STABILITY OF FUNCTIONAL DIFFERENTIAL EQUATIONS
V. B. KOLMANOVSKII Institute for Problems in Mechanics Academy of Sciences of USSR Moscow, USSR
AND
V. R. NOSOV Institute for Problems in Mechanics Academy of Sciences of USSR Moscow, USSR
1986
ACADEMIC PRESS, INC. Harcourt Brace Jovanovich, Publishers

London Orlando San Diego * New York Montreal . Sydney . Tokyo * Toronto Austin

COPYRIGHT 0 1986 BY ACADEMIC PRESS INC. (LONDON) LTD ALL RIGHTS RESERVED. NO PART OF THIS PUBLICATION MAY BE REPRODUCED OR TRANSMITTED IN ANY FORM OR BY ANY MEANS, ELECTRONIC OR MECHANICAL, INCLUDING PHOTOCOPY, RECORDING, OR ANY INFORMATION STORAGE AND RETRIEVAL SYSTEM, WITHOUT PERMISSION IN WRITING FROM THE PUBLISHER.
ACADEMIC PRESS INC. (LONDON) LTD. 2428 Oval Road LONDON NWI 7DX
United States Edition published by
ACADEMIC PRESS, INC. Orlando, Florida 32887
British Library Cataloguing in Publication Data Kolmanovski:, V. B. Stability of functional differential equations. (Mathematics in science and engineering, 1 . Functional differential equations I . Title 11. Nosov, V. R. 111. Series 515.3’5 QA372 Library of Congress Cataloging in Publication Data Kolmanovski;, Vladimir Borisovich. Stability of functional differential equations. (Mathematics in science and engineering) Includes index. 1. Functional differential equationsNumerical solutions. 2. Stability. I. Nosov, V. R. 11. Title. 111. Series. QA372.K78 1985 515.3’5 859118 ISBN 0124179401 (hardcover) (alk. paper) ISBN 012417941X (paperback) ( a k . paper) PRINTED IN THE UNITED STATES OFAMhKlCA
66878889
9 8 1 6 5 4 3 2 1
Contents xi
Preface Chapter 1. Theoretical foundations of functional differential equations 1. Time lags in engineering 1.1. 1.2. 1.3. 1.4. 1.5. 1.6. 1.7.
Reactor (technological lag) Rolling mill (transportation lag) Ship stabilization (informational lag) Manual control Turbojet engine Nuclear reactor Infeed grinding model
2. Aftereffects in physical, biological and other systems 2.1. Microwave oscillator 2.2. Systems with lossless transmission lines 2.3. Relativistic dynamics allowing for the finiteness of interaction speed 2.4. Population dynamics models 2.5. Model of immune response (predatorprey model) 2.6. Distribution of albumin in the bloodstream 2.7. Some economic applications
3. Initialvalue problem for functional differential equations 3.1. 3.2. 3.3. 3.4. 3.5. 3.6. 3.7.
General equations with delay RFDE with bounded delay RFDE with unbounded delay NFDE Two useful lemmas Estimations of solutions growth Methods for solving an FDE
4. Linear equations 4.1. 4.2. 4.3. 4.4. 4.5.
General form of a linear RFDE A priori solution estimate Linear NFDEs FDEs with small delays Variationofconstant formula
5. Linear autonomous equations 5.1. Characteristic function
V
1 1 3 4 5 6 8 8 9 9 10 13 13 15 16 17 18 18 19 21 21 23 24 25 25 25 26 27 21 28 31 31
vi
Contents 5.2. Zeros of characteristic function 5.3. Representation of a solution as an infinite series 5.4. Linear autonomous NFDEs 6.
Feedback systems with aftereffects
6.1. 6.2. 6.3. 6.4.
Singleloop feedback systems Effects of little delays Analysis of singleloop feedback systems Twoloop systems with delay
Chapter 2. Stability of retarded equations 1 . General definitions of stability 1.1. 1.2. 1.3. 1.4.
Definition of Liapunov stability Uniform Stability Asymptotic stability Other stability definitions. Stability under delay disturbances
2. Linear autonomous retarded equations 2.1. Statement of the problem 2.2. General theorems
3. Methods of stability investigation of linear autonomous RFDEs 3.1. 3.2. 3.3. 3.4. 3.5.
Analytical methods (Pontriagin’s and Chebotarev’stheorems) Method of D subdivision. Vyshnegradskii diagrams Michailov, Nyquist and integral frequency criteria Stability for arbitrary delays, Tsypkin criterion General function (2.1 1)
4. Stabilization system of the tworeflector antenna 5. Liapunov direct method for equations with delay
6.
5.1. Application of Liapunov functions. Razumikhintype theorems 5.2. Method of Liapunov Krasovskii functionals for equations with bounded delay 5.3. Equations with unbounded delay 5.4. Stability in the first approximation and under steady acting perturbations 5.5. Case of a nonpositive derivative 5.6. Global stability 5.7. Survey of other results. Exponential stability Construction of Liapunov  Krasovskii functionals 6.1. Linear autonomous equations 6.2. Stability of a chemical reactor closed by a P controller 6.3. Scalar nonlinear equations 6.4. Nonlinear equations of second order
7. Stability of nuclear reactors
7.1. Singletemperature reactor with convective feedback 7.2. Stability of two interconnected reactors
8. Mathematical models in immunology 8.1. Models of virus disease 8.2. Analysis of model (8.2), (8.3) 8.3. Discussion
31 33
33 35 35 36
38 39
44
44
46 46 47
48
48
50 53
53
55
60 64 66 67 71 72 72
74
75 76
78 79
80 80 81 83 85 87 87 87 90 90
91 94
vii
Contents
9. Design of adaptive controller for retarded systems 9.1. Scalar equations 9.2. Delay adjustment 9.3. Multidimensional systems
10. Stability of viscoelastic bodies
10.1. Constitutive equations of viscoelastic bodies 10.2. Dynamic stability of a viscoelastic bar
11. Absolute stability of systems with delay 12. Stability of linear periodic equations 12.1. General stability theorem 12.2. Particular classes of RFDEs 12.3. Floquet theory
Chapter 3. Stability of neutral functional differential equations 1. Stability of linear autonomous NFDEs
1.1. General stability theorems 1.2. Methods of stability investigation of linear autonomous NFDEs
2. 3.
4.
Stability of aeroautoelasticity equations
2.1. Equations of unsteady motion of an elastic rigid body 2.2. Stability of aeroautoelastic equations
Liapunov direct method for neutraltype equations
3.1. 3.2. 3.3. 3.4. 3.5. 3.6. 3.7.
Introduction Some definitions General theorems Global stability Stability of the functional inequalities Equations with bounded delay Examples
Construction of degenerated functionals for concrete systems
4.1. Stability of a chemical reactor closed by a PD controller 4.2. Stability of onedimensional nonlinear systems 4.3. Use of degenerate functionals for stabilityinvestigationsof RFDEs
5. Instability of neutraltype equations
5.1. Statement of the problem 5.2. Influence of the choice of an admissible class of disturbances on stability 5.3. Instability conditions 5.4. Connection between instability of NFDEs and ordinary differential equations 5.5. Other instability conditions 5.6. Instability of equations with bounded delay 5.7. Distributed selfoscillatorysystems
6.
Asymptotic properties of neutraltype equations 6.1. 6.2. 6.3. 6.4. 6.5.
Some introductory remarks Basic definitions Stability of invariant sets Instability Interconnection between functional and differential equations
96 96 100 101 103 103 104 106 108 108 108 110
113 113 116 120 120 123 126 126 126 127 128 129 130 131 132 132 134 140 142 142 143 144 146 147 148 150 151 151 152 153 154 155
...
Contents
Vlll
7.
Liapunov functionals depending on derivatives
156 156 158 8. Stability and boundedness of linear nonhomogeneous equations 159 8.1. Connection between stabilitv and boundedness 159 8.2. Boundedness of derivatives 159 9. Linear periodic equations 160 9.1. Some examples 160 9.2. Stability theorem 162 7.1. General stability theorem 7.2. Examples
Chapter 4. Stability of stochastic functional differential equations 1. Some prerequisites from the theory of stochastic retarded functional differential equations Statement of the initialvalue problem 1.2. Existence theorem 1.3. ItG's formula 1.4. Equations for moments 1.1.
2.
Formulation of stability problems for SRFDEs. Liapunov direct method
2.1. 2.2. 2.3. 2.4.
3.
4.
Basic definitions of stability Asymptotic p stability Exponential p stability SRFDEs with random delays
Stability of scalar stochastic equations 3.1. Preliminaries 3.2. Case of autonomous linear part 3.3. Case of nonautonomous linear part
Stability of secondorder equations
4.1. General assumptions 4.2. Stability conditions 4.3. Stability domains of linear autonomous systems obtained by the Liapunov direct method
5. Stationary and periodic solutions of stochastic retarded equations 5.1. 5.2. 5.3. 5.4.
Existence of stationary solutions Relation between stability and stationarity of solutions Periodic solutions Ergodic properties of stationary solutions
6.
Stability with respect to the first approximation
7.
Neutraltype stochastic functional differential equations
8.
Stability of linear autonomous equations
6.1. General theorem 6.2. Scalar equations
7.1. Definition of SNFDEs 7.2. Existence theorem 1.3. Stability conditions
8. I . Systems of linear equations
164 164 165 166 166 167 167 168 170 170 171 171 171 174 175 175 176 180 187 187 187 189 190 191 191 192 193 193 194 194 196 196
ix
Contents
8.2. Corollary for nthorder equations 8.3. Necessary and sufficient stability conditions of scalar equations
198 199
References
202
Index
215
Preface This book provides an introduction to the structure and stability properties of solutions of functional differential equations (FDEs).The purpose of this book is to introduce mathematicians, physicists, engineers and other scientists to this subject and to give some ideas about the many applications of FDEs. The beginning reader may find it useful to learn the main results, corollaries and examples from the theory and applications of FDEs,and great pains have been taken to make the text accessible to both engineers and mathematicians. Our guiding idea is clarity of exposition; thus, proofs are usually given unless precluded by excessive length or complexity, in which case the appropriate reference is cited. It is intended that the notes and references should help the reader to explore the journal literature further. The book consists of a number of more or less independent sectionsdesigned to illustrate the most important results of FDE theory and their applications. A student with a good knowledge of calculus should find all of it accessible, but the material of the book should also be appropriate for twosemester graduatelevel courses. The hereditary systems (or the systems with aftereffect, with time lag or with delay) are of great theoretical interest and form an important class as regards their applications. This class of systems is described by functional differential equations, which are also called differential equations with deviating arguments. According to Myshkis [ 156(3)] thefunctional dzferential equation is an equation involving the function x(t) of one scalar argument t (called time) and its derivatives for several values of argument t. Among functional differential equations one may distinguish some special classes of equations retardedfunctional diferential equations, neutral functional dzferential equations, diferentialdifference equations,functional or dzflerenceequations, etc. The following abbreviations are constantly used: FDEs RFDEs NFDEs SFDEs SRFDEs
functional differential equations retarded functional differential equations neutral functional differential equations stochastic functional differential equations stochastic retarded functional differential equations xi
xii
Preface
