
period-doubling bifurcation at (y, .X.)= (0, 0). Is it possible to determine the
limit of this solution as Ej ----> 0, j = 1, 2 ... , m, from Theorem 7.3?
If we scale the Ej as Ej = w~j 1 , j = 1, 2, ... , m, and let x1 = y,
Xj = axj 1±j_ 1+xj_ 1, j = 2, 3, ... , m, then we obtain equivalent equations
that are a special case of (7.23). It is possible to use Theorem 7.3 to prove
the following.
Theorem 7.4. Consider the Equation (7.28) with Ej = wj 1 , j = 1, 2, ... , m.

If the scalar map y ----> g(y, .X.) undergoes a generic period-doubling bifurca-
tion at (y,.X.) = (0, 0) and f(x,.X.) = col(g(x1,.X.), g(x1,.X.)), x = col(x1,x2),
then the map x ----> f(x, .X.) undergoes a generic period-doubling bifurcation
at (x, .X.) = (0, 0) and the corresponding system (7.23) for (7.28) satisfies
(HI) - (H 5 ) with r 0 = 2:.;'= 1 aj 1 . Therefore, the conclusions in Theorem
7.3 are valid.
Let us now consider a further generalization of Equation (7.23) consist-

ing of a matrix delay differential equation coupled with a matrix difference
equation, a so-called hybrid system
E±(t) + Ax(t) = Af(y(t), .X.)

(7.29)
y(t) = g(x(t- 1), y(t- 1), .X.),
where E > 0, A are small real parameters, x E ffim, y E m,n are vectors, the
m x m matrix A has an inverse and the functions f(y, .X.) and g(x, y, .X.) are
smooth vector-valued functions.
ForE= 0, we obtain the map on ffin defined by
(7.30)
Suppose that Equation (7.30) undergoes a generic supercritical period dou-

bling at A = 0 with the period-two points being dl>,, d2 >.. In IRm x m,n,
we have, forE= 0, the square wave (xE(t), yE(t)), t E ffi, which alternately
takes on the values (f(dn, .X.), G>.(d 1>.)) and (f(d 2>., .X.), G>.(d 2>.)) on inter-
vals of length one. Since the bifurcation is supercritical, this function will
be stable if we impose a few additional conditions on the function f. There-
fore, one would expect that there is a solution of Equation (7.29) forE small
that will be close to the square wave. Under appropriate hypotheses, this is
true. The first step in attacking the problem is to use scaling to eliminate
the parameter E. However, for the results that have been obtained so far,
the next step in the proof is completely different from the preceding one.
It involves methods more functional analytical in nature and uses concepts
of exponential dichotomies. See the supplementary remarks for references

that contain a precise statement with the proof.
Other problems more general than Equation (7.29) arise in a very
natural way. For example, suppose that we consider a scalar equation with
two delays:
(7.31) Ei:(t) = -x(t) + f(x(t- 1), x(t- O"), .A)

where O" :::=: 1. There is essentially nothing known about the relationship be-
tween the solutions of (7.31) forE> 0 small and the corresponding solutions
of the difference equation:
(7.32) x(t) = f(x(t- 1), x(t- O"), .A).
If O" is irrational, then Equation (7.32) is an infinite-dimensional problem

and, of course, none of the ideas mentioned seem to shed much light on
the problem. On the other hand, if O" is rational, then Equation (7.32) can
be considered as a map on a finite-dimensional space and, at least, we
may speak of generic period doubling. It certainly would be of interest to
know something about the implications for E = 0. One possible approach
would be the following. For simplicity, suppose that O" = 2. If we define
x(t- 1) = y(t), then Equation (7.32) is equivalent to the hybrid system:
Ei:(t) + x(t) = f(x(t- 1), y(t- 1), .A)

(7.33)
y(t) = x(t- 1).
If we could extend these theories to these systems, then we will have at
least solved the problem of period doubling.
12.8 Averaging
In this section, we review some of the results on the application of the

method of averaging for RFDE. We begin with a brief review of the results
and methods that are used in ODE.
12.8.1 Averaging in ODE
For E > 0 a small parameter, we consider the ODE

. t
(8.1) x = f(-,x)
E
where j(T + 1, x) = j(T, x) for all T, x, is continuous in (t, x) and is con-

tinuously differentiable in x. It is possible to extend the following remarks
to functions f with a more general dependence on T (for example, almost
periodic in T uniformly for x in bounded sets), and we discuss this case
12.8 Averaging 397
only to avoid technical difficulties. Along with Equation (8.1), we consider

also the averaged equation
(8.2) Y= fo(y),
where
(8.3) fo(Y) = 11 f(T,y)dT.
If E > 0 is sufficiently small, it is possible to make a transformation of

variables in Equation (8.1), which is periodic in t of period 1 and close to
the identity, to obtain a new ODE for which the vector field is close to the
averaged vector field. More specifically, if we let x = z + EU( ~, z), where
u(s,x)= 1 8
[f(T,x)-fo(x)]dT,
then
. t
(8.4) z = fo(z) + g( -, z, E),
E
where
(8.5) g(T,z,O) =0, g(T,z,E) =g(T+1,z,E).
The first classical result on averaging asserts that we can keep a solu-
tion of Equation (8.1) close to a special function associated with a solution
of the averaged equation for as long as we want if we choose E sufficiently
small. More precisely, we have
Theorem 8.1. Let x(t) be a solution of Equation (8.1) with x(O) = x 0 and
let y(t) be a solution of Equation (8.2) with y0 = y0 (x 0 ) chosen to satisfy
the equation Xo =Yo+ EU(O, y 0 ). If y(t) is bounded fort ~ 0, then, for any
7] > 0, L > 0, there is an Eo= Eo(7J,L) > 0 such that for 0 < E <Eo, we
have
(8.6) \x(t)- x*(t)\ ::; 7] for 0::; t::; L,
where x*(t) = y(t) + EU(~, y(t)).
It is also possible to obtain some qualitative results that are valid on

the infinite time interval [0, oo), for example,
Theorem 8.2. If Equation (8.2) has a hyperbolic equilibrium point x 0 , then

there is an E-periodic solution x* (t, E) of Equation (8.1), x( T, 0) = x 0 , which
is hyperbolic and has the same stability properties as the equilibrium point
xo of Equation (8.2).
If Equation (8.2) has a hyperbolic periodic orbit"(, the Equation (8.1)

has a hyperbolic invariant manifold M, C IR x IRn such that M 0 = IR x "(,
the cross section Met of M, at time t is E-periodic in t; that is, there exists
a hyperbolic invariant torus of Equation (8.1).
We remark that if we lett f--+ Et in Equation (8.1), we obtain the more

classical equation
(8.7) x= Ef(t, x)
this is encountered so often in the theory of nonlinear oscillations. Of course,

these results can be easily translated to the solutions of Equation (8. 7) and
the averaged equation
(8.8) iJ = Efo(y).
12.8.2 Averaging in RFDE
It is possible to extend the results of the previous section to the RFDE
(8.9)
where f(T + 1, ¢) = f(T, ¢) for all (T, ¢) E IR x C, is continuous in (T, ¢)

and is continuously differentiable in ¢. The averaged equation is
(8.10) y(t) = fo(Yt),

where
(8.11)
To be able to obtain results similar to those in the previous section, we

will use the variation-of-constants formula inC, considering RFDE (8.9) as
a perturbation of the zero vector field:
(8.12) x(t) = o, t > 0.

If we let T0 (t) be the semigroup on C generated by Equation (8.12),
(8.13) Tt(t)¢(B)={¢(t+B) fort+B:::;O,

0 ¢(0) for t + B > 0,
then the solution of RFDE (8.9) with initial value ¢ at t = 0 can be repre-
sented as
(8.14) Xt = To(t)¢+ 10
t s
d[Ko(t,s)]f(-,xs)
E
12.8 Averaging 399
where K 0 (t, s)(O) = J; Xo(t + 0- a) da, -r ~ 0 ~ 0 and

Xo(t) = { 0, t < 0;
I, t ~ 0.
Fix T > 0 and define the following transformation of variables Xt = F Zt

with F: BC([O,T],C) --7 BC([O,T],C) and
(8.15) Fv(t) = v(t)- EAo 1 0

t T
d[K(t, T)]u( -, v(T))
E
t
+ w( -, v(t))
E
where u(s, ¢) = J;[j(T, ¢)- fo(¢)] dT and A 0 is the generator associated

with Equation (8.12).
It is easy to verify that the transformation Xt = Fzt given by (8.15) is
well defined, periodic of period 1, and close to the identity; that is, there is
a constant C and an Eo > 0 such that for 0 ~ E ~Eo, the difference
sup lzt- Xtl < EC.

O~t~T
Next we derive the integral equation for Zt when Xt satisfies Equation

(8.14). If we substitute (8.15) into (8.14) and rewrite the expression, we
obtain
Zt = To(t- s)zo +lot d[Ko(t,T)]fo(z,.)

(8.16)
+ 1
0
t
d[Ko(t,T)]n(-,z,.)
T
where xo = zo + w(O, zo) and
T T dv T T
(8.17) n( -, v) = -ED¢u( -, v) dt
E E
+ f( -,Fv)-
E
f( -, v).
E
To prove that Yt is a solution of the averaged equation up to terms of order
E, it remains to analyze the nonlinearity n(t, v). A simple estimate yields
Lemma 8.1. For v E BC 1 ([0, T], C) there is a constant C > 0 such that
(8.18) 1
I
0
t T
d[Ko(t, T)] N( -, v(T), E) dTI ~ C(Eivll
E
+ IFv- vi).
So if Xt = F Zt, then Zt is a solution of the averaged equation up to terms

of order E. As a first application, we compare the solution Xt of Equation
(8.9) with xo = ¢ E C with the approximate solution x; ~f Fy; where y;
is a solution of the averaged equation with y0 = 'ljJ and¢= 'ljJ + w(O, '1/J).
Theorem 8.3. If for¢ E C, the solution y; of the averaged equation (8.10)

with y 0 = '1/J and¢= '1/! + w(O, '~/!) is uniformly bounded fort :::0: 0, then for
any TJ and L, there is an Eo such that for 0 < E < Eo, the difference
(8.19) for 0 ~ t ~ L.
where x; = :F y;.
With the transformation theory mentioned earlier, we can use the
methods from the theory of invariant manifolds to obtain
Theorem 8.4. If y 0 is a hyperbolic equilibrium point of the averaged equation

(8.10), then there exist positive constants Eo, TJ, such that for 0 < E ~ Eo,
there is an £-periodic solution x*(t, E) of Equation (8.9), x*(·, 0) =yo, which
is hyperbolic, has the same stability properties as y 0 , and is unique in the
set { x E 1Rn : lx- Yo I < TJ }.
If y 0 is hyperbolic and uniformly asymptotically stable, then the unique
£-periodic solution is hyperbolic, and uniformly asymptotically stable and
there are positive constants p, C, 1 such that if x( t, ¢) ( resp. y( t, ¢)) is the
solution of (8.9) (resp. (8.10)) through (0, ¢) and 1¢- Yo I < p, then, for
t :::0: 0, we have
lx(t, c/J)- x*(t, E)l ~ Ce- 1 t,
(8.20)
lx(t,¢)- y(t,¢)1 < TJ·
It is possible also to consider attractors.
Theorem 8.5. If the averaged equation (8.10) has a local attractor A 0 , then
there is an Eo > 0 such that for 0 < E ~ Eo, the Poincare map for Equation
(8.9) has a local attractor A, and dist (A" A 0 ) --+ 0 as E--+ 0.
As an example illustrating the results, we consider the equation

x(t- r)
(8.21) x(t) = -x(t) + b 1 +xt-rn
( )
where n is a fixed even integer and b > 0 is a parameter.

The solution map is point dissipative and thus there exists a global
attractor A,. It is known (at least numerically) that for n :::0: 8, there exists
a b0 > 0 such that for b :::0: b0 , there is some chaotic motion on A. Let us
consider the following class of rapidly oscillating disturbances of (8.21):
. acos(t/E) + x(t- r)
(8.22) x(t) = -x(t) + b1 + (cos(t/E) + x(t- r))n'
where a > 0 is a constant that measures the energy of the perturbation.
It is possible to prove the following result: For E > 0 sufficiently small, the
12.9 Infinite delay 401
attractor A€ of the Poincare map for Equation (8.22) is just a singleton

provided that a > max{2b, 3}. This result implies that high-frequency per-
turbations can eliminate complicated motion on the attractor. The proof
of the result consists of averaging Equation (8.22), estimating the result-
ing nonlinear vector field, and using a Razumikhin-type theorem to obtain
the existence of a globally attracting equilibrium point for the averaged
equation. Theorem 8.5 completes the proof.
Averaging also has been discussed for the equation
(8.23) x(t) = Ef(t, x),

where E > 0 is a small parameter. If fo is defined by Equation (8.11), then
the averaged equation is the ODE
(8.24) y(t) = Efo(iJ), f)( B)= y for() E [-r, 0].

Results similar to Theorem 8.4 are available, but the proofs that have been
given follow a different approach. If we consider RFDE (8.23) as a pertur-
bation of
(8.25) x(t) = o · xt,

then the decomposition inC, Xt = iz(t) + Wt,
i(B) =I, the identity, for
() E [-r, 0], for the linear equation (8.25) implies that Wt approaches zero
faster than any exponential. We can now use the invariant manifold theory
to show that the flow for RFDE (8.23) in any given bounded set is equivalent
to the flow defined by an 0 DE
Z = Eg(t, z, E), g(t, z, 0) = f(t, z).

The classical averaging procedure can be applied to this ODE.
If we lett~---+ t/E in RFDE (8.23) and let x(t/E) = y(t), then we obtain
the equation
y(t) = f(!,Yt,€),
E
where Yt,€(8) = y(t+EB), () E [-r, 0]. This is an equation with a small delay,
but it is rapidly oscillating in t. Therefore, it reasonable to expect that it
should be possible to obtain these results for RFDE (8.23) by using the
transformation theory that we used for (8.9), but this has not been done.
12.9 Infinite delay
Suppose that 0 ::::; r ::::; oo is given. If x : [0' - r, 0' +A) ____, 1Rn, A > 0, is a
given function, then for each t E [s, 0' +A), () E [-r, 0], we define, as usual,
Xt(B) = x(t +B). It is understood here that [0'- r, 0' +A) = ( -oo, 0' +A)
if r = oo. In the theory of RFDE,
(9.1) x(t) = J(t, xt),
the choice of the space for the initial data, the phase space, is never com-
pletely clear. For each particular application, a decision is made that is
believed to reflect the important aspects of the problem under investiga-
tion. In the case of finite delay (r < oo), the solution of (9.1) is required
to be a continuous function for t 2: a. Therefore, after one delay interval
r, the state Xt, t 2: a + r, belongs to the space of continuous functions. As
a consequence, the choice of the phase space is not so important from the
point of view of the qualitative theory. However, in specific applications, it
is convenient to have other phase spaces. For example, in control theory, the
space L 2 ((-r,O),IRn) x IRn is frequently encountered. In this setting, the
problem is formulated in a Hilbert space, which leads to the adaptability of
many classical results to RFDE. This has proved to be particularly useful in
linear control and identification problems. On the other hand, for nonlinear
problems, we do not have, at this time, a theory in this space that can be
used to develop a general qualitative theory. This is probably due to the
fact that the requirement that the solutions are differentiable with respect
to the initial data puts very severe requirements on the nonlinearities.
If the delay interval is infinite, then the state Xt at time t always
contains the initial data. As a consequence, the introduction of a new phase
space in a particular application requires a new and separate development
of the theory. On the other hand, it is possible to give an abstract and
axiomatic definition of a phase space for which many of the fundamental
and desired properties hold. 'vVe present an axiomatic framework that will
permit the development of the fundamental theory of existence, uniqueness,
continuation, continuous dependence, differentiability with respect to initial
data and parameters, etc. In addition, we need the abstract properties to
imply something about the global behavior of orbits; for example, when are
bounded orbits precompact, when is stability in IRn equivalent to stability
in the function space, etc? In this way, we will gain a better understanding
of the equations and at the same time avoid too much repetition.
We first remark that our axioms prevent the norm in the space from
imposing any differentiability properties on the initial functions. In the ap-
plications, it is convenient at times to require the initial functions to belong
to a Banach space of functions that have some derivatives with specified
properties. However, if we consider all differential equations whose right-
hand sides are continuous or continuously differentiable in such a space,
then the equations will be of neutral type: that is, the derivatives of the in-
dependent variable will contain delays. The theory for such systems should
be developed, but it will require more sophistication than the one described
later for RFDE.
The phase space B for RFDE with infinite delay is a linear space, with
a seminorm l·ls mapping ( -oo, OJ into the finite-dimensional Banach space
E = IRn or <Cn. The first two axioms on B are motivated by the fact that
we want a solution of (9.1) to be continuous to the right of the initial time

and we desire certain continuity properties of the solutions.
(A) There is a positive constant Hand functions K, M : IR+ ----+ IR+,
with K continuous and M locally bounded, such that for any u E IR,
a > 0, if x : ( -oo, u +a) ----+ E, Xa E B, and x is continuous on
[u,u+a), then for every t E [u,u+a), the following conditions hold:
(i) Xt E B,
(ii) lx(t)IE::; HlxtiB,
(iii) lxtiB::; K(t- u) sup{ lx(s)IE: u::; s::; t} + M(t- u)lxaiB·
(AI) For the function x in (A), Xt is a B-valued continuous function for
t E [u,u+a).
We remark that the elements in a space B satisfying these axioms may
satisfy 1¢ - '¢liB = 0 and the functions may not be pointwise equal on
( -oo, OJ. However, from condition A(ii), if 1¢- '¢liB= 0, then ¢(0) = '¢1(0).
Let us give examples of spaces that satisfy (A) and (AI). For any
continuous positive function g on ( -oo, OJ, let
where
l¢1g ~fsup{ ~:~:~ 1 : -oo < B::; 0}.
Let
Ue9 = { ¢ E e 9 : 1!_ is uniformly continuous on ( -oo, OJ}
g
Le9 = { ¢ E e 9 : lim ¢((BB)) exists in IRn }.

