NATO Science Series
A Series presenting the results of scientific meetings supported under the NATO Science
Programme.
The Series is published by IOS Press, Amsterdam, and Springer in conjunction with the NATO
Public Diplomacy Division
SubSeries
I. Life and Behavioural Sciences
II. Mathematics, Physics and Chemistry
III. Computer and Systems Science
IV. Earth and Environmental Sciences
IOS Press
Springer
IOS Press
Springer
The NATO Science Series continues the series of books published formerly as the NATO ASI Series.
The NATO Science Programme offers support for collaboration in civil science between scientists of
countries of the EuroAtlantic Partnership Council. The types of scientific meeting generally supported
are “Advanced Study Institutes” and “Advanced Research Workshops”, and the NATO Science Series
collects together the results of these meetings. The meetings are coorganized bij scientists from
NATO countries and scientists from NATO’s Partner countries – countries of the CIS and Central and
Eastern Europe.
Advanced Study Institutes are highlevel tutorial courses offering indepth study of latest advances
in a field.
Advanced Research Workshops are expert meetings aimed at critical assessment of a field, and
identification of directions for future action.
As a consequence of the restructuring of the NATO Science Programme in 1999, the NATO Science
Series was reorganised to the four subseries noted above. Please consult the following web sites for
information on previous volumes published in the Series.
http://www.nato.int/science
http://www.springer.com
http://www.iospress.nl
Series II: Mathematics, Physics and Chemistry – Vol. 205
Delay Differential Equations
and Applications
edited by
O. Arino
University of Pau, France
M.L. Hbid
University Cadi Ayyad,
Marrakech, Morocco
and
E. Ait Dads
University Cadi Ayyad,
Marrakech, Morocco
Published in cooperation with NATO Public Diplomacy Division
Proceedings of the NATO Advanced Study Institute on
Delay Differential Equations and Applications
Marrakech, Morocco
9–21 September 2002
A C.I.P. Catalogue record for this book is available from the Library of Congress.
ISBN10
ISBN13
ISBN10
ISBN13
ISBN10
ISBN13
1402036469 (PB)
9781402036460 (PB)
1402036450 (HB)
9781402036453 (HB)
1402036477 (ebook)
9781402036477 (ebook)
Published by Springer,
P.O. Box 17, 3300 AA Dordrecht, The Netherlands.
www.springer.com
Printed on acidfree paper
All Rights Reserved
© 2006 Springer
No part of this work may be reproduced, stored in a retrieval system, or transmitted
in any form or by any means, electronic, mechanical, photocopying, microfilming,
recording or otherwise, without written permission from the Publisher, with the exception
of any material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work.
Contents
List of Figures
xiii
Preface
xvii
Contributing Authors
xxi
Introduction
M. L. Hbid
xxiii
1
History Of Delay Equations
J.K. Hale
1
Stability of equilibria and Lyapunov functions
2
Invariant Sets, Omegalimits and Lyapunov functionals.
3
Delays may cause instability.
4
Linear autonomous equations and perturbations.
5
Neutral Functional Diﬀerential Equations
6
Periodically forced systems and discrete dynamical systems.
7
Dissipation, maximal compact invariant sets and attractors.
8
Stationary points of dissipative ﬂows
Part I General Results and Linear Theory of Delay Equations in
Finite Dimensional Spaces
2
Some General Results and Remarks on Delay Diﬀerential Equations
E. Ait Dads
1
Introduction
2
A general initial value problem
2.1
Existence
2.2
Uniqueness
2.3
Continuation of solutions
2.4
Dependence on initial values and parameters
2.5
Diﬀerentiability of solutions
v
1
3
7
10
12
16
20
21
24
29
31
31
33
34
35
37
38
40
vi
DELAY DIFFERENTIAL EQUATIONS
3
Autonomous Functional Diﬀerential Equations
Franz Kappel
1
Basic Theory
1.1
Preliminaries
1.2
Existence and uniqueness of solutions
1.3
The Laplacetransform of solutions. The fundamental
matrix
1.4
Smooth initial functions
1.5
The variation of constants formula
1.6
The Spectrum
1.7
The solution semigroup
2
Eigenspaces
2.1
Generalized eigenspaces
2.2
Projections onto eigenspaces
2.3
Exponential dichotomy of the state space
3
Small Solutions and Completeness
3.1
Small solutions
3.2
Completeness of generalized eigenfunctions
4
Degenerate delay equations
4.1
A necessary and suﬃcient condition
4.2
A necessary condition for degeneracy
4.3
Coordinate transformations with delays
4.4
The structure of degenerate systems
with commensurate delays
Appendix: A
Appendix: B
Appendix: C
Appendix: D
46
54
55
59
68
71
71
90
101
104
104
109
110
110
116
119
References
137
Part II Hopf Bifurcation, Centre manifolds and Normal Forms for Delay Diﬀerential Equations
4
Variation of Constant Formula for Delay Diﬀerential Equations
M.L. Hbid and K. Ezzinbi
1
Introduction
2
3
Variation Of Constant Formula Using SunStar Machinery
2.1
Duality and semigroups
2.1.1 The variation of constant formula:
2.2
Application to delay diﬀerential equations
2.2.1 The trivial equation:
2.2.2 The general equation
Variation Of Constant Formula Using Integrated Semigroups
Theory
3.1
Notations and basic results
3.2
The variation of constant formula
41
41
41
44
124
127
129
131
132
141
143
143
145
145
146
147
147
149
149
150
153
Contents
5
Introduction to Hopf Bifurcation Theory for Delay Diﬀerential
Equations
M.L. Hbid
1
Introduction
1.1
Statement of the Problem:
1.2
History of the problem
1.2.1 The Case of ODEs:
1.2.2 The case of Delay Equations:
2
The Lyapunov Direct Method And Hopf Bifurcation: The Case
Of Ode
3
The Center Manifold Reduction Of DDE
3.1
The linear equation
3.2
The center manifold theorem
3.3
Back to the nonlinear equation:
3.4
The reduced system
4
Cases Where The Approximation Of Center Manifold Is
Needed
4.1
Approximation of a local center manifold
4.2
The reduced system
6
An Algorithmic Scheme for Approximating Center Manifolds
and Normal Forms for Functional Diﬀerential Equations
M. Ait Babram
1
Introduction
2
Notations and background
3
Computational scheme of a local center manifold
3.1
Formulation of the scheme
3.2
Special cases.
3.2.1 Case of Hopf singularity
3.2.2 The case of Bogdanov Takens singularity.
4
Computational scheme of Normal Forms
4.1
Normal form construction of the reduced system
4.2
Normal form construction for FDEs
7
Normal Forms and Bifurcations for Delay Diﬀerential Equations
T. Faria
1
Introduction
2
Normal Forms for FDEs in Finite Dimensional Spaces
2.1
Preliminaries
2.2
The enlarged phase space
2.3
Normal form construction
2.4
Equations with parameters
2.5
More about normal forms for FDEs in Rn
3
Normal forms and Bifurcation Problems
3.1
The BogdanovTakens bifurcation
3.2
Hopf bifurcation
vii
161
161
161
163
163
164
166
168
169
172
177
179
182
183
188
193
193
195
199
202
209
209
210
213
214
221
227
227
231
231
232
234
240
241
243
243
246
viii
DELAY DIFFERENTIAL EQUATIONS
4
5
Normal Forms for FDEs in Hilbert Spaces
4.1
Linear FDEs
4.2
Normal forms
4.3
The associated FDE on R
4.4
Applications to bifurcation problems
Normal Forms for FDEs in General Banach Spaces
5.1
Adjoint theory
5.2
Normal forms on centre manifolds
5.3
A reactiondiﬀusion equation with delay and Dirichlet
conditions
References
Part III Functional Diﬀerential Equations in Inﬁnite Dimensional
Spaces
253
254
256
258
260
262
263
268
270
275
283
8
285
A Theory of Linear Delay Diﬀerential Equations in Inﬁnite Dimensional Spaces
O. Arino and E. S´
anchez
1
Introduction
285
1.1
A model of ﬁsh population dynamics with spatial diffusion (11)
286
1.2
An abstract diﬀerential equation arising from cell population dynamics
288
1.3
From integrodiﬀerence to abstract delay diﬀerential equations (8)
292
1.3.1 The linear equation
292
1.3.2 Delay diﬀerential equation formulation of system (1.5)(1.6)
295
1.4
The linearized equation of equation (1.17) near nontrivial steadystates
297
1.4.1 The steadystate equation
297
298
1.4.2 Linearization of equation (1.17 ) near (n, N )
1.4.3 Exponential solutions of (1.20)
299
1.5
Conclusion
303
2
The Cauchy Problem For An Abstract Linear Delay Diﬀerential Equation
303
2.1
Resolution of the Cauchy problem
304
2.2
Semigroup approach to the problem (CP)
306
310
2.3
Some results about the range of λI − A
3
Formal Duality
311
3.1
The formal adjoint equation
313
316
3.2
The operator A∗ formal adjoint of A
3.3
Application to the model of cell population dynamics
317
3.4
Conclusion
320
4
Linear Theory Of Abstract Functional Diﬀerential Equations
Of Retarded Type
320
321
4.1
Some spectral properties of C0 semigroups
ix
Contents
4.2
4.3
4.4
Decomposition of the state space C([−r, 0]; E)
A Fredholm alternative principle
Characterization of the subspace R (λI − A)m for λ in
324
326
(σ\σe ) (A)
326
4.5
5
Characterization of the projection operator onto the
subspace QΛ
4.6
Conclusion
A Variation Of Constants Formula For An Abstract Functional
Diﬀerential Equation Of Retarded Type
5.1
The nonhomogeneous problem
5.2
Semigroup deﬁned in L(E)
5.3
The fundamental solution
5.4
The fundamental solution and the nonhomogeneous
problem
5.5
Decomposition of the nonhomogeneous problem
in C([−r, 0]; E)
9
The Basic Theory of Abstract Semilinear Functional Diﬀerential
Equations with NonDense Domain
K. Ezzinbi and M. Adimy
1
Introduction
2
Basic results
3
Existence, uniqueness and regularity of solutions
4
The semigroup and the integrated semigroup in the autonomous
case
5
Principle of linearized stability
6
Spectral Decomposition
7
Existence of bounded solutions
8
Existence of periodic or almost periodic solutions
9
Applications
References
Part IV More on Delay Diﬀerential Equations and Applications
10
Dynamics of Delay Diﬀerential Equations
H.O. Walther
1
Basic theory and some results for examples
1.1
Semiﬂows of retarded functional diﬀerential equations
1.2
Periodic orbits and Poincar´e return maps
1.3
Compactness
1.4
Global attractors
1.5
Linear autonomous equations and spectral
decomposition
1.6
Local invariant manifolds for nonlinear RFDEs
1.7
Floquet multipliers of periodic orbits
1.8
Diﬀerential equations with statedependent delays
331
335
335
336
337
338
341
344
347
347
350
354
372
381
383
385
391
393
399
409
411
411
411
416
418
418
419
423
425
435
x
DELAY DIFFERENTIAL EQUATIONS
2
3
4
5
Monotone feedback: The structure of invariant sets and
attractors
2.1
Negative feedback
2.2
Positive feedback
Chaotic motion
Stable periodic orbits
Statedependent delays
11
Delay Diﬀerential Equations in Single Species Dynamics
S. Ruan
1
Introduction
2
Hutchinson’s Equation
2.1
Stability and Bifurcation
2.2
Wright Conjecture
2.3
Instantaneous Dominance
3
Recruitment Models
3.1
Nicholson’s Blowﬂies Model
3.2
Houseﬂies Model
3.3
Recruitment Models
4
The Allee Eﬀect
5
FoodLimited Models
6
Regulation of Haematopoiesis
6.1
MackeyGlass Models
6.2
WazewskaCzyzewska and Lasota Model
7
A Vector Disease Model
8
Multiple Delays
9
Volterra Integrodiﬀerential Equations
9.1
Weak Kernel
9.2
Strong Kernel
9.3
General Kernel
9.4
Remarks
10
Periodicity
10.1 Periodic Delay Models
10.2 Integrodiﬀerential Equations
11
StateDependent Delays
12
Diﬀusive Models with Delay
12.1 Fisher Equation
12.2 Diﬀusive Equations with Delay
12
WellPosedness, Regularity and Asymptotic Behaviour of Retarded Diﬀerential Equations by Extrapolation Theory
L. Maniar
1
Introduction
2
Preliminaries
3
Homogeneous Retarded Diﬀerential Equations
436
437
439
451
456
468
477
477
478
479
481
483
484
484
486
487
488
489
491
491
493
493
495
496
498
500
502
504
505
505
507
511
514
514
515
519
519
521
525
Contents
xi
13
Time Delays in Epidemic Models: Modeling and Numerical Considerations
J. Arino and P. van den Driessche
1
Introduction
2
Origin of time delays in epidemic models
2.1
Sojourn times and survival functions
2.2
Sojourn times in an SIS disease transmission model
3
A model that includes a vaccinated state
4
Reduction of the system by using speciﬁc P (t) functions
4.1
Case reducing to an ODE system
4.2
Case reducing to a delay integrodiﬀerential system
5
Numerical considerations
5.1
Visualising and locating the bifurcation
5.2
Numerical bifurcation analysis and integration
6
A few words of warning
Appendix
1
Program listings
1.1
MatLab code
1.2
XPPAUT code
2
Delay diﬀerential equations packages
2.1
Numerical integration
2.2
Bifurcation analysis
539
540
540
541
544
548
548
549
550
550
551
552
555
555
556
557
557
557
558
References
559
Index
579
539
17 10.5 10.24 10.15 10.10 10.29 413 413 417 417 420 422 423 424 425 427 428 430 431 432 432 434 438 439 441 442 443 445 446 447 447 449 450 451 453 xiii .2 10.13 10.26 10.12 10.8 10.27 10.22 10.14 10.3 10.9 10.6 10.20 10.4 10.23 10.16 10.1 10.28 10.25 10.19 10.11 10.List of Figures 10.21 10.7 10.18 10.
5107.147.81.10 Aperiodic behavior of the solutions of the MackeyGlass model (6. Here λ = 0. c = 1. g = 0. 481 11. x0 = 4.2.2 The periodic solution of the Hutchinson’s equation (2.2).xiv DELAY DIFFERENTIAL EQUATIONS 10.30 453 10.000226. m = 10 and τ = 20. Here r = 0. a periodic solution bifurcated from x∗ = 1. d = 0.1. m = 10 and τ = 6.00.1).15. z = 0.7 11.1).31 454 10. 487 The steady state of the delay model (4. b = 1.3 11.8).1. x0 = 4.00. 486 Numerical simulations in the houseﬂies model (3. and τ = 15. δ = 0. 489 The steady state of the delay foodlimited model (5. K = 1.3) is stable for small delay (τ = 8) and unstable for large delay (τ = 12. 492 11.4 11. (i) When τ = 8.36 464 10. (ii) When τ = 11. Here the parameter values b = 1. Here r = 0.1). 481 Numerical simulations for the Hutchinson’s equation (2. k = 0. the steady state x∗ = 1 is stable. δ = 0. a = 01. τ = 5 were reported in Taylor and Sokal (1976). K = 1.2. 482 Oscillations in the Nicholson’s blowﬂies equation (3.1 The bifurcation diagram for equation (2.1) is attractive. τ = 0. 491 Oscillations in the MackeyGlass model (6. Here P = 8.5 11.5. Here λ = 0.34 459 10.6 11. Here a = 1.37 465 10.38 466 11. c = 0.9 11.15. g = 0. and τ = 15.35 463 10. 486 Aperiodic oscillations in the Nicholson’s blowﬂies equation (3.475. Here P = 8.33 459 10.2. 492 .175.1).32 457 10.1).8 11. a = 01.1).1).
using dde23. (a) The steady state (a) is stable when τ1 = 15 and τ2 = 5 and (b) becomes unstable when τ1 = 15 and τ2 = 10.4 − 3. 11. (c) and (d): Forward bifurcation case. with parameters as in the text.15. here for both cases τ = 5. parameters as in text. K = 1.1).8.3.0 13. a1 = 0.13 (a) Weak delay kernel and (b) strong delay kernel.2). 11.2 The transfer diagram for the SIV model. a2 = 0.8(a < b). 11. 13. the zero steady state u = 0 is asymptotically stable.1) is globally stable.12 For the two delay logistic model (8.1541. 11.15. a Hopf bifurcation occurs.14 The steady state of the integrodiﬀerential equation (9.8. When a = 3. b = 4. b = 1.11 Numerical simulations for the vector disease equation (7.1). and (ii) a = 9. K = 1.1. b = 4.4 Bifurcation diagram and some solutions of (4. 11. 13. Here r = 0.15 The steady state x∗ = K of the integrodiﬀerential equation (9.3) with r = 0. (a) and (b): Backward bifurcation case.00.5 Value of I ∗ as a function of ω by solving H(I. xv 494 496 498 500 502 512 515 542 545 551 553 554 555 . 13. ω) = 0.00 and τ (x) = a + bx2 . 13.8(a > b).1).17 The traveling front proﬁles for the Fisher equation (12.065.16 Numerical simulations for the statedependent delay model (11. (i) a = 5.15.3 Possible bifurcation scenarios.75. Here r = 0. parameters as in the text except that σ = 0.25.1. parameters as in the text.10) losses stability and a Hopf bifurcation occurs when α changes from 0.List of Figures 11. When a = 5. 13. choose r = 0. c = 2.00. Here D = r = K = 1.1 The transfer diagram for the SIS model. b = 1. K = 1. the positive steady state u∗ is asymptotically stable for all delay values.6 Plot of the solution of (6.15.3).65 to 0. 11.
The school was supported by ﬁnancements from NATOASI (Nato advanced School). adapted to delay diﬀerential equations. to equations having in addition terms involving past values of the solution.Preface This book groups material that was used for the Marrakech 2002 School on Delay Diﬀerential Equations and Applications. The school was held from September 921 2002 at the Semlalia College of Sciences of the Cadi Ayyad University. and the classic approach for the resolution of these equations by perturbation. a training programme was devised that may be summarized by the following three keywords: the Cauchy problem. Marrakech. including ordinary diﬀerential equations and semilinear evolution equations. from Monday through Friday. taught by block of two units. such equations where the only independent variable is the time variable. the International Centre of Pure and Applied Mathematics (CIMPA.30 am to 6. France) and Cadi Ayyad University. The delay introduces speciﬁc diﬀerences and diﬃculties which are taken into account in the progression of the course. the ﬁrst week having been devoted to “ordinary” delay diﬀerential equations. Nice. the variation of constants formula. Courses were divided into units of 45mn duration. such as the diﬀusionreaction equation. plenary lectures (3) and communications (9).30 pm. 47 participants and 15 instructors originating from 21 countries attended the school. in addition. The school was intended for mathematicians willing to acquire some familiarity with delay diﬀerential equations or enhance their knowledge on this subject. Morocco. The aim was indeed to extend the basic set of knowledge. 8. Financial limitations only allowed support for part of the people from Africa and Asia who had expressed their interest in the school and had hoped to come. and a 25mn break between two blocks. This deﬁnes the general method for the resolution of semilinear evolution equations. local study of equilibria. In order to achieve this goal. with a short 5mn break between two units within a block. such as for example the diﬀusionreaction equations arising in morphogenesis or the BelouzovZhabotinsky chemical reaction. only the ﬁnite dimension was xvii . The activity of the school consisted in courses.
Aside the training on the basic theory of delay diﬀerential equations. As many events of this nature at that time. which represents an entirely new line of research.xviii DELAY DIFFERENTIAL EQUATIONS considered. The problem gives rise to a diﬀerential equation with deviated arguments (both retarded and advanced). the course by John MalletParet during the ﬁrst week discussed very recent results motivated by the problem of determining wave fronts in lattice diﬀerential equations. The main part of the courses given in the school are reproduced as lectures notes in this book. HansOtto Walther presented results regarding existence and description of the attractor of a scalar delay diﬀerential equation. Three plenary conferences usefully extended the contents of the ﬁrst week courses. During the second week. Arino and M. J. This book is dedicated to him. A quick description of the contents the book is given in the general introduction. as well as on partial diﬀerential equations with delay. He wanted this book to be published. Hbid .L. Also. He unfortunately passed away on September 29. and did a lot to that eﬀect. during the second week. 2003. this school was under the scientiﬁc supervision of Ovide Arino. attention was focused on “ordinary” delay diﬀerential equations in inﬁnite dimensional vector spaces.
This book is dedicated to the memory of Professor Ovide Arino .
Professor Emeritus. Georgia Institute of Technology. British Columbia. Mohammed Ait Babram. Portugal Moulay Lhassan Hbid. Assistant Professor. xxi . Khalil Ezzinbi. University of Manitoba. Morocco. Mostafa Adimy. Marrakech. Julien took over edition of the book after Ovide’s death. Canada. Professor. Cadi Ayyad University. Centre de Bondy. Morocco. Marrakech. University of Victoria. Julien Arino. Professor. Professor. Atlanta. Cadi Ayyad University of Marrakech. P. Professor. van den Driessche. Professor. Winnipeg. Professor. Professor. France. University of Pau. Victoria. Hale. Jack K. Assitant Professor. University of Lisboa. USA. France. Manitoba. Ovide Arino. Canada. and Universit´e de Pau. Tereza Faria. Morocco. Cadi Ayyad University.Contributing Authors Elhadi Ait Dads. Institut de Recherche pour le Developpement. Marrakech. Cadi Ayyad University. Morocco.
Spain. University of Miami. University Polytecnica de Madrid.xxii DELAY DIFFERENTIAL EQUATIONS Franz Kappel. PostDoctoral Fellow. Professor. (Technical realization of the book). Cadi Ayyad University. Lahcen Maniar. Eva Sanchez. Marrakech. Professor. University of Graz. Marrakech. Cadi Ayyad University. University of Giessen. . Said Boulite. Shigui Ruan. Germany. Morocco. USA. Austria. Professor. Morocco. Professor. Professor. HansOtto Walther.
Hale. contrary to the ﬂow associated with a smooth ordinary diﬀerential system of equations. The third part corresponds to the contributions of Professors O.L. This part groups the contributions of Professor T. The results and conjectures presented by Ait Dads are classical and are for most of them taken from a recent monograph by J. An example of such exotic properties. It consists of four parts. L. Ait Babram and Professor M. Faria. Sanchez. which is a local diﬀeomorphism for all times. Hbid. Hale and S. Adimy and K. Their inclusion in the initiation to DDE proposed by Ait Dads is mainly intended to allow readers to get some familiarity with the subject and open their horizons and possibly entice their appetite for exploring new avenues. degenerates in ﬁnite time and can even vanish in ﬁnite time. Arino. The second part deals with some qualitative theory of DDE : normal forms. Ruan. S. M. the semiﬂow associated with a DDE does not extend backward in time. centre manifold and Hopf bifurcation theory in ﬁnite dimension. Ait Dads. L. The ﬁrst part concerns some general results on the quantitative aspects of nonlinear delay diﬀerential equations.O.Introduction M. he then derives a number of typical properties of DDE and their solutions. discussed in Ait Dads’s lectures. Kappel. Arino. The last part contains the contributions of Professors H. preceded by an overview by Professor J. Maniar and J.K. Ezzinbi. Verduyn Lunel on the subject. xxiii . It is devoted to discussions on quantitative and qualitative aspects of functionnal diﬀerential equations (FDE) in inﬁnite dimension. Faria. Ait Dads’s contribution deals with a direct method to provide an existence result. and a linear theory of delay diﬀerential equations (DDE) by Professor F. E. Doctor M. Walther. Many such properties are not yet understood and would certainly deserve to be thoroughly investigated. is the fact that. Hbid This book is devoted to the theory of delay equations and applications. T. by Professor E.
Even if it is possible to proceed by direct methods. that is to say. that of completeness. The typical framework here is a DDE deﬁned in an open subset of the state space. The Lyapunov function is determined recursively in the form of a Taylor expansion. The part undertaken jointly by M. namely. more recently. Changes are expected near values of the parameter for which the equilibrium is a center. This method consists in looking for a Lyapunov function associated with the ordinary diﬀerential equation obtained by restricting the DDE to a center manifold. Faria in her lecture notes. the direct Lyapunov method. rather a family of such equations dependent upon one parameter.xxiv DELAY DIFFERENTIAL EQUATIONS In his lecture notes. The delay introduces its own problems in that case. Hbid and M. which possesses for each value of the parameter a known equilibrium (the socalled “trivial equilibrium”): one studies the stability of the equilibrium and the possible changes in the linear stability status and how these changes reﬂect in the local dynamics of the nonlinear equation. A ﬁrst application of the linear and the semi linear theory presented by Ait Dads and Kappel is the study of bifurcation of equilibria in nonlinear delay diﬀerential equations dependent on one or several parameters. The same issue. dependent on the nature of the delay and. The method was presented both in the case of a delay diﬀerential equation and also in the case when the equation is the sum of a delay diﬀerential equation . The method presented by Faria is an extension to this inﬁnite dimensional frame of the wellknown method of normal forms. Franz Kappel presents the construction of the elementary solution of a linear DDE using the Laplace transform. was dealt with by T. useful in the study of spectral properties of DDEs. when is the vector space spanned by the eigenvectors total (dense in the state space)? This issue is tightly connected with another delicate and still open one.L. Ait Babram deals with a panorama of the best known methods. The very complete and elaborate lecture notes he provided for the course are in fact an extension of the ones written on the occasion of the ﬁrst school. in the context of partial diﬀerential equations with delay. Kappel also dealt with a fundamental issue of the linear theory of delay diﬀerential equations. the Laplace transform provides an explicit expression of the elementary solution. and these problems have given rise to a variety of approaches. that is. the existence of “small” solutions (solutions which approach zero at inﬁnity faster than any exponential). Kappel taught during the ﬁrst school on delay diﬀerential equations held at the University of Marrakech in 1995. then concentrates on a method elaborated within the dynamical systems group at the Cadi Ayyad University. on the dimension of the underlying space of trajectories. This course extends the one that Prof.
pointing out the signiﬁcance of the delay. elaborated in collaboration with Professor Eva Sanchez. The equations under investigation range from delay diﬀerential equations deﬁned by a bounded “delay” operator to equations in which the “delay” operator has a domain which is only part of a larger space (it may be for example the sum of the Laplace operator and a bounded operator). . and Professor Jianhong Wu. which extends to inﬁnite dimensions the classical linear theory. Arino.. as it is treated in the monograph by Hale and Lunel. The lecture notes written by Professor HansOtto Walther are composed of two independent parts: the ﬁrst part deals with the geometry of the attractor of the dynamical system deﬁned by a scalar delay diﬀerential equation with monotone feedback. The lectures notes by Professors O. to neutral type equations in which the delay appears also in the time derivative. Canada. Both Prof. they constitute a state of the art of the treatment of the Cauchy problem in the frame of linear functional differential equations. Altogether. both in the autonomous and the non autonomous cases. and give a generic scheme to approximate the center manifolds in both cases of singularies (Hopf. and L. Tibor Krisztin. K. S. Both a negative and a positive feedback were envisaged by Walther and his coworkers. The course could only give a general idea of the general procedure that was followed in proving those results and was mainly intended to elicit the interest of participants. Ruan provides a thorough review of models involving delays in ecology. The methods presented range from the classical theory of strongly continuous semigroups to extrapolation theory. Ezzinbi and M. Hungary. In collaboration with Dr. the last one being more than 200 pages long.INTRODUCTION xxv and a diﬀusion operator. to inﬁnite delay. very detailed global results on the geometric nature of the attractor and the ﬂow along the attractor were found. also including the theory of integrated semigroups and the theory of perturbation by duality.). Faria and Dr. Fields Institute. Adimy and Ezzinbi dealt with a general neutral equation perturbed by the Laplace operator. Maniar present approaches along the line of the semigroup theory. . Toronto. Ait Babram discuss the BogdanovTakens and the Hopf bifurcation singularities as examples. Arino presented a theory. These lectures prolong in the framework of inﬁnite dimension the presentations made during the ﬁrst part by Ait Dads and Kappel in the case of ﬁnite dimensions. The second part of Walther’s lecture notes is devoted to a presentation of very recent results obtained by Walther in the study of statedependent delay diﬀerential equations. HopfHopf. from the University of Szeged. BogdanovTakens. These results have been obtained within the past ten years or so and are presented in a number of articles and monographs. Adimy. In his lecture notes.
of fundamental importance in applications where it notably justiﬁes robustness of model representations. stability loss and the corresponding changes in the dynamical features of the problem. The Morse decomposition. semilinear equations.xxvi DELAY DIFFERENTIAL EQUATIONS Most of his concern is about stability. The methods of resolution covered most of the currently known ones. DDE with inﬁnite delay. just brieﬂy reviewed the “structural stability” approach. equations with delay dependent upon the starter. The proliﬁc theory of oscillations for DDE was not even mentioned. the linear theory in ﬁnite and in inﬁnite dimensions. The methods used by Ruan are those developed by Faria and Magalhaes in a series of papers. J. nor the DDE with impulses which are an important example in applications. All this however is just a small portion of the theory of DDE. Dr. the theory of monotone (with respect to an order relation) semi ﬂows for DDE which plays an important role in applications to ecology (cooperative systems) was considered only in the scalar case (the equation with positive feedback in Walther’s course). Various types of functional diﬀerential equations are considered in addition to the usual DDE: neutral delay equations. including the geometry of the attractor. the LyapunovRazumikin method brieﬂy alluded to in the introductory lectures by Hale. Various aspects of the theory of delay diﬀerential equations are presented in this book. etc. a breakthrough accomplished during the 19851995 decade by MalletParet and coworkers is just mentioned in Walther’s course. which have been extensively described by Faria in her lectures. Arino discusses the issue of delay in models of epidemics. Delay diﬀerential equations have become a domain too wide for being covered in just one book. starting from the direct method. stochastic DDE. The lecture notes touched a variety of issues. as well as the integrated semigroup or the socalled sunstar approach. the semigroup approach. the Hopf and BogdanovTakens singularities. We might name many subjects which haven’t been or have just been brieﬂy mentioned in lectures notes: the second Lyapunov method for the study of stability. including the Cauchy problem. .
o` u les mouvements sont de nature r´eversible. Schmitt (1911) for references and some properties of linear equations). c’est` adire par l’´etat ` a un instant donn´e et a l’instant inﬁniment voison. ` Les ´etats ant´erieurs n’y intervenant pas. At the international conference of mathematicians.Chapter 1 HISTORY OF DELAY EQUATIONS J. Some of the early work originated from problems in geometry and number theory.K. The subject gained much momentum (especially in the Soviet Union) after 1940 due to the consideration of meaningful models of engineering 1 O. Delay Differential Equations and Applications. (eds. (1928) discussed the integrodiﬀerential equations that model viscoelasticity.edu Delay diﬀerential equations. sont donc inapplicables aux ˆetres vivants. qui apporteront la part de l’h´er´edit´e. USA hale@math. Volterra (1909). Picard (1908) made the following statement in which he emphasized the importance of the consideration of hereditary eﬀects in the modeling of physical systems: Les ´equations diﬀ´erentielles de la m´ecanique classique sont telles qu’il en r´esulte que le mouvement est d´etermin´e par la simple connaissance des positions et des vitesses.gatech. he wrote a fundamental book on the role of hereditary eﬀects on models for the interaction of species. In (1931). Nous pouvons r´ever d’´equations fonctionnelles plus compliqu´ees que les ´equations classiques parce qu’elles renfermeront en outre des int´egrales prises entre un temps pass´e tr`es ´eloign´e et le temps actuel. l’h´er´edit´e y est un vain mot.). diﬀerential integral equations and functional diﬀerential equations have been studied for at least 200 years (see E. © 2006 Springer. Hale Georgia Institute of Techology Atlanta. L’application de ces ´equations o` u le pass´e ne se distingue pas de l’avenir. Arino et al. . 1–28.
as well as many others (see Kolmanovski and Myshkis (1999)). it was natural to use the Laplace transform. Naturally. heat ﬂow. physiology. During the 1950’s. Halanay (1966). Krasovskii (1959). At the present time. For the linear equations with constant coeﬃcients. the mode of thought on FDE at the time was contrary to the new approach and sometimes not easily accepted. epidemiology. much of the subject can be considered as well developed as the corresponding one for ordinary diﬀerential equations (ODE). distributed networks. The genesis of the modern theory evolved from the consideration of the latter problem. played a very important role in the modern analytic and geometric theory of FDE. in my opinion. With our present knowledge of FDE. combustion.2 DELAY DIFFERENTIAL EQUATIONS systems and control. For the stability of equilibria. However. it was important to understand the extent to which one could apply the second method of Lyapunov (1891). During the last 50 years. neural networks. These books give a clear picture of the subject up to the early 1960’s. A new approach . It is probably true that most engineers were well aware of the fact that hereditary eﬀects occur in physical systems. In these lectures. nuclear reactors. but this eﬀect was often ignored because there was insuﬃcient theory to discuss such models in detail. I describe a few problems for which the method of solution. it is diﬃcult not to wonder why most of the early papers making connections between these two subjects were not written long ago. Bellman and Cooke (1963). This led to expansions of solutions in terms of the eigenfunctions and the convergence properties of these expansions. interaction of species. mechanics. microbiology. It took considerable time to take an idea from ODE and to ﬁnd the appropriate way to express this idea in FDE. learning models. the topics chosen are subjective and another person might have chosen completely diﬀerent ones. Most research on functional diﬀerential equations (FDE) dealt primarily with linear equations and the preservation of stability (or instability) of equilibria under small nonlinear perturbations when the linearization was stable (or unstable). there was considerable activity in the subject which led to important publications by Myshkis (1951). Stochastic eﬀects are also being considered but the theory is not as well developed. the theory of functional diﬀerential equations has been developed extensively and has become part of the vocabulary of researchers dealing with speciﬁc applications such as viscoelasticity.
0 attracts elements in U uniformly. Results concerning existence. 0) = 0 t→∞ that is. he would have written f (x(t + θ)) with the understanding that he meant a functional. there is a δ > 0 such that for any ϕ ∈ C with ϕ < δ. as well as the dependence on parameters. If r ≥ 0 is a given constant. be deﬁned by xt (θ) = x(t + θ). The point 0 is said to be a local attractor if there is a neighborhood U of 0 such that lim dist(x(t. ϕ) = ϕ(θ) for θ ∈ [−r. θ ∈ [−r. The point 0 is said to be stable if. 0]. then there is a solution x(t. We remark that the notation in (0. let C = C([−r. αϕ ). If f is C k . . α ). such that x0 (θ) = x(θ. let xt ∈ C. We begin the discussion with retarded functional diﬀerential equations (RFDE) with continuous initial data. a right hand derivative at t = 0 and satisﬁes (0. Furthermore. Rn ) and. t → αϕ ϕ 1. Following Lyapunov. α) → Rn . are essentially the same as for ODE with a few additional technicalites due to the inﬁnite dimensional character of the problem. for any ε > 0. α). then x(t. U ). The point 0 is asymptotically stable if it is stable and there is b > 0 such that ϕ < b implies that x(t. Stability of equilibria and Lyapunov functions One of the ﬁrst problems that occurs in diﬀerential equations is to obtain conditions for stability of equilibria. uniqueness and continuation of solutions. α). we have x(t. where in (0.3 History Of Delay Equations was necessary to obtain results which were diﬃcult if not impossible to obtain in the classical way. 0]. ϕ) < ε for t ≥ −r..1). If f is continuous and takes bounded sets into bounded sets. ϕ) of (0. If f : C → Rn is a given function. then the solution becomes unbounded as − .1) with initial value ϕ at t = 0 is a continuous function deﬁned on an interval [−r. ϕ) → 0 as t → ∞.1) for t ∈ [0. t ∈ [0. ϕ) through ϕ which exists on a maximal interval [−r. 0]. a retarded FDE (RFDE) is deﬁned by the relation x (t) = f (xt ) (0.1) is the modern one and essentially due to Krasovskii (1956). that is. a zero of f . α > 0.1) If ϕ ∈ C is given. then a solution x(t. Deﬁnition 1 Suppose that 0 is an equilibrium point of (0. if x : [−r.1). The notation above was introduced by Hale (1963). ϕ) is C k in ϕ in ([0. ϕ) has a continuous derivative on (0.) is C k+1 and x(. α). α). α > 0. x(t. if αϕ < ∞. . it is reasonable to make the following deﬁnition.
This is a consequence of the analyticity of (1.an eigenvalue with a positive real part). If the RFDE is nonlinear.)I). ϕ) < ε for t ≥ 0 (resp.2) have real parts < 0 (see Bellman and Cooke (1963) for detailed references). ∆(λ) = λI − f (exp −λ.2) may have inﬁnitely many solutions. Equation (1. it is necessary and suﬃcient to have each λ with real part < 0.. there is a solution of the form exp (λt) c for some nonzero nvector c if and only if λ satisﬁes the characteristic equation detD(λ) = 0.2) in λ and the fact that Reλ → −∞ if λ → ∞. (1. The eigenvalues play an important role in stability of linear systems. but there can be only a ﬁnite number in any vertical strip in the complex plane. For linear retarded equations. If the RFDE is nonlinear and 0 is an equilibrium and the zero solution of the linear variational equation is not asymptotically stable and there is no eigenvalue with positive real part. Proposition 1 An equilibrium point of (0.. if 0 is an equilibrium with all eigenvalues with negative real parts (resp. then classical approaches using a variation of constants formula and Gronwall type inequalities can be used to show that 0 is asymptotically stable (resp. asymptotically stable). then classical methods give no information.4 DELAY DIFFERENTIAL EQUATIONS In this deﬁnition. In this case. f : C → Rn a continuous linear functional. one can use essentially the same proof as for ODE to obtain the following important result. The veriﬁcation of this property in a particular example is far from trivial and much research in the 1940’s and 1950’s was devoted to giving various methods for determining when the λ satisfying (1.1). If there is an eigenvalue with positive real part. This is no restriction since. if 0 is stable (resp. xt (. then the origin is unstable. notice that the closeness of initial data is taken in C whereas the closeness of solutions is in Rn . unstable) for the nonlinear equation (see Bellman and Cooke (1963)). it is quite natural to attempt to adapt the well known methods of Lyapunov to RFDE. For asymptotic stability..1) is asymptotically stable if and only if it is a local attractor. Two independent approaches to this problem were given in the early 1950’s by Razumikhin (1956) and Krasovskii (1956). ϕ) → 0 as t → ∞). . that is. then xt (. As a consequence of the smoothing properties of solutions of (0.2) The numbers λ are called the eigenvalues of the linear equation.
We have seen that it belongs to this region. a > 0.4) is ≤ 0 for those functions with the property that ϕ≤ϕ(0). Let us indicate a few details. but is a function from C to R.1).5 History Of Delay Equations The approach of Razumikhin (1956) was to use Lyapunov functions on Rn . On the other hand. for any 0 < δ < ε.3) V (ϕ(0)) = ∂x . then the function V (ϕ(0)) is not a function from Rn to R. we can deﬁne V (x(t. there is a function ϕ with norm < δ and a time t1 > 0 such that x(t1 .4) is stable independent of the delay. If we chose V (x) = x2 /2. show that V (ϕ(0)) < 0 for a class of functions which at least includes functions with the property that ϕ > ϕ(0).4) is stable. Bellman and Cooke (1963) or Hale and Lunel (1993)). Of course. To illustrate how such conditions arise in a natural way. ϕ) = ε and x(t + θ. In this example.4) x (t) = −ax(t) − bx(t − r) where a > 0 and b are constants. a > 0. if 0 is not stable. then V is positive deﬁnite from R to R and (x(t. . As a consequence of this remark. ϕ)) and compute the derivative along the solution evaluated at ϕ(0) as . then we would know that the zero solution of (1.5) If we knew that the right hand side of (1.5) were ≤ 0. we cannot expect that the derivative in (1. It can be shown that it is suﬃcient to have the class satisfy ϕ≤pϕ(0) for some constant p > 1. 0). If the RFDE is not an ODE. 0 is stable if b≤a. b)space coincides with the region for which the origin of (1. The origin is not asymptotically stable if a + b = 0 since 0 is an eigenvalue. To use such a method. It is clear that this can be done if b≤a. Therefore. ϕ) < ε for all θ ∈ [−r. ∂V (ϕ(0)) f (ϕ) ≡ G(ϕ). (1. Is it possible to use the Lyapunov function V (x) = x2 /2 to prove this? For asymptotic stability. it is only necessary to ﬁnd conditions on b for which the right hand side of (1. for example. We remark that this region in (a. this can never be true for all functions in C in a neighborhood of zero.3) is negative for all small initial data ϕ. then there is an ε > 0 such that. one must consider a restricted set of initial data which is relevant for the consideration of stability. ϕ)) = −ax2 (t) − bx(t)x(t − r) V1. it is possible to show that all eigenvalues have negative real parts if b < a. As a consequence. ϕ) is a solution of (0. If V : Rn → R is a given continuously diﬀerentiable function and x(t. we must . consider the example (1.4 (1. but to show that it coincides with this region requires more eﬀort (see.
. v strictly increasing. then the point 0 is stable. At about the same time as Razumikhin. v(s) positive for s > 0. v. w:[0. ϕ ∈ C. 0]. Theorem 1 (Razumikhin) Suppose that u. if V (ϕ(θ))≤V (ϕ(0)). In addition. A simple computation shows that . ϕ)) − V (ϕ)]≤ − w(ϕ(0)) t→0 t then 0 is stable. The state space for FDE should be the space Csince this is the amount of information that is needed to determine a solution of the equation. 0]. If there is a continuous function V : C → R such that u(ϕ(0))≤V (ϕ)≤v(ϕ). R such that u(x)≤V (x)≤v(x). then 0 is asymptotically stable. V (ϕ) = ϕ (0) + µ 2 −r where µ is a positive constant. ∞) → [0. but.. it was not the way in which FDE were being discussed. θ ∈ [−r. V (ϕ) = −(a − µ)ϕ2 (0) − bϕ(0)ϕ(−r) − µϕ2 (−r). in addition.4) with 0 1 2 ϕ2 (θ)dθ. If there is a continuous function V :Rn → . ∞) → [0. v(s) positive for s > 0. (Krasovskii) Suppose that u. and V (ϕ(0))≤ w(ϕ(0)). The idea now seems quite simple. Let us apply the Theorem of Krasovskii to the example (1. u(s). at the time. if there is a continuous nondecreasing function p(s) > s for s > 0 such that . x ∈ Rn . u(0) = v(0) = 0. then 0 is asymptotically stable.6 DELAY DIFFERENTIAL EQUATIONS Razumikhin (1956) gave general results in the spirit of the above example to obtain suﬃcient conditions for stability and asymptotic stability using functions on Rn . u(s). If. Krasovskii (1956) was also discussing stability of equilibria and wanted to make sure that all of the results for ODE using Lyapunov functions could be carried over to RFDE. This is the observation made by Krasovskii (1956). θ ∈ [−r. 1 V (ϕ) = lim sup [V (xt (. We state a result on stability. w : [0. v. V (ϕ(0))≤w(ϕ(0)) if V (ϕ(θ))≤p(V (ϕ(0))). . ∞) are continuous nondecreasing functions. w(s) > 0 for s > 0. u(0) = v(0) = 0. He was then able to extend the complete theory of Lyapunov by using functionals V : C → R. ∞) are continuous nonnegative nondecreasing functions. Theorem 2 .
then {T (t). then it is asymptotically stable. Invariant Sets. 4(a − µ)µ ≥ b2 and positive deﬁniteness if the inequalities are replaced by strict inequalities. I personally had been thinking about delay equations in the 1950’s and reading the RAND report of Bellman and Danskin (1954). t. 2. This could not have been done without working in the phase space C and exploiting some properties of linear systems which will be mentioned later. that is T (0) = I. . My ﬁrst works were devoted to understanding the neighborhood of an equilibrium point (stable and unstable manifolds) and to deﬁning invariant sets in a way that could be useful for applications. t ≥ 0.. T (t + τ ) = T (t)T (τ ). The methods there did not seem to be appropriate for a general development of the subject. For more details. If we ﬁnd the region in parameter space for which this quadratic form is nonnegative. which gives the region of stability as a ≥ b and asymptotic stability as a > b. To obtain the largest region in the (a. generalizations and examples. Shimanov (1959) gave some interesting results on stability when the linearization has a zero eigenvalue. The condition for nonnegativeness is a ≥ µ > 0. we should choose µ = a/2. ϕ) through ϕ ∈ C at t = 0 is deﬁned for all t ≥ 0. ϕ). which is the same result as before using Razumikhin functions. the solution x(t. If we deﬁne the family of transformations T (t) : C → C by the relation T (t)ϕ = xt (.History Of Delay Equations 7 The right hand side of this equation is a quadratic form in (ϕ(0). Following this idea. then the origin is stable. see Hale and Lunel (1993). b) space for which these relations are satisﬁed. In 1959. Omegalimits and Lyapunov functionals. t ≥ 0} is a semigroup on C. τ ≥ 0. Kolmanovski and Myshkis (1999). In the present section. ϕ(−r)). For simplicity. let us suppose that. for every ϕ ∈ C. it was a revelation when Lefschetz gave me a copy of Krasovskii’s book (in Russian). it is best to describe invariance since the ﬁrst important application was related to stability. If it positive deﬁnite. I began to work very hard to try to obtain interesting results for RFDE on concepts which were well known to be important for ODE. The suggestion made by Krasovskii that one should exploit the fact that the natural state space for RFDE should be C opened up the possibility of obtaining a theory of RFDE which would be as general as that available for ODE.
Ua : V (ϕ) = 0} and M is the maximal invariant set in Wa . there is a preimage and. n = 1. it is possible to deﬁne negative orbits through ϕ. If Ua is connected. if T (t)A = A. for each ϕ.} ⊂ B and tn → ∞ as n → ∞ such that T (tn )ϕn → ψ as n → ∞. is ω(ϕ) = ∩ Cl(γ + (T (τ )ϕ)). This is easily seen from the ODE x = x − x3 . Also. 2. The following result is consequence of the fact that T (r) is completely continuous. ϕ∈B x ∈ R. the solution x(t.1). for any ϕ ∈ Ua for which γ + (ϕ) is bounded. so is M . Wa = {ϕ ∈ . C with V (ϕ)≤0 for all ϕ ∈ C. A function ψ ∈ ω(B) if and only if there exist sequences {ϕn . The following result is a natural generalization of the classical LaSalle invariance principle for ODE. If A is invariant. then ω(B) is a indexxcompact invariant set which attracts B under the ﬂow deﬁned by (0. A set A in C attracts a set B in C if dist(T (t)B. This is not possible for all ϕ ∈ C since a solution of (0. The ω−limit set of ϕ ∈ C. for t ≥ r.8 DELAY DIFFERENTIAL EQUATIONS The smoothness of T (t)ϕ in ϕ is the same as the smoothness of f and. A) → 0 as t → ∞. T (t) is a completely continuous operator since. then M = ω(Ua ) is compact invariant and attracts Ua under the ﬂow deﬁned by (0. Theorem 4 : (Hale. 1963) Let V be a continuous scalar function on . then. denoted by ω(B).1) becomes continuously diﬀerentiable for t ≥ r. . .1) and is connected if B is connected. there may not be a unique negative orbit through ϕ ∈ A. ϕ) is diﬀerentiable for t ≥ r.. Theorem 3 : If B ⊂ C is such that γ + (B) is bounded. τ ≥0 Note that a function ψ ∈ ω(ϕ) if and only if there is a sequence tn → ∞ as n → ∞ such that T (tn )ϕ → ψ as n → ∞. We remark that ω(B) may not be equal to ∪ ω(ϕ). t ≥ 0}. is ω(B) = ∩ Clγ + (T (τ )B)). then. τ ≥0 The ω−limit set of a subset B in C. A set A ⊂ C is an invariant set for T (t). we have ω(ϕ) ⊂ M . for any ϕ ∈ A. denoted by ω(ϕ).. If Ua is a bounded set. t ≥ 0. Deﬁnition 2 :The positive orbit γ + (ϕ) through ϕ is the set {T (t)ϕ. t ≥ 0. thus. If Ua = {ϕ ∈ C : V (ϕ)≤a}.
we need to investigate the set where V = 0. (LevinNohel) 1). we will use their clever choice of a Lyapunov function.a ≥ 0. a(r) = 0.9 History Of Delay Equations To the author’s knowledge. To do this. a≤ 0. ω(ϕ) is an rperoidic orbit in C deﬁned by an rperiodic solution of the ODE y + g(y) = 0.. invariant set in the set where V = 0 coincides with those bounded solutions of the equation y + g(y) = 0. Theorem 5 :. then. we observe that any solution of (2.6) is given by 0 0 . V (ϕ) = G(ϕ(0)) − 2 −r θ then the derivative of V along the solutions of ( 2. If a is not a linear function. for any ϕ ∈ C. We outline the proof of Hale (1963) which is independent of the one given by Levin and Nohel. ω(ϕ) is a zero of g. a ≥ 0. x (t) + a (0)g(x(t)) = −a (r) 0 −r g(ϕ(s))ds + 0 −r 0 a (−θ) g(ϕ(s))ds dθ. 1 1 0 V (ϕ) = a (r)[ g(ϕ(s))ds]2 − a (−θ) g(ϕ(s))ds]2 dθ. . for any ϕ ∈ C. If a is a linear function. we see that the largest . . However. The hypotheses imply that V (ϕ)≤0 and γ + (ϕ) is bounded for each ϕ ∈ C. θ . With this relation and the fact that V (ϕ)≤0. 2 2 −r −r θ .6) x (t) = − r −r where the zeros of g are isolated and x G(x) = g(s)ds → ∞ as x → ∞ 0 . If 0 2 1 0 a (−θ) g(ϕ(s))ds dθ. 2). To apply the above theorem. then. these concepts for RFDE were frist introduced in Hale (1963) and were used to give a simple proof of convergence for the LevinNohel equation 1 0 a(−θ)g(x(t + θ))dθ (2.6) also must satisfy the equation .
If a is a linear function.10 DELAY DIFFERENTIAL EQUATIONS which satisfy the property that 0 g(y(t + θ))dθ = 0. This shows that ω(ϕ) belongs to the set of periodic orbits generated by rperiodic solutions of the ODE. The set C−1 = {ϕ ∈ C : ϕ(0) = −1} is the translation of a codimension one subspace of C and is invariant under the ﬂow deﬁned by (3. any solution with initial data in C−1 is equal to the constant function −1 for t ≥ 1. Jones (1962) gave a procedure for determining the existence of a periodic solution of delay diﬀerential equations which has become a standard tool in this area. this implies that y is periodic of period r and there is a constant k such that y(t) = kt+(a periodic function of period r). An equation was also encountered in prime number theory by Wright (1955) which had the same property. The linear variational equation about x = −1 has only the eigenvalue α > 0 and is. then there is an s0 such that a (s0 ) = 0 and there is an interval Is0 of s0 such that a (s) = 0 for s ∈ Is0 . t ∈ R. In his study of the control of the motion of a ship with movable ballast. Furthermore. t ∈ R. Delays may cause instability. we have the conclusion in part 1 of the theorem. 3. The equation (3.7). we observe that y is periodic of period s for every s ∈ .7) If a is not a linear function. if a (r) = 0.8). therefore. Integrating the second order ODE from −s to 0 and using (2.8) where α > 0 is a constant.8) has two equilibria x = 0 and x = −1. . unstable. I describe this for the equation of Wright x (t) = −αx(t − 1)(1 + x(t)). Minorsky (1941) (see also Minorsky (1962)) made a realistic mathematical model which contained a delay (representing the time for the readjustment of the ballast) and observed that the motion was oscillatory if the delay was too large. if a (s) = 0 (2. 0 then we must be concerned with the solutions of the ODE for which −r g(y(t+θ))dθ = 0. To prove that the ω−limit set is a single periodic orbit requires an argument using techniques in ODE which we omit. It was many years later that S. −r 0 −s g(y(t + θ))dθ = 0. Since the zeros of g are isolated and ω(ϕ) is connected.Is0 . y is a constant and ω(ϕ) belongs to the set of zeros of g. Boundedness of y implies that y(t) is rperiodic. Thus. (3. As before.
ϕ). If ϕ ∈ C+ . Numerical computations suggested that there should be such a solution with simple . λ + α exp(−λ) = 0. As a consequence of these remarks. As a consequence of some later remarks.8) are the solutions of the equation. 2 π It is reasonable to conjecture that A = {0} for 0 < α < . then the solution x(t) = −1 for all t ≥ 0 for which it is deﬁned. then. equation ( 3. there is the following interesting result. The sets C− = {ϕ ∈ C : ϕ(0) < −1} and C+ = {ϕ ∈ C : ϕ(0) > −1} are positively invariant under the ﬂow deﬁned by (3.8) through a point ϕ ∈ C with ϕ(0) = −1. We indicate the ideas in the original proof of Jones (1962). for each ϕ ∈ C+ . For 2 π α > . t−1 x(s)ds} (3.9) 1 + x(t) = [1 + x(t0 )] exp(−α t0 −1 This proves the assertion.11 History Of Delay Equations The set C−1 serves as a natural barrier for the solutions of (3.18). there is a unique pair of eigenvalues with maximal real part > 0. t ≥ 0. α). invariant and attracts bounded sets of C. each solution exists for all t ≥ 0 and (3. A is compact.. if x(t) is a solution of (3.8).8) deﬁnes a semigroup Tα (t) : C+ → C+ . then (3. for any t0 for which x(t0 ) = −1 and any t ≥ t0 . ϕ) → −∞ as t → ∞. It is possible to show that. The eigenvalues have negative π real parts for 0 < α < . a pair of purely imaginary eigenvalues for 2 π α = with the remaining ones having negative real parts and. In fact. where Tα (t)ϕ = xt (.9) shows that a solution x(t) cannot become unbounded on a ﬁnite interval. that is. On the other hand. The eigenvalues of the linearization about zero of (3. π Theorem 6 : (Jones (1962)) If α > . It also is clear that Tα (t) is a bounded map for each t ≥ 0.8) has a periodic 2 solution oscillating about 0 and with the property that the distance between zeros is greater than the delay. Wright (1955) proved that every solution approaches zero as t → ∞ if 0 < α < exp(−1) and this implies that {0} is the compact global attractor for (3. but this has 2 not been proved.8).8). this implies the existence of the compact global attractor A of (3. Yorke (1970) extended this result (even for more general equations) to the interval 0 < α < 3/2. there is a t0 (ϕ. we discuss this equation in the subset C+ of C. If ϕ ∈ C− . it is not diﬃcult to show that x(t. α) such that Tα (t)ϕ≤ exp α − 1 for t ≥ t0 (ϕ. In fact. if a solution has initial data ϕ for which ϕ(0) = −1.
The main diﬃculty in the proof of the theorem is that 0 ∈ K is a ﬁxed point of Pα and one wants a nontrivial ﬁxed point. Such an approach limits the extent to which one can obtain many of the interesting properties that are similar to the ones for ODE. A nontrivial ﬁxed point of Pα yields a periodic solution of (3. 0]} Tα (t1 )ϕ ∈ K. the spectrum is only point spectrum plus possibly 0. Many reﬁnements have been made of this method using interesting ejective ﬁxed point theorems (which were discovered because of this problem) (see Browder (1965). Let K ⊂ C be deﬁned as K = Cl{nondecreasing functions ϕ ∈ C: ϕ(−1) = 0. One can show that this map is completely continuous. The inﬁnitesimal generator has compact resolvent with only point spectrum.18) with the properties stated in the above theorem.12 DELAY DIFFERENTIAL EQUATIONS zeros and the distance from a zero to the next maximum (or minimum) is ≥ the delay. for example. most researchers discussed linear autonomous equations and perturbations of such equations by considering properties of the solution x(t. and even after. van Giles. The complete theory of the linear case was developed by Hale (1963) (see also Shimanov (1965)). ϕ(θ) > 0. Nussbaum (1974)). Grafton (1969) showed that the use of unstable manifold theory could be of assistance in the veriﬁcation of ejectivity. Prior to Krasovskii. The point 0 ∈ K is unstable and Jones was able to use this to show eventually that he could obtain the desired ﬁxed point from Schauder’s ﬁxed point theorem. Lunel and Walther (1991). The point spectrum of the generator determines the point spectrum of the semigroup by exponentiation. ϕ) in Rn . the point 0 is ejective. Shimanov (1959) exploited this fact to discuss the stability of an equilibrium point for a nonlinear equation for which the linear variational equation had a simple zero eigenvalue and the remaining ones had negative real parts. we are content to give only an indication of the results with a few applications. See. This deﬁnes a Poincar´e map Pα :ϕ ∈ K maps to Tα (t1 )ϕ ∈ K if we deﬁne Pα 0 = 0. θ ∈ ( −1. This suggests a decomposition theory into invariant subspaces similar to ones used for ODE. a linear autonomous equation generates a C 0 semigroup on C which is compact for t ≥ r. Linear autonomous equations and perturbations. Diekmann. In the approach of Krasovskii. . Therefore. In the above problem. Hale and Lunel (1993). 4. Since linear equations are to be discussed in detail in this summer workshop.
13
History Of Delay Equations
Consider the equation
x (t) = Lxt ,
(4.10)
where L : C → Rn is a bounded linear operator. Equation (4.10) generates a C 0 semigroup TL (t), t ≥ 0, on C which is completely continuous
for t ≥ r. Therefore, the spectrum σ(TL (t)) consists only of point spectrum σp (TL (t)) plus possible zero.
The inﬁnitesimal generator AL is easily shown to be given by
D(AL ) = {ϕ ∈ C 1 ([−r, 0], Rn ) : ϕ (0) = Lϕ}, AL ϕ = ϕ
(4.11)
The operator AL has compact resolvent and the spectrum is given
σ(AL ) = {λ : (λ) = 0}, (λ) = λI − L exp(λ.)I}.
(4.12)
In any vertical strip in the complex plane, there are only a ﬁnite number of elements of σ(AL ). If λ ∈ σ(AL ), then the generalized eigenspace
Mλ of λ is ﬁnite dimensional, say of dimension dλ . If φλ = (ϕ1 , ......., ϕdλ )
is a basis for Mλ , then there is a dλ × dλ matrix Bλ such that
AL φ = φBλ , [φ] = Mλ ,
where [φ] denotes span.
From the deﬁnition of AL , it is easily shown that Mλ is invariant
under TL (t) and
TL (t)φ = φ exp (Bλ t) , t ∈ R.
There is a complementary subspace Mλ⊥ ⊂ C of Mλ such that
C = Mλ ⊕ Mλ⊥ , TL (t)Mλ⊥ ⊂ Mλ⊥ , t ≥ 0,
(4.13)
Of course, such a decompostion can be applied to any ﬁnite set of elements of σ(AL ).
In applications of this decompostion theory, it is necessary to have a
speciﬁc computational method to construct the complementary suspace.
This was done in detail by using an equation which is the ‘adjoint’ of
(4.10) (see, for example, Hale (1963), Shimanov (1965), Hale (1977),
Hale and Lunel (1993), Diekmann, van Giles, Lunel and Walther (1991)).
Consider now a perturbed linear system
x (t) = Lxt + f (t),
(4.14)
where f : R → Rn is a continuous function. In ODE, the variation
of constants formula plays a very important role in understanding the
eﬀects of f on the dynamics. For RFDE, a variation of constants formula
is implicitly stated in Bellman and Cooke (1963), Halanay (1965). It is
14
DELAY DIFFERENTIAL EQUATIONS
not diﬃcult to show that there is a solution X(t) = X(t, X0 ) of (4.10)
through the n × n discontinuous matrix function X0 deﬁned by
X0 (θ) = 0, θ ∈ [−r, 0) ,X0 (0) = I, θ = 0.
(4.15)
With this function X(t) and deﬁning Xt = T (t)X0 , it can be shown that
the solution x(t) = x(t, ϕ) of (4.14) through ϕ is given by
t
xt = T (t)ϕ +
T (t − s)X0 f (s)ds, t ≥ 0.
(4.16)
0
The equation (4.16) is not a Banach integral. For each θ ∈ [−r, 0], the
equation (4.16) is to be interpreted as an integral equation in Rn .
The equation (4.16), interpreted in the above way, was used in the
development of many of the ﬁrst fundamental results in RFDE (see, for
example, Hale (1977)). The Banach space version of the variation of
constants formula makes use of sunreﬂexive spaces and there also is a
theory based on integrated semigroups (see Diekmann, van Giles, Lunel
and Walther (1991) for a discussion and references).
With (4.16) and the above decomposition theory, it is possible to make
a decomposition in the variation of constants formula. We outline the
procedure and the reader may consult Hale (1963), (1977) or Hale and
Lunel (1993) for details.
Let x(t) = x(t, ϕ) be a solution of (4.14) with initial value ϕ at t = 0,
choose an element λ ∈ σ(AL ) and make the decomposition as in (4.13),
letting
⊥
⊥
(4.17)
ϕ = ϕλ + ϕ⊥
λ X0 = X0,λ + X0,λ , xt = xt,λ + xt,λ
The decomposition on X0 needs to be and can be justiﬁed.
If we apply this decomposition to (4.16), we obtain the equations
t
xt,λ = TL (t)ϕλ +
T (t − s)X0,λ f (s)ds
0
⊥
x⊥
t,λ = TL (t)ϕλ +
t
0
⊥
T (t − s)X0,λ
f (s)ds
(4.18)
If we use the fact that Tλ (t)φλ = φ exp(Bλ t) and let xt,λ = φy(t), and,
for simplicity in notation, x⊥
t,λ = zt , then we have
y(t) = exp(Bλ t)a +
t
exp Bλ (t − s)Kf (s)ds
0
zt = TL (t)z0 +
0
t
⊥
T (t − s)X0,λ
f (s)ds
(4.19)
15
History Of Delay Equations
where a = y(0) and X0,λ = φK with K being a dλ × dλ matrix. The
ﬁrst equation in (4.19) is equivalent to an ODE
y = Bλ y + Kf (t)
(4.20)
with y(0) = a.
At the time that this decomposition in C was given for the nonhomogeneous equation, it was not readily accepted by many people that
were working on RFDE. The main reason was that we now have the solution expressed as two variable functions φy(t) and zt of t and, if f
= 0,
then neither of these functions can be represented by functions of t + θ,
θ ∈ [−r, 0]. Therefore, neither function can satisfy an RFDE. Their sum
yields a function xt which does satisfy this property.
In spite of the apparent discrepancy, it is precisely this type of decomposition that permits us to obtain qualitative results similar to the
ones in ODE. The ﬁrst such result was given by Hale and Perello (1964)
when they deﬁned the stable and unstable sets for an equilibrium point
and proved the existence and regularity of the local stable and unstable
manifolds of an equilibrium for which no eigenvalues lie on the imaginary axis; that is, they proved that the saddle point property was valid
for hyperbolic equilibria. The proof involved using Lyapunov type integrals to obtain each of these manifolds as graphs in C (see, for example,
Hale (1977), Hale and Lunel (1993) or Diekmann, van Giles, Lunel and
Walther (1991)). We now know that further extensions give a more complete description of the neighborhood of an equilibrium including center
manifolds, foliations, etc.
Another situation that was of considerable interest in the late 1950’s
and 1960’s was the consideration of perturbed systems of the form
x (t) = Lxt + M (t, xt ),
(4.21)
where, for example, M (t, ϕ) is continuous and linear in ϕ and there is a
function a(t) such that
M (t, ϕ)≤a(t)ϕ, t ≥ 0,
(4.22)
and the function ais small in some sense. The problem is to determine
conditions on ato ensure that the behavior of solutions of (4.21) are
similar to the solutions of (4.10).
For example, Bellman and Cooke (1959) considered the following
problem. Suppose that xis a scalar, λ is an eigenvalue of (4.10)and there
are no other eigenvalues with the same real parts. Determine conditions
on aso that there is a solution of (4.21) which asymptotically as t → ∞
behaves as the solution exp λt of (4.10). Their approach was to replace
16
DELAY DIFFERENTIAL EQUATIONS
a solution x(t) of (4.14) by a function w(t) ∈ Rn by the transformation
x(t) = (exp λt) w(t) in Rn . In this case, the function w(t) will satisfy a
RFDE and the problem is to show
∞ that w(t) → 0 as t → ∞. In the case
where a(t) → 0 as t → ∞ and 0 a(t)dt may be ∞, it was necessary
to impose several additional conditions on a to ensure that w(t) → 0 as
t → ∞. Some of these conditions seemed to be artiﬁcial and were not
needed for ODE.
Another way to solve this type of problem is to make the transformation in C given by xt = exp(λt)zt and determine conditions on zt so that
zt → 0 in C as t → ∞. In this case, the function zt does not satisfy an
RFDE. On the other hand, some of the unnatural conditions imposed
by Bellman an Cooke (1959) were shown to be unnecessary (see Hale
(1966)).
5.
Neutral Functional Diﬀerential Equations
Neutral functional diﬀerential equations have the form
x (t) = f (xt , xt );
(5.23)
that is, x (t) depends not only upon the past history of x but also on
the past history of x . Let X, Y be Banach spaces of functions mapping
the interval [−r, 0] into Rn . We say that x(t, ϕ, ψ) is a solution of (5.23)
through (ϕ, ψ) ∈ X × Y , if x(t, ϕ, ψ) is deﬁned on an interval [−r, α),
α > 0, x0 (., ϕ, ψ) = ϕ, x0 (., ϕ, ψ) = ψ and satisﬁes (5.23) in a sense
consistent with the deﬁnitions of the spaces X and Y .
For example, if X = C([−r, 0], Rn ), Y = L([−r, 0], Rn ), then the
solution should be continuous with an integrable derivative and satisfy
(5.23) almost everywhere.
The proper function spaces depend upon the class of functions f that
are being considered. For a general discussion of this point, see Kolmanovskii and Myskis (1999).
From my personal point of view, it is important to isolate classes
of neutral equations for which it is possible to obtain a theory which
is as complete as the one for the RFDE considered above. Hale and
Meyer (1967) introduced such a general class of neutral equations which
were motivated by transport problems (for example, lossless transmission lines with nonlinear boundary conditions) and could be considered
as a natural generalization of RFDE in the space C. We describe the
class and state a few of the important results. The detailed study of
this class of equations required several new concepts and methods which
have been used in the study of other evolutionary equations (with or
without hereditary eﬀects); for example, damped hyperbolic partial dif
17
History Of Delay Equations
ferential equations and other partial diﬀerential equations of engineering
and physics.
Suppose that D, f : C → Rn are continuous and D is linear and
atomic at 0; that is,
0
[dµ(θ)]ϕ(θ)
Dϕ = ϕ(0) −
−r
and the measure µ has zero atom at 0. The equation
d
Dxt = f (xt )
dt
(5.24)
is called a neutral FDE (NFDE). If Dϕ = ϕ(0), we have RFDE.
For any ϕ ∈ C, a function x(t, ϕ) is said to be a solution of (5.24)
through ϕ at t = 0 if it is deﬁned and continuous on an interval [−r, α),
α > 0, x0 (., ϕ) = ϕ, the function Dxt (., ϕ) is continuously diﬀerentiable
on (0, α) with a right hand derivative at t = 0 and satisﬁes (5.24) on
[0, α].
It is important to note that the function D(xt (., ϕ)) is required to be
diﬀerentiable and not the function x(t, ϕ).
Except for a few technical considerations, the basic theory of existence, uniqueness, continuation, continuous dependence on parameters,
etc. are essentially the same as for RFDE.
As before, if we let TD,f (t)ϕ = xt (., ϕ) and suppose that all solutions
are deﬁned for all t ≥ 0, then TD,f (t), t ≥ 0, is a semigroup on C with
TD,f (t)ϕ a C k function in ϕ if f is a C k function.
It is easily shown that the inﬁnitesimal generator AD,f of T (t) is given
by
D(AD,f ) = {ϕ ∈ C 1 ([−r, 0], Rn ) : D(ϕ ) = f (ϕ), AD,f ϕ = ϕ . (5.25)
Hale and Meyer (1967) considered (5.24) when the function f was linear and were interested in the determination of conditions for stability
of the origin with these conditions being based upon properties of the
generator AD,f . It was shown that, if the spectrum of AD,f was uniformly bounded away from the imaginary axis and in the left half of the
complex plane, then one could obtain uniform exponential decay rate for
solutions provided that the initial data belonged to the domain of AD,f .
Of course, this should not be the best result and one should get these
uniform decay rates for any initial data in C. This was proved by Cruz
and Hale (1971), Corollary 4.1, and a more reﬁned result was given by
Henry (1974).
Due to the fact that the solutions of a RFDE are diﬀerentiable for
t ≥ r, the corresponding semigroup is completely continuous for t ≥ r.
18
DELAY DIFFERENTIAL EQUATIONS
Such a nice property cannot hold for a general NFDE since the solutions
have the same smoothness as the initial data. On the other hand, there is
a representation of the semigroup for a NFDE as a completely continuous
perturbation of the semigroup generated by a diﬀerence equation related
to the operator D.
To describe this in detail, suppose that
Dϕ = D0 ϕ +
0
−r
B(θ)ϕ(θ)dθ, D0 ϕ =
∞
Bk ϕ(−ρk )ϕ(−rk )
(5.26)
k=0
where B(θ) is a continuous n × n matrix, Bk , k = 0, 1, 2,...... is an n × n
constant matrix, r0 = 0, 0 < rk ≤r, k = 1, 2, ......
Let CD0 be the linear subspace of C deﬁned by
CD0 = {ϕ ∈ C : D0 ϕ = 0}
and let TD0 (t), t ≥ 0, be the semigroup on CD0 deﬁned by TD0 (t) =
yt (., ϕ), where y(t, ϕ) is the solution of the diﬀerence equation
D0 yt = 0, y0 = ϕ ∈ CD0 .
(5.27)
Theorem 7 : (Representation of solution operator}) There are a
bounded linear operator ψ:C → CD0 and a completely continuous operator UD,f (t) : C → C, t ≥ 0, such that
TD,f (t) = TD0 (t)ψ + UD,f (t), t ≥ 0.
(5.28)
Using results in Cruz and Hale (1971), the representation theorem
was proved in Hale (1970) for the situation where Dis an exponentially
stable operator. The proof in that paper clearly shows that it is only
required that D0 is an exponentially stable operator. Henry (1974) also
gave such a representation of the semigroup for (5.24).
In the statement of the above theorem, there is the mysterious linear
operator ψ : C → CD0 . Let us indicate how it is constructed. For RFDE,
D0 ϕ = ϕ(0) and ψϕ = ϕ − ϕ(0), which is just a translation of the initial
function in order to have ψϕ ∈ CD0 . For general D0 , one ﬁrst chooses
a matrix function φ = (ϕ1 , ......, ϕn ), ϕj ∈ C, j ≥ 1, so that D0 (φ) = I,
the identity and then set ψ = I − φD0 .
The representation (5.28) shows that the essential spectrum
σess (TD,f (t)) of the operator TD,f (t) on C coincides with the essential
spectrum σess (TD0 (t)) of TD0 (t) on CD0 . It can be shown (see Henry
(1987)) that the spectrum σ(TD0 (t)) of TD0 (t) coincides with its essen
19
History Of Delay Equations
tial spectrum and that
σess (TD0 (t))=Cl exp(λt) : det∆D0 (λ)=0; ∆D0 (λ)=
∞
Bk exp(λrk )
k=1
(5.29)
We say that the operator D0 is exponentially stable if σ(TD0 (t)) for t > 0
is inside the unit circle with center zero in the complex plane. In this
case, there are constants KD0 ≥ 1, αD0 > 0 such that
TD0 (t)ϕ ≤ KD0 exp −αD0 t , t ≥ 0, ϕ ∈ CD0 .
(5.30)
As a consequence, there is a constant K1 such that
TD0 (t)ψϕ ≤ KD0 K1 exp −αD0 t, t ≥ 0, ϕ ∈ C.
(5.31)
We remark that, if D0 is exponentially stable, it is possible to ﬁnd an
equivalent norm in C such that KD0 K1 = 1; that is, TD0 (t)ψ is a strict
contraction for each t > 0. In this norm, relation (5.28) implies that
TD,f (t) is the sum of a strict contraction and a completely continuous
operator for each t > 0.
For the RFDE, Dϕ = D0 ϕ = ϕ(0) and D0 is exponential stable since
the solutions of the diﬀerence equation x(t) = 0 on {ϕ ∈ C : ϕ(0) = 0} is
identically zero for t ≥ r and has the spectrum of the semigroup consisting only of the point 0. This fact implies that the semigroup is compact
for t ≥ r as we have noted before. The representation (5.28) says more
since it gives properties of the semigroup on the interval [0, r] as the
sum of an exponentially decaying semigroup and a completely continuous semigroup. We make an application of this later when we discuss
periodic solutions of time varying systems with periodic dependence on
time.
If D0 ϕ = ϕ(0) − aϕ(−r), then D0 is exponentially stable if and only
if a < 1 and ress (TD0 (t)) = at .
Stability of an equilibrium point of a NFDE is deﬁned in the same
way as for RFDE. We remark that, for general NFDE, an equilibrium
point being asymptotically stable does not imply that the equilibrium
point is a local attractor. In fact, it is possible to have a linear system
d
Dxt = Lxt
dt
for which every solution approaches zero as t → ∞, 0 is stable and is
not a local attractor. In this case, the essential spectral radius of the
semigroup (and therefore of TD0 (t)) must be equal to one for all t ≥ 0.
In fact, if there is a t1 > 0 such that it is less that one, then the above
20
DELAY DIFFERENTIAL EQUATIONS
representation formula implies that, if the solutions approach zero as
t → ∞, then the spectrum of the semigroup for t > 0 must be inside the
unit circle and 0 would be a local attractor.
The example
d
[x(t) − ax(t − 1)] = −cx(t)
dt
with a = 1 and c > 0 has the property just stated. It is an interesting
exercise to verify this fact.
Much of the theory for RFDE has been carried over to NFDE if the
operator D0 is exponentially stable. However, there are many aspects
that have not yet been explored (see Hale and Lunel (1993), Hale, Magalhaes and Oliva (2002)).
6.
Periodically forced systems and discrete
dynamical systems.
If X is a Banach space and T : X → X is a continuous map, we obtain
a discrete dynamical system by considering the iterates T n , n ≥ 0, of the
map. Positive orbits, ω−limit sets, invariance, etc. are deﬁned the same
way as for continuous semigroups.
In the context of the present lectures, we can obtain a discrete dynamical system in the following way. If the RFDE or NFDE is nonautonomous with the time dependence being τ periodic, we can deﬁne the
Poincar´e map π : C → C which takes the initial data ϕ ∈ C to the solution xt (., ϕ) at time τ . The map π deﬁnes a discrete dynamical system.
Fixed points of π correspond to periodic solutions of the equation with
the same period as the forcing.
In his study of the periodically forced van der Pol equation, Levinson
(1944) attempted to determine the existence of a periodic solution of
the same period as the forcing. To do this, he introduced the concept of
point dissipativeness described below and was able to use the Brouwer
ﬁxed point theorem to prove that some iterate of the Poincar´e map had
a ﬁxed point, but he could not prove that the map itsef had a ﬁxed
point.
Massera (1950) gave an example of a 2dimensional ODE with τ periodic coeﬃcients which had a 2τ periodic solution, but did not have
a τ periodic solution. In addition, all solutions were bounded.
This shows that dissipation is necessary if the diﬀerential system of
Levinson is to have a τ periodic solution. There was a successful solution
to this problem which we describe in the next section.
21
History Of Delay Equations
7.
Dissipation, maximal compact invariant sets
and attractors.
We have seen above that dissipation can perhaps be beneﬁcial. In
many important applications, there is dissipation which forces solutions
with large initial data away from inﬁnity; that is, ‘inﬁnity is unstable.’
It is important to understand the implications of such a concept for both
autonomous and nonautonomous problems.
In this section, we introduce the new concepts that are needed for
continuous dynamical systems deﬁned by a C 0 semigroup T (t), t ≥ 0,
as well as discrete dynamical systems {T n , n = 0, 1, 2, ........} deﬁned by
a map T on a Banach space X. To write things in a uniﬁed way, we let
+ denote either the interval (−∞, ∞) or the set Z = {0, ±1, ±2, .....},
+ (resp.
− ) the nonnegative (resp. nonpositive) subsets of
. We
can now write the continuous and discrete dynamical systems with the
notation T (t), t ∈
+ .
Deﬁnition 3 : (Dissipativeness). The semigroup T (t), t ∈
+ is said
to be point dissipative (resp. compact dissipative) (resp. bounded dissipative) if there is a bounded set B in X such that, for any ϕ ∈ X
(resp. compact set K in X) (resp. bounded set U in X), there is a
t0 = t0 (B, ϕ) ∈
+ (resp. t0 = t0 (B, K)) (resp. t0 = t0 (B, U )) such that
T (t)ϕ ∈ B (resp. T(t)K ⊂ B) (resp. T (t)U ⊂ B) for t ≥ t0 , t ∈
+ .
As remarked earlier, point dissipativeness was introduced by Levinson (1944). In ﬁnite dimensional space, all of the above concepts of
dissipativeness are the same.
Deﬁnition 4 : Maximal compact invariant set A set A ⊂ X is said to
be the maximal compact invariant set for the dynamical system T (t), t ∈
G+ , if it is compact invariant and maximal with respect to this property.
Deﬁnition 5 : Compact global attractor A set A ⊂ X is said to be the
compact global attractor if it is compact invariant and for any bounded
set B ⊂ X, we have
lim
dist(T (t)B, A) = 0.
t∈G+ ,t→∞
For ODE in Rn , Pliss (1966) proved that point dissipativeness implied the existence of a compact global attractor. Using this fact and an
asymptotic ﬁxed point theorem of Browder (1959) for completely continuous maps, he was able to prove that a dissipaative nonautonomous
ODE which is τ periodic in thas a τ periodic solution; thus, answering in the aﬃrmative the problem of Levinson (1944). For RFDE with
22
DELAY DIFFERENTIAL EQUATIONS
the period τ greater than the delay, Jones (1965) and Yoshizawa (1966)
used the same asymptotic ﬁxed point theorem to prove the existence
of a τ periodic solution. Recall that τ greater than the delay makes
the Poincar´e map completely continuous. In the interval [0, r], it is not
completely continuous, but we have seen above that it is the sum of
a contraction and a completely continuous operator. We give an abstract ﬁxed point theorem later which will allow us to conclude that
the Poincar´e map has a ﬁxed point and, therefore, there is a τ periodic
solution, which generalizes the result of Pliss (1966) to RFDE.
In a fundamental paper, Billoti and LaSalle (1971) proved the existence of a compact 2 global attractor if T (t), t ∈
+ is point dissipative
and there is a t1 ∈
for which T (t1 ) is completely continuous. In the
discrete case, this implied that there is a periodic point of period t1 .
Hale, LaSalle and Slemrod (1973) extended these results to the class
of asymptotically smooth dynamical systems deﬁned below. This class of
dynamical systems includes NFDE with an exponentially stable D0 operator as well as many other dynamical systems, including those deﬁned
by many dissipative partial diﬀerential equations.
Deﬁnition 6 : (Asymptotically smooth) A dynamical system T (t), t ∈
+ , on a Banach space X is said to be asymptotically smooth if, for any
bounded set B in X for which T (t)B ⊂ B for t ∈
+ , there is a compact
set J in X such that
lim
dist(T (t)B, J) = 0.
t∈G+ ,t→∞
This deﬁnition stated in a diﬀerent but equivalent way is due to Hale,
LaSalle and Slemrod (1973). It is very important to note that a dynamical system can be asymptotically smooth and there can be a bounded
set for which the positive orbit is unbounded; for example, the dynamical system deﬁned by the ODE x = x. The deﬁnition contains the
conditional expression ‘if T (t)B ⊂ B, t ∈
+ .’
Proposition 2 :. If T (t), t ∈
+ , is asymptotically smooth and B is
a bounded set for which there is a t1 ∈
+ such that γ + (T (t1 )B) is
bounded, then ω(B) is a compact invariant set. It is connected for continuous dynamical systems if B is connected.
It is obvious that, if there is a t1 ∈
+ , t1 > 0 such that T (t1 ) is
completely continuous, then T (t), t ∈
+ , is asymptotically smooth. it
also is not diﬃcult to prove that, if T (t) = S(t) + U (t) where U (t) is
completely continuous for all t ∈
+ , and S(t) is a linear semigroup with
spectral radius α(t) → 0 as t → ∞, then T (t), t ∈
+ , is asymptotically
smooth.
History Of Delay Equations
23
In particular, the semigroup associated with RFDE and NFDE with
D0 exponentially stable are asymptotically smooth. For similar equations with τ periodicity in time, the Poincar´e map is asymptotically
smooth.
The following result is essentially due to Hale, LaSalle and Slemrod
(1973) with some minor reﬁnements in its statement.
Theorem 8 : For the dynamical system T (t), t ∈
+ , if there is a
nonempty compact set K that attracts compact sets of X and A =
∩ T (t)K, then A is independent of K and
t∈+
i) A is the maximal, compact, invariant set,
ii) A is connected if X is a Banach space.
iii) A is stable and attracts compact sets of X.
iv) For any compact set K, there is a neighborhood UK of K and a
t0 = t0 (K) such that γ + (T (t0 )UK )is bounded.
If, in addition, the dynamical system is asymptotically smooth, then
v) A is also a local attractor.
vi) If C is any subset of X for which there is a t0 = t0 (C) such that
γ + (T (t0 )C) is bounded, then A attracts C.
vii) In particular, if, for any bounded set B in X, there is a t0 = t0 (B)
such that γ + (T (t0 )B) is bounded, then A is the compact global attractor.
We also can state the following result.
Theorem 9 : A dynamical system T (t), t ∈
+ , on X has a compact
global attractor if and only if
i) T (t), t ∈
+ , is asymptotically smooth.
ii) T (t), t ∈
+ , is point dissipative.
iii) For any bounded set B in X, there is an t0 = t0 (B) such that
γ + (T (t0 )B) is bounded.
The suﬃciency is proved in the following way. From (i) and (iii),
ω(B) is a compact invariant set which attracts B for any bounded set
B. Since (ii) is satisﬁed, it follows that the compact invariant set which
is the ω−limit set of the bounded set in the deﬁnition of point dissipative
attracts each compact set of X. Theorem 16 implies the existence of the
compact global attractor. Conversely, if the compact global attractor
exists, then we must have (ii) and (iii) satisﬁed. Furthermore, if B
is a bounded set such that T (t)B ⊂ B, t ∈ G+ , then ω(B) must be
compact and attract B, which implies that the dynamical system is
asymptotically smooth.
Cholewa and Hale (2000) have shown that there is an asymptotically
smooth dynamical system and a bounded set B for which ω(B) is compact and invariant and yet γ + (B) is unbounded.
Theorem 10 : Suppose that the dynamical system T (t). 8. if there is the compact global attractor). of some results which will be stated below. As a consequence. Then (1) The attractor A has ﬁnite capacity c(A). Theorem 11 : 1) If an autonomous RFDE or NFDE with D0 exponentially stable is point dissipative.24 DELAY DIFFERENTIAL EQUATIONS This shows that (iii) in Theorem 17 cannot be replaced by γ + (B) bounded for each bounded set B. if the dynamical system has a maximal compact invariant set (in particular. We remark that. t ≥ G ¸ . If the compact global attractor exists. There is an example of Jones and Yorke (1969) in R3 of an ODE for which all solutions are bounded and yet there is no equilibrium point. 2) If a nonautonomous RFDE or NFDE with periodic coeﬃcients and D0 exponentially stable is point dissipative. Stationary points of dissipative ﬂows As we have remarked earlier. then the most interesting part of the ﬂow deﬁned by the dynamical system occurs on the attractor even though the transient behavior to a neighborhood of the attractor can be very important in a speciﬁc application. then there is a residual set of the set of all continuous projections on Son which the projection of the ﬂow onto S is onetoone. see Hale. Magalhaes and Oliva (2002)). In spite of this. it was of interest to determine the existence of ﬁxed points for the Poincar´e map associated with nonautonomous FDE with τ periodic dependence on t. (2) If S is a linear subspace of X with dimension ≥ 2c(A)+1. then the above properties hold for this set. we have the following surprising result of Maln´e (1981) (for the ﬁrst part for ﬁnite Hausdorﬀ dimension and X a Hilbert space. the result is due to MalletParet (1976)) (for a complete proof. The ﬁrst part of the theorem implies that the Hausdorﬀ dimension of A is ﬁnite dimension and the second part says that A generically can be embedded into ﬁnite dimensional subspaces of X if the dimension is suﬃciently large. then there is an equilibrium point. then there is a ﬁxed point of the Poincar´e map. There are examples of FDE for which the ﬂow on the attractor can be very complicated and the attractor itself is not a manifold. we can state the following result (see Hale and Lopes (1973)). . on a Banach space X has a compact invariant set A and has the property that the derivative of T (t)ϕ with respect to ϕ is the sum of a contraction and a completely continuous operator for each t > 0 and ϕ ∈ A.
for any bounded set B in X. compact dissipative is assumed. Minh and Shin (2002) for RFDE with inﬁnite delay and the lectures on abstract evolutionary functional diﬀerential equations at this workshop. Suppose that X is a Banach space. Naito. Hino. it is assumed only that the Poincar´e is point dissiaptive and in Theorem 20. We can only refer to the reader to the books of Diekmann. The Kuratowski measure α(B) of noncompactness of a bounded set B in X is deﬁned by α(B) = inf {d : B has a ﬁnite cover of diameter < d}. Hino. However. Hale. . the system must be compact dissipative.1). Theorem 12 : If X is a Banach space and T : X → X is an αcontraction which is compact dissipative. A map T : X → X is said to be an αcontraction if there is a constant k ∈ [0. therefore. It is not known if point dissipative is equivalent to compact dissipative for αcontractiing maps. Final Remarks We have only touched upon a few of the topics in FDE due to limited space. We now describe the asymptotic ﬁxed point theorem that is used to prove the above result. Murakami and Naito (1991).especially. Notice that part (2) has no restriction on the delay as in the results mentioned previously of Jones (1965) and Yoshizawa (1966). Massatt (1983) proved that point dissipativeness is equivalent to compact dissipativeness. Hale and Lunel (1993). the development of the qualitative theory.History Of Delay Equations 25 This says that some type of dissipation is necessary to obtain the conclusion in (0. then there is a ﬁxed point of T. Kolmanovski and Myshkis (1999). stability of the ﬂow on the attractor with respect to the vector ﬁeld as well as detailed investigation of the ﬂow on the attractor for speciﬁc types of equations that occur frequently in the applications. In Theorem 19. the results presented set the stage for much research in the last 35 years . Magalhaes and Oliva (2002) for FDE on Rn . The following result was discovered independently and with diﬀerent proofs by Nussbaum (1972). We have remarked earlier that point dissipative for RFDE implies the existence of the commpact global attractor and. we have α(T B)≤kα(B). We need a few remarks to see why Theorem 19 is a consequence of Theorem 20. We remark that the motivation for this ﬁxed point theorem came form the above problem. Lunel and Walther (1991). For NFDE with an exponentially stable D0 operator. van Gils. 1) such that. Wu (1996) for partial diﬀerential equations with delay. Hale and Lopes (1973). The proof is nontrivial and uses dissipativeness in two spaces.
Cooke (1963) Diﬀerential Diﬀerence Equations. R. J. 110. J.K. Hale.351. (1959) On a generalization of the Schauder ﬁxed point theorem. Hale (1971) Exponential estimates and the saddle point property for neutral functional diﬀerential equations. References Bellman.A. M. J. Appl. Diekmann. 32. (1963) A stability theorem for functional diﬀerential equations. 575578. 6. 1.K. Bull. 291303. 50. Hale. J. (1965) Suﬃcient conditions for stability and instability of autonomous functional diﬀerential equations. Cruz. 34. F. 3 (revision and expansion of 1971 Edition). Duke Math. Circ. Oscillations. J. 159176. RAND Corp. (1977) Theory of Functional Diﬀerential Equations. Browder. and J. Browder. J.W. Functional.K. Dynamics of Continuous. Sci. Stability. Bellman. Time Lags. and K. O.26 DELAY DIFFERENTIAL EQUATIONS It is clear that the subject is alive and is a good area of research both in theory and applications. Hale (2000) Some counterexamples in dissipative systems. Springer. 87109. Soc. 267288. R. 10821089. Mem. 331. Diﬀerential Eqns. (1954) A Sruvey of the Mathematical Theory TimeLag. van Gils. (1969) A periodicity theorem for autonomous functional diﬀerential equations. Danskin. Complex.L. Appl. LaSalle (1971) Periodic dissipative processes. F. S. Hale. Jr. Math. and K. J. J.K. Bellman. J. Report R256.. 942946. Press. and Nonlinear Analysis. and J. and J. Sci. Math. Billotti. Diﬀerential Eqns. Rend. (1965) A further generalization of the Schauder ﬁxed point theorem. Walther (1991) Delay Equations. S.K. Palermo (2)15. (1966) Diﬀerential Equations. Discrete and Impulsive Systems 7. Hale. Proc. Santa Monica CA. J. (1966) Linearly asymptotically autonomous functional diﬀerential equations.O. Anal. Halanay. (1963) Linear functional diﬀerential equations with constant coeﬃcients. A.K. Contributions to Diﬀerential Equations 2. Math. 452482. Math. Springer. Cholewa. Sci. and J. R. Appl. Retarded Control and Hereditary Processes. 6.A. Math. and H.P. J. J. 35. .L. Nat. Lunel. Soc. Acad.. 291319. Acad. Cooke (1959) Asymptotic behavior of solutions of diﬀerential diﬀerence equations. Am. Duke Math. 26.M. Hale. Academic Press. R.E.K. Amer. Grafton.
(1987) Topics in analysis. Mem. 351375. S. Slemrod (1973) Theory of a general class of dissipative processes. D. Press. Publ. Levinson. J. MalletParet. Nat. Krasovskii.V. 177191. 898. J. 67. V.R. and A. Krasovskii.S. 13 391402. 2984. Appl. Univ. Y.. G..History Of Delay Equations 27 Hale. Henry. J. Magalh˜ aes. Contributions to Diﬀerential Equations 3. Henry. and S. and M. Appl. Hale. Anal. Notes Math. Lect. Aut. (1944) Transformation theory of nonlinear diﬀerential equations of second order. Hino. Man´e. Naito.K. . Kluwer Acad. and J. Hale.P. LaSalle. and K. Hale. Soc. J. (1959) Stability of Motion. P. Lyapunov’s second method for investigating stability. Kolmanovski. and C. Y. Anal.K. Math. Oliva (2002) Dynamics in Inﬁnite Dimensions. Lopes (1973) Fixed point theorems and dissipative processes. 230242. A. Am. J. J. D. Taylor and Francis. (1974) Linear autonomous neutral functional diﬀerential equations. Math. (1976) Negatively invariant sets of compact maps and an extension of a theorem of Cartwright. Hino. Acad. Studies 17(1947).K. J. Hale. Math. 326333. Sbornik 40. (1981) On the dimension of the compact invariant sets of certain nonlinear maps. Sci. Minh. M. J. 136137. J. 39.. J. Pub. Diﬀerential Eqns.K. and W. Stanford Univ. 724737. 1963). J. Translation with additions of the 1959 Russian edition. London. Perell´ o (1964) The neighborhood of a singular point for functional diﬀerential equations. (1983) Attractivity properties of αcontractions. Math. 106128. Lyapunov. Diﬀerential Eqns. 99. Annals of Math. Springer. Appl.M. Lunel (1993) Introduction to Functional Diﬀerential Equations. N.. (1956) Theory of A.K. 331348. In Lecture Notes in Math. Meyer (1967) A class of functional diﬀerential equations of neutral type. 76. 5. Naito (1991) Functional Diﬀerential Equations with Inﬁnite Delay. Shin (2002) Almost Periodic Solutions of Diﬀerential Equations in Banach Spaces. Proc. Hale. 435450. R. Barcelona 39. T. (1962) The existence of periodic solutions of f (x) = −αf (x− 1)[1 + f (x)]. (1970) A class of neutral equations with the ﬁxed point property. 45. Massatt. Diﬀerential Eqns 48. N. N. Math. Mat. Myshkis (1999) Introduction to the Theory and Applications of Functional Diﬀerential Equations. J. Math.. L. 15. (1891) Probl`eme g´en´eral de la stabilit´e du mouvement.S. 1473 SpringerVerlag. Originally published in 1891. Murakami. N.K. and T. Hale. J. J. Sci. Jones. Second Edition. Appl. Sci. 47. Diﬀerential Eqns 22. Annales Math. J.K. and O. Springer.
. (1974) Periodic solutions of some nonlinear autonomous functional diﬀerential equations. S. Math. 500512. Prikl. V. Inc. S. 6687. J. Shimanov. 20. Picard. E Actes du IVe congr`es international des Math´ematiciens. ¨ (1908) La mathematique dans ses rapports avec la physique. Minorsky. GauthierVillars. B. ¨ Schmitt. Lincei 18. (1941) Control problems. A.N. E. Math. (1956) On the stability of systems with a delay. R. Pura Appl. (1972) Some asymptotic ﬁxed point theorems. Minorsky. V. (1950) The existence of periodic solutions of systems of diﬀerential equations. Princeton. D. Sci. 17 457475. Volterra. Translation of the 1964 Russian text. 119. (1966) Nonlocal problems in the theory of nonlinear osciallations. 836844. Ann. 10. N. Nussbaum. (1951) Lineare Diﬀerentialgleichungen mit nacheilenden Argumentom. Franklin Institute232:6. Math. V. Math.349375. 519551. N. 1955. E.N.. Academic Press.28 DELAY DIFFERENTIAL EQUATIONS Massera. 295. Berlin. Van Nostrand Co. Translation of the 1951 Russian edition. (1911) Uber eine Klasse linearer funktionaler Diﬀerentialgleichungen. Soc. 7. Am. J. . Atti Reale Accad. (1962) Nonlinear Oscillations. Duke Math. [Russian] Diﬀerentialniye Uravnenija 1.M. Math. 499524. Razumikhin. 263306. Theory and Application of Partial Functional Diﬀerential Equations Appl. Pliss. Myshkis. Wiss. Volterra. Shimanov. Wu. J. J.L. (1965) On the theory of linear diﬀerential equations with retardations. (1931) Th´eorie Math´ematique de la Lutte pour la Vie. Rome. 102116. Ann. [Russian] Prikl. Meh. J. 23. Mat. Wright. V. 249298. J. 194. D. Meh.A. Paaris. Math. (1928) Sur la th´eorie math´ematique des ph´enom`enes h´er´editaires. 171. Volterra. R. Reine Angew. (1955) A nonlinear diﬀerentialdiﬀerence equation. Mat. Trans. 70. Springer.S. (1909) Sulle equazioni integrodiﬀerenziali della teorie dell’elasticita. Pures Appl. Nussbaum. (1959) On the stability in the critical case of a zero root with time lag. Deutscher Verlag.
In diﬀerential diﬀerence equations.P. Assume that the length of the juvenile period is exactly h units of time for each individual. . Such equations 31 O. Introduction In applications. Arino et al. Example 1 A retarded functional diﬀerential equation. the past exerts its inﬂuence in a signiﬁcant manner upon the future. than by ordinary diﬀerential equations.Chapter 2 SOME GENERAL RESULTS AND REMARKS ON DELAY DIFFERENTIAL EQUATIONS E. or more generally functional diﬀerential equations. B. Assume that adults produce oﬀspring at a per capita rate α and that their probability per unit of time of dying is µ. So the time variation of the population density N involves the current as well as the past values of N . (eds. rN (t−h) meaning that newborns become adults with some delay. Then the dynamics of N can be described by the diﬀerential equation dN (t) = −µN (t) + rN (t − h) dt (1. 31–40. Delay Differential Equations and Applications. Ait Dads D´epartement de Math´ematiques. Let N (t) denote the density of adults at time t.1) which involves a nonlocal term. Assume that a newborn survives the juvenile period with probability ρ and put t = αρ. Many models are better represented by functional diﬀerential equations. Imagine a biological population composed of adult and juvenile individuals. Facult´e des Sciences Universit´e Cadi Ayyad. the future behavior of many phenomena are assumed to be described by the solutions of an ordinary diﬀerential equation. Implicit in this assumption is that the future behavior is uniquely determined by the present and independent of the past. Fax 212 44 43 74 09 Morocco 1. © 2006 Springer.). 2390 Marrakech.
Thus the method of steps and elementary theory of ordinary diﬀerential equations provide us with a very simple existence and uniqueness proof in this example. Remarks 1) For any continuous function ϕ deﬁned on [−h. denoted x(ϕ).32 DELAY DIFFERENTIAL EQUATIONS are called Retarded Functional Diﬀerential Equations (RFDE) or. 0].2) coincide at t = 0. 0] and then to use ( 1.1) by N (θ) = ϕ(θ) for − h ≤ θ ≤ 0 where ϕ is a given function. N (t) = ϕ(0) exp(−µt) + r 0 Using this expression we can give an expression for N on the interval [h. 2) The solution x(ϕ) has a continuous time derivative for t > 0. The most convenient (though not the most natural from a biological point of view) manner to do this is to prescribe N on the interval [−h. 3) For a given ϕ on [−h. dt The solution x(ϕ) is smoother than the initial data. all over the interval [0. etc. but not at t = 0 unless ϕ(θ) has a left hand derivative at θ = 0 and dϕ (0) = −µϕ(0) + rϕ(−h). h] t exp(−µ(t − τ ))ϕ(τ − h)dτ. So we supplement ( 1. Equation ( 1. To determine a solution past time t = 0. Explicitly. h] we have the same problem: in order to integrate the equation past some time t ∈ [0. Delay Equations. 2h].1) for t ≥ 0. Example 2 The solutions t → sin π 2 t+ 1 2 and t → cos π 2 t+ 1 2 of the equation dx π = − x(t − 1) dt 2 (1. say x(ϕ)(t) . if x(ϕ)(t) is deﬁned for t ≤ −h. ∞]. we need to prescribe the value of N at time −h.1) describes the changes in N . we need to prescribe the value N (t − h). since the following example agree that this condition is not enough to determine completely the solution. 0]. alternatively. In fact. the solution x(ϕ)(t) of (1. we then have for t ∈ [0.2) need not be deﬁned for t ≤ −h.2) on [−h. In fact. and we can see that it is not enough to give the value at the point −h. there is a unique solution x of (1. h]. So we have to prescribe a function on an interval of length h.
Let σ ∈ R. A > 0 such that x ∈ C ([σ − r. σ + A] . ϕ) is a solution of ( 2. x(t − r(t))) dt as well as dx (t) = dt for 0 ≤ r(t) ≤ r 0 g(t. Let f : R×C → Rn be a given function. we let C = C ([−r. ϕ).3) is a very general type of equation and includes diﬀerentialdiﬀerence equations of the type dx (t) = f (t. θ. x(t − r)). ϕ) = ϕ. −r≤θ≤0 A > 0 and x ∈ C ([σ − r. x(t + θ))dθ.3) with initial value at σ or simply a solution of ( 2. x(t). is a solution of ( 2. b] = [−r. Rn ) and designate the norm of an element ϕ in C by ϕ = sup ϕ(θ). for t ≥ σ (2. σ + A] . Rn ). A more general form of a delay diﬀerential equation is as follows: dx = F (t. σ + A] . we let xt ∈ C. ϕ) + h(t). σ + A]. 0]. Rn ) and x satisﬁes ( 2. ϕ) if there is an A > 0 such that x(σ. dt 2. xt ). be deﬁned by xt (θ) = x(t + θ). Equation (2. ε > 0.3) dt and xσ = ϕ. denote C ([a. x(t). b] . σ + A] and xσ (σ. then ϕ(θ) must have a continuous ﬁrst derivative for θ ∈ ]−ε. If [a. A general initial value problem Given r > 0.3) on [σ − r.33 General Results and Remarks on DDE is deﬁned for t ≥ −h − ε. for − r ≤ θ ≤ 0. Rn ). then for any t ∈ [σ. Deﬁnition 7 x is said to be a solution of ( 2. the Banach space of continuous functions mapping the interval [a. .3) through (σ. −r If f (t.3) for t ∈ [σ. σ + A] for a given σ ∈ R and a given ϕ ∈ C we say that x = x(σ. In such a case we say that x is a solution of ( 2. A functional diﬀerential equation is given by the following relation dx (t) = f (t.3) if there are σ ∈ R. 0[. 0] . ϕ) = L(t.3) on [σ − r. b] into Rn with the topology of uniform convergence.
σ + α] . σ By the Gronwall lemma xt (. ϕ) ∈ D. ∃δ > 0. ϕ). 0]. Proof.3) through (σ. it is called homogeneous if h ≡ 0. b > 0. we have x(t + θ)− x(s + θ) < ε. σ + α] . ϕ) is equivalent to solving: t f (s.3). there exists a solution of equation ( 2. Then.3) is an autonomous one. Consequently for t. ϕ) = g(ϕ). t − s < δ. f (t. Proposition 3 If f is at most aﬃne i.. For any (σ.e. Let ϕ ∈ C. ϕ) = (ϕ + bβ) exp aβ < ∞. ∀ϕ. φ ≤ a φ+b with a.3) through (σ. we say that the equation is a linear delay diﬀerential equation. If f (t. By integrating the equation (2. ϕ) = ϕ + (a xs  + b)ds σ and xt (. Rn ). it is uniformly continuous and thus ∀ε > 0.. then there exists a global solution i.1 Existence Lemma 2 [12] If x ∈ C ([σ − r. ﬁnding a solution of equation ( 2. Since x is continuous on [σ − r. ϕ) → L(t. β[. one has t f (s. and assume that the solution is deﬁned only on [α. xs )ds t ≥ σ and xσ = ϕ. Theorem 1 [12] Let D be an open subset of R × C and f :D → Rn be continuous. x(t) = ϕ(0) + σ 2.. ϕ). xt is a continuous function of t for t ∈ [σ. ∞[ . σ + α]. equation (2. σ + α]. ϕ) = ϕ + a t xs  ds + bβ. ϕ) is deﬁned on [α.34 DELAY DIFFERENTIAL EQUATIONS in which L is linear in ϕ and (t. such that x(t) − x(s) < ε if t − s < δ. xs )ds x(t) = ϕ(0) + 0 which gives t x(t. . s in [σ. the solution x(σ. then. Lemma 1 [12] Let σ ∈ R and ϕ ∈ C be given and f be continuous on the product R × C.e. Proof. ∀θ ∈ [−r.
35 General Results and Remarks on DDE On the other hand .
.
.
dx .
.
sup .
(t).
.
note it xβ . then equation ( 2. t∈[0. let us consider the following counterexamples: 1) There may be two distinct solutions of (2.3) on [σ − r. ϕ) in a compact subset containing the trajectories (t. 2. ϕ) be lipschitzian with respect to ϕ in every compact subset of D.3) has a unique solution passing through (σ.β[ dt and gives that the solution is uniformly continuous on [0. β + ε] . which contradicts the maximality of the solution. σ + α] with xσ = ϕ = yσ . Bβ as deﬁned in the proof of theorem . Then t t≥σ x(t) − y(t) = σ (f (s.3) has at least one solution deﬁned on [0. xs ) − f (s. Choose α such that kα < 1. Consider Iα . ∞) and they coincide on (0. for t ∈ Iα one has: t k xs − ys  ds ≤ kα sup xs − ys  . x(t) − y(t) ≤ σ σ≤s≤t And this implies that x(t) = y(t) for t ∈ Iα . Proof. then it satisﬁes the property in the proposition below. t ∈ Iα . The following example was given by A. The uniqueness may not hold if the function is not locally lipschitzian. and equation (2.. and suppose x and y be two solutions of (2. β[ and this implies that lim xt (. ys ))ds. For this. Then. xt ) and (t. If (σ.2 Uniqueness Theorem 3 [12] Let D be an open subset of R × C and suppose that f : D → Rn be continuous and f (t. < ∞. t→β Let us consider the following delay diﬀerential equation dy = f (t. xσ − yσ = 0 Let k be the Lipschitz constant of f (t. ϕ) ∈ D. yt ). ∞) . ϕ). Corollary 2 If f is lipschitzian with respect to the second variable. β + ε] for some ε > 0. ϕ) exists and is ﬁnite. yt ) for t ≥ β dt yβ = xβ ∈ C this last equation has at least one solution on [β.3) deﬁned on (−∞. .
t < 0. In fact. and x(t) =≡ 0. for θ ∈ [−1. Let x ∈ [−1. xt  = x(t − 1) = −(t − 1)3 and dx (t) = −3t2 It is easy to verify that dt −3t2 = f ((t − 1)3 ) for t < 0. 1] is smooth. it is clear that x satisﬁes the equation for t ≥ 0. 2) d x(t) = x(t − σ(x(t))) (2. 0] . Let r = 1. dx dt (t) = −3t . 0] such that ϕ + x < ρ. then G(ϕ + x) − G(ϕ) ≤ k x which implies . 0] . 0 ≤ s ≤ 1. f (s) = −3 ( 3 s − 1) . Note that the right hand side of (2. ∞) . ϕ2  < ρ √ Let ϕ(θ) = (−1+ 1 + θ). and consider the equation dx (t) = f (xt ) .36 DELAY DIFFERENTIAL EQUATIONS √ 2 Hausrath. σ (0) = 0 and σ(0) = 1. the 2 function x(t) = −t3 . where is a positive constant such that ϕ < ρ . R) and G is not locally lipschitz in a neighborhood of zero . since x ≤ 1 for t ≥ −1.4) dt where σ : R → [0. In fact assume that there exist positive constants k and ρ such that G(ϕ1 ) − G(ϕ2 ) ≤ k ϕ1 − ϕ2  for ϕ1  . t ≥ 0. Also. s > 1. dt The function x ≡ 0 is a solution of this equation on (−∞. f (s) = 0. Since x is monotone decreasing for t ≤ 0.4) can be written as G(ϕ) = ϕ(−σ(ϕ(0)) for ϕ ∈ C([−1.
.
.
.
.
(σ(x) − 1) .
.
.
.
.
.
1 − σ(x) + x.
≤ k x and .
.
≤ (1 + k). .
x 1 − σ(x) .
The uniqueness has been proved for lipschitzian initial data ϕ. The equation d x(t) = x(t − x(t)) dt . Therefore the right hand side of equation (2. the standard argument for uniqueness can not be applied in this example. Letting x approaches zero. see [26] However .4) is not locally lipschitz near zero. we obtain a contradiction.
4 In fact one has t − x1 (t) = − t2 .3) has a maximal solution deﬁned on [−r. a] is the maximal interval of existence of the solution x. t→β . Continuation of solutions Deﬁnition 8 Suppose f in equation ( 2. and in fact the related equation d x(t) = −ax(t − σ(x(t))) dt was studied theoretically by Cooke. if f is a bounded function. for t ∈ [0. 1] . A solution x is noncontinuable if no such continuation exists.3) on an interval [σ.3) on [σ. respiration..3) is continuous. coincides with x on [σ − r. then equation ( 2.3 = ϕ(t − x1 (t)) and − x2 (t)) for t ∈ [0. b] . b] . a] .5) has two solutions namely x1 (t) = t + t2 and x2 (t) = t. a > σ. β[ with if β < ∞ =⇒ lim xt (. d x(t) = −a(t)x(t)−λb(t)f (x(t−σ(x(t)))) where λ is a positive parameter dt has been proposed as models for a variety of physiological processes and conditions including production of blood cells. we say x is a continuation of x if there is a b > a such that x is deﬁned on [σ − r.5) x(θ) = θ + 1 Then equation (2. 1] . a] . If x is a solution of equation ( 2. and t − x2 (t) = 0. 3) The following counter example explains more the situation d dt x(t) = x (2. and cardiac arrhythmias. Theorem 4 Furthermore on the hypotheses of the precedent theorem.General Results and Remarks on DDE 37 arises in models of crystal growth. and satisﬁes Equation ( 2. that is the interval [σ. ϕ) = ∞. it follows that 4 x1 (t) = 1 + 2t x2 (t) = 1 = ϕ(t 2.
38 DELAY DIFFERENTIAL EQUATIONS Proof. ϕ) < ∞ then there exists N such that .. β[ . By steps. we can integrate the equation (2. xt (.3). ϕ) is deﬁned. with dx = f (t. xt ) and from the boundedness of f .. β[ . be the maximal interval on which x(. we have dt . Assume that lim xt (.. let [−r. ϕ) ≤ N. ∀t ∈ t→β [0.
.
.
dx .
.
sup .
(t).
.
ϕ1 ) − x(t. x(s. Let ψ ∈ C( [−r. 0). ϕ2 ))ds. . ϕ2 ) = ϕ1 (0) − ϕ2 (0)+ t 0 (f (s.. x(. If (σ.. then. Rn ) deﬁned by ψ = xβ . for any /U closed bounded set U in R×C. f : Ω → Rn be globally lipschitz and completely continuous: that is.ϕ2 ) the associated solutions. ϕ) exists. the application ϕ → xt (. b] . ϕ1 ) − f (s.. From the corollary 2 and the fact that f is lipschitzian with respect to the second variable f (t. and x is a noncontinuable solution of equation ( 2. one has x(t.β[ then x is uniformly continuous on [0. ϕ) be lipschitzian with respect to ϕ in every compact subset of D. Let ϕ1 . < ∞. f is continuous and takes closed bounded sets of Ω into bounded sets of Rn .3) on [σ − r. which contradicts the maximality of x. lim xt (.4 Dependence on initial values and parameters Theorem 6 Let D be an open subset of R×C and suppose that f : D → Rn be continuous and f (t. xt ) ∈ for tU ≤ t < b. there exists ε > 0 such that the equation dy = f (t. Theorem 5 [12]Suppose Ω be an open set in R × C.ϕ1 ). x(s. which t→β we denote by xβ . Then. β[ . U in Ω. The following lemma is needed before we proceed in that direction.. 2. β[ . β + ε] . So. Proof. We now consider the existence of solutions of (2. β + ε] . ϕ2 ∈ C. ϕ ≤ k ϕ + f (t. yt ) for t ≥ β dt yβ = xβ ∈ C has at least one solution on [β. and x(. ϕ) is continuous lipschitz. the recollement of x and y gives a solution deﬁned on [α.3) for all t ≥ −r. there is a tU such that (t. under the existence theorem. dt t∈[0. ϕ) ∈ D.
Then. φ0 ) − φ0 ≤ r∗ (t. u(t)) dt through (t0 .Then. u0 ) through every point (t0 . t0 . φ) ≤ g(t. We merely state. For t ∈ J and φ ∈ Cρ . 0). R+ ). φ0 ) of equation ( 2. ϕ2 )) ds. 0) on the common interval of existence of x(t. 2ρ] . ϕ1 ) − x(t. Suppose ﬁnally that the solutions u(t. 0). t0 . u0 ). . t0 . Then. φ0 ) and r∗ (t. ϕ. φ0 ).General Results and Remarks on DDE 39 x(t. Rn ). φ) . t0 . ϕ − φ) where g ∈ C (J × [0. we can prove the following theorem on dependence on parameters. 0) is the maximal solution of du = G(t. φ ≤r Suppose that r∗ (t. one has x(t. φ ∈ C1 f (t. we have : xt (t0 . we put G(t. ϕ1 ) − x(s. 0).3) are unique and continuous with respect to the initial values (t0 . t0 . By the Gronwall’s lemma. ϕ1 ) − x(t. ϕ2 ) = ϕ1 − ϕ2  + t k 0 x(s. ϕ) − f (t. Theorem 7 [22] Let f ∈ C (J × Cρ . if x(t. t0 . Using the arguments of the precedent theorems. xt ) dt with φ0 as an initial value at t = t0 . Assume that u(t) ≡ 0 is the only solution of the scalar equation du = g(t. u0 ) exist for t ≥ t0 and are continuous with respect to the initial values (t0 . Rn ) and for t ∈ J. which will be used subsequently. Lemma 3 [22] Let f ∈ C (J × Cρ . ϕ2 ) ≤ ϕ1 − ϕ2  exp (kt) We shall ﬁrst prove the following lemma. φ0 ) is any solution of dx (t) = f (t. the solutions x(t0 . r) = max f (t. u(t)) dt through (t0 .
5 Diﬀerentiability of solutions In precedent section suﬃcient conditions were given to ensure that the solution x(σ. φ. ϕ. t0 + a]. ϕ. (2. µ0 )(t) be a solution of dx (t) = f (t. φ0 . the derivative of x with respect to ϕ. f )(σ) = I. then the solution x(σ. ϕ. Rn ) satisﬁes the nonhomogeneous variation equation z (t) = Dϕ f (t. ϕ. ϕ. φ. (2. ϕ. f ).φ). Df x(σ. f )(σ) = 0. ϕ. φ ∈ C. f )zt + g(t. xt (σ. µ)(t) deﬁned on some interval [t0 . Then. µ) − f (t. f )(t) is a linear operator from C to Rn . ϕ. Dϕ x(σ. µ ∈ Rm f (t. the diﬀerential equation dx (t) = f (t. p ≥ 1. φ. Theorem 9 [12]If f ∈ C p (Ω. let C p (Ω.3) through (σ. Rn ) and. f ). Rn ).7) . f ) for t in any compact set in the domain of deﬁnition of x(σ. p ≥ 0 designate the space of functions taking Ω into Rn that have bounded continuous derivatives up through order p with respect to ϕ in Ω. for µ = µ0 let x0 (t) = x0 (t0 . t0 + a] such that x(t) − x0 (t) < ε for t ∈ [t0 . xt (σ. ∃δ > 0 such that for every µ satisfying µ − µ0  < δ (ε).40 DELAY DIFFERENTIAL EQUATIONS Theorem 8 [22]Let f ∈ C (J × Cρ × Rm . Furthermore. ϕ. Df x(σ. for each t ≥ σ. Rn ). If Ω is an open set in R × C. ϕ) is unique and continuously diﬀerentiable with respect to (ϕ. for each t ≥ σ. µ0 ). xt (σ. Dϕ x(σ. ϕ. f ) of the ( 2. ϕ. φ0 . and Df x(σ. µ) dt admits a unique solution x(t) = x(t0 . and for t ∈ J. µ) ≤ k ϕ − φ . f )g(t) for each g in C p (Ω. f ). f )(t) is a linear operator from C p (Ω. ∀ ε > 0. µ0 ) µ→µ0 uniformly in (t. f )ψ(t) for each ψ in C satisﬁes the linear variational equation y (t) = Dϕ f (t. µ) = f (t. f )yt . Rn ) into Rn .3) depends continuously on (σ. dt with an initial function φ0 at t0 existing for t ≥ t0 . ϕ. ϕ. Assume that lim f (t. ϕ. and Dϕ x(σ. xt . 2. ϕ. xt .6) Also. the identity. f )). ϕ. f ) on a (2. In this section some results are given on the diﬀerentiability with respect to (σ.
© 2006 Springer. the function t → L(xt ) is continuous for t ≥ 0. 1. A 8010 GRAZ (AUSTRIA) franz. Heinrichstrasse 36. which implies that x(t) is continuously 41 O.Chapter 3 LINEAR AUTONOMOUS FUNCTIONAL DIFFERENTIAL EQUATIONS Franz Kappel Institute for Mathematics and Scientiﬁc Computing. ∞) → C. is deﬁned by xt (θ) = x(t + θ). Since t → xt is a continuous mapping [0. .).1 Basic Theory Preliminaries Throughout these notes r is a ﬁxed constant. t ≥ 0. Arino et al. University of Graz. (1. 0 ≤ t < α. 41–139. is a continuous function. (1. −r ≤ t ≤ 0. ∞) → Cn such that t x(t) = φ(0) + L(xτ ) dτ. t ≥ 0. φ) and is a continuous function [−r. −r ≤ θ ≤ 0. We denote by C the Banach space of continuous functions [−r. 0 x(t) = φ(t).at 1.kappel@unigraz. 0] → Cn with norm φ = sup−r≤θ≤0 φ(θ). α) → Cn . Let L be a bounded linear functional C → Cn and choose φ ∈ C. α > 0.1) (1. Then we shall consider the following Cauchy problem: x(t) ˙ = L(xt ). ∞) will be denoted by x(t) = x(t. 0 ≤ r < ∞. (eds. then xt ∈ C. Delay Differential Equations and Applications. x0 = φ.2) A solution of problem (1. If x : [−r. where  ·  is any vector norm in Cn .1).2) on [0.
if Φ = (φ1 . ∞) → Cn . (1. 0). it is clear that taking t = 0 as initial time is no restriction of generality. (1. . . . f ) of (1.. By the Riesz representation theorem (cf. for instance [15]) that η can be represented as η = η1 + η2 + η3 . .1). .5) 0 0 x(t) = φ(t). In addition to problem (1. to assume that the extension is normalized in the following sense: η(θ) = 0 for θ ≥ 0. ∞) and that (1. It will be useful to extend the deﬁnition of η to all of R and..1) is autonomous. = sup{L(φ)  φ = 1}.. . f ) is absolutely continuous on bounded subintervals of [0.j=1. of course).. x0 = φ ∈ C. then L(Φ) := L(φ1 ).7) From the theory of functions of bounded variation it is well known (cf. −r −r (1.3) is satisﬁed almost everywhere on [0. φ. i. . η(θ) = η(−r) for θ ≤ −r. The norm of L will be denoted by .42 DELAY DIFFERENTIAL EQUATIONS diﬀerentiable for t ≥ 0 and equation (1.2) we also shall consider the nonhomogeneous problem x(t) ˙ = L(xt ) + f (t). ∞) → Cn such that t t L(xτ ) dτ + f (τ ) dτ. Moreover.. .1) is satisﬁed on [0. (1. for instance) there exists a matrix η = (ηij )i. φn ) is an n × nmatrix with columns φj ∈ C. . [31]. . Since equation (1. t ≥ 0. . ∞) is a continuous function [−r. It is clear that x(t. 0] such that n 0 n 0 L(φ) = col φj (θ) dη1j (θ).3). ∞).4) where f is a locally integrable function [0. (1.e. L(φn ) . in addition. φj (θ) dηnj (θ) =: j=1 −r j=1 0 [dη(θ)] φ(θ).n of bounded variation on [−r.4) on [0. A solution x(t) = x(t.3) (1. x(t) = φ(0) + (1. . ∞) (at t = 0 the derivative is understood to be the righthand derivative. φ.6) φ ∈ C. η is lefthand continuous on (−r. . −r ≤ t ≤ 0. t ≥ 0.
. . The corresponding matrix valued .. . the Ak ’s are constant n × nmatrices and θ → A(θ) is integrable in [−r. −r φ ∈ C. p.. φ ∈ C. k=1 η 1 being a function of bounded variation is equivalent to ∞ Ak  < ∞. 0].43 Autonomous Functional Diﬀerential Equations where η 1 is a saltus function of bounded variation with at most countably many jumps on [−r. L(φ) = m k=0 Ak φ(−rk ) + 0 A(θ)φ(θ) dθ. k=1 where  ·  can be any matrix norm. 0]. For η 2 we deﬁne A(θ) by θ 2 η (θ) = A(θ) dθ. 0 If η 3 ≡ 0. i. then −r [dη 3 (θ)] φ(θ) cannot be transformed to a Lebesgue integral or to a series.j=1. η 2 is an absolutely continuous function on [−r. we have 0 A(θ) dθ < Since the functions ηij −r ∞. 0].e. . 1 ) ≡ 0 and let −r . η 3 is non constant.n . −r φ ∈ C. where 0 = r0 < r1 < · · · < rm = r. Then Ak = (akij )i. For a concrete example see [15. −r . 0]. continuous and has derivative η˙ 3 = 0 almost everywhere on [−r... 0] and η 3 is either zero or a singular function of bounded variation on [−r. For most situations it is suﬃcient to consider the special case where η 3 ≡ 0 and η 1 has only a ﬁnite number of jumps. where akij = ηij k k ij 0 [dη 1 (θ)] φ(θ) = −r ∞ Ak φ(−rk ). Deﬁne of those points in [−r. 0 Then 0 [dη 2 (θ)] φ(θ) = −r 0 A(θ)φ(θ) dθ. 0] where at least one ηij 1 (−r + 0) − η 1 (−r − 0). 457]. −r ≤ θ ≤ 0. be an enumeration Consider η 1 = (ηij 1 2 1 has a jump. 2 are of bounded variation. In view of (1.7) we can assume η 1 (0) = η 2 (0) = η 3 (0) = 0.
. ..44 DELAY DIFFERENTIAL EQUATIONS function η 1 is given by 0 −A0 −(A0 + A1 ) η 1 (θ) = .
m−1 − k=0 Ak − .
φ. We ﬁrst prove some estimates which will be useful later. .4) on [0. f ) be a solution of (1. t ≥ 0.1). (1. For t ≥ 0 and −r ≤ θ ≤ 0 we immediately get from (1. (1.9) 0 Proof. Uniqueness of solutions is an immediate consequence of Lemma 1.8) now follows from Gronwall’s inequality. f ) ≤ φ + t f (τ ) dτ et . t + θ ≥ 0. f ) ≤ φ + 0 and consequently L(xt (φ. Lemma 1. for −rm < θ ≤ −rm−1 .m A k=0 k 1. (1.8) xt (φ. xτ dτ. Existence and uniqueness of solutions In this section we consider existence and uniqueness of solutions for the problems (1.1. t ≥ 0. Then we have the estimate t f (τ ) dτ et . (1. t ≥ 0.2 for θ ≥ 0. ∞). for −r2 < θ ≤ −r1 . for −r1 < θ < 0. (1. Therefore we have xt ≤ φ + 0 t f (τ ) dτ + t t + θ ≤ 0.5) the estimates t+θ t+θ xτ dτ + f (τ ) dτ x(t + θ) ≤ φ(0) + 0 0 t t f (τ ) dτ.1 Let x(t) = x(t.3). for θ ≤ −rm . ≤ φ + xτ dτ + 0 0 and x(t + θ) = φ(t + θ) ≤ φ.3).2) and (1. 0 The estimate (1.4).
TT is a contraction on C0 (0. i. ∞) → Cn is a solution of (1.3). Proof. 0 ≤ t ≤ T. It is clear that TT C0 (0. ∞).e.3).4) if and only if the restriction of x to any interval [0. f )(t) = xt (φ. f ). 0 ≤ t ≤ T . TT y (t) = φ(0) + 0 0 in the Banach space C0 (0. Cn ).4) on [0. A continuous function [−r..8).2 For any φ ∈ C and any f ∈ L1loc (0. C) deﬁned by σ(φ. T ) be the unique ﬁxed point of TT on C0 (0. the function y˜ is deﬁned by y˜(t) = φ(t) for −r ≤ t ≤ 0 and y˜(t) = y(t) for 0 < t ≤ T . Proposition 1. T . T . which proves TT y − TT zγ < − y − zγ . 0≤t≤T This implies t TT y (t) − TT z (t) ≤ y − zγ e−γτ dτ < − y − zγ e−γt γ 0 for 0 ≤ t ≤ T . we have ˜ yτ − z˜τ = sup ˜ y (τ + θ) − z˜(τ + θ)eγ(τ +θ) e−γ(τ +θ) −r≤θ≤0 −γτ ≤e sup y(t) − z(t)eγt = e−γt y − zγ . T . T . T . 0≤t≤T y ∈ C0 (0. Cn ). Cn ) → C(0. Cn ). (1. Cn ) of all continuous functions y : [0. T . z ∈ C0 (0. ∞) → Cn such that x(t) = φ(t) for −r ≤ t ≤ 0 and x(t) = y(t. T ) for 0 ≤ t ≤ T and any T > 0. Let y(·. T . ∞. γ By choice of γ we have 0 < −/γ < 1. For y. is bounded and linear. T . Cn ). T > 0. T . T .Autonomous Functional Diﬀerential Equations 45 Proposition 1. Proof. 0 Since y˜(t) − z˜(t) = 0 for −r ≤ t ≤ 0 and γ < 0. Cn ) there exists exactly one solution of problem (1. (1. Cn ) with the equivalent norm yγ = sup y(t)eγt . Cn ) ⊂ C0 (0.3 For any T > 0 the mapping σ : C × L1 (0. T . . Cn ). For y ∈ C0 (0. Boundedness of σ is an obvious consequence of (1. is a ﬁxed point of the operator TT deﬁned by t t L(˜ yτ ) dτ + f (τ ) dτ. T ]. We choose a negative constant γ < − and supply C0 (0. Cn ) we get t TT y (t) − TT z (t) ≤ ˜ yτ − z˜τ dτ. It is clear that there exists exactly one function x : [−r. T ] → Cn with y(0) = φ(0) supplied with the supnorm.
we deﬁne ∆(λ) = λI − and p(λ. .1). . for any solution of (1.1).1) shows that also x(t) ˙ has a Laplacetransform at least in Re λ > given by λˆ x(λ) − φ(0)1 . The fundamental matrix Throughout this section we consider the homogenous problem (1. φ) = 0 exists at least for Re λ > .2). for λ ∈ C.2). Thus we have ∆(λ)ˆ x(λ) = p(λ. and λ ∈ C the function ej (λ) ∈ C is deﬁned by ej (λ)(θ) = θj λθ e . t ≥ 0.3) . (1.3 DELAY DIFFERENTIAL EQUATIONS The Laplacetransform of solutions. . φ). j! −r ≤ θ ≤ 0. φ) ≤ φet . (1. the existence of a Laplacetransform for x(t) ˙ implies the same for x(t) (see Theorem A.1) we get for Re λ > ∞ 0 e−λt [dη(θ)] x(t + θ) dt λˆ x(λ) − φ(0) = −r 0 = 0 −r 0 = ∞ [dη(θ)] 0 0 [dη(θ)] −r e−λt x(t + θ) dt λ(θ−τ ) e 0 φ(τ ) dτ + −r θ eλθ dη(θ) · x ˆ(λ). For j = 0. Therefore the Laplacetransform ∞ e−λt x(t. φ) = φ(0) + [dη(θ)] −r 0 eλθ dη(θ) −r 0 Re λ > .10) 1 In fact. In this case inequality (1. θ In order to save space we introduce the following notation.8) implies x(t.46 1. 1. Taking Laplacetransforms on both sides of equation (1. (1. where. φ) dt x ˆ(λ) = x ˆ(λ. the integral converging absolutely in this half plane.9) together with equation (1. Similarly inequality (1. Interchanging Stieltjes integration with improper Riemann integration is justiﬁed by Fubini’s theorem. φ ∈ C. 0 eλ(θ−τ ) φ(τ ) dτ.
0] ηij for Re λ ≥ 0. φ) can be written as 0 p(λ. 0] is deﬁned by θ f (θ − τ )g(τ ) dτ. φ). φ) exists at least in the set {λ ∈ C  det ∆(λ) = 0} and.13) j=0 the series converging absolutely. φ) = φ(0) − L(e0 (λ) ∗ φ) = φ(0) + 0 eλθ dη(θ). λ −r Since 0 −r e dηij (θ) ≤ λθ var[−r. eλθ dη(θ) ≤ λ −r λ α This means that. φ).4 a) The Laplaceintegral for x(t) = x(t.47 Autonomous Functional Diﬀerential Equations Furthermore. for any ε > 0 the series is uniformly convergent for Re λ ≥ K + ε. is absolutely convergent at least for Re λ > . For λ = 0 we may write 1 0 λθ ∆(λ) = λ I − e dη(θ) . b) The Laplacetransform x ˆ(λ) = x ˆ(λ. for functions f. λ λ −r (1. there exists a constant K > 0 such that for any α > 0 1 0 K K ≤ for Re λ ≥ α. the matrix ∆−1 (λ) is given by ∞ j 1 1 0 λθ −1 ∆ (λ) = e dη(θ) . Moreover. where. φ) = φ(0) − [dη(θ)] e0 (λ) ∗ φ (θ) −r = φ(0) − L(e0 (λ) ∗ φ). h(θ) = 0 With these notations p(λ. λ ∈ C. −r ≤ θ ≤ 0. We summarize our results in Theorem 1. φ ∈ C. −r 0 0 [dη(θ)] −r θ eλ(θ−τ ) φ(τ ) dτ. . C) the convolution h = f ∗ g on [−r. for λ ∈ C. 0. φ ∈ C. is given by x ˆ(λ) = ∆−1 (λ)p(λ.0] ηij e−r Re λ var[−r. φ) of x(t. for Re λ > K. g ∈ L1 (−r.11) (1.12) (1. for Re λ < 0. ∆(λ) = λI − L(e I) = λI − λ· p(λ. on this set.
The same is true for p(λ. The Laplaceintegral for Y (t) is absolutely convergent in some half plane Re λ > β. Y (t) can be represented as t H(τ ) dτ. The Laplacetransform x ˆ(λ) therefore is a meromorphic function in C. ∆−1 (λ) = λ λ −r j=0 the series being uniformly and absolutely convergent in each half plane Re λ ≥ K + ε. ∞). Note that the elements of ∆(λ) and therefore also det ∆(λ) are entire functions.48 DELAY DIFFERENTIAL EQUATIONS Moreover. there exists a constant K > 0 such that ∞ j 1 1 0 λθ e dη(θ) for Re λ > K. Moreover. Y (t) = I + where H(t) = . which is locally absolutely continuous on [0. t ≥ 0.5 ∆−1 (λ) is the Laplacetransform for a function Y (t). φ). Theorem 1. ε > 0. Therefore the zeros of det ∆(λ) are isolated on C and consequently the elements of ∆−1 (λ) only can have poles in C.
j = 1. . λ = 0.6 The function 1 ˆ h(λ) = λ 0 −r eλθ dη(θ). the Laplaceintegral j ˆ . . t ≥ 0. . so that t ∞ t H(τ ) dτ = hj (τ ) dτ. 2. . is the Laplacetransform of h1 (t) = −η(−t). 0 j=1 0 The proof of this theorem is based on the following lemma: Lemma 1.∞ 0 j=1 hj (t) and 2 hj (t) = (−1)j η(−t) ∗ · · ∗ η(−t). · j = 1. . t ≥ 0. Consequently. converging absolutely for Re λ > 0. . . 2. jtimes The series for H(t) is uniformly and absolutely convergent on bounded tintervals. h(λ) 2“ ∗ ” here denotes the convolution of functions deﬁned on [0. . ∞) .
Integration by parts together with η(0) = 0 gives 1 λ 0 1 eλθ dη(θ) = − e−λr η(−r) − η(θ)eλθ dθ λ −r −r r ∞ 1 −λr −λθ η(−θ)e dθ = − η(−θ)e−λθ dθ. In order to show that ∆−1 (λ) is a Laplacetransform we apply Theorem A. are absolutely convergent for Re λ > 0.7). h(λ) j=1 ˆ j is a Laplacetransform with By Lemma 1. . Proof. . Also the Laplaceintegrals for hj (t).Autonomous Functional Diﬀerential Equations 49 is the Laplacetransform of the jtimes iterated convolution hj (t) of h1 (t). (1. Theorem A. The rest of the proof follows from standard results on Laplacetransforms of convolutions (cf.15) . = − e η(−r) − λ 0 0 0 ˆ Absolute convergence of the Laplacewhich proves the result for h. Thus condition (i) of Theorem A.9 is satisﬁed for any α0 > 0.5. 2.9 to ˆ H(λ) = ∞ ˆ j. Proof of Theorem 1. We claim that for any α > 0 ∞ 0 e−αt hj (t) dt ≤ 0 ∞ j e−αt h1 (t) dt . It remains to prove that ∞ j=1 ∞ e−α0 t hj (t) dt (1. j = 1. integral for Re λ > 0 is clear because η(−θ) = η(−r) for θ ≥ r. .6 each of the functions h(λ) the Laplaceintegral converging absolutely in Re λ > 0.14) 0 is convergent for an α0 > 0. .
the Laplaceintegral for H(t) is absolutely convergent for Re λ > β (note that α0 could be arbitrarily > β). O(tj−1 ) as t → ∞. Then we get ∞ ∞ t e−αt hj (t) dt = e−αt h1 (t − τ )hj−1 (τ ) dτ dt 0 0 0 t ∞ e−αt h1 (t − τ ) hj−1 (τ ) dτ dt ≤ 0 ∞ ∞ 0 e−α(t−τ ) h1 (t − τ )e−ατ hj−1 (τ ) dt dτ = ∞ 0 ∞ τ e−ατ hj−1  dτ e−ατ h1 (τ ) dτ = 0 0 ∞ j −ατ ≤ e h1 (τ ) dτ .15) is true for j − 1. Since η(−t) is bounded on t ≥ 0. t ≥ 0.9 implies that H(λ) is the Laplacetransform of H(t) = ∞ hj (t). α 0 Therefore.50 DELAY DIFFERENTIAL EQUATIONS This is proved by induction. 0 Interchanging the order because h1  ∗ hj−1  = sequence from the fact β = sup0≤t<∞ h1 (t) we of integration is justiﬁed by Fubini’s theorem. Assume that (1. say η(−t) ≤ a. Setting get ∞ β e−αt h1 (t) dt ≤ . we get. which in turn is a conthat h1 (τ ) is constant for τ ≥ r. we have convergence of the series (1. ˆ Theorem A. j=1 the series being absolutely convergent.14) if we choose α0 > β. Moreover.