If the value of the highest derivative at time t depends only on the values of lower derivatives at preceding times t 6,  03 < 6 5 0, we get an RFDE. Retarded functional differential equations describe those systems or processes whose rate of change of state is determined by their past and present states. If the rate of change of state depends not only on the past states but also on the past rates of change of state, then such systemsare governed by an NFDE. A functional or difference equation is an equation which involves the values of the function x(t)at different times t 8,  ~0 < 8 < co or at times ti, j=l,2,. . . . Historically the first FDEs were considered by Euler, Bernoulli, Lagrange, Laplace, Poisson and others in the eighteenth century [ 1921. The sources of FDEs for these mathematicians were various geometrical problems. In the nineteenth century and at the beginning ofthe twentieth century, FDEs were studied very occasionally. The situation changed radically in the 1930s and 1940s. At that time a number of important scientific and technical problems demanded that their adequate description take into account existing different delays. The first problems of this type were considered by Volterra (viscoelasticity, 1909 [230(l)]; predatorprey models, 1928 1931 [239(2), 239(3)]), Kostyzin (problems of mathematical biology, 1934 [ 1 lo]), Callender and Stevenson (instability in systems with time lag, 1936 [32]),an editorial article in Engineer (damping effect of time lag, 1937 [225]),Gorelik (microwaveoscillator, 1939 [69]), Voznesensky and Solodovnikov (influence of hydroshock on turbine stability 1934, 1941 [243,219( l)]), Bogomolov (feedbacksystem for a hydroelectric power station, 1941 [25]), Sievert (reaction of cells to xray irradiation, 1941 [208]), Minorsky (ship stabilization, 1942 [ 149(2)]),Andronov and Mayer (time lag in feedback systems, 1946 [4]), Kabakov and Sokolov (control process for steam pressure, 1946 [93, 2 17]), Gerasimov (feedback system for heat transfer, 1949 [62]) and many others. The appearance of such practically important problems stimulated the interest of mathematicians in investigatingthese not very wellknown equations. In 1939 Tychonov [227] considered the functional equation of the Volterra type. In 1942 Pontryagin [ 1751 obtained some fundamental results about the distribution of quasipolynomialzeros. In the early 1940s Chebotarev [37] published a number of papers devoted to the RouthHurwitz problem for quasipolynomials.The rapid growth of the theory of the FDE began in the 1950s.The paper of Myshkis [ 156(l)] has an important effect on the development of the FDE. In his paper for the first time the initialvalue problem was correctly formulated. Myshkis introduced a general class of linear RFDEs and studied some basic properties of their solutions. These results have been stated in the book [ 156(3)], which was the first entirely
+
+
Preface
...
Xlll
devoted to the theory of the RFDE. At the same time papers by Bellman [ 171, Tsypkin [232(1)232(3)], Neimark [ 160(l)], Leontjev [ 1251, Wright [247(1)247(3)] and Zubov [252(l)] appeared in which linear autonomous equations and their stability were studied in detail. Stability investigationsof these equations are closely connected with study of special analytical functions representing some types of quasipolynomials. The distribution of the quasipolynomial zeros was investigated by several different methods analytical (Pontriagyn, Chebotarev), frequency criteria (Kabakov, Tsypkin) and the method of D subdivision (Meiman, Neimark). The stability conditions of general linear autonomous equations with distributed bounded delay were obtained by Krasovskii with semigroup methods. It was important for the theory of equations with deviating argument that El’sgol’tz [58(3)] and Kamenskii [95( l)] distinguished different classes of the class of NFDEs. The use of Liapunov’s direct method for investigation of stability criteria of equations with delay encountered some principal difficulties [58(l)]. Krasovskii suggested the use of functionals defined on equations’ trajectories instead of Liapunov functions. In particular, for equations with bounded delay, Krasovskii proved the inversion theorem and theorems about stability in the first approximation and under steadyacting perturbations [ 1 14(1) 1 14(6)].Another approach to the stability problems for equations with delay based on a modification of the Liapunov function method was proposed by Razumikhin [ 180(1), 180(2)]. More recently, a number of new important results in stability theory were obtained a generalization of Liapunov’s direct method for NFDEs and SFDEs, stability, criteria for linear periodic equations, instability theorems, critical cases, etc. Our first chapter is devoted to the theoretical basis of the FDE with examples from different pure and applied problems in which delays were taken into account. The second chapter deals with the stability of RFDEs. First, linear autonomous systems and corresponding quasipolynomials are studied; then stability of feedback systems for the case of a tworeflector antenna is considered. Attention is given to Liapunov’s direct method and to the construction of concrete Liapunov Krasovskii functionals. On the basis of these functionals the stability of nuclear reactors is studied. Other problems presented concern some mathematical models of virus diseases and the design of adaptive controllers with time lag. Finally the stability conditions of viscoelastic systems and absolute stability criteria for retarded control systems are derived. The third chapter is devoted to the stability of NFDEs. After the theoretical investigation of stability of linear autonomous equations, the interesting stability problems of unsteady motion in a continuous medium (for a model of aeroautoelasticity) are considered. Further, a particular method for stabil
xiv
Preface
ity investigation of NFDEs proposed by the authors is described. This method is based on the use of degenerated functionals. Different stability and instability theorems and examples are given. The conditions of selfexciting oscillation in distributed systems are established and finally the asymptotic behaviour of solutions of NFDEs is considered. Chapter 4 is devoted to SFDEs. Some aspects of the theory of these equations and general methods of stability investigation are formulated. Stability conditions of specific systems are given and the asymptotic properties of solutions of SFDEs are studied. We use double enumeration of theorems and formulae inside the same chapter and triple enumeration when they are referred to in other chapters. We provide an adequate but not complete reference list. Some of the publications cited include rather extensive references to the original papers. We would like to thank the many people to whom we are deeply indebted, particularly, M. A. Krasnoselskii, A. D. Myshkis, M. I. Monastyrskii, L. E. El’sgol’tz, R. Z. Hasminskii, A. M. Zverkin, L. E. Shaichet and D. D. Bainov for their most valuable help during the preparation of this volume.
Chapter 1
Theoretical Foundations of Functional Differential Equations
$1. TIME LAGS I N ENGINEERING
In this section we shall give some examples of engineering systems which require taking into account the systems aftereffects for their accurate qualitative and quantitative descriptions. 1.1. Reactor (Technological Lag)
Reactors and mixing tanks are standard elements of chemical engineering systems [143,203]. Consider the system shown in Fig. 1.1. The reaction vessel (or mixing tank) is filled with two liquids through pipes A and B. There is a tap on pipe B. After a chemical reaction (or mixing), the liquid leaves the vessel through pipe C. The pH value of the liquid in pipe C is measured with a pickup. The pickup signal transformed by a controller reaches an actuator that sets the tap into motion. The physical and chemical processes in the reactor are distinguished by their complexity. Usually a simplified reactor model is used which describes the reactor by the transfer function W,(s) = K exp(sh)/(Ts 1). The factor exp( sh) is introduced in the transfer function to take into account the fact that a change in the volumetric rate of liquid entering the system causes a change in the amount of liquid leaving the system only after a time h. In reality, a certain time is needed for the mixing of liquids in the vessel, for a chemical reaction and for the transportation of liquid from pipe B to C . This time is assumed to be h. The time lag can be large enough. For example, the delay that arises in the description of the process of cellulose cooking is about 3 min, the delay in the description of the limestone drying process in rotating furnaces reaches several tens of minutes C681.
+
1
2
1. Theoretical Foundations of Functional Differential Equations
I
Liquid
I
I
Liquid
I
Actuator
T a l  D  7 ,
Reoctor
Controller
C
W
>
Fig. 1.1. Feedback system for a chemical reactor.
Such a time lag can be called a technological lag. We come across technological lags in chemical engineering (production of sulfuric acid or glass, drying, and so on [68,203], in thermal power [62,173], in economic models [33,60,102,174], etc. The stability of reactors closed by different controllers with regard for hysteresis and saturation in an actuator is investigated in Chapter 2 $6 and Chapter 3 §4.
Fig. 1.2. Scheme of rolling on the continuous rolling mill.
3
$1. Time Lags in Engineering
rus rr Fig. 1.3. Determination of equivalent time lag: curve 1, response on the unit step; curve 2, response on the ramp input.
1.2. Rolling Mill (Transportation Lag) The regulation of the thickness of the outgoing plate is one of the most important problems in designing a rolling mill feedback system [56]. Figure 1.2 represents a scheme of rolling on the continuous rolling mill. In this figure the following notation is used: Rolling rates V,, V, after the first and last stands change from 0.51.5 m/sec (strip settingup rate) to 2535 m/sec (operative rolling rate); TR1, TR2 are input and output plate thickness regulators; TP1, TP2 are plate thickness pickups; EDS represents stand electric drives; EDP means electric drives of housing screw; and L,, L, are 1.5 m. the distances from actuators EDS and EDP to pickups equal to The adequate description of the rolling mill requires taking into account a time lag z equal to z = z1 + zee+ tep. Here z, = Li/V,, the transportation lag which varies from 0.05 to 1 sec; t,, is the equioalent time lag of stand electric drive equal to 0.3 to 0.4 sec; and zePis the equivalent time lag of pickup equal to 0.1 to 0.3 sec. Time lags z, and rep may be found by using transient responses. The method of their determination is illustrated by Fig. 1.3 in which the responses on the unit step and ramp inputs are shown (for details see Draljuk and Sinaiskii [56] and Rotach [l89]). Both types of lags often occur in other problems also. The transportation lag is usually encountered when substances, energy or signals are transmitted to certain distances. It is necessary to take this type of lag into account in heat power engineering (tranportation time of fuel from tank to furnace), in astronautics (time of wave propagation from the Earth to the Moon) and so on.