0--+-oo g
Leo = { ¢ E e 9 :
. ¢(B)
hm -(B) = 0 }.
9
g
0--+-oo
For the special case where g( B) = e-"fO, where 1 > 0 is a constant, we define
e"~ ~r Le9 • For g = 1, we obtain the following classical spaces:
Be= { ¢ E e(( -oo, OJ, IRn): sup I¢( B) I< oo}
BU = { ¢ E Be:¢ is uniformly continuous on ( -oo, OJ}

Le = { ¢ E Be: lim ¢(B) exists in IRn}
0--+-oo
Leo = { ¢ E Be : lim ¢(B) = 0}

0-->-00
It is possible to prove that the spaces Ue9 , Le9 , Leg with the function
g nonincreasing satisfy the axioms (A) and (AI). In particular, this is true
for the spaces BU, Le, Le0 , and e"~. The space Be satisfies (A) but not
(AI).
If we now suppose, for example, that f(t, ¢) in (9.1) is continuous in

lR x B, continuously differentiable in¢, is locally bounded, and the space B
satisfies axioms (A) and (A 1 ), then for any (a, f) E lR x B, it is possible to
prove the local existence and uniqueness of the solution x(t, a,¢) of (9.1),
defined on an interval to the right of a and X a (·,a,¢) = ¢. Furthermore, the
solution is continuously differentiable in¢. We define T(t, a)¢= Xt(·, a,¢)
and refer to T(t, s) as the solution operator of (9.1). For simplicity in the
presentation, we assume that T(t, a) is defined for all t 2:: a. From the
assumptions on fin (9.1), the mapping T(t,a) is a bounded map for each
t ;::: s.
To describe some further properties of the solution operator, we need
some more axioms.
(B) The space B is complete.
We say that a sequence of functions ¢n E B converges compactly on ( -oo, OJ
to a function ¢ on ( -oo, OJ if the sequence converges uniformly on compact
subsets of ( -oo, OJ.
(C 1 ) If { ¢n } c B is a Cauchy sequence in B with respect to the seminorm
and if ¢n converges compactly to ¢ on ( -oo, OJ, then ¢ is in B and
l¢n - ¢1B --+ 0 as n --+ oo.
For¢ E B, the symbol¢ denotes the equivalence class { '¢: 1'¢- ¢1B =
0 } and f3 denotes the quotient space { ¢ : ¢ E B }, which becomes a normed
linear space with the norm l¢lg = I¢1B· Axiom (B) is equivalent to saying
that B is a Banach space.
It is rather surprising that one of the basic properties of the map T( t, s)
fort> sis determined by the trivial RFDE in 8:
(9.2) ±(t) = o.
We assume that B satisfies the axioms (A), (AI), (B), and (CI). Let
S(t), S(O) = I, be the solution operator of (9.2) and let S0 (t) be the re-
striction of S(t) to the closed subspace
(9.3) Bo = {¢ E B: ¢(0) = 0 }.
The operator 8 0 ( t) : 8 0 --+ 8 0 and satisfies the inequality
(9.4) ISo(t)IB :'S M(t),
where M(t) is the function in axiom (A(iii)). Let S(t) and So(t), t 2:: 0, be,
respectively, the induced operators on f3 and B0 • These are Ca-semigroups
of operators.
Let us recall that the a-measure ofnoncompactness of a bounded linear
operator A on a Banach space X is defined by a( A) = inf{ k : a(AB) :'S
ka(B) for all bounded sets B C X}. Also recall that re(A) denotes the
radius of the essential spectrum of A. Let
1 ~
(9.5) f3 = lim -log a(S(t)).
t-HXJ t
An important result is the following.
Lemma 9.1. re(S(t)) = ef3t::; !So(t)ls for all t ~ 0.
As an example, we remark that it can be shown that
~ g(B+t)
(9.6) re(S(t))::; sup{ g(B) : -oo < e::; t} if B = C9
(9.7) if B = C,.
Lemma 9.2. Suppose that B satisfies the axioms (A), (A1), (B), and (CI)
and K(t) is bounded fort~ 0. Then the solution operator T(t, a) of Equa-
tion (9.1) can be written as
(9.8) T(t, a)¢= S(t- a)¢+ U(t, a)¢, t ~a,
where the operator U(t, a) on B is completely continuous.
Lemma 9.2 is a consequence of our axioms and the representation of

U(t, a) as
{
0, e
if t + < a,
[U(t,a)¢](B) = J:+l1 f(s,T(s,a)¢)ds, ift+B >a.
It is interesting to consider the implications of Lemma 9.2 for the

existence of compact global attractors for autonomous equations. The same
remarks will hold for the Poincare map of an equation that is periodic in
time. Suppose that f E C 1(B,IRn) is a locally bounded map and consider
the autonomous equation
(9.9) x(t) = f(xt)
on the space B satisfying all of the previous axioms. Let T(t) be the solution
operator of Equation (9.9) with T(O) = I.
Theorem 9.1. Suppose that B satisfies the axioms (A), (AI), (B), and (CI)
and K(t) is bounded fort ~ 0 and M(t) ---+ 0 as t ---+ oo. If T(t) is point
dissipative and positive orbits of bounded sets are bounded, then there is a
compact global attractor for Equation (9.9).
The hypothesis on M(t), (9.1), Lemma 9.1, and (9.4) imply that
re(S(t)) ---+ 0 as t ---+ oo. From Lemma 9.2, we infer that T(t), t ~ 0, is
an a-contraction. One also shows that the w-limit set of any bounded set
is a compact invariant set and then the conclusion follows from Theorem
4.3.3.
The corresponding result for the Poincare map of a periodic system
will yield the existence of a compact global attractor and the existence of
a fixed point. Therefore, there will be a periodic solution of the RFDE of
the same period as the coefficients of the vector field.
The theory of linear equations can be developed on the Banach space
f3. To be somewhat more specific, suppose that B satisfies the axioms (A),
(AI), (B), and (C 1 ) and consider the linear autonomous equation
(9.10)
where L: B-+ Eisa bounded linear operator. Equation (9.10) generates a

strongly continuous semigroup T(t), t ::=: 0, on B. Let A be the infinitesimal
generator. We can define the operators T(t) and A on B induced by the
operators T(t) and A and given by the formulas T(t)(/y = T(t)¢, .A¢= A¢
for all ¢ E ¢. Then T(t) is a strongly continuous semigroup on Band A is
the infinitesimal generator.
The type number of T(t) is denoted by Ct£ and is given by
1 ~ 1 ~
(9.11) Ct£ lim -log IT(t)ls = inf -log IT(t)ls
= t-+oo t t>O t
and the spectral radius ra(T(t)) of T(t) is given by etDiL. If we let Pu(A)
denote the point spectrum of A, then it is possible to show that
(9.12)
Furthermore,
Ct£ = max{,6, sup { Re ..\ : ..\ E Pa(A)} },
where ,6 is given in (9.5).

It is possible to continue in this way to obtain all of the decomposition
theory of Chapter 7 and, therefore, we have at our hands all of the ma-
chinery for the local theory that we had for the case of finite delay. We do
not pursue this any further and recommend the references for details and
further references.
The definition of functional differential equations on manifolds as given

in Section 1 (as well as many of the examples) is due to Oliva [1,2,3]. A
more complete presentation and proofs of many of the results are in Hale,
Magalhaes, and Oliva [1]. Theorem 1.2 was stated by Kurzweil [4] (the
first complete proof was given by Mallet-Paret [2]). Theorem 1.3 was first
proved by Kurzweil [1,2,4] where he also presented other interesting results

for RFDE near ordinary differential equations (see also Kurzweil [3,5]).
Theorem 1.4 is due to Oliva [4]. For the complete proof of the assertion
in Example 1.7 about the limit of Tp(t) as indicated in the text, see Hale
and Raugel [2]. Another proof is in Hale, Magalhaes, and Oliva [1]. For a
different proof of Theorem 1.3 and generalizations that permit the consid-
eration of structural stability and generic one-parameter bifurcations near
equilibrium points, see Magalhaes [5,6, 7].
Mallet-Paret [4] proved that the compact attractor has finite Hausdorff
dimension in a separable Hilbert space. Maiie [1] proved the more general
results in Theorems 2.1 and 2.2. Theorems 2.3 and 2.4 are due to Mallet-
Paret [4] and generalize the corresponding results of Cartwright [1,2] for
ordinary differential equations.
The stability result (Theorem 3.1) on Morse-Smale systems and a proof
were given in an unpublished work of Oliva [5] and is reproduced in Hale,
Magalhaes, and Oliva [1]. Theorem 3.3 is due to Hale and Rybakowski [1].
Some good references for the generic theory of ordinary differential
equations are Abraham and Robbin [1], Markus [1], Nitecki [1], Peixoto [1]
and Smale [1]. A very readable proof of the Kupka-Smale theorem is given
by Peixoto [2]. The theorem was first proved by Kupka [1] and Smale [2],
but Markus [2] had previously announced some partial results.
The generic theory of RFDE initiated from the important contribu-
tion of Oliva [1]. In this paper, he began the generalization of the Kupka-
Smale theorem by proving that the sets Q0 (K) and 91 (K) of RFDE(F)
with F E Xk such that the critical points are nondegenerate and hyper-
bolic, respectively, are open and dense. Oliva proved (officially announced
in Oliva [2]) the result that the sets 93; 2(T, K) and Q2(T, K) of RFDE(F)
with F E Xk with nondegenerate and hyperbolic, respectively, periodic
orbits in K of period ::::; T are open. For the completion of the proof of
Theorem 4.1, it was, therefore, necessary to prove density, which is the
most difficult part. Mallet-Paret [1] proved the density by the ingenious
proof outlined in the text. The generic results on the differential difference
equation x(t) = f(x(t), x(t- 1)) also are due to Mallet-Paret [3].
It is of interest to note that ideas from generic theory have been used by
Chow and Mallet-Paret [2] to define an index for periodic orbits of RFDE
as was done by Fuller [1] for ordinary differential equations. Fuller's index
can be used to obtain a new class of periodic solutions of certain equations,
for example, the equation
x(t) = -[ax(t- 1) + (Jx(t- 2)]f(x(t)).
The generic theory for NFDE has received very little attention. We
mention the paper of de Oliveira [1] in which he proved that the sets 90 (K)
and gl (K) are generic.
Theorem 5.1 is due to Nussbaum [1]. Theorems 5.2 and 5.3 are due
to Hale and Oliva [1]. Theorems 5.4 and 5.5 and Propositions 5.2 and 5.3
are due to Sternberg [1]. Theorem 5.6 is a consequence of an abstract lower

semicontinuity result for gradient systems due to Hale and Raugel [1].
Theorems 6.1-6.3 and the remarks on periodic solutions in Section 6
are due to Mallet-Paret [5]. For linear equations, the observation that the
number of zeros per unit interval (the discrete Liapunov functional V) does
not increase with time goes back to Mishkis [1]. The same property holds
also for scalar parabolic partial differential equations in one space variable
(see Nickel [1], Matano [1]).
Cao [1] has generalized the definition of the discrete Liapunov func-
tional in the text in such a way as to be able to characterize the small
solutions (those that approach zero faster than any exponential) of lin-
ear nonautonomous equations. Cao [2] has used this functional also to
show that there can be no small solutions of an analytic delay differen-
tial equation x(t) = F(x(t), x(t -1)) provided that aF(x, y)jay is not zero
if (x, y) = (0, 0).
Dynamical systems for which there exist such discrete Liapunov func-
tions (generalizing the number of zeros of a function) have many very in-
teresting properties. For example, for the ordinary differential equation
x(t) = f(x), x E IRn, for which the matrix af(x)jax is of Jacobi type
(the matrix is tridiagonal with the off-diagonal elements positive), there is
a discrete Liapunov function that is given by the number of sign changes
in the vector x. This property can be used to prove that hyperbolicity of
equilibrium points implies that the stable and unstable manifolds intersect
transversally (Fusco and Oliva [1]). Oliva, Kuhl, and Magalhaes [1] have
extended this result to diffeomorphism with oscillatory Jacobians.
We refer to the papers of Fiedler and Mallet-Paret [1], Fiedler [1] and
Fusco and Oliva [1] and the references therein for other aspects of this
exciting area of research.
Kaplan and Yorke [2] were the first to use the projection of an orbit
of a delay differential equation onto a plane to obtain the existence of a
periodic orbit and some of the asymptotic properties of special solutions.
This was done for slowly oscillating solutions and a very special nonlinear
equation.
Fiedler and Mallet-Paret [2] showed that the connecting orbits Cff
exist for N < N* and McCord and Mischaikow [1] gave the general result
that Cjf i= 0 for all K < N.
As remarked in the text, the flow on a Morse set AN may be very
complicated. The numerical experiments of Mackey and Glass [1], Farmer
[1], Chow and Green [1] and Hale and Sternberg [1] clearly indicate this
fact. The numerical computations of Hale and Sternberg [1] were designed
to test the hypothesis that the chaotic motion was a consequence of the
creation of a transversal intersection of the stable and unstable manifolds
of a periodic orbit. Some theoretical results exhibiting classes of delay dif-
ferential equations that possess a hyperbolic periodic orbit with its stable
and unstable manifolds having nonempty transversal intersection may be
found in Walther [10], an der Heiden and Walther [1] and Hale and Lin [1].
Equation (7.1) has served as a model for many applications, including
physiological control systems (Glass and Mackey [1], an der Heiden and
Mackey [1], Lasota [1], Mackey and Glass [1], Mackey and an der Heiden
[1], Wazewska-Czyzewska and Lasota [1]), optically bistable devices and the
transmission of light through a ring cavity (Berre et al. [1], Derstine et al.
[1,2], Gibbs et al. [1], Hopf et al. [1], Ikeda [1], Ikeda, Daido, and Akimoto
[1], Ikeda, Kondo, and Akimoto [1], Ikeda and Matsumoto [1], Malta and
Ragazzo [1]) and population dynamics (Blythe et al. [1], Gurney et al. [1],
Hoppensteadt [1]).
Propositions 7.1 and 7.2 are due to Ivanov and Sharkovsky [1]. Theorem
7.1 is due to Mallet-Paret and Nussbaum [1,2] (see also Mallet-Paret and
Nussbaum [4]). They also give an explanation of the Gibbs' phenomenon
mentioned in the text. For a given function f, it is a nontrivial task to
verify that the hypotheses of Theorem 7.1 are satisfied. Mallet-Paret and
Nussbaum [3] have given ranges of the parameters for which the hypotheses
are satisfied for each of the following functions:
h(x) = J.L- x 2 , h(x) = x 3 - J.LX, h(x) = -J.L[sin (x + o:)- sino:],

J.l-Xv
fs(x) = x.A + 1' X;::: 0.
Theorem 7.2 for the supercritical case was conjectured by Chow and
Mallet-Paret [3] and proved by Chow and Huang [1] by a method different
from the one outlined in the text. The proof in the text is due to Chow,
Hale, and Huang [1]. The subcritical case is due to Hale and Huang [2]. All
of the results mentioned in Section 7 for the matrix case are due to Hale
and Huang [3].
Equation (7.28) has served as a model of transmitted light through
ring cavities with several chambers (Vallee, Dubois, Cote, and Delisle [1],
Valee and Marriott [1]) as well as some problems in physiology (an der
Heiden [1]). The system (7.29) has been used by Ikeda [1], Ikeda, Daido,
and Akimoto [1] as a model of a ring cavity containing a nonlinear dielectric
medium for which part of the transmitted light is fed back into the medium.
For some precise results on the existence of periodic solutions of (7.29) with
a supercritical period doubling for (7.30), see Chow and Huang [1].
One of the difficulties in the proofs outlined in Section 7 is the deter-
mination of the first few terms in the Taylor series of the vector field on
a center manifold. In the papers referred to earlier, there is a general pat-
tern that is followed to do these computations, but it is perhaps not easily
recognized by a nonexpert. Recently, Faria and Magalhaes [1,2] have devel-
oped the theory of normal forms for functional differential equations and
have given systematic methods for the computation of the normal forms.
These methods can be used for the determination of the approximate vector
field on the center manifold. They have used these methods to discuss the
Bogdanov-Takens singularity and the Hopf bifurcation.
Since the flow for an RFDE evolves in an infinite-dimensional space,

it is perhaps to be expected that all of the complications that occur in
ordinary differential equations must appear in an RFDE. Of course, this is
true if we make the dimension of the RFDE very large. On the other hand,
if the dimension of the RFDE is fixed, say at n = 1, it is not clear that this
is the case. In fact, this is not the case if the RFDE is a differential delay
equation with one delay. It is therefore interesting to investigate in more
detail the types of flows that can be realized by RFDE of fixed dimension.
Not too much is known, but there are some local results. To be more precise,
let us suppose that the linear scalar RFDE
(10.1) ±(t) = LXt

has m eigenvalues (counting multiplicity) on the imaginary axis and let us
consider the perturbed linear system
(10.2) x(t) = Lxt + f(xt)

where f is a C 00 -function from C to IR and is small. For each given J, we
can determine a center manifold CM(f) of (10.2). The flow on CM(f) is
determined by an ordinary differential equation
(10.3) iJ =By+ Y(f, y),

where y E IRm, the eigenvalues of them x m matrix B are purely imaginary
and coincide with the eigenvalues of (10.1) that are on the imaginary axis
and them-vector function Y(f, ·) vanishes when f = 0. The problem is to
determine the range of the mapping f E C 00 (C,IR) f--t Y(f, ·) E IRm; that
is, describe those vector fields that can be realized on the center manifold.
Since the RFDE is a scalar equation, it is possible to show that (10.3)
is equivalent to an mth-order scalar equation:
(10.4) z(m) + a1z(m- 1) + · · · + amz = G(f, z(m- 1), ... , z).

The results of Hale [25,26] imply that for any given k-jet }kg of a function
g : IRm --. IR, there exists a function f : C --. IR such that the k-jet
of G(f, Zm, ... , zl) coincides with }kg. Furthermore, there exist constants
r 1 < r2 < ... < rm-1 ::; r and a function F : IRm --. IR such that f(¢) =
F(¢(0), ¢( -rl), ... , ¢( -rm- 1)), that is, a differential delay equation. It was
asserted in Hale [25,26] that all vector fields could be realized, but the proof
only yields the above information. Rybakowski [4,5], using the Nash-Moser
implicit function theorem, has shown that every vector field can be realized
in the following sense: for every m ?: 17 and every cm+l 5 -function g, g :
IRm --. IR, with g(O) = 0 there is a em-delay differential equation F with
m- 1 delays, F(O) = 0, such that G(F, z(m-1), ... , z) = g(z(m-1), ... , z).
For the case in which (10.1) and (10.2) are FDE in IRn, Faria and
Magalhaes [3] extended the results of Hale [25,26] by showing that all k-
jets on a center manifold can be realized if n is larger or equal to the
largest number of Jordan blocks associated with each of the eigenvalues

of the matrix B. The proof uses their theory of normal forms (Faria and
Magalhaes [1,2]). Under these assumptions, Rybakowski [5] has shown that
all vector fields can be realized under the same differentiability assumptions
mentioned in the previous paragraph.
For scalar equations Faria and Magalhiies [4] use the theory of normal
forms to determine the restrictions that are imposed on the vector fields
when the number of delays in the nonlinearity is less than m - 1. There
are no restrictions for the generic Hopf bifurcation or the Bogdanov-Takens
singularity with a double-zero eigenvalue. There are restrictions when there
are two purely imaginary and one zero eigenvalue.
For the classical method of averaging in ODE, the reader may consult,
for example, Bogoliubov and Mitropolsky [1] or Hale [21]. The averaging
theory for RFDE with rapid oscillations in the time variable and the details
of the example (8.16) are due to Hale and Verduyn Lunel [1,2]. The theory
also is applicable to parabolic partial differential equations. The results for
Equation (8.17) were first given by Hale [5] and extended earlier work of
Halanay [3].
The first axiomatic approach for equations with infinite delay was given
by Coleman and Mizel [1,2,3] (see also Coleman and Dill [1], Coleman and
Mizel [4,5], Coleman and Owen [1], Lima [1], and Leitman and Mizel [1,2,3])
for a special class of fading memory spaces. The beginnings of the general
abstract theory of phase spaces appeared for the first time in Hale [27], but
there were only a few axioms, no proofs, and, therefore, several points of
confusion and omission. The more complete theory was developed by Hino
[1,2,3], Naito [1,2,3], Hale [20], Hale and Kato [1], Schumacher [1,2], Shin
[1,2]. The recent book of Hino, Murakami, and Naito [1] contains almost all
of the earlier works as well as a more extensive theory and applications to
stability theory and the existence of periodic and almost periodic solutions.
This work contains an extensive bibliography to the other literature and
methods. The presentation in the text also is based on this book. Makay
[2] has given further interesting remarks on the determination of stability
by using Liapunov functionals. For a theory of dependence of solutions on
the memory function, see Hines [1] and references therein.
There are many problems in FDE that we have not addressed in these
notes that certainly are important and deserve to be studied in detail. Par-
tial differential equations for which there are delays in time occur frequently
in modeling. For the basic theory of existence and uniqueness of solutions;
see, for example, Fitzgibbon [1], Thieme [1], Webb [1,2,3], and the references
therein. For some of the models that occur in age-dependent populations,
interesting new ideas are required in the development of the theory (see
Thieme [1], Webb [4]).
Travis and Webb [1,2], Mitropolskii and Fodcuk [1], Mitropolskii and
Korenevskii [1,2], and Dombrovskii [1] have used the generalization to such
equations of the decomposition theory of Chapter 7 and made applications
to stability theory and the existence of invariant manifolds. Memory [1,2]

has given a complete theory of stable and unstable manifolds near an equi-
librium point.
In ecological models, we encounter systems of the form
(10.5)
in a smooth bounded domain D with boundary conditions. We consider

only the nonnegative solutions of Equation (10.5). With an appropriate
generalization of the notion of negative feedback and restrictions on the
coefficients, this equation defines a monotone dynamical system. An exten-
sive theory has been developed in this direction (for results and further
references, see, Martin and Smith [1,2], Smith [3], and Smith and Thieme
[1,2,3]). For similar situations where the coefficients are periodic or almost
periodic, see Tang and Kuang [1,2].
A special case of Equation (10.5) is the scalar equation
(10.6) Ut- dilu(x, t) = u(x, t)[1- u(x, t- T)]

with either Dirichlet or Neumann boundary conditions. In case D = (0, 1)
(one space dimension), Luckhaus [1] has shown that (10.6) is point dissi-
pative in the L 2 -norm for all choices of the positive parameters d, T. For a
smooth bounded general domain D and a fixed value of d, he also proved the
same result if T < To, where To is sufficiently small. Friesecke [1] has shown
that there are positive constants d0 , To such that (10.6) is point dissipative
if d > d0 and T <To. He proves also the following surprising fact if the diffu-
sion coefficient is too small and the delay is too large: If dim (D) :2: 2, then
there exist positive constants d1, 81 such that (10.6) is not point dissipative
in the region
{(d,T):d<d1,T> 81 }.
y'81(d1- d)
In this region of parameters for (d, T) and in the set of nonnegative so-
lutions, there is an open nonempty subset of initial values for which the
L 1 -norm of each solution tends exponentially to oo as t ---+ oo.
Another variant of population models is a one-dimensional diffused
version of Wright's equation:
= -( 2 + J.L)u(x, t- 1)[1 + u(x, t)]