This proves uniform convergence ∞ of j=1 hj (t) on bounded intervals. . . (1. hj+1 (t) ≤ a(at)j /j!. t ≥ 0. 1. . Convergence of the Laplaceintegral for H(t) in Re λ > β implies that . λ λ −1 which implies that ∆ (λ) is the Laplacetransform of t H(τ ) dτ. . ∞). To ﬁnish the proof we have to observe that 1 ˆ 1 ∆−1 (λ) = I + H(λ).16) Y (t) = I + 0 This also shows that Y (t) is locally absolutely continuous on [0. j = 0.
7 a) Y (t) is the unique solution of t Y (t) = I − Y (t − τ )η(−τ ) dτ. We next relate Y (t) to equation (1. i = 1.1). Y (0) = I and Y (t) = 0 for −r ≤ t < 0.5 t Y (t − τ )η(−τ ) dτ. Proof of Theorem 1. ∞). 0 which is (1.6 and Theorem 1.7.21 and Deﬁnition 1.18) holds everywhere on [r. r] and Y (t) is the unique solution of Y˙ (t) = L(Yt ) a. 0 H(τ ) dτ The precise abscissa of absolute convergence for the Laplaceintegral of Y (t) will be determined later (see Theorem 1.17).51 Autonomous Functional Diﬀerential Equations t = o(eαt ) as t → ∞ for any α > β (Theorem A.19) t ≥ 0. Theorem 1.e. on t ≥ 0. 0 which by Gronwall’s inequality implies Y1 (t) ≡ Y2 (t). I = Y (t) + (1. the Laplacetransform of Y˙ (t) exists and is given by λ∆−1 (λ) − I.2) and therefore we have absolute convergence of the Laplaceintegral of Y (t) in Re λ > β. ∞) with the exception of a countable subset of [0. t ≥ 0. a) From the identity ˆ I = ∆−1 (λ)∆(λ) = ∆−1 (λ) − ∆−1 (λ)h(λ) we get using Lemma 1. then Y1 (t) − Y2 (t) ≤ t sup η(θ) −r≤θ≤0 Y1 (τ ) − Y2 (τ ) dτ. t ≥ 0. (1. 0 b) L(Yt ) = −r[dη(θ)]Y (t + θ) is deﬁned on [0. Since Y˙ (t) = H(t) a. are two solutions of (1.20) .17). If Yi (t). for t ≥ 0. 2. From the identity λ∆ −1 (λ) − I = ∞ j=1 −1 ˆ ˆ ˆ j = h(λ) λ∆ (λ) − I + h(λ) h(λ) (1.19). the function Y (t) is continuously diﬀerentiable for t ≥ r and equation (1.18) Since the columns of Yt are in C for t ≥ r.17) 0 Y (t) = 0 for t < 0.e. (1.
Consequently we have Y1 (t) ≡ Y2 (t). are two solutions of (1. If Yi (t). i = 1.22) . 0] and η is continuous at −t. (1.19) and (1. 2. on t ≥ 0.e. [dη(θ)]Y (t + θ) = −t max(−t. on t ≥ 0. we get the assertion concerning L(Yt ) and equation (1.20) we can also use the identities 1 −1 ˆ I = ∆−1 (λ) − h(λ)∆ (λ) λ and ˆ ˆ + h(λ). λ∆−1 (λ) − I = λ∆−1 (λ) − I h(λ) we see that Y (t) is also the unique solution of Y (t) = I − t η(−τ )Y (t − τ ) dτ.18). −r Since η is discontinuous on an at most countable set. then 0 [dη(θ)]Y (t + θ) = −t 0 [dη(θ)]Y (t + θ). = −η(−t) − −t Therefore Y˙ (t) = 0 0 [dη(θ)]Y (t + θ) a. 1. Remarks. Since instead of (1.21) If −t is in [−r. 0 Y (t) = 0.18).−r) (1.1) with φ = 0.e.21) implies (1.52 DELAY DIFFERENTIAL EQUATIONS we get Y˙ (t) = − t η(τ − t)Y˙ (τ ) dτ − η(−t) a. then the columns of Y1 (t)−Y2 (t) are solutions of the homogeneous equation (1. t < 0. 0 Integration by parts gives t t η(τ − t)Y˙ (τ ) dτ = η(τ − t)dY (τ ) 0 0 t = η(0)Y (t) − η(−t)Y (0) − [dητ (τ − t)]Y (τ ) 0 0 [dη(θ)]Y (t + θ). t ≥ 0.
17). Y (t) is called the fundamental matrix of (1. respectively.1).8 are satisﬁed. As before we get (1.e. We write ∞ 1 0 λθ ˆ j.6. 3. ˆ e dη(θ) + h(λ) H(λ) = λ −r j=2 The ﬁrst term is a Laplacetransform by Lemma 1. Instead of applying Theorem A. (1.6 on convolutions and just diﬀerentiate equation (1. This proves that the assumptions of Theˆ orem A.8 The matrix ∆(λ) and the function det ∆(λ) are called the characteristic matrix and the characteristic function of equation (1. on t ≥ 0. If we start with (1. for any α > 0. b) one could also use Theorem A.23).18). the estimate ∞ ∞ K2 K j j ˆ h(λ) ≤ = λ λ(λ − K) j=2 j=2 for Re λ ≥ α and λ > K. Using (1.53 Autonomous Functional Diﬀerential Equations and Y˙ (t) = 0 Y (t + θ)dη(θ) −r Y (0) = I and Y (t) = 0 a.23) on t < 0. on t ≥ 0.5.e.9 for the proof of Theorem 1.1). In analogy to the theory of ordinary diﬀerential equations we introduce the following notions: Deﬁnition 1.16).12) we get. . the Laplaceintegral converging absolutely for Re λ > 0. Therefore H(λ) is a Laplacetransform of some function H(t). For the proof of Theorem 1.22) we get (1. This gives t ˙ Y˙ (t − τ )η(−τ ) dτ + η(−t) 0 = Y (t) + 0 0 [dY (t + τ )]η(τ ) + η(−t) = Y˙ (t) + = Y˙ (t) − −t 0 −t Y (t + τ ) dη(τ ) = Y˙ (t) − 0 −r Yt (θ) dη(θ) a. because this in addition provided us an explicit representation of Y (t) in terms of the matrix η.5. which is equation (1. ˆ 2.9 to H(λ) we could also proceed as follows. We decided to use Theorem A.
1) is continuously diﬀerentiable on [0. b). ∆(λ)φ(0) + p(λ. dη(θ) = ∆(λ)φ(0) + L(φ) + λ −r θ If in addition (ii) is true. 0]. 0]. φ) and continuous on (0. Then x(t.26) ∆−1 (λ)p(λ. Relation (1. ˙ φ) = x(t. φ) λ ˙ λˆ x(λ) − φ(0) = ∆−1 (λ)p(λ. but not necessarily differentiable at 0. φ) = This shows 1 ˙ .1). ∞) if and only if (i) φ ∈ C 1 ([−r. ˙ (ii) φ(0) = L(φ). Therefore we have ∞ e−λt x(t. Cn ). Then integration by parts gives 0 λ(θ−u) 1 0 ˙ e φ(u) du dη(θ) p(λ.54 1. φ) dt − φ(0) = λx∗ (λ) − φ(0) λ 0 ∞ ∞ t = e−λt y(t) dt = λ e−λt y(τ ) dτ dt 0 0 0 .26) and Theorem 1. φ) be a solution of equation (1. This is in general not true in case of functional diﬀerential equations. show that λˆ x(λ) − φ(0) is ˙ which is exponentially bounded the Laplacetransform of y(t) = x(t. φ) where x ˆ(λ) = as before denotes the Laplacetransform of x(t. ∞). t ≥ −r. φ) = φ(0) + λ −r θ 1 0 1 0 λθ (1. A solution of (1. even if the initial function is continously diﬀerentiable on [−r. φ). (1. ∞). if conditions (i) and (ii) hold. then p(λ. Suppose that (i) is true. Moreover. φ). (1.4. Theorem 1. φ) has a continuous derivative on [−r.24) Proof.4 DELAY DIFFERENTIAL EQUATIONS Smooth initial functions In case of an ordinary linear autonomous diﬀerential equation the derivatives of solutions are again solutions of the equation.9 Let x(t. φ).25) e dη(θ) φ(0) + − dη(θ) φ(θ) λ −r λ −r 0 0 λ(θ−u) 1 ˙ e φ(u) du. then ˙ x(t.
Remarks. 0 This shows that x(t. φ). relation (1. then 0 [dη(θ)]φ(t + θ) g(t) = L(φt ) = −r (1. this proves that x(t.55 Autonomous Functional Diﬀerential Equations for Re λ suﬃciently large. φ) has a continuous derivative on [0.1) (see Theorem 1. Rn ). Now suppose that x(t. ∞) and is continuous. associated with equation (1. φ ∈ dom A. t ≥ 0. This ˙ φ) at t = 0 implies φ ∈ C 1 ([−r. ∞). 2. 1. On the other hand relation (1. t ≥ 0.10 The function 0 [dη(θ)] e0 (λ) ∗ φ (θ). gˆ(λ) = − −r is the Laplacetransform of the function −t [dη(θ)]φ(t + θ) g(t) = −r 0 λ ∈ C. If we extend the deﬁnition of φ by φ(t) = 0 for t > 0. t ≥ 0. φ) has a continuous derivative on [−r. By continuity of x(t. ∞) and ˙ x(t. x(t. ˙ φ) = L(φ).e.24) means AT (t)φ = T (t)Aφ. φ) − φ(0) = y(τ ) dτ. t ≥ −r. Since Aφ = φ˙ for φ ∈ dom A. φ. which is well known for C0 semigroups. φ) resp. which implies (note that x(t.23). ˙ φ) exists on [−r. on [0. for 0 ≤ t ≤ r. ∞). Lemma 1. Since solutions of equation (1. φ and f. ˙ we get φ(0) = x(0−. 0]. Conditions (i) and (ii) characterize the domain of the inﬁnitesimal generator A corresponding to the solution semigroup T (t). ∞). φ) is continuous) t x(t. f ) in terms of the fundamental matrix and the data η. 1. ˙ φ) = x(0+. of course. for t > r. φ) a.26) shows that ˙ y(t) = x(t.8) are continuous on [−r. ˙ φ) = x(t.27) .5 The variation of constants formula In this section we prove a representation formula for x(t.
If φ(0) = 0. The upper limit of the integral can be replaced by min(t. φ) = ∆ (λ)φ(0) − ∆ (λ) −r From this (1. The representation (1. t)]φ(θ). x ˆ(λ) = ∆ (λ)p(λ. Using Fubini’s theorem we get 0 −θ 0 [dη(θ)] e0 (λ) ∗ φ (θ) = [dη(θ)] e−λτ φ(τ + θ) dτ − −r r = 0 e−λτ −τ −r 0 ∞ [dη(θ)]φ(τ + θ) dτ = −r e−λτ g(τ ) dτ.10. where Φ(θ. (1. which according to Riesz’ theorem must have a 0 representation −r[dθ Φ(θ. For the application of the convolution theorem (Theorem A. φ) is given by 0 −1 −1 −1 [dη(θ)] e0 (λ) ∗ φ (θ). 0 Theorem 1. then (1.2) is given by t Y (t − τ )L(φτ ) dτ. Proposition 1.7) we have to note that the Laplaceintegrals of Y (t) and L(φt ) are absolutely convergent in some right half plane. φ) = Y (t)φ(0) + 0 where L(φτ ) is deﬁned in Lemma 1.27) holds for all t ≥ 0. r] (in fact the subset where η(−θ) is discontinuous). Proof. 0 −t [dη(θ)]φ(t + θ) = [dη(θ)]φ(t + θ) −r −r provided η is continuous at −t. φ → 0 Y (t − τ )L(φτ ) dτ deﬁnes a bounded linear functional C → Cn .28) x(t. t)]φ(θ). for ﬁxed t ≥ 0. Proof. t For ﬁxed t ≥ 0. 0 −r .10. for φ(0) = 0 and 0 ≤ t ≤ r.27) of g follows from the fact that.12 For any φ ∈ C we have 0 t Y (t − τ )L(φτ ) dτ = [dθ Φ(θ. r). is a function of bounded variation in θ.28) follows by using Lemma 1.1).56 DELAY DIFFERENTIAL EQUATIONS for all t ≥ 0 with the exception of an at most countable subset of [0. t ≥ 0.11 (Variation of Constants Formula) The solution of problem (1. t). (1. The Laplacetransform of x(t.
−∞ 0 −∞ It is clear that the function s(τ ) is Borelmeasurable on R.t−r) t ∈ R. . τ ) = η(σ − t + τ ) for σ. ν) = V (σ. τ ) → σ + τ − t on R2 and η on R. t − r) ≤ σ ≤ t. t − r). dν p(σ. We also change.1) to the integral t t 0 Y (t − τ )L(φτ ) dτ = 0 [dτ η(σ − t + τ )]φ(τ ) dσ. τ )dk(σ) = −∞ V (σ. τ ) = 0 σ−t τ V (σ. Then we have ∞ ∞ t Y (t − τ )L(φτ ) dτ = s(τ )dτ p(σ. k(max(0. t) = t Y (t − σ)η(θ − σ) dσ. max(0. for −r ≤ τ ≤ 0. −r ≤ θ ≤ 0. for this proof the deﬁnition of η at points of discontinuity in [−r. max(0. k is of bounded variation on bounded intervals and τ → p(σ.σ−t+τ ] η for τ ≥ 0. p(σ. t − r)) for σ ≤ max(0. τ ) = − dν p(σ. 0 Y (τ ) dτ k(σ) = k(t) for σ ≥ t.1 we have to investigate σ−t+τ τ dη(θ) = var[σ−t.57 Autonomous Functional Diﬀerential Equations where Φ(θ. τ ) dk(σ). ν) = −var[σ−t+τ. 0 Proof. see also Theorem C. φ(τ ) s(τ ) = 0 for τ > 0. t) we apply the nonsymmetric Fubini theorem by Cameron and Martin ([6]. In order to get the explicit expression for Φ(θ. φ(−r) for τ ≤ −r. 0) such that η is righthand continuous on R. τ ) is of bounded variation on bounded intervals for all σ ∈ R. t ≥ 0. Borelmeasurability of p on R2 follows from Borelmeasurability of the mappings (σ. 0 By deﬁnition of k we get t ∞ V (σ.σ−t] η for τ < 0. τ ∈ R.t−r) t−r−σ We deﬁne the functions σ for max(0. In order to verify condition (i) of Theorem C. τ )Y (σ) dσ < ∞.
1 is satisﬁed. f (·. max(0. f (t. The estimates (1. t) would diﬀer only by a constant.t−r) for −r ≤ θ ≤ 0 and t ≥ 0.29) . We also have ∞ s(τ ) dτ p(σ. t1 )) ≤ 0 t1 x(t. For any t1 > 0 we deﬁne f (t) for 0 ≤ t ≤ t1 .3) we consider for the moment x(t) ˙ = L(xt ) + f (t.e.8) and (1. t ≥ 0.0] η. r) in the second integral can be replaced by t.58 DELAY DIFFERENTIAL EQUATIONS This shows that condition (i) of Theorem C. 0. x(t. σ ∈ R. t) = Y (t − σ)η(θ − σ) dσ 0 max(0. 0 (1. f (·. which does not matter. Instead of equation (1.r) Y (σ)η(σ − t + θ) dσ = Φ(θ. ˙ 0. τ ) s(τ ) −∞ 0 −∞ t 0 = −r dτ Y (σ)η(σ − t + τ ) dσ φ(τ ).t−r) i. t ≥ 0. We next turn to the nonhomogeneous problem (1. max(0.3). τ ) dk(σ) t ≤ φvar[−r. Therefore we get t ∞ ∞ dτ Y (t − τ )L(φτ ) dτ = [dk(σ)]p(σ. t ≥ 0.1 holds.4). The upper limit min(t. If is suﬃcient to consider the case φ = 0 only. −∞ and ∞ −∞ ∞ −∞ s(τ ) dτ p(σ.0] η Y (τ ) dτ < ∞. τ ) ≤ φvar[−r. (1.t−r) This proves t min(t. also condition (ii) of Theorem C. t1 ) = 0 for t > t1 . t1 )) ≤ f (τ ) dτ et . t1 ).. One easily checks that the resulting function Φ(θ.9) imply t1 f (τ ) dτ et .
t1 )) and x(t) ˙ exist. e−λt f (t) dt. t ≥ 0. the solution of problem (1. on t ≥ 0. φ) = eλt ∆−1 (λ)p(λ. t ≥ −r. f (·. t > 0.59 Autonomous Functional Diﬀerential Equations Therefore the Laplacetransforms of x(t) = x(t. Theorem A. 1.32) Therefore formula (1.7. b) implies that. Taking Laplacetransforms on both sides of equation (1. t ≥ 0. the function x(t. φ) dλ.e.6 The Spectrum As we have seen in Section 1.4) is given by t Y (t − τ )f (τ ) dτ.3).32) with a = φ(0) and therefore is inﬂuenced only by φ(0). One is the solution of (1. is the unique solution of x(t) ˙ = L(xt ) a.2) is absolutely convergent for Re λ > (see Proposition 1.3 the Laplaceintegral for any solution x(t) = x(t.13 For any φ ∈ C and f ∈ L1loc (0. t1 )) = x(t.33) with a = 0 and f (t) = L(φt ) and depends only on the past history φ [−r. (1.4). 0. f ) for 0 ≤ t ≤ t1 and t1 > 0 was arbitrary.34) x(t. x(0) = a and x(t) = 0 for −r ≤ t < 0. φ) of (1. φ) is the sum of two functions. (1. on t ≥ 0.33) where a = φ(0) and f (t) = L(φt ). a) = Y (t)a. we have proved the following theorem: Theorem 1. Therefore the complex inversion formula (cf. for any a ∈ Cn .0) .31) x(t. Cn ). Theorem 1.29) we get where fˆ(λ.28) also shows that the solution x(t. 0. t1 ) = t1 x ˆ(λ) = ∆−1 (λ)fˆ(λ. φ. t1 ).10) is applicable for any γ > and gives (1.28) can be viewed as the variation of constants formula for the nonhomogeneous problem x(t) ˙ = L(xt ) + f (t) a. This implies t Y (t − τ )f (τ . Formula (1.30) 0 t ≥ 0. the Laplaceintegrals converging absolutely for Re λ > . (1.1). x(t) = (1. t1 ) dτ. f (·. the other is the solution of (1. (1. φ) + 0 Remark. (1. (γ) . ∞. 0 Since x(t. x(0) = a and x(t) = 0 for −r ≤ t < 0. f ) = x(t.e. 0.
34) we can choose any γ > sup{Re λ  ∆−1 (λ)p(λ. c) There exists a constant ρ > 0 such that (see the shaded areas in Figure 1) σ(L) ⊂ {λ  Re λ ≥ 0 and λ ≤ ρ} ∪ {λ  Re λ < 0 and λern Re λ ≤ ρ}. then det ∆(λ) is a polynomial in λ of degree n. If σ(L) is ﬁnite. we have 0 < κ0 ≤ k0 .1). then λ0 is pole of some order κ0 for ∆−1 (λ) (i. Since the poles of ∆−1 (λ)p(λ. we shall investigate ﬁrst the location of these zeros. 2πi T →∞ −T (γ) We shall show in the next section that instead of taking γ > in (1.. φ) occur under the zeros of det ∆(λ). λ0 is pole of at least one element of ∆−1 (λ) and κ0 is the maximal order occuring).1) and ρ(L) = C \ σ(L) is the resolvent set of equation (1. Here and in the 2 following we use the notation T 1 lim f (λ) dλ = f (γ + iτ ) dτ. We ﬁrst state some basic facts about σ(L): Proposition 1.15 a) σ(L) is nonempty and all points of σ(L) are isolated in C. .14 The set σ(L) = {λ ∈ C  det ∆(λ) = 0} is called the spectrum of equation (1. φ) has a pole at λ}.e. b) If λ0 ∈ σ(L)is a zero of det ∆(λ) with multiplicity k0 . Moreover.60 DELAY DIFFERENTIAL EQUATIONS 1 For t = 0 the integral gives φ(0) and zero for t < 0. Deﬁnition 1.
Then this is also true for det ∆−1 (λ) and det ∆(λ) det ∆−1 (λ). n. . (1. j = 1. Let λ0 ∈ σ(L) be 1 adj∆(λ) it is clear a zero of multiplicity k0 .11) therefore implies that there exists a constant δ > 0 such that δ for Re λ ≥ 0. the elements of σ(L) are isolated points and zeros of ﬁnite multiplicity. a) Since det ∆(λ) is an entire function. We have to prove that λ0 is indeed a pole for ∆−1 (λ). (1. . . . n − 1. This leads to the contradiction 0 = det ∆(λ0 ) det ∆−1 (λ0 ) = limλ→λ0 det ∆(λ) det ∆−1 (λ) = 1.61 Autonomous Functional Diﬀerential Equations Proof. The estimate (1. If λ >. . i.35) where each αk (λ) is a ﬁnite sum of products of at most n of the functions 0 λθ −r e dηij (θ). . Therefore we have the estimates n λ−k  det ∆(λ) ≥ λ 1 − δ n for Re λ ≥ 0 k=1 and n e−rn Re λ λ−k  det ∆(λ) ≥ λn 1 − δ ≥ λn 1 − δ k=1 n ern Re λ λ −k for Re λ < 0. because the elements of adj∆(λ) are entire functions. .36) αk (λ) ≤ −rn Re λ δ for Re λ < 0. e k = 0. k=1 Note that e−n Re λ > 1 for Re λ < 0. From ∆−1 (λ) = det ∆(λ) that λ0 is a pole of order κ0 ≤ k0 . . b) We may write det ∆(λ) = λn + αn−1 (λ)λn−1 + · · · + α1 (λ)λ + α0 (λ). Suppose that ∆−1 (λ) is holomorphic at λ0 .
.ρ for Re λ ≥ 0 and λern Re λ > ρ for Re λ < 0 with ρ > 0 satisfying nk=1 ρ−k = 1/δ. r = . µ. µ < 0. We also see that ν behaves like ±ρe−rnµ as µ → −∞. ν = ±ρe−rnµ 1 − 2 e2rnµ ρ for λern Re λ = ρ. then we have 1/2 µ2 . In Figure 1 we show this curve for ρ = 1 and r = .35. then  det ∆(λ) > 0. If we set λ = µ + iν.5 resp. ν real.
where p(λ) is a polynomial of degree ≥ 0 and α ∈ C.37). Thus α has to be real.1). Theorem B. λ∈ / σ(L).62 DELAY DIFFERENTIAL EQUATIONS c) It remains to prove the result for ﬁnite σ(L). −r −r In view of Proposition 1. Thus we have α = 0 and pn = 1. lim eαρ = ρ→∞ pn  where pn = 0 is the coeﬃcient of λn in p(λ). p(iω) The righthand side of this inequality is bounded for ω ≥ n0 . ω real. the question arises if det ∆(λ) = det(λI − A0 ) provided σ(L) is ﬁnite. Therefore σ(L) is not empty.37) is only possible if p(λ) is a polynomial of degree n. Taking absolute values we get  det ∆(λ) . a). Suppose that σ(L) is empty or ﬁnite. which shows that det ∆(λ) is a function of exponential type (see Deﬁnition B. Therefore the assumption on σ(L) implies (cf. But (1. n0 suﬃciently large. for λ = iω. eRe(αλ) = p(λ) This together with (1. Suppose α is not real. Then (1. With σ = arg α we choose λk = keiζk where ζk = −σ + π/2 or ζk = −σ − π/2 such that Re(αλk ) = 0 and Re λk ≥ 0.37) implies that 1 .15. For k → ∞ we see that (1.37) j=0 for Re λ ≥ 0 and λ suﬃciently large. We can prove this under an additional assumption on η: . gives −ω Im α e ≤ ω + δ n n−1 ω j=0 j 1 .36) and the representation (1.7) det ∆(λ) = p(λ)eαλ . Then 0 0 eλθ dη(θ) = A0 + eλθ dη0 (θ). 0 ) by η (θ) = η(θ) for θ < 0 and by η (0) = Deﬁne the matrix η0 = (ηij 0 0 η(0−).36) implies λ −δ n n−1 −1 λ p(λ) j ≤e Re(αλ) ≤ λ +δ n j=0 n−1 λj p(λ)−1 (1. We set A0 = −η(0−). whereas the lefthand side is not.35) for det ∆(λ) we get  det ∆(λ) ≤ λn + · · · + λ + 1 ernλ for λ ∈ C. Using (1.
17 For any α ∈ R the set σ(L) ∩ {λ ∈ C  Re λ ≥ α} is ﬁnite. .0] ηij . If σ(L) is ﬁnite.18 Let α ∈ R be given. c). 0 0 (θ). j = 1. The In each product at least one factor is of the form −r eλθ dηij assumption on η0 implies 0 0 0 eρθ dηij (θ) ≤ e−ρδ var[−r. j = 0. b) There exists a positive constant M = M (α) such that p(λ. . ρ > 0. i. n − 1. (1.35) can be written as αj (λ) = aj + βj (λ). then det ∆(λ) = det(λI − A0 ).15. .39) implies that. .39) −r From q(λ) = βn−1 (λ)λn−1 + · · · + β1 (λ)λ + β0 (λ) = det ∆(λ) − det(λI − A0 ) we see that q(λ) is a polynomial of degree ≤ n − 1. Then the following is true: a) There exist positive constants K = K(α) and β = β(α) such that ∆−1 (λ) ≤ K λ for λ ≥ β and Re λ ≥ α. φ) ≤ M φ for Re λ ≥ α and φ ∈ C. . An immediate and important consequence of Proposition 1.63 Autonomous Functional Diﬀerential Equations Proposition 1.38) where the aj are the coeﬃcients of det(λI − A0 ) and the βj (λ) are ﬁnite 0 sums of ﬁnite products involving elements of A0 and of −r eλθ dη0 (θ). for a constant K > 0.16 Assume that there exists a δ > 0 such that η0 (θ) = η0 (−δ) for θ ∈ [−δ. . 0]. q(ρ) ≤ Ke−ρδ ρn−1 for ρ ≥ 1. We shall need the following estimates: Lemma 1. The coeﬃcients αj (λ) in (1. This implies that q(λ) is the zero polynomial. n. . (1. . and the fact that σ(L) cannot have an accumulation point in C is the following result: Corollary 1. Proof. . On the other hand the estimate (1.
64 DELAY DIFFERENTIAL EQUATIONS Consequently. . φ) = ∆−1 (λ)p(λ. φ) ≤ (1. .22).36) shows that for some constant σ0 = σ0 (α) > 0 we have αk (λ) ≤ σ0 for Re λ ≥ α and k = 0.40) for α < 0. φ) deﬁnes a bounded linear n functional C → C for all λ ∈ σ(L). 1/α) φ for α ≥ 0. . From (1. 1 + min(r. Then n λ−k  det ∆(λ) ≥ λn 1 − σ0 for Re λ ≥ α. b) Indeed. The names for ωL and ωL. Proof. φ) of (1. 1/α)φ α(θ−τ ) e0 (λ) ∗ φ (θ) ≤ φ e dτ ≤ reαr φ θ for α ≥ 0.11) we infer that there exists a constant K = K(α) such that adj∆(λ) ≤ K n−1 λ 2 for Re λ ≥ α and λ ≥ β. k=1 Therefore there is a constant β = β(σ0 ) such that 1  det ∆(λ) ≥ λn 2 for Re λ ≥ α and λ ≥ β.φ will be justiﬁed later (see Corollary 1.2). 0 min(r. p(λ. . φ)} is called the exponential type of the solution x(t. φ) ≤ φ + e0 (λ) ∗ φ for all λ ∈ C and φ ∈ C.1). for Re λ ≥ α.φ ≤ ωL for all φ ∈ C. for Re λ ≥ α (note that θ − τ ≤ 0). the number ωL. It is clear that ωL. . φ → x ˆ(λ.φ = sup{Re λ  λ is a pole of ∆−1 (λ)p(λ. For any φ ∈ C. Furthermore. n − 1. 1 + reαr φ Deﬁnition 1. p(λ. for α < 0. (1.1). Therefore we have ∆−1 (λ) ≤ K λ for Re λ ≥ α and λ ≥ β. a) The estimate (1. This proves that.19 The number ωL = sup{Re λ  λ ∈ σ(L)} is called the exponential type of equation (1.
we have ωL. (1. φ) ≤ K λ for γ ≤ Re λ ≤ γ0 and  Im λ ≥ ρ˜.42) (γ) Proof. x(t.41) (γ) b) For any γ > ωL. φ) dλ = 0.φ < γ ≤ γ0 and < γ0 .44) . According to the deﬁnition of ωL. 2. Y (t) = eλt ∆−1 (λ) dλ for t > 0.φ we have for all T > 0 eλt ∆−1 (λ)p(λ. φ) = eλt ∆−1 (λ)p(λ. (1.18 there ˜ and ρ˜ such that exist positive constants K ˜ λt −1 e ∆ (λ)p(λ. (1. We only prove part b). This shows eλt ∆−1 (λ)p(λ. φ) dλ = 0. (1. Of course.43) Γ where Γ is the closed contour depicted in Figure 2. φ) dλ for t > 0.φ .20 a) For any γ > ωL . lim T →∞ Γi i = 1.65 Autonomous Functional Diﬀerential Equations Proposition 1. By Lemma 1.
which implies that the Laplaceintegral ˜ x(t. Theorem 1. By choice of γ0 .φ is the smallest number such that for any ε > 0 ˜ = K(ε) ˜ there exists a constant K ≥ 1 such that for any ψ ∈ C with ωL. Theorem A.φ − γ)/2 we have γ + ε < ωL. the Laplaceintegral for x(t.φ . t ≥ 0. We now are able to prove that the numbers ωL and ωL. This proves (1. Indeed.46) is true for a γ < ωL.ψ ≤ ωL. φ) dλ.ψ ≤ ωL. ψ) = eλt ∆−1 (λ)p(λ.φ cannot be replaced by a smaller number.10) eλt ∆−1 (λ)p(λ. assume that (1. ψ) = ∆−1 (λ)−(λ−ωL.φ . (γ) According to the identity ∆−1 (λ)p(λ. (1.φ +ε)t ˜ x(t. φ) converges absolutely for Re λ > γ. φ) is absolutely convergent for Re λ ≥ γ0 and therefore (cf.21 a) The number ωL is the smallest number such that for any ε > 0 there exists a constant K = K(ε) ≥ 1 such that Y (t) ≤ Ke(ωL +ε)t for t ≥ 0. x(t. φ) for t > 0. ψ) . for ε = (ωL. ψ) ≤ Kψe for t ≥ 0. φ) ≤ K(ε)φe for x(t. t > 0. ψ) dλ.φ (ωL. We now choose a ψ ∈ C with ωL.43) and (1. ψ) −r + (λ − ωL.φ . (1. For ε > 0 we set ωL. φ) = (γ0 ) By (1.5.φ ) ∆ (λ) −ωL. respectively. t > 0. a contradiction to the deﬁnition of ωL.45) b) The number ωL.φ determine the exponential growth of Y (t) and x(t. Proof.φ and (γ+ε)t .44) also (γ) eλt ∆−1 (λ)p(λ. φ) dλ exists and equals x(t. φ). where β is the number appearing in Theorem 1.46) There always exists a φ ∈ C such that ωL.φ I + eλθ dη(θ) p(λ.φ )−1 p(λ.42) x(t.φ = ωL . Then. Again we just prove b).φ )−1 I p(λ.φ +ε = γ and get from (1. In the proof concerning Y (t) we have to choose γ0 > β.66 DELAY DIFFERENTIAL EQUATIONS where Γ1 and Γ2 are the horizontal parts of Γ (see Figure 2).φ )−1 p(λ. ψ) 0 −1 −1 = (λ − ωL. We show ﬁrst that ωL. ψ)+(λ−ωL.42).
φ t ∗L(ψt ) (see Lemma 1. Therefore we have t eωL.φ (t−τ ) L(ψτ ) dτ ≤ min(t.47) and (1. ψ) dλ −r (γ) and I2 = eλt (λ − ωL.φ  + N M ψ = 2 ψeγt .11) there exist positive constants K = K(γ). I2 = ψ(0)eωL.φ )−1 ψ(0) is the Laplacetransform of eωL.φ t this is trivial and for the convolution it is valid because it is true for both factors.φ (t−τ ) L(ψτ ) dτ. N = N (γ) and β such that for λ = γ + iτ λt e (λ − ωL.φ )−1 L(e0 (λ) ∗ ψ) is the Laplacetransform of the convolution eωL.r) eωL.φ )−1 ∆−1 (λ) −ωL.48) I2  ≤ eωL. With respect to I2 we observe that (λ − ωL. t > 0. Therefore we get (1.φ t L(ψτ ) dτ ≤ κ2 eωL. All Laplaceintegrals are converging absolutely for Re λ > ωL.φ t ψ(1 + κ2 ) for t ≥ 0. M = M (γ).φ t ψ(0) and −(λ − ωL.18 and (1.φ I + 0 eλθ dη(θ) p(λ.φ t + 0 For the integral we get (note that L(ψτ ) = 0 for τ > r) t eωL.φ ) ∆ (λ) −ωL. (γ) By Lemma 1.φ . .67 Autonomous Functional Diﬀerential Equations we have to estimate λt −1 −1 I1 = e (λ − ωL.47) where κ1 = κ1 (γ) is some positive constant. For eωL. This implies that I1  ≤ κ1 ψ eγt . 0 where κ2 is a positive constant which does not depend on ψ. (1. ψ) dλ.φ t ψ.φ I + eλθ dη(θ) p(λ. the desired result follows from (1.φ (t−τ ) L(ψτ ) dτ 0 0 r ≤e ωL. t ≥ 0. Since x(t.φ )−1 p(λ. ψ) = I1 + I2 . ψ) 0 −r ˜ K 1 K ωL.48).10). ≤ eγt τ  τ  τ τ  ≥ β.
23 a) The inﬁnitesimal generator A of S is given by dom A = {φ ∈ C 1 (−r. t The proof of this corollary is an immediate consequence of Theorem 1.1) if and only if det ∆(λ0 ) = 0 and b is a nonzero solution of ∆(λ0 )b = 0. Since x(t. φ ∈ dom A.68 DELAY DIFFERENTIAL EQUATIONS An easy calculation shows that x(t) = eλ0 t b is a nontrivial solution of equation (1. i. Theorem 1. we also have T (t + s) = T (t)T (s). Aφ = φ .. c) The resolvent operator of A is given by ψλ = (λI − A)−1 φ. T ]. φ) = ωL. which proves that S is a C0 semigroup. T ≥ −r. From Proposition 1. Since xt+s (φ) = xt (xs (φ)) for t. Therefore. it is clear that xt (φ) − φ = sup x(t + θ. Cn )  φ (0) = L(φ)}.3 (for f ≡ 0) we see that T (t) is a bounded linear operator. φ ∈ C. lim sup t→∞ 1 ln x(t. φ ∈ C. the family T = T (t) t≥0 is a semigroup. 1.7 The solution semigroup For any t ≥ 0 we deﬁne the operator T (t) : C → C by T (t)φ = xt (φ). φ) − φ(θ) → 0 −r≤θ≤0 as t ↓ 0. we have at least one solution whose exponential type is Re λ0 (note that L(eλ0 t ) = (λ − λ0 )−1 ). (1.e.21.22 We have lim sup t→∞ 1 ln Y (t) = ωL t and. φ) is uniformly continuous on intervals [−r. 0. for any φ ∈ C. s ≥ 0. for any λ0 ∈ σ(L).φ . s ≥ 0. Corollary 1. t. b) The spectrum σ(A) of A is all point spectrum and is given by σ(A) = σ(L) = {λ ∈ C  det ∆(λ) = 0}.49) .
For the moment let B denote the inﬁnitesimal generator of T .69 Autonomous Functional Diﬀerential Equations where λθ ψλ (θ) = ψλ (0)e + 0 eλ(θ−s) φ(s) ds. Cn ). −r ≤ θ ≤ 0. In this case we have ˆ(λ. θ ψλ (0) = ∆(λ)−1 p(λ. is equivalent to ψ(θ) = ψ(0)eλθ . We ﬁrst compute the resolvent operator for the operator deﬁned by (1. The equation (λI − A)ψλ = φ. φ) = x ˆ(λ. and ∆(λ)ψ(0) = 0). Then we have (see for instance [28]) (λI − B)−1 φ = 0 ∞ e−λt T (t)φ dt. ψ ∈ dom A. Therefore we have shown that λ ∈ σ(A) if and only if det ∆(λ) = 0. equivalently. ψλ ∈ dom A. . θ The function ψλ is in C 1 (−r. −r ≤ θ ≤ 0. Thus parts b) and c) are proved. λ ∈ σ(A) implies that λI − A is not injective ((λI−A)ψ = 0. We have to choose ψλ (0) such that ψλ ∈ dom A. ∆(λ)ψλ (0) = φ(0) + L · This equation can be solved for ψλ (0) if and only if det ∆(λ) = 0. It remains to prove that A is the inﬁnitesimal generator of T . i. b) we see that φ → ψλ (0) deﬁnes a bounded linear functional C → Cn . implies ψλ (θ) = λψλ (θ) − φ(θ). λψλ (0) − φ(0) = L(ψλ ) = L(e I)ψλ (0) + L λ· 0 eλ(·−s) φ(s) ds · or. Proof. φ ∈ C. 0 λ(·−s) e φ(s) ds = p(λ.49). ψλ (0) = ∆(λ)−1 p(λ. φ ∈ C. φ). Re λ > ωL . and λθ ψλ (θ) = ψλ (0)e 0 + eλ(θ−s) φ(s) ds. φ).e. φ) = x From Lemma 1.. −r ≤ θ ≤ 0. φ). Moreover.18. This implies that φ → ψλ is a bounded linear operator C → C. 0.
−r ≤ θ ≤ 0. . A = B.e.70 DELAY DIFFERENTIAL EQUATIONS where the integral is understood as an improper Riemann integral. which proves the result. −r ≤ θ ≤ 0. 0] we have ∞ ∞ −λt e xt (φ) dt (θ) = e−λt x(t + θ. m(φ) = −r Then we have m ∞ −λt e 0 ∞ m e−λt xt (φ) dt 0 ∞ 0 x(t + θ. φ)e + θ = (λI − A)−1 φ (θ). is a function in C. This proves (λI − A)−1 = (λI − B)−1 . 0 Proof. Let m be a bounded linear functional on C which is represented by the vector µ of bounded variation on [−r. i. φ) dµ(θ) dt. φ) dt θ 0 0 λθ eλ(θ−s) φ(s) ds =x ˆ(λ. 0]. 0 Re λ > ωL . φ) dt. On the other hand a simple integration using also part c) shows that ∞ ∞ ∞ −λt −λt e T (t)φ dt (θ) = e xt (φ) dt (θ) = e−λt x(t + θ. φ) dt = e−λt x(t + θ. 0 φ(θ)dµ(θ). Using Fubini’s theorem we get 0 ∞ ∞ e−λt x(t + θ.. φ) dt 0 for all bounded linear functionals m on C. φ ∈ C. = xt (φ) dt = 0 ∞ −r It is not diﬃcult to prove that θ → 0 e−λt x(t + θ. φ) dt 0 0 0 ∞ 0 = eλ(θ−s) φ(s) ds + eλθ eλt x(t. For the proof of A = B we have also used the following lemma: Lemma 1. Re λ > ωL . Re λ > ωL . φ) dµ(θ) dt. φ) dt dµ(θ) m −r 0 ∞ 0 0 = −r 0 Thus we have shown that ∞ −λt m e xt (φ) dt = m 0 e−λt x(t + θ.24 For any θ ∈ [−r. ∞ e−λt x(t + ·. φ) dt.
. . is a solution of (1. bk ) ∈ Cn(k+1) . λ0 ∈ C. ··· ˜bk+1 = col (b0 . . (ii) ∆ where ˜ k+1 (λ0 ) = ∆ ∆(λ0 ) 0 . ˜ k+1 (λ0 )˜bk+1 = 0. ··· ··· . is a solution of (1. k. . . . ··· . . 1 1! ∆ (λ0 ) . 0 1 (k) k! ∆ (λ0 ) ∈ Cn(k+1)×n(k+1) . .. b1 ∈ Cn . . .71 Autonomous Functional Diﬀerential Equations 2. . .1) on R if and only if (i) λ0 ∈ σ(L). . ... ..1 Generalized eigenspaces Consider equation (1. b0 . −r which in turn is equivalent to ∆(λ0 )b1 = 0. be given. . Eigenspaces 2.1) if and only if det ∆(λ0 ) = 0 and ∆(λ0 )b0 = 0.. b0 ∈ Cn . λ0 ∈ C. 1 1! ∆ (λ0 ) ∆(λ0 ) . ∆(λ0 )b0 + ∆ (λ0 )b1 = 0 0 (observe that ∆ (λ0 ) = I − −r θeλ0 θ dη(θ)). . A function x(t) = eλ0 t b0 . Similarly a function x(t) = eλ0 t b0 + tb1 ... The function k tj λ0 t bj ≡ 0 (2.1) if and only if 0 t∆(λ0 )b1 + I − θeλ0 θ dη(θ) b1 + ∆(λ0 )b0 ≡ 0.1) x(t) = e j! j=0 is a solution of (1. In general we have: Proposition 2. .. . . . j = 0. .1 Let λ0 ∈ C and bj ∈ Cn .. −r φ ∈ C...1) with 0 L(φ) = [dη(θ)]φ(θ). 0 ˜bk+1 = 0. . . . .
φ) = e k tj j=0 j! bj . is called the generalized eigenspace of (1. a) Assume that. φ) is x ˆ(λ. ∆ The solution x(t. φ) is given by λ0 t x(t. i. where ∆ 0 0 Pλ0 = φ ∈ C φ(θ) = e λ0 θ (2.3 Let λ0 ∈ σ(L) be given. for a φ ∈ Pλ0 . j! j=0 ˜ κ (λ0 )˜bκ = 0 . we have k ≥ κ0 .2 Let λ0 ∈ σ(L) be given. For λ0 ∈ σ(L) let k0 be the multiplicity of λ0 as a root of det ∆(λ) and κ0 be the order of λ0 as a pole of ∆−1 (λ). Pλ0 = φ ∈ C ∃k ≥ 0 : φ(θ) = eλ0 θ j! j=0 ˜ ˜ where ∆k+1 (λ0 )bk+1 = 0 −r ≤ θ ≤ 0. b) dim Pλ0 = nκ0 − rank ∆ 0 Proof. .9).1) with ˜bk satisfying (ii) of Proposition 2.1) corresponding to λ0 is given by κ 0 −1 θj bj .2) ˜ κ (λ0 ). φ(θ) = eλ0 θ k θj j=0 j! bj . The Laplacetransform of x(t.. φ) = k j=0 3 Below 1 bj .e. −r ≤ θ ≤ 0.1) corresponding to λ0 .1 is called a generalized eigenfunction of equation (1. The subspace k θj bj . A function x(t) given by (2.1) corresponding to λ0 .3 Proposition 2. We ﬁrst show that in the deﬁnition of Pλ0 we can set k = κ0 − 1. a) The generalized eigenspace of (1.72 DELAY DIFFERENTIAL EQUATIONS Deﬁnition 2. We know already that 0 < κ0 ≤ k0 . (λ − λ0 )j+1 we shall prove that κ0 cannot be replaced by a smaller number (see Corollary 2. ˜ k+1 (λ0 )˜bk+1 = 0.
. φ)). From ∆ ˜ κ (λ0 )˜bκ = 0. . 0). we deﬁne bk+1 = bk+2 = · · · = ˜ k+1 (λ0 )˜bk+1 = 0 we get that also bκ0 −1 = 0. . φ) = ∆−1 (λ)p(λ. From ∆ diately get ˜ κ (λ0 )˜bκ = 0. . ∆ 0 0 .e. we must have bκ0 = bκ0 +1 = · · · = bk = 0. . ∆ 0 0 If φ ∈ Pλ0 is deﬁned with k < κ0 − 1. ˜bk = col(b0 . . ˜ k+1 (λ0 )˜bk+1 = 0 we immei. bκ0 −1 . 0. φ) of order at most κ0 (observe that x ˆ(λ.Autonomous Functional Diﬀerential Equations 73 Since λ0 can be a pole of x ˆ(λ. .. .
be elements in Pλ0 and b) Let φi (θ) = eλ0 θ j=0 j! bj .κ0 −1 θj (i) . . m. i = 1. . .
αm ∈ C. . The equation m i=1 αi φi = 0 is equivalent to κ 0 −1 j=0 m θj (i) αi bj ≡ 0. . . . . . . . . . . ˜bκ0 ∈ ker ∆ 0 independent. ρ = 1. dρ . . . κ0 − 1. . . ρ = 1. . ˜ κ (λ0 ) we have to use an elementary In order to determine rank ∆ 0 divisor theory for holomorphic matrices. . θ2 /2!. with 0 ≤ d 1 < · · · < d and mρ > 0. We start with a fundamental theorem. i=1 The last equations are equivalent to m αi˜b(i) κ0 = 0. . φm ∈ Pλ0 are linearly indepen(1) (m) ˜ κ (λ0 ) are linearly dent if and only if the vectors ˜bκ0 . Theorem 2. . . then there exist integers . . j! i=1 which by linear independence of the functions 1. . i=1 This proves that the elements φ1 . . . . The result on dim Pλ0 follows immediately. . . . . . . If det K(λ) ≡ 0 in V . θ. .4 Let K(λ) be an n × nmatrix which is holomorphic in a neighborhood V of zero. . θκ0 −1 /(κ0 − 1)! is in turn equivalent to m (i) αi bj = 0 for j = 0. . . . . mρ . choose α1 .
H0 = 0. . λd M (λ) in U . (ii) F (λ). Proof. . ρ = 1. . (iv) the numbers . n1 ≥ 0. . dρ are uniquely determined by properties (i) – (iii). . . The power series expansion of K(λ) around 0 has the following form: K(λ) = λn1 H(λ). G(λ) are holomorphic in U and det F (λ) ≡ det G(λ) ≡ 1 in U. (iii) Mρ (λ). G(λ) and a neighborhood U of zero such that (i) M (λ) = F (λ)K(λ)G(λ) = diag λd1 M1 (λ).74 DELAY DIFFERENTIAL EQUATIONS n × nmatrices F (λ). mρ . . . . is an mρ × mρ matrix. which is holomorphic in U with det Mρ (λ) = 0 in U . .
Case 2. Assume that det H0 = 0. Let X. 0 0 where H 00 is an m1 × m1 matrix with det H00 = 0. d1 = n1 . Y be n × nmatrices with det X = det Y = 1 such that H00 0 XH0 Y = . In this case we have F (λ) ≡ G(λ) ≡ I. Assume that det H0 = 0. = 1. Case 1. Observing H(λ) = . Let m1 = rank H0 . From H0 = 0 and det H0 = 0 we see that 0 < m1 < n. because K(λ) = λn1 H(λ) satisﬁes already (i)–(iii). m1 = n. j where H(λ) = ∞ j=0 λ Hj .
. j−1 H we get H0 + λ ∞ j j=1 λ XH(λ)Y = H00 0 H11 (λ) H12 (λ) . +λ 0 0 H21 (λ) H22 (λ) where the Hij (λ) are matrices which are holomorphic in V . We set M1 (λ) = H00 + λH11 (λ).
d2 = n 1 + n 2 + 1 > d1 . . F1 (λ)K(λ)G1 (λ) = 0 λn1 +1 K2 (λ) ˜ . . n − 1. m2 = n − m1 . It remains to prove (iv). Since a (ν + 1) × (ν + 1)minor is a linear combination of ν × νminors (Laplace expansion theorem for determinants). From the of 0. . (ii) There exists at least one ν × νminor of N (λ) such that q(0) = 0.4 leave the numbers βν invariant. ν = 1. j=0 (2) If det H0 = 0.75 Autonomous Functional Diﬀerential Equations ˜ From det H00 = 0 we conclude that det M1 (λ) = 0 in a neighborhood U ˜ and is holomorphic there. The matrix K2 (λ) is holomorphic where det F1 (λ) ≡ det G1 (λ) ≡ 1 in U ˜ in U and has a power series expansion K2 (λ) = λ n2 ∞ (2) λj H j (2) with H0 = 0. . For an n × nmatrix N (λ) which is holomorphic in a neighborhood U of zero we introduce the numbers βν (N ). we see that n 0 λ 1 M1 (λ) . where q(λ) is holomorphic in U . In a ﬁnite number of steps we arrive at M (λ) as given in the theorem. λ ∈ U. we set m2 = rank H0 and proceed as above in Case 2. ν = 1. n. we have βν+1 (N ) ≥ βν (N ). as follows: (i) Any ν × νminor of N (λ) has the form λβν (N ) q(λ). The following lemma shows that equivalence transformations of the form appearing in Theorem 2. Therefore M1 (λ)−1 exists in U equation I 0 M1 λH12 M1 0 I −λM1−1 H12 = . . (2) (2) If det H0 = 0. −1 λH21 λH22 0 λK2 I 0 I −λH21 M1 where K2 (λ) = H22 (λ) − λH21 (λ)M1 (λ)−1 H12 (λ). =2 and M2 (λ) = K2 (λ). . . In order to do this we have to introduce the notion of determinantal divisors for a holomorphic matrix. . then we have d1 = n1 . .
. ....4... . kν of K(λ).. . Using the special form of M (λ)... ν = 1.kν . . . . . βm1 = m1 d1 . j = 1.kν denotes the ν × νminor containing the elements in rows j1 .. .. λ ∈ U ....... ... βm−1 +1 = −1 mρ dρ + d ... Starting from K(λ) = F (λ)−1 M (λ)G(λ)−1 we get βν (K) ≥ βν (M ). λd M (λ) we see that (observe det Mj (0) = 0.j ... ) β1 = d1 . βm1 +1 = m1 d1 + d2 .. ..αν K σ1 .kν α1 . . . . ..4. .. . . . bm1 +m2 = m1 d1 + m2 d2 .jν Proof. . then the Bin´etCauchy formula gives (see for instance [8]) j1 .. . ...5 Let K(λ) be an n×nmatrix being holomorphic in a neighborhood U of 0. . Consequently we have βν (M ) ≥ βν (K).. . . ..jν k1 . .. . βn−1 = ρ=1 −1 mρ dρ + (m − 1)d ..kν = λβν (K) p(λ). . where p(λ) is holomorphic at zero. n. β2 = 2d1 .5 we conclude that βν := βν (K) = βν (M ).σν G k1 . .jν α1 . From Lemma 2. βn ρ=1 = ρ=1 mρ dρ .. 1≤α1 <···<αν ≤n 1≤σ1 <···<σν ≤n This shows that M j1 .. n.σν ν F = M kj11 . d λ 1 M1 (λ) λd2 M2 (λ) M (λ) = 0 ν = 1. where F (λ) and G(λ) satisfy (ii) of Theorem 2.. . . .. .76 DELAY DIFFERENTIAL EQUATIONS Lemma 2... Let M (λ) = F (λ)K(λ)G(λ). j1 ... ... 0 . . .... Then we have βν (M ) = βν (K). ..αν σ1 .. jν and columns k1 . Proof for (iv) of Theorem 2. If K k1 ..
. compute the sequence β2 − β1 ≤ β3 − β2 ≤ · · · ≤ βn − βn−1 . . . m−1 are those indices. . . dρ are uniquely determined by the numbers β1 . From this it is clear that the numbers mρ . . k = 1. . . ρ = 1. Given β1 . . m2 . . . n − 1. . βn . Indeed. βn . The numbers m1 . − 1 (the number of strict inequalities in the above sequence is − 1). The . . βmρ +1 − βmρ > βmρ − βmρ −1 . . m−1 . . . . we have to proceed as follows: Step 1. . . . . where a strict inequality occurs in this sequence. where the “>”sign is valid exactly for k = m1 .77 Autonomous Functional Diﬀerential Equations From this we also see that (using d1 < d2 < · · · < d ) βk+1 − βk ≥ βk − βk−1 . . . .
. . ρ = 1. . . . The following lemma relates the matrices B−µ B−µ+1 · · · B−µ+k−1 .. ∈ Cnk×nk . are called the determinantal divisors of K(λ) at λ = 0. . . βn . . . . n. . . . . Remark. . mρ . .. .4). ν = 1. n.. . . n. k = m−1 + 1. Step 2. − 1. . . . . number m is given by n − −1 ρ=1 mρ . εν (λ) = λβν −βν−1 . The polynomials δν (λ) := λβν . .. B−µ+1 0 ··· 0 B−µ (2. . . . dρ . ν = 1. . The numbers dρ are given by (m0 := 0) dρ = βk − βk−1 . n − 1. . are uniquely determined by β1 . . mρ . . From the relations between the numbers βν we conclude that δν  δν+1 and εν  εν+1 . . . whereas the polynomials ε1 (λ) ≡ 1. ˜k (λ0 ) = B . are the elementary divisors of K(λ) at λ = 0. d = βk − βk−1 . ν = 2. ρ = 1. . . 0 .. . . . we have proved (iv).3) . Since the numbers . . . . k = mρ−1 + 1. which in turn are uniquely determined by the equivalence class of K(λ) (with respect to equivalence transforms satisfying (ii) of Theorem 2. ..
Then according to Lemma 2.. ˜k (λ0 ) in several steps.4 (with λ replaced by λ−λ0 of course). The proof is obvious if we compute the Laurent series expansions on both sides of H(λ) = F (λ)G(λ) around λ0 .6 will be needed below.. Then the following two statements are equivalent: (i) H(λ) = F (λ)G(λ) in a neighborhood U of λ0 .. . ∞ B(λ) = (λ − λ0 )j Bj . . .. . . ··· ··· . and the Laurent series of the matrix B(λ) around λ0 . 1 1! M (λ0 ) M (λ0 ) ∈ Cnk×nk . . . 0 1 (k−1) (λ ) 0 (k−1)! M .6 for µ = ν = 0. . ... det F˜k (λ0 ) = det F (λ0 ) The general form of Lemma 2... Note that in this case k etc. .. . . . . where M (λ0 ) 0 ˜ k (λ0 ) = M . G(λ) and H(λ) be n × nmatrices with Laurent series F (λ) = ∞ (λ − λ0 ) Fj . For the proof of Theorem 2. . Let M (λ) be the matrix We now compute rank B obtained from B(λ) according to Theorem 2. . . . ˜ k (λ0 ) for all k = 1. . . . . . 2.6 we have ˜k (λ0 ) = rank M ˜ k (λ0 ). . . . 0 1 1! M (λ0 ) .6 Let F (λ). .. 2. rank B k = 1. 2. ··· ··· . j G(λ) = j=−µ ∞ (λ − λ0 )j Gj j=−ν H(λ) = ∞ (λ − λ0 )j Hj j=−µ−ν around λ = λ0 . ..78 DELAY DIFFERENTIAL EQUATIONS k = 1. j=−µ Lemma 2.. ˜ k (λ0 ) = F˜k (λ0 )G (ii) H Proof.4 we only need Lemma 2. . One also has to observe that the coeﬃcient matrix of λ−µ+j−1 is the block appearing in the jth diagonal of F˜k (λ0 ) etc.
7 k−1 1 (j) ˜ k (λ0 ) = rank M (λ0 ) ⊗ Rkj . then we have ˜ k (λ0 ) = M k−1 Rkj ⊗ j=0 1 (j) M (λ0 ). . 0 and R1 = (1). .. . 118]) P −1 (S ⊗ T )P = T ⊗ S. k×k 0 ∈ C 1 ··· . j=0 Proof. . . . . . . . . Lemma 2. . Rk = . . . . Since det P = 0.. Y is deﬁned by X ⊗ Y = xij Y .. 0 ··· ··· 0 . . . .. .. . . 2. j! k = 1. . the result is established. . where the Kronecker product of two matrices X = (xij ).. . . rank M j! k = 1. 3. . p.. . . . .. . For any positive integers k. 2.79 Autonomous Functional Diﬀerential Equations If we introduce the matrices 0 1 0 . n there exists an nk×nk permutation matrix P such that for any matrices S ∈ Cn×n and T ∈ Ck×k we have (see for instance [9. ··· for k = 2.
...n .. k−1 1 (j) (λ ) ⊗ Rj it is useful to introIn order to compute rank 0 j=0 j! M k duce the nk × nkmatrix M (λ0 I + Rk ) = mij (λ0 I + Rk ) i.. where M (λ) = (mij (λ))i...n and mij (λ0 I + Rk ) is deﬁned by the .j=1.j=1..
n. (ij) power series expansion of mij (λ) around λ0 . mij (λ) = ∞ µ=0 αµ (λ − λ0 )µ . i. . That means (observe that Rkµ = 0 for µ ≥ k) mij (λ0 I + Rk ) = k−1 αµ(ij) Rkµ . . M (λ0 I + Rk ) = j! j=0 . j = 1. µ=0 (ij) Since we have αµ = (µ) 1 µ! mij (λ0 ). . we get k−1 1 (j) M (λ0 ) ⊗ Rkj . .
. dρ ..80 DELAY DIFFERENTIAL EQUATIONS Observing the special form of M (λ) we get M (λ0 I + Rk ) = diag (Im1 ⊗ Rkd1 )M1 (λ0 I + Rk ). . .. ∗ 0 · · · 0 Mρ (λ0 ) = k rank Mρ (λ0 ) = kmρ we see that d d rank(Imρ ⊗ Rkρ )Mρ (λ0 I + Rk ) = rank(Imρ ⊗ Rkρ ). . . . d1 . . ρ = 1. 2 . . . k − 1. (Im ⊗ Rkd )M (λ0 I + Rk ) . . Thus we have rank M (λ0 I + Rk ) = d rank(Imρ ⊗ Rkρ ) = ρ=1 d mρ rank Rkρ . Then we have rank M (λ0 I + Rk ) = q mρ (k − dρ ). . . . . If d1 > 0. . Theorem 2. . . . . mρ . ρ=1 Thus we have proved the following theorem. .. . . From k−1 1 Rkj ⊗ Mρ(j) (λ0 ) rank Mρ (λ0 I + Rk ) = rank j! j=0 ∗ Mρ (λ0 ) ∗ · · · ... 0 = rank . for dρ = 1. For k > d1 deﬁne q = max{ρ ∈ N  dρ < k}. 2. . ρ=1 Obviously we have d rank Rkρ 0 = k − dρ for dρ ≥ k. . . . . then rank M (λ0 I + Rk ) = 0 for k = 1. . .8 Let B(λ) be an n × nmatrix which is holomorphic at λ0 and let . be the numbers uniquely determined by .
. . Then we have 0 for k = 1. Denote by k0 the multiplicity of λ0 as a root of det B(λ) and by κ0 the order of λ0 as a pole of B(λ)−1 .81 Autonomous Functional Diﬀerential Equations B(λ) according to Theorem 2.4. . . d1 (if d1 > 0). .
ρ=1 The latter equation follows from the special form of M (λ). .q m (k − d ) for k = d + 1.9 We have nk . . . Corollary 2. Concerning the result for k ≥ d we have to note that mρ = n and ρ=1 mρ dρ = k0 . . for k ≥ d . . Furthermore. from M (λ)−1 = G(λ)−1 B(λ)−1 F (λ)−1 and the fact that G(λ)−1 . nk − k0 Moreover. . − 1. F (λ)−1 are holomorphic at λ0 we see that κ0 = d . . . we have κ0 = d . d . ρ q q+1 ρ=1 ρ ˜k (λ0 ) = rank B q = 1.
˜ k (λ0 ) = k .
. for k = dq + 1. We only have to observe that nk − q mρ (k − dρ ) = k ρ=1 mρ − ρ=1 =k ρ=q+1 q mρ (k − dρ ) ρ=1 mρ + q ρ=1 mρ dρ . κ0 in the representation (2. (if d1 > 0). . dim ker ∆ mρ + qρ=1 mρ dρ ρ=q+1 k for k = 1. . . − 1. . . . Proof. . The proof is obvious. for k = κ0 . 0 In particular we have dim Pλ0 = k0 . . Furthermore. . . . . .2) of Pλ0 cannot be replaced by a smaller number. . κ0 + 1. d1 . . dq+1 . . q = 1.
which proves the claim on κ0 .82 DELAY DIFFERENTIAL EQUATIONS ˜ k (λ0 ) = (d −1)m + For k = κ0 −1 = d −1 we get dim ker ∆ k0 − m < k0 . .
ψ) = − col pk−1 (λ0 . .5) we have φ˙ − λ0 φ = k−2 j=0 ej (λ0 )bj+1 + ek−2 (λ0 ) ∗ ψ (2. . Theorem 2. . The results of this section can also be interpreted for the solution semigroup T and its inﬁnitesimal generator A. ψ) λ=λ0 .5) j=0 is in dom Ak (for the deﬁnition of ej see (1. ψ).1 is true and λ0 is a zero of ﬁnite multiplicity of det ∆(λ). p0 (λ0 . That dim Pλ0 = k0 was ﬁrst proved in [24].4) has a solution ˜bk . The presentation given here is based on [20]. bk−1 ) is a solution of (2.6) . (ii) The equation ˜ k (λ0 )˜bk = −˜ pk (λ0 . . For the elementary divisor theory for holomorphic matrices see [14]. where pj (λ0 .5). 2.10)). . . . b) If ψ ∈ range(A − λ0 I)k . a) The following statements are equivalent: (i) ψ ∈ range(A − λ0 I)k . ψ) ∆ (2.10 Let ψ ∈ C and λ0 ∈ σ(A) be given. . ψ) = 1 dj p(λ. An easy calculation shows that for φ given by (2.4). The results presented above are true whenever Proposition 2. 1. Proof.−1 ρ=1 mρ dρ = Remarks. where ˜bk = col(b0 . . j! dλj (iii) There exist vectors b0 . . bk−1 ∈ Cn such that the function φ= k−1 ej (λ0 )bj + ek−1 (λ0 ) ∗ ψ (2. then ψ = (A − λ0 I)k φ for any φ which is given by (2. . .
7) and (2. ψ). This proves φ ∈ dom Ak if and only if (2.7) j! j=0 In this case equation (2. k − 2.6) just means (A − λ0 I)φ = k−2 ej (λ0 )bj+1 + ek−2 (λ0 ) ∗ ψ. ψ) = − 0 [dη(θ)] ek−1 (λ0 ) ∗ ψ (θ). −r ˙ From (2. . j! (2. ψ). 0 j λ0 θ θ e dη(θ) = I − ∆ (λ0 ) for j = 1. (2. Therefore we have φ ∈ dom A if and only if k−1 1 (j) ∆ (λ0 )bj = −pk−1 (λ0 . . Moreover. i. In k − 1 steps we obtain that φ is in dom Ak−1 if and only if k−1−i j=0 1 (j) ∆ (λ0 )bj+i = −pk−1−i (λ0 .83 Autonomous Functional Diﬀerential Equations and L(φ) = − k−1 1 (j) ∆ (λ0 )bj + λ0 b0 + b1 − pk−1 (λ0 . (2. 3. −r for j = 2. Therefore χ is in dom A if and only if ∆(λ0 )bk−1 = −p0 (λ0 . if and only if (2. ψ) + ψ(0).e. . . .8) j=0 Therefore φ is in dom A2 if and only if (2. j! j=0 Here we have used λ0 I − ∆(λ0 ) for j = 0. ψ). −∆(j) (λ0 ) and pk−1 (λ0 .9) Moreover.9) holds also for i = k − 1. j=0 Analogously we see that (A − λ0 I)φ ∈ dom A if and only if k−2 1 (j) ∆ (λ0 )bj+1 = −pk−2 (λ0 . ˙ = λ0 bk−1 +ψ(0) and L(χ) = −∆(λ0 )bk−1 + For χ we get χ˙ = λ0 χ+ψ. ψ).6) we get φ(0) = λ0 b0 + b1 . . ψ).4) is satisﬁed. j! i = 0. we have χ := (A − λ0 I)k−1 φ = e0 (λ0 )bk−1 + e0 (λ0 ) ∗ ψ. .8) are satisﬁed.. we . χ(0) λ0 bk−1 − p0 (λ0 . .
with ∆ ej (λ0 )bj j=0 Proof. . k = 1. . i. dθ This implies that φ is given by (2. be the numbers associated with ∆(λ) at λ0 according to Theorem 2. if there exists a basis φ0 . j = 1.j . .3. .3 we see that Pλ0 = ker(A − λ0 I)k . Thus we have established that (iii) is equivalent to (ii) and that (iii) implies (i). . φk−1 of Z (k = dim Z) such that Aφ0 = λ0 φ0 . . . for general k. j = 1. The result follows immediately from Theorem 2.11 Let φ ∈ C and λ0 ∈ σ(A) be given. Then φ is solution of the diﬀerential equation k d − λ0 φ = ψ. The proof of Proposition 2. Since φ ∈ dom Ak . . . . dρ and mρ . Then we have φ ∈ ker(A − λ0 I)k if and only if φ= k−1 ˜ k (λ0 )˜bk = 0. . .10. Aφj = λ0 φj + φj−1 . is a subspace cyclic with respect to A with dim Zρ. for ψ = 0. . Assume that ψ ∈ range(A − λ0 I)k . .12 Let λ0 ∈ σ(L) and let .e. 2. From the results given above it is possible to determine the decomposition of Pλ0 into a direct sum of subspaces cyclic with respect to A. the vector ˜bk solves (2.j = dρ .4. .10) We call a subspace Z of Pλ0 cyclic with respect to A. shows that dim ker(A− ˜ k (λ0 ). From Proposition 2. Corollary 2. Then Pλ0 = mρ ! ! Zρ. . Theorem 2. .. . ρ = 1. (2. This proves that (i) implies (iii). κ0 + 1. mρ . k − 1.5). ψ = (A − λ0 I)k φ for some φ ∈ dom Ak .84 DELAY DIFFERENTIAL EQUATIONS have (A − λ0 I)χ = (A − λ0 I)k φ in this case. . .j .4). b). .9 gives also the dimenλ0 I)k = dim ker ∆ sion of ker(A − λ0 I)k . b). k = κ0 . . . . ρ=1 j=1 where Zρ. . Therefore Corollary 2. .
we can − λ0 I)d . . . Since dim ker(A − λ0 I)d − dim ker(A − λ0 I)d −1 = m . j = 1. m .85 Autonomous Functional Diﬀerential Equations Proof. . such that choose . .
We put j=1 αj ψj.0 . . . the subspace () () Z. . () By deﬁnition of the φj. . .i = (A − λ0 I)d −i ψj. .i = (A − λ0 I)d −1−i ψj.i−1 .i ’s we see that φj. d − 1. m . j = 1. .i ∈ ker(A − λ0 I)i+1 . d − 1. This follows from () (A − λ0 I)φj. φj.d −1 is cyclic. m . . . . .j = span φj. . . For each j = 1. . . ∈ d ker(A −1 implies α = · · · = α 1 m = 0. j = .0 = (A − λ0 I)d ψj.m m elements ψj. . . (A − λ0 I)φj. . ∈ ker(A − λ0 I) φj. = φj. = 0. . . () i = 0. () () () i = 1. .
() i 1.i ∈ ker(A − λ0 I) implies α1 = · · · = . Furthermore. m . . . m j=1 αi φj. . .
d − 1. from 0 = m j=1 . Indeed. . () αj (A − λ0 I)i φj.i = αm = 0. . . . i = 0.
m .
j = span φj. .i + j=1 (−1) βj φj. Furthermore. . .i 1. d−1 − 1 and m j=1 j = 1. . .−1 ∈ ker(A − λ0 I)d−1 . . . In the next step we observe dim ker(A − λ0 I)d−1 − dim ker(A − λ0 I)d−1 −1 = m + m−1 and choose m−1 elements ψj. (−1) are cyclic. such that m m−1 αj (A − λ0 I)d −d−1 ψj. .d−1 −1 . . j = 1. . .0 . . . we have φj. We set (−1) φj. i = 0. As in the ﬁrst step we see that the spaces (−1) (−1) Z−1. . m−1 . . . + j=1 βj ψj. . .i we see that j=1 αj (A − λ0 I) j=1 αj ψj. .m d −1 ψ d −1 . . . . i = 0. . . φj. . d−1 − 1. m−1 .i ∈ ker(A − λ0 I) which by choice of the ψj. j. ∈ ker(A − λ0 I)i+1 for j = m−1 () αj φj. m−1 . .i ’s implies α1 = · · · = αm = 0. m−1 .−1 ∈ ker(A − λ0 I)d−1 −1 j=1 implies α1 = · · · = αm = β1 = · · · = βm−1 = 0. j = 1. . .i = (A − λ0 I)d−1 −1−i ψj.−1 .i ∈ ker(A − λ0 I)i .
0 . . . . .dρ −1 .i j=1 m−1 αj (A − λ0 I)d −1 ψj. . + j=1 βj (A − λ0 I)d−1 −1 ψj. ρ = 1. . . j=1 In a ﬁnite number of steps we obtain the cyclic subspaces (ρ) (ρ) Zρ. This follows from 0= m αj (A − + j=1 = m m−1 () λ0 I)i φj.86 DELAY DIFFERENTIAL EQUATIONS implies α1 = · · · = αm = β1 = · · · = βm−1 = 0. φj. The .−1 . .i (−1) βj (A − λ0 I)i φj. . . . j = 1.−1 ∈ ker(A − λ0 I)d−1 −1 . . + j=1 βj ψj. mρ . . j=1 which means m m−1 d −d−1 αj (A − λ0 I) ψj.j = span φj.
. . i = 0.i φj. m1 . . .i . . . d − 1. . . (1) φj. . .11) ρ=1 j=1 i=0 . . i = 0. . d−1 − 1. . i = 0.i . ρ=1 mρ dρ = k0 = dim Pλ0 elements () φj. . . are linearly independent. (2. . Assume that mρ dρ −1 (ρ) (ρ) αj. . d1 − 1.i = 0. . . m−1 . j = 1. . . j = 1. m . . .i . . This can be seen as follows. . j = 1. (−1) φj. . .
.d −2 φj.d −2 = (A − λ0 I)ψj. ∈ ker(A − λ0 I)d −1 . .d −2 ψj.d −1 φj. () () d −1 and thus α() This implies m j=1 αj.d −2 ∈ ker(A − λ0 I)d −2 () () j=1 () or (observe φj. j = 1.d −1 ∈ ker(A − λ0 I) j.11) implies m αj.d −1 = 0. () . ) m j=1 αj. . . Then (2. m .
13 Let k0 be the multiplicity of λ0 as a root of det ∆(λ).d−1 −1 j=1 αj. i = d−1 .e. . all cyclic subspaces of Pλ0 are of dimension one resp.d−1 −1 = 0. . . (3) κ0 = k0 .d−1 −1 ψj. . . . . d − 3.12) .d−1 −1 = 0. m−1 . j = 1. . j = 1. .d−1 −1 (A − λ0 I)d −d−1 ψj. () j=1 m−1 + αj. i. . . ρ = 1. Proposition 2. mρ . It also proves that dim Zρ. . . (ii) k0 = n. mρ . m . . Pλ0 itself is cyclic. .−1 ∈ ker(A − λ0 I)d−1 −1 (−1) j=1 () (−1) and consequently αj. are a basis for Pλ0 . j = 1. . . . . (2. m .d−1 −1 φj. . j = 1.d −2 = 0. .87 Autonomous Functional Diﬀerential Equations () Consequently we have αj. (2) We have rank ∆(λ0 ) = n − 1. . i = 0.i = 0 for ρ = 1. (3) κ0 = 1 (κ0 is the order of the pole of ∆−1 (λ) at λ0 ). dρ − 1. . .i = 0. . . . . . (ρ) Continuing in this way we see that αj. j = 1.11) implies that m m−1 () () αj. . m . . .j = dρ . .. . . i = 0. . .d−1 −1 φj. . . . Then (2. . ∆(λ0 ) = 0 and rank ∆ (λ0 ) = n. . . (2) One of the following two conditions is satisﬁed: (i) k0 < n and rank ∆ (λ0 ) = n − k0 . a) The following statements are equivalent: (1) All subspaces cyclic with respect to A of Pλ0 are onedimensional. . Analogously we see () that αj. . dρ − 1. Of special interest are the extreme cases.d−1 −1 j=1 m = + (−1) (−1) αj. which proves that the elements φj.i . b) The following statements are equivalent: (1) Pλ0 is cyclic with respect to A. and αj. . j = (ρ) 1.
b) Pλ0 is cyclic if and only if m = 1 and d = k0 (= dim Pλ0 ). In view of 0 ≤ d1 < d2 < · · · < d this is equivalent to either (i ) = 2 and d1 = 0. i. . because k0 < n and rank ∆(λ0 ) = rank M (λ0 ) = m1 = n − k0 in this case. This implies = 2. n. a) All cyclic subspaces to be of dimension 1 is equivalent to dρ ≤ 1 for all ρ = 1. which implies either = 1.88 DELAY DIFFERENTIAL EQUATIONS Proof. (λ − λ0 )M2 (λ) . (i ).. d1 = 0 and d2 = 1. . . m2 = n − k0 and d2 = 1.8).e. (ii ) holds. In case of (ii ) the matrix M (λ) is given by M (λ) = (λ − λ0 )M1 (λ) which shows that k0 = n. Thus we have shown that (ii ) implies (ii). . In case of (i ) the matrix M (λ) in Theorem 2. then d = 1. i. βk0 +1 ≥ 1 and βn = k0 .4 is given by M (λ) = diag M1 (λ). which together with βn = n implies βi = i. Therefore condition (i ) implies condition (i).e. If on the other hand κ0 = 1. . Observing ∆1 (λ0 ) = ∆ (λ0 ) and M1 (λ0 ) = F (λ0 )∆ (λ0 )G(λ0 ) we see that rank ∆ (λ0 ) = n.13) . i. . From m1 d1 + m2 d2 = k0 and m1 + m2 = n we immediately get m2 = k0 and m1 = n − k0 . Since βm1 < · · · < βn (observe 0 ≤ d1 < d2 · · · < d ).e. From M (λ0 ) = F (λ0 )∆(λ0 )G(λ0 ) we get that also ∆(λ0 ) = 0. n. we must have βn−k0 +i = i. (ii ).e. Since κ0 = d (see Theorem 2. . In this case we have β1 < · · · < βn . conditions (i ) or (ii ) imply κ0 = 1. (2. then βn = n because of rank ∆ (λ0 ) = n and d1 > 0 because of ∆(λ0 ) = 0. i = 1. then the numbers βk deﬁned in the proof of Theorem 2. d2 = 1. . If on the other hand (i) is true.. . i.. . M (λ0 ) = 0 and rank M1 (λ0 ) = n.. . i = 1. This shows = 1 and d1 = 1. If on the other hand (ii) is true. . or (ii ) = 1 and d1 = 1.4 are given by β1 = · · · = βk0 = 0. . (i ) holds. This shows m1 = k0 and d1 = 0. But then we have ∆(λ) = (λ − λ0 )∆1 (λ). or = 2. d1 = 1.
89 Autonomous Functional Diﬀerential Equations .
For n = 1 this means β1 = βn > 0 and consequently = 1. Suppose conversely that rank ∆(λ0 ) = n − 1. If (2. In the ﬁrst case we have n = m1 = 1. i.e. m1 = n − 1. In the second case we get m1 = n − m2 = n − 1. d1 = 0. .. This implies −1 ρ=1 mρ dρ = 0 which is equivalent to either = 1 or = 2 and d1 = 0. m1 = 1 and d1 = k0 (observe k0 = m1 d1 in this case). For n > 1 we get β1 = · · · = βn−1 = 0 and βn = k0 . m2 = 1 and d2 = k0 . This implies = 2. ∆(λ) is scalar with ∆(λ0 ) = 0. then κ0 = d = k0 . then d = k0 . If conversely κ0 = k0 .13) holds. In both cases we have rank M (λ0 ) = rank ∆(λ0 ) = n − 1.
. n − 1. (2. . . i = 1. and vice versa. 0 .. . 0 (θ) ηn−1. . . . i. . . .. . . −r ≤ θ ≤ 0. . . φn ) satisﬁes φi+1 = φ˙ i . .. . y (n−1) (t) . .. . y (n−1) (θ) = φn (θ). . 0 0 ··· 0 λ −1 −c0 (λ) · · · ··· ··· −cn−1 (λ) .13) holds. . 0 ··· ··· ··· ··· η1 (θ) ··· .14) with initial data y(θ) = φ1 (θ). .. . and the matrix η(θ) by 0 η12 (θ) 0 .j+1 (θ) = 0 for θ < 0 and = 1 for θ ≥ 0.. .. . ˙ .14) Such an equation can always be written as a system.. . t ≥ −r.n ηn (θ) where ηj. As a special case we want to consider an nth order equation y (n) (t) = n−1 0 j=0 −r t ≥ 0. . j = 1. . We just deﬁne x(t) = col y(t). . .. . . n − 1..e.. . . η(θ) = . . . 0 ··· 0 . . . This and ρ=1 mρ dρ = k0 imply m = 1. . From λ −1 0 · · · 0 . . .. then the ﬁrst coordinate y(t) = x1 (t) is a solution of equation (2. (2..1) with η(θ) as given above such that φ = col(φ1 . y(t). . ∆(λ) = . . . y (j) (t + θ) dηj+1 (θ). . .. . .. . If x(t) is a solution of system (1.
−r one sees immediately that det ∆(λ) = λn − n−1 λj cj (λ). According to Proposition 2. . then ˜ k (λ0 )˜bk = 0 for some k < k0 . Indeed. if bk0 −1 = 0 for all solutions of this equation. i. Obviously we have rank ∆(λ0 ) = n − 1 for any zero λ0 of det ∆(λ0 ).10).90 DELAY DIFFERENTIAL EQUATIONS where cj (λ) = 0 j = 0. The vector bk0 −1 = 0 satisﬁes ∆(λ0 )bk0 −1 = 0. . . The special form of the matrix ∆(λ0 ) implies that the ﬁrst component of bk0 −1 is nonzero (otherwise we would have bk0 −1 = 0). Thus we have shown that the ﬁrst . . b) we have κ0 = k0 .e.14) has at least one solution y(t) = p(t)eλ0 t . Proof. . (2. . a contradiction to (2. eλθ dηj+1 (θ).14 If λ0 is a root of multiplicity k0 of det ∆(λ) as given in (2. all solutions of this equation satisfy ∆ k Pλ0 = ker(A − λ0 I) for some k < k0 = κ0 .15) j=0 Proposition 2. . Therefore ˜ k (λ0 )˜bk = 0 has a solution ˜bk = col(b0 .13. t ∈ R.. bk −1 ) with the equation ∆ 0 0 0 0 bk0 −1 = 0. where p(t) is a polynomial of degree k0 − 1. . n − 1.15) then equation (2.
Since y(t) = x1 (t) solves (2.2 Projections onto eigenspaces In this sections we shall deﬁne continuous projections πλ0 : C → Pλ0 . λ=λ0 t ≥ −r. .14). 2. x1 (t) = eλ0 t γ (k0 − 1)! t ≥ 0. λ0 ∈ σ(L).1) (resp. Proposition 2. the solution semigroup T ).15 Let λ0 ∈ σ(L) be given and assume that the function g(λ) is holomorphic at λ0 . The following result will give the motivation for the deﬁnition of πλ0 . Then the function x(t) = Res eλt ∆−1 (λ)g(λ).k0 −1 tj component x1 (t) of x(t) = eλ0 t j=0 j! bj is of the form tk0 −1 + q(t) . where the subspaces are invariant with respect to equation (1. the proof is ﬁnished. such that C = range πλ0 ⊕ ker πλ0 is a decomposition of C. where γ = 0 and q(t) is a polynomial of degree ≤ k0 − 2.
. λ−1 ˜κ D upper diagonals of ∆ 0 0 in the series expansion ∆(λ)∆−1 (λ) ≡ I around λ0 which are all zero.1 proves the result. D−κ0 +1 0 ··· 0 D−κ0 Since we have ∆(λ)∆−1 (λ) ≡ I in a neighborhood of λ0 . we get from Lemma 2. . λ=λ0 is in Pλ0 . .. .. . . of course) ˜ κ = 0. . . . . Then the function eλt ∆−1 (λ)g(λ) has at most a pole of order ≤ κ0 at λ0 .. cκ0 −1 ) ∈ Cnκ0 . As usual let κ0 denote the order of the pole of ∆−1 (λ) at λ0 . . . κ0 − 1. .17) D κ0 = . . ∈ Cnκ0 ×nκ0 . ∆ 0 0 which in view of Proposition 2. .16) where c˜κ0 = col(c0 . . This can be written as ˜ κ g˜κ . . ˜κ D (2..18) ∆ 0 0 This together with (2. t ≥ −r. . . where cj = D−j−1 g0 + · · · + D−κ0 gκ0 −j−1 . −r ≤ θ ≤ 0. . c˜κ0 = D 0 0 (2. . . .1) corresponding to λ0 and the function θ → Res eλθ ∆−1 (λ)g(λ).91 Autonomous Functional Diﬀerential Equations is an generalized eigenfunction of (1. .. 0 ˜ (2.6 (observe that the matrices appearing in the main and the ˜ κ are the coeﬃcient matrices of λ−κ0 . . From the expansions ∆−1 (λ) = ∞ (λ − λ0 )j Dj and g(λ) = j=−κ0 ∞ (λ − λ0 )j gj j=0 we get Res eλt ∆−1 (λ)g(λ) = eλ0 t λ=λ0 κ 0 −1 j t j=0 j! cj .. g0 ) ∈ Cnκ0 and D1 D−κ0 D−κ0 +1 · · · . .16) implies ˜ κ c˜κ = 0. Proof. . g˜κ0 = col(gκ0 −1 . j = 0.
16 Let λ0 ∈ σ(L) and φ ∈ C be given. Fix t ≥ 0 and deﬁne y(s) = x(s.4. xt (φ)) = x(t + s. then ∆−1 (λ)p(λ. We deﬁne πλ0 : C → Pλ0 by πλ0 φ = Res e0 (λ)∆−1 (λ)p(λ. φ) dλ. φ) (see Theorem 1. λ=λ0 Proof. φ) is the Laplacetransform of x(t. −r ≤ θ ≤ 0. deﬁne the subspace Qλ0 by Qλ0 = ker πλ0 . i. s ≥ 0. φ) dτ. φ).e.19) Γ where Γ is any positively oriented circle around λ0 such that there is no other element of σ(L) on Γ or in its interior. φ). b)). φ). Therefore we have (observe that x ˆ(t. xt (φ)) = e−λτ x(t + τ .92 DELAY DIFFERENTIAL EQUATIONS If we take g(λ) = p(λ. Theorem 2. t ≥ 0. λ=λ0 This can also be written as 1 πλ0 φ = 2πi φ ∈ C. (2. φ) = ∆−1 (λ)p(λ. a) πλ0 is a continuous projection C → Pλ0 . yˆ(λ) = ∆ 0 The second term on the righthand side of this equation is an entire function. b) The subspace Pλ0 is invariant with respect to the solution semigroup T in the sense that for any φ ∈ Pλ0 the solution x(t. φ). Then we have πλ0 xt (φ) (θ) = Res eλ(t+θ) ∆−1 (λ)p(λ. −r ≤ θ ≤ 0. λ=λ0 λ=λ0 −r ≤ θ ≤ 0. Lemma 2. πλ20 = πλ0 and range πλ0 = Pλ0 . xt (φ)). φ) dτ 0 t = eλt x ˆ(λ. φ) − eλt e−λτ x(τ .17 Let λ0 ∈ σ(L) be given. φ) can be extended . For λ0 ∈ σ(L). e0 (λ)∆−1 (λ)p(λ.. φ). λ=λ0 A short calculation yields −1 ∞ (λ)p(λ. φ)) Res eλθ ∆−1 (λ)p(λ. According to the deﬁnition of πλ0 we have πλ0 xt (φ) (θ) = Res eλθ ∆−1 (λ)p(λ. xt (φ)) = Res eλ(t+θ) ∆−1 (λ)p(λ.
φ) is holomorphic at λ0 }. . .19). (2. In order to prove boundedness we ﬁx Γ in (2. From Lemma 1. The subspace Qλ0 is positively invariant with respect to T . Moreover. φ is given as . . p(λ. ∞) and xt (φ) ∈ Pλ0 for t ∈ R. we have C = Pλ0 ⊕ Qλ0 . Therefore we have (ρ = radius of Γ) πλ0 φ ≤ 2πρKM φ. φ ∈ C. b). for a constant M > 0. By continuity of ∆−1 (λ) on Γ there exists a constant K > 0 such that e0 (λ)∆−1 (λ) ≤ K for λ ∈ Γ. . . φ) ≤ M φ for λ ∈ Γ. a) It is obvious that πλ0 is linear. κ0 − 1. Proof. .20) c) The subspace Qλ0 is characterized as Qλ0 = {φ ∈ C  x ˆ(λ) = ∆−1 (λ)p(λ. and range(A − λ0 I)k Qλ0 for k = 0. we get. . d) We have Qλ0 = range(A − λ0 I)k for k = κ0 .18. κ0 + 1.93 Autonomous Functional Diﬀerential Equations to (−∞. In order to prove range πλ0 = Pλ0 we choose a φ ∈ Pλ0 . .
. Then x ˆ(λ. 0 −1 φ = κj=0 ej (λ0 )bj . κ 0 −1 j=0 ej (λ0 )bj = φ. j=0 which gives πλ0 φ = Res e0 (λ) λ−λ0 κ 0 −1 (λ − λ0 )−j−1 bj = j=0 This proves range πλ0 = Pλ0 and πλ20 = πλ0 . φ) = κ 0 −1 (λ − λ0 )−j−1 bj .
94 DELAY DIFFERENTIAL EQUATIONS .
φ) dτ + eλt x ˆ(λ. Together with its Laplacetransform it is given by x(t. (λ − λ0 )−j−1 bj . In order to prove positive invariance of Qλ0 we use c). λ=λ0 λ=λ0 −r ≤ θ ≤ 0. φ = κj=0 ej (λ0 )bj with ∆ 0 0 exists on R. t ≥ 0. which is proved below. s ≥ 0. φ) = xt (φ)(θ). c) If ∆−1 (λ)p(λ. j=0 By Lemma 2. This implies πλ0 xt (φ) = 0 for all t ≥ 0. φ ∈ Qλ0 . φ) is holomorphic at λ0 . φ) = ∆−1 (λ)p(λ. yˆ(λ) = 0 t Since the ﬁrst term on the righthand side is entire. i. xt (φ) ∈ Qλ0 . This equation is also true for t < 0.. λ=λ0 . Then the Laurent series of ∆−1 (λ)p(λ. the solution b) For φ ∈ Pλ0 . φ) = κ 0 −1 j! bj . i. 0 −1 ˜ κ (λ0 )˜bκ = 0. −r ≤ θ ≤ 0. then obviously πλ0 φ = 0. φ) = eλ0 (t+θ) λ=λ0 j! j=0 = x(t + θ. Using this we get as in the case t ≥ 0 that πλ0 xt (φ) = xt (φ). t ∈ R. φ) = eλ0 t κ 0 −1 j t j=0 x ˆ(λ. This implies πλ0 φ = Res e0 (λ)∆−1 (λ)p(λ. at λ0 . φ) = eκ−1 (λ0 )q−κ + · · · + e0 (λ0 )q−1 ≡ 0. λ = λ0 ..e. φ) ds = e e−λτ x(τ .16 we have κ 0 −1 (t + θ)j πλ0 xt (φ) (θ) = Res eλ(t+θ) x bj ˆ(λ. Fix t < 0 and deﬁne y(s) = x(t + s. φ) = x(s. t ≥ 0. φ) for φ ∈ Qλ0 is holomorphic at λ0 . Since x ˆ(λ. 0 < κ ≤ κ0 . xt (φ)). φ) = 0 λ=λ0 for all t ≥ 0 and −r ≤ θ ≤ 0. φ). The Laplacetransform yˆ(λ) of y(s) is given by 0 ∞ −λs λt e x(t + s.20) follows from C = range πλ0 ⊕ range(I − πλ0 ).e. The representation (2. φ) is of the form (λ − λ0 )−κ q−κ + (λ − λ0 )−κ+1 q−κ+1 + · · · with q−κ = 0. we get Res eλ(t+θ) ∆−1 (λ)p(λ. φ). xt (φ)) = Res eλ(t+θ) ∆−1 p(λ. Assume now that ∆−1 (λ)p(λ. we get Res eλθ ∆−1 (λ)p(λ. φ) has a pole of order κ.
21) imply ∆ −1 (λ)p(λ. φ) + hk (λ.3) (k) x ˆ(λ. i. ∞) and x(t. j=0 . φ) is ktimes continuously diﬀerentiable on [−r. k φ ∈ dom A implies that x(t.e. If k ≥ κ0 we see that ∆−1 (λ)p(λ. i. φ(k) ). If ψ ∈ Qλ0 . φ) is holomorphic at λ0 . φ k )=λ ∆ −1 (λ)p(λ. where hk (λ. ψ) = ∆(λ)q(λ) in a neighborhood of λ0 or.21) It is easy to see that φ ∈ dom A implies that x(t. . i 1 (j) ∆ (λ0 )qi−j = pi (λ0 . i=0 This and (2. φ ) j k k j=0 j j ∆−1 (λ)p(λ. φ −1 )=∆ (k) (λ)p(λ. ψ= k k j j=0 (j) (−1)k−j λk−j 0 φ . ψ ∈ Qλ0 by part c). φ) (−1)k−j λk−j λ 0 = (λ − λ0 )k ∆−1 (λ)p(λ. equivalently. 1. φ) is continuously ˙ By induction we see that diﬀerentiable on [−r. φ) + hk (λ. φ ∈ / Qλ0 . (2. j! i = 0. d) Let ψ = (A − λ0 I)k φ for some φ ∈ dom Ak . φ). . φ) − k−1 λi φ(k−1−i) (0). . Therefore we have p(λ.. ψ) is holomorphic at λ0 . ∞) and x(k) (t. ψ). Taking Laplacetransforms we get (observe Theorem A.e.. φ). ψ) is holomorphic at λ0 . ψ) = k k j=0 = −1 (j) (−1)k−j λk−j 0 ∆ (λ)p(λ.. ˙ φ) = x(t. . φ) = x(t. then q(λ) = ∆−1 (λ)p(λ.95 Autonomous Functional Diﬀerential Equations i.e.
. . . κ0 − 1. . Suppose that range(A − λ0 I)k = Qλ0 for a k < κ0 . This shows that equation (2. . By Theorem 2. Then ψ ∈ range(A − λ0 I)k = Qλ0 and ψ ∈ ker(A − λ0 I)κ0 −k ⊂ Pλ0 which by C = Pλ0 ⊕ Qλ0 implies ψ = 0. . . . . .10 solution ˜bk = − col(q " k−1 . 2.4) has the k = 1. . Take φ ∈ ker(A − λ0 I)κ0 such that ψ = (A − λ0 I)k φ = 0. a contradiction. . i where q(λ) = ∞ i=0 (λ − λ0 ) qi . . . This ﬁnishes the proof of Q range(A − λ I) this implies ψ ∈ ∞ 0 λ0 = k=1 k range(A − λ0 I) . Note that ker(A − λ0 I)k ker(A − λ0 I)κ0 for k = 0. q0 ) for any k . . . . κ0 + 1. k = κ0 .
t ∈ R.18 Let x(t) = x(t. Cn ).1). (2.4) with φ ≡ 0 and f ∈ L1loc (0. the deﬁnition of πλ0 implies y (λ). be the solution of (1.30). f ) (θ) = Res eλ(t+θ) ∆−1 (λ)fˆ(λ. .3). t) λ=λ0 for −r ≤ θ ≤ 0. πλ0 xt1 (0. (1. f (·. 0. −1. πλ0 Yt = Dt . f ) = Res e0 (λ)ˆ λ=λ0 where the Laplacetransform yˆ(λ) is given by ∞ e−λt y(t) dt yˆ(λ) = 0 ∞ t1 λt1 −λt λt1 e x(t. t1 )) dt − e e−λt x(t. We ﬁx t1 > 0 and deﬁne f (t. 0.23) then (2. f (·.24) Since ∆−1 (λ) is the Laplacetransform of the fundamental matrix Y (t). If.1) and correspondingly the columns of Dt are in Pλ0 . t1 )) dt. j = −κ0 . . t1 ) = f (t) for 0 ≤ t ≤ t1 and = 0 for t > t1 . t1 ) − eλt1 0 where we have used (1. t) = t ≥ 0. f (·. we have Dt (θ) = Res eλ(t+θ) ∆−1 (λ). (1.18) implies that the columns of D(t) are generalized eigenfunctions of equation (1. λ=λ0 −r ≤ θ ≤ 0. . Furthermore. t1 )). Then we have (2. we deﬁne λ0 t D(t) = e κ 0 −1 j t j=0 j! D−j−1 . t ≥ 0. f ). . t ≥ 0. t1 )) dt =e 0 0 t1 e−λt x(t. 0. 0 Proof. for the matrices Dj . = eλt1 ∆−1 (λ)fˆ(λ.2) with φ = xt1 (0. t ≥ 0. where t ˆ e−λτ ) f (τ ) dτ. is the solution of the homogeneous problem (1. 0. because the second term on the righthand side is holomorphic at zero. Since y(t) = x(t + t1 . . f ). Dt can be considered as the projection of Yt into Pλ0 . (2. f (·.96 DELAY DIFFERENTIAL EQUATIONS For the nonhomogeneous equation we have: Proposition 2. ∞. The result now follows. 0. f (λ.22) πλ0 xt (0. t ≥ 0. t ≥ 0.
(2. where 1 dj ˆ (−1)j t j −λ0 τ fj (t) = f (λ. j! −r ≤ θ ≤ 0. (1.25) With these notations we get κ 0 −1 (t + θ)j λ0 (t+θ) gj (t).16 and the proof of Proposition 2. . . . . j=0 where cj = D−j−1 p0 + · · · + D−κ0 pκ0 −j−1 . . . κ0 − 1. . f ) (θ) = t eλ0 (t+θ−τ ) 0 −1 κ 0 −1 κ 0 t = eλ0 (t+θ−τ ) 0 t 0 = j=0 k=j eλ0 (t+θ−τ ) (−τ )k−j (t + θ)j f (τ ) dτ (k − j)!j! k 0 −1 κ 1 k! D−k−1 (t + θ)j (−τ )k−j f (τ )dτ k! (k − j)!j! j=0 k=0 = D−k−1 0 −1 κ (t − τ + θ)k D−k−1 f (τ ) dτ k! k=0 t Dt−τ (θ)f (τ ) dτ.1). With these expressions we get for πλ0 xt (0. j = 0. . g˜κ0 (t) = col(g0 (t). t ≥ 0.97 Autonomous Functional Diﬀerential Equations Note however. πλ0 xt (0. . . . We set f˜κ0 (t) = col(fκ0 −1 (t). . that the columns of Yt are not in C for 0 ≤ t < r.2) we get from Lemma 2. (k − j)! j = 0. . j! dλj j! 0 and ˜ κ f˜κ (t). gκ0 −1 (t)) = D 0 0 (2. . κ0 −1. t ≥ 0. t) λ=λ0 = τ e f (τ ) dτ.25) implies gj (t) = κ 0 −1 D−k−1 fk−j (t) k=j t λ0 t e = 0 0 −1 κ D−k−1 k=j (−τ )k−j f (τ ) dτ. f ) (θ) = e j! j=0 Equation (2. . f0 (t)). . j = 0.26) . 0 For the solution of the homogeneous problem (1. . .15 κ 0 −1 t + θ)j λ0 (t+θ) πλ0 xt (φ) (θ) = e cj . −r ≤ θ ≤ 0. f ) πλ0 xt (0. . κ0 − 1. t ≥ 0. .
f ∈ L1loc (0. + 0 In particular we have. . 1 dj 0 0 pj = p(λ. . t ≥ 0. . . for φ ∈ C. φ) λ=λ0 = j 1 j! dλ [dη(θ)] (θ − τ )j eλ0 (θ−τ ) φ(τ ) dτ j! −r θ for j = 1.27) t Dt−τ f (τ ) dτ. . Cn ) and λ0 ∈ σ(L) be given. πλ0 φ (θ) = D(θ)φ(0) + 0 −r ≤ θ ≤ 0. j = 0. ∞. . 0 Thus we have established the following result: Proposition 2. 2. . Using Fubini’s theorem we get 0 0 0 τ j λ0 (s−τ ) j λ0 τ [dη(s)] (s − τ ) e φ(τ ) dτ = τ e [dη(s)]φ(s − τ ) dτ −r s −r −r r (−τ )j e−λ0 τ L(φτ ) dτ. . t ≥ 0. 1 (t − τ + θ)k D−k−1 L(φτ ) dτ k! −r ≤ θ ≤ 0. .19 Let φ ∈ C. Then we have r Dt−τ L(φτ ) dτ πλ0 xt (φ.98 DELAY DIFFERENTIAL EQUATIONS and 0 0 [dη(θ)] eλ0 (θ−τ ) φ(τ ) dτ φ(0) + −r θ for j = 0. r D(−τ + θ)L(φτ ) dτ. f ) = Dt φ(0) + 0 (2. = 0 Equation (2. 1.26) implies κ κ 0 −1 0 −1 (t + θ)j (t + θ)j λ0 (t+θ) λ0 (t+θ) πλ0 xt (φ) (θ) = e D−j−1 φ(0) + e j! j! j=0 × κ 0 −1 k=j 1 D−k−1 (k − j)! = Dt (θ)φ(0) + r (−τ )k−j e−λ0 τ L(φτ ) dτ 0 λ0 (t−τ +θ) e = D(t + θ)φ(0) + r j=0 0 κ 0 −1 k=0 r Dt−τ (θ)L(φτ ) dτ.
. the restriction of T to Pλ0 is generated by a bounded linear operator. . k0 . . bkj 0 ) ∈ Cn×k0 . Furthermore. φ(k0 ) (θ) . t ≥ 0. . constitute a basis for Pλ0 . .3. j=0 where ˜bνκ0 = col(bν0 .28) and (1. . We deﬁne ˜ κ = (˜b1κ . be a basis for ker ∆ 0 of Proposition 2. ˜bkκ0 ) ∈ Cnκ0 ×k0 . i. . j = 0. κ0 . . . which we shall compute below. f ) is formally obtained from (1. Φ 0 0 0 Φj = (b1j . .. . . . .e. . .99 Autonomous Functional Diﬀerential Equations Remark. . ˜ κ (λ0 ).equation(2. . is the solution semigroup of an ordinary diﬀerential equation.31) by applying the projection πλ0 and observing that πλ0 Yt = Dt . Since Pλ0 is invariant with respect to the solution semigroup T and is ﬁnite dimensional. . b)) the functions φ(ν) = κ 0 −1 ej (λ0 )bνj . Φλ0 (θ) = φ(1) (θ). bνκ0 −1 ). ν = 1. k0 . ν = 1. . −r ≤ θ ≤ 0. Then (see the proof Let ˜bνκ0 . . The above result shows that the projection πλ0 xt (φ. . . . . . . .27) for f ≡ 0 allows to deﬁne the continuation of z(t) = πλ0 xt (φ) (0) to negative t. .
.e. b)). . . i. j=0 On the other hand we have φ = πλ0 φ = Res e0 (λ)∆−1 (λ)p(λ. . so that (2.17. . κ0 − 1.. . .28) has a unique solution. . ej (λ0 )cj with c˜κ0 = D 0 0 j=0 where c˜κ0 = col(c0 . φ = Φλ0 a = κ 0 −1 ej (λ0 )Φj a. j = 0. . . . cκ0 −1 ). .28) ˜ κ = k0 . This implies cj = Φj a. p0 ) (see (2. . . . 0 −1 ˜ κ = col (Φ0 . . ej (λ0 )Φj and Φ Then we have Φλ0 = κj=0 0 0 For φ ∈ Pλ0 let a ∈ Ck0 be the coordinate vector of φ with respect to the basis φ(1) .4 This 4 This also follows from S(t)Pλ0 ⊂ Pλ0 . ˜κ a = D Φ 0 0 0 (2. Note that rank Φ 0 From Pλ0 = ker(A − λ0 I)κ0 and (A − λ0 I)κ0 Aφ = (A − λ0 I)κ0 +1 φ + λ0 (A − λ0 I)κ0 φ for φ ∈ Pλ0 we see immediately that APλ0 ⊂ Pλ0 .26) for θ = 0). . . . t ≥ 0 (compare Theorem 2. p˜κ0 = col(pκ0 −1 . of course. φ(k0 ) . Φκ −1 ). or equivalently ˜ κ p˜κ . φ) = λ=λ0 κ 0 −1 ˜ κ p˜κ . . .
∞. . The matrix Bλ0 is the solution of (2. Since Dt ∈ Pλ0 . t ∈ R. . Cn ) let y(t) be the coordinate vector of πλ0 xt (φ.29) and Hλ0 solves D−1 . . . t ∈ R (see (2. 0 0 0 Φκ0 −1 0 (2.30) where πλ0 φ = Φλ0 a. y(0) = a. The coordinate vector a of πλ0 φ can be computed from ˜ κ p˜κ . This equation is equivalent to Φ1 . Then y(t) is the solution of y(t) ˙ = Bλ0 y(t) + Hλ0 f (t).24)). . . ˜κ + ˜ κ Bλ = λ0 Φ Φ . φ) λ=λ0 . for an n × k0 matrix H(t). κ0 − 1. we have Dt = Φλ0 H(t). f ) with respect to Φλ0 . ˜ κ Hλ = Φ . t ≥ 0. t ∈ R..100 DELAY DIFFERENTIAL EQUATIONS implies AΦλ0 = λ0 Φλ0 + κ 0 −2 ej (λ0 )Φj+1 = Φλ0 Bλ0 j=0 for some k0 × k0 matrix Bλ0 . f ) = Φλ0 y(t).. i. H(t) = eBλ0 t Hλ0 . 0 0 D−κ0 . Theorem 2. t ∈ R. . . i. . πλ0 xt (φ. j = 0. where Hλ0 is determined by D(θ) = Φλ0 (θ)Hλ0 . t ≥ 0. −r ≤ θ ≤ 0.29) which can be uniquely solved for Bλ0 . ˙ we get H(t) = Bλ0 H(t). . ˜κ a = D Φ 0 0 0 j d where p˜κ0 = col(pκ0 −1 .e..20 Let λ0 ∈ σ(L) and a basis Φλ0 of Pλ0 be given. From d d d ˙ D(t + θ) = Φλ0 (θ)H(t) D(t + θ) = Φλ (θ)H(t) = dt dθ dθ 0 = φλ0 (θ)Bλ0 H(t).. For φ ∈ C and f ∈ L1loc (0. p0 ). pj = dλ j p(λ. −r ≤ θ ≤ 0. t ∈ R. (2.e.
We set PΛα = N # i=1 Pλi . i = 1. . . . κ0 − 1. . . equivalently. From Proposition 2. BλN . .. .3 j = 0. N .c)) QΛα = φ ∈ C  ∆−1 (λ)p(λ. ΦλN . .31) is ﬁnite for any α ∈ R. . Equation (2. D−j−1 = Φj Hλ0 . which proves (2. i = 1. Let λ1 . t ≥ 0. . 2.101 Autonomous Functional Diﬀerential Equations Proof. (2. πλ0 φ = Φλ0 Hλ0 φ(0) + 0 r e−Bλ0 τ Hλ0 L(φτ ) dτ . . QΛα = N $ Qλi .e. . If we deﬁne ΦΛα = Φλ1 . . The equation for Hλ0 follows from D(θ) = κ 0 −1 ej (λ0 )(θ)D−j−1 = Φλ0 Hλ0 = κ 0 −1 j=0 ej (λ0 )(θ)φj Hλ0 j=0 for −r ≤ θ ≤ 0 or.17 the set Λα = {λ ∈ σ(L)  Re λ ≥ α} (2.28). The equation for a given in the theorem is just (2. where φλi is a basis for Pλi . .17. .27) implies r e−Bλ0 τ Hλ0 L(φτ ) dτ Φλ0 y(t) = Φλ0 eBλ0 t Hλ0 φ(0) + 0 t eBλ0 (t−τ ) Hλ0 f (τ ) dτ. .30). φ) is holomorphic for Re λ ≥ α . . . + Φλ0 so that a = Hλ0 φ(0) + 0 t 0 i. . . . BΛα = diag Bλ1 . N .32) i=1 Then C = PΛα ⊕ QΛα and (see Theorem 2. y(t) = eBλ0 t a + 0 eBλ0 (t−τ ) Hλ0 f (τ ) dτ . Exponential dichotomy of the state space According to Corollary 1. . λN be the elements in Λα and denote by ki the multiplicity of λi . . .19 and the deﬁnition of H(t) we see that r e−Bλ0 τ Hλ0 L(φτ ) dτ .
21. . for any γ > 0. respectively.36) . γ) ≥ 1 such that.31) and (2.21 For α ∈ R let Λα and QΛα . . t ≤ 0.19) that ωL. πΛ α φ = i=1 Then there exist constants γ > 0 and K = K(α. γ) ≥ 1 such that x(t.φ +γ)t ≤ Kψe(α−γ)t . πΛα xt (φ) ≤ KπΛα φe(α−γ)t . PΛα be deﬁned by (2.33) then y(t) solves t ∈ R. there exists a K1 ≥ 1 such that y(t) ≤ K1 y(0)e(α−γ)t . Let φ ∈ QΛα be such that ωL. e−(α−γ)r )) This implies (with K (α−γ)t ˜ .ψ . deﬁne the projection πΛα : C → PΛα by N πλi φ.32). Therefore. xt (ψ) ≤ Kψe ψ ∈ QΛα . . b) there exists a constant K = K(α.17 it also follows that there exists a γ > 0 such that Re λ ≤ α − 2γ for all λ ∈ σ(L) \ Λα .φ = maxψ∈QΛα ωL. λN . (2. for φ ∈ PΛα . (2. ΦΛα y(0) = φ. y(t) ˙ = BΛα y(t). . t ≤ 0. Thus we have proved the following theorem: Theorem 2. From Corollary 1. The only eigenvalues of BΛα are λ1 . t ≥ 0. ψ) ≤ Kψe(ωL. y(t) ∈ Ck1 +···+kN by xt (φ) = ΦΛα y(t). for any φ ∈ C. (2. Furthermore. Then according to Theorem 1. (I − πΛα )xt (φ) ≤ K(I − πΛα )φe(α−γ)t .33).102 DELAY DIFFERENTIAL EQUATIONS and. t ≥ 0.34) is equivalent to xt (φ) ≤ K1 φe(α−γ)t . t ∈ R. then (2.34) where y(t) is deﬁned in (2. If we deﬁne the norm on Ck1 +···+kN by z = Φα z. (2. ψ ∈ QΛα .35) t ≥ 0. φ ∈ PΛα . This implies (see Deﬁnition 1. φ ∈ C. ˜ = K max(1. t ≤ 0.φ ≤ α − 2γ for all φ ∈ QΛα .
b) Equation (1. ∆(λ0 ) = 0 and rank ∆ (λ0 ) = n (k0 is the multiplicity of λ0 ).1) which parallels the corresponding result for ordinary diﬀerential equations. (I − πΛ0 )xt (φ) = xt (φ). t ≥ 0.21.22 a) Equation (1.1) is asymptotically stable if and only if it is exponentially stable. i. and . t ≥ 0.1) is called exponentially stable if there exist constants K ≥ 1 and β > 0 such that. Theorem 2. either k0 < n and rank ∆(λ0 ) = n − k0 or k0 = n.21 allows to reduce a problem for functional diﬀerential equations to a problem for ordinary diﬀerential equations. We choose α = 0 in Theorem 2..103 Autonomous Functional Diﬀerential Equations In many cases Theorem 2.14) condition (ii) can be replaced by (ii ) Re λ0 = 0 and λ0 is a simple root of det ∆(λ).. xt (φ) ≤ Kφe−βt .1) is called stable if and only if there exists a constant K ≥ 1 such that. As an example we prove below a characterization of stability resp. for all φ ∈ C. xt (φ) ≤ Kφ.1) is stable if and only if for all eigenvalues λ0 ∈ σ(L) the following two conditions are satisﬁed: (i) Re λ0 ≤ 0. Deﬁnition 2. Proof. t ≥ 0. b) Equation (1. then Λ0 = ∅.e. asymptotic stability for equations (1. If all eigenvalues satisfy condition (i).1) is called asymptotically stable if and only if it is stable and lim xt (φ) = 0 for all φ ∈ C. which is the case if and only if condition (i) is satisﬁed. for all φ ∈ C. (ii) If Re λ0 = 0 then all subspaces of Pλ0 which are cyclic with respect to A are onedimensional. i.23 a) Equation (1. In case of the scalar nth order equation (2. t→∞ c) Equation (1.e.
(1.5 In this section we shall investigate solutions of problem (1. . If on the other hand (1. This is equivalent to ωL. φ) ≤ xt (φ) ≤ Kφeαt . The solution with initial function φ ∈ Pλ0 is bounded on t ≥ 0 if and only if all cyclic subspaces of Pλ0 are onedimensional. φ) = −∞ t→∞ t t respectively to limt→∞ eβt x(t.1) is asymptotically stable. then (i) must hold. Small Solutions and Completeness 3. φ) = 0 for all β ∈ R. If Λ0 = ∅ and (ii) holds. Obviously we have $ QΛα . This motivates the following deﬁnition: 5 Note that P itself is invariant in the sense of Theorem 2.1 Small solutions We deﬁne the subspaces $ Q= Qλ and P = λ∈σ(L) ! Pλ λ∈σ(L) of C. then there exists a solution of the form beλ0 t with b = 0 satisfying ∆(λ0 )b = 0. b). φ) = lim ln x(t.21 implies that for any α ∈ R there exists a constant K = K(α) ≥ 1 with x(t.17. Then Theorem 2. Asymptotic stability implies Re λ0 < 0. Indeed.104 DELAY DIFFERENTIAL EQUATIONS the inequality (2. t ≥ 0.φ = lim sup t→∞ 1 1 ln x(t.2) with φ ∈ Q.1). Q= α∈R Let φ ∈ Q be given. 3.13.36) implies exponential stability. a). Then the result on stability follows immediately from Proposition 2. then there is a ﬁnite number of eigenvalues on the imaginary axis. if Re λ0 is an eigenvalue. From invariance of Qλ and Pλ it follows that the closed subspaces Q and P are positively invariant with respect to the solution semigroup T .
φ) = 0 t→∞ for all β ∈ R.2 Let φ ∈ C.1).1 A solution x(t. 1] the solution is given by t 1 Aφ(τ − 1) dτ = Ab(t − 1)2 . φ = 0. φ) = x(1. (1.1 and the obvious fact that the Laplacetransform of a small solution is an entire function it follows that Q is the subspace of initial values for small solutions of equation (1.1) if and only if there exists a t1 > −r such that x(t. But A does not have nonzero eigenvalues and Ab = 0. 0]. φ) is a small solution of equation (1. 2 where Ab = 0. Note that φ(θ) = 0 on [−r. ρ→∞ ρ ξ = lim sup . x(t.2) is called a small solution if and only if x(·. Theorem 3. We choose 1 φ(θ) = Ab + bθ. Then the following is true: a) x(t.1) behave like the solution constructed in the example above. φ) = 0 on [0. 1 ≤ t ≤ 2. φ) + 1 t 1 2 = A b (τ − 2)2 dτ = 0. 2 1 This implies x(t. On [0. t ≥ 0. It is easy to see that there exist equations with Q = ∅. φ) = 0 for all t ≥ t1 . 0 ≤ t ≤ 1. with an n × nmatrix A such that A = 0 and A2 = 0 (which implies n ≥ 2). b) Deﬁne the numbers 1 ln  det ∆(−ρ) ρ ρ→∞ 1 σ = lim sup ln  adj ∆(−ρ). 1). because φ(θ0 ) = 0 would imply Ab = − 12 θ0 b. ∞) and lim eβt x(t. 2] we have t Ax(τ − 1. The main result of this section is that all small solutions of equations of the form (1. φ) = 0 for t ≥ 1 and consequently φ ∈ Q.105 Small solutions and Completeness Deﬁnition 3. Take for instance x(t) ˙ = Ax(t − 1). From the considerations before Deﬁnition 3. φ) ≡ 0 on [−r. φ) dτ x(t. So neither θ0 < 0 nor θ0 = 0 is possible. −r ≤ θ ≤ 0. be given. φ) of problem (1. On [1. φ) = φ(0) + 2 0 which shows x(t.1).
φ ∈ Q or equivalently ∆−1 (λ)p(λ.40)) p(λ. φ) = 0 for all t ≥ t1 is given by tφ = lim sup ρ→∞ 1 ln adj ∆(−ρ)q(−ρ. we have the estimate −r < tφ ≤ σ − ξ.0] ηij . φ) is a small solution of equation (1. Since the quotient ∆−1 (λ)p(λ. then x(t. φ) ≤ Kφ 1 . φ) be a small solution. (3. i. φ) = ∆−1 (iω)p(iω. φ) ≡ 0. i. φ) and det ∆(λ) are entire functions of exponential type. φ) is also entire.2) Proof.. λ ∈ C. For p(λ. φ) gives also a precise characterization of the minimal interval which contains the support of x(t. (3.3) with positive constants a0 and a1 . φ) is entire. φ) = φ(0)− [dη(θ)] −r θ λ(θ−τ ) e 0 φ(τ ) dτ +λ −r e−λτ φ(τ ) dτ.1) where 0 q(λ.1.e.2. −r λ ∈ C.1) then the number tφ = min t1  x(t. it is also of exponential type by Proposition B. b) Theorem B. −r Therefore we have adj ∆(λ) ≤ a0 eλ(n−1)r and  det ∆(λ) ≤ a1 eλnr . of . λ ∈ C. Thus the functions adj ∆(λ)p(λ. An application of Theorem B. φ) ≡ 0. ρ (3.6 proves the result in case x ˆ(λ. φ). If x ˆ(λ. By Lemma 1. λ ∈ C.18 we have the estimate ∆−1 (λ)p(λ. ω + 1 ω ∈ R. φ) − ξ. Moreover. φ) ≡ 0 on t ≥ 0 by Theorem A.. From (1. which shows that x ˆ(iω.e. φ) is square integrable on R.106 DELAY DIFFERENTIAL EQUATIONS If x(t. φ) ≤ 1 + reλr φ.6 applied to x ˆ(λ.11) we get the estimate 0 eλθ dηij (θ) ≤ eλr var[−r. Conversely let x(t. then trivially it is a small solution. φ) we obtain (compare the proof of the estimate (1. φ) = 0 for t ≥ t1 > −r. a) If x(t.
The ﬁrst two estimates are immediate consequences of (1. . φ). φ) is an entire function of exponential type. ρ In the remaining part of the proof we show that the exponential type ξ of det ∆(−ρ) can be split oﬀ in the formula for tφ . where 0 q(λ. φ). ) as ω → ∞. in case x ˆ(λ. φ) ≤ 1 + r + ωr φ. yˆ(λ) ≡ 0. φ) ρ ρ→∞ 1 = r + lim sup ln ∆−1 (−ρ)q(−ρ. φ) ≤ O(ω) n−1 as ω → ∞. the function y(t) has compact support on [0. Since x(t. t ≥ 0. yˆ(λ) = e x −r which can be written as yˆ(λ) = e−λr ∆−1 (λ)q(λ. φ). φ) = p(λ. This proves tφ = t0 − r = lim sup ρ→∞ 1 ln ∆−1 (−ρ)q(−ρ. −r Since x ˆ(λ. because φ ≡ 0. Moreover. φ) is a small solution.107 Small solutions and Completeness tφ > 0.6 implies that y(t) = x(t − r.11) for Re λ = 0 and the deﬁnition of ∆(λ). φ) ≡ 0. Theorem B. ρ→∞ ρ t0 = hyˆ(π) = lim sup The number t0 cannot be replaced by a smaller one. The third estimate follows from q(iω. φ) = 0 for t ≥ t0 . φ) + ∆(λ) = φ(0) − e−λτ φ(τ ) dτ −r 0 [dη(θ)] −r θ λ(θ−τ ) e −r 0 φ(τ ) dτ + λ e−λτ φ(τ ) dτ. φ). ∞). ω ∈ R. An easy calculation shows that 0 −λr −λr ˆ(λ. as ω → ∞. Since we want to characterize tφ also for small solutions with tφ < 0. where 1 ln erρ ∆−1 (−ρ)q(−ρ. we have to investigate the Laplacetransform of y(t) = x(t − r. φ) + e e−λτ φ(τ ) dτ. We shall need the following estimates:  det ∆(iω) = O(ωn ) adj ∆(iω) ≤ O(ω q(iω. the same is true for yˆ(λ).
This implies that lim sup ρ→∞ 1 −1 ln ∆ (−ρ)q(−ρ.3) and q(−ρ.5 the indicator functions of det ∆(λ) and g(λ) are given by hdet ∆(λ) (α) = −hdet ∆(λ) (π) cos α. 3π/2] the ‘lim sup’ in the deﬁnition of the indicator function can be replaced by ‘lim’. . 3π/2] we have lim sup ρ→∞ 1 1 −1 1 iα ln ∆ (−ρ)q(−ρ. ρ > 0. By Theorem B. ρ which is equation (3. It remains to prove the estimate (3.108 DELAY DIFFERENTIAL EQUATIONS The above estimates show that the integrals ∞ + ∞ + ln  det ∆(iω) ln g(iω) dω and dω 2 2 1+ω −∞ −∞ 1 + ω exist. where we have set g(λ) = adj ∆(λ)q(λ. for a dense subset of [π/2. We have hg(λ) (π) ≤ lim sup ρ→∞ 1 1 ln adj ∆(−ρ) + lim sup ln q(−ρ. Moreover.2). φ) = hg(λ) (π) − hdet ∆(λ) (π). This shows that for α in a dense subset of [π/2. φ).1). φ) ρ ρ ρ→∞ ≤ σ + 0. φ) = lim sup ln g(ρe ) ρ det ∆(ρeiα ) ρ→∞ ρ 1 = lim sup ln g(ρeiα ) − ln  det ∆(ρeiα ) ρ→∞ ρ 1 = lim sup ln g(ρeiα ) − lim ln  det ∆(ρeiα ) ρ→∞ ρ→∞ ρ = − hg(λ) (π) − hdet ∆(λ) (π) cos α. φ) ≤ 1 + r + ρr φ. hg(λ) (α) = −hg(λ) (π) cos α for π/2 ≤ α ≤ 3π/2. where we have used (3.
2) it is obvious that Q ⊂ ker T (t) for t ≥ σ − ξ.6 The eigenfunctions of (1. from (3. in [12] it is proved that α = σ − ξ is the exact number such that (3.109 Degenerate delay equations 3. φ) is a small solution of equation (1. Then we have range T (t) = P for all t ≥ σ − ξ. If φ ∈ ker T (t1 ) for a t1 ≥ 0.1) and φ − ψn → 0 as n → ∞. An analogous result on the range of T (t) is much more diﬃcult to prove: Theorem 3. An immediate consequence of Theorem 3. . Proof. ρ ρ→∞ ρ Then we have ker T (t) = Q for all t ≥ σ − ξ. x(t.1) is of central importance for the problem of completeness of generalized eigenfunctions of equation (1. On the other hand.5 is the following characterization of completeness: Corollary 3. for any φ ∈ C there exists a sequence ψn . 2. The proof is based on duality theory for equation (1. Moreover. n = 1. i.. . which in turn is equivalent to ξ = nr.1). i. then T (t)φ = T (t−t1 )T (t1 )φ = 0 for all t ≥ t1 .4) For a proof of this theorem we refer to [12.1) and therefore we have φ ∈ Q.1) are complete in C if and only if Q = {0}.e.2 Completeness of generalized eigenfunctions The subspace Q of initial functions for small solutions of equation (1.4. Deﬁnition 3. Section 7.3 We say that the generalized eigenfunctions of equation (1.5 Let ξ and σ be deﬁned as in Proposition 3.4 Let ξ = lim sup ρ→∞ 1 1 ln  det ∆(−ρ) and σ = lim sup ln  adj ∆(−ρ).4) holds for t ≥ α..1) and the fact that the existence of small solutions to the equation which corresponds to the adjoint semigroup can be characterized in the same way as for (1. . .1) are complete in C if and only if P = C.8] and the literature quoted there. . The following result on ker T (t) is easy to prove: Proposition 3.1) in C. (3. such that each ψn is a ﬁnite sum of generalized eigenfunctions of equation (1.e.
the question is. In this chapter we only present a few important fact on degenerate systems and refer to [18]. For the rest of the chapter we assume that the delay equations we consider have complex valued righthand sides. 6 For . The proof of necessity would require that the solutions of the delay equation at any time t ≥ 0 span all of Rn . q = 0. A ﬁrst result is the following: a vector a ∈ Cn resp.1 A necessary and suﬃcient condition In this section we restrict ourselves to degenerate equations of type (1. we call system (1. DELAY DIFFERENTIAL EQUATIONS Degenerate delay equations 4. if there exists a t1 > 0 and a vector q ∈ Rn .110 4. Deﬁnition 4. Some parts of the treatment of the degeneracy problem parallel the investigations of small solutions in Section 10. a property which for ODEsystems clearly is satisﬁed. For ODEsystems a necessary and suﬃcient condition for nullcontrollability is that a certain matrix – the so called controllability Grammian – has full rank. such that q T x(t. Weiss posed the question if an analogous phenomenon is also possible for equations of the form x(t) ˙ = A0 x(t) + A1 x(t − r) More precisely. a matrix A ∈ Cn×n the conjugate transposed is denoted by a∗ resp. Weiss considered in his 1967 paper [30] controllability and in particular nullcontrollability for linear control systems with time delay in the state equation.1) is called degenerate if and only if there exists a vector q ∈ Cn \ {0} such that6 lim q ∗ eγt x(t.1 System (1. Simple examples (see for instance [10. For delay systems of retarded type this condition is only suﬃcient.1) If this condition holds. L.1) degenerate with respect to q. p. for references and constructions of degenerate systems.1) and refer for nonautonomous equations to [18] for results and literature. [19] for further results. φ) = 0 for all solutions of the diﬀerencediﬀerential equation and all t ≥ t1 . (4. L. φ) = 0 t→∞ for all γ ∈ R and all φ ∈ C. 41]) show that there exist nonautonomous linear scalar equations such that all solutions are zero after some ﬁxed timeinterval. by A∗ .
φ) = 0 for t ≥ tφ .111 Degenerate delay equations Proposition 4. Choose φ ∈ C such that q ∗ ∆−1 (λ)p(λ. This implies ∞ e−γt q ∗ (t.1) there exists an M > 0 with e(−γ+1)t q ∗ x(t. Proof.e. φ) is squareintegrable on λ = iω. q x(t. φ) dt. ω ∈ R. Assume that this is the case. The smallest t1 with this properties is called the degeneracy time of system (1.2 we see that 1 tφ = lim sup ln q ∗ adj ∆(−ρ)p(−ρ.18 with α = 0 we get the estimate ∗ −1 q ∆ (λ)p(λ. φ) ≡ 0 for all φ ∈ C. φ) ρ and cannot be replaced by a smaller number.2) If this condition holds. φ) dt = 0 0 In view of (4.2 System (1. a contradiction. For any γ ∈ R we have ∞ ∞ −γt ∗ e e(−γ+1)t e−t q ∗ x(t.. This proves that q ∗ ∆−1 (λ)p(λ.1) holds and ﬁx φ ∈ C. The number tφ is given by tφ = lim sup ρ→∞ 1 ∗ −1 ln q ∆ (−ρ)p(−ρ.1) is degenerate with respect to q ∈ Cn \{0} if and only if q ∗ ∆−1 (λ)p(λ. φ) − ξ. φ λ for Re λ ≥ 0 and λ suﬃciently large. φ) is entire. φ) ≤ M . Next assume that (4. Then we have q ∗ x(t. (4. i. φ) = 0 for all t ≥ 0. φ) dt ≤ M. Then the PaleyWiener theorem (see Theorem B. then there exists a t1 > 0 such that q ∗ x(t. φ) = 0 for all φ ∈ C and all t ≥ t1 . φ) is entire for all φ ∈ C. which for t = 0 implies q = 0.6) implies that there exists a tφ > 0 such that q ∗ x(t. t ≥ 0. ρ→∞ ρ .1). φ) ≡ 0. We cannot have q ∗ ∆−1 (λ)p(λ. q ∗ ∆−1 (λ)p(λ. Assume ﬁrst that (4. φ) converges absolutely for all λ ∈ C. 0 Thus we have shown that the Laplaceintegral for q ∗ x(t. From Lemma 1. φ) ≤ q const.2) holds. As in the proof of Theorem 3.
φ) ρ 1 1 ≤ lim sup ln q ∗ adj ∆(−ρ) + lim sup ln eρr (r + 1)φ ρ→∞ ρ ρ→∞ ρ 1 = lim sup ln q ∗ adj ∆(−ρ) + r.3) 0 < t1 ≤ nr.2 has the disadvantage that it refers to all initial functions φ ∈ C.112 DELAY DIFFERENTIAL EQUATIONS where ξ is deﬁned in Theorem 3.2. ρ ρ→∞ Using the estimate (1. (4.36) we easily get lim sup ρ→∞ 1 ∗ ln q adj ∆(−ρ) ≤ (n − 1)r.5): Corollary 4. Using the estimate p(−ρ. 0 = h0 < · · · < hm = r.1) has the form x(t) ˙ = m Aj x(t − hj ). If system (1.1) cannot be degenerate.4) j=0 then the condition of the corollary is equivalent to det Am = 0. then the equation (1.3 If ξ = nr. The result of Proposition 4. In the following theorem we eliminate this shortcoming.1) is degenerate then (4. Proof. . The proof of Proposition 4. In case equation (1.2 provides also an estimate for the degeneracy time: (4. The following result will be improved below (see Corollary 4. ρ Thus we have established that 0 < tφ ≤ (n − 1)r + r − ξ = nr − ξ for all φ ∈ C.3) shows that ξ < nr. φ) ≤ 1 + reρr φ for ρ ≥ 0 we obtain lim sup ρ→∞ 1 ∗ ln q adj ∆(−ρ)p(−ρ.
4 (Kappel [18]) Let q ∈ Cn \{0} be given. −r ≤ θ ≤ 0.5 a) Let q ∈ Cn \ {0} be given. (4.113 Degenerate delay equations Theorem 4.1). t ≥ 0. Then q ∗ ∆−1 (λ)g(λ) has a pole at λ0 . Consequently q ∗ x(t) cannot be zero for t ≥ t1 .5) Proof. Corollary 4. Suppose (4. φ) = q ∗ Y (t − τ )L(φτ ) dτ = 0. φ) = Y (t)φ(0) + 0 Using (4. In view of Proposition 4. From q ∗ x(t) := Res eλt q ∗ ∆−1 (λ)g(λ).4 we have that q ∗ ∆−1 (λ) is entire. and consequently φτ = 0 for τ ∈ (t − t1 . because eλt q ∗ ∆−1 (λ)g(λ) has a pole at λ0 .1) is degenerate with respect to q if and only if q ∗ ∆−1 (λ) is entire. t−t1 because we have τ + θ > t − t1 − r ≥ 0.6) b) If ξ ≥ (n − 1)r then equation (1.11) t Y (t − τ )L(φτ ) dτ. λ=λ0 t ∈ R.2 it is clear that (4. Then equation (1. then according to Theorem 4. For φ ∈ C we have (see Theorem 1. Since q ∗ x(t) is real analytic it cannot have an accumulation point in R of zeros.6) we get for t ≥ t1 + r t ∗ q x(t.1) is degenerate with respect to q if and only if there exists a t1 > 0 such that7 q ∗ Y (t) = 0 for t ≥ t1 . (4. Thus equation (1. Proposition 2. λ=λ0 t ∈ R. Proof. If on the other hand equation (1. Then there exists a λ0 ∈ σ(L) such that λ0 is a pole also of q ∗ ∆−1 (λ). x(t. t]. is a generalized eigenfunction corresponding to λ0 .15 implies that x(t) := Res eλt ∆−1 (λ)g(λ). In order to prove necessity let us assume that q ∗ ∆−1 (λ) is not entire. Then equation (1.6) is true. . we see that q ∗ x(t) is not identically zero. 7Y (·) denotes the fundamental matrix solution of (1.1) is degenerate with respect to q.5) is suﬃcient.1) is not degenerate. Choose a function g(λ) which is holomorphic at λ0 and satisﬁes g(λ0 ) = 0.1) is not degenerate.
. equivalently. Therefore we can apply the PaleyWiener and get q ∗ Y (t) = 0 for ∗ −1 theorem t ≥ td = lim supρ→∞ (1/ρ) ln q ∆ (−ρ) . Therefore condition(4. ∞). . Proof. ψ) = ∗ −1 q ∆ (λ)p(λ. q ∗ x(t. Since lim sup(1/ρ) ln q ∗ adj ∆(−ρ) ≤ (n − 1)r.4 is the starting point for various constructions leading to degenerate functional diﬀerential equations of the form (4. i. Then there exist a ψ ∈ C and a sequence (tn )n∈N ⊂ [0. φ) is analytic on R.1) cannot be degenerate with respect to any q ∈ Cn \ {0}. The number td cannot be replaced by a smaller one. ψ) = 0. . ω ∈ R. on [0. n = 1.7) q ∗ φ(θ) = 0.. ψ) cannot be the Laplacetransform of a function with . ∞) with tn → ∞ such that q ∗ x(tn .7). This implies q ∗ Y (t) = 0. Proposition 4. Furthermore the function q ∗ x ˆ(λ.1). φ) = 0. The next result shows that degeneracy is in fact a property of the generalized eigenspaces of system (1.2 we see that td = lim sup ρ→∞ 1 ∗ ln q adj ∆(−ρ) − ξ. By continuity of Y (t) we get q ∗ = q ∗ Y (0) = 0. .5) (see [19]). φ) ≡ 0 for all φ ∈ P . system (1.6 System (1. . . ρ In order to prove b) we assume that ξ ≥ (n − 1)r. Assume that equation (1. . n = 1. a contradiction.e. ρ→∞ we get the estimate td ≤ (n−1)r −(n−1)r = 0. Therefore we only have to prove suﬃciency of condition (4. −r ≤ θ ≤ 0. 2.7) we see that ψ ∈ / P . for all φ ∈ P = λ∈σ(L) or. .114 DELAY DIFFERENTIAL EQUATIONS But q ∗ ∆−1 (λ) is of exponential type and squareintegrable on λ = iω. Note that t → x(t. If q ∗ ∆−1 (λ) ≡ 0 then we have q ∗ Y (t) = 0 a. t ≥ −r.1) is not degenerate with respect to q.7) is equivalent to the following: For all φ ∈ P there exists a sequence (tn )n∈N ⊂ R which has an accumulation point on R and q ∗ x(tn . As in the proof for Theorem 3. By (4. 2. Taking t = 0 we get q = 0.1) is degenerate with respect to q ∈ Cn \{0} if and only if ! Pλ (4.e. Theorem 4.