4
1. Theoretical Foundations of Functional Differential Equations
1.3. Ship Stabilization (Informational Lag) Consider the simplest model of ship stabilization [160(2)]. Assume that ship dynamics is described by
[email protected]+h4=K$, I>O, K>0, h2O (1.1) where cp is the ship deviation angle and $ the turning angle of the rudder. Assume also that the change of rudder angle $ is governed by the automatic helmsman rule T & + J /= t i t + B [ , T>O (1.2) Here ( is a measured value of the ship deviation angle. In practice ( ( t ) # q ( t ) because it is impossible to measure ship deviation instantaneously. Assume that = q(t  7) (1.3) Under these hypotheses the ship with an automatic helmsman is described by an RFDE
+
+
TZiji(t) ( T h  Z)+(t)  h+(t) + K/
[email protected](t Z) KUq(t  Z) = 0 (1.4) One of the important problems is the choice of automatic helmsman parameters a and /? which guarantee the stability of a closed system (1.4). Notice that the measure rule (1.3) is often replaced by
+ ((t) = d t )
(1.5) Such replacement sometimes yields true qualitative conclusions. In Figure 1.4 the full line limits the stability domain for the measure rule (1.3) and the dotted line for rule (1.5) [160(2)]. The time lag arising in this problem can be called an informational lag. It may occur in a number of control systems which contain subsystems of Z&)
a
I Fig. 1.4. Stability domains: solid line, measure rule (1.3); dashed line, measure rule (1.5).
5
$1. Time Lags in Engineering
Controller
Informational lag es~
w,(S)
Fig. 1.5. Control system with informational lag.
transfer and processing of information. An informational lag is usually modelled by a unit with transfer function exp( sz) (see Fig. 1.5).
1.4. Manual Control In manmachine systems it is important to have a model of human response [22,24,201]. Figure 1.6 shows a structural scheme of manual control for a singleloop system. Here Wr(s), W,(s)and W,(s) are transfer matrices of the signal reproduction system, of the plant and of the measure system, respectively, and n, are external noises. Man in this scheme is represented by a nonlinear unit with input y ( t ) and output c(t). Man perceives input signals by image, sound and acceleration channels. Man's output signal is a muscle response. Experimental researches [24,201] shows that in first approximation man's transfer function is given by WXs) =
+
K exp(st) 1 asT, l+sTl l+sT,
Here lag t evaluates the time interval necessary for the signal to reach the human central nervous system and then the muscles. This lag lies in the interval 0.15 to 0.25 sec, and it is almost constant for every man. Time I "1
I "3
Fig. 1.6. Stochastic model of manual control.
6
1. Theoretical Foundations of Functional Differential Equations
Table 1.1 Signal frequency
TI
T2
T
a
K
0.16 0.32 0.48
25 9.1 5
0.67 0.2 0.09
0.15 0.20 0.25
0.08 0.055 0.067
100
40 15
constant TI characterizes the limitation of muscle reaction speed and approximately equals 10sec. The factor a takes into account the degree of instruction, experience of the operator or type of problem. The process of instruction is more prolonged than is the process of control. For this reason it is assumed that factor a in (1.6) is a constant. Table 1.1 [24] gives experimental average values of parameters for transfer function (1.6) for different frequencies of input signal. More adequate than (1.6) is a model which contains a noise n(t) that does not correlate with the input signal (see Fig. 1.6). Such a system is usually described by an SRFDE (see Chapter 4). A nonlinear model of manual control is represented in Fig. 1.7 [24]. This model allows for a dead zone of operation and dynamics of the central nervous system and muscle response.
Fig. 1.7. Nonlinear model of manual control.
1.5. Turbojet Engine Modern turbojet engines have to operate over a large range of speeds and altitudes. The main purpose of the turbojet engine is to provide speed adequate for flight. The turbojet engine contains the following basic elements: a variablegeometry supersonic air intake, a turbocompressor with principal combustion chamber, an afterburner and a propulsive nozzle [23,41,136,159,202]. In the turbojet engine, control factors include the following: G, (the fuel quantity entering the principal combustion chamber), F , (the propulsive nozzle section area), and F,, (the crosssectional area of the overcross flap in air intake). Controlled variables are n (the turbocom
7
41. Time Lags in Engineering
Turbojet engine
Air intake
a2
Fig. 1.8. Structural scheme of turbojet engine with air intake.
pressor rotation frequency), T, (the gas temperature before the turbine), and a: (the intradiffuser parameter); a: = P,/Pi, where P , is static pressure after the shock wave and Pi is impact pressure in the air intake. Variables P (inlet engine pressure) and G, (air quantity in combustion chamber) reflect the interaction between engine and air intake (see Fig. 1.8).Linearized equations of the turbojet engine are [23,136,159,202]
+ xl(t) = K15x5(t) + K16U6(t)+ K 1 7 U 7 ( t ) Ti2(')
+ X2(t) = KZ6(M26u6(t) + u6(t>) K 2 7 U 7 ( t )
Tei3(t) + X3(t) = K35(M35Mt) + K36(M36u6(t)
qa4(t) + X4(t) = K43X3(t  51)
Ta5(t)+ X5(t) = K53X3(t  5 2 )
+ x,(t)) + U6(t))
+ K48Us(t + K58us(t
(1.7) 51)
 52)
Here xlx5 and u6u8 denote increments of n, T,, G,, E , P , G,, F, and F,,, respectively (for example, xl(t) = (n(t)  nsr)/rzsr).Also, Kij and M , denote different positive constants and z 1 and z 2 represent time lag in intake versus are time flow and time lag along flow, respectively. Further, T, and constants of the engine and the intake. It is not possible to neglect 5 , and z2 in (1.7) because TJT,E 1 and T ~ / T 1. Sometimes one uses the simplified model in which the intake is described by the transfer function
Here T~ = l/a(1  M), where 1 is the distance from the overcross flap to the pressure pickup P,; a is a sound velocity and M is a Mach number in the subsonic part of the intake.
8
I. Theoretical Foundations of Functional Differential Equations
1.6. Nuclear Reactor Nuclear reactors are complex feedback systems, described by different models [67,70].Some of these models use RFDEs. For example, in [70] the following model is investigated: N ( t ) = [ aT(t  h)