7r
(10.7) Ut- duxx
in the interval (0, 1) with Neumann boundary conditions. Yoshida [1] and
Morita [1] have discussed the existence and stability of the spatially inde-
pendent periodic orbit that arises through a Hopf bifurcation as a function
of the parameters d, p,. For a fixed value of p, > 0, Morita [1] has shown
that this solution is unstable for d < d0 , with d0 sufficiently small. Memory
[3] has shown there is a positive constant d 1 > d0 at which another Hopf
bifurcation from zero occurs, resulting in an unstable, spatially varying,
periodic solution. She also shows how to destabilize the original periodic
solution (as dis decreased) before this bifurcation occurs by replacing the
term u(x, t- 1) by u(x, t- 1) + hu 3 (x, t- 1) for appropriate h. This shows
that the global attractor can exhibit interesting dynamic behavior.
Recently, stochastic RFDE have received some attention. The method
of averaging has been extended to the case of stochastic evolutionary equa-
tions by Seidler and Vrkoc [1], and Maslowski, Seidler, and Vrkoc [1,2]. The
results here also overlap with the averaging procedure of Section 12.8.
For questions in stochastic RFDE related to the topics discussed in
these notes-existence, uniqueness, stability, Liapunov exponents, variation-
of-constants formula, stable manifolds-see Ito and Nisio [1], Mizel and
Trutzer [1], Mohammed [1,2], Mohammed and Scheutzow [1], Mohammed,
Scheutzow, and Weizsiicker [1] and Scheutzow [1] and the extensive refer-
ences in Scheutzow [2].
Freidlin [1], Freidlin and Wentzell [1], and Ventsel and Freidlin [1] (same
people) have given an extensive theory of large deviations for Gaussian pro-
cesses with values in Hilbert spaces. They have applied these results to the
study of random perturbations of ordinary differential equations. In partic-
ular, for an ordinary differential equation with a globally stable equilibrium
point 0 and any neighborhood V of 0, they use quasi-potentials to determine
the most likely point of escape from V. Langevin, Oliva, and de Oliveira
[1] have extended such results to random perturbations of RFDE. Similar
results have been given for NFDE by de Oliveira [1]. Galves, Langevin, and
Vares [1] have considered similar problems for maps when the attractor is
one-dimensional with three fixed points (similar problems for differential
equations had been considered by Freidlin and Wentzell [1]).
In recent years, there have been many papers devoted to the oscillatory
properties of the solutions of scalar delay differential equations. For linear
autonomous equations, it is a general rule that a necessary and sufficient
condition for solutions to be oscillatory is that no roots of the characteristic
equation be real. In the nonautonomous case, only sufficient conditions
have been given. There are also several results on autonomous nonlinear
equations that are related to stability and instability properties of solutions.
The reader may consult the book of Gyori and Ladas [1] and the proceedings
of a recent conference (Graef and Hale [1]) for details and references.
Appendix
Stability of characteristic equations
The purpose of this appendix is to give methods for dete.rmining when the
roots of a characteristic equation are in the left half-plane. The most general
results are due to Pontryagin [1] for the zeros of characteristic equations
of the form P(z, ez) = 0 where P(x, y) is a polynomial in x, y. Pontryagin
gave necessary and sufficient conditions for all solutions of P(z, ez) to lie
in the left half-plane. To obtain the results, he extended the methods used
in proving the Routh-Hurwitz criterion for the zeros of a polynomial to be
in the left half-plane. We state the results of Pontryagin without proof and
give applications to a few specific equations.
Suppose P(z,w) is a polynomial in z,w,
r s
(A.l) P(z,w) = L LamnZmWn.
m=Dn=O
We call arsZrws the principal term of the polynomial if ars -/:- 0 and if for
each other term amnZmwn with amn -:f. 0, we have either r > m, 8 > n,
or r = m, 8 > n, or r > m, 8 = n. Clearly, not every polynomial has a
principal term.
If w = ez, then P(z, ez) = 0 corresponds to the characteristic equation
for the scalar differential difference equation
(A.2)
The equation P(z, ez) could also correspond to a matrix system of differ-
ential difference equations. One important thing to notice is that the only
characteristic equations that can be discussed by the methods of this ap-
pendix are those for which the delays have ratios that are rational. One can
then change the time variable to obtain integer delays.
In Equation (A.2), let
x(t) = Yl(t), dx(t)/dt = Y2(t), ... , dr-lx(t)/dtr-l = Yr(t).
Then Equation (A.2) can be written as the system

Appendix: Stability of characteristic equations 415
il1-1(t) = Y1(t), j = 2,3, ... ,r

r-1
L arniJ(t + n) = L L amnY(t + n).

8 8
(A.3)
n=O m=On=O
To say that P(z, w) has a principal term is equivalent to saying that System
(A.3) is a neutral differential difference equation according to the definition
in Chapter 9. For neutral equations, we have previously remarked that all
zeros of P(z, ez) = 0 must have real parts bounded above. The fact that
these equations are the only ones for which this is true is a consequence of
Theorem A.l. If the polynomial P(z, w) has no principal term, then the
equation P(z, ez) = 0 has an infinity of zeros with arbitrarily large real
parts.
The basic results for applications are the next theorems.
Theorem A.2. Let Ll(z) = P(z, ez) and suppose P(z, w) is a polynomial
with principal term ar 8 ZrW 8 • All of the zeros of .d(z) have negative real
parts if and only if
(i) The complete vector .1( iy) rotates in the positive direction with a pos-
itive velocity for y ranging in ( -oo, oo).
(ii) For y E [-2k1f, 2k1f], k 2: 0 an integer, there is an Ek ---+ 0 ask ---+ oo
such that .d(iy) subtends an angle 4k1fs + 1rr + Ek·
Theorem A.3. Let Ll(z) = P(z, ez) where P(z, w) is a polynomial with prin-
cipal term. Suppose .d(iy), y E IR is separated into its real and imaginary
parts, .d(iy) = F(y) + iG(y). If all zeros of .d(z) have negative real parts,
then the zeros of F(y) and G(y) are real, simple, alternate, and
(A.4) G'(y)F(y)- G(y)F'(y) >0

for y E JR. Conversely, all zeros of Ll(z) will be in the left half-plane pro-
vided that either of the following conditions is satisfied:
(i) All the zeros of F(y) and G(y) are real, simple, and alternate and
Inequality (A.4) is satisfied for at least one y.
(ii) All the zeros of F(y) are real and for each zero, Relation (A.4) is
satisfied.
(iii) All the zeros of G(y) are real and for each zero, Relation (A.4) is
satisfied.
One other result is needed for the applications. Suppose f(z, u, v) is a

polynomial in z, u, v with real coefficients that has the form
416 Appendix: Stability of characteristic equations
r s
(A.5) f(z,u,v) = L Lzm¢~l(u,v)
m=On=O
where ¢~) (u, v) is a homogeneous polynomial of degree n in u, v. The prin-
cipal term in the polynomial f(z,u,v) is the term zr¢~s)(u,v) for which
either r > m, s > n or r = m, s > n or r > m, s = n for all other terms in
(A.5).
Let zr ¢~s) denote the principal term of f(z, u, v) in (A.5), let ¢~s) (u, v)
denote the coefficient of zr in f(z, u, v),
8
¢~s)(u,v) = L¢~n)(u,v),
n=O
and let
4>~8 )(z) = ¢~s)(cosz,sinz).
Theorem A.4. Let f(z, u, v) be a polynomial with principal term zr ¢~s) (u, v).
If E is such that 4>~s) (E + iy) =f. 0, y E IR, then, for sufficiently large integers
k, the function F(z) = f(z,cosz,sinz) will have exactly 4ks + r zeros in
the strip -2k7r + E:::; Re z :::; 2k7r +E. Consequently, the function F(z) will
have only real roots if and only if, for sufficiently large integers k, it has
exactly 4ks + r roots in the strip - 2kn + E:::; Re z :::; 2kn + E.
The following result is due to Hayes [1] with the proof based on Bellman
and Cooke [1].
Theorem A.5. All roots of the equation (z + a)ez + b = 0, where a and bare
real, have negative real parts if and only if
a> -1
(A.6) a+b>O
b < ( sin ( - a cos (
where ( is the root of (=-a tan(, 0 < ( < n, if a =f. 0 and ( = n/2 if
a= 0.
Proof. If Ll(z) = (z + a)ez + b; Ll(iy) = F(y) + iG(y), y E IR, then
F(y) =a cosy- ysiny + b

(A.7)
G(y) = asiny + ycosy.
Necessity. From Theorem A.3, the zeros of G(y) must be real and simple.
If g(y,u,v) = uy + av, G(y) = g(y,cosy,siny), then the function g>~s)(z)
in Theorem A.4 is cos z and we may take E = 0. For k sufficiently large,
Theorem A.4 implies the function G(y) has exactly 4k+ 1 zeros for -2kn:::;
y :::; 2k7r.
Appendix: Stability of characteristic equations 417
The equation G(y) = 0 is equivalent to the equation
(A.8) y = -atany.
We must have a =I= -1, for otherwise, the equation G(y) = 0 has a triple root
at y = 0, which contradicts Inequality (A.4). If a< -1, then there is only
one root in [-1r, 1r] and, exactly one root in any interval [n1r, (n + 1)7r], n =/=
0, -1. Therefore, there are exactly 4k- 1 roots on the interval [-2k7r, 2k7r]
for any integer k. This contradicts Theorem A.4 since we should have 4k+ 1.
Therefore, a > -1.
If a> -1, then there is exactly one root Yn in each interval (n1r, 1r(n +
1)) and no other roots except the root y = 0 for n = 0, n = -1. Let Yo = 0.
Let us now check Inequality (A.4) at the zeros of G. To do this, first observe
that G' (y) is given by
G'(y) =a cosy+ cosy- ysiny.
If y = 0, then G'(O)F(O) = (a+ 1)(a +b) > 0 implies a+ b > 0. It is easy

to verify that, for any y = Yn =/= 0, (siny)G'(y) = -y + ~ sin2y. Therefore,
(siny)G'(y) > 0 for ally E (O,oo): Since Yl E (0,1r), we have siny1 > 0.
Therefore, G' (y 1) < 0 and Relation (A.4) is valid at y = y1 if and only
if F(yl) < 0. This is precisely the relation b < y1 siny1 - acosy 1 . This
completes the proof of the necessity.
Sufficiency. Assume a > -1. As before, Theorem A.4 implies all the roots
of G(y) = 0 are real. Observe that Yn = Y-n and, thus, we need only check
Condition (iii) of Theorem A.3 for n ~ 0. For y = 0, we have already
observed that G'(O)F(O) =(a+ 1)(a +b) > 0 if a> -1, a+ b > 0. We also
have observed that (siny)G'(y) < 0 for ally E (O,oo). Observe first that
the last of Inequalities (A.6) is equivalent to b < y' a 2 + (2. At any y = Yn
with n odd, we have
F(y) = y'a 2 +y2 +b if a> 0

F(y) = -y'a2 + y 2 + b if a< 0.
If a> 0 and (A.6) is satisfied, then F(y) > 0. But, if a> 0, then sin y2k+l <
0, G'(Y2k+l) > 0 and Condition (A.4) is satisfied. If a< 0 and Conditions
(A.6) are satisfied, then F(y) < 0. But, if a < 0, then sin y2k+ 1 > 0,
G'(Y2k+l) < 0 and Condition (A.4) is satisfied. The roots y 2k of G are
treated in a similar manner to complete the proof of the theorem. D
For the equation (z+a)ezr +b = 0, Boese [1] has given a more explicit
stability chart in terms of r < r 0 (a, b).
Theorem A.6. All roots of the equation (z 2 +az)ez + 1 = 0 have negative real
parts if and only if a> (sin()/( where (is the unique root of the equation
( 2 =cos(, 0 < ( < 1rj2.
418 Appendix: Stability of characteristic equations
Proof. If L1(z) = (z 2 + az)ez + 1, L1(iy) = F(y) + iG(y), y E IR, then
F(y) = -y 2 cosy- aysiny + 1,

(A.9)
G(y) = -y 2 siny+aycosy.
If g(y,u,v) = -y 2 v + yu, then G(y) = g(y,cosy,siny) and the function

<P~s)(z) in Theorem A.4 is (-siny). Therefore, we may take E = Jr/2 in
Theorem A.4. From Theorem A.4, all zeros of G(y) are real if and only if
there are 4k+2 real zeros of G(y) in the interval [-2k1r+1r /2, 2k1r+1r /2] for
k a sufficiently large integer. Observe that G(O) = 0 and y -1- 0, G(y) = 0
is equivalent to the equation
(A.10) y =a cot y.
Therefore, G(y) = 0 has 4k + 2 roots on the interval [-2k7r + (Ir/2), 2k7r +

(Ir/2)] if and only if a> 0. Theorem A.4 implies the zeros of G(y) are real
if and only if a > 0.
Necessity. If all zeros of L1(z) are in the left half-plane, then we must have
Inequality (A.4) satisfied at the zeros of G(y); that is, G'(y)F(y) > 0 for
ally such that G(y) = 0. Since
G'(y) = -(2 + a)ysiny- (y 2 - a) cosy

we have G'(O)F(O) =a> 0. Using Equation (A.10), we obtain
G ' (y ) = -y sin y ( y 2 +a 2 +a )
a
F(y) = 1- ysiny (y2 + a2).
a
Therefore, the sign of G' (y )F(y) is determined by the sign of the expression
. 2 .
h(y) = sm y (y2 +a2)- smy.
a2 ay
From Equation (A.10), we observe that
siny
h(y) - 1 - -
ay
at any zero y -1- 0 of G(y). Thus, we must have a> (sin y)jy for all solutions
y of Equation (A.10). It is clear that this requires restricting a so that a >
(siny 1 (z))/y 1 (a) where YI(a) is the unique root of G(y) = 0, 0 < y < Ir/2.
Let ( be the unique root of ( 2 = cos(, 0 < ( < 1r /2. One can now check
that a > ( is equivalent to the last statement. This proves necessity.
Sufficiency. One easily reverses the steps and the proof of the theorem is
complete. D
Bibliography
Abraham, R. and J. Robbin [1] Transversal Mappings and Flows. Benjamin, 1967.
Alt, W. [1] Some periodicity criteria for functional differential equations. Manu-
scripta math. 23 (1978), 295-318; [2] Periodic solutions of some autonomous
differential equations with variable time delay. Lect. Notes Math. 730 (1979),
16-31.
an der Heiden, U. [1] Periodic solutions of a nonlinear second order differential
equation with delay. J. Math. Anal. Appl. 70 (1979), 599-609.
an der Heiden, U. and M. C. Mackey [1] The dynamics of production and de-
struction: analytic insight into complex behavior. J. Math. Biol. 16 (1982),
75-101.
an der Heiden, U. and H.-0. Walther [1] Existence of chaos in control systems
with delayed feedback. J. Differential Eqns. 47 (1983), 273-295.
Artola, M. [1] Sur les perturbations des equations d'evolution. Application a des
problemes de retard. Annales Ec. Norm. Sup. 2 (1969), 137-253.
Artstein, Z. [1] On continuous dependence of fixed points of condensing maps.
Dynamical Systems-An International Symposium, 73-76, Academic Press,
1976.
Asner, B. A. [1] New constructions for pointwise degenerate systems. IFAC, 6th
Triennial World Conf. 1 (1975), 9.6.1-9.6.5.
Anser, B. A. and A. Halanay [1] Pointwise degenerate second-order delay dif-
ferential systems. Anal. Univ. Bucharesti, Mat. Mec. 22 (1973), 45-60; [2]
Algebraic theory of pointwise degenerate delay differential systems. J. Dif-
ferential Eqns. 14 (1973), 293-306; [3] Delay-feedback using derivatives for
minimal time linear control processes. J. Math. Anal. Appl. 48 (1974), 257-
262; [4] Non-controllability of time-invariant systems using one-dimensional
linear delay feedback. Rev. Roumaine Sci. Tech. Ser. Electrotechnique et
Energetique 18 (1973), 283-293.
Avellar, C. E. and J. K. Hale [1] On the zeros of exponential polynomials. J.
Math. Anal. Appl. 13 (1980), 434-452.
Babin, A. B. and M. I. Vishik [1] Attractors in Evolutionary Equations (in Rus-
sian) Nauka, Moscow, 1989.
Bailey, H. R. and E. B. Reeve [1] Mathematical models describing the distribution
of 1131 -albumin in man. J. Lab Clin. Med. 60 (1962), 923-943.
Bailey, H. R. and M. Z. Williams [1] Some results on the differential difference
equation i:(t) = L:o A;x(t- T;). J. Math. Anal. Appl. 15 (1966), 569-587.
420 Bibliography
Baiocchi, C. [1] Teoremi di esistenza e regolarita per certe classi di equazioni

differenziali astratte. Ann. Math. Pura Appl. (4) 72 (1966), 365-418; [2]
Sulle equazioni differenziali astratte lineari del primo e del secondo ordine
negli spazi di Hilbert. Ann. Mat. Pura. App. (4) 76 (1967), 233-304.
Banks, H. T. [1] Modeling and Control in the Biomedical Sciences. Lecture Notes
in Biomathematics, Vol. 6, Springer-Verlag, 1975;
Banks, H. T. and G. Kent [1] Control offunctional differential equations to target
sets in functions space. SIAM J. Control10 (1972), 567-593.
Banks, H. T. and A. Manitius [1] Projection series for retarded functional differen-
tial equations with applications to optimal control problems. J. Differential
Eqns. 18 (1975), 296-332.
Barbu, V. and S. Grossman [1] Asymptotic behavior of linear integral differential
systems. Trans. Amer. Math. Soc. 173 (1972), 277-289.
Barnea, D. I. [1] A method and new results for stability and instability of au-
tonomous functional differential equations. SIAM J. Appl. Math. 17 (1969),
681-697.
Bartosiewicz, z. [1] Density of images of semigroup operators for linear neutral
functional differential equations. J. Differential Eqns. 38 (1980), 161-175.
Bellman, R. and K. Cooke [1] Differential Difference Equations. Academic Press,
1963; [2] Stability theory and adjoint operators for linear differential differ-
ence equations. Trans. Amer. Math. Soc. 92 (1959), 470-500; [3] Asymptotic
behavior of solutions of differential difference equations. Mem. A mer. Math.
Soc. 35 (1959); [4] On the limit solutions of differential difference equations
as the retardation approaches zero. Proc. Nat. Acad. Sci. 45 (1959), 1026-
1028.
Bellman, R. and J. M. Danskin [1] A survey of the mathematical theory of time
lag, retarded control, and hereditary processes. The Rand Corporation, R-
256, 1954.
Berre, M. L., Ressayre, E., Tallet, A. and H. M. Gibbs, High dimension chaotic
attractors of a nonlinear ring cavity. Phys. Rev. Lett. 56 (1986), 274-277.
Bhatia, N. and 0. Hajek [1] Local Semi-Dynamical Systems. Lecture Notes in
Math., vol. 90, Springer-Verlag, 1969.
Billotti, J. E. and J. P. LaSalle [1] Periodic dissipative processes. Bull. Amer.
Math. Soc. 6 (1971), 1082-1089.
Biroli, M. [1] Solutions presque periodiques d'une equation et d'une
inequation parabolique avec terme de retard nonlineaire. I, II, III Atti Accad.
Naz. Lincei Rend. Cl. Sci. Fis. Mat. Natur. (8) 48 (1970), 576-580; ibid (8)
49 (1970), 23-26 (1971); ibid (8) 49 (1970), 175-179 (1971).
Blythe, S. P., Nisbet, R. M. and W. S. C. Gurney [1] Instability and complex
dynamic behavior in population models with long time delays. Theor. Pop.
Biol. 2(1982), 147-176.
Boas, R. [1] Entire Functions. Academic Press, New York, 1954.
Boese, F. G. [1] Some stability charts and stability conditions for a class of
difference-differential equations. Z. Angew. Math. Mech. 67 (1987), 56-59.
Boffi, V. and R. Scozzafava [1] Sull'equazione funzionale lineare
j'(x) = -A(x)f(x- 1), Rend. Math. e Appl. (5) 25 (1966), 402-410; [2]
A first-order linear differential difference equation with N delays. J. Math.
Anal. Appl. 17 (1967), 577-589.
Bogoliubov, N. N. andY. A. Mitropolsky [1] Asymptotic Methods in the Theory
of Nonlinear Oscillations, Gordon and Breach, New York, 1961.
Bibliography 421
Brayton, R. [1] Nonlinear oscillations in a distributed network. Quart. Appl. Math.