Then φ = limk→∞ (φk + ψk ). . −r ≤ θ ≤ 0. . Then the following is true: a) We have 1 0 < td ≤ lim sup ln q ∗ adj ∆(−ρ) − ξ. As in the proof of Theorem 2. Proposition 4. Since q ∗ x(t. λ=λ0 Since this function is real analytic it cannot have an accumulation point of zeros in R. φk + ψk ) = 0 for t ≥ lim supρ→∞ (1/ρ) ln q ∗ adj ∆(−ρ) − ξ.7 Let system (1. Let ﬁrst φ ∈ P ⊕ Q and set φ = φP + φQ . ψ).2 show that the degeneracy time depends on the small solutions only. t ≥ 0. . Thus we have shown that q ∗ x From Proposition 4. 2. k = 1. ψ) has a pole at λ0 . . 2. we get also q ∗ x(t. 2. ρ→∞ It is obvious that ξ = lim supρ→∞ (1/ρ) ln  det ∆(−ρ) = 0 in case that det ∆(λ) is a polynomial in λ only. ψ) ≡ 0 for all ψ ∈ P . then 1 0 < td ≤ lim sup ln q ∗ adj ∆(−ρ) . . we conclude that q ∗ Res eλt ∆−1 (λ)p(λ. We investigate the function y(t) = q ∗ x(t − r. ψ) ≡ 0 on R. φ) = q ∗ ∆−1 (λ)p(λ. φ) = 0 for t ≥ lim sup(1/ρ) ln q ∗ adj ∆(−ρ) − ξ. Proposition 4. ρ→∞ ρ Proof.1) be degenerate with respect to q ∈ Cn \ {0} and denote by td the degeneracy time of system (1. .1) is degenerate with respect to q. ρ→∞ ρ Next choose φ ∈ C = P ⊕ Q. which is a contradiction to ˆ(t.17. φP ) = 0 for t ≥ 0. . for a sequence (t˜n )n∈N ⊂ [0. . Consequently q ∗ x(t. . ∞) with t˜n → ∞. Then we have q ∗ x(t. k = 1. φ) is entire. φQ ). ψ) = 0. . φQ ) for t ≥ −r. . It is clear that λ0 ∈ σ(L). because q ∗ x(t. Therefore ψ(θ) = Resλ=λ0 eλθ ∆−1 (λ)p(λ. φQ ) = 0 for t ≥ lim sup ln q ∗ adj ∆(−ρ) − ξ. is a function in Pλ0 such that q ∗ x(t˜n .2 we see that 1 q ∗ x(t. ψk ∈ Q. where φk ∈ P . φ) = q ∗ x(t. c).2 we conclude that system (1. As in the proof of Theorem 3. .7). (4. . n = 1.1). Assume that q ∗ x ˆ(λ.115 Degenerate delay equations bounded support. ρ→∞ ρ b) If det ∆(λ) is a polynomial in λ only.6 and the characterization of small solutions as given in Theorem 3. φP ∈ P . φQ ∈ Q.
i = 1. If k = 1 the delays are commensurate and (4. . 0 = h0 < · · · < hm = r. . j = 1.8) can now be written as x(t) ˙ = m k Aj x t − βj. . . . Let r˜1 . . j = 1. (4.i ri . . . m}. . . k. i = 1. . We set β0. .i ∈ Q. generate K and are linearly independent. i=1 where αj. k. .i ∈ Z. m. We can only prove these results for diﬀerencediﬀerential equations with a ﬁnite number of commensurate delays. . . i = 1. . . i = 1. j = 1. . Then we have hj = k αj. . r˜k ∈ K be linearly independent elements which also generate K. . . k.i ≥ 0. . elements ri = α˜ We have k βj.i ’s. . . . . . . . . j = 1. . . Denote by α the least common multiple o the denominators of the αj. Equation (4. Then we explain the meaning of this necessary condition for the structure of the degenerate delay equation. . .i = 0.2 DELAY DIFFERENTIAL EQUATIONS A necessary condition for degeneracy In this section we ﬁrst prove a necessary condition for degeneracy of delay equations in terms of a speciﬁc factorization of the characteristic function.10) .8) j=0 View R as an (inﬁnite dimensional) vector space of the ﬁeld Q of rational numbers and denote by K the positive cone generated by8 h1 .i r˜i . .116 4. . (4. We start with diﬀerencediﬀerential equations with a ﬁnite number of delays: x(t) ˙ = m Aj x(t − hj ). m. . . hm . . αj.9) has the form x(t) ˙ = m Aj x(t − j h) j=0 8K = {α1 h1 + · · · + αm hm  αi ∈ Q and αi ≥ 0. hj = i=1 where now βj. . m. . . i = 1. j=0 (4. Then also the ri . .i ri . m. .9) i=1 In case k ≥ 2 we call the delays noncommensurate. k.
−1 q adj λI − A(µ) = p1 (λ. . i ∈ N. . e−λh .10) we have with µ = e−λh A(µ) = m Aj µj . e−λrk and det ∆(λ) = χA λ. k.. However. e−λrk . µ) . µ1 . . λ and µ as indeterminates. µ) p1 (λ. µ) . . . p0 (λ.1 · · · µk j.. . j=0 where µi = e−λri . . .. µk ) = m β β Aj µ1 j. . i = 1. . µ) = χA (λ. . . . µ = e−λh . .1 · · · e−λrk j. In case of system (4. .k . . . . . . . µ) . e−λr1 .117 Degenerate delay equations with 0 = 0 < 1 < · · · < m . In the following we shall treat λ. . µ). . m.k . . . .k j=0 m β β Aj µ1 j. µk ] by A(µ1 . . . occasionally we have to remember that µi = e−λri . It is not diﬃcult to see that ∆(λ) = λI − = λI − = λI − m k Aj exp −λ βj. . Our next goal is to prove Zverkin’s necessary condition for degeneracy of systems of the form (4. .. µ). µ) p0 (λ. . j=0 ∆(λ) = λI − A e−λh and det ∆(λ) = χA λ. . ∗ This implies −1 = q λI − A(µ) ∗ pn (λ.1 · · · µk j. .i ri j=0 i=1 m β β Aj e−λri j. . . j=0 Then we have ∆(λ) = λI − A e−λr1 .10). i = 1. i = 1. . µk resp. . We deﬁne A ∈ Cn×n [µ1 . pn (λ. k resp. We introduce the notation p0 (λ.
µ) = λκ + q˜(λ. ν = 1.118 DELAY DIFFERENTIAL EQUATIONS We have the following claim: If system (4. µ)R1 (λ. . Consequently we also have ∆(λ) = s0 (λ)s1 (λ. . . then pj (λ.1 with m = 2 we see that there exist R1 . µ) = w(λ)d(λ. µ) with λdegree q˜ < κ. .9 (Zverkin [32]) Assume that system (4. µ) with some monic polynomial s1 ∈ C[λ. µ] and a w ∈ C[λ] such that p(λ.10) is degenerate. . e−λh has inﬁnitely many diﬀerent zeros λ1 . We shall need the following lemma on greatest common divisors of polynomials in C[λ. . i. This implies q = 0. a contradiction. q ∈ C[λ. Proof. . By assumption (ii) also p λ. . λ2 . µ)/∂µ ≡ 0 and (iii) p(λ. Then d must depend on µ. R2 ∈ C[λ. µ) ≡ 0 for at least one j ∈ {1. Theorem 4. then the last equation implies w(λν )d(λν ) = 0. . . (see [2]).8 Let p. µ) Then we have ∂d/∂µ ≡ 0. (ii) ∂p(λ. Then there exists a monic polynomial s0 ∈ C[λ] with degree s0 ≥ 1 such that det(λI − A(µ)) = s0 (λ)s1 (λ. d really depends on µ. µ) + q(λ. We assume that (i) q(λ. . d = d(λ). Lemma 4. If d does not depend on µ. Because of assumption (i) the function q λ. which is not possible. This implies q ∗ (λI − A(µ))−1 ≡ 0 and – taking µ = e−λh – q ∗ ∆−1 (λ) ≡ 0 in C. . µ]. . n}. µ).10) is degenerate with respect to q. Note that d ≡ 0 because p ≡ 0 and q ≡ 0. e−λh has at least the zeros λ1 . q(λ.. µ)R2 (λ. 2. (4. . From Lemma D. µ]. . µ) is entire. e−λh ). µ] \ {0} and let d be a greatest common divisor. λ2 .e.11) In order to prove the claim suppose that p1 ≡ · · · ≡ pn ≡ 0. .
e−λh )/p0 (λ. Since pj0 (λ.10 Let S(µ) = S0 +· · ·+Sk µk be a unimodular matrix in Cn×n [µ] with inverse S −1 (µ) = T0 +· · ·+Tk µk .8 and get p0 (λ. µ) of a matrix A(µ) ∈ Cn×n [µ] we can transform A by a similarity transform with a unimodular matrix to a blocktriangular form.e.11) we have pj0 ≡ 0 for some j0 ∈ {1. In this section we want to investigate the meaning of this transformation for the delay equation (4.5 states that corresponding to a factorization of the characteristic polynomial χA (λ. e−λh )/s0 (λ. Thus we have ∂s0 /∂µ ≡ 0.4. Cn ) → C(−∞. We follow the algebraic proof given in [18]. Deﬁnition 4. t ∈ R. ∞.13) x(t) ˙ = A(σ)x (t) =: L(xt ) . But then dd˜ would be a mon divisor d˜ of sj0 and s0 with ∂ d/∂µ common divisor of pj0 and p0 . If p0 (λµ) = det(λI − A(µ)) does not depend on µ. A˜ ∈ Cn×n [µ] be given with A˜ = S −1 AS. i. .3 Coordinate transformations with delays Theorem D. then(4. by Theorem 4. π −1 be the mappings deﬁned by (4. µ) = d(λ. a contradiction with the deﬁnition of d because d˜ is nonconstant. . µ)sj0 (λ. For x. n}. ∞. µ). where d is a greatest common divisor of p0 and pj0 .. ∞. Cn ). Since. t ∈ R.119 Degenerate delay equations Proof. y ∈ C(−∞. q ∗ ∆−1 (λ) is entire for some q ∈ Cn \ {0}. Moreover we have ∂d/∂µ ≡ 0. π −1 : C(−∞. µ) and pj0 (λ. Let A.12) If we deﬁne the shift σ : C(−∞. Cn ) → C(−∞. e−λh ) = sj0 (λ. µ)s0 (λ. π −1 y = S(σ)y. . . From this it is also clear that π and π −1 are indeed inverses of each other.12) Furthermore let (4. x. Cn ) we deﬁne the linear mappings π. y ∈ C(−∞. µ) = d(λ. ∞. (4. ∞. Cn ) by (σx)(t) = x(t − h). then there is nothing to be proved. 4. Cn ) by (πx)(t) = T0 + · · · + Tk x(t − kh) and (π −1 y)(t) = S0 + · · · + Sk y(t − k h).10) corresponding to A.12) can also be written as πx = S −1 (σ)x. Lemma 4. where S ∈ Cn×n [µ] is unimodular and let π. Assume that ∂p0 /∂µ ≡ 0. we can apply Lemma 4. According to the claim (4. e−λh ) is entire.8 implies that in case ∂s0 /∂µ ≡ 0 there exists a greatest com˜ ≡ 0. s0 ∈ C[λ]. ∞.
. ∞. (4. j! t ∈ R. . In order to give more detailed information on the transformations π and π −1 we introduce some notation: ˜ ∆(λ) := λI − A˜ e−λh = S −1 e−λh ∆(λ)S e−λh .13) resp. .11 a) If x ∈ C(−∞. . Then it is easy to see that y is diﬀerentiable at least for t ≥ t0 + kh and y(t) ˙ = S −1 (σ)x˙ (t) = S −1 (σ)A(σ)x (t) ˜ (t). ˜ be given and let.12 Let λ0 ∈ σ(L) = σ(L) x(t) = eλ0 t tj a. for some a ∈ Cn .. t ≥ t0 + kh. 1! ∆ (λ) ˜ 0 ··· 0 ∆(λ) cj ∈ Cn . . . i.. ∞. .. . . A. πL1 = L2 . Then we have (πx)(t) = eλ0 t j tκ κ=0 1 dj−κ −1 −λ0 h S e a. j = 0. Cn ) be a solution of (4.e. ∞.. . ∆k+1 (λ) = . k. P˜λ0 . b) Let L1 resp.14) at least for t ≥ t0 + kh. y which are solutions of (4. = S −1 (σ)A(σ)S(σ)y (t) = A(σ)y b) In view of part a) of the theorem it is clear that πL1 ⊂ L2 . . For y ∈ L2 we have x = π −1 y ∈ L1 and πx = y.14) ˜ Then we have the be the delay systems corresponding to A resp. L2 denote the subspaces of C(−∞. . a) let x ∈ C(−∞. . Set y(t) = T0 x(t) + · · · + Tk x(t − kh).. Then π maps L1 onetoone onto L2 . . .13) for t ≥ t0 .14) corresponding to λ0 .14) on R. ˆ ∈ Cnk×nk . ck ).e.120 DELAY DIFFERENTIAL EQUATIONS and ˜ ˜ t) y(t) ˙ = A(σ)y (t) =: L(y (4.. .13) for t ≥ t0 . cˆk+1 = col(c0 . Proof. . i.. generalized eigenspace of equation (4. Cn ) consisting of all x resp. following result: Proposition 4. then y = πx solves (4. x is diﬀerentiable for t ≥ t0 and x(t) ˙ = A(σ)x (t). 1 ˜ .. 1 ˜ 1 ˜ (k) ˜ ∆(λ) 1! ∆ (λ) · · · k! ∆ (λ) . Cn ) solves (4. Lemma 4. . t ≥ t0 . κ! (j − κ)! dλj−κ · t ∈ R. . 0 .
e..121 Degenerate delay equations Proof.13 Choose λ0 ∈ σ(L) = σ(L) λ0 t Xk (λ0 ) = e k−1 j t j=0 j! ˜ k (λ0 )˜bk = 0 bj  ∆ k−1 j λ0 t t ˜ ˆ k (λ0 )ˆ Xk (λ0 ) = e ck = 0 . . . . . ˜ 0 )˜bk = 0. . 2. . .12 we get λ0 t (πx)(t) = e j k−1 tκ j=0 κ=0 = eλ0 t 1 dj−κ −1 −λ0 h e bj S κ! (j − κ)! dλj−κ k−1 k−1 κ t κ=0 κ! j=κ dj−κ −1 −λ0 h 1 S e bj . Observing j κ j! κ!(j − κ)! = the deﬁnition of π implies: k 1 (t − h)j eλ0 (t−h) T a (πx)(t) = j! =0 j k j κ 1 −λ0 h λ0 t t (−h)j−κ e T a =e κ j! =0 j tκ = eλ0 t κ=0 j = eλ0 t κ=0 κ! κ=0 k 1 e−λ0 h (−h)j−κ T a (j − κ)! =0 1 dj−κ −1 −λ0 h t−κ e a. ˜ and deﬁne the subspaces Lemma 4. Let x(·) ∈ Xk (λ0 ) be given. S κ! (j − κ)! dλj−κ t ∈ R. i. cj  ∆ j! k = 1. Then π maps Xk (λ0 ) onetoone onto X Proof. x(t) = eλ0 t k−1 j t j=0 j! bj . (j − κ)! dλj−κ . ∆(λ From Lemma 4. k = 1. j=0 ˜ k (λ0 ). . 2.
6 applied to ∆(λ) = S −1 e−λ0 h ∆(λ)S e−λ0 h implies (see (2. .11 we see that ker(λ0 I − A) = ˜ = {y [−˜r. which is satisﬁed in a neighborhood of λ0 .. . S (j − κ)! dλj−κ κ = 0. .0]  y ∈ X ˜ k (λ0 )}. κ0 + 1.3) for the deﬁnition of S˜k (λ0 )) ˜ k (λ0 )S˜k (λ0 ). so that ˜ k (λ0 )˜bk = 0. 0 . −λ h −1 e 0 0 ··· 0 S Lemma 2. .16) we get From ∆ ˜ k (λ0 )S˜k (λ0 )T˜k (λ0 )˜bk . ˆ k (λ0 ) = T˜k (λ0 )∆ ∆ (4. where the matrix Tk (λ0 ) is deﬁned by −1 −λ h 1 dk−1 S −1 e−λ0 h e 0 · · · · · · (k−1)! S dλk−1 . k − 1. . ˜ k (λ0 ) ⊂ πXk (λ0 ) ⊂ X π π −1 X Remark. . i.. ˆ 0 )ˆ ck = T˜k (λ0 )∆ ∆(λ ˜ k (λ0 ). [−˜ r. . P˜λ0 = ker(λ0 I − A) k = κ0 . . .13) resp. . . . . .15) T˜k (λ0 ) = .122 DELAY DIFFERENTIAL EQUATIONS We set cκ = k−1 j=κ dj−κ −1 −λ0 h 1 e bj . . ..6 applied to S e−λh S −1 e−λh ≡ I. .. In order to prove that πx ∈ X ˜ ˜ ˜ It is easy to see that cˆk = Tk (λ0 )bk . 0] denote the delay interval for system (4. . We {x [−r.. By Corollary 2. πXk (λ0 ) = X ˜ k (λ0 ). (4..e. 0] resp. (4. ˆ k (λ0 )ˆ ck = T˜k (λ0 )∆ ∆ Lemma 2. Let [−r.. Consequently we have X π we obtain π X ˜ k (λ0 ).14). ˜ k.16) ˜ k (λ0 )˜bk = 0 and (4. Taking π −1 instead of Thus we have shown that πXk (λ0 ) ⊂ X −1 ˜ k (λ0 ) = ˜ k (λ0 ) ⊂ Xk (λ0 ). ˆ k (λ0 )ˆ ˜ k (λ0 ) we have to show that ∆ ck = 0.0]  x ∈ Xk (λ0 )} and ker(λ0 I − A) also have seen in Section 7 that Pλ0 = ker(λ0 I − A)k . . gives S˜k (λ0 )T˜k (λ0 ) = I ∈ Cnk×nk .
φ) = .κ0 −1 j (θ /j!)bj . φ(θ) = eλ0 θ j=0 ˜ ˜ −r ≤ θ ≤ 0. If we agree to extend functions φ ∈ Pλ0 . to all of R by setting x(t. with ∆κ0 (λ0 )bκ0 = 0.
123 Degenerate delay equations .
of (4. .κ0 −1 j eλ0 t j=0 (t /j!)bj . we can take Z˜ρ. mρ with mρ > 0. φ) [−˜r. .j = dρ . . which exist according to ˜ ˜ −λh ) at = λI − A(e Theorem 2. Then there exist uniquely determined numbers ≥ 1.17) ρ=1 j=1 where the subspaces Zρ. . . .j is Acyclic and Z˜ρ. ˜ j = 1. that π ˙ for φ ∈ Pλ ˙ φ) = x(·. ρ = 1. where % % P = λ0 ∈σ(L) Pλ0 . mρ . . . ρ = 1. P˜ = λ0 ∈σ(L) P˜λ0 .17).j of P˜λ0 satisfy (i) dim Zρ. φ) ˙ φ) [˜r.5 and the proof for part (iv) of Theorem 2. . ker(λ0 I − A) The next result states that a coordinate transform with delays given by a unimodular matrix does not change the structure of the generalized eigenspaces as given in Theorem 2. . = dt dt dt .0] = πφ = Aπφ. .4 for ∆(λ) = λI − A(e−λh ) resp. πAφ = π φ˙ = πx(·. Proof. . for ∆(λ) λ0 are the same. . . mρ . .j and P˜λ0 = mρ ! ! Z˜ρ.0] = π x(·.j . ρ = 1. Z˜ρ. . . mρ .14). (ii) Zρ. .j . P˜λ the S ∈ Cn×n [µ]. . . . dρ . Using the facts that the action of A and of A˜ is diﬀerentiation. . as follows: πφ = πx(·.4 we see that the numbers . (4. φ) and π commute with diﬀerentiation and that x(·. φ) [−˜r. .j is Acyclic.0] .12.0] = πx(·. 0 (see Theorem 1.13) resp. . . j = 1. . .9) we get. 2. j = 1. . . ρ = 1. . . for φ ∈ Pλ0 .14 Let A(µ) = S −1 (µ)A(µ)S(µ) with a unimodular matrix ˜ denote by Pλ resp. From Lemma 2. dρ .j = dim Z˜ρ. ˜ Theorem 4.0] d d d ˜ πx(· : φ) [−˜r. then we can deﬁne a mapping π : P → P˜ .13 implies that π also maps ker(λ0 I − A)k onetoone onto ˜ k . . . .j = πZρ. This establishes the decomposition (4. For λ0 ∈ σ(L) = σ(L) 0 0 generalized eigenspace of system (4. ρ = 1. . . mρ . . .j of Pλ0 resp. Moreover. k = 1. . . The result of Lemma 4. satisfying 0 ≤ d1 < · · · < d such that Pλ0 = mρ ! ! ρ=1 j=1 Zρ. . Then it is clear that π maps P onetoone onto P˜ . ˙ [−˜r.
i. . .2 −1 ˜ A = S AS = 0 A2 with A1 ∈ Cn1 ×n1 [µ]. k − 1. k − 1. χA2 = p2 .12). . 4. Then there exists a unimodular matrix S ∈ Cn×n [µ] such that A1 A1. k = dim Zρ. j = 1. Z˜ = πZ is also cyclic with dim Z˜ = dim Z = k.5) suggest that degenerate systems with commensurate delays can be transformed to a system which consists of two coupled subsystems. µ)p2 (λ) with monic and relatively prime polynomials the polynomial p1 does not contain a nontrivial factor independent of µ.15 Assume that system (4. and χA1 = p1 . . . .4 The structure of degenerate systems with commensurate delays The necessary condition for degeneracy of systems with commensurate delays given in Zverkin’s theorem together with the triangularization theorem for polynomial matrices over the ring C[µ] (see Theorem D. . j = 1. .j . Aφj = λ0 φj − φj−1 . For φ˜j = πφj . µ) = p1 (λ. A2 ∈ Cn2 ×n2 [µ]. .j = dρ and choose a basis φ0 . n2 = degree p2 > 0.124 DELAY DIFFERENTIAL EQUATIONS Set Z = Zρ.13) is degenerate with respect to q ∈ Cn \ {0} and that in the factorization χA (λ. where n1 = λ– degree p1 . . where one of these subsystems is degenerate and has a polynomial of λ only as its characteristic function. . . (see the paragraph before Theorem 2.. Loosely speaking. .e. The precise statement is the following: Theorem 4. . j = 0. Remark. . φk−1 of Z satisfying Aφ0 = λ0 φ0 . It is easy to see (but tedious to formulate) that all results presented in this section hold also with appropriate notational changes for the case of noncommensurate delays. . k − 1. ˜ 0 = πAφ0 = λ0 πφ0 = λ0 φ˜0 and A˜φ˜j = Aπφ ˜ j = we get A˜φ˜0 = Aπφ πAφj = π(λ0 φj − φj−1 ) = λ0 φ˜j − φ˜j−1 . degenerate systems with commensurate delays are those which are similar (under transformations with delays) to ordinary diﬀerential equations. .
= λI − A(µ) 0 λI − A2 (µ) Consequently we get −1 ˜ (λ) = q ∗ W1 (λ). q ∗ S e−λh ∆ where −1 . e +V e λI − A2 e−λh . . µ). µ) = sj0 (λ. µ)d(λ. the delay system corresponding to A2 (µ) is degenerate with respect to a q2 ∈ Cn2 \ {0}.18) is entire. 0) ∈ Cn2 . µ). . Without restriction of generality we can assume that q = col(0. Suppose that q ∗ U (µ) ≡ 0. µ). W1 (λ) = U e−λh λI − A1 e−λh −λh −λh −1 −λh W2 (λ) = U e B1. According to Theorem 4. Otherwise we choose a invertible matrix T ∈ Cn×n such that q ∗ T = (0. rn1 (λ. µ) in the representation −1 = q ∗ U (µ) λI − A1 (µ) 1 r1 (λ. By Zverkin’s result it is clear that n2 > 0.125 Degenerate delay equations Moreover.4 the function q ∗ ∆−1 (λ) is entire.18) q ∗ S e−λh S −1 e−λh ∆−1 (λ)S e−λh = q ∗ S e−λh ∆ is entire. −1 ˜ is given by The matrix λI − A(µ) ' & −1 −1 λI − A (µ) B (λ. Then we see that at least one rj0 (λ. V ) where U ∈ Cn×n1 [µ] and V ∈ Cn×n2 [µ]. p1 (λ. It remains to prove that we can choose S such that the delay system corresponding to A2 (µ) is degenerate with respect to q2 . Since the matrix (4. the delay system corresponding to A1 (µ) cannot be degenerate. whereas the delay system corresponding to A1 (µ) is not degenerate. µ) does not contain a nontrivial factor which depends in λ only. µ) 1 1. This implies that also −1 ˜ (λ) (4. q2 ) with q2 = col(1. . We set S = (U. Since p1 (λ. As in the proof of Theorem 4. . we see that also q ∗ W1 (λ) is entire. µ) = s0 (λ)d(λ. . Proof. .5 ˜ has the there exists a unimodular matrix S ∈ Cn×n [µ] such that A(µ) properties given in the theorem. . W2 (λ) .2 ˜ −1 . According to Theorem D. . 0. µ) p1 (λ.9 we see that rj0 (λ. . µ) is not identically zero. q2∗ ) and replace A(µ) by T −1 A(µ)T .2 λ.
Set v(µ) = v1 (µ). q ∗ V e−λh λI − A2 e−λh −1 . .19) is entire. a contradiction. Since the vector (4. µ) has a nontrivial factor depending on λ only. 0. T (µ)A(µ)T (µ) = 0 P (µ)A2 (µ)P −1 (µ) . Consequently (see Theorem 4. we see that also −1 −λh −1 P e q2∗ P e−λh λI − A2 e−λ is entire.2 (µ)P −1 (µ) −1 . i. .. If instead of S(µ) we take the unimodular matrix T (µ) = S(µ) diag(I. P −1 (µ)). . . Consequently we get −1 ˜ (λ) = 0. v(µ) ∗ P (µ) = . But then p1 (λ. q ∗ V (µ) = v(µ). .19) ˜ −1 (λ) ≡ We have q ∗ V (e−λh ) ≡ 0.126 DELAY DIFFERENTIAL EQUATIONS From λ – degree rj0 < λ – degree p1 we conclude that degree s0 > 0. . v(µ)) is the (n1 + 1)st row of S(µ). q ∗ S e−λh ∆ (4. ∗ Using the special form of q2 we obtain q ∗ V (µ) = v(µ) = q2∗ P (µ). This and q ∗ U (µ) ≡ 0 imply that (0. vn2 (µ) and let w be a greatest common divisor ˜ and det S(µ) ∈ C \ {0} imply of v1 . This implies −1 −1 q2∗ P (µ) λI − A2 (µ) P −1 (µ) = q ∗ V (µ) λI − A2 (µ) P −1 (µ). which implies q = 0. which contradicts the assumption made on p1 .4) the delay system corresponding to A˜2 (µ) := P (µ)A2 (µ)P −1 (µ) is degenerate with respect to q2 . Then S(µ) = w(µ)S(µ) w = α ∈ C \ {0}. . vn2 .. .. . . Under this condition we can ﬁnd a unimodular matrix P ∈ Cn2 ×n2 [µ] which has v as its ﬁrst row. . .e. Let v(µ) be the (n1 + 1)st row of V (µ). Otherwise we would have q ∗ S(e−λh )∆ 0. . This proves that q ∗ U e−λh ≡ 0. then the matrix T −1 (µ)A(µ)T (µ) is given by A1 (µ) A1.
If the Laplaceintegral for f converges absolutely at λ0 . Re λ > σc (f ). (A.1) exists uniformly for λ ∈ D. ∞. It is clear that −∞ ≤ σc ≤ σa ≤ ∞.1 Let fˆ and gˆ be Laplacetransforms of functions f and g. By the uniform convergence of the Laplaceintegral in sectors the function ∞ e−λt f (t) dt. but refer to the standard literature (see for instance [7]).e.1) 0 We say that the Laplaceintegral for ωf converges at λ if the limit in (A. Therefore the domain of convergence of the Laplaceintegral is an open right half plane plus a subset of the boundary.127 Laplacetransforms Appendix: A LaplaceTransforms In this appendix we collect some results on Laplacetransforms. C) and λ ∈ C. If fˆ = gˆ in some right half plane. . except equality everywhere. Therefore the domain of absolute convergence of the Laplaceintegral is an open or closed right half plane. 0 fˆ is called the Laplacetransform of f . fˆ(λ) = 0 is holomorphic in the half plane Re λ > σc (f ). The derivatives of fˆ are given by dj fˆ (λ) = (−1)j dλj ∞ e−λt tj f (t) dt. converges absolutely at λ if limω→∞ 0 e−t Re λ f (t) dt exists and converges uniformly on a set D ⊂ C if the limit in (A. With a few exceptions we do not provide proofs. are possible for these inequalities. Re λ > σc (f ). All cases. on t ≥ 0. Of fundamental importance is the following uniqueness result: Theorem A. Then the Laplaceintegral of f at λ is deﬁned by ∞ ω e−λt f (t) dt = lim ω→∞ 0 e−λt f (t) dt. 0 ≤ ε < π/2. then it is uniformly convergent in any sector {λ ∈ C   arg(λ − λ0 ) ≤ ε}. respectively.1) exists. We deﬁne the abscissa of convergence by σc (f ) = inf{σ ∈ R  The Laplaceintegral for f converges at σ}. We deﬁne the abscissa of absolute convergence by σa (f ) = inf{σ ∈ R The Laplaceintegral for f is absolutely convergent at σ}. Frequently we shall denote by fˆ or L(f ) also the function obtained by analytic continuation. then f (t) = g(t) a. If the Laplaceintegral converges at λ0 . Let f be in L1loc (0. so it also does in Re λ ≥ λ0 . In particular it converges in Re λ > Re λ0 . We shall also use the notation L(f ) = fˆ.
Theorem A. ∞. C). Then σa (f1 ∗ f2 ) ≤ x0 and L(f1 ∗ f2 ) = L(f1 )L(f2 ). The function f whose existence is guaranteed by the above proposition is called the convolution of f1 and f2 .e. Moreover. f (t) = o eλ0 t A. Then f = f1 ∗ f2 is locally absolutely continuous on t ≥ 0 and f˙ = f˙1 ∗ f2 + f1 (0)f2 a. 2. If f1 is bounded on bounded intervals or if f1 and f2 are in L2loc (0. then f = f1 ∗ f2 is continuous on t ≥ 0. then it is of interest to know if it is a Laplacetransform. Then the integral Proposition t f (t − τ )f2 (τ ) dτ exists a. Theorem A. t Theorem A. ∞. on t ≥ 0. Theorem A. 1 gˆ(λ) = fˆ(λ) λ and O(1) if λ0 = 0.e. 2. g(t) = if λ0 > 0 o eλ0 t as t → ∞. t ≥ 0.3 Let f be locally absolutely continuous on t ≥ 0 and suppose that σc (f˙) < ∞. .7 Let fi ∈ L1loc (0. f = f1 ∗ f2 . i = 1.5 Let fi ∈ L1loc (0. Assume that the Laplaceintegral for f is convergent at λ0 ∈ R.4 Let fi ∈ L1loc (0. be given. C). C). on t ≥ 0 and the function f deﬁned by this integral. ∞. We shall need the following results on smoothness of f1 ∗ f2 . If a function fˆ(λ) is holomorphic in some right half plane.128 DELAY DIFFERENTIAL EQUATIONS It is clear that taking Laplacetransforms is a linear operation. ∞. diﬀerentiation and convolution. i = 1. and by f (t) = 0 otherwise is in L1loc (0. for any real λ0 . Theorem A. C). We next collect results on the behavior of the Laplacetransform with respect to integration. 2. C). 0 1 where it exists.8 Let fˆ be holomorphic in Re λ > x1 and assume that (i) for any ε > 0 and any δ > 0 there exists a K > 0 such that fˆ(λ) ≤ ε for Re λ ≥ x1 + δ and λ ≥ K.2 Let f ∈ L1loc (0. Moreover. ∞. C). be given and suppose that σa (f1 ) ≤ x0 and σa (f2 ) ≤ x0 . The operation “ ∗ ” is commutative and associative. be given. Then σc (f ) ≤ σc (f˙) and L(f˙)(λ) = λL(f )(λ) − f (0).6 Let f1 be locally absolutely continuous on t ≥ 0 and f2 be in L1loc (0. C) with σc (f ) < ∞ and set g(t) = 0 f (τ ) dτ . we have as t → ∞. ∞. ∞. Concerning Laplacetransforms of convolutions we have: Theorem A. i = 1. Then the Laplaceintegral for g converges in Re λ > 0 if λ0 = 0 and in {λ ∈ C  Re λ > λ0 } ∪ {λ0 } if λ0 > 0. where the Laplaceintegral for f˙ converges.
t ≥ 0. Theorem A. . . . for any x > x1 is given by f (t) = eλt fˆ(λ) dλ.9 Let fˆν be the Laplacetransforms of fν .129 Entire Functions ∞ (ii) and the integral f (x + iy) dy exists for all x > x1 . . −∞ Then fˆ is the Laplacetransform of a function f which. . . . (ii) ν=1 0 . A result of more constructive nature is contained in the following theorem. . ν = 1. 2. (x) x+iy Here (x) means limy→∞ (2πi)−1 x−iy . 2. and assume that there exists an x0 ∈ R such that (i) x0 ≥ σa (fν ). ν = 1. ∞ ∞ e−x0 t fν (t) dt is convergent.
.
∞
ˆ
Then the series ∞
ν=1 fν converges uniformly in Re
λ∞≥ x0 and the series ν=1 fν (t)
absolutely a.e. on t ≥ 0. If we deﬁne f (t) =
ν=1 fν (t), then σa (f ) ≤ x0 and
ˆ
fˆ(λ) = ∞
ν=1 fν (λ) for Re λ ≥ x0 .
Finally we quote one version of the complex inversion formula for Laplacetransforms:
Theorem A.10 Let f ∈ L1loc (0, ∞; C) with σa (f ) < ∞ be given. Then for any
γ ≥ σa (f ) and any t ≥ 0 where f is of bounded variation in a neighborhood of t the
following formula is valid:
1
f (t + 0) + f (t − 0)
if t > 0,
λt ˆ
e f (λ) dλ = 21
f (0+)
if t = 0.
2
(γ)
For t < 0 the integral is always zero.
Appendix: B
Entire Functions
In this appendix we quote some facts on entire functions, which are used in the
section on small solutions. As standard references we cite [5, 23].
Deﬁnition B.1 A function f : C → C is called an entire function of exponential
type if and only if f is holomorphic in C and there exist constants α, β ≥ 0 such that
f (λ) ≤ αeβλ
The number
τf = lim sup
ρ→∞
for all λ ∈ C.
1
ln M (ρ),
ρ
where M (ρ) = maxλ=ρ f (λ), is called the exponential type of f .
130
DELAY DIFFERENTIAL EQUATIONS
It is trivial that the sum and the product of two entire functions of exponential
type are again such functions. Less trivial is the result on quotients:
Proposition B.2 If f and g are entire functions of exponential type, then f /g is
also an entire function of exponential type provided it is entire.
For a proof of this proposition see for instance [23, p. 24].
A detailed description of the growth behavior of an entire function of exponential
type at inﬁnity given by the indicator function:
Deﬁnition B.3 Let f be an entire function of exponential type. The function hf :
[0, 2π] → R deﬁned by
1
hf (θ) = lim sup ln f (ρeiθ )
ρ→∞ ρ
is called the indicator function of f .
For a detailed representation of the general properties of indicator functions see
for instance [23, Ch. 1, Sect. 15] or [5, Ch. 5]. We just mention the following result:
Proposition B.4 Let f be an entire function of exponential type. If f ≡ 0, then hf
is ﬁnite and continuous.
If we have information on the behavior of f on a vertical line, we obtain a very
precise information on the indicator function. For our purpose very useful is the
following result (cf. [23, p. 243] or [5, p. 116]):
Theorem B.5 Let f ≡ 0 be a function of exponential type. If
∞ +
ln f (iω)
dω < ∞,
1 + ω2
−∞
(B.1)
then
hf (θ) = −k cos θ,
θ ∈ [π/2, 3π/2],
and, for θ in a dense subset of [π/2, 3π/2],
hf (θ) = lim
ρ→∞
1
ln f (ρeiθ ).
ρ
The function ln+ is deﬁned by ln+ ξ = max(0, ln ξ), ξ > 0. Condition (B.1) can be
replaced by (see [23, p. 251])
∞
f (iω)2 dω < ∞.
(B.2)
−∞
Of fundamental importance for us is the following characterization of those entire
functions of exponential type which are twosided Laplacetransforms of functions
with compact support.
Theorem B.6 (PaleyWiener) Let f ≡ 0 be an entire function of exponential type.
Then the following two statements are equivalent:
(i) There exist constants H , H ∈ R with −H < H and a squareintegrable F with
support in [−H , H] such that
H
f (λ) =
e−λt F (t) dt, λ ∈ C.
−H
131
An Unsymmetric Fubini Theorem
(ii)
∞
−∞
f (iω)2 dω < ∞.
Moreover, if we take H = hf (0) and H = hf (π), then the interval [−H , H] cannot
be replaced by a smaller one.
For a proof of this theorem see [23, p. 387], [5, p. 103] or [7, Vol. III, pp. 238, 241].
Finally, Hadamard’s factorization theorem assumes the following form for functions
of exponential type (see [23, p. 24] or [5, p. 22]):
Theorem B.7 Let f be an entire function of exponential type and let (an ) be the
ﬁnite or inﬁnite sequence of nonzero zeros of f . Moreover, assume that λ = 0 is a
zero of multiplicity m. Then there exist numbers α, β ∈ C such that
f (λ) = λm eα+βλ
(
1−
n
λ pλ
e ,
an
λ ∈ C,
where p = 0 or p = 1.
Appendix: C
An Unsymmetric Fubini Theorem
For the convenience of the reader we state in this appendix the unsymmetric Fubini
theorem of [6]. In the following it is always assumed that functions which are of
bounded variation on bounded intervals are normalized by the requirement to be
righthand continuous.
Theorem C.1 Assume that the functions k : R → R, s : R → R and p : R × R → R
satisfy:
a) k is of bounded variation on bounded intervals.
b) s is Borel measurable on R.
c) p is Borel measurable on R × R and, for kalmost all σ ∈ R, the function p(σ, ·) is
of bounded variation on bounded intervals.
Deﬁne
τ
dν p(σ, ν)
for τ ≥ 0,
0
V (σ, τ ) =
0
dν p(σ, ν) for τ < 0.
−
τ
If
(i)
∞
−∞
V (σ, τ )dk(σ) < ∞
for all τ
and
(ii)
∞
−∞
∞
−∞
s(τ )dτ
dk(σ)
∞
−∞
∞
−∞
V (σ, τ )dk(σ) < ∞ or
s(τ ) dτ p(σ, τ ) < ∞,
132
DELAY DIFFERENTIAL EQUATIONS
then the other integral under (ii) also exists and
∞
∞
∞
s(τ )dτ
p(σ, τ )dk(σ) =
dk(σ)
−∞
−∞
−∞
∞
s(τ )dτ p(σ, τ ).
−∞
Appendix: D
Results on Polynomial Rings and Modules
over Polynomial Rings
In this appendix we present the basic deﬁnitions and results concerning the theory
of polynomial rings and modules over polynomial rings. We give proofs for those
results where a proof is not available in the literature in the form needed here.
We denote by C[λ1 , . . . , λm ] the ring of polynomials in the indeterminates λ1 , . . .
. . . , λm , m ≥ 1, over the ﬁeld C. The ring C[λ] is a Euclidean ring and consequently
also a principal ideal domain. The rings C[λ1 , . . . , λm ] with m ≥ 2 are not even
principal ideal domains, but unique factorization domains. Therefore the concepts
”greatest common divisor” and ”least common multiple” are well deﬁned in all of
these rings. If r1 , r2 are polynomial in C[λ], then there exists a greatest common
divisor d ∈ C[λ] which can be computed by the Euclidean algorithm. Moreover there
exist polynomials R1 , R2 ∈ C[λ] with
d = r1 R1 + r2 R2 .
This result in C[λ] has the following analogue in the rings C[λ1 , . . . , λm ], m ≥ 2:
Lemma D.1 Let r1 , r2 ∈ C[λ1 , . . . , λm ], m ≥ 2, be given and let d be a greatest
common divisor of r1 , r2 . Then there exist polynomials R1 , R2 ∈ C[λ1 , . . . , λm ] and
a polynomial w ∈ C[λ1 , . . . , λm−1 ] such that
r1 (λ1 , . . . , λm )R1 (λ1 , . . . , λm ) + r2 (λ1 , . . . , λm )R2 (λ1 , . . . , λm )
= w(λ1 , . . . , λm−1 )d(λ1 , . . . , λm ).
A proof of this result can be found in [29].
Next we state some deﬁnitions and results concerning module theory. As general
references for module theory we quote [1], [3], [4], [16], [27].
Let M be an additive, commutative group and R a commutative ring with multiplicative unit 1. Then M (more precisely (M, R)) is called a module over R if and
only if there is a scalar multiplication (r, x) ∈ R × M → rx ∈ M satisfying
r(x1 + x2 ) = rx1 + rx2 ,
r ∈ R, x1 , x2 ∈ M,
(r1 + r2 )x = r1 x + r2 x,
r1 , r2 ∈ R, x ∈ M, and
1 · x = x,
x ∈ M.
The modules which are most similar to vector spaces are the free modules. A module
M over R is called a free module if and only if there exists a basis e1 , . . . , en ∈ M
for M , i.e., every element x ∈ M has a unique representation x = α1 e1 + · · · + αn en ,
αi ∈ R, i = 1, . . . , n, and α1 e1 + · · · + αn en = 0 implies α1 = · · · = αn = 0 (linear
independence). The modules Cn [λ1 , . . . , λn ] are free modules, a basis being e1 =
Results on Polynomial Rings and Modules over Polynomial Rings
133
col(1, 0, . . . , 0), . . . , col(0, . . . , 0, 1) (canonical basis). A set U is called a submodule
of M if and only if it is again a module over R, i.e., for α1 , α2 ∈ R and x1 , x2 ∈ U we
have also α1 x1 + α2 x2 ∈ U .
We shall need results on the modules M = Cn [λ1 , . . . , λm ] over the ring R =
C[λ1 , . . . , λm ]. The module Cn [λ, . . . , λm ] consists of all nvectors col(x1 , . . . , xn )
with xi ∈ C[λ1 , . . . , λm ], i = 1, . . . , n. The module Cn [λ] is special among the modules
Cn [λ1 , . . . , λm ] by the following property:
Lemma D.2 Every submodule of C[λ] is a free module.
This result in fact is valid for free modules over principal ideal domains (see [16, p. 78]
or [27, p. 109]).
An important question in module theory is the following: If X is a submodule
of M , does there exist another submodule Y of M such that M = X ⊕ Y ? We
know that the answer is ”yes” for vector spaces. But even for the module Cn [λ] the
answer in general is ”no”. However we have also a positive result. In order to state
this result we need a deﬁnition. For a1 , . . . , ak ∈ Cn [λ] let S = span(a1 , . . . , ak ) =
{α1 a1 + · · · + αk ak  αi ∈ C[λ], i = 1, . . . , k}. Obviously S is a submodule of Cn [λ],
the submodule generated by a1 , . . . , ak .
Proposition D.3 Let A ∈ Cn×n [λ] (i.e., A is an n × nmatrix with elements in
C[λ]). Then there exists a submodule X of Cn [λ] such that
Cn [λ] = ker A ⊕ X.
We can take X = span(x1 , . . . , xk ) for any elements x1 , . . . , xk ∈ Cn [λ] such that
Ax1 , . . . , Axk is a basis for range A.
This result is a special case of the corresponding result for epimorphisms where the
range is a free submodule (see [27, p. 108]). Note that by Lemma D.2 the submodule
range A is a free module. Take a basis y1 , . . . , yk of range A and elements x1 , . . . , xk ∈
Cn [λ] with yi = Axi , i = 1, . . . , k. Observe that x1 , . . . , xk are linearly independent,
because otherwise also y1 , . . . , yk would be linearly dependent and could not be a
basis for range A.
For matrices A ∈ Cn×n [µ1 , . . . , µm ] we deﬁne the characteristic polynomial χA ∈
C[λ, µ1 , . . . , µm ] by
χA (λ, µ1 , . . . , µm ) = det λI − A(µ1 , . . . , µm ) .
For A ∈ Cn×n [µ1 , . . . , µm ] and p ∈ C[λ, µ1 , . . . , µm ] we set (with some abuse of
notation)
p(A) = p(A, µ1 , . . . , µm ) ∈ Cn×n [µ1 , . . . , µm ].
Then as in vector space theory we have (see for instance [27, p. 246])
Theorem D.4 (CayleyHamilton) For every matrix A ∈ Cn×n [µ1 , . . . , µm ], m ≥
1, we have
χA (A) = 0.
On the basis of Proposition D.3 we can prove the following result on blocktriangularization of matrices in Cn×n [µ]:
Theorem D.5 ([19, Theorem 2.1]) Let A ∈ Cn×n [µ] be given and assume that
χA = p1 p2 ,
134
DELAY DIFFERENTIAL EQUATIONS
where p1 , p2 ∈ C[λ, µ] are relatively prime (i.e., 1 is a greatest common divisor) and
monic (i.e., the coeﬃcient of the largest power of λ is 1). Set ni = degree pi and
assume ni > 0, i = 1, 2. Then there exists a unimodular matrix S ∈ Cn×n [µ] (i.e.,
det S(µ) ≡ c ∈ C \ {0} or, equivalently, also S −1 (µ) ∈ Cn×n [µ]) such that
˜ = S −1 AS = A1 A1,2 ∈ Cn×n [µ],
A
0
A2
χAi = pi ,
i = 1, 2.
In general we cannot ﬁnd S such that A1,2 = 0.
Proof. We set C = p1 (A) and choose a basis e˜1 , . . . , e˜r of ker C (observe that
ker C is a free module), where 0 < r < n. Choose e˜r+1 , . . . e˜n ∈ Cn [µ] such that
C e˜r+1 , . . . , C e˜n is a basis for range C. Then according to Proposition D.3 we have
Cn [µ] = ker C ⊕ span e˜r+1 , . . . , e˜n .
Note that as in vector space theory all bases of Cn [µ] have the same number of
elements.
Let x ∈ ker C, i.e., we have Cx = 0. Since A and C commute, this implies CAx =
ACx = 0, i.e., Ax ∈ ker C. Consequently ker C is Ainvariant. Let S ∈ Cn×n [µ]
denote the matrix corresponding to the change of bases e1 , . . . , en → e˜1 , . . . , e˜n and
˜ the representation of the endomorphism x → Ax with respect to the
denote by A
basis e˜1 , . . . , e˜n . Then we have A˜ = S −1 AS. In view of the invariance of ker C with
˜ must have the form given in the theorem.
respect to this endomorphism we see that A
It remains to prove χA1 = p1 . Then χA = χA1 χA2 automatically implies χA2 = p2 .
Let C(µ) denote the ﬁeld of rational functions of µ with coeﬃcients in C and Cn (µ)
the vector space of all nvectors with elements in C(µ). Then we can view A and C
also as matrices in Cn×n (µ). Let x ∈ Cn (µ) with Cx = 0 be given. We can write x as
1
x
˜, s ∈ C(µ), x
˜ ∈ Cn [µ],
s
where s is the least common multiple of the denominators of the coordinates of x.
Then Cx = 0 also implies C x
˜ = 0, i.e.,
x=
x
˜=
r
αi e˜i ,
αi ∈ C[µ], i = 1, . . . , r.
i=1
This implies x = ri=1 αi /s e˜i . We conclude that e˜1 , . . . , e˜r is also a basis for ker C
in C(µ). But then vector space theory implies χA1 = p1 .
That in general we have A1,2 = 0 is demonstrated by the following example. Let
µ 1
A=
0 0
with χA (λ, µ) = (λ − µ)λ, p1 (λ, µ) = λ − µ, p2 (λ, µ) = λ. From
A1 A1,2
˜
A=
0
A2
and χA1 = λ − µ, χA2 = λ we conclude that A1 = (µ) and A2 = (0). Assume that
we can choose a unimodular S ∈ C2 [µ] such that A1,2 = (0) and set
t1,1 t1,2
s1,1 s1,2
−1
S=
, S =
.
s2,1 s2,2
t2,1 t2,2
Results on Polynomial Rings and Modules over Polynomial Rings
Then a simple calculation shows that
µ
µ 0
= S −1
0
0 0
135
1
S
0
is equivalent to
t1,1 s1,2 µ + s2,2 = 0,
t2,1 s1,2 µ + s2,2 = 0,
t2,1 s1,1 µ + s2,1 = 0,
t1,1 s1,1 µ + s2,1 = µ.
(D.1)
(D.2)
(D.3)
(D.4)
Case 1: t1,1 = 0.
Equation (D.4) implies µ = 0, a contradiction.
Case 2: s1,2 µ + s2,2 = 0.
Then equations (D.1) and (D.2) are satisﬁed. Assume that s1,1 µ + s2,1 = 0. Then
(D.4) implies µ = 0, which is impossible. Thus we have s1,1 µ + s2,1 = 0. From (D.3)
we then get t2,1 = 0, i.e., we have
t1,1 t1,2
S −1 =
.
0
t2,2
Consequently we have s2,1 = 0. But this and s1,2 µ + s2,2 = 0 imply s2,2 = 0, so that
det S(µ) ≡ 0.
Thus we have shown that neither case is possible, i.e., we never can get A1,2 = 0.
Analogously we can see that A cannot be transformed to diag(0, µ).
The result of Theorem D.5 does not carry over to Cn [µ1 , . . . , µm ] in case m ≥ 2.
The following example was given in [21]. Let
2
µ
µν
A(µ, ν) =
∈ C2×2 [µ, ν]
µν ν 2
2
be given. Then we have χA (λ, µ, ν) =λ λ − µ2 −
ν =: p1 (λ, µ, ν)p2 (λ, µ, ν). We
have p1 (A) = A and ker p1 (A)= span col(ν, −µ) . Assume that there exists a free
submodule X = span col(α, β) such that
C2 [µ.ν] = ker p1 (A) ⊕ X.
This implies that the vectors col(1, 0) and col(0, 1) (which constitute a basis for
C2 [µ.ν]) are linear combinations of col(ν, −µ) and col(α, β), i.e., there exist polynomials pi , qi ∈ C[µ, ν], i = 1, 2, such that
1 = νp1 + αp2 ,
(D.5)
0 = −µp1 + βp2 ,
(D.6)
0 = νq1 + αq2 ,
(D.7)
1 = −µq1 + βq2 .
(D.8)
From (D.6) we see that µ divides βp2 .
136
DELAY DIFFERENTIAL EQUATIONS
Case 1: µ  p2 .
p2 , p˜2 ∈ C[µ, ν], and in view of (D.6) µ(−p1 + β p˜2 ) = 0. ConseThis implies p2 = µ˜
quently we have p1 = β p˜2 . Then (D.5) implies
1 = p˜2 (νβ + µα),
which is impossible.
Case 2: µ  β.
˜ β˜ ∈ C[µ, ν], and get from (D.8)
In this case we have β = µβ,
˜ 2 ),
1 = µ(−q1 + βq
which again is impossible.
Thus we have shown that ker p1 (A) is not a direct summand in C2 [µ, ν]. For p2
we have
2
−ν
µν
p2 (A) =
µν −µ2
and ker p2 (A) = span col(µ, ν) . Analogously as above we prove that also ker p2 (A)
2
is not a direct summand of C [µ, ν].
H. Appl. T. K. Mannheim 1966. Basel 1950/55/56. Cooke. Doetsch. Modules. Functional Diﬀerential Equations. Academic Press. K. An unsymmetric Fubini theorem. [12] J.. J. I – III. J. Livre II: Alg`ebre. [8] F. P. ´ ements de Math´ematique. Berlin 1993. Deutscher Verlag der Wissenschaften. Bourbaki. Soc. Matrizenrechnung. Livre II: Alg`ebre. [11] Hale. Cameron and W. and M. Introduction to Functional Diﬀerential Equations. [5] R. Bull. New York 1954. El´ Chapitre II: Alg`ebre Lin´eaire. Hermann. [13] D. Vol. Groups. 177191. New York 1963. Buchsbaum.P. Anal. Paris 1970. I. Auslander and D. M. Bibliographisches Institut [10] J. LaSalle. Bourbaki. [4] N. Academic Press. Verduyn Lunel. Small solutions of linear autonomous functional diﬀerential equations. New York 1971. New York 1974.. Birkh¨ auser Verlag. L. ´ ements de Math´ematique. Vols. [6] R. Math. 121–125. Math. [3] N. 137 . R.39. Berlin 1965. Hale and S. Henry. Boas. Paris 1964. Jr.References [1] M. Springer Verlag. Matrizenrechnung. Harper and Row. Martin. Gr¨ obner. 47 (1941). Amer.K. Rings. Bellman and K. [7] G. Hale. Springer Verlag. J. 8 (1970). DiﬀerentialDiﬀerence Equations. J. Mannheim. Handbuch der LaplaceTransformation. El´ Chapitre VII: Modules sur les Anneaux Principaux. Gantmacher. Hermann. Diﬀerential Eqs. [2] R. 494 – 501. Entire Function. Slemrod (1973) Theory of a general classof dissipative processes. [9] W.
[16] N. Institute f¨ ur Mathematik. 1953.138 DELAY DIFFERENTIAL EQUATIONS [14] K. PhDThesis. Kappel. Providence. Jacobson. Levinger. Linear autonomous functional diﬀerential equations in the state space C. J. [18] F. MalletParet. J. Vol. Landsberg. W. 134 – 147. Wimmer. Stromberg. Algebra I. de Gruyter. Diﬀerential Eqs. Springer Verlag. Centrum voor Wiskunde en Informatica. [19] F. Remmert. R. [22] V. Pazy. Theorie der algebraischen Funktionen einer Variablen. Lineare Algebra I. Lectures in Abstract Algebra. [27] E. New York 1974.Roger D. Princeton. 1902. J. K. Kappel. Perron. An elementary divisor theory for autonomous linear functional diﬀerential equations. . Diﬀerential Eqs.. Paraskevopoulos. On degeneracy of functionaldiﬀerential equations. Amsterdam 1988. 21 (1976). Kappel and H. 612 – 619. in ”Proceedings of the Liverpool Singularities Symposium II”. New York 1971. 1983. 99 – 126. N. Exponential type calculus for linear delay equations. Academic Press. January 1984. LCDS 9316 December (1993). 24 (1977). [28] A. 250 – 267. SpringerVerlag. Math. Lunel. 1964. Oeljeklaus and R. J. [23] B. van Nostrand. Universit¨ at und Technische Universit¨ at Graz. (1969). [15] E. theory and applications.Lakshmikantham and S. Hewitt and K. Kappel. Khabbaz and G. V. Amer. Levin. Soc. [20] F. J. Diﬀerential Eqs. M. Distribution of Zeros of Entire Functions. Leipzig. [29] O. Normal forms for analytic matrix valued functions. [24] B. Springer Verlag. [17] F. 4 (1968). New York. volume 1. [26] J. Berlin 1932. Hensel and G.Leela: Diﬀerential and integral inequalities. 34 – 1984. Diﬀerential Eqs. 22 (1976). Nussbaum and P. Semigroups of Linear Operators and Applications to Partial Diﬀerential Equations.. I. Real and Abstract Analysis. II: Linear Algebra. ”Periodic Solutions for Functional Diﬀerential Equations with Multiple StateDependent Time Lags. J.. Degenerate diﬀerencediﬀerential equations: algebraic theory. [21] S. Stengle. A folk theorem in functional diﬀerential equations. Technical Report No. Berlin 1969. Springer Verlag. [25] S.
[32] A. Berlin 1974. . Functional Analysis. Yosida. Springer Verlag. Control 5 (1967). M. On the controllability of delaydiﬀerential systems. On the pointwise completeness of systems with delay. [31] K. Weiss. Zverkin. 575 – 587. 430 – 436. SIAM J.REFERENCES 139 [30] L. Diﬀerencial’nye Uravnenija 9 (1973).
Delay Differential Equations and Applications. the solution reads t TL (t − s)X0 f (s)ds xt = (1. Ezzinbi University Cadi Ayyad . ezzinbi@ucam. Morocco. for θ < 0. Hbid University Cadi Ayyad . In the literature. Rn ). Morocco. Arino et al. Introduction The variation of constants formula for delay diﬀerential equations has for some time been a puzzling part of the theory. X0 (0) = IdRn and TL (t) is the semigroup associated with the homogeneous equation on the space C = C ([−r. © 2006 Springer. Marrakech.2) 0 in which X0 is the matrixvalued function deﬁned by X0 (θ) = 0.ma K.ma 1.L. hbid@ucam.Chapter 4 VARIATION OF CONSTANT FORMULA FOR DELAY DIFFERENTIAL EQUATIONS M. 143–159. Formula (1. . 0] .). Marrakech. (eds.1) (1.ac.2) indicates that TL (t) is evaluated at X0 although this function (only considering the column vectors) is not continuous and so is not in the space where the semigroup is deﬁned. For the general inhomogeneous linear equation dx = Lxt + f (t) dt with initial value x0 = 0. this odd fact is discretely overlooked by 143 O.ac.
it will be discussed and used in diﬀerent context by M. The inconvenience caused by this apparent inconsistency is particularly visible when dealing with semilinear equations and stability or bifurcation issues.Ezzinbi in the third part of this book . M. Arino proposed a solution close to the one proposed in .Heijmans and H. ﬁrst. One should note that a rigorous variation of constants formula exists: it is enough to use the fundamental solution U (t) instead of TL (t)X0 .A. In the next section. An important consequence of both of these theories is to give new and interesting perspectives of DDE.Thieme (1985).J.M. H. the approach of embedding the problem set in a larger space containing not only continuous functions but also such functions as X0 . and a number of important results are to be found in their book.use extrapolation theory to discuss the same problem. which makes it diﬃcult to use in the study of dynamical features of the solutions.The last section of this chapter will be devoted to show how the theory of integrated semigroups can be used to provide a rigorous variation of constants formula. Ph. that is. that is. O.144 DELAY DIFFERENTIAL EQUATIONS using the expression ”formal”. a space intermediary between the space C and its bidual.Phillips(1957). in a completely diﬀerent context.Hille and R.Gyllenberg. Another approach burgeoned in the mideighties (198485) within a group of ﬁve people. that is. We will come to this theory last part of this chapter. not published). It seems that it had been reviviﬁed by H.Adimy and K. the theory of integrated semigroups. a thorough study of the relationships between the dual and the adjoint semigroups has been performed by E. to work in the space C 1 instead of C. Recently. Maniar.Clement. We will now give a very short summary of dual semigroups and we will show how it can be used to provide a rigorous variation of constants formula in the case of delay diﬀerential equations.S. the solution x reads x(t) = t U (t − s)f (s)ds 0 The shortcoming of the above formula is that it is not expressed at the level of the phase space of the equation. In terms of U (t).R. .Amann in a completely diﬀerent context. This approach was not new. A ﬁrst step in the direction of resolving this inconsistency was made by Chow and MalletParet In a later work (1980. These approaches are in fact very close to one which emerged a few years later.Diekmann. Rhandi et al .
we deﬁne 0 1 (2. In the general case ∗ T (t) is not strongly continuous. It is strongly continuous when X is reﬂexive. T ∗ (t) is a semigroup on the dual space X ∗ of X. 145 Variation Of Constant Formula Using SunStar Machinery 2.1 Duality and semigroups Let A be the inﬁnitesimal generator of a C0 semigroup (T (t))t≥0 in a Banach space X.Variation of constant formula for DDE 2. Following Hille and Phillips (1957)[21].3) X .
= x∗ ∈ X ∗ : lim T ∗ (t)x∗ − x∗ = 0 t→0 X .
It is a closed subspace of X ∗ . (pronunciated as X − sun) is the space of strong continuity of T ∗ (t). One can determine X .
in terms of A∗ . In fact.4) X . we have ([21]): (2.
this result shows that X . = D(A∗ ) (this is the closure with respect to the strong topology of X ∗ .
The restriction of T ∗ (t) to X . ⊇ D(A∗ )).
denoted T . .
is a strongly continuous semigroup of linear bounded operators on the space X . (t).
. The inﬁnitesimal generator of T .
(t) is the operator denoted A.
. to the domain ' & D(A. which is the restriction of A∗ .
) = x∗ ∈ D(A∗ ) : A∗ x∗ ∈ X .
D(A.
One can repeat the construction starting from X . ) is weak* dense in X ∗ .
and T .
that is. (t). we can the space X .
∗ from thee injection of X .
we have X . in X ∗ .
with continuous injection.∗ contains X ∗∗ . Since X ∗∗ contains X . X can be considedred as subspace of X .
and.∗ . in addition X is a closed subspace of X .
By the same operations as above we deﬁne X .∗ .
.
0 1 2 2 X .
.
= x.
∗ ∈ X .
∗ : lim 2T .
∗ (t)x.
∗ − x.
5) t→0 and T .∗ 2 = 0 (2.
.
(t) is the restriction of T .
∗ (t) to X .
.
Deﬁnition: X is called sunreﬂexive with respect to A if and only if X = X . .
.
. remark 1 This equality should be understood as identiﬁcation : we notice earlier that X can be identiﬁed to a closed subspace of X .
∗ . In the case of sunreﬂexivity this space X .
.
we have: A. In the sunreﬂexive case. .
.
= A and T .
.
. (t) = T (t).
an operator A0 which is the inﬁnitesimal generator of a C0 −semigroup T0 (t) We assume that X is sun reﬂexive with respect to A0 .Let be B ∈ £(X .146 DELAY DIFFERENTIAL EQUATIONS 2.1 The variation of constant formula:. Given a Banach space X .1.
∗ ).6) B is a perturbation of A0 . At ﬁrst sight this pertubation doesnt make sense since B takes values in a space bigger than X. The only way to make it meaningful is to look for x(t) in a bigger space in fact in the space X . We consider the following equation dx = A0 x(t) + Bx(t). dt (2.
Assuming that x(t) lies in this space. one can write equation (2.6) in the form t T0.∗ .
7) has each initial value x0 one and only one x(t) deﬁned for all t ≥ 0. Using a method of successive approximations. one can show that the equation (2. We denote T (t)x0 = x(t). T (t) is a semigroup strongly continuous in X .7) x(t) = T0 (t)x0 + 0 where x(0) = x0 .∗ (t − s)Bx(s)ds (2.
The proof goes through the next lemma: Lemma 18 Let f : [0. We will now show that T (t) is a C0 − semigroup on X. +∞[ → X .∗ .
Then. t t→ T0.∗ be normcontinuous.
∗ (t − s)f (s)ds 0 is a normcontinuous X .
.
an operator A0 which is the inﬁnitesimal generator of a C0 −semigroup T0 (t) We assume that X is sun reﬂexive with respect to A0 .  valued function Theorem 1 Given a Banach space X .Let be B ∈ L(X .
we have ' & . let A be the inﬁnitesimal generator of T (t). for each x0 in X. +∞[ with values in X. Then. a unique solution deﬁned on [0. Then.7 has. Theorem 2 Under the same assumptions as the above theorem. continuous and such that x(0) = x0 . equation2. The map T (t)x0 = x(t) is continuous on X.∗ ).
∗ (2.8) D(A) = x ∈ D(A.
∗ 0 ) : A0 x + Bx ∈ X Ax = A.
9) .∗ 0 x + Bx (2.
Finally. D(A) = D(A.147 Variation of constant formula for DDE X is sunreﬂexive with respect to A.
∗ 0 ) and + B. A = A.
In the sequel we present some properties of the unperturbed semigroup {T0 (t). 2. dϕ . ϕ(t + θ) if t + θ ≤ 0. 0] . IR).2. dt x(θ) = ϕ(θ). A∗0 (f ) = g (= df ). The semigroup {T0 (t). ϕ >= df (τ )ϕ(−τ ).2 Application to delay diﬀerential equations 2. which is given by 3 (T0 (t)) ϕ(θ) = ϕ(0) if t + θ ≥ 0. 0 Most of the following results are borrowed from Diekmann and van Gils (1990). Its generator A∗0 veriﬁes. IR) . g ∈ N BV and g(1) = 0 . t ≥ 0} related to the trivial equation: dx = 0. . ). We have just adapted them to the speciﬁc equation considered here. +∞[. t ≥ 0} is generated by A0 ϕ = ϕ with & ' . Lemma 19 [33] The semigroup T0∗ (t) is given by the formula (T0∗ (t)f )(τ ) = f (t + τ ) for τ > 0. D(A0 ) = ϕ ∈ C 1 ([−1. which will be considered next. D(A∗0 ) 3 = f : f (t) = f (0+) + t 4 g(τ )dτ for t > 0. with the pairing given by +∞ < f. ϕ(0) = 0 (ϕ = dt Let IE ∗ be represented by BV ([0. 0 and for f ∈ D(A∗0 ). dt .1 The trivial equation:.∗ 0 We have all the ingredients necessary for the application to the delay diﬀerential equations.
148 DELAY DIFFERENTIAL EQUATIONS Lemma 20 (Diekmann et al. g ∈ L1 (R+ ) and g(σ) = 0 for σ ≥ 1 0 IE .) IE = 3 t f : f (t) = f (0+ ) + 4 g(τ )dτ for t > 0.
g) = c + gL1 . IR) equipped with the norm (c. 1].)The semigroup T0. The space is isometrically isomorphic to R × L1 ([0. Lemma 21 (Diekmann et al.
(t) is given by the formula T0.
g(t + . (t)(c. g) = t c+ g(τ )dτ. 0 Its generator A.) .
0 veriﬁes D(A.
0 ) = {(c. g ∈ AC(IR+ )} . g). and A.
g). g) = (g(0). 0 (c. We represent IE . where AC(IR+ ) is the space of absolutely continuous functions on R+ .
∗ by IR × L∞ ([−1. ϕ) = sup(α. (α. IR) equipped with the norm (α. and the pairing: 0 g(τ )ϕ(τ )dτ. ϕL∞ ). 0]. g). < (c. ϕ) >= cα + −1 Lemma 22 [33] The semigroup T0.
∗ (t) is given by T0.
Its generator A. α if t + s > 0. ϕαt ) 3 where ϕαt (s) = ϕ(t + s) if t + s ≤ 0. ϕ) = (α.∗ (t) (α.
∗ 0 satisﬁes D(A.
ϕ ∈ Lip(α)} and A. ϕ) .∗ 0 ) = {(α.
∗ 0 (α. ϕ) . Taking the closure of D(A. Here Lip(α) denotes the subset of L∞ (IR+ . IR) whose elements contain as a class in L∞ a Lipschitz continuous function which assumes the value α at τ = 0. ϕ) = (0.
we lose the Lipschitz condition but continuity remains.∗ 0 ). Lemma 23 [33] IE .
.
= {(α. ϕ continuous and ϕ(0) = α} . ϕ) .
We now consider the equation dx = L(xt ) dt The equation leads to the following ”formal” equation d(xt ) = A0 xt + X0 L(xt ) dt The map X0 ∈ IE .149 Variation of constant formula for DDE 2.2 The general equation.2.
so if t → x(t) is continuous.∗ . t → X0 L(xt ) is continuous with values in IE .
∗ . We can then write the following ﬁxed point problem in t T0.
∗ (t − s)X0 L(xs )ds xt = T0 (t)x0 + 0 The solution leads to a semigroup T (t) whose generator A is deﬁned by Aϕ = A0 ϕ + X0 L(ϕ) with ' & D(A) = ϕ ∈ D(A.
for establishing a variation of constants formula for functional diﬀerential equations in ﬁnite dimensional spaces. L is a continuous linear functional from C := C ([−r. 0] . and [8] an approach has been developed.10) is : given ϕ ∈ C. Variation Of Constant Formula Using Integrated Semigroups Theory In [4].∗ 0 ) : A0 ϕ + X0 L(ϕ) ∈ IE This corresponds to ' & D(A) = ϕ ∈ C 1 . −R ≤ θ ≤ 0. 0] . ϕ (0) = L(ϕ) 3. This approach will be discussed in this section to obtain the variation of constant formula for the following delay diﬀerential equations: dx (t) = L(xt ) + f (t). 0] by xt (θ) = x(t + θ). The initial value problem associated with equation (3. the function deﬁned on [−r. based on the theory of integrated semigroups. (3. as usual. R ) . where xt denotes. +∞[ into Rn . Rn ) into Rn and f is a function from [0. [5]. . t ≥ 0.10) dt n x0 = ϕ = C ([−r.
h > 0. s ≥ 0 S(s)S(t) = 0 . A family (S(t))t≥0 of bounded linear operators S(t) on X is called an integrated semigroup if the following conditions are satisﬁed: (i) S(0) = 0.10) and the nonlinear equation dx (t) = L (xt ) + f (t. deﬁne a strongly continuous translation semigroup (T (t))t≥0 in the space C := C ([−r. c ∈ Rn and (X0 c) (θ) = X0 (θ)c} as an integrated semigroup of operators and we derive its consequences regarding the nonhomogeneous linear equation (3. In this paper. The expression T (·)X0 in the integral is not strictly deﬁned. (iii) for any t.1 Notations and basic results We start by giving some basic terminology.150 DELAY DIFFERENTIAL EQUATIONS to ﬁnd a continuous function x : [−R. xt ). since X0 is not in C and T (·) acts on C. s (S(t + τ ) − S(τ ))dτ. 0] . for t ≥ 0. (ii) for any x ∈ X. and results that will be needed in the sequel. (3. deﬁnitions. t ≥ 0. in Rn . (3. The initial value problem associated with equation (3.12) dt x0 = ϕ ∈ C.10) on [0. We know that the solutions of the linear autonomous retarded functional diﬀerential equation dx (t) = L(xt ). Deﬁnition 20 [9] Let X be a Banach space.10) can be extended to the space C ⊕ X0 . h[. 3. satisfying equation (3. The following deﬁnitions are due to Arendt. diﬀerentiable on [0. h[ .10) can be written formally as an integral equation t T (t − s)X0 f (s)ds. Rn ) . t ≥ 0. xt = T (t)ϕ + 0 where X0 denotes the function deﬁned by X0 (θ) = 0 if θ < 0 and X0 (0) = IdRn . h[ → Rn . we prove that the semigroup (T (t))t≥0 associated to equation (3. where X0 = {X0 c.11) dt x0 = ϕ. S(t)x is a continuous function of t ≥ 0 with values in X.
and there exists a unique operator A satisfying (ω. but R(λ) is injective if and only if (S(t))t≥0 is nondegenerate. If (S(t))t≥0 is an integrated semigroup exponentially bounded. for all t. τ ] . then +∞ e−λt S(t)dt exists for all λ with the Laplace transform R(λ) := λ 0 e(λ) > ω. s ∈ [0. continuous. +∞) ⊂ ρ(A) (the resolvent set of A) such that R(λ) = (λI − A)−1 . if there exist constants M ≥ 0 and ω ∈ IR such that S(t) ≤ M eωt for all t ≥ 0. and there exists a strongly continuous exponentially bounded family (S(t))t≥0 of lin +∞ −1 ear bounded operators such that S(0) = 0 and (λI−A) = λ e−λt S 0 (t)dt for all λ > ω. Deﬁnition 22 [9] An operator A is called a generator of an integrated semigroup. for all t ≥ 0. for all e(λ) > ω. if S(t)x = 0. Proposition 26 [9] Let A be the generator of an integrated semigroup (S(t))t≥0 . if there exists ω ∈ IR such that ]ω. . +∞[ ⊂ ρ(A). for all τ > 0. From [9]. 0 0 An important special case is when the integrated semigroup is locally Lipschitz. we know that every locally Lipschitz continuous integrated semigroup is exponentially bounded. Then for all x ∈ X and t ≥ 0 t t S(s)xds ∈ D(A) and S(t)x = A S(s)xds + tx. The operator A is called the generator of (S(t))t≥0 . there exists a constant k(τ ) > 0 such that S(t) − S(s) ≤ k(τ ) t − s .. Moreover (S(t))t≥0 is called nondegenerate.151 Variation of constant formula for DDE Deﬁnition 21 [9] An integrated semigroup (S(t))t≥0 is called exponentially bounded. implies that x = 0. Deﬁnition 23 [9]An integrated semigroup (S(t))t≥0 is called locally Lipschitz continuous if. In this case R(λ) satisﬁes the following expression R(λ) − R(µ) = (µ − λ)R(λ)R(µ). We have the following deﬁnition.
(iii) u(t) = x + A 0 0 From this deﬁnition.13) dt u(0) = x ∈ X. +∞[ ⊂ ρ(A) and 2 2 ' & sup (λ − ω)n 2(λI − A)−n 2 . we 1 t+h have u(t) ∈ D(A). because u(t) = lim u(s)ds and h→0 h t t+h u(s)ds ∈ D(A). locally Lipschitz continuous integrated semigroup. Then the problem ( 3. t ≥ 0. The following theorem shows that the HilleYosida condition characterizes generators of locally Lipschitz continuous integrated semigroups. +∞[ → X is an integral solution of the equation ( 3. (i) A is the generator of a nondegenerate. x ∈ D(A) is a necessary condition t for the existence of an integral solution of (3. for t ≥ 0. pour t ≥ 0. if there exists M > 0 and ω ∈ IR such that ]ω. X) t u(s)ds ∈ D(A). λ > ω ≤ M. for t ≥ 0. we say that u : [0. we give some results for the existence of particulars solutions of the following Cauchy problem du (t) = Au(t) + f (t). u(t) = S (t)x + dt 0 .152 DELAY DIFFERENTIAL EQUATIONS Deﬁnition 24 [9] We say that a linear operator A satisﬁes the HilleYosida (HY ) condition. (ii)A satisﬁes the (HY ) condition . +∞[ . In particular. n ∈ N. (3. +∞.13). we deduce that for an integral solution u. In the sequel.13) if the following assertions are true (i) u ∈ C( [0. x ∈ D(A) and f : [0. Theorem 3 [9] The following assertions are equivalent. Deﬁnition 25 [32] Given f ∈ L1loc (0. +∞[ → X is a continuous function. where A satisﬁes the (HY ) condition.13) has a unique integral solution which is given by d t S(t − s)f (s)ds. (ii) 0 t t u(s)ds + f (s)ds. X) and x ∈ X. for all t > 0. Theorem 4 [22]Suppose that A satisﬁes the (HY ) condition.
c : [−r. θ ∈ [−r. . θ ∈ [−r. +∞[ → Rn by 3 ϕ(t) if t ∈ [−r. Rn ). ∆(λ)−1 exists in L(Rn ) . 0] x(t) = (T (t)ϕ) (0) if t > 0. deﬁne x : [−r. for t ≥ 0. We know. 0] . ϕ ∈ C. where L(Rn ) is the space of linear bounded operators on Rn and ∆(λ) = λI − Lλ . 0] . 0] → Rn is deﬁned by eλ· c (θ) = eλθ c. we deﬁne the bounded linear operator Lλ : Rn → Rn by Lλ (c) = L(eλ.Variation of constant formula for DDE 153 where S(t) is the integrated semigroup generated by A. Moreover u satisﬁes t ωt e−ωs f (s) ds). Then x is the unique solution of ( 3.11) and T (t)ϕ = xt . We now present our ﬁrst main result. For each complex number λ. from [48] that the resolvent set ρ(A) of A is given by ' & ρ(A) = λ ∈ C. We consider the linear functional diﬀerential equation (3. ϕ (0) = L(ϕ) is the inﬁnitesimal generator of a strongly continuous semigroup (T (t))t≥0 on C satisfying the translation property 3 ϕ(t + θ) if t + θ ≤ 0 (T (t)ϕ) (θ) = (T (t + θ)ϕ) (0) if t + θ > 0. 0] . c).2 The variation of constant formula Let L be a linear bounded operator from C into Rn . for each ϕ ∈ C. u(t) ≤ M e (x + 0 3.11). for t ≥ 0. for c ∈ Rn . Furthermore. t ≥ 0. where eλ. Theorem 5 [48]The operator Aϕ = ϕ with domain ' & D(A) = ϕ ∈ C 1 ([−r. The following result is known.
with a ∈ Rn .λ>ω where R(λ. This equation splits into two 3 (λI − A)ψ = ϕ ∆(λ)a = c. a. such that: sup ρ(A) ⊃ (ω. where ψ ∈ D(A) and a ∈Rn .154 DELAY DIFFERENTIAL EQUATIONS 5 of the operator A to the space Theorem 6 The continuous extension A C ⊕ Xo . (i) D(A) 5 and for n ≥ 1 (ii) (ω. Rn ) . In particular. and the decomposition ψ5 = ψ + eλ. 5 such that 5 = ϕ + X0 c. It suﬃces to show that A By the HilleYosida theorem. Lemma 24 For λ >6 ω. 5 satisﬁes the (HY ) condition . for λ > ω. +∞) {(λ − ω)n R(λ. there exists ω ∈ lR. is unique. (λI − A) 5 5 ψ can be written as ψ5 = ψ + eλ· a. Let ψ5 ∈ D(A) deduce that ψ ∈ Ker(l) = D(A). c) ∈ C × Rn . R(λ. A)n } < +∞ n∈N. and Aϕ D(A) generates a locally Lipschitz continuous integrated semigroup (S(t))t≥0 t on C ⊕ X0 such that S(t)(C ⊕ X0 ) ⊂ C and S(t)ϕ = T (s)ϕds. If we put ψ = ψ5 − eλ· ∆(λ)−1 l(ψ). 5 then we 5 and λ > ω. for 0 ϕ ∈ C. . 0] . We look for ψ5 ∈ D(A). We have 5 (λI − A)(ψ + eλ· a) = ϕ + X0 c. Let ϕ 5 ∈ C ⊕ X0 . deﬁned by 5 = ϕ + X0 (L(ϕ) − ϕ (0)). 5 = C 1 ([−r. ϕ 5 ψ5 = ϕ. A) = (λI − A)−1 . c ∈ Rn . Proof of lemma. where eλ· = eλ· c. We consider the following operator 5 → Rn l : D(A) ψ → l(ψ) = ψ (0) − L(ψ). 7 6 7 & ' 5 = D(A) ⊕ eλ· . +∞) ⊂ ρ(A) 5 n (ϕ + X0 c) = R(λ. We need the following lemma. A)n−1 eλ· ∆−1 (λ)c . A) for every (ϕ. A)n ϕ + R(λ. ∆(λ) is invertible on L(Rn ).
in this case. A)n ϕ + R(λ. λ Without loss of generality. consider the nonhomogeneous functional diﬀerential equation (3. ∆−1 (λ)c .λ>ω0 5 is the generator of a locally Lipschitz We conclude by theorem 3. I − Lλ λ λn λ n≥0 So. (λI − A) Repeating this procedure. we can assume that λ > 0. On the other hand. A) This completes the proof of the lemma. that A continuous integrated semigroup. λ − L Using the relation (3. (λI − A) 5 −1 (ϕ + X0 c) = (λI − A)−1 ϕ + eλ· ∆(λ)−1 c.14). we have 2 2 2 1 2 L 2 Lλ 2 ≤ < 1. 2 −1 2 2 2 1 1 2 2 . (3. A)n−1 eλ. we obtain 2 2 2 5 n2 sup 2(λ − ω0 )n R(λ. . for t ≥ 0 dt x0 = ϕ ∈ C.14) R(λ. A) 2 < ∞. Now. ω). for λ > max(L . ω).10) dx (t) = L (xt ) + f (t). we have for every n ≥ 1 5 n (ϕ + Xo c) = R(λ. 2λ 2 λ Hence the operator I − λ1 Lλ is invertible and −1 1 n 1 = L . n∈N.155 Variation of constant formula for DDE 5 −1 exists and It follows that for λ > ω. for λ > L . one has 1 ∆(λ) = λI − Lλ = λ I − Lλ . 2≤ 2 I − Lλ 2 1 − λ1 L 2 λ and 2 −1 2 2∆ (λ)2 ≤ 1 . where ω0 = max(L .
10) has a unique solution x on [−r. for t ≥ 0 dt x0 = ϕ.16) xt = T (t)ϕ + ( S(t − s)X0 f (s)ds).15). dθ 0 we deduce that t t 5 xs ds = xt − ϕ + X0 L( xs ds) − x(t) + ϕ(0) . such that x0 = ϕ. (3. x(t) = ϕ(0) + L( xs ds) + 0 It follows that 0 t t 5 f (s)ds. for t ≥ 0.15) Theorem 7 If f ∈ C([0. for t ≥ 0.10) from 0 to t. u(t) = ϕ + A( u(s)ds) + X0 0 0 We obtain that u is an integral solution of (3. So this solution can be The operator A written as t d S(t − s)X0 f (s)ds . integrating equation (3. It suﬃces to show that the function u deﬁned by u(t) = xt . for t ≥ 0. +∞[ which is given by t d (3.10). (3. Let x be the solution of (3.17) is an integral solution of equation (3.156 DELAY DIFFERENTIAL EQUATIONS and the nonhomogeneous associated Cauchy problem du 5 (t) = Au(t) + X0 f (t). 5 satisﬁes the (HY ) condition. xt ). (3. u(t) = T (t)ϕ + dt 0 Consider the nonlinear equation dx (t) = L(xt ) + f (t. then equation ( 3.15). A 0 0 On the other hand. for t ≥ 0 dt u(0) = ϕ ∈ C. Rn ). we get t t f (s)ds. By using the relation t d xs ds = xt − ϕ. +∞[ . dt 0 Proof.18) .
157
Variation of constant formula for DDE
and the initial value problem associated
du
5
(t) = Au(t)
+ X0 f (t, u(t)), for t ≥ 0
dt
u(0) = ϕ,
(3.19)
where f : [0, +∞[ × C → Rn is Lipschitz with respect to the second
variable and continuous.
We know that equation (3.18) has one and only one solution x which
is deﬁned on [−r, +∞[ by
t
t
f (s, xs )ds, for t ≥ 0.
x(t) = ϕ(0) + L( xs ds) +
0
0
Substituting f (t, xt ) for f (t) in formula 3.16, equation (3.18) can be
written as a ﬁxed point equation.
We deduce the following result.
Theorem 8 Under the above conditions, the solution x of equation
( 3.18) can be written as
t
d
S(t − s)X0 f (s, xs )ds , for t ≥ 0.
xt = T (t)ϕ +
dt
0
Appendix
In this section, we prove that the integrated semigroup (S(t))t≥0 associated to the equation (3.11) can be written as a perturbation of the
integrated semigroup (S0 (t))t≥0 associated to the trivial equation
dx
(t) = 0, for t > 0,
(3.20)
dt
x0 = ϕ ∈ C.
The generator A0 of the C0 semigroup associated to equation (3.20) is
given by
&
'
D(A0 ) = ϕ ∈ C 1 ; ϕ (0) = 0 , and A0 ϕ = ϕ .
By theorem 6, we obtain:
80 of the operator A0 deﬁned
Corollary 9 The continuous extension A
on
C ⊕ Xo by
80 ) = C 1 ([−r, 0] , Rn ) , and A
80 ϕ = ϕ − X0 ϕ (0),
D(A
158
DELAY DIFFERENTIAL EQUATIONS
generates a locally Lipschitz continuous integrated semigroup (S0 (t))t≥0
on
t
T0 (s)ϕds, for
C ⊕ X0 such that S0 (t)(C ⊕ X0 ) ⊂ C and S0 (t)ϕ =
0
ϕ ∈ C.
Theorem 10 The integrated semigroup (S(t))t≥0 associated with equation ( 3.11) is given by
t
S0 (t − s)X0 L(T (s)ϕ)ds, t ≥ 0 and ϕ ∈ C.
S(t)ϕ = S0 (t)ϕ +
0
Proof. For every function ϕ ∈ C, consider the nonhomogenuous
Cauchy problem
du
50 u(t) + h(t), for t ≥ 0
(t) = A
(3.21)
dt
u(0) = 0,
where h : [0, +∞[ → C ⊕ X0 is given by
h(t) = ϕ + X0 (L(S(t)ϕ).
50 satisﬁes the (HY ) condition. Hence, by
In view of corollary 9 , A
theorem 4 the nonhomogeneous Cauchy problem (3.21) has an integral
solution u given by
t
d
S0 (t − s)h(s)ds , for t ≥ 0.
u(t) =
dt
0
On the other hand, we have
5
S(t)ϕ = A
t
S(s)ϕds + tϕ.
0
This implies that
t
d
( S(s)ϕds) = S(t)ϕ − tϕ, for ϕ ∈ C.
dθ 0
By applying
5 =A
50 ψ + X0 (L(ψ)), for ψ ∈ C 1 ,
Aψ
we obtain
t
t
50 ( S(s)ϕds) + X0
L (S(s)ϕ) ds + tϕ,
S(t)ϕ = A
0
0
159
Variation of constant formula for DDE
so,
50
S(t)ϕ = A
t
t
S(s)ϕds +
h(s)ds.
0
0
Hence, the function t → S(t)ϕ is an integral solution of (3.21). By
uniqueness, we conclude that S(t)ϕ = u(t), for all t ≥ 0.
Consequently,
t
d
S0 (t − s)X0 (L(S(s)ϕ) ds .
S(t)ϕ = S0 (t)ϕ +
dt
0
We deduce that
S(t)ϕ = S0 (t)ϕ +
d
dt
and
t
S0 (s)X0 (L(S(t − s)ϕ)ds ,
0
S(t)ϕ = So (t)ϕ +
t
S0 (t − s)X0 (L(T (s)ϕ)ds.
0
Remark: 1) All results of this section can be obtained if we consider
the following abstract functional diﬀerential equation
dx
(t) = Bx(t) + L(xt ) + f (t),
t ≥ 0,
(3.22)
dt
x0 = ϕ,
where B is generator inﬁnitesimal of C0 semigroup (T (t))t≥0 on a Banach
space E.
2) The approach developed in the ﬁrst section, based on a certain
class of weakly * continuous semigroups on a dual Banach space (see
Clement et al.) ([27]), for the study of delay diﬀerential equations in
ﬁnite dimensional spaces cannot be used in inﬁnite dimensional spaces
because we do not have necessarily C
C.
Chapter 5
INTRODUCTION TO HOPF BIFURCATION
THEORY FOR DELAY DIFFERENTIAL
EQUATIONS
M.L. Hbid
University Cadi Ayyad , Marrakech, Morocco.
hbid@ucam.ac.ma
1.
Introduction
1.1
Statement of the Problem:
Consider the family of equations:
G(α, x) = 0
(1.1)
where x ∈ X , α ∈ Λ, G : Λ × X → X ; X is a Banach space; Λ
is a topological space of parameters. Suppose that equation (1.1) has
for each α, a root translated to the origin x = 0, qualiﬁed as a trivial
solution, that we have G(α, 0) = 0 for each α ∈ Λ.
The problem here is to determine non trivial solutions of (1.1).
Deﬁnition 26 We say that (α0 , 0) is a bifurcating point of solutions of
( 1.1) if each neighborhood of (α0 , 0) in Λ × X contains a point (α, x)
satisfying x
= 0 and G(α, x) = 0.
Deﬁnition 27 A bifurcation branch emanating from (α0 , 0) is a map
b deﬁned from the interval I = [0, a[ ⊂ IR into Λ × X such that b :
s → b(s) = (α(s), x(s)) satisfying x(s)
= 0 for s
= 0, limx(s) = 0,
s→0
limα(s) = α0 ,and G(α(s), x(s)) = 0.
s→0
 The graph of the map b is also called the bifurcating branch (or
branch of bifurcation).
 Λ × {0} is the trivial branch.
161
O. Arino et al. (eds.), Delay Differential Equations and Applications, 161–191.
© 2006 Springer.
162
DELAY DIFFERENTIAL EQUATIONS
 A bifurcation method is any method that could be used to exhibit
a branch of bifurcation for a family of equations.
Bifurcation theory studies persistence and exchange of qualitative
properties of dynamical systems under continuous perturbations. A typical case, which we will present here, is the case of systems depending
continuously on a single parameter. We will deal with retarded diﬀerential equations of the form
dx(t)
= F (α, xt )
dt
(1.2)
with F : R × C → Rn , F of class C 2 and F (α, 0) = 0 ∀α ∈ R and C =
C([−r, 0] , Rn ) the space of continuous functions from [−r, 0] into Rn . As
usual, xt is the function deﬁned from [−r, 0] into Rn by xt (θ) = x(t + θ)
, r ≥ 0 (r could be inﬁnite ).
We will be concerned with the nature of the equilibrium point 0 (stable
or unstable) with respect to the values of the parameter α. What happens
if at such a point α0 , we observe a change in the stability (that is to say,
a transition from stability to instability when the system crosses the
value of the parameter α0 ).
Such a change may in fact correspond to important changes of the
dynamics.
Examples of systems depending on a parameter are abundant in physics,
chemistry, biology and...etc. Parameters can be the temperature, resistance, reaction rate or the mortality rate, the birth rate, etc. Under
small variations of the parameter, such systems may lose stability, more
pecisely, the ”trivial” equilibrium may lose stability, and restabilize near
another equilibrium or a closed orbit or a larger attractor. The transition can be continuous, gentle and smooth, with the new equilibrium
state emerging in the vicinity of the ”trivial” one, or it can be of a discontinuous nature, with the new equilibrium being far from the ”trivial”
one. The ﬁrst case corresponds to a local bifurcation, the second one is
a global bifurcation.
In local bifurcations, one can distinguish two types of bifurcations:
I) the system leaves its equilibrium state and reaches a new ﬁxed
equilibrium state.
II) the system goes from an equilibrium state to an invariant subset made generally of several equilibriums and curves connecting them,
closed orbits, or tori, etc. The most elementary situation is the Hopf
bifurcation, characterized by the onset of a closed orbit, starting near
the trivial equilibrium, which is the phase portrait of a periodic solution
with period close to some ﬁxed number (Hopf 1942).
In this course, we will restrict our attention to Hopf bifurcation.
Introduction to Hopf Bifurcation theory DDE
163
Deﬁnition 28 (α0 , 0) ∈ R × C is said to be a Hopf bifurcation point
for equation ( 1.2) if every neighbourhood of this point in R × C contains
a point (α, ϕ), with ϕ
= 0 such that ϕ is the initial data of a periodic
solution, with period near a ﬁxed positive number, of equation ( 1.2) for
the value of the parameter α.
We will now make the following assumptions:
(H0 ) F is of class C k , for k ≥ 2, F (α, 0) = 0 for each α, and the map
(α, ϕ) → Dϕk F (α, ϕ) sends bounded sets into bounded sets.
(H1 ) The characteristic equation
det∆(α, λ) := λId − Dϕ F (α, 0) exp(λ(.)Id)
(1.3)
of the linearized equation of (1.2) around the equilibrium v = 0 :
dv(t)
= Dϕ F (α, 0)vt
dt
(1.4)
has in α = α0 ≥ 0 a simple imaginary root λ0 = λ(α0 ) = i , all the
others roots λ satisfy λ
= mλ0 for m = 0, ±2, ±3, ±4, ...
((H1 ) implies notably that the root λ0 lies on a branch of roots λ =
λ(α) of equation (1.3), of class C k−1 )
(H2 ) λ(α) being the branch of roots passing through λ0 , we have
∂
eλ(α) α=α0
= 0
dα
(1.5)
Theorem 1 Under the assumptions (H0 ), (H1 ) and (H2 ), there exist
constants R, δ > 0, and η > 0 and functions α(c), ω(c) and a periodic
function with period ω(c), u∗ (c), such that a) All of these functions are of
class C 1 with respect to c, for c ∈ [0, R[,α(0) = α0 , ω(0) = ω0 , u∗ (0) = 0
; b) u∗ (c) is a periodic solution of ( 1.2), for the parameter value equal
α(c) and period ω(c); c) For  α − α0 < δ and  ω − 2π < η, any ω−
periodic solution p, with p < R, of ( 1.2) for the parameter value α,
there exists c ∈ [0, R[ such α = α(c), ω = ω(c) and p is, up to a phase
shift, equal to u∗ (c).
1.2
History of the problem
1.2.1
The Case of ODEs:.
A bifurcation result similar to
theorem1 was ﬁrst established by Hopf (Hopf (1942), Absweigung einer
periodischen l¨osung eines diﬀerential systems. Berichen Math. Phys.
Ki. S¨
ach. Akad. Wiss Leipzig 94, 122) in the case of ODE, assuming
that the function is analytic with respect to both the state variable and
the parameter. Hopf obtained α(c), ω(c) and u∗ (c) analytic in c at 0.
164
DELAY DIFFERENTIAL EQUATIONS
In the seventies, Hsu and Kazarinoﬀ, Poore, Marsden and McCracken
(1976) and others discuss in their works the computation of important
features of the Hopf bifurcation, especially the direction of bifurcation
and dynamical aspects (stability, attractiveness, etc), both from theoretical and numerical standpoints. A very important new achievement
was the proof by Alexander and Yorke of what is known as the global
Hopf bifurcation theorem, which roughly speaking describes the extent
of the branch. The theory was also extended towards allowing further
degeneracies (more than two eigenvalues crossing theimaginary axis, or
multiplicity higher than one, etc) leading notably to the development
of the generalized Hopf bifurcation theory (Bernfeld et al (1980, 1982),
Negrini and Salvadori (1978), M.L. Hbid (Th`ese de Doctorat 1987)).
After the work done in the seventies, one might consider that the
Hopf bifurcation theory for ODEs is essentially closed (proof of theorem
of existence of periodic solutions, computation of direction of bifurcation,
estimation of elements of bifurcation, stability of the bifurcating branch).
1.2.2
The case of Delay Equations:.
 Chafee in 1971 was
the ﬁrst to prove the local Hopf bifurcation theorem for DDE. Chafee’s
theorem gives both existence and stability: for this, it is necessary to
make further assumptions in addition to the extension of the conditions
stated in the Hopf bifurcation theorem for ODEs. The ﬁrst proof of a
stricto sensu extension of the Hopf bifurcation theorem to DDE is attributed, in historical notes of Hale’s book, to Chow and MalletParet
(1974). It is probably right to say that the machinery, at the functional
analytic level, that was necessary for such an extension, was there, ready
to be used, by the beginning of the seventies, and was indeed used independently by several groups, here and there. I will now quote some
of the work done starting in the midseventies. This list should not be
considered exhaustive in whatever way:
 Hale in his book (1977) gives a proof of the theorem (1), based on
some of his earlier work in the case of ODE.
 Chow and MalletParet (1978) use the averaging method to determine the stability and the amplitude of the bifurcating periodic orbit.
 Arino (1980 ) (not published) discusses one of the issues entailed
by DDEs, namely a lack of regularity when the parameter is the delay.
This problem shows up notably in the ﬁxed point formulation for the
determination of periodic solutions. The method elaborated in Arino
(1980) goes through the adaptation of the implicit function theorem.
 Schumacher(1982) proves the Hopf bifurcation theorem for a strongly
continuous nonlinear semigroup deﬁned in a closed subset of a Banach
space.
Introduction to Hopf Bifurcation theory DDE
165
 Stech (1985) uses the LyapunovSchmidt reduction method and generalizes a proof given by De Oliveira and Hale (1980) in the case of
ODEs to inﬁnite delay diﬀerential equations. He also gives a computational scheme of bifurcation elements via an asymptotic expansion of
the bifurcation function.
 Staﬀans (1987) shows the theorem in a case analogous to Stech’s in
the case of neutral functional diﬀerential equations. His method is based
on the Lyapunov Schmidt reduction method.
 Diekmann and van Gils (1990) reformulate the problem as an integral
equation. Using the theory of perturbation by duality, these authors
prove a centre manifold theorem for DDE and deduce the existence of
bifurcating periodic solutions via a reduction on centre manifold.
 Adimy (1991) proves the Hopf bifurcation theorem using the integrated semigroup theory.
 Talibi (1992) uses an approximation method to obtain the existence
of periodic solutions.
 M.L. Hbid (1993) uses a reduction principle to a centre manifold,
the Lyapunov direct method and the Poincar´e procedure to prove the
theorem.
 T.Faria and L.Magalhaes (1995) extend the normal form theory to
DDE and apply the theory to Hopf bifurcations of such equations.
In a more recent past, the Hopf bifurcation theory has been considered
in the case of functional partial diﬀerential equations, delay diﬀerential
equations in inﬁnite dimensional spaces. This subject will be presented
in T.Faria’s lectures. The theory has also been extended to the case of
diﬀerential equations with statedependent delay.
The following statement of the local Hopf bifurcation theorem is given
by J.Hale
Theorem 2 (J.K. Hale, Theory of functional diﬀerential equations, 1977)
Suppose F (α, φ) has continuous ﬁrst and second derivatives with respect
to α, φ, F (α, 0) = 0 for all α, and Hypothesis (H1 ) and (H2 ) are satisﬁed. Then there are constants a0 > 0, α0 > 0, δ0 > 0, functions
α(a) ∈ R, ω(a) ∈ R, and an ω(a)− periodic function x∗ (a), with all
functions being continuously diﬀerentiable in a for a < a0 , such that
x∗ (a) is a solution of equation ( 1.2) with
x∗0 (a)Nα = Φα(a) y ∗ (a),
x∗0 (a)Qα = z0∗ (a)
(1.6)
where y ∗ (a) = (a, 0)T + o(a), z0∗ (a) = o(a) as a → 0. Furthermore,
for α near 0 and ω − 2π < δ0 , every ω− periodic solution of equation
( 1.2) with xt  < δ0 must be of the type exact for a translation in phase.
166
DELAY DIFFERENTIAL EQUATIONS
The approach we want to illustrate is based on two steps: ﬁrst, a
reduction to a centre manifold, where the DDE reduces to an ODE; then,
the study of the bifurcation of periodic solutions on the reduced system.
For the reduction, we have followed the approach of Diekmann and van
Gils (1991), which itself is an adaptation to DDE of a method originally
proposed by Vanderbauwhede in the context of ODE. Once reduction
has been obtained, one has the problem of investigating the dynamics
near a critical point: nonlinearities play a crucial role there, and the
diﬃculty lies in the fact that the nonlinearities are only known implicitly.
The method we present here consists in computing a Lyapunov function
using the Poincar´e procedure.
2.
The Lyapunov Direct Method And Hopf
Bifurcation: The Case Of Ode
Before discussing the behavior of the reduced ordinary diﬀerential
system (3.28), let us recall the framework of the generalized Hopf bifurcation and hasymptotic stability related to the Poincar´e procedure (see
[14],[21][20]). We consider the planar system of diﬀerential equations
dx1
dt = α(µ)x1 − β(µ)x2 + R1 (µ, x1 , x2 )
dx2
= α(µ)x2 − β(µ)x1 + R2 (µ, x1 , x2 )
dt
(2.7)
α(µ), β(µ) ∈ C k+1 (]−µ, µ[ , R and R1 , R2 ∈ C k+1 ((−µ, µ)×B 2 (a), R) with
dα(0)
= 0, R1 (µ, 0, 0) =
k integer k ≥ 3 such that α(0) = 0, β(0) = 1 and
dµ
R2 (µ, 0, 0) = 0 and Dx R1 (µ, 0, 0) = Dx R2 (µ, 0, 0) = 0
Introducing polar coordinates x1 = ρ cos θ , x2 = ρ sin θ, we have
dρ
∗
∗
dt = α(µ)ρ + R1 (µ, ρ, θ) cos θ + R2 (µ, ρ, θ) sin θ
(2.8)
ρdθ
= β(µ)ρ + R2∗ (µ, ρ, θ) cos θ − R1∗ (µ, ρ, θ) sin θ
dt
∗ (µ, ρ, θ) = R (µ, ρ cos θ, ρ sin θ)
where R1,2
1,2
Let
R∗ (µ, ρ, θ) cos θ − R1∗ (µ, ρ, θ) sin θ
W (µ, ρ, θ) = β(µ) + 2
ρ
W (µ, 0, θ) = β(µ)
(2.9)
x2 ) i = 1. c) Deﬁnition 29 [21]The function V(µ.11) exists in[0. θ) cos θ − R2∗ (µ. 2π] Since α(0) = 0. then for any µ ∈ ]−µ. Deﬁnition 30 [21] Let h be an integer.9) and (2. θ0 . ρ. θ0 . it is easily seen by (2. x2 ). a[ × [0. x2 ) + Θ(x1 . x2 ) dt .f2h (x1 .11) (µ. θ0 ) = ρ0 (2. Σ ∈ C B 2 (a).Introduction to Hopf Bifurcation theory DDE 167 For every ρ0 ∈ [0.10) When the function ρ(θ. µ[ and c ∈ [0.. This solution will be denoted by ρ(θ. θ) dθ W (µ. x2 ) + Σ(x1 ... ρ0 ). ρ(θ. x2 ) dt where fi (x1 . For µ = 0.7) passing through (ρ0 .2 (R1 and R2 are introduced in formula ( 2. a[ and µ are suﬃciently small. x2 ) = Ri (0. system ( 2. a[ . θ) sin θ dρ = = R(µ. x2 ) + . ρ.c) = ρ(θ. ρ.13) dx 2 = x1 + f22 (x1 . µ. the solution x1 = x2 = 0 of the system dx1 dt = −x2 + f12 (x1 . θ0 . x1 . θ0 ∈ R. hcompletely unstable) if i) For every Θ. R of order greater than h in (x1 . the orbit of (2. µ[ × [0. a[ the solution of (2. x2 ) (2. θ) ∈ ]−µ. µ0 . c) − c is called a displacement function of ( 2. ρ0 ) has been determined.12) is said to be hasymptotically stable (resp.7) will be obtained by integrating the following equation dθ = W (µ.7). x2 ) + . complete information of the solution of (2. 2π].7). x2 ) (2.. µ0 . µ0 .. θ0 ) will be represented by the solution ρ(θ. ρ0 ) of the problem α(µ) + R1∗ (µ. ρ. θ) ρ(µ.. ρ. h ≥ 3. θ) dt (2. µ.7) can be written in the form: dx1 dt = −x2 + f1 (x1 .12) dx 2 = x1 + f2 (x1 .f1h (x1 . The solution x1 = x2 = 0 of ( 2.10) that when a ∈ [0.
Remark 2 The Lyapunov function F is searched using the Poincar´e procedure in the form h F (x) = Fj (x).> 0).12) has the form: • F (4. Then. The following assertions are equivalent: 1) The solution x1 = x2 = 0 of ( 2. > 0). hcompletely unstable). .12) is either 3asymptotically stable or 3completely unstable. µ near 0. The Center Manifold Reduction Of DDE Let us ﬁrst very brieﬂy draw the perspective of the center manifold theorem. system ( 2. x2 ). 0)xt . h ≥ 3. . µ[ .12). . Consider a general nonlinear DDE dx = f (p.12) is hasymptotically stable (resp. ∂ h V (0. where V is the displacement function of ( 2.6) = Gh x h+1 + o( x h+1 ) with Gh < 0 (resp. the local behavior of the solutions near 0 is dependent upon the type of the linearized equation dx = Dϕ f (p.7) has exactly one periodic solution if α(µ)Gh < 0 and no periodic solution if α(µ)Gh > 0. 0) = 0 for all p in a given region. there exists a real number µ > 0 such that for µ ∈ ]−µ. 3) There exists a Lyapunov function F such that the derivative of F along the solution of ( 2. deﬁned in deﬁnition 29. h − 1} Proposition 27 [21] Let h be an integer. dt in which p represents a parameter (possibly.168 DELAY DIFFERENTIAL EQUATIONS is asymptotically stable (resp. dt . 3. 0) ∂ i V (0. Theorem 3 [21](Generalized Hopf bifurcation theorem) Assume that the solution x1 = x2 = 0 of ( 2.completely unstable) ii) Property i) not satisﬁed when h is replaced by any integer m ∈ {2. where the Fj s are homogeneous pˆ olynomials of j=1 degree j in R2 . 0) = 0 for 1 ≤ i ≤ h − 1 and < 0 2) One has ∂ci ∂ch (resp. a vector). xt ). . x = (x1 . Assuming that f (p.
that is.Introduction to Hopf Bifurcation theory DDE 169 which can itself be described in terms of the characteristic equation (providing the eigenvalues of the inﬁnitesimal generator of the linear equation) det(λI − Dϕ f (p. for example. including DDE.) ⊗ I)) = 0. well known in the case of ODE. solutions which remain close enough to the origin either approach the origin at +∞ or escape from the origin. This scenario. In the hyperbolic case. the situation that arises when for some p = πc. It is a consequence of the socalled saddlepoint theorem which can be found. We are concerned here with the non hyperbolic case. has been extended to many inﬁnitely dimensional cases. 0)(exp(λ. in Hale’s books on DDE. In particular. no periodic solution can be found in the vicinity of the origin in this situation. as long as the equation has no root with zero real part. that is to say.
each simple.15) . the characteristic equation has a pair of imaginary roots. and no other multiple of these numbers (including 0) is a root of the equation. we will consider the following example. ϕ) = −bϕ(−1) + ε(ϕ(−1))3 + o(ϕ3 ). ε. An example of such an equation is the logistic delay equation in the case of an odd nonlinearity 1 − x(t − 1) dx = x(t) dt 1 + x(t − 1) (Change the variable from x to y = ln(x)). Equation (3. the scalar delay diﬀerential equation dx = f (b. dt (3.1 The linear equation .∗ . b >> ε . xt ). dt (3. Amongst the rich variety of cases. 3. The linear equation is dx = −bx(t − 1). the characteristic equation has a root with zero real part. b and ε are two real parameters. we choose one of the two simplest ones: we assume that at p0 . To be more speciﬁc. The vector of parameters is p = (b.14) is an example of equations that has been considered by many authors. b > 0.14) where f (b. ε). ε.
1].18) . 2 Lemma 27 The characteristic equation ( 3. (3. that is λ + be−λ = 0. −1 if θ = −1 if − 1 < θ ≤ 0. σ(A) = P σ(A) and λ ∈ σ(A) if and only if λ satisﬁes λ + be−λ = 0. Lemma 26 All the roots of equation ( 3. The characteristic equation is obtained by substitution of x(t) = eλt x0 (x0 ∈ C) into the linear equation.170 DELAY DIFFERENTIAL EQUATIONS There exists a scalar function η(θ) with bounded variation such that −bϕ(−1) = In fact. the point spectrum) of A.17) −1 def for every pair ϕ ∈ E = C([−1. 2 The adjoint equation: The formal dual product 0 0 < ψ.16) Lemma 25 If we denote σ(A) (resp.2 = ±i and all other roots have negative real parts if 2 π b= . η(θ) may be chosen as 3 0 η(θ) = −b 0 [dη(θ)] ϕ(θ).16) have negative real part if π and only if 0 < b < . ψ ∈ E∗ = C([0. ϕ >= ψ(0)ϕ(0) − b def 0 ψ(s + 1)ϕ(s)ds (3. P σ(A)) the spectrum (resp.16) has two purely imagiπ nary roots λ1. R). ϕ >= ψ(0)ϕ(0) + ψ(s − θ)dη(θ)ϕ(s)ds −1 leads to θ < ψ. 0]. R). The adjoint equation is: 0 dy(s) y(s − θ)dη(θ) = by(s + 1) = dt −1 (3.
2} . The bilinear form < ψj .19) < ψj .18). 0 π1 π π π Let Λ = −i . j. We know that if x is a solution of (3.18) associated with Λ. 2 (3.15) and Ψ∗ = (ψ1 . I is the identity matrix. Φ) = I. then we have A Φ = ΦB where 0 − π2 . ϕk >= ψj (0)ϕk (0) − 2 −1 By deﬁning a new basis. 2 2 2 ψ2 (s) = 2µ0 (cos π s + π sin π s). σ) designate the element of E∗ deﬁned by y t (s) = y(s + t) f or 0 ≤ s ≤ 1. ϕ2 ) is a basis for the generalized eigenspace P = PΛ of (3. π2 1+ 2 We have the decomposition E = P ⊕ Q.Introduction to Hopf Bifurcation theory DDE 171 If y is a solution of (3. then (y t . ϕ >= 0. still denoted Ψ.18) on (−∞. ϕk > j. 2 reads π 0 ψj (θ + 1)ϕk (θ)dθ. for PΛ∗ we can make (Ψ. σ]. 2 2 2 1 . a2 ). j = 1. ψ1∗ (s) = sin s and ψ2∗ (s) = cos s are two independent solutions 2 2 of (3. xt ) is a constant on [τ. We have: Ψ = col(ψ1. ϕ) it follows that T (t)ϕ − ΦeBt a0 tends to 0 exponentially as t tends to . ψ2 ) where ψ1 (s) = 2µ0 (sin π s + π cos π s). Denote Q = QΛ = {ϕ :< ψj . σ + 1) with σ > τ .15) on τ − 1 < t < +∞ and y is a solution of (3. a0 = col(a1 . ψ2 ) is a basis for the generalized eigenspace PΛ∗ of (3. Simi2 π π larly. ϕ2 (θ) = cos θ for −1 ≤ 2 2 2 2 π θ < 0 are two independent solutions of (3. that is.15) in the case b = . a0 = (ψ. Then ϕ1 (θ) = sin θ. +i . σ + 1[ then we let y t for each t ∈ (−∞. Φ = (ϕ1 . k = 1. B= π 0 2 Therefore T (t)Φ = ΦetB . ϕP = Φa0 . 0 π π π a1 = µ0 πϕ(0) + µ0 π −1 (cos 2 s − 2 sin 2 s)ϕ(s)ds (3. Since ϕQ = ϕ − ϕP .18) on an interval ]−∞. k = 1. Hence any ϕ ∈ E can be written as ϕ = ϕP + ϕQ where ϕP = Φa0 .20) a = 2µ πϕ(0) + µ π 0 ( π cos π s + sin π s)ϕ(s)ds 2 0 0 2 2 −1 2 with µ0 = If A is the inﬁnitesimal generator of T (t).
Using the variation of constants formula.20). which leads to a decomposition of the state space E = EC ⊕ ES which corresponds to I = πC + π S . There is part of the method which is general and follows in fact the approach by Vanderbauwhede. I will ﬁrst state the basic problem.15) approaches a periodic function given by π π a1 sin t + a2 cos t 2 2 where a1 and a2 are given by (3. taking a general equation: dx = Lxt + F (xt ) dt (3.172 DELAY DIFFERENTIAL EQUATIONS +∞. DF (0) = 0. Moreover. I assume that the linear part L has only a stable and a center part.21) reads as t T . This means that any solution of (3. and part which is implicated by the delay: essentially. equation (3. we go through the derivation of the centre manifold made by Diekmann and van Gils. 3. the treatment of the variation of constant formula using the sunstar extension.2 The center manifold theorem Here.21) We assume that F is smooth. for every ϕ ∈ E. F (0) = 0.
∗ (t − s)F (xs )ds xt = T (t)ϕ + 0 which decomposes onto the subspaces Ec and Es as follows t πC (xt ) = TC (t)ϕC + πC T .
∗ (t − s)F (xs )ds 0 t πS (xt ) = TS (t)ϕS + πS T .
∗ (t − s)F (xs )ds. the second identity can be written starting from any initial point σ t πS T . 0 Looking for solutions deﬁned on the whole real axis and uniformly bounded on their domain.
∗ (t − s)F (xs )ds. πS (xt ) = TS (t − σ) (xσ )S + σ .
while the integral has a limit. the term TS (t − σ) (xσ )S approaches 0. which yields an expression where the projection ϕS has been eliminated t πS T .173 Introduction to Hopf Bifurcation theory DDE Letting σ → −∞.
∗ (t − s)F (xs )ds. πS (xt ) = −∞ This leads to the following expression for xt t .
∗ xt = TC (t)ϕC + πC T (t − s)F (xs )ds + t −∞ 0 πS T .
(3. thus it is necessary to extend it to the space EC in order to use it in the above expressions.22) In fact. The fact that it can be done in some natural way is proved in the paper by Diekmann and van Gils.∗ (t − s)F (xs )ds. the projector πC is deﬁned on E. who use the following . the above formulas are not correct.
∗ used for the extension) formula. (with the notation πC 1 .
∗ (zI − A.
∗ )−1 dz πC = 2πi Γ (Γ is a contour enclosing the imaginary roots and no other root) The following expression is established 3 0 4 1 .
∗ zθ −1 z(η−s) β+L πC (β. ) and (β. h) ∈ E. h) = e (zI − Lz ) e h(s)ds dz. Lz = L (ez.23) with the same meaning for Γ. 2πi Γ η (3.
∗ ∼ = Rn × .
Rn ). In terms of πC xt = TC (t)ϕC + 0 . 0] . (3.∗ .22) reads L∞ ([−1.
it has the same range as πC .24) makes sense.∗ πC t ∗ ∗ πC T (t − s)F (xs )ds + t −∞ ∗ ∗ I − πC T (t − s)F (xs )ds (3. 2) choosing a suitable state space in which the ﬁxed point problem deﬁned by (3. . in fact. there are still two steps: 1) localizing the equation using a truncation. In order to set up the full equation for the center manifold.24) Note that is still a projection onto a ﬁnite dimensional subspace.
∗ is an extension of πC to the space and we have Proposition 28 πC .
Φ > ϕ2 for any Φ = (β. h) ∈ E. Φ > ϕ1 + < ψ2 .∗ Φ =< ψ1 .
∗ πC ψj and ϕj (j = 1. 2) are respectively the basis of the space E∗c and Ec . .
We ﬁrst consider the case when h = 0.174 DELAY DIFFERENTIAL EQUATIONS Proof. formula (3.23) reduces to 1 1 ∆ (z) . Then.
. πc. ezθ πc (β.π i2 −i π2 π π e ∆(i ) ∆(−i ) e 2 2 + β. 0) = 2πi Γ 2πi Γ p (z) where ∆ (z) is the cofactor matrix of the matrix zI − Lz and p(z) = det(zI − Lz ).∗ zθ −1 e (zI − Lz ) β dz = βdz. we obtain . Using the residue theorem.
R ∆ i 2 π π2 range of the linear operators ∆ i and ∆ i . with π π i I − Li π2 vj = 0 and wj i I − Li π2 = 0. j = 1. 2 . 2 2 π Since i is of multiplicity 1. 0) = p (i π2 ) p (−i π2 ) π In the case we are considering here. So. respectively. 2 of Ec and E∗c may be written in the form ϕj = eiθ vj and ψj = eiθ wj . we have 2 π R ∆ i = {v1 } 2 π 0 1 R ∆ i = w1 2 π π and R ∆ i denotes the which implies w1 v1 = 0. 2 .π i2 −i π2 e e v w v w 1 1 2 2 πc.∗ (β. . we deduce that ∆ i = v1 w1 . ±i are of multiplicity 1 and the 2 basis functions ϕj and ψj j = 1. 2 π 2 Then.
0) = + β. p (i π2 ) p (−i π2 ) and Hence.∗ (β. πc.
23) yields 3 0 4 1 . 0].24) Let ϕ ∈ E = C([−1. (3. β)ϕ1 + (w2 . β)ϕ2 . Formula (3.∗ (β. 0) = (w1 . R) be such that ϕ(0) = 0. with wj = ψj (0).
πc (0. ϕ) = 2πi Γ θ .∗ zθ −1 zθ zs e (zI − Lz ) L e e ϕ(s)ds dz.
we have πc. can be represented in terms of the formal dual product using formula (3. Therefore.23). on the other hand. which.Introduction to Hopf Bifurcation theory DDE 175 We may observe that the right hand side of the above formula is equal to P0 ϕ.
R). ϕ) ∈ E.∗ (0. ϕ(0) = 0} is dense in L1 ([−1. 0]. the above formula holds for any (0. ϕ)ϕ1 + (ψ2 . ϕ) = (ψ1 . ϕ)ϕ2 . such that ϕ(0) = 0. Since the space IE 0 = {ϕ ∈ E.
Combining the above two results.∗ . h) ∈ E. we deduce that for any (β.
∗ we have : πc.
h) = (ψ1 .The proof of Proposition 33 is complete. The spectrum σ(A) is a pure point spectrum. (β. where <. h))ϕ2 . h))ϕ1 + (ψ2 . Ec is a two dimensional space on which T .17). > is the extension of the formal dual product given by formula (3.∗ (β. (β.
Moreover the decomposition is an exponential dichotomy on R. for s ≥ 0 and x ∈ E0 . that is for any β > 0.∗ (t) can be extended to a group on R. 2 2 2 2 2T (s)x. there exists a positive constant k = k(β) such that T (s)x ≤ k exp(β s) x .
∗ 2 ≤ k exp((γ + β) s) 2x.
∗ 2 . for s ≥ 0 and x.
∗ ∈ E.
where ' & γ = sup Reλ.∗ s . λ ∈ P σ(A.
∗ ) and Reλ < 0 . Observe that we have the same properties for the space R2 × E.
∗ . that is 2 R2 × E.
∗ = R2 ×Ec ⊗ E.
∗ s and R × Ec is a ﬁnite dimensional space .
For a moment.∗ invariant by T (t) T (t). let us consider the general non homogeneous equation associated with (3.35). that is: t T .
(3. E) is a Banach space when endowed with the norm f ν = sup (exp (−νt)) f (t) < ∞.25) v(t) = T (t − s)v(s) + s The Center manifold theorem. R .we write BC(R.∗ (t − s)h(s)ds. BC ν (R. IE) is the space of all continuous function f from R into E such that sup (exp (−νt)) f (t) < ∞. E). R For ν = 0. Let us recall a few deﬁnitions Deﬁnition 31 [33] BC ν (R.
R2 × E.176 DELAY DIFFERENTIAL EQUATIONS Deﬁnition 32 [33] We deﬁne K = BC ν (R.
∗ ) by: t t .
∗ .
∗ (Kh) (t) = T (t − s)πc h(s)ds + T .
∗ (t − s)πs.
πc. ν ∈ [0.∗ h(s)ds. (3.26) 0 0 for t ∈ R.26). −γ[. In formula ( 3.
∗ and πs.
respectively. .∗ are.
∗ . the projections of (R2 × E.
∗ ) on the subspaces (R2 × IEc ) and IE− Proposition 29 [33] For each ν ∈ ]0. R2 ×IE . −γ[ 1) K is a bounded linear mapping from BC ν (R.
25) in this space with vanishing IEc− component at t = 0. 2) (I − πc. Kh is the unique solution of equation ( 3.∗ ) into BC ν (R. R2 × IE).
∗ )K is a bounded linear mapping from BC ν (R. R2 × IE .
we denote by . Let ξ be a C ∞ −function deﬁned from R+ into Rsuch that 1) ξ(y) = 1 for 0 ≤ y ≤ 1 2) 0 ≤ ξ(y) ≤ 1 for 1 ≤ y ≤ 2 3) ξ(y) = 0 for y ≥ 2 For δ > 0. R2 × IE).∗ ) into BC ν (R.
∗ (I − πc.
∗ )x πc x ξ . hδ = F(x)ξ δ δ hδ is a map from R2 × IE .
R2 × IE) into BC ν (R.∗ into itself. R2 × E. We still denote by hδ the map from BC ν (R.
In other words.27) Theorem 4 [33] Assume that δ is small enough. Deﬁnition 33 [33] The centre manifold Wc is the image of the map C : R2 × E0 −→ R2 × E deﬁned by C(ϕ) = u∗ (ϕ)(0). we can deﬁne Wc as the graph of the map C : R2 × Ec −→ E. ϕ). For υ ∈ ]0. R2 ×Ec ) such that u = u∗ (ϕ) is the unique solution in BC ν (R. ϕ) = T (. R2 × IE) × (R2 × Ec ) into BC ν (R. −γ[. Then there exists a C k mapping u∗ deﬁned from R2 ×Ec into BC ν (R. R2 × E) of equation u = G(u.∗ ) deﬁned by (hδ (f ))(t) = hδ (f (t)).. R2 × E) by G(u. we deﬁne G from BC ν (R.)ϕ + Khδ (u) (3.
∗ s deﬁned by C(ϕ) = (πs.
∗ (u∗ (ϕ)))(0). .
Introduction to Hopf Bifurcation theory DDE Now let 177 y(t) = (πc.
Then y satisﬁes the equation: t T .∗ (u∗ (ϕ)))(0)(t).
∗ (t − s)πc.
y(t) = T (t)y(0) + (3.28).∗ hδ (C(y(s)))ds. we obtain dy = Ay(t) + πc.28) 0 Diﬀerentiating (3.
∗ hδ (C(y(t))). ε. dt (3.29) We write y(t) = (b. y(t) ∈ R2 × Ec . Therefore C(y(t)) ∈ R2 × Ec ⊕ E. z(t)) ∈ R2 × Ec where z(t) ∈ Ec .
ε. z(t)) ∈ R2 × E. M(b.∗ s Then we can write: C(y(t)) = (b. z(t))) where M(b. ε. ε. z(t).
consider the nonlinear delay equation: dx(t) = −bx(t − 1) + εx3 (t − 1) + o (xt )3 . 3. we are it with two trivial equations dt dt concerned in the sequel with the system of delay equations dx = Lxt + N (b.30) may be written in the form: 0 dx(t) = dη(s)x(t − s) + g(ε.30) .31) dt −1 g(ε.3 Back to the nonlinear equation: Now. dt . ε. xt ). ϕ ∈ X = E. dt db (3. ϕ) = εϕ3 (−1) + o (ϕ)3 with the initial condition: x(θ) = ϕ(θ) for −1 ≤ θ ≤ 0.30) (3. we supplement dε db = 0 and = 0. xt ).32) = 0. Precisely. (3. dt (3. Returning to the nonlinear delay equation (3. dt dε = 0.∗ s .
178 DELAY DIFFERENTIAL EQUATIONS with N (b. xt ) = −(b − π )x(t − 1) + g(ε. xt ) 2 and π Lxt = − x(t − 1).32) may be interpreted as an evolution equation of the form: d (xt ) = Axt + r∗ N (b. dt dε = 0. dt db (3. 2 Equation (3. dt where A is the inﬁnitesimal generator of the semigroup {T (t)}.33) = 0. ε. xt ). ε. solution dx = Lxt and r.
∗ = (I. of the linear equation dt We note that the map ϕ −→ r. 0).
xt ) is deﬁned from E into E. ε.∗ N (b.
If we identify ϕ ∈ E with (ϕ(0). ϕ) ∈ E.∗ .
.
dt F is a continuous function deﬁned from R3 × E into E. ε.34) = 0. then equation (3. xt ). db (3. dt dε = 0.33) may be written in the form: ! du ∗ ∗ dt = A0 + r L u + F (b.
The variation of constant formula in this case is t T .∗ and u = xt .
36) . 0) and T (b. in our case. ε. the ordinary equation (3.29) becomes.35) v(t) = T (t − s)v(s) + s with v(t) = (b. Bz(t) + β (Φz (t) + M(b. ε. z(t)) (−1))3 ψ(0) dβ dt dε dt = 0 = 0 (3. ε. ε.∗ (t − s)F(v(s))ds (3. u(t)) = (b. z(t)) (−1)) + ε (Φz (t) +M(b. Hence. u(t)). z(t)) (−1))3 + dz(t) = dt o (Φz (t) + M(b. ε. u(t)). F(v(t)) = (F (b. 0. ε. ε. T (t)ϕ).
ε. z(t)) (−1) + 2µ0 ε (−z (t) + M(b. we will be concerned with the ordinary diﬀerential equation in R2 parametrized by β and ε (β and ε are in the neighborhood of 0). z) is Note that M(b. We ﬁrst concentrate on the case β = 0 ( β is deﬁned in section 3. ε.39) where M(ε. ε. satisfying g(ε. ∂M(β. z(t)) (−1) + 2µ0 ε (−z (t) + M(b. since g is a C ∞ function with respect to ϕ. ε. z(t)))3 = 0 = 0 (3. ε. z(t)))3 + 2 o (−z1 (t) + M(b. z(t)) (−1))3 + o (−z1 (t) + M(b. z(t)) (−1) + εµ0 π (−z1 (t) + M(b. Dg(ε. z(t))(−1)]3 = π2 z1 (t) + 2εµ0 [−z1 (t) + M(ε. In this case. dt C k with k ≥ 2. ε. our ODE is reduced to d z (t) dt 1 d z (t) dt 2 = − π2 z2 (t) + εµ0 π [−z1 (t) + M(ε. ε. ε. z(t)) (−1))3 + o (−z1 (t) + M(b. z(t)))3 ( π2 + µ0 πβ)z1 (t) − 2βµ0 M(b. ε. ε. z) → 0 as z → 0. From now on.4 The reduced system We are now in a position to discuss the behavior of the reduced ordinary diﬀerential system. o( z ) that is z . z(t)))3 ( π2 + µ0 πβ)z1 (t) − 2βµ0 M(b. 0) = 0. z(t))(−1)]3 + o [−z1 (t) + M(ε. z(t))(−1) is a Since M(β. ε. z(t))(−1)]3 (3. z(t))(−1)]3 + o [−z1 (t) + M(ε.37) d M(b. z(t)(−1)). M(β. ε. ε. 2 So we obtain: d z1 (t) dt d z2 (t) dt dβ dt dε dt = = π − z2 (t) + M(b. 0) = 0. ε.179 Introduction to Hopf Bifurcation theory DDE with ψ(0) = µ0 π 2µ0 . 0) = 0 and ∂z M(β. β = 0 corresponds to b = π2 ). z(t)))3 (3. Φ= π 1+ 2 −1 0 and β = b − π . 0)z = z and z −→ M(b. The next section will be devoted to the existence of periodic solutions of the equation d z1 (t) dt d dt z2 (t) = = − π2 z2 (t) + M(b. ε. µ2 = 1 2.38) 3. 0) = 0. 0) z = z. ε. z(t)) (−1) + εµ0 π (−z1 (t) + M(b. ε. ε. ε. ε. z(t)))3 +o (−z1 (t) + M(b. z(t)(−1)) = M(0.
Proof of Proposition 4.is asymptotically stable if ε > 0 and unstable if ε < 0. Gh > 0 ). let us derive its consequences on equation (3. z2 ) = − z2 + z2 − εµ0 πz13 − 2εµ0 z13 .2. when the latter is either 3asymptotically stable or 3completely unstable.39) is the reduced equation of equation (3.40). Therefore.39) is the system d 3 π dt z1 (t) = − 2 z2 (t) − εµ0 π [z1 (t) ] (3. Proof of Corollary 5 Equation (3. it enjoys the same stability properties as equation (3. ( F3 and F4 are two homogeneous polynomials of degree 3 and 4 respectively).40) is 3asymptotically stable (resp.14). z2 ) 4 (3.40). z2 ) = Gh ( z12 + z22 )2 + o (z1 .39) is a perturbation of order o( z 3 ) of equation (3. z2 ) = − π2 z2 ∂F + π2 z2 ∂F − εµ0 πz13 ∂F − 2εµ0 z13 ∂F + − π2 z2 ∂z2 ∂z1 ∂z2 π ∂F4 ∂z1 4 4 + 2 z2 ∂z2 − εµ0 πz13 ∂F − 2εµ0 z13 ∂F − 2εµ0 πz14 − 4εµ0 z13 z2 .40) is 3asymptotically stable if ε > 0 (3completelyunstable if ε < 0). Since equation (3. z2 ) = z12 + z22 + F3 (z1 . We have π ∂F π ∂F • ∂F ∂F F (z1 .2 Proving that the origin of (3. z2 ) + F4 (z1 .14) on the centre manifold we deduce from the fundamental centre manifold theorem that the two equations have the same stability properties. ∂z1 ∂z2 ∂F4 ∂z1 .is asymptotically stable if ε > 0 and unstable if ε < 0. 2 ∂z1 2 ∂z2 ∂z1 ∂z2 that is • 3 3 3 4 F (z1 . Before we proceed to the proof of proposition 4. that is the origin of (3.14). 3completelyunstable) is equivalent to determining a Lyapunov function F of the form F (z1 .14).180 DELAY DIFFERENTIAL EQUATIONS A natural candidate to be the principal part of system (3. z2 ) such that • F (z1 . Corollary 5 The origin of equation ( 3.40) 3 π d dt z2 (t) = 2 z1 (t) − 2εµ0 [−z1 (t) ] Proposition 30 The origin z = 0 of ( 3.41) with Gh < 0 (resp.
For ε = 0. a2 = µ0 ε. F4 = a1 z14 + a2 z13 z2 + a3 z12 z22 + a4 z1 z23 + a5 z24 . This means that system (3. the origin x = 0 is 5asymptotically stable (respectively. Formula (3.41). π 2 2 We observe that G4 (ε) < 0 if ε > 0 and G4 (ε) > 0 if ε < 0. we arrive at Let T denote the operator −z π T3 F3 = 0. 3completely unstable) if a > 0 (resp. a5 a3 . a2 a3 − 2a1 a4 − a2 2a5 − a3 −a4 = π2 G4 + 4εµ0 = π4 µ0 ε 4 = 3π G4 = 0 = π2 G4 from which we deduce 3 G4 (ε) = − µ0 επ. Then. We know from [14] that Tj is a bijection from Pj to Pj when j is odd.42) yields a system of linear equations. 4 (3.Introduction to Hopf Bifurcation theory DDE 181 ∂ ∂ + z1 and Tj its restriction to ∂z1 ∂z2 the space Pj of homogeneous polynomials of degree j. (ai ∈ R. a3 ∈ R. a1 = a3 − µ0 ε. a < 0). 2 2 π T4 F4 = G4 z12 + z22 + 2µ0 επz14 + 4µ0 εz13 z2 2 (3. ∂z1 ∂z2 F4 is a homogeneous polynomial of degree 4. we can choose a3 = 0 and the Lyapunov function F is F (z1 .42) and ∂F4 ∂F4 2 − εµ0 z13 − 2εµ0 z13 = o z12 + z22 . that is F3 = 0. Therefore. equating terms with the same degree on both sides of (3. ϕ) = −bϕ(−1)+ε (ϕ(−1))3 +a (ϕ (−1))5 . z2 ) = z12 + z22 − 2 5 3 µ0 εz14 + µ0 εz13 z2 + µ0 εz1 z23 . β = 0 and a = 0. Proposition 31 Suppose f (b. . for 1 ≤ i ≤ 5). 2 2 2 2 π We know from [14] that the homogenous polynomials of even degree j are j determined up to the addition of constant terms z12 + z22 . ε.40) is 3asymptotically stable if ε > 0 and 3completely unstable if ε < 0.43) 3 5 1 2 1 a4 = µ0 ε.