E(N(t)  N,,)]N(t)
?(t) = r [ N ( t )  N,,  T ( t ) ] ,
a, E r > O
Here N ( t ) denotes reactor power, N,, is a stationary reactor power, T(t)is a deviation of reactor temperature from the stationary value and h is a time lag. One of the important problems of nuclear reactor analysis is a study of their stability under different disturbances. Some stability results are given in Chapter 2 $7 (see also Gorjachenko [70]).
1.7. Infeed Grinding Model* A mathematical model of grinding has to take into account the action of mz), dissipative forces ( F J , elastic forces (Fc), grinder element masses (ml, stiffness of the piece and grinding wheel contact and other phenomena. The grinding process is schematically shown in Fig. 1.9. As control factors we take the crossfeed rate V and the angular velocity of the piece rotation w1 (roundfeed). Output variables are the radial F , and tangential F , components of the cutting effort. In the linear model of grinding one uses the transfer functions
+
Here z1 = 2710~, zz = 2n(w Aw)'and ai, b,, ci, di, hi, gi are constants. The time lag appearance in this model is conditioned by the dependence of the grinding process on the surface state at the previous rotation. Relations (1.9) show that grinding can be described by an NFDE of fifth order. Other cutting models using RFDEs have been studied by Shil'man [204].
* From
Michelkevich and Chabanov [146].
9
$2. Aftereffects In Physical, Biological and Other Systems
Fig. 1.9. Scheme of infeed grinding.
$2. AFTEREFFECTS IN PHYSICAL, BIOLOGICAL
AND OTHER SYSTEMS 2.1. Microwave Oscillator According to Gorelik [69] the equation
+
%(t) 2rx(t)
+ w2x(t) + 2qx(t  1) =
&X3(t 
1)
(2.1)
describes the vacuumtube oscillator (Fig. 1.10) and takes into account the time lag caused by the finiteness of the time that the electrons require to pass
Lc Fig. 1.10. Microwave vacuumtube oscillator.
L+
10
1. Theoretical Foundations of Functional Differential Equations
ZfT)
 C