24 (1976), 289--301; [2] Small signal stability criterion for electrical networks
containing lossless transmission lines. IBM J. Res. Dev. 12 (1968), 431-440.
Brayton, R. and R. A. Willoughby [1] On the numerical integration of a symmetric
system of differential difference equations of neutral type. J. Math. Anal.
Appl. 18 (1967), 182-189.
Browder, F. [1] On a generalization of the Schauder fixed-point theorem. Duke
Math. J. 26 (1959), 291-303; [2] A further generalization of the Schauder
fixed-point theorem. Duke Math. J. 32 (1965), 575-578.
Brumley, W. E. [1] On the asymptotic behavior of solutions of differential differ-
ence equations of neutral type. J. Differential Eqns. 7 (1970), 175-188.
Burd, V. S. and Ju. S. Kolesov [1] On the dichotomy of solutions of functional
differential equations with almost periodic coefficients. Sov. Math. Dokl. 11
(1970), 1650-1653.
Burton, T. A. [1] Uniform asymptotic stability in functional differential equa-
tions. Proc. Am. Math. Soc. 38 (1978), 195-200; [2] Volterra Integral and
Differential Equations. Academic Press, New York, 1983; [3] Stability and
Periodic Solutions of Ordinary and Functional Differential Equations, Aca-
demic Press, New York, 1983.
Cao, Y. [1] The discrete Lyapunov function for scalar differential delay equations.
J. Differential Eqns. 87 (1990), 365-390; [2] Non-existence of small solution<;
for scalar differential delay equations. ?reprint 1992.
Carr, J. [1] Applications of Centre Manifold Theory. Springer-Verlag, New York,
1981.
Cartwright, M. L. [1] Almost-periodic flows and solutions of differential equations.
Proc. London Math. Soc. (3) 17 (1967), 355-380. Corringenda (3) 17 (1967),
768; [2] Almost-periodic differential equations and almost-periodic flows, J.
Differential Eqns. 5 (1969), 167-181.
Cesari, L. [1] Nonlinear oscillations in the frame of alternative problems, vol. I,
29-50. Dynamical Systems-An International Symposium. Academic Press,
1976; [2] Alternative methods in nonlinear analysis, Intern. Conf. Diff. Eqns.
Los Angeles, September 1974, 95-148. Academic Press, 1975.
Chafee, N. [1] The bifurcation of one or more closed orbits from an equilibrium
point of an autonomous differential equation. J. Differential Eqns. 4 (1968),
661-679; [2] A bifurcation problem for functional differential equations of
finitely retarded type. J. Math. Anal. Appl. 35 (1971), 312-348.
Charrier, P. [1] Linear delay differential systems-Controllability to a function
space. 31-33 of Proc. Symp. Differential-Delay and Functional Equations.
Univ. Warwick, July 1972.
Chary, K. S. R. [1] A note on functional differential equations of neutral type.
Proc. Nat. Acad. Sci. India Sect. A 43 (1973), 271-278.
Chow, S. [1] Remarks on one-dimensional delay differential equations. J. Math.
Anal. Appl. 41 (1973), 426-429; [2] Existence of periodic solutions of au-
tonomous functional differential equations. J. Differential Eqns. 15 (1974),
350-378.
Chow, S.-N. and D. Green, Jr. [1] Some results on singular delay-differential
equations. In Chaos, Fractals and Dynamics (P. Fischer and W. R. Smith,
Eds.) Marcel Dekker, 1986, 161-182.
422 Bibliography
Chow, S. and J. K. Hale [1] Strongly limit compact maps, Funk. Ekv. 17 (1974),
31-38; [2] Periodic solutions of autonomous equations. J. Math. Anal. Appl.
66 (1978), 495-506.
Chow, S.-N., Hale, J. K. and W. Huang [1] From sine waves to square waves in
delay equations. Proc. Roy. Soc. Edinburgh 120A (1992), 223-229.
Chow, S.-N. and W. Huang [1] Singular perturbation problems for a system of
differential difference equations. Submitted to J. Differential Eqns.
Chow, S.-N. and J. Mallet-Paret [1] Integral averaging and bifurcation. J. Dif-
ferential Eqns. 26 (1977), 112-159; [2] The Fuller index and global Hopf
bifurcation. J. Differential Eqns. 29 (1978), 66-85; [3] Singularly perturbed
delay differential equations. In Coupled Nonlinear Oscillators (Eds. J. Chan-
dra and A. Scott), North Holland, 1983.
Chow, S.-N. and H.-0. Walther [1] Characteristic multipliers and stability of
periodic solutions of i:(t) = g(x(t- 1)). Trans. Am. Math. Soc. 307 (1988),
127-142.
Clement, Ph., Diekmann, 0., Gyllenberg, M., Heijmans, H.J.A.M. and H.R.
Thieme [1] Perturbation theory for dual semigroups. I. The sun-reflexive
case. Math. Ann. 277 (1988), 709-725.
Coffman, C. V. and J. J. Schaffer [1] Linear differential equations with delays. Ex-
istence, uniqueness, growth, and compactness under natural Caratheodory
conditions. J. Differential Eqns. 16 ( 1974), 26-44; [2] Linear differential equa-
tions with delays: Admissibility and conditional exponential stability. J. Dif-
ferential Eqns. 9 (1971), 521-535.
Coleman, B. D. and H. Dill [1] On the stability of certain motions of incompress-
ible materials with memory. Arch. Rat. Mech. Anal. 30 (1968), 197-224.
Coleman, B. D. and V. J. Mizel [1] Norms and semigroups in the theory of fading
memory. Arch. Rat. Mech. Anal. 2 (1966), 87-123; [2] On the general theory
of fading memory. Arch. Rat. Mech. Anal. 29 (1968), 18-31; [3] On the
stability of solutions of functional differential equations. Arch. Rat. Mech.
Anal. 30 (1968), 173-196; [4] A general theory of dissipation in materials
with memory. Arch. Rat. Mech. Anal. 27 (1968), 255-274; [5] Existence of
entropy as a consequence of asymptotic stability. Arch. Rat. Mech. Anal. 25
(1967), 243-270.
Coleman, B. D. and D. R. Owen [1] On the initial-value problem for a class of
functional differential equations. Arch. Rat. Mech. Anal. 55 (1974), 275-299.
Comincioli, V. [1] Ulteriori osservazioni sulle soluzioni del problema periodico per
equazioni paraboliche lineari con termini di perturbazioni. !st. Lombardo
Acad. Sci. Lett. Rend. A 104 (1970), 726-735.
Cooke, K. L. [1] Functional differential equations: Some models and perturbation
problems. Differential Equations and Dynamical Systems (Proc. Int. Symp.
Mayaguez, P. R. (1965), 167-183. Academic Press, 1967; [2] Linear func-
tional differential equations of asymptotically autonomous type. J. Differ-
ential Eqns. 7 (1970), 154-174; [3] The condition of regular degeneration for
singularly perturbed linear differential difference equations. J. Differential
Eqns. 1 (1965), 39-94.
Cooke, K. and J. Ferreira [1] Stability conditions for linear retarded functional
differential equations. J. Math. Anal. Appl. 96 (1983), 480-504.
Cooke, K. and D. Krumme [1] Differential difference equations and nonlin-
ear initial-boundary-value problems for linear hyperbolic partial differential
equations. J. Math. Anal. Appl. 24 (1968), 372-387.
Bibliography 423
Cooke, K. L. and K. R. Meyer [1] The condition of regular degeneracy for singular
perturbed systems of linear differential difference equations. J. Math. Anal.
Appl. 14 (1966), 83-106.
Cooke, K. L. and S. M. Verduyn Lunel [1] Distributional and small solutions for
linear time-dependent delay equations, Vrije Universiteit Amsterdam Rap-
port WS-393 (1992), to appear Differential Integral Eqn.
Cooke, K. L. and J. A. Yorke [1] Equations modelling population growth, eco-
nomic growth, and gonorrhea epidemiology, 35-55 in Ordinary Differential
Equations, L. Weiss, Ed., Academic Press, 1972.
Coppel, W. A. [1] Stability and Asymptotic Behavior in Differential Equations.
Heath Mathematical Monographs, Boston, 1965.
Corduneanu, C. [1] Integral Equations and Stability of Feedback Systems. Aca-
demic Press, 1973; [2] Sur la stabilite des systemes perturbes a argument
retarde. An. Sti. Univ. "Al. I. Cuza" Ia§i Sect. I. a Mat. {N.S.) 11 (1965),
99-105.
Cruz, M. A. and J. K. Hale [1] Existence, uniqueness and continuous dependence
for hereditary systems. Annali Mat. Pura Appl. (4) 85 (1970), 63-82; [2]
Stability of functional differential equations of neutral type. J. Differential
Eqns. 7 (1970), 334-355; [3] Asymptotic behavior of neutral functional differ-
ential equations. Arch. Rat. Mech. Anal. 34 (1969), 331-353; [4] Exponential
estimates and the saddle-point property for neutral functional differential
Cunningham, W. J. [1] A nonlinear differential difference equation of growth.
Proc. Nat. Acad. Sci. U.S.A. 40 (1954), 709-713.
Dafermos, C. [1] On the existence and asymptotic stability of solutions to the
equations of linear thermoelasticity. Arch. Rat. Mech. Anal. 29 (1968), 241-
271; [2] An abstract Volterra equation with applications to linear viscoelas-
ticity. J. Differential Eqns. 7 (1970), 554-569; [3] Semiflows associated with
compact and uniform processes. Math. Sys. Theory 8 (1974), 142-149.
Darbo, G. [1] Punti uniti in transformazioni a condominia non compatto. Rend.
Sem. Math. Univ. Padova 24 (1955), 84-92.
Datko, R. [1] An algorithm for computing Liapunov functionals for some differen-
tial difference equations. In Ordinary Differential Equations (L. Weiss, Ed.),
(1972), 387-398, Academic Press; [2] Linear autonomous neutral differential
equations in a Banach space. J. Differential Eqns. (1977), 258-274; [3] Re-
marks concerning the asymptotic stability and stabilization of linear delay
differential equations. J. Math. Anal. Appl. 111 (1985), 571-584.
Delfour, M. C. and A. Manitius [1] The structural operator F and its role in the
theory of retarded systems I. J. Math. Anal. Appl. 73 (1980), 466-490; [2]
The structural operator F and its role in the theory of retarded systems II.
J. Math. Anal. Appl. 74 (1980), 359-381.
de Nevers, K. and K. Schmitt [1] An application of the shooting method to
boundary-value problems for second-order delay equations. J. Math. Anal.
Appl. 36 (1971), 588-597.
Derstine, M. W., Gibbs, H. M., Hopf, F. A. and D. L. Kaplan [1] Bifurcation gap
in a hybrid optically bistable system. Phys. Rev. A, 26 (1982), 3720-3722;
[2] Alternate paths to chaos in optical bistability. Ibid 27 (1983), 3200-3208.
Diekmann, 0. [1] Volterra integral equations and semigroups of operators. MC
Report TW 197 Centre for Mathematics and Computer Science, Amsterdam;
[2] A duality principle for delay equations. In Equadiff 5 (M. Gregas, Ed.)
Teubner Texte zur Math. 47 (1982), 84-86; [3] Perturbed dual semigroups
424 Bibliography
and delay equations. In Dynamics of Infinite Dimensional Systems (S.-N.

Chow and J. K. Hale, Eds.) Springer-Verlag, Series F: 37 (1987), 67-74.
Diekmann, 0. and S. A. van Gils [1] The center manifold for delay equations in
the light of suns and stars. Preprint, 1990.
Diekmann, 0., van Gils, S. A., Verduyn Lunel, S.M. and H.O. Walther [1] Delay
Equations: Complex, Functional, and Nonlinear Analysis. Springer-Verlag,
New York, to appear.
Diestel, J. and J. J. Uhl [1] Vector Measures. Amer. Math. Soc., Math. Surveys
15, 1973.
Doetsch, G. [1] Handbuch der Laplace-Transformation Band I, Birkhiiuser, Basel,
1950.
Dombrovskii, V. A. [1] The stability of periodic solutions of systems with dis-
tributed parameters and lag. [Russian] Ukrain. Mat. Z. 24 (1972), 161-170.
Driver, R. D. [1] A functional differential system of neutral type arising in a two--
body problem of classical electrodynamics. 474-484. Nonlinear Differential
Equations and Nonlinear Mechanics. Academic Press, 1963; [2] Existence
and continuous dependence of solutions of a neutral functional differential
equation. Arch. Rat. Mech. Anal. 19 (1965), 149-166; [3] Ordinary and Delay
Differential Equations. Springer-Verlag, New York, 1977.
Dugundji, J. [1] An extension of Tietze's theorem. Pac. J. Math. 1 (1951), 353-
367.
Dunkel, G. [1] Single-species model for population growth depending on past
history. Seminar on Differential Equations and Dynamical Systems, 92-99.
Lecture Notes in Math. vol. 60. Springer-Verlag, 1968.
El'sgol'tz, L. E. [1] Qualitative Methods in Mathematical Analysis. Trans. Math.
Mono., val. 12, Amer. Math. Soc. 1964; [2] Introduction to the Theory of
Differential Equations with Deviating Arguments. Holden-Day, 1966.
El'sgol'tz, L. E. and S. P. Norkin [1] Introduction to the Theory and Applicaton of
Differential Equations with Deviating Arguments. Translated from Russian
by J. L. Casti, Academic Press, 1963.
Ergen, W. K. [1] Kinetics of the circulating fuel nuclear reaction. J. Appl. Phys.
25 (1954), 702-711.
Faria, T. and L. T. Magalhaes [1] Normal forms for retarded functional differen-
tial equations and applications to Bogdanov singularity. Preprint 1991; [2]
Normal forms for retarded functional differential equations with parameters
and applications to Hopf bifurcation. Preprint 1991; [3] Realization of or-
dinary differential equations by retarded functional differential equations in
neighborhoods of equilibrium points. Preprint 1992; [4] Restrictions on the
possible flows of scalar retarded functional differential equations in neigh-
bourhoods of singularities. Preprint 1992.
Farkas, M. and G. Stepan [1] On perturbation of the kernel in infinite delay
systems. ZAMM 72 (1992), 153-156.
Farmer, J. D. [1] Chaotic attractors of an infinite dimensional dynamical system.
Phys. D 4 (1982), 366-393.
Fennell, R. and P. Waltman [1] A boundary-value problem for a system on non-
linear functional differential equations. J. Math. Anal. Appl. 26 (1969), 447-
453.
Fiedler, B. [1] Discrete Lyapunov functionals and w-limit sets. lvf AN 23 (1989),
415-431.
Bibliography 425
Fiedler, B. and J. Mallet-Paret [1) A Poincare-Bendixson theorem for scalar re-

action diffusion equations. Arch. Rat. Mech. Anal. 107 (1989), 325-345; [2)
Connections between Morse sets for delay differential equations. J. Reine
Angew. Math. 397 (1989), 23-41.
Fink, A. M. [1), Almost Periodic Differential Equations. Lect. Notes Math. vol.
377 (1974), Springer-Verlag.
Fitzgibbon, W. E. [1) Semilinear functional differential equations in Banach space.
J. Differential Eqns. 29 (1978), 1-14.
Fodcuk, V. I. [1) Integral varieties for nonlinear differential equations with re-
tarded arguments. [Russian) Ukrain. Mat. Z. 21 (1969), 627-639; [2) Inte-
gral manifolds for nonlinear differential equations with retarded arguments.
[Russian) Differentialniye Uravenija 6 (1970), 798-808.
Franke, J. M. and H. W. Stech [1), Extensions of an algorithm for the analy-
sis of nongeneric Hopf bifurcations, with applications to delay-differential
equations. In Lect. Notes Math. 1475 (1991), 161-175.
Freidlin, M. J. [1) Semilinear PDE and limit theorems for large deviations.
?reprint 1991.
Freidlin, M. J. and A. D. Wentzell [1) Random Perturbations of Dynamical Sys-
tems. Springer-Verlag, 1985.
Friesecke, G. [1) Exponentially growing solutions for a delay diffusion equation
with negative feedback. J. Differential Eqns. 98 (1992), 1-18.
Fuller, F. B. [1) An index of fixed-point type for periodic orbits. Am. J. Math. 89
(1967), 133-148.
Fusco, G. and W. M. Oliva [1) Jacobi matrices and transversality. Proc. Royal
Soc. Edinb. 109A (1988), 231-241.
Galves, A., Langevin, R. and M. E. Vares [1), Tunneling for randomly perturbed
Morse-Smale systems. IMPA (1989). ?reprint.
Gantmacher, F. R. [1) Theory of Matrices, vol. II. Chelsea, 1960.
Gerstein, V. M. [1) On the theory of dissipative differential equations in a Banach
space. [Russian) Funk. Anal. i Prilozen, 4 (1970), 99-100.
Gerstein, V. M. and M. A. Krasnoselskii [1) Structure of the set of solutions of
dissipative equations [Russian) Dokl. Akad. Nauk SSSR 183 (1968), 267-269.
Gibbs, H. M., Hopf, F. A., Kaplan, D. L. and R. L. Shoemaker [1) Observation
of chaos in optical bistability. Phys. Rev. Letters. 46 (1981), 474-477.
Ginzburg, R. E. [1) An application of the method of Liapunov functions to the
investigation of oscillations in nonlinear systems with lag. [Russian) Differ-
entialniye Uravnenija 7 (1971), 1903-1905; [2) Oscillations of linear systems
with autonomous self-regulating lag [Russian) Differentialniye Uravnenija 6
(1970), 1257-1264.
Glass, L. and M. C. Mackey [1) Pathological conditions resulting from instabilities
in physiological control systems. Ann. New York Acad. Sci. 316 (1979), 214-
235.
Gohberg, I. C. and E. I. Sigal [1) An operator generalization of the logarithmic
residue theorem and the theorem of Rouche. Mat. Sb. 84 (1971) 609-629
(Russian) (Math. USSR Sb. 13 (1971), 603-625).
Gopalsamy, K. [1) Stability and Oscillations in Delay Differential Equations of
Population Dynamics. Kluwer Academic Publishers, Dordrecht, 1992.
Graef, J. R. and J. K. Hale [1) Oscillation and Dynamics in Delay Equations.
Contemporary Math. 129, Amer. Math. Soc., Providence, 1992.
426 Bibliography
Grafton, R. (1] A periodicity theorem for autonomous functional differential equa-