Because the linear part of this system is independant of the parameter ε.39) is a perturbation of system (3. Reconsider the delay equation x (t) = f (x (t − 1)) .45). z(t))(−1)] +aµ0 π [−z1 (t) + M(β. z(t))(−1)]5 The candidate to be the principal part of (3. Assume that f satisﬁes the following assumptions 3 for some k ≥ 3 f ∈ Ck (IR) (H) : f (0) = 0. Cases Where The Approximation Of Center Manifold Is Needed Let f be a real function deﬁned on IR.44) is the system: d π 5 dt z1 (t) = − 2 z2 (t) − aµ0 πz1 (t) d π 5 dt z2 (t) = 2 z1 (t) − 2aµ0 z1 (t) (3. F has the sign of −a.44) d π z (t) = 2 z1 (t) + 2εµ0 [−z1 (t) + M(β.39) has no more than one periodic solution. (4.40). Thus. we can only concude that system (3.46) . ε. The reduced ordinary diﬀerential system associated to (3.45) is . From Proposition 30 and generalized Hopf bifurcation theorem. ε. z(t))(−1)]5 (3.2 System (3. ε. Its derivative along the solutions of (3. we show that the function given by F (z1 . 4 . near the origin. z2 ) = z12 + z22 − 4 11 10 5 aµ0 z16 − aµ0 z15 z2 + µ0 az13 z23 + aµ0 πz1 z25 3π 4 3 4 is a Lyapounov function of system (3. the condition given in terms of the sign of α (µ) Gh (see Theorem 3) is not fulﬁled. which gives the conclusion of the proposition Remark 4. 3 3 5 F (z1 .182 DELAY DIFFERENTIAL EQUATIONS Proof. Theorem 3. ε. 4.45) Using once again the Poincar´e procedure. z2 ) = − aµ0 π z12 + z22 + o z12 + z22 . z(t))(−1)]3 dt 2 +2aµ0 [−z1 (t) + M(β.14) becomes d 3 π dt z1 (t) = − 2 z2 (t) + εµ0 π [−z1 (t) + M(β.
f (0) = 0.48) . IR2 ). The presence of a quadratic term in equation (4. z (t)) (−1))3 +o (Φ (−1) z (t) + M (β.47) for the value γ = . equation (4. 4. z (t)) (−1))3 = 0.183 Introduction to Hopf Bifurcation theory DDE Equation (4. we refer to Diekmann and van Gils [33]. The determination of the Lyapunov function leads us to give an explicit approximation of the homogeneous part of degree two of the local center manifold near zero π associated to equation (4.46) may be written as x (t) = −γx (t − 1) + αx2 (t − 1) + εx3 (t − 1) + o x3 (t − 1) .47) where γ = −f (0) .47) has a local center manifold M : U ×V (β. given by dz (t) dt dβ (t) dt = Bz (t) + Ψ (0) [−β (Φ (−1) z (t) + M (β. ξ) . In the last section. 2 It is also possible to extend the procedure to compute a center manifold for a general autonomous functional diﬀerential equation. z (t)) (−1)) +α (Φ (−1) z (t) + M (β. ξ) −→ Xs −→ M (β. For the most recent and satisfactory presentation of this result. z (t)) (−1))2 +ε (Φ (−1) z (t) + M (β. also in several other works where it is assumed that the nonlinearity is odd. α = 12 f (0) and ε = 3!1 f (0) .47) is reduced to an ordinary diﬀerential one.47) creates an additional diﬃculty compared to the situation dealt with in the last section. So. where U (resp.1 Approximation of a local center manifold Let us now brieﬂy describe what the local center manifold theorem tells us. (4. On this center manifold the delay equation (4. If we use the method developed before. V ) is a neighborhood of zero in IR ( resp. (4.
b. We denote by h2 (ξ) = aξ12 + bξ1 ξ2 + cξ22 . we know that one can assume that the center manifold to be Ck . . for each ξ ∈ V. (4. 0 − Denote by h (ξ) = M (0. we know however that the Taylor expansion at any order is unique. 0] and π a (0) + a (−1) = α 2 π b (0) + b (−1) = 0 2 c (0) + π c (−1) = 0. and of course h(0) = 0. b. Under the regularity assumption made on f .49) the homogeneous part of degree two of h. The Taylor expansion of h near ξ = 0 yields h (ξ) = aξ12 + bξ1 ξ2 + cξ22 + o ξ2 in which a. b.47). Even though center manifolds are not unique. Theorem 6 Let a. be the coeﬃcients of h2 . c are elements of XS . 2 (4.50) c (θ) = − π b (θ) 2 for θ ∈ [−1. h is a function with values in Xs . c) is a solution of the following system of equations: π b (θ) + Φ (θ) Ψ (0) α a (θ) = 2 b (θ) = −πa (θ) + πc (θ) (4. b and c.51) where Φ and Ψ are as in section 2 and α is the coeﬃcient of (x(t − 1))2 in equation ( 4. Then (a. The following lemma provides us with an eﬀective way to determine the coeﬃcients a. which implies that Dh(0) = 0. ξ) . We also know that the center manifold is tangent to Xc .184 where DELAY DIFFERENTIAL EQUATIONS 0 B= π 2 π 2 . c ∈ Xs .
1] and θ ∈ [−1. denote by z (t) the solution of the ordinary diﬀerential equation dz (t) = Bz (t) + Ψ (0) g (Φ (−1) z (t) + h (z (t)) (−1)) dt (4.55) On the other hand. it follows that ∂h (ξ) (θ) = ∂θ ∂h (ξ) (θ) [Bξ + Ψ (0) g (Φ (−1) ξ + h (ξ) (−1))] ∂z + Φ (θ) Ψ (0) g (Φ (−1) ξ + h (ξ) (−1)) .56) ∂θ 2 The homogeneous part of degree two with respect to ξ of (4. the function t −→ Φz (t) + h (z (t)) is a solution of equation (4.185 Introduction to Hopf Bifurcation theory DDE Proof : For every ξ ∈ V .19).47) with γ = 2 diﬀerentiable at t = 0. the local semi ﬂow t −→ Φz (t) + h (z (t)) generated π on the center manifold is by the delay equation (4. (4.53) By diﬀerentiating equation (4. So.52) z (0) = ξ.55) and (4. is given by ∂h2 (ξ) (θ) = ∂θ ∂h2 (ξ) (θ) [Bξ] + Φ (θ) Ψ (0) g2 (Φ (−1) ξ) ∂z (4. ∂h (ξ) (t + θ) + Φ (t + θ) Bξ ∂θ = ∂h (z (t)) (θ) Φ (θ) + ∂z × [Bz (t) + Ψ (0) g (Φ (−1) z (t) + h (z (t)) (−1))] . implies that it veriﬁes the translation property. (4. letting t go to 0. from above . ∂θ 2 which implies ∂h (ξ) π (0) = − h (ξ) (−1) + g (Φ (−1) ξ + h (ξ) (−1)) . This. for each t ∈ [0. So.56)respectively. in particular.57) . −t] (ie : t + θ ≤ 0). (4. ∂ π [Φξ + h (ξ)] (0) = − [Φ (−1) ξ + h (ξ) (−1)]+g (Φ (−1) ξ + h (ξ) (−1)) . then. π π where g (x) = x + f (x) = x + −γx + αx2 + εx3 + o x3 .47).53) with respect to t and rearranging terms. if z is a solution of equation (3. 2 2 The ﬁrst property of a center manifold is that it is invariant with respect to the original semiﬂow.54) So. (4. Φ (θ) z (t) + h (z (t)) (θ) = Φ (t + θ) ξ + h (ξ) (t + θ) .
0) ds. 0. conditions (4. 2 2 (4. c) are unique as coeﬃcients of the quadratic part of the Taylor expansion of h. The following proposition provides us with uniqueness of the solution of that system of equations.51) they satisfy determines these coeﬃcients in a unique way.50) is given by 0 X (θ) = exp (Hθ) X (0) + S (θ) .59). Proof : Put π 0 a 2 and X = b .51) is unique.60) − sin (πθ) cos (πθ) sin (πθ) exp (Hθ) = 1 1 1 2 [1 − cos (πθ)] − 2 sin (πθ) 2 [1 + cos (πθ)] then. Remark We already know that (a.58) ∂θ 2 Then. b. the general solution of (4.57) and (4. 0 The boundary condition (4. 0) . −π 0 π H= π c 0 0 2 By the use of the variation of constant formula. 0 . 0 .50) and (4.50) with the boundary conditions ( 4.186 DELAY DIFFERENTIAL EQUATIONS and π ∂h2 (ξ) (0) = − h2 (ξ) (−1) + g2 (Φ (−1) ξ) .51) follow respectively from (4. (4. if we note that g2 (x) = αx2 . where S (θ) = θ exp (H (θ − s)) col (Φ (s) Ψ (0) α . if we notice that 1 1 1 2 [1 + cos (πθ)] 2 sin (πθ) 2 [1 − cos (πθ)] .58). it is not obvious from the beginning that the system of equations (4. by virtue of (4.50) and (4. Proposition 32 The solution of the ordinary diﬀerential equation ( 4. (4.59) Hence. it follows that X (0) is a solution of the linear system DX (0) = e.51) reads π π H + exp (−H) X (0)+col (Φ (s) Ψ (0) α .61) . 0)+ S (−1) = col (α. (4. However.
− . Corollary 8 If a.51) hold. . 1 2 Hence. Therefore.60) of exp (Aθ). π Since the matrix D is invertible. b. X (0) = col 5π 3 5π 3π 5π 3 So. c) are the coeﬃcients of the quadratic part of the Taylor expansion of h if and only if (a. by substituting the expressions of a (0) .50) and (4. Corollary 7 Let a. it follows that 4µ0 α 4µ0 α 4µ0 α 8µ0 α 2µ0 α 8µ0 α − .61) it follows that 4α 2αµ0 4α 8αµ0 6α 4αµ0 − .Introduction to Hopf Bifurcation theory DDE where 187 0 1 1 D = −2 −1 2 1 −1 0 and e = −S (−1) + col 2α (1 − 2µ0 ) . 5π 3π Proof : By virtue of the expression (4. b. (a. the inverse matrix of D 2 1 D−1 = 2 5 3 is given by −1 3 −1 −2 .51) . b. c ∈ Xs are a (−1) b (−1) c (−1) the coeﬃcients of h2 . − . 0 . hence. b (0) and c (0) in the equation (4. the result follows immediately .61) .51) . − . c) is the solution of the system of equations (4.50) and (4. 0 . − S (−1) = col 3 3π 3 3π 3 3π Moreover. c ∈ C. from (4. b. existence and uniqueness of X (0) are ensured from (4. existence and uniqueness of the solution of (4. then = 6α 4αµ0 − 5π 3π = − = 4α 4αµ0 + 5π 3 4α 8αµ0 − .
∞[ and 3completely unstable for α ∈ ]−α0 (ε) . π suppose that β = 0 ( i. Proof: Coming back to proposition (27)[21] it suﬃces to prove the existence of a map F satisfying 2 2 . it is enough to consider the system dz1 dt dz2 dt π = − z2 + µ0 απ (−z1 + h (z) (−1))2 − πµ0 εz13 2 = (4.63) is 3asymptotically stable . we mainly deal with 3asymptotic stability and 3complete unstability of the trivial solution z ≡ 0 of (4. then there is α0 (ε) ∈ IR+ such that the trivial solution z ≡ 0 of (4.63) is 3asymptotically stable for α ∈ ]−∞.63) π z1 + 2µ0 α (−z1 + h (z) (−1))2 − 2µ0 εz13 . To this end. γ = ). equation (4. −α0 (ε)[ ∪ ]α0 (ε) . First. α0 (ε)[ .64) .48) becomes 2 dz1 dt dz2 dt π = − z2 + µ0 απ (−z1 + h (z) (−1))2 2 +µ0 επ (−z1 + h (z) (−1))3 +o (−z1 + h (z) (−1))3 (4. it follows that i) If ε > 0 and α ∈ IR . In the sequel. (z1 . then the trivial solution z ≡ 0 of (4. z2 ) = G3 z12 + z22 + o z12 + z22 F(4. 2 Proposition 33 Under the standing hypothesis (H). Then.62) = π z1 + 2µ0 α (−z1 + h (z) (−1))2 2 +2µ0 ε (−z1 + h (z) (−1))3 +o (−z1 + h (z) (−1))3 .188 4.63) (4.e.62).2 DELAY DIFFERENTIAL EQUATIONS The reduced system We are now in position to study the reduced system (4. ii) If ε < 0 .48) .
z2 ) = π π π T2 F2 + T3 F3 + T4 F4 2 2 2 ∂F2 ∂F3 ∂F4 µ0 απ (−z1 + h (z) (−1))2 − πµ0 εz13 + + + ∂z1 ∂z1 ∂z1 + ∂F2 ∂F3 ∂F4 + + ∂z2 ∂z2 ∂z2 2µ0 α (−z1 + h (z) (−1))2 − 2µ0 εz13 . b. which The expression of F(4. where a.65). In [14]. the authors have proved that such a map may be computed using the Poincar´e procedure.63) is equivalent to the three relations : (i) T2 F2 = 0. Without loss of generality. ∂F2 ∂F2 π +2 (ii) T3 F3 + π µ0 αz12 = 0. z2 ) is desired to be of the form (4. we shall exploit relations (i). we put r = 1. F2 (z) = r z12 + z22 . Let F be a functional having the form (4.63) (z1 . where Tj is a linear operator deﬁned on the space of homogeneous polynomials of degree j by Tj P = −z2 ∂P ∂P + z1 .(ii) and (iii) to derive a Lyapunov function F : Relation (i) implies that for r ∈ IR. c ∈ Xs denote the coeﬃcients of h2 (z).64) . b (−1) and c (−1) . F (z) = 4 Fj (z) . m2 and m3 respectively a (−1) . 3π 1 3 . 2 ∂z1 ∂z2 (iii) G3 z14 + 2G3 z12 z22 + G3 z24 = π ∂F ∂F µ0 αz12 T4 F4 + π 3 + 2 3 2 ∂z1 ∂z2 ∂F ∂F −2µ0 αP (z) z1 − µ0 εz13 . Then F(4. + π 2 +2 2 ∂z1 ∂z2 where P (z) = h2 (z) (−1) . For notational convenience.65) j=1 where Fj is a homogeneous polynomial of degree j. Now.189 Introduction to Hopf Bifurcation theory DDE with G3 ∈ IR \ {0} . (4. we denote by m1 . Relation (ii) implies that 2 3 2 3 2 F3 (z) = 4µ0 α z − z1 z2 − z2 . ∂z1 ∂z2 (z1 .
. the relation (iii) is equivalent to π 2 a2 − 2π (2µ0 αm1 + εµ0 ) π (2a3 − 4a1 ) − 4µ0 α (πm2 + 2m1 ) − 4εµ0 − 8π (µ0 α)2 2 π (3a4 − 2a2 ) − 4µ0 α (2m2 + πm3 ) − 16 (µ0 α)2 2 π (4a5 − 2a3 ) − 8m3 µ0 α 2π − a4 2 which implies that = G3 = 0 = 2G3 = 0 = G3 . α) = G3 = 2 4 (4. α) ∈ S1 and G3 (ε. ii)By virtue of the previous proposition. we distinguish two main parts S1 and S2 of the parameter space IR2 ( see Fig 1 ). from the study of the sign of where λ = 5 5π G3 (ε.190 DELAY DIFFERENTIAL EQUATIONS Finally. ii) Equation (4. α) > 0. the following assertions hold i) Equation (4.67) G3 (ε. α). Hence.66) So. there exist r0 = r0 (ε. for ε ∈ IR− . −2µ20 α2 + Theorem 9 Assume that hypothesis (H) is satisﬁed. r0 ) if βG3 (ε. G3 (ε.47) has exactly one periodic solution in B (0. if we deﬁne F4 (z) = a1 z14 + a2 z13 z2 + a3 z12 z22 + a4 z1 z23 + a5 z24 . the assertions of the proposition follow immediately. Then.47) has no periodic solution in B (0. where δ = − 4λ given above. (4. r0 ) if βG3 (ε. 3π 3π − µ0 m1 − µ0 (2m2 + πm3 ) α − εµ0 . α) > 0 and β0 = β0 (ε. α) with respect to the parameters (ε.66). it follows that 3π εµ0 . for each (ε. α) = λα2 − 4 µ0 8 − 13 < 0. then. Remarks: i) The map α0 in the previous ? proposition 33 is deﬁned by 3πµ0 and λ is the constant α0 (ε) = δ ε. α) < 0. α) > 0 for (ε. α) ∈ S1 ∪ S2 . α) > 0 such that for every β ∈ ]−β0 . by substituting the values of mi given in corollary 8 in (4. α) ∈ S2 . α) < 0 for (ε. such that G3 (ε. β0 [ .
In conclusion. α. or ε < 0 and for all α. we can state as a corollary the following immediate consequences of theorem 9 regarding the direction of bifurcation.which completes the proof of the ﬁrst assertion . ε) G3 (ε. sign (Reλ (β.62) is 3asymptotically stable if G3 (ε.Introduction to Hopf Bifurcation theory DDE 191 Proof : By virtue of proposition 33.68) is equivalent to βG3 (ε. (4. The results on the direction of bifurca2 tion are as follows : the quadratic term acts in the same way as in the negative feedback case. Then. while the cubic term has an opposite eﬀect. then. α) > 0 and β0 = β0 (ε. for example. α) > 0 and β0 = β0 (ε. However. relation (4.48) has no periodic solution in B (0. by the generalized Hopf bifurcation theorem 3. r0 ) if Reλ (β. α) > 0. positive feedback). Subcriticality is caused by the cubic term and occurs only when the coeﬃcient of this term is negative and large enough in absolute value. Hence. the bifurπ cation of nontrivial periodic solutions of (4. ε) G3 (ε. r0 ) if Reλ (β. by theorem 3. 3π bifurcation occurs at γ = − . Reλ (β. α) < 0. ε) = 12 µ0 πβ. In fact. it follows that. α. Similar computations have been performed in the case γ < 0 (that is. ε) is the real part of the two eigenvalues of the linear part of (4. ε)) = sign (β) .which yields the second assertion. α. there exists r0 = r0 (ε.47) arising from γ = is 2 i) subcritical if ε < 0 and α is small enough. if β = 0. The proof of the second assertion is similar.68) where Reλ (β.However. ii) supercritical if either ε < 0 and α is larger than some value. α) ∈ S1 ). Thus. α. α) < 0 ( in other words if (ε. α) > 0 such that the system (4. α) < 0. α. the quadratic term pushes the bifurcation to the right. Corollary 10 Assume that hypothesis (H) is satisﬁed. no matter its sign is. α) > 0 such that the system (4. There is little change in the formulae.48) has exactly one periodic solution in B (0. . the system (4. there exists r0 = r0 (ε.62) .
this theory can be traced back to the work done a hundred years ago by Poincar´e [14]. Ait Babram University Cadi Ayyad . © 2006 Springer. Introduction Normal forms theory is one of the most powerful tools in the study of nonlinear dynamical systems. the principal difﬁculty in developing normal form theory is the fact that the phase space is not ﬁnite dimensional. In the context of ﬁnitedimensional ordinary diﬀerential equations (ODEs). the authors in [44] succeeded to a simpler inﬁnite dimensional ODE so deﬁned as normal form of the original FDE. (eds. 193–227. thus hopefully greatly simplifying the dynamical analysis. the authors have considered an FDE as an abstract ODE in an adequate inﬁnitedimensional phase space which was ﬁrst presented in the work of Chow and MalletParet [30]. Concerning functional diﬀerential equations (FDEs). babram@ucam. The basic idea of normal form consists of employing successive.ac. Morocco. is due to T. .ma 1. [44]. Faria et al. Marrakech. in particular. for fear of loosing the explicit relationships between the coeﬃcients in the normal form obtained and the coeﬃcients in the original FDE. The ﬁrst work in this direction. their method provides 193 O. nearidentity. In that paper. qualitatively equivalent to the original system in the vicinity of a ﬁxed equilibrium point. nonlinear transformations. Delay Differential Equations and Applications.Chapter 6 AN ALGORITHMIC SCHEME FOR APPROXIMATING CENTER MANIFOLDS AND NORMAL FORMS FOR FUNCTIONAL DIFFERENTIAL EQUATIONS M. in such a way to overcome this diﬃculty.). One of [44] s proposal is that. Arino et al. and through a recursive process of nonlinear transformations. for stability and bifurcation analysis. This inﬁnite dimensional ODE is then handled in a similar way as in the ﬁnite dimensional case. which leads us to a diﬀerential equation in a simpler form.
The problem have been considered in the previous papers [2] and [1]. and we consider some applications of the computation of center manifolds to the special cases of the Hopf singularity and . we are concerned with FDEs having a general singularity. However. We assume that the existence of a local center manifold is ensured by ﬁnitely many eigenvalues with zero real part. by means of a recursive procedure. This point of view will be presented in detail later. with the normal form of an FDE. The chapter is organized as follows: The preliminary section 2 provides us with elementary notions and a background about the theory of FDEs (for more details see [58]). we derive an algorithmic scheme which allows us.The result generalizes singularities considered in both of the works [2] and [1].194 DELAY DIFFERENTIAL EQUATIONS us an eﬃcient algorithm for approximating normal forms for an FDE directly without computing beforehand a local center manifold near the singularity. This allows. a simpler scheme that greatly saves computational times and computation memory. in most cases. In this work. This is accomplished in three steps. the only information that can be extracted at the expense of nontrivial algorithmic procedures. Finally. in the third step we present a direct computing scheme of normal forms for FDEs (in the sense of the deﬁnition given in [44]). we will show that the known of Taylor coeﬃcients of the center manifold is necessary for the algorithm to go on. to compute at each step the term of order k ≥ 2 of the Taylor expansion of a local center manifold : In fact. First. In the ﬁrst step we consider the reduced system on a known local center manifold for which an algorithmic scheme is developed. Our motivations for this work are two folds. we prove that the coefﬁcients of the homogeneous part of degree k of a local center manifold satisfy an initial value problem in ﬁnite dimension. in a recursive way. This gives us idea about information needed from a local center manifold in order to compute terms of the normal form for the reduced system. Note that the computation of Taylor expansion is. by expliciting steps of the algorithm given in [4]. The issue of this chapter is to present a new. whose parameters depend only on terms of the same order in the considered FDE and the terms of lower order of a local center manifold already computed. In section 3. eﬃcient and algorithmic formulation of the problem of computing local center manifolds and normal forms associated to an FDE. we present an algorithm which provides us. This characterization is then explored to derive a computational scheme of a local center manifold. we state a result which gives us an analytic characterization of a local center manifold. in the second step. but only in approximating local center manifold associated to an FDE with special singularities (namely Hopf and BogdanovTakens singularities). Secondly.
0] . for φ ∈ C. An example related to the Bogdanov Taknes singularity is given in the end of this section. we refer the reader to [53] for notations and general results on the theory of FDEs in ﬁnite dimension spaces.. φ) .We reminder the reader that a concrete example relatedb to the Hopf singularity had been discussed in the last chapter. 0] and t ∈ [σ. Together with (1) . σ + a] . 0].. Section 4 details the normal form construction for both of the reduced system and the original FDE. σ + a] into Rn . Rn ) the Banach space of continuous functions from [−r. for t ≥ 0 where u is the solution of (2. Note that the Riesz representation allows us to represent the operator L as 0 dη (θ) φ (θ) L (φ) = −r where η is an n × n matrix valued function of bounded variation in θ ∈ [−r. for every θ ∈ [−r. then we denote by ut the element of C deﬁned by ut (θ) = u (t + θ) .2) with u0 = φ.2) with initial function φ at zero. φ) the unique solution of equation (2.1) dt where L is a bounded linear operator from C into Rn and f is a sufﬁciently smooth function mapping C into Rn such that f (0) = 0 and Df (0) = 0 (Df denotes the Fr´echet derivative of f ).2) If we denote by u (. We denote by C = C ([−r. Denote by A the inﬁnitesimal generator of (T (t))t≥0 . It is known that the spectrum σ (A) of A . then equation (2.0] taking [σ − r. Notations and background In this section. a) . If u is a continuous function θ∈[−r.2) determines a C0 semigroup of bounded linear operators given by T (t) φ = ut (. a) . 2. Let r ≥ 0 and n ∈ N .Approximating center manifolds and normal forms for FDE 195 Bogdanov Takens singularity respectively. 0] to Rn endowed with the supremum norm φ = sup φ (θ) . we consider the linearized equation near zero du (t) = Lut dt (2. Our main concern throughout this paper is with the autonomous FDE of the form du (t) = Lut + f (ut ) (2. into Rn such that x0 = φ and satisfying (1) for t in (0. A solution u = u (φ) of (1) through a point φ in C is a continuous function taking [−r..
.... . 0 . . σm−1 λm (2.. . φj )0≤i. . . ψm ) of the dual space PΛ∗ in C ∗ such that Ψ. m} this set of eigenvalues. Let m ∈ N be the number of solutions of (2.. With the this bilinear form we can compute the basis Ψ = col (ψ1 . . Φ = (ψi . according to the hypothesis the form 0 λ1 σ 1 0 λ2 . . the phase space C is decomposed by Λ as C = PΛ ⊕ QΛ .. . 1} such that the semiﬂow generated by (2..j≤m = IRn .. . where PΛ is the m dimensional generalized eigenspace associated to elements of Λ and QΛ is its unique T (t) invariant complement subspace in C.. φ ...5) where C ∗ = C ([0. the subspace QΛ is given in a compact form: QΛ = {φ ∈ C : Ψ. To explicit the subspace QΛ and the projection operator associated with the above decomposition of C we need to deﬁne the so called adjoint bilinear form as ψ.196 DELAY DIFFERENTIAL EQUATIONS coincides with its point spectrum σp (A) and it consists of those λ ∈ C which satisfy the characteristic equation 0 p (λ) = det∆ (λ) = 0 (2..3) with zero real part. we assume that the following hypothesis holds: (H) Λ = ∅... If we denote by Φ = (φ1 . φ) = 0} . Moreover. the subspace PΛ is then written as PΛ = {Φx : x ∈ Rm } .. there exists a matrix B of ··· . 0 . Rn∗ ) and Rn∗ is the ndimensional space of row vectors.2) restricted to PΛ is a linear ODE whose matrix is exactly B.. 0 ··· ··· (H) . ϕ = ψ (0) ϕ (0) − 0 −r θ ψ (s − θ) dη (θ) ϕ (s) ds. φm ) a basis of PΛ . ψ2 . 0 ... φ2 . counting their multiplicities. .3) where ∆ (λ) = λIRn − −r eλθ dη (θ) .. . and. . Denote by Λ = {λi ... i = 1. B= . r] . φ = col(ψ1 . Using the formal adjoint theory of Hale [53]... for ϕ ∈ C and ψ ∈ C ∗ 0 (2. ψm .4) with σi ∈ {0. Throughout this paper. . .
1) for which they proved that.1) near the steady state: more precisely. Using the variation of constants formula given in [53]. δ[ . Dh (0) = 0 and there exists a neighborhood V of zero in Rm such that. The classical deﬁnition of a local center manifold associated with the set Λ of eigenvalues in the imaginary axis is given as: Deﬁnition 34 : Given a C 1 map h from Rm into QΛ . the uniqueness of the Taylor expansion of these manifolds is ensured.6) x (0) = z. δ[ and it is given by ut = Φx (t) + h (x (t)) . Note that the lack of uniqueness makes local center manifolds hard to determine.1) is given by t T (t − s) X0 f (us ) ds. the authors in [53] have considered equation (2. .Approximating center manifolds and normal forms for FDE 197 Now consider the nonlinear equation (2. where x (t) is the unique solution of the ordinary diﬀerential equation dx (t) = Bx (t) + Ψ (0) f (Φx (t) + h (x (t))) dt (2. the solution ut given above may be decomposed as ut = Φx (t) + y (t) . fortunately. for each z ∈ V. the behavior of orbits of (2. It is important to observe that in general y (t) does not satisfy the translation property (ie y (t) (θ) = y (t + θ) (0)). where x (t) is an element of Rm and y (t) ∈ QΛ . This tool allows us to give a complete description of the local dynamics of equation (2. 0[ According to the decomposition of C endowed by the set Λ of eigenvalues. there exists δ = δ (ξ) > 0 and the solution x of (2. under hypothesis (H) . However. which is necessarily an mdimensional ODE.1) in C can be completely described by the restriction of the ﬂow to a local center manifold associated with Λ.1) . for t ≥ 0 ut = T (t) u0 + 0 where X0 = X0 (θ) is deﬁned by 3 IRn X0 (θ) = 0 for θ = 0 for θ ∈ [−r. the nonlinear term of the reduced system can be explicitly computed up to the order k + 1. the graph of h is said to be a local center manifold associated with equation (1) if and only if h (0) = 0. for t ∈ [0. a solution of (2. we have the existence of a local center manifold. After computing the Taylor expansion of a center manifold up to an order k.1) with initial data Φz + h (z) exists on the interval ]−δ − r. By the use of a ﬁxed point theorem.
0[ and with a jump discontinuity at 0.6) is said to be the reduced system of (2. the authors have presented a method for computing normal forms for FDEs. It is exactly the space BC of the functions from [−r. In terms of the function X0 deﬁned previously.7) dt 5 = φ +X0 [Lφ − φ (0)] with D A 5 = C 1 . 0] into Rn uniformly continuous on [−r. To develop a normal form theory for FDEs.1) . The abstract ODE in BC associated with the FDE (2. According to the above decomposition of BC. X0 = Ψ (0) .5) can be extended in a natural way to C ∗ × BC by Ψ. if we identify the space BC to C × Rn then it can be endowed with the norm φ + X0 αBC = φC + αRn to be a Banach space.198 DELAY DIFFERENTIAL EQUATIONS with B is the matrix deﬁned in (2.1) can be written in the form du (t) 5 (t) + X0 f (u (t)) = Au (2. a local center manifold is the graph of a given function mapping a neighborhood of zero in PΛ into QΛ which is tangent to PΛ and locally invariant under the semiﬂow generated by equation (2. the solution u can be written u (t) = Φx (t) + y (t) .1). The bilinear form in where Aφ C ∗ × C deﬁned in (2. The adequate phase space for this situation is the one introduced in [55] and used in [28] for the application of averaging methods to FDEs. Also. φ + Ψ (0) α] for every φ ∈ C and α ∈ Rn leading to the decomposition BC = PΛ ⊕ Ker (π) Let u be a solution of (2. The ODE (2.4) .1) is written as an abstract ODE. we can deﬁne the projection operator π : BC −→ PΛ such that π (φ + X0 α) = Φ [Ψ. it was necessary. In [44]. an element ψ ∈ BC can be written as ψ = φ + X0 α where φ ∈ C and α ∈ Rn . Geometrically speaking. in the ﬁrst step.7). to enlarge the phase space C in such a way that (2. and consequently.
10) +Φ (θ) Ψ (0) f (Φz + h (z)) . (2. for each . and ∂h (z) (0) = Lh (z) + f (Φz + h (z)) . Computational scheme of a local center manifold In this section. necessary and suﬃcient analytic conditions on a given function h under which its graph is a local center manifold of (2. we are particularly interested in obtaining normal forms for equations giving the ﬂow on local center manifolds.7) is equivalent to the system commutes with A d dt x (t) = Bx (t) + Ψ (0) f (Φx (t) + y (t)) (2. then. we have ∂ ∂h (z) (h (z)) (θ) = (θ) [Bz + Ψ (0) f (Φz + h (z))] ∂θ ∂z (3. In the context of FDEs. Theorem 1 Let h be a map from Rm into QΛ with h (0) = 0 and Dh (0) = 0.1).8) d 51 y (t) + (I − π) X0 f (Φx (t) + y (t)) y (t) = A dt 51 is the restriction of A 5 to Q1 interpreted as an operator acting where A in the Banach space Ker (π) . A necessary and suﬃcient conditions on the graph of h to be a local center manifold of equation (2. at ﬁrst.11) Proof. the equation (2. ∂θ where z (t) is the unique solution of (2. there exists a neighborhood V of zero in R2 such that.1). In applications. As π where x (t) ∈ Rm and y (t) ∈ Ker (π) ∩ D A 5 in C 1 . (3. typical equations to be considered for the computation of normal forms read as • x = Bx + Ψ (0) f (Φx + h (x)) . we derive. Necessity : If h is a local center manifold of equation (2.Approximating center manifolds and normal forms for FDE 199 5 = QΛ ∩ C 1 Q1 .6) . This will be exploited in the second step to develop the computational scheme of terms in the Taylor expansion of h. for each z ∈ V. 3.9) where h is a local center manifold.1) is that there exists a neighborhood V of zero in Rm such that.
15) +Φ (θ) [Ψ (0) f (Φξ + h (ξ))] . The relations Φ (θ) z (t) + h (z (t)) (θ) = Φ (t + θ) ξ + h (ξ) (t + θ) . we get ∂h (z (t)) ∂ (h (ξ)) (t + θ) = Φ (θ) + (θ) Φ (t + θ) Bξ + ∂θ ∂ξ × [Bz (t) + Ψ (0) f (Φz (t) + h (z (t))) Let t go to zero. for t ∈ ]−δ. ∂h (ξ) ∂ (h (ξ)) (θ) = (θ) [Bξ + Ψ (0) f (Φξ + h (ξ))] ∂θ ∂ξ (3. δ[ and it is given by xt = Φz (t) + h (z (t)) . Moreover.14) so.12) with respect to t. by using the proporety LΦξ = ΦBξ. the relation (3. On the other hand. this semiﬂow has a backward extension to ]−δ. there exists δ = δ (ξ) > 0 such that the solution of (2. one can see that there exists a continuous function y : ]δ − r.1) in a local center manifold. 0].12) and Φ (θ) z (t) + h (z (t)) (θ) = Φ (0) z (t + θ) + h (z (t + θ)) (0) . δ[ . for t ∈ ]−δ. for t > 0. δ[ . Then. ∂θ (3. it follows from (3. (3.1) with initial data Φξ + h (ξ) exists on the interval ]−δ − r. δ[ −→ Rn such that yt = Φz (t) + h (z (t)) . Suﬃciency : From the formulae (3. . for t + θ ≤ 0 and 0 ≤ t < δ. for t + θ ≥ 0 and 0 ≤ t < δ (3.13) . Then.200 DELAY DIFFERENTIAL EQUATIONS ξ ∈ V.11) follows immediately. by diﬀerentiating relation (3.13) that L (Φξ + h (ξ)) + f (Φξ + h (ξ)) = ΦBξ + ∂h (ξ) (0) .13) are immediately deduced from the translation property of the semiﬂow t −→ xt = Φz (t) + h (z (t)) generated by equation (2.12) and (3.
15) and substituting for the right hand side of (3. it remains to prove that y is a solution of (2.1) . which completes the proof of the theorem. ∂h (ξ) +G (Φξ + h (ξ)) ΦBξ + ∂θ . And in the case of neutral functional diﬀerential equations of the form d [D (xt ) − G (xt )] = Lxt + F (xt ) .15) into (3.201 Approximating center manifolds and normal forms for FDE So. (3.16) ∂ (h (z (t))) (θ) ∂θ ∂h(z (t)) d y(t) = Φ(0)Bz (t) + (0) dt ∂θ = L (yt ) + f (yt ) . A necessary and suﬃcient conditions on the graph of h to be a local center manifold of equation (3. In fact. we have ∂ (h (z)) (θ) ∂θ = ∂h (z) (θ) [Bz + Ψ (0) F (Φz + h (z)) + BΨ (0) G (Φz + h (z))] ∂z +Φ (θ) [Ψ (0) f (Φz + h (z)) + BΨ (0) G (Φz + h (z))] . for each z ∈ V. we obtain (3. ∂h (z (t)) dz (t) d y(t) = Φ(0) + (0) dt ∂ξ dt Letting ξ = z (t) and θ = 0 in (3. the analytic characterization of a local center manifold is given in the following corollary: Corollary 2 Let h be a map from Rm into QΛ with h (0) = 0 and Dh (0) = 0. and D ∂h (ξ) ∂θ = Lh (ξ) + F (Φξ + h (ξ)) . F and G are suﬃciently smooth functions mapping C into Rn such that F (0) = G (0) = 0 and F (0) = G (0) = 0 ( F and G denote the Frechet derivative of F and G respectively ).16).17) is that there exists a neighborhood V of zero in Rm such that.17) dt where L and D are bounded linear operators from C into Rn .
To this end consider the analytic characterization given in the above Theorem. the term hk ∈ Vkm (QΛ ) is computed from the terms of lower orders already computed (ie (hj )2≤j≤k−1 ) and from the terms of lower order of f. Hence.j. . .1) < q (m. pm − qm is positive.18) h (z) = j≥2 and f (φ) = fj (φ) (3. we denote by Djm the set of parameters Djm = q = (q1 . . cq ∈ Y Vjm (Y ) = q=j where q = deﬁned as m i=1 qi . . . .j. qm ) and p = (p1 . where j = card Djm and q (m. . . q2 . . . the polynomial (hj )j≥2 .202 3. . . q2 . the space Djm reads 0 1 Djm = q (m. zm ) with coeﬃcients in Y.j. z2 . . .j. . we have p < q if the ﬁrst nonzero diﬀerence p1 − q1 . In the sequel of this work we adopt the following notation : for j ∈ N and Y a normed space. .19) j≥2 The term hj (z) denotes the homogeneous part of degree j with respect to z of h. . . . . We will show that at the step k ≥ 2. In other words cq z q : q ∈ N m . . Without loss of generality we can assume that h and f can be written respectively in the form: hj (z) (3. pm ) in Djm . Vjm (Y ) denotes the space of homogeneous polynomials of degree j in m variables z = (z1 . j .1 DELAY DIFFERENTIAL EQUATIONS Formulation of the scheme Now we are in position to derive our algorithmic scheme for approximating a local center manifold associated with (2. p2 . . . . Moreover. in a recursive way. The goal now is to develop our algorithmic scheme for computing.1) .2) < · · · < q (m. . p2 − q2 .j ) . qm ) ∈ N m : q = m qi i=1 endowed with the following order : for q = (q1 .i) : i = 1.
21) for all i ≥ 2. . H (z) = f (Φz) + Df (Φz) h (z) + i! i≥2 So. according to the representations (3.22) h + i≥2 i!1 j≥2 Di fj (Φz) l≥2 l . Using the Taylor expansion of H (z) = f (Φz + h (z)) near h = 0 results in 1 Di f (z) [h (z)]i . . .+lj =i−(p−j) l1 l2 (3.10) is given by ∂hk (z) ∂hk (z) = Bz + Fk (z) .h + j=2 j! p=j D fp (Φz) lj (z) l1 +l2 +.18) and (3. the space Dkm is written as & ' Dkm = q i : i = 1. By dropping the parameters k and m in the above notation for simplicity of formulae.. k and as a consequence the space Vkm (Y ) can be written k i c i z q : ci ∈ Y . An exact formula of Hi is given by the following proposition : Proposition 34 Under the standing hypothesis (H).. .. it follows that i−1 Dfj (Φz) hi+1−j (z) Hi (z) = fi (Φz) + j=2 1 0 [ 2i ] 1 i−j j h h .. k − 1 are already performed (ie the terms (hj )2≤j≤k−1 are known). the above equation reads j≥2 Hj (z) = i≥2 fi (Φz) + i≥2 j≥2 Dfj (Φz) hi (z) i (3.. .Approximating center manifolds and normal forms for FDE 203 Fix k ∈ N.. Proof. Assume that steps of order 2.. . Vkm (Y ) = i=1 The homogeneous part of degree k with respect to z of equation (3..19). (3.20) ∂θ ∂z where Fk (z) is an element of Vkm (C) given by Fk (z) = k−1 ∂hk−i+1 (z) i=2 ∂z Ψ (0) Hi (z) + ΦΨ (0) Hk (z) where Hi ∈ Vim (Rn ) is the homogeneous part of degree i with respect to z of H (z) = f (Φz + h (z)) .
then it is easy to see that i i Dl (z) Bz = µi z q + qji σj z pj . .. 0] (3.20) in the space of homogeneous polynomial in z = (z1 . . .204 DELAY DIFFERENTIAL EQUATIONS Finally comparing the same order terms of (3. ak2 .20) i q an adequate formulation of computing and (3.24) in the basis z 1≤i≤k the term hk (z) ∈ Vkm (QΛ ) will be derived i i an1element of Dkm . First.24) ∂θ In the sequel of this section we will be concerned only with the equations (3. qm If we denote by Ii = j ∈ N : pij = q i + ej+1 − ej ∈ Dkm .25) dθ . q2i . i=1 where the aki are elements of QΛ . By identifying coeﬃcients of the both sides of (3. λ = kj=1 qji λj . akk . j∈Ii where µi = q i .. Remark 3 Fk (z) (θ) is an element of Vkm (Rm ) . (3.24) . ..22) yields us the result of the proposition.20) and (3.11). we put hk (z) = k i aki z q and X k = col ak1 . & ' for all i ∈ 1. Moreover we have pij > q i . The above proposition provides us with a clear relationship between Fk (z) and the terms of h (already computed) as well as the terms of the nonlinearity f of the original FED In the same way as above. (3. . we obtain ∂hk (z) (0) = Lhk (z) + Hk (z) .... . z2 . So. . . zm ) of degree k with coeﬃcients in QΛ it leads dX k (θ) = Ak X k (θ) + F k (θ) .23) i=1 where Fik are elements of C.. . Remark 4 Let l (z)0= z q with q i = q1i . the homogenous polynomial Fk (z) can be written as k F (z) (θ) = k i Fik (θ) z q . k and j ∈ Ii .. for θ ∈ [−r. by comparing the korder terms of (3. So according to the above remark and by identifying the two members of (3. according to the above notation. .
(3. . 0] where θ k S (θ) = (3. Mkk . 0] .... . . . . F2k (θ) . . If we put Hk (z) = k i Mik z q .. ··· 0 ··· .. 0 So. for θ ∈ [−r. consider the second equation (3..24) . . . .. .29) Bk = Ak − L (exp (·Ak )) (3.. Fkk (θ) .. . Secondly. . where The variation of constants formula applied to the ordinary diﬀerential equation (3. . i=1 then. (3. . the equation (3. for Mjk ∈ Rn .31) where. . it follows that the vector X k (0) is a solution of the linear system Bk X k (0) = E k ..26) F k (θ) = col F1k (θ) . M2k .25) yields X k (θ) = exp (θAk ) X k (0) + S k (θ) .205 Approximating center manifolds and normal forms for FDE where Ak = and µ1 IRn 0 µ2 IRn . . . ..30) E k = M k − F k (0) − LS k ..24) reads dX k (0) − LX k = M k .. in a vector representation. by substituting the above expression of X k into the relation (3. .28) exp ((θ − s) Ak ) F k (s) ds.. and ... ..27).. . 0 0 µk−1 IRn 0 µk IRn . ··· . dθ (3. for θ ∈ [−r.. 0 ..27) M k = col M1k . .. . (3.
.. ∆ µ . 0] dθ (3.. we have shown that the part of degree k of the function h satisﬁes the following initial value problem : k dX (θ) = Ak X k (θ) + F k (θ) .34) . .. . for θ ∈ [−r. . . As a consequence. . exp ·µk IRn D (3.. . Bk X k (0) = E k In which F k and E k are determined in terms h2 . a solution of (3. 0 ··· 0 . .. . X k = 0.33) k with X k ∈ (QΛ ) . . . · · · · · · . Substituting (3.. .) is the characteristic matrix deﬁned in the preliminary section.. .. .. . .. ..33) must satisfy Ψ... The only way to do that is the formal adjoint theory deﬁned in the6 preliminary 7 section. . . . fk of f. . . 0 0 .Ak ) is the n (k + 1)×m (k + 1) matrix given by C Ck = D Ψ. . ··· ∆ µk where ∆ (. . C 0 Ψ. . Bk = (3.32) .. hk−1 of h and the terms f2 . . Remark 5 It is noted that the matrix Bk need not to be nonsingular. S k and Ck = Ψ.. More precisely.1) ..30) yields the matrix Bk in the form: 1 0 0 ··· ··· 0 ∆ µ 2 .26) into (3. . exp ·µ1 IRn 0 C D Ψ.. exp ·µ2 IRn ··· 0 . . . . .206 DELAY DIFFERENTIAL EQUATIONS Then. in order to compute the initial data X k (0) of the problem (2. the vector X k (0) should be a solution of a system of two systems Ck X k (0) = N k and Bk X k (0) = E k 6 7 where N k = − Ψ. So.. ∆ µk−1 0 . we should take into account the abstract condition X k ∈ (QΛ )k . . Summarizing. . · · · · · · .. . . . . exp (.
for θ ∈ [−r. it suﬃces to prove that Y k (0) = 0 is the unique solution of the two systems Bk Y k (0) = 0 and Ck Y k (0) = 0. it satisﬁes equation (3. E k by ( 3.Approximating center manifolds and normal forms for FDE According to the above remark. one can see that equation (3.k is such a family of coeﬃcients then. To this end.35). in which the F k are given by ( 3. k .··· . our interest will be focused in the well posedness of problem (3. of the polynomials of degree k with coeﬃcient in the basis z q i=1.k be the family of the coeﬃcients of the graph representation ofi any local center manifold associated with equation (2.31) and Bk by ( 3. it is suﬃcient to prove that Y k ≡ 0 is the unique solution of the problem dY k (θ) = Ak Y k (θ) ..35) Finally.. the system (3. ai i=1.23). Fix i ∈ 1. 2. That’s why. To this end.··· . k .35) .. Bk X k (0) = E k 207 (3. (3. .k in QΛ .··· . no additional information or equation is necessary for the computation the kth term in the Taylor expansion of h. 0] dθ and Ck X k (0) = N k . 2. 0] dθ (3.30). Since e µi = that µi ∈ . The following result provides us with the complete scheme of computing the homogeneous part of degree k of a local center manifold. we have to prove that the kernel of the bloc matrix Ck zero. one can see that if kProof. for θ ∈ [−r. the ai i=1.··· k . we distinguish & 'two cases:& In the ﬁrst ' case we assume / Λ for all i ∈ 1.36) k k Bk Y (0) = 0 and Ck Y (0) = 0...1) . In other words.33) reads dX k (θ) = Ak X k (θ) + F k (θ) . . Then.35) leads to a unique set of coeﬃcients that is. In view of the above computations.k are uniquely determined by the equation (3. Theorem 3 Let aki i=1.35) .35) is all what can be extracted from the coeﬃcients of hk ..37) Bk is exactly So. What remains to be seen is that equation (3.
Then it is easy to see that Dk is invertible.. So. 'In the second case we assume that for all i in a subset Jk ⊂ 1... Or Bk = Ker (Bk ) ∩ Ker (Ck ) = {0} ..38) Ψ. the value µi coincides with an element of Λ.. exp ·µ b = 0 Note that µi is a complex element with zero real part. Consider the mapping α : Jk i −→ α (i)6 ∈ {1. So... . the boundary conditions (3.38) .. Then φα(i) (θ) = exp θµi and ψα(i) . 2. . we will show that b =1 0 for all i ∈ 1.µ ∈ Λ then ∆ µ b = 0 implies that exp ·µ 6 b is an 7 eigenfunc1 tion A associated to the eigenvalues µ = λα(1) . In view of (3. Then.. ... 2. if µ 1 1 and b1 = 0 and if not 1 1 ie. Ker Ck This implies that Y k (0) = 0 and thus Y k ≡ 0 is the unique solution of the problem & (3. bk an element of Cnk such that Bk b = 0 and i Ck b&= 0. where Dk is a matrix that can be constructed by mixing terms of Bk and those of Ck .. In prove that the kernel of the block matrix Ck fact.. Proof..34) .208 DELAY DIFFERENTIAL EQUATIONS 0. . When n = 1 (the scalar case). Ψ. & ' Lemma 28 Assume that for i ∈ Jk ⊂ 1. if we denote by Dk the matrix obtained from Bk and & Ck by replacing ' the rows i ∈ Jk of Bk with the rows (i + α (i)) ∈ 1. .37) can be written as Dk Y k (0) = 0. . So.37) is equivalent to Dk Y k (0) = 0. in view of the form (3. b2 .30) of the matrix Bk . k . k .32) and (3. by virtue 1 1 ∈ 1 is nonsingular / Λ then ∆ µ of (3.. In the nonscalar case (n > 1) ... we have ∆ µi = 0 for i ∈ Jk . exp ·µ1 b1 = 0 . it is diﬃcult to guess exactly the rows that we should eliminate from the matrix Bk and the ones that we can bring from the matrix Ck in order to build the invertiblematrix Dk . the characteristic matrix ∆ µi is nonsingular.. µi coincides with an element of Λ. it follows that b satisﬁes 3 1 1 ∆ 6 µ b =1 0 1 7 (3. To conclude the proof of the theorem we need the following lemma. φα(i) = 1. then.. In an ' induction procedure. But it is easy to Bk is reduced to zero.. 2. there exist a nonsingular nk × nk matrix Dk such that the boundary condition (3. 2.. it follows that Ker (Bk ) = {0} .36) . 7 m} . Thus. k .Or. 2. let b = col b1 . such that λα(i) = µi . mk of Ck respectively..
(3. akk . one can prove that the problem (3.35) can be written as dX k (θ) = Ak X k (θ) + F k (θ) .Approximating center manifolds and normal forms for FDE 209 7 6 leads to ψα(i) . for θ ∈ [−r..1 Case of Hopf singularity. Remark 6 In the same way as above. 0] .. . 3. in view of (3.we have dX k (θ) k = AX (θ) + F k (θ) ..34) .2 Special cases. 0] dθ (3. . by the use of elementary linear algebra tools. we suppose that Λ = {±ωi}. 2. we can prove that bj+1 = 0.32) and (3... then. exp ·µ and in the same spirit as above. the problem of computing terms of a local center manifold can be formulated in resolving a sequence of ﬁnite dimensional ODEs. This achieves the proof. j} . If we assume that bi = 0 for i ∈ {1. Thus Rank Ck =nk and consequently. This formulation makes the problem of computing coeﬃcients of a local center manifold in suitable form easy to manipulate in symbolic or numerical computations.. 3. In this case. In conclusion. · · · · · · . we have shown that Ker(Bk )∩Ker(Ck )={0}.2. where ω = 0 By virtue of assumption (H)..39) k k Dk X (0) = R where Dk is the matrix given in the above lemma and Rk is a complex vector whose components are a mixing of those of the vectors E k and N k . k − 1 .40) . exp ·λα(i) b1 = 0 which is not true except if b1 = 0. for θ ∈ [−r. i=0 where the aki are elements of QΛ . we prove that for θ ∈ [−r. we have 3 j+1 j+1 0 7 ∆ 6 µ b j+1 = j+1 b =0 Ψ. it follows that dim PΛ = 2. If we put hk (z) = k aki z1k−i z2i and X k = col ak0 . Summariz Bk ing. one can build a matrix Dk satisfying the lemma. Let ' & j ∈ 1. 0] dθ Bk X k (0) = E k and Ck X k (0) = N k ..
.. exp(−(k − 2j))iω)IRn . . 0 0 0 0 0 (k − 1) IRn . . In this case. where 0 kIRn 0 Ak = . . .. i=0 where the aki are elements of QΛ . . . ··· 0 IRn . .then for θ ∈ [−r.... k}) . By virtue of assumption (H). .. .1). ··· ··· 0 ··· .. ··· ··· ··· . · · · . akk .2. which ensures the existence of a center manifold associated with (2. . . . .. . . . . 0] .... . 0 0 0 . . .. 0 . j ∈ {0. .. we have dX k (θ) = Ak X k (θ) + F k (θ) . . . . If we put hk (z) = k aki z1k−i z2i and X k = col ak0 . . j ∈ {0.. .. k}) and Ck = Diag (Ψ.. .. · · · · · · ... · · · . .. ... 0 0 . 0}. .. ... . . we assume that zero is a double eigenvalue: Λ = {0. .. 0] dθ Bk X k (0) = E k and Ck X k (0) = N k .. . 0 . . .2 The case of Bogdanov Takens singularity. . . . 0 . ··· . 3. 0 (k − 2)iωIRn 0 0 kiωIRn . 0 . for θ ∈ [−r. Bk = Diag (∆(−(k − 2j))iω). . it follows that dim PΛ = 2. . .210 DELAY DIFFERENTIAL EQUATIONS where Ak = −kiωIRn 0 0 0 −(k − 2)iωIRn 0 ... . . .
By using the above lemma. C1k−1 ∆(1) (0) .. So...41) where ε is a real positive number assumed to be small.. k−1 . .. R) .. k ≥ 3. . θC 1 . (3. m! m=0 In a recursive way. λ is a real parameter and f is an element of C k (R. 0 k−1 k k k−1 0 θ Ck IRn θ Ck−1 IRn · · · · · · C0 IRn Proof : The matrix Ak is nilpotent. . It is a singularly perturbed delay equation given by εx (t) = −x (t) + f (x (t − 1) . . . .. In this section we consider as an application of our computational scheme an example treated by Hale and Huang [56]. Example. 0 . . in view of the study of qualitative structure of the ﬂow on the center manifold. In particular. . where k−j+1 m!Cm .. for some a. 0 k−1 (k−1) ∆ (0) · · · · · · C00 ∆ (0) Ckk ∆(k) (0) Ck−1 . . λ) = − (1 + λ) x + ax2 + bx3 + o x3 . Bk = . λ) . . . ..211 Approximating center manifolds and normal forms for FDE and Bk is a consequence of the following lemma: Lemma 29 The exponential matrix exp(θAk ) is given by 0 ··· ··· 0 C0k IRn . . exp(θAk ) = . which completes the proof of the lemma . where. one can easily prove that (Ak )m = (αij IRn )1≤ij≤n(k+1) ... . .. θC1k IRn C0k−1 IRn . if i − j = k αij = 0 . . it is easy to verify that exp(θAk ) = k θm (Ak )m . . b ∈ R. . . if i − j = k. . more speciﬁcally f (x. . . . one can see that 0 ··· ··· C0k ∆ (0) k (1) C1 ∆ (0) C0k−1 ∆ (0) . they need to know its Taylor expansion up to the second order terms. the authors need more precise information about a center manifold. . . ...
for (θ. 0] dθ B2 X 2 (0) = E 2 and C2 X 2 (0) = N 2 . In view of the assumed smoothness on the function f. with B = 0 1 0 0 . equation (3.1 that the computation of X 2 can be obtained by solving the diﬀerential equation dX 2 (θ) = A2 X 2 (θ) + F 2 (θ) .42) has a local center manifold. dt (3.42) with (r. then.212 DELAY DIFFERENTIAL EQUATIONS With the change of variables w1 (t) = x (−εrt) and w2 (t) = x (−εrt + 1 + εr) . (3. λ) = (1.42) If r = 1 and λ = 0. We showed in subsubsection 3. Φ = IR2 . for some a2i ∈ QΛ and χ (ξ) = o ξ2 . 0]×[0. and h (ξ) = a20 ξ12 +a21 ξ1 ξ2 +a22 ξ22 +χ (ξ) . Put X 2 = col a20 . The above equation has zero as a double root and the remaining roots have a negative real part.41) reads dw1 = rw1 (t) − rf (w2 (t − 1) . This ensures the existence of a two dimensional center manifold associated to the equation (3. a22 . at least of class C 3 . λ) dt dw2 = rw2 (t) − rf (w1 (t − 1) . 1] .2. for µ ∈ C. s) ∈ [−1.43) .42) is (µ − 1)2 − e−2µ = 0.42) with r = 1 and λ = 0. Base of PΛ and QΛ are respectively s −s −1 − 13 − θ and Ψ = . the characteristic equation associated to (3. Φ= 1 −1 1 1 3 +θ It is easy to verify that Ψ. 0) Throughout this section we use the notation presented in section 2 for the equation (3. λ) . the equation (3. a21 . and AΦ = ΦB. for θ ∈ [−r.
the relation (3. we present an algorithm for computing normal forms for FDEs. Moreover. . we will consider the computation of the parameters of the above system. we have a20 (θ) = a 2 1 1 2 1 1 θ 1 θ 11 + + . Now. 4. a21 (θ) = a θ + . for θ ∈ [−1.Approximating center manifolds and normal forms for FDE where B2 is the 6 × 6 real matrix −1 −1 −1 −1 2 2 B2 = 2 2 0 −1 −1 0 213 given by 0 0 0 0 0 0 0 0 −1 −1 0 0 −1 −1 0 0 1 1 −1 −1 1 1 −1 −1 . On the other hand. 6 F 2(θ) 7 = 0. exp (·A2 ) = 1 0 0 0 0 −1 1 −1 −1 1 −1 1 12 12 6 6 2 2 1 −1 −1 1 0 0 3 3 2 2 then by replacing the ﬁrst (resp. 3 3 9 9 is easy to see that the 6 × 6 real matrix G2 is given by −1 1 0 0 0 0 2 2 0 0 0 0 0 0 1 −1 −1 1 0 0 3 3 2 2 C2 = Ψ. . Computational scheme of Normal Forms In this section. a22 (θ) = a . we .30) leads to E = 4a 4a −4a −4a . In the context of ODEs. . X 2 (0) = col 2 2 3 3 36 36 and for θ ∈ [−1. the key idea is based on computing a change of variables that allows us to transform the original ODE into an equation with a simpler analytic expression. . 0] . 0] and as a consequence. 3 2 3 36 1 1 This conﬁrms the result established Hale and Huang in [56]. . for θ ∈ 2 [−1. by the third) line of the matrix C2 . . the boundary condition C2 X 2 (0) = N 2 and B2 X 2 (0) = E 2 imply that X 2 (0) is given in the unique way by a a a a 11a 11a .23) we have F 2 (ξ) (θ) = 0. S = 0. −a. 0] . . According to relation (3. we have 2 2 N = − Ψ. So. it col −a. by Lemma 29. the third) line of B2 by the ﬁrst (resp. In this section.
1) .1) as an abstract inﬁnite dimensional ODE of the form (2.8) .8) . if G = i≥2 Gi then Gi is known for all i ≥ 2 once (hi )i≥2 are founded. in conventional way. In the second subsection. Assume that the eﬀects of the above change of variables on the reduced system (4. we assume that a local center manifold is known and we deal with developing.46) is of the form • x = Bx + F (x) . we have elaborated an algorithmic procedure allowing us to compute at any order k the terms of the Taylor expansion of h. Without loss of generality h may be assumed in the form hi .44) are known at any order. in an algorithmic way. a computational scheme of normal form for the ODE (2.9) representing the restriction of the semiﬂow generated by (2. Note that. If we assume that (hi )i≥2 are already computed.1).1 Normal form construction of the reduced system Let h be a mapping from Rm into QΛ whose graph is a local center manifold associated with the equation (2. In the previous section. then terms of the ODE (4. Our approach is shared out in two shutters. our method consists in elaborating a scheme of computing nonlinear transformations that greatly simplify the (2. In the ﬁrst subsection. DV (0) = 0.1) (in the sense of deﬁnition presented in [44]).45) where V be a mapping from Rm into Rm such that V (0) = 0. Consider a nonlinear transformation of the form x = x + V (x) (4. In other words.44) where G (x) = Ψ (0) H (x) . it is the most information that can extracted from h with the impossibility of obtaining the closed form. in general.1). on this local center manifold is described by the following mdimensional ODE • x = Bx + G (x) .214 DELAY DIFFERENTIAL EQUATIONS will be concerned with developing. 4. we deal with the equation (2. the notion of normal forms for FDEs of the form (2.1) . h= i≥2 The ﬂow of the FDE (2. and we will show that this scheme leads us to simpler inﬁnite dimensional ODE which coincides with the normal form of (2.46) . (4. (4.
47) Vjm (Rm ) = Im Mj1 ⊕ Im Mj1 c where the complementary space Im Mj1 is not uniquely determined. Then. with Let P j be the projection associated the c above decomposition.Approximating center manifolds and normal forms for FDE 215 Our main goal now is to develop an algorithmic scheme providing us with adequate nonlinear transformations V that transform (2. we put Vi and F = Fi .. are exactly the Lie brackets that appear in computing normal forms for ﬁnite dimensional ODEs. the nonlinearity F is simpler than H). Bx + k≥2 k≥2 .48) in Proof.47) is given by Fk (x) = Gk (x) − Mk1 Vk (x) k−1 + i=2 {DGk+1−i (x) Vi (x) − DVi (x) Fk+1−i (x)} 1 0 [ k2 ] 1 k−j j D G (x) V V . V = i≥2 i≥2 In the sequel.+lj =k−(i−j) l1 l2 j! (4. We denote by Mj1 the operator deﬁned by 1 Mj p (x) = Dp (x) Bx − Bp (x) . the recursive formula for computing the coeﬃcients of the normal form associated to the decomposition (4. Substituting the change of variables (4. The operators Mj1 . If 1 1 we choose bases for Im Mj and Im Mj .6) into a simpler equation (4.45) into (4. for all p ∈ Vjm (Rm ) .. we will deal with developing a the recursive eﬃcient approach for computing the kth order normal form Fk and the kth order nonlinear transformation Vk in terms of (Gi )2≤i≤k as well as the terms (Fi )2≤i≤k−1 already computed..44) results Gk (x) = [I + DV (x)] Bx + Fk (x) .46) (ie.V + j=2 i lj (x) . Theorem 4 Fix k ≥ 2. We start by the following Deﬁnition: Deﬁnition 35 Let j ≥ 2. then. Without loss of generality. each element q of Vjm (Rm ) can be split into two parts P j q and I − P j q: P j q can be c spanned by the basis of Im Mj1 and I − P j q by that of Im Mk1 .. deﬁned above. It is well known that the space Vjm (Rm ) may be decomposed as c (4. i=j l1 +l2 +.
according to the equations (4.216 DELAY DIFFERENTIAL EQUATIONS which can be rearranged as k≥2 Fk (x) = k≥2 Gk (x + V (x)) + k≥2 BVk (x) − k≥2 DVk (x) Bx − DV (x) F (x) k k≥2 k≥2 k Then we use Taylor expansion of Gk (x + V (x)) near V = 0 to rewrite the above equation as F (x) = G (x) + DG (x) V (x) k k≥2 k k≥2 k k≥2 j≥2 j 1 + k≥2 Mk Vk (x) − DV (x) F (x) k k≥2 k≥2 k j 1 j . one can compute an adequate nonlinear transformation Vk such that and which leads us to 5k . by virtue of (4. we choose a suitable complementary space to the situation to be handled.23) given in the above theorem can be rewritten in a compact form 5 k − M 1 Vk . c Remark 7 It has been observed that the subspace Im Mk1 is not uniquely determined.+lj =k−(i−j) 1 2 Then. Moreover.51) In other words. Fk = I − P k G (4.50) and (4. one can note that the nonlinear transformation Vk depends on c the choice of Im Mk1 ... Consequently the normal forms are not unique. (4. In applications.47)..48).49) Fk = G j where 5 k (x) = G Gk (x) + k−1 i=2 {DGk+1−i 0(x) Vi (x) − DVi (x) Fk+1−i (x)} 1 k 1 k+j j + j=2 j! i=j D Gi (x) V V .51) ..49) and (4. This achieves the proof of the Theorem.V (x) l l l j l1 +l2 +. Mk1 Vk = P k G (4. + k≥2 j≥2 j! D Gk (x) l≥2 Vl (x) Finally comparing terms degree k in the above equation yields the formula (4. we can ﬁnd a change of variables that aﬀect the reduced system by taking away nonlinear (called non resonant terms) that terms are in the range subspaces Im Mk1 and conserving only terms (called c resonant terms) that are in the complementary subspace Im Mk1 . The formula (3. .50) 5k .
1) then F and V satisfy DV (x) Bx − BV (x) = Ψ (0) f Φ [x + V (x)] + 5 h (x) − DV (x) F (x) − F (x) (4. However. h (x) . the following Theorem illustrates another approach that combines center manifold and normal forms schemes into one step to simultaneously obtain a compact form illustrating the close relationship between the normal forms F . we need the following lemma: . Proof.53) . we must take into account the algorithm developed in section 3 and which provides us with homogeneous parts of the mapping h.Approximating center manifolds and normal forms for FDE 217 Remark 8 In view of the above Theorem. it has observed from the Proposition (34) that the homogeneous polynomial Gj can be given in terms of (hi )2≤i≤j−1 .45) .23) . So. the associated nonlinear transformations V. So.45) into (2. in view of (F N 1) it where f51 (x) = Ψ (0) f Φ [x + V (x)] + 5 follows that [I + DV (x)] [Bx + F (x)] = Bx + BV (x) + f51 (x) .9) with a nonlinear transformation of the form (4. As a consequence.6) results in • [I + DV (x)] x = Bx + BV (x) + f51 (x) .53) 5 where h (x) = h (x + V (x)) .52) . which can be rearranged as DV (x) Bx − BV (x) = f51 (x) − DV (x) F (x) − F (x) . if x is a solution of the equation ( 2. the complete scheme for computing normal forms for (2. Substituting the nonlinear transformation (4.52) and 515 D5 h (x) B−A h (x) = (I − π) X0 f Φ [x + V (x)] + 5 h (x) −D5 h (x) F (x) . in order to present the complete scheme for computing normal forms for (2. a local center manifold h and the coeﬃcients of the original FDE (2.6) . (3. In order to save computational time and computer memory.6) should be a combination of the formulae (3. Theorem 5 Eﬀect the reduced system (2.21) .39) and (3.1) . To prove (4. (4. Hence. it is noted that the kth order term Fk of the normal form depends upon the known of (Fj )2≤j≤k−1 and the results (Vj )2≤j≤k as well as (Gj )2≤j≤k−1 . This gives the relation (4.
54) satisﬁed by h results in 51 h (x + V (x)) = f52 (x) Dh (x + V (x)) Bx + BV (x) + f51 (x) − A h (x) .45) into the equation (4. More exactly. This achieves the proof of the theorem.1) is that there exists a neighborhood V of zero in Rm such that 51 h (z) = (I − π) X0 f (Φz + h (z)) − Dh (z) Ψ (0) f (Φz + h (z)) Dh (z) Bz − A for z ∈ V (4. So. Thus.53) in the above theorem results in Fk (x) = Hk1 (x) − Mk1 Vk (x) (4. If we denote by Mj2 the operators deﬁned as 2 51 p (x) for all p ∈ Vjm Q1Λ .52) already proved. Substituting the nonlinear transformation (4.53) cannot be founded. by the aid of the with f52 (x) = (I − π) X0 f Φ [x + V (x)] + 5 formula (4.54) Proof. the closed form solutions of (4. A necessary and suﬃcient conditions on the graph of h to be a local center manifold of equation (2. approximate solutions may be assumed in the form of 5 5 h= hi i≥2 where 5 hi is an element of Vim Q1Λ . the above equation reads 51 h (x + V (x)) = f52 (x) . Hk1 (x) and Hk2 (x) are respectively the kth homogenous part of the functions h (x) − DV (x) F (x) (4.56) where Hk1 (x) and Hk2 (x) are homogeneous polynomial of degree k.218 DELAY DIFFERENTIAL EQUATIONS Lemma 30 Let h be a map from Rm into QΛ with h (0) = 0 and Dh (0) = 0.57) H 1 (x) = f 1 Φ [x + V (x)] + 5 . comparing the same order in the equations (4. Dh (x + V (x)) [I + DV (x)] [Bx + F (x)] − A which is equivalent to 515 D5 h (x) B − A h (x) = f52 (x) − D5 h (x) F (x) . Mj p (x) = Dp (x) Bx − A then.52) and (4. In general.55) and Mk25 hk (x) = Hk2 (x) (4.
. [ k2 ] 1 k−j j 2 (x) D f (Φx) W W . So if we assume depend into the terms (Vi ) 2≤i≤k−1 2≤i≤k−1 that the lower order terms in V and 5 h are already computed.. 5 k for all k ≥ 2. we proceed in the following way: We consider the equation (4. (4. The above result illustrates clearly howtheelements Hk1 (x) and Hk2 (x) and 5 hi . Proof. then. for all φ ∈ C.W + j=2 l l l i l +l +.58) where f 1 (φ) = Ψ (0) f (φ) and f 2 (φ) = (I − π) X0 f (φ) . then.Approximating center manifolds and normal forms for FDE 219 and H 2 (x) = f 2 Φ [x + V (x)] + 5 h (x) − D5 h (x) F (x) .47) of the space Vjm (Rm ) ..55) .+lj =k−(i−j) Wl1 Wl2 . one can easily obtain the above relations..Wlj (x) i=j D fi (Φx) j! = 1 k−1 0 2 Dfk+1−i fk2 (Φx) + i=2 (Φx) Wi (x) + D5 hi (x) Fk+1−i (x) 1 0 . An exact formula of H 1 (x) It is easy to see that Hk1 = H k and Hk2 (x) is given in the following result.. .. In the same spirit of the the proof of the theorem () . in order to compute an adequate nonlinear transformation Vk that greatly simpliﬁes the reduced system. by the aid of a decomposition of the form (4..+l =k−(i−j) 1 2 j i=j j! 1 2 j and Hk2 (x) where Wl = ΦVl + 5 hl for all l ≥ 2. Theorem 6 The recursive formula deriving Hk1 (x) and Hk2 (x) is given by Hk1 (x) = 1 k−1 0 1 Dfk+1−i fk1 (Φx) + i=2 (Φx) Wi (x) − DVi (x) Fk+1−i (x) 1 0 [ k2 ] 1 k−j j 1 + j=2 l1 +l2 +. Removing non resonant terms and letting resonant terms results in Mk1 Vk = P k Hk1 and Fk = I − P k Hk1 .. the homogeneous polynomial Hk1 (x) is spliced into two parts the resonant and the non resonant one.
56) of the scheme. Identifying the two sides of the equation (4. . Let us now consider the equation (4. It is noted that. In fact.. Ker (π) and Rn . k leads us to 5k 5 k (θ) + F5k (θ) . .. cki 1≤i≤k and αik 1≤i≤k are respectively elements of Q1Λ .. from the fact that Hk2 ∈ Vkm (Ker (π)) . Put X = col b1 . once Hk1 is known. i=1 where the coeﬃcients bki 1≤i≤k . In conclusion.56) 1 0 i according the canonical basis xq : i = 1. solving equation (4.56) . So. 0] = Ak X dθ 5k 5k 5 k (0) = E 5 k (0) = N Bk X and Ck X where D C 5 k = − Ψ.56) will be needed for computing the kth order term 5 hk which allows us to pass to the next step k + 1 where the computation of the (k + 1)th order normal form Fk+1 and the (k + 1)th order nonlinear transformation as well as 5 hk+1 .57) one can easily observe that 5 hk and Hk2 are respectively elements of Vkm Q1Λ and Vkm (Ker (π)) .. for θ ∈ [−r. the coeﬃcients cki and αik should satisfy the k k k k k k 5 relation ci = ΦΨ (0) αi for i = 1. if we combine the adjoint bilinwhere M 1 2 k ear form (cf. b2 . αk . bk and F5k = col ck1 . 0] dX (θ) = Ak X dθ 5 k (0) dX 5 k ∈ (QΛ )k 8k 5k = M X −L X dθ 8k = col αk .... Then.. according to the notation presented in section 3.. polynomials 5 hk and Hk2 are written as 5 hk (x) = k i=1 bki xq i and Hk2 (x) = k i=1 i cki xq + X0 k i αik xq .. ck2 . In view of the deﬁnition of 5 h and the formula (4.56) is necessary for passing from a step to step. S5k N (4. αk . . The above problem reads 5 k (θ) dX 5 k (θ) + F5k (θ) .220 DELAY DIFFERENTIAL EQUATIONS Remark 9 One can ask a question about the usefulness of the second equation (4.59) . for θ ∈ [−r. ckk .. section 2) and the variation of constant formula tools. . .. k.. the equation (4.
y) (4. the authors have been concerned with solving.61) is all what we need for computing normal formes for the reduced system (2.38) is equivalent to 5 k (θ) dX 5 k (θ) + F5k (θ) . we will be concerned with equation (2.1) .56) which lends itself to numerical or symbolic computations by means of a recursive process.62) • 51 y + f 2 (x. In view of 51 in the deﬁnition of M 2 (j ≥ 2). 5 k and N vectors E In conclusion. we claim that the sequence of equations (4. for which an algorithmic scheme for computing normal forms is derived.55) and (4. it is known that equation (2. the authors in [44] proved that A Mj2 is nonsingular for all j ≥ 2. The equation (3. 0] = Ak X dθ (4. Proposition 35 Let Ak and Dk are the matrices appearing in the algorithm of approximation of a local center manifold (cf. However. In view of the preliminary section. the diﬃculties entailed by the operator A j the authors in [44] have been restricted to the study of the spectrum of 51 .60) and (4. their derivation does not allow them to compute or at least approximate the inverse of Mj2 . equations similar to (4.2 Normal form construction for FDEs In this subsection.Approximating center manifolds and normal forms for FDE and 8k − F5k (0) − LS5k 5k = M E with S5k = θ 221 (4.6) at any order.59).60) exp ((θ − s) Ak ) F k (s) ds 0 Finally.56) . the previous result provides us with a suitable form of (4. In contrast. for θ ∈ [−r. section 3).55) and (4. And under non resonance conditions. Remark 10 In the work [44]. at each step. y) y = A . we have the following us with the algorithm result which provides of computing coeﬃcients bki 1≤i≤k in terms of cki 1≤i≤k and αik 1≤i≤k .1) can viewed as an inﬁnite dimensional ODE given by • 1 x = Bx + f (x. 4.61) 5 k (0) = R 5k Dk X 5k is a vector whose components are a mixing of the ones of the where R 5 k respectively deﬁned in (4.
65) and (4. y) + DV 2 (x) Bx + g 1 (x.64) with g 1 and g 2 satisfying g 1 (x.63) yields equations (4.64) Our goal in the sequel is to develop an iterative procedure for ﬁnding adequate transformations V 1 and V 2 that greatly simplify equation (4. by substituting them into equation (4. assuming that the change of variables (4.1).62) and (4. y) = f 1 x + V 1 (x) . As has been deﬁned in [44]. We ﬁrst start by stating the following Theorem that illustrates us how. the transformed equation (4. the above change of variables (4. y) y = A (4.63) results in • • x = I + DV 1 (x) x • • • y = y + DV 2 (x) x So. y) = (I − π) X0 f (Φx + y) .62) . y) y = A Finally combining the above formula with (4. y) • 51 y + g 2 (x.45) transforms equation (4.62) . diﬀerentiating the relations of (4. y) = Ψ (0) f (Φx + y) and f (x. First.63) . A conventional approach for computing normal forms consists in ﬁnding adequate nonlinear transformations V 1 and V 2 that provide us.63) y = y + V 2 (x) results in • x = Bx + g 1 (x.64) associated to that change of variables is called normal form of (2.62) with the change of variables (4. Eﬀecting the equation (4.65) 2 51 V 2 (x) x + V 1 (x) . y + V 2 (x) −DV 2 (x) g 1 (x. y) • 51 y + g 2 (x.62) into equation of the form (F N 2) yields • x = I + DV 1 (x) Bx + g 1 (x.66) . with a simpler form of the both .222 DELAY DIFFERENTIAL EQUATIONS 1 2 where f (x. y + V 2 (x) − DV 1 (x) g 1 (x. y) − DV 1 (x) Bx − BV 1 (x) (4. aﬀects equation (4. This ends the proof of the Theorem.66) Proof. y)− DV 2 (x) Bx − A and g 2 (x. Theorem 7 Eﬀecting the equation (4. y) = f (4.62) .45) results in equation of the form(4.62) with nonlinear transformations of the form 3 x = x + V 1 (x) (4.
+lj =k−(i−j) Wl1 Wl2 ..66) cannot be founded. y) = 0 1 1 Dfi2 (Φx + y) Wk+1−i (x) + DVi1 (x) gk+1−i fk2 (Φx + y) + k−1 (x. y) Dfi1 (Φx + y) Wk+1−i (x) + DVi1 (x) gk+1−i i=2 0 1 [ k2 ] 1 k−j j 1 + j=2 l1 +l2 +.. y) − DV 2 (x) g 1 (x. Refereing to the results stated in the previous subsection. Denote by f5 (x. the closed form of V 1 and V 2 respectively solutions of (4.68) and gk2 (x. V1 = i≥2 i≥2 Our task is then reduced only to the computation terms of the Taylor 1 2 expansion respectively of V 1 and V 2 ... y) = f5k (x.66) results in 51 gk1 (x. y) − Mk1 Vk1 (x) (4. y) are respectively the kth order term (with 1 51 respect to the arguments x and y) of the functions f5 (x. expanding and comparing the both sides of equations (4. y) and f5k (x. Fix y = 0 in formulae (4.Approximating center manifolds and normal forms for FDE 223 of g 1 and g 2 .. y) i=2 1 0 [ k2 ] 1 k−j j 2 + j=2 l1 +l2 +.. one can ﬁnd adequate nonlinear transforma .68) . approximate solutions may be assumed in the form of Vi1 and V 2 = Vi2 .65) and (4. y + V (x) . y) is given by 51 f5 k (x.70) where Wl (x) is the homogeneous part of degree l of W (x) = ΦV 1 (x) + V 2 (x) . y) = f5 (x.69) and 52 f5 k (x. y) = 1 0 1 fk1 (Φx + y) + k−1 (x. So. however. y) − Mk2 V 2 (x) (4.Wlj (x) i=j D fi (Φx + y) j! (4. y) 51 52 where f5k (x. In general. y) . y) and f (x.67) 2 5 5 = f k (x.67) and (4.Wlj (x) i=j D fi (Φx + y) j! (4.65) and (4.. y) re 2 spectively the function f x + V 1 (x) .+lj =k−(i−j) Wl1 Wl2 . Thus. we have the following proposition: 52 51 Theorem 8 The recursive expression of f5k (x.. y) − 2 52 5 1 1 DV (x) g (x. y) and f5k (x. y) = f5 (x. More precisely. y + V 2 (x) and 2 1 2 f x + V (x) . y) and f5 (x.
providing us with terms of the normal form for FDEs at any order and the associated nonlinear transformations: Theorem 9 Consider an FDE of the form (2. we are in position to state the following Theorem summarizing the results for the new recursive and computationally eﬃcient approach. 0) = 0 for x. 0) (4.73) for all k ≥ 2. To be explicit. Now. 0) . y) k≥2 k with gk1 and gk2 satisfying (4.65) .1): Deﬁnition 36 The normal form for (2. y small and k ≥ 2.73) gk2 (x.72) Mk2 Vk2 2 5 5 (x) = f k (x. a local center manifold satisﬁes y = 0 and the normal form (in the usual sense for ODEs) of the reduced system of (2.71) 51 gk1 (x. with the aid of the previous Deﬁnition. (4.1) on this invariant manifold is given by the mdimensional ODE • x = Bx + g 1 (x.73) and gk2 (x.8) . 0) = I − P k f5k (x. Moreover. And thus similar to the one given in [44]. (4. y) (4. If we consider (2. (4.71) . 0) (4. y) y• = A 51 y + g 2 (x.75) .7)) relative to the decomposition (4.1) with non hyperbolic steady state satisfying the hypothesis (H) .1) as an inﬁnite dimensional ODE of the form (2. then there exists formal nonlinear transformations x = x + V 1 (x) and y = y + V 2 (y) such that x• = Bx + 1 k≥2 gk (x.72) and (4.1) (or equation (2. 0) .224 DELAY DIFFERENTIAL EQUATIONS tions Vk1 and Vk2 satisfying and 51 Mk1 Vk1 (x) = P k f5k (x.47) is an equation of the form x• = Bx + 1 k≥2 gk (x. we state the following deﬁnition.74) y• = A 51 y + g 2 (x. of normal forms for (2. y) k≥2 k with gk1 and gk2 satisfying (4.71) . (4. (4. 0) = 0.
78) . As a consequence.77). + j=2 l1 +l2 +. DV 2 (x) Bx − A So. if we observe that Hi1 (x) = Ψ (0) fi (Φx) for i ≥ 2 then..69) results in 5 f5 k (x. it follows that gk1 (x. one can see that the ﬁrst part of the Theorem is a direct consequence of properties of the operators Mk1 and Mk2 stated in the previous subsection combined with the equations (4. This reads 51 V 2 (x) = f 2 x + V 1 (x) .1) .76) y = V 2 (x) .77) V 2 (x) = 5 In the other hand. 0) = Fk (x) .. it is known that V 2 (x) is a solution of (4.Wlj i=j D fi (Φx) j! So. (4. Let us to prove the second part of the Theorem. Proof: Substituting y = 0 in the change of variables (4. the above equation can be written as 515 l (x) = G (x) . To this end we need the following lemma: Lemma 31 The equation y = 0 yields the equation of a local invariant center manifold of (2. we need to show that y = 0 provides us with this center manifold. by applying the implicit functions Theorem to the above equation. In the second step. one can prove the existence of a function h mapping a neighborhood V of zero in Rm into Ker (π) such that y = l (x) . the equation (4.66) .71) . 0) 1 = & 1 ' 1 Dfi (Φx) Wk+1−i (x) + DVi1 (x) gk+1−i fk1 (Φx) + k−1 (x) i=2 1 0 [ k2 ] 1 k−j j 1 (x) . D5 l (x) Bx − A (4. substituting y = 0 into (4.73) . (4.75) is exactly the normal form of the reduced system of (2.+lj =k−(i−j) Wl1 Wl2 .65) . V 2 (x) − DV 2 (x) F (x) .63) yields 3 x = x + V 1 (x) (4.. by virtue of (4.Approximating center manifolds and normal forms for FDE 225 Proof: After what was done before the statement.1) in the local center manifold considered. (4. So.. This will be done in two steps: First. This implies that l (x) = l x + V 1 (x) .
where Gk (x) is the homogeneous part of degree k of the function G (x) deﬁned above. combining (4. This will be done is a recursive way: In one hand.56) leads to 5 lk (x) = 5 hk (x) . we should prove that the kth parts 5 lk (x) and 5 hk (x) respectively of 5 l (x) and 5 h (x) coincide for all k ≥ 2.226 DELAY DIFFERENTIAL EQUATIONS where G (x) = (I − π) X0 f Φ x + V 1 (x) . and since the operator is nonsinguand M225 h2 (x) . it is noted that 5 l2 (x) 2 5 5 and h2 (x) are respectively solutions of M2 h2 (x) = (I − π) X0 f2 (x) l2 (x) = (I − π) X0 f2 (x). for k =2. Assuming that lar then it follows that 5 l2 (x) = 5 5 5 li (x) = hi (x) for 2 ≤ i ≤ k − 1 results in Hk2 (x) = Gk (x) . . This ends the proof of the Theorem. Thus. In the other hand.78) and (4. 5 l (x) − D5 l (x) F (x). To achieve the proof of the Lemma.
pt 1. In fact. Universidade de Lisboa Campo Grande 1749016 Lisboa Portugal tfaria@ptmat. and early XX century with the works of Liapunov and Birkhoﬀ. 227–282.ul. and CMAF. Arino et al. a brief explanation of the method for ODEs is given. called a normal form. around the origin. 227 O. Eq.Chapter 7 NORMAL FORMS AND BIFURCATIONS FOR DELAY DIFFERENTIAL EQUATIONS T.f. this idea is very old. f (0) = 0. (eds. Introduction The key idea of the normal form (n. x(t) ˙ = Bx + f (x) (1. going back to the late XIX century with the works of Poincar´e on Celestial Mechanics. all the eigenvalues of B have nonzero real parts.fc. which has the same qualitative behaviour of the original equation.. is of interest mostly in the case of a nonhyperbolic equilibrium x = 0. © 2006 Springer. Faculdade de Ciˆencias. Df (0) = 0 and f ∈ C k (k ≥ 2).e.1) where B is an n × n constant matrix. Bibikov and Br’juno gave signiﬁcant contributions to the ﬁeld. the theorem of HartmanGrobman assures that. Delay Differential Equations and Applications. Consider an autonomous nonlinear ODE in Rn with an equilibrium at zero.1) is topologically equivalent to the linearization x˙ = Bx.) technique is to transform a nonlinear diﬀerential equation into an equation with a simpler analytic expression. (1. Before developing a normal form theory for delay diﬀerential equations (DDEs). if x = 0 is a hyperbolic equilibrium. Faria Departamento de Matem´ atica.f. . In the framework of ordinary diﬀerential equations (ODEs).). i. More recently (1970’s). The application of the method of n.
leading to 1 1 1 gj−1 (x) + f˜j (x) + h.f. . j − 1 have already been performed.o. . Mj : Vjn (Rn ) → Vjn (Rn ). one can choose Uj in such a way that gj ∈ Im (Mj )c . Uj ]. x ¯˙ = B x ¯ + g2 (¯ 2 j! x ¯ ∈ Rn . (1. for Y a Banach space. x˙ = Bx + g2 (x) + · · · + 2 (j − 1)! j! where j!1 f˜j (x) are the terms of order j obtained after the changes of variables in previous steps. A change of variables of the form x=x ¯+ 1 Uj (¯ x). with coeﬃcients in Y .3) with the new terms of order j given by gj = f˜j − [B. j! with Uj ∈ Vjn (Rn ).t. at each step j. .t. . Hence.228 DELAY DIFFERENTIAL EQUATIONS Write the Taylor formula for f . (1.4) permit to write gj = f˜j − Mj Uj .o. The method of n. . Mj U = [B. The operators associated with these changes of variables. Uj ] is the Lie bracket [B. f (x) = k 1 fj (x) + o(xk ). transforming it into 1 1 x) + · · · + gj (¯ x) + h. 2 ≤ j ≤ k. Here. all the nonrelevant terms of order j are eliminated from the equation. such that. xn ) ∈ Rn . and h. . U ].2) is then applied to the above equation. . Uj ](x) = DUj (x)Bx − BUj (x). j! j=2 where fj ∈ Vjn (Rn ) is the jth Fr´echet derivative of f at zero. where [B. . consists of a recursive process of changes of variables. . (1. we use the notation Vjn (Y ) to denote the space of homogeneous polynomials of degree j in n variables x = (x1 .t stands for higher order terms.o. At a step j. assume that steps of orders 2.
tangent at zero to the centre space Rp for the linearization of (1. f 2 : Rn → Rm are C k functions (k ≥ 2) such that f 1 (0. Let Pj : Vjn (Rn ) → Im (Mj ) be the canonical projection associated with a decomposition Vjn (Rn ) = Im (Mj ) ⊕ Im (Mj )c . y ∈ Rm . We can actually compute recursively the function h(x) = 12 h2 (x) + · · · giving the equation of W c (see e. Re λ(C) = 0. (1. . 0) = 0. associated with a decomposition Vjn (Rn ) = Ker (Mj ) ⊕ Ker (Mj )c . 0) = 0. since it requires to compute the centre manifold beforehand. y). h(x)). Eq.1). (1. 0) = 0. Uj deﬁned by (1.g. and therefore write the equation for the ﬂow on such manifold: x˙ = Bx + f 1 (x.1) are called resonant terms of order j. p + m = n. the method of n. Df 1 (0. The terms j!1 gj (x) ∈ Im (Mj )c that cannot be eliminated from (1. thus Uj = (Mj )−1 Pj f˜j .8).7). Consider now an equation (1.Normal Forms And Bifurcations For DDE 229 where Im (Mj )c is a complementary space (in general not uniquely determined) of Im (Mj ). 0) = 0.6). where V is a neighbourhood of 0 ∈ Rp . y) y˙ = Cy + f 2 (x.3) is called a normal form for Eq. This approach is not very eﬃcient. up to j order terms.8).1) in the form 3 x˙ = Bx + f 1 (x. Df 2 (0.f.6) In order to have Uj must be chosen so that Mj Uj = Pj f˜j . (1. (1. (1. and that the qualitative behaviour of the ﬂow near zero is determined by its behaviour on the centre manifold. For gj . for (1. f 2 (0. In order to apply the normal form procedure to the equation for the ﬂow on the centre manifold of (1.f. (1. Re λ(B) = 0.9) After this ﬁrst step. (1.7) where (Mj )−1 is a right inverse of Mj . x ∈ Rp .5) gj = P rojIm (Mj )c f˜j = (I − Pj )f˜j . (1. described above is applied to compute a n. The centre manifold theorem tells us that there is a centre manifold W c = {(x. there are two possible approaches. f 1 : Rn → Rp . x ∈ V }. [24]).8) where B is a p × p matrix.9). y) : y = h(x). C is an m × m matrix.
2) is replaced by (1. .230 DELAY DIFFERENTIAL EQUATIONS Another procedure ([29. simultaneously. (1. Uj2 ∈ Vjp (Rm ).8) on the centre manifold and.8) is given by the system 3 x= ˙ Bx (1. This is achieved by considering a sequence of change of variables x=x ¯+ 1 1 U (¯ x). y ∈ Rm . . and Mj2 : Vjp (Rm ) → Vjp (Rm ). Note that the linearization of (1. k Eq. Following the procedure described above. U ]. with the operators Mj1 deﬁned as before. gj2 = f˜j2 − Mj2 Uj2 . . a change of variables is used to project the original equation (1. some diﬃculties arise in working in an inﬁnitedimensional phase space.11) 1 2 where the new terms gj = (gj .13) y= ˙ Cy.8) is transformed into an equation in the form 3 1 1 gk (x. x˙ = Bx + g21 (x. . Clearly. where now the change of variables (1. (1. although this is not suﬃcient to decompose a linear DDE into a decoupled system of linear DDEs similar to (1. y) + · · · + k! gk (x. j! j (1. j! j y = y¯ + 1 2 U (¯ x). . y) + h. to eliminate the nonresonant terms of order j from the equation giving the ﬂow on the centre manifold. .t. y ∈ Rm .1. . (Mj2 U )(x) = DU (x)Bx − CU (x). at each step j.12) will be deﬁned in an inﬁnitedimensional space Vjp (P c ). Hence a decomposition similar to (1.10) where Uj1 ∈ Vjp (Rp ). gj ) of orders j. for P the centre space for its linearization at zero. A second diﬃculty comes from the fact that the operator Mj2 in (1. .5) is not obvious. . are given by gj1 = f˜j1 − Mj1 Uj1 . (1.t. Mj1 U = [B. Chapter 12]) is based on the idea that.o. [58]). j = 2. y) + h. x ∈ Rp . a complementary space P c for P can be obtained by using the formal adjoint theory for linear DDEs in Rn (cf. x ∈ Rp . such as the natural phase space for a retarded functional diﬀerential equation (FDE) in Rn . where x ∈ Rp ≡ P .13). This latter procedure provides a general setting to be naturally extended to inﬁnitedimensional spaces. after steps j = 2.o.10). .12) Remark 1. First of all.8). However. and y in a complementary space P c for P . y) + · · · + k! 1 2 2 y˙ = Cy + g2 (x. k. it is not clear how to decompose the linear part of an FDE in Rn so that it can be written as (1.
ϕ(0) ˙ = L(ϕ)}. and F is a C k function (k ≥ 2). 0]. Rn∗ ).1 Preliminaries For Rn∗ the ndimensional vector space of row vectors. 0 θ (ψ. ϕ ∈ C. L : C → Rn is a bounded linear operator. 0] to Rn equipped with the sup norm. 2. C := C([−r. Aϕ = ϕ. and σ(A) = σP (A) = {λ ∈ C : det∆(λ) = 0}. Normal Forms for FDEs in Finite Dimensional Spaces In the next sections. u(t) ˙ = L(ut ) + F (ut ) (t ≥ 0). 0]. Consider also the linearized equation at zero. ψ ∈ C ∗ . since we are mainly interested in applying n. u(t) ˙ = L(ut ). r]. where I is the n×n identity matrix.1) where r > 0. and the formal duality (·. in general formal Taylor series will be considered. ut ∈ C is deﬁned by ut (θ) = u(t+θ) for θ ∈ [−r. to the study of singularities with possible bifurcations that are generically determined up to a ﬁnite jet. DF (0) = 0. Consider autonomous semilinear FDEs of retarded type in Rn (or Cn ) and with an equilibrium point at the origin. we refer the reader to [58] for notation and general results on the theory of retarded FDEs in ﬁnite dimensional spaces. the function g = 12 g2 + 3!1 g3 +· · · in the n. with F (0) = 0. A : D(A) ⊂ C → C.231 Normal Forms And Bifurcations For DDE Remark 1. deﬁne C ∗ = C([0. For the sake of exposition. is not necessarily analytic. in applications this is not a problem. ∆(λ) := L(eλ· I) − λI.f. Even if f is analytic.f. Rn ) is the Banach space of continuous mappings from [−r. 2. Let A be the inﬁnitesimal generator for the C0 semigroup deﬁned by the ﬂow of (2. ϕ) = ψ(0)ϕ(0) − ψ(ξ − θ)dη(θ)ϕ(ξ)dξ. ·) in C ∗ × C. However.2).2. −r 0 . ˙ where D(A) = {ϕ ∈ C : ϕ˙ ∈ C. (2. We recall that A has only the point spectrum.2) and its characteristic equation det∆(λ) = 0. (2.
ϕ) = 0 for all ψ ∈ P ∗ }. such that (Ψ. L(ϕ) = −r A∗ of A is deﬁned as the inﬁnitesimal genThe formal adjoint operator erator for the solution operator of the adjoint equation in C ∗ . In BC. there exists a p×p real matrix B with σ(B) = Λ. An adequate phase space to accomplish this is the space BC (see [30. y(t) ˙ =− −r Fix a nonempty ﬁnite set Λ = {λ1 . (I is the n × n identity matrix). and the dual space P ∗ is the generalized eigenspace for A∗ associated with the eigenvalues in Λ. Eq. ∃ lim ψ(θ) ∈ Rn }. 44. t ≤ 0. .232 DELAY DIFFERENTIAL EQUATIONS where η : [−r. Chapter 7]). where P is the generalized eigenspace associated with the eigenvalues in Λ. −r ≤ θ < 0 X0 (θ) = I. . ˙ for θ ∈ [−r. Using the formal adjoint theory of Hale (see [58. λs } ⊂ σ(A). For dual bases Φ and Ψ of P and P ∗ respectively. The elements of BC have the form ψ = ϕ+X0 α. 0] → Rn × Rn is a function of bounded variation such that 0 dη(θ)ϕ(θ). Q = {ϕ ∈ C : (ψ. . 0). ˙ = ΦB and −Ψ ˙ = BΨ. p = dim P . θ = 0. α ∈ Rn . ϕ ∈ C. ﬁrst it is necessary to enlarge the phase space C in such a way that (2.1) is written as an abstract ODE. Let v(t) = ut ∈ C. where 3 0. 0] → Rn  ψ is continuous on [−r.2 The enlarged phase space To develop a normal form theory for FDEs.1) takes the form dv dt (0)= L(v) + F (v) dv dt (θ)= v(θ). Φ) = Ip . . 0). 0 y(t − θ)dη(θ). (2. 45]) deﬁned by BC := {ψ : [−r. that satisﬁes simultaneously Φ 2. the phase space C is decomposed by Λ as C = P ⊕ Q. so that BC is identiﬁed with C × Rn with the norm ϕ + X0 α = ϕC + αRn . . θ→0− with the sup norm.
X0 α) = ψ(0)α. v = Φx + y. y ∈ C 1 ∩ Ker π = C 1 ∩ Q := Q1 . ϕ ∈ C. π is a continuous projection on BC that commutes with ˜ = C 1.Normal Forms And Bifurcations For DDE 233 with the notation dv ˙ θ ∈ [−r.3) is decomposed as a system of abstract ODEs in Rp × Ker π ≡ BC.1) as where D(A) dv ˜ + X0 F (v).3) becomes π A˜ = Aπ Φ dx dy ˜ ˜ + πX0 F (Φx + y) + (I − π)X0 F (Φx + y) + = AΦx + Ay dt dt ˙ + Ay ˜ + ΦΨ(0)F (Φx + y) + (I − π)X0 F (Φx + y). is therefore extended to a canonical projection π : BC → P. ˜ = ϕ˙ + X0 [L(ϕ) − ϕ(0)]. Deﬁne an extension of the dθ (θ) = v(θ). We then write (2. α ∈ Rn . x ∈ Rp . Aϕ ˙ ˜ = C 1 := {ϕ ∈ C  ϕ˙ ∈ C}. α ∈ Rn . ˜ inﬁnitesimal generator A.5) Writing v ∈ C 1 according to the above decomposition. . 0]. with x ∈ Rp (p = dim P ). ϕ). The canonical projection of C onto P associated with the decomposition C = P ⊕ Q. ϕ) + Ψ(0)α .3) which is the abstract ODE in BC associated with (2. ϕ → ϕP = Φ(Ψ.1. (2. denoted by A. the decomposition C = P ⊕ Q yields a decomposition of BC by Λ as the topological direct sum BC = P ⊕ Ker π. Eq. (2. we also extend the duality (·. π(ϕ + X0 α) = Φ (Ψ.6) y˙ = AQ1 y + (I − π)X0 F (Φx + y). where AQ1 is the restriction of A˜ to Q1 interpreted as an operator acting in the Banach space Ker π. to prove the following lemma. In a natural way. = Φx Therefore. Lemma 2. = Av dt (2. (ψ. and since ˜ in C 1 . (2.4) ˜ and integration by parts. it is easy By using the deﬁnition of π and A. as x˙ = Bx + Ψ(0)F (Φx + y) (2. ·)C ∗ ×C to a bilinear form on C ∗ × BC by deﬁning ψ ∈ C ∗ . A˜ : C 1 ⊂ BC → BC. A˜ on D(A) Since π is a continuous projection. y ∈ Q1 .1). (2.
which is tangent to the centre space P for (2. x ∈ V. Therefore.f.f. for the ODE giving the ﬂow on such manifold. W c = {ϕ ∈ C : ϕ = Φx + h(x). the deﬁnition of the linear operator AQ1 in the second equation of (2.8) to a n. for equations giving the ﬂow on centre manifolds.2.1).8) The goal now is to develop an algorithm of n. we write the ODE for the ﬂow on W c as x˙ = Bx + Ψ(0)F (Φx + h(x)). .7) where V is a neighbourhood of 0 ∈ Rp . (2.2) associated with the eigenvalues in Λ.6) was not adequate. the spectrum of AQ1 plays an important role in the construction of normal forms. Consider the formal Taylor expansion of F .6) looks like (1.13).e.. we turn our attention to the case where Λ is the set of eigenvalues of A on the imaginary axis. and for the sake of exposition.2) at zero. F (v) = 1 Fj (v). we can prove [44] that Lemma 2. Λ = {λ ∈ σ(A) : Re λ = 0} = ∅.234 DELAY DIFFERENTIAL EQUATIONS Note that the linear part of system (2. up to a ﬁnite order k. where P is the generalized eigenspace of (2. Chapter 10] and references therein) assures that there is a pdimensional invariant manifold for (2. j! j≥2 v ∈ C. the ﬁnite dimensional variable x and the inﬁnite dimensional variable y are decoupled in the linear part of the equation. so it is crucial to restrict A˜ to a linear operator in the Banach space Ker π. 2. σ(AQ1 ) = σP (AQ1 ) = σ(A) \ Λ. In this case. (2. (2.. We also point out that the enlarged phase space BC was used by Chow and MalletParet [30] to derive system (2. we can consider any nonempty ﬁnite set Λ ⊂ σ(A) for which there is an invariant manifold tangent to P at zero.f. which aﬀects simultaneously the linearization of the centre manifold and reduces the ODE (2. In fact. i. for x ∈ V }. In this setting. However. In fact.f. since the decomposition (2.6). the method described here is applied to compute a n. we are particularly interested in obtaining n. as we will mention.3 Normal form construction In applications.5) was not explicitly given in [30]. From Eq.6). The centre manifold theorem ([58.
y) − [DUj1 (x)Bx − BUj1 (x)] gj2 (x. where we denote by f˜j = (f˜j1 . Section 1): for j. are obtained by a recursive process of changes of variables. y) obtained after the previous transformations of variables. y) = f˜j2 (x. xp ). with the norm  q=j cq xq  = q=j cq Y . where we adopt the following notation (cf. y). n. x ¯ ∈ Rp .9) with fj := (fj1 . . y¯) + 1 1 (U (¯ x). y¯). y) + h. fj2 ). y) + j! fj (x.t. with coeﬃcients in Y . f˜j2 ) the terms of order j in (x. (2.12) x. p ∈ N and Y a normed space. y) y˙ = AQ1 y + j≥2 j!1 fj2 (x. y ∈ Q1 . j ≥ 2.9) into the normal form x. 235 (2. . cq ∈ Y }. deﬁned by fj1 (x. Assume that steps of orders 2. Eq.11j ) where x. j − 1 have already been performed. This recursive process transforms (2. . That is.t.10) As for autonomous ODEs in Rn . leading to 1 ˜1 1 1 x˙ = Bx + j−1 =2 ! g (x.. x ∈ Rp . y) + h. . y¯˙ = AQ1 y¯ + j≥2 j!1 gj2 (¯ where gj := (gj1 . Uj2 : Rp → Q1 are homogeneous polynomials of degree j in x ¯. 1 ˜2 1 2 y¯˙ = AQ1 y + j−1 =2 ! g (x. Uj2 (¯ x)). the terms of order j ≥ 2 are computed from the terms of the same order and from the terms of lower orders already computed in previous steps. .f. j ≥ 2. At a step j. j! j (2. x = (x1 . y) = f˜j1 (x.6) becomes x˙ = Bx + j≥2 j!1 fj1 (x. y¯) x ¯˙ = B x ¯ + j≥2 j!1 gj1 (¯ (2. y) = (¯ x. y) = (I − π)X0 Fj (Φx + y). y) + j! fj (x. given by gj1 (x. y. Vjp (Y ) denotes the space of homogeneous polynomials of degree j in p variables.o. . eﬀect a change of variables that has the form (x. Vjp (Y ) = { q=j cq xq : q ∈ Np0 . (2. y) = Ψ(0)Fj (Φx + y). .Normal Forms And Bifurcations For DDE where Fj is jth Fr´echet derivative of F . Uj2 ∈ Vjp (Q1 ). y) − [DUj2 (x)Bx − AQ1 (Uj2 (x))]. gj2 ) are the new terms of order j. .o. fj2 (x. At step j. y¯ ∈ Q1 and Uj1 : Rp → Rp . Uj1 ∈ Vjp (Rp ). .
y¯) + O(¯ xk ) be the change of variables obtained after the sequence of transformations (2. with h(¯ x) = o(¯ xk ). Pj2 ) is the canonical projection of Vjp (Rp ) × Vjp (Ker π) onto Im Mj . so for y = 0 one can deﬁne gj (x. c Consequently the ﬂow on W is given by (after dropping the bars) x˙ = Bx + k gj1 (x. We are particularly interested in the situation of Mj2 surjective and onetoone. y) = Θ(¯ x.3. Assume that the operators Mj2 . where V is a neighbourhood of 0 ∈ Rp .f. The remaining terms (resonant terms) are in general not determined in a unique way. o) = (I − Pj )f˜j (x. The operators Mj1 .7) is transformed into c W = {ϕ ∈ C : ϕ = Φ¯ x + h(¯ x).f. j ≥ 2. Clearly.11j ). depending on the choices of complementary spaces Im (Mj )c for Im (Mj ). are onetoone and onto. y) = f˜j (x. (in the usual sense of n. .15) is in n. Mj2 : Vjp (Q1 ) ⊂ Vjp (Ker π) → Vjp (Ker π) (Mj2 h2 )(x) = Dx h2 (x)Bx − AQ1 (h2 (x)).f. deﬁned by the Lie brackets. for ﬁnitedimensional ODEs. (2. called nonresonant terms. (Mj1 h1 )(x) = Dh1 (x)Bx − Bh1 (x). 0). y) = Θ(¯ x. y¯) = (¯ x. Uj2 ). 0) + o(xk ).15) j=2 and the change (x. The inﬁnitedimensional part in the transformation formulas is handled through the operators Mj2 . the local centre manifold W c for (2.13) With Uj = (Uj1 . are the operators that appear in Section 1 for computing n. 2 ≤ j ≤ k.1) deﬁned in (2.236 DELAY DIFFERENTIAL EQUATIONS Deﬁne the operators Mj = (Mj1 . y) − Mj Uj (x). by Mj1 : Vjp (Rp ) → Vjp (Rp ). Let (x. the ranges of Mj1 . x∈V. it is clear that gj (x. for x ¯ ∈ V }. Mj2 ). Then.14) where I is the identity operator on Vjp (Rp ) × Vjp (Ker π) and Pj = (Pj1 . Theorem 2. j ≥ 2. for ODEs). 2 ≤ j ≤ k. Mj2 contain exactly the terms that can be removed from the equation. y¯) can be chosen so that (2. (2. (2.
¯ = (λ1 . 237 (2.8) up to second order terms. y¯) + h. . this implies that h2 = 0. k 1 2 y¯˙ = AQ1 y¯ + j=2 j! gj (¯ x. Since M22 is onetoone. .o. The spectra of the operators Mj2 are given in the theorem below.. . λ) Q (ii) σ(Mj2 ) = σP (Mj2 ). With the above notations. Theorem 2. j ≥ 2. 2 2 2 From (2. multiplicities. equation 1 x. . M22 h2 = 0. that is. y¯) + h. for ODEs in Rp . q ∈ Dj }. are closed operators in the Banach space Vjp (Ker π). it is straightforward to prove that Mj2 . where: Dj = {q ∈ Np0 : q = j}. so Mj2 are onetoone and onto if and only if 0 ∈ σ(Mj2 ). we have 1 1 1 Dh2 (x)[Bx + O(x2 )] + O(x3 ) = AQ1 (h2 (x)) + g22 (x. and thus y = o(x2 ). . λp are the eigenvalues in Λ. (2.16). since the operators Mj1 coincide with the operators in the algorithm for computing n. Let Eq. because M22 is onto. λp ) where λ1 .16) we get 1 Dh(x)x˙ = AQ1 (h(x)) + g22 (x. . (See also [45] for the situation of DDEs with parameters. counting λ ¯ = q 1 λ1 + · · · + qp λp .4. The complete proof of this theorem can be found in [44]. . Write h(x) = 12 h2 (x) + o(x2 ). qp ). 0) + h.t x ¯˙ = B x ¯ + g21 (¯ 2 is a n. For y = h(x).Normal Forms And Bifurcations For DDE Proof.o. Hence Dh2 (x)Bx − AQ1 (h2 (x)) = 0. 0) + O(x3 ).6) be transformed into x ¯˙ = B x ¯ + kj=2 j!1 gj1 (¯ x.) Here we give only a sketch of the proof. q = q1 + · · · + qp for q = (q1 . . The rest of the proof follows by induction.t. λ) Proof. y) = Θ(¯ x. . Thus. it is a closed operator. from (2. we conclude that g22 (x. . .t. ¯ − µ : µ ∈ σ(A1 ). Since A is the inﬁnitesimal generator of a C0 semigroup. This implies that the extension A˜ is also a closed oper˜ Q1 is an operator acting in the ator. .o. (q.f. therefore AQ1 is also closed.14). for (2. (i) σP (Mj2 ) = {(q. On the other hand. 0) = 0. Note that the restriction AQ1 = A Banach space Ker π. h(x)) + O(x4 ). . y¯).f.16) through the change (x. 2 Using again (2.
BC ≡ Rp × Ker π. . proving that αI − Mj2 is surjective. Chapter 12]. qp ) and = (1 . λ)−α ∈ ρ(AQ1 ) p q for all q ∈ Dj . there is a formal change of variables of the form (x. y) = (¯ x. .18) . Conditions (2. 0. 0.2 it follows immediately that 0 ∈ σP (Mj2 ) = σ(Mj2 ) if and only if (2. Fix g = q∈Dj gq x ∈ Vj (Ker π). (2.17j ) are called nonresonance conditions of order j. such that: x. . . In this order. . . . relative to the set Λ. . . one can prove (i).13) are onetoone and onto if and only if ¯ = µ. y¯) + O(¯ x2 ). .f. From (i).6. we say that q < if the ﬁrst nonzero diﬀerence q1 − 1 . 0) = 0. for any q ∈ Dj it is possible to construct h(q) = h x ∈Dj such that [(αI − Mj2 )h(q) ] = g for ≤ q. j ≥ 2. Corollary 2. 0) is the ﬁrst element in Dj . we get (αI − Mj2 )h(q) = g. it is suﬃcient to show that if α ∈ σP (Mj2 ). then the operator αI − Mj2 is surjective. for ODEs). The operators Mj2 deﬁned in (2.238 DELAY DIFFERENTIAL EQUATIONS By adapting the algebraic arguments for ﬁnite ODEs in [29. . (j. . 0). . p ) in Dj . By induction in Dj ordered as above. . For q the last element of Dj .9) is transformed into Eq. In order to prove this. consider the set Dj with the following order: for q = (q1 . (2.4 and Lemma 2. . λ) for all µ ∈ σ(A) \ Λ and q ∈ Dj .17j ) Proof. .5. j) is the last element. . (iii) the ﬂow on it is given by the ODE in Rp x ¯˙ = B x ¯+ 1 g 1 (¯ x.17j ) holds. From Theorem 2. qp − p is positive. To prove (ii). (2. we deduce that µq := (q. and (0. Theorem 2. . (i) Eq. (2. .12) where gj2 (¯ (ii) a local centre manifold for this system at zero satisﬁes y¯ = 0. ¯ Let α ∈ σP (Mj2 ). Then. Let Λ = {λ ∈ σ(A) : Re λ = 0} = ∅. . (q. and BC be decomposed by Λ. j! j j≥2 which is in normal form (in the usual sense of n.
f. Uj2 (x) = (Mj2 )−1 f˜j2 (x. 2 ≤ j ≤ k. (2. The terms gj1 (x. If the nonresonance conditions (2. Let Λ ⊂ σ(A) be a nonempty ﬁnite set. we have f˜21 = f21 . For the choice Λ = {λ ∈ σ(A) : Re λ = 0}.. with Uj = (Uj1 . and (Mj1 )−1 : Im(Mj1 ) → Ker (Mj1 )c is the right inverse of Mj1 for a decomposition Vjp (Rp ) = Ker (Mj1 ) ⊕ Ker (Mj1 )c . Theorem 2. for ﬁnite dimensional ODEs: i.e.f. procedure is justiﬁed if the nonresonance conditions hold. Eq. 0) = (I − Pj1 )f˜j1 (x. From (2. the statements in the above theorem are valid for the invariant manifold for (2. Remark 2. from which we .9) (or Eq. j ≥ 2. clearly the nonresonance conditions (2. are satisﬁed.17j ) hold for j ≥ 2. we can choose Uj2 such that gj2 (x. 0). (2. 0) in the n. On the other hand.f.12) is said to be a normal form for Eq.1) associated with Λ. (2. 0) and Uj1 (j ≥ 2) are chosen according to the method of n. y) = f˜j (x.3 implies that the centre manifold is given by equation y¯ = 0 (up to a certain order k.16). j ≥ 2. respectively. Therefore.8. relative to Λ up to korder terms. j ≥ 2. Note that Pj1 and (Mj1 )−1 depend on the choices of complementary spaces for Im (Mj1 ) and Ker (Mj1 ) in Vjp (Rp ).14). and the nonresonance conditions of order j. y) − Mj Uj (x). provided that this manifold exists. gj1 (x. Im (Mj2 ) = Vjp (Ker π) and Ker (Mj2 ) = {0}.f. gj2 ) are deﬁned by gj (x. according to the following scheme: in the ﬁrst step (j = 2). Deﬁnition 2. and that the ﬂow on this manifold is given by (2.1)) relative to Λ if gj = (gj1 . are recursively computed in terms of the coeﬃcients of the original FDE. hold. if F is C k smooth). where Pj1 : Vjp (Rp ) → Im(Mj1 ) is the canonical projection associated with a decomposition Vjp (Rp ) = Im(Mj1 ) ⊕ Im(Mj1 )c . 0).Normal Forms And Bifurcations For DDE 239 Proof. Theorem 2.17j ) hold for any j ≥ 2. 0) = f˜j2 (x. which is an ODE in n. (2. Remark 2.17j ). For other choices of ﬁnite nonempty sets Λ ⊂ σ(A)..18).9. this n. if F ∈ C k for some k ≥ 2. Uj2 ). 0) − (Mj2 Uj2 )(x) = 0. f. Uj1 (x) = (Mj1 )−1 Pj1 f˜j1 (x. If the nonresonance conditions (2.7. a n. one can compute Eq.10.
0). 0) = (I − P21 )f21 (x. where α ∈ V. etc.20) u(t) ˙ = L(α)(ut ) + F (ut . (2.4 Equations with parameters For studying bifurcation problems.20) as u(t) ˙ = L0 (ut ) + (L(α) − L0 )(ut ) + F (ut . α ∈ V. Its linearization around the origin is u(t) ˙ L0 (ut ) . Introducing the parameter α as a variable by adding α˙ = 0. 2 3! the term of order j in the variables (u.19) from which we ﬁnally obtain g31 (x. Writing the Taylor expansions 1 L(α)(u) = L0 u + L1 (α)u + L2 (α)u + · · · 2 1 1 F (u. Rn ). As usual. and let C = P ⊕ Q be decomposed by Λ. 0]. = 0 α(t) ˙ Note that the term (L(α) − L0 )(ut ) is no longer of the ﬁrst order. with the operator norm. 0)U21 (x) + (Dy f21 )(x. for all α ∈ V . Rn ) denotes the space of bounded linear operators from C to R. L(C. we need to consider situations with parameters: (2. Clearly the phase space for (2. 0)U22 (x) − (Dx U21 )(x)g21 (x. 2 (2. and 3 f˜31 (x. α) (α(t) ˙ = 0). α) for (2. k ≥ 2. α).21) is C([−r.21) where L0 := L(0). α) = 0. 2. we must compute U21 (x) = (M21 )−1 P21 f21 (x. α) + · · · . α) = F2 (u. 0). (j − 1)! j! .21) is given by 1 1 Lj−1 (α)u + Fj (u. 0) = (I − P31 )f˜31 (x. and L : V → L(C. V is a neighbourhood of zero in Rm . u ∈ C. 0)]. 0). F : C × V → Rn are respectively C k−1 and C k functions. α). α) + F3 (u. Rn+m ). 0) = f31 (x. 0). U22 (x) = (M22 )−1 f22 (x.240 DELAY DIFFERENTIAL EQUATIONS compute g21 (x. F (0. we write (2. in the second step (j = 3). D1 F (0. α) = 0. 0) + [(Dx f21 )(x. since α is taken as a variable. Consider a ﬁnite nonempty set Λ ⊂ σ(A) such that 0 ∈ Λ if 0 ∈ σ(A).
10) for the equation (2. Λ as the set of eigenvalues of A on the where D(A) imaginary axis. y.5) yields x˙ = Bx + Ψ(0)[F (Φx + y) + G (Φx + y)(Φx˙ + y)] ˙ y˙ = AQ1 y + (I − π)X0 [F (Φx + y) + G (Φx + y)(Φx˙ + y)]. α)].f. and G.24) ˜ = C 1 . L are bounded linear operators from C to Rn .4) still commutes in C 1 with the operator A˜ given by (2. dt where ut ∈ C. µ(0) = 0. fj2 (x. on centre . D. with 0 Dφ = φ(0) − −r d[µ(θ)]φ(θ) and µ is nonatomic at zero.5 More about normal forms for FDEs in Rn In [99]. 2. the terms (2.24). y ∈ Q1 .f. ˙ x ∈ Rp . in [99] the inﬁnitesimal generator general A for the semigroup deﬁned by the solutions of the linear equation d D(ut ) = L(ut ) dt was naturally extended to A˜ : C 1 ⊂ BC → BC. k ≥ 2. Since the projection π in (2.23) D(ut ) − G(ut ) = L(ut ) + F (ut ). (2.24). α) = Ψ(0)[jLj−1 (α)(Φx + y) + Fj (Φx + y. and consider C = P ⊕ Q and BC = P ⊕ Ker π decomposed by Λ. Following the ideas above.1 and 2.22) The above procedure can now be applied to compute n. α)]. α) = (I − π)X0 [jLj−1 (α)(Φx + y) + Fj (Φx + y. where AQ1 is the restriction of A˜ to Q1 interpreted as an operator acting in the Banach space Ker π. y. F are C k functions . where now A˜ is as in (2. the n.f.11). Aϕ ˙ (2. The method for computing n.2 are true in this framework.23) is written as the abstract ODE ˜ + X0 [F (v) + G (v)v]. ˙ v˙ = Av and decomposition (2. ˜ = ϕ˙ + X0 [L(ϕ) − Dϕ].6) in BC read now as fj1 (x. See [45] for details. Mj2 by the formal expression (2. Deﬁne again the operators Mj1 . for (2. Lemmas 2.g.241 Normal Forms And Bifurcations For DDE Therefore. in BC (2. Fix e.13). theory presented here was generalized to construct n.f. for neutral FDEs of the form d (2.
λ) for all µ ∈ σ(A) \ Λ. the normal form on the centre manifold for (2. It would also be very important to generalize the above method to construct n. ϕ ∈ C. and used to deduce a natural decomposition of C. If the new coordinates allow us to decompose also the enlarged phase space BC. j ≥ 2 are satisﬁed. (2. around periodic orbits for autonomous retarded FDEs on C = C([−r. (2. it may be useful to use the work of Hale and Weedermann [59].242 DELAY DIFFERENTIAL EQUATIONS manifolds (or other invariant manifolds) for neutral equations (2.26). The major diﬃculty when trying to generalize the theory of n. . In [35]. q ∈ Dj . (2. xt ).26) is that its linear part is nonautonomous. condition L(t) ≡ L is very restrictive. the theory was developed for the particular case of equations (2.27) Clearly.26) with autonomous linear part. to an FDE of the form (2.26) where L(t) = Df (pt ) and F (t. ϕ)dt. xt ). However. where 1 F0 (ϕ) = ω ω F (t. (q. Through the change u(t) = p(t) + x(t). If the nonresonance conditions ¯ + ik = µ. With the above notations. let Λ = {λ ∈ σ(A) : Re λ = 0} and assume Λ is nonempty.26) can be written as an abstract ODE in BC. Suppose that p(t) is an ωperiodic solution of (2. k ∈ Z.25) becomes x(t) ˙ = L(t)xt + F (t. in this case.1). 0 To treat the general periodic case (2. ϕ) = f (pt + ϕ) − f (pt ) − Df (pt ) are ωperiodic functions in t. where a suitable system of coordinates was established. then (2.25) with f ∈ C k (k ≥ 2).27) coincides with the normal form for the averaged equation x(t) ˙ = Lxt + F0 (xt ). Theorem 2.f. the work in [35] provides a strong result.23) is then developed in [99] along the lines of the method for FDEs (2.25). 0]. (2. Rn ) u(t) ˙ = f (ut ).11.f. x(t) ˙ = Lxt + F (t.
In C = C([−r. Let L(0) = L0 . P ∗ = span Ψ. 0 1 B= .1 The BogdanovTakens bifurcation We start by studying the BogdanovTakens singularity for a general scalar FDE. r]. with one or two discrete delays. ψ1 (0) − sψ2 (0) . ˙ = ΦB.f. 0]. α = 0. For the linearization at u = 0. θ ∈ [−r. 0 0 (3. 103]. 40. 47. 102. 0]. 57. BC = P ⊕ Ker π be decomposed by Λ. with (Ψ. and for its dual P ∗ : Φ(θ) = (1. ϕ ∈ C. α).g. With the notations in Section 2. (3.1) where α = (α1 . α2 ) ∈ V ⊂ R2 . 243 Normal forms and Bifurcation Problems In this section we illustrate the application of the method of n. consider u(t) ˙ = L(α)ut + F (ut . In the framework of FDEs in Rn .f. 41. α) = 0 for α∈V. F : C × V → R is C 2 . for other examples of applications of n. L : V → L(C. L0 (θ2 ) = 0. Φ ψ2 (0) = −L0 (θ2 /2)−1 . we assume: (H1) λ = 0 is a double characteristic value of u(t) ˙ = L0 ut : L0 (1) = 0. to the study of BogdanovTakens and Hopf bifurcations for general scalar FDEs. R). Φ) = I if ψ1 (0) = L0 (θ2 /2)−2 L0 (θ3 /3!). 3. L0 (θ) = 1. Ψ(s) = ψ2 (0) P = span Φ. (H2) all other characteristic values of u(t) ˙ = L0 ut have nonzero real parts Above and throughout this section. to bifurcation problems see e. α) = 0.Normal Forms And Bifurcations For DDE 3. θ). D1 F (0.2) . we often abuse the notation and write L0 (ϕ(θ)) for L0 (ϕ). R) is C 1 . Choose bases Φ and Ψ for the the centre space P of u(t) ˙ = L0 ut . V a neighbourhood of zero. [38. ﬁx Λ = {0}. s ∈ [0. and let C = P ⊕ Q. F (0. n ≥ 2. −Ψ ˙ = BΨ for Also.
On the other hand. 0 0 0 0 0 0 = span . from the deﬁnition of M21 and the choice of S. Consider the Taylor expansions L(α) = L0 + L1 (α) + O(α2 ). 0. . . α) = F2 (Φx. . 0.. α) = Ψ(0)[2L1 (α)(Φx) + F2 (Φx. ψ2 (0) 2 Write F2 (Φx. x21 x1 x2 x1 α1 x1 α2 x2 α1 x2 α2 .1) x1 x2 + A(0. for α = 0 the 2dimensional ODE on the centre manifold has a BogdanovTakens singularity.. 0)}.6 (see also Remark 2. x2 ) ∈ R . 2 From Theorem 2.13).. . . and (conf. α) = 2 2 1 ψ1 (0) {L1 (α)(ϕ1 x1 + ϕ2 x2 ) + F2 (ϕ1 x1 + ϕ2 x2 . α1 α2 α22 α12 Note that 1 1 f (x. 0 0 0 . .o. f21 (x. ∂p2 p2 p 2 ∂x1 x2 Decomposing V24 (R2 ) = Im (M21 ) ⊕ Im (M21 )c . 29].4 . .2) x22 . α). 0. α) = F2 (u. M2 = ∈ V24 (R2 ). . .9). .22)) g21 (x. a possible choice for Im (M21 )c is 3 S := Im (M21 )c .244 DELAY DIFFERENTIAL EQUATIONS Thus. 2 2 where x = (x1 . we have . . α) + h. 0. 0) = A(2. . 0.∂p1 x − p p1 2 2 1 p1 ∂x 1 . we deduce that the projection on S of the elements of the canonical basis . α) + O(u3 + αu2 ). up to quadratic terms on the centre manifold of the origin near α = 0 is given by 1 x˙ = Bx + g21 (x. α)]. α) = P rojIm (M21 )c f21 (x. . (2. .t.0) x21 + A(1. From the deﬁnition of M21 in (2.f. 1 F (u. Consequently its dynamics around u = 0 are generically determined by its quadratic terms [24. we deduce that the n.
. .0) + ψ2 (0)A(1. For α small. . We summarize the above computations in the following result: Theorem 3.3) and λ1 =ψ2 (0)L1 (α)(1) λ2 =ψ1 (0)L1 (α)(1) + ψ2 (0)L1 (α)(θ).5) x˙ 2 = λ1 x1 + λ2 x2 + B1 x21 + B2 x1 x2 + h.1) is given by the n. B2 . 2 x1 x2 x1 αi x2 αi x2 x1 (i = 1. the ﬂow on the centre manifold for (3. λ2 are linearly independent.t (3. let B1 < 0 and B2 > 0. (3. The bifurcation diagram for the ﬂow on the centre manifold depends on the signs of B1 . . 0 0 0 0 0 . 2). . 2) are. . λ2 and the coeﬃcients B1 . Furthermore. 2 2 x2 αi x1 x2 x1 αi x2 x1 . α = 0. f. B1 x21 + B2 x1 x2 1 B1 = ψ2 (0)A(2.4) We remark that the bifurcation parameters λ1 .1. α) = 2 2 with 0 λ1 x1 + λ2 x2 0 + . if B1 B2 = 0. .0) 2 1 B2 = ψ1 (0)A(2.o. B2 and the new bifurcating parameters λ1 . .4). 2) (i = 1. B2 are explicitly computed in terms of the coeﬃcients in the original equation (3. . the new bifurcating parameters λ1 . For instance. (i = 1. λ2 are as in (3. . (up to quadratic terms) x˙ 1 = x2 + h. 2) . 0 0 2x1 x2 0 x2 αi 0 0 0 0 0 . Also. (3. x1 x2 0 x1 αi x2 αi x21 Hence 1 1 g (x. .1) 2 (3. . .245 Normal Forms And Bifurcations For DDE for V24 (R2 ). 0 0 0 0 0 0 0 0 0 0 . Then in .3) and (3.o. where the coeﬃcients B1 . 2 .t.1). respectively. (i = 1. 0.5) exhibits a generic BogdanovTakens bifurcation from u = 0. .
21). c ∈ R.7) where α ∈ V . V a neighbourhood of zero in R. Example 3. 0)/2 = (a+b+c)x21 −(b+2c)x1 x2 + cx22 . α2 = a2 − 1. (3. ∆ (0) = 0 and ∆ (0) = 0. a. L1 (α)(ϕ) = α1 ϕ(0) − α2 ϕ(−1) and F (ϕ) = aϕ2 (0) + bϕ(0)ϕ(−1) + cϕ(−1)2 . λ2 < 0. we deduce that the ODE on the centre manifold of the origin near α = 0 is given in n. Consider the scalar DDE u(t) ˙ = a1 u(t) − a2 u(t − 1) + au2 (t) + bu(t)u(t − 1) + cu2 (t − 1).2. ∆(0) = 0. α2 ) ∈ R2 . We get ψ1 (0) = 2/3. and nonlinear terms L1 (α)(ut ) + F (ut ). α) ∈ C × R2 given by L0 (ut ). a2 . by (3. L : V → L(C. with the linear terms in (ut . 0]. One can check that λ = 0 is a double characteristic value.4). In the phase space C = C([−r. and consider dual basis Φ. b.g. and F2 (Φx. Fix Λ = {0}. α). Ψ as above. . is given by ∆(λ) := λ − a1 + a2 e−λ = 0. R) is C 2 . R). i. F (0. (3.2). so it has the form (2. α) = 0. The characteristic equation for its linearization at zero.5) with B1 = 2(a + b + c). Applying formulas (3. λ1 = 2(α1 − α2 ).2 2 B2 = (2a − b − 4c) 3 2 λ2 = (α1 + 2α2 ). for ϕ ∈ C = C([−1. 3. if and only if a1 = a2 = 1. α = (α1 .f. (See e. where L0 (ϕ) = ϕ(0) − ϕ(−1). and F : C × V → R is C 3 . ψ2 (0) = 2 from (3.6) where a1 . 3 Hopf bifurcation We study the generic Hopf bifurcation for the general scalar case. consider u(t) ˙ = L(α)ut + F (ut .e. 0].. α) = 0 for α ∈ V .6) becomes u(t) ˙ = u(t)−u(t−1)+α1 u(t)−α2 u(t−1)+au2 (t)+bu(t)u(t−1)+cu2 (t−1). Rescaling the parameters by setting α1 = a1 − 1.3)(3. [24. the Hopf bifurcation curve H and the homoclinic bifurcation curve HL lie in the region λ1 > 0. and both the homoclinic loop and the periodic orbit are asymptotically stable.246 DELAY DIFFERENTIAL EQUATIONS the (λ1 . u(t) ˙ = a1 u(t) − a2 u(t − 1). D1 F (0. with H to the left of HL. R). λ2 )bifurcation diagram. 29]). (3.
on the center manifold up to third order terms given by 1 1 x˙ = Bx + g21 (x.f. (H2) for α = 0. α)+h.t. α)+ F3 (ϕ. ϕ2 (θ)) = (eiωθ . = ψ2 (s) ψ2 (0)eiωs P = span Φ. the operators Mj1 have a diagonal matrix representation in the canonical basis of Vj3 (C2 ). −iω} as C = P ⊕ Q. Write the Taylor expansions L(α) = L0 +αL1 + α2 1 1 L2 +h. j ≥ 2. γ (0) = 0 (Hopf condition). (3. Vj3 (C2 ) = Im (Mj1 ) ⊕ Ker (Mj1 )..o. Φ) = I if ψ1 (0) = (1 − L0 (θeiωθ ))−1 . α) + g31 (x. where the dual bases Φ. 0. ω := ω(0) > 0. Applying the method of n. Ψ for P.o.8) ˙ = ΦB. C is decomposed by Λ = {iω. assume that: (H1) there is a pair of simple characteristic roots γ(α)±iω(α) for u(t) ˙ = L(α)ut crossing transversely the imaginary axis at α = 0: γ(0) = 0. α) = F2 (ϕ. 2 3! . 2 2 3! The Hopf condition assures that a Hopf bifurcation occurs on a two dimensional centre manifold of the origin. −Ψ ˙ = BΨ for Note that Φ iω 0 B= .247 Normal Forms And Bifurcations For DDE For u(t) ˙ = L(α)(ut )..t.. we have the n. in particular. α) + h. with (Ψ.. ψ2 (0) = ψ1 (0). e−iωθ ). −r ≤ θ ≤ 0 0 ≤ s ≤ r. With the notations above. there are no other characteristic roots with zero real parts. F (ϕ.t.f. P ∗ can be chosen as Φ(θ) = (ϕ1 (θ). ψ1 (0)e−iωs ψ1 (s) ∗ P = span Ψ. and using complex coordinates. 0. Ψ(s) = . The generic Hopf bifurcation is determined up to cubic terms. 0 −iω Since B is a diagonal matrix.o. because of the use of complex coordinates.
0) x21 + A(1.2) x1 x22 + A(0. α) = Ψ(0)[ L2 (Φx + y) + F3 (Φx + y. α) = 2 2 1 1 g (x. q2 ) ∈ N20 . y.9) . = j.1) x21 x2 + A(1. α) = F2 (Φx. α)]. Hence. g31 are the second and third order terms in (x. . . 0. 0) = A(2. for l + q1 + q2 1 0 e1 = . α). for S := span 0 x1 x22 1 1 1 f2 (x. α) = 0 for all α ∈ R. y. Thus. α)]. . D1 F (0. Ker (M21 ) = span 0 x2 α 4 3 2 0 0 x1 α2 x1 x2 1 Ker (M3 ) = span . α) = P rojS f˜31 (x. α). are given by 1 1 B1 x1 α g (x. 3! The operators Mj1 are deﬁned by Mj1 (αl xq ek ) = iω(q1 − q2 + (−1)k )αl xq ek . 0. k = 1. 0. 2 1 P rojKer(M31 ) f˜31 (x. j = 2. α) = Ψ(0)[αL1 (Φx + y) + F2 (Φx + y. 1 0 2 0 x1 x2 . α). 0 0 x1 x22 x2 α2 Note that g31 (x. 0) + O(xα2 ). we write F2 (Φx. α) = P rojKer(M31 ) f˜31 (x. B1 x2 α 2 2 with B1 = ψ1 (0)L1 (ϕ1 ). the second order terms in (α. 0) = A(3. x) of the n.1) x1 x2 + A(0.0) x31 + A(2.f. For the present situation. α) = . 3! 3 2 3! Since F (0. q = (q1 . α) = 0. respectively: 1 1 g (x. (3. 0. l ∈ N0 . 0.2) x22 F3 (Φx. α) = 3! 3 1 P rojKer(M21 ) f21 (x. α) = (I − π)X0 [αL1 (Φx + y) + F2 (Φx + y.q1 ) = A(q1 .248 DELAY DIFFERENTIAL EQUATIONS where g21 . 0. 0.q2 ) . 2. 2 2 α2 1 1 1 f (x. 2 2 1 2 1 f2 (x. α)].3) x32 where A(q2 . 3. y. 0 1 4 3 0 x1 α . e2 = . 0. .