L
Radiator
03
Reflector
 R .
Amplifier

/A
4
V (sound velocity)
Fig. 1.11. Radio system with transportation time lag.
from the cathode to the anode of a vacuum tube. For a similar problem in Rubanik [193] an equation of the Van der Pol type (2.2) is used:
+
%(t)
[email protected])

f(x(t

z ) ) i ( t  z)
+ X(t) = 0
(2.2)
The greatest interest in investigation of Eqs. (2.1) and (2.2) is the search for periodic solutions and their stability studies [9,66]. An equation similar to (2.1) has been obtained by Minorsky [149(2)]. Minorsky dealt with the system shown on Fig. 1.11. For current I ( t ) he has obtained LCijt)
+ RCj(t) + Z ( t ) + AKLI'
(
t   = BK3Li3(t

28/44
The interaction of coupled oscillators is described by Starik [221] as y(t)
+
+ 62 sin cp(t  z,)]y(t)
[02
+
=  ~ [ h j ( t ) yy3(t  z2)1
Z @ ( t ) = E[L(@(t))  Hcp(t)  a,y2(t  t3)cos
cp(t)
 (r2 sin q(t)  (r3 cos cp(t)]
The feedback system stability for the twomirror spherical antenna is investigated in Chapter 2 $4. 2.2. Systems with Lossless Transmission Lines Neutral functional differential equations are frequently used for the study of distributed networks containing lossless transmission lines. In Bogomolov [25] and Solodovnikov [219(1)] NFDEs appeared in studies of feedback systems for the hydraulic turbine. Let us consider in detail one example of this type. Let the system consist of a long electrical line (cable) of length I, one end of which is switched on a power source E with resistance R, while the other end is switched on an oscillating circuit formed by a condenser C , and a nonlinear element, the voltampere characteristic of which is i = g(u) (Fig.
$2. Aftereffects In Physical, Biological and Other Systems
11
I
+
/=g(v)
I_c1
g(v)
1.12).Let L and C denote the linear inductance and capacitance of a long line, respectively, and assume that the line is lossless. The processes in such a system are described by the hyperbolic partial differential equations [28,108(8), 147,1481 Li,(x, t ) =

u,(x, t),
CUt(X,t ) =
O<x n/2, then there exists a periodic solution not equal to a constant. In this last case the population is said to be ecologically stable, since it can exist for an infinitely long time. Kolesov has considered Eq. (2.8) in a number of papers [104,250]. In particular, he has obtained the asymptotic formulae solutions of Eq. (2.8) for aL  4 2 4 1 and for aL % 4 2 . A more complicated population model, stated by Kolesov, is i ( t ) = r{ 1

k  ' [ a , x ( t  h) + a,x(t

h  l)])x(t),
a1
+ a, = 1
It is shown [250] that for h z 1 a small a2 > 0 (existence of survivals reaching longevity) implies a decrease of oscillations, but for h z 2 any number of longlivers implies an increase of oscillations. Thus it permits us to explain the reasons for more intensive oscillations in Iakoutia (a region in the north of the USSR) of hare quantity in comparison with nutria quantity. Other generalizations of Eq. (2.8) are available [lo, 40,103,224,239(3)] as are other ecological models [lo, 14(2), 50, 71, 2461.
15
$2. Aftereffects In Physical, Biological and Other Systems
2.5. Model of Immune Response (PredatorPrey Model) The following model of immune response is developed in [55]. Let g(t) denote the number of antigens penetrating an organism and a(t) the number of antibodies. Assume that the rate of antibody production is proportional to g(t  7). Here z is a mean time for organism reaction on the antigens. For real diseases z is great enough (34 days). In linear approximation the dynamics of the immune process is described by g(t) = W t )