tions. J. Differential Eqns. 6 (1969), 87-109; [2] Periodic solutions of certain
Leinard equations with delay. J. D·ifferential Eqns. 11 (1972), 519-527; [3]
Periodic solutions of Leinard equations with delay: Some theoretical and ex-
perimental results, 321-334, in Delay and Functional Differential Equations
and Their Applications. Ed. K. Schmitt, Academic Press, 1972.
Granas, A. [1] The theory of compact vector fields and some of its applications
to topology of function spaces (I). Rozprawy Mat. 30 (1962), 93 pp.
Greiner, G. and M. Schwarz [1] Weak spectral mapping theorems for functional
differential equations. J. Differential Eqns. 94 (1991), 205-216.
Grimm, L. J. and K. Schmitt [1] Boundary-value problems for differential equa-
tions with deviating arguments. Aequationes Math. 4 (1970), 176-190; (2]
Boundary-value problems for delay differential equations. Bull. A mer. Math.
Soc. 74 (1968), 997-1000.
Grimmer, R. and G. Seifert [1] Stability properties of Volterra integrodifferential
equations. J. Differential Eqns. 19 (1975), 142-166.
Grippenberg, G., Londen, S.-0. and 0. Staffans (1] Volterra Integral and Func-
tional Equations, Encyclopedia of Mathematics and Its Applications. Cam-
bridge University Press, 1990.
Gromova, P. S. and A. M. Zverkin (1] On trigonometric series whose sums are
continuous unbounded functions on the real axis-Solutions of equations
with retarded arguments. [Russian] Differentialniye Uravnenija 4 (1968),
1774-1784.
Grossberg, S. [1] A prediction theory for some nonlinear functional differential
equations. I. Learning of lists. J. Math. Anal. Appl. 21 (1968), 643-694.
II. Learning of patterns. Ibid. 22 (1968), 422-490; [2] Learning and energy-
entropy dependence in some nonlinear functional differential equations. Bull.
Amer. Math. Soc. 75 (1969), 1238-1242.
Grossman, S. F. [1] Stability in n-dimensional differential delay equations. J.
Math. Anal. Appl. 40 (1972), 541-546.
Gumowski, I. [1] Sur le calcul des solutions periodiques de !'equation de Cherwell-
Wright. C.R. Acad. Sci. Paris. Ser. A-B 268 (1969), A157-A159.
Gurney, W. S. C., Blythe, S. P. and R. M. Nisbet [1] Nicholson's blowflies revis-
ited. Nature 287 (1980), 17-21.
Gyori, I. and G. Ladas, Oscillation Theory of Delay Differential and Delay Dif-
ference Equations with Applications. Oxford Univ. Press, 1991.
Habets, P. [1] Singular perturbations of functional differential equations. Proc.
Vllth Int. Conf. on Nonlinear Oscillations, Berlin, Sept. 1975.
Hadeler, K. P. and J. Tomiuk [1] Periodic solutions of difference differential equa-
tions. Arch. Rat. Mech. Anal. 65 (1977), 87-95.
Hahn, W. [1] On difference differential equations with periodic coefficients. J.
Math. Anal. Appl. 3 (1961), 70-101.
Haddock, J. R., Krisztin, T. Terjecki, J. and J. H. Wu (1] An invariance principle
of Lyapunov-Razumikhin type for neutral functional differential equations.
J. Differential Eqns. To appear.
Halanay, A. [1] Differential Equations, Stability, Oscillations, Time Lags. Aca-
demic Press, 1966. Teoria Calitativa a Ecuatilior Diferentiale (Rumanian],
Editura Acad. Rep. Populaire Romine, 1963; (2] Almost periodic solutions
for a class of nonlinear systems with time lag. Rev. Roumaine Math. Pures
Bibliography 427
Appl. 14 (1969), 1269-1276; [3] On the method of averaging for differential

equations with retarded argument. J. Math. Anal. Appl. 14 (1966), 70-76.
Halanay, A. and J. Yorke [1] Some new results and problems in the theory of
differential delay equations. SIAM Review 13 (1971), 55-80.
Halanay, A. and D. Wexler [1] Qualitative Theory of Sampled-Data System. [Ru-
maniam] Editura Acad. RSR, Bucharest, 1968. Russian edition in 1971 pub-
lished by Mir, Moscow.
Hale, J. K. [1 J Forward and backward continuation for neutral functional differen-
tial equations. J. Differential Eqns. 9 (1971), 168-181; [2] Sufficient condi-
tions for stability and instability of autonomous functional differential equa-
tions. J. Differential Eqns. 1 (1965), 452-482; [3] Asymptotic behavior of the
solutions of differential difference equations. Proc. Int. Symp. Nonlin. Vibra-
tions, IUTAM, Kiev, 1963, vol. 2. 409-426; [4] Linear functional differential
equations with constant coefficients. Contributions to Differential Equations
2 (1963), 291-319; [5] Averaging methods for differential equations with re-
tarded arguments. J. Differential Eqns. 2 (1966), 57-73; [6] Linearly asymp-
totically autonomous functional differential equations. Rend. Oirc. Palermo
(2) 15 (1966), 331--351; [7] Applications of Alternative Problems. CDS Lec-
ture Notes 71-1, Brown Univ., Div. of Appl. Math., 1971; [8] Behavior near
constant solutions of functional differential equations. J. Differential Eqns.
15 (1974), 278--294; [9] Critical ca:>es for neutral functional differential equa-
tions. J. Differential Eqns. 10 (1971), 59-82; [10] Solutions near simple
periodic orbits of functional differential equations. J. Differential Eqns. 9
(1970), 126-183; [11] Functional Differential Equations. Appl. Math. Sci.,
vol. 3, Springer-Verlag, 1971; [12] Continuous dependence of fixed points of
condensing maps. J. Math. Anal. Appl. 46 (1974), 388-394; [13] Parametric
stability in difference equations. Bol. Un. Mat. It. (4) 11 Supp. (1975), 209-
214; [14] Smoothing properties of neutral equations. An. Acad. Brasil Ci. 45
(1973), 49-50; [15] A cla~s of neutral equations with the fixed-point property.
Proc. Nat. Acad. Sci. U.S.A. 67 (1970), 136-137; [16] a-contractions and dif-
ferential equations. Equations Differentielles et Fonctionelles Nonlineaires,
15-42. Hermann, Paris, 1973; [17] Oscillations in neutral functional differen-
tial equations. In Nonlinear Mechanics. C.I.M.E., June 1972; [18] Stability of
linear systems with delays. In Stability Problems. C.I.M.E., June 1974; [19]
Functional differential equations of neutral type. Dynamical Systems-An
International Symposium, 179-194. Academic Press, 1976; [20] Functional
differential equations with infinite delays. J. Math. Anal. Appl. 48 (1974),
276-283; [21] Ordinary Differential Equations, Wiley, 1969; [22] Functional
Differential Equations, Springer-Verlag, 1977; [23] Asymptotic Behavior of
Dissipative Systems, Amer. Math. Soc., 1988; [24] Nonlinear oscillations in
equations with delays. In Nonlinear Oscillations in Biology (Ed. F. C. Hop-
pensteadt), Lecture in Applied Math. 17(979), 157-185. Am. Math. Soc.;
[25] Flows on centre manifolds for scalar functional differential equations.
Proc. Royal Soc. Edinb. lOlA (1985), 193-201; [26] Local flows for func-
tional differential equations. Contemporary Math. 56 (1986), 185-192; [27]
Dynamical systems and stability. J. Math. Anal. Appl. 26 (1969), 39- 69;
[28] Introduction to dynamic bifurcation. In Lect. Notes Math. 1057 (1984),
106-151. Springer-Verlag.
Hale, J. K. and W. Huang [1] Global geometry of the stable regions for two delay
differential equations. J. Math. Anal. Appl.. to appear; [2] Period doubling
in singularly perturbed delay equations. J. Differential Eqns. to appear; [3]
Square and pulse waves in matrix delay differential equations. Dynamic Sys-
tems and Applications 1 (1992), 51-70.
428 Bibliography
Hale, J. K., Infante, E. F. and F.-S. P. Tsen [1] Stability in linear delay equations.
J. Math. Anal. Appl. 105 (1985), 533-555.
Hale, J. K. and A. F. Ivanov [1] On a high order differential delay equation. J.
Math. Anal. Appl., to appear.
Hale, J. K. and A. Ize [1] On the uniform asymptotic stability of functional
differential equations of neutral type. Proc. Amer. Math. Soc. 28 (1971),
100-106.
Hale, J. K. and J. Kato [1] Phase space for retarded equations with infinite delay.
Funk. Ekvacioj 21 (1978), 11-41.
Hale, J. K. and L. A. C. Ladeira [1] Differentiability with respect to delays.
J. Differential Eqns. 92 (1991), 14-26; [2] Differentiability with respect to
delays for a retarded reaction-diffusion equation. Nonlinear Anal., to appear.
Hale, J. K., LaSalle, J. P. and M. Slemrod [1] Theory of a general class of dissi-
pative processes. J. Math. Anal. Appl. 39 (1972), 177-191.
Hale, J. K. and X.-B. Lin [1] Examples of transverse homoclinic orbits in delay
equations. Nonlinear Anal. 10 (1986), 693-709; [2] Symbolic dynamics and
nonlinear semiflows. Annali de Mat. pura appl. (IV) CXLIV (1986), 229-260.
Hale, J. K. and 0. Lopes [1] Fixed-point theorems and dissipative processes. J.
Hale, J. K., Magalhaes, L. and W. Oliva [1] An Introduction to Infinite Dimen-
sional Dynamical Systems-Geometric Theory, Springer-Verlag, 1984.
Hale, J. K. and P. Martinez-Amores [1] Stability in neutral equations. J. Nonl.
Anal. Theory Meth. Appl. 1 (1977), 161-172.
Hale, J. K. and J. Mawhin [1] Coincidence degree and periodic solutions of neutral
Hale, J. K. and K. R. Meyer [1] A class of functional equations of neutral type.
Mem. Amer. Math. Soc., No. 76, 1967.
Hale, J. K. and W. Oliva [1] One-to-oneness for linear retarded functional differ-
ential equations. J. Differential Eqns. 20 (1976), 28-36.
Hale, J. K. and C. Perell6 [1] The neighborhood of a singular point for functional
differential equations. Contributions to Differential Equations 3 (1964), 351-
375.
Hale, J. K. and G. Raugel [1] Lower semicontinuity of attractors of gradient
systems and applications. Ann. Mat. pura appl. (IV) CLIV (1989), 281-326;
[2] Convergence in gradient-like systems and applications. J. Appl. Math.
Phys. (ZAMP) 43 (1992).
Hale, J. K. and K. Rybakowski [1] On a gradient-like integra-differential equation,
Proc. Roy. Soc. Edinburgh 92A (1982), 77-85.
Hale, J. K. and J. Scheurle [1] Smoothness of bounded solutions of nonlinear
evolution equations. J. Differential Eqns. 56 (1985), 142-163.
Hale, J. K. and N. Sternberg [1] Onset of chaos in differential delay equations. J.
Camp. Phys. 77 (1988), 271-287.
Hale, J. K. and S. M. Verduyn Lunel [1] The effect of rapid oscillations in the
dynamics of delay equations. In Mathematical Population Dynamics (Eds.
0. Arino, D. E. Axelrod, and M. Kimmel) (1991), 211-216, Marcel-Dekker;
[2] Averaging in infinite dimensions. J. Integral Equations Appl. 2 (1990),
463-494.
Harband, J. [1] On the asymptotic stability of solutions to an equation of car
following. Tech. Rpt. Math. 13, Univ. Negev, August 1972.
Bibliography 429
Hastings, S. P. [1] Backward existence and uniqueness for retarded functional

Hausrath, A. [1] Stability in the critical case of purely imaginary roots for neutral
functional differential equations. J. Differential Eqns. 13 (1973), 329-357.
Hayes, N. D. [1] Roots of the transcendental equation associated with a certain
differential difference equations. J. London Math. Soc. 25 (1950), 226-232.
Henry, D. [1] Linear autonomous neutral functional differential equations. J. Dif-
ferential Eqns. 15 (1974), 106-128; [2] Small solutions of linear autonomous
functional differential equations. J. Differential Eqns. 8 (1970), 494-501; [3]
The adjoint of a linear functional differential equation and boundary-value
problems. J. Differential Eqns. 9 (1971), 55-66; [4] Linear autonomous neu-
tral functional differential equations. J. Differential Eqns. 15 (1974), 106-
128; [5] Geometric Theory of Semilinear Parabolic Equations, Lect. Notes
Math., Vol. 840, 1981, Springer-Verlag; [6] Variedades invariantes perto dum
ponto fixo. Lecture Notes, Univ. Sao Paulo, Brasil, 1983; [7] Topics in Anal-
ysis. Pub. Mat. UAB 31 (1987), 29-84.
Hetzer, G. [1] Some applications of the coincidence degree fork-set contractions
to functional differential equations of neutral type. Comment. Math. Univ.
Carolinae 16 (1975), 121-138.
Hille, E. and R. Phillips [1] Functional Analysis and Semigroups. Amer. Math.
Soc. Colloq. Publ., vol. 31, 1957.
Hines, W. [1] Ph.D. thesis, Georgia Institute of Technology, 1992.
Hino, Y. [1] Continuous dependence for some functional differential equations.
Tohoku Math. J. 23 (1971), 565-571; [2] Asymptotic behavior of solutions of
some functional differential equations. Tohoku Math. J. 22 (1970), 98-108;
[3] On stability of some functional differential equations. Funkcial. Ekvac.
14 (1971), 47-60.
Hino, Y., Murakami, S. and T. Naito [1] Functional Differential Equations with
Infinite Delay, Lect. Notes Math. 1473, Springer-Verlag, 1991.
Hirsch, M., Pugh, C. and M. Shub [1] Invariant Manifolds, Lect. Notes Math 583,
Springer-Verlag, 1977.
Hopf, E. [1] Abzweigung einer Periodischen Losung eines Differential Systems,
Berichen Math. Phys. Kl. Siich. Akad. Wiss. Leipzig 94 (1942), 1-22.
Hopf, F. A., Kaplan, D. L., Gibbs, H. M. and R. L. Shoemaker [1] Bifurcation to
chaos in optical stability. Phys. Rev. A, 25 (1982), 2172-2182.
Hoppensteadt, F. C. [1] Mathematical theories of population: Demographics, ge-
netics and epidemics. Regional Conf. in Appl. Math. 120 (1975), SIAM,
Philadelphia.
Hoppenstadt, F. and P. Waltman [1] A problem in the theory of epidemics, I
and II, Math. Biosciences 9 (1970), 71-91. MR 44, 7083; Ibid. 12 (1971),
133-145.
Horn, W. A. [1] Some fixed-point theorems for compact mappings and flows on
a Banach space. Trans. Amer. Math. Soc. 149 (1970), 391-404.
Huang, W. [1] Generalization of Liapunov's theorem in a linear delay system. J.
Math. Anal. Appl. 142 (1989), 83-94.
Huang, Y. S. and J. Mallet-Paret [1] Asymptotics of the spectrum for linear peri-
odic differential delay equations. ?reprint 1992; [2] The infinite dimensional
version of the Floquet theorem. ?reprint 1992.
Hughes, D. K. [1] Variational and optimal control problems with delayed argu-
ment. J. Opt. Theory Appl. 2 (1968), 1-14.
430 Bibliography
Ikeda, K. [1] Multiple-valued stationary state and its instability of the transmitted
light in a ring cavity. Opt. Comm. 30 (1979), 257-261.
Ikeda, K., Daido, H. and 0. Akimoto [1] Optical turbulence: chaotic behavior of
transmitted light from a ring cavity. Phys. Rev. Lett. 45 (1980), 709-712.
Ikeda, K., Kondo, K. and 0. Akimoto [1] Successive higher harmonic bifurcations
in systems with delayed feedback. Phys. Rev. Lett. 49 (1982), 1467-1470.
Ikeda, K. and K. Matsumoto [1] High dimensional chaotic behavior in systems
with time delayed feedback. Physica 29D (1987), 223-235.
Imaz, C. and Z. Vorel [1] Generalized ordinary differential equations in Banach
space and applications to functional equations. Bol. Soc. Mat. Mexicana 10
(1966), 47-59.
Infante, E. F. and W. B. Castelan [1] A Liapunov functional for a matrix difference
differential equation. J. Differential Eqns. 29 (1978), 439-451.
Infante, E. F. and M. Slemrod [1] Asymptotic stability criteria for linear systems of
differential difference equations of neutral type and their discrete analogues.
J. Math. Anal. Appl. 38 (1972), 399-415.
Israelson, D. and A. Johnson [1] Application of a theory for circummutations to
geotropic movements. Physiologia Plantorium. 21 (1968), 282-291; [2] Phase-
shift in geotropical oscillations-A theoretical and experimental study. Phys-
iologia Plantorium. 22 (1969), 1226-1237.
Ito, K. and M. Nisio [1] On stationary solutions of a stochastic differential equa-
tion. J. Math. Kyoto Univ. 4 (1964), 1-75.
Ivanov, A. F. and A. N. Sharkovsky [1] Oscillations in singularly perturbed delay
equations. Dynamics Reported (New Series) 1 (1991), 165-224.
Ize, A. [1] Asymptotic stability implies uniform asymptotic stability in periodic
retarded equations. Personal communication; [2] Linear functional differen-
tial equations of neutral type asymptotically autonomous. Ann. Mat. Pura
Appl. {4) 96 (1973), 21-39.
Ize, A. and N. de Molfetta [1] Asymptotically autonomous neutral functional
differential equations with time dependent lag. Dynamical Systems-An In-
ternational Symposium, 127-132, Academic Press, 1976.
Johnson, A. and H. G. Karlsson [1] A feedback model for biological rhythms. I.
Mathematical description and basic properties of the model. J. Theor. Biol.
36 (1972), 153-174.
Jones, G. [1] Hereditary structure in differential equations. Math. Systems Theory,
1 (1967), 236-278; [2] Stability and asymptotic fixed-point theory. Proc. Nat.
Acad. Sci. U.S.A. 53 (1965), 1262-1264; [3] The existence of critical points
in generalized dynamical systems, 7-19. Seminar on Differential Equations
and Dynamical Systems. Lecture Notes in Math., vol. 60, Springer-Verlag,
1968; [4] The existence of periodic solutions off' (x) = -af(x- 1) [1 + f(x)].
J. Math. Anal. Appl. 5 (1962), 435-450; [5] On the nonlinear differential
difference equation f'(x) = -af(x- 1)[1 + f(x)]. J. Math. Anal. Appl. 4
(1962), 440-469; [6] Periodic motions in Banach space and applications to
functional differential equations. Contributions to Differential Equations 3
(1964), 75-106.
Jones, G. and Yorke, A. [1] The existence and nonexistence of critical points in
bounded flows. J. Differential Eqns. 6 (1969), 238-246.
Kaashoek, M.A. and S.M. Verduyn Lunel [1] Characteristic matrices and spectral
properties of evolutionary systems. Trans. Amer. Math. Soc. 334 (1992),
479-517; [2] An integrability condition on the resolvent for hyperbolicity of
Bibliography 431
the semigroup. J. Differential Eqns., to appear.