11) (3. U21 (x. 0) + (Dy f21 )(x.249 Normal Forms And Bifurcations For DDE Now we determine the cubic terms for α = 0. 0) 1 = iω ψ1 (0)(A(2. 3 2 .1) − ψ1 (0) ( A(2. . α)α=0 is determined by the equation (M22 U22 )(x. 0. 0. 0. 2 (3.1) x1 x2 − 13 A(0.0) x21 − A(1. 0)U21 (x. . 0. 0)U21 (x.10) where U21 (x. From (2. ψ2 (0)( 13 A(2. . .12) .0)  + A(1. 0) = f22 (x. 0) = U22 (x. 0) ψ1 (0)A(2. The elements of the canonical basis for V23 (C2 ) are 2 2 2 x1 x2 x1 α x2 α x1 x2 α . 0 0 0 0 0 0 0 0 0 0 0 0 .2 . 0).2) x22 ) . α)α=0 = (M21 )−1 P rojIm(M21 ) f21 (x. .0) x21 + A(1.2) x1 x22 and P rojS [(Dx f21 )(x. 2 .− 2 . 0) = (M21 )−1 f21 (x. 0)U22 (x. . 0) and U22 (x.0) A(1. 0) 3 + P rojS (Dx f21 )(x. 0.2) x22 ) A few computations show that P rojS f31 (x. 0)] with C1 x21 x2 . . −3 2 .− . 0) = U21 (x. − 1 2 . x1 x2 x1 α x2 α x21 x2 α2 1 M21 given respectively by with images under iω 2 2 2 x x 0 x α x x1 α . . 0. 0) . 0 x1 x2 x1 α α2 x1 x2 Hence.1) x1 x2 − A(0. −2 2 0 0 0 0 0 0 0 0 0 0 0 0 .1) ) .19). 0) = P rojS f31 (x.1) x21 x2 = ψ2 (0)A(1. we have g31 (x. 0. ω 3 (3. . 0. . = C1 x1 x22 i 2 2 2 2 2 C1 = ψ1 (0)A(2. . 0.
0) = C2 x21 x2 . (3. Let H2 : C × C → C be the bilinear symmetric form such that F2 (u. 0). Thus. u). Using the deﬁnition of M22 . where h20 . or.13) . 0). h˙ 20 − 2iωh20 = A(2. h11 = h11 .0) and h˙ 11 = A(1. to ˙ h(x) − Dx h(x)Bx = ΦΨ(0)F2 (Φx. h02 ∈ Q1 . 0. 0) ˙h(x)(0) − L0 h(x) = F2 (Φx. From this system. AQ1 and π. Then.α=0 (h) = 2Ψ(0)H2 (Φx. 0).. 0)U22 (x. 0) is equivalent to Dx h(x)Bx − AQ1 (h(x)) = (I − π)X0 F2 (Φx. 0)]. Deﬁne h = h(x)(θ) by h(x) = U22 (x. 0) is the quadratic form associated with H2 (u.1) ΦΨ(0) h˙ 11 (0) − L0 (h11 ) = A(1. in other words. h). 0) = H2 (u. v)). h11 ) + H2 (ϕ2 .e. we obtain P rojS [(Dy f21 )h](x. 0. and two initial value problems (IVPs). 0. d h. we deduce that equation (M22 h)(x) = f22 (x.250 DELAY DIFFERENTIAL EQUATIONS It remains to determine P rojS [(Dy f21 )(x. and write h(x) = h20 x21 + h11 x1 x2 + h02 x22 . u∈C (i.1) .0) ΦΨ(0) h˙ 20 (0) − L0 (h20 ) = A(2. F2 (u. C 2 x1 x22 with C2 = 2ψ1 (0)[H2 (ϕ1 . we deduce where we use the notation h˙ = dθ h02 = h20 . h11 . (Dy f21 )y=0. h20 )]. α) = F2 (u.
1)] 3! 2L0 (1) ψ1 (0)A 1 (2. 0) = 3! 3 B2 x11 x2 . w2 = ρ sin ξ. Solving the two IVPs above.15) becomes ρ= ˙ K1 αρ + K2 ρ3 + O(α2 ρ + (ρ. α)). (3. where x1 = w1 − iw2 .f.17) A(1.15) with B1 . and the stability of the nontrivial periodic orbits is determined by the sign of K2 (see e. the ﬂow is given in polar coordinates by Eq. + ψ1 (0)e − ψ1 (0)e h11 (θ) = A(1.14) (3.16) ˙ −ω + O((ρ. the n.18) . B2 given by (3.0) −iωθ 2iωθ + Re H (e .10)(3.14). x2 = w1 + iw2 . 0.9). K2 explicitly given in terms of the original FDE as K1 = Re B1 . A Hopf bifurcation occurs from α = 0 on a 2dimensional local centre manifold of the origin. (3. with K1 = Re (ψ1 (0)L1 (eiωθ )) K2 = (3.16). + x˙ = Bx + B1 x2 α B2 x1 x22 B2 = (3.251 Normal Forms And Bifurcations For DDE Hence. 2 2 2iω − L0 (e2iωθ ) (3. Eq. e ) .1) 1 Re [ψ1 (0)A(2.1) + (C1 + C2 ).3. ξ= with the coeﬃcients K1 .g. K2 = Re B2 .13) we get 1 1 g (x. B 2 x1 x22 with 1 1 ψ1 (0)A(2. Recall that the generic Hopf bifurcation corresponds to the situation K2 = 0.1) − L0 (1) iω Theorem 3. Changing to real coordinates w. α)4 ) (3.1) ] − Re [ψ1 (0)H2 (eiωθ . w1 = ρ cos ξ. the direction of the bifurcation is determined by the sign of K1 K2 . from (3. and then to polar coordinates (ρ.0) iω 3iω 2iω − L0 (e2iωθ ) 1 1 iωθ −iωθ . 3! 4 Therefore. On this manifold. ξ). we obtain e2iωθ ψ1 (0)eiωθ ψ1 (0)e−iωθ − − h20 (θ) = A(2. reads as B1 x1 α B2 x21 x2 + O(xα2 + x4 ). (3. [30]).
K1 < 0. u(t) ˙ = −au(t − 1). Set Λ = {iωN . the bifurcating periodic orbits are always stable inside the centre manifold.19) occurs from u = 0. E. we have Λ = {iπ/2. On the other hand. −iωN }.22) > 0. 2 N ∈ N0 . and the Hopf bifurcation is supercritical.16). thus the Hopf bifurcation is subcritical. For ψ(0) N . α) = −aN v(0) v(−1). (3. −iπ/2}. α) + F3 (v.3.20) λ + ae−λ = 0. (3. α) = F2 (v. a = aN .4.20) if and only if a = aN and ω = ωN . α) = −αv(0)v(−1). ψ1 (0) = 1−iω 1+ω 2 N L(α)v = L0 v + L1 (α)v.g. ωN = π + N π. a generic Hopf bifurcation for (3.19) is written as u(t) ˙ = −aN u(t − 1) + αu(t − 1) + F (ut . K1 > 0 if N even. Write the Taylor formulas as in (3. 2 3! where L0 (v) = −aN v(−1).252 DELAY DIFFERENTIAL EQUATIONS Example 3. (3.f. at the ﬁrst bifurcating point a = a0 = π/2 (N = 0). (3. K2 = K1 = 2 4+π 5(4 + π 2 ) . 1 1 F (v. 1 + a2N ωN [(−1)N − 3ωn ] < 0. where aN = (−1)N ωN . Consider the wellknown Wright equation u(t) ˙ = −au(t − 1)[1 + u(t)].8). L1 (α)v = −αv(−1) and 12 F2 (v. With α = a − aN . we obtain the n.. α).19) which has been studied by many authors (see [30. and 2π π(2 − 3π) < 0. Thus. (3. is (3. with aN .21) where F (v. Applying Theorem 3. If N is odd. α) = −(aN + α)v(0)v(−1). There is a pair of (simple) imaginary roots ±iω for (3. F3 (v. thus K1 K2 < 0. on the center manifold of the origin near a = aN given in polar coordinates by Eq. Since K2 < 0 for all N ∈ N0 . α). K2 = 5(1 + a2N ) K1 = for N ∈ N0 . 45] and references therein). The characteristic equation for the linearization around zero.
neural networks.. Equations involving both time delays and spatial diﬀusion have been increasingly used in population dynamics. where v stands for the density of the population spread over the interval [0. y(t) = bx(t) − 1. X). Write ut ∈ C for ut (θ) = u(t + θ). such as Neumann or Dirichlet conditions. x)[1 + v(t. This RDequation is often used as a model in population dynamics. Considering also a spatial variable x ∈ [0. DF (0) = 0. L : C → X is a bounded linear operator. Normal Forms for FDEs in Hilbert Spaces In the following. known as the Hutchinson equation. (4. and other ﬁelds. and consider FDEs in X given in abstract form as u(t) ˙ = AT u(t) + L(ut ) + F (ut ) (t ≥ 0).4) y(t) ˙ = −ay(t − r)[1 + y(t)]. the model becomes x(t) ˙ = ax(t)[1 − bx(t − r)]. x)]. x ∈ (0.e. We are particularly interested in equations (4. Translating the positive equilibrium 1/b to the origin and eﬀecting a scaling. i. should be added. −r ≤ θ ≤ 0. 1]. we obtain the model ∂2v ∂v (t. Example 3. X is a real or complex Hilbert space with inner product < ·.253 Normal Forms And Bifurcations For DDE 4.1). to which boundary conditions. 0] to X with the sup norm. where a is the growth rate of the population and K = 1/b the “carrying capacity”. since they include reactiondiﬀusion (RD) equations with delays appearing in the reaction terms. x) = d 2 − av(t − r. A typical example of a RDequation with delays is the socalled logistic equation with delays and spatial diﬀusion. In order to simplify the notation.1) where AT : D(AT ) ⊂ X → X is a linear operator. we get the Wright equation (conf. d is the diﬀusion rate that measures the internal migration of the population. Taking into account the maturation period r > 0 of the population. and diﬀusion terms. and F : C → X is a C k function (k ≥ 2) such that F (0) = 0. X) (r > 0) is the Banach space of continuous maps from [−r. . disease modelling. L ∈ L(C. In the framework of ODEs. 1]. b. · > and C = C([−r. 0]. 1). ﬁx R as the scalar ﬁeld. and a. ∂t ∂x t > 0. consider the logistic equation x˙ = ax(1 − bx). r have the same meaning as before.
∆(λ)y = λy − AT y − L(eλ· y). {µk }∞ k=1 (H4) the subspaces Bk of C.2) we assume in this section the following hypotheses (see [73. since it almost imposes that L is a scalar multiplication.1) it is possible to associate an FDE in Rn .f. the situation where (H1)(H4) hold will be studied separately in this section since it provides a strong result: eventually under additional conditions.1 DELAY DIFFERENTIAL EQUATIONS Linear FDEs For the linearized equation about the equilibrium zero u(t) ˙ = AT u(t) + L(ut ).2) in Banach spaces see [101. u(t) = eλt y. form an orthonormal basis for X. y ∈ D(AT ) \ {0}. on invariant manifolds for (4. satisfy L(Bk ) ⊂ span{βk }.2) of the exponential type. (4. Nevertheless. k ∈ N.2) if and only if λ. Hypothesis (H4) roughly requires that the operator L does not mix the spatial variations described by eigenvectors βk of AT . Bk := {ϕβk  ϕ ∈ C([−r. Clearly u(t) as above is a nontrivial solution of (4. For a weaker version of (H4) see [36]. y satisfy the characteristic equation ∆(λ)y = 0. we start by looking for solutions of the linear equation (4. method for FDEs in Rn described in Section 2 without further computations. 0].1). This condition is very restrictive in terms of applications. whose n. on invariant manifolds coincide with n. R)}. Chapter 3]. y ∈ D(AT ). normal forms are developed for FDEs in Banach spaces without imposing (H3)(H4). where the characteristic operator ∆(λ) is deﬁned by ∆(λ) : D(AT ) ⊂ X → X. In Section 5 of the present text. This allows us to apply to (4. For a straightforward generalization of these assumptions to equations (4.f.f. up to some order. λ ∈ C. 104]): (H1) AT generates a C0 semigroup of linear operators {T (t)}t≥0 on X. As for retarded FDEs in Rn . (H3) the eigenfunctions {βk }∞ k=1 of AT . 101.1) the n. with corresponding eigenvalues . (H2) T (t) is a compact operator for t > 0. (4.3) . for (4.254 4. 77.
With this identiﬁcation. Φk )k = I .2) as in [73. Pk = spanΦk . only the elements of Λ) with the eigenvalues of certain scalar FDEs. such that Λ = {λ ∈ C : λ is a solution of (4. . For 1 ≤ k ≤ N . the projection Pk ⊕ Qk → Pk . r]. On Bk . We describe brieﬂy an adjoint theory for FDEs (4. for some k ∈ {1. for each α ∈ R. and decompose C([−r. βk >)k βk . that the spectrum of A is reduced to the point spectrum and that. R∗ ) × C([−r.5k ) with characteristic equation given by (4. 0]. Q = Ker π. and let P be the centre space for (4.5k ). ϕ → Φk (Ψk . Since (H4) states that L does not mix the modes of eigenvalues of AT . . For each k ∈ N. u)k induces a projection C → Pk βk . R) by Λk := {λ ∈ C : λ satisﬁes (4. where Pk is the centre space for (4. k ∈ N. N }}. we ﬁx Λ = {λ ∈ σ(A) : Re λ = 0} = ∅. 0]. (4.4k ) with Re λ = 0. and from (H3) X can be decomposed by {βk }∞ k=1 . deﬁne (·. R) associated with (4. assuming the above hypotheses. Pk∗ = spanΨk .2).4k ) λ − µk − Lk (eλ· ) = 0 (k ∈ N). 101. we deduce that equation ∆(λ)y = 0 is equivalent to the sequence of characteristic equations (4. Again. . dimPk = dimPk∗ := mk . where Lk : C([−r.4k ). (Ψk . < ϕ(·). the linear equation u(t) ˙ = AT u(t) + L(ut ) is equivalent to the FDE on R z(t) ˙ = µk z(t) + Lk zt . R) = Pk ⊕ Qk . for k ∈ N. there exists N ∈ N. 104]) that its eigenvalues are exactly the roots of the characteristic equation (4. in the following way. 77. 92. 77]. It is wellknown ([73. the standard adjoint theory for retarded FDEs can be used to decompose C.5k ) and Bk is an mk × mk constant matrix.3). .255 Normal Forms And Bifurcations For DDE Let A be the inﬁnitesimal generator associated with the semiﬂow of the linearized equation (4. Φ˙ k = Φk Bk . the set σ(A) ∩ {λ ∈ C : Re λ ≥ α} is ﬁnite. The main idea is to relate the eigenvalues of the inﬁnitesimal generator A (in this case. . 0]. ·)k as the adjoint bilinear form on C([0.1 (see [58]): C([−r. We use the above decompositions to decompose C by Λ: C =P ⊕Q . u → Φk (Ψk . Since Λ is a ﬁnite set. 0]. for the sake of applications. P = Im π .4k ) and Re λ = 0} as in Section 2. R) → R are deﬁned by Lk (ψ)βk = L(ψβk ).2).
256 DELAY DIFFERENTIAL EQUATIONS N where dim P = k=1 mk := M and π : C −→ P is the canonical projection deﬁned by πφ = N Φk (Ψk . . using the above decomposition. α ∈ X}. In BC. v = Av dt (4. −r ≤ θ < 0. with Q ⊂ Ker π and P ⊂ C01 . 1 ≤ k ≤ N. One can prove that π is continuous and commutes with A˜ in C01 . Denote by X0 the function deﬁned by X0 (0) = I. (4. < φ(·). βk >)k βk . with xk (t) ∈ R. we ﬁrst enlarge the phase space so that (4. d ˜ + X0 F (v).6) where A˜ is the extension of the inﬁnitesimal generator A deﬁned by A˜ : C01 ⊂ BC −→ BC ˜ = C01 := {φ ∈ C : φ˙ ∈ C. α ∈ X. still denoted by π. k=1 4. along the lines of Section 2. φ(0) ∈ D(AT )} D(A) ˙ ˜ = φ˙ + X0 [L(φ) + AT φ(0) − φ(0)]. βk > βk .1) can be written as an abstract ODE. X0 (θ) = 0 . by deﬁning π(X0 α) = N Φk Ψk (0) < α. For v ∈ C01 . y ∈ C01 ∩ Ker π := Q1 . the enlarged space BC is decomposed by Λ as BC = P ⊕ Ker π. and let BC ≡ C × X = {φ + X0 α : φ ∈ C. where I is the identity operator on X. we write v(t) = N k=1 Φk xk (t) + y(t). Aφ We further consider the extension of the projection π : C → P to BC.1) reads as an abstract ODE.2 Normal forms As for retarded FDEs in Rn . k=1 By using π.
writing the Taylor expansion for F. (4. AQ1 : Q1 ⊂ Ker π → Ker π. (4. dt N p=1 where AQ1 is the restriction of A˜ to Q1 . Mj2 associated with the sequence of changes of variables are still given by (2. following the approach in [77]. v ∈ C. . . it is possible to develop a normal form theory for FDEs in X based on the theory in Section 2. BN ). Deﬁning B = diag (B1 .7). . βk > . . y) 1 2 j≥2 j! fj (x. p=1 d y =AQ1 y + (I − π)X0 F Φp xp βp + y . . the operators Mj1 . . .1) written in the form (4. . . F (v) = j! j≥2 where Fj is the jth Fr´echet derivative of F . x˙ =Bx + Ψ(0) < F (Φx) .6) is decomposed as a system of abstract ODEs on RM × Ker π. with linear and nonlinear parts separated and with ﬁnite and inﬁnite dimensional variables also separated in the linear term: N Φp xp βp + y). y =AQ1 y + (I − π)X0 F (Φx) . For (4. . y) . + y . but now AQ1 and π are as deﬁned above in this subsection. . . +y . N. βk > βk .. .7) Recall that the existence of a local centre manifold for (4. y ∈ Q1 . . k=1 βN β1 d . . . . the above system is written in a simpler form as β1 N . Eq. . Φ = diag (Φ1 . . .7) reads as x˙ =Bx + y˙ =AQ1 y + 1 1 j≥2 j! fj (x. ΦN ) and Ψ = diag (Ψ1 . . x˙ k =Bk xk + Ψk (0) < F ( k = 1. 1 Fj (v) . dt βN x = (x1 . xN ) ∈ RM .1) tangent to P at zero was proven under (H1)(H4) in [73].257 Normal Forms And Bifurcations For DDE Hence. ΨN ). (4. On the other hand. Formally.13).
4. βN The main results in Section 2 are still valid in this setting.3 The associated FDE on RN . RN ) → RN deﬁned by β1 N . it is natural to associate an FDE in RN with the original FDE (4.8). fj (x.11) N N where x(t) = xk (t) k=1 . (4. fj (x. formally we relate (4. 0]. . for the nonlinearity G : CN := C([−r. i = 1. (4. G(φ) = < Fj φ . (4. y) = (I − π)X0 Fj (Φx) . .6).5k ). N ). φN ) ∈ C([−r. . On the other hand. + y .10) R(φ) = (µk φk (0) + Lk (φk ))N k=1 . (4.9). RN ) (cf.1) by Λ at zero: x(t) ˙ = R(xt ) + G(xt ). Details and examples of applications can be found in [37]. the linearization of the original FDE in RN (2.1). 2. 0].9) However.2. 0]. y) = Ψ(0) < Fj (Φx) . . . y ∈ Q1 .7) and (4. y)..k k=1 .7) with the DDE in RN 3 x˙ = Bx + Ψ(0)G(Φx + y) (4. k = 1. The following FDE in CN := C([−r. Deﬁnition 4. RN ) deﬁned by (4.. given here by β1 N 1 . in order to derive the linear terms in (4.7) with Eq.1. (4. (2.258 DELAY DIFFERENTIAL EQUATIONS with the terms fji (x. with R ∈ L(CN . If we compare Eq. the operators A˜ and π are deﬁned diﬀerently in (4. j ≥ 2. . In fact.1) should have the form u(t) ˙ = R(ut )..8) y˙ = AQ1 y + (I − π)X0 G(Φx + y). . βk > k=1 βN β1 2 . . xt = xt. R are deﬁned by (4. + y . . . .8) from the procedure in Section 2. RN ) is said to be the FDE in RN associated with Eq. βk > k=1 βN (4.10). x ∈ RM . for φ = (φ1 . and G.
Normal Forms And Bifurcations For DDE 259 The relevance of this associated FDE in RN is summarized in the next theorem. and 2! 3! P roj span{β1 . In the present framework.2. between the second order terms of F and the eigenspaces for AT . the n. φ ∈ C([−r. up to second order terms.f. Now recall that second order terms of n.11) to be the same.f. The above result is not completely surprising. [37] Consider Eq. it is natural to expect that additional nonresonance conditions of second order should be imposed. are computed by looking at second order terms in the equation for x ∈ RN . The role of (H5) is to impose such nonresonance conditions.. φ ∈ C([−r. since it enables us to reduce the computation of n. are obtained by using not only second and third order terms in the equation for x ∈ RN . then. the n.1) and (4. (i) if F is of class C 2 .11) are the same up to k order terms. algorithm for (4. 1 1 (ii) if F is of class C 3 . for a suitable change of variables. Theorem 4. for FDEs in Hilbert spaces X to the computation of n. in order to conclude that the n. for FDEs in ﬁnite dimensional spaces. Of course. F (v) = F2 (v) + F3 (v) + o(v3 ). and assume (H1)(H4). βp >= 0 . R).βN } DF2 (u)(φβj ) = 0 . we could derive other nonresonance conditions between the jorder terms (j ≤ k − 1) of F and the eigenspaces for AT .f.. for the ﬂow on the centre manifold for both equations (4.f. Note that assumption (H5) can also be written as < DF2 (u)(φβj ).f. Moreover.1) and (4.f.f. on the centre manifold for both (4. Therefore. even if (H5) fails. In fact. but also second order terms in the equation for y ∈ Ker π. up to third order terms. j > N and all u ∈ P. On the other hand. (H5) then. it is not surprising to expect the quadratic terms in the n.. Theorem 4.1) and (4.11) are the same. The cubic terms of n. 0]. relative to Λ for (4..1) with the corresponding operators appearing in the computation of n. Mj2 of the n. 0].1). R). u ∈ P. The proof of this theorem is based on the possibility of identifying the operators Mj1 .11) is the same. in the sense that it imposes nonresonances inside the linear part of the equation.11). between L and the operator AT . in order to derive the same result for cubic terms. for a suitable change of variables.2 is particularly relevant. on the centre manifold for both (4. for 1 ≤ p ≤ N.f.f.f. note that (H4) can be interpreted as nonresonance conditions of ﬁrst order. on the centre manifold for both (4. one can prove . (4.1) and (4. for j ≥ N +1.
x)].g. 2 ] is not a Banach algebra. 2 ] [73. 2 ] ⊂ R. Consider the Wright equation with diﬀusion and Neumann conditions ([37. For the choice e.1). 73. x) =d − au(t − 1. x). and take as the Banach state space the fractional power space Xβ = D(Aβ ).260 DELAY DIFFERENTIAL EQUATIONS [37] that the associated FDE by Λ (4. and restrict the space of initial conditions that are chosen in a natural “initialhistory” space [85. e. 2 ] or X = W 2. x) =d + a(x)u(t. t ≥ 0. π) (4. Xβ ) as the phase space.4 Applications to bifurcation problems Population dynamics models involving memory and diﬀusion in one spatial variable are often given by equations of the form ∂ 2 u(t.3. x) + b(x)u(t − r. Another possibility is to consider a fractional power (AT )β of AT (0 < β < 1). ∂t ∂x2 ∂u(t. As an application. theory to a 2dimensional RDequation with two delays. X). x) ∂u(t. This theory is well established since the work of Henry [63]. One can restrict the state space to an appropriate space of functions invariant under products. x ∈ (0. the problem is not wellposed. x) ∂u(t. t > 0. 77. 4. X = L2 [1 . 94]. since L2 [1 .2 [1 . see [40]. with either Neumann or Dirichlet boundary conditions on [1 .11) still carries much information that can be used when computing n. there are several approaches. 0]. with norm vβ = (AT )β v. Another procedure is to consider X = L2 [1 .12) . x ∈ (1 . and Cβ = C([−1. Example 4. since that is not the aim of the n. 2 ] . X = C[1 . 54. However. 77.f. 0) = = 0 .f. 101]. and initial condition u0 = ϕ ∈ C = C([−r. theory. 86]. For an application of the n. 2 ]. 104]): ∂ 2 u(t. 2 ). ∂x ∂x t > 0. we shall consider the Wright equation with diffusion. x)u(t − r. See also [ [46.g. π) ∂u(t. the question of existence of solutions will not be addressed here. x) ∂t ∂x2 + c0 (x)u(t. 93. x)[1 + u(t. for (4. In order to guarantee existence of solutions.f. 0].
. π) : dv = 0 at x = 0. The eigenvalues of AT = d∆ are µk = −dk 2 . R) → R in (4.16) . . (4. One can easily sponding normalized eigenfunctions βk (x) = cos(kx) 2.14k ) Yoshida [104] showed that for a < π/2. where F (v. √ > 2 π π π π . i.15) x(t) ˙ = L0 (xt )+ < F xt β0 . there exists a unique pair of characteristic values on the imaginary axis. Since the Hopf condition holds [104].12) is given by d u(t) = d∆u(t) + L(a)(ut ) + f (ut .40 ). α) = −αv(−1) − ( π2 + α)v(0)v(−1). Let X = W 2. β0 > = − < αx(t − 1) √ + ( + α)x(t) √ x(t − 1) √ .14k ). α . D(∆) = {v ∈ W dx L(a)(v) = −av(−1). 2 π (4. (4. there is a pair of eigenvalues crossing transversely the imaginary axis at α = 0. with correcos(kx) .e. π) and C = C([−1. π} and where ∆ = ∂x 2 . We have 1 1 π 1 1 < F (xt β0 . k ∈ N0 . β0 >. f (v. ±iπ/2.Normal Forms And Bifurcations For DDE 261 where d > 0. β0 = √1π . 1.2 (0.13) 2 ∂ 2. Since L(ψβ0 ) = − π2 ψ(−1)β0 .2 (0. α). thus the zero solution is asymptotically stable. the characteristic equation is equivalent to the sequence of characteristic equations λ + ae−λ + dk 2 = 0 . At the critical point a = π/2.2 check that (H1)(H4) hold. (k = 0. all roots have negative real parts. a) = −av(0)v(−1). is deﬁned by π L0 (ψ) = − ψ(−1). For the linearization of (4. have negative real parts. . 0]. . a Hopf bifurcation occurs on a 2dimensional centre manifold. a) dt (4. With the above notations (except that here k ∈ N0 instead of k ∈ N).15) becomes x(t) ˙ = −( 1 π + α)x(t − 1)[1 + √ x(t)]. for a = π/2 and k = 0. In abstract form in C. N = 0. corresponding to the eigenvalue µ0 = 0. Let Λ = {iπ/2. the operator L0 : C([−1. and that all the other characteristic values of (4. α = 0 is (4. a > 0. ). k ≥ 0.13) about the equilibrium u = 0. −iπ/2}. rescale the parameter by setting a = π/2 + α. µ0 = 0. 1 π = −x(t − 1)[α + ( + α) √ x(t)] 2 π Thus. that are simple roots. (4. X). 2 The associated FDE (on R) by Λ at the equilibrium point u = 0. 0].
AT : D(AT ) ⊂ X → X is linear. Eﬀecting the scaling x → √1π x. write u = Φ0 cβ0 . where now X is a general Banach space. x(t) 1 < D1 F2 (u. space P0 for the linear FDE in R (4. X).e. Before developing a normal form theory on an invariant manifold for FDEs (5. i.e.12). DF (0) = 0.19). from Theorem 4. For (3. it is necessary to establish two necessary technical tools: a formal adjoint theory for linear equations. β0 >= 0. 0].1) ([43]). K2 are given by (3.19) for the scaling x → √1π x.22). α ∈ R. and the existence and smoothness of centre manifolds for nonlinearly perturbed equations (5. R). with stable nontrivial bifurcating periodic solutions.16).2 we deduce the n.. ψ ∈ C([−1. (4. 2 v ∈ C. We prove now that hypothesis (H5) is satisﬁed.1) in the phase space C := C([−r. and F : C → X is C k (k ≥ 2) with F (0) = 0. K2 as in (3. Normal Forms for FDEs in General Banach Spaces Consider again a semilinear FDE (4. where K1 ..1). c ∈ C2 . He have F2 (v. β0 > 2 √ π = − αψ(−1) + Φ0 (0)cψ(−1) + Ψ0 (−1)cψ(0) < βk . β0 > − < u(0)ψ(−1) + u(−1)ψ(0) βk .1) u(t) ˙ = AT u(t) + L(ut ) + F (ut ). This implies (H5). 0].22). (5. on the centre manifold for (4. ξ) by (3. 2 for all k ≥ 1. on the centre manifold is given in polar coordinates (ρ. the n. β0 > 2 π = −α < ψ(−1)βk . α) = −αv(−1) − π v(0)v(−1).12) without further computations: 1 ρ˙ =K1 αρ + K2 ρ3 + O α2 ρ + (ρ. α) . Thus.16) is the Wright equation (3. . 2 with K1 . 5.262 DELAY DIFFERENTIAL EQUATIONS i. where c ∈ C2 and Φ0 is a basis for the centre ˙ = L0 (xt ). α)4 π π ξ˙ = − + O (ρ.f.f.50 ). Then. a generic supercritical Hopf bifurcation occurs on the centre manifold for (4. For u ∈ P . α)(ψβk ).
We start by recalling the results in [92]. On the other hand. ˙ D(A) = {ϕ ∈ C : ϕ˙ ∈ C. since they include RDequations with delays. ϕ(0) ∈ D(AT ). a complete formal adjoint theory was developed in [43]. the theory in [92] was quite incomplete and more hypotheses were assumed. 0] → L(X. L(ϕ) = −r Under (H1)(H3). contrary to the situation studied in Section 4. From the point of view of the applications we have in mind.2). and its inﬁnitesimal generator A.1 263 Adjoint theory Consider a linear FDE in C of the form u(t) ˙ = AT u(t) + L(ut ). . In [16].3) where the characteristic operator ∆(λ) : D(AT ) ⊂ X → X is deﬁned by ∆(λ)x = AT x + L(eλ· x) − λx. (5. Consider the characteristic equation (4. However. ϕ(0) ˙ = AT ϕ(0) + L(ϕ)}. ϕ ∈ C.2) where AT . x ∈ D(AT ) \ {0} (5. ∆(λ)x = 0. (H3) there exists a function η : [−r. L are as in (5. The adjoint theory for abstract FDEs was initiated in [92]. deﬁned by the mild solutions of (5. (H2) T (t) is a compact operator for t > 0. following some ideas in [16] and [92]. we are not imposing any conditions which relate L with the eigenvalues for AT . U (t) : C → C.1) are more interesting.2) with initial condition ϕ ∈ C. whereas in our setting we will be dealing with compact operators.Normal Forms And Bifurcations For DDE 5.1). we shall assume only the following hypotheses: (H1) AT generates a C0 semigroup of linear operators {T (t)}t≥0 on X. Arino and Sanchez [16] developed a complete formal adjoint theory for equations u(t) ˙ = L(ut ). Consider the C0 semigroup of linear operators {U (t)}t≥0 . U (t)ϕ = ut (ϕ).3) for (5. Aϕ = ϕ. equations in the form (5. X) of bounded variation such that 0 dη(θ)ϕ(θ). In this section. the authors worked only with characteristic values that are not in the essential spectrum in order to obtain Fredholm operators. Note that. and will be presented here.
1. dη(θ)ϕ(ξ) > dξ << α. X ∗ ). ϕ >>=<< α.4) where Im [(A − λI)m ] is a closed subspace of C and Ker [(A − λI)m ] is the generalized eigenspace for A associated with λ. 0] → L(X ∗ . and deﬁne the operators θj eλθ x .264 DELAY DIFFERENTIAL EQUATIONS Travis and Webb [92] proved that U (t) is a compact operator for each t > r and σ(A) = σP (A). Thus. t ≤ 0. j ∈ N0 . · > is the usual where ∗ duality in X × X. Let λ ∈ C. the ascent and descent of A − λI are both equal to m. denoted as usual by Mλ (A). Let A∗ be the inﬁnitesimal generator associated with the ﬂow for (5. η ∗ : [−r. we now introduce some auxiliary operators.5) We use the formal duality << ·. The operator A∗ is called the formal adjoint of A. (5. r]. In particular. and < ·. L∗ : C ∗ → X ∗ is given by L∗ (α) = −r dη ∗ (θ)α(−θ). for λ ∈ σ(A). m ∈ N. Our purpose now is to write Im [(A − λI)m ] in terms of a (formal) adjoint theory.2) as α(t) ˙ = −A∗T α(t) − L∗ (αt ). ϕ(0) > − −r C∗ C([0. it follows that. where m is the order of λ as a pole of the resolvent R(λ. As seen in Section 4. C = Ker [(A − λI)m ] ⊕ Im [(A − λI)m ]. Chapter 7] and [16]. λ is a characteristic value if and only λ is an eigenvalue of A.g. Ljλ (x) = L j! . X ∗ ). and η ∗ (θ) is the adjoint for η(θ). and. in the sense that a Fredholm alternative result should be established. (5. ϕ ∈ D(A). Ljλ : X → X. ϕ >>=< α(0). · >> in C ∗ × C introduced in [92] as the bilinear form 0 θ < α(ξ − θ). A) [90].6) where αt ∈ C ∗ . Deﬁne the formal adjoint equation to (5. Ker (A − λI) = {eλ· x : x ∈ Ker ∆(λ)}. for α ∈ D(A∗ ). Inspired by the work of Hale [58. [84]). with Mλ (A) ﬁnite dimensional.2). Aϕ >> . that will be used ﬁrst of all to get an explicit characterization of the null spaces and ranges for the operators (A − λI)m and (A∗ − λI)m . there is x ∈ D(AT ) \ {0} such that u(t) = eλt x is a solution of (5. r]. (5.6). since << A∗ α. in this case. αt (s) = α(t + s) for s ∈ [0. A∗T is the adjoint operator 0 for AT . 0 X∗ := is the dual of X. From general results on C0 semigroups and compact operators (see e.
79]. clearly λ ∈ ρ(A) if and only if 0 is not an eigenvalue of ∆(λ). Proof. . ∈ Ker ((L(m) )∗ ) . with . r].... . Let λ ∈ C and m ∈ N. Then: (i) Ker [(A − λI)m ] = 1 (m) Ker (Lλ ) . Rλ (ψ) = θ λ(θ−ξ) (θ − ξ)ψ(ξ)dξ −L 0 e θ ψ(0) − L 0 eλ(θ−ξ) ψ(ξ)dξ . 0 0 . 0 m−1 j=0 u0 . . . Hence. Moreover. m−1 (−s)j −λs ∗ xm−j−1 .. Let λ ∈ C. λ x∗m−1 The proof of this theorem follows from direct computations. . which generates a compact C0 semigroup. This also implies that 0 ∈ ρ(∆(λ)) iﬀ 0 ∈ σP (∆(λ)).2. . . = . . Lm−2 λ . . with (iii) Ker [(A∗ − λI)m ] = j=0 j! e ∗ x0 1 . The characteristic operator ∆(λ) generates a compact C0 semigroup of linear operators on X.4). ∆(λ) L1λ − I 0 0 . with the bounded linear operator L0λ − λI. On the other hand. Lemma 5.. . . . (m) m : C → X .1. . λ ∈ σ(A) if and only if 0 ∈ σ(∆(λ)). ∈ um−1 (m) (m) (ii) ψ ∈ Im [(A − λI)m ] iﬀ 0Rλ (ψ) ∈ Im (Lλ ).3. s ∈ [0. Lλ and L2λ . From (5. . p. Theorem 5. θ ∈ [−r. m ∈ N. 0 ∆(λ) (m) Rλ m−1 θ −L 0 eλ(θ−ξ) (θ−ξ) ψ(ξ)dξ (m−1)! . ∆(λ) is the sum of AT . ∆(λ) generates a compact semigroup of linear operators on X [84.265 Normal Forms And Bifurcations For DDE (m) Lλ (m) : X m → X m . Then: (m) (m) (i) if µ ∈ ρ(∆(λ)). θj λθ j! e uj .. .. We now establish some preliminary lemmas. Lemma 5. Hence λ ∈ ρ(A) iﬀ 0 ∈ ρ(∆(λ)). then µ ∈ ρ(Lλ ) and (Lλ − µI)−1 is a compact operator.. Lm−1 ∆(λ) L1λ − I λ 0 ∆(λ) L1λ − I . 58] for similar proofs. See [16. . . so it is omitted. . . 0].
.1 and Lemma 5.2. (5. From Lemma 5. dim Ker [(A − λI)m ] = dim Ker [(A∗ − λI)m ].9).2 we conclude that this operator generates a compact C0 semigroup. dim Mλ (A) = dim Mλ (A∗ ). We are now in the position to state a “Fredholm alternative” result. and from Schauder’s (m) theorem we conclude that its adjoint [(Lλ − µI)−1 ]∗ is also a compact operator. The rest of the proof of (i) follows by induction.3. m ∈ N.266 DELAY DIFFERENTIAL EQUATIONS (m) (ii) Im (Lλ ) is a closed subspace of X m .7). As an immediate consequence of Theorem 5. For the proof of (ii). we use (i) and general spectral properties of compact operators. Let λ ∈ C. dim Ker (Lλ ) = dim Ker ((Lλ )∗ ) (m) (m) (5. One can prove that I ). For m = 1. and the ascents of both operators A − λI and A∗ − λI are equal. In particular.4.8) where [(Lλ )∗ − µI]−1 = [(Lλ − µI)−1 ]∗ .8) and (5. Then.7) Proof. µ I (m) (m) Ker ((Lλ )∗ ) = Ker ([(Lλ )∗ − µI]−1 + ). we obtain (5. we have Lλ = ∆(λ). In particular. That is. Let λ ∈ σ(A). and for λ ∈ σP (A). m ∈ N. we (m) have that (Lλ − µI)−1 is a compact operator. See [43] for details. µ) = [∆(λ) − µI]−1 is a compact operator. 0 ∈ σ(∆(λ)). Then. we can now state the following: Theorem 5. Fix any (m) µ ∈ ρ(∆(λ)).4.9) µ µ From (5. σP (A) = σP (A∗ ). Thus. Let λ ∈ C. then Mλ (A∗ ) = Ker [(A∗ − λI)m ].3 we deduce that µ ∈ ρ(Lλ ). its resolvent R(∆(λ).5. (1) Proof. µ (m) (m) Ker (Lλ ) = Ker ((Lλ − µI)−1 + (5. Let µ ∈ ρ(∆(λ)). Lemma 5. and from Lemma 5. From Lemma 5. Also from Lemma 5. if Mλ (A) = Ker [(A − λI)m ]. (m) (m) (m) dim Ker ((Lλ − µI)−1 + I I (m) ) = dim Ker ([(Lλ − µI)−1 ]∗ + ).
ψ ∈ Im [(A − λI)m ] iﬀ Rλ (ψ) ∈ Im (Lλ ). we denote Im [(A − λI)m ] = Ker [(A∗ − λI)m ]⊥. . λs } ⊂ . ψ >> . Then ψ ∈ Im [(A − λI)m ] if and only if << α. (m) ∗ )T ∈ (X ∗ )m one By using the formal duality. Rλ (ψ) >= 0 (m) for all Y ∗ ∈ Ker ((Lλ )∗ ). ⊥ where Pλ := Mλ (A) and Qλ := Mλ (A∗ ) .1. From Lemma 5. Let λ ∈ σ(A). . (m) hence they coincide with the annihilator of the adjoint operator (Lλ )∗ . Ker [(A − λI)m ] = Mλ (A). . . For the particular case of m being the ascent of A − λI. . . ψ >>= 0 for all α ∈ Ker [(A∗ − λI)m ]. thus ψ ∈ Im [(A − λI)m ] iﬀ < Y ∗ . Im (Lλ ) = Ker ((Lλ )∗ )⊥ . . Rλ (ψ) > m−1 ∗ < yj .4) we obtain the decomposition of the phase space C by λ ∈ σ(A) given as (5. . for Y ∗ = (y0∗ . m ∈ N.10) C = Pλ ⊕ Qλ .1.3. . · >>. e−λs α= (m − j − 1)! j j=0 ∗ )T ∈ Ker ((Lλ )∗ ). (m) (m) (m) (m) From Theorem 5. the ranges of Lλ are closed subspaces. .6. ym−1 proves that < Y ∗ . from (5. where the orthogonality “⊥” is relative to the formal duality << ·.. we deduce that α ∈ Ker [(A∗ − λI)m ] if and only if m−1 (−s)m−j−1 ∗ y .6. (m − j − 1)! j Using again Theorem 5. (m) Proof. (m) From the Fredholm alternative result in Theorem 5. ψ >>= 0 for all α ∈ Ker [(A∗ − λI)m ]. i.e. The above arguments imwith Y ∗ = (y0∗ . L =− (m) j=0 = m−1 j=0 << e−λs θ eλ(θ−ξ) 0 (θ − ξ)m−j−1 ∗ ψ(ξ)dξ > + < ym−1 . ψ(0) > (m − j − 1)! (−s)m−j−1 ∗ y .267 Normal Forms And Bifurcations For DDE Theorem 5. ym−1 ply that ψ ∈ Ker [(A − λI)m ] if and only if << α. . . It is straightforward to generalize this decomposition to a nonempty ﬁnite set Λ = {λ1 .
ϕ >>= 0. Ψ are normalized dual bases for bases of P. which can be interpreted as a perturbation of (5. [43. Ψλs P = span Φ P ∗ = span ΨΛ where p = dim P = dim P ∗ . p × p matrix such that Φ .268 DELAY DIFFERENTIAL EQUATIONS σ(A). Ψ for P.2). ϕ >>= 0 for all α ∈ Ker [(A∗ − λI)m ] and ϕ ∈ Ker [(A − µI)r ]. Dh(0) = 0. and B is the ˙ = ΦB. there is a centre manifold Wc = {ϕ ∈ C : ϕ = Φx + h(x).. P ∗ . . P is the centre space for (5. For applications to bifurcation problems. −Ψ ˙ = BΨ.7. and assume that Λ = {λ ∈ σ(A) : Re λ = 0} is nonempty.2 Normal forms on centre manifolds Consider now Eq. By using the decomposition (5.1) with F a C k (k ≥ 1) function. . and h : V → Q is a C k smooth function such that h(0) = 0. x ∈ V }. λ ∈ Λ. where Φ. we will as usual ﬁx Λ = {λ ∈ σ(A) : Re λ = 0}. the existence of a C 1 smooth pdimensional centre manifold for (5. . for all ∈ P ∗ } Q= P and P ∗ = Mλ1 (A∗ ) ⊕ · · · ⊕ Mλs (A∗ ). 5.11).1). on Wc the ﬂow is given by the ODE x˙ = Bx+ < Ψ(0).2). One also proves that for any m. where V is a neighbourhood of 0 ∈ Rp .10).. r ∈ N. it is therefore possible to choose normalized bases Φ. . such that << Ψ. P ∗ . The following centre manifold theorem states that actually the centre manifold is C k smooth. Φ >>= Ip . The decompositions (5. . in [69]. By using the Fredholm alternative result.11) where P = Mλ1 (A) ⊕ · · · ⊕ Mλs (A) ∗ ⊥ = {ϕ ∈ C : << α. Φλs ). Assuming F is a C 1 function. Theorem 5. then << α.2). Ψλ1 Ψ = . (5.2) tangent to P at zero was proven. 69] For (5. if λ. x ∈ V. With the above notations. yield a decomposition of C by Λ as C = P ⊕ Q. µ ∈ σ(A) and λ = µ. and p = dim P . Φ = (Φλ1 . Q are invariant subspaces under A and (5. (5. Clearly P. F (Φx + h(x)) >.
π(ϕ + X0 α) = Φ << Ψ. y(t) = (I − π)v(t) ∈ Ker π ∩ C01 = Q ∩ C01 := Q1 . process are still given by (2. v(t) >>∈ Rp . ϕ >> + < Ψ(0). Recall the deﬁnition of the extension A˜ of the inﬁnitesimal generator A. for FDEs in Rn must be established.14) At each step j ≥ 2. Formally. ˜ Hence BC is decomposed Again. y) where now the terms of order j are given by fj1 =< Ψ(0).13) becomes x˙ = Bx + y˙ = AQ1 y + 1 1 j≥2 j! fj (x. (5. See [39] for complete proofs. results similar to the ones proven for the case of n. (5. . now follows along lines similar to the ones in Section 2. Uj2 (¯ x)).6) in the enlarged phase space BC introduced in Section 4. Uj2 ∈ Vjp (Ker π)) is performed. where AQ1 = A The algorithm of n. fj2 = (I − π)X0 F (Φx + y). by Λ as BC = P ⊕ Ker π. π : BC → P. so that all the nonresonant terms of degree j vanish in the transformed equation. y) 1 2 j≥2 j! fj (x. F (Φx + y) >.13).f. α > . Eq.12). v ∈ C. (4. the operators appearing in the n.1) be written as the abstract ODE (4.f. y) = (¯ x. (U (¯ j! j (Uj1 ∈ Vjp (Rp ). however the operators π and AQ1 are as deﬁned in this subsection.13) y˙ = AQ1 y + (I − π)X0 F (Φx + y) ˜ Q1 is as in Section 4. Of course. 1 Fj (v). for these operators. (5.f. y¯) + 1 1 x). Deﬁne now an extension of the canonical projection C = P ⊕ Q → P .Normal Forms And Bifurcations For DDE 269 Let Eq. a change of variables of the form (x. Writing the formal Taylor expansion of F .12) For v(t) ∈ C01 decomposed according to (5. F (v) = j! j≥2 system (5. v(t) = Φx(t) + y(t) with x(t) =<< Ψ. F (Φx + y) > (5.6) is equivalent to the system 3 x˙ = Bx+ < Ψ(0). π commutes with A˜ on C01 = D(A).
π) = 0. since the equilibria Uk (x) are spatially nonconstant. which is locally stable if rk max{Uk (x) : x ∈ [0.270 DELAY DIFFERENTIAL EQUATIONS 5. and there is a unique positive spatially nonconstant equilibrium Uk .15) where k > 0 and r > 0. π]. r)y = 0.17) ∆(k.16) was ﬁrst studied by Green and Stech [50].16) are assumed. π] = {y ∈ X  y. U (x.3 A reactiondiﬀusion equation with delay and Dirichlet conditions For the Hutchinson equation with diﬀusion.3. X = L [0.15) with Neumann conditions ∂U ∂x (0. it was proven that the equilibrium U = 0 is a global attractor of all positive solutions if k ≤ 1. They also proved that there is a sequence (rkn )n∈ν0 of Hopf bifurcation points. ∂t ∂x2 t > 0. ∂t ∂x2 t ≥ 0. .15) is particularly diﬃcult when Dirichlet conditions (5. ∂ 2 U (t. t) = − ku(x. y(0) = y(π) = 0}. t ≥ 0.15)(5. r) = D2 + k(1 − Uk ) − ke−λr Uk − λ. we consider now Dirichlet conditions U (t. the characteristic equation associated with the linearization about Uk (x) is a second order diﬀerential equation with nonconstant coeﬃcients.16) ∂U For (5. x ∈ (0. 0) = U (t. the translation u = U − 1 transforms (5. where ∆(k. π). X. and H0 = H0 [0. 2 ∂ 2 2 2 ˙ y¨ ∈ with D2 = ∂x 2 . The study of (5. (5. there exists r(k) > 0 such that Uk is locally stable if the delay r is less than r(k). where rk0 = r(k) is the ﬁrst bifurcation point. In the next statement. λ.15) into ∂ 2 u(x. and unstable for r > r(k). t) = ∂x (π. subject to constraints given by the boundary conditions. we summarize the results in [23]. (5. In this case. the characteristic equation is given by (5. The main problem derives from the fact that. x ∈ (0. already studied in Example 4. x) = + kU (t. π]} < π/2 in [65]. t ≥ 0. U = 0 becomes unstable. Model (5. t)) ∂u(x. t) ≡ 1 is the unique positive steady state. This stability criterion was improved to rk max{Uk (x) : x ∈ [0. the authors showed that for k − 1 > 0 small. In fact. t − r)[1 + u(x. λ. In [23]. t) = 0. π]} < 1. t)]. π). y ∈ H02 \ {0}. x)[1 − U (t − r. In [50]. x)]. x) ∂U (t. for k > 1.
if and only if ν = νk . π] as k → 1+ . θk + 2nπ π rkn = as k → 1+ . the eigenvectors yk . are all known only implicitly. Since we want to use the delay r as the bifurcating parameter. and the critical values rkn . we emphasize that the stationary solutions Uk . (5.17) for r = rkn . ˙ D(Ar ) = {φ ∈ C 1 ([−r.f. r). k ∗ ) a Hopf bifurcation arises from the equilibrium Uk as the delay r increases and crosses the critical points r = rkn .16) are written in C as u(t) ˙ = rD2 u(t) + Lr (ut ) + f (ut . [23] There exists a k ∗ > 1 such that for each ﬁxed k ∈ (1. the eingenvalues iνk . x) = U (rt. where details of the results below can be found. νk = (k − 1)hk . the eigenvalue problem (5. r > 0. n ∈ N0 . y). ·) − Uk ∈ X. it is convenient to normalize the delay by the timescaling. ¯ (t. (t ≥ 0) (5. techniques. x). y¯k . . n ∈ N0 .15)After translating Uk to the origin. . In the above theorem. f (ϕ.271 Normal Forms And Bifurcations For DDE Theorem 5. The aim now is to determine the stability of the periodic solutions near the positive steady state Uk arising from Hopf. with hk → 1 as k → 1+ . The proof of this statement was given in [42] by using n. and for k ∈ (1. r) = −rkϕ(0)ϕ(−1). Furthermore. y ∈ H02 \ {0}. k ∗ ). for u(t) = U (5. U ¯ (t. In [23] it was stated without proof that these periodic solutions are stable on the centre manifold.18) where for k > 1 ﬁxed we denote Lr (ϕ) = rk[(1−Uk )ϕ(0)−Uk ϕ(−1)]. Fix k > 1. X) as the phase space. 0]. ϕ ∈ C. θk → νk 2 yk → sin(·) in L2 [0. and choose C := C([−1. the characteristic values λ = iνk are simple roots of (5. X) : φ(0) ∈ H02 . The linearized equation around the equilibrium Uk is given by u(t) ˙ = rD2 u(t) + Lr (ut ). (5. n ∈ N0 . 0].8. φ(0) = r[D2 + k(1 − Uk )]φ(0) − rkUk φ(−r)}.19) whose solution operator has inﬁnitesimal generator Ar deﬁned by ˙ Ar φ = φ. y = cyk (c = 0). −iνk . as the delay crosses rkn . where r = rkn . with ν ≥ 0. r.17) has a solution (iν. n ∈ N0 .
and B = iλkn 0 0 . α). α) + 3!1 F3 (ut . With the usual notations. α) > (5. α). α) = 12 F2 (ut . F (ϕ. set Λ = Λk. method. in BC decomposed by Λ. where (5. n ∈ N0 . For the linearization u(t) Lrkn ut .18) as u(t) ˙ = rkn D2 u(t) + Lrkn ut + F (ut . P ∗ = span Ψ.272 DELAY DIFFERENTIAL EQUATIONS Due to the timescaling. Λ = {iλkn . where x(t) ∈ C2 . α). λ. r)y = 0. kn where π Skn = 0 [1 − krkn e−iθk Uk (x)]yk2 (x)dx. . For the Taylor expansion F (ut . and decompose C = P ⊕ Q and BC = P ⊕ Ker π by Λ via the formal adjoint theory in Section 5.20) We now describe brieﬂy the framework for the application of the n. One must compute the n. For r = rkn . α) = αD2 ϕ(0) + Lα (ϕ) + f (ϕ. −iλkn . P ∗ as Φ(θ) = [φ1 (θ) φ2 (θ)] = [yk eiλkn θ y¯k e−iλkn θ ] (−1 ≤ θ ≤ 0) .1. Ψ(s) = 1 y¯k eiλkn s ψ2 (s) S P = span Φ. for the 2dimensional ODE giving the ﬂow on the centre manifold for (5.n as the set of eigenvalues on the imaginary axis.19) becomes ∆(k.21) is written as x˙ = Bx+ < Ψ(0). the characteristic equation for (5.f. (5. y(t) ∈ Q1 =Q∩ C01 . rkn + α). −iλkn }.18) at the bifurcating points r = rkn . we can choose normalized dual bases Φ. y ∈ H02 \ {0} ∆(k. λ. we have F2 (ut . r) = r[D2 + k(1 − Uk ) − ke−λ Uk ] − λ.21) Lrkn (ϕ) = rkn k[(1 − Uk )ϕ(0) − Uk ϕ(−1)]. Ψ for P. Let k > 1 and n ∈ N0 be ﬁxed. Consider the scaling r = rkn + α. α) = 2α[D2 u(t) + k(1 − Uk )u(t) − kUk u(t − 1)] − krkn u(t)u(t − 1). α) = −3!kαu(t)u(t − 1).1 −iλkn s y e k (s) ψ S 1 kn = (0 ≤ s ≤ 1).f. ˙ = rkn D2 u(t) + F3 (ut . F (Φx + y. and write (5.22) y˙ = A˜Q1 y + (I − π)X0 F (Φx + y. (5.
yk2 >< yk .f. α)). n and are given in terms of the coeﬃcients of the original equation. yk2 > y¯k e−iλkn θ }. with k ∈ (1. yk2 > yk eiλkn θ + < y¯k . k ∗ ) and n ∈ N0 . y ¯ >   < y k k k k 3 Skn 2 2krkn < yk . α)4 ) ˙ −σn + O((ρ. n be ﬁxed. and 1 1 < yk . n) and K2 = K2 (k. [42] Let k. 1 1 < yk .f.Normal Forms And Bifurcations For DDE 273 The application of the n. (h20 (−1) + h20 (0)eiθk )¯ C2 = C2 (k. n) depend on k. y  >  + . n) = − and h20 . K1 = Re A1 . h11 ∈ Q1 are the solutions of the IVPs. h11 is a linear ODE that can be easily solved. n) + C2 (k. ξ) by ρ= ˙ K1 αρ + K2 ρ3 + O(α2 ρ + (ρ. n) = iνk 1 < yk . n)] 4 and C1 = C1 (k. The ﬁrst equation of the system for both h20 .9. yk 2 > yk eiλkn θ + < y¯k . n) = 8k2 rkn i 1 Re(eiθk )e−iθk < yk . yk > Skn 1 A2 = A2 (k. respectively. the second equation requires the resolution of . method to (5. However.22) leads to the following result: Theorem 5. yk 2 > y¯k e−iλkn θ } h˙ 11 = −4krkn Re(eiθk ){ Skn S kn h˙ 11 (0) − rkn {D2 h11 (0) + k(1 − Uk )h11 (0) − kUk h11 (−1)} = −4krkn Re(eiθk )yk 2 . Note that the coeﬃcients K1 = K1 (k. A1 = A1 (k. (h11 (−1) + h11 (0)e−iθk )yk > Skn yk > . n) = [C1 (k. h˙ 20 − 2iλkn h20 = −2krkn e−iθk { Skn S kn h˙ 20 (0) − rkn {D2 h20 (0) + k(1 − Uk )h20 (0) − kUk h20 (−1)} = −2krkn e−iθk yk2 . + < yk . up to cubic terms for the ﬂow on the centre manifold of the origin near α = 0 is given in polar coordinates (ρ. yk 2 > νk Sk2n 1 1 iθk 2 2 2 2 2 − 2(Re(e ))  < y . The n. ξ= where α = r − rkn . K2 = Re A2 .
Since the solutions cannot be explicitly computed. Note that these results hold for k > 1 close to 1. the signs of the meaningful constants K1 and K2 have to determined by indirect methods. n) < 0. π]. k ∗ ) and n ∈ N0 . whose conclusions we give in the next theorem.10. [42] Fix k ∈ (1. Therefore a supercritical Hopf bifurcations occurs at r = rkn . Then for 0 < k − 1 << 1. An interesting question is whether the results persist for large k > 1. n) > 0 and K2 (k. Theorem 5. This requires hard computations. with stable nontrivial periodic solutions on the centre manifold.274 DELAY DIFFERENTIAL EQUATIONS a PDE with nonconstant coeﬃcients and Dirichlet boundary conditions on [0. . K1 (k.
Ait Babram. Arino and R. M. Acad. Ezzinbi. Approximation scheme of a center manifold for functionnal diﬀerential equations. Vector valued laplace transforms and cauchy problems. Arendt. O.L. [2] M. [7] M. M. Arino and M. (preprint). Sci.Acad. Adimy et K. Computational scheme of a center neutral manifold for functionnal diﬀerential equations. 258 (2001).R.R. Paris S´erie I 311 (1990). 177 (1993). Hbid. Ait Babram. x(t − ω)). Soc. [4] M. 213 (1997). 396–414. [6] M. [3] M. Math. 554–572. Benkhalti. Arino. Anal. Periodic solutions for x (t) = f (x(t).. Integrated semigroups and functional diﬀerential equation in inﬁnite dimension. Sci. London Math. J. (preprint). Integrated semigroups and delay diﬀerential equations. Paris S´erie I 318 (1994). Acad. Proceeding of Royal Society of Edinburgh 109A (1989). Adimy. 423–428. Arendt. Ezzinbi. Arino. Adimy. and O. J. 125–134. vol 54. 529–534. C. C. Ait Babram. [10] W. R. [9] W. Resolvant positive operators and integrated semigroup.L. Israel J. 1987. Appl. Bifurcation de Hopf locale par les semi groupes int´egr´es. 3. Sci. Adimy and K. Anal.L. 327–352. Adimy. Appl. Hopf global bifurcation for retarded diﬀerential equations using integrated semigroups.References [1] M. to appear in C. Anal. Math. Existence of periodic solutions for a delay diﬀerential equation via an approximation of center manifold. 321–349. Hbid. Hbid and O. Math 59 (1987). Appl. J. Equations de type neutre et semi groupes int´egr´es. [11] O. [8] M. 277 . 245–260. [5] M. Math. Proc.
Nonlinear Anal. [19] R. 361–385. [17] O. Diﬀerential Equations 124 (1996). Sanchez.R. 509–520. [23] S. Annali di Matematica pura and applicata 4 (1982). Appl. Hbid Periodic solution of retarded diﬀerential systems close to ordinary ones. Stability and Hopf bifurcation for a population delay model with diﬀusion eﬀects. [24] J. 1305–1320. Bernfeld. 9 (1996).D. 4 (1996). Poincar´e procedure for an ordinary diﬀerential system perturbed by a functional term. .A. Convergence theorems for integrated semigroup.L. Diﬀ. Quart. Arino and A. [15] O. Huang. Int. Canad. Salvadori. [14] O. Cherif. J. Busenberg and W. Int. Appl. Carr.S. 2 (1994). Academic Press. Bernfeld and L. Dufour. Existence of periodic solutions for a delay diﬀerential equation via the Poincare procedure. [21] S. Arino and M. New York (1963). 1981. An exact formula of period 4solution of x’(t)=λf (x(t − 1)) which bifurcate at λ = π2 . Busenberg and B.D. 5 (1992). 80–107. D. [18] J. 135–155. Eq. 14 (1990). T.L. Hbid. 547–571. Hbid. Int. A variation of constants formula for an abstract functionaldiﬀerential equation of retarded type. Salvadori.S. Arino and M. 125–148. D.R. Math. Arino and E. 191 (1995). Wu.E. Nonlinear Anal. “Applications of Center Manifold Theory”. Arino and M.L. Anal. Belair and S. Diﬀ. J. P. Linear theory of abstract functional diﬀerential equations of retarded type.E. Diﬀ. SpringerVerlag.L. Generalized Hopf bifurcation and hasymptotic stability. math.M. Stability in a three dimensional system of delay diﬀerential equations. [20] S. [16] O. Arino et E. Cooke. New York. Eqns. Quasiinvariant manifolds.276 DELAY DIFFERENTIAL EQUATIONS [12] O. “Diﬀerential diﬀerence equations”. [22] S. stability and generalized Hopf bifurcation. Negrini and L. J. Bellman and K. Eq. Sanchez. TMA 4 (1980). 2 (1989). [13] O. 4 (1996).
O. A bifurcation problems for functional diﬀerential equation of ﬁnitely retarded type. Amer. Diﬀerential operators with nondense domains.K. Existence of periodic solutions of autonomous functional diﬀerential equations.N. 350–378. 1991. Chafee. Sup. Diﬀerential Equations 15 (1974). Fields Institute Communications 21 (1999). 143–158. 31 (1977).Stewart (eds). I. 1982.N. Part II. [38] T. Cl´ement. Chow and J.A. van Gils. Faria. 129–149. Integral averaging and bifurcation. Warwick. An. 112–159. 277 (1987). Gyllenberg. [31] J. Appl. “Methods of Bifurcation Theory”. J.N. Math. 352 (2000). J. 122–141. Chafee. Exact formula for periodic solutions of dx(t+1) dt 3 −αx(t) + bx (t). 709–725. Singularity Theory and its Application. J. M. [35] T. Normal forms and Hopf bifurcation for partial diﬀerential equations with delays. Proc. The center manifold for delay equations in light sun stars. Faria. [36] T. 35 (1971). Chow. Math. 2217– 2238. Soc.Roberts. H. Da Prato and E. [27] P. Chow and J. Faria. = [34] P. (4) 14 (1988). I. Cushing. [26] N. Normal forms for periodic retarded functional diﬀerential equations. Diﬀerential Equations 26 (1977). Faria. Scuola Norm. [29] S. Bifurcations aspects for some delayed population models with diﬀusion. [37] T.M 1464. J. Trans. 27 (1978). Math. MalletParet. Biosci. J. On a planar system modelling a neuron network with memory. Math. Thieme.N. Ann. Diekmann. Hale. Springer L.M. 285– 344.J. [33] O. M. Periodic solutions of two species interactions models with lags. 1989. . Sci. Math. 21–46. Roy. Diekmann and S. Generalized Hopf bifurcation and perturbation in full neighbourhood of a given vector ﬁeld. Soc.REFERENCES 277 [25] N. Diﬀerential Equations 168 (2000). J. [28] S.N. [30] S. Applied Math and Physics (ZAMP) 37 (1986). Pisa Cl. Dormayer. Edinburgh 127A (1997). Indiana Univ. SpringerVerlag. Perturbation theory for dual semigroups. Heijmans and H. Sinestrari. The sunreﬂexive case. New York. [32] G. Anal.M. 143–156.R.
[48] A. T. Eds. 357– 372. [45] T. Cont. SIAM J. 201–224. J. in “Oneparameter semigroups of positive operators” (R. Faria and L. [43] T. SpringerVerlag. Math. Diﬀerential Equations 29 (1978). [50] D. W. Magalh˜ aes. Faria. Grabosch and U. [49] R.). (1986). Fields Institute Communications 31 (2002). On the Study of Singularities for a Planar System with Two Delays. Physica D 159 (2001). Smoothness of center manifolds for maps and formal adjoints for semilinear FDEs in general Banach spaces. Appl. Academic Press (1981). J. [41] T. pp 219–232. Berlin/New York. Diﬀerential Equations 6 (1980). A semigroup approach to retarded diﬀerential equations. [42] T. Green and H.278 DELAY DIFFERENTIAL EQUATIONS [39] T. . ed. Disc.Nagel. J. Huang. Normal forms for semilinear functional diﬀerential equations in Banach spaces and applications.). Faria and W. Huang and J. Faria and L. 19–28. Cooke. [44] T. Semilinear Equations functional diﬀerential equations in Banach spaces. Stability and bifurcation for a delayed predatorprey model and the eﬀect of diﬀusion. Systems 7 (2001). Grafton. Impul. Part II. Faria. Diﬀusion and hereditary eﬀects in a class of population models. J.B. Faria. Equations 122 (1995). 173–203. Normal forms for retarded functional diﬀerential equations with parameters and applications to Hopf singularity. 1–14. Cont. [46] W. 215–232. 254 (2001). Math. 181–200. Giannakopoulos and A. Zapp. A periodicity theorem for autonouous functional differential equation. Anal. Vol. Fitzgibbon. 125–141. Magalh˜ aes. [47] F. pp. Epidemics and Population Problems (S. 1184. Lecture Notes in Mathematics. Busenberg and C. Stability of periodic solutions arising from Hopf bifurcation for a reactiondiﬀusion equation with time delay. Faria. Normal forms for retarded functional diﬀerential equations and applications to BogdanovTakens singularity. Anal. J. Bifurcations in a planar system of diﬀerential delay equations modeling neural activity. [40] T. T. in Diﬀerential Equations and Applications in Ecology. 34 (2002). Wu. 433– 463. Disc. Dyn. Diﬀerential Equations 122 (1995). Dyn. Systems A 10 (2003). Moustakas. Diﬀerential. 155–176. Stech.
12 (2000). “Geometric Theory of Semilinear Parabolic Equations”. Hale. Critical cases for neutral functional diﬀerential equations.K.M. 1–30. Rabat.. [56] J. J. Huang.K.K.D. Tanaka. Vol. Solutions of x (t) = −g(x(t−1)) approach the KaplanYorke orbits for odd sigmoid g. Diﬀ. 87–95. Hbid. Providence. Periodic solutions of diﬀerence diﬀerential equations. Weedermann. New York/Berlin. Hadeler and J. 1–23. [52] K. Square and pulse waves with two delays. N. On perturbations of delaydiﬀerential equations with periodic orbits. Philips. Contribution a` l’´etude des perturbations des ´equations diﬀ´erentielles `a retard.REFERENCES 279 [51] R. [53] J. [59] J. Soc. Maroc (1992). and Y. Cambridge University Press. 1 (1977).M. [61] M. Th`ese de Doctorat d’Etat de l’Universit´e Mohamed V. Diﬀerential Equations 10 (1971). and Anal. Diﬀerential Equations 197 (2004). Math. ArcH.K. [58] J. “Functional analysis and semigroups”. Cambridge (1981).Kazarinoﬀ. Tomiuk. 36–53. “Introduction to Functional Diﬀerential Equations”. Hille and S.L. [60] B. Diﬀerential Equations 118 (1995). J. Appl. Verduyn Lunel. Large diﬀusivity and asymptotic behavior in parabolic systems. Theory of Functional Diﬀerential Equations. Diﬀerential Equations 114 (1994). 840. Hale and S. Grimmer. Henry.K. [62] A.I. Lecture Notes in Math. J. Hale and M. Period Doubling in singularly perturbed delay equations. 219–246. 1981. (1957).P. Wan. J. Math. Dynam. Hale. [54] J. 59–82. Amer. . Rat. Herz.V. [55] J. NewYork (1993).K. Hassard. Hale. [63] D. 72 (1979). SpringerVerlag. Anal. Eq. J. “Theory and applications of Hopf bifurcation”.K. Hale and W. Existence of periodic solution of functional diﬀerential equation. J. J. Diferential Equations 118 (1986).M. 455–466. R. SpringerVerlag. Hale and S.H. Springer Verlag (1977). [57] J. [64] E.D.
Nonlinear Anal. Ordinary diﬀerential equations which yields periodic solutions of diﬀerential delay equations. J. Kappel. Kunish and W. Edinburgh 122A (1992). Princeton University Press. J. Stable and unstable manifolds for partial functional diﬀerential equations. MalletParet. Differential Equations 50 (1983). Memory. Lin. Marrakech June 1995. 1999. 163–183. Wu. (1976). Huang. McCracken. [69] T. Walther and J. Centre manifolds for partial diﬀerential equations with delays. H. Roy. Nonlinear Anal. . Diﬀ. New York. New Jersey. [71] K. Necessary conditions for partial diﬀerential equations with delay to generate a C0 semigroup. 48. “Shape. Schappacher. Kunish and W. Diﬀerential Equations 4 (1968). Levinger. Generic periodic solutions of functional diﬀerential equation. [75] J. Soc. J. Appl. 49–79. J. Attractivity and Hopf bifurcation. Wu. Smoothness and Invariant Stractiﬁcation of an Attracting Set for Delayed Monotone Positive Feedback”.W. (thesis) T¨ ubingen (1986). The Fields Institute Monograph Series. The Hopf bifurcation and its applications. Variation of constants formula for partial diﬀerential equations with delay. Schappacher. Yorke. Anal. 612–619. [72] B. TMA 16 (1991). [73] X. of Math Studies 17. SpringerVerlag.E. Integ. [68] H. [78] P. 123–142.C. [76] J.O. Equations 4 (1991). cours de l’´ecole d’´et´e sur les ´equations diﬀerentielles `a retard. 1974. “Probl`eme g´en´eral de la stabilit´e du mouvement”. A Folk theorem in functional diﬀerential equations.M. J. Lyapounov. Ann. Linear autonomous functional diﬀerential equations. [66] J.M. J. Salvadori. Math. On asymptotical stability for a linear delay equations. American Mathematical Society.280 DELAY DIFFERENTIAL EQUATIONS [65] W. So and J. Proc. TMA 5 (1981). Negrini and L. Kaplan and J. 237–254. (1947). Integrated semigroups. Nonlinear Anal. [67] F. [74] A.W.H.A 3 (1979). [77] M. Kellerman. Diﬀerential Equations 25 (1977). [70] K. 1303–1316. Marsden and M. Providence. T. 131–142. Krisztin.
Integrated semigroups and integrated solutions to abstract cauchy problems. . Travis and G. Integrated semigroups and their applications to the abstract Cauchy problems. 238 (1978).M. 58 (1977). Appl. Periodic solutions of some nonlinear functional diﬀerential equations. Math. [81] R. [85] W. Stech. Eqs. Mat. Anal. 152 (1990). New York. Rabat. [86] W. Pazy. Maroc (1992). J. Contribution a` l’´etude du probl`eme de bifurcation de Hopf locale dans le cadre des ´equations diﬀ´erentielles `a retard.F. Periodic solutions of some nonlinear functional diﬀerential equations. Takens. J.C. 200 (1974). 109 (1985). Math. Math. 198 (1996). Talibi. Paciﬁc J. Existence and stability for partial functional diﬀerential equations. Diﬀerential Equations 14 (1973). “Semigroups of Linear Operators and Apllications to partial diﬀerential Equations”. 416–447. [80] R. Unfolding of certain singularity of vector ﬁeld . J. Anal.D. 135 (1988). 843–876.H. 476–493. Doctorat d’Etat de l’Universit´e Mohamed V. 263–338. Thieme. Soc. Summers. Soc. Linearized stability for abstract diﬀerential equations with delay. J. Lay. Ruess. [84] A.REFERENCES 281 [79] F. Ann. Willey. SpringerVerlag. Taylor and D.C. The Hopf bifurcation: a stability result and applications. Anal. [83] R. 1980.D. Nussbaum. Pura Appl. Amer. “Introduction to Functional Analysis”. Amer. Trans. Nussbaum. 101 (1974). Math. Math. generalized Hopf bifurcation. Advances in Diﬀ. [82] R. Nussbaum. Appl. [92] C. [91] H. [90] A. 4 (1999). 395–418.M. New York. 310–336. 1983. J. II. Math. 111–155. Existence and stability of solutions to partial functional diﬀerential equations with delay. The range of periods of periodic solutions of x’(t)=αf (x(t − 1)). Appl. 472–491. [89] F. Trans. A global Hopf bifurcation theorem of functional differential systems. Ruess and W. Webb.W. Neubrander. 360– 394. Math. Nussbaum. Anal. Diﬀerential Equations 14 (1973). [87] H. 139–164. Appl. J.E. [88] H.
“Theory and Applications of Partial Functional Diﬀerential Equations”. Z. Mathematical and Computer Modelling 30 (1999). 117–138. Walther.M. Fields Institute Communications 29 (2001). Expositions Dynam. [100] E. A theorem on the amplitudes of periodic solutions of diﬀerential delay equations with application to bifurcation. 194 (1955).O. Trans. Wu. Synchronization and stable phaselocking in a network of neurons with memory. S. 56 (1976). Reine Angew. Weedermann. T. Faria and Y. . Math. 12 (1982). Bifurcation from solutions in functional diﬀerential equations. “Theory of Nonlinear AgeDependent Population Dynamics”. [95] A. [104] K. 361–368. Travis and G. Webb. [98] G. 66–87.C. A non linear diﬀerential diﬀerence equation. Iooss. Amer. Report. Appl.C. New York.S. New York. Wu. Marcel Dekker. Diﬀerential Equations 176 (2001). 182 (1983).F.282 DELAY DIFFERENTIAL EQUATIONS [93] C. J. 495–510. [101] J. Yoshida.F. 321–348. Walther. [96] H. J. J. Vanderbauwhede and G. Partial diﬀerential equations with deviating arguments in the time variable. [103] D. stability and compactness in the αnorm for partial functional diﬀerential equations. Existence. Webb. (1985). Ruan. J. Dynam. Wright. The Hopf bifurcation and its stability for semilinear diﬀusion equations with time delay arising in ecology. 1996. Travis and G.F. Center manifold theory in inﬁnite dimension. Xiao and S. Math. Hiroshima Math. Math.) 1 (1992). Math. Huang. Multiple bifurcations in a delayed predatorprey system with nonmonotonic functional response. Diﬀerential Equations 29 (1978). [102] J. Soc.O. J. 240 (1978). Systems (N. Normal forms for neutral functional diﬀerential equations. Webb. [99] M. 396–404. SpringerVerlag. 129–143. Anal. [94] C. 397–409. [97] H.
[36]). which joins spatial diﬀusion with demographic reaction.upm. This is the situation in the ﬁrst example that we are going to present. When thinking about extending functional diﬀerential equations from ﬁnite to inﬁnite dimensions. which is a model proposed by A.– France. Industriales . But it is not the only situation one can think of: we also present a second example which arises from a model of cell proliferation ﬁrst introduced by M. esanchez@math. c.ird. (eds.– 28006 Madrid. While this technique oﬀers the advantage of allowing explicit integration in terms of the initial function. ovide. Introduction This section is intended to motivate the interest in abstract diﬀerential delay equations (DDE) by means of examples.S.es 1. El Idrissi ([11]).– IRD.I.– Spain. Arino et al.etsii. it has also some deﬁciencies specially when looking at nonlinear perturbations 285 O.– E.– 93143 Bondy cedex. Calsina and O.T. Kimmel et al. Av. one of the ﬁrst and main examples which comes to mind is the case of evolution equations combining diﬀusion and delayed reaction. Henri Varagnat.arino@bondy. .).fr E. 285–346. it is viewed as an integrodiﬀerence equation and solved using a stepbystep procedure ([2]. Delay Differential Equations and Applications.Chapter 8 A THEORY OF LINEAR DELAY DIFFERENTIAL EQUATIONS IN INFINITE DIMENSIONAL SPACES O. S´ anchez Departamento de Matem´ aticas. [3]. 32. 2. Arino UR G´eodes. Centre de Bondy. In most treatments. Jos´e Guti´errez Abascal. © 2006 Springer. Several modiﬁed versions of this model were proposed and studied later. in [24].
the model is formulated in terms of the following state variables: i) u(a. reads as: ∂u ∂ ∂u + (a. 1.1 A model of ﬁsh population dynamics with spatial diﬀusion ([11]) Let us consider a ﬁsh population divided into two stages. z) is the density of the adult population at depth z at time t. ii) v(t. The functions mi (r) are mortality rates for the young ﬁsh (i = 1) and the adults (i = 2). t. r(0) = r0 . z)  ∗ l g(r(t)) − h L = u(a. The system ∂u (a. iii) r(t) is the maximum available resource at time t. distributed within the water column. z) − m2 (r(t))v(t. l] is the age. Then.) 0 z∗ = . z ∗ ] is the position in the water column. The adults are structured by the position they had when turning adult. positive on [0.1) In the above system: The function k(z) represents the mixing coeﬃcient due to water turbulence. t. . t. ξ) dξ 0 = 0 = u0 (a. oriented positively from the water surface to the bottom. 0. . and z ∈ [0. A crucial assumption is that only the youngsters may move in the water column randomly. 0) ∂z u(a.. t. t. z r(t) 0 b(z. at depth z. z) = v0 (z) . ξ)v(t. r.286 DELAY DIFFERENTIAL EQUATIONS and dynamical features. z ∗ ) and k(z ∗ ) = 0. where l < +∞ is the age of maturity. They are smooth and decreasing with a positive inﬁmum. z) is the density of young ﬁsh of age a. v. v(t. t. t. young and adults. It is continuous. z) ∂t r (t) u(0. z) ∂a ∂v (t. (1. their movement being described by a diﬀusion coeﬃcient. which is represented by an interval of the real line. z) dadz. at time t. z) = k(z) (a. which are better studied in the framework of diﬀerential equations than in the one of integral equations. a ∈ [0. We introduce the following variables: t ∈ R is the chronological time. v(0. t. t. z) ∂t ∂z ∂z = u(l. z) ∂u (a. z) . z) − m1 (r(t))u(a. z) of equations satisﬁed by u. where z ∗ is the depth.
L is a positive linear functional which gives the total potential pressure of the ﬁsh on the resource. z) = w0 (z) in terms of the fundamental solution of the problem. By concentrating on age cohorts. The function g(r) is smooth and decreasing and changes signs at some critical value rc . It may be interpreted as the logistic which would govern the dynamics of the resource in the absence of consumption by ﬁsh. z) = v0 (t. z. for 0 ≤ t ≤ l r(t) = r0 (t) . ξ) 0 dξ 0 −m2 (r(t))v(t. z) ∂t − = e 0 m1 (r(t+s)) ds −l  z∗ . with h(0) = 0.287 A theory of linear DDE in inﬁnite dimensional spaces The function b = b(z. ξ) 0 0 .1) leads to the following ∂v (t. the eggs produced at any level have a density lower than the water and go to the surface as a result of Archimedes law. b(ξ. for some physiological reason.e. z) . ζ) dζ K(l. for t > l r (t) = (g(r(t)) − h[L(p(t). z. z) = [k(z)w(t. while the function h is a bounded smooth and increasing function. i. ξ) is the fertility rate of the adults combined with a mixing process of the newborns throughout the water column: it can be for example that. z) . v(t. and using the representation of the solution to the Cauchy problem for the parabolic equation ∂ ∂w (t. z. . z∗ b(ξ. v(t. z) = 0 system (1. ζ) dζ dξ dzda.))])r(t) . for 0 ≤ t ≤ l for t > l (1. which models a limiting eﬀect on resource accessibility caused by overcrowding. z)] ∂t ∂z ∂w (t. ξ)w0 (ξ) dξ w(t.2) where p(t) is given by l z∗ p(t) = 0 − −l m1 (rt (s)) ds e 0 0  z∗ z∗ K(a. groups of individuals born at the same time. 0) = 0 ∂z w(0. namely ([89]): z∗ K(t. ζ)v(t − l. . ζ)v(t − a.
2) are also solutions of (1. The model considers the cells only at two distinct stages: . systems (1. these quantities are smooth functions of v(t.) and r(t − . we can see that each of them is made up of a combination of expressions involving the values of the unknown functions v or r either at time t or at some earlier time: the integral in the exponential term is expressed in terms of the values of r over the interval [t − l. Of course.1).2) when the problem to be solved is (1. choosing a pair of functions v0 . we could not do it using system (1. r0 arbitrarily.2).2 An abstract diﬀerential equation arising from cell population dynamics In the model elaborated by M. The diﬀerence with the initial formulation is that system (1. The second and main reason is that real systems are normally expected to converge asymptotically towards their attractor where the transients entailed by a particular initial condition vanish: in this respect.1) involves partial derivatives with respect to z and so requires additional regularity on the functions v and r to make sense. .) or v(t − .1) and (1. we can always solve (1. 1. We would have to use (1.2) are indistinguishable. Notice that (1. in most cases. Under some reasonable assumptions on the coeﬃcients. by setting up a ﬁxed point problem which can be shown to be a strict contraction for time small enough. after division (by mitosis) is divided in some stochastic way between the two daughter cells. l] and we would then be able to solve the equation for t > l by using (1. Kimmel et al. while the study of qualitative properties of (1.2). Only some of the solutions of (1. . It is a system of delay diﬀerential equations: even if it is formulated in a space of inﬁnite dimensions..1) to build up the solution on the interval [0.1) has been formulated from t = 0 while (1. On the other hand.1)? There are at least two reasons: ﬁrst of all. at least conceptually.1). This means that if we were to solve the equation with initial value given at t = 0.). So what is the point of looking at (1.2) is much simpler than the one for (1. cells are classiﬁed according to some constituent which. at least theoretically. it does not involve an unbounded operator and can be solved.1).2) directly.288 DELAY DIFFERENTIAL EQUATIONS Looking at the right hand side of the two equations in system (1. ζ). It is in fact a delay diﬀerential equation deﬁned on a space of functions. easier than that of a system of partial diﬀerential equations. what we will obtain is not a solution of (1.2) only holds for t > l.2) is.2). t] while the other integral uses v(t − l. and reported in [24]. the Cauchy problem associated with system (1.
. x) dtdx. the mother cells). a function f (x. x) dtdx. x): t2 x2 m(t. just after division. the number of cells going through mitosis (that is. y) > 0 if and only if d1 y < x < d2 y. b]. Ψ(+∞) > 0. y) such that: x2 f (x. continuously diﬀerentiable and such that inf a≤x≤b Ψ (x) > 0 on each interval [a. with a size in some size interval [x1 . x): t2 x2 n(t.A theory of linear DDE in inﬁnite dimensional spaces 289 a) At mitosis. x2 ]. with a size in some size interval [x1 . t1 x1 b) At the beginning of the cycle. the number of cells born during the time interval [t1 . 0 < d1 < 1/2 < d2 < 1. Φ(x) > x for x > 0. t2 ]. 0 < a ≤ b < +∞. b]. t1 x1 The distribution of sizes from mother to daughter is modelled by a conditional probability density. is determined in terms of a density function m(t. y) dx = probability for a daughter cell born from a mother cell x1 of size y to have size between x1 and x2 . in some time interval [t1 . x2 ] is determined in terms of a density function n(t. such that f (x. d2 . where Ψ is positive. Other conditions on f include a support property: it is assumed that there exist d1 . 2) The size at division in terms of the initial size y = Φ(x). where Φ is increasing. The model is completed by two functions describing respectively: 1) The duration of a cycle in terms of the initial size T = Ψ(x). Φ is continuously diﬀerentiable and such that inf a≤x≤b Φ (x) > 0 on each interval [a. t2 ]. decreasing. 0 < a ≤ b < +∞.
Then.4). large enough. (1. y) dy.3) is explicitly determined in terms of the initial function p. 0 < A1 ≤ A2 < +∞. Note that for such cells the cell cycle length is bounded above by the number r = Ψ(A1 ). x) = 2 0 The assumption on the support of f induces a support property for n. Assuming that d1 Φ (0) > 1 and d2 lim sup x→+∞ Φ(x) < 1. moreover. x ∈ (A1 . Ψ(A2 )) × (A1 . that is. let us point out that the righthand side of (1. A2 ) by a straight integration of the initial function.3) n(t.3) yields: n(t. t ≥ 0: nt (θ. the solution can be computed on the set (0. that is. A2 ). and s < t. 0) × (A1 . then equation (1. for all 0 < t < Ψ(A2 ). and A2 < +∞. taking nψ(A2 ) as a new initial function. −r ≤ θ < 0 . A2 as we just said.3) yields a positive semigroup in the space E = L1 ((−r. the size interval [A1 . all generations of cells born from cells with initial size in [A1 . with n0 = p. x we deduce from the above remark that for any A1 > 0 small enough.3)) for all t ≥ 0. x) = 0 for all x outside an interval [A1 .4) Assuming all the above conditions hold. A2 ] is invariant. determines a strongly continuous semigroup of positive operators on E ([84]) which. x) . (1. A2 ] will have size at birth in the same interval. where n veriﬁes equation (1. As a brief comment on the original proof of this result. the one parameter family of maps which to any initial function p ∈ E associates the function denoted nt . Eliminating m from the two equations leads to +∞ f (x. choosing A1 . Φ(y)) n (t − Ψ(y). if we assume that n(s. one can derive two fundamental equations relating the state functions n and m (see [24]). x) = 0 for x < d1 Φ(A1 ) or x > d2 Φ(A2 ).290 DELAY DIFFERENTIAL EQUATIONS From these basic assumptions. . that is. So. it was proved in [24] that equation (1. A2 )) . are compact for t large enough. A2 ]. r being given by formula (1. x) := n(t + θ.
The expression λσ (N (t − τ )) models a control mechanism which can be interpreted as a rate of defective division. λ is a coeﬃcient of magnitude 0 < λ < 1. Thus. by induction. in very restricted situations.5) 0 where +∞  . 2Ψ(A2 )) × (A1 . computing the solutions is almost a trivial task. assumed to be C 1 . In the case τ > 0. Φ(y)) n (t − Ψ(y). For this purpose. the case τ = 0 was studied. We will mention only one such model ([5]): n(t.5) may change. (1. the diﬀerence equation approach is not wellsuited. then. A reasonable conjecture is that these large oscillations are preceded by small periodic oscillations taking place in the vicinity of the nontrivial equilibrium.3) were introduced. [4]. zero loses its stability and a new global equilibrium arises. x) = 2λσ (N (t − τ )) +∞ f (x.6) t−Ψ(x) N (t) represents the number of cells present at time t. . Note that the principle of integration of such equations is the same as in the case of diﬀerence equations. the solutions converge to zero. It was proved that for λ less than some value. 0 < σ(x) < 1. stability properties of equilibria of equation (1. In [5] some global results regarding the existence of slowly oscillating periodic solutions were obtained. As λ increases from 0 to 1. due to a local Hopf bifurcation. A2 ) veriﬁes equation (1. In [3]. the problem is far more complicated.291 A theory of linear DDE in inﬁnite dimensional spaces the solution can be computed on (Ψ(A2 ). while. A2 ). (1.3) for all t > 0. This rate is shaped by the function σ. [5]. One of the main motivations of our eﬀort in the direction of abstract delay equations lies in the fact that we aimed at investigating the above conjecture. it can be computed on each set (kΨ(A2 ). t N (t) = n(s. the only not completely obvious thing being that the function deﬁned by putting end to end the pieces corresponding to each set (kΨ(A2 ). above this value. and decreasing. (k + 1)Ψ(A2 )) × (A1 . nonlinear extensions of equation (1. y) dy. x) ds 0 dx. σ being unchanged. A2 ) for each k ≥ 1. In [3]. (k + 1)Ψ(A2 )) × (A1 .
∂y from which we deduce the following expression for ∂1 n: ∂ 1 [n (t − Ψ(y). it involves a theory that is not available yet. We will now derive two diﬀerential equations from equation (1. the second one). Although it is probably the best framework in the sense that it leads to the same semigroup as equation (1. one in which the solutions of interest determine a proper subset. Taking the derivative with respect to t on both sides of equation (1.– Consider a solution n of equation (1. stability and bifurcation of equilibria in the larger system reﬂect similar properties for the smaller one. ∂t 0 We have ∂ [n (t − Ψ(y). Φ(y)) (∂1 n) (t − Ψ(y).3. The computations we will perform are mostly formal. We will use the notations ∂1 n (resp. This fact amply justiﬁes in our view the study of this system which. Assume that n can be diﬀerentiated with respect to time and size.3). The theory underlying such an equation is the theory of abstract delay diﬀerential equations with unbounded operators. However. a retarded version of the LotkaVolterra equation of demography ([42]). is much simpler than the smaller one. y) dy.3). They could be justiﬁed afterwards. y) + ∂2 n (t − Ψ(y). The ﬁrst one leads to a partial diﬀerential equation with retarded arguments.292 1. [7]). (∂1 n) (t − Ψ(y).3 DELAY DIFFERENTIAL EQUATIONS From integrodiﬀerence to abstract delay diﬀerential equations ([8]) 1.3). Retarded partial diﬀerential equation formulation of equation (1. The main shortcoming of the second equation is that it leads to a dynamical system much bigger than the one under study. The second one is a diﬀerential equation with retarded arguments and a bounded operator.3). y) . ∂2 n) to denote the function obtained by taking the partial derivative of n with respect to its ﬁrst variable (resp. y) .1 The linear equation. y)] = −Ψ (y)∂1 n (t − Ψ(y). for example. on the other hand. y)] − ∂2 n (t − Ψ(y). It ﬁts into the framework within which the extension of the theory of delay diﬀerential equations will be developed later ([58]. some of the dynamical properties of the larger system also hold for the smaller one. we arrive at +∞ ∂n =2 f (x.3). y) = − Ψ (y) ∂y .
y) dy 2 Ψ (y) 0 corresponds to ∂l/∂a(t. that is. we obtain +∞ f (x. and integrate it to get back to equation (1. y) dy ∂t ∂y Ψ (y) 0 +∞ f (x. (1. A2 ]. an extension of the Lotkavon Foerster equation for agedependent populations ([42]).7) is a sort of functional transport equation. y) dy. Φ(y)) n (−Ψ(y). in order to come back to equation (1. then.3). y) dy −2 ∂y Ψ (y) 0 . x)) in the cell population model ([13]). y) dy. it is necessary to complement the partial diﬀerential equation with the boundary condition +∞ f (x. we start from equation (1. Φ(y)) ∂ ∂n (t. x) = 0 for x < A1 or x > A2 .7) exists inL . which yields the following +∞ ∂n ∂ f (x. x) whose support in x is contained in [A1 . Φ(y)) +2 ∂2 n (t − Ψ(y). The term +∞ f (x. x) = 2 0 Equation (1. y)] dy ∂t Ψ (y) ∂y 0 +∞ f (x. x) = 2 n (t − Ψ(y). Φ(y)) ∂2 n (t − Ψ(y).8) Ψ (y) 0 If.7). A ) 1 2 ∂y Ψ (y) then we can use integration by parts on the ﬁrst integral in the above expression of ∂n/∂t. n(t. x) = 2 0 So. Φ(y)) ∂2 n (t − Ψ(y). we obtain: +∞ f (x. a) in the classical SharpeLotka model of demography.9) n(0. or ∂/∂x (g(x)p(t. x) = −2 [n (t − Ψ(y). (1. y)] dy. y) dy.3). Φ(y)) [n (t − Ψ(y). y) − n (−Ψ(y). Φ(y)) (t. or the BellAnderson equation ([13]) for sizedependent cell populations. x) − n(0. If we assume further that ∂ f (x. in fact.A theory of linear DDE in inﬁnite dimensional spaces 293 Inserting the righthand side of the above in the expression of ∂n/∂t. The quantity +∞ ∂ f (x. Φ(y)) n (t − Ψ(y). we remind the reader that we work with functions n(t. n(t. (1. Φ(y)) 1 2 (A . +2 Ψ (y) 0 At this point. conversely.
Φ(y)) f (y. We summarize the relationship between the integrodiﬀerence equation and the P.3).294 DELAY DIFFERENTIAL EQUATIONS accounts for the mortality term.11) provided that we assume ∂ f (x.D.9). integration by parts using the fact that n(t. ·) has support in [A1 . as follows: Proposition 36 Assume all the above stated conditions on f . equation ( 1. and taking the derivative of (1. Ψ (y) (1.11): +∞ +∞ (Lp)(x) := 4 0 0 ∂ ∂y f (x. Delay diﬀerential equation formulation of equation (1. Φ(z)) n (t − Ψ(y) − Ψ(z). Ψ (y) ∂y Once again. z) dzdy. Φ(y)) f (y. 0]. Φ(y)) f (y. =4 ∂t 0 0 The quantity ∂1 n (t − Ψ(y) − Ψ(z). Φ(z)) n (t − Ψ(y) − Ψ(z). A2 )3 . including condition given by formula ( 1. x) = 4 ∂t +∞ 0 +∞ 0 ∂ ∂y f (x. Φ(z)) p (−Ψ(y) − Ψ(z). z) dzdy. z) = − 1 ∂ [n (t − Ψ(y) − Ψ(z).7)( 1. n(t. A2 ]. A2 ) −→ L1 (A1 . we can express n (t − Ψ(y). Φ(y)) f (y. z)] . Φ(y)) f (y. ∂y Ψ (y) Denoting L the linear operator: L : C [−2r. Φ(z)) ∂1 n (t − Ψ(y) − Ψ(z).10) on both sides. A2 ) deﬁned by the righthand side of equation (1. y) in integral form. Then.– If in equation (1.7). we arrive at +∞ +∞ ∂n f (x.3) is equivalent to the system ( 1. z) dzdy.E.3) we let t > r. Ψ (y) . Φ and Ψ. z) dzdy. yields the desired equation: ∂n (t. L1 (A1 . x) = 4 0 0 (1. This leads to the equation +∞ +∞ f (x.10) Assuming that n can be diﬀerentiated with respect to t. z) can be interpreted as the derivative with respect to y: ∂1 n (t − Ψ(y) − Ψ(z). Φ(z)) exists in L1 (A1 .
1. n is a solution of (1. conversely.5) we let t > r. Φ(y)) f (y.3) for all t > t0 . integrating the equation on both sides from 0 to t we obtain. Φ(y)) f (y.295 A theory of linear DDE in inﬁnite dimensional spaces equation (1.(1.5)Note that in the case of system (1. Φ(z)) n (−Ψ(y) − Ψ(z). If. Φ(z)) × 0 ×σ (N (t − Ψ(y) − τ )) n (t − Ψ(y) − Ψ(z).12) Equation (1. than equation (1.12) is an example of what we call abstract linear delay differential equations. z) dzdy. 0 (1.3). This leads to the equation: n(t. Φ(z)) n (t − Ψ(y) − Ψ(z).2 Delay diﬀerential equation formulation of system (1. y) in integral form.3) for all t in (t0 . x) = 4 0 f (x. z) dzdy 0 +∞ +∞ −4 0 f (x.6) . t0 + r).11) reads ∂n = L(nt ). in equation (1.15) . x) = 4λ2 σ (N (t − τ )) 0 +∞ +∞ f (x.12). mum delay is τ + r.3).13) then the same equality holds for all t > 0. we can express n (t − Ψ(y). that is +∞ +∞ n(t. x) = +∞ +∞ 4 0 f (x. z) dzdy . x) = 4 0 f (x. If.6). x) − n(0.3. Φ(z)) n (t − Ψ(y) − Ψ(z). Φ(y)) f (y. the maxi(1.14) and. 0 If we assume that +∞ +∞ n(0. z) dzdy . A very partial converse result in this respect is the following: if n is a solution of equation (1. n satisﬁes (1. for some t0 . and leads to a much bigger set of solutions than the one determined by (1. however. after integration by parts: n(t. ∂t (1. Φ(z)) n (−Ψ(y) − Ψ(z). 0 (1. Φ(y)) f (y.14) Equation (1. then n satisﬁes (1.5)(1. Φ(y)) f (y. z) dzdy.14) is strictly more general.
which in particular implies that N is locally Lipschitz continuous. x) ∂t = n(t.5)). x) σ (N (t − τ )) +∞ +∞ f (x. (1. Φ(z)) × 4λ2 σ (N (t − τ )) ∂y Ψ (y) 0 0 ×σ (N (t − Ψ(y) − τ )) n (t − Ψ(y) − Ψ(z).17). y)] dy n(t. =− Ψ (y) ∂y Integration by parts with respect to y of the last integral on the righthand side of (1. equation (1. z)] ∂t 1 ∂ [σ (N (t − Ψ(y) − τ )) n (t − Ψ(y) − Ψ(z). Φ(y)) f (y. Assuming f bounded on bounded rectangles. and: +∞ N˙ (t) = [n(t.14) in the following way: A2 t−Ψ(A1 ) f x. z) dzdy. and taking the derivative with respect to t on both sides of (1. Φ(y)) f (y. x)] dx. Using the above expression and formula (1. x) − n (t − Ψ(x). T ] × R+ .14). x) σ (N (t − τ )) 0 +∞ +∞ ∂ f (x. we obtain: ∂n (t.17) 0 We note the same property as in the previous computation: ∂ [σ(N (t − Ψ(y) − τ ))n(t − Ψ(y) − Ψ(z). x) ∂t = +∞ σ (N (t − τ )) [n(t − τ. ·) is deﬁned on R (for example.296 DELAY DIFFERENTIAL EQUATIONS Assuming that n can be diﬀerentiated with respect to t. Φ(y)) f (y. z)] dzdy. (1. z) dzdy.15) yields the following ∂n (t.16) ∂t We point out the fact that N can indeed be diﬀerentiated. In this case. (T > 0).18) Assume that n(t. z)] . Assuming moreover that f is C 1 . Φ(z)) × +4λ2 σ (N (t − τ )) σ (N (t − τ )) N˙ (t − τ ) 0 0 ∂ × [σ (N (t − Ψ(y) − τ )) n (t − Ψ(y) − Ψ(z). we conclude that n is bounded and continuous on each set [−T. n is a periodic solution of (1. Φ(z)) +4λ2 σ (N (t − τ )) × ∂y Ψ (y) 0 0 ×σ (N (t − Ψ(y) − τ )) n (t − Ψ(y) − Ψ(z).15) transforms this integral into +∞ +∞ ∂ f (x.14). n satisﬁes equation (1. y) − n (t − τ − Ψ(y). Φ(z) × A1 t−Ψ(A2 ) . (1. it is possible to write the integral in equation (1. Φ Ψ−1 (t − s) f Ψ−1 (t − s).
A theory of linear DDE in inﬁnite dimensional spaces
297
×σ (N (s − τ )) n (s − Ψ(z), z) (Ψ−1 ) (t − s) dsdz ,
which implies that n is locally Lipschitz in t, uniformly in x, in other
words, there exists N , a negligible set of tvalues, such that n(t, x) is
diﬀerentiable for all t not in N , and all x ∈ (A1 , A2 ). In conclusion, we
have the following
Lemma 32 If n is a solution of ( 1.5), deﬁned for every t in R, then n
is diﬀerentiable in t with ∂n/∂t continuous in R×R, satisfying equation
( 1.17).
Equation (1.17) is a nonlinear delay diﬀerential equation on the space
with maximum delay equal to 2r + τ . As in the linear case,
it is obvious, by construction, that the solutions of the system (1.5)(1.6)
are solutions of equation (1.17), but equation (1.17) has a much bigger
set of solutions than system (1.5)(1.6). The motivation for embedding
the latter system into a bigger one is that the framework of diﬀerential
equations seems to be more suited to the study of dynamical features
such as stability than the framework of integral equations.
L1 (A1 , A2 ),
1.4
The linearized equation of equation (1.17)
near nontrivial steadystates
Equation (1.17) is the most complete of all the equations considered
here. It can be reduced to equation (1.11) by assuming that 2λσ(N ) =
1. If the delay τ = 0, we obtain the equation studied in [3]. Our
program regarding equation (1.17) is to start from the steadystates and
investigate their stability and the loss of stability in terms of λ. While
we already know from [3] that stability passes from a constant steadystate to another one when τ = 0, we expect that in the delay case stable
oscillations may take place. This fact has been shown under several
restrictions in [5]: existence of periodic slowly oscillating solutions has
been demonstrated for a range of values of the parameters. However, the
result is not explicitly related to any bifurcation of steadystate periodic
solution. We believe however that there is a strong possibility that such
a bifurcation occurs at these values of λ where some eigenvalues of the
linearized equation cross the imaginary axis (away from zero).
1.4.1
The steadystate equation.
Note that any function
n(x) of x only veriﬁes equation (1.17) trivially. If n = n, we have
N (t) = N =
+∞
Ψ(x)n(x) dx.
0
298
DELAY DIFFERENTIAL EQUATIONS
From them we will only retain those which are nonnegative solutions of
the system (1.5)(1.6). Equation (1.5) leads to
+∞
n(x) = 2λσ(N )
f (x, Φ(y)) n(y) dy.
0
Integrating the above equation on both sides and using the fact that
f (·, Φ(y)) is a probability density, we obtain
2λσ(N ) = 1
and
+∞
n(x) =
f (x, Φ(y)) n(y) dy.
(1.19)
0
Equation (1.19) is, notably, dealt with in detail in [2]. It is a typical example of application of the PerronFrobenius theory of positive operators
on ordered Banach spaces.
Equation (1.19) has a unique positive solution n1 such that
+∞
0
n1 (x) dx = 1.
In terms of n1 , n = kn1 , where k is a constant. The constant k satisﬁes
the relation
+∞
2λσ k
Ψ(x)n1 (x) dx = 1.
(1.20)
0
Assuming that
σ(0) >
1
2
equation (1.20) has, for each λ >
;
σ(+∞) = 0,
1
, one and only one root k = k(λ).
2σ(0)
The function k is increasing.
1.4.2
Linearization of equation (1.17) near (n, N ) .
The
linearization is obtained by formally diﬀerentiating equation (1.17) with
respect to every occurrence of n as if it were an independent variable
and multiplying the result by the same occurrence of the variation ∆n.
This yields the following:
∂
(∆n)(t, x)
∂t
=
σ (N )
n(x)
σ(N )
+∞
[(∆n)(t − τ, y) − (∆n)(t − τ − Ψ(y), y)] dy
0
+∞ +∞
∂ f (x, Φ(y)) f (y, Φ(z))
+4λ2 σ(N )
×
∂y
Ψ (y)
0
0
× σ (N )∆N (t − Ψ(y) − τ )n(z)
+σ(N )∆n (t − Ψ(y) − Ψ(z), z) dzdy.
A theory of linear DDE in inﬁnite dimensional spaces
299
The equation can be simpliﬁed a little bit. In view of relation (1.19),
we have
+∞ +∞
0
0
∂
∂y
+∞
f (x, Φ(y)) f (y, Φ(z))
∂ f (x, Φ(y)) n(y)
n(z)dzdy
=
dy.
Ψ (y)
∂y
Ψ (y)
0
This, in turn, leads to
+∞
0
0
0
+∞
=
+∞
=
f (x, Φ(y)) f (y, Φ(z))
∆N (t − Ψ(y) − τ ) n(z) dzdy
Ψ (y)
0
∂ f (x, Φ(y)) n(y)
∆N (t − Ψ(y) − τ ) dy
∂y
Ψ (y)
+∞
∂
∂y
f (x, Φ(y)) n(y)∂1 (∆N ) (t − Ψ(y) − τ ) dy
+∞
=
0
f (x, Φ(y)) n(y) ×
+∞
×
[(∆n) (t − Ψ(y) − τ, z) − (∆n) (t − Ψ(y) − Ψ(z) − τ, z)] dz dy.
0
We arrive at the following equation
σ (N )
∂
n(x)
(∆n)(t, x) =
∂t
σ(N )
+2λσ (N )
×
+∞
+∞
+
0
+∞
[(∆n)(t − τ, y) − (∆n) (t − τ − Ψ(y), y)] dy
0
f (x, Φ(y)) n(y) ×
0
[(∆n) (t − Ψ(y) − τ, z) − (∆n) (t − Ψ(y) − Ψ(z) − τ, z)] dz dy
0
+∞ +∞
0
∂
∂y
f (x, Φ(y)) f (y, Φ(z))
(∆n) (t − Ψ(y) − Ψ(z), z) dzdy.
Ψ (y)
(1.21)
1.4.3
Exponential solutions of (1.20) .
The study of the
stability of n as a solution of equation (1.17) amounts to the study of
the growth rate of the semigroup associated to equation (1.20). In the
case of a compact or eventually compact semigroup, this amounts to
determining exponential solutions (exp zt)x. Then, the growth rate is
negative if sup Re z < 0, and it is positive if sup Re z > 0. Bifurcation
of equilibria occurs at the values of parameters where the growth rate
changes signs. It can be shown that the semigroup associated to equation
(1.20) is not eventually compact, but its restriction to the elements of
300
DELAY DIFFERENTIAL EQUATIONS
C ([−r, 0]; X) satisfying (1.13) is. This remark motivates the study of
exponential solutions of equation (1.20).
Let ν(t, x) = ezt p(x) be a solution of equation (1.20). Substituting
for ∆n in (1.20) yields
ezt p(x)
zp(x)
=
A2
σ (N )
1 − e−zΨ(y) p(y) dy
n(x)
σ(N )
A1
A2
+2λσ (N )
f (x, Φ(y)) n(y) ×
e−zτ
A1
A2
×
−z(Ψ(y)+τ )
e
A1
A2 A2
+
A1
A1
∂
∂y
−zΨ(u)
1−e
p(u) du dy
f (x, Φ(y)) f (y, Φ(u)) −z(Ψ(y)+Ψ(u))
p(u)dudy.
e
Ψ (y)
(1.22)
To abbreviate, we will call such an equation a spectral equation. Note
that the righthand side of the above equation vanishes for z = 0. For
this value of z, every function p satisﬁes the equation. Equation (1.21)
is quite complicated. We will not study it directly. Instead, we will
now introduce the eigenvalue equation associated with the linearization
of equation (1.5). Then, we will show that both equations are partially
equivalent to each other.
Ideally, one could expect that this part of the spectrum is the part
brought up by the passage from the integral to the diﬀerential formulation, and the remaining part of the spectrum reﬂects the spectrum
of the integral equation only. The spectral equation associated to the
linearization of equation (1.5) near n is
. .

A2
−zΨ(y)
1
−
e
dy n(x)
p(y)
p(x) = e−zτ 2λσ (N )
z
A1
A2
f (x, Φ(y)) e−zΨ(y) p(y) dy.
(1.23)
+
A1
One can easily verify that every solution of equation (1.22) is also a
solution of equation (1.21). In the following Lemma, we state a partial
converse of this result.
Lemma 33 Denote σ1 (resp. σ2 ) the set of complex numbers z, z
= 0,
Re z ≥ 0, such that equation ( 1.21) (resp.) ( 1.22) ) has a nonzero
solution for the value z. Then, σ1 = σ2 .
A theory of linear DDE in inﬁnite dimensional spaces
301
Moreover, for each z ∈ σi , let N i z ( i = 1, 2) be the space of functions
p such that p is a solution of ( 1.21) (resp. ( 1.22) ) for the value z.
Then, N 1 z = N 2 z .
Proof.– For each z
= 0, we introduce the operators
A2
−zΨ(y)
1
−
e
Bz (p) := 2λσ (N )
p(y)
dy,
z
A1
A2
(Lz p) (x) :=
f (x, Φ(y)) e−zΨ(y) p(y) dy.
A1
Then equation (1.22) is
p(x) = e−zτ n(x)Bz (p) + (Lz p) (x),
and equation (1.21) reads
p(x) = e−zτ n(x)Bz (p) + (Lz q) (x),
where
q(x) = e−zτ n(x)Bz (p) + (Lz p) (x).
We need only prove that p, a solution of (1.21) for some z
= 0, with
Re z ≥ 0, satisﬁes (1.22) for the same value of z. But this is equivalent
to proving p = q. Since we have
p(x) − q(x) = (Lz (p − q)) (x),
and the equation
p(x) = (Lz p) (x)
can be satisﬁed with p
= 0 only if Re z < 0, the Lemma follows.
We will now brieﬂy discuss equation (1.22).
For z = 0, equation (1.22) reduces to
A2
p(y)Ψ(y) dy n(x)
p(x) = 2λσ (N )
A1
A2
+
f (x, Φ(y)) p(y) dy,
A1
which has p = 0 as the only solution.
In fact, integrating the above equation on both sides on (A1 , A2 ), we
obtain
A2
p(y)Ψ(y) dy = 0,
2λσ (N )
A1
302
DELAY DIFFERENTIAL EQUATIONS
which implies that
A2
p(x) =
f (x, Φ(y)) p(y) dy,
A1
and therefore, in view of (1.19):
p(x) = kn(x).
So,
A2
A2
p(y)Ψ(y) dy = k
0=
n(y)Ψ(y) dy.
A1
A1
This, with the fact n > 0, implies that k = 0, that is p = 0.
For z
= 0, integrating equation (1.22) on both sides on (A1 , A2 ), we
obtain
A2
A2
−zΨ(y)
−zτ
dy 2λσ (N )e
p(y) 1 − e
n(y) dy − z = 0.
A1
A1
(1.24)
Equation (1.24) breaks down into two equations
A2
p(y) 1 − e−zψ(y) dy = 0,
A1
zτ
ze
(1.25)
A2
= 2λσ (N )
n(y) dy.
(1.26)
A1
Assuming (1.25) holds, equation (1.22) yields, with z
= 0:
A2
p(x) =
f (x, Φ(y)) e−zΨ(y) p(y) dy,
A1
which can be satisﬁed with p
= 0 only if Re z < 0.
So, instability from nonzero roots crossing the imaginary axis can only
come from roots of equation (1.26). We have the following result:
For each λ > 0, equation (1.24) has a pair of nonzero roots crossing
the imaginary axis at each value τ = τk , where
τk =
π/2 + 2kπ
A
2λσ (N ) A12 n(y) dy
,
k ∈ N.
Once equation (1.24) has been solved for z
= 0, in order to have a full
solution of equation (1.21), it remains to solve equation (1.22) for p.
The equation can be written in the form
(I − K)p = 0,
A theory of linear DDE in inﬁnite dimensional spaces
303
where K is a compact operator. The fact that
A2
([I − K]p) (x) dx = 0
A1
for every p ∈ X, implies that Im (I − K) is a proper subspace of X, and
therefore, by virtue of a well known property of compact operators, the
null set of I − K is not reduced to 0.
1.5
Conclusion
The construction of a theory of delay diﬀerential equations in inﬁnite
dimensional spaces is motivated by the study of some integrodiﬀerence
equations arising in some models of cell population dynamics. The introduction of a nonlinearity in these models to account for possible selfregulating control mechanism of cell number suggested possible oscillations. The integrodiﬀerence nature of the model did not allow the use
of any known standard techniques to exhibit such oscillations.The observation that was made that this system (and its solutions) can in fact
be viewed as solutions of a larger system of delay diﬀerential equations
in an inﬁnite dimensional space led to the idea to extend this frame the
machinery for the study of stability that has been developed for delay
diﬀerential equations in ﬁnite dimensional spaces.
Let us conclude with a few comments about oscillations. Oscillation
with respect to time of the number of cells is a recognized fact in several
examples, among which is the production of red blood cells ([4], [17]).
This phenomenon can also be observed in cultured cells. It used to be a
common belief that such oscillations reﬂect the role of delays at various
stages of the life of cells. A lot of eﬀorts were made during the seventies
in the study of homogeneous nonstructured models with the goals of
ﬁrst proving existence of periodic oscillations, after that counting them
and looking at their stability and other properties. Amongst those who
contributed to these works let us mention [12], [22] and [41]. In contrast
to this, very little has been done in the case of structured models. One of
the reasons is that there is no general theory for the study of dynamical
properties of such equations. The theory on abstract D.D.E. that we
present here is a step in this direction.
2.
The Cauchy Problem For An Abstract Linear
Delay Diﬀerential Equation
Our interest in functional diﬀerential equations in inﬁnite dimensions
stemmed from the study of the dynamical properties of such equations
304
DELAY DIFFERENTIAL EQUATIONS
and the remark we made at some point that it is possible to associate
to any of them a delay diﬀerential equation on an inﬁnite dimensional
vector space. We start with this theory for the linear and autonomous
case and this section is devoted to the resolution of the Cauchy problem.
We start setting up the general framework of the equations under
study. The notation C stands for the space C([−r, 0]; E) (r > 0) of the
continuous functions from the interval [−r, 0] into E, E being a real
Banach space, endowed with the sup norm
ϕC := sup ϕ(θ)E .
θ∈[−r,0]
Let L : C −→ E be a bounded linear operator. We are considering the
following Cauchy problem for the delay equation deﬁned by the operator
L:
For each ϕ ∈ C, to ﬁnd a function x ∈ C([−r, +∞[; E), x ∈ C 1 ([0, +
∞[; E) such that the following relations hold
3
x (t) = L(xt ) , for all t > 0
(CP )
x0 = ϕ
As usual, xt denotes the section at t of the function x, namely xt (θ) :=
x(t + θ), θ ∈ [−r, 0]. The Cauchy problem is wellposed if it has one and
only one solution for each initial value ϕ.
2.1
Resolution of the Cauchy problem
There are several ways to solve problem (CP). The most elementary
one is to write the equation in integral form
t
L(xs ) ds , t > 0.
(2.1)
x(t) = x(0) +
0
Deﬁne the function
ϕ0
: [−r, +∞[−→ E by
3
ϕ(t) if t ∈ [−r, 0]
0
ϕ (t) :=
ϕ(0) if t ≥ 0.
The change of x by the unknown function
y(t) := x(t) − ϕ0 (t)
transforms equation (2.1) into
t
t
L(ys ) ds +
L(ϕ0s ) ds
y(t) =
0
0
(2.2)
A theory of linear DDE in inﬁnite dimensional spaces
305
which can be formulated as an abstract equation in a suitable framework.
To this end, for each α > 0 let us deﬁne
E0α := {y ∈ C([0, +∞); E) ; y(0) = 0 , yα := sup e−αt y(t)E < +∞}
t≥0
which is a Banach space with respect to the norm · α .
Let Q be the bounded linear operator deﬁned by
t
0
0
L(˜
ys ) ds
Q : Eα −→ Eα ; (Qy)(t) :=
0
where y˜ is the extension by zero of function y to [−r, 0]. In terms of this
operator, equation (2.2) reads as:
t
L(ϕ0s ) ds.
(Id − Q)(y)(t) =
0
The norm of Q on
E0α
satisﬁes the following estimate
Qyα ≤
L
yα
α
so, for α > L, Q is a strict contraction. This implies that equation
(2.1) has one and only one solution deﬁned on [−r, +∞[. This proves
that the Cauchy problem (CP) is wellposed.
As usual, bearing in mind that the initial state is the function ϕ and
the state at time t is the function xt , we introduce the solution operator
T (t) deﬁned by
T (t)ϕ := xt , t > 0.
It is easy to check that the family {T (t)}t≥0 is a strongly continuous
semigroup of linear bounded operators on C. An obvious property of
this semigroup is that
3
ϕ(t + θ)
if t + θ ≤ 0
[T (t)ϕ] (θ) =
[T (t + θ)ϕ] (0) if t + θ ≥ 0
that is, {T (t)}t≥0 is a translation semigroup.
Straightforward computation leads to the following operator as the
inﬁnitesimal generator of the semigroup:
Aϕ := ϕ˙
(2.3)
˙
= L(ϕ)}.
D(A) = {ϕ ∈ C 1 ([−r, 0]; E) ; ϕ(0)
(2.4)
with domain
306
2.2
DELAY DIFFERENTIAL EQUATIONS
Semigroup approach to the problem (CP)
Semigroups deﬁned on some functional space of the type F(J; E),
where J is an interval and F is some class of functions (continuous
functions, Lp functions, etc) and with inﬁnitesimal generator A deﬁned
by the derivative on J, have been investigated by several authors after
a pioneering work by A. Plant [34]. Extension of Plant’s results to
abstract delay diﬀerential equations can be found in [16] and in [73].
The following result is borrowed from [73].
Theorem 1 The operator A deﬁned by ( 2.3) with domain given by
( 2.4) is the inﬁnitesimal generator of a strongly continuous semigroup
{S(t)}t≥0 on C satisfying the translation property
3
ϕ(t + θ)
if t + θ ≤ 0
(S(t)ϕ)(θ) =
(S(t + θ)ϕ)(0) if t + θ > 0
t > 0, θ ∈ [−r, 0], ϕ ∈ C. Furthermore, for each ϕ ∈ C, deﬁne x :
[−r, +∞) −→ E by
3
ϕ(t)
if t ∈ [−r, 0]
x(t) :=
(S(t)ϕ)(0) if t > 0.
Then x is the unique solution of (CP) and S(t)ϕ = xt , t > 0.
Remark 1.– The last part of the theorem states that in fact S(t) =
T (t), t ≥ 0, which we can assert as soon as we know that {S(t)}t≥0 and
{T (t)}t≥0 are both associated with the same inﬁnitesimal generator.
Proof.– In order to illustrate the type of computations and ideas involved in delay equations, we will brieﬂy sketch the proof of the theorem.
First of all, we will check that A is the inﬁnitesimal generator of a
strongly continuous semigroup. Using the LumerPhillips version of the
HilleYosida theorem, we have to prove that:
i) A is a closed operator with a dense domain.
ii) The resolvent operator
R(λ, A) := (λI − A)−1
exists, for all real λ large enough and there exists a real constant
ω such that
(λI − A)−1 ≤
1
λ−ω
;
λ > ω.
307
A theory of linear DDE in inﬁnite dimensional spaces
We skip (i), which is obvious. In order to prove (ii), given f ∈ C we have
to solve
(λI − A)ϕ = f ; ϕ ∈ D(A).
(2.5)
The resolution of (2.5), which is straightforward, leads to
0
eλ(θ−s) f (s) ds.
ϕ(θ) = eλθ ϕ(0) +
(2.6)
θ
In order to verify that ϕ ∈ D(A), we have to check the boundary condition for ϕ:
0
λ·
λ(·−s)
e
f (s) ds .
(2.7)
λϕ(0) − f (0) = L e ϕ(0) +
·
Let us introduce the following bounded linear operators
Lλ : E −→ E
;
Lλ (z) := L(ελ ⊗ z)
where ελ ⊗ z ∈ C([−r, 0]; E) is deﬁned by
(ελ ⊗ z)(θ) := eλθ z
θ ∈ [−r, 0]
,
and Sλ : C −→ E deﬁned by
Sλ (f ) = f (0) + L
0
λ(·−s)
e
f (s) ds .
(2.8)
·
Lemma 34 For each λ ∈ C, the operator Sλ is onto.
Proof.–
Let us consider the family of bounded linear operators {Σλ,k }k=1,2,... ⊂ L(E)
deﬁned by
Σλ,k (z) := Sλ (χk ⊗ z)
where
(χk ⊗ z)(θ) := χk (θ)z
,
θ ∈ [−r, 0]
and {χk }k=1,2,... is the family of real functions deﬁned by
kθ
+1 , 0
, θ ∈ [−r, 0].
χk (θ) := max
r
It is easily seen that for each λ, the sequence of operators Σλ,k converges
in operator norm towards the identity. Therefore Σλ,k is invertible for k
large enough.
on any set of the type Ωa := {z ∈ C .9) It holds that Sλ ◦ Υλ. we can build a right inverse of Sλ . A).k ∈ L(E. A)(f )(θ) = eλθ ∆(λ)−1 Sλ (f ) + 0 eλ(θ−s) f (s) ds. This proves our claim. C) by ∀b ∈ E .k (b). we have Re λ ≥ a =⇒ Lλ ≤ M. λ−M .k (b) := χk ⊗ Σ−1 λ. 0]: R(λ.10) θ To complete the proof of the ﬁrst part of the theorem it only remains to obtain a convenient estimate for the norm of R(λ. there exists a0 ∈ R such that ∆(λ) is invertible all over regions of the complex plane Ωa with a ≥ a0 . So. Let us deﬁne M := sup Lλ . if we take a > 0 and a ≥ a0 large enough.308 DELAY DIFFERENTIAL EQUATIONS In terms of Σλ. Straightforward computation shows that the operator Lλ is uniformly bounded on each right halfplane of C (that is.k . (2. it holds that ϕ(0) = ∆(λ)−1 Sλ (f ) and. (2. we have the following characterization for λ to belong in the resolvent set of the operator A: Proposition 37 A necessary and suﬃcient condition for λ ∈ ρ(A) (resolvent set of A) is that the operator ∆(λ) deﬁned by ∆(λ) := λI − Lλ be invertible. In terms of operators Lλ and Sλ . let us deﬁne Υλ. To this end. So. Assuming that λ ∈ ρ(A). Υλ. for θ ∈ [−r. Re z ≥ a} for some a). we will have λ ≥ a =⇒ Lλ <1 λ and then ∆(λ)−1 ≤ 1 . Re λ≥a0 For any a ≥ a0 .k = IE where IE is the identity operator on E.
Note that the evaluation map ϕ → [S(t)ϕ] (0) is continuous and even locally uniformly continuous with respect to t. Therefore . A) ≤ max 1 + . λ λ−M λ Since the quantity inside brackets is a convex combination of two ﬁxed values. we assume that ϕ ∈ D(A). λ λ−M λ To conclude. Sλ (f ) ≤ 1 + λ Combining these estimates in formula (2. 0] What we have to verify is that S(t)ϕ = xt . t ≥ 0. a basic semigroup property ensures that S(t)ϕ ∈ D(A) for all t ≥ 0. So. and in fact it is also locally uniformly continuous with respect to t. it is bounded by the maximum of both quantities. From this remark it follows that we only have to prove the property on a dense subset of initial values. To this end.A theory of linear DDE in inﬁnite dimensional spaces 309 We can also readily check the following estimate for Sλ (f ) L f C . A) ≤ 1 λ−ω for λ > ω. if we choose ω so that ω > max(a0 . A)f ] (θ)E ≤ e + 1−e f C . from which we deduce L 1 1 R(λ.– It remains to check that the semigroup {S(t)}t≥0 enjoys the translation property. Translation semigroup property. 0]: L 1 λθ λθ 1 1+ [R(λ. for each ϕ given in C. for θ ∈ [−r. Then. It will extend by continuity to all initial values.10) we obtain. so the map ϕ → xt is continuous for each t ≥ 0. from this point on. . let us deﬁne the function 3 [S(t)ϕ] (0) if t ≥ 0 x(t) = ϕ(t) if t ∈ [−r. L + M ) we will have the desired inequality: R(λ.
We now consider two situations: i) t + θ ≤ 0. θ) and can be diﬀerentiated on this domain. Formula (2. 0]. θ). ii) t + θ ≥ 0. θ) = −t. max(−t.2. we give h the values θ and 0. the function h → [S(t + h)ϕ] (θ − h) makes sense as long as t + h ≥ 0 and θ − h ≤ 0. h ≥ max(−t. according to the deﬁnition of x. In this case. we have the following results. θ) and h ≥ max(−t.– First we suppose that R (λI − Lλ ) is a closed subspace of E and let {ψn }n=1.– It follows immediately from (2. Lemma 36 The subspace R (λI − A) is closed in C if and only if the subspace R (λI − Lλ ) is closed in E. Lemma 35 For any λ ∈ C. therefore. Giving h the values −t and 0.. The derivative is equal to d d S(t + h)ϕ (θ − h) − [S(t + h)ϕ] (θ − h) = 0.7) combined. This time. that is. In both cases.6) shows that the determination of ϕ such that (λI − A)ϕ = f amounts to determining ϕ(0) and formula (2. S(t)ϕ = xt . t ≥ 0. Given a pair (t.. Proof. we have that f ∈ R (λI − A) if and only if Sλ (f ) ∈ R (λI − Lλ ). θ). which leads to [S(t)ϕ] (θ) = [S(t + θ)ϕ] (0) for t + θ ≥ 0. we have [S(t)ϕ] (θ) = x(t + θ). .6) and (2.. dt dθ This implies that [S(t + h)ϕ] (θ − h) is a constant for each ﬁxed pair (t. θ ∈ [−r. 2. as desired. ⊂ R (λI − A) be a convergent sequence with ψ = limn→∞ ψn .7) gives a necessary and suﬃcient condition for existence of ϕ(0) which is precisely the condition stated in the lemma.310 DELAY DIFFERENTIAL EQUATIONS we have that both t → [S(t)ϕ] (θ) and θ → [S(t)ϕ] (θ) are diﬀerentiable. Proof. we obtain [S(t)ϕ] (θ) = ϕ(t + θ) for t + θ ≤ 0.3 Some results about the range of λI − A In line with Proposition 37.
Let us consider the space C ∗ := C([0. 0]. E ∗ ) × C([−r. Select one such map Υλ. . we have Sλ (fn ) ∈ R (λI − Lλ ). Denote S([0. By construction. r]. We are going to deﬁne a continuous bilinear form denoted << α.. we have that Sλ (ψn ) ∈ R (λI −Lλ ) and then Sλ (ψ) ∈ R (λI−Lλ ). A function f : [0.2. Therefore. 2. . . . We have also lim Sλ (fn ) = e . we have fn ∈ R (λI − A) and Sλ (fn ) = (λI − Lλ )(bn ). ϕ >> on the product C([0. We assume that R (λI −A) is closed and we want to show that it implies that R (λI −Lλ ) is closed. converges to some element e ∈ E. n = 1. Formal Duality In this section we present an extension of the formal duality theory to D.k (e). r].A theory of linear DDE in inﬁnite dimensional spaces 311 From Lemma 35. . E ∗ ) the space of simple functions. r]..k (e)] ∈ R(λI − Lλ ).k (e) ∈ R (λI − A). r]. 3. . in view of Lemma 35. . E) which will be interpreted as a formal duality. n→∞ R (λI − A) being closed yields that Υλ.. Ap ∈ Σ with pi=1 Ai = [0. n→∞ lim fn = Υλ. Lemma 35 again can be invoked to conclude that e = Sλ [Υλ. E ∗ ) where E ∗ is the topological dual space of E.E.k [(λI − Lλ )(bn )] . Deﬁne the sequence fn := Υλ. r] −→ E ∗ is called simple if thereEexist two ﬁnite collections x∗1 . r]. .2. The presentation follows the work made in [58].k as deﬁned by (2. x∗p ∈ E ∗ and A1 . Now we are going to demonstrate the converse. To this end. Let {bn }n=1.D.. in inﬁnite dimensional spaces.. . ) Ai Aj = ∅ such that f= p x∗i χAi i=1 where χA is the characteristic function of A and Σ is the Borel algebra on [0. . .. Again Lemma 35 enables us to conclude that ψ ∈ R (λI−A) and so R(λI − A) is closed in C. The proof is complete.9). we recall that Sλ is surjective and that in Lemma 34 we have exhibited a family of right inverses. . be a sequence in E such that the sequence {(λI − Lλ )(bn )}n=1.
It is interesting to specify the formal duality for α ∈ C([0.1) Proof.2.. E ∗ ) and ϕ ∈ C([0. L p ∗ i=1 xi χAi χAi (ξ − θ)ϕ(ξ) dξ > θ i=1 where α = and E. r] : E ∗ ) × C([−r. ϕ(0) > + p < x∗i . E ∗ ) deﬁned by (f ⊗ u∗ ) (s) := f (s)u∗ . E) and we call this the formal duality associated with the operator L. r]. θ (3.. βi (n) ∗ u ⊗ χAi (n) n=1. ϕ >>=< u . . pn i=1 converges to f ⊗ u∗ . E ∗ )×C([−r. Since  << α. We consider the function f ⊗ u∗ ∈ C([0. ϕ >> is independent of the representation chosen for α as a linear combination of characteristic functions. We restrict our extension to the product C([0. r]. E) we deﬁne the bilinear form << α. 0 and < ·. 0]. r]: f = lim n→∞ pn βi (n) χAi (n) i=1 and therefore the sequence of functions in S([0.312 DELAY DIFFERENTIAL EQUATIONS Deﬁnition 37 For any α ∈ S([0. r]. Lemma 37 Let f ∈ C([0.. r]. r]. s ∈ [0. f (0)ϕ(0) > + < u . 0]. r]) and u∗ ∈ E ∗ . r]) ⊗ E ∗ . ϕ >>:=< α(0).– The function f is representable as the limit of a uniformly convergent sequence of simple real functions deﬁned in [0. Then ∗ ∗ ∗ 0 << f ⊗ u . r]. E ∗ ). · > denote the usual duality between E ∗ The value << α. ϕ >>  ≤ (1 + rL) αϕ there exists a unique continuous extension of this bilinear form to the completion of S([0. E) where both spaces are equipped with the sup norm. L f (ξ − θ)ϕ(ξ) dξ >.
χAi (n) (ξ − θ)ϕ(ξ) dξ βi (n) χAi (n) (ξ − θ)ϕ(ξ) dξ . m(A)(x) > . L(f ) > . 0]. E ∗∗ ) is the space of the bounded linear operators deﬁned ˆ the Borel on E with values in E ∗∗ .2) The formal adjoint equation Before proceeding to the construction of this equation let us remind some well known results about the integral representation of bounded linear operators deﬁned on C([−r.313 A theory of linear DDE in inﬁnite dimensional spaces By continuity of the formal duality.8) the above formula reads as << ε−λ ⊗ u∗ . 0]. L =< (f ⊗ u∗ )(0). 0]. x∈E A∈Σ ∀x∗ ∈ E ∗ .0] f d mx∗ =< x∗ .1 (3. 3. which satisﬁes [−r. We refer to [14] for the general theory. ϕ >> ∗ =< (f ⊗ u )(0). L  (n) ∗ 0 θ 0 θ . Any bounded linear operator L : C([−r. ϕ >>=< u∗ .L  < u∗ . Sλ (ϕ) > . E). E ∗∗ being the bidual of E. f (ξ − θ)ϕ(ξ) dξ > . ˆ . . f (0)ϕ(0) > + < u∗ . << ελ ⊗ u . E) −→ E determines a ˆ −→ L(E. E ∗∗ ) of bounded semivariation unique vector measure m : Σ and such that for all f ∈ C([−r. θ In particular. and Σ algebra on [−r. there exists a vector measure mx∗ : Σ deﬁned by < mx∗ (A). 0]. ϕ(0) > + lim n→∞ pn i=1 pn < βi u . introducing the notation ελ (θ) := eλθ . ˆ −→ E ∗ For each x∗ ∈ E ∗ . L θ In terms of the operator Sλ deﬁned in (2. ϕ(0) > + < u . we obtain << f ⊗ u∗ . ϕ(0) > + lim n→∞ i=1 0 =< u∗ . x >:=< x∗ . > >. ϕ >>=< u .0] where L(E. E) we have fd m L(f ) = [−r. we have 0 ∗ ∗ ∗ λ(ξ−θ) e ϕ(ξ) dξ > .
E ). u >:=< u∗ . p. it is convenient to obtain the expression of L∗ for the elements in C([0. r]) ⊗ E ∗ . L(fˆ ⊗ u) > where fˆ(θ) := f (−θ). x >  x ≤1 sup  < x∗i . i=1 ˜ is Lemma 38 If the vector measure m is of bounded variation. E ∗ ) −→ E ∗ for any f ∈ Next we deﬁne the linear operator L p S([0. i = 1. .– Let f = i=1 xi χAi be a simple function with xi ≤ 1. u∈E (3. . r]) ⊗ E ∗ . . Just as we did in the case of the formal duality. E ∗ ). r]. . x ≤1 ˜˜ Under this hypothesis. Then ˜ ) = L(f p mx∗i (−Ai ) ≤ i=1 ≤ p p mx∗i (−Ai ) i=1 m(−Ai ) ≤ v(m) ([−r. p ∗ ∗ Proof. 0]) < +∞ i=1 where v(m) means the variation of m and we have used that mx∗i (−Ai ) = = sup  < mx∗i (−Ai ).3) .314 DELAY DIFFERENTIAL EQUATIONS ˜ : S([0. . Deﬁnition 38 The operator L∗ is the restriction to the space C ∗ of the ˜ ˜ extension operator L. r]. L∗ (f ⊗ u∗ ) ∈ E ∗ is the linear form deﬁned by < L∗ (f ⊗ u∗ ). r]. r]. E ∗ ) equipped with the of the operator L sup norm and we are able to deﬁne the formal adjoint operator of the operator L. f = i=1 x∗i χAi by ˜ ) := L(f p mx∗i (−Ai ). there exists a unique continuous extension L ˜ to the completion of S([0. Calculations very similar to the ones above for Lemma 37 show that Lemma 39 For each f ⊗ u∗ ∈ C([0. then L ∗ continuous with respect to the sup norm in S([0. m(−Ai )(x) >  ≤ m(−Ai ).