QdtMt)
(2.9)
b(t) = Ag(t  T)  Rg(t)a(t)  Ea(t)
System (2.9) has two stationary points: I = (0,O) and I1 = [ K / Q , K E / ( C A Q  KR)]. Point I is always unstable. The stability of point I1 depends essentially on the delay z. For small z point I1 is asymptotically stable. This means biologically that the disease does not progress, the number of antigens is practically constant and the organism becomes an infection carrier. For great z there is a periodic solution in System (2.9) which can be interpreted as a chronic disease. The numerical simulations show that the oscillation amplitude decreases with the decrease of z. The time interval when the number of antigens is nearly zero (from 7 . to 8.5. lop4 of the stationary value) decreases and reaches 0.5 to 0.7 of the oscillation period. Such an interval corresponds to almost complete convalescence, but there always exists a danger of relapse. More complicated models of immune response have been developed by Marchuk [21, 139(1)139(3)]. In Chapter 2 $8 we shall investigate them in detail, Other immunological models are also available [lo, 14(2), 26,27, 1353. Equations similar to (2.9) have been considered by Volterra [239( 3)], Wangersky and Cunningham [245] and others [SO, 1041. One paper [245] deals with the equations i ( t ) =
[email protected]) cx(t)y(t)  b X 2 ( t )
)'(t) =  e y ( t )
+ c,x(t  h)y(t  h)
and [239(3)] with the equations i(t)=
[
b , x ( t ) 1  c , , x ( t )  c12
y(t

s) d a 4
I rm
bi > 0,
cij > 0,
da,(s) > 0,
da,(s) = 1, J O
i = 1, 2
16
1. Theoretical Foundations of Functional Differential Equations
9''q
I
C
Fig. 1.14. Mechanical model of the bloodstream.
The modification of the LotkaVolterra model has been proposed [ 10,1041 i ( t ) = r,[1
+ cr(1

y(t)/K,)  x(t  h,)/K,]x(t)
3 ( t ) = rzCx(t)/K1  Y(t  h,)/K,lY(t)
2.6. Distribution of Albumin in the Bloodstream* The following mechanical model can be used to explain the distribution of albumin in the bloodstream. Consider the system shown in Figure 1.14 in which A is a unity volume tank filled with water which circulates through three pipes with volumetric flow rates A A , and A, and with flow times T,, T, and T3.Assume that at t = 0; 1 g of dye is introduced into tank A , which is mixed instantaneously. Thus, the initial dye concentration in the tank is unity and the initial dye concentration in the pipes is zero. Dyed water leaves the system through a pipe C with volumetric rate L, and pure water enters the system at B with the same flow rate. It is assumed that no mixing takes place in the pipes, that the diffusion is neglected and that the flow velocity is assumed to be constant over a cross section of any pipe. Let x ( t ) denote the dye concentration in the tank at time t . A material balance on dye in the tank leads to the RFDE
,,
i(t) =
+ + + + +
+
(A, A , A3 L ) x ( t ) A , x ( t A , x ( t  T,) A,x(t  T3)

TI)
The initial conditions describing the dye concentration are given by x ( t ) = 0, T3 I t < 0, ~ ( 0 = ) 1. Other biomedical models have been proposed [14(2), 139(3)].To describe
* From
Bailey and Reeve [ i l l .
17
$2. Aftereffects In Physical, Biological and Other Systems
20
c
I
1900
1920
I
I
1940 1960 Year
I
1980
Fig. Z.15. Total petrol extraction in the USA (in millions of tonnes per day): solid line, real extraction; crossmarks, prognosis based on dependence of lag on the time of extraction.
the behaviour of the central nervous system in the process of study, Shimbell [206] has proposed the following model: X ( t ) = k [ x ( t )  x(t 
l)][N  ~ ( t ) ] , t > 1
k(t) = k x ( t ) [ N  x ( t ) ] ,
(2.10)
0 I t 5 1, x ( 0 ) = xg
The stability of Eq. (2.10) is investigated in detail in Chapter 3 95. 2.7. Some Economic Applications Retarded functional differential equations are used in Kobrinskii and Kus'min [1023 for description of different economic processes. The equation j ( t ) = ky[t  z(t)] allows the influence of production time on characteristics of the process and is called a reproduction equation. The delay z(t) in the reproduction equation changes with the age or size of the system. The reproduction equation can be used for analysis and prognosis of economic characteristics. As examples, two prognoses are shown in Figs. 1.15 and 1.16 [102]. For an economic interpretation of Eq. (2.7), see Kobrinskii and
18
1. Theoretical Foundations of Functional Differential Equations
220 210
8 v)
._ a
20c 19c
u 0
5
180
c 0 F
17C 16C 157 117
70.E
I
I
I
I
I
I
t950 60
I
I
I
I
I
I
I
I
I
70 71 75
65 Year
Fig. 1.16. Production of machine tools in USSR (in thousands of pieces): solid line, real production; crossmarks, prognosis based on reproduction equation.
Kus’min [102]. Other economic models which take into account the aftereffects may also be found [33,60, 102,1741.
53. INITIALVALUE PROBLEM FOR FUNCTIONAL DIFFERENTIAL EQUATIONS 3.1. General Equations with Delay In this section we study equations with unknown functions ~ ( tE )R , depending on a scalar argument t. Here R , is any real ndimensional normed vector space with norm Ix(t)l. Let M denote a complete metric space of continuous functions cp(0) mapping the interval ( GO, 01 into R , with the metric p. As M we can take the space C [  GO, 01 of continuous functions cp: (  co,01 +R , with a countable system of seminorm IIcpllj and a metric p : t~ Cj, 01, j = 1 , 2 , , . .
llcpllj = maxlcp(t)l, a,
p ( ~$1 , =
1 2’1Iq j=1