Kamenskii, G. A. [1] On the inverse operator of the shift operator over trajectories
of equations with retardation. [Russian] Trudy Sem. Teorii Diff. Urav. Otkl.
Argumentom. 9 (1975), 87-92.
Kamenskii, G. A., Norkin, S. B., and L. E. El'sgol'tz [1] Certain directions in the
development of the theory of differential equations with retarded arguments.
[Russian] Trudy Sem. Tear. Diff. Urav. Otkl. Argumentom. 6 (1968), 3-36.
Kaplan, J. and J. Yorke [1] Ordinary differential equations which yield periodic
solutions of differential delay equations. J. Math. Anal. Appl. 48 (1974), 317-
325; [2] On the stability of a periodic solution of a delay differential equation.
SIAM J. Math. Anal. 6 (1975), 268-282.
Kappel, F. [1] Some remarks to the problem of degeneracy for functional differen-
tial equations, 463-472 of Janssens, Mawhin, Rouche, Ed., Equations Differ-
entielles et Fonctionnelles Non Lineaires. Hermann, 1973; [2] Degeneracy of
functional differential equations, 434-448. Proc. Int. Conf. on Diff. Eqns. H.
A. Antosiewicz, Ed. Academic Press, 1975; [3] On degeneracy of functional
differential equations. J. Differential Eqns. 22 (1976), 250-267; [4] Laplace-
transform methods and linear autonomous functional differential equations.
Ber. math.-stat. Sektion Forschungszentrum Graz. No. 64 (1976).
Kappel, F. and H. K. Wimmer [1] An elementary divisor theory for autonomous
linear functional differential equations. J. Differential Eqns. 21 (1976), 134-
147.
Kappel, F. and K. P. Zhang [1] A neutral functional differential equation with
nonatomic difference operator. J. Math. Anal. Appl. 113 (1986), 311-346.
Kato, J. [1] On Liapunov-Razumikhin type theorems for functional differential
equations. Funcial. Ekvac. 16 (1973), 225-239; [2] On the existence of 0-
curves II, Tohoku Math. J. 19 (1967), 126-140; [3] Asymptotic behaviors in
functional differential equations. Tohoku Math. J. (2) 18 (1966), 174-215.
Klein, E. [1] An application of nonlinear retarded differential equations to the
circummutation of plants. M.Sc. thesis, Brown University, June 1972.
Kobyakov, I. I. [1] Conditions for negativity of the Green's function of a two-point
boundary value problem with deviating arguments. [Russian] Differential-
niye Uravenija 8 (1972), 443-452.
Kolmanovskii, V.B. and V.R. Nosov [1] Stability of Functional Differential Equa-
tions. Academic Press, San Diego, 1986.
Kolmanovskii, V.B. and A. Myshkis [1] Applied Theory of Functional Differential
Equations. Kluwer Academic Publishers, Dordrecht, 1992.
Konovalov, Ju. P. [1] The almost-periodic solutions of quasiharmonic systems
with time lag. [Russian] Izv. Vyss, Ucebn. Zaved. Matematika {89) (1969),
62-69.
Kovac, Ju. I. and L. I. Savcenko [1] A boundary-value problem for a nonlinear
system of differential equations with retarded arguments. [Russian] Ukrain.
Mat. Z. 22 (1970), 12-21.
Koval, B. 0. and E. F. Carkov [1] Necessary and sufficient conditions of the
absolute asymptotic stability of linear systems of differential equations with
constant lag. [Russian] Dopovidi Akad. Nauk Ukrain. RSR Ser. A (1972),
506-509, 573.
Krasnoselskii, M. A. [1] The Operator of Translation Along the Trajectories of
Differential Equations. Amer. Math. Soc., Providence, RI, 1968; [2] Positive
Solutions of Operator Equations, Noordhoff, Groningen, 1964.
432 Bibliography
Krasovskii, N. [1] Stability of Motion, Moscow, 1959. Translation, Stanford Uni-

versity Press, 1963; [2] On the stabilization of unstable motions by additional
forces when the feedback loop is incomplete. [Russian] Prikl. Mat. Mek. 27
(1963), 641-663. TPMM 971-1004.
Kuang, Y. and H. L. Smith [1] Slowly oscillating periodic solutions of autonomous
state-dependent delay equations. Nonlinear Anal., to appear; [2] Periodic so-
lutions of differential delay differential equations with threshold-type delays.
In Oscillations and Dynamics in Delay Equations, Contemporary Math. 129
(1992), 153-176.
Kupka, I. [1] Contribution ala Theorie des champs gem3riques. Contributions to
Differential Equations 2 (1963), 457-484. Ibid. 3 (1964), 411-420.
Kuratowski, C. [1] Surles espaces complets. Fund. Math. 15 (1930), 301-309.
Kurzweil, J. [1] Invariant manifolds, I. Comm. Math. Univ. Carolinae 11 (1970),
309-336; [2] Invariant manifolds for flows. Differential Equations and Dy-
namical Systems. Academic Press, 1967, 431-468; [3] Invariant manifolds in
the theory of functional differential equations. Int. Conf. Nonlinear Osc. E.
Berlin, September 1974; [4] Global solutions of functional differential equa-
tions. In Lecture Notes in Math. vol. 144, Springer-Verlag, 1970; [5] On a
system of operator equations. J. Differential Eqns. 11 (1972), 364-375.
Kwapisz, M. [1] On the existence of periodic solutions of functional differen-
tial equations. [Russian] Trudy Sem. Teor. Diff. Urav. Otk. Arg. 7 (1969),
43-54; [2] On quasilinear differential difference equations with quasilinear
conditions. Math. Nach. 43 (1970), 215-222.
Ladyzenskaya, 0. A. [1] A dynamical system generated by the Navier-Stokes
equation. Zap. Nauchn. Sem. Leningrad Otdel. Mat. Inst. Steklov. (LOMI)
27 (1972), 91-115; [2] On the determination of minimal global attractors for
the Navier-Stokes and other partial differential equations. Russian Mathe-
matical Surveys 42:6 (1987), 27-73.
Laksmikantham, V. and S. Leela [1] Differential and Integral Inequalities, vol. 2.
Academic Press, 1969.
Langevin, R., Oliva, W. M. and J. C. F. de Oliveira [1] Retarded functional dif-
ferential equations with white noise perturbations. Univ. Sao Paulo (1989).
Preprint.
LaSalle, J. P. [1] A study of synchronous asymptotic stability. Annals of Math-
ematics, 65 (1957), 571-581; [2] An invariance principle in the theory of
stability. International Symposium on Differential Equations and Dynamical
Systems. Academic Press, 1967, 277.
Lasota, A. [1] Ergodic problems in biology. Asterique 50 (1977), 239-250.
Leitman, M. J. and V. J. Mizel [1] On linear hereditary laws. Arch. Rat. Mech.
Anal. 38 (1970), 45-68; [2] On fading memory spaces and hereditary integral
equations. Arch. Rat. Mech. Anal. 55 (1974), 18-51; [3] Asymptotic stability
and the periodic solutions of x(t) + J~ a(t- s)g(s, x(s)) ds = f(t). J. Math.
Anal. Appl. 66 (1978), 606-625.
Leung, A. [1] Periodic solutions for a prey-predator differential delay equation.
Preprint.
Levin, J. J. and J. Nohel [1] On a nonlinear delay equation. J. Math. Anal. Appl.
8 (1964), 31-44.
Levinger, B. W. [1] A folk theorem in functional differential equations. J. Differ-
ential Eqns. 4 (1968), 612-619.
Levinson, N. [1] Transformation theory of nonlinear differential equations of the
Bibliography 433
second order. Annals of Math. 45 (1944), 724-737; [2] A second-order differ-

ential equation with singular solutions. Ann. Math. 50 (1949), 126-153.
Lillo, J. C. [1 J Backward continuation of retarded functional differential equa-
tions. J. Differential Eqns. 17 (1975), 349-360; [2] Oscillatory solutions of
the equation y'(x) = m(x)y(x- n(x)). J. Differential Eqns. 6 (1969), 1-36;
[3] Periodic differential difference equations. J. Math. Anal. Appl. 15 (1966),
434-441; [4] The Green's function for periodic differential difference equa-
tions. J. Differential Eqns. 4 (1968), 373-384; The Green's function for nth-
order periodic differential difference equations. Math. Sys. Theory 5 (1971),
13--19; [5] Periodic perturbations of the nth-order differential difference equa-
tions. Amer. J. Math. 94 (1972), 651-675.
Lima, P. [1] Hopf bifurcation in equations with infinite delays. Ph.D. thesis, Brown
University, Providence, R. I., 1977.
Littlewood, J. E. [1] On nonlinear differential equations of the second order: IV.
The general equation y + kf(y)y + 9(y) = bkp(¢), ¢ = t +a. Acta Mathe-
matica, 98 (1957).
Lizano, M. [1] The sunflower equation. ?reprint.
London, W. P. and J. A. Yorke [1] Recurrent epidemics of measles, chickenpox,
and mumps I: Seasonal variation in contact rates. Amer. J. Epid. 98 (1973),
453-468. II: Systematic differences in rates and stochastic effects. Amer. J.
Epid. 98 (1973), 469-482.
Lopes, 0. [1] Stability and forced oscillations in nonlinear distributed networks.
?reprint; [2] Existencia e estabilidade de oscila<;oes forcadas de equa<;6es
diferenciais funcionais. Dep. Mat. Inst. Cienc. Mat. Sao Carlos, Brasil, 1975.
Tese para o Concurso de Livre Docencia; [3] Periodic solutions of perturbed
neutral differential equations. J. Differential Eqns. 15 (1974), 70-76; [4] Pos-
itive eigenvectors and periodicity theorem for autonomous retarded equa-
tions. Trabalhos de Matematica, no. 85, Universidade de Brasilia, July 1974;
[5] Asymptotic fixed-point theorems and forced oscillations in neutral equa-
tions. Ph.D. Thesis, Brown University, Providence, RI, June 1973; [6] Forced
oscillations in nonlinear neutral differential equations. SIAM J. Appl. Math.
29 (1975), 196-207.
Luckhaus, S. [1] Global boundedness for a delay differential equation. Trans. AMS
294 (1986), 767-774.
MacCamy, R. C. [1 J Exponential stability for a class of functional differential
equations. Arch. Rat. Mech. Anal. 40 (1971), 120-138; [2] Nonlinear Volterra
equations on a Hilbert space. J. Differential Eqns. 16 (1974), 373-393; [3]
Stability theorems for a class of functional differential equations. SIAM J.
Appl. Math. 30 (1976), 557-576.
Mackey, M. C. and U. an der Heiden [1] Dynamical diseases and bifurcations:
Understanding functional disorders in physiological systems. Funkt. Biol.
Med. 156 (1982), 156-164.
Mackey, M. C. and L. Glass [1] Oscillation and chaos in physiological control
systems. Science 197 (1977), 287-289.
Magalhaes, L. T. [1] Exponential estimates for singularly perturbed linear func-
tional differential equations. J. Math. Anal. Appl. 103 (1984), 443-460; [2] In-
variant manifolds for singularly perturbed linear functional differential equa-
tions. J. Differential Eqns. 54 (1984), 310-345; [3] Convergence and bound-
ary layers in singularly perturbed linear functional differential equations. J.
Differential Eqns. 54 (1984), 295-310; [4] The asymptotics of solutions of
singularly perturbed functional differential equations: Distributed and con-
434 Bibliography
centrated delays are different. J. Math. Anal. Appl. 105 (1985), 250-257; [5]
Invariant manifolds for functional differential equations close to ordinary dif-
ferential equations. Funk. Ekv. 28 (1985), 57-82; [6] Persistence and smooth-
ness of hyperbolic invariant manifolds for functional differential equations.
SIAM J. Math. Anal. 18 (1987), 670-693; [7] Shadow ordinary differential
equations for retarded functional differential equations with small delays.
?reprint; [8] The spectra of linearization of dissipative infinite dimensional
dynamical systems. ?reprint.
Mahaffy, J. M. [1] Geometry of the stability region for a differential equation with
two delays. ?reprint 1990.
Makay, G. [1] On the asymptotic stability in terms of two measures for func-
tional differential equations. Nonlinear Anal. 16 (1991), 721-727; [2] On the
asymptotic stability of the solutions of functional differential equations with
infinite delay. J. Differential Eqns., to appear.
Mallet-Paret, J. [1] Generic periodic solutions offunctional differential equations.
J. Differential Eqns. 25 (1977), 163-183; [2] Generic and qualitative proper-
ties of retarded functional differential equations. Meeting Func. Diff. Eqns.
Braz. Math. Soc. Sao Carlos, July 1975; [3] Generic properties of retarded
functional differential equations. Bull. Amer. Math. Soc. 81 (1975), 750-752;
[4] Negatively invariant sets of compact maps and an extension of a theorem
of Cartwright. J. Differential Eqns. 22 (1976), 351-348; [5] Morse decom-
positions for delay-differential equations. J. Differential Eqns. 72 (1988),
270-315.
Mallet-Paret, J. and R. Nussbaum [1] Global continuation and asymptotic be-
havior for periodic solutions of a differential-delay equation. Ann. Mat. pura
appl. (IV) CXLV (1986), 33-128; [2] Global continuation and complicated
trajectories for periodic solutions of a differential delay equation. In Proc.
Symp. Pure Math. 45 (1986), 155-167. Am. Math. Soc.; [3] A differential-
delay equation arising in optics and physiology. SIAM J. Math. Anal. 20
(1989), 249-292; [4] A bifurcation gap for a singularly perturbed delay equa-
tion. In Chaotic Dynamics and Fractals (Eds. M. F. Barnsley and S. G.
Demko (1986), 263-286. Academic Press; [5] Boundary layer phenomena for
differential-delay equations with state-dependent time lags, I. Arch. Rat.
Mech. Anal. 120 (1992), 99-146; [6] Ibid, II. ?reprint 1991.
Malta, C. P. and C. G. Ragazzo [1] Bifurcation structure of scalar differential
delayed equations. Int. J. Bif. Chaos 1 (1991), 657-665.
Mane, R. [1] On the dimension of the compact invariant sets of certain nonlinear
maps. In Lecture Notes in Math. 898 (1981), 230-242, Springer-Verlag.
Manitiu~, A. [1] Completeness and F-completeness of eigenfunctions associated
with retarded functional differential equations, J. Differential Eqns. 35
(1980)' 1-29.
Mansurov, K. [1] Stability of linear systems with delay. [Russian] Studies in Dif-
ferential Equations and Their Applications, 190-199. lzdat. "Nauka," Alma-
Alta, 1965.
Markus, M. [1] Lectures in Differentiable Dynamics. Regional Conference Series
in Math., No. 3 (1971); [2] Generic properites of differential equations. Int.
Symp. on Nonlinear Differential Equations and Nonlinear Mechanics. Aca-
demic Press, 1963, p. 22.
Marriott, C. Vallee, R. and C. Delisle [1] Analysis of a first order delay differential
equation containing two delays. Phys. Rev. A, 40 (6) (1989), 3420-3428.
Bibliography 435
Marsden, J. E. and M. F. MacCracken (1] The Hopf Bifurcation and Its Applica-
tions. Appl. Math. Sci., Vol. 19, Springer-Verlag.
Martin, R. H. and H. L. Smith [1] Abstract functional differential equations and
reaction diffusion systems. Trans. Amer. Math. Soc. to appear; (2] Reaction
diffusion systems with time delays: Monotonicity, invariance, comparison and
convergence. J. Reine Angew. Math. XXX (1990).
Martynyuk, D. I. (1] Lectures on the Theory of Stability of Solutions of Systems
with Retardations. (Russian] Inst. Mat. Akad. Nauk. UK. SSR, Kiev, 1971.
Maslowski, B., Seidler, J. and I. Vrkoc (1] An averaging principle for stochastic
evolution equations II. Math. Bohemica 116 (1991), 191-224; (2] Integral
continuity and stability for stochastic hyperbolic equations. Cesk. Akad. Ved,
Mat. Ostav. Preprint 1991.
Matano, H. (1] Nonincrease of the lap number of a solution for a one dimensional
semilinear parabolic equation. J. Fac. Sci. Univ. Tokyo Sect. IA Math. 29
(1982), 401-441.
Mawhin, J. [1] Periodic solutions of nonlinear functional differential equations. J.
Differential Eqns. 10 (1971), 240-261; (2] Nonlinear perturbations of Fred-
holm mappings in normed spaces and applications to differential equations.
Lecture notes, Univ. de Brasilia, May 1974; (3] Topology and nonlinear
boundary-value problems, vol. I, 51-82. Dynamical Systems-An Interna-
tional Symposium. Academic Press, 1976.
McCord, C. and K. Mischaikow (1] On the global dynamics of attractors for scalar
delay equations. CDSNS Ga. Tech Preprint 1992.
Medzitov, M. (1] Two-point boundary-value problem for a nonlinear differential
equation with retarded argument. (Russian] Trudy Sem. Tear. Diff. Urav.
Otk. Arg. 'l (1969), 178-182.
Melvin, W. R. (1] A class of neutral functional differential equations. J. Differen-
tial Eqns. 12 (1972), 524-534; (2] Topologies for neutral functional differential
equations. J. Differential Eqns. 13 (1973), 24-32; (3] Some extensions of the
Karsnoselskii fixed-point theorems. J. Differential Eqns. 11 (1972), 335-348;
(4] Stability properties of functional differential equations. J. Math. Anal.
Appl. 48 (1974), 749-763.
Memory, M. (1] Stable and unstable manifolds for partial functional differen-
tial equations. Nonlinear Anal., to appear; (2] Invariant manifolds for par-
tial functional differential equations. In Mathematical Population Dynamics
(Eds. 0. Arino, D. E. Axelrod, and M. Kimmel) (1991), 223-232, Marcel-
Dekker; (3] Bifurcation and asymptotic behavior of solutions of a delay dif-
ferential equation with diffusion. SIAM J. Math. Anal. 20 (1989), 533-546.
Mikolajska, Z. (1] Une remarque sur les solutions bornees d'une equation differo-
differentielle nonlineaire. Ann. Polan. Math. 15 (1964), 23-32.
Miller, R. K. (1] Nonlinear Volterra Integral Equations. Benjamin, 1971; (2] Linear
Volterra integrodifferential equations as semigroups. Funkcialaj Ekvaciaj, 11
(1974), 39-55.
Miller, R. K. and G. Sell [1] Topological dynamics and its relation to integral
equations and nonautonomous systems. In Dynamical Systems (Eds. Cesari,
Hale, Lasalle), Vol. 1, Academic Press, 1976, 223-249.
Minorsky, N. [1] Self-excited oscillations in dynamical systems possessing retarded
actions. J. Appl. Mech. 9 (1942), 65-71; [2] Nonlinear Oscillations, D. Van
Nostrand Company, Inc., Princeton, 1962.
Minsk, A. F. [1] The second method of Liapunov for equations of neutral type.
[Russian] Trudy Sem. Tear. Diff. Urav. Otkl. Arg. 6 (1968), 78-109; [2] Ab-
436 Bibliography
solute stability of nonlinear systems of automatic regulators of neutral type.