Deﬁnition 39 The formal adjoint equation associated to (CP) is α(s) ˙ = −L∗ (αs ) . . But since < α(t). L t+θ Therefore d << αt . L (f (t)x(t + θ)) >=< α(t). 0].5) for all x∗ ∈ N (L∗λ − λI). x(t) > + < α(t). xt >> = < α (t). (j) In particular. 0]. a < b.5) on ] − ∞. x (t) > dt + < x∗ . x (t) > − < α(t). u > (j) and then (j) . x(t) > + < x . E ∗ ) is a solution to the formal adjoint equation if α ∈ C 1 (] − ∞. b] and that x(t) is a solution of (CP) on [a. E) deﬁned by (f ⊗ u)(θ) := f (θ)u θ ∈ [−r. It is easy to check that α(s) := e−λs ⊗ x∗ . xt >> is constant for all t ∈ [a. . E ∗ ) and satisﬁes (3. L (f (t − θ)x(t)) > − < x∗ . L (f (t)x(t + θ)) > . 2. u > (j) = < u∗ . s ≤ 0 is a solution of (3. L(ελ ⊗ u) > (j) := < u∗ . = < α(t). xt (0) > + < x . 0] . s ≤ 0. 0]). for the functions ελ deﬁned by (j) ελ (θ) := θj λθ e j! . Then << αt . x (t) > − < x∗ . j = 1. +∞[. L(xt ) >= 0 .4) we reach the result < L∗ (ˆ ελ ⊗ u∗ ). Indeed 0 ∗ ft (ξ − θ)xt (ξ) dξ > << αt .5) for all s ≤ 0. we have denoted by f ⊗ u the element of C([−r. θ ∈ [−r. Suppose that α(t) := f (t)x∗ is a solution of (3.315 A theory of linear DDE in inﬁnite dimensional spaces For each f ∈ C([−r. Lλ (u) >=< (Lλ )∗ (u∗ ). (3. (3. . xt >> = < αt (0). L∗ (ˆ (j) (j) The existence of the operator L∗ allows us to deﬁne a new linear functional diﬀerential equation associated with problem (CP). u∈E ελ ⊗ u∗ ) = (Lλ )∗ (u∗ ). . 0]. r]. b]. u ∈ E. L θ t ∗ f (ω − θ)x(ω) dω > .5) A function α ∈ C(] − ∞.
Sλ (ψ) = ψ(0) + L 0 . ∀ϕ ∈ D(A). 3. From Lemma 37 and after an adequate integration by parts we obtain for α := f ⊗ x∗ ∈ D(A∗ ): << α. θ ∗ λ(s−θ) e ψ(s) ds . Proof. with a dense domain contained in C ∗ . but let us notice that it is not the adjoint of Sλ . r]. x(t) > = < α (t). xt >> = < α (t). We get to θ ∗ ∗ λ(s−θ) e ψ(s) ds . θ ∈ [0.2 The operator A∗ formal adjoint of A In the sequel we assume that the vector measure m associated to L is of bounded variation and L∗ has been deﬁned. (λI − Lλ ) (ϕ(0)) = ψ(0) + L 0 The right hand side of the above formula is similar to the operator Sλ deﬁned in (2. We will denote it Sλ . x(t) > + < L∗ (ft ⊗ x∗ ).– The solution of (λI − A∗ )ϕ = ψ with ψ ∈C ∗ is θ −λθ eλ(s−θ) ψ(s) ds . Aϕ >>=<< A∗ α. that is. E ∗ ) . D(A∗ ) := {α ∈ C 1 ([0. x(t) > + < L∗ (αt ). the operator A∗ deﬁned by A∗ (α) := −α˙ with domain ˙ = −L∗ (α)}. r] ϕ(θ) = e ϕ(0) + 0 where ϕ(0) is to be determined so that ϕ ∈ D(A∗ ). x(t) >= 0.8).316 DELAY DIFFERENTIAL EQUATIONS we have d << αt . α(0) A∗ is linear and closed. ϕ >> . L (f (t − θ)x(t)) > dt = < α (t). x(t) > + < x∗ . Proposition 38 The spectra σ(A) and σ(A∗ ) of operators A and A∗ satisfy the equality σ(A) = σ(A∗ ). Deﬁnition 40 We call the formal adjoint operator of A relative to the formal duality.
x∗ ∈ N (µI − Lµ ∗ )}. Notice that N (λI − A) = {ελ ⊗ x . A∗ respectively. Also. Proof. both the functions n and ∆n(t. and λ = µ. x ∈ N (λI − Lλ )} N (µI − A∗ ) = {ε−µ ⊗ x∗ . the integrals on the righthand side of (1. E := L1 (A1 . Therefore the result follows. ϕ >>= µ << α. Aϕ >>= λ << α. ϕ >> therefore the condition λ = µ implies that << α. ϕ >>=<< α. it is convenient to express the equation in a more general setting.20) are restricted to this interval. 3. << A∗ α. Proposition 39 The subspace R (λI − A∗ ) is closed in C ∗ if and only if R (λI − A) is closed in C. A2 ]. ϕ >>= 0 for all α ∈ N (µI −A∗ ). let us mention that if λ. With the notations r := 2Ψ(0) + τ . A2 ) . ·) have their support contained in some compact interval [A1 . Therefore. µ are eigenvalues of A. ϕ >>= 0. namely. We have seen above that λ ∈ σ(A) if and only if λ ∈ σ(Lλ ) and it is well known that σ(Lλ ) = σ(Lλ ∗ ) ([39]). ϕ ∈ N (λI −A). there is a support property for solutions of (1.– Arguments similar to the ones used in the proofs of Lemmas 35 and 36 lead to a) For any λ ∈ C we have that α ∈ R (λI − A∗ ) if and only if Sλ (α) ∈ R (λI − Lλ ∗ ).3 Application to the model of cell population dynamics We will now see how the theory developped in this section applies to equation (1.A theory of linear DDE in inﬁnite dimensional spaces 317 Since Sλ is onto. First of all.5).20). To this end. Indeed. then << α. The proof is concluded by recalling the known result that R (λI − Lλ ) is a closed subspace in E if and only if R (λI − Lλ ∗ ) is closed in E ∗ . b) The subspace R (λI −A∗ ) is closed in C ∗ if and only if R (λI −Lλ ∗ ) is closed in E ∗ . we conclude that λ ∈ σ(A∗ ) if and only if λ ∈ σ(Lλ ∗ ).
y) dy ∂t A1 A1 0 A2 + g(x. u. 0]. 0) × (A1 . E ∗ = L∞ (A1 . such that g(x. . v ∈ E.20) is of the form A2 A2 ∂ν (t. with (L(ϕˆ ⊗ v)) (x) = b(x)ϕ(τ ) 0 A1 A2 v(y) dy + c(x) ϕ(τ + Ψ(y))v(y) dy A1 A2 + g(x. y ∈ (A1 . A2 ) × (−r. y)ϕ(−u)v(y) dydu. L(ϕˆ ⊗ v) > = A2 ∗ b(x)v (x) dx ϕ(τ )v(y) dy A1 A1 A2 A2 + ∗ c(x)v (x) dx ϕ(τ + Ψ(y))v(y) dy A1 A2 A1 A2 + A1 A1 0 r ∗ g(x. A2 0 g(x. A2 )). M ≥ 0. L(ϕˆ ⊗ v) >. L∗ is determined using formula (3.318 DELAY DIFFERENTIAL EQUATIONS equation (1. (3. A2 ). A2 ). −u. u. y)v (x)ϕ(u) dudx v(y) dy. r]). y) dy + c(x)ν(t − τ − Ψ(y).e. A2 ). x) = b(x)ν(t − τ. y) dudx ≤ M. 0 < u < r . y. A1 −r The map L : C ([−r. we have < L∗ (ϕ ⊗ v ∗ ). y) dydu. y) dydu. for a. y) ∈ (A1 . y)ϕ(u. A2 )2 . u. The assumptions on f yield the following for g: There exist 0 < r < r. v >=< v ∗ . −r ∗ A2 A1 A2 < v .e. u) = 0. (x. v ∗ ∈ E ∗ . for ϕ ∈ C ([0. In fact.6) −r A1 where b and c belong to L1 (A1 . y) dy (Lϕ)(x) = b(x) A1 0 A1 A2 g(x. + −r A1 Here. for a. g is in L1 ((A1 .5) is given by A2 A2 ϕ(−τ. u. y)ν(t + u. y) dy + c(x) ϕ(−τ − Ψ(y). E) −→ E associated with equation (3.3).
x) dξdx dy. z) dξ dy dzdu dy. Using formula (3. r]. 0 v (y)c(y) ζ(ξ − θ)ϕ(ξ. u. ϕ >> A2 = ϕ∗ (0. u A1 Here. 0]. y) ϕ(ξ. x) dξ A1 = θ A1 v (y)b(y) + v ∗ (y)L A1 A2 + v ∗ (y) ζ(ξ + τ + Ψ(u))ϕ(ξ. −u. y) dy A2 +  ∗ A2 . we obtain << ϕ∗ . z) dξ ϕ∗ (ξ + τ. y)ϕ(ξ. x) dx A1 A2 A1 r + A1 g(x. r]. y)ϕ∗ (u. −u. the formula extends to C ([0. 0 The above formula can be extended to the space C ([0. y)v ∗ (x)ϕ(u) dudx. ·) dξ (y) dy −τ A1 A1 0 ζ(ξ + τ )ϕ(ξ. 0 We will now determine the formal dual product. y) dy A1 A2 A2 0 ∗ c(y) + A1 A2 0 ϕ (ξ + Ψ(u) + τ. u) dξdu dy −Ψ(u)−τ A1 A2 0 g(y. E ∗ ) × C ([−r. y)ϕ(0. x) dx + c(x)ϕ∗ (τ + Ψ(y). dzdudy . u. E ∗ ) for ζ ⊗v ∗ in the above formula. y)ϕ(0. y) dy + A1 A2 ∗ A2 A2 A2 0 dx dy (ζ ⊗ v ∗ )(0. r]. u) dξdu 0 −Ψ(u)−τ A1 A2 0 g(y. E). we have << ζ ⊗ v ∗ . z) −r A1 ζ(ξ − u)ϕ(ξ. E ∗ ) as: A2 A2 ∗ ∗ ∗ (L ϕ )(y) = b(x)ϕ (τ.1).319 A theory of linear DDE in inﬁnite dimensional spaces Identifying the right hand side with the product of L∗ (ϕ ⊗ v ∗ ) and v yields A2 A2 ∗ ∗ ∗ L (ϕ ⊗ v )(y) = ϕ(τ ) b(x)v (x) dx + ϕ(τ + Ψ(y)) c(x)v ∗ (x) dx A1 A2 r + A1 A1 g(x. Substituting ϕ∗ ∈ C ([0. ϕ >> = A2 ζ(0) v ∗ (y)ϕ(0. y)ϕ(ξ. x) dudx. z) + A1 −r A2 + A1 A2 u 0 b(y) A1 A1 −τ ϕ∗ (ξ − u.
there is a basic feature of the semigroup in ﬁnite dimensions which will generally be lacking: the eventual compactness of the semigroup.D. We have associated to the problem a dual product between C and C ∗ := C([0.1) is well posed and the general solution gives rise to a strongly continuous semigroup of bounded linear operators. we will use these notions in the study of the spectral decomposition associated to abstract D. to better investigate the asymptotic behavior of solutions and the stability of the system. Since we are in the inﬁnite dimensional case. in the applications we have in view. This goes through a decomposition of the state space C into the sum of a stable and an unstable part. t>0 (4. and in many situations all amounts to looking at the real part of the roots of a characteristic equation.E.E. however.1) where L : C −→ E is a bounded linear operator and E is a Banach space.D. the sum of the most unstable and a more stable part. 4.E. In the next section.: x (t) = L(xt ) . Hale in [74].D. [75] and others for D. First of all. For most computations. Eventual compactness reduces the problem to the study of the eigenvalues and eigenvectors of the generator. In as much as possible. we follow the method elaborated by J. for example. We have now a formal dual product as well as a formal adjoint equation.4 DELAY DIFFERENTIAL EQUATIONS Conclusion The material presented in this section is mainly preparatory.D. In the general situation. E ∗ ) and an adjoint equation. One way to look at the dual product is to consider that it provides a dynamic system of coordinates which allows. r] . . Linear Theory Of Abstract Functional Diﬀerential Equations Of Retarded Type Let us recall that the subject of this section is the abstract D. As a result.320 3. We know that the Cauchy problem for (4. or at least not at all as easy to use as in ﬁnite dimensions. or more generally.E. it is in general necessary to embed the equation in a much larger space with no nice way to identify the original problem in the larger one. it is suﬃcient to use it for special functions in the formal dual space for which such a description is feasible. A second crucial diﬀerence which will appear when dealing with spectral decomposition is the lack of compactness in general D. it is in general impossible to describe the dual product to a large extent. not much could be said. However. in ﬁnite dimensions. One of the problems when going from ﬁnite to inﬁnite dimensions is that the representation of a linear functional in the inﬁnite dimensional framework is not always possible.
that is. It is based on a review paper [1]. a situation that has been encountered by several authors in the study of population dynamics models in the eighties ([42]). that is. in addition to the one we have just given ([10]): The essential spectrum of A.A theory of linear DDE in inﬁnite dimensional spaces 321 it happens often that the most unstable part of the spectrum leads to a decomposition of the state space and allows the same study as in the case of eventual compactness. The spectrum of A can be subdivided into three disjoint subsets σ(A) = σP (A) ∪ σC (A) ∪ σR (A) where i) σP (A) is the point spectrum. [35]. but is dense in E. the set of λ ∈ C such that λI − A is not injective. [77]. [21]. ii) The generalized eigenspace associated to λ. The notion of essential/non essential spectrum however is a general one which was introduced in the sixties in connection with the study of partial diﬀerential equations. 4. is the set of λ ∈ C. . that is. There are other ways to divide the spectrum. ii) σC (A) is the continuous spectrum. such that λI − A is not boundedly invertible. We will in fact consider another one. iii) σR (A) is the residual spectrum. ρ(A) := C/σ(A) is the resolvent set. is the set of λ ∈ σ(A) such that at least one of the following holds: i) R (λI − A) is not closed. The spectrum of A. N (λI −A) := ∪∞ m=1 Ker (λI − A)m is inﬁnite dimensional. This is when the most unstable part of the spectrum of the semigroup is non essential. iii) λ is a limit point of σ(A). is not E. the set of λ ∈ C such that λI − A is injective and the range of λI − A. We refer the reader to [15]. denoted by R (λI − A). The complementary set in C. The next section presents a short survey of the subject. E being a Banach space.1 Some spectral properties of C0 semigroups We start considering a linear closed operator A with dense domain D(A) ⊂ E. denoted σ(A). [37] for the general theory. the set of λ ∈ C such that λI − A is injective and R (λI − A) is not dense in E. σe (A).
This theorem extends partially to the case of an unbounded linear operator. while obviously it cannot arise from the spectrum of A. This fact can be used if some information about the essential spectrum of the semigroup is avalaible. with some abuse. In this case. the residual spectrum as well as the non essential spectrum of A and T (t) correspond to each other in the simplest way possible. But no such relationship can be asserted in general for the continuous spectrum or the essential spectrum. E ∈ L(E) and let f : Ω ⊂ C −→ C be an holomorphic function with σ(B) ⊂ Ω. If A is the inﬁnitesimal generator of a strongly continuous linear semigroup {T (t)}t≥0 (that can be. For example. λ ∈ σJ (A) where J can be either P (the point spectrum) or R (the residual spectrum) or N E (the nonessential spectrum). and the notation σ ∗ indicates that we are not counting 0 which might be a spectral value of T (t) (and is indeed for t > 0 in the case of compact semigroups). 0 1 σJ∗ (T (t)) = etλ . t > 0. The spectral theorem states that f (B) can be deﬁned as a bounded linear operator on E and also: σ(f (B)) = f (σ(B)). We refer to Theorem 3 in [1] for some results on this issue. Let B be a bounded linear operator on some Banach space E. i. These values are the ones which matter regarding the stability issue: they are non essential and thus emanate from the generator. or there are spectral values of T (t) which exceed in magnitude. the ones in the essential spectrum. which provides exponential stability of the semigroup. either the whole spectrum of T (t) is located in the interior of the unit disk.322 DELAY DIFFERENTIAL EQUATIONS The complementary set. for some t > 0. .e. In this direction the following partial result can be shown: exp(tσe (A)) ⊂ σe (T (t)) . then the point spectrum. viewed as a generalized exponential operator exp(tA)). Examples of spectral values of the semigroup (other than 0) which do not arise from the spectrum of the generator can be found in the literature ([21]). if we know that it is contained in the interior of the unit disk. (σ\σe ) (A) is called the nonessential spectrum. namely. There are relationships between the two partitions of σ(A).
The above deﬁnition of measure of noncompactness can be extended to bounded operators B ∈ L(E): 4 3 α(B(U )) . The crucial point here is that the generator does not help at all at this point. which we recall in the following. Then we will proceed to the next step. λ∈σe (B) The computation of the essential spectral radius is a little more involved.A theory of linear DDE in inﬁnite dimensional spaces 323 This result shows that it is important to have means to estimate the essential spectrum of an operator and possibly also of a semigroup of operators. The spectral radius of B is the number r(B) := sup λ. Let B be a linear bounded operator deﬁned on the Banach space E. The measure of noncompactness of U is the number α(U ) deﬁned by: 3 4 U can be covered by a ﬁnite number of α(U ) := inf ε > 0 . α(B) := sup α(U ) . U ⊂ E bounded and α(U ) = 0 . it has in fact to do with the distance of the operator to the set of compact operators. thus the essential spectral radius is equal to 0. n→∞ Accordingly. We will end this section providing some estimates of the essential spectrum. the essential spectrum reduces to {0}. that is to say. In particular. We will now quote the few results known on this issue. It uses the notion of measure of noncompactness for sets and operators. the essential spectral radius of B is deﬁned by: re (B) := sup λ. subsets of E of diameter less than ε. the spectral decomposition associated with the non essential spectrum. We will see that it works much the same as in ﬁnite dimensions. the following is obvious: α(U ) = 0 if and only if U has a compact closure. Nussbaum ([32]). Notice that for compact operators. The following formula was proved by R. Let U be a bounded subset of E. λ∈σ(B) There is a formula for the computation of the spectral radius which does not use the spectrum: r(B) = lim B n 1/n .
the following quantity may also be of interest.1). denoted s(A). It is the quantity ω1 deﬁned by: ω1 := lim log α [T (t)] . n→∞ Finally. λ ∈ σ(A)} . E) Let σ(A).2 Decomposition of the state space C([−r. [42]). From the general theory about operator reduction for isolated points of the spectrum ([77]. Accordingly we can deﬁne the essential growth bound of the semigroup. sup Re λ . 4. then the growth bound will be determined by the nonessential spectrum of the generator.324 DELAY DIFFERENTIAL EQUATIONS We can now state the formula for the essential spectral radius ([32]): re (B) = lim α(B n )1/n . It is the quantity ω0 deﬁned by: ω0 := lim log T (t). λ∈(σ\σe )(A) If we can compute or at least estimate ω1 and show that ω1 < ω0 or ω1 < s(A). ω0 = max ω1 . t→+∞ We also have the expected relationship re (T (t)) = eω1 t . σe (A) be the spectrum and the essential spectrum respectively of the inﬁnitesimal generator A of the semigroup {T (t)}t≥0 deﬁned by the solutions to (4. The spectral bound of a linear closed operator A. is deﬁned by: s(A) := sup {Re λ . we obtain the following theorem which yields the decomposition of C into a direct sum: . . 0]. The potential interest of the theory that will be presented next lies in the following formula ([42]): . that is: ∀t > 0 r(T (t)) = eω0 t . t→+∞ The connection between the growth bound and the spectral radius is what the intuition suggests it should be. Corresponding to the spectral radius of an operator is the growth bound of a semigroup {T (t)}t≥0 .
BΛ := diag (Bλ1 . . .E. s. . The only eigenvalue of Bλj is λj . where Φλj is a basis of the generalized eigenspace associated to λj . The same theory applies for a ﬁnite subset of (σ\σe ) (A) and gives a very clear description of the geometric behaviour of the solutions of (4. . . N (λj I − A)mj . λs } is any ﬁnite subset of the nonessential spectrum of A. . with dim N (λj I − A)mj = qj < +∞. we have ϕ = Φλ a for some qvector a and the solution is deﬁned by xt = T (t)ϕ = T (t)Φλ a = Φλ eBλ t a .A theory of linear DDE in inﬁnite dimensional spaces 325 Theorem 2 Let λ ∈ (σ\σe ) (A). . We summarize these results in the following theorem. and Bλj is a constant matrix such that AΦλj = Φλj Bλj . Therefore Φλ (θ) = Φλ (0)eBλ θ .1) in ﬁnitedimensional spaces remains valid without essential modiﬁcations. let PΛ := N (λ1 I − A)m1 ⊕ · · · ⊕ N (λs I − A)ms . Moreover A is completely reduced by this decomposition. Moreover. Let Φλ = (ϕ1 . Some of the properties owe to translation semigroup features enjoyed by abstract D. and let ΦΛ := (Φλ1 . Theorem 3 Suppose Λ = {λ1 . ϕq ) be a basis for N (λI − A)m . Bλs ). There is a q × q ˙ λ = Φλ Bλ and λ is the unique eigenvalue constant matrix Bλ such that Φ of Bλ . and are valid in a much more general context. . . θ ∈ [−r.D. such that C = PΛ ⊕ QΛ . . j = 1. 0] and also T (t)Φλ = Φλ eBλ t . . if the initial value ϕ of (4. t > 0. A restricted to N (λI − A)m is bounded with spectrum {λ} and the subspaces N (λI − A)m . .1) belongs to N (A − λI)m .1). The description of the fundamental solution matrix associated with a ﬁnite dimensional invariant subspace done in [74] for problem (4. . . t > 0. Furthermore. . Then there exists a subspace QΛ of C invariant under A and {T (t)}t≥0 . R (λI − A)m are invariant under the semigroup {T (t)}t≥0 . Then λ is an eigenvalue of A and for some positive integer m we have C = N (λI − A)m ⊕ R (λI − A)m where N (λI − A)m is the generalized eigenspace of A with respect to λ and dim N (λI − A)m = q < +∞. . . (σ\σe ) (A). . Φλs ).
3 A Fredholm alternative principle What we want to do next is to use the formal duality presented in the previous section to obtain an explicit characterization for the projection operator on the subspace QΛ . we have that R (λI − A) is closed. we have < x∗ . Therefore we can identify an element in the range of λI − Lλ in terms of the null space of the natural adjoint of this operator. 4. for any initial value ϕ = ΦΛ a where a is a constant vector of dimension q1 + · · · + qs . Then. f >>= 0 for all α = ε−λ ⊗ x∗ . which holds if and only if < x∗ . with x∗ ∈ N (λI − (Lλ )∗ ) where (Lλ )∗ is the adjoint of operator Lλ . f >>. according to formula (3. given f in C. We say that A is reduced by Λ. those computations are in some sense valid in general. The computations we are going to present now are similar to the ones performed in ﬁnite dimensions. Lemma 35 tells us that f ∈ R (λI − A) if and only if Sλ (f ) ∈ R (λI − Lλ ).326 DELAY DIFFERENTIAL EQUATIONS and the operator A is completely reduced by this decomposition. Sλ (f ) >=<< ε−λ ⊗ x∗ .– Since λ ∈ (σ\σe ) (A). The ﬁrst result is a Fredholm alternative principle for the characterization of the range of λI − A. ϕ >>= 0 for all x∗ ∈ N (λI − (Lλ )∗ ). But. the solution of the Cauchy problem associated to ( 4. Furthermore. Proposition 40 Let λ ∈ (σ\σe ) (A). 4. But. in the case when λ is an essential spectral value. The main point is that the equation which is originally set up in C will .1) is deﬁned by xt = T (t)ϕ = T (t)ΦΛ a = ΦΛ eBΛ t a . Then. the equation (λI − A)ψ = ϕ has a solution if and only if << ελ ⊗ x∗ . t ≥ 0. In fact.4 Characterization of the subspace R (λI − A)m for λ in (σ\σe ) (A) The fact that λ is non essential ensures that R (λI −A)m is closed with a ﬁnite codimension. and there is an integer m0 such that R (λI − A)m is the same for all m ≥ m0 . the computation will not be conclusive. Proof. Now. Sλ (f ) >= 0 for all x∗ ∈ N (λI − (Lλ )∗ ). f ∈ R (λI − A) if and only if << α.2). We can state Proposition 40 in a Fredholm alternative principle form: Let λ ∈ (σ\σe ) (A). so Lemma 36 gives that R (λI − Lλ ) is closed in E. which completes the proof of the proposition.
)T means the transpose vector operators Lλ (m) ∈ L(E m ) deﬁned by: λI − Lλ I − Lλ (1) . 0 0 .. . um−1 )T = Sλ (m) (f ) where (. 0 λI − Lλ in which Lλ (j) ∈ L(E). . . Lλ (m) (u0 . . . . Sλ (f ) := . . The operators Sλ (m) : C −→ E are deﬁned by 0 m−1 λ(θ−ξ) (ξ − θ) f (ξ) dξ e L (m − 1)! θ 0 m−2 λ(θ−ξ) (ξ − θ) L f (ξ) dξ e (m − 2)! θ .2) amounts to proving that each ϕ(k) satiﬁes the boundary condition ϕ˙ (k) (0) = L(ϕ(k) ). 0 λ(θ−ξ) L e (ξ − θ)f (ξ) dξ θ 0 λ(θ−ξ) −ψ(0) + L e f (ξ) dξ θ . u ∈ E and ελ (j) are deﬁned in (3. . . Introducing the notation d k ϕ(θ) .4). m − 1... . . By direct calculation. .327 A theory of linear DDE in inﬁnite dimensional spaces reduce to an equation in the product space E m . which. . . θ ∈ [−r... in view of the expression (4.. 0] eλ(θ−ξ) j! (m − 1)! θ j=0 (4.. . θ ∈ [−r.. In fact. k = 0. . um−1 are arbitrary elements of E which must be determined so that ϕ ∈ D ((λI − A)m ). . (m) . (m) . . I − Lλ (1) λI − Lλ . 0 λI − Lλ . Lλ (j) (u) := L(ελ (j) ⊗ u). . .2) where u0 .. 0] ϕ(k) (θ) := λI − dθ we have that ϕ ∈ D ((λI − A)m ) if and only if ϕ(k) ∈ D(A).. . := Lλ . 0 0 and we have introduced the −Lλ (m−2) −Lλ (m−1) −Lλ (m−3) −Lλ (m−2) .. the solution to (λI − A)m ϕ = f is given by: 0 m−1 θj (ξ − θ)m−1 ϕ(θ) = eλθ uj + f (ξ) dξ . the problem reduces to an algebraic equation to be satisﬁed by the uk .
we need to see that R Lλ (m) is a closed subspace of E m . N (λI − A) is spanned by the functions ελ ⊗ u with u ∈ N (λI − Lλ ). We wish to locate R Lλ (m) by means of the adjoint operator Lλ In order to achieve this.. Proposition 41 Let λ ∈ (σ\σe ) (A). Lemma 41 If λ ∈ (σ\σe ) (A). by assumption. one obtains that if Since R (λI − A)m = Sλ (m) R Lλ (m) is a closed subspace in E m . Proof. Lemma 42 Let A1 . Therefore. then λI − Lλ is a Fredholm operator. However. . then dim N (λI − Lλ ) < +∞. Am be Fredholm operators on E. we can see that Sλ (R (λI − A)) = R (λI − Lλ ) and Sλ−1 (R (λI − Lλ )) = R (λI − A).328 DELAY DIFFERENTIAL EQUATIONS One can check readily that Lλ (j) 1 (u) = j! d dλ j (Lλ ) (u). The proof is an immediate consequence of the two next lemmas. Finally. So. from Lemma 36 we infer that R (λI − Lλ ) is a closed subspace in E. . we have proved that R (λI −Lλ ) is closed with a ﬁnite codimension. (m) ∗ . the converse statement is of a bigger interest to us. From what precedes we deduce that Sλ (M ) ∩ R (λI − Lλ ) = {0} and Sλ (M )+R (λI −Lλ ) = E. The operator A(m) deﬁned on the product space E m by ∗ A1 . 2. A(m) := .– Since λ ∈ (σ\σe ) (A). λI − Lλ is a Fredholm operator. On the other hand. R (λI − A) has a ﬁnite codimension. Then Lλ (m) is a Fredholm operator for each m = 1. which can be done since. Select a ﬁnite dimension vector space M complementing R (λI − A) in C. Therefore. then R (λI − A)m is closed in C. using Lemma 34 and Lemma 35. . . we have proved the following result Lemma 40 f ∈ R (λI − A)m if and only if Sλ (m) (f ) ∈ R Lλ (m) . We begin its analysis proving the following proposition. No other result similar to the one stated in Lemma 35 is known. . 0 Am . −1 R Lλ (m) .
Proof. δj (resp. ελ (j) (0)ϕ(0) > + < u∗ .. we may deﬁne G. . . A(m) D = + := U + ∆. ·. Proceeding exactly the same way on the left. Multiplying the above expression on both sides by U −1 leads to A(m) DU −1 = I + ∆U −1 . . The product A(m) D can be decomposed as follows I ∗ δ1 0 .. x∗m−1 )T ∈ ∗ N Lλ (m) . γj ) in L(E) such that Aj ◦ Dj = I + δj (resp. A(m) is a Fredholm operator. 1 ≤ j ≤ m. Notice that << εˆλ ⊗ u∗ . so Sλ (m) (f ) ∈ R and we conclude that R Lλ Lλ (m) if and only if < X ∗ . Assuming that the Aj . with the δj and the γj having a ﬁnite rank. Gj ◦ Aj = I + γj ). .329 A theory of linear DDE in inﬁnite dimensional spaces where the operators ∗ are in L(E). . Sλ (m) (f ) >= 0 for all X ∗ = (x∗0 . L 0 (j) θ ελ (j) (θ − ξ)ϕ(ξ) dξ > . ϕ >>=< u∗ ..– Fredholm operators can be characterized as follows: let F ∈ L(E). . is a ﬁnite rank operator. is a Fredholm operator. where ∆ Therefore. Deﬁne the following operator on E m : ∗ D1 . 0 Dm Accordingly. ∆U −1 has a ﬁnite rank. . Then. We return to the problem set before the statement of Proposition 41 (m) is closed. Gj . Obviously. . we can also determine an operator G5 in L(E m ) such that ˜ 5 (m) = I + ∆ GA ˜ is a ﬁnite rank operator. where δ and γ are ﬁnite rank operators. we can determine Dj . D := . are Fredholm operators. F is Fredholm if and only there exist two operators D and G in L(E) such that F ◦ D = I + δ and G ◦ F = I + γ. 0 I 0 δm The ﬁrst matrix U in the righthand side is invertible and the second one.
f >>= 0. . . (x∗0 . (j) j=0 Let us consider the subspace Kλ (m)∗ := α= m−1 εˆλ ⊗ x∗m−j−1 . . In fact. . We summarize the results achieved up to now in the following proposition. thus we have that dim N Lλ dim Kλ (m)∗ = p. . . . . Rλ (m) (f ) >=<< m−1 εˆλ ⊗ x∗m−j−1 . . choosing a basis Φ∗λ := (ϕ∗1 . Therefore ∗ < X . . f >>= 0 for each α ∈ Kλ (m)∗ . . Therefore ∗ Φ∗λ (θ) = eBλ θ Φ∗λ (0) . . . . . . . . x∗m−1 )T ∈ N (j) Lλ (m) ∗ . . Moreover N (λI − A)m = ϕ= m−1 ελ (j) ⊗ uj . . ∗ also implies that Notice that (x∗0 . This fact implies that elements of this subspace are solutions of a linear O. um−1 )T ∈ N j=0 and then dim N (λI − A)m = dim N Lλ (m) Lλ (m) < +∞. . Proposition 42 Let λ ∈ (σ\σe ) (A) and let m be a positive integer. . ϕ∗p )T = B ∗ Φ∗ Φ λ λ λ λ where Bλ∗ is a constant p × p matrix and λ is the only eigenvalue of this matrix. x∗0 .330 DELAY DIFFERENTIAL EQUATIONS where the notation f stands. operator. r]. we have ˙ ∗ = (ϕ˙ ∗1 . θ ∈ [0. (m) = p for some p < +∞ and then. x∗m−1 )T ∈ N Lλ (m) ∗ and then it is easy to prove by Lλ (m) (0.D. Then f ∈ R (λI −A)m if and only if << α. . for f (−s). . ϕ˙ ∗p )T = B ∗ (ϕ∗1 . .E. . j=0 Since Lλ (m) is a Fredholm ∗ the same is true for its adjoint. x∗m−2 )T ∈ N direct calculation that the subspace Kλ (m)∗ is diﬀerentiation invariant. . as usual. ϕ∗p )T of Kλ (m)∗ . . (u0 . .
5 331 Characterization of the projection operator onto the subspace QΛ We recall the conclusion of Theorem 2. ∗ (m) (m) Recall that p = dim N Lλ . λq )T ∈ N (M ). . . Φλ >>= [<< ϕ∗i . Φ∗λ = (ϕ∗1 ...3) If (λ1 . . . . λ1 ψ1 + · · · + λq ψq >>= 0 for all α∗ in Kλ (m)∗ and Proposition 42 implies that λ1 ψ1 + · · · + λq ψq ∈ R (λI − A)m . . .p... . it is ﬁrst convenient to investigate the relationship between the numbers p. . . ψj >>]i.. q = dim N Lλ but in general. ϕ >>= 0 for all α ∈ Kλ (m)∗ . Ψλ >>:= [<< αi∗ . We wish to ﬁnd a suitable coordinate system which serves in characterizing the projection operator onto the subspace R (λI − A)m . To this end. ϕq ).p. In particular we can choose two new bases Φλ := (ϕ1 .. q. for λ ∈ (σ\σe ) (A). implying that q ≤ p.q = [δij ]i. then << α∗ . . (4. ϕq ). Let Ψλ = (ψ1 . Φ∗λ := such that the constant p × q matrix satisﬁes (ϕ∗1 . .j=1. Therefore λi = 0. . . and make up the constant p × q matrix M : Ψ∗λ M :=<< Ψ∗λ . there exists a positive integer m for which a direct sum decomposition of the following type holds: C = N (λI − A)m ⊕ R (λI − A)m with dim N (λI − A)m = q < +∞ and ϕ ∈ R (λI − A)m if and only if << α. i = 1. . . . . We summarize all of this in a theorem: Theorem 4 If λ ∈ (σ\σe ) (A).p. αp∗ )T a basis for Kλ (m)∗ . . . ϕ∗p )T << Φ∗λ . .. . ψq ) be a basis for the subspace N (λI − A)m and = (α1∗ . .q where δij is the Kronecker symbol.. . .j=1. Thus we conclude that M has rank q. . then dim N (λI − A)m ≤ dim Kλ (m)∗ and there exist two bases Φλ = (ϕ1 .q . even for Fredholm operators.. .p.. we have p = q.j=1.. .... . Φλ >>= [δij ]i. We now show that q ≤ p and next we will obtain some suﬃcient conditions for p = q. . N (M ) = {0}. ϕ∗p )T of the subspaces N (λI − A)m and Kλ (m)∗ respectively such that << Φ∗λ . .. .q .j=1. . q. Also dim Kλ (m)∗ = p < +∞. But we also have λ1 ψ1 + · · · + λq ψq ∈ N (λI − A)m and then λ1 ψ1 +· · ·+λq ψq = 0. that is.A theory of linear DDE in inﬁnite dimensional spaces 4. ϕj >>]i.
.332 DELAY DIFFERENTIAL EQUATIONS Moreover. 0]. ˙ ϕ >>= 0. we say that the formal duality is non degenerate if the equality << α. j = 1. Φλ >> = B ∗ << Φ∗ . ˙ ϕ >>=< u∗ . we have << α. Notice that for p = q the last relation reduces to Bλ∗ = −Bλ . Bλ∗ deﬁned by Φ λ Bλ∗ Φ∗λ . . ϕK >>= 0 if i > q. Therefore ˙ λ >> ˙ ∗ . Φλ >>= − << Φ∗ . Lemma 43 If the formal duality is non degenerate then p = q. Proof. ϕ˙ >> + << α. Φλ >>= Iq 0 where Iq is the identity q × q matrix. L fˆ ⊗ ϕ(0) > + < u∗ . Φλ >> Bλ . p. for α ∈ Kλ (m)∗ and ϕ ∈ D(A). . ˙ λ = Φλ Bλ . Φ˙ ∗ = Next we relate the matrices Bλ . implies that α = 0. we obtain −Bλ N ∗ Bλ = 0 P where N and P are two matrices of the adequate dimensions and 0 is a zero submatrix. . for all ϕ ∈ C.– As usual. ϕ >>= 0. Φ << Φ λ λ λ λ = − << Φ∗λ . . << α. ϕK >>= λi if i ≤ q and << ∗ ϕi . Since << Φ∗λ . r]). ϕI >>= 0. f (0)ϕ(0) > where fˆ(θ) = f (−θ). It is easy to check that for α = f ⊗ u∗ . Thus. ϕI ∈ R (λI − A)m and << ϕ∗j . f ∈ C 1 ([0. q ∗ Also ϕK = i=1 λi ϕi with << ϕi . u∗ ∈ E ∗ and for all ϕ ∈ D(A). E) we have a unique decomposition ϕ = ϕK + ϕI with ϕK ∈ N (λI − A)m . ϕ˙ >> + << α. for each ϕ ∈ C([−r. Next we state some suﬃcient conditions to have p = q.
. 0 0 0 . Since the formal duality is non degenerate. ϕ > >= 0.. λ = integers m. . If the (m) operator Hλ or some of its iterates is compact.. . Finally we relate the equality p = q with some compactness properties on the operator Lλ (m) . . Proof. .. ... Lλ Lλ (m) . . µ.. . we must have µ1 α1∗ + · · · + µp αp∗ = 0 and then N M T = {0}. . := . r and α ∈ Kλ (m)∗ . << µ1 α1∗ + · · · + µp αp∗ . let (µ1 . Then. . the operator λI + J (m) is invertible and then (m) (m) .. . ϕ ∈ N (µI − A)r . the Bezout identity ensures the existence of two . Hλ . (x − λ)m and (x − µ)r are relatively prime. Lλ (m) = λI + J (m) − Hλ = (λI + J (m) ) I − (λI + J (m) )−1 Hλ (m) The operators J (m) and Hλ conmute and from well known general results about the spectrum of compact operators ([35].A theory of linear DDE in inﬁnite dimensional spaces 333 With the above notations. 0 0 0 0 I . µ be given in (σ\σe ) (A). 0 I 0 0 0 .– Given that the two polynomials in x. .25) we have Lemma 44 Let λ ∈ (σ\σe ) (A) be an eigenvalue of A. (1) 0 Lλ 0 . J (m) := . .. 0 0 (1) (m−2) (m−1) Lλ Lλ Lλ .3) ). Lλ 0 0 .. µp )T ∈ N M T (M deﬁned in (4. it holds that << α... Introduce the operators 0 I 0 .. The converse is not true in general. so p = q. .... . .. Lλ (m−3) (m−2) 0 Lλ .. then p = q.. 4. . This implies that p ≤ q. 0 Lλ If λ = 0. 0 0 ... For any positive Lemma 45 Let λ. There are examples for ﬁnite dimensional spaces E such that p = q and the formal duality is degenerate.. Th. λ = 0. . . ϕ >>= 0 for all ϕ ∈ R (λI − A)m and also for all ϕ ∈ C....
ΦΛ >>= [δij ]i. ϕPΛ >>=<< ΦΛ . any ϕ ∈ C may be written as ϕ = ϕPΛ + ϕQΛ with << Φ∗Λ .k=1. Φl >> is zero for k = l. Q (d/dθ) (µI − d/dθ)r ϕ >> = 0. << Φ∗Λ . j = 1....334 DELAY DIFFERENTIAL EQUATIONS polynomials P (x). Q(x) such that d r d d d m + µI − P Q I = λI − dθ dθ dθ dθ where d/dθ is the diﬀerentiation operator. . Therefore there exist two bases ΦΛ .. J := << Φ∗j . We are now able to characterize the subspace QΛ by an orthogonality relation associated with the formal duality.. Keeping all the above notations. . Φk >> j. ϕ >> = << α.. ϕQΛ >>= 0 and ϕPΛ = ΦΛ a where a is a constant vector of dimension q1 + · · · + qs such that a ∗ ∗ ∗ << ΦΛ . E) stated in Theorem 3. Φ∗Λ of the subspaces PΛ . has rank equal to qj and also the matrix << Φ∗k . ϕ >>= 0}. λs } be a ﬁnite subset of non essential points of σ(A) and consider the decomposition of the space C([−r. . we deﬁne PΛ∗ := Kλ1 (m1 )∗ ⊕ · · · ⊕ Kλs (ms )∗ . qj ≤ pj . Let Λ = {λ1 . 0].s has rank equal to q1 + · · · + qs .j=1.. Next. From the results stated above we know that each constant pj × qj matrix Jj :=<< Φ∗j . . . s. Moreover. .q1 +···+qs . . Then the matrix J of order (p1 + · · · + ps ) × (q1 + · · · + qs ). Kλj (mj )∗ respectively. Φj >>.. Φ∗j be bases of the subspaces N (λj I − A)mj . (λI − d/dθ)m P (d/dθ) ϕ >> + << α. . ΦΛ > satisﬁes << Φ∗Λ . PΛ∗ respectively such that the constant matrix << Φ∗Λ . ΦΛ > a = 0 . To this end. so that << α. let Φj . we have proved the following: Theorem 5 Consider the direct sum decomposition C = PΛ ⊕ QΛ .p1 +···+ps . Finally we have found a characterization of the projection onto the subspace QΛ . ϕ >>=<< ΦΛ . P (d/dθ) ϕ >> + << α. Then. QΛ = {ϕ ∈ C . (µI − d/dθ)r Q (d/dθ) ϕ >> = << (d/dθ + λI)m α.
j = 1. Bernier and A. [38]. 5. we make no hypotheses on the functional term or on the space. Webb [94]. Moreover. In contrast. In most cases. . on attainability completeness or degeneracy. Many of the considerations of this author are similar to ours. A Variation Of Constants Formula For An Abstract Functional Diﬀerential Equation Of Retarded Type The aim of this section is to obtain a variation of constants formula for an abstract linear nonhomogeneous retarded functional diﬀerential equation. the equation is of the form 0 dη(s)x(t + s) x (t) = Ax(t) + −r where A generates a C0 semigroup on a Banach space X and dη is a suitable restricted Stieltjes measure with values in L(X) (see [78]. Travis and G. 0).C. Nakagiri [29]. Nakagiri includes results on the spectral theory of such equations and the characterization of some generalized eigenspaces in terms of the solutions of an adjoint equation.A theory of linear DDE in inﬁnite dimensional spaces 335 and 0 is the zero vector of dimension p1 + · · · + ps − (q1 + · · · + qs ) . it is assumed that the space X is reﬂexive. We will follow the work made in [7]. to quote a few). [30]. . The work done by S. . Notice that if pj = qj . s. then ϕPΛ = ΦΛ a = ΦΛ << Φ∗Λ . Manitius. [79].F. . [31]. Increasingly elaborate extensions of earlier work by C. D ∈ L1 ((−r. Another important motivation is related to control theory. . 4. L(X)). which takes the form 0 0 dη(s)ϕ(s) = Ai ϕ(−τi ) + D(s)ϕ(s) ds −r −r with Ai ∈ L(X). ϕ >> . [27].6 Conclusion Several extensions of Hale’s theory of functional diﬀerential equations to inﬁnite dimensions exist in the literature. starting from the one given by C. to the case of partial diﬀerential equations with delays have been given by S. The results diﬀer in that some restrictions are imposed by Nakagiri on the measure. [9]. The main motivation is the study of partial diﬀerential equations with ﬁnite or inﬁnite delay.
it is easy to see that Π(t. +∞[. E) ∩ C 1 ([0. Regarding the existence we observe that if xH is a solution of (CP) and xP is a solution of the particular nonhomogeneous problem 3 x (t) = L(xt ) + f (t) . E) −→ C([−r. t ∈ [0. If we extend ϕ by zero to the interval [−r. E) deﬁned by 3 0. Then it is enough to prove the existence of xP . 0]. +∞[. τ ∗ [. (r > 0). x ∈ 1 C ([0. E) −→ E.– For each ﬁxed T > 0. Proof. 0]. +∞[. +∞[. ϕ) := ϕ(t + θ) . if t + θ < 0 Π(t.1 DELAY DIFFERENTIAL EQUATIONS The nonhomogeneous problem Let us consider the Cauchy problem for the nonhomogeneous retarded functional equation: 3 x (t) = L(xt ) + f (t) .1) on [0. T ]. E) which satisﬁes (NH) for t ≥ 0. t ≥ 0 (N H) x0 = ϕ where L : C([−r.336 5. E). let us consider the space C0 ([0. E is a Banach space. ϕ) = ϕt and then (5. t ≥ 0 x(t) = 0 and also 0 t x(t) = Kx(t) + f (s) ds 0 . T ]. A solution of this problem is a function x ∈ C([−r. 0]. T ] . f : [0. ϕ(0) = 0} and the operator Π : [0. E). E). But we already know existence and uniqueness of xH for each initial value ϕ ∈ C([−r.1) x0 = 0 then x = xH + xP satisﬁes (NH). E) which satisﬁes ( 5. τ ∗ [. Proposition 43 Given f ∈ C([0. if t + θ ≥ 0. 0]. E) . Uniqueness of solution for (NH) follows immediately from that of the homogeneous problem. T ]. E). is a bounded linear operator. there exists a unique function xP ∈ C([−r. T ]. x)) ds + f (s) ds . τ ∗ [−→ E (τ ∗ > 0) is a continuous function and ϕ ∈ C := C([−r.1) is equivalent to the integral problem t t L(Π(s. t ≥ 0 (5. T ] × C0 ([0. E) := {ϕ ∈ C([0.
Since KuC([0. 2.E) we obtain the following estimations for the iterates of K: Kn ≤ Ln T n n! . AΦ ˜ := {Φ ∈ C([−r. L(E)) −→ L(E) L(Φ)(b) ˜ L := L(Φ ⊗ b). t ≥ 0 V (t) = L(V V0 = Φ (5. T ]. Proposition 44 The semigroup {T (t)}t≥0 associated to the solution of (CP) is related to the semigroup {T˜(t)}t≥0 by T˜(t)Φ (θ)(b) = T (t) (Φ ⊗ b) (θ) . L(E)). . .2 Semigroup deﬁned in L(E) For each Φ ∈ C([−r. n = 1. Ku(t) = 0 It is enough to prove that I − K is invertible. .E) ≤ T LuC([0. 0]. 0]. E) be the function deﬁned by (Φ ⊗ b) (θ) := Φ(θ)(b) .– Let {U (t)}t≥0 be the family of bounded linear operators deﬁned in E by: U (t)(b) := T (t) (Φ ⊗ b) (0) . Proof. and the result follows easily. L(E)) . 0]. Also. L(E)). let Φ ⊗ b ∈ C([−r. 0]. 5. . 0]. θ ∈ [−r. 0] and then consider the bounded linear operator ˜ : C([−r.2) has a unique solution for each initial value Φ ∈ C([−r. L(E)) a strongly continuous translation semigroup {T˜(t)}t≥0 such that T˜(t)Φ = Vt . t ≥ 0. The inﬁnitesimal generator of this semigroup is ˜ := Φ ˙ .T ].337 A theory of linear DDE in inﬁnite dimensional spaces where K is the bounded linear operator deﬁned on C0 ([0.T ]. Theorem 1 can be used to conclude that the Cauchy problem 3 ˜ t) . ∀b ∈ E . t ≥ 0. 0]. E) by t L(Π(s. this solution deﬁnes on C([−r. Φ(0) ˙ ˜ D(A) = L(Φ)}. b ∈ E. θ ∈ [−r. 0]. u)) ds.
L(E)). 0]. So. 5. we look at the equation. t ≥ 0. which. L(E)) satisﬁes (5. the function t −→ U(t)e is continuous in R+ . deﬁned by U(t)e := (xe ) (t) . Also. t ≥ 0 x0 = 0. the function U ∈ C([0. θ ∈ [−r. 0]. by deﬁnition of the semigroup {T˜(t)}t≥0 and uniqueness of the so˜ in C([−r. Deﬁnition 41 The fundamental solution of problem (NH) is the family of operators {U(t)}t≥0 . Ut = ˜ T (t)Φ. 0]. Observe that U(0) = I and U(t)e = L(xt e ) + e. t + θ ≥ 0. by density of D(A) equality extends to all Φ. we have: Ut (θ)(b) = T (t + θ) (Φ ⊗ b) (0) = T (t) (Φ ⊗ b) (θ) and U = Φ in [−r. Also. and then ˜ t )(b).338 DELAY DIFFERENTIAL EQUATIONS Since {T (t)}t≥0 is a translation semigroup. Finally. for t > 0.3 The fundamental solution Let e ∈ E be given. Next.2) . θ ∈ [−r. We denote by xe (t) the unique solution of the problem 3 x (t) = L(yt ) + e . t ≥ 0. let Φ ∈ D(A) prove that Φ ⊗ b ∈ D(A) for all b ∈ E. lution of equation (5. it is easy to prove that for each e ∈ E ﬁxed. U(t) : E −→ E. t + θ ≤ 0.3) 0 We now deﬁne the fundamental solution of (NH). The proof of the proposition is complete. in integral form. U (t)(b) = T (t)A (Φ ⊗ b) (0) = AT (t) (Φ ⊗ b) (0) = L (T (t) (Φ ⊗ b)) = L(U That is. 0]. +∞[.2). Ut (θ)(b) = Φ(t + θ)(b) = (Φ ⊗ b) (t + θ) = T (t) (Φ ⊗ b) (θ). reads as: t e (x )(t) = L(xes ) ds + et . t ≥ 0. .2). ˜ be the initial value in problem (5. for t > 0. We have ˜ t )(b) = L(Ut ⊗ b) = L (T (t) (Φ ⊗ b)) L(U . (5. It is easy to Now.
. More precisely. .– Let t1 .0] t = L xs e C ds + te. Proposition 45 The function t −→ V(t) is continuously diﬀerentiable from R+ into L(E). t ≥ 0.339 A theory of linear DDE in inﬁnite dimensional spaces Lemma 46 For all t ≥ 0. 0 Therefore U(t)e ≤ Lxt e C + e ≤ e 1 + tLet L . from (5. Proof.– Given e ∈ E. We are now going to prove the continuity in R+ of the function t −→ U(t) and to this end we need to introduce another family of operators which is related to the fundamental solution by integration. 0 This implies that ([74]): t xt C ≤ teexp e L ds = teet L . let {V(t)}t≥0 be the family of operators deﬁned by: V(t) : E −→ E We have V(t)e = t V(t)e := xe (t). t2 ∈ R+ : V(t1 ) − V(t2 ) = ≤ 2 2 sup V(t1 )e − V(t2 ) = sup 2 2 e≤1 sup e e≤1 ≤ t2 t1 e≤1 t2 t2 t1 2 2 U(s)e ds2 2 U(s) ds t1 1 + sLesL ds −→ 0 (t1 − t2  → 0).3) we have: t+θ e e L sup x s C ds + (t + θ)e xt C ≤ 0 θ∈[−t. we have U(t) ∈ L(E). (xe ) (s) ds = 0 and also V(t)e ≤ e t t U(s)e ds 0 t 1 + sLes L ds U(s) ds ≤ e 0 0 which proves that V(t) ∈ L(E) . Proof.
5) An immediate consequence of Proposition 45 is the continuity of the fundamental solution: Corollary 6 The function t −→ U(t) is continuous from R+ into L(E). Vt (θ) = 0. t ≥ 0 U(t) = L 0 U0 = X0 .340 DELAY DIFFERENTIAL EQUATIONS The above inequality yields local Lipschitz continuity. These sections are related to V(t) by: t+θ Vt (θ) = V(t + θ) = 0 U(s) ds = and therefore we can easily show: t Uτ (θ) dτ .5) we conclude that the fundamental solution satisﬁes the Cauchy problem: t ˜ Uτ dτ + I . into C([−r. we arrive at t ˜ s ) ds + tI . 0[. the extension is not continuous but we can deﬁne the sections of U(t). Since V(0) = 0. the same holds for the map t −→ L(V Observe that xes (θ) = xe (s + θ) = V(s + θ)(e) = Vs (θ)e = (Vs ⊗ e) (θ) hence ˜ (Vs ) (e). 0] by 0. The fundamental solution satisﬁes a retarded equation which is formally similar to (5. we can confrom which.2). t ≥ 0. 0 From (5. Therefore. L(xes ) = L (Vs ⊗ e) = L Substituting the righthand side of the above formula for L(xes ) in (5.3). Since U(0) = I. 0]. local Lipschitz continuity is preserved. To see this. (5. We extend V to the interval [−r. t≥0 . It also holds for the map t −→ Vt from R+ ˜ t ). we extend by zero the fundamental solution to the interval [−r.4) L(V 0 ˜ t ). t −θ Uτ (θ) dτ . 0]. L(E)). if t+θ <0 if t+θ ≥0 θ ∈ [−r. in view of local Lipschitz continuity of L(V clude that V(t) is continuously diﬀerentiable for t ≥ 0 and it holds: ˜ t) + I V (t) = L(V . t ≥ 0 V(t) = (5.
A theory of linear DDE in inﬁnite dimensional spaces 341 where X0 (θ) := 0 if θ ∈ [−r. x (t) = 0 Proof. t ≥ 0. E). In this case we have d d [V(t − s) (f (s))] = V(t − s) (f (s)) + V(t − s) f (s) . X0 (0) := I. we can prove the continuous dependence of xP upon the function f . . Hence it is enough to choose f ∈ C 1 ([0. ds ds We deﬁne y(t) := = t t d U(t − s) (f (s)) ds = − V(t − s) (f (s)) ds 0 0 ds t V(t − s) f (s) ds . 0]. t ≥ 0 V(t) (f (0)) + 0 and y(t) = 0 in [−r. 0[. we may write formally: Ut = T˜(t)X0 5.1) in terms of the fundamental solution. t ≥ 0.– By arguments similar to those employed in Lemma 46. t ≥ 0 L(U U(t) = 0 U(0) = X0 and also formally in the diﬀerential form: 3 ˜ t) . Then we have: t P U(t − s) (f (s)) ds . U (t) = L(U U0 = X0 . Theorem 7 Let xP be the solution of ( 5. t>0 Using the semigroup {T˜(t)}t≥0 with some abuse. This retarded functional equation can be written formally in the integral form: t ˜ τ ) dτ + I . The fundamental solution and the nonhomogeneous problem We keep the above notations and we are going to express the solution of the particular nonhomogeneous problem (5.4 .1). T ]. T > 0.
t+θ t+θ We recall that the solution of (NH) is x = xH + xP and so we obtain for each initial value ϕ ∈ C([−r.– We do not carry out the details of the calculations which are based on t t Ut−s (θ) (f (s)) ds = U(t + θ − s) (f (s)) ds = 0. . which completes the proof of the theorem. We now prove that y satisﬁes the same equation as xP : t y (t) = V (t) (f (0)) + V (t − s) f (s) ds 0 t ˜ (Vt−s ) f (s) ds + f (t) ˜ t ) (f (0)) + L = L(V 0 t = L Vt ⊗ f (0) + Vt−s ⊗ f (s) ds + f (t). Vt ⊗ f (0) + 0 Substituting yt for the lefthand side of the above expression in the formula of y (t) we obtain: y (t) = L(yt ) + f (t) together with y(t) = 0 for t ≤ 0. E) the following formula: t U(t − s) (f (s)) ds . t ≥ 0. Lemma 47 The sections of the solution xP are given in terms of the fundamental solution by t P Ut−s ⊗ f (s) ds . but we do not assure the diﬀerentiability of y in t = 0. 0 On the other hand. +∞[. we can easily check that t Vt−s ⊗ f (s) ds = yt . 0]. since y(0) = 0.342 DELAY DIFFERENTIAL EQUATIONS Observe that y is continuous in [−r. (x )t = 0 Proof. t ≥ 0. t ≥ 0 x(t) = T (t)ϕ(0) + 0 and also t xt = T (t)ϕ + 0 Ut−s ⊗ f (s) ds . so y(t) = xP (t). from the deﬁnition of y for t ≥ 0 and the fact that both y and V are identically for t < 0.
xH . 0] ϕ(t) ˜ := ϕ(0) if t > 0. 0 Finally. Finally we can write t ˜ I) ˜ ds (ϕ(0)) + U(s)L( x(t) = I + 0 0 min(t.r) t U(t − s) (Lϕ(0)) ds + U(t − s) (Lϕs ) ds x(t) = ϕ(0) + 0 0 where ϕ(0) after L in the ﬁrst integral is to be considered as a constant function. t > 0. y(t) = 0 Coming back to the original unknown function x. E). we obtain: t U(t − s) (Lϕ˜s ) ds . +∞[. t > 0 y0 = 0. 0]. in terms of the fundamental solution U(t).r) U(t − s) (Lϕs ) ds . Given ϕ ∈ C([−r. t > 0.343 A theory of linear DDE in inﬁnite dimensional spaces Formally: t xt = T (t)ϕ + T˜(t − s)X0 ⊗ f (s) ds. deﬁned by x(t) = y(t) + ϕ(t) ˜ we have that y satisﬁes the nonhomogeneous problem: 3 y (t) = L(yt ) + L(ϕ˜t ) . let us deﬁne ϕ˜ ∈ C([−r. we can also express the solution of the homogeneous equation. we arrive at min(t. E): 3 ϕ(t) if t ∈ [−r. −r ≤ t ≤ +∞ and using this expression in the quantity under the integral. Changing the unknown function x into y. Theorem 7 applies and yields the following expression t U(t − s) (L(ϕ˜s )) ds . x(t) = ϕ(0) + 0 Rewriting ϕ˜ in the form ϕ(t) ˜ = ϕ(t) + ϕ(0) .
L(g(t)xt ) > = < y(t). x(t) > + < u . b] we have: t << yt .– This can be done by calculations quite similar to those employed for the homogeneous problem. 0]. Then. xt >>=<< y0 .344 DELAY DIFFERENTIAL EQUATIONS ˜ := I. Integrating. 5.5 Decomposition of the nonhomogeneous problem in C([−r. f (s) > ds << Φ∗Λ . 0 Proof. 0] −→ L(E) deﬁned by I(θ) Comparing this representation with the solution of the homogeneous Cauchy problem associated with a constant initial data. xt >>= e−BΛ t << Φ∗Λ . is deﬁned. xt >> = < y(t). x(t) ˙ > dt + < u∗ . xt >>=< y(t). Hence we don’t carry out the details. We have 0 ∗ g(t + ξ − ∗)x(t + ξ) dξ > << yt . f (t) > . with a < b. Lemma 48 Let x(t) be a solution to (NH) deﬁned in [a. formally adjoint to L. 0]. ϕ >> + 0 . f (s) > ds. L ∗ and then d << yt . for all t ∈ [a. b]. deﬁned in ] − ∞. Then t ∗ ∗ < eBΛ (s−t) Φ∗Λ (0). x0 >> + < y(s). L(gˆt ⊗ x(t)) > − < u∗ . E). E) We keep the notations of sections 3 and 4. the lemma follows. ˙ x(t) > + < y(t). +∞[ and let y(t) = g(t)u∗ be a solution of the formally adjoint equation y(t) ˙ = −L∗ (yt ) t≤0 . we conclude that the ﬁrst term is the restriction of the semigroup {T (t)}t≥0 to the constant functions. In what follows we accept that the measure m associated with the operator L is of bounded variation and therefore the operator L∗ . with I˜ : [−r. Proposition 46 Let xt be the section of the solution to (NH) corresponding to the initial value ϕ ∈ C([−r.
b > . 0]. f (s) > ds. f (s) > ds 0 t < eBΛ (t−s) Φ∗Λ (0). . xt >> by diﬀerentiation we obtain ˙ Z(t) = −BΛ∗ Z(t)+ < Φ∗Λ (0). . .345 A theory of linear DDE in inﬁnite dimensional spaces where BΛ∗ := diag(Bλ∗1 . f (s) > ds. Since b∈E . if we deﬁne Z(t) =<< ΦΛ ∗ . j = 1. . ∗ Proof. x0 >> + 0 Also. ϕ >> + < eBΛ (s−t) Φ∗Λ (0). ϕ >> + ΦΛ eBΛ (t−s) < Φ∗Λ (0). xt >>= e−BΛ t << Φ∗Λ . f (s) > ds. f (t) > . ϕ >> + 0 From this we can obtain the projection of the solution xt on the subspace PΛ in terms of the semigroup {T (t)}t≥0 : t xt P = ΦΛ eBΛ t << Φ∗Λ . L(E)) be the operator deﬁned by X0 P := ΦΛ < Φ∗Λ (0). . · > that is. << e ΦΛ . s and hence BΛ∗ = −BΛ . b >) . section of the solution of (NH). f (s) > ds 0 t T (t − s)ΦΛ < Φ∗Λ (0). = eBΛ t << Φ∗Λ .– We apply Lemma 48 to yt = eBΛ t Φ∗Λ and then t ∗t ∗ ∗ BΛ ∗ < eBΛ s Φ∗Λ (0). we have xt = xt P + xt Q . Let us suppose that that pj = qj . . We apply the decomposition of the space C([−r. T˜(t)X0 P (θ)(b) = T (t) X0 P ⊗ b (θ) = T (t) X0 P (θ)(b) = T (t) (ΦΛ (θ) < Φ∗Λ (0). 0]. X0 P (θ)(b) := ΦΛ (θ) < Φ∗Λ (0). . . E) stated in theorems of section 4 and for xt . = T (t)ϕP + 0 Let X0 P ∈ C([−r. xt >>=<< ΦΛ . xt Q ∈ QΛ where xt P = ΦΛ a and t ∗ ∗ a = << Φ∗Λ . θ ∈ [−r. Bλ∗s ). 0]. xt P ∈ PΛ .
. the formal adjoint operator L∗ is not deﬁned but the last decomposition formula remains still valid. In fact. Proposition 46 can be proved directly. 0 Notice that if the measure m is not of bounded variation. so that the last decomposition formulas are also valid.346 DELAY DIFFERENTIAL EQUATIONS we ﬁnally obtain the characterization of the projection xt P in terms of the semigroups {T (t)}t≥0 . {T˜(t)}t≥0 and the operator X0 P : t P P xt = T (t)ϕ + T˜(t − s)X0 P ⊗ f (s) ds.
to ﬁnd a 347 O. X) denotes the space of continuous functions from [−r. 64000 Pau. we mean an evolution equation of the type du (t) = A0 u(t) + F (t. where A0 : D(A0 ) ⊆ X → X is a linear operator.). F is a function from [0. Avenue de l’universit´e. (eds. Suppose that r > 0 is a given real number. X). Marrakech.ac. Delay Differential Equations and Applications.fr 1. 0] to X with the uniform convergence topology and we will use simply CX for C ([−r. Introduction Let (X. France adimy@univpau. . X). For u ∈ C ([−r. Adimy D´epartement de Math´ematiques Appliqu´ees. Morocco.Chapter 9 THE BASIC THEORY OF ABSTRACT SEMILINEAR FUNCTIONAL DIFFERENTIAL EQUATIONS WITH NONDENSE DOMAIN K. Arino et al. ERS 2055. C ([−r. +∞) × CX into X and ϕ ∈ CX are given.1) is the following : given ϕ ∈ CX . 0] . ut ). ·) be an inﬁnite dimensional Banach space and L(X) be the space of bounded linear operators from X into X. t ≥ 0. −r ≤ θ ≤ 0. ezzinbi@ucam. let ut denote the element of CX deﬁned by ut (θ) = u(t + θ).ma M.1) dt u0 = ϕ. Universit´e de Pau. (1. 0] . © 2006 Springer. b] . b > 0 and t ∈ [0. Ezzinbi University Cadi Ayyad . By an abstract semilinear functional diﬀerential equation on the space X. The initial value problem associated with (1. b]. 347–407.
1) for t ∈ [0. the space C 1 with null value on the boundary is nondense in the space of continuous functions. from restrictions made on the space where the equation is considered (for example. ∀ n ∈ N. for example by using the variationofconstants formula ([94]. periodic continuous functions. [96]) 3 u(t) = T0 (t)ϕ(0) + ϕ(t). ω0 ∈ IR such that if λ > ω0 (λI − A0 )−1 ∈ L(X) and 2 2 2(λ − ω0 )n (λI − A0 )−n 2 ≤ M0 . and the references therein. Nondensity occurs. it is sometimes convenient to take initial functions with more restrictions. In this case one can prove existence and uniqueness of a solution of (1. it is necessary to impose additional restrictions. us )ds.348 DELAY DIFFERENTIAL EQUATIONS continuous function u : [−r. for t ∈ [0. H¨ older continuous functions) or from boundary conditions (e. if F = 0. [79]). A0 is an operator verifying (i) and (ii). equivalently when (i) D(A0 ) = X. h > 0. In the case where the mapping F in Equation (1.1). for every ϕ ∈ CX . One can refer for this to [64] for more details. t 0 T0 (t − s)F (s. Webb [95]. A case which is easily handled is when F takes their values in D(A0 ). if t ≥ 0. [83]). Memory ([82]. h) such that u(t) ∈ D(A0 ).1) is equal to zero. In all of the quoted papers. h) → X. On the other hand. h) and u0 = ϕ. (ii) there exist M0 . Fitzgibbon [72]. in many situations.. Wu [96]. But. h) and u satisﬁes the evolution equation of (1. There are many examples in concrete situations where evolution equations are not densely deﬁned. For related results. 0] . Kunish and Schappacher ([78]. Only hypothesis (ii) holds. It is wellknown (see for example [94] and [96]) that the classical semigroup theory ensures the well posedness of Problem (1.1) when A0 is the inﬁnitesimal generator of a C0 semigroup of bounded linear operators (T0 (t))t≥0 in X or. the problem can still be handled by using the classical semigroups theory because A0 generates a strongly continuous semigroup in the space D(A0 ). the integrated semigroups theory allows the range of the operators F to be any subset of X. Let us now brieﬂy discuss the use of integrated semigroups.g. In the applications. . if t ∈ [−r. diﬀerentiable on [0. see for example Travis and Webb [94].
a. The problem ( 1. T ] × CX → X is deﬁned by F (t. T ] × CX → X is deﬁned by F (t. ut (·. t ∈ [0. l] . l] . u(0) = 0 . (1. a) ∈ [−r.2) where ϕ is a given function on CX := C ([−r. ∆u ∈ C Ω.2) as an abstract semilinear functional diﬀerential equation 3 t ∈ [0. ut (·. . (t. T ]. and F : [0. t ∈ [0. Of course. IR . a)). (1. a) ∈ [0. the operator A0 satisﬁes (ii) but the domain D(A0 ) is not dense in X and so. a. and F : [0. T ] . ϕ)(a) = f (t. x)) for t ∈ [0.3). 0) = 0. T ] . a) = ϕ(θ. T ] . × [0. X). u(θ. x ∈ Ω. where 3 & ' D(A0 ) = u ∈ C 1 ([0. a). ∆ is the Laplace operator in the sense of distributions on Ω and ϕ is a given function on CX :=C([−r. θ ∈ [−r. ϕ)(x) = f (t.4) can be reformulated as the abstract semilinear functional diﬀerential equation ( 1. x) = ∆u(t. ∂t ∂a u(t. u(θ. ϕ ∈ CX and x ∈ Ω. a) + (t. ∂t u(t. IR). x) + f (t. A0 does not generate a C0 −semigroup. T ] . we can reformulate the partial diﬀerential problem ( 1. with X = C ([0. IR) . V (t) = A0 V (t) + F (t. l] . (1. Example 4 Consider the reactiondiﬀusion equation with delay described by ∂u (t. ϕ ∈ CX and a ∈ [0.3) V0 = ϕ ∈ CX . there are many other examples encountered in the applications in which the operator A0 satisﬁes only (ii) (see [64]). X). l]. x. t ∈ [0. In the two examples given here. ϕ(·. x) = ϕ(θ. x) = 0. with 3 & ' D(A0 ) = u ∈ C Ω. x ∈ Ω. l] . T ] . ϕ(·. x ∈ ∂Ω. 0] .Basic theory of abstract DFE with nondense domain 349 Example 3 Consider the model of population dynamics with delay described by ∂u ∂u (t. (θ. IR . a) = f (t. IR and u = 0 on ∂Ω . By setting V (t) = u(t. x). 0] . with X = C Ω. ·). Vt ).4) where Ω ⊂ IRn is a bounded open set with regular boundary ∂Ω. x)). 0] × [0. A0 u = −u . T ]. x. 0] . A0 u = ∆u. a)) for t ∈ [0.
We start with a few deﬁnitions. we proceed to establish the main results. the existence of bounded. We prove the existence. [88] and [90]) to study. we give a short review of the theory of integrated semigroups and diﬀerential operators with nondense domain. s (iii) for any t. We prove that the solutions generate a nonlinear strongly continuous semigroup.1). In Section 4. . unstable and center. Basic results In this section. In Section 6. the solutions are shown to generate a locally Lipschitz continuous integrated semigroup. the existence of a direct sum decomposition of a state space into three subspaces : stable. S(t)x is a continuous function of t ≥ 0 with values in X. As a consequence of the results established in Section 6. In the end. the stability of Equation (1. in the nonlinear autonomous case. we show in the linear autonomous case. we use a principle of linearized stability for strongly continuous semigroups given by Desh and Schappacher [65] (see also [87].350 DELAY DIFFERENTIAL EQUATIONS We will study here the abstract semilinear functional diﬀerential equation (1. which satisﬁes a compactness property. Deﬁnition 42 [56] Let X be a Banach space. 2.1) in the case when the operator A0 satisﬁes only the HilleYosida condition (ii). These results give natural generalizations of results in [94] and [96]. regularity and continuous dependence on the initial condition. uniqueness. A family (S(t))t≥0 ⊂ L(X) is called an integrated semigroup if the following conditions are satisﬁed : (i) S(0) = 0. (ii) for any x ∈ X. In the linear case. We derive a variationofconstants formula which allows us to transform the integral solutions of the general equation to solutions of an abstract Volterra integral equation. we consider the autonomous case. s ≥ 0 S(s)S(t) = (S(t + τ ) − S(τ ))dτ. if there exist constants M ≥ 0 and ω ∈ IR such that S(t) ≤ M eωt for t ≥ 0. which are semigroup invariants. In Section 5. After providing some background materials in Section 2. periodic and almost periodic solutions is established in the sections 7 and 8. A natural generalized notion of solutions is provided in Section 3 by integral solutions. we give some examples. 0 Deﬁnition 43 [56] An integrated semigroup (S(t))t≥0 is called exponentially bounded.
0 remark 2 If an operator A is the generator of an integrated semigroup (S(t))t≥0 . +∞) ⊂ ρ(A) (the resolvent set of A) such that R(λ) = (λI − A)−1 . for all λ > ω. If (S(t))t≥0 is an integrated semigroup. Then for all x ∈ X and t ≥ 0. and in the case when (S(t))t≥0 is nondegenerate.Basic theory of abstract DFE with nondense domain 351 Moreover (S(t))t≥0 is called nondegenerate if S(t)x = 0. S(t)x = tx + 0 . t t S(s)xds ∈ D(A) and S(t)x = A S(s)xds + tx. for all x ∈ D(A). for all t ≥ 0. t S(s)Ax ds. 0 Proposition 47 [56] Let A be the generator of an integrated semigroup (S(t))t≥0 . then ∀λ ∈ IR. there exists a unique operator A satisfying (ω. Deﬁnition 44 [56] An operator A is called a generator of an integrated semigroup. A − λI is the generator of the integrated semigroup (Sλ (t))t≥0 given by t −λt Sλ (t) = e S(t) + λ e−λs S(s) ds. and there exists a strongly continuous exponentially bounded family (S(t))t≥0 of linear bounded operators such that S(0) = 0 and (λI − A)−1 = +∞ λ e−λt S(t)dt. exponentially bounded. then +∞ e−λt S(t)dt exists for all λ with the Laplace transform R(λ) := λ 0 Re(λ) > ω. R(λ) is injective if and only if (S(t))t≥0 is nondegenerate. t ≥ 0 S(t)x ∈ D(A) and and AS(t)x = S(t)Ax. for all Re(λ) > ω. +∞) ⊂ ρ(A). This operator A is called the generator of (S(t))t≥0 . R(λ) satisﬁes the following expression R(λ) − R(µ) = (µ − λ)R(λ)R(µ). We have the following deﬁnition. implies that x = 0. if there exists ω ∈ IR such that (ω. 0 0 Moreover.
if for all τ > 0 there exists a constant k(τ ) > 0 such that S(t) − S(s) ≤ k(τ ) t − s . . we give some results for the existence of solutions of the following Cauchy problem du (t) = Au(t) + f (t). where A satisﬁes the condition (HY). n ∈ N. Theorem 2 [81] The following assertions are equivalent. λ > ω ≤ M. (2. In this case. that (S(t))t≥0 is exponentially bounded. X) satisfying u(t) ∈ D(A) (t ∈ [0. τ ] . without being densely deﬁned. T ] where T > 0. Deﬁnition 46 [81] We say that a linear operator A satisﬁes the HilleYosida condition (HY) if there exist M ≥ 0 and ω ∈ IR such that (ω. Moreover.352 DELAY DIFFERENTIAL EQUATIONS Corollary 1 [56] Let A be the generator of an integrated semigroup (S(t))t≥0 . +∞) ⊂ ρ(A) and 2 2 ' & (HY) sup (λ − ω)n 2(λI − A)−n 2 . s ∈ [0.1) hold. The following theorem shows that the HilleYosida condition characterizes generators of locally Lipschitz continuous integrated semigroups.1) on [0. An important special case is when the integrated semigroup is locally Lipschitz continuous (with respect to time). we understand a function u ∈ C 1 ([0. Then for all x ∈ X and t ≥ 0 one has S(t)x ∈ D(A). (ii) A satisﬁes the condition (HY). Deﬁnition 45 [81] An integrated semigroup (S(t))t≥0 is called locally Lipschitz continuous. for x ∈ X. In the sequel. T ] . we know from [81]. The following result is due to Da Prato and Sinestrari. S(·)x is rightsided diﬀerentiable in t ≥ 0 if and only if S(t)x ∈ D(A). In that case S (t)x = AS(t)x + x. for all t. By a solution of Problem (2. t ≥ 0. (i) A is the generator of a locally Lipschitz continuous integrated semigroup.1) dt u(0) = x ∈ X. T ]) such that the two relations in (2.
Furthermore. for t ≥ 0. we say that u : [0. (iii) u(t) = x + A 0 0 ¿From this deﬁnition. and for each t ∈ [0. X).1) if the following assertions are true (i) u ∈ C([0. x ∈ D(A) is a necessary condition t for the existence of an integral solution of (2. Problem (2. for all t > 0. u(t) ≤ M e 0 . +∞) → X is a continuous function.1) may still have an integral solution. we deduce that for an integral solution u. (ii) 0 t t u(s)ds + f (s)ds. x ∈ D(A) and f : [0. because u(t) = lim u(s)ds and h→0 h t t+h u(s)ds ∈ D(A). t (b) f (t) = f (0) + 0 g(s) ds for some Bochnerintegrable function g. In particular. for t ≥ 0. the function u satisﬁes the inequality t ωt −ωs x + e f (s) ds . following the work of Da Prato and Sinestrari [64]. for t ≥ 0. u(t) = S (t)x + dt 0 where S(t) is the integrated semigroup generated by A. +∞) → X is an integral solution of ( 2. Deﬁnition 47 [64] Given f ∈ L1loc (0. Theorem 4 [61] Suppose that A satisﬁes the condition (HY). we 1 t+h have u(t) ∈ D(A).1) has a unique integral solution which is given by d t S(t − s)f (s)ds. u(t) ≤ M e 0 In the case where x is not suﬃciently regular (that is. (c) Ax + f (0) ∈ D(A). x is just in D(A)) there may not exist a strong solution u(t) ∈ X but.1). +∞) . x ∈ D(A) such that (a) A satisﬁes the condition (HY). Then there exists a unique solution u of Problem ( 2. pour t ≥ 0. X) and x ∈ X. T ]. t u(s)ds ∈ D(A). T ] → X.Basic theory of abstract DFE with nondense domain 353 Theorem 3 [64] Let A : D(A) ⊆ X → X be a linear operator. Then the problem ( 2. f : [0. T ] t ωt −ωs x + e f (s) ds .1) on the interval [0. +∞.
by using the next result and Theorem 2. dt u0 = ϕ ∈ CX . In view of the remark following Deﬁnition 44. u (t) = (Aut ) (0) . We know from Theorem 2 that A0 is the generator of a locally Lipschitz continuous integrated semigroup (S0 (t))t≥0 on X. Aϕ = ϕ . Consider ﬁrst the linear Cauchy problem 3 t ≥ 0. we will sometimes assume without loss of generality that ω0 = 0. This is 3 t ≥ 0. T ] × CX → X is a continuous function. We can show. where 3 & D(A) = ϕ ∈ C 1 ([−r. t ≥ 0. This problem can be reformulated as a special case of an abstract semilinear functional diﬀerential equation with delay. +∞) ⊂ ρ(A0 ) and 2 2 ' & sup (λ − ω0 )n 2(λI − A0 )−n 2 . X) . Throughout this work. t + θ < 0. t + θ ≥ 0. Proposition 48 The operator A is the generator of a locally Lipschitz continuous integrated semigroup on CX given by 0 ϕ(s) ds + S0 (t + θ)ϕ(0). ϕ(0) ∈ D(A0 ). ' ϕ (0) = A0 ϕ(0) . θ . 0] . (1. u (t) = A0 u(t). 3. t 0 S(t − s)f (s)ds is diﬀerentiable Existence.354 DELAY DIFFERENTIAL EQUATIONS Note that Theorem 4 also says that with respect to t. n ∈ N. uniqueness and regularity of solutions We restate the problem du (t) = A0 u(t) + F (t. u0 = ϕ. that the operator A satisﬁes the condition (HY). we assume that A0 satisﬁes the HilleYosida condition on X : (HY) there exist M0 ≥ 0 and ω0 ∈ IR such that (ω0 . λ > ω0 ≤ M0 . θt+θ (S(t)ϕ) (θ) = ϕ(s) ds. ut ). u0 = ϕ ∈ CX . (and (S0 (t))t≥0 is exponentially bounded).1) where F : [0.
In order to prove that A is the generator of (S(t))t≥0 . Consider the equation (λI − A) ϕ = ψ. and we are looking for ϕ ∈ D(A). If t + θ ≤ 0 and s + θ ≤ 0. θ ∈ [−r. τ ].355 Basic theory of abstract DFE with nondense domain for t ≥ 0. s ∈ [0. The above equation reads λϕ(θ) − ϕ (θ) = ψ(θ). . τ ] and ϕ ∈CX . because (S0 (t))t≥0 is Lipschitz continuous on [0. s+θ then (S (t) ϕ) (θ) − (S(s)ϕ) (θ) ≤ k (t + θ) ϕ(0) − (s + θ) ϕ . 0] . θ ∈ [−r. Consider τ > 0. Whose solutions are such that 0 eλ(θ−s) ψ(s) ds. This implies that S (t) − S(s) ≤ (k + 1) t − s . where ψ is given in CX . Proof. If t + θ ≥ 0 and s + θ ≥ 0. ϕ(θ) = eλθ ϕ(0) + θ θ ∈ [−r. If t + θ ≥ 0 and s + θ ≤ 0. It follows immediately that there exists a constant k := k(τ ) > 0 such that (S (t) ϕ) (θ) − (S(s)ϕ) (θ) ≤ k t − s ϕ(0) . t. 0] . we obtain (S (t) ϕ) (θ) − (S(s)ϕ) (θ) = S0 (t + θ)ϕ(0) + 0 ϕ(u)du. we calculate the spectrum and the resolvent operator of A. 0] and ϕ ∈CX . we have (S (t) ϕ) (θ) − (S(s)ϕ)(θ) = S0 (t + θ)ϕ(0) − S0 (s + θ)ϕ(0). It is easy to see that (S(t))t≥0 is an integrated semigroup on CX . It may be concluded that (S(t))t≥0 is locally Lipschitz continuous. we have (S (t) ϕ) (θ) − (S(s)ϕ) (θ) = 0.
+∞) ⊂ ρ(A0 ). On the other hand. Integrating by parts the ﬁrst expression. +∞) ⊂ ρ(A) and (λI − A)−1 ψ (θ) = eλθ (λI − A0 )−1 ψ(0) + 0 eλ(θ−s) ψ(s)ds. So.1) using integrated semigroups. 1 (λI − A)−1 ϕ (θ). λ So. from the formula stated in Proposition 48. (0. We shall prove that this extension determines a locally Lipschitz continuous integrated semigroup on C5X . 0] and λ > 0. for each λ > 0. θ for θ ∈ [−r.356 DELAY DIFFERENTIAL EQUATIONS ϕ is in D(A) if ϕ(0) ∈ D(A0 ) and ϕ (0) = A0 ϕ(0). not larger than ω0 = 0. that is ϕ(0) ∈ D(A0 ) and (λI − A0 ) ϕ(0) = ψ(0). Therefore. Therefore. . The proof of Proposition 48 is complete Our next objective is to construct an integrated version of Problem (1. the above equation has a solution ϕ(0) = (λI − A0 )−1 ψ(0). where X0 = {X0 c. we know that (0. c ∈ X and (X0 c) (θ) = X0 (θ)c} and X0 denotes the function deﬁned by X0 (θ) = 0 if θ < 0 and X0 (0) = IdX . − A0 )−1 ϕ(0) . We obtain +∞ 0 e−λt (S (t) ϕ) (θ) dt −θ +∞ t+θ 0 = 0 e−λt θ ϕ(s)ds dt + −θ e−λt θ ϕ(s)ds dt +∞ + −θ e−λt S0 (t + θ)ϕ(0) dt. it is clear that t → (S(t)ϕ) (θ) has at most exponential growth. A is related to (S(t))t≥0 by the formula which characterizes the inﬁnitesimal generator of an integrated semigroup. λθ = − eλ = = = eλθ λ 0 eλθ λ +∞ −λs 0 −λs ϕ(s)ds + 0 e S0 (s)ϕ(0)ds θ e 0 −λs ϕ(s)ds + (λI θ e . We need to extend the integrated semigroup (S(t))t≥0 to the space C5X = CX ⊕ X0 . it yields +∞ 0 e−λt (S (t) ϕ) (θ) dt 0 ϕ(s)ds + λ1 θ eλ(θ−s) ϕ(s)ds θ +∞ λ(θ−s) λθ 0 + eλ θ ϕ(s)ds + 0 e S0 (s)ϕ(0)ds. By assumption on A0 . one can deﬁned its Laplace transform. for λ > 0.
and the decomposition is clearly unique. c ∈ D(A0 ). 5 ∈ D(A) 5 and λ > 0.Basic theory of abstract DFE with nondense domain 357 5 5 Proposition 49 The family of operators (S(t)) t≥0 deﬁned on CX by 5 S(t)ϕ = S(t)ϕ. D(A) 5 = ϕ + X0 (A0 ϕ(0) − ϕ (0)) . c) ∈ CX × X . Aϕ Proof. Using the same reasoning as in the proof of Proposition 48. we consider the following operator 5 →X l : D(A) ϕ → l(ϕ) = A0 ϕ(0) − ϕ (0). ϕ(0) ∈ D(A0 ) . X) . for c ∈ X. for ϕ ∈ CX and 3 S0 (t + θ)c. Lemma 1 For λ > C 0. 5 S(t)X0 c (θ) = 0. one D has C D 0 1 5 = D(A) ⊕ eλ· . if t + θ ≥ 0. Setting Ψ = Ψ−e 5 λ· (λI −A0 )−1 l(Ψ). (eλ· c)(θ) = eλθ c . +∞) ⊂ ρ(A) −1 +∞ 5 ϕ 5 ϕ 5 = λ 0 e−λt S(t) 5 dt. for λ > 0 and ϕ 5 ∈ C5X . 0] . λI − A For this. 5 one can show that (S(t)) t≥0 is a locally Lipschitz continuous integrated 5 semigroup on CX . where eλ· = eλ· c. is a locally Lipschitz continuous integrated semigroup on C5X generated 5 deﬁned by by the operator A ' & 3 5 = ϕ ∈ C 1 ([−r. +∞) ⊂ ρ(A) 5 −1 (ϕ + X0 c) = (λI − A)−1 ϕ + eλ· (λI − A0 )−1 c. we need the following lemma. 5 we deduce Let Ψ that Ψ ∈ Ker(l) = D(A). Proof of the lemma. λI − A . (ii) Consider the equation 5 ϕ + eλ· c = ψ + X0 a. (i) D(A) 5 and (ii) (0. The proof will be completed by showing that 5 and (0. if t + θ ≤ 0. (λI − A) for every (ϕ. For the proof of (i).
+∞ = λ −θ e−λt S0 (t + θ)cdt. t 1 t U (t−s)U (h)G(s) ds = lim U (t−s)(Aλ G(s)) ds. c) . T ] → E (0 < T ). λI − A This complete the proof of the lemma We now turn to the proof of the proposition. we may restrict our attention to the case when ϕ 5 = X0 c. a ∈ X. −1 5 (ϕ + X0 c) = (λI − A)−1 ϕ + eλ· (λI − A0 )−1 c. we have 5 −1 X0 c (θ) = eλθ (λI − A0 )−1 c. Lemma 2 Let (U (t))t≥0 be a locally Lipschitz continuous integrated semigroup on a Banach space (E. D(A)) and G : [0. Then. for t ∈ [0. and we are looking for (ϕ. This yields (λI − A) ϕ + λeλ· c − λeλ· c − X0 (A0 c − λc) = ψ + X0 a. All we want to show −1 1 5 is that ϕ5 can be expressed as the Laplace transform of λI − A λ 5 ϕ. the function K : [0. c = (λI − A0 )−1 a. which has the solution 3 ϕ = (λI − A)−1 ψ.358 DELAY DIFFERENTIAL EQUATIONS where (ψ. a Bochnerintegrable function. ϕ ∈ D(A). (λI − A) +∞ = λeλθ 0 e−λt S0 (t)c dt. Consequently. T ] → E deﬁned by t U (t − s)G(s) ds K(t) = 0 is continuously diﬀerentiable on [0. T ] . (i) K (t) = lim h→0 h 0 λ→+∞ 0 . c ∈ D(A0 ). which completes the proof of the proposition We will need also the following general lemma. +∞) ⊂ ρ(A). S(t) 5 In view of the decomposition and what has been already done for A. +∞ 5 = λ 0 e−λt S(t)X 0 c (θ) dt. a) is given. T ] and satisﬁes. ·) generated by (A. In this case. 5 (0. ψ ∈ CX .
in the second integral of the righthand side. Theorem 4 implies that K is continuously diﬀerentiable on [0. On the other hand.Basic theory of abstract DFE with nondense domain 359 with Aλ = λ (λI − A)−1 . Proof. It follows that. U (t)U (h)x = 0 for t. T ] K (t) = lim h→0 0 t 1 U (t − s) U (h)G(s) ds. 0 U (t + h − s) − U (t − s) = U (t − s)U (h). and for the prove of (ii) and (iii). for t ∈ [0. h . [76] and [92]. u = h1 (s − t). h ≥ 0 and x ∈ E. (ii) K(t) ∈ D(A). T ] . we know by the deﬁnition of an integrated semigroup that t (U (s + h)x − U (s)x) ds. This yields that the function t → U (t)U (h)x is continuously diﬀerentiable on [0. T ]. h t 0 This implies that 1 lim h→0 h Hence 1 K (t) = lim h→0 h But t+h U (t + h − s)G(s) ds = 0. we have t 1 K (t) = lim (U (t + h − s) − U (t − s)) G(s) ds h→0 h 0 1 t+h + U (t + h − s)G(s) ds . t (iii) K (t) = AK(t) + 0 G(s) ds. we obtain 1 1 t+h U (t + h − s)G(s) ds = U (h(1 − u))G(t + hu) du. h t If we put. x ∈ E. see [61]. and satisﬁes U (t)U (h)x = U (t + h)x − U (t)x. Furthermore. t t (U (t + h − s) − U (t − s)) G(s) ds. for each h ≥ 0.
see [93] We are now able to state a ﬁrst result of existence and uniqueness of solutions. such that ϕ(0) ∈ D(A0 ). 5 0 S(t − s)X0 F (s. Then. of . T ] and ϕ. where A is the generator of the integrated semigroup (S(t))t≥0 on CX .1) y(t) = S (t)ϕ + dt 0 5 with (S(t))t≥0 given in Proposition 48 and S(t) in Proposition 49. we deduce from Corollary 1 that S(·)ϕ is diﬀerentiable and (S (t))t≥0 can be deﬁned to be a C0 semigroup on D(A). T ] . y t ∈ [0. y n (t) = S (t)ϕ + d dt t n−1 (s)) ds. ψ) ≤ L ϕ − ψ . there exists a unique function y : [0. Since ϕ(0) ∈ D(A0 ). for t ∈ [0. T ] × CX → X be continuous and satisfy a Lipschitz condition F (t. the proof of this theorem is standard. for given ϕ ∈ CX . ϕ) − F (t. ψ ∈ CX .360 DELAY DIFFERENTIAL EQUATIONS For the prove of the last equality of (i). (3. we have ϕ ∈ D(A). Theorem 5 Let F : [0. n ≥ 1. t ∈ [0. T ] . T ] t ∈ [0. t≥0 Proof. where L is a positive constant. Using the results of Theorem 4. T ] → CX which solves the following abstract integral equation d t5 S(t − s)X0 F (s. y(s)) ds. Then. Let (y n )n∈N be a sequence of continuous functions deﬁned by y 0 (t) = S (t)ϕ.
there exists α ≥ 0 such that . the continuity of F and S (·)ϕ.
By virtue 0 .
F (s. y (s).
for s ∈ [0. T ] . using Theorem 4. Then. we obtain t . ≤ α.
.
.
1 .
.
y (t) − y 0 (t).
≤ M0 .
y 0 (s).F (s.
0 Hence . ds.
1 .
.
y (t) − y 0 (t).
≤ M0 αt. In general case we have .
n .
.
y (t) − y n−1 (t).
≤ M0 L t.
.
.
y n−1 (s) − y n−2 (s).
0 So. n . ds.
.
n .
y (t) − y n−1 (t).
n! . ≤ M0n Ln−1 α t .
T ] and n→∞ y : [0.Basic theory of abstract DFE with nondense domain 361 Consequently. we introduce the function v deﬁned by .1). In order to prove that y is a solution of Equation (3. the limit y := lim y n (t) exists uniformly on [0. T ] → CX is continuous.
.
.
.
d t5 .
v(t) = .
y(s)) ds.y(t) − S (t)ϕ − S(t − s)X0 F (s.
.
. dt 0 We have v(t) ≤ ≤ ≤ .
.
.
.
y(s)) ds. 5 − s)X0 F (s.
.
. .
y(t) − y n+1 (t).
+ .
.
y n+1 (t) − S (t)ϕ − d t S(t dt 0 .
.
.
.
y n (s))) ds. y(s)) − F (s. 5 − s)X0 (F (s.
.
. .
y(t) − y n+1 (t).
+ .
.
d t S(t dt 0 .
.
.
y(t) − y n+1 (t).
T ] . p=n This implies that v(t) ≤ (1 + M0 L) α L ∞ p=n (M0 L)p+1 tp+1 tn+1 α + (M0 L)n+1 . t 5 (ii) y(s)ds ∈ D(A). we have y(t) − y (t) = n ∞ y p+1 (t) − y p (t) . y(s))ds. + M0 L t y(t) − y n (t) . . T ] . T ] → CX of the abstract integral equation ( 3. (iii) y(t) = ϕ + A 0 0 for t ∈ [0. t ≥ 0. (i) y ∈ C([0. To show uniqueness. y (t) = Ay(t) y(0) = ϕ ∈ CX . i. y(t)). (p + 1)! L (n + 1)! Consequently we obtain v = 0 on [0. for t ∈ [0. Then t y(s) − z(s) ds. 0 Moreover.1) is the unique integral solution of the equation 3 5 + X0 F (t. for n ∈ N.1). T ]. suppose that z(t) is also a solution of Equation (3. 0 t t 5 y(s)ds + X0 F (s. CX ). y(t) − z(t) ≤ M0 L 0 By Gronwall’s inequality. T ] Corollary 6 Under the same assumptions as in Theorem 5. z = y on [0. T ] .e. the solution y : [0.
e. us )ds.3) Furthermore. if we consider the function u : [−r. (3. for t ∈ [0. X). If we deﬁne the function f : [0. T ]. T ] . T ] . the solution y of the integral equation ( 3. for t ∈ [0. . y(s)) ds . 0] the translation property 3 y(t + θ)(0) if t + θ ≥ 0. ut ≤ M0 ϕ + 0 Conversely. we can use Theorem 4 and Theorem 5 to prove this result Corollary 7 Under the same assumptions as in Theorem 5. T ] → X deﬁned by 3 y(t)(0) if t ≥ 0. Then. for t ∈ [0. u is the unique integral solution of the problem ( 1. T ] . for t ∈ [−r. 0] . y(t) ≤ M0 ϕ + 0 Proof. X) is also the unique solution of d t S0 (t − s)F (s. t u(s)ds ∈ D(A0 ). then the function t → ut is a solution of Equation ( 3. (ii) 0 t t u(s)ds + F (s. 0] . us )ds. we have for t ∈ [0. for t ∈ [−r.1) satisﬁes. T ] t F (s. T ] → C5X by f (s) = X0 F (s. (i) u ∈ C([−r. we have.1). (3. where the operator A Furthermore.2) The function u ∈ C([−r.1). T ] and θ ∈ [−r. y(t)(θ) = ϕ(t + θ) if t + θ ≤ 0. us ) ds . ϕ(0) + A0 (iii) u(t) = 0 0 ϕ(t). S0 (t)ϕ(0) + u(t) = dt 0 ϕ(t). Moreover. T ] .1).362 DELAY DIFFERENTIAL EQUATIONS 5 is given in Proposition 49. t F (s. y(s)). if u is an integral solution of Equation ( 1. i. T ] . for t ∈ [0. for t ∈ [0. u(t) = ϕ(t) if t ≤ 0.
¿From Proposition 48 and Proposition 49. Then.1). us ) ) ds. T ] ut − u t ≤ M0 eLt ϕ − ϕ . T ] into X is called an integral solution of Equation ( 1. This yields t v(t) ≤ M0 ϕ − ϕ + F (s. for t ∈ [0.3). respectively. This is just a consequence of the last inequality stated in Corollary 7. y(s)) ds. . ϕ ∈ CX .2) or equivalently u satisﬁes ( 3. t 5 − s)X0 F (s. we obtain 0 S(t (S (t)ϕ) (θ) = ϕ(t + θ). u be the functions given by Corollary 7 for ϕ. we obtain t ≤ M0 eLt ϕ − ϕ ut − u We give now two results of regularity of the integral solutions of (1. y(s)) (θ) ds = 0 and and for t + θ ≤ 0. ut − u 0 By Gronwall’s inequality.1) if the function t → ut satisﬁes Equation ( 3. The second part of the corollary follows from Corollary 6 Remark 8 A continuous function u from [−r. Corollary 9 Assume that the hypotheses of Theorem 5 are satisﬁed and let u. us + v(s)) − F (s. we have for t+θ ≥ 0 d t+θ S0 (t + θ − s)F (s.1) has been centered near u. us ) − F (s. 0 this is t +L t ≤ M0 ϕ − ϕ us − u s ds .363 Basic theory of abstract DFE with nondense domain Proof. dt 0 with v(t) = u t − ut . u s ) ds . y(t)(θ) = S0 (t + θ)ϕ(0) + dt 0 = y(t + θ)ϕ(0). Hence y(t)(θ) = ϕ(t + θ). After the equation (3.1) or equivalently u satisﬁes ( 3. we obtain the equation d t5 − ϕ) + v(t) = S (t) (ϕ S(t − s)X0 (F (s. Proof.
Dϕ F (t. T ] and satisﬁes the Cauchy problem 3 5 + X0 F (t. ψ) ≤ L ϕ − ψ . w(t) = ϕ + 0 We will show that w = y on [0. y(s)) + Dϕ F (s. ψ) ≤ β ϕ − ψ . S(t . t ∈ [0. Dt F (t. y is continuously diﬀerentiable on [0. T ] and satisﬁes the Cauchy problem ( 1.364 DELAY DIFFERENTIAL EQUATIONS Theorem 10 Assume that F : [0. is continuously diﬀerentiable on [−r. T ] . for given ϕ ∈ CX such that ϕ(0) ∈ D(A0 ). for t ∈ [0. ϕ) − Dt F (t. γ ≥ 0 such that F (t. ϕ ∈ CX . + 0 S(t = ϕ + S(t)ϕ + t 0 5 − s)X0 (Dt F (s. y(s))v(s)) ds. y(s))v(s)) ds. y(s))v(s)) ds. T ] → CX which solves the following integral equation d t5 v(t) = S (t)ϕ + S(t − s)X0 (Dt F (s. ψ ∈ CX .1) such that y(0) = ϕ. ϕ (0) ∈ D(A0 ) and ϕ (0) = A0 ϕ(0) + F (0. Proof. for all t ∈ [0. y(s)) + Dϕ F (s. T ] → CX be the function deﬁned by t v(s) ds. u(t) = ϕ(t) if t < 0. We deduce from Theorem 5 that there exists a unique function v : [0.1). y(s)) + Dϕ F (s. we obtain t w(t) = ϕ + 0 S (s)ϕ ds t 5 − s)X0 (Dt F (s. Let w : [0. dt 0 such that v(0) = ϕ . Moreover. ϕ) − Dϕ F (t. T ] → CX be the solution of the abstract integral equation ( 3. Let y be the solution of Equation (3. ψ) ≤ γ ϕ − ψ . Then. ϕ) − F (t. T ] . where Dt F and Dϕ F denote the derivatives. T ] such that y(0) = ϕ. β. Then. the function u deﬁned on [−r.1) on [0. Using the expression satisﬁed by v. T ] by 3 y(t)(0) if t ≥ 0. T ] × CX → X is continuously diﬀerentiable and there exist constants L. let y : [0. ϕ). T ]. y(t)). y (t) = Ay(t) y(0) = ϕ. T ] and ϕ.
ϕ). y(s))) ds t 5 − s)X0 (Dt F (s. we deduce w(t) − y(t) ≤ M0 t (L + β + γ v(s)) w(s) − y(s) ds. On the other hand. w(s))) v(s) ds. + 0 S(t = d t 5 S (t)ϕ + dt 0 S(t − s)X0 F (s. w(s))ds t 5 − 0 S(t − s)X0 (Dt F (s. ϕ). This implies that 5 5 Aϕ 5 + S(t)X 5 = S(t) S(t)ϕ = S(t)ϕ 0 F (0. T ]. w(s)) + Dϕ F (s. T ] . y(s)) − Dt F (s. t ∈ [0. w(s))) v(s) ds. ϕ). Using Corollary 1.Basic theory of abstract DFE with nondense domain 365 5 and ϕ (0) = A0 ϕ(0) + F (0. 5 5 Aϕ+g(0) 5 The assumptions of Theorem 10 imply that ϕ ∈ D(A). + 0 S(t We obtain w(t) − y(t) = d t 5 dt 0 S(t − s)X0 (F (s. This implies that y is continuously diﬀerentiable on [0. z (t) = Az(t) (3. we deduce that 5 S(t)ϕ = S (t)ϕ − ϕ + S(t)X 0 F (0. we have d t 5 dt 0 S(t − s)X0 F (s. Using Theorem 3. w(s))v(s)) ds t 5 − s)X0 (Dt F (s. w(s))] v(s)ds + S(t)X 5 S(t 0 F (0. y(s)) + Dϕ F (s. ϕ). Consider now the function g : [0. T ] → C5X deﬁned by g(t) = X0 F (t. w(s))) ds t 5 − s)X0 (Dϕ F (s. Furthermore. w(s)) + Dϕ F (s. w(s)) ds = = t 0 d dt t 0 5 S(s)X 0 F (s.4) z(0) = ϕ. Then d t 5 w(t) = S (t)ϕ + dt 0 S(t − s)X0 F (s. T ]. So. w(s)) − F (s. T ]. we have ϕ ∈ D(A) 5 then ϕ = Aϕ + X0 F (0. y(t)) and consider the Cauchy problem 3 5 + g(t). w(s))ds t 5 + 0 S(t − s)X0 (Dt F (s. y(s))v(s)) ds. y(s)) − Dϕ F (s. ∈ D(A) and g is continuously diﬀerentiable on [0. 0 By Gronwall’s inequality. ϕ). y(s)) − Dϕ F (s. y(s)) − Dt F (s. we conclude that w = y on [0. w(s))) ds + 0 S(t t 5 + 0 S(t − s)X0 (Dϕ F (s. w(t − s)) ds 5 − s)X0 [Dt F (s. we .
for t ∈ [0.4). T ] and ψ. s ∈ [0. The second part of the theorem is a consequence of Corollary 6 Assume that T > r and A0 : D(A0 ) ⊆ X → X satisﬁes (with not necessarily dense domain) the condition π there exist β ∈ 2 . ϕ) ≤ L (t − sα + ψ − ϕ) for t. us )ds.366 DELAY DIFFERENTIAL EQUATIONS deduce that there exists a unique solution on [0. T ] by 3 y(t)(0) if t ≥ 0. T ]. lim A0 h→0 h t Furthermore. ut ).5) (nondensely deﬁned) on X and there exist a constant L > 0 and α ∈ ]0. for given ϕ ∈ CX . we obtain u(t) ∈ D(A0 ) and u (t) = A0 u(t) + F (t.5) is stronger than (HY). We have the following result. we know that this solution is given by d t5 z(t) = S (t)ϕ + S(t − s)g(s)ds. ϕ ∈ CX . 2(λI − A0 ) 2 ≤ M λ The condition (3. T ] . then 2 ∈ C − {0} and −1 2 0 2 . Then. Theorem 11 Suppose that A0 satisﬁes ( 3. If we consider the function u deﬁned on [−r. 0 We have also the existence of t+h 1 u(s) ds = u (t) − F (t. By Theorem 4. T ] . the operator A0 is closed. ψ) − F (s. the integral solution . 1[ such that F (t. u(t) = ϕ(t) if t < 0. π and M0 > 0 such that if λ (3. we have t 0 t u(s)ds + u(t) = ϕ(0) + A0 0 u(s)ds ∈ D(A0 ) and t F (s. T ] of Equation (3. such that ϕ(0) ∈ D(A0 ). Then. ut ). By virtue of Corollary 7. dt 0 Theorem 5 implies that z = y on [0. for t ∈ [0.5) 2 arg λ < β.
We know. we obtain that the function u is γ−H¨older continuous on [ε. Theorem 12 Suppose that F : [0. Consequently. T ] and satisﬁes u(t) ∈ D(A0 ). T ] . p. We deduce from ([89]. T ] . ϕ2 ∈ CX and ϕ1 . we deduce that u is continuously diﬀerentiable on (r. [91]). T ] → X deﬁned by g(t) = F (t. T ] .4. ut ). . ϕ2 ) ≤ C0 (α) ϕ1 − ϕ2 . T ] for each ε > 0 and γ ∈ ]0. +∞) × CX → X is continuous and satisﬁes the following locally Lipschitz condition : for each α > 0 there exists a constant C0 (α) > 0 such that if t ≥ 0. for t.5. u(t) = e ϕ(0) + 0 Using (Theorem 3.487) that A0 is the generator of an analytic semigroup (not necessarily C0 −semigroup) deﬁned by 1 A0 t = eλt (λI − A0 )−1 dλ. which exists by virtu of Theorem 5. ut ). T ] u (t) = A0 u(t) + F (t. On the other hand. We prove now the local existence of integral solutions of Problem (1. Proof. T ] . [91]). s ∈ (r. u (t) = A0 u(t) + F (t. ut ). T ] → g(t) is locally H¨ older continuous. T ] of Equation (1). there exists L1 ≥ 0 such that ut − us ≤ L1 t − sγ . for t. 1[ . p. By virtue of (Theorem 4. Hence. we have g(t) − g(s) ≤ L(t − sα + ut − us ). T ] is continuously diﬀerentiable on (r. from ([77]. ϕ2 ≤ α then F (t. T ] and satisﬁes u(t) ∈ D(A0 ).Basic theory of abstract DFE with nondense domain 367 u of Equation (1) on [0.1) under a locally Lipschitz condition on F . u (t) ∈ D(A0 ) and for t ∈ (r. t > 0 e 2πi +C where +C is a suitable oriented path in the complex plan. and consider the function g : [0. ϕ1 . Let u be the integral solution on [0. u (t) ∈ D(A0 ) and for t ∈ (r. for t ∈ [0. ϕ1 ) − F (t. the function t ∈ (r.4 and 4. s ∈ [0.106) that t A0 t eA0 (t−s) g(s) ds.
for all s ∈ [−r. T1 ] . 0) ≤ αC0 (α) + sup F (s. T1 ] . we have t ∈ [0.T1 ] Let ϕ ∈ CX . ϕ) − u(s. X) . us )ds.. Consider the mapping H : Zϕ → C ([−r. u(t. s∈[0. s∈[0. ϕ(0) ∈ D(A0 ) and t ∈ [0. X) deﬁned by H (u) (t) = S0 (t)ϕ(0) ϕ(t). Tϕ > 0. ϕ) ≤ C0 (α) ϕ + F (t. deﬁned on [−r. Tψ [ and u(s. we have F (t. 0] and sup u(s) − ϕ(0) ≤ 1 . We will show that H (Zϕ ) ⊆ Zϕ . for suitable constants M0 and ω0 . and a unique integral solution u(. T1 ] . in the sense that if ϕ ∈ CX . ψ(0) ∈ D(A0 ) and ϕ − ψ < ε. for ψ ∈ CX . or t→Tϕ− Moreover. Let u ∈ Zϕ and t ∈ [0. then there exist positive constants L and ε such that. Tϕ [ and either Tϕ = +∞ lim sup u(t. ϕ) = +∞. ϕ) of Equation ( 1. X) : u (s) = ϕ (s) if s ∈ [−r. there exists a maximal interval of existence [−r. ϕ(0) ∈ D(A0 ). for each α > 0 there exists C0 (α) > 0 such that for ϕ ∈ CX and ϕ ≤ α.368 DELAY DIFFERENTIAL EQUATIONS Let ϕ ∈ CX such that ϕ(0) ∈ D(A0 ). s∈[0. 0] .T1 ] where C ([−r. Proof. for t ∈ [0. d + dt t S0 (t − s)F (s. T1 ] . X) is endowed with the uniform convergence topology. Tϕ [ . 0) .1). 0) . Tϕ [ . T1 ]. Let T1 > 0. ϕ) is a continuous function of ϕ.T1 ] Consider the following set Zϕ = {u ∈ C ([−r. 0 for t ∈ [−r. t] . T1 ] . α = ϕ+1 and c1 = αC0 (α)+ sup F (s. we have. Note that the locally Lipschitz condition on F implies that. Then. It’s clear that Zϕ is a closed set of C ([−r. ψ) ≤ L ϕ − ψ .
t .
.
d .
.
H (u) (t) − ϕ(0) ≤ S0 (t)ϕ(0) − ϕ(0) + .
S0 (t − s)F (s. us )ds.
.
. ≤ S0 (t)ϕ(0) − ϕ(0) + M0 eω0 t 0 . us ) ds. dt 0 t e−ω0 s F (s.
that ω0 > 0. t . without loss of generality. Then.369 Basic theory of abstract DFE with nondense domain We can assume.
.
F (s. H (u) (t) − ϕ(0) ≤ . us ) ds.
S0 (t)ϕ(0) − ϕ(0).
for s ∈ [0. 0) . us ≤ α. T1 ] and F (s. Consider T1 > 0 suﬃciently small such that . we obtain u(s) ≤ α. ≤ c1 . and α = ϕ + 1. Then. for s ∈ [−r. + M0 eω0 t 0 Since u(s) − ϕ(0) ≤ 1. T1 ] . for s ∈ [0. us ) ≤ C0 (α) us + F (s. T1 ] .
&.
' sup .
S0 (s)ϕ(0) − ϕ(0).
+ M0 eω0 s c1 s < 1. s∈[0. v ∈ Zϕ and t ∈ [0. for t ∈ [0. On the other hand. T1 ]. T1 ]. we deduce that H (u) (t) − ϕ(0) ≤ S0 (t)ϕ(0) − ϕ(0) + M0 eω0 t c1 t < 1.T1 ] So. let u. H (Zϕ ) ⊆ Zϕ . we have . Hence.
.
t .
.
d .
us ) − F (s. S0 (t − s)(F (s. vs ))ds.
.
H (u) (t) − H (v) (t) = . .
M0 eω0 T1 C0 (α)T1 ≤ M0 eω0 T&.X) . then 1c T . ≤ M0 eω0 t 0 t ω t 0 us − vs ds. ≤ M0 e C0 (α) 0 ≤ M0 eω0 T1 C0 (α) T1 u − vC([−r. dt 0 t F (s.T1 ]. vs ) ds. Note that α ≥ 1. us ) − F (s.
1 1 .
' ≤ sup .
S0 (s)ϕ(0) − ϕ(0).
.. Let u(. Tϕ > 0. ϕ) be the unique integral solution of Equation (1. deﬁned on it’s maximal interval of existence [0. + M0 eω0 s c1 s .1) has one and only one integral solution which is deﬁned on the interval [−r. s∈[0. H has one and only one ﬁxed point u in Zϕ .1). T1 ]. H is a strict contraction in Zϕ . Then. We conclude that Equation (1. Tϕ [ . So.T1 ] < 1.
0]. If t + θ ≥ 0. ϕ) < +∞. and θ ∈ [−r.370 DELAY DIFFERENTIAL EQUATIONS Assume that Tϕ < +∞ and lim sup u(t. ϕ) ≤ α. t + h ∈ [0. for t ∈ [−r. we obtain . Let t. Tϕ [ . Tϕ [. Then. there t→Tϕ− exists a constant α > 0 such that u(t. h > 0.
.
ϕ) − u(t + θ. ϕ) ≤ . u(t + θ + h.
(S0 (t + θ + h) − S0 (t + θ))ϕ(0).
.
t+θ+h .
t+θ .
d .
d + .
.
. us (. us (. ϕ))ds − S0 (t + θ − s)F (s. S0 (t + θ + h − s)F (s. ϕ))ds..
.
dt 0 dt 0 . .
t+θ+h .
2 2.
.
.
d .
≤ 2 S0 (t + θ)2 .
S0 (h)ϕ(0) − ϕ(0).
+ .
.
ϕ))ds.. S0 (s)F (t + θ + h − s. ut+θ+h−s (.
.
dt t+θ .
t+θ .
.
d .
+ .
.
. ϕ))) ds. ϕ) − F (t + θ − s. ut+θ−s (. S0 (s) (F (ut+θ+h−s .
.
−r≤σ≤0 Consequently. ϕ) ≤ β(h) exp C0 (α)M0 eω0 Tϕ Tϕ . ut+h (θ.. +M0 e 0 where δ(h) = sup u(σ + h. Tϕ [ . it follows ut+h (. ut+h (. ϕ) − ut (θ. ϕ) − ut (θ. . with .. −r≤σ≤0 By Gronwall’s Lemma.. h0 [ . for h ∈ ]0. ϕ) − us (. ϕ) − u(σ. ϕ) − us (. ϕ) ds. ϕ) . ϕ) − u(σ. for t.. h ∈ ]0. t + h ∈ [0. ϕ) − ut (.. ϕ) ds. ϕ) − ut (. ϕ) ≤ M0 eω0 Tϕ S0 (h)ϕ(0) − ϕ(0) + M0 eω0 Tϕ c1 h t us+h (. ϕ) . ϕ) ≤ sup u(σ + h. h0 [ ut+h (θ. +M0 eω0 Tϕ C0 (α) 0 If t + θ < 0.. ϕ) ≤ δ(h) + M0 eω0 Tϕ (S0 (h)ϕ(0) − ϕ(0) + c1 h) t ω T ϕ 0 C0 (α) us+h (. Let h0 > 0 suﬃciently small such that. dt 0 This implies that.
.
β(h) = δ(h) + M0 eω0 Tϕ .
S0 (h)ϕ(0) − ϕ(0).
Tϕ [ . u(. ϕ) exists. which contradicts the maximality of [0. one can show the same result for h < 0.. . that lim u(t. ϕ) can be extended to Tϕ . + c1 h . Using the same reasoning. t→Tϕ− Consequently. It follows immediately.
1).. we obtain for s ∈ [0. 0] . ϕ1 ) − Dϕ F (t..6) we deduce the dependence continuous with the initial data. ψ) ≤ α. ψ) ≤ M0 eω0 t ϕ − ψ + M0 eω0 t C0 (α) s uσ (. 1 := C 1 ([−r. ϕ1 ) − Dt F (t. Let ε ∈ ]0. ϕ1 ) − F (t.371 Basic theory of abstract DFE with nondense domain We will prove now that the solution depends continuously on the initial data.. (3.. Tϕ [. for given ϕ ∈ CX ϕ(0) ∈ D(A0 ). Thus. us (. ϕ) −r≤s≤t and c(t) = M0 eω0 t exp M0 eω0 t C0 (α)t .. ϕ) − us (. ϕ) − us (. for σ ∈ [0. We put α = 1 + sup u(s. ψ) ≤ α for σ ∈ [0. ϕ2 ≤ α then F (t. Dϕ F (t. let u(. ψ(0) ∈ D(A0 ) such that ϕ − ψ < ε. Tϕ [ → X be the unique integral solution of Equation ( 1. ϕ2 ) ≤ C1 (α) ϕ1 − ϕ2 . for all s ∈ [0. ϕ (0) ∈ D(A0 ) and ϕ (0) = Aϕ(0) + F (ϕ). for s ∈ [0. This completes the proof of Theorem Theorem 13 Assume that F is continuously diﬀerentiable and satisﬁes the following locally Lipschitz condition: for each α > 0 there exists a constant C1 (α) > 0 such that if ϕ1 . Proof. X) such that Then. ϕ) is continuously diﬀerentiable on [−r. Dt F (t. If we suppose that T0 < t.. T0 ≥ t and t < Tψ . 0 By Gronwall’s Lemma. Let T0 = sup {s > 0 : uσ (. s] . ϕ) : [−r.1). Tϕ [ and satisﬁes Equation ( 1.. T0 ] . ψ) ≤ c(t) ϕ − ψ . ϕ(0) ∈ D(A0 ) and t ∈ [0. Let ϕ ∈ CX . We have ψ ≤ ϕ + ε < α. s]} . us (. we deduce that us (. Then. ϕ2 ) ≤ C1 (α) ϕ1 − ϕ2 . ϕ2 ∈ CX and ϕ1 . The proof is similar to the proof of Theorem 10 . T0 ] . Furthermore. ψ) ≤ α.... ϕ) − uσ (. ψ) ≤ c(t)ε + α − 1 < α. ψ) dσ. then using the inequality (3.6) This implies that us (. u(. It follows that T0 cannot be the largest number s > 0 such that uσ (. t]. 1[ such that c(t)ε < 1 and ψ ∈ CX . ϕ2 ) ≤ C1 (α) ϕ1 − ϕ2 .
372 4. for all ϕ1 . ϕ2 ∈ E. Moreover. v) (T (t))t≥0 satisﬁes. that is. T (t)(ϕ1 ) − T (t)(ϕ2 ) ≤ M eγt ϕ1 − ϕ2 . Then.2) ϕ(t + θ). ii) T (t + s) = T (t)T (s). (T (t)(ϕ))(θ) = (4. (4. We recall the following deﬁnition. iii) for all ϕ ∈ E. y(t) = S (t)ϕ + dt 0 Let us consider the operator T (t) : E → E deﬁned. for all t. and vi) there exists γ > 0 and M ≥ 0 such that. DELAY DIFFERENTIAL EQUATIONS The semigroup and the integrated semigroup in the autonomous case Let E be deﬁned by CX E := D(A) 0 1 = ϕ ∈ CX : ϕ(0) ∈ D(A0 ) . T (t) is continuous from E into E. +∞) → X satisfying the following equation d t5 for t ≥ 0. iv) for all t ≥ 0. By virtue of Theorem 5. Proposition 50 Under the same assumptions as in Theorem 5. F : CX → X. The proof of this proposition is standard. there exists for each ϕ ∈ E a unique continuous function y(. for t ≥ 0. 0] and ϕ ∈ E. T (t)(ϕ) is a continuous function of t ≥ 0 with values in E. Proof. that is i) T (0) = Id. s ≥ 0. by T (t)(ϕ) = y(t. the translation property 3 (T (t + θ)(ϕ))(0). if t + θ ≥ 0. Deﬁnition 48 A C0 −semigroup (T (t))t≥0 on a Banach space (X. θ ∈ [−r.. the family (T (t))t≥0 is a nonlinear strongly continuous semigroup of continuous operators on E. ·) is called compact if all operators T (t) are compact on X for t > 0. if t + θ ≤ 0. ϕ) : [0.1) S(t − s)X0 F (y(s)) ds. . for t ≥ 0 and ϕ ∈ E. ϕ). we have the following result. Throughout this section we will suppose the hypothesis of Theorem 5 except that we require F to be autonomous.
Let H = {S (t)x. T + c] . s ∈ [c. ϕλ )(θ). we have t + θ > 0. On the other hand. deﬁne yλ by yλ (t)(θ) = y(t. λ ∈ Λ} is a precompact subset of (D(A). λ ∈ Λ} is still a precompact set of D(A). λ ∈ Λ} a set of continuous functions from the ﬁnite interval [0. ·). The proof is similar to the proof of a fundamental result of Travis and Webb [94]. Then. we have. λ ∈ Λ h c is contained in the closed convex hull of (t − c)H. . (A.CX ) for every t > r. Using the same reasoning as in the proof of Lemma 2. the nonlinear semigroup (T (t))t≥0 is compact on (E. one can prove that H is precompact in (D(A). x ≤ kN }. x ∈ D(A).5 [94]. ·) → (D(A0 ). the convex hull of H is precompact. 0]. D(A)) its generator. Proof. the set 3 4 1 S (s) S(h)Gλ (t − s). Letting h tend to zero. Let {ϕλ . for c < t ≤ T + c and λ ∈ Λ. ·). Hence the proof is complete Theorem 14 Assume that S0 (t) : (D(A0 ). Hence. where k is the Lipschitz constant of S(·) on [0. Proof. from Lemma 2 1 t−c S (t − s)S(h)Gλ (s) ds. . t] . ·) is compact for each t > 0. T ] and N is a bound of B. ·) → (D(A). ·) is compact for each t > 0 and the assumptions of Theorem 5 are satisﬁed. T + c]. T ] into B. T + c] ﬁxed. For each λ ∈ Λ. the set {Pλ (t).Basic theory of abstract DFE with nondense domain 373 Lemma 49 Let (S(t))t≥0 a locally Lipschitz continuous integrated semigroup on a Banach space (X. c > 0 a constant and d t−c S(t − s)Gλ (s) ds. for each h > 0 small enough and t ∈ (c. λ ∈ Λ h is contained in H. {Gλ . Pλ (t) = lim h0 h 0 and. {Pλ (t). Then. ·). λ ∈ Λ} be a bounded subset of E and let t > r. for each t ∈ (c. For θ ∈ [−r. B a bounded subset of X. t ∈ [c. the set 4 3 t 1 S (s)S(h)Gλ (t − s) ds. Then. Pλ (t) = dt 0 Assume that S (t) : (D(A).
we show that the family {yλ (t). say K. Observe that . s ∈ [0. one shows that {F (yλ (s)). we obtain t+θ 0 S0 (t + θ − s)F (yλ (s)) ds. Using the Lipschitz condition on F . t] and λ ∈ Λ} is bounded by a constant.374 DELAY DIFFERENTIAL EQUATIONS Then. 1 t+θ = S0 (t + θ)ϕλ (0) + lim S0 (t + θ − s)S0 (h)F (yλ (s)) ds. for each t > r. λ ∈ Λ} is equicontinuous. Let λ ∈ Λ. 0 < c < t − r and −r ≤ θ < θ ≤ 0. h0 h 0 yλ (t)(θ) = S0 (t + θ)ϕλ (0) + d dt First.
.
.
.
.
.
.
≤ .
.
S (t + θ)ϕλ (0) − S (t + θ)ϕ λ (0).
.
.
yλ (t)(θ) − yλ (t)(θ) 0 0 .
.
.
.
1 .
t+θ .
+ lim .
S0 (t + θ − s)S0 (h)F (yλ (s)) ds.
.
h 0 h .
t+θ .
.
.
t+θ .
1.
.
+ lim .
S0 (t + θ − s) − S0 (t + θ − s) S0 (h)F (yλ (s)) ds.
.
h 0 h .
t+θ−c .
.
.
t+θ−c .
1.
.
+ lim .
S0 (t + θ − s) − S0 (t + θ − s) S0 (h)F (yλ (s)) ds.
. .
h 0 h .
0 2 2 .
.
2 .
.
2 ≤ 2S0 (t + θ) − S0 (t + θ) 2 ϕλ (0) + M0 eω0 t kK .
θ − θ.
] s∈[0.t+θ−c . + c2M0 eω0 t kK 2 2 2 2 + t sup 2S0 (t + θ − s) − S0 (t + θ − s)2 kK.
.
.
.
If we take .
θ − θ.
t] → L(D(A0 )) for 0 < a < t. 4 3 t+θ d S0 (t + θ − s)F (yλ (s)) ds. small enough.) : [a. λ ∈ Λ dt 0 is precompact. for each t > r and θ ∈ [−r. 0]. the claimed equicontinuity of {yλ (t). 0]. λ ∈ Λ} is precompact. On the other hand. Observe that for 0 < c < t + θ and λ ∈ Λ . the family {S0 (t + θ)ϕλ (0). for each t > r and θ ∈ [−r. it follows from the uniform continuity of S0 (. We will show that. λ ∈ Λ}.
.
.
1 .
.
t+θ .
S (t + θ − s)S (h)F (y (s)) ds 0 λ 0 . ≤ cM0 eω0 t kK.
h .
t+θ−c Then. if h tends to zero we obtain .
.
t+θ .
.
d .
S0 (t + θ − s)F (yλ (s)) ds.
.
≤ cM0 eω0 t kK. .
3) . dt t+θ−c (4.
λ ∈ Λ dt 0 is precompact in X. λ ∈ Λ . This fact together with (4. X) . 0] . t ≥ 0. for t ≥ 0. ϕ). ϕ (0) = A0 ϕ(0) + L (ϕ)} ..5) z(t) = S(t)ϕ + S(t − s)X0 L(z(s)) ds . if 0 < c < t + θ and λ ∈ Λ. Proof. t+θ−c S0 (t + θ − s)F (yλ (s)) ds = 0 t+θ 375 S0 (s)F (yλ (t + θ − s)) ds. Pϕ = ϕ . 0]. ϕ(0) ∈ D(A0 ). (4. dt 0 Theorem 15 The family of operators (U (t))t≥0 deﬁned on CX by U (t)ϕ = z(t. then 4 3 t+θ−c d S0 (t + θ − s)F (yλ (s)) ds. It is clear that there exists a unique function z := z(. ϕ) : [0. for each t > r and θ ∈ [−r.Basic theory of abstract DFE with nondense domain For 0 < c < t + θ and λ ∈ Λ. dt 0 Using the ArzelaAscoli theorem.4) u0 = ϕ. L . we obtain the result of Theorem 14 Consider now the linear autonomous functional diﬀerential equation 3 u (t) = A0 u(t) + L(ut ). Given ϕ ∈ CX .3) yields the precompactness of the set 4 3 t+θ d S0 (t + θ − s)F (yλ (s)) ds. is a locally Lipschitz continuous integrated semigroup on CX generated by the operator P deﬁned by & D(P) = ϕ ∈ C 1 ([−r. Consider the operator 5 : CX → C5X L deﬁned by 5 (ϕ) = X0 L(ϕ). +∞) → CX which solves the following abstract integral equation t d 5 (4. c By Lemma 49. where L is a continuous linear functional from CX into X.
Then. d z(t) = S(t)ϕ + dt t 5 S(t − s)X0 L (V (s)ϕ) ds . 5 ∈ L(CX . 0 On the other hand. G (ϕ) = G It is easy to see that G = P. the nonhomogeneous Cauchy problem dz 5 (t) = Az(t) + h(t). 0 . 5 D(G) 5 = A 5 + L. is the generator of a locally Lipwhere A 5 satisﬁes (HY) 5X . which is deﬁned by : Let us introduce the part G of G 0 1 5 Gϕ 5 ∈ CX . if we consider. On the other hand. because A schitz continuous integrated semigroup on C 5X ). D(G) = ϕ ∈ D(G). z(t) = dt 0 t t d d 5 5 S(t − s)ϕ ds + S(t − s)X0 L (V (s)ϕ) ds . h(t) = ϕ + L(V By Theorem 4. Proposition 47 gives t V (t)ϕ = P V (s)ϕ ds + tϕ.6) has a unique integral solution z given by t d 5 S(t − s)h(s) ds . (4. 5 (ϕ) . the nonhomogeneous Cauchy problem (4. where h : [0. one can prove that the operator G 5 deﬁned in CX by 3 5 = D(A).376 DELAY DIFFERENTIAL EQUATIONS 5 Using a result of Kellermann [80]. = dt dt 0 0 Then.6) dt z(0) = 0. C and L 5 in CX . 5 G 5 is deﬁned in Proposition 49. for each ϕ ∈ CX . +∞[ → C5X is given by 5 (t)ϕ). P is the generator of a locally Lipschitz continuous integrated semigroup (V (t))t≥0 on CX . for t ≥ 0.
we obtain 5 V (t)ϕ = A t t V (s)ϕ ds + X0 L V (s)ϕ ds + tϕ. Then we have U (t) = V (t) on CX .6). the function t → V (t)ϕ is an integral solution of Equation (4. for ϕ 5 ∈ C5X . 5 V (t)ϕ = A 0 t t V (s)ϕ ds + h(s) ds. ϕ (0) ∈ D(A0 ) ϕ (0) = A0 ϕ(0) + L (ϕ)} . ϕ) 5 which is a continuous mapping from 5 (t) deﬁned on [0. for every t > r. ϕ) 5 . for all t ≥ 0. See [93] Corollary 17 Under the same assumptions as in Theorem 14. 0 D(B) = ϕ ∈ C 1 ([−r. Corollary 16 B is the inﬁnitesimal generator of the C0 −semigroup (U (t))t≥0 on E. the linear C0 −semigroup (U (t))t≥0 is compact on E. the integral equation t d 5 − s)X0 L(z(s)) ds . 0] . ϕ(0) ∈ D(A0 ). 0 0 Hence. Moreover. we have 5 + X0 L (ψ) .7) has a unique solution z := z(. for t ≥ 0. the family of operators U C5X by t≥0 5 (t)ϕ U 5 = z(t. +∞) → CX . Pψ = Aψ Then. the abstract integral equation ( 4. X) .377 Basic theory of abstract DFE with nondense domain Moreover. By uniqueness. 0 So. Bϕ = ϕ . Then. Proposition 51 Given ϕ 5 ∈ C5X . Thus the proof of Theorem 15 Let B be the part of the operator P on E.7) dt 0 One can show the following result.. To deﬁne a fundamental integral solution Z(t) associated to Equation (4. for ψ ∈ D(P). we conclude that V (t)ϕ = z(t).4). Proof. 5 ϕ S(t z(t) = S(t) 5+ (4. we consider.
we deﬁne the linear operator ∆(λ) : D(A0 ) → X by ∆(λ)x := λx − A0 x − L eλ· x . We have. 0] → CX . Let ω > max(0.4) is deﬁned as dim Ker∆(λ). We will call λ a characteristic value of Equation (4. This yields . For each complex number λ. for λ > ω0 ∆(λ) = (λI − A0 ) I − (λI − A0 )−1 L(eλ· I) . Consider the equation Kλ ≤ 5 λ· a) = X0 c. (λI − G)(e where c ∈ X is given and we are looking for a ∈ D(A0 ). The multiplicity of a characteristic value λ of Equation (4. λ − ω0 Hence the operator ∆(λ) is invertible for λ > ω. Corollary 18 There exits ω ∈ IR.4) if there exists x ∈ D(A0 )\{0} solving the characteristic equation ∆(λ)x = 0. x ∈ D(A0 ). for λ > ω. 5 D(G) 5 5 Gϕ = A + X0 L(ϕ). We have the following result. where eλ· x : [−r. θ ∈ [−r. is deﬁned for x ∈ X by (note that we consider here the comlexiﬁcation of CX ) eλ· x (θ) = eλθ x. such that for λ > ω and c ∈ X. ω0 + M0 L) and Kλ = (λI − A0 )−1 L(eλ· I). 0] . M0 L < 1. (λI − G) Proof. Then.378 DELAY DIFFERENTIAL EQUATIONS is a locally Lipschitz continuous integrated semigroup on C5X generated 5 deﬁned by by the operator G 3 5 = D(A). one has 5 −1 (X0 c) = eλ· ∆(λ)−1 c.
for Corollary 19 For each c ∈ X. Then λθ −1 e ∆(λ) c=λ +∞ −λs e 5 U (s)(X0 c) (θ) ds = λ +∞ −θ 0 So. by 5 (t)(X0 c)(0) Z(t)c = U Corollary 20 Z(t) is the fundamental integral solution of Equation ( 4. that is +∞ −1 e−λt Z(t) dt.379 Basic theory of abstract DFE with nondense domain λeλ· c − λeλ· c + X0 λc − A0 c − L eλ· c = X0 a. t + θ ≥ 0. for c ∈ X −1 5 λI − G (X0 c) = eλ· ∆(λ)−1 c. Corollary 21 If c ∈ D(A0 ) then the function t → Z(t)c e−λs Z(s+θ)c ds. for λ > ω.)(X0 c) satisﬁes. −1 ∆(λ) +∞ c=λ e−λt Z(t)c dt 0 It is easy to prove the following result.4). 5 λI − G −1 (X0 c) = eλ· ∆(λ)−1 c 5 (. for λ > ω c = ∆(λ)−1 a. Consequently. We have. the following translation property 5 (t + θ)(X0 c) (0). U 5 (t)(X0 c) (θ) = U 0. Then. for t ≥ 0 and c ∈ X. t + θ ≤ 0. 0]. the function U t ≥ 0 and θ ∈ [−r. ∆(λ) = λ 0 Proof. . We can consider the following linear operator Z(t) : X → X deﬁned.
)(X0 ). we have 0 5 = D(P) = ϕ ∈ CX . Moreover. The fundamental solution Z (t) is deﬁned only for c ∈ D(A0 ) and is discontinuous at zero.9) where f : [0. y (t) = Gy(t) y(0) = ϕ. ∆(λ) c = 0 Hence the name of fundamental integral solution. or its integrated form t d 5 S(t − s)X0 L [(y(s)) + f (s)] ds . solution Z(t) or equivalently U We construct now a variationofconstants formula for the linear nonhomogeneous system u (t) = A0 u(t) + L(ut ) + f (t). 5 = D(G) 5 and G 5 =A 5 + X0 L. Theorem 22 For any ϕ ∈ CX . D(A) integral solution of the equation 3 5 + X0 f (t). (4. t ≥ 0. t ≥ 0. such that ϕ(0) ∈ D(A0 ). Then. D(G) 1 ϕ(0) ∈ D(A0 ) = E. +∞) → X deﬁned by d t5 (4. . For these reasons we use the fundamental integral 5 (. y is the On the other hand.380 DELAY DIFFERENTIAL EQUATIONS is diﬀerentiable for all t > 0 and we have +∞ −1 e−λt Z (t)c dt. y (t) = Ay(t) y(0) = ϕ.8) for t ≥ 0. t ≥ 0. It follows immediately from Theorem 5 and Corollary 6 that.9) has a unique solution y which is the integral solution of the equation 3 5 + X0 [L(y(t)) + f (t)] . y(t) = U (t)ϕ + dt 0 satisﬁes Equation ( 4.10) U (t − s)X0 f (s) ds.9). y(t) = S (t)ϕ + dt 0 (4. +∞) → X is a continuous function. Equation (4. the function y : [0. for ϕ ∈ CX and ϕ(0) ∈ D(A0 ). Proof.