ll/IIj.(1
+ IIq  $IIj)’
19
93. InitialValue Problem for Functional Differential Equations
Another example of M is a normed space CS[  m,O] of continuous and 0. bounded functions on ( cc,01 with norm IIqlle = sup)q(8)(, co < 8 I An important example of M is a space C [  h, 01 of continuous functions q(8) mapping the interval [h,O] into R, with the usual norm llqll = maxlq(8)I,h I 8 I 0. A number of space M is given by normed spaces, coordinated with the equations (see Chapter 2 92 and Kolmanovskii [108(3)]),and also by spaces “with memory” [4244,473. Designate QH a sphere in the space M
Denote x, = x,(8) = x ( t
(3.1)
=(~PEM ~ ,( ~ 3i: 0 ff) )
QH
+ 8 ) an element of M . We say that the relation
i ( t ) = f ( t , x,),
x, = X(t
+ e)
(3.2)
is a general retarded functional differential equation on M and denote it by an RFDE. In (3.2) f : R , x M + R, is a given function (map) and the overdot 0, then Eq. (3.2) is an represents the righthand derivative. If cc < 0 I RFDE with unbounded delay. If  cc <  h I 8 i: 0, then (3.2) is an RFDE with bounded delay. Let to be an initial moment and let ( ~ ( 8E) M be a given function. The initialvalue problem (or Cauchy problem) for Eq. (3.2) is as follows. Define a t o and which has function x ( t ) which satisfies Eq. (3.2) for t 2 to and for t I the initial data X,” =
x(t0
+ 8) = q(8)
(3.3)
3.2. RFDE with Bounded Delay The simplest RFDE is i ( t ) = f [ t , ~ ( t )x(t ,  h)],
h >0
(3.4)
Equation (3.4) is also called a differential difference equation. The solution x ( t ) of the initialvalue problem (3.4), (3.3) can be obtained by a step method. This method permits us to define a continuous solution x ( t ) of problem (3.4), (3.3) by successive integration of ordinary differential equations. Replacing x(t  h) for t E [to, to + h] in (3.4) by q ( t  h), we obtain the ordinary differential equation i(t)= f[t,~
( t )V,( t  h)],
to I t I to
+h
Having defined the solution x ( t ) on the interval [ t o , t o + h), we can continue in a similar way and define x ( t ) on [to h, to 2h], [to + 2h, t o 3h] and so
+
+
+
20
1. Theoretical Foundations of Functional Differential Equations
on. For example, the solution of the equation i ( t ) = ax(t  l), t 2 0, with the initial condition x,(0) = 1 , t I 0, is given by the formula N
an[t  ( n  l)h]"(n!)',
x(t) =
N
=
[th'1
+ 1.
n=O
Here [s] denotes the greatest integer in s. The step method is not universal. For example, it cannot be applied to the equations with timevarying delays which vanish in some points: k(t) = f [ t , x ( t ) , x(t  h(t)],
h(t) 2 0, h(ti) = 0,
ti
> to
But the step method can be used for general equations (3.2) if f ( t , x,) depends only on x ( t ) and x(0) for t9 I E < 0. In some cases the step method allows us to obtain different theorems of existence, uniqueness and smoothness of solutions and other qualitative properties. Unlike ordinary differential equations, the solution of (3.4), (3.3) even for infinitely differentiable functions cp and f has, in general, a discontinuity at the point t = t o . Really, i ( t , + 0) = f[O, cp(0)l and is not always equal to k(to  0) = @(O). At point t = to + h the solution has, in general, a discontinuity in the second derivative, but its first derivative is by all means continuous, and so on. Thus the solution becomes more smooth from step to step. This process is called a smoothing of solutions. Let us outline another characteristic peculiarity of RFDEs. For the RFDE (3.4), a solution exists to the right of the initial point t o and does not exist, in general, to the left of t o . Consider, for example, the equation k(t) = x ( t

l),
t 2 1;
x(t) =
+
0I t I1
cp(t),
(3.5)
Rewritten in the form x ( t ) = i ( t l ) , this equation means that on [  1,0], x ( t ) = +(t 1 ) ; on [  2 ,  13, x ( t ) = i ( t 1 ) = @(t 2); and so on. Solution x ( t ) is expressed by q(")(t)and m + 00 when t +  00. If q(t) is only
+
+
+
continuous, then the solution x ( t ) does not exist to the left o f t = 0. If there exists q'"'(t) but cp("+')(t)does not exist, then the solution x ( t ) of (3.5) exists only on [  m, 01. In the case of Eqs. (3.4) or (3.2) the situation is more complicated still. The solution to the left of the initial data is sometimes referred to as the backward continuation of a solution. The problem of backward continuation is incorrect in the Hadamard sense. Consider again Eq. (3.5). Assume that the initial function cp(t) is not known exactly; i.e., we know only that the function @(t)= q(t) 6(t) with some estimate 6, of error 6(t), Ild(t)ll Id,.Under these assumptions it is impossible to determine x ( t ) to the left of t = 0 with precision 6, by the method described previously in this section. In fact, if we let 6(t) = 6, sin ot,then x ( t ) = @ ( t )= @ ( t )+ 6,o cos ot and IIx(t)  +(t)ll = 6,w + 00 if o + oc). In this case it is necessary to apply a certain regularization method for backward continuation. Notice also
+
$3. InitialValue Problem for Functional Differential Equations
21
that some applied problems require defining the initial state of the system from the known current state [107]. Let us give some known results [59,81(4), 156(3)] about existence, uniqueness and continuous dependence of RFDE solutions with bounded delay:
A function x ( t ) is said to be a solution of (3.6) on [ t o ,to + S), 6 > 0 if x E C[to  h, t o + 6 ) and x ( t ) satisfies Eq. (3.6) for all t E [ t o , to + 6). We assume that i ( t o ) = f (t, cp).
Theorem 3.1. Suppose that 0 G R x C[h,O] is open and that the map f (t, cp) is continuous and satisjies a local Lipschitz condition in cp. If (to, cp) E 0, then for some 6 > 0 there exists a unique solution of problem (3.6) on the interval [ t o ,to + 8). This solution depends continuously on initial data.
3.3. RFDE with Unbounded Delay Consider problem (3.2), (3.3) with unbounded delay. First assume that the initial function q(0) E C [  co,O]. The definition of the solution is the same as in the case of bounded delay. Let the map f ( t , cp), f : R , x C [  00, 01 +R , be continuous and satisfy a local Lipschitz condition in cp:
If ( 4 cp)
 f ( t , 11/11 I Wcp, $1
for all cp, 11/ from some subset 0 c C [  m,O] and for all t E [to, to + 6). Under these assumptions, and using a fixedpoint theorem, it is possible to prove local existence, uniqueness and continuous dependence (in the sense of metric p ) of the solution of problem (3.2), (3.3) on initial data. Other theorems can be obtained if the initial function is taken from a subspace of C [  co,O]. For example, assume that cp(@ E CB[  co,01 but that C B [  m,O] is a subspace of C [  co,01. Thus the set of continuous maps from CB[  co,01 into R , is more powerful than the same set for C[  co,01. This fact implies that the conditions of existence and uniqueness for cp E CB[  co,O] will be less restrictive than for cp E C[  co,01. Other spaces of initial functions are considered in Corduneanu and Lakshmikantham [47]
3.4. NFDE Investigations of NFDEs are of great interest but are more complicated than RFDEs. Consider the NFDE
22
1. Theoretical Foundations of Functional Differential Equations
An initialvalue problem (Cauchy problem) for N F D E (3.7) can be stated as follows: Given an initial function cp and an initial moment t o , find a solution x : (  co,t o + 6) + R, such that
x,, = x(to + e) = cp(e),
it =, i(to
+ 01 = 4(e)  a < e I o
(3.8)
The initial function q ( B ) in (3.8) is absolutely continuous. The simplest NFDE is a neutral diferential diference equation i(t) =
f [t, ~ ( t )~, ( t h), i ( t