[Russian] Trudy Sem. Teor. Diff. Urav. Otkl. Argum. 7 (1969), 92-106.
Mishkis, A. D.[1] Lineare Differentialgleichungen mit nacheilenden Argumentom,
Deutscher Verlag. Wiss. Berlin, 1955. Translation of the 1951 Russian edi-
tion Linear Differential Equations with Retarded Aryuments. [Russian] Iz-
dat. "Nauka" Moscow, 1972. MR 50, 5135; [2] General theory of differential
equations with a retarded argument. Amer. Math. Soc. Transl. no. 55 (1951).
Uspehi Mat. Nauk (N.S.) 4 (33) (1949), 99-141.
Mishkis, A. D. and L. E. El'sgol'tz [1] The state and problems of the theory
of differential equations with perturbed arguments. [Russian] Uspehi Mat.
Nauk 22 (134) (1967), 21-57.
Mitropolskii, Ju. A. and V. I. Fodcuk [1] Asymptotic methods of nonlinear me-
chanics applied to nonlinear differential equations with retarded arguments.
[Russian] Ukrain. Mat. Z. 18 (1966), 65-84.
Mitropolskii, Yu. and D. G. Korenevskii [1] The investigation of nonlinear oscilla-
tions in systems with distributed parameters and time lag. [Russian] Mathe-
matical Physics, No. 4 (1968), 93-145. Naukova Dumka, Kiev, 1968; [2] An
application of asymptotic methods to systems with distributed parameters
and lag. [Russian] Prikladna. Meh. 5 (1969), vyp. 4, 1969.
Mizel, V. J. and V. Trutzer [1] Stochastic hereditary equations: Existence and
asymptotic stability. J. Integral Equations 7 (1984), 1-72.
Mohammed, S. E. A. [1] Stochastic Functional Differential Equations. Pitman,
Boston, 1984; [2] The Lyapunov spectrum and stable manifolds for stochastic
linear delay equations. Stochastics 29 (1990), 89-131.
Mohammed, S. E. A. and M. K. R. Scheutzow [1] Lyapunov exponents and sta-
tionary solutions for affine stochastic delay equations. Stochastics 29 (1990),
259-283.
Mohammed, S. E. A., Scheutzow, M. K. R. and H. v. Weizsacker [1] Hyperbolic
state space decomposition for a linear stochastic delay equation. SIAM J.
Control and Opt. 24 (1986), 543-551.
Morita, Y. [1] Destabilization of periodic solutions arising in delay diffusion sys-
tems in several space dimensions. Japan J. Appl. Math. 1 (1984), 39-65.
Moreno, C.J. [1], The zeros of exponential polynomials. I. Comp. Math. 26 (1973),
69-78.
Mosjagin, V. V. [1] A boundary-value problem for a differential equation with
lagging argument in a Banach space. [Russian] Leningrad Gos. Ped. Inst.
Ucen. Zap. 387 (1968), 198-206.
Myjak, J. W. [1] A boundary-value problem for nonlinear differential equations
with a retarded argument. Ann. Polan. Math. 27 (1973), 133-142.
Naito, T. [1] Integral manifolds for linear functional differential equations on some
Banach space. Funkial. Ekvac. 13 (1970), 199-213; [2] On autonomous lin-
ear functional differential equations with infinite retardations. J. Differential
Eqns. 21 (1976), 297-315; [3] Adjoint equations of autonomous linear func-
tional differential equations with infinite retardations. Tohoku Math. J. 28
(1976), 135-143.
Neustadt, L. W. [1] On the solutions of certain integral-like operator equations.
Existence, uniqueness, and dependence theorems. Arch. Rat. Mech. Anal. 38
(1970)' 131-160.
Neves, A.F., H. Ribeiro and 0. Lopez [1] On the spectrum of evolution operators
generated by hyperbolic systems. J. Funct. Anal. 67 (1986), 320-344.
Bibliography 437
Nickel, K. [1] Gestaltaussagen iiber Losungen papabolischer Differentialgleichun-

gen. J. Reine Angew. Math. 211 (1962), 78-94.
Nitecki, Z. [1] Differentiable Dynamics. MIT Press, 1972.
Noonburg, V. W. [1] Bounded solutions of a nonlinear system of differential delay
Norkin, S. B. [1] Differential Equations of the Second Order with Retarded Argu-
ments. [Russian] Moscow; 1965. English translation by K. Schmitt and L. J.
Grimm. Trans. Math. Mono. vol. 31. Amer. Math. Soc. 1972.
Nosov, V. R. [1] Linear boundary-value problems with small lags. Diff. Urav.
3 (1967), 1025-1028; [2] Periodic solutions of systems of linear equations of
general form with retarded arguments. [Russian] Differentialniye Uravnenija
7 (1971), 639-650; [3] Periodic solutions of quasilinear functional differen-
tial equations. [Russian] Izv. Vyss. Ucebn. Zved. Matematica 1973 (132),
55-62; [4] Periodic solutions of quasilinear functional differential equations.
[Russian] Izv. Vyss. Ucebn. Zaved. Mat. 1975 (132), 55-62.
Nussbaum, R. [1] Periodic solutions of analytic functional differential equations
are analytic. Mich. Math. J. 20 (1973), 249-255; [2] Some asymptotic fixed-
point theorems. Trans. Amer. Math. Soc. 171 (1972), 349-375; [3) Periodic
solutions of some nonlinear autonomous functional differential equations.
Ann. Math. Pura Appl. 10 (1974), 263-306; [4] Periodic solutions of some
nonlinear autonomous functional differential equations, II. J. Differential
Eqns. 14 (1973), 368-394; [5] A global bifurcation theorem with applications
to functional differential equations. J. Functional Anal. 19 (1975), 319-339;
[6] The range of periods of periodic solutions of x'(t) = -a.f(x(t- 1)). J.
Math. Anal. Appl. 58 (1977), 280-292; [7] Global bifurcation of periodic so-
lutions of some autonomous functional differential equations. J. Math. Anal.
Appl. 55 (1976), 699-725; [8] Existence and uniqueness theorems for some
functional differential equations of neutral type. J. Differential Eqns. 11
(1972), 607-623; [9] A Hopf global bifurcation theorem for retarded func-
tional differential equations. Trans. Amer. Math. Soc. 238 (1978), 139-163;
[10] Periodic solutions of nonlinear autonomous functional differential equa-
tions. Lect. Notes Math. 730 (1979), 283-325; [11] Differential delay equa-
tions with two time delays. Mem. Am. Math. Soc. 16, 1978.
Oliva, W. M. [1] Functional differential equations on compact manifolds and an
approximation theorem. J. Differential Eqns. 5 (1969), 483-496; [2] Func-
tional differential equations-Generic theory. Proc. Int. Symp. Diff. Eqn.
Dyn. Syst., Brown University, August 1974, Dynamical Systems-An Inter-
national Symposium, vol. I, 195-209. Academic Press, 1976; [3] Some open
questions in the geometric theory of retarded functional differential equa-
tions. Proc. 10th Braz. Colloq. Math., Poc;os de Caldas, July 1975; [4] The
behavior at infinity and the set of global solutions of retarded functional
differential equations. Meeting on Func. Diff. Eqns. Braz. Math. Soc., Sao
Carlos, July 1975; [5] Stability of Morse-Smale maps. ?reprint, 1982.
Oliva, W. M., Kuhl, N. M. and L. T. Magelhaes [1] Diffeomorphism of lRn with
oscillatory Jacobian. Personal communication.
de Oliveira, J. C. F. [1] Quasipotencial relativo a un poc;o. Univ. Sao Paulo (1989).
?reprint.
Onuchic, N. [1] On a criterion of instability for differential equations with time
delay, 339-342. Periodic Orbits, Stability and Resonance, Ed. by G. E. 0.
Giacaglia, Reidel Publ., Dordrecht, 1970; [2] On the asymptotic behavior of
the solutions of functional differential equations. In Differential Equations
and Dynamical Systems, 223-233, Academic Press, 1967; [3] Asymptotic be-
438 Bibliography
havior of a perturbed linear differential equation with time lag on a product

space. Ann. Mat. Pura Appl. (4) 86 (1970), 115-124.
Pavel, N. [1] On the boundedness of the motions of a periodic process. Atti. Accad.
Naz. Lincei Rend. Cl. Sci. Fis. Mat. Natur. {8} 54 (1973), 25-33.
Pazy, A. (1] Semigroups of Linear Operators and Applications to Partial Differ-
ential Equations. Springer-Verlag, 1983.
Pecelli, G. (1] Dichotomies for linear functional differential equations. J. Differ-
ential Eqns. 9 (1971), 555-579.
Peixoto, M. (1] Qualitative theory of differential equations and structural stability.
Differential Equations and Dynamical Systems. Academic Press, 1967, 469-
480; (2] On an approximation theorem of Kupka-Smale. J. Differential Eqns.
3 (1967), 214-227.
Perell6, C. (1] Cualidades del retrato fase que se preservanal introducir un retraso
pequefio en el tiempo. Acta Mexicana Ci. Teen. 3 (1969), 12-30; [2] Periodic
solutions of differential equations with time lag containing a small parameter.
J. Differential Eqns. 4 (1968), 160-175; (3] A note on periodic solution of
nonlinear differential equations with time lag. In Differential Equations and
Dynamical Systems; 185-187, Academic Press, 1967.
Perez, J. F., Malta, C. P. and F. A. B. Coutinho (1] Qualitative analysis of
oscillations in isolated populations of flies. J. Theor. Biol. 71 (1978), 505-
514.
Pesin, Ya. B. (1] On the behavior of a strongly nonlinear differential equation
with retarded argument. Diff. Urav. 10 (1974), 1025-1036.
Pitt, H. R. (1] A theorem on absolutely convergent Fourier series. J. Math. Phys.
16 (1937), 191-195.
Pliss, V. A. (1] Nonlocal Problems of the Theory of Nonlinear Oscillations, Aca-
demic Press, 1966. (Translation of 1964 Russian edition.)
Pontryagin, L. S. (1] On the zeros of some elementary transcendental functions.
(Russian] Izv. Akad. Nauk SSSR, Ser. Mat. 6 (1942), 115-134. English trans-
lation in Amer. Math. Soc. Transl. {2} 1 (1955), 95-110.
Popov. V. M. (1] Pointwise degeneracy of linear, time invariant, delay differen-
tial equations, J. Differential Eqns. 11 (1972), 541-561; (2] Delay-feedback,
time-optimal, linear time-invariant control systems. Ordinary Differential
Equations, 1971 NRL-MRC Conference, Academic Press, 1972, 545-552.
Prokopev, V. P. and S. N. Shimanov (1] On stability in the critical case of two zero
roots for systems with lags. (Russian] Differentialniye Uravnenija 2 (1966),
453-462.
Ragazzo, C. G. and C. P. Malta (1] Mode selection on a differential equation with
two delays arising in optics. ?reprint 1990.
Razumikhin, B. S. (1] On the stability of systems with a delay. (Russian] Prikl.
Mat. Meh. 20 (1956), 50Q-512; [2] Application ofLiapunov's method to prob-
lems in the stability of systems with a delay. (Russian] Automat. i Telemeh.
21 (1960), 74Q-749.
Razvan, V. (1] Stabilitatea Absoluta a Sistemelor Automate cu Intirziere. Ed.
Acad. Rep. Soc. Romania, Bucharest, 1975; (2] Absolute stability of a class
of control processes described by functional differential equations of neu-
tral type. Equations Differentielles et Fonctionelles Nonlineaires. Hermann,
1973.
Reissig, R., Sansone, G., and R. Conti (1] Nichtlineare Differential Gleichungen
Hoherer Ordnung. Cremonese, 1969.
Bibliography 439
Repin, Yu. M. [1] Quadratic Liapunov functionals for systems with delay. [Rus-
sian] Prikl. Mat. Meh. 29 (1965), 564-566; TPMM 669-67; [2] On conditions
for the stability of systems of differential equations for arbitrary delays.
Uchen. Zap. Ural. 23 (1960), 31-34.
Rubanik, V. P. [1] Oscillations of Quasilinear Systems with Retardation. [Russian]
Nauk, Moscow, 1969.
Ruiz-Clayessen, J. [1] Effect of delays on functional differential equations. J. Dif-
ferential Eqns. 20 (1976), 404-440.
Rybakowski, K. P. [1] On the homotopy index for infinite-dimensional semifiows,
Trans. Amer. Math. Soc. 269 (1982), 351-382; [2] An introduction to homo-
topy index theory in noncompact spaces, Appendix in Hale, Magalhaes, and
Oliva [1]; [3] The Homotopy Index Theory on Metric Spaces with Applica-
tions to Partial Differential Equations, Universitext, Springer-Verlag, 1987;
[4] Realization of arbitrary vector fields on center manifolds for parabolic
Dirichlet BVPs. J. Differential Eqns. to appear; [5] Realization of arbitrary
vector fields on center manifolds for functional differential equations. J. Dif-
ferential Eqns., to appear.
Sabbagh, L. D. [1] Variational problems with lags. J. Optimization Theory Appl.
3 (1969), 34-51.
Sacker, R. J. and G. R. Sell [1] Existence of dichotomies and invariant splittings
for linear differential equations. J. Differential Eqns. 15 (1974), 429-458.
Sadovskii, B. N. [1] Limit compact and condensing operators. [Russian] Uspehi
Math. Nauk 27 (163) (1972), 81-146. Russian Math. Surveys, 85-146.
Saupe, D. [1] Global bifurcation of periodic solutions of some autonomous differ-
ential delay equations. In Forschungsschwerpunkt Dynamische Systeme Rep.
No. 71, Univ. Bremen, Bremen, July, 1982; [2] Accelerated PL-Continuation
methods and periodic solutions of parameterized differential delay equations,
Ph.D. dissertation, Univ. Bremen, Bremen, 1982 (German).
Schaffer, J. J. [1] Linear differential equations with delays: Admissibility and
conditional exponential stability, II. J. Differential Eqns. 10 (1971), 471-
484.
Scheutzow, M. K. R. [1] Qualitative behavior of stachastic delay equations with a
bounded memory. Stochastics 12 (1986), 41-80; [2] Stationary and periodic
stochastic differential systems. Habilitationsschrift, Univ. Kaiserslautern,
1988.
Schumacher, K. [1] Existence and continuous dependence for differential equations
with unbounded delay. Arch Rat. Mech. Anal. 67 (1978), 315-335; [2] Dy-
namical systems with memory on history spaces with monotonic seminorms.
J. Differential Eqns. 34 (1979), 44Q-463.
Seidler, J. and I. Vrkoc [1] An averaging principle for stochastic evolution equa-
tions. Casopis pest. mat. 111 (1990), 240-263.
Seifert, G. [1] Global asymptotic stability for functional differential equations.
Personal communication; [2] Positive invariant closed sets for systems of de-
lay differential equations. J. Differential Eqns. 22 (1976), 292-304; (3] Pos-
itively invariant closed sets for systems of delay differential equations. J.
Differential Eqns. 22 (1976) 292-304.
Sell, G. [1] Lectures on Topological Dynamics and Differential Equations, Van
Nostrand-Reinhold, Princeton, N. J., 1971; [2] Linear Differential Systems,
Lecture Notes, Univ. Minnesota, 1975.
440 Bibliography
Sentebova, E. Yu. [1] Method of quasilinearization for ordinary differential

equations with retarded arguments. [Russian] Differentialniye Uravnenija
8 (1972), 2260-2263.
Shimanov, S. N. [1] On the instability of the motion of systems with retardations.
[Russian] Prikl. Mat. Meh. 24 (1960), 55-63; TPMM 7Q-81; [2] On stability
in the critical case of a zero root with time lag. [Russian] Prikl. Mat. Meh.
23 (1959), 836-844; [3] On the theory of linear differential equations with
retardations. [Russian] Differentialniye Uravnenija 1 (1965), 102-116; [4] On
the theory of linear differential equations with periodic coefficients and time
lag. [Russian] Prikl. Math. Meh. 27(1963), 450-458; TPMM 674-687; [5] On
the vibration theory of quasilinear systems with time lag. [Russian] Prikl.
Mat. Meh. 23 (1959), 836-844.
Shin, J. S. [1] An existence theorem of functional differential equations with in-
finite delay in a Banach space. Funk. Ekvacioj 30 (1987), 225-236; [2] Exis-
tence of solutions and Kamke's theorem for functional differential equations
in Banach spaces. J. Differential Eqns. 81 (1989), 294-312.
Sigueira Marconato, S.A. and C.E. Avellar [1] Difference equations with delays
depending on time. Applicable Analysis, to appear.
Silkowskii, R. A. [1] Star-shaped regions of stability in hereditary systems. Ph.D.
thesis, Brown University, Providence, RI, June 1976.
Slater, M. and H. S. Wilf [1] A class of linear differential difference equations.
Pac. J. Math. 10 (1960), 1419-1427.
Slemrod, M. [1] A hereditary partial differential equation with application in
the theory of simple fluids. ?reprint; [2] Nonexistence of oscillations in a
nonlinear distributed network. J. Math. Anal. Appl. 36 (1971), 22-40.
Smale, S. [1] Differentiable dynamical systems. Bull. Amer. Math. Soc. 73 (1967),
747-817; [2] Stable manifolds for differential equations and diffeomorphisms.
Ann. Scuola Nomale Sup. Pisa 18 (1963), 97-116.
Smith, H. L. [1] On periodic solutions of delay integral equations modeling epi-
demics and population growth. Ph.D. thesis, Univ. of Iowa, May 1976; [2]
Structures polulation models, threshold-type delay equations and functional
differential equations. In Delay and Differential Equations, World Scientific,
1992, 57-64; [3] Monotone semifiows generated by functional differential
Smith, H. L. and H. R. Thieme [1] Quasi convergence and stability for strongly
order preserving semifl.ows. SIAM J. Math. Anal. 21 (1990), 673-692; [2]
Convergence for strongly order preserving semifl.ows. SIAM J. Math. Anal.
22 (1991); [3] Strongly order preserving semifl.ows generated by functional
Solodovnikov, V. V. [1] Techniceskaya Kibernatika. [Russian] vol. 2, Chap. XI,
Moscow, Masinostroenie, 1967.
Somolinos, A. [1] Stability ofLurie-type functional equations. J. Differential Eqns.
26 (1977), 191-199.
Staffans, O.J. [1] A neutral FDE with stableD-operator is retarded. J. Differential
Eqns. 49 (1983), 208-217; [2] Semigroups generated by a neutral functional
differential equation. SIAM J. Math. Anal. 17 (1986), 46-57.
Starik, L. K. [1] Coupled quasilinear oscillating systems with a source of energy
involving retarded connections. [Russian] Trudy Sem. Teorii Diff. Urav. Otk.
Arg. 3 (1965), 119-132.
Bibliography 441
Stech, H. W. [1] The Hopf bifurcation: stability result and application. J. Math.
Anal. Appl. 11 (1979), 525-546; [2] Nongeneric Hopf bifurcations in func-
tional differential equations. SIAM J. Appl. Math. 16 (1985), 1134-1151.
Stephan, B. H. [1] On the existence of periodic solutions of z'(t) = -az(t- r +
~-tk(t, z(t))) +F(t). J. Differential Eqns. 6 (1969), 408-419; [2] Periodic solu-
tions of differential equations with almost constant time lag. J. Differential
Eqns. 8 (1970), 554-563.
Sternberg, N. [1] One-to-oneness of the solution map in retarded functional dif-
ferential equations, J. Differential Eqns. 85 (1990), 201-213; [2] A Bartman-
Grohman theorem for a class of retarded functional differential equations J.
Math. Anal. Appl. submitted; [3] A Hartman-Grobman theorem for maps.
Partial and Functional Differential Equations, Longman, 1992.
Stokes, A. P. [1] On the stability of integral manifolds of functional differential
equations. J. Differential Eqns. 9 (1971), 405-49; [2] A Floquet theory for
functional differential equations. Proc. Nat. Acad. of Sci. U.S.A. 48 (1962),
1330-1334; [3] On the approximation of nonlinear oscillations. J. Differen-
tial Eqns. 12 (1972), 535-558; [4] On the stability of a limit cycle of an
autonomous functional differential equation. Contributions to Differential
Equations 3 (1964), 121-139; [5] Some implications of orbital stability in
Banach spaces. SIAM J. Appl. Math. 11 (1969), 1317-1325; [6] Local coordi-
nates around a limit cycle of functional differential equations. J. Differential
Eqns. 24 (1977), 153-172.
Strygin, V. V. [1] The bifurcation of quasistationary periodic solutions of differ-
ential difference equations. [Russian] Differentialniye Uravnenija 10 (1974),
1332-1334.
Taboas, P. [1] Periodic solutions of a planar delay equation. Proc. Royal Soc.
Edinburgh or New Directions in Differential Equations and Dynamical Sys-
tems, 573-589, Royal Soc. Edinburgh, 1991.
Tang, B. and Y. Kuang [1] Asymptotic behavior for a class of delayed nonau-
tonomous Lotka Volterra type equations. Preprint 1991; [2] Existence and
uniqueness of periodic solutions of periodic functional differential systems.
Preprint 1991.
Taylor, A. F. [1] Introduction to Functional Analysis. John Wiley and Sons, 4th
ed., 1964.
Temam, R. [1] Infinite Dimensional Dynamical Systems, Springer-Verlag, 1988.
Thieme, H. R. [1] Semiflows generated by Lipschitz perturbations of non-densely
defined operators. Preprint 1990.
Travis, C. C. and G. F. Webb [1] Partial differential equations with deviating ar-
guments in the time variable. Preprint; [2] Existence and stability for partial
differential equations. Trans. Amer. Math. Soc. 200 (1974), 395-418.
Tychonov, A. [1] Surles equations fonctionelles de Volterra et leurs applications
a certains problemes de la physique mathematique. Bull. de l'Univ. d'Etat
de Moscou, Ser. Intemat. Sect. A, 1 (1938), 1-25.
Vallee, R., Dubois, P. Cote, M. and C. Delisle [1] Phys. Rev A36 (1987), 1327.
Vallee, R. and C. Marriott [1] Analysis of an Nth-order nonlinear differential
delay equation. Phys. Rev. A39 (1989), 197-205.
Ventsel, A. D. and M. I. Freidlin [1] On small random perturbations of dynamical
systems. Russian Math. Surveys 25 (1970), 1-55.
Verduyn Lunel, S. M. [1] A sharp version of Henry's theorem on small solu-
tions. J. Differential Eqns. 62 (1986), 266-274; [2] Exponential type calculus
442 Bibliography
for linear delay equations. Centre for Mathematics and Computer Science,
Tract No. 57, Amsterdam, 1989; (3] Series expansions and small solutions
for Volterra equations of convolution type. J. Differential Eqns. 85 (1990),
17-53; [4] The closure of the generalized eigenspace of a class of infinitesi-
mal generators. Proc. Roy. Soc. Edinburgh Sect. A 117 (1991), 171-192; (5]
Small solutions and completeness for linear functional differential equations.
In Oscillations and Dynamics in Delay Equations, (Ed. J.R. Graef and J.K.
Hale), Contemporary Mathematics 129, Amer. Math. Soc. 1992; [6] About
completeness for a class of unbounded operators, Report Ga Tech, 1993; [7]
Series expansions for functional differential equations. In preparation 1993.
Volterra, V. [1] Sur la theorie mathematique des phenomenes hen§ditaires. J.
Math. Pures Appl. 7 (1928), 249-298; [2] Theorie Mathematique de la Lutte
pour la Vie, Gauthier-Villars, Paris, 1931; [3] Sulle equazioni integrodifferen-
ziali della teorie dell'elasticita, Atti Reale Accad. Lincei 18 (1909), 295.
Walther, H. 0. (1] Existence of a nonconstant periodic solution of a nonlinear
nonautonomous functional differential equation representing the growth of
a single species population. J. Math. Biol. 1 (1975), 227-240; (2] On a tran-
scendental equation in the stability analysis of a population growth model.
J. Math. Biol. 3; (3] Stability of attractivity regions for autonomous func-
tional differential equations. Manuscripta Math. 15 (1975), 349-363; (4], Uber
Ejektivitat und periodische Losungen bei autonomen Punktionaldifferential-
gleichungen mit verteilter Verzogerung, Habilitationsschrift, Univ. Miinchen,
1977; [5] On instability, w-limit sets and periodic solutions of nonlinear au-
tonomous differential delay equations. Lect. Notes Math. 730 (1979), 489-
503; (6] Density of slowly oscillating solutions of ±(t) = - f(x(t- 1)). J.
Math. Anal. Appl. 79 (1981), 127-140; [7] Bifurcation from periodic solutions
in functional differential equations. Math. Z. 182 (1983), 269-325; (8] An in-
variant manifold of slowly oscillating solutions for ±(t) = -p,x(t)+ f(x(t-1)).
J. Reine Angew. Math. 414 (1991), 67-112; (9] Unstable manifolds of peri-
odic orbits of a differential delay equation. In Oscillations and Dynamics in
Delay Equations, (Eds. J.R. Graef and J.K. Hale), Contemporary Mathe-
matics 129, Amer. Math. Soc. 1992; [10] Homoclinic solutions and chaos in
x(t) = f(x(t- 1)). Nonlinear Anal. 5 (1981), 775-788.
Waltman, P. [1] Deterministic Threshold Models in the Theory of Epidemics.
Lecture Notes in Biomathematics, Vol. 1, Springer-Verlag, 1974.
Waltman, P. and J. W. Wong (1] Two-point boundary-value problems for nonlin-
ear functional differential equations. Trans. Amer. Math. Soc. 164 (1972),
39-54.
Wangersky, P. J. and W. J. Cunningham [1] Time lag in prey-predation popula-
tion models. Ecology 38 (1957), 136-139; (2] Time lag in population models.
Cold Spring Harbor Symposia on Quantitative Biology 22 (1957).
Wazewska-Czyzewska, M. and A. Lasota (1] Mathematical models of the red cell
system (in Polish). Matematyka Stosowana 6 (1976), 25-40.
Webb, G. F. [1] Autonomous nonlinear functional differential equations and non-
linear semigroups. J. Math. Anal. Appl. 46 (1974), 1-12; [2] Asymptotic
stability for abstract functional differential equations. Proc. Amer. Math.
Soc. 54 (1976), 225-230; (3] Functional differential equations and nonlinear
semigroups in £P-spaces. J. Differential Eqns. 20 (1976), 71-89; (4] Theory of
Nonlinear Age-Dependent Population Dynamics. Marcel Dekker, New York
1985.
Weiss. L. [1] On the controllability of delay differential systems. SIAM J. Control
5 (1967), 575-587.
Bibliography 443
Wexler, D. [1] Solutions periodiques des systems lineaires a argument retarde. J.