. E = ϕ ∈ CX . F (0) = 0 and F is satisﬁes the Lipschitz condition F (ϕ1 ) − F (ϕ2 ) ≤ L ϕ1 − ϕ2 . We make the following hypothesis : F is continuously diﬀerentiable. t ≥ 0.1) in the autonomous case. t ≥ 0. for each ε > 0. t ≥ 0. Principle of linearized stability In this section. and let (U (t))t≥0 be the corresponding linear C0 −semigroup on E. ϕ) is the unique integral solution of Equation (5. for ϕ ≤ δ. is the linear semigroup U (t). (5. (5.1) and 1 0 ϕ(0) ∈ D(A0 ) . Proof. for all ϕ1 . t ≥ 0. ϕ2 ∈ CX .2).)ϕ is diﬀerenand the part of G tiable in t ≥ 0 and Theorem 4 implies d t5 U (t − s)X0 f (s) ds y(t) = U (t)ϕ + dt 0 5. Consider the linearized equation of (5. .1) corresponding to the Fr´echetderivative Dϕ F (0) = F (0) at 0 : du (t) = A0 u(t) + F (0)ut . Then t → U (. where L is a positive constant. ϕ). We know that the family (T (t))t≥0 is a nonlinear strongly continuous semigroups of continuous operators on E. It suﬃces to show that. be deﬁned by T (t)(ϕ) = ut (. Let T (t) : E → E..Basic theory of abstract DFE with nondense domain 381 5 in CX is the operator P.1). for t ≥ 0. Proposition 52 The Fr´echetderivative at zero of the nonlinear semigroup T (t). associated to Equation ( 5. we give a result of linearized stability near an equilibrium point of Equation (1.2) dt u0 = ϕ ∈ CX . where u(.1) dt u0 = ϕ. associated to Equation ( 5. that is du (t) = A0 u(t) + F (ut ). there exists δ > 0 such that 2 2 2T (t)(ϕ) − U (t)ϕ2 ≤ ε ϕ .
382 DELAY DIFFERENTIAL EQUATIONS We have .
.
T (t)(ϕ) − U (t)ϕ = sup .
(T (t)(ϕ)) (θ) − U (t)ϕ (θ).
θ∈[−r. .0].
.
= sup .
(T (t)(ϕ)) (θ) − U (t)ϕ (θ).
0 It follows that 2 2 2T (t)(ϕ) − U (t)ϕ2 ≤ M0 eω0 t t . . for t + θ ≥ 0 d (T (t)(ϕ)) (θ) − U (t)ϕ (θ) = dt t+θ S0 (t + θ − s) F (T (s)(ϕ)) − F (0)(U (s)ϕ) ds.0] and. t+θ≥0 θ∈[−r.
.
e−ω0 s .
F (T (s)(ϕ)) − F (0)(U (s)ϕ).
ds 0 and T (t)(ϕ) − U (t)ϕ t .
.
≤ M0 e−ω0 s .
F (T (s)(ϕ)) − F (U (s)ϕ)).
ds 0 t .
.
−ω0 s .
.
we deduce that for ε > 0. eω0 t 0 By virtue of the continuous diﬀerentiability of F at 0. there exists δ > 0 such that t . + e F (U (s)ϕ) − F (0)(U (s)ϕ) ds .
.
e−ω0 s .
F (U (s)ϕ) − F (0)(U (s)ϕ).
0 On the other hand. we obtain t . ds ≤ ε ϕ . for ϕ ≤ δ.
.
e−ω0 s .
F (T (s)(ϕ)) − F (U (s)ϕ).
ds ≤ L 0 t 2 2 e−ω0 s 2T (s)(ϕ) − U (s)ϕ2 ds. 0 By Gronwall’s Lemma. for each t ≥ 0 Deﬁnition 49 Let Y be a Banach space. 0 Consequently. and x0 ∈ W . 2 2 2T (t)(ϕ) − U (t)ϕ2 ≤ M0 eω0 t ε ϕ + L t 2 2 e−ω0 s 2T (s)(ϕ) − U (s)ϕ2 ds . t ≥ 0. we obtain 2 2 2T (t)(ϕ) − U (t)ϕ2 ≤ M0 ε ϕ e(LM0 +ω0 )t . We conclude that T (t) is diﬀerentiable at 0 and Dϕ T (t)(0) = U (t). (V (t))t≥0 a strongly continuous semigroup of operators V (t) : W ⊆ Y → W.
. unstable and center. for all x1 . Spectral Decomposition In this part. the spectrum σ(U (t)) is a countable set and is compact with the only possible accumulation point 0 and if µ = 0 ∈ σ(U (t)) then µ ∈ P σ(U (t)). then x0 is an exponentially asymptotically stable equilibrium of (V (t))t≥0 . the existence of a direct sum decomposition of the state space 1 0 ϕ(0) ∈ D(A0 ) E = ϕ ∈ CX .e. Theorem 23 Suppose that the zero equilibrium of (U (t))t≥0 is exponentially asymptotically stable. If the zero equilibrium of (Y (t))t≥0 is exponentially asymptotically stable.Basic theory of abstract DFE with nondense domain 383 an equilibrium of (V (t))t≥0 (i. the following results. We assume that the semigroup (T0 (t))t≥0 on D(A0 ) is compact. we show in the linear autonomous case (Equation (4. It follows from the compactness property of the semigroup U (t). Theorem 24 (Desh and Schappacher [65]) Let (V (t))t≥0 be a nonlinear strongly continuous semigroup of type γ on a subset W of a Banach space Y . Corollary 25 [89] For each t > r. k ≥ 1 such that V (t)x − x0 ≤ ke−µt x − x0 for all x ∈ W with x − x0 ≤ δ and all t ≥ 0. Then. (Y (t))t≥0 is a strongly continuous semigroup of bounded linear operators on Y. The equilibrium x0 is called exponentially asymptotically stable if there exist δ > 0. 6. . for t > r. on E. V (t)x0 = x0 . t ≥ 0.e.4)). into three subspaces : stable. x2 ∈ W and assume that x0 ∈ W is an equilibrium of (V (t))t≥0 such that V (t) is Fr´echetdiﬀerentiable at x0 for each t ≥ 0. then zero is exponentially asymptotically stable equilibrium of (T (t))t≥0 . V (t)x1 − V (t)x2 ≤ M eγt x1 − x2 . (where P σ(U (t)) denotes the point spectrum). i. for all t ≥ 0). which are semigroup invariants. µ > 0. We have the following result. with Y (t) the Fr´echetderivative at x0 of V (t). The proof of this theorem is based on the following result.
On the other hand. then for all ϕ ∈ E. ϕ ∈ E. sup e ω1 := inf ω ∈ lR : t≥0 The compactness property of the semigroup (see [84]) implies that ω1 = s1 (B) := sup {eλ : λ ∈ P σ(B)} .4). Then. Proof. Proof. then there exists x = 0 and a complex λ such . there exists n such that eλn = s1 (B). if β is a given real number then there exists only a ﬁnite number of λ ∈ P σ(B) such that eλ > β. We can give now an exponential estimate for the semigroup solution. Moreover. Then.384 DELAY DIFFERENTIAL EQUATIONS Corollary 26 [89] There exists a real number δ such that eλ ≤ δ for all λ ∈ σ(B). then eλ ≤ δ. This implies that ϕ(θ) = eλt ϕ(0) and ϕ (0) = A0 ϕ(0) − L(ϕ) with ϕ(0) = 0. then we have ω1 = s1 (B) < 0 and the stability holds. for γ > 0 there exists a constant k(γ) ≥ 1 such that 2 2 2U (t)ϕ2 ≤ k(γ)e(δ+γ)t ϕ . We deduce that λ is a characteristic value of Equation (4. the sequence (eλn )n is stationary. we deduce that {λn : eλn > β} is ﬁnite. there exists β such that eλn > β for n ≥ n0 with n0 large enough.4). So. Theorem 27 Let δ be the smallest real number such that if λ is a characteristic value of Equation ( 4. We will prove now the existence of λ ∈ P σ(B) such that eλ = s1 (B). If δ = 0. Proposition 53 Assume that δ is a real number such that eλ ≤ δ for all characteristic values of Equation ( 4. Let ω1 be deﬁned by 3 4 −ωt 2 2 2U (t)2 < +∞ . then there exists ϕ ∈ E such that U (t)ϕ → +∞ as t → +∞. U (t)ϕ → 0 as t → +∞. From Corollary 26. Assume that δ < 0. for all t ≥ 0. If δ > 0. Consequently. This completes the proof of Proposition 53 The asymptotic behavior of solutions can be now completely obtained by the characteristic equation.4). Then. Let (λn )n be a sequence in P σ(B) such that eλn → s1 (B) as n → +∞. if λ ∈ P σ(B) then there exists ϕ = 0 ∈ D(B) such that Bϕ = λϕ. If δ < 0. ∆(λ)ϕ(0) = 0. If δ = 0 then there exists ϕ ∈ E\{0} such that U (t)ϕ = ϕ for all t ≥ 0.
385 Basic theory of abstract DFE with nondense domain that eλ = 0 and ∆(λ)x = 0. there exist three linear subspaces of E denoted by S.1]. S. γ U (t)PCN ϕ ≤ M e 3 t PCN ϕ . For the convenience of the reader. λ ∈ P σ(B) and eλt ∈ P σ(U (t)). there exists ϕ = 0 such that U (t)ϕ = eδt ϕ → +∞. Then. 2 2 This implies that U (t)ϕ = 2eλt ϕ2 = ϕ . and its integrated form ut = U (t)ϕ + d dt t 0 5 (t − s)X0 f (s)ds. 7. there exists ϕ = 0 such that U (t)ϕ = eλt ϕ. Assume now that δ > 0. U (t)PS ϕ ≤ M e−γt PS ϕ . This completes the proof of Theorem 27 Using the same argument as in [96]. (ii) U S and CN are ﬁnite dimensional. Then. (7. B(U S) ⊂ U S and B(CN ) ⊂ CN .3. for t ∈ lR. U S and CN . [section 3. U for t ≥ 0. there exists M = M (γ) > 0 such that U (t)PU S ϕ ≤ M eγt PU S ϕ . = {λ ∈ P σ(B) : eλ > 0} . t ≥ 0. Consequently. (v) for any 0 < γ < α := inf {eλ : λ ∈ P σ(B) and eλ = 0} . U S and CN are called stale.8) mentioned in the introduction. U (t) (S) ⊂ S. Theorem 3. U (t) (CN ) ⊂ CN. Then. t ∈ lR.1) . P σ(B CN ) = (iii) P σ(B U S ) {λ ∈ P σ(B) : eλ = 0} . for t ≥ 0. PU S and PCN are projections of E into S. Existence of bounded solutions We reconsider now the equation (4. we obtain the following result. as t → +∞. U S and CN respectively. respectively. Theorem 28 Suppose that X is complex. t ≤ 0. dt u0 = ϕ ∈ CX . Then. we restate this equation du (t) = A0 u(t) + L(ut ) + f (t). there exists x = 0 and a complex λ such that eλ = δ and ∆(λ)x = 0. where PS . t ≥ 0. such that E = S ⊕ U S ⊕ CN and (i) B(S) ⊂ S. unstable and center subspaces of the semigroup (U (t))t≥0 . (iv) U (t) (U S) ⊂ U S.
U (t − s)B for t ≥ 0. (7. ϕ ∈ S. in the hyperbolic case the center subspace CN is reduced to zero. Theorem 28 implies that. We give now the ﬁrst main result of this section. we have & π(B) = x : lR− → X. U (t)ϕ ≤ M eγt ϕ . We need the following deﬁnition. Thanks to Lemma 2. t ≤ 0. for any f ∈ B. ϕ ∈ U S.386 DELAY DIFFERENTIAL EQUATIONS where f is a continuous function form lR into X.2) ut = U (t)ϕ + lim λ→+∞ 0 8λ : X0 → CX is deﬁned by where the operator B −1 8λ X0 c = λ λI − G 5 B (X0 c) = λeλ· ∆(λ)−1 c. B(lR+ ) or B(lR). (i) for B = B(lR− ). lim lim λ→+∞ 0 . then the space E is decomposed as E = S ⊕ US and there exist positive constants M and γ such that U (t)ϕ ≤ M e−γt ϕ . t ≥ 0. Let B represent B(lR− ). we have the following result. Moreover. t ≤ 0} and (Kf )t = S t 8λ X0 f (τ ) dτ + lim U (t − τ ) B s→−∞λ→+∞ s t U S 8λ X0 f (τ ) U (t − τ ) B dτ. Corollary 29 If the semigroup (U (t))t≥0 is hyperbolic. Thus. there is a unique solution Kf ∈ B of Equation ( 7. the set of bounded continuous functions from lR− . Deﬁnition 50 We say that the semigroup (U (t))t≥0 is hyperbolic if σ(B) ∩ ilR = ∅. which is equivalent to (7.1) : t 8λ X0 f (s)ds.8). Theorem 30 Assume that the semigroup (U (t))t≥0 is hyperbolic. we will use the following integrated form of Equation (4. Then.2) (for some ϕ ∈ E) such that πKf = 0 and K : B → B is a continuous linear operator.4) (for any ϕ ∈ E) which are in B. there exists ϕ ∈ U S such that x(t) = (U (t)ϕ) (0). Let π : B → B be a projection onto the integral solutions of Equation ( 4. lR+ or lR respectively to X. c ∈ X.
4) in S. we have & π(B) = x : lR+ → X. let ϕ ∈ S and u(. x : lR → X. the function u can be decomposed as S + uSt . Theorem 28 implies that the space U S is ﬁnite dimensional and ' & U −(t) (U S) ⊆ U S. Since us (. t ≤ 0 ⊆ π(B(lR− )).387 Basic theory of abstract DFE with nondense domain (ii) For B = B(lR+ ).2) in B(lR− ). Thanks to Corollary 29 we have ut (. there exists ϕ ∈ S such that x(t) = (U (t)ϕ) (0). s ≤ t. Assume that there is a t ∈ (−∞.. ϕ) be the integral solution of Equation (4. we have π(B) = {0} and (Kf )t = S 8λ X0 f (τ ) dτ + U (t − τ ) B s→−∞λ→+∞ s t U S 8λ X0 f (τ ) U (t − τ ) B dτ. we have ut (. f ) be a solution of Equation (7. ut = uU t . Therefore. ϕ. ϕ) = 0. there exists ϕ ∈ U S such that x(t) = (U (t)ϕ) (0). ϕ) = 0. 0] such that ut (. and (Kf )t = lim λ→+∞ lim t ≥ 0} S 8λ X0 f (τ ) dτ + U (t − τ ) B 0 t U S 8λ X0 f (τ ) lim U (t − τ ) B dτ. which is bounded on lR− . Let f ∈ B(lR− ) and u = u(.. t ≤ 0} . Then. Conversely. In the same manner. Then.. ϕ) is bounded. there exists ϕ ∈ U S such that x(t) = (U (t)ϕ) (0). ϕ). one can prove the same relations for B(lR+ ) and B(lR)... with initial value ϕ ∈ E... lim lim lim t lim s→+∞λ→+∞ s Proof. ϕ) . Then. ϕ) = U (t − s)us (. & π(B(lR− )) ⊆ x : lR− → X... ϕ) ≤ M e−γ(t−s) us (. we deduce that us (. t s→+∞λ→+∞ s (iii) For B = B(lR). for any s ∈ (−∞. t).
ϕU S . 0) = 0. we get B 2 S 2 2 2 8λ X0 f (τ ) 2 ≤ M e−γ(t−τ ) M λ 2U (t − τ ) B sup f (τ ) ... ϕU S .2) in B(lR− ) satisfying the estimates (7.3) for s ≤ t ≤ 0.6) We have proved that t U S 8λ X0 f (τ ) U (t − τ ) B dτ + ut = U (t)ϕU S + lim λ→+∞ 0 t S 8λ X0 f (τ ) dτ. (7. for t ≤ 0.0] Conversely. for t ≤ 0. + γ (7.388 DELAY DIFFERENTIAL EQUATIONS S ∈ U S and uS ∈ S are given by where uU t t t U S US US 8λ X0 f (τ ) U (t−τ ) B dτ.7) is a solution of Equation (7. (7. f ). lim lim s→−∞λ→+∞ s (7.7) and let K : B(lR− ) → B(lR− ) be deﬁned by Kf = (I − π) u(.. we deduce that t S S 8λ X0 f (τ ) dτ.8) τ ∈(−∞.. f ) be deﬁned by (7.4) since U (t) is deﬁned on U S for all t ∈ lR.8) for every ϕ ∈ E. f ) = u(.. Let u = u(.0] t ≤ 0. u(.5) U (t − τ ) B ut = lim lim s→−∞λ→+∞ s 8λ X0 = λeλ· ∆−1 (λ). (7.6) and (7. We can easily verify that ut (. 0) + u(. U (t − τ ) B for t. By Corollary 29. the following estimate 2 t U S 2 2 2 8 2≤ 2U (t)ϕU S + lim U (t − τ ) B X f (τ ) dτ 0 λ 2 2 λ→+∞ 0 2 US2 MM γt 2 2 Me ϕ sup f (τ ) . ϕU S . s ∈ lR. we can verify that the expression (7. By Lemma 1.. Thus. 0.0] Consequently. ut = U (t−s)us + lim λ→+∞ s uSt = U (t − s)uSs + lim λ→+∞ s t S 8λ X0 f (τ ) dτ. (I − π) u(. 2 2 λ − ω1 τ ∈(−∞. γ τ ∈(−∞. 0. U (t − τ ) B for t ≤ 0.7) We obtain also. (7. . 0) = U (t)ϕU S . ϕU S . we get 2 S2 MM 2ut 2 ≤ sup f (τ ) . ϕU S . f )..
u(n) + lim U (t − τ ) B ) dτ + τ λ→+∞ 0 t U S 8λ X0 g(τ. +∞) → X with ut ≤ h for t ≥ 0 and uS0 = ϕ. ϕ2 ∈ E. By Theorem 30. Proposition 54 Assume that the semigroup (U (t))t≥0 is hyperbolic and the assumption (H1) holds. if ϕ ≤ ε then 2ut 2 < h for t ≥ 0. h[ It is clear that u0 2 2 2 (n) 2 small enough such that. ϕ2 ≤ h and t ∈ lR. uτ ) U (t − τ ) B dτ. Equation ( 7.2) for every f ∈ B(lR− ). 0) = 0 for t ∈ lR and there exists a nondecreasing function α : [0. deﬁned by = U (t)ϕ (n+1) ut U (t)ϕ (0) ut = lim S t 8λ X0 g(τ. Let ϕ ∈ S. dt and its integrated form t 8λ X0 g(s.10) where g is continuous from lR × CX into X. Kf satisﬁes Equation (7.9) (7. We will assume that (H1) g(t. Then. ϕ1 ) − g(t.389 Basic theory of abstract DFE with nondense domain Therefore. This completes the proof of Theorem 30 We consider now the nonlinear equation du (t) = A0 u(t) + L(ut ) + g(t. we can choose h > 0 and ε ∈ ]0. K is a continuous linear operator on B(lR− ). +∞) → X of the following equation t S 8λ X0 g(τ. h[ such that for any ϕ ∈ S with ϕ ≤ ε. it suﬃces to establish the existence of a bounded solution u : [−r. lim lim s→+∞λ→+∞ s Let (u(n) )n∈N be a sequence of continuous functions from [−r. For B(lR+ ) and B(lR) the proofs are similar. Moreover. +∞) with lim α(h) = 0 and h→0 g(t. Proof. ut ). πK = 0 and K is explicitly given by Theorem 30. ϕ2 ) ≤ α(h) ϕ1 − ϕ2 for ϕ1 . . ϕ1 . there exists h > 0 and ε ∈ ]0. τ ) s→+∞λ→+∞ s (n) S = ϕ. +∞) → [0. uτ ) dτ + U (t − τ ) B ut = U (t)ϕ + lim λ→+∞ 0 t U S 8λ X0 g(τ. us )ds.10) has a unique bounded solution u : [−r. U (t − s)B ut = U (t)ϕ + lim λ→+∞ 0 (7. u(n) lim U (t − τ ) B dτ. +∞) to X.
v(t) ≤ 2h 1 + γ γ k=n+1 Consequently. uτ ) 8λ X0 g(τ. +∞) n→+∞ and u is continuous on [−r. +∞). v = 0 on [0. u(n) lim +2 U (t − τ ) B − B dτ 2 τ ) 2 λ→+∞ 2 0 2 t U S U S 2 2 2 (n) 8λ X0 g(τ. In order to prove that u is a solution of Equation (7. we introduce the following notation 2 t S 2 8 2 U (t − τ ) Bλ X0 g(τ. the limit u := lim u(n) exists uniformly on [−r. +∞). uτ ) U (t − τ ) B dτ 2 − lim lim 2. we have 2 2 2 2 t 2 (n) 2 (n+1) (n) 2 (n−1) 2 − ut 2 ≤ 0 M M e−γ(t−τ ) α(h) 2uτ − uτ 2 dτ + 2ut 2 2 +∞ 2 2 (n) (n−1) M M eγ(t−τ ) α(h) 2uτ − uτ 2 dτ. t So. 2ut γ t ≥ 0. Moreover. we have ut − (n+1) ut +∞ = (k+1) ut (k) − ut . s→+∞λ→+∞ s We obtain v(t) ≤ 2 2 2 (n+1) 2 2ut − ut 2 2 t S S 2 2 2 8λ X0 g(τ. γ 2 Consequently. uτ ) lim +2 lim U (t − τ ) B − B dτ 2 2 s→+∞ 2. by induction we get n 2 2 2M M α(h) 2 (n+1) (n) 2 − ut 2 ≤ 2h . We choose h > 0 such that 1 2M M α(h) < . k=n+1 It follows that +∞ k 2M M α(h) 2M M α(h) .390 DELAY DIFFERENTIAL EQUATIONS On the other hand.10). λ→+∞ s Moreover. ut < h for t ≥ 0 and uS0 = ϕ. uτ ) dτ v(t) = 2ut − U (t)ϕ − lim 0 λ→+∞ U S 2 t 2 8λ X0 g(τ. . uτ ) 8λ X0 g(τ.
Deﬁnition 51 A function g ∈ B(lR) is said to be almost periodic if and only if the set {gσ : σ ∈ lR} . As a consequence of the existence of a bounded solution we obtain the following result. .10). This proves the uniqueness and completes the proof 8. then the bounded solution of Equation ( 7. Then. 2wt − ut 2 ≤ 2h γ as n → +∞. Let u be the unique bounded solution of Equation (7. where gσ is deﬁned by gσ (t) = g(t + σ).391 Basic theory of abstract DFE with nondense domain To show the uniqueness suppose that w is also a solution of Equation (7. then the bounded solution of Equation ( 7. We ﬁrst recall a deﬁnition. s→+∞λ→+∞ s This implies n 2 2 2M M α(h) 2 (n+1) 2 → 0. uτ ) dτ 2 2. we are concerned with the existence of periodic (or almost periodic) solutions of Equations (7.2) is also almost periodic. Theorem 32 Assume that the semigroup (U (t))t≥0 is hyperbolic. If the function f is ω−periodic. This completes the proof of the corollary We are concerned now with the existence of almost periodic solution of Equation (7. is relatively compact in B(lR).2). for t ∈ lR.2) and (7. wτ ) − B 8λ X0 g(τ. + ω). The uniqueness property implies that u = u(. + ω) is also a bounded solution of Equation (7.2).2). wτ ) + 2 lim lim U (t − τ ) − Bλ X0 g(τ.2) given by Theorem 30 is also ω−periodic. u(n) 2 lim B U (t − τ ) ) dτ 2 τ 2λ→+∞ 2 0 2 t U S U S 2 2 2 (n) 8 8 2 Bλ X0 g(τ. If the function f is almost periodic. Existence of periodic or almost periodic solutions In this section. The function u(. Proof. Corollary 31 Assume that the semigroup (U (t))t≥0 is hyperbolic.10) with wt < h for t ≥ 0. 2 2 2 (n+1) 2 2wt − ut 2≤ 2 t S S 2 2 2 8λ X0 g(τ.
We will assume the followings. This completes the proof of the theorem. t∈R. Equation ( 7. we deduce that Q({fσ : σ ∈ lR}) is relatively compact in B(lR).2) is also almost periodic. Hv1 − Hv2 ≤ K1 K2 v1 − v2 . for σ ∈ lR.2) with f = g(. By the continuity of the operator Q. We can see that there exists a positive constant K2 such that v1 . This means that for each ε > 0 and any compact set K of CX there exists lε > 0 such that every interval of length lε contains a number τ with the property that sup g(t + τ. Proof. Then. ϕ) < ε. vt ) is also almost periodic. where u is the unique almost periodic solution of Equation (7. ϕ2 ) ≤ K1 ϕ1 − ϕ2 . This implies that if the function f is almost periodic. Let AP (lR) be the Banach space of almost periodic functions from lR to X endowed with the uniform norm topology. t ∈ lR. + lim lim s→+∞λ→+∞ s Then. v. Consider the operator H deﬁned on AP (lR) by Hv = u. ϕ1 ) − g(t. ). ϕ1 . . we obtain (Qf )σ = Q (fσ ) .ϕ∈K We know from [63] that if the function g is almost periodic in t uniformly in any compact set of CX and if v is an almost periodic function. ϕ) − g(t.. We are concerned now with the existence of almost periodic solutions of Equation (7. Q is a bounded linear operator from AP (lR) into B(lR). ϕ2 ∈ CX . then the function t → g(t. We have the following result.10) has a unique almost periodic solution.10). for t ∈ lR.392 DELAY DIFFERENTIAL EQUATIONS Proof. Deﬁne the operator Q by t S 8λ X0 f (τ ) (0) dτ U (t − τ ) B (Qf ) (t) = lim lim s→−∞λ→+∞ s t U S 8λ X0 f (τ ) U (t − τ ) B (0) dτ . By a sample computation. (H2) g is almost periodic in t uniformly in any compact set of CX . then the bounded solution of Equation (7. if in the assumption (H3) K1 is chosen small enough. (H3) hold. Proposition 55 Assume that the semigroup (U (t))t≥0 is hyperbolic and (H2). v2 ∈ AP (lR). (H3) g(t.
where A0 : D(A0 ) ⊆ X → X is the linear operator deﬁned by 3 & ' D(A0 ) = u ∈ C 1 ([0. such that ϕ(0. a). IR). u(0) = 0} = X. ρ(A0 ) = C. T ] . t ∈ [0. Equation (9. a) + (t. and F : [0. a ∈ [0. Moreover. 0) = 0. So. T ] × CX and satisﬁes a Lipschitz condition F (t. then the map H is a strict contraction in AP (lR). . This completes the proof of the proposition 9. we restate the equations. u(0) = 0 . a) = ϕ(θ. for t ∈ [0. T ] . T ] . a ∈ [0. t ∈ [0. with L ≥ 0 constant. l] . a. l] . a) = f (t. V (0) = ϕ. 2 2 2 −1 2 for λ > 0. Example 1 ∂u ∂u (t. This gives an almost periodic solution of Equation (7. T ] × CX → X is the function deﬁned by F (t. We have the following result. T ] . t ∈ [0. 2(λI − A0 ) 2 ≤ λ1 .1) u(t. where T. a)). l] . Applications We now consider the two examples mentioned in the introduction. ϕ) − F (t. A0 u = −u .10). IR) . ϕ(·. 0] . a)).1) reads 3 V (t) = A0 V (t) + F (t. ·). IR) . l] . l > 0. θ ∈ [−r. T ] and ϕ. Then. X) and X = C ([0. 0) = 0. ψ) ≤ L ϕ − ψ . ϕ ∈ CX and a ∈ [0. ∂t ∂a (9. By setting U (t) = u(t. ut (·. For the convenience of the reader. ϕ ∈ CX := C ([−r. Theorem 33 Assume that F is continuous on [0. u(θ. this implies that A0 satisﬁes (HY) on X (with M0 = 1 and ω0 = 0). for a given ϕ ∈ CX . l] . 0] . a. ψ ∈ CX . Vt ).Basic theory of abstract DFE with nondense domain 393 If K1 is chosen such that K1 K2 < 1. t ∈ [0. l] . ϕ)(a) = f (t. We have D(A0 ) = {u ∈ C([0. H has a unique ﬁxed point in AP (lR).
(S0 (t)x) (a) = 0a a−t x(s) ds. uτ −a+t (·. we obtain 3 a x(s) ds. IR) and 0 u(s. . In view of the deﬁnition of (S0 (t))t≥0 . uτ −a+t (·. a) = ϕ(0.1). y ∈ D(A0 ) and x ∈ X. a) ds + if t ∈ [0. a ϕ(0. x (a) = y(a) = a e−λt x(a − t) dt. a) = ϕ(t. 0)ds = 0. ut (·. if a ≤ t. if t ≤ 0. if a ≥ t. l] . a) ds. 0 Integrating by parts one obtains a −1 −λa x(t)dt + (λI − A0 ) x (a) = e 0 a −λt a e 0 a−t By uniqueness of Laplace transform. Proof. ϕ(t. u is the unique integral solution of the partial diﬀerential equation ( 9.1). a).2) Furthermore. a. t t (i) u ∈ CX . T ] → CX which solves the integral equation (3. (9. The assumptions of Theorem 33 imply that ϕ ∈ CX and ϕ(0. if a ≥ t ≥ 0. It suﬃces to calculate each term of the integral equation (3. 0] . τ )) dτ. for t ∈ [0.e. Consequently. solving the equation (λI − A0 ) y = x. we deduce that there exists a unique function v : [0. a − t) + a−t f (τ − a + t . a) + ∂ ∂a t t u(s. x(s)ds dt. f (s. (λI − A0 ) x (a) = λ 0 On the other hand. τ . τ . Let (S0 (t))t≥0 be the integrated semigroup on X generated by A0 . we obtain −1 (λI − A0 ) where λ > 0. if a ≤ t. τ )) dτ. u(t. we have +∞ −1 e−λt (S0 (t)x)(a) dt. T ] . T ] . 0 0 if t ∈ [−r. a). T ] → X solution of the following initial value problem a 0 f (τ − a + t . from Theorem 5.1). (ii) u(t. i. 0 u(s.394 DELAY DIFFERENTIAL EQUATIONS there exists a unique function u : [0. ·)ds ∈ C 1 ([0. ·) ∈ D(A0 ).
·) = ϕ(t. t+θ ϕ(s. 0) = 0 implies that S(·)ϕ ∈ C 1 ([0. . a) = = t+θ (S0 (t + θ − s) G (s)) (a) ds. CX ) and we have if a ≤ t + θ. τ ) dτ. a). 0 0 t+θ−a a−t−θ+s For a ≥ t + θ ≥ 0. a) = dt ϕ(t + θ. d ϕ(0. Let G : [0. T ] . The assumption ϕ(0. a − t − θ). a) = θ ϕ(s. a ≥ t + θ ≥ 0. we obtain a 0 θ ϕ(s. CX ). 0 a (S(t)ϕ) (θ. if t ≤ 0. dK a G(τ − a + t + θ)(τ ) dτ. T ] → X deﬁned by 3 v(t)(0. u(t. ·). t+θ a 0t+θ−a a G (s) (τ ) dτ ds + G (s) (τ ) dτ ds. if a ≥ t + θ ≥ 0. Corollary 7 implies that ut = v(t). Remark that the condition ϕ(0. (S(t)ϕ) (θ. T ] → C5X deﬁned by t S50 (t − s)X0 G(s) ds. if t ≥ 0. T ] . ·).395 Basic theory of abstract DFE with nondense domain Using Proposition 48. t + θ ≤ 0. a) ds + 0 ϕ(0. we have K(t)(θ. 0 G(τ − a + t + θ)(τ ) dτ. 0. we obtain t+θ K(t)(θ. we have K(t)(θ. if t + θ ≤ 0. a) = a−t−θ dt 0. Hence. a) = 0. a) ds. The derivation of K is easily obtained a a ≤ t + θ. K(t) = 0 For a ≤ t + θ. if a ≥ t + θ ≥ 0. 0) = 0 is necessary to have S(·)ϕ ∈ C 1 ([0. a) = 0 a G (s) (τ ) dτ ds. a−t−θ+s and for t + θ ≤ 0. T ] → X (T > 0) be a Bochnerintegrable function and consider the function K : [0. a) ds + a−t−θ ϕ(0. θ if a ≤ t + θ. if t + θ ≤ 0. τ ) dτ. If we consider the function u : [−r.2). (t)(θ. u is the unique solution of (9.
x)) for t ∈ [0. ut (·. ·) ∈ C 1 ([0. IR . a) + (0. for given ϕ ∈ CX such that ∂ϕ ∈ CX . 0] . a. with X = C Ω. l] and is equal to the unique solution of Problem ( 9. a) = f (0. T ] × CX → X is deﬁned by F (t. A0 u = ∆u. a)). x) = 0. l] . u(θ. ∆ is the Laplace operator in the sense of distributions on Ω and ϕ is a given function on CX := C([−r. V (0) = ϕ. One can use Theorem 10 (all the assumptions of this theorem are satisﬁed). ∂t ∂a the solution u of Equation ( 9. T ] .3) can be reformulated as an abstract semilinear functional diﬀerential equation 3 V (t) = A0 V (t) + F (t. 0] .3) where Ω ⊂ IRn is a bounded open set with regular boundary ∂Ω. x)). ψ) ≤ L ϕ − ψ . l] . Dϕ F (t. x ∈ Ω. ∂t u(t. x. ∆u ∈ C Ω. Problem (9. for a ∈ [0. ψ) ≤ β ϕ − ψ . x) + f (t. ∂t ∂ϕ (0. t ∈ [0. x) = ϕ(θ. 0) = 0 and ϕ(0. and F : [0. x). . ϕ)(x) = f (t. ψ) ≤ γ ϕ − ψ .2) is continuously diﬀerentiable on [0. 0) = ∂t ∂ϕ ∂ϕ (0. (9. with & ' D(A0 ) = u ∈ C Ω. ϕ(·.1). Dt F (t. Vt ). T ] . γ ≥ 0 such that F (t. x ∈ Ω. X). β. ϕ(·. ϕ) − Dt F (t. ϕ) − F (t. Example 2 ∂u (t.396 DELAY DIFFERENTIAL EQUATIONS The second part of Theorem 33 follows from Corollary 7 Theorem 34 Assume that F is continuously diﬀerentiable and there exist constants L. x ∈ ∂Ω. IR) . x. ϕ(0. ϕ ∈ CX and x ∈ Ω. t ∈ [0. x) = ∆u(t. Then. θ ∈ [−r. T ] × [0. T ] . IR and u = 0 on ∂Ω . IR . t ∈ [0. ϕ) − Dϕ F (t. Proof. T ] .
·) = 0 on ∂Ω and ∂ϕ (0. and x ∈ Ω if t ∈ [−r. x. u = 0 on ∂Ω . for x ∈ Ω. Moreover ρ(A0 )⊂ (0. ut (·. x). T ] → X of the partial diﬀerential equation ( 9. +∞) 2 2 2 −1 2 for λ > 0. x)). x. ψ ∈ CX .397 Basic theory of abstract DFE with nondense domain & ' We have D(A0 ) = u ∈ C Ω. t t (i) u ∈ CX . 0] . ·)ds ∈ C Ω. IR . x) ds + 0 f (s. ∂t There is a unique function x deﬁned on [−r. ·) ∈ C Ω. this implies that A0 satisﬁes (HY) on X (with M0 = 1 and ω0 = 0). ψ) ≤ γ ϕ − ψ . we obtain the following theorems (the all assumptions are satisﬁed). T ] × CX and satisﬁes a Lipschitz condition F (t. ·)ds = 0. Dϕ F (t.e. x) + f (0. ϕ) − F (t. on ∂Ω.3) on [0. ∂t ϕ(0. u(t. Then. such that ∆ϕ(0. x) = ∆ϕ(0. ϕ) − F (t. ϕ) − Dϕ F (t. ∆ϕ(0. ψ) ≤ L ϕ − ψ . for given ϕ ∈ CX such that ∂ϕ ∈ CX . for a given ϕ ∈ CX . T ] . if t ∈ [0. Theorem 36 Assume that F is continuously diﬀerentiable and there exist constants L. ψ) ≤ L ϕ − ψ . ψ) ≤ β ϕ − ψ . x) ds. T ] and ϕ. ·) = 0. ·) = ∂ϕ ∂t (0.3). β. ∆ 0 u(s. x) = ϕ(t. Theorem 35 Assume that F is continuous on [0. (ii) t t ϕ(0. 0] × Ω and satisﬁes Equation ( 9. . Dt F (t. i. = X. IR and 0 u(s. T ] × Ω. 2(λI − A0 ) 2 ≤ λ1 . x) + ∆ 0 u(s. Using the results of Section 3. ϕ(·. Then. with L ≥ 0 constant. IR . and x ∈ Ω. there exists a unique integral solution u : [0. T ] × Ω. ϕ) − Dt F (t. on ∂Ω. t ∈ [0. γ ≥ 0 such that F (t. such that x = ϕ on [−r.
Vol. Kimmel (1989) Asymptotic behavior of a nonlinear functionalintegral equation of cell kinetics with unequal division. J. Arino and E.9. 30: 437456. Biol. Mortabit (1992) A periodicity result for a nonlinear functional integral equation.6: 13051320. S´ anchez (1996) A variation of constants formula for an abstract functional diﬀerential equation of retarded type. Diﬀ. [2] O. 401 . SIAM J. Some consequences. N. Vol. [8] O. Models and Methods in Applied Sci. Math. SIAM Rev. J. N. [6] O. S´ anchez (1998) An integral equation of cell population dynamics formulated as an abstract delay equation. Math. Manitius (1978) On semigroups in Rn × Lp corresponding to diﬀerential equations with delay. Arino and E. Kimmel (1991) Asymptotic behaviour of nonlinear semigroup describing a model of selective cell growth regulation. J. Arino and M. [4] O. and Appl. J. Arino and A. 29: 289314. Math.4: 445476. N. Arino and E. Biol. Kimmel (1987) Asymptotic analysis of a cell cycle model based on unequal division. Bernier and A. Biol. J.4: 713735.1: 128145. 191: 547571.34. and Int. S´ anchez (1995) Linear Theory of Abstract Functional Diﬀerential Equations of Retarded Type. 27: 341354. [7] O.References [1] O. 30: 897914. Anal. 47. Arino (1992) Some spectral properties for the asymptotic behavior of semigroups connected to population dynamics. [3] O. Vol. Math. Canad. Arino and M.8. Vol. Arino and M. Equat. Math. [5] O. Appl. N. Math. [9] C. Math.
Anal. Springer. Hille (1948) Functional analysis and semigroups. and Applications. Soc. Am. J. Math. 311: 357390. Trans. [20] J.R. [11] A. Grabosch and H. Hale (1977) Functional Diﬀerential Equations. Hale and S. Schwartz (1958) Linear Operators. [23] T. Soc. Anal. Verduin Lunel (1993) Introduction to Functional Diﬀerential Equations. Part I: General Theory. 5: 435450. Math. [14] J. OneParameter Semigroups of Positive Operators. Heijmans and H. H. Colloq. Math. Diekmann. Diﬀ. Amer.M. Providence. RI. Mathematical Surveys N. A.S. [16] A. Soc. Math. Math. Springer. Grabosch (1989) Translation semigroups and their linearizations on spaces of integrable functions.M.400 DELAY DIFFERENTIAL EQUATIONS [10] F. Grabosch and U. 26: 112159. Math. Interscience. XXXI. Pub. Chow and J. SpringerVerlag. Equat.J. 131. Amer. J. and Appl.E. R. Marcel Dekker: 189210. MalletParet (1977) Integral averaging and bifurcation. Browder (1961) On the spectral theory of elliptic diﬀerential operators. [21] E. Dunford and J. development and maturation of red blood cells: a mathematical model. [18] A. J. Math. Lect.J. Nagel Ed. Uhl (1977) Vector Measures. [15] N. Kato (1966) Perturbation Theory for Linear Operators. 223: 808826. Heijmans (1991) Production.Ann. . Mathematical Population Dynamics. J. 1184 (219232). [13] O. Thieme (1984) On the stability of the cell size distribution. 142: 22130. Lect. Notes in Math. [19] J. Jones (1962) The existence of periodic solutions of f (x) = −αf (x − 1){1 + f (x)}. El Idrissi (1999) Asymptotic behaviour of an agestructured population model and optimal maturation age.15. Notes in Pure and Applied Mathematics. Biol. 19: 227248. [17] A.N.J.S. Vol. Diestel and J. SpringerVerlag. Moustakas (1986) A semigroup approach to retarded diﬀerential equations. [12] S. Calsina and O. [22] G.
N. J. [30] S.) (1986) OneParameter Semigroups of Positive Operators. Manitius (1980) Completeness and Fcompleteness of eigenfunctions associated with retarded functional diﬀerential equations. 1184. J. Traganos (1984) Analysis of a cell cycle model based on unequal division of metabolic constituents to daughter cells during cytokinesis. O. Kunish and W.T. Arino and M.REFERENCES 401 [24] M. Nussbaum (1970) The radius of the essential spectrum. TMA. Duke Math.4. [31] S. Lect. Diﬀ. Kimmel (1991) Noncompact semigroups of operators generated by cell kinetics models. O. Pazy (1983) Semigroups of linear operators and applications to partial diﬀerential equations. Nakagiri (1988) Structural properties of functional diﬀerential equations in Banach spaces. Anal. Schappacher (1981) Variation of Constants Formula for Partial Diﬀerential Equations with Delay. J. Appl. Biol. Math. 37: 473478. Diﬀ. 35: 129. 110: 637664. [29] S. [33] A. Schappacher (1983) Necessary conditions for partial diﬀerential equations with delay to generate C0 semigroups. Nakagiri (1987) Pointwise completeness and degeneracy of functional diﬀerential equations in Banach spaces I: General time delays. Plant (1977) Nonlinear semigroups of translations in Banach space generated by functional diﬀerential equations. Arino and F. Nakagiri (1981) On the Fundamental Solution of DelayDiﬀerential Equations in Banach Spaces. Math. [28] R. Appl. Equations. [32] R. 5: 123142. 50: 4979. Osaka J. Math. N. J. Nagel (Ed. 25: 353398. . Kimmel. J. Kunish and W. [35] W. Diﬀerential and Integral Equations. 60: 6774. [25] K. SpringerVerlag. Nonlinear Anal. [26] K. 127: 492529. S´ anchez.6: 12331249.Y. Springer Verlag. [27] A. McGrawHill. 41: 349368. Z. J. Anal. Equations. Rudin (1973) Functional Analysis. Vol. Notes Math. Darzynkiewicz. Equations. [34] A. Theor. Diﬀ. J. [36] E.
Diﬀ.F. Amer. Dynamic Systems and applications. Local existence and linearized stability for partial functional diﬀerential equations. [42] G. Ezzinbi. Travis and G.F. (1999).Mathematical Analysis and . (1993). [39] A. Trans. 177. A class of linear partial neutral functional diﬀerential equations with nondense domain. Shumacher (1985) On the resolvent of linear nonautonomous partial diﬀerential equations. 7. (1996). [44] M. [48] M. Journal of . R. Webb (1974) Existence and Stability for Partial Functional Diﬀerential Equations. 147. Schaefer (1971) Topological Vector Spaces. Adimy et K. Webb (1985) Theory of Nonlinear AgeDependent Population Dynamics. Some results on existence. Acad. S´erie I.Applications. (to appear 2004). Journal of Diﬀerential Equations. Soc. 285332. J. Math. [40] C.O. Ezzinbi. Sci. 200: 395418. Journal of Diﬀerential Equations and Dynamical Systems Vol. [38] K. Paris.H. 59: 355387. (1998). Ezzinbi. Walther (1975) Existence of a nonconstant periodic solution of a nonlinear nonautonomous functional diﬀerential equation representing the growth of a single species population. Adimy et K.323. Adimy and K. 389404. asymptotic behavior and stability of solutions for a class of partial neutral functional diﬀerential equations. 125134. Integrated semigroups and delay diﬀerential equations. [47] M Adimy and K. Ezzinbi. J. Taylor (1967) Introduction to Functional Analysis. Existence and linearized stability for partial neutral functional diﬀerential equations. [46] M Adimy and K. Equations. Math. (1998). [43] M.Adimy. t. [45] M. Semigroupes int´egr´es et ´equations diﬀ´erentielles `a retard en dimension inﬁnie. Springer Verlag.402 DELAY DIFFERENTIAL EQUATIONS [37] H. Number 4. C. Marcel Dekker Inc. 1: 227240. Ezzinbi. [41] H. October. 371417.C. John Wiley. Biol. 481486. Hiroshima Mathematical Journal. 7. .
Laklach. Theory. http://ejde. K. Ezzinbi. Boundedness and almost periodicity for some statedependent delay diﬀerential equations. [54] M. 67. 48. H. 91112..Ezzinbi and M. 9. C. Ezzinbi. Sci. [51] M Adimy. Paris. Bouzahir and K.Acad. (1987). [50] M Adimy. Winter (2001).Arendt. [55] E. (2004). Israel J. 1. H. (2001).Bouzahir and K. Ezzinbi. 438461. (2002). Existence for a class of partial functional diﬀerential equations with inﬁnite delay.54. 2002. Methods and Applications. Local existence for a class of partial neutral functional diﬀerential equations with inﬁnite delay. Local existence and stability for a class of partial functional diﬀerential equations with inﬁnite delay. Spectral decompostion for some neutral partial functional diﬀerential equations. (1987). (2000). Local existence and global continuation for a class of partial neutral functional diﬀerential equations..math. Journal of Nonlinear Analysis. Electronic Journal of Diﬀerential Equations. 321349.Arendt. Journal of Mathematical Analysis and Applications.REFERENCES 403 [49] M Adimy. Vol. Vol 46. Vol. .swt. London Math. Adimy. (2002). Vector valued Laplace transforms and Cauchy problems. Existence and stability for some partial neutral functional diﬀerential equations with inﬁnite delay. 957962. Ezzinbi and M. Ezzinbi. [56] W. 59. [52] M. [53] M.4. (to appear 2004). H.R. 113. Journal of Diﬀerential Equations and Dynamical Systems. Ezzinbi. No . No . Adimy.edu . 327352. S´erie I. Ait Dads and K. 323348. 294.330. Bouzahir and K. Issue 2. Journal of Nonlinear Analysis. Adimy. 134. t. Proc. Vol. Canadian Applied Mathematics Quarterly. Bouzahir and K. K. Methods and Applications. No . Soc. H. Resolvant positive operators and integrated semigroup.Math. 3. [57] W. vol. Theory. Laklach.
H. Existence and stability for some partial functrional differential equations with inﬁnite delay. Perturbation theory for dual semigroups. 709725. Ann. vol.Sanchez.M. .Da Prato and E. SpringerVerlag.A. [63] C. the sunreﬂexive case.R. Diﬀerential and Integral Equations. Int. Corduneanu. Diﬀ. Jazar. S. 191. Delay Equations. Eq.Cl´ement. Complex and Nonlinear Analysis.swt. Linearized stability for nonlinear semigroups. Linear theory of abstract functional diﬀerential equations of retarded type. http://ejde. (1995). 285344. Van Gils. No .uszeged. (1995). [61] S.Sinestrari. M.) pp. Electronic Journal of Qualitative Theory of Diﬀerential Equations. in “Diﬀerential Equations in Banach Spaces” (A. (1987). (2003). Functional.Arino et E. [60] K. 2003.. Pisa Cl.hu/ejqtde/. Journal of Applied Mathematics and Stochastic Analysis. Electronic Journal of Differential Equations. Diﬀerential operators with nondense domains. Ezzinbi. (1987). New York / Berlin (1986).Wu. Walther. (to appear 2004). 277. 5. 14. [62] P. May..Mathematical Analysis and Applications. Benkhalti.. http://www. Chelsea . 3. Math. (2004). Heijmans and H. Desch and W.math. [66] O. Eds. Sci.J. Schappacher.Thieme. Lecture Notes in Math.Busenberg and B. Almost Periodic Functions. Diekmann. Norm. 116. Ann. 112. New criteria for the existence of periodic and almost periodic solutions for some evolution equations in Banach spaces. Convergence theorems for integrated semigroup. Verduyn Lunel and H. SpringerVerlag.edu. 509520. [68] K.math.Gyllenberg.404 DELAY DIFFERENTIAL EQUATIONS [58] O.6. 113. Periodic solutions for some partial functional diﬀerential equations. 547571. [59] R. Benkhalti and K.Favini and E. (to appear 2004). No . O.Obrecht. (1992). New York (1989). 1223. Existence and stability for some partial neutral functional diﬀerential equations.Sup. Vol. O.. 2nd edition. 6173. [64] G. Ezzinbi and R. Scuola. M.Diekmann. [65] W. Vol. [67] K. S. Journal of . Ezzinbi and M.. Ezzinbi. A. I.
ed.REFERENCES 405 [69] K. [71] K. (1989). Periodic solutions of NonDensely deﬁned delay evolution equations. (2002). Berlin/New York.Rhali.Ezzinbi anf J. 2.Kunish and W. Kellermann. [74] J. [77] T. 29. [76] M. 10271038. New York. Variationofconstants formula for partial diﬀerential equations with delay. (1983). Vol. 113. 1184. [75] J. Berlin (1966). [81] H. 4979. 15.Fitzgibbon. p 219232. Liu. Integrated semigroup. 1. 160180.). Dissertation. 123142.E. 114.Hale. Vol. Positivity and stability for some partial functional diﬀerential equations. Journal of Applied Mathematics and Stochastic Analysis. (2002). Nonlinear Anal.Grabosch and U.Kellermann and M. J.Kato. (1981). Ezzinbi and J. New York (1993).Lunel. SpringerVerlag. .. [78] K.Anal. Springer. (1986). (1978). T¨ ubingen. (1989). (1977).Schappacher. in ” OneParameter Semigroups of Positive Operators (R. No .H.K.Kunish and W.Hieber.Nagel. Vol. [72] W. Lecture Notes in Mathematics. Perturbation theory for linear operators. Theory of functional diﬀerential equations SpringerVerlag. J. [79] K. Necessary conditions for partial differential equations with delay to generate a C0 −semigroups. Integrated semigroups and diﬀ. H.E 50. TMA 5.Fun. 10. 113123. Dissertation. SpringerVerlag.Moustakas. [70] K. (1986).Schappacher. Integrated Semigroups. Diﬀerential Equ. Ezzinbi and S. 15.L. [73] A. (2003). 36. [80] H. operators on Lp . J. Liu. Nondensely deﬁned evolution equations with non local conditions. Nonlinear Diﬀerential Equations and Applications. A semigroup approach to retarded diﬀerential equations. .D.K. Mathematical and Computer Modeling. Introduction to functional diﬀerential equations. Semilinear functional diﬀerential equations in Banach space.Hieber.Hale and S. No .
Arino. Methods and Applications. (1988). 131142. Diﬀ. [86] F.27. Journal of . Oneparameter Semigroups of Positive Operators. 111155. Summers. Parrrot. (1996). On the abstract Cauchy problem of parabolic type in spaces of continuous functions. MarcelDekker. Nagel and E. Journal of Mathematical Analysis and Applications.Mathematical . (1991).E.Thieme. Theory. H. O. SpringerVerlag. 310336.. Linearized stability and irreducibility for a functional diﬀerential equation. SIAM J.. Sinestrari. 167186.Mathematical . [85] R. 223232.Memory.Analysis and . 152. Paciﬁc.Pazy. Math. Semiﬂows generated by Lipschitz perturbations of nondensely deﬁned operators. 2.Axelrod and M. Semigroups of linear operators and applications to partial diﬀerential equations. Notes in Mathematics. Math. 198. Integrated semigroups and their applications to the abstract Cauchy problems. [93] H. Nagel.M.E. 170182. and Integ. 1666.6. (1986). 416447. Springer. (1996).. Nonlinear hyperbolic Volterra integrodiﬀerential equations. Anal. (1990). [92] H. No.Sinestrari. Ruess and W. [83] M. Journal of . Positivity and a principle of linearized stability for delay diﬀerential equations.E.Kimmel) (1991). [89] A. In Mathematical Population Dynamics (Eds.Neubrander. Equations 3. 16. 1184.Thieme. TMA. [88] M. Lect. (1990). [91] E. No. Integr. 107 (1985). Parrrot. 10351066. Equa.2. Vol. (1989). Integrated semigroups and integrated solutions to abstract Cauchy problems.Analysis and . [84] R.Memory. Stable and unstable manifolds for partial functional diﬀerential equations. D. (1992). Berlin (1983). Diﬀ. Invariant manifolds for partial functional diﬀerential equations. J. J.Applications. Nonlinear Analysis. . [90] W. 135. 23. Linearized stability for abstract diﬀerential equations with delay. Nonlinear Anal. 649661.Applications. [87] M.406 DELAY DIFFERENTIAL EQUATIONS [82] M.
REFERENCES 407 [94] C. [96] J. SpringerVerlag.Webb. 395418. Trans.F. Asymptotic stability for abstract functional diﬀerential equations.Travis and G. Existence and stability for partial functional diﬀerential equations. [95] G. 200. Soc. (1974). Amer. Math.C.Soc. Amer. Theory and Applications of Partial Functional Diﬀerential Equations. Proc. . (1996). Math. 54 (1976). 225230.Webb.F. Wu.
© 2006 Springer. Some results require that f is continuously diﬀerentiable. te ).1 Semiﬂows of retarded functional diﬀerential equations Let h > 0. 0]. 0]. In the sequel we shall assume that U ⊂ C is open and that f is locally Lipschitz continuous. t] ⊂ I. This implies 411 O.Walther@math.). diﬀerentiable on (t0 . 411–476. Autonomous retarded functional diﬀerential equations on C (RFDEs) are equations of the form x(t) ˙ = f (xt ). −∞ < t0 < te ≤ ∞. Solutions x : [t0 − h.Chapter 10 DYNAMICS OF DELAY DIFFERENTIAL EQUATIONS H. with φ = max−h≤s≤0 φ(s). for a given map f : C ⊃ U → Rn . the segment xt : [−h. Rn ). 0]. te ) → Rn . (eds. 0] → Rn is the shifted restriction of x to the initial interval [−h.O. . Let C denote the Banach space C([−h. of the previous RFDE are continuous. Basic theory and some results for examples Proofs for most of the basic results presented in this part are found in the monograph [57].unigiessen. 1. 2 D 35392 Giessen Germany HansOtto.de 1. The space C will serve as state space. and satisfy the RFDE for t0 < t < te . See also [105]. Delay Differential Equations and Applications. xt (s) = x(t + s) for s ∈ [−h. Arino et al. n ∈ N. For x : R ⊃ I → Rn and t ∈ R with [t − h. Walther Mathematisches Institut Universit¨ at Giessen Arndtstr.
The condition ξ g(ξ) > 0 for all ξ = 0 corresponds to delayed positive (and instantaneous negative) feedback. Solutions are easily found via the method of steps: For φ ∈ C given. (1. The equation x(t) ˙ = −µ x(t) + g(x(t − 1)) (1. In case of negative feedback and µ ≥ 0. which have any two zeros spaced at a distance larger than the delay h = 1. is a solution of the RFDE which satisﬁes the initial condition and has the property that any other solution of the same IVP is a restriction of xφ . evaluate successively the variationofconstants formulae t −µ(t−n) x(n) + e−µ(t−s) f (x(s − 1))ds. n = 1. This yields the maximal solution xφ : [−1. Example. with x0 = φ. Let µ ∈ R and a continuously diﬀerentiable function g : R → R be given.If g(0) = 0 then x : R t → 0 ∈ R is a constant solution (equilibrium solution). . A maximal solution xφ : [t0 − h. and f (φ) = −µ φ(0) + g(φ(−1)) for all φ ∈ C. (1.412 DELAY DIFFERENTIAL EQUATIONS that their right derivative at t0 exists and equals f (xt0 ). te (φ)) → Rn of the initial value problem (IVP) x(t) ˙ = f (xt ). te ) (including on R) are diﬀerentiable and satisfy the RFDE everywhere.1). x(t) = e n for nonnegative integers n and n ≤ t ≤ n + 1. with h = 1. Solutions on (−∞. How is slowly oscillating behaviour reﬂected in the state space C ? Segments xt of slowly oscillating solutions belong to the set S ⊂ C \ {0} of . x0 = φ ∈ U.1) is an RFDE. ∞) → R of the IVP for eq. For µ > 0 and ξ g(ξ) < 0 for all ξ = 0 eq. initial data with at most one zero deﬁne slowly oscillating solutions.1) models a combination of delayed negative feedback and instantaneous negative feedback with respect to the zero solution.
2. and 0 ≤ φ. or φ ≤ 0. It can be shown that S is homotopy equivalent to the circle. ∞) · S ⊂ S which says that S is a wedge. a]. (−1)j φ(t) ≤ 0 on [a. 0] so that 0 ≤ (−1)j φ(t) on [−1. S is not convex. g bounded from below or from above. φ = 0. or there are j ∈ {0.1. 1} and a ∈ [−1. the set R of data whose maximal . See [240]. initial data with at most one change of sign. and g (ξ) < 0 for all ξ ∈ R almost every initial function φ ∈ C deﬁnes a solution which is eventually slowly oscillating. We have cl S = S ∪ {0}. The set S is positively invariant.e.Dynamics of Delay Diﬀerential Equations 413 Figure 10. Let us call a solution eventually slowly oscillating if for some t its zeros z ≥ t are spaced at distances larger than the delay h = 1. The main result of [177] establishes that in case µ ≥ 0. Figure 10. More precisely. for all φ ∈ S and all t ≥ 0. 0]. xφt ∈ S. More precisely. and (0.. i.
Remark. ∞) × U : t < te (φ)}. Then compute a slowly oscillating periodic solution for a suitable smooth function g which is positive and constant on some interval (−∞. Back to the general case. In a simple case. In case of positive feedback. (1. φ) = φ and . ∞). F (t. φ) ∈ [0. See [137]. Recall the algebraic semiﬂow properties F (0. 0 < b < a. the convex cones of nonnegative and nonpositive data are positively invariant. Exercises.414 DELAY DIFFERENTIAL EQUATIONS solutions are not eventually slowly oscillating is closed and a graph T ⊃ dom → Q with respect to a decomposition of C into a closed subspace T of codimension 2 and a 2dimensional complementary subspace Q. Deﬁne and compute a ‘slowly oscillating periodic solution’ of eq. The maximal solutions of the IVP with t0 = 0 deﬁne a continuous semiﬂow F : Ω → U. −a] and linear and decreasing on some interval [−b.1). ∞) × U . This shows that the set C \ R of initial data of eventually slowly oscillating solutions is open and dense in C.1) for µ = 0 and the step function g = −sign. For certain bounded but nonmonotone functions g and for certain µ > 0 there exist rapidly oscillating periodic solutions with attracting orbits [116].(1. Problem. This means that for such g and µ the previous statement is false. The domain Ω is open in [0. which was excluded from the considerations up to here. they are easily found. φ) = xφt and Ω = {(t. Does the density result hold in case µ = 0 for all g which satisfy the negative feedback condition ? We shall see that in case of negative feedback slowly oscillating periodic solutions play a prominent role in the dynamics generated by eq.
Also. φ)). ·) are linear.χ with the maximal solution v φ. For certain t > 0. φ)) = F (t+s. (1. Solutions of the equation y(t) ˙ = L(t. Example. t ≥ 0. ∞). te (φ)) → Rn of an IVP associated with the nonlinear RFDE above the variational equation along x is the previous linear equation with L(t.1) with g constant on a nontrivial interval I. 0 ≤ t < te (F (s.1). Ωt may be empty. F (s. Suppose f is continuously diﬀerentiable. te ) × C → Rn . as there are no backward solutions for nondiﬀerentiable initial data. and for all φ ∈ Ωt . t0 < te ≤ ∞.Dynamics of Delay Diﬀerential Equations 415 F (t.2) with α = g (0). initially diﬀerent solutions may become identical on some ray [t. ψ) = Df (xt )ψ for t0 ≤ t < te (φ). For a given maximal solution x = xφ : [t0 − h. Before discussing diﬀerentiability with respect to initial data a look at nonautonomous linear delay diﬀerential equations is convenient: Let L : [t0 . yt ) are deﬁned as before. Example. Therefore solutions of the backward IVP are in general not uniquely determined. te ). DFt (φ)χ = vtφ. The variational equation along the zero solution of eq. te (φ)) → Rn of the IVP v(s) ˙ = Df (F (s. In case f is continuously diﬀerentiable all solution operators Ft on nonempty domains are continuously diﬀerentiable. ψ ∈ C. (1. Remark. φ) whenever φ ∈ U. or linearization at 0. are deﬁned on the open sets Ωt = {φ ∈ U : (t. is v(t) ˙ = −µ v(t) + α v(t − 1) (1. χ ∈ C. . Eq. v0 = χ. be continuous and assume all maps L(t. The solution maps Ft = F (t. initial data φ = ψ with range in I and φ(0) = ψ(0). ·). Maximal solutions exist and are deﬁned on [t0 − h.χ : [−h. Let g(0) = 0. φ))vs . A ﬂow does not exist in general. 0 ≤ s < te (φ). φ) ∈ Ω}.
x0 )1 ∈ Recall that L is the kernel of a continuous linear functional φ∗ . te (φ)) t → F (t. . with D1 F (t. I. An application of the Implicit Function Theorem close to the solution (p. The Poincar´e return map P : H ∩ N φ → F (τ (φ). Notice that for 0 ≤ t < h the solution segments are in general only continuous and not diﬀerentiable. φ) − x0 ) = 0. φ) ∈ H is continuously diﬀerentiable (with respect to the C 1 submanifold structures on H ∩ N and H).e. φ)1 = x˙ φt . p]} ⊂ U at x0 = xp is a hyperplane H = x0 + L. 1. D1 F (p.2 Periodic orbits and Poincar´ e return maps Let a periodic solution x : R → Rn of the RFDE be given.. with a closed subspace L of codimension 1. φ) ∈ U exist and are given by the derivatives of the solution segments. φ) ∈ H is equivalent to the equation φ∗ (F (t. x0 ) yields an open neighbourhood N of x0 in C and a continuously diﬀerentiable intersection map τ :N →R so that τ (x0 ) = p and F (τ (φ). Assume f is continuously diﬀerentiable. not necessarily minimal. so that / L = Tx0 H. The relation F (t. φ) ∈ H for all φ ∈ N.416 DELAY DIFFERENTIAL EQUATIONS On the open subset of Ω given by t > h the semiﬂow F is continuously diﬀerentiable. Let p > h be a period of x. A transversal to the periodic orbit o = {xt : t ∈ [0. for t > h the tangent vectors of the ﬂowline [0.
. are given by prφ ◦ DFτ (φ) (φ)χ where prφ : C → L is the projection along the (tangent) vector / L.4. η = D1 F (τ (φ). In particular. DP (x0 )χ = prx0 DFp (x0 )χ.417 Dynamics of Delay Diﬀerential Equations Figure 10. φ)1 = x˙ φτ(φ) ∈ We have prφ ψ = ψ − φ∗ ψ η. φ∗ η Figure 10. φ ∈ H ∩ N and χ ∈ Tφ H = L. The derivatives DP (φ)χ.3.
A (compact) global attractor A for F is a compact invariant set A ⊂ M which attracts bounded sets in the sense that for every bounded set B ⊂ M and for every open set V ⊃ A there exists t ≥ 0 such that F (s. B) for all t > 0. Therefore standard tools of dynamical systems theory on ﬁnitedimensional manifolds. ∞) × M → M on a complete metric space M ..418 1. not necessarily invertible maps in Banach spaces. It follows that all these maps are not diﬀeomorphisms. B) ⊂ V for all s ≥ t. (F0 = idU is not compact. which had originally been developped for diﬀeomorphisms only. Compactness follows by means of the AscoliArz`ela Theorem from simple apriori bounds on solutions and their derivatives. 1. A global attractor contains each bounded set B ⊂ M which satisﬁes B ⊂ F (t. due to dim C = ∞). local invariant manifolds.g. and the shadowing lemma associated with a hyperbolic set. (1. Consider for example eq. e. Also Poincar´e maps are often compact. It follows from the exercise that a global attractor A contains every compact invariant set. Existence of global attractors is in some cases easy to verify. Example. Eq. A complete ﬂowline is a curve u : R → M so that F (t.4 Global attractors Consider a semiﬂow F : [0. s ∈ R.3 DELAY DIFFERENTIAL EQUATIONS Compactness Often the solution maps Ft . had to be generalized in a suitable way for diﬀerentiable. u(s)) = u(t + s) for all t ≥ 0. t ≥ h. A coincides with the set of all values of all bounded complete ﬂowlines. A set I ⊂ M is invariant if each x ∈ I is a value of a complete ﬂowline u with u(R) ⊂ I. Exercise. This is equivalent to Ft (I) = I for all t > 0. . are compact in the sense that they map bounded sets into sets with compact closure.1) with g bounded from above or from below. inclination lemmas. Furthermore. (1) with µ > 0 and g : R → R injective and bounded. and that it is uniquely determined.
So the spectrum σ of G consists of a countable number of isolated eigenvalues of ﬁnite multiplicities.Dynamics of Delay Diﬀerential Equations 419 Then all solution maps are injective. The set + F (t. α < − π2 (negative feedback). and the complexiﬁcation of G has compact resolvents.5 Linear autonomous equations and spectral decomposition For L : C → Rn linear continuous. of the generator. The semiﬂow is now given by the C0 semigroup of solution operators Tt : C φ → ytφ ∈ C.2) with µ = 0. (1. Example: For eq. An important property is that any halfplane given by an equation Re λ > β for β ∈ R contains at most a ﬁnite number of eigenvalues. The Ansatz y(t) = eλt c for a solution leads to a characteristic equation for the eigenvalues. The solution operators are compact for t ≥ h. consider the RFDE y(t) ˙ = Lyt . DG = {φ ∈ C : φ continuously diﬀerentiable. there exist pairs of complex conjugate eigenvalues with positive real part. . Examples. For r > 1 + (sup g)/µ the open ball Cr in C with radius r > 0 and center 0 is positively invariant under the semiﬂow and absorbs every ﬂowline in the sense that for every φ ∈ C there exists t ≥ 0 with F (s. t ≥ 0. Cr ) A= t≥0 is the global attractor [137]. λ + µ − α e−λ = 0.2). which are real or occur in complex conjugate pairs. (1. and F1 is compact. and there are no real eigenvalues. φ) ∈ Cr for all s ≥ t. 1. The domain of its generator G : DG → C is ˙ = Lφ}. φ(0) ˙ The solution behaviour is determined by the spectrum and Gφ = φ. In case µ > 0 and α > 0 (positive feedback) the eigenvalue with largest real part is positive and all other eigenvalues form complex conjugate pairs. For eq.
5.420 DELAY DIFFERENTIAL EQUATIONS Figure 10. .
σ<0 of the eigenvalues with positive. For the leading eigenvalues. and that the semigroup is not analytic. β) ≥ 0. The semigroup is called hyperbolic is there are no eigenvalues on the imaginary axis. The unstable. C c consists of segments of solutions with polynomial growth.. as well as generalized eigenspaces C<β associated with all eigenvalues in a left halfplane given by an equation Re λ < β. with k = k(L. i. On the complementary subspace C<β . C u consists of segments of solutions on the line which decay to 0 exponentially as t → −∞ and whose segments grow exponentially as t → ∞. Tt φ ≤ k eβ t φ. respectively. C s of G are the realiﬁed generalized eigenspaces given by the sets σ>0 . Solutions starting in C s decay exponentially to 0 as t → ∞.Dynamics of Delay Diﬀerential Equations 421 Rapid growth of the imaginary parts of the eigenvalues as Re λ → −∞ shows that G is not sectorial.β = eBt .e. center. (1. generalized eigenspaces of an eigenvalue with respect to G and B coincide. Examples of other spectral decompositions of C. t ∈ R. there is a leading pair of eigenvalues. with a linear vectorﬁeld B : C≥β → C≥β . On the ﬁnitedimensional generalized eigenspace C≥β given by the set σ≥β of all eigenvalues satisfying Re λ ≥ β the semigroup induces a group of isomorphisms Tt. and stable subspaces C u . For convenience we choose β so that there are no eigenvalues on β + iR. σ0 . with β ∈ R. C c . . In the case µ ≥ 0 and α < 0 (negative feedback). We have C = C u ⊕ C c ⊕ C s. t ≥ 0. Consider eq. Im λ < π.2). there exists β ∈ R so that the sum of the multiplicities of the eigenvalues with Re λ > β is 2. and negative real part. zero. The realiﬁed generalized eigenspaces associated with a real eigenvalue or a pair of complex conjugate eigenvalues are positively invariant under the semigroup. The spectrum of B is σ≥β .
. Later we shall consider the decomposition C = C2 ⊕ C2< where C2< = C<β . (1. Figure 10. for the space T containing the domain dom of the map representing the thin set R of segments of rapidly oscillating solutions and for the target space Q of this map.2). C2< and C2 are here deﬁned by the variational equation (1.422 DELAY DIFFERENTIAL EQUATIONS and all solutions with segments in the 2dimensional space C2 = C>β are slowly oscillating.6.2) along the zero solution of eq. we actually have T = C2< and Q = C2 . See [240]. (1. In the result from [177] on density of initial data for eventually slowly oscillating solutions of the nonlinear eq. Of course. Solutions on the line with segments in realiﬁed generalized eigenspaces associated with other complex conjugate pairs of eigenvalues are not slowly oscillating. We have C2 ⊂ S ∪ {0} and S ∩ C2< = ∅.1).
and 0 ∈ U is a stationary point of the semiﬂow: F (t. Then the zero function on the line is a solution. Choose β > 0 so that the 3 eigenvalues with largest real parts satisfy β < Re λ. with α suﬃciently large. See [137]. C c . C s . For simplicity. respectively. 1. and there are no eigenvalues on β + iR. N c . 0) = 0 for all t ≥ 0. The linear variational equation along the zero solution is given by L = Df (0). Figure 10. N s of 0 in the spaces C u . The solutions on the line with segments in C3 = C>β have at most 2 zeros in any interval of length h = 1. Later we shall consider the decomposition C = C3 ⊕ C3< where C3< = C<β . ws : N s → C u ⊕ C c with ﬁxed point 0 and zero derivatives at 0 whose graphs W ∗ = {χ + w∗ (χ) : χ ∈ N ∗ } . and continuously diﬀerentiable maps wu : N u → C c ⊕ C s . there are at least 3 eigenvalues in the open right halfplane.Dynamics of Delay Diﬀerential Equations 423 In the positive feedback case µ ≥ 0 and α > 0. and they decay to 0 as t → −∞ faster than t → eβ t .7. There exist open neighbourhoods N u . wc : N c → C u ⊕ C s . φ0 = 0. and f (φ0 ) = 0.6 Local invariant manifolds for nonlinear RFDEs Assume f is continuously diﬀerentiable.
asymptotic stability of the linearized equation implies the same for the equilibrium of the original nonlinear equation. . Then there exists a 3dimensional leading local unstable manifold W3 .1) with µ > 0 and α > 0 suﬃciently large (positive feedback). i. one obtains the Principle of Linearized Stability. 0] which tend to 0 as t → −∞ faster than t → eβ t .e.. The local unstable and stable manifolds are unique (in a certain sense) while local center manifolds are not. te ) → Rn which decay to 0 exponentially as t → −∞.e.424 DELAY DIFFERENTIAL EQUATIONS are locally positively invariant under the semiﬂow F . It is also possible to begin with a splitting of σ not at β = 0. Figure 10. The local unstable manifold W u consists of segments of solutions x : (−∞.. so that there are at least 3 eigenvalues in the open right halfplane. See [137]. Consider again eq. ∞) which remain in a neighbourhood of 0 and tend to 0 exponentially as t → ∞ belong to W s . Conversely. a locally positively invariant C 1 submanifold of C with T0 W3 = C3 which consists of segments of solutions on intervals (−∞. All solutions which start in W s decay to 0 exponentially as t → ∞. This means that there is a neighbourhood N of 0 so that ﬂowlines starting in N ∩ W ∗ remain in W ∗ as long as they stay in N . all such solutions with all segments in a neighbourhood of 0 in C have their segments in W u . In case all eigenvalues have negative real part. (1. and the segments of all solutions on [−h.8. i. All bounded solutions on the line which remain suﬃciently small have all their segments in W c .
the periodic orbit o is exponentially stable with asymptotic phase. an eigenvector in C is x˙ 0 = x˙ p = D1 F (p. which means in particular that there are a neighbourhood N ⊂ U of o. x0 )1. and let a periodic solution x : R → Rn and a period p > h of x be given. The number 1 is a Floquet multiplier.7 Floquet multipliers of periodic orbits Assume f is continuously diﬀerentiable. 1. Remark: If a global attractor exists then local unstable manifolds and their forward extensions by the semiﬂow belong to the global attractor. γ > 0. and if there are no other Floquet multipliers on the unit circle. constants k ≥ 0. x0 ). In case of hyperbolicity and if there are no Floquet multipliers outside the unit circle. The periodic orbit is called hyperbolic if 1 is an eigenvalue with 1dimensional generalized eigenspace. and a function θ : N → R so that for all φ ∈ N and all t ≥ 0. . If there are Floquet multipliers outside the unit circle then the periodic orbit is unstable.9. φ) − xt+θ(φ) ≤ k e−γ t . They can be used to describe the behaviour of ﬂowlines close to the periodic orbit o = {xt : 0 ≤ t ≤ p}.Dynamics of Delay Diﬀerential Equations 425 Figure 10. The associated Floquet multipliers are the nonzero points in the spectrum (of the complexiﬁcation) of the monodromy operator M = D2 F (p. F (t.
Without hyperbolicity. In the negative feedback case. the Floquet multipliers are squares of the eigenvalues of V . How can one obtain results about Floquet multipliers ? If µ = 0 and if g is odd and bounded with g (0) suﬃciently large then eq. v˙ 1 = λ−1 g (x(t))v2 . Then the segments v0 . Substitute v0 = λ−1 v2 in the last diﬀerential equation. with the additional symmetry x(t) = −x(t − 2).φ of the variational equation along x satisfy v˙ 2 = g (x(1 + t))v1 . one ﬁnds. v2 of the solution v = v x0 . v˙ 1 = g (t)v0 and v2 = λ v 0 .426 DELAY DIFFERENTIAL EQUATIONS If P is a Poincar´e return map on a transversal H = x0 + L and if x0 is a ﬁxed point of P then the spectral properties of DP (x0 ) : L → L and those of the monodromy operator are closely related. In the most general case. however. to the best of my knowledge. slowly oscillating periodic solutions of minimal period 4. v1 ) satisﬁes v˙ 2 = g (x(1 + t))v1 . 1 becomes an eigenvalue of DP (x0 ) with a realiﬁed generalized eigenspace of reduced dimension. due to the formula DP (x0 ) = prx0 ◦ M L. a characteristic equation is found in the following way: Suppose V φ = λφ. For the latter. It follows that the associated monodromy operator has V = D2 F (2. The simplest case is that M is compact and hyperbolicity holds. φ = 0.1) has periodic solutions with rational periods [120]. Then (v2 . v1 . Then the Floquet multipliers λ = 1 and their multiplicities coincide with the eigenvalues of DP (x0 ) and their multiplicities. x0 ) as a root. (1. among others. The compactness assumption can be relaxed. several natural questions concerning monodromy operators and Poincar´e maps seem not yet discussed in the literature.
more general characteristic equations for periodic solutions with arbitrary rational periods were developped. The latter were used in the ﬁrst proof that chaotic behaviour of slowly oscillating solution exists for certain smooth nonlinearities g of negative feedback type [153. and the nonlocal boundary conditions v2 (−1) = v1 (0). 154]. 60]. See also [59. In cases where period and delay are incommensurable characteristic equations for Floquet multipliers seem unknown.10. In recent work. which depends on the parameter λ through the fundamental matrix solution and through the boundary conditions. Using a fundamental matrix solution of the ODE system one ﬁnds that a certain determinant. the period in this case is 5/4.1) with µ > 0 and g representing positive feedback are hyperbolic. . A ﬁrst application was to show that for certain µ > 0 and special nonlinearities g of negative feedback type the orbits of some slowly oscillating periodic solutions with minimal period 3 are hyperbolic and stable [211]. ﬁrst in a proof of bifurcation from a continuum of period4solutions of a oneparameter family of delay diﬀerential equations [238]. This is the desired characteristic equation. Though there is no explicit formula and the dependence on the underlying periodic solution x seems rather involved it was possible to make use of the characteristic equation.Dynamics of Delay Diﬀerential Equations 427 Figure 10. v1 (−1) = (v0 (0) =)λ−1 v2 (0). It is planned to extend the method in order to establish that certain unstable rapidly oscillating periodic solutions of eq. (1. must vanish. and later in the search for hyperbolic stable [43] and unstable [115] period4solutions.
ws : Ns → E u ⊕ E c . wc : Nc → E u ⊕ E s . Then E is the direct sum of the associated realiﬁed generalized eigenspaces E u . Figure 10. on. E c . Assume dim E u < ∞. and stable linear spaces. and continuously diﬀerentiable maps wu : Nu → E c ⊕ E s . Suppose for simplicity that f (0) = 0. With respect to an equivalent norm on E the isomorphism on the unstable space is an expansion and the map on the stable space is a contraction.11.428 DELAY DIFFERENTIAL EQUATIONS Local invariant manifolds of continuously differentiable maps in Banach spaces We also need local invariant manifolds for continuously diﬀerentiable maps f : E ⊃ U → E on open subsets of Banach spaces (over R). . E c . i. respectively. Ns of 0 in E u . E s . so that the local unstable. There are open neighbourhoods Nu . of the unstable. The derivative Df (0) induces isomorphisms of the unstable and center spaces and an endomorphism of the stable space. center. center. dim E c < ∞. Nc . and stable manifolds W∗ = {x + w∗ (x) : x ∈ N∗ } are locally positively invariant with respect to f and have further properties analogous to the locally invariant manifolds for the semiﬂow introduced above.e. and that the (compact) spectrum of the derivative Df (0) is the disjoint union of its compact parts outside. with w∗ (0) = 0 and Dw∗ (0) = 0. respectively. E s . and inside the unit circle..
n=0 The proof employs a method due to Vanderbauwhede and van Gils [231]. Show that local unstable and stable manifolds of solution operators coincide with local unstable and stable manifolds for the semi . we have local invariant manifolds for the solution maps Ft at the ﬁxed point 0 as well as for the semiﬂow. See [103] for local unstable and stable manifolds. Wu consists of endpoints of trajectories (xj )0−∞ of f which tend to 0 as j → −∞. and wsc (Nsc ) ⊂ Nu so that the set Wsc = {z + wsc (z) : z ∈ Nsc } satisﬁes f (Wsc ∩ N ) ⊂ Wsc and ∞ + f −n (Nsc + Nu ) ⊂ Wsc . There exist convex open bounded neighbourhoods Nsc of 0 in E c ⊕ E s . Using the equivalent norm one can construct Wu and Ws in such a way that f induces a diﬀeomorphism from a convex open neighbourhood of 0 in Wu onto Wu and maps Ws into Ws . any trajectory (xj )0−∞ in some neighbourhood of 0 in E which for j → −∞ converges to 0 at a certain rate belongs to Wu . and a continuously diﬀerentiable map wsc : Nsc → E u with wsc (0) = 0. and there exist k ≥ 0 and q ∈ (0.are called hyperbolic if there is no spectrum on the unit circle. Exercise.Dynamics of Delay Diﬀerential Equations 429 For example. Nu of 0 in E u . The ﬁxed point 0 . 1) so that xj ≤ k q −j x0 along all such trajectories. N of 0 in U . Theorem. With respect to the new norm the inverse diﬀeomorphism and the map induced on Ws are contractions. Conversely. say φ0 = 0 for simplicity. and compare the results in the appendices in [137].and the linear map Df (0) . In case of a semiﬂow F as before and a stationary point. which to my knowledge and surprise was not available elsewhere in the literature. The latter contain a proof of the following natural result on local centerstable manifolds. Then E c = {0}. Dwsc (0) = 0.
not the iterated vectors themselves. but not necessarily unstable spaces into unstable spaces. Hyperbolic sets. E) and constants k ≥ 1. E = Df (x)Eux + Esf (x) . Concerning subsequent unstable spaces it is only required that the projections of transported vectors grow. . shadowing. There exist center manifolds Wc for solution maps which are not locally positively invariant under the semiﬂow [133]. and can be veriﬁed for certain Poincar´e return maps which do not have hyperbolic sets in the classical sense. 1) so that the unstable spaces Eux = pr(x) and the stable spaces Esx = (id − pr)(x). The situation for center manifolds is more subtle. Consider an open set U ⊂ E and a continuously diﬀerentiable map f : U → E. respectively. and chaotic trajectories of continuously differentiable maps in Banach spaces Let E denote a real Banach space.12. This is weaker than in the classical deﬁnition for diﬀeomorphisms in ﬁnitedimensional spaces. q ∈ (0. Figure 10. pr(f n (x)) ◦ Df n (x)v ≥ k −1 q −n v for all n ∈ N. Df n (x)v ≤ k q n v for all n ∈ N. So the derivative maps stable spaces into stable spaces.430 DELAY DIFFERENTIAL EQUATIONS ﬂow. v ∈ Esx . 220] and [149]. k. See [219. v ∈ Eux . x ∈ H have the following properties: Df (x)Esx ⊂ Esf (x) . The triple (pr. A hyperbolic set H of f is a positively invariant subset of U together with a uniformly continuous and bounded projectionvalued map pr : H → Lc (E. q) is called hyperbolic structure.
. Suppose f has a hyperbolic ﬁxed point p with local unstable manifold Wu and local stable manifold Ws . . yn+1 − f (yn ) < δ for all integers. It asserts that under a weak additional smoothness condition on the map f .e. See [219. Figure 10. we have xm ∈ Wu . provided the hyperbolic set H is rich enough for this. Very roughly. Assume also the transversality condition that there exists n0 ∈ N so that for all integers m < −n0 and n > n0 .Dynamics of Delay Diﬀerential Equations 431 For our hyperbolic sets the Shadowing Lemma remains valid in full generality. is accompanied by a unique trajectory (wn )∞ −∞ of f . xn ∈ Ws . the Shadowing Lemma makes it possible to establish complicated trajectory behaviour by prescribing complicated pseudotrajectories. 149]. in the diﬀeomorphism case) for which one can ﬁnd a hyperbolic structure is the following. i. and Df n−m (xm )Txm Wu is a closed complement of Txm Ws . A fundamental scenario going back to Poincar´e (1899. wn − yn ≤ for all integers.13. The shadowing trajectory does in general not belong to H. for a given hyperbolic set H and for every suﬃciently small > 0 there exists δ > 0 so that each δpseudotrajectory (yn )∞ −∞ in H. x0 = p. xn → p as n → ∞. and there is a homoclinic trajectory (xn )∞ −∞ .
Consider the space of symbols ΣJ = {1.g.15. e. for chaotic interval maps. . This occurs. Under these conditions the homoclinic loop HL = {p} ∪ {xn : n ∈ Z} is a hyperbolic set [220]. By means of shadowing one can then prove that chaotic trajectories close to the homoclinic loop HL exist. Figure 10. including elementary facts from symbolic dynamics. Let J ∈ N. By the way.432 DELAY DIFFERENTIAL EQUATIONS Figure 10. . The precise statement of the result requires preparations. it is not excluded that eventually xn = p.14. . .. J} with the discrete topology. The product space SJ = (ΣJ )Z .
. . . is a homeomorphism which conjugates the shifts σf V and σJM : α ◦ σf V = σJM ◦ α. As f may not be invertible. wn ∈ V0 and wn+1 ∈ / V0 imply wn+1 ∈ V1 . M }. . . . an = j if and only if wn ∈ Vj . and for every (wn )n∈Z ∈ Tf V . J − 1} imply wn+1 ∈ Vj+1 . . . the Bernoulli shift induces a homeomorphism which still has dense trajectories and inﬁnitely many periodic orbits. This diﬃculty is circumvented by the following construction. .Dynamics of Delay Diﬀerential Equations 433 of biinﬁnite sequences in the symbol space is compact. J. VJ of U with the following properties: p ∈ V0 . . Let M ∈ N be given and consider the compact subset SJM ⊂ SJ which consists of all symbol sequences composed of pieces of at least M consecutive symbols 0. This is a prototype of chaotic behaviour. . Tf V is a closed subset of V Z . Consider the set Tf V of complete trajectories (wn )n∈Z ∈ V Z of f in a given open subset V ⊂ U . wn ∈ VJ implies wn+µ ∈ V0 for all µ ∈ {1. On SJM . Theorem [220]. and the Bernoulli shift σJ : SJ → SJ given by σJ ((an )n∈Z ) = (an+1 )n∈Z is a homeomorphism with dense trajectories and inﬁnitely many periodic orbits. . the map σf V : Tf V (wn )n∈Z → (wn+1 )n∈Z ∈ Tf V is a homeomorphism. i. equipped with the product topology from V Z . The shift induced by f on its biinﬁnite trajectories in V . The map α : Tf V → SJM given by α((wn )n∈Z ) = (an )n∈Z . HL ⊂ 0 Vj = V . it can not be expected that on given subsets V ⊂ U the map f itself is conjugate to the homeomorphism σJM . alternating with pieces of the form 12 .e. .. wn ∈ Vj and j ∈ {1. There exist positive integers J and M ≥ 2 and mutually disjoint open EJ subsets V0 . . .
434 DELAY DIFFERENTIAL EQUATIONS Figure 10.16. .
. 0]. no matter how smooth g and r are. h] is assumed to be deﬁned on a subset U ⊂ C. for some h > 0. where r : U → [0. 0] → Rn is the evaluation map given by ev(φ. s)1 = φ˙ exists. Therefore the existence and smoothness results designed for IVPs of RFDEs on C are not applicable.8 435 Diﬀerential equations with statedependent delays Modelling sometimes leads to equations of the form x(t) ˙ = g(x(t − r(xt ))) with a map g : Rn ⊃ dom → Rn and a delay functional r which is deﬁned on some set of functions φ : [−h. . 0] → Rn . s) = φ(s). E. h]. the right hand side may also depend on x(t) and on further delayed values of x. The problem is that such f in general does not satisfy the previous smoothness hypotheses. A ’reason’ for this may be seen in the fact that the middle composite ev is not smooth: Lipschitz continuity of ev would imply Lipschitz continuity of elements φ ∈ C. Of course. and ev : C × [−h. . The diﬀerential equation with statedependent delay rewritten as a RFDE is x(t) ˙ = f (xt ) with f = g ◦ ev ◦ (id × (−r)).Dynamics of Delay Diﬀerential Equations 1.g. If C is replaced with the smaller Banach space C 1 = C 1 ([−h. Rn ) of continuously diﬀerentiable functions φ : [−h.Let us see why the associated IVP does not ﬁt into the framework presented before. with the usual norm given by ˙ φ1 = φ + φ. more complicated versions occur also. 0] → Rn and has values in [0. Diﬀerentiability would imply that D2 ev(φ.
for a reasonable class of RFDEs covering diﬀerential equations with statedependent delays. p : C 1 φ → φ(0) 1 satisﬁed on C . both continuously diﬀerentiable. 2. for g and r : U → [0. 249]. The maximal solution xφ : [−h. it seems necessary to restrict the state space further [247. s)χ = Ev(χ. Suppose this IVP is wellposed. with D1 Ev(φ. which in general is not satisﬁed on open subsets of the space C 1 . if we now look for solutions of the IVP with initial values in the open subset U ⊂ C 1 . U ⊂ C 1 open. and either g (ξ) < 0 for all ξ ∈ R (negative feedback) or g (ξ) > 0 for all ξ ∈ R . Monotone feedback: The structure of invariant sets and attractors Consider eq.1). At t = 0 we get ˙ φ(0) = x(0) ˙ = f (x0 ) = f (φ). So. with µ ≥ 0 and with a continuously diﬀerentiable function g : R → R which satisﬁes g(0) = 0. 0] → Rn is continuously diﬀerentiable. (Notice ˙ ∈ Rn the equation is trivially however that for f = p. h].and diﬀerentiable solution operators . s)1 = φ(s). (1. the resulting map f : U → Rn is continuously diﬀerentiable. Hence xφ itself would be continuously diﬀerentiable. x(t) ˙ = −µx(t) + g(x(t − 1)).) In order to obtain a continuous semiﬂow . a new diﬃculty arises. te (φ)) → Rn would have continuously diﬀerentiable segments. However. te (φ)) t → xφt ∈ C 1 would be continuous. s) ˙ and D2 Ev(φ.436 DELAY DIFFERENTIAL EQUATIONS then the smoothness problem disappears since the restricted evaluation map Ev : C 1 × [−h. What to do and what can be achieved will be the topic of the last of these lectures. and the ﬂowline [0.
If however µ ≥ 0 and −g (0) is suﬃciently large then the leading pair of eigenvalues is in the open right halfplane. e. It may happen that AS = {0}. and there exists a leading 2dimensional local unstable manifold W2 of the nonlinear eq. W2 consists of segments of bounded slowly oscillating solutions on the line. The strict monotonicity properties are essential for all results explained in this part. Recall also that S contains the 2dimensional leading realiﬁed generalized eigenspace C2 of the linearization at 0. Assume also that g is bounded from below or from above. AS is a proper subset of A. which contains all segments of slowly oscillating solutions and absorbs the ﬂowlines of eventually slowly oscillating solutions. In the nontrivial case AS = {0} the attractor AS has the following properties [240.1) which is tangent to C2 at 0. Example: µ = 0 and −1 < g .. while S ∩ C2< = ∅.g. and is therefore contained in AS .1 Negative feedback Recall the positively invariant wedge S = S ∪ {0} ⊂ C of data with at most one change of sign. Equations of this type arise in various applications.Dynamics of Delay Diﬀerential Equations 437 (positive feedback). . ∞) × C → C induces a continuous semiﬂow FS on the complete metric space S. in general. A compact global attractor A of F exists as well. 250]. Soon we shall use the decomposition C = C2 ⊕ C2< as a coordinate system. 2. which has a compact global attractor AS ⊂ S. There exists a map a : C2 ⊃ DS → C2< so that AS = {φ + a(φ) : φ ∈ DS }. (1. The semiﬂow F : [0. in neural network theory.
with solution maps continuously diﬀerentiable on the interior part {φ + a(φ) : φ ∈ DS } of AS . I.e. The periodic orbits are nested. and its boundary in C2 is the trace of a simple closed continuously diﬀerentiable curve. The semiﬂow FS induces a ﬂow on AS . All ﬂowlines of the ﬂow except the trivial one are given by slowly oscillating solutions on the line. the closure DS is homeomorphic to the compact unit disk in the Euclidean plane. or vice versa.438 DELAY DIFFERENTIAL EQUATIONS DS ⊂ C2 is an open subset of C2 containing 0. the restriction to the interior is continuously diﬀerentiable. one needs 0 = p2 (φ − ψ) for φ = ψ in AS . The fact that AS is a graph with respect the decomposition of C above means that the spectral projection p2 : C → C2 along C2< is injective on AS . The map a : DS → C2< is continuously diﬀerentiable. and the minimal periods are given by 3 successive zeros. or are heteroclinic from 0 to a periodic orbit. I.e... with the stationary point 0 in the interior of each periodic orbit. A tiny but basic ingredient of the proof is the following consideration. Figure 10. The zeros of the underlying periodic solutions are simple. The aperiodic orbits in AS \ {0} wind around the stationary point 0 and are heteroclinic from one periodic orbit to another.17. The manifold boundary bd AS = AS \ {φ + a(φ) : φ ∈ DS } = {φ + a(φ) : φ ∈ ∂DS } of AS is the orbit of a slowly oscillating periodic solution. and each boundary point has an open neighbourhood N in C2 so that the restriction to N ∩ int DS extends to a continuously diﬀerentiable map on N . .
A ﬁrst remark is that all solution maps Ft are injective. 242]. It is not diﬃcult to show that in case µ > 0 and g bounded the semiﬂow has a global attractor. The result in [177] yields that all ﬂowlines of F starting in the open dense set C \ R converge to the disk AS as t → ∞. The positively invariant set Σ = {φ ∈ C : (xφ )−1 (0) is unbounded} separates the domain of absorption into the interior of the positively invariant cone K from the domain of absorption into the interior of −K. Figure 10. The semiﬂow is monotone with respect to the ordering on C given by the cone K of nonnegative initial data.18. it is suﬃcient to have AS − AS ⊂ S ∪ {0} = S. 136. So. we can speak of the parts of C above and below Σ. in case µ = 0 every maximal . 2. In [239] explicit conditions on µ and g are given so that the manifold boundary is the only periodic orbit in AS .2 Positive feedback Let us describe in greater detail the results obtained in [137. and AS = A.Dynamics of Delay Diﬀerential Equations 439 As S does not intersect the nullspace C2< of p2 . One of the ﬁrst results in [137] is that the separatrix Σ is a Lipschitz graph over a closed hyperplane in C. Graph representations of stable manifolds associated with periodic orbits in AS have been obtained in [241]. However.
W3 is tangent to C3 at 0 and contained in W U . More structure appears when the 3 leading eigenvalues have positive real part. g (0). and all other eigenvalues have negative real part. ∞) × (W U ∩ B)). These are the closure of the unstable set W U of all segments of solutions on the line which decay to the stationary point 0 as t → −∞. This situation can be characterized by an explicit inequality which involves µ. Consider the decomposition C = C3 ⊕ C3< and the associated leading 3dimensional local unstable manifold W3 at 0. one being positive and one negative. The second additional assumption then is g (ξ − ) < µ and g (ξ + ) < µ. and a largest negative argument ξ − where g and µ id intersect. . What can be said about the structure of the closed invariant set W ? In case µ > 0 two mild additional assumptions are needed. The forward extension W = F ([0. The ﬁrst additional assumption is g(ξ) < µ for ξ outside a bounded neighbourhood of 0. then W U consists of 0 and of the segments of two solutions on the line. which implies that W U is contained in every neighbourhood of the compact set A. ξ This statement and the inequality before combined show that there exist a smallest positive argument ξ + where g and µ id intersect. and subsets thereof. yielding W U ⊂ A = A and W U ⊂ A. If µ and g (0) are such that only one eigenvalue of the generator of the semigroup given by the variational equation along the zero solution has positive real part. and thus a global attractor does not exist. ∞) × W3 ) of W3 is an invariant subset of W U . Notice that in case a global attractor A exists.440 DELAY DIFFERENTIAL EQUATIONS solution which starts in in the interior of K ∪ (−K) is unbounded. and elementary functions only. Assume that this inequality holds. W U ⊂ F ([t. So it is natural to look for substitutes of a global attractor which are present in all cases of interest. necessarily W U ⊂ A since for every ball B and for every t ≥ 0. They are related to the fact that the hypothesis about the eigenvalues implies g (0) > µ.
. and the restriction of w to int DW is continuously diﬀerentiable. The set W is invariant. and consists of the trace of a simple closed continuously diﬀerentiable curve and its interior. and continuous mappings w : DW → E and wΣ : DΣ → G1 ⊕ E such that W = {χ + w(χ) : χ ∈ DW }. and there is no other periodic orbit in W . a closed complementary space E of G3 in C. W ∩ Σ = {χ + wΣ (χ) : χ ∈ DΣ }. DW = ∂DW ∪ int DW . The map wΣ is continuously diﬀerentiable (in the sense explained above for the map a). About the shape of W and the dynamics in W the following has been proved. ξ + deﬁne hyperbolic semigroups. Figure 10. The domain DΣ is homeomorphic to the closed unit disk in the Euclidean plane. respectively. The set o = (W ∩ Σ) \ (W ∩ Σ) = {χ + wΣ (χ) : χ ∈ ∂DΣ } is a periodic orbit. and the semiﬂow deﬁnes a continuous ﬂow FW : R × W → W . with all associated eigenvalues in the open left halfplane. The restriction of FW to R × W is continuously diﬀerentiable. Also.Dynamics of Delay Diﬀerential Equations 441 It follows that the variational equations along the constant solutions given by ξ − . For W and for the part W ∩ Σ of the separatrix Σ in W there are graph representations: There exist subspaces G2 ⊂ G3 of C of dimensions 2 and 3. W = {χ + w(χ) : χ ∈ int DW }.19. a compact set DΣ ⊂ G2 and a closed set DW ⊂ G3 . The open annulus (W ∩ Σ) \ {0} consists of heteroclinic connections from 0 to o. a complementary space G1 of G2 in G3 .
The further properties of W and the ﬂow FW are diﬀerent in the cases µ > 0 and µ = 0. For every φ ∈ W . W is compact and contains the stationary points ξ− and ξ+ in C given by the values ξ − and ξ + . This means that the restriction to int DW is continuously diﬀerentiable. φ) → o as t → −∞. and that each boundary point except χ− .20. Consider χ− and χ+ given by ξ− = χ− + w(χ− ) and ξ+ = χ+ + w(χ+ ). φ) → ξ− as t → ∞. which send the manifold boundary bd W = W \ W = {χ + w(χ) : χ ∈ ∂DW } and ∂DW onto the unit sphere S 2 ⊂ R3 . Combining some of the results stated above. FW (t.442 DELAY DIFFERENTIAL EQUATIONS For every φ ∈ W . For µ > 0. The set ∂DW \ {χ− . There exist homeomorphisms from W and from DW onto the closed unit ball in R3 . The points φ ∈ W \ Σ above the separatrix Σ form a connected set and satisfy FW (t. ξ− ≤ φ ≤ ξ+ . . χ+ } is continuously diﬀerentiable. one obtains for µ > 0 the following picture: W is a smooth solid spindle which is split by an invariant disk in Σ into the basins of attraction towards the tips ξ− and ξ+ . and all φ ∈ W \ Σ below the separatrix Σ form a connected set and satisfy FW (t. FW (t. χ+ } is a 2dimensional continuously diﬀerentiable submanifold of G3 . φ) → ξ+ as t → ∞. and the restriction of w to DW \ {χ− . φ) → 0 as t → −∞. χ+ has an open neighbourhood N in G3 so that the restriction of w to N ∩ int DW extends to a continuously diﬀerentiable map on N . respectively. Finally. Figure 10. for every φ ∈ bd W diﬀerent from ξ− and ξ+ .
φ) → o as t → −∞. The ﬁrst steps exploit the monotonicity of the semiﬂow. There exist homeomorphisms from W and from DW onto the solid cylinder {z ∈ R3 : z12 + z22 ≤ 1} which send bd W and ∂DW onto the cylinder S 1 × R ⊂ R3 . The points φ ∈ W \ Σ above the separatrix Σ form a connected set and satisfy xφ (t) → ∞ as t → ∞. which go back to [171] and in a special case to [236. These tools are used to characterize . FW (t. Finally. which was introduced by MalletParet [171]. given by monotone solutions x : R → R. A powerful tool for the investigation of ﬁner structures is a version of the discrete Lyapunov functional V : C \ {0} → N0 ∪ {∞} counting sign changes of data φ ∈ C. Among others we obtain that in case µ > 0 the set W is contained in the order interval between the stationary points ξ− and ξ+ . Related are apriori estimates for the growth and decay of solutions with segments in level sets and sublevel sets of V . and the points φ ∈ W \ Σ below Σ form a connected set and satisfy xφ (t) → −∞ as t → ∞. and that there are heteroclinic connections from 0 to ξ− and ξ+ . 237]. We turn to aspects of the proof.Dynamics of Delay Diﬀerential Equations 443 In case µ = 0 the sets W and DW are unbounded.21. for every φ ∈ bd W . The boundary ∂DW is a 2dimensional continuously diﬀerentiable submanifold of G3 . and w is continuously diﬀerentiable. Figure 10.
. become obvious. and that this restriction is continuously diﬀerentiable. and a map from a subset of the realiﬁed 2dimensional generalized eigenspace L of the complex conjugate pair of eigenvalues next to λ into the complementary space C1 ⊕ C3< which represents W ∩ Σ. and FW is continuously diﬀerentiable on the manifold R × W . Instead. The last facts make it possible to introduce global coordinates on W and W ∩ Σ. It is not hard to deduce that W is given by the restriction of w to an open set. the semiﬂow extends to a ﬂow FW : R × W → W . c(φ))tr and 1 c(φ)eλ t . 2}) and V −1 (2). φ(−1). and injectivity on W for a continuous linear evaluation map Π3 : C → R3 . turn out to be locally Lipschitz continuous. Our attempts to obtain the second map failed. Moreover.444 DELAY DIFFERENTIAL EQUATIONS the invariant sets W \ {0} and (W ∩ Σ) \ {0} as the sets of segments xt of solutions x : R → R which decay to 0 as t → −∞ and satisfy V (xt ) ≤ 2 for all t ∈ R and V (xt ) = 2 for all t ∈ R. On W . It is not diﬃcult to show injectivity for the continuous linear evaluation map Π2 : C φ → (φ(0). C3 = C1 ⊕ L as a framework for graph representations. Guided by the results on negative feedback equations it seems natural to expect that there exists a map from a subset of C3 = T0 W3 into C3< which represents W . G1 ⊂ G3 . however. respectively. respectively. respectively. The map Π3 is given by Π3 φ = (φ(0). φ(−1))tr ∈ R2 on W ∩ Σ. The inverse maps of the restrictions of Π2 and Π3 to W ∩Σ and W . so that representations by maps w and wΣ with domains in G3 and G2 and ranges in complements E of G3 in C and E ⊕ G1 of G2 in C. 1+ where p1 is the spectral projection onto the realiﬁed onedimensional generalized eigenspace C1 of the leading eigenvalue λ associated with the linearization at 0. (p1 φ)(t) = g (0)e−λ The next step leads to the desired graph representation. R3 ⊃ Π3 W and R2 ⊃ Π2 (W ∩ Σ) are embedded in a simple way as subspaces G3 ⊃ G2 into C. A map of the ﬁrst kind had been constructed for W by Ammar in case µ = 0 [11]. nontrivial diﬀerences of segments in W and W ∩ Σ belong to V −1 ({0. Therefore we abandoned the decomposition C = C3 ⊕ C3< .
We consider a transversal Y of o at some point p0 ∈ o and a Poincar´e return map with domain in Y and ﬁxed point p0 . It is shown that pieces of W ∩ Σ in Y are open sets in the transversal intersection of W with a local centerstable manifold of the Poincar´e return map. in the level set V −1 (2). We use the fact that there is a heteroclinic ﬂowline in W ∩ Σ from the stationary point 0 to the orbit o. The study of the linearized stability of the periodic orbit o is closely related to earlier work in [153] and to apriori results on Floquet multipliers and eigenspaces for general monotone cyclic feedback systems with delay which are due to MalletParet and Sell [176]. A ﬁrst idea how to show that the graph W ∩ Σ ⊂ V −1 (2) ∪ {0} is continuously diﬀerentiable might be to consider the family of 2dimensional local invariant submanifolds with tangent space L at the stationary point 0 ∈ C. and to look for a member formed by heteroclinics connecting 0 with the periodic orbit o.e. Diﬀerentiability and continuity of the derivative at 0 .Dynamics of Delay Diﬀerential Equations 445 Phase plane techniques apply to the coordinate curves in R2 (or G2 ) which correspond to ﬂowlines of FW in the invariant set W ∩ Σ. is at most 2dimensional. Figure 10. The approach in [137] is quite diﬀerent. in order to show that precisely one Floquet multiplier lies outside the unit circle. It follows that W ∩ Σ is given by the restriction of wΣ to an open subset of G2 .22. It also follows that the center space of the monodromy operator M = D2 F (ω. as well as the identiﬁcation of Π2 (W ∩ Σ) with the interior of the trace Π2 o. and yield the periodic orbit o = (W ∩ Σ) \ (W ∩ Σ). The investigation of the smoothness of the part W ∩ Σ of the separatrix Σ in W and of the manifold boundary bd W = W \ W begins with a study of the stability of the periodic orbit o. p0 ∈ o and ω > 0 the minimal period. p0 ).. Then we use the ﬂow FW to obtain continuous diﬀerentiability of the set (W ∩ Σ) \ {0}. i.
ξ+ } as forward extension of a local un . the manifold boundary without the stationary points ξ− . The next step achieves the continuous diﬀerentiability of a piece of W in a transversal H of the periodic orbit o. In case there is a forward trajectory in Wcs \ Wc which converges to the ﬁxed point we construct positively invariant open subsets of Wcs . in the sense stated before. For µ = 0 the full manifold boundary has the same property. ξ+ } and W ∩ Σ are continuously diﬀerentiable. Figure 10. The long proof of these facts involves the charts Π2 and Π3 and uses most of the results obtained before. coincides with the forward semiﬂow extension of a local unstable manifold of the period map Fω at a ﬁxed point p0 ∈ o.. Having established continuous diﬀerentiability of W ∩ Σ it is then shown in case µ > 0 that the set bd W \ {ξ− .446 DELAY DIFFERENTIAL EQUATIONS and the relation T0 (W ∩ Σ) = L follow by other arguments which involve apriori estimates and an inclination lemma. In case µ > 0 the identiﬁcation of bd W \ {ξ− .e. Using the ﬂow FW we then derive that W \ {ξ− . ξ+ }. ξ+ . We construct a continuously diﬀerentiable graph over an open set in a plane X12 ⊂ H which extends such a piece of W ∩ H close to a point p0 ∈ o ∩ H beyond the boundary. The ﬁnal steps lead to the topological description of W . A technical detail of the approach just described concerns centerstable manifolds Wcs and onedimensional center manifolds Wc at ﬁxed points of continuously diﬀerentiable maps. i.23.
Dynamics of Delay Diﬀerential Equations Figure 10.24.25. 447 . Figure 10.
Analogously L+ is the realiﬁed 2dimensional generalized eigenspace of the next pair of eigenvalues. Let C1+ denote the leading realiﬁed onedimensional generalized eigenspace associated with the variational equation along the solution R t → ξ + ∈ R. and involves subsets of boundaries of neighbourhoods of 0. 0] → C. We point out that the proof of this topological property relies on the smoothness of the set W \ {ξ− . The inclusion W ⊂ ξ+ − K implies that the tangent cone T0 W is the singleton {0}. Then a generalization due to Bing [29] of the Schoenﬂiess theorem [210] from planar topology is employed to obtain homeomorphisms from W onto the closed unit ball in R3 . and C3<+ is given by the remaining eigenvalues.448 DELAY DIFFERENTIAL EQUATIONS stable manifold is used to deﬁne homeomorphisms from bd W onto the unit sphere S 2 ⊂ R3 . In order to construct these smooth boundaries we have to go back to the variationofconstants formula for RFDEs in the framework of sundual and sunstar dual semigroups [57]. The application of Bing’s theorem requires to identify the bounded component of the complement of the set Π3 (bd W ) ∼ = S2 in R3 as the set Π3 W . and to verify that Π3 W is uniformly locally 1connected. in bd W ending at ξ+ . In case µ = 0 the construction of the desired homeomorphism from W onto the solid cylinder is diﬀerent but uses Bing’s theorem as well. ξ+ in W which are transversal to the ﬂow FW . In [242] the singularity ξ+ is described in terms of the outward tangent cone T+ bd W . We have C1+ = R · η+ with η+ ∈ C a segment of the associated positive decreasing exponential solution of the variational equation. ξ+ }. This means that for every > 0 there exists δ > 0 so that every closed curve in a subset of Π3 W with diameter less than δ can be continuously deformed to a point in a subset of Π3 W with diameter less than . deﬁned as the set of tangent vectors c (0) ∈ C of continuously diﬀerentiable curves c : (−. What can be said about W close to the attracting stationary points ξ− . > 0. therefore W and bd W are not smooth at ξ+ . ξ− . ξ+ ? Consider ξ+ . .
One may ask for diﬀeomorphisms instead of homeomorphisms from DW onto the solid cylinder in case µ = 0. . as φ tends to ξ+ . might be relaxed. χ+ } onto the unit ball without its north and south poles in case µ > 0. Convergence of ﬁnitedimensional subspaces of C is here deﬁned by the gap topology. Let us state some open problems. the periodic orbit o is the only one of the semiﬂow. Recall the question whether W ∩Σ can be represented by a smooth map from a subset of L into the complementary space C1 ⊕ C3< . which is given by the Hausdorﬀ distance of the compact intersections of two such spaces with the unit sphere. The equations covered include examples from neural network theory. ξ+ hyperbolic.Dynamics of Delay Diﬀerential Equations 449 and in terms of the set LT P of limit points of tangent planes Tφ bd W . The assumption g (ξ − ) < µ.26. The ﬁrst and basic ingredient of the proof is the uniqueness result. g (ξ + ) < µ. We prove T+ bd W = [0. Figure 10. and from DW \ {χ− . For narrower classes of nonlinearities g and for the parameters µ > 0 and g (0) > 0 in a certain smaller range than before. which makes the semigroups associated with the stationary points ξ− . φ = ξ+ . The proofs employ apriori estimates and inclination lemmas. ∞) · η+ and show that LT P coincides with the set of planes in C1+ ⊕ L+ containing C1+ . and the set W is in fact the global attractor of F [136].
Also connections from periodic orbits to 0 become possible. If A is strictly larger than A≤2 then we have to expect a ﬁnite number of analogues of the smooth disk A2 . forming the manifold boundary of A2 . from periodic orbits and 0 to ξ− and ξ+ .450 DELAY DIFFERENTIAL EQUATIONS How is the global attractor organized in case W is a proper subset? We add a few remarks about the last question. given by higher even values of V . a smooth disklike submanifold A2 in A which extends W ∩ Σ beyond o and contains at least one additional periodic orbit. Suppose now that A is not conﬁned to the order interval {φ ∈ C : ξ− ≤ φ ≤ ξ+ }. a section of A≤2 containing ξ− and ξ+ might have the structure shown by a sliced onion. and a more complicated variety of heteroclinic connections. from 0 to periodic orbits and to ξ− and ξ+ . ξ+ }.27. together with the stationary point 0. A Morse decomposition of A similar to the one constructed by MalletParet [171] should be useful to describe a part of these heteroclinic connections. Let us ﬁrst consider situations where A is a subset of the order interval {φ ∈ C : ξ− ≤ φ ≤ ξ+ }. These submanifolds should subdivide the 3dimensional subset A≤2 of A formed by 0 and the segments of all solutions x : R → R with ξ − ≤ x(t) ≤ ξ + and V (xt ) ≤ 2 for all t ∈ R into invariant layers. The segments of the solutions x : R → R with V (xt ) = 2 for all t ∈ R may form. the unstable sets of the periodic orbits in A2 should be analogues of bd W \ {ξ− . It may be the case that A = A≤2 . between periodic orbits in diﬀerent disks. namely 2dimensional invariant submanifolds given by heteroclinic connections from the periodic orbit to ξ− and ξ+ . and there are zeros ξ ∗ of g − µ id below ξ − and above ξ + with . between periodic orbits in the same disk. In this case. Figure 10.
3. for functions g which satisfy the negative feedback condition. We may assume y(−1) = 0 < y(−1). which yields existence of chaotic solution behaviour for eq. One out of many possibilities is that the part of A in a certain neighbourhood of a stationary point ξ∗ ∈ C given by such a value ξ ∗ looks just as we began to sketch it for the case A ⊂ {φ ∈ C : ξ− ≤ φ ≤ ξ+ }. ˙ The linear hyperplane H = {φ ∈ C : φ(−1) = 0} . The result in [154] was not the ﬁrst one about existence of chaotic solutions of delay diﬀerential equations. but the ﬁrst concerning slowly oscillating solutions of equations with smooth nonlinearities. Chaotic motion This part describes steps of the proof in [153. x(t) ˙ = g(x(t − 1)). The nonlinearities considered have extrema.1) with µ = 0. Recall that monotonicity implies that the attractor AS of almost all solutions is a disk. Figure 10. It follows that the period map F4 has onedimensional local center and unstable manifolds at its ﬁxed point y0 = η. which excludes complicated behaviour of slowly oscillating solutions. whose orbit o ⊂ C is unstable and hyperbolic.Dynamics of Delay Diﬀerential Equations 451 g (ξ ∗ ) > µ. (1.28. 154]. This multiplier is real and simple. Nevertheless the proof begins with odd functions g which satisfy g (ξ) < 0 for all ξ ∈ R. with exactly one Floquet multiplier λ outside the unit circle [115]. For a function of this type there is a slowly oscillating periodic solution y : R → R with the symmetry y(t) = −y(t − 2) and minimal period 4.
y is increasing on the initial interval. in a carefully chosen convex open neighbourhood N ⊂ H of η on which the original map P is also a Poincar´e return map with respect to the modiﬁed delay diﬀerential equation.452 DELAY DIFFERENTIAL EQUATIONS is transversal to o at η. z0 ∈ This is done as follows. −δ + 1] as a solution z ∗ : R → R of the new equation. 0]. but aﬀects the forward continuation of z(−∞. These modiﬁcations look rather complicated and have several extrema: 3 local maxima and 4 local minima. with a nonlinearity g ∗ of negative feedback type which coincides with g in a neighbourhood of the interval y(R). ∞) preserves y and the restriction of z to the interval (−∞. whose trajectories all translate into slowly oscillating solutions of a delay diﬀerential equation as before. and P (N ∩ Ws ) ⊂ N ∩ Ws . The key to all is now a simple observation: Any modiﬁcation g ∗ of g in the interval (y(0) + . For an associated Poincar´e return map the ﬁxed point η is hyperbolic with onedimensional local unstable manifold Wu and local stable manifold Ws of codimension 1 in H. Information about the eigenspace of the Floquet multiplier λ reveals that Wu contains the segment z0 of a solution z : R → R of the original equation with z(0) > y(0) so that for some > 0 and δ > 0. Recall the abstract result for maps in Banach spaces which guarantees chaotic trajectories in a neighbourhood of a homoclinic loop. i.with / o. Modiﬁcations g ∗ are designed in such a way that the curves t → zt∗ intersect the hyperplane H close to η at some t∗ > −δ + 1. z(t) < y(0) + on (−∞. z(t) > y(0) + on (−δ. The plan is to verify the hypotheses of this result for a modiﬁcation of the map P . .e. and y(0) = η = max y. zt → o as t → ∞. −δ).. −δ + 1] as solutions. The ﬁrst major step is to ﬁnd such functions g ∗ so that there is a solution z : R → R which is homoclinic with respect to o. Intersections can be achieved above and below the local stable manifold Ws .
z ∗ (t) = z(t) on (−∞. t0 ]. Figure 10. Two such modiﬁcations g ∗ are connected by a oneparameterfamily. Continuity yields a member so that the intersection of the curve t → zt∗ with H is on Ws . Redeﬁne P on int B ∗ using the previous intersection map.29. Choose a small closed ball B ∗ ⊂ N P (zt∗0 ) ∈ / B ∗ for j < 0 and so that one has an centered at zt∗0 = zt0 = ζ.Dynamics of Delay Diﬀerential Equations 453 Figure 10.30. Now the map P is modiﬁed. and keep P on N \ B ∗ un . with zt∗j ∈ intersection map on B ∗ which has the value t∗ − t0 at ζ. The neighbourhood N has the property that it contains a backward trajectory (zt∗j )0−∞ ∈ Wu of P so that / N . and the desired nonlinearity and homoclinic solution are found.
Then the ﬁxed point η and the points zt∗j . ’Suitable’ means that the criterion should be suﬃciently explicit in terms of the nonlinearity g ∗ and the periodic solution y so that it can actually be applied. and that solutions v : [−1. zt∗∗ .31. The monotonicity of g on the range of y implies that the set S of data with at most one change of sign is again positively invariant. This freedom is needed for the remaining part of the proof. The ﬁrst step towards the transversality criterion concerns the variational equation along y. j ∈ N. P j (zt∗∗ ). η = y0 . which is to verify the transversality condition.454 DELAY DIFFERENTIAL EQUATIONS changed. In order to verify the direct sum decomposition of H we therefore need a suitable criterion for a vector in H not to be tangent to the local stable manifold Ws . j ≤ 0. with D(P ∗ )n−m (φm )Tφm Wu ⊕ Tφm Ws = H. the points φm and φn of the homoclinic trajectory of P ∗ belong to Wu and Ws . respectively. that there exists n0 ∈ N so that for all integers m < −n0 and n > n0 . In our situation. v(t) ˙ = g (y(t − 1))v(t − 1) and the monodromy operator M = DF4 (η). ∞) → R starting in S are eventually slowly oscillating. form a homoclinic loop of the new map P ∗ . Figure 10. The previous construction is in fact carried out for families of oneparameterfamilies of nonlinearities. The realiﬁed generalized eigenspace C≥ of the 2 leading Floquet multipliers on and . namely. codim Ws = 1.
Dynamics of Delay Diﬀerential Equations 455 outside the unit circle is contained in S. χ ∈ C \ Tφ Ws (η) if and only if v φ. ∞) → R. So. ∞) t → a x˙ φ (t) + b v φ. which is satisﬁed by v φ. χ ∈ H \ Tφ W s if and only if for every pair of reals (a. ∞) → R. the function (0. 0). but not yet for the right map and only at a particular point of the local stable manifold. in other words. dim C≥ = 2. is that for any φ ∈ Ws and any χ ∈ H. In a next step it is shown that for any φ ∈ Ws (η) and χ ∈ C. . With regard to the proof. i.The main tool used in the proof is an inclination lemma in a nonhyperbolic situation. and if a solution v is eventually slowly oscillating then χ = v0 ∈ C \ Z. Initial data in Z deﬁne solutions v : [−1. and the asymptotic phase for ﬂowlines starting in Ws . The ﬁnal form of the criterion.The proof exploits the relation between derivatives of the Poincar´e map and the period map. The last part of the proof that chaotic motion exists is the search for a subclass of nonlinearities g ∗ so that for all nonzero χ ∈ D(P ∗ )n (ζ)Tζ Wu . This yields that the coeﬃcient function in the variational equation along xφ : [−1. . the initial value is transversal to Z. b) = (0.χ is eventually slowly oscillating. Here we have a ﬁrst suﬃcient condition for transversality.χ (t) ∈ R is eventually slowly oscillating.. Recall that the local stable manifold Ws (η) of the period map F4 at η satisﬁes Tη Ws (η) = Z. ∞) → R with segments in C \S.χ . for the Poincar´e map P . Let Z denote the complementary subspace given by the remaining spectrum of M .e. Notice that for t > 1 the function considered here satisﬁes the variational equation along the solution xφ : [−1. approaches a shifted copy of the coeﬃcient function in the variational equation along the periodic solution y. the linear stable space of M . notice that for φ close to η the spaces Tφ Ws (η) and Z are close. Z ∩ S = ∅. Flowlines starting in Ws (η) converge to o as t → ∞ with asymptotic phase.
1). It has not yet been proved that the MackeyGlass and LasotaWazewskaCzyzewska equations generate chaotic ﬂowlines. A reason for this is that g ∗ was chosen in such a way that the curve t → zt∗ through a point in Wu close to η is driven far away from o and then returns to Ws within a short interval of length approximately 3. The nonlinearities for which the proof just described works look much more complicated than the unimodal functions in versions of eq. As in part 1. in a ’Cantor dust’. Let us mention some open problems. thin set in state space. whereas numerical experiments indicate that many if not almost all solutions are complicated for a variety of nonlinearities. (1. 4.1) with µ > 0 which are known from applications. (1. In a more recent paper LaniWayda was able to modify the method and obtain a result for nicerlooking unimodal nonlinearities [150]. Poincar´eBendixson theorems). and by more geometric arguments (planar dynamics in the attractor of almost all solutions. by related index and continuation arguments. notably by means of the ejective ﬁxed point principle.456 DELAY DIFFERENTIAL EQUATIONS and for all suﬃciently large integers n the suﬃcient condition in the previous ’transversalitybyoscillation’ criterion is fulﬁlled. by Schauder’s theorem directly. The result presented here establishes chaotic dynamics only in a small. These are equations of the form (1. Stable periodic orbits In many cases the dynamics of nonlinear autonomous delay diﬀerential equations is structured by periodic orbits. Much less could be achieved concerning stability and uniqueness of periodic orbits. Existence of periodic orbits has been shown by several methods. for µ ≥ 0 and nonlinearities g : R → R of negative feedback type which are continuous or better. Other recent results on chaotic solutions of delay diﬀerential equations are due to LaniWayda [151] and LaniWayda and Szrednicki [152]. Consider eq. the method of steps yields maximal solutions x : [−1.1) with µ > 0 and unimodal analytic nonlinearities. x(t) ˙ = −µ x(t) + g(x(t − 1)). ∞) → R of the associated IVP and a continuous .
∞). xp ≥ β and x(p) = β.32. i. The segment xp is a ﬁxed point of the solution map Fp and deﬁnes a periodic solution which has period p and is slowly oscillating. It is then easy to design nonlinearities g as above so that all segments xt of solutions x = xφ starting at φ ≥ β with φ(0) = β return after some time p > 0. . 2. β > 0 and if initial data φj ∈ C. 1].Dynamics of Delay Diﬀerential Equations 457 semiﬂow F . Moreover. with φj (t) ≥ β coincide at t = 0 then the method of steps shows immediately that the corresponding solutions coincide on [0. Recall that. there are special equations for which it is extremely easy to ﬁnd slowly oscillating periodic solutions whose orbits in C seem very stable and attractive. The underlying observation for this is the following. They depend only on the values g(β) and φ(0).. j = 1. in some contrast to the preceding remarks. Figure 10. for every ψ ∈ C close to φ = xp the solution xψ has segments which after some time merge into the periodic orbit o = {xt : t ∈ R}. and consequently for all t ≥ 0. If g is constant on some ray [β.e.
0]. and the little information for nonlinearities which are close to these but not necessarily constant on any interval. ). (1. ∞). ) of continuous odd functions g : R → [−b. b] which satisfy −a − < g(ξ) < −a + for all ξ ≥ β. β] the function g is close to −a sign and a sign. Proposition. β). ) is steep on (0. respectively.458 DELAY DIFFERENTIAL EQUATIONS This expresses extremely strong attraction towards the periodic orbit. The general results on existence mentioned at the beginning may give existence. q) and reaches the level −β at t = q.g = xφ of eq. but nothing beyond. When convenient the associated semiﬂow will be denoted by Fg . detailed results in special cases as before. The method of steps yields sharp estimates of the solutions also on the next intervals of unit length. derivatives can be estimated with . however. For β > 0 and > 0 consider nonlinearities in the set N (β. Lipschitz continuity always played a role. and close to −a sign on [β.1) has not been exploited. g ∈ N (β. The important property is Lipschitz continuity. What has just been said is strictly limited to equations where the nonlinearity g is constant on some interval. 1] solutions starting in A(β) lie in a narrow tongue between 2 decreasing exponentials. On A(β) a return map can be found for β and suﬃciently small. there exists q = qg (φ) > w + 1 so that xφ. Initial data will be taken from the closed convex sets A(β) = {φ ∈ C : φ(t) ≥ β on [−1. Poincar´e return maps on transversals to o are locally constant. small and for g ∈ N (β. the derivative at the ﬁxed point on o is zero. Furthermore.g increases on (w + 1. (1) is strictly below −β on [w. Moreover. 1) such that for β. There exists w ∈ (0. Fix µ > 0. φ(0) = β}. w + 1]. For β and small. (In case of related equations with statedependent delay. ∞) and on (−∞. The proof relies on the fact that on [β. φ ∈ A(β) the solution xφ. which in earlier work on eq. its spectrum is the singleton {0} and the periodic orbit is hyperbolic. Using this and the variationofconstants formula one sees that on (0. b > a > 0. What happens if g is only close to such a function but not constant on any nontrivial interval? The approach sketched above fails. In the sequel a rather elementary method from [243] is presented which in a sense ﬁlls the gap between the easily accessible.).
33.34. Figure 10.Dynamics of Delay Diﬀerential Equations Figure 10. 459 .
(1) has been used. our objective is more than mere existence. (Observe that for all φ ∈ A(β) whose solutions coincide on [w. ) contains Lipschitz continuous functions g so that Lβ (g) is arbitrarily small.1+w (φ). (1) with the symmetry x(t) = −x(t + qg (φ)) for all t ∈ R. w + 1] the argument qg (φ) is the same . The return map R = Rg : A(β) → A(β) is then given by R(φ) = −Fg (qg (φ). ψ) ∈ A(β) where s = sg (ψ) is deﬁned by the equation w + 1 + s = qg (φ) for all φ ∈ A(β) with ψ = Fg. due to Schauder’s theorem. w + 1].1 to A(β). .w to Fg.1 (A(β)). and minimal period p = 2q. φ). ).1+w (A(β)) ψ → −Fg (sg (ψ).it depends only on the behaviour of the solutions on [w. This is facilitated by a decomposition of R = Rg into the restriction of Fg. We derive Lipschitz estimates of the composites of the return map Rg for g ∈ N (β. Notice that up to here only continuity of the nonlinearities in eq. Notice that for each g ∈ N (β.460 DELAY DIFFERENTIAL EQUATIONS suﬃcient accuracy. L(g) ≥ a− β is large for β > 0 small.) The Lipschitz constant L(Rg ) will be estimated by an expression which involves the Lipschitz constant L(g) of g and the Lipschitz constant Lβ (g) = L(g[β. The next objective is to ﬁnd an upper estimate of the Lipschitz constant L(R) = sup φ=ψ R(φ) − R(ψ) ≤∞ φ − ψ of R. followed by the restriction of Fg. ∞)). However. Fixed points exist. On the other hand. each N (β. ). This will be exploited later. with β and suﬃciently small. Notice that g being odd any ﬁxed point φ of Rg deﬁnes a slowly oscillating periodic solution x : R → R of eq. and ﬁnally followed by the map Fg.
and arguments as in the proof of the estimate of L(Fg. ∞).1 (ψ)(t) = xφ. L(Fg. β = xφ.w Fg. 1 + w + sg (ψ)]. and traverse the interval [−β. µ The proof uses the triangle inequality. The resulting equation for sg (ψ) and for some sg (ψ). The diﬀerence to the previous proposition is due to the fact that arguments of Fg. Proposition. are subtracted from each other.g (1 + t) − xψ.Dynamics of Delay Diﬀerential Equations 461 Proposition.1+w (A(β))) ≤ L(sg ) · [µ (1 − a (1 − e−µ )) + max{b. φ ∈ A(β).w+1 (φ) with φ ∈ A(β) there is an equation deﬁning qg (φ) and sg (ψ).w may diﬀer at t = 0.g (1 + w + sg (ψ)). Proposition. ·)Fg. The last term is evaluated by the variationofconstants formula on [1 + w.1+w (φ). ψ in A(β) the variationofconstants formula says Fg.1 (φ)(t) − Fg. 0 and that the arguments of g belong to [β. a + }] + 1 + Lβ (g). L(Fg (sg (·). the preceding proposition.g (qg (φ)) = xφ.1 A(β)) ≤ Lβ (g). β] where g is steep. Proposition.1 (A(β))) ≤ 1 + w L(g).g (1 + t) =0−0+ 1+t e−µ(1+t−s) (g(φ(s − 1)) − g(ψ(s − 1)))ds. L(sg ) ≤ 1 + eµ Lβ (g) a − − β µ eµ The proof begins with the fact that for ψ = Fg. ψ = Fg. Further elementary manipulations lead to the desired estimate.w Fg. One time unit later large deviations of one solution from another become possible. L(Fg. .1 (A(β)). The proof is almost obvious from the facts that for φ. In the same way one ﬁnds the next result. Next we need an estimate of the Lipschitz constant of the intersection map sg .
β] where g is monotone and steep. small. . not necessarily monotone. ·) . So this term does not become large for suitable g ∈ N (β. which are steep close to 0 and ﬂat away from 0. γ > 0 large. β). The improved Lipschitz estimates yield contracting return maps (and hyperbolic stable periodic orbits) for the scaled hyperbolic tangens above. and which arise in applications. β] → [−b. Remarks. This involves sharp estimates of the divergent behaviour of ﬂowlines one time unit after the underlying solutions traverse the interval [−β. e. ). and may not be diﬀerentiable everywhere.. The nonlinearities g for which Rg is a contraction are not necessarily constant on any interval. b] which satisﬁes −a − < g(β) < −a + . ∞) and Lβ (g) so small that the bound for L(Rg ) is less than 1. and for other nonlinearities. The orbits of the periodic solutions obtained from the ﬁxed points of the contracting return maps attract all ﬂowlines starting in A(β). −a + ) on (β. The resulting estimate for the Lipschitz constant of the return map is L(Rg ) ≤ Lβ (g) (1 + w L(g)) L(Fg (sg (·). Continue this function to an odd map on R with values in (−a − . . used in neural network theory. ) so that Rg is a contraction. First ﬁx β and so that the smallness hypothesis for the previous estimate of L(Rg ) holds for all g ∈ N (β.1+w (A(β))). ). the function g = − tanh(γ·).g. ·)Fg. may not have limits at inﬁnity.) is bounded in terms of Lβ (g) and constants for β. The negative feedback condition may be violated in (−β. for g = − arctan(γ·). . one can show that in case g is continuously diﬀerentiable the periodic orbits are hyperbolic and stable. In [244] the approach described above is reﬁned for nonlinearities g ∈ N (β. A disadvantage of the previous Lipschitz estimate of L(Rg ) is that it is not easy to verify for given analytic nonlinearities. Then choose a Lipschitz continuous odd function g : [−β. on the way of ﬂowlines back to −A(β). like. The small constant Lβ (g) compensates diverging solution behaviour allowed by the factor 1 + w L(g). A consequence is that there exist g ∈ N (β.462 DELAY DIFFERENTIAL EQUATIONS Notice that L(Fg (sg (·). γ > 0 large. Examples are found as follows. ) which are monotone close to 0.
xφ. 1] is formed by the solutions y> = y>. Let z< = z<. Let g ∈ N (β.β. and y< = y<.g (t) ≤ −β on [w.. ) consider the return map Rg : A(β) → A(β) and decompose it as before.) ≤ 4 + L(g) (a + )(µβ + a − ) 1 log . 1]. y(0) = β. ) is almost everywhere diﬀerentiable.Dynamics of Delay Diﬀerential Equations 463 The sharpened estimates are obtained as follows. g decreasing on D ∩ (β− .β.. For β and small and g ∈ N (β. L(Fg. and let β− = β−.g . β). denote the zeros of y< and y> . of the IVPs y˙ = −µ y − a ± . and z> = z>. Recall that a Lipschitz continuous function g ∈ N (β. 0) and increasing on D ∩ (0.β.1 (A(β)). Obviously. The objective is to ﬁnd a better estimate for the middle composite Fg.35. ) be given. For β not too large. with g (ξ) < 0 on D ∩ (β− .w ).w Fg. Figure 10.β. φ ∈ A(β). Deﬁne w = w(β.β. on [0. 0 < w < 1. Proposition. The tongue containing xφ. ) to be the argument where y> reaches the level y = −β. We need the hypothesis that there is a set D ⊂ R whose complement has Lebesgue measure 0. = y< (w) < −β. µ (a − )(µβ + a + ) . respectively. β).
0) and increasing on (0.. 0 ≤ s ≤ w..w .w (φ) − Fg.g (Fg. . β). 0 w g(φ(s − 1)) − g(ψ(s − 1)))ds..). for β and small. Notice that they vanish for = 0 ! This indicates that L(Fg. .w ).ds = g(x(s)) − g(z(s))ds. ∞) → R which start in A(β). ≤ φ − ψ + 0 In the last integral we have φ = x1 and ψ = z1 with solutions x : [−1....36. For −1 ≤ t ≤ 0 and 0 < w + t.w (ψ))(t) = xφ. Remark. 0 0 In order to estimate the integrand. ψ. ..g w (t) − xw (t) = w+t (φ(0) − ψ(0))e−µ(w+t) + e−µ(w+t−s) (g(φ(s − 1)) − g(ψ(s − 1)))ds.464 DELAY DIFFERENTIAL EQUATIONS Figure 10. β) which is decreasing on (β− . ∞) → R and z : [−1.. Let φ and ψ in Fg.) may not be too large even if L(g) is large. g  g(x(s)) − g(z(s)) =  Is ≤ sup g  · x(s) − z(s) ≤ sup . consider the interval Is with endpoints x(s) and z(s).1 (A(β)) be given. Obviously.. .. φ − ψ Is . The factors following L(g) replace the constant w in the former estimate of L(Fg. The proof of the proposition begins with the extension of g to a function on (β− . w w .
. s < z< ) the monotonicity property of g is applicable. y> (s)] ⊂ [β− . Is Figure 10. . We have z< z> w w . Now the last integrals can be estimated as follows. In the remaining case w + t ≤ 0.g (Fg. . z> < s).465 Dynamics of Delay Diﬀerential Equations and β− ≤ y< (s) ≤ min{x(s).w (φ)−Fg.w (ψ))(t) = xφ. The endpoints z< and z> of the range in the second integral can be computed as functions of β.. a. µ.g w (t)−xw (t) = φ(w+t)−ψ(w+t) ≤ φ−ψ. sup g  = −g (0). β]. or Is ⊂ [y< (s). z(s)} ≤ max{x(s). z(s)} ≤ y> (s) ≤ β. 0 0 z< z> −1 The substitution s = y< (ξ) in the ﬁrst integral leads to the upper bound −1 (ξ) in the third integral leads 1 for this integral. The last relation shows that in case 0 < y< (s) (i.e.. Is In case y> (s) < 0 (i. The substitution s = y> to the upper bound 2 for it. clearly ψ. We ﬁnd sup g  ≤ −g (y< (s)). Is In case 0 ∈ Is . ≤ φ − ψ · ( g ◦ y<  + g (0) 1+ g ◦ y> )..e. sup g  ≤ −g (y> (s)).37.
∞). a + }] + 1 + Lβ (g)]. µ a − − µβ e µ The middle factor comes from the estimate of the middle composite Fg. Figure 10.466 DELAY DIFFERENTIAL EQUATIONS Remark. at least for large γ? We have Lβ (gγ ) = γ Lγβ (g). ·[ How can one verify for given nonlinearities g that the upper bound for L(Rg ) is less than 1? Let µ > 0 and an odd bounded Lipschitz continuous function g : R → R be given so that g is negative and increasing on the complement of a set of measure 0 in [0. It can be shown that the Lipschitz estimate of the proposition is optimal in the limit (β. and consider the family of functions gγ = g(γ·).. L(Rg ) ≤ Lβ (g) [4 + L(g) β 3µ ] a2 1 + eµ Lβ (g) a · [µ (1 − (1 − e−µ )) + max{b. Suppose a > 0.. γ > 0. β∗ (γ) → 0. ) and L(Rgγ ) < 1. 0). ) → (0. ). ∞) → (0. Then g has a limit −a ≤ 0 at +∞. ) small and g ∈ N (β. Proposition. Lβ (g) is small. Recall that for suitable g ∈ N (β.. . Continuing as in [243] one ﬁnds that for (β. How can β = β(γ) and = (γ) be chosen so that gγ ∈ N (β.38. ∞) so that for γ → ∞. ). Suppose there exists a function β∗ : (0.w .
Then there exists γ∗ > 0 so that for all γ > γ∗ .. Examples. For the latter. Special cases are given by g = − tanh. ∞) ξ → ξ r g (ξ) is bounded. A somewhat stronger result for oneparameterfamilies of nonlinearities gγ which are at least continuously diﬀerentiable was obtained earlier by Xie [255]. In the approach described here the condition that the nonlinearities g are odd is not essential. I. and = a + g(γ β). γ Lγβ∗ (γ) (g) → 0. 96] and to an equation with statedependent delay [245]. Remarks. Rgγ is a contraction. Suppose in addition that for some r > 32 . He proved that for γ large the orbits of given periodic solutions are hyperbolic and stable. Choose s ∈ (−1. . D ∩ (0.467 Dynamics of Delay Diﬀerential Equations γ β∗ (γ) → ∞. ). 0) with 1 1+s < r and 1 + 1 2+2s < r. β > 0. Then the estimate of L(Rgγ ) has to be used. The proof is a study of the leading Floquet multipliers and relies on very careful apriori estimates of periodic solutions. Consider µ > 0 and g as before. that g is Lebesgue integrable. The function gγ is Lipschitz continuous and satisﬁes the monotonicity hypothesis required for the last proposition. we have gγ ∈ N (β. no other method to obtain stable periodic motion seems presently available. γ Lγβ∗ (γ) (g) γ β∗ (γ) [a + g(γ β∗ (γ))] → 0. and that ξ g (ξ) → 0 as ξ → ∞. For every γ > 0. In order to prove this one sets b = 2a. Variants of the method have also been applied to systems of delay diﬀerential equations [246. and the behaviour of g at inﬁnity is less restricted than in the examples mentioned above. no monotonicity properties of g are required. The hypotheses needed in [255] are that the smooth function g has limits at inﬁnity.e. Then the map β∗ : γ → γ s satisﬁes the hypotheses of the last proposition. g = − arctan.
e. b). b) which are long enough. h < b. for L : C → Rn linear continuous.1) x(t) ˙ = f (xt ) with a continuously diﬀerentiable map f : C 1 ⊃ U → Rn . X = Xf = {φ ∈ U : φ(0) Notice that X is a nonlinear version of the positively invariant domain ˙ = Lφ} {φ ∈ C 1 : φ(0) of the generator G of the semigroup deﬁned by the linear autonomous RFDE y(t) ˙ = Lyt on the larger space C. and the global attractor if the latter is present. The associated IVP is not wellposed on U . b) → Rn . however. Notice also that in case of a locally Lipschitz continuous map f∗ : U∗ → Rn .equilibria. Proposition. This requires an additional condition on f . the set Xf∗ absorbs all ﬂowlines on intervals [−h. periodic orbits. Suppose X = ∅ and (P1) each derivative Df (φ). . all solutions x : [−h. φ ∈ U. of the RFDE x(t) ˙ = f∗ (xt ) satisfy ˙ = f∗ (φ)} for h ≤ t < b. DELAY DIFFERENTIAL EQUATIONS Statedependent delays Diﬀerential equations with statedependent delay can often be written in the form (5. Continuous ﬂowlines may in general only be expected for data in the subset ˙ = f (φ)} ⊂ C 1 . Then X is a continuously diﬀerentiable submanifold of C 1 with codimension n.. has a continuous linear extension De f (φ) : C → Rn . local unstable manifolds. xt ∈ Xf∗ = {φ ∈ U∗ ∩ C 1 : φ(0) I. U open. In order to have a semiﬂow on X with diﬀerentiability properties we need that X is smooth. In particular. b ≤ ∞ .468 5. Xf∗ contains all segments of solutions on intervals (−∞. U∗ ⊂ C open.
˙ = 1. te (φ)). 1). For all φ ∈ X. p : C 1 φ → φ(0) Then X = (p − f )−1 (0). and χ ∈ Tφ X. We also consider solutions on unbounded intervals (−∞. R ⊂ D(p − f )(φ)C 1 . and eq. φ) → xφt ∈ X. (5.1) which start at points φ ∈ X deﬁne a continuous semiﬂow * [0. Suppose (P1) holds. (5. φ ∈ X. te ).χ . and (P2) for each φ ∈ U there exist a neighbourhood V (in C 1 ) and L ≥ 0 so that f (ψ) − f (χ) ≤ Lψ − χ for all ψ ∈ V.1) is satisﬁed. Ω = φ∈X and all solution maps Ft : Ωt φ → F (t. Then the maximal solutions xφ : [−h. Proof of this for n = 1: By (P1). are surjective. DFt (φ)χ = vtφ. It is enough to show that all derivatives D(p − f )(φ).) A solution of eq. te (φ)) → Rn of eq. are continuously diﬀerentiable. te ) → Rn . with t0 ∈ R and t0 < te ≤ ∞. X = ∅. Maximal solutions of IVPs are deﬁned as in part I. Hence There exists χ ∈ C 1 ∩ Cδ with χ(0) 0 < χ(0) ˙ − Df (φ)χ = D(p − f )(φ)χ. Theorem. ∅ = Ωt = {φ ∈ X : (t. φ) ∈ X. for some δ > 0. χ ∈ V. Consider the continuous linear map ˙ ∈ Rn . De f (φ)Cδ ⊂ (−1. te (φ)) × {φ}. φ) ∈ Ω}. xt ∈ U .469 Dynamics of Delay Diﬀerential Equations The proof is easy.1) is deﬁned to be a continuously diﬀerentiable map x : [t0 − h. so that for t0 ≤ t < te . (The Implicit Function Theorem then yields that close to φ the set X is given by a map from a subset of the nullspace N of D(p − f )(φ) into a complement of N in C 1 . (5. t ∈ [0. F : Ω (t.
Examples show how eq. for certain classes of diﬀerential equations with statedependent delay: First the delays are frozen at the given equilibrium φ. and Paraskevopoulos [175]. like the principle of linearized stability for diﬀerential equations with statedependent delay due to Cooke and Huang [48]. v0 = χ. for a class of diﬀerential equations with statedependent delay.χ : [−h. Almost the same property occurred earlier [175]. It is closely related to the earlier idea of being locally almost Lipschitzian in work of MalletParet.470 DELAY DIFFERENTIAL EQUATIONS with a continuously diﬀerentiable solution v φ. te (φ)) → Rn of the IVP v(t) ˙ = Df (Ft (φ))vt . had to be proven without knowledge of linearizations at the equilibrium. . Property (P1) is a special case of a condition used in Krisztin’s work on smooth unstable manifolds [132]. In [139] a complete metric space analogous to X serves as a state space for neutral functional diﬀerential equations. (a) is simply v(t) ˙ = De f (φ)vt . under the name almost Frechet diﬀerentiability. Property (P2) was used in [245] in a proof that for certain diﬀerential equations with statedependent delay stable periodic orbits exist. or results on local unstable manifolds [131. Hbid. then the resulting RFDE with constant delays is linearized. Notice that the Lipschitz estimate in (P2) involves the smaller norm · ≤ · 1 on the larger space C ⊃ C 1 . 132]. Nussbaum. It is mentioned in [164] that X is a Lipschitz manifold. Eq. Comments. (a) is obtained heuristically. Louihi. Former results about solutions close to an equilibrium. (a) is related to the actual variational equation v(t) ˙ = Df (φ)vt for data in Tφ X ⊂ C 1 : Eq. The diﬃculty was circumvented by means of an associated linear autonomous RFDE (a) on the large space C. and Arino [164] identiﬁed the set X as the domain of a generator in the context of nonlinear semigroup theory.
Wc .suitable . Ws of the solution maps Ft at ﬁxed points. At a stationary point φ the local unstable and stable manifolds Wu . under the hypothesis that the associated linear autonomous RFDE (a) is hyperbolic. for the semigroup (T (t))t≥0 on C associated with the last RFDE and for its generator G. in particular. The theorem yields continuously diﬀerentiable local invariant manifolds W u . 0 ≤ t ≤ T. center. by a continuously diﬀerentiable map x : [−h. at stationary points φ of the semiﬂow. In the last case the tangent spaces of the local invariant manifolds at φ are the unstable. respectively. x should also be continuously diﬀerentiable with respect to φ. Krisztin [132]. C c. Tφ X = dom(G) and DFt (φ)χ = T (t)χ on Tφ X. and restricted stable spaces C u. x(t) = φ(0) + 0 x0 = φ ∈ X. Hyperbolicity is also necessary for a recent result of Arino and Sanchez [16] which captures saddle point behaviour of solutions close to equilibrium for a class of diﬀerential equations with statedependent delay. and C s ∩ Tφ X of the generator G. An analogous result for . Local unstable manifolds at stationary points of diﬀerential equations with statedependent delay were found earlier by H. Ws coincide with local unstable and stable manifolds W u .471 Dynamics of Delay Diﬀerential Equations Also. The essential part of the proof of our theorem is to solve the equation t f (xs )ds. T ] → Rn . for φ ∈ X given. Krishnan [131] and T.center manifolds seems unknown. W s of the semiﬂow F . Is the Contraction Mapping Principle with parameters applicable? How does the operator given by the ﬁxed point equation depend on φ? We ﬁrst rewrite the ﬁxed point equation so that the dependence of the .
1 small. u(t) = 0 with parameter φ ∈ X. Now let some φ0 ∈ X be given.472 DELAY DIFFERENTIAL EQUATIONS integral on φ becomes explicit. L may be large!) ≤ L max u(s) − u(s). T ] given by ˆ = φ(0) + φ(0)t ˙ φ(t) on [0. 1 is given by the norm on C0T ˙ yC 1 = max y(t) + max y(t). u). ·) are contractions with respect to the norm on 1 . 0 ≤ t ≤ T. ˆ Then u belongs to the Banach space C 1 of continuously Set u = x − φ. u ∈ C0T and 0 ≤ t ≤ T with T > 0 small. v = C0T A(φ. For 0 ≤ t ≤ T . u). 0T diﬀerentiable maps y : [−h. 0T −h≤t≤T −h≤t≤T u and φˆ satisfy ˆ = φ(0) + u(t) + φ(t) t f (us + φˆs )ds 0 and ˆ = φ(0) + φ(0)t ˙ φ(t) = φ(0) + t f (φ) (since φ ∈ X) t f (φ)ds. T ]. T ] → Rn with y(t) = 0 on [−h. = φ(0) + 0 1 this yields the ﬁxed point equation For u ∈ C0T t (f (us + φˆs ) − f (φ))ds. The property (P2) is used to show that the maps A(φ. deﬁne A(φ. u)(t) by the right hand side of the last equation. 0≤s≤T . 0]. with a contraction factor independent of φ: Let v = A(φ. ˙ = f (ut + φˆt ) − f (ut + φˆt ) ≤ Lut − ut v(t) ˙ − v(t) (due to (P2). For φ ∈ C 1 let φˆ denote the continuously diﬀerentiable extension to [−h. For φ ∈ X close to φ0 .
This completes the essential step in the proof of the theorem. From this a position is computed (which is not necessarily the true position). u(s) − u(s) = 0 − 0 + 0T 0 Hence ˙ ≤ L T u − uC 1 . −w < 0 the position of the obstacle. v(t) − v(t) =  t (f (us + φˆs ) − f (us + φˆs ))ds ≤ L T max us − us 0≤s≤t 0 ≤ L T u − uC 1 . We give an example which is based on elementary physics. and µ > 0 a friction constant.g.473 Dynamics of Delay Diﬀerential Equations We exploit the fact that the last term does not contain derivatives: For 0 ≤ s ≤ T. Consider an object on a line which attempts to regulate its position x(t) by echo.) For 2 L T < 1 the map A(φ. The formula deﬁning A shows that A is continuously diﬀerentiable. The reﬂected signal is detected and the signal running time is measured. v(t) ˙ − v(t) 0T Also. The acceleration is given by a function a : R → R.. 0T (Property (P2) is not necessary here. an equilibrium position in a certain distance from the obstacle). Let c > 0 denote the speed of the signals. Alternatively the local Lipschitz continuity of f with respect to the norm on C 1 can be used to ﬁnd a suitable upper estimate. ·) which is continuously diﬀerentiable with respect to φ. The object emits a signal which is then reﬂected by an obstacle. v(t) its velocity. The computed position is followed by an acceleration towards a preferred position (e. s ˙ (u(r) ˙ − u(r))dr ≤ T u − uC 1 . It follows that for each φ the ball contains a ﬁxed point uφ of A(φ. One ﬁnds a closed ball which is mapped into itself by each map A(φ. ·) becomes a contraction. and s(t) the running time of the signal which has . Let x(t) denote the position of the object at time t. p(t) the computed position. one may think of negative feedback: a(0) = 0 and ξ a(ξ) < 0 for ξ = 0. ·).
Then necessarily 2w + 2w+ = h. The formula deﬁning p(t) yields the true position if x(t) = x(t − s(t)). take h as just deﬁned. Elementary estimates imply that it is also Lipschitz continuous with respect to the norm on C = C([−h. φ˙1 (t) < c for −h ≤ t ≤ 0}. which holds at least at equilibria. . c The model has not yet the form (3) considered in the theorem. h]. The Implicit Function Theorem shows that the resulting map σ : U → [0. Each φ ∈ U determines a unique solution s = σ(φ) of 1 s = (φ1 (−s) + φ1 (0) + 2w). Let w+ > 0. In order to reformulate the model. 0]. R2 ). R2 ). consider the space C 1 = C 1 ([−h. φ2 ) ∈ C 1 : −w < φ1 (t) < w+ .474 DELAY DIFFERENTIAL EQUATIONS been emitted at time t − s(t) and whose reﬂection is detected at time t. Restrict attention further to bounded solutions with −w < x(t) < w+ and x(t) ˙ < c. h] is continuously diﬀerentiable. The model equations then are x(t) ˙ = v(t) v(t) ˙ = −µv(t) + a(p(t)) c p(t) = s(t) − w 2 c s(t) = x(t − s(t)) + x(t) + 2w Here only solutions with −w < x(t) are considered. c as the right hand side of this ﬁxed point equation deﬁnes a contraction on [0. 0 < s(t) ≤ and the open convex subset U = {φ = (φ1 . 0].
0].475 Dynamics of Delay Diﬀerential Equations I. 0] (ψ. 0) + Da( σ(φ) − w) Dσ(φ)χ 2 2 ﬁnally shows that f has property (P1). and for bounded solutions as above the model can be rewritten in the form of eq. Assume the response function a : R → R is continuously diﬀerentiable. Suppose a(0) = 0. and σ(0) = 2w . Then φ = 0 is a stationary point.e. Dσ(φ)χ = χ1 (−σ(φ)) + χ1 (0) Ev1 (χ1 . Veriﬁcation of property (P1). R) × [−h. 0]. c The previous calculations show that the variational equation from the theorem is the system v˙1 (t) = v2 (t) v˙2 (t) = −µv2 (t) + a (0) v1 (t − 2w/c) + v1 (t) . −σ(φ)) + Ev1 (χ1 . t ∈ [−h. Then the map f : U → R2 given by f1 (φ) = φ2 (0). 0]. (5. t) → ψ(t) ∈ R. −σ(φ)) − c with the evaluation map Ev1 : C 1 ([−h. and c f2 (φ) = −µφ2 (0) + a( σ(φ) − w) = −µφ2 (0) + a 2 φ1 (−σ(φ)) + φ1 (0) 2 is continuously diﬀerentiable and has property (P2). is the restriction of a continuous linear evaluation functional on the space C([−h. It follows that Dσ(φ) : C 1 → R has a continuous linear extension De σ(φ) : C → R.. ˙ φ˙1 (−σ(φ)) − c Ev1 (φ1 . 2 . The formula c c Df2 (φ)χ = −µEv1 (χ2 . 0) = . an analogue of property (P2) holds. Each Ev1 (·. R). For each φ ∈ U and all χ ∈ C 1 . t).1).
476 DELAY DIFFERENTIAL EQUATIONS Notice that the heuristic approach (freeze the delay. A ﬁrst Hopf bifurcation theorem for diﬀerential equations with statedependent delay is due to M. (P1∗ ) holds for the example. Can the approach be generalized so that one obtains higher order derivatives for semiﬂows given by diﬀerential equations with statedependent delay? For suitable parameters stable periodic motion far away from equilibrium was established in [248]. Back to general results. Notice that continuous diﬀerentiability of the semiﬂow (for t > h) is needed in order to have Poincar´e return maps for periodic orbits. Eichmann [62]. The stronger condition that the map U φ → De f (φ) ∈ L(C. The semiﬂow F from the theorem is continuously diﬀerentiable for t > h provided f satisﬁes (P1) and the condition that (P1∗ ) the map U × C (φ. Further work. which was initiated by R. among others. then linearize) yields the very same system. . χ) → De f (φ)χ ∈ Rn is continuous. Driver [61]. A related but in several respects more complicated problem is the description of the motion of two charged particles. (P1) and (P1∗ ) imply (P2). The solution maps Ft are compact for t ≥ h under certain other hypotheses on f . Open Problems. see the forthcoming survey article [107] and [134]. The latter are satisﬁed by the example for suitable functions a. For continuously diﬀerentiable center manifolds of the semiﬂow.D. Rn ) be continuous is not satisﬁed by the example. Remarks.
we shall review various delay diﬀerential equations models arising from studying single species dynamics. Gopalsamy (1992). Delay Differential Equations and Applications. © 2006 Springer. Then x(t + ∆t) − x(t) = bx(t)∆t − dx(t)∆t.miami. Kuang (1993) and MacDonald (1978) for discussions of general delayed biological systems. we obtain ∗ Research was partially supported by NSF grant DMS0412047 and a Small Grant Award from the University of Miami. In this survey. respectively. Introduction Time delays of one type or another have been incorporated into biological models to represent resource regeneration times. t + ∆t]. reaction times. Arino et al. delay diﬀerential equations exhibit much more complicated dynamics than ordinary diﬀerential equations since a time delay could cause a stable equilibrium to become unstable and cause the populations to ﬂuctuate. We refer to the monographs of Cushing (1977a).). Ruan∗ Department of Mathematics University of Miami PO Box 249085 Coral Gables.edu 1. FL 331244250 USA ruan@math. where ∆t > 0.Chapter 11 DELAY DIFFERENTIAL EQUATIONS IN SINGLE SPECIES DYNAMICS S. Dividing by ∆t and letting ∆t approach zero. . maturation periods. 477–517. (eds. In general. 477 O. feeding times. etc. Let x(t) denote the population size at time t. on the time interval [t. let b and d denote the birth rate and death rate. by many researchers.
1) where r = b − d is the intrinsic growth rate of the population.1) with an initial population x(0) = x0 is given by (1. we obtain (x(0) = x0 ) x(t) = x0 K . the process of reproduction is not instantaneous.3). Such a population growth. The solution of equation (1. but by the amount of food available when the eggs were forming.4) If x0 < K. Solving (1. some time before they pass into the broad pouch. x0 − (x0 − K)e−rt (1. approaching K asymptotically as t → ∞.3). in Daphnia a large clutch presumably is determined not by the concentration of unconsumed food available when the eggs hatch. the positive equilibrium x = K of the logistic equation (1.3) by separating the variables. the population grows. (1. 2. Hutchinson’s Equation In the above logistic model it is assumed that the growth rate of a population at any time t depends on the relative number of individuals at that time. lim x(t) = K for solution x(t) of (1. dt (1.1). Taking the fact that resources are limited into account. Verhulst (1836) proposed the logistic equation x dx = rx 1 − . when x is small the population grows as in the Malthusian model (1. when x is large the members of the species compete with each other for the limited resources. In model (1.3) is globally stable. that is. For example. In fact.3) dt K where r(> 0) is the intrinsic growth rate and K(> 0) is the carrying capacity of the population. but it cannot go on forever. x = K is called an equilibrium of equation (1. Thus. the population decreases. In practice.2) x(t) = x0 ert .2) represents the traditional exponential growth if r > 0 or decay if r < 0 of a population. If x0 > K. due to Malthus (1798). the population remains in time at x = K. If x0 = K. may be valid for a short period.3) with any initial t→∞ value x(0) = x0 . again approaching K asymptotically as t → ∞.478 DELAY DIFFERENTIAL EQUATIONS dx = bx − dx = rx. Between this time of determination and the time of hatching many newly hatched animals may have been liberated from the brood pouches of other Daphnia in . The function (1.
0] implies that all solutions x(t) of (2. If in addition there is a δ0 > 0 such that φ(t) − x∗  ≤ δ0 on [−τ.3) which is a transcendental equation. then x∗ is called t→∞ asymptotically stable. which shows that x = 0 is unstable with exponential growth. Let X = x − K. 0] satisfy x(t) − x∗  < for all t ≥ 0.3) have negative real parts. (2. 2. τ > 0 is a constant.479 DDE in Single Species Dynamics the culture. dt (2. 0]. 0] implies lim x(t) = x∗ . Hutchinson (1948) assumed egg formation to occur τ units of time before hatching and proposed the following more realistic logistic equation x(t − τ ) dx = rx(t) 1 − .1) is given by x(θ) = φ(θ) > 0. 0]. An equilibrium x = x∗ of (2.2) We look for solutions of the form X(t) = ceλt . Then. where c is a constant and the eigenvalues λ are solutions of the characteristic equation λ + re−λτ = 0. r dX = −rX(t − τ ) − X(t)X(t − τ ).3).1) is stable if for any given > 0 there is a δ > 0 such that φ(t) − x∗  ≤ δ on [−τ. In fact. where φ is continuous on [−τ. θ ∈ [−τ.1) dt K where r and K have the same meaning as in the logistic equation (1.1 Stability and Bifurcation The initial value of equation (2. Small perturbations from x = 0 satisfy the linear equation dx dt = rx.1) has equilibria x = 0 and x = K. Equation (2. x = K is asymptotically stable if all eigenvalues of (2. (2. By the linearization theory. Notice that equation (2. We thus only need to consider the stability of the positive equilibrium x = K. so increasing the population. .1) is often referred to as the Hutchinson’s equation or delayed logistic equation. the linearized equation is dX = −rX(t − τ ).1) with initial value φ on [−τ. in an extreme case all the vacant spaces K − x might have been ﬁlled well before reproduction stops. dt K Thus.
(2. all roots and all other roots have negative real parts.3) becomes λ+r = 0 and the eigenvalue λ = −r < 0 is a negative real number. Theorem 1 (i) If 0 ≤ rτ < π2 . By the continuity. . We have the transversality condition dµ = r2 > 0. . which implies that ν o τk = π + 2kπ.1) is asymptotically stable. In other words. we have τk = 2kπ π + . then the positive equilibrium x = K of equation ( 2. k = 0. say τ0 . The periodic solutions exist for rτ > π2 and are stable. 1. dτ τ =τk τ = τ0 = We have just shown the following conclusions. if λ changes from −r to a value such that Reλ = µ > 0 when τ increases. Denote λ(τ ) = µ(τ ) + iν(τ ) the root of equation (2. When 0 < τ < 2r of (2. . .480 DELAY DIFFERENTIAL EQUATIONS Set λ = µ + iν. .3) have strictly negative real parts. 1.3) must have a pair of purely imaginary roots ±iν0 . . 2. (ii) If rτ > π2 . ν(τk ) = ν0 . We seek conditions on τ such that Reλ changes from negative to positive. 2. .4) Notice that when τ = 0. then x = K is unstable. a Hopf bifurcation occurs at x = K. .3) has a pair of purely imaginary roots ±ir. k = 0. there must be some value of τ. Separating the real and imaginary parts of the characteristic equation (2.3) satisfying µ(τk ) = 0. we obtain µ + re−µτ cos ντ = 0. . k = 0. Suppose such is the case. which are simple π . that is. (iii) When rτ = π2 . the characteristic equation (2. ν − re−µτ sin ντ = 0. Then we have cos ν0 τ = 0. 1. at which Reλ(τ0 ) = µ(τ0 ) = 0. when π . . k = 0. . 1. 2r equation (2. 2. ν0 = ν(τ0 ). 2. . 2 Noting that νo = r. . periodic solutions bifurcate from x = K.3). 2r r Therefore. . the characteristic equation (2.
If T is the period then x(t + T ) = x(t) for all t.1).2). the stability of a periodic solution means that if a perturbation is imposed the solution returns to the original periodic solution as t → ∞ with possibly a phase shift. By (iii).2 Wright Conjecture The Hutchinson’s equation (2. x K 0 Figure 11. it is 4τ (see Fig. 11. 2.1.3. 11.1. the Hutchinson’s equation (2.1).1) can have periodic solutions for a large range of values of rτ . τ The bifurcation diagram for equation (2. (1981)).2. Roughly speaking. the product of the birth rate r and the delay τ .1) can be written as dy = −ry(t − τ )[1 + y(t)] dt .481 DDE in Single Species Dynamics x K π/2r 0 Figure 11. where the solid curves represent stability while the dashed lines indicate instability. t1 t 1+τ t 1+2τ t 1+3τ t 1+4τ 1 t The periodic solution of the Hutchinson’s equation (2. Numerical simulations are given in Fig. The period of the solution at the critical delay value is 2π ν0 (Hassard et al. The above theorem can be illustrated by Fig. thus. 11.
Here r = 0.3. Letting t = τ t. Wright then conjectured that the zero solution of (2.5) is globally stable if α < π/2.6 0. some attention has been paid to the study of equation (2. a periodic solution bifurcated from x∗ = 1.6 1.4 1.8 tau=11 1. See also Kuang (1993) for further results and more references. y(t) = y(τ t¯).2 1 tau=8 0.time 200 250 300 Figure 11. we obtain dy = −αy(t − 1)[1 + y(t)]. Jones (1962a. Recently. Sugie (1992) showed that the zero solution of (2. K = 1.8 0. For example.5) oscillate if α > 1/e and do not oscillate if α < 1/e. Walther (1975). dt Denoting α = rτ and dropping the bars. see Hadeler and Tomiuk (1977).00. Wright (1955) showed that the zero solution of (2. Kakutani and Markus (1958) proved that all solutions of (2.5) is globally stable if α < 3/2. the steady state x∗ = 1 is stable.5) is asymptotically stable if α < π/2 and unstable if α > π/2.482 DELAY DIFFERENTIAL EQUATIONS x(t) 1. Hale and Verduyn Lunel (1993). we know that the zero solution of (2. (i) When τ = 8.5) By Theorem 1. Numerical simulations for the Hutchinson’s equation (2.4 0 50 100 150 t .5) when α = α(t) is a positive continuous function. dt (2. which is still open.6) . etc. Kaplan and York (1975). by assuming y(t) = −1 + x(t)/K.15. 1962b) studied the global existence of periodic solutions for α > π/2. we have d y(t) = −rτ y(t − 1)[1 + y(t)]. Naussbaum (1974). (ii) When τ = 11.1). 2 (2. For further research on existence of nonconstant periodic solutions.5) with α = α(t) is uniformly stable if there is a constant α0 > 0 such that α(t) ≤ α0 < 3 for all t ≥ 0.
then the steady state x∗ = ymptotically stable for all delay τ ≥ 0.7) are called 32 −stability criteria. r) to R and denote xt (θ) = x(t + θ). we can show that it is asymptotically stable for any initial value. Theorem 2 (i) If a1 ≥ a2 .3 Instantaneous Dominance Consider a logistic equation with a discrete delay of the form dx = rx(t)[1 − a1 x(t) − a2 x(t − τ )]. There is a positive equilibrium 1 . xt (θ)) = x − x − x ln ∗ + ξ x −τ . Choose a Lyapunov function of the form 0 x ∗ ∗ [xt (θ)]2 dθ. globally stable. 1 a1 +a2 of ( 2. (2.7) Stability conditions such as (2. dt (2.8) is Proof. then the steady state x∗ = globally stable. τ0 ) and 2 unstable when τ > τ0 .483 DDE in Single Species Dynamics Chen et al.8) where a1 and a2 are positive constants. that is. A Hopf bifurcation occurs at x∗ when τ passes through τ0 . For further related work. 1 a1 +a2 is as (ii) If a1 < a2 . Yu (1996) and the references therein. one can prove the following results. In fact. we refer to Kuang (1993). α(s)ds ≤ α0 < 2 t−1 (2. (1995) improved condition (2. a2 r a2 − a21 1 such that x∗ = a1 +a is asymptotically stable when τ ∈ [0.6) and (2. θ ∈ [−τ. 0].9) V (x(t). Employing similar x∗ = a1 +a 2 arguments. then there is a critical value τ0 given by a22 − a21 a1 + a2 τ0 = 2 arcsin . The above result indicates that if a1 ≥ a2 . Suppose x is a continuous function from [−τ. 2. Theorem 3 If a1 > a2 . if the instantaneous 1 is asymptotically term is dominant. that is.6) to the following: t 3 for t ≥ 1. which is stable when there is no delay. then the steady state x∗ = a1 +a 2 stable for all delay τ ≥ 0.
484 DELAY DIFFERENTIAL EQUATIONS where ξ > 0 is follows: dx dt Then we have dV = dt (2.1) can be used to explain several experimental situations. Rewrite equation (2. τ is approximately the time for a larva to mature into an adult. K is set by the food level available. Over a period of nearly two years Nicholson recorded the population of ﬂies and observed a regular basic periodic oscillation of about 3540 days.1 dV dt (2. Let x(t) denote the population of sexually mature adults. the intrinsic growth rate of the population. To apply the Hutchinson’s equation (2. dt To express R(t) we have to consider the populations of all the various stages in the lifehistory of the species concerned and make the following assumptions: (i) all eggs take exactly τ time units to develop into sexually mature adults. (2. 3.10) a constant to be determined. so result follows.8) as = rx(t)[−a1 (x(t) − x∗ ) − a2 (x(t − τ ) − x∗ )]. . then the delay is about 9 days: the actual delay is about 15 days. including Nicholson’s (1954) careful experimental data of the Australian sheepblowﬂy (Lucila cuprina). choose ξ = 12 ra1 . (1980) tried to modify Hutchinson’s equation. If a1 > a2 .10) is negatively deﬁnite and the Recruitment Models Nicholson’s Blowﬂies Model The Hutchinson’s equation (2.10) dx x − x∗ + ξ[(x(t) − x∗ )2 − (x(t − τ ) − x∗ )2 ] dt x = −{(ra1 − ξ)[x(t) − x∗ ]2 + ra2 [x(t) − x∗ ][x(t − τ ) − x∗ ] +ξ[x(t − τ ) − x∗ ]2 }. To overcome the discrepancy in estimating the delay value. Notice that Nicholson’s data on blowﬂies consist primarily of observations of the time variation of adult population.1). Gurney et al. Then the rate of change of x(t) is the instantaneous rate of recruitment to the adult population R(t) minus the instantaneous total death rate D(t) : dx = R(t) − D(t). If we take the observed period as 40 days. The only unknown parameter is r. 3.
11. 11. . where P is the maximum per capita daily egg production rate. (3. (iii) the probability of a given egg maturing into a viable adult depends only on the number of competitors of the same age. and δ is the per capita daily adult death rate. Testing Nicholson’s data. thus R(t + τ ) = θ(x(t)) = P x(t) exp[−x(t)/x0 ]. Assume that the per capita adult death rate has a time and density independent value δ. Assume that the average per capita fecundity drops exponentially with increasing population. there is a critical value of the time delay. As in the Hutchinson’s equation. The additional assumption that the total death rate D(t) is a function only of the instantaneous size of the adult population D(t) = φ(x(t)) = δx(t) enables the entire population dynamics to be expressed in the delay diﬀerential equation x(t − τ ) dx = P x(t − τ ) exp − − δx(t). becomes unstable when it is greater the value. of limitcycle type (see Fig.1) dt x0 There is a positive equilibrium x∗ = x0 ln(P/δ) if the maximum possible per capita reproduction rate is greater than the per capita death rate. that is. Gurney et al. x0 is the size at which the blowﬂies population reproduces at its maximum rate.DDE in Single Species Dynamics 485 (ii) the rate at which the adult population produces eggs depends only on its current size.1) not only provides selfsustaining limit cycles as the Hutchinson’s equation did. The positive equilibrium is stable when the delay is less the critical value. and there are oscillations. These imply that the rate of recruitment at time t + τ is a function only of the instantaneous size of the adult population at time t. but also gives an accurate measurement of the delay value as 15 days.4). equation (3. See Fig. The period of the cycles is set mainly by the delay and adult death rate. Moving deeper into instability produces a number of successive doublings of the repeated time until a region is reached where the solution becomes aperiodic (chaotic). (1980) showed that the ﬂuctuations observed by Nicholson are quite clear. High values of P τ and δτ will give large amplitude cycles. if P > δ.5.
400 Here P = x(t) 25 20 15 10 5 0 0 50 100 150 200 t . Gy¨ ori and Troﬁmchuk(2002). Oscillations in the Nicholson’s blowﬂies equation (3.486 DELAY DIFFERENTIAL EQUATIONS x(t) 60 50 40 30 20 10 0 0 50 100 150 200 t . etc. Kulenovi´c et al. Taylor and Sokal (1976) proposed the delay equation dx = −dx(t) + bx(t − τ )[k − bzx(t − τ )].5.1).1). (1992). and τ = 15. and τ = 15.4. dt (3. δ = 0. 3. Equation (3. see Nisbet and Gurney (1982). δ = 0. Here P = 8.time 250 300 350 400 Figure 11.2) .175. Aperiodic oscillations in the Nicholson’s blowﬂies equation (3. 8.2 Houseﬂies Model To describe the oscillations of the adult numbers in laboratory populations of houseﬂies Musca domestica.475.time 250 300 350 Figure 11. Smith (1995). x0 = 4. x0 = 4. So and Yu (1994).1) is now refereed to as the Nicholson’s blowﬂies equation.
11.time 200 250 300 Figure 11. i. where k > 0 is the maximum eggadult survival rate. (1982) proposed a general single species population model with a time delay dx = R(x(t − τ )) − Dx(t).2) has never been carried out.. k − bzx(t − τ ) represents the eggtoadult survival rate.147. τ = 0. However.e. an adult population of size x. dt (3. Numerical simulations in the houseﬂies model (3.000226. 3. the time delay τ > 0 is the length of the developmental period between oviposition and eclosion of adults. aperiodic oscillations have not been observed. where b > 0 is the number of eggs laid par adult. z = 0.2). τ = 5 were reported in Taylor and Sokal (1976). Here the parameter values b = 1.81. d > 0 denotes the death rate of adults. so at time t − τ the number of new eggs would be bx(t − τ ). unlike the Nicholson’s model. and z is the reduction in survival produced by each additional egg. d = 0.487 DDE in Single Species Dynamics x(t) 1400 1200 1000 800 600 400 200 0 0 50 100 150 t .3) where R and D represent the rates of recruitment to. see Brauer and CastilloCh´ avez (2001). and death rate from. For a linear analysis of the model. Notice that when there is no time delay.6. then the equation becomes the familiar logistic equation. k = 0. and τ > 0 is the maturation period. numerical simulations indicate that its dynamics are very similar to that of the Nicholson’s blowﬂies equation (see Fig.5107. The number of eggs laid is assumed to be proportional to the number of adults. .3 Recruitment Models Blythe et al. Though analytical analysis of equation (3.6). where x(t) is the number of adults.
the phenomenon in which reproduction rates of individuals decrease when density drops below a certain critical level is now known as the Allee eﬀect. etc. Some parasitic insects aggregate so that they can overcome the defense mechanism of a host. for example x(t − τ ) bx2 (t − τ ) 1− R(x(t − τ )) = x(t − τ ) + x0 K as in Beddington and May (1975). altruism. In other . c > 0. Karakostas et al. such as R(x(t − τ )) = P x(t − τ ) exp[−x(t − τ )/x0 ] in the Nicholson’s blowﬂies equation. The Allee Eﬀect The logistic equation was based on the assumption that the density has a negative eﬀect on the percapita growth rate. bees. etc. See also Cao and Gard (1995).488 DELAY DIFFERENTIAL EQUATIONS This equation could exhibit very complex dynamic behavior for some functions R. dt (4. termites. Freedman and Gopalsamy (1986) studied three classes of general single species models with a single delay and established criteria for the positive equilibrium to be globally stable independent of the length of delay. Aggregation and associated cooperative and social characteristics among members of a species were extensively studied in animal populations by Allee (1931). However.. (1992). Such cooperative processes have a positive feedback inﬂuence since individuals have been provided a greater chance to survive and reproduce as density increase. when the density of the population is not small. However. In the model.1) where a > 0. For example some species form hunting groups (packs. A number of social species such as ants. have developed complex cooperative behavior involving division of labor. Gopalsamy and Ladas (1990) proposed a single species population model exhibiting the Allee eﬀect in which the per capita growth rate is a quadratic function of the density and is subject to time delays: dx = x(t)[a + bx(t − τ ) − cx2 (t − τ )]. etc. and b are real constants. some species often cooperate among themselves in their search for food and to escape from their predators. τ ≥ 0. 4. prides. Fish and birds often form schools and ﬂocks as a defense against their predators.) to enable them to capture large prey. the time delay is not necessarily destabilizing (see also Rodr´ıguez (1998)). for some other functions. the positive feedback eﬀects of aggregation and cooperation are dominated by densitydependent stabilizing negative feedback eﬀects due to intraspeciﬁc competition.
The steady state of the delay model (4.6 0 20 60 40 80 100 t . 2 then the equilibrium x∗ attracts all positive solutions of ( 4.1) is attractive. Theorem 4 If 3 τ x∗ (2cx∗ − b) ≤ .8 1.2) dt where p. The following result is a corollary of the main results of Liz et al.6 2. (4.1). and discussed oscillation and global attractivity in the solutions. c = 0. If the delay is suﬃciently large. See also Cao and Gard (1995).4 2. FoodLimited Models Rewriting the logistic equation (2. q are positive constants. b = 1. Ladas and Qian (1994) generalized (4.8 2.2 3 2.2 2 1.1) oscillate about the positive equilibrium. (2003).7. solutions of equation (4.1) to the form dx = x(t)[a + bxp (t − τ ) − cxq (t − τ )]. Here a = 1.1) as x 1 dx . Equation (4. τ = 0. intraspeciﬁc mutualism dominates at low densities and intraspeciﬁc competition dominates at higher densities. 11.2.489 DDE in Single Species Dynamics x(t) 3.7). words.5.1) has a positive equilibrium √ b + b2 + 4ac ∗ .time Figure 11. the positive equilibrium is globally attractive (see Fig. =r 1− K x(t) dt . 5.4 3. x = 2c Gopalsamy and Ladas (1990) showed that under some restrictive assumptions.
see Gopalsamy et al. The dynamics are very similar to the Hutchinson’s model. .2) and obtained the delayed foodlimited model K − x(t − τ ) dx = rx(t) . Note that F/T is not usually equal to x/K.8). dt c1 > 0.2) has also been discussed by Hallam and DeLuna (1984) in studying the eﬀects of environmental toxicants on populations. Equation (5. Qian (1996). c2 ≥ 0. For other related work on equation (5. x dt c1 K If we let c = c2 /c1 ≥ 0. (1993). So and Yu (1995). Gopalsamy et al.490 DELAY DIFFERENTIAL EQUATIONS we can see that the average growth rate of a population is a linear function of its density. It is assumed that F depends on the density x (that is being maintained) and dx/dt (the rate of change of the density) and takes the following form: F = c 1 x + c2 dx . x = K and T = F. When saturation is attained.3) and its generalizations. Thus. etc. the above equation can be simpliﬁed to the form K − x(t) 1 dx(t) =r . and T is the rate at which the population uses food when it is at the equilibrium K. 11.1) x dt T where F is the rate at which a population of biomass x consumes resources. (1990a). Smith argued that the per capita growth rate of a population is proportional to the rate of food supply not momentarily being used.1) becomes c1 x + c2 dx 1 dx dt =r 1− . (1988) introduced a time delay τ > 0 into (5. (5. dx/dt = 0.3) dt K + rcx(t − τ ) They studied global attractivity of the positive equilibrium x∗ = K and oscillation of solutions about x∗ = K (see Fig. (5. In experiments of bacteria cultures Daphnia magna Smith (1963) found that the average growth rate (1/x)(dx/dt) is not a linear function of the density. This results in the model: F 1 dx =r 1− . T = c1 K and equation (5. Grove et al.2) x(t) dt K + rcx(t) which is referee to as the foodlimited population model. (5.
15.4 tau=8 0. There is a feedback from the blood to the bone marrow. Assume that the cells are lost at a rate proportional to their concentration. Regulation of Haematopoiesis Haematopoiesis is the process by which primitive stem cells proliferate and diﬀerentiate to produce mature blood cells. It is driven by highly coordinated patterns of gene expression under the inﬂuence of growth factors and hormones. called erythropoiesis.8 0.time 150 200 Figure 11.8 tau=12. where g has dimension (day−1 ).1 MackeyGlass Models Let c(t) be the concentration of cells (the population species) in the circulating blood with units cells/mm3 .8 1.491 DDE in Single Species Dynamics x(t) 2 1. gc. Assume the ﬂux of cells into the blood stream depends on the cell concentration at an earlier time. The principal factor stimulating red blood cell product is the hormone produced in the kidney.6 1. Here r = 0. The steady state of the delay foodlimited model (5.6 0. K = 1. c(t − τ ). 6. where τ is the delay.3) is stable for small delay (τ = 8) and unstable for large delay (τ = 12.00. which in turn causes an increase in the release of the blood elements from the marrow. When the level of oxygen in the blood decreases this leads to a release of a substance.8).4 1. The regulation of haematopoiesis is about the formation of blood cell elements in the body. 6. About 90% of the erythropoiesis is secreted by renal tubular epithelial cells when blood is unable to deliver suﬃcient oxygen.2 1 0.2 0 50 100 t .8. c = 1. . say. Abnormalities in the feedback are considered as major suspects in causing periodic haematopological disease. White and red blood cells are produced in the bone marrow from where they enter the blood stream. After the reduction in cells in the blood stream there is about a 6day delay before the marrow releases further cells to replenish the deﬁciency.
.1).10. Here λ = 0.times 400 500 600 Figure 11.1 0. a = 01. a = 01.05 0.1).10).9.02 0 0 100 200 300 t .12 0. the following delay model for the blood cell population dc λam c(t − τ ) = m − gc(t).9). m. 11. a. among others. g = 0.1.1 0.11 0. Aperiodic behavior of the solutions of the MackeyGlass model (6. dt a + cm (t − τ ) (6.04 0. Oscillations in the MackeyGlass model (6.09 0. 11.1) by Mackey and Glass (1977) (see also Mackey and Milton (1988)) indicate that there is a cascading sequence of bifurcating periodic solutions when the delay is increased (see Fig. c(t) 0.06 0. g = 0.times 400 500 600 Figure 11.2.08 0.04 0. Here λ = 0.06 0.492 DELAY DIFFERENTIAL EQUATIONS c(t) 0. m = 10 and τ = 6.1.03 0. When the delay is further increased the periodic solutions becomes aperiodic and chaotic (see Fig. The numerical simulations of equation (6.1) where λ.14 0. g and τ are positive constants.07 0. m = 10 and τ = 20.12 0.2.08 0.02 0 100 200 300 t . Mackey and Glass (1977) suggested.
(1992). (1990b). Global attractivity in the MackeyGlass model (6. and by the host to that vector. where d is the infective rate of the vectors. Within the vector there is an incubation period τ before the disease agent can infect a host. 7. The disease is transmitted to the host by an insect vector. Thus. Infection of the host is assumed to proceed at a rate (e) proportional to encounters between uninfected host and vectors capable of transmitting the disease.1) and the LasotaWazewska model (6. Kuang (1992) and Gy¨ ori and Troﬁmchuk (1999). p and γ are positive constants related to the production of redblood cells per unit time and τ is the time required to produce a redblood cells. Karakostas et al.2 493 WazewskaCzyzewska and Lasota Model Another wellknown model belongs to WazewskaCzyzewska and Lasota (1976) which takes the form dN = −µN (t) + pe−γN (t−τ ) .1990) and Fowler and Mackey (2002). Liz et al.1) . A Vector Disease Model Let y(t) denote the infected host population and x(t) be the population of uninfected human. ex(t) · dy(t − τ ) = by(t − τ )[1 − y(t)]. Infection leads neither to death. Based on these assumptions. (2002) study these models when the delay is inﬁnite. µ is the probability of death of a redblood cell. Assume that the total host population is constant and is scaled so that x(t) + y(t) = 1. assumed to have a large and constant population. See also Arino and Kimmel (1986). (6.DDE in Single Species Dynamics 6. Other types of delay physiological models can be found in Mackey and Milton (1988. (7. b is the contact rate.2) has been studied by Gopalsamy et al. dt where a > 0 is the cure rate.2) dt where N (t) denotes the number of redblood cells at time t. So the population of vectors capable of infecting the host is z(t) = dy(t − τ ). and recovery to proceed exponentially at a rate c. immunity or isolation. Cooke (1978) proposed a delay model dy = by(t − τ )[1 − y(t)] − ay(t).
b = 4.11.15 0.8. then the steady state solution u1 = (b − a)/b is asymptotically stable and the set {φ ∈ ([−τ.1).8(a < b).11. . (ii) If 0 ≤ a < b.1) when b = b(t) is a positive periodic function. R) : 0 ≤ φ(θ) ≤ 1 for − τ ≤ θ ≤ 0} is a region of attraction. Numerical simulations are given in Fig.8(a > b). 0].494 DELAY DIFFERENTIAL EQUATIONS u(t) 0. b = 4. Using the Liapunov functional method. Numerical simulations for the vector disease equation (7.35 0. The stability results indicate that there is a threshold at b = a.1 0.time 40 50 60 Figure 11. Theorem 5 For the vector disease model ( 7. the positive steady state u∗ is asymptotically stable for all delay values.45 0.4 0. we have the following: (i) If 0 < b ≤ a. the zero steady state u = 0 is asymptotically stable.1). 11.25 0. When a = 5. R) : 0 < φ(θ) ≤ 1 for − τ ≤ θ ≤ 0} is a region of attraction. When a = 3.5 0. There is no nonconstant periodic solutions in the region 0 ≤ u ≤ 1. he obtained the following results on global stability of the steady states. then the steady state solution u0 = 0 is asymptotically stable and the set {φ ∈ ([−τ.3 a<b 0. If b > a. If b ≤ a. the proportion of infectious individuals tends to an endemic level u1 = (b − a)/b as t becomes large.05 a>b 0 0 10 20 30 t . 0].8. here for both cases τ = 5. then the proportion u of infectious individuals tends to zero as t becomes large and the disease dies out. Busenberg and Cooke (1978) studied the existence of periodic solutions in the vectorhost model (7.2 0.
τ1 and τ2 are positive constants.1).1) becomes ˙ X(t) = −(1 + X(t))[A1 X(t − τ1 ) + A2 X(t − τ2 )]. Then (8. Multiple Delays Kitching (1977) pointed out that the life cycle of the Australian blowﬂy Lucila cuprina has multiple time delay features which need to be considered in modelling its population.1) where r. Using the results in Li et al. The linearized equation of (8.1) and observed stable limit cycles when τ2 /τ1 is large. Gopalsamy (1990) obtained stability conditions for the positive equilibrium.2) where A1 = ra1 x∗ . Mahaﬀy et al. Braddock and van den Driessche (1983) described some linear stability regions for equation (8. (1993)). Equation (8. a2 . Very rich dynamics have been observed in such equations (Hale and Huang (1993). τ 1 ].2) at X = 0 is ˙ X(t) = −A1 X(t − τ1 ) − A2 X(t − τ2 ). etc. Theorem 6 If one of the following conditions is satisﬁed: 3π (i) A1 < A2 and τ1 > 0 such that 2τπ1 < A22 − A21 < 2τ .1) has a positive equilibrium x∗ = 1/(a1 + a2 ). a1 . Let x(t) = x∗ (1 + X(t)). A2 = ra2 x∗ . Braddock and van den Driessche (1983) proposed the two delay logistic equation (see also Gopalsamy (1990)) dx = rx(t)[1 − a1 x(t − τ1 ) − a2 x(t − τ2 )]. Based on this observation. (1995)). (1995)). medical models (B´elair et al. epidemiological models (Cooke and Yorke (1973)). (8. where τ 1 = (A21 − A22 )− 2 arcsin 1 (iii) A1 = A2 and τ1 > . we can obtain the following theorem on the stability and bifurcation of equation (8. 1 (ii) A2 < A1 and τ 1 > π 2(A1 +A2 ) such that τ1 ∈ [ 2(A1π+A2 ) . Other equations with two delays appear in neurological models (B´elair and Campbell (1994)). (1999). physiological models (Beuter et al.495 DDE in Single Species Dynamics 8. dt (8.
such that when τ2 = τ20 the twodelay equation ( 8. 1 2A1 . . (A21 − A22 )/A21 . then there is a τ20 > 0.1) undergoes a Hopf bifurcation at x∗ = 1/(a1 + a2 ).
For the two delay logistic model (8.time 150 200 Figure 11.2 (a) 1 0. choose r = 0.1).496 DELAY DIFFERENTIAL EQUATIONS x(t) 1. a1 = 0.1) means that the regulatory eﬀect depends on the population at a ﬁxed earlier time t − τ .6 0 50 100 t .4 1.25.3) dt i=1 Their global stability conditions very much depend on the negative. rather than at the present time t. a Hopf bifurcation occurs.8 0.75. Attempts to apply logistic models to these experiments were often unsuccessful because populations died out. This results in an equation with a distributed delay or an inﬁnite delay. The ﬁrst work using a logistic equation with distributed delay was by Volterra (1934) with extensions by Kostitzin (1939).4) dt Ki i=1 9. (a) The steady state (a) is stable when τ1 = 15 and τ2 = 5 and (b) becomes unstable when τ1 = 15 and τ2 = 10.6 (b) 1. a2 = 0. (8. instantaneously dominated constant a. many experiments were performed with laboratory populations of some species of small organisms with short generation time. One of . In a more realistic model the delay eﬀect should be an average over past populations. It would be interesting to determine the dynamics of the multipledelay logistic equation without the negative instantaneously dominated term (see Kuang (1993)) n dx x(t − τi ) = rx(t) 1 − .12. (8. Lenhart and Travis (1986) studied the global stability of the multiple delay population model n dx = x(t) r + ax(t) + bi x(t − τi ) . Volterra Integrodiﬀerential Equations The Hutchinson’s equation (2.15. In the 1930’s.
respectively.1) reduces to the discrete delay logistic equation 1 t x(t − τ ) dx = rx(t) 1 − . T = 0 It follows that if G(u) = δ(u − τ ). called the delay kernel. Usually the delay kernel is normalized so that ∞ G(u)du = 1. Volterra (1934) used an integral term or a distributed delay term to examine a cumulative eﬀect in the death rate of a species. depending on the population at all times from the start of the experiment.1) dt K −∞ where G(t). The weak kernel qualitatively indicates that the maximum weighted response . (n = 2). then T = τ . n an integer.2) where α > 0 is a constant. −∞ then equation (9.3) remains an equilibrium in the presence of time delay. Two special cases. (n − 1)! n = 1. If G(u) is the Dirac function δ(τ − t). with the average delay T = n/α. . is a weighting factor which indicates how much emphasis should be given to the size of the population at earlier times to determine the present eﬀect on resource availability. The model is an integrodiﬀerential equation 1 t dx = rx 1 − G(t − s)x(s)dx . δ(t − τ − s)x(s)dx = rx(t) 1 − dt K −∞ K The average delay for the kernel is deﬁned as ∞ uG(u)du. (9. the discrete delay. 0 In this way we ensure that for equation (9. .497 DDE in Single Species Dynamics the causes was the pollution of the closed environment by waste products and dead organisms. where ∞ δ(τ − s)f (s)ds = f (τ ). are called weak delay kernel and strong delay kernel. G(u) = αe−αu (n = 1). (9. .1) the equilibrium of the instantaneous logistic equation (1. G(u) = α2 ue−αu . We usually use the Gamma distribution delay kernel G(u) = αn un−1 e−αu . 2.
1) is called stable if given any > 0 there exists a δ = δ() > 0 such that φ(t) − x∗  ≤ δ for t ∈ (−∞.13. 0] implies that any solution x(t) of (9. then x∗ is called t→∞ asymptotically stable.13).4) . 11..1) and (9. = rx(t) 1 − K −∞ dt (9. continuity and continuation about solutions to such a kind of integrodiﬀerential equations. If in addition there exists a constant δ0 > 0 such that φ(t) − x∗  ≤ δ on (−∞.1) is x(θ) = φ(θ) ≥ 0. −∞ < θ ≤ 0. The initial value for the integrodiﬀerential equation (9. 9. An equilibrium x∗ of equation (9.3) exists and satisﬁes x(t)− x∗  < for all t ≥ 0. 0] implies lim x(t) = x∗ . Following Volterra (1931) or Miller (1971).498 DELAY DIFFERENTIAL EQUATIONS (b) (a) α 0 u Figure 11. we can obtain existence. 1 t dx −α(t−s) αe x(s)ds . i. (9. On the other hand the strong kernel means that the maximum inﬂuence on growth rate response at any time t is due to population density at the previous time t − T (see Fig. 0 T u (a) Weak delay kernel and (b) strong delay kernel.3) where φ(θ) is continuous on (−∞. let us ﬁrst consider the equation with a weak kernel. of the growth rate is due to current population density while past densities have (exponentially) decreasing inﬂuence.1 Weak Kernel To determine the stability of x∗ = K.e. 0]. uniqueness.
6) is (x∗ .8).7) α 1 2 ± α − 4αr. y ∗ ). let X = x − x∗ .8) Choose a Liapunov function as follows V (x.9) Along the solutions of (9. 2 2 < 0.14). The characteristic equation of the linearized system is given by λ2 + αλ + αr = 0. dt Notice that the positive equilibrium of system (9.1). approach (x∗ . To determine the stability of (x∗ . it follows that solutions of system (9. Reλ1.5) −∞ Then the scalar integrodiﬀerential equation (9.1) as follows: dx = − r x(t)(y(t) − y ∗ ). 11.6) dy = αx(t) − αy(t). Y = y − y ∗ . . y) = x − x∗ − x∗ ln x r (y − y ∗ )2 . y ∗ ) = (K.2 cally stable. we have dx x − x∗ r r dy dV = + (y − y ∗ ) = − (y − y ∗ )2 < 0. K dt dy = α(x(t) − x∗ ) − α(y(t) − y ∗ ). K). dt (9. Theorem 7 The positive equilibrium x∗ = K of the logistic equation ( 9. Therefore.499 DDE in Single Species Dynamics Using the linear chain trick (Fargue (1973) and MacDonald (1978)). (9. x∗ = K is globally stable.8). x(t) → x∗ as t → ∞. Rewrite (9. which implies that x∗ = K is locally asymptotiλ1.4) is equivalent to the following system of two ordinary diﬀerential equations dx = rx(t) 1 − 1 y(t) . deﬁne t y(t) = αe−α(t−s) x(s)ds. The above analysis can be summarized as the following theorem. and hence of (9. In fact. y ∗ ) as t → ∞. which has roots (9. dt dt x αK dt K Since the positive quadrant is invariant. + ∗ x 2αK (9. K dt (9.2 = − Therefore.1) with a weak kernel is globally stable (see Fig.
y ∗ .15.2 0 50 100 t .time 150 200 Figure 11.10) dt K −∞ To use the linear chain trick. Then equation (9. dt (9. K = 1.500 DELAY DIFFERENTIAL EQUATIONS x(t) 1. z(t) = −∞ t −∞ α2 (t − s)e−α(t−s) x(s)ds.1) with a strong kernel.6 0. (9. 9. y ∗ . K.8 alpha=1. Considering the linearization of (9.4 0.11) which has a positive equilibrium (x∗ . z ∗ ). we obtain the characteristic equation (9.10) is equivalent to the system dx = rx(t) 1 − 1 z(t) . We shall see that the logistic equation with a strong kernel exhibits richer dynamics similar to the logistic equation with a constant delay. K dt dy = αx(t) − αy(t).e. . deﬁne t −α(t−s) y(t) = αe x(s)dx. dt dz = αy(t) − αz(t).1) is globally stable. The steady state of the integrodiﬀerential equation (9. The result indicates that if the delay kernel is a weak kernel. i. ∗ ) = (K.12) λ3 + a1 λ2 + a2 λ + a3 = 0..14.25 0. K). the logistic equation with distributed delay has properties similar to the instantaneous logistic equation.2 Strong Kernel Consider the logistic equation (9.125 1 0.11) at (x∗ . Here r = 0.00. dx 1 t = rx(t) 1 − α2 (t − s)e−α(t−s) x(t − s)ds .2 alpha=0.
11) and hence equation (9. (9. (9. a3 = rα2 . When T = T0 = 4/r.501 DDE in Single Species Dynamics where a1 = 2α. and the periodic solutions exist for T > 4/r and are orbitally stable. r Therefore.10) exhibits a Hopf bifurcation as the average delay T passes through the critical value T0 = 4/r (Marsden and McKraeken (1976)). The RouthHurwitz criterion says that all roots of the equation (9.12) has a negative real root λ1 = −r and a pair of purely imaginary roots λ2.12) such that u(T0 ) = 0. Inequality (9.15) T < . Clearly. We can verify that 8 dµ = r2 > 0. 2 Thus.3 = ±i 2r . a Hopf bifurcation occurs at x∗ = K and a family of periodic solutions bifurcates from x∗ = K. As the (average) delay is increased through a critical value the positive equilibrium passes from stability to instability. 11.15).10) is asymptotically stable if the average delay T = 2/α < 4/r and unstable if T > 4/r. Denote by λ(T ) = µ(T ) + iν(T ) the complex eigenvalue of (9. Theorem 8 The positive equilibrium x∗ = K of equation ( 9. a2 = α2 . the period of the bifurcating solutions is νπ0 = 2π r . the equilibrium x∗ = K is stable for “short delays” (T < 4/r) and is unstable for “long delays” (T > 4/r). (9. . Thus. exhibits a typical bifurcation phenomenon.14) then becomes 4 (9. We thus have the following theorem regarding equation (9. the equilibrium x∗ = K is stable if α > r/2 and unstable if α < r/2.13) rα2 > 0. a1 = 2α > 0. a1 a2 − a3 > 0.12) have negative real parts if and only if the following inequalities hold: a1 > 0. When T = 4/r. the logistic equation with a strong delay kernel. Note that the average delay of the strong kernel is deﬁned as T = 2/α.16) thus implies that system (9.16) dT 4 5 T=r The transversality condition (9. The last inequality becomes r (9. accompanied by the appearance of stable periodic solutions (see Fig. a3 = a3 > 0. ν(T0 ) = r/2.10).14) α> . just like the logistic equation with a discrete delay.
9.17) (9.10) losses stability and a Hopf bifurcation occurs when α changes from 0. . r = K of ( 9.5 3 2.5 alpha=0.time 300 350 400 450 500 Figure 11.065.18) have negative real parts. Here r = 0.00. Let X = x − K. Theorem 9 If 0 then x∗ ∞ 1 sG(s)ds < .15.65 to 0.17).18) 0 If all eigenvalues of the characteristic equation (9. The steady state x∗ = K of the integrodiﬀerential equation (9.1) is asymptotically stable.1) with a general kernel. Then (9.1) can be written as t t dX = −r G(t − s)X(s)ds + rX(t) G(t − s)X(s)ds.065 4 3.5 alpha=0. that is.5 0 0 50 100 150 200 250 t . the equilibrium x∗ = K of (9.1).3 General Kernel Now consider the stability of the equilibrium x∗ = K for the integrodiﬀerential equation (9.502 DELAY DIFFERENTIAL EQUATIONS x(t) 4. dt −∞ −∞ The linearized equation about x = K is given by t dX G(t − s)X(s)ds = −r dt −∞ and the characteristic equation takes the form ∞ G(s)e−λs ds = 0.5 2 1.65 1 0. K = 1.15. then the solution X = 0 of (9. is asymptotically stable. λ+r (9.
Img(ε + iν) is seen to increase . ε) that constitutes the boundary of the region {λ ε ≤ Reλ ≤ a. 0 Because a > r. that g(λ) does not encircle 0 as λ traverses Γ(a. g(λ) must assume at least one positive real value (and no negative values) as λ travels clockwise from ε + ia to ε − ia along Γ. Since the roots of (9. In fact. we have ∞ g(µ + ia) = µ + ia + r G(s)e−(µ+ia)s ds. Img(µ + ia) is always positive here. By continuity. one can show that Img(µ − ia) is negative along the segment λ = µ − ia. ε ≤ µ ≤ a. ε). we may conclude that every real value assumed by g(λ) along this segment must be positive. Under the assumption. it follows that 0 ∞ −(a+iν)s G(s)e ds ≤ ∞ G(s)ds = 1.18) coincide with the zeros of the function ∞ g(λ) = λ + r G(s)e−λs ds. we have ∞ g(a + iν) = a + iν + r G(s)e−(a+iν)s ds. Along the segment of Γ given by λ = a + iν. Since the zeros of g(λ) are isolated. ε ≤ µ ≤ a. 0 we may apply the argument principle to g(λ) along the contour Γ = Γ(a. The argument principle now states that the number of zeros of g(λ) contained in the region bounded by Γ is equal to the number of times g(λ) wraps Γ around the origin as λ traverses Γ.503 DDE in Single Species Dynamics Proof. Finally. it suﬃces to show for all small ε > 0 and all large a > r. 0 A similar argument shows g(λ) to assume no real value along this path. 0 Since a > 0. we may choose a and ε so that no zeros of g(λ) lie on Γ. consider the path traced out as λ = ε + iν increases from ε − ia to ε + ia.) Thus. Similarly. −a ≤ µ ≤ a. Along the segment of Γ given by λ = µ + ia. −a ≤ Im ≤ a. 0 < ε < a}. (A zero of g(λ) of multiplicity m is counted m times.
9. is a contradiction since the positivity of α shows g(λ) to have no real positive zeros. ∞ d G(s)e−εs sin(νs)ds ν+r dν ∞ 0 = 1+r sG(s)e−εs cos(νs)ds 0 ∞ sG(s)ds ≥ 1−r d Img(ε + iν) = dν 0 > 0. g(λ) does not encircle the origin.1) is the following equation t dx G(t − s)x(s)ds . the real value assumed by g(λ) along this last segment must be positive. that real value is nonzero. In fact. predicting exactly one zero λ0 of g(λ) inside the region bounded by Γ. It follows immediately that g(λ) assumes precisely one real value along this last segment of Γ.4 Remarks In studying the local stability of equation (9. (λ + α)2 0 which is equation (9. αe−(λ+α) sds = λ + λ+r λ+α 0 which is equation (9. Therefore.18) becomes ∞ α2 r λ+r α2 se−(λ+α) sds = λ + = 0.12). This completes the proof. One of the varieties of equation (9. If G is a weak kernel. If G is a strong kernel. we applied the socalled linear chain trick to transform the scalar intergrodiﬀerential equation into equivalent a system of ﬁrst order ordinary diﬀerential equations and obtained the characteristic equations (9. Since no zero of g(λ) lies on Γ. Since α and G are real. forcing λ0 to be real.7).18).7) and (9. (9. however. This. g(λ) would have wrapped Γ once about the origin. then (9.19) = x(t) a − bx(t) − c dt −∞ .504 DELAY DIFFERENTIAL EQUATIONS monotonically with ν.1) with weak and strong kernels. then (9. Thus.18) becomes ∞ αr = 0.12). the zeros of g(λ) occur in complex conjugate pairs. Assuming it to be negative. It should be pointed out that these characteristic equations can be derived directly from the characteristic equation (9.
b ≥ 0. Pianka (1974) discussed the relevance of periodic environment to evolutionary theory. It has been suggested by Nicholson (1933) that any periodic change of climate tends to impose its period upon oscillations of internal origin or to cause such oscillations to have a harmonic relation to periodic climatic changes. Rosen (1987) noted the existence of a relation between the period of the periodic carrying capacity and the delay of the logistic equation. Landman (1980) showed that the exists a positive a∗ such that for a = a∗ . a steady state becomes unstable and oscillatory solutions bifurcate for a near a∗ . MacDonald (1978) and references cited therein. 2 (10.1) which is globally attractive with respect to all other positive solutions. Theorem 10 Suppose that 0 nω 3 r(s)ds ≤ . 10. (10. See Corollary 13. We refer to Miller (1966). c ≥ 0.1) dt K(t) where r(t + ω) = r(t). seasonal eﬀects of weather. Cushing (1977a). Namely. mating habits. For example. K(t + ω) = K(t) for all t ≥ 0. 10. etc. Stability and bifurcation of equation (9. food supplies. Zhang and Gopalsamy (1990) assumed that the intrinsic growth rate and the carrying capacity are periodic functions of a period ω and that the delay is an integer multiple of the period of the environment.2) . b+c = 0. they considered the periodic delay diﬀerential equation of the form x(t − nω) dx = r(t)x(t) 1 − .1 Periodic Delay Models Nisbet and Gurney (1976) considered a periodic delay logistic equation and carried out a numerical study of the inﬂuence of the periodicity in r and K on the intrinsic oscillations of the equation such as those caused by the time delay.).. See also Simpson (1980). They proved the following result on the existence of a unique positive periodic solution of equation (10.g. Periodicity If the environment is not temporally constant (e.505 DDE in Single Species Dynamics where a > 0.19) when other coeﬃcients (not necessarily the average delay) are chosen as bifurcation parameters.19) have been studied by many researchers. then the parameters become time dependent. It should be pointed out that bifurcations can occur in equation (9.
The coincidence degree theory (Gaines and Mawhin (1977)) has also been used to establish the existence of periodic solutions in periodic models with general periodic delays. etc. if . they proved the following theorem. See. Freedman and Wu (1992) considered the following singlespecies model with a general periodic delay dx = x(t)[a(t) − b(t)x(t) + c(t)x(t − τ (t))]. and when the population size is not so small.4) dt where the net birth rate a(t) > 0. quite a few papers have been produced by reconsidering the delayed models which appeared in the previous sections with the assumption that the coeﬃcients are periodic.4) has a positive ωperiodic solution Q(t). the positive feedback is a(t) + c(t)x(t − τ (t)) while the negative feedback is b(t)x(t). growth is proportional to the size.2) (Yan and Feng (2003)). (1996)). the periodic WazewskaCzyzewska and Lasota model (6. and the delay τ (t) ≥ 0 are continuously diﬀerentiable.1)(Lalli and Zhang (1994)) and (4. ωperiodic functions on (−∞. the periodic foodlimited model (5. Moreover. the selfinhibition rate b(t) > 0. Using ﬁxed point theorem and Razuminkin technique. This model represents the case that when the population size is small. continuously diﬀerentiable solution K(t). ∞).3) (Gopalsamy et al. ωperiodic. Then equation ( 10. for example. Such circumstance could arise when the resources are plentiful and the reproduction at time t is by individuals of at least age τ (t) units of time.1) (Saker and Agarwal (2002)). (1990a)).506 DELAY DIFFERENTIAL EQUATIONS Then the periodic delay logistic equation ( 10.1) has a unique positive solution x∗ (t) and all other solutions of ( 10. t→∞ (10. φ(0) > 0.2) (Greaf et al. (10. However uniqueness is not guaranteed and stability can be obtained only when the delays are constant (Li (1998)). the reproduction rate c(t) ≥ 0. Theorem 11 Suppose that the equation a(t) − b(t)K(t) + c(t)K(t − τ (t)) = 0 has a positive. In all these papers. the delays are assumed to be integral multiples of periods of the environment.3) Following the techniques of Zhang and Gopalsamy (1990). the periodic Nicholson’s blowﬂies model (3.1) corresponding to initial conditions of the form x(θ) = φ(θ) ≥ 0. the periodic Allee eﬀect models (4. 0] satisfy lim x(t) − x∗ (t) = 0. φ ∈ C[−nω.
(10.4). Notice that in equation (10. then Q(t) is globally asymptotically stable with respect to positive solutions of ( 10. where Γ = {x ∈ Cω : a − c(G ∗ x) > 0}. Since a > 0. the periodic delays can have either a stabilizing eﬀect or a destabilizing one.6) dt −∞ where a > 0. It is still an open problem to study the dynamics. that is.507 DDE in Single Species Dynamics b(t) > c(t)Q(t − τ (t))/Q(t) for all t ∈ [0. 10. Bardi and Schiaﬃno (1982) considered the integrodiﬀerential equation (9.4).5) dt K(t) where τ (t) a positive periodic function. x(t) ≡ 0 belongs to Γ. b(t) has to be greater than zero. ω].2 Integrodiﬀerential Equations Periodic logistic equations with distributed delay have been systematically studied in Cushing (1977a). a = ω 0 The convolution of the kernel G and a bounded function f is deﬁned by t G(t − s)f (s)ds. t ∈ R. So Theorem 11 does not apply to the periodic delay logistic equation x(t − τ (t)) dx = r(t)x(t) 1 − (10.6) is a ﬁxed point of the operator B : Γ → Cω deﬁned by (Bx)(t) = u(t). For a ∈ Cω . deﬁne the average of a as 1 ω a(s)ds. such as existence. for the periodic delay logistic equation (10. uniqueness and stability of periodic solutions and bifurcations. (G ∗ f )(t) = −∞ Observe that an ωperiodic solution of (10. As Schley and Gourley (2000) showed. b > 0. Γ is not empty. that is. R) denote the Banach space of all ωperiodic continuous functions endowed with the usual supremum norm x = sup x(t).5). Let Cω = Cω (R.1) when the coeﬃcients are periodic. . t dx = x(t) a(t) − b(t)x(t) − c(t) G(t − s)x(s)ds . Deﬁne u0 (t) = (B0)(t). c ≥ 0 are ωperiodic continuous functions on R and G ≥ 0 is a normalized kernel. depending on the frequency of the periodic perturbation.
which implies that v(t) ≥ 0 for all t ≥ t0 . Claim IV. lim inf (G ∗ z)(t) ≥ l − . we have αi = bui because ui (t)(i = 1. If t > t . Setting v(t) = u1 (t) − u2 (t). t→∞ t→∞ . Thus. Let z be a bounded continuous function on R. Claim III. Then lim inf (v(t) − u(t)) > 0 implies lim inf ((Bu)(t) − v(t)) > 0. then c(G ∗ v) = v(G ∗ c) . Then α1 (t) ≥ α2 (t). u2 (t0 ) ≤ u1 (t0 ). By the periodicity of v(t). we have (j+1)ω +∞ ω c(t) G(t − s)v(s)dsdt c(G ∗ v) = jω j=−∞ 0 +∞ = ω j=−∞ 0 +∞ = ω c(t) G(t − s − jω)v(s)dsdt 0 ω (1−j)ω v(t) −jω j=−∞ 0 G(t − s)c(s)dsdt = v(G ∗ c) . then Bx2 ≤ Bx1 . Let l = lim inf t→∞ z(t). Choose > 0 and pick t such that z(t) > l − for any t > t . we deduce that bu1 ≥ bu2 and for some t0 ∈ R. Claim II. t Hence. If x1 and x2 belong to Γ with x1 ≤ x2 . 2) are periodic. Then lim inf (G ∗ z)(t) ≥ lim inf z(t). In fact. we have Bx2 ≤ Bx1 .6).508 DELAY DIFFERENTIAL EQUATIONS Claim I. t→∞ t→∞ t→∞ t→∞ We only prove the ﬁrst inequality. t→∞ which implies the ﬁrst inequality. we have t t G(t − s)z(s)ds + G(t − s)z(s)ds (G ∗ z)(t) = −∞ ≥ inf z(t) t t t −∞ G(t − s)ds + (l − ) t G(t − s)ds. Let u ∈ Γ and let v(t) > 0 be the solution of (10. In fact. we have v(t) ˙ ≥ (α1 (t) − b(t)(u1 (t) + u2 (t)))v(t). let αi (t) = a(t) − c(t)(G ∗ xi )(t) and ui (t) = (Bxi )(t) for t ∈ R(i = 1. if we deﬁne G(t) = 0 for t < 0. If v and c belong to Cω . Since αi (t) = u˙ i (t)/ui (t) + b(t)ui (t). t→∞ t→∞ lim sup(v(t) − u(t)) < 0 implies lim sup(Bu)(t) − v(t)) < 0. lim sup(G ∗ z)(t) ≤ lim sup z(t). 2).
t t 1 t exp{ β(θ)dθ}c(s)v(s)ds. the periodicity of u0 (t) and Claim II imply that a = bu0 > c(G∗u0 ) .509 DDE in Single Species Dynamics We prove the ﬁrst statement. t0 where γ3 > 0 is a suitable constant. Deﬁne z(t) = w(t) − v(t). If b(t) > (G ∗ c)(t) (10. then equation ( 10. As u0 > 0.6) has a unique positive ωperiodic solution x∗ (t) which is globally asymptotically stable with respect to all solutions of equation ( 10. Because w(t)/w(t) ˙ is periodic and its average is zero. 0]. b(t)v(t) is positive and bounded. there exists a t0 ∈ R. Thus. Then lim inf t→∞ z(t) ≥ γ3 /γ which implies the ﬁrst statement. Because of Claim III. Since u˙ 0 (t)/u0 (t) = a(t) − b(t)u0 (t). such that z(t) ˙ > (w(t)/w(t) ˙ − b(t)v(t))z(t) + lc(t)v(t)/2 for all t > t0 . Proof. Let l = lim inf t→∞ (v(t) − u(t)). Let w(t) = (Bu)(t). θ ∈ (−∞. ω]. t ∈ R. we can see that t t0 β(s)ds > γ1 − tγ2 . Then w(t) is a solution of w(t) ˙ = a(t)w(t) − b(t)w(t)2 − c(t)w(t)(G ∗ u)(t) while v(t) ˙ = a(t)v(t) − b(t)v(t)2 − c(t)v(t)(G ∗ v)(t). we have Bu0 ≤ u0 . where γ1 and γ2 > 0 are constants. φ(0) > 0.7) for any t ∈ [0. that is. Theorem 12 Suppose a > 0. Then z(t) ˙ = (a(t) − b(t)w(t) − c(t)(G ∗ u))z(t) + c(t)v(t)(G ∗ (v − u))(t) = (w(t)/w(t) ˙ − b(t)v(t))z(t) + c(t)v(t)(G ∗ (v − u))(t). z(t) > z(t0 ) exp{ β(s)ds} + 2 t0 t0 s where β = w(t)/w(t) ˙ − b(t)v(t). for any v ∈ Cω satisfying . Therefore.6) under initial condition x(θ) = φ(θ). t z(t) > γ3 exp((s − t)γ)ds = (γ3 /γ)(1 − exp((t0 − t)γ2 )).
Deﬁne un (t) = (B n u0 )(t) = (Bun−1 )(t).. by Claim I.510 DELAY DIFFERENTIAL EQUATIONS 0 < v ≤ u0 . we have 0 < Bu0 ≤ Bv ≤ u0 . To prove the global stability. u˙ + (t) = (a(t) − c(t)(G ∗ u− )(t))u+ (t) − b(t)u+ (t)2 . v(b − G ∗ c) = 0. we have a − c(G ∗ u+ ) − bu− = a − c(G ∗ u− ) − bu+ followed by the fact that ln u+ and ln u− are periodic. Then we have c(G ∗ v) = bv . n→∞ that u− (t) If we can show = u+ (t) = u∗ (t). 1.. Now by Claim II we have c(G ∗ v) = v(G ∗ c) . we have u˙ − (t) = (a(t) − c(t)(G ∗ u+ )(t))u− (t) − b(t)u− (t)2 . 2. Then u− (t) = lim u2n+1 (t) and u+ (t) = lim u2n (t) u− (t) n→∞ ≤ u+ (t). Since 0 < B 2 0 = Bu0 . Moreover. By the deﬁnition. In fact. By the monotonicity and uniform boundedness of {un } we have the L2 −convergence of both u2n+1 and u2n and their derivatives. the set Γ0 = {v ∈ Cω : 0 < v ≤ u0 } ⊂ Γ is invariant under B. Bu0 ≤ Bv ≤ u0 ⇒ Bu0 ≤ B 2 v ≤ B 2 u0 ⇒ B 3 u0 ≤ B 3 v ≤ B 2 u0 and by induction B 2n+1 u0 ≤ B 2n+1 v ≤ B 2n u0 .6) satisﬁes lim inf t→∞ v(t) > 0. we have v(t) ˙ < a(t)v(t) − b(t)v(t)2 and lim sup(v(t) − (Bu)(t)) ≤ 0. which implies that v ≡ 0 by the assumption (10. Therefore. Hence. Taking the limits. n = 0.6).7). B 2n+1 u0 ≤ B 2n+2 v ≤ B 2n+2 u0 . ﬁrst we show that any solution v(t) of equation (10. we have u˙ n (t) = (a(t) − c(t)(G ∗ un−1 )(t))un (t) − b(t)un (t)2 . . Let v(t) = u+ (t)− u− (t). t→∞ . exist with 0 < ∗ it is easy to see that u (t) is the unique ﬁxed point of B. Dividing them by u− (t) and u+ (t) respectively. u∗ (t) is a unique periodic solution of the equation (10. Hence. we know that {B 2n+1 u0 } is increasing and {B 2n u0 } is decreasing.
19) (with constant coeﬃcient) is globally stable with respect to positive solutions of ( 9. then the positive equilibrium x∗ = a/(b + c) of equation ( 9. by Claim III. Karakostas (1982) and the references therein.19). If a. b and c are real positive constants. For other related work on periodic logistic equations with distributed delay.1) leads to a statedependent delay model of the form b(x(t)) − b(x(t − L[x(t)]) dx = . Corollary 13 If b > c and G satisﬁes the above assumptions.511 DDE in Single Species Dynamics Choose > 0 so that u(t) = u0 (t) + ∈ Γ. Cushing (1977a). Assume that the lifespan L of individuals in the population is variable and is a function of the current population size. we have (B´elair (1991)) t b(x(s))ds. 11. If we take into account the crowding eﬀects. dt 1 − L [x(t)] b(x(t − L[x(t)]) (11.2) . then condition (10. (11. Cohen and Rosenblat (1982). Assume that the number of births is a function of the population size only. This is the main result in Miller (1966). Since (B 2n+1 u0 )(t) < v(t) < (B 2n u0 )(t) for large t. lim inf (v(t) − (B 2n+1 u0 )(t)) > 0. Since the population size x(t) is equal to the total number of living individuals. The birth rate is thus density dependent but not age dependent. we have lim inf t→∞ v(t) > 0. Thus. StateDependent Delays Let x(t) denote the size of a population at time t. then L(·) is a decreasing function of the population size.1) x(t) = t−L[x(t)] Diﬀerentiating with respect to the time t on both sides of equation (11. t→∞ Since (Bu)(t) is strictly positive and periodic. t→∞ t→∞ Given ε > 0.7) becomes b > c. we refer to Bardi (1983). By Claim IV we have lim inf (v(t) − (Bu)(t)) ≥ . lim inf t→∞ (u0 (t) − v(t)) > 0 and by induction. lim sup(v(t) − (B 2n u0 )(t)) < 0. choose n such that u∗ (t) − ε < (B 2n+1 u0 )(t) < (B 2n u0 )(t) < u∗ (t) + ε. it follows that the sequence {B j u} tends to u∗ uniformly as j → ∞.
1).00 and τ (x) = a + bx2 .16.2) but clearly it may not necessarily be a solution of (11.2) is not equivalent to the integral equation (11.1. Choose r = 0.1541 1. We observe stability of the equilibrium x = K = 1 for small amplitude of τ (x) and oscillations about the equilibrium for large amplitude of τ (x) (see Fig. 11.1) and τ (x) = a + bx2 .16) show that the logistic model with a statedependent delay x(t − τ (x(t)) .1). Numerical simulations for the statedependent delay model (11. (11.2) but the reverse is not true. Recently. K = 1.6 0 50 100 150 t . Numerical simulations (see Fig.16).6 1.15. 11. K = 1. great attention has been paid on the study of statedependent delay equations. See also Cooke and Huang (1996). and (ii) a = 9.time 200 250 300 Figure 11. Consider εx (t) = −x(t) + f (x(t − r(x(t)))) and assume that (11.3 for the Hutchinson’s model (2.8 0.4) .15.1541.2 a=5 1 0.2) has been studied by B´elair (1991). 11. (1970) to model the control of the bone marrow stem cell population which supplies the circulating red blood cell population.1).1. See also Mackey and Milton (1990). The asymptotic behavior of the solutions of (11. any constant function is a solution of (11.3) with r = 0.3) x (t) = rx(t) 1 − K has similar dynamics to the Hutchinson’s model (2.4 1. b = 1.512 DELAY DIFFERENTIAL EQUATIONS x(t) 1.00 as in Fig. (i) a = 5. b = 1. Note that statedependent delay equation (11.8 a=9. It is clear that every solution of (11.1) is a solution of (11. In fact. Statedependent delay models have also been introduced by Kirk et al.
Walther (2002) for existence of periodic solutions. 2003) studied the shape of general periodic solutions of the equation (11. MalletParet et al. MalletParet and Nussbaum (1996.4) is called an slowly oscillating periodic (SOP) solution if there exist numbers q1 > 1 and q2 > q1 + 1 such that = 0. then uf (u) < 0 for all u ∈ [−B. q 1 = 0.DDE in Single Species Dynamics 513 (H1) A and B are given positive real numbers and f : [−B. (1994). and to Ait Dads and Ezzinbi (2002) and Li and Kuang (2001) for almost periodic and periodic solutions to statedependent delay equations. r : [−B. Magal and Arino (2000). (2002) for semigroup property of the solutions. MalletParet and Nussbaum (1992) proved the following theorem. to Bartha (2001) and Chen (2003) for convergence of solutions. A] → [−B. (2001) for a brief review on statedependent delay models. t = 0. (H1 ) B is a positive real number and r : [−B. ∞) → R is a locally Lipschitz map with r(0) = 1. t = q2 and x(t + q2 ) = x(t) for all t. A] → R is a Lipschitz map with r(0) = 1 and r(u) ≥ 0 for all u ∈ [−B. be the unique solution of cos ν0 = −1/k and deﬁne λ0 = ν0 / k 2 − 1. Kuang and Smith (1992a. A periodic solution x(t) of (11. A]. . 1992b). (1998). r(u) ≥ 0 for all u ≥ −B and r(−B) = 0. x(t) < t < q2 .4) and their limiting proﬁle as ε → 0+ . See also Arino et al. Theorem 14 Assume that f and r satisfy (H1)(H2) or (H1 )(H2 ). < 0. (H2 ) f : R → R is a locally Lipschitz map. (H2) For A and B in (H1). We refer to Arino et al. π/2 √< ν0 < π. to Louihi et al. A]. Then for each λ > λ0 the equation x (t) = −λx(t) + λf (x(t − r(x(t)))) has an SOP solution xλ (t) such that −B < xλ (t) < A for all t. Let ν0 . u = 0. = 0. and if B is as in (H1 ) and A = sup{f (u) : −B ≤ u ≤ 0}. Bartha (2003). Suppose that f is in C 1 near 0 and f (0) = −k < −1. t = q1 . A]. A] is a Lipschitz map with xf (x) < 0 for all x ∈ [−B. 0 < t < q1 . x = 0. to Krisztin and Arino (2001) for the existence of two dimensional attractors. > 0.
When the domain is small. In terms of randomness. Fisher (1937) proposed the model to investigate the spread of an advantageous gene in a population. Recall that for the spatially uniform logistic equation u ∂u = ru 1 − . chemicals and so on. cells. When this microscopic irregular movement results in some macroscopic regular motion of the group. t) represent the population density at location x and time t and the source term f represents the birthdeath process. . (a) In a ﬁnite domain with zeroﬂux (Neumann) boundary conditions.1 Fisher Equation Let u(x. the conclusions still hold (Fife (1979) and Britton (1986)). bacteria. the onedimensional scalar reactiondiﬀusion equation takes the form (Fisher (1937) and Kolmogorov et al. ∂t ∂x K a < x < b. but it losses its stability when the domain exceeds a certain size and a nontrivial steadystate solution becomes asymptotically stable. u = K is no longer a solution to the problem. How does the introduction of the diﬀusion aﬀect these conclusions? The answer depends on the domain and the boundary conditions. spreads as a whole according to the irregular motion of each particle. u = 0 is asymptotically stable. With the logistic population growth f = ru(1−u/K) where r is the linear reproduction rate and K the carrying capacity of the population.17). (c) In an inﬁnite domain the Fisher equation has travelling wave solutions (see Figure 11. (12.514 12. 0 ≤ t < ∞. (b) Under the Dirichlet conditions. DELAY DIFFERENTIAL EQUATIONS Diﬀusive Models with Delay Diﬀusion is a phenomenon by which a group of particles.1) which is called the Fisher equation or diﬀusive logistic equation. In this case the behaviour of solutions depends on the size of the domain. the phenomenon is a diﬀusion process. ∂t K the equilibrium u = 0 is unstable while the positive equilibrium u = K is globally stable. for example animals. 12. (1937)) ∂2u u ∂u = D 2 + ru 1 − . diﬀusion is deﬁned to be a basically irreversible phenomenon by which particles spread out within a given space according to individual random motion.
Feng and Lu (1996). Gourley and Chaplain (2002) considered the case when the delay is ﬁnite. For instance. Simpson (1980). The diﬀusive logistic equations with a distributed delay of the form ∂u(x. Tesei (1980). Tesei (1981). Schiaﬃno and A. . Busenberg and Huang (1996). etc.6 0. c = 2. the diﬀusive logistic equations with a discrete delay of the form ∂u(x. Redlinger (1985).2) ∂t ∂x2 K with either Neumann or Dirichlet boundary conditions have been investigated by Green and Stech (1981). t) ∂ 2 u(x.2 Diﬀusive Equations with Delay In the last two decades. Gourley and Britton (1993). Bonilla and Li˜ na´n (1984).515 DDE in Single Species Dynamics u(x.2 0 0 5 10 z=xct 15 20 Figure 11.t) 1 0.3) have been studied by Schiaﬃno (1979).3) with diﬀusion. researchers have studied the combined eﬀects of diﬀusion and various delays on the dynamics of the models mentioned in previous sections. Lin and Khan (1982).17. Recently. For example. Freedman and Zhao (1997). Here D = r = K = 1. t) 1 − (12. etc. for the foodlimited model (5. Morita (1984). Huang (1998). Yamada (1993). t) 1 t =D + ru(x.1).4 0. t) u(x. Luckhaus (1986). Faria and Huang (2002).0 12. Pao (1997).4 − 3. The traveling front proﬁles for the Fisher equation (12.8 0. diﬀusive biological models with delays have been studied extensively and many signiﬁcant results have been established. Britton (1990). t) 1 − G(t − s)u(x. Gopalsamy and Aggarwala (1981). t) ∂ 2 u(x. Yoshida (1982). t − s)ds ∂t ∂x2 K −∞ (12. t − τ ) =D + ru(x.
x) = d∆u(t.4) with Dirichlet boundary conditions has been investigated by So and Yang (1998). Gourley and Ruan (2000) studied various local and global aspects of Nicholson’s blowﬂies equation with inﬁnite delay ∂u = d∆u − τ u(x. x)+b[1−u(t. y ∈ Ω. y) is a positive continuous function in its variables t ∈ R. x. 0 Ω Ruan and Xiao (2004) studied the stability of the steady states and proved the following results.7) Ω where u(t.6) Rn 0 0 Gourley (2000) discussed the existence of travelling waves in equation (12. x. t) + βτ u(x. and the convolution kernel F (t. x.5) for (x. where Ω is either all of or some ﬁnite domain. s. x)] ∂t t −∞ F (t−s. x) is the normalized spatial density of infectious host at time t ∈ R+ in location x ∈ Ω.516 DELAY DIFFERENTIAL EQUATIONS Feng and Lu (2003) assumed that the time delay in an integral multiple of the period of the environment and considered the existence of periodic solutions. t) ∈ Ω × [0. t − 1) exp[−u(x. Some numerical and bifurcation analysis of this model has been carried out by So. Wu and Yang (1999) and So and Zou (2001). s. (12. s) ds exp − t −∞ F (t − s)u(x. t) + βτ ∂t t −∞ F (t − s)u(x. and the kernel satisﬁes F (t) ≥ 0 and the conditions ∞ ∞ F (t) dt = 1 and t F (t) dt = 1. ∞).5). s ∈ R+ . The diﬀusive Nicholson’s blowﬂies equation ∂u = d∆u − τ u(x. s) ds (12. . t − 1)] ∂t (12. Ruan and Xiao (2004) considered the diﬀusive integrodiﬀerential equation modeling the hostvector interaction ∂u (t. x)−au(t. Ω is an open bounded set in RN (N ≤ 3). which is normalized so that ∞ F (t. y)dyds. y)dyds = 1. They studied the global attractivity of the positive steady state of the equation. y)u(s. and Gourley and So (2002) investigated the dynamics when the delay is nonlocal. (12. Davidson and Gourley (2001) studied the model with inﬁnite delay. x.
s. t t − s − t−s ∂u = d∆u(t. the existence of traveling waves has been established.7) becomes the vector disease model (7. The existence of traveling front solutions show that there is a moving zone of transition from the diseasefree state to the infective state. Notice that when F (t. and u does not depend on the spatial variable. ∞) and the kernel is a local strong kernel. then u0 = 0 is the unique steady state solution of ( 12. When x ∈ (−∞. 2 ∂t −∞ τ (12. (ii) If 0 ≤ a < b. x)] e τ u(s. The author would like to thank Stephen Cantrell for comments and corrections on an earlier version of the paper.7) in ¯ M = {u ∈ E : 0 ≤ u(x) ≤ 1. x ∈ Ω} and it is globally asymptotically stable in C((−∞. i.e. x) − au(t. where τ > 0 is a constant.1) and Theorem 15 reduces to Theorem 5 obtained by Cooke (1978). then equation (12.8) where (t.DDE in Single Species Dynamics 517 Theorem 15 The following statements hold (i) If 0 < b ≤ a. x)ds. x) + b[1 − u(t. x. 0]. .8) has a traveling wave solution connecting u0 = 0 and u1 = (b − a)/b. 0]. Theorem 16 For any τ > 0 suﬃciently small there exist speeds c such that the system ( 12. y) = δ(x − y)δ(t − s − τ ). x) ∈ R+ × Ω. We refer to the monograph by Wu (1996) for a systematic treatment of partial diﬀerential equations with delay. M ). where u0 is unstable and u1 is globally asymptotically stable in C((−∞. then there are two steady state solutions in M : u0 = 0 and u1 = (b − a)/b. M ). Acknowledgements.
s]. 63.ac. it has been studied the wellposedness of the inhomogeneous autonomous retarded differential equations (IRDE) (L(t) ≡ L). τ ∈ [s − r. for instance. Delay Differential Equations and Applications. (L(t))t≥0 is a family of bounded linear operators from the space C([−r. REGULARITY AND ASYMPTOTIC BEHAVIOUR OF RETARDED DIFFERENTIAL EQUATIONS BY EXTRAPOLATION THEORY L. and xt is the function deﬁned on [−r. 0]. To proﬁt from the big development of the semigroup theory and its tools (as the perturbation technique). [1. Marrakesh 40000.ma 1.P. 0]. f : IR+ −→ E. It has been also studied the asymptotic behaviour and the regularity properties of the solutions of homogeneous retarded equations (HRDE) (f ≡ 0) as well as the asymptotic behaviour of the solutions of the inhomogeneous equations with respect to the inhomogeneous term f . θ ∈ [−r. and has been treated by using diﬀerent techniques and methods. B. 105. (eds. maniar@ucam.). The equation (IRDE) has attracted many authors. (IRDE) x(τ ) = ϕ(τ − s). Morocco. 2390. Arino et al. In these papers. 44.Chapter 12 WELLPOSEDNESS. x (t) = Ax(t) + L(t)xt + f (t). 227. we cite. Maniar Department of Mathematics. Introduction Consider the following inhomogeneous nonautonomous retarded differential equation t ≥ s. . Faculty of Sciences Semlalia. E) into E. where A is a linear (unbounded) operator on a Banach space E. 3. 226. 253]. many authors thought to transform the retarded equation (IRDE) to an abstract Cauchy problem (un519 O. © 2006 Springer. 64. 0] by xt (θ) = x(t + θ). 519–538.
94. and then the ones of the inhomogeneous autonomous retarded equations. 0]. A := 0 Aδ0 − δ0 . also using the extrapolation theory. and 0 L(t) B(t) := . [178] the wellposedness of the inhomogeneous autonomous retarded equations and. t ≥ 0. introduced by R. The ﬁrst remark that one can do is that X0 := D(A) = {0}×C([−r. Rhandi. has showed the existence of mild solutions of the inhomogeneous autonomous Cauchy problem (CP)f . are not in and the values of the inhomogeneous term 0 X0 . which means that the operator A is not densely deﬁned. 0 0 t ≥ 0. together with A. E) = X .520 DELAY DIFFERENTIAL EQUATIONS retarded equation) in a product Banach space X := E × C([−r. t ≥ 0. the regularity and the asymptotic behaviour of the solu . 10]. G¨ uhring. One remarks also that the ranges of the perturbation B(t). [17]. d 0 dτ D(A) := {0} × φ ∈ C 1 ([−r. Schnaubelt have studied [55. Thieme [224]. 0]. H. Amir [8. one cannot use the standard theory of semigroups to solve neither the perturbation problem (CP)0 nor the inhomogeneous Cauchy problem (CP)f . we studied the regularity of the solutions to the homogeneous autonomous equations. Using the extrapolation theory. A. To study the wellposedness of the homogeneous nonautonomous retarded diﬀerential equations (HRDE). E) as follows U (t) = AU (t) + B(t)U(t) + f (t) . E) : φ(0) ∈ D(A) . Rhandi [199] has proposed the use of the socalled extrapolation theory. R. with A. 0]. F. 9. we have also treated. With B. R¨ abiger and R. G. 93] the wellposedness and the asymptotic behaviour of the general inhomogeneous nonautonomous retarded equations (IRDE). we have studied the asymptotic behaviour of solutions to the inhomogeneous autonomous retarded equations. t ≥ s. Hence. Usually by the extrapolation theory. Nagel [63]. 92. 0 (CP )f U(s) = 0 . operators f (t) . B´ atkai. ϕ where. by the introduction of a new variation of constants formula. Our contribution in this book consists in the study of the wellposedeness.
perturbation results. These last results will be crucial to study the regularity and the asymptotic properties of these solutions. the result is shown in [187.Extrapolation theory for DDE 521 tions of the homogeneous nonautonomous retarded diﬀerential equations (HRDE). We study the existence of solutions. [1]..1) From the HilleYosida theorem. and that these solutions are given by DysonPhillips expansions and satisfy variation of constants formulas. 199]. ∞) ⊂ ρ(A) and (λ−w)n R(λ. [8].4]) the following result. have been also shown by many authors. We recall that a linear operator (A. we present a perturbation of HilleYosida operators result. and was successfully used in these last years for many purposes. 186]. but the results concerning the asymptotic behaviour are in our knowledge new. the part (A0 . [200] and [201]. [186]. D(A0 )) of A in X0 := (D(A). [227] and [253]. Then. we deﬁne the extrapolation spaces and the extrapolated semigroups for HilleYosida operators. More precisely. Nagel [63. Our document is organized as follows: In section 2. and present some of their important properties. At the end.g. using the extrapolation theory. Preliminaries In this section we introduce the extrapolation spaces and extrapolated semigroups and give all needed abstract results of the extrapolation theory. see. we recall some deﬁnitions and important results from the extrapolation theory. [178]. e. [92]. [109. The regularity results. to study the homogeneous nonautonomous retarded diﬀerential equations (HRDE). For more details we refer to the book [63] and [17. D(A)) be a HilleYosida operator on a Banach space X. 2.2. D(A)) on a Banach space X is called a HilleYosida operator if there are constants M ≥ 1 and w ∈ R such that (w.g. 186]. Nagel [63]. 12.. Section 3 is devoted to the application of the above abstract results. [226]. we have (cf. presented in this section. (2. using alternative diﬀerent techniques. [6]. · ) given . Thm. The extrapolation theory was introduced by R. [190]. e. Proposition 1 Let (A.. introduced by R. see e. [105]. [199].g. A)n ≤ M for all λ > w and n ∈ N.
Notice that HilleYosida operators on reﬂexive Banach spaces are densely deﬁned. 1. t ≥ 0. From now on. We have the following properties (see [186.522 DELAY DIFFERENTIAL EQUATIONS by D(A0 ) := {x ∈ D(A) : Ax ∈ X0 } A0 x := Ax for x ∈ D(A0 ) generates a C0 semigroup (T0 (t))t≥0 on X0 . As an immediate consequence of Proposition 2. since T0 (t)x−1 = T0 (t)R(λ0 . for a ﬁxed λ0 > ω. we have Corollary 3 Let (A. . It is strongly continuous and we denote by (A−1 . Then (A∗ .3. The extrapolated semigroup (T−1 (t))t≥0 is the unique continuous extension of (T0 (t))t≥0 to X−1 . (d) X is continuously embedded in X−1 and R(λ. A) for λ ∈ ρ(A−1 ) = ρ(A). D(A−1 )) its generator. Prop. D(A∗ )) generates the C0 semigroup (T ∗ (t))t≥0 on X ∗ . D(A)) be a HilleYosida operator on a reﬂexive Banach space X. A−1 ) is an extension of R(λ. t ≥ 0. x ∈ X0 . (c) A−1 : X0 → X−1 is the unique continuous extension of A0 . Then D(A) is dense in X and hence A generates a C0 semigroup . Moreover. we have ρ(A) = ρ(A0 ) and (λ − A0 )−1 is the restriction of (λ − A)−1 to X0 for λ ∈ ρ(A). A0 )x. (A. A0 )x ≤ M eωt x−1 . 1. D(A)) be the generator of a C0 semigroup (T (t))t≥0 on a reﬂexive Banach space X.4]). Proposition 2 Let (A. (b) D(A−1 ) = X0 . From the resolvent equation it is easy to see that diﬀerent λ ∈ ρ(A) yield equivalent norms. x ∈ X0 . Theorem 4 We have the following: (a) T−1 (t)L(X−1 ) = T0 (t)L(X0 ) . D(A)) will be a HilleYosida operator on a Banach space X. We deﬁne on X0 the norm x−1 := R(λ0 . Thm. The completion X−1 of X0 with respect to · −1 is called the extrapolation space.
The proof of the following perturbation result relies on the fundamental Proposition 5. Proposition 5 Let f ∈ L1loc (R+ . s) = Id for t ≥ r ≥ s ≥ 0 and (b) {(t. s s t T−1 (t − r)f (r) dr ≤ M1 t eω(t−r) f (r) dr.523 Extrapolation theory for DDE (e) A0 and A are the parts of A−1 in X0 and X. X)). For contraction semigroups (T0 (t))t≥0 . D(C(t)) = {x ∈ D(A) : Ax + B(t)x ∈ X0 }. We end this section by a nonautonomous perturbation result of HilleYosida operators. (2. due to A. we need only a particular version of this result.3) . The following result. +∞) t −→ 0 continuous. Theorem 6 The expansion U (t. X). M1 = 1 and ω = 0. concerning the convolutions of (T−1 (t))t≥0 and Xvalued locally integrable functions. A−1 )X0 = D(A0 ). s). s) : t ≥ s ≥ 0} (t. For the proof. First. and is the key to show the next theorem. see [63]. is one of the fundamental results of the extrapolation theory. t T−1 (t − r)f (r) dr ∈ X0 and (i) for all t ≥ s ≥ 0. the operator C(t) is deﬁned by C(t) = A + B(t). For all t ≥ 0.2) s M1 is a constant independent of t and f . s) and U (s. s) = ∞ n=0 Un (t. t T−1 (t − s)f (s)ds ∈ X0 is (ii) The function [0. let C(t) denote the part of the perturbed operator A + B(t) in X0 . X) for all t ≥ 0. s) → U (t. and t −→ B(t)x is a bounded continuous function on IR+ for all x ∈ X0 . Consider now a perturbation family B(·) ∈ Cb (R+ . s)x is continuous for all x ∈ X. Then. Rhandi [199]. respectively. i. For the aim of this document. recall that a family (U (t. (2. (f ) R(λ0 ..e. s))t≥s≥0 of bounded linear operators on X is called evolution family if (a) U (t. By deﬁnition. r)U (r. B(t) ∈ L(X0 . Ls (X0 . s) = U (t.
8) n=0 S0 (t) := T0 (t) and Sn+1 (t) := 0 t T−1 (t − σ)BSn (σ)dσ. In addition. and satisﬁes the variation of constants formula t T−1 (t − σ)BS(σ)xdσ. s) := T0 (t−s) and Un+1 (t. x ∈ X0 . D(C(t)))t≥0 . S(t)x = T0 (t)x + 0 (2. t ≥ 0. (U (t. s)x dσ.7) generates a C0 semigroup (S(t))t≥0 . t −→ U (t. x ∈ X0 . n ≥ 0. s)x is diﬀerentiable on [s. In the autonomous case. (2.5]. B(t) ≡ B. s) := T−1 (t−σ)B(σ)Un (σ. s))t≥s≥0 satisﬁes the variation of constants formula t U (t. s)x dt for all t ≥ s.6) Moreover.X) .5) for t ≥ s ≥ 0 and c := supτ ≥0 B(τ )L(X0 . n ≥ 0.10) . U (t. s)dσ. s) : t ≥ s ≥ 0} and deﬁnes an evolution family on X0 . satisfying U (t. In this case we say that the evolution family is generated by the family (C(t). Thm. t ≥ s ≥ 0.4) converges in the uniform operator topology of L(X0 ) uniformly on compact subsets of {(t. s (2. From Theorem 6. s)x = (A + B(t))U (t. t ≥ 0. (2. s) ≤ M e(ω+M c)(t−s) (2. the operator D(C) = {x ∈ D(A) : Ax + Bx ∈ X0 } C = A + B. 4. for all x ∈ D(C(s)). s)x = T0 (t−s)x+ T−1 (t−σ)B(σ)U (σ. ∞). see [187. if t −→ B(t)x is diﬀerentiable on IR+ for all x ∈ X0 then. s)x ∈ D(C(t)) and d U (t.1. s (2.9) (2. the semigroup (S(t))t≥0 is given by the DysonPhillips expansion S(t) = ∞ Sn (t).524 DELAY DIFFERENTIAL EQUATIONS where t U0 (t.
(CP )0 U(s) = ϕ0 . which satisﬁes the equation (HRDE). Here. A is a HilleYosida operator on a Banach space E. Before showing the equivalence of (HRDE) and (CP)0 . To do this. E0 ) to E. d 0 dτ D(A(t)) := {0} × φ ∈ C 1 ([−r.525 Extrapolation theory for DDE 3. x(t) = ϕ(t − s). for every t ≥ 0. with δ0 is the mass of Dirac concentrated at 0. t ≥ s. E0 ) : φ(0) ∈ D(A) . and by a mild solution a continuous function x : [−r. E0 ). satisfying t S0 (t − s)ϕ(0) + s S−1 (t − σ)L(σ)xσ dσ. (HRDE) xs = ϕ. 0] . t ≥ s. A(t) := 0 Aδ0 − δ0 + L(t) . (3. s − r ≤ t ≤ s. 0]. Homogeneous Retarded Diﬀerential Equations In this section we apply the extrapolation results from the previous section to solve the homogeneous retarded diﬀerential equations x (t) = Ax(t) + L(t)xt . we study the wellposedness of the last Cauchy problem (CP)0 .1) To get the purpose of this section. +∞) −→ E0 . Let S := (S0 (t))t≥0 the C0 semigroup generated by the part of A in E0 := D(A). 0]. A := d 0 dτ . (L(t))t≥0 is a family of bounded linear operators from Cr := C([−r. t ≥ s. +∞) −→ E0 . By a (classical) solution of (HRDE). we design a continuously differentiable function x : [−r. we write the operator A(t) as the sum of the operator 0 Aδ0 − δ0 . where. δ0 ϕ := ϕ (0) for all ϕ ∈ C 1 ([−r. and study the regularity and the asymptotic behaviour of their solutions. and (S−1 (t))t≥0 its extrapolated semigroup. we associate to the equation (HRDE) an equivalent Cauchy problem on the product Banach space X := E ×Cr U (t) = A(t)U(t).
τ It is clear that the function ϕ1 belongs to Cr . we propose to use the extrapolation theory to study the wellposedness of the Cauchy problem (CP)0 . A)[y + g(0)] + eλ(τ −σ) g(σ) dσ. . We conclude that 0 λτ ϕ1 (τ ) = e R(λ. From the equation (λ − A)( ϕ01 ) = ( yg ) on the space X . 0]. with ϕn (τ ) = n−1 i=0 eλτ (−τ )i 1 R(λ. 0]. we have 0 eλ(τ −σ) g(σ) dσ. E0 ) : φ(0) ∈ D(A) . This result has been shown in [199] when A generates a C0 semigroup. Proof. As we have remarked at the introduction. For this reason. g ∈ Cr and n ≥ 1.2) is satisﬁed for n = 1.526 DELAY DIFFERENTIAL EQUATIONS D(A) := {0} × φ ∈ C 1 ([−r. A)n−i [y + g(0)] + i! (n − 1)! 0 (σ − τ )n−1 eλ(τ −σ) g(σ) dσ τ (3. A)n ( yg ) = ( ϕ0n ). X0 := D(A) = {0} × Cr which means that the operator A is not densely deﬁned. 0]. ϕ1 (τ ) = eλτ ϕ1 (0) + τ and Aϕ1 (0) − ϕ1 (0) = y. 0 0 t ≥ 0.2) for all τ ∈ [−r. we obtain ϕn (τ ) = n−1 i=0 0 eλτ 1 (−τ )i R(λ. 0]. y ∈ E. τ Since A is a HilleYosida operator. and the operator 0 L(t) B(t) := . A)n−i [y + g(0)] + i! (n − 1)! (σ − τ )n−1 eλ(τ −σ) g(σ) dσ. Lemma 7 The operator A is of HilleYosida and R (λ. To this purpose. By induction. 0]. and the relation (3. τ ∈ [−r. it follows that τ ∈ [−r. τ ∈ [−r. we need to show that the operator A is of HilleYosida.
A)n ≤ M (λ − ω)n for all λ > ω. (3. It is clear then that the operator A0 := d . by Proposition 1. E0 ) : φ(0) A0 = d 0 dτ ∈ D(A). ϕ (0) = Aϕ(0) . t + θ ≤ 0. E0 ) : φ(0) ∈ D(A). t ≥ 0. ϕ (0) = Aϕ(0) dτ generates also a C0 semigroup (T0 (t))t≥0 . and one can easily show that it is given by ϕ(t + θ). 0]. D(A0 ) = {0} × φ ∈ C 1 ([−r. Therefore. S0 (t + θ)ϕ(0). 0]. and this achieve the proof. and it generates a C0 semigroup (T0 (t))t≥0 .527 Extrapolation theory for DDE ϕn (τ ) ≤ n−1 [(−λ + ω)τ ]i M λτ + ye (λ − ω)n i! g[ 0 M eλτ (λ − ω)n i=0 n−1 i=0 M [(−λ + ω)τ ]i + i! (n − 1)! (−σ)n−1 eσ(λ−ω) dσ] τ ≤ M [y + g]. T0 (t) = 0 T0 (t) . 0]. we have I 0 . (λ − ω)n The last inequality comes from 0 n−1 λτ [(−λ + ω)τ ]i M M M e (σ − τ )n−1 e(λ−ω)(τ −σ) dσ = + (λ − ω)n i! (n − 1)! τ (λ − ω)n i=0 for τ ∈ [−r. The part of A in X0 is the operator 0 0 .3) (T0 (t)ϕ) (θ) = t + θ > 0. n ≥ 1. R(λ. Hence as A0 is a diagonal matrix operator. D(A0 ) := φ ∈ C 1 ([−r.
D(AL (t)) : = φ ∈ C 1 ([−r. s) ϕ = T0 (t−s)ϕ + (3. the part of the operator (A + B(t)) in X0 is the operator 0 0 .7) . the functions t −→ B(t) ϕ0 . The family (AL (t).6) s for all t ≥ s and ϕ ∈ Cr . E0 ) : φ(0) ∈ D(A). by the perturbation result. t ≥ s. A)L(σ)U (σ. 0]. the function t −→ L(t)ϕ is continuously diﬀerentiable.4) n=0 where t U0 (t. Therefore. the family of the parts of operators (A + B(t)) in X0 generates an evolution family (U(t. which satisﬁes the variation of constants formula U (t. s) ϕ0 dσ U(t. by identiﬁcation of the elements of X0 and those of Cr . Proposition 8 Assume that (H) hold. s) := s T−1 (t−σ)B(σ)Un (σ. s) = ∞ Un (t.5) and which satisﬁes the variation of constants formula t 0 0 T−1 (t − σ)B(σ)U(σ. for each t ≥ 0. E0 ) : φ(0) ∈ D(A). are bounded from X0 into X and from (H). The linear operators B(t). (3. s)ϕ dσ.528 DELAY DIFFERENTIAL EQUATIONS To study now the existence of classical solution of (HRDE). (3. s))t≥s≥0 on X0 given by the DysonPhillips expansion U(t. s)ϕ = T0 (t−s)ϕ+ lim λ→∞ s t T0 (t − σ)eλ· λR(λ. s)dσ. s) := T0 (t−s) and Un+1 (t. s). ϕ ∈ Cr . One can see also that. we assume: (H) For all ϕ ∈ Cr . (3. D(AL (t)))t≥0 of operators deﬁned by d AL (t) := . (A + B(t))/X0 = d 0 dτ with the domain D((A + B(t))/X0 ) = {0} × φ ∈ C 1 ([−r. n ≥ 0. 0]. are continuously differentiable. we get the following result. s))t≥s≥0 on the space Cr. Theorem 6. ϕ (0) = Aϕ(0) + L(t)ϕ} . Thus. t ≥ 0. ϕ (0) dτ = Aϕ(0) + L(t)ϕ} generates an evolution family (U (t.
ϕ). D(AL (t)))t≥0 . it satisﬁes (3. s − r ≤ t ≤ s. s.7). x(t. As the operators (A + B(t))/X0 . lim λ→∞ n ≥ 0..6) and extrapolation results. ϕ) := U (t. U(t. by Lemma 7 we obtain the variation of constants formula (3. Moreover. (3. By the variation of constants formula (3. .4)(3.s)ϕ and (U (t. are diagonal matrix operators on X0 = {0} × Cr . s))t≥s≥0 is an evolution family on the space Cr generated by (AL (t). A) 0 ϕ L(σ)U (σ.5). t ≥ s. Also by the same argument. s) ϕ0 = U (t. s) = = T0 (t − s) and Un+1 (t. s) t T0 (t − σ)eλ· λR(λ.e. t > s. we obtain (3. (3. we can show that 0 for all t ≥ s anϕ ∈ Cr . s. A) 0 T0 (t−s)ϕ λ→∞ + lim λ→∞ s s T−1 (t − σ)B(σ)U(σ.s)ϕ = 0 T0 (t−s)ϕ = t + lim λR (λ. s)dσ. In the following proposition. s)ϕ(0).10) is the mild solution of (HRDE).8)(3. i.s)ϕ 0 dσ dσ. Proposition 9 Let ϕ ∈ Cr and s ≥ 0.8). t ≥ s. s)ϕ = xt (·. s) t T0 (t − σ)λR (λ. A)L(σ)Un (σ.1). we give the correspondence between the mild solutions of the retarded equation (HRDE) and those of the Cauchy problem (CP)0 .(3. from the relations (3. Hence. we have: (i) The function deﬁned by ϕ(t − s).8) n=0 where U0 (t. x satisﬁes (3.11) U (t. s) = ∞ Un (t.529 Extrapolation theory for DDE and is given by the DysonPhillips series U (t. s).9) s Proof. t ≥ 0. we obtain that 0 U (t.
e. L(t) = L for all t ≥ 0. and the assertion (i) is proved. x(t. s. A)L(σ)U (σ.. s)ϕdσ if t − s + τ > 0. λ→∞ s t+τ S (t − s + τ )ϕ(0) + S−1 (t + τ − σ)L(σ)U (σ. The part (a) has been obtained also by many authors. [63]. Let the function ϕ(t − s). t > s. this implies that x(·. dτ D(AL ) := ϕ ∈ C 1 ([−r. [105] and [253]. s. x(t. s)ϕ(τ ) = T0 (t − s)ϕ(τ ) + lim t λ→∞ s t+τ T0 (t − σ)eλ· λR(λ. In the particular case of autonomous retarded diﬀerential equations. s. ϕ) dσ. A)L(σ)U (σ. 0 (3. s)ϕdσ if t − s + τ > 0. Theorem 10 (a) The operator AL := d .12) = s ϕ(t − s + τ ) if t − s + τ ≤ 0 for all τ ∈ [−r.s. ϕ) = S0 (t − s)ϕ(0) + s t−s S−1 (t − σ)L(σ)U (σ. s − r ≤ t ≤ s. i.g. Let τ ∈ [−r. 0] and ϕ ∈ Cr . The assertion (ii) can also be deduced from the above relations. we have the following theorem. From the extrapolation results. s. s)ϕ dσ(τ ) S0 (t + τ − σ)λR(λ. we have U (t. s)ϕds if t − s + τ ≤ 0. s)ϕdσ.12). Hence. From the equality (3.530 DELAY DIFFERENTIAL EQUATIONS (ii) If x(·. A)L(σ)xσ (·. s)ϕ(0). [1].e. Proof. ϕ (0) = Aϕ(0) + Lϕ . one can obtain easily the relation (3.ϕ) T0 (t − σ)eλ· λR(λ. s.7) and the deﬁnition of (T0 (t))t≥0 . t ≥ s. ϕ) is the mild solution of (HRDE) then t −→ is the mild solution of the Cauchy problem (CP)0 . 0] and 0 ≤ s ≤ t. s.1). A)L(σ)U (σ. the variation of constants formula (3. Thus. E0 ) : ϕ(0) ∈ D(A). s)ϕdσ+ S0 (t − s + τ )ϕ(0) + lim λ→∞ s t = + lim eλ(t+τ −s) λR(λ. ϕ) := U (t. and xt (·. ϕ) satisﬁes (3. ϕ) = T0 (t − s)ϕ + lim λ→∞ s t 0 xt (·. λ→∞ t+τ t ϕ(t − s + τ ) + lim eλ(t+τ −s) λR(λ. A)L(σ)U (σ. 0].11).
the operator (A + B)/X0 generates a C0 semigroup T := (T (t))t≥0 .13) T (t)ϕ(0). t ≥ 0. T is given by the DysonPhillips series T (t) = ∞ n=0 Tn (t). and I 0 T (t) = . From Section 2. one can see that the operator (AL . Hence.531 Extrapolation theory for DDE generates a strongly continuous C0 semigroup T := (T (t))t≥0 on the space Cr . A)LTn−1 (s)ϕ ds. −r ≤ t ≤ 0. t ≥ 0. . S−1 (t + τ − s)LTn−1 (s)ϕds if t + τ > 0. D((A + B)/X0 ) = {0} × φ ∈ C 1 ([−r. E0 ) : φ(0) ∈ D(A). t+τ ≤0 (3. D(AL )) generates also a C0 semigroup T := (T (t))t≥0 on Cr . T (t)ϕ = xt (b) The semigroup T is also given by the DysonPhillips series T (t) = ∞ t ≥ 0. t ≥ 0. the semigroup T is given by for all ϕ ∈ Cr and t ≥ 0. The part of A + B in X0 = {0} × Cr is the operator given by (A + B)/X0 = 0 0 0 d dτ . Tn (t). ϕ (0) = Aϕ(0) + Lϕ . one has: (i) the solution x of (HRDE) is given by ϕ(t). n=0 where Tn (t)ϕ := lim λ→∞ 0 or Tn (t)ϕ(τ ) = t+τ 0 0 if t T0 (t − s)eλ· λR(λ. t ≥ 0. 0 T (t) Moreover. Moreover.14) for all ϕ ∈ Cr and n ≥ 1. x(t) := (3. from the form of (A + B)/X0 . Proof. 0]. (ii) If x is the solution of (HRDE).
(ii) If S is immediately compact then T (t) is also compact for all t > r. we have t+τ S0 (t + τ − s)λR(λ. A)LTn−1 (s)ϕds if t + τ > 0. we need the following results. are convolutions between the extrapolated semigroup (S−1 (t))t≥0 and Evalued functions. Now. to get our aim. The DysonPhillips series obtained in the above theorem will be now used to study the regularity properties of the semigroup T solution of the retarded equation (HRDE). λ→∞ 0 Thus. A)LTn−1 (s)ϕ ds. [63] and [253].14). 0].3). for instance. Tn (t)ϕ = lim λ→∞ 0 and from (3. one can see easily as in the nonautonomous case that t T0 (t − s)eλ· λR(λ. λ→∞ t t+τ lim eλ(t+τ −s) λR(λ.532 DELAY DIFFERENTIAL EQUATIONS where 0 ) := Tn (t)( ϕ0 ) := ( Tn (t)ϕ 0 t T−1 (t − s)BTn−1 (s)( ϕ0 ) ds. (i) If {t : t > 0} t → S0 (t) ∈ L(E0 ) is continuous (or S is t immediately norm continuous) then {t : t > 0} t −→ 0 S−1 (t − s)LG(s) ds ∈ L(Cr . Then. the terms Tn of the series. A)LTn−1 (s)ϕds if t + τ ≤ 0. As. (ii) If the operator S0 (t) is compact for all t > 0 (or S is immediately compact) then the operator t S−1 (t − s)LG(s) ds 0 deﬁned from Cr to E0 is compact for all t ≥ 0. Theorem 12 (i) If S is immediately norm continuous then {t : t > r} t → T (t) ∈ L(Cr ) is also a continuous function. E0 ) is continuous. showed also. t ≥ 0.14). Lemma 11 [17] Let G ∈ C(IR+ . see (3. The assertions (i)(ii) are particular cases of Proposition 9. From this and Lemma 7. Ls (Cr )). we can announce the following regularity results. in [1]. for τ ∈ [−r. . we obtain the relation (3. A)LTn−1 (s)ϕds+ lim λ→∞ 0 t Tn (t)ϕ(τ ) = + lim eλ(t+τ −s) λR(λ.
we have also that S is immediately norm t+θ continuous. The function f is said to be asymptotically almost periodic if H(f ) is relatively compact in BU C(IR+ . Consequently.533 Extrapolation theory for DDE Proof. X0 ). see [129] for more details.4. More precisely. e. the function IR+ t → Tn (t) ∈ L(Cr ) is continuous for all n ≥ 1. and then the subset {θ → Tn (t)(θ) ϕ in some bounded set of Cr } of Cr is equicontinuous. From Lemma 11. X0 ). t ≥ 0. we show that the solution of (HRDE) IR+ t −→ x(t) has the same asymptotic behaviour. The compactness of Tn (t) for all t > 0 and n ≥ 1 follows ﬁnally from ArzelaAscoli theorem. as the series (3. X0 ). we obtain the second assertion. First. the set of functions {θ → 0 S−1 (t+ θ−s)LTn−1 (s)ϕ ds ϕ in some bounded set of Cr } is equicontinuous. boundedness. called the hull of f . Hence. and then Tn (t)(θ) is also a compact operator from Cr to E0 for all θ ∈ [−r. By [63. We end this section by studying the robustness of the asymptotic behaviour of the solutions to nonautonomous retarded equation (HRDE) with respect to the term retard. Theorem II. (3. .29].15) n=1 and that this series converge in L(Cr ) uniformly in compact intervals of IR+ . asymptotic almost periodicity. Let t > 0 and n ≥ 1. by Lemma 11 (i).15) converges in the uniform operator topology of Cr . the set of all translates. 0]. let us recall the following deﬁnitions: For a function f ∈ BU C(IR+ . By the deﬁnition of the semigroup (T0 (t))t≥0 it is easy to show that t → T0 (t) is norm continuous for t > r if we consider (i). the space of bounded and uniformly continuous functions from IR+ into X0 . We have seen in Theorem 10 that T (t) = ∞ n=0 Tn (t) = T0 (t) + ∞ Tn (t).. Treat now the assertion (ii).14) and Lemma 11 (i).g. From the relations (3. Hence. is H(f ) := {f (· + t) : t ∈ IR+ }. the assertion (i) is obtained. as the map IR+ t −→ S0 (t)ϕ(0). Assume that S is immediately norm continuous. and that the operator T0 (t) is compact for all t > r in the case of (ii). and Eberlein weakly asymptotically almost periodic if H(f ) is weakly relatively compact in BU C(IR+ . we have that t+θ S−1 (t + θ − s)LTn−1 (s) ds 0 is a compact operator from Cr to E0 for all t + θ ≥ 0.
n ∈ N and t ≥ s ≥ s0 +∞ L(σ)Un (σ.8). s) : t ≥ s ≥ s0 }. (H2) there exist s0 ≥ 0 and a constant 0 ≤ q < 1 such that ∞ L(s + t)T0 (t)ϕdt ≤ qϕ for all ϕ ∈ Cr and s ≥ s0 . s)ϕdσ ≤ q n+1 ϕ (3. Suppose now that the estimate holds for n − 1.17) is only our assumption (3. and operator invariant if M ◦ f ∈ E for every f ∈ E and M ∈ L(X0 ).16) 0 Under the above hypotheses. by using Fubini’s theorem we have . Let ϕ ∈ Cr and t ≥ s ≥ s0 . The closed subspace E is said to be homogeneous if it is translation biinvariant and operator invariant. Proposition 13 (i) For all ϕ ∈ Cr . t ≥ 0. s))t≥s≥0 is uniformly bounded and U (t.8) converges uniformly on all the set {(t. For n = 0. Moreover. Proof. From the relation (3. the evolution family (U (t. we assume that: (H1) the C0 semigroup (S0 (t))t≥0 is of contraction.16).18) (ii) The series (3. the estimate (3.534 DELAY DIFFERENTIAL EQUATIONS A closed subspace E of BU C(IR+ .17) s and Un (t. we have the following results. s)ϕ ≤ q n ϕ. where M ◦ f is deﬁned by (M ◦ f )(t) = M (f (t)). s) ≤ 1 1−q for all t ≥ s ≥ s0 . (3. X0 ) is said to be translation biinvariant if for all t ≥ 0 f ∈ E ⇐⇒ f (· + t) ∈ E. For our purpose. (3.
s)ϕdσ σ t = L(σ) lim T0 (σ − τ )eλ· λR(λ.535 Extrapolation theory for DDE t L(σ)Un (σ. A)L(τ )Un−1 (τ. and other asymptotic properties of this mild solution we need the following Lemma. To obtain the asymptotic almost periodicity. the boundedness of the mild solution of (HRDE) is also obtained. from (H1). A)g(τ )dτ ! ! ! ≤ CgL1 . s)ϕdτ dσ s λ→∞ s s ≤ lim t λ→∞ s  σ L(σ)T0 (σ − τ )eλ· λR(λ. In the above proposition we obtain the boundedness of the evolution family (U (t. E). E). it is clear that ! ! t ! ! λ· ! lim T0 (t − τ )e λR(λ. A)L(τ )Un−1 (τ. 0 Hence. Lemma 14 Let g ∈ L1 (IR+ . The assertion (ii) follows then by this estimate (3.8) and the ﬁrst estimate (3.17). For g ∈ L1 (IR+ . since (T0 (t))t≥0 is bounded.16) and the induction hypothesis t t L(σ)Un (σ. s)ϕdσ ≤ q L(τ )Un−1 (τ. s)ϕdτ dσ s t = lim λ→∞ s t−τ L(σ + τ )T0 (σ)eλ· λR(λ. If t −→ T0 (t)ϕ belongs to E for all ϕ ∈ Cr then the function t T0 (t − τ )eλ· λR(λ.18). s)ϕdτ s s ≤ q n+1 ϕ. Proof. A)L(τ )Un−1 (τ.18) follows also from (H1). and this gives the estimate (3. s)ϕdσdτ. the relation (3. λ→∞ 0 . The inequality (3.17). and thus from (3. s))t≥s≥0 . the inequality (3.13). A)g(τ )dτ IR+ t −→ T0 ∗ g(t) := lim λ→∞ 0 belongs to E.
16) is satisﬁed then. E). where x(·. the space of all bounded continuous functions. it is suﬃcient to show this result for simple functions. b ≥ a ≥ 0. t −→ xt+s (·. Cr ). This achieves the proof.17) for all s ≥ s0 . Cr ) and the linear operator g −→ T0 ∗g is bounded from L1 (IR+ . ϕ) is the mild solution of (HRDE). s)ϕ = U (t + s + s0 . From Proposition 9 and the relation (3. Furthermore. xt+s (·. ϕ) is also in E for all ϕ ∈ Cr and s ≥ 0. Let g := 1(a. E). s)ϕ belongs to L1 (IR+ . We can now state the following main asymptotic behaviour result. by Lemma 14.8) and the estimate (3. s. one can write U (t + s0 + s. the function IR+ t −→ xt+s (·. A)L(s + σ)U (s + σ. then as shown above t −→ U (t + s0 + s. A)x dτ. A)x dτ b T0 (b − τ )eλ· λR(λ. s)ϕ. s. .b) ⊗ x. Hence. s + s0 )U (s + s0 . s. s)ϕ belongs to E for all s ≥ s0 and ϕ ∈ Cr . = T0 (t) lim λ→∞ a λ→∞ a Hence. E) into BC(IR+ . T0 ∗ g(t + b) = lim b T0 (t + b − τ )eλ· λR(λ. and this achieve the proof. As s + s0 ≥ s0 . since E is translation biinvariant and t −→ T0 (t)ϕ belongs to E for every ϕ ∈ Cr . For all s ≥ 0 and t ≥ 0. If t −→ T0 (t)ϕ belongs to E for all ϕ ∈ Cr . s. s + s0 )ϕ belongs to E and by the translation biinvariance of E. s)ϕ belongs to E. we conclude that T0 ∗ g(·) ∈ E. and the condition ( 3. ϕ) = U (t + s. s. the linearity and the density. ϕ) = U (t + s. We have. A)L(s + σ)U (s + σ.536 DELAY DIFFERENTIAL EQUATIONS which implies that for every g ∈ L1 (IR+ . s)ϕ dσ.19) As t −→ T0 (t)ϕ belongs to E. Theorem 15 Assume that (H1) and (H2) hold. the Cr valued function IR+ t −→ xt+s (·. Proof. we have t T0 (t − σ)eλ· λR(λ. By this boundedness. ϕ) = T0 (t)ϕ+ lim λ→∞ t ≥ 0. t ≥ 0 λ→∞ 0 belongs to E as well. 0 (3. and this follows from (3. s)ϕ dσ. T0 ∗g ∈ BC(IR+ . it is suﬃcient to that the function f from IR+ to Cr t f (t) := lim T0 (t − σ)eλ· λR(λ.11). it is suﬃcient to show that the function g(·) := L(· + s)U (· + s. x ∈ E and t ≥ 0.
for every ε > 0 there is l(ε) > 0 and K ≥ 0 such that each interval of length l(ε) contains a τ for which this inequality S0 (t + τ )ϕ(0) − S0 (t)ϕ(0) ≤ ε holds for all t. then we can hope that they have the same asymptotic behaviour. when t → ∞. see [22]. one has (T0 (t + τ )ϕ)(θ) − (T0 (t)ϕ)(θ) = S0 (t + θ + τ )ϕ(0) − S0 (t + θ)ϕ(0) ≤ S0 (t − r + τ )ϕ(0) − S0 (t − r)ϕ(0) <ε for all θ ∈ [−r. In the following lemma. If one has (1). Lemma 16 Let ϕ ∈ Cr . Assume now that we have (2). Let now t > K + r. the function IR+ t −→ T0 (t)ϕ has the same property in BU C(IR+ . 0]. E0 ) for all θ ∈ IR. or ∞ e−αs S0 (· + s)ϕ(0) ds ex(3) uniformly ergodic. the limit lim α α→0+