Widder, D.V. [1] The Laplace Transform. Princeton University Press, Princeton,
1946.
Winston, E. and J. A. Yorke [1] Linear delay differential equations whose solutions
become identically zero. Rev. Roum. Math. Pures Appl. 14 (1969), 885-887.
Wright, E. M. [1] A functional equation in the heuristic theory of primes. The
Mathematical Gazette, 45 (1961), 15-16; [2] A nonlinear differential differ-
ence equation J. Reine Angew. Math. 194 (1955), 66-87; [3] Linear differen-
tial difference equations. Proc. Camb. Phil. Soc. 44 (1948), 179-185.
Xie, X. [1] The multiplier equation and its application to S-solutions of a differ-
ential delay equation. J. Differential Eqns. 95 (1992), 259-280.
Yorke, J. A. [1] Noncontinuable solutions of differential delay equations. Proc.
Amer. Math. Soc. 21 (1969), 648-652; [2] Asymptotic stability for one-
dimensional differential delay equations. J. Differential Eqns. 7 (1970), 189-
202.
Yoshida, K. [1] Functional Analysis, 6th ed. Springer-Verlag, 1980.
Yoshida, K. [1] The Hopf bifurcation and its stability for semilinear diffusion
equations with time delay arising in ecology. Hiroshima Math. J. 12 (1982),
321-348.
Yoshizawa, T. [1] Stability Theory by Liapunov's Second Method. Math. Soc.
Japan, 1966; [2] Stability Theory and the Existence of Periodic Solutions and
Almost-Periodic Solutions. Applied Math. Sciences, vol. 14, 1975. Springer-
Verlag.
Ziegler, H. J. [1] Sur une note de M. C. Perella sur des solutions periodiques
d'equations fonctionnnelles differentielles. C. R. A cad. Sci. Paris Ser. A -B
267 (1968), A783-A785.
Zivotovskii, L. A. [1] Absolute stability of the solutions of differential equations
with retarded arguments. [Russian] Trudy Sem. Tear. Diff. Urav. Otkl. Arg.
7 (1969), 82-91.
Zmood, R. B. and N. H. McClamroch [1] On the pointwise completeness of dif-
ferential difference equations. J. Differential Eqns. 12 (1972), 474-486.
Zverkin, A. W. [1] Dependence of the stability of solutions of linear differential
equations with lagging argument upon the choice of the intial moment. [Rus-
sian] Vestnik Mos. Univ. Ser. Mat. Meh. Astr. Fiz. Him. 5 (1959), 15-20;
[2] The pointwise completeness of systems with lag. [Russian] Differential-
niye Uravnenija 9 (1973), 430-436, 586-587; [3] Expansion of solutions of
differential difference equations in series. [Russian] Trudy Sem. Tear. Diff.
Urav. Otkl. Argumentom 4 (1967), 3-50; [4] The connection between bound-
edness and stability of solutions of linear systems within infinite number of
degrees of freedom. [Russian] Differentialniye Uravnenija 4 (1968), 366-367;
[5] The completeness of a Floquet-type system of solutions for equations with
retardation. [Russian] Differentialniye Uravnenija 4 (1968), 474-478; [6] Ap-
pendix to Russian Translation of Bellman and Cooke, Differential Difference
Equations, Mir, Moscow, 1967.
Zverkin, A. M., Kamenskii, G. A. Norkin, S. B. and L. E. El'sgol'tz [1] Differential
equations with retarded arguments I. [Russian] Uspehi Mat. Nauk 17 (1962),
77-164; II, Trudy Sem. Tear. Diff. Urav. Otkl. Argumentom, 2 (1963), 3-49.
Index
A-stable, 372 critical point, see equilibrium point

algebraic multiplicity, 198
ascent, 78
difference operator, 59
attracting set, 107, 119
dynamical system, 101
attractive point
discrete, 101
uniform, 338
integral set of, 104
attractor, see global attractor
invariant set of, 104
atomic, 52, 53
eigenvalue, 198
backward continuation, 51 of finite type, 198
bifurcation normal, 198
subcritical, 388 ejective point, 335
supercritical, 388 evolutionary system, 173
bifurcation equation, 188 backward, 176
equivalent, 72
Caratheodory condition, 58 exponential type, 75
characteristic exponential dichotomy, 190
equation, 17, 198 equilibrium point
matrix, 201 hyperbolic, 302
exponent, 237 nondegenerate, 377
multiplier, 237, 245
multiplicity, 237 fixed-point
value, 200, 203 hyperbolic, 372
complete, 220 fixed-point theorem, 43
completely continuous operator, 43 Floquet multiplier, 237
cone functional differential equation
truncated, 34 7 neutral, 59
connecting orbits, 382 autonomous, 59
continuation, 44 homogeneous, 59
contraction, 48 linear, 59
uniform, 48 solution of, 59
Index 445
retarded, 38 mapping
autonomous , 39 asymptotica lly smooth, 110
homogeneous, 39 bounded, 69, 90
linear, 39 locally, 69
solution of, 38, 58 uniformly, 90
completely continuous, 43
locally, 89
fundamental solution, 19, 30, 172 conditionally, 89, 113
a-contractin g
conditionally, 113
generalized eigenspace, 198, 220
condensing
generic, 376
conditionally, 113
global attractor, 108, 120, 121
convergent, 126
minimal, 374
dissipative, 110, 120
guiding function, 165
method of steps, 14
gradient-like, 381
Morse decomposition, 381
graph norm, 202
Morse-Smale map, 373
Morse-Smale flow, 374
Hausdorff dimension, 369 multiplier, see characteristi c
Jordan chain, 203 negative feedback, 360

canonical system, 204 non-atomic
uniformly, 256
nonwanderin g
infinitesimal generator, 193 point, 367
invariant set, 104 set, 367
Kuratowski measure, 108 orbit, 101

positive, 119, 290
Laplace-Stieltjes transform, 76
Liapunov function, 143, 293, 383 period module, 371
limit capacity, 370 periodic point, 373
limit set hyperbolic, 373
w-, 102, 119, 290 periodic orbit, 318
a-, 102, 119, 290 Floquet multiplier, 318
Lipschitz graph, 305 hyperbolic, 318
lower semicontinuous, 380 index, 318
nondegenera te, 324
stable set, 318
manifold local, 319
local center, 313 local synchronized, 321
local center stable, 313 unstable set, 318
local center unstable, 314 local, 319
local synchronized, 321
446 Index
phase space, 402 local, 304

process, 100 local strongly, 313
asymptotically smooth, 123 synchronized, 320
autonomous, 101 local, 320
critical point, 102 uniformly, 121
dissipative, 123 square wave, 387
equilibrium, 102 Stieltjes-Volterra kernel, 256
integral of the, 104 solution of RFDE, 38
w-periodic, 100, 121 bounded, 131
invariant set of, 104 uniformly, 131
convergent, 127 ultimately, 131
uniformly, 131
stable, 130
residual set, 57
asymptotically, 130
uniformly, 130
semigroup, 68, 119 uniformly, 130
adjoint, 195
asymptotically smooth, 120
tangent, 305
a-contraction, 120
topological space,
strongly continuous, 193
dimension, 369
transposed, 196
finite dimensional, 369
slowly oscillating, 359
trajectory, 101
small solution, 74
p-periodic, 102
nontrivial, 74
smoothness on the measure, 52
solution map, 56, 68 unstable set, 304
spectral radius, 213 local, 304
spectrum, 197 local strongly, 313
continuous, 197 synchronized, 320
point, 197
residual, 197
stable D operator, 274 Volterra
kernel, 169
stable set, 108, 120, 121, 304
resolvent, 169
asymptotically, 108, 120, 121
uniformly, 108, 120, 121
Index 447
List of symb ols
A(F), 367 dim H(K), 370

B(A), 202 .C(X, Y), 52
co J, 109 "Y+(a,x), 101
C([a, b], 1Rn), 38 "Y+(B), 119
c(K), 370 KCr(X , X), 372
C(V, 1Rn), 40 D(F), 367
C 0 (V; 1Rn), 40 (-' . ), 211, 268
CP(f?, JRn), 48 IR, 38
c, 38 T+(a,x), 101
Cl G, 143 w(a, x), 102
G, 143
(continued from page ii)
52. Chipot: Variational Inequalities and Flow in Porous Media.

53. Majtla: Compressible Fluid Flow and System of Conservation Laws in Several Space Variables.
54. Wasow: Linear Turning Point Theory.
55. Yosida: Operational Calculus: A Theory of Hyperfunctions.
56. Chang/Howes: Nonlinear Singular Perturbation Phenomena: Theory and Applications.
57. Reinhardt: Analysis of Approximation Methods for Differential and Integral Equations.
58. Dwoyer/Hussaini/Voigt (eds): Theoretical Approaches to Turbulence.
59. Sanders/Verhulst: Averaging Methods in Nonlinear Dynamical Systems.
60. Ghil/Childress: Topics in Geophysical Dynamics: Atmospheric Dynamics, Dynamo Theory and
Climate Dynamics.
61. Sattinger/Weaver: Lie Groups and Algebras with Applications to Physics, Geometry, and Mechanics.
62. LaSalle: The Stability and Control of Discrete Processes.
63. Grasman: Asymptotic Methods of Relaxation Oscillations and Applications.
64. Hsu: Cell-to-Cell Mapping: A Method of Global Analysis for Nonlinear Systems.
65. Rand/Armbruster: Perturbation Methods, Bifurcation Theory and Computer Algebra.
66. Hlavacek/Haslinger/Necasl/Lov(sek: Solution of Variational Inequalities in Mechanics.
67. Cercignani: The Boltzmann Equation and Its Applications.
68. Temam: Infinite Dimensional Dynamical Systems in Mechanics and Physics.
69. Golubitsky/Stewart/Schaeffer: Singularities and Groups in Bifurcation Theory, Vol. II.
70. Constantin/Foias/Nicolaenko/Temam: Integral Manifolds and Inertial Manifolds for Dissipative Partial
Differential Equations.
71. Catlin: Estimation, Control, and the Discrete Kalman Filter.
72. Lochak/Meunier: Multiphase Averaging for Classical Systems.
73. Wiggins: Global Bifurcations and Chaos.
74. Mawhin/Willem: Critical Point Theory and Hamiltonian Systems.
75. Abraham/Marsden/Ratiu: Manifolds, Tensor Analysis, and Applications, 2nd ed.
76. Lagerstrom: Matched Asymptotic Expansions: Ideas and Techniques.
77. Aldous: Probability Approximations via the Poisson Clumping Heuristic.
78. Dacorogna: Direct Methods in the Calculus of Variations.
79. Hernandez-Lerma: Adaptive Markov Processes.
80. Lawden: Elliptic Functions and Applications.
81. Bluman/Kumei: Symmetries and Differential Equations.
82. Kress: Linear Integral Equations.
83. Bebernes/Eberly: Mathematical Problems from Combustion Theory.
84. Joseph: Fluid Dynamics of Viscoelastic Fluids.
85. Yang: Wave Packets and Their Bifurcations in Geophysical Fluid Dynamics.
86. Dendrinos/Sonis: Chaos and Socio-Spatial Dynamics.
87. Weder: Spectral and Scattering Theory for Wave Propagation in Perturbed Stratified Media.
88. Bogaevski/Povzner: Algebraic Methods in Nonlinear Perturbation Theory.
89. O'Malley: Singular Perturbation Methods for Ordinary Differential Equations.
90. Meyer/Hall: Introduction to Hamiltonian Dynamical Systems and the N-body Problem.
91. Straughan: The Energy Method, Stability, and Nonlinear Convection.
92. Naber: The Geometry of Minkowski Spacetime.
93. Colton/Kress: Inverse Acoustic and Electromagnetic Scattering Theory.
94. Hoppensteadt: Analysis and Simulation of Chaotic Systems.
95. Hackbusch: Iterative Solution of Large Sparse Systems of Equations.
96. Marchioro/Pulvirenti: Mathematical Theory of Incompressible Nonviscous Fluids.
97. Lasota/Mackey: Chaos, Fractals and Noise: Stochastic Aspects of Dynamics, 2nd ed.
98. de Boor/Hollig/Riemenschneider: Box Splines.
99. Hale!Lunel: Introduction to Functional Differential Equations.
100. Arnol 'd!Constantin/Feigenbaum/Golubitsky/1oseph/Kadano[f!Kreiss/McKean/Marsden/Teman:
Trends and Perspectives in Applied Mathematics.
101. Nusse/Yorke: Dynamics: Numerical Explorations.
102. Chossat/looss: The Couette-Taylor Problem.

Introduction To Functional Differential Equations, J Hale, Springer, 1993,458s PDF

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Introduction To Functional Differential Equations, J Hale, Springer, 1993,458s PDF

Diunggah oleh

Hak Cipta:

Format Tersedia

Applied Mathematical Sciences

1. John: Partial Differential Equations, 4th ed.

(continued following index)

Springer Science+Business Media, LLC

Mathematics Subject Classification (1991): 34K20, 34A30, 39A10

Library of Congress Cataloging-in-Publication Data

Printed on acid-free paper.

© 1993 Springer Science+Business Media New York

Production managed by Natalie Johnson; manufacturing supervised by Jacqui Ashri.

1. Linear differential difference equations . . . . . . . . . . . . . . . . . . . . 11

2. Functional differential equations: Basic theory . . . . . . . . . . 38

3. Properties of the solution map.............................. 67

3.6 Compactness and representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

4. Autonomous and periodic processes ....................... 100

5. Stability theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

6. General linear systems .............................. ......... 167

7. Linear autonomous equations .............................. . 193

8. Periodic systems .................... .................... ...... 236

9. Equations of neutral type .................... ............... 255

10. Near equilibrium and periodic orbits .................... . 302

11. Periodic solutions of autonomous equations ............. 331

12. Additional topics .................... .................... .... 364

12.9 Infinite delay .................... .................... .......... 401

Appendix: Stability of characteristic equations .............. 414

In many applications, one assumes the system under consideration is

x(t) = -cxx(t- 1)[1 + x(t)]

in his study of the distribution of primes. Variants of this equation have

x(t) =-ex [1° 1

for the growth of a single species.

x(t) = [E1- ''!lY(t)- j_or F1(8)y(t + 8)d8] x(t)

y(t) = [ -El + 'Y2x(t) + j_or F2(8)x(t + B)d()] y(t)

x(t) = ax(t) [m ~x(t)] - bx(t)y(t)

S(t) = -(3(t)S(t)[2'Y + S(t- 14)- S(t- 12)] + 'Y

where I represents the number of infectious individuals and g is a non-

A more general equation describing the spread of disease taking into

i:(t) = -l~r a(t- u)g(x(u)) du

x(t) + ai:(t)- f(x(t- r))i:(t- r) + x(t) = 0

y(t) + [w 2 + EAsin¢(t- rl)]y(t) = -E[hy(t) + '"YY3 (t- r 2 )]

In the theory of optimal control, Krasovskii [2] has studied extensively

u(t) = /_or d[ry(t, B)]y(t +B)+ /_or d[JL(t, B)]u(t +B).

connection through the partial differential equations (considered as a con-

with the boundary conditions

v(x, t) = ¢(x- st) + '1/J(x + st)

i(1, t)- Ki(1, t- ~) =a-~ v(1, t)- K v(1, t- ~)

The prescription for passing from a linear partial differential equation

¢(1-st)= b(t)- K'lj;(st- 1).

v(1, t) = w(t)- Kw(t- r)

In his consideration of shunted transmission lines, Lopes [2] encountered

x(t) + wfx(t) = Eh (x(t), x(t), y(t), y(t)) + 'Ydi(t- r),

x(t) = JI(t,x(t),x(g(t))) + h(t,x(t),x(g(t)))x(g(t))

special case of the retarded equation if we replace t by - T and investigate

Ut(x, t)- .du(x, t) + R(x, t)u(x, w(t)) = f(x, t)

x(t) = g(t, x(t- 1), x(t- 2), ... , x(t- N)),

so important in problems in economics and models of heredity.

x(t) = f(t) +lot a(t, s, x(s)) ds

In this chapter, we discuss the simplest possible differential difference equa-

1.1 Differential and difference equations

(1.2) det (M- A)= 0.

The coefficients of the polynomials Pj are determined from the generalized

(1.3) x =Ax+ f(t)

The derivation is easily obtained by making the transformation of variables

(1.5) x(t) = Ax(t- 1) + Bx(t- 2)

(1.6) z(t) = Cz(t- 1),