h)]
(3.9)
Here f is a given continuous function. The step method can be used in order to find the solution of (3.9). It is easy to verify that there does not exist a smooth solution even for the simplest NFDEs with an arbitrary smooth initial function. If the initial function satisfies a socalled splicing condition @(O) = f [ t o , cp(B), 43(d)], then the solution of (3.7), (3.8) has a continuous derivative. Solutions of N F D E (3.9) do not smooth, unlike solutions of RFDEs. Notice that if it is possible to obtain i ( t  h) from (3.9), then the solution of (3.9), (3.8) always exists, not only to the right of initial interval [ t o  h, t o ] , but also to the left of it. The “splicing” condition is too restrictive for a number of problems. If it does not hold, then the solution of (3.7), (3.8) must be understood in some generalized sense [l, 253(2)]. Introduce a space L,[ h, 01 of functions cp(8), cp: R , + R,, which are measurable and essentially bounded on [ h, 01. Define the norm IIcpll, in L , [  h , O ] by the formula Ilcpll, = ess suplcp(O)1,  h < 8 I 0. A function x ( t ) is said to be a solution of problem (3.7), (3.8) if x ( t ) is absolutely continuous on [ t o  h, t o + 6) for 6 > 0, i ( t ) E L , [ t o  h, to 6) and x ( t ) satisfies (3.7) and (3.8) almost everywhere. Another equivalent definition of solution of (3.7), (3.8) is given in Achmerov
+
PI.
Assume that the map f ( l ,cp, $), f : R , x C[ h, 01 x L , [  h, 01 + R, is continuous for some open set R c R , x C[  h, 01 x L , [  h, 01 and is Lipshitzian in cp. Then
I f ( 4 91,$1  f(4 (PZ, $11 I L l l ~, (~211,
( t , ~ 1( ,~
2 ) ~ (3.10) a
The map f ( t , cp, $) satisfies a modijied Lipschitz condition in $ with the constant less than 1. It means that for any functions $,(#) and $,(8) such that t,bl(Q)= $2(0),  h I 0 I E < 0 for some E > 0, we have
I f ( 4 cp,
*A
K < 1,

f ( t , cp,
$2)1
I
Kll*1  $ 2 l l m
(4 cp, $11, (4 cp, $z)E R
(3.1 1)
Theorem 3.2 [253(2)]. Let R G R, x C [  h, 01 x L , [  h, 01 be open and let the map f ( t , cp, $) satisfy the hypothesis formulated above. I f ( t o , cp, 4) E R,
23
$3. InitialValue Problem for Functional Differential Equations
then there exists a unique solution of (3.7), (3.8) on the interval [ t o ,to + 6) for some 6 > 0. This solution depends continuously on initial data.
An interesting class of NFDE is (dldt)Cx(t) G(t, X J l ~ , , ( 8 )= ( ~ ( 8 ) ,
= F(t, X t ) ,
h
t 2 to
(3.12)
I8 I 0
Here maps G, F : R , x C [  h, 01 + R , are continuous. A function x ( t ) is said to be a solution of(3.12) if x ( t ) E C [ t o  h, t o + S ) , 6 > 0, x(t) is equal to q ( 8 ) for t E [to  h, to], the function x ( t )  G(t,x,) is continuously differentiable and this function satisfies Eq. (3.12) for all t E [to, to + 6).
Theorem 3.3 [81(4),831. Let cp(8) E C[  h, 01, and let the maps G and F be continuous in an open set Q G R , x C[  h, 01. T h e map F ( t , cp) is Lipschitzian in cp: IF(t, cp)

F(4
$11
I
Ucp  $11,
( t , cp), ( t , $)En
The map G(t, cp) satisfies in Q a modified Lipschitz condition in cp with a constant less than 1: IG(4 cp)  G(t, $11
K 0. This solution depends continuously on initial data. Other existence and El, 5,30,81(4), 95(2)1.
uniqueness
theorems
are
also
available
3.5. Two Useful Lemmas Lemma 3.1. Let a scalarfunction x ( t ) satisfy an inequality i ( t ) s  yx(t) + PIlxJ, t 2 t o , x , = x ( t 8),  h I 8 5 0 . Here y , B are constants such that y > B 2 0. Then Ix(t)l I IIxt,I/ exp[A(t  to)], t o I t, where A is a unique positive root ofthe equation i= y  j? exp(ih) [79(4)].
+
Lemma 3.2 (GronwallBellman inequality) [181. Let scalar continuous functions x(t), and g ( t) 2 0 satisfy the inequality x ( t ) I a(t)
+
1:
g(s)x(s)ds,
t20
Here a ( t ) is a nondecreasingfunction. Then x ( t ) I cr(t) exp[ro g(s) d s ] . For the proofs of Lemmas 3.1 and 3.2, see, for example, Kolmanovskii and Nosov [108(5)].
24
1. Theoretical Foundations of Functional Differential Equations
3.6. Estimations of Solutions Growth
Let us derive an a priori estimate on how the solution x(t) of problem (3.7), (3.8) grows. Let the map f be measurable and let
with the initial function q(8) E CB[  00, 01 and @(8)E Lm[ m,O], Ilcpll = ess sup1 q ( O ) l ,  GO < 8 I 0. Lemma 3.3. Ifthere exists a solution x(t) ofproblem (3.7), (3.Q thenfor some positive constants C, C4 the inequality holds:
Ix(t)I ProoJ
+ I I q I I m + 11cPIIm)expC~4(t  to],
C,(1
t2
(3.14)
to
From (3.7), (3.8) it follows that Ii(t)l
ri(t) = Ci(1
I r1(t) + C2r2(t),
+ I l d l m ) + C211dIm + Ci supIx(~)l, to I
r2(t) = sup\i(t)\,
T
(3.15)
t 2 to to
I
T
I t
4t
Since the righthand part of the inequality (3.15) does not decrease, we have r 2 ( t )I r l ( t )
+ C2r2(t)
or
r2(t) I ( 1  C2)'r,(t)
(3.16)
Integrating (3.7) from to to t and taking into account (3.13), (3.15), and (3.16), we obtain
1: +1
I x(t>l I I x(t0)l + I (x(to)l
I f(s, xs, xs)l ds
rl(s)[l
+ (1  C 2 )  ' 1 ds
The inequality (3.1 5) yields SUP IX(T)l I Ix(t0)I
+ ( t  t,)Cf + (1  G
 l I
105.ISi
. [ ~ 1 ( 1 +
SUP
I I ~ I +I c~2 1) 1 4 1 1 + m
1
c , c ~(1+ ~ 2 )  1 1
Ix(T)~ds
i0SISS
This inequality and Lemma 3.2 imply the estimate (3.14). Lemma 3.3 is proved.
25
$4. Linear Equations
3.7. Methods for Solving an FDE
This section is devoted to the brief discussion of certain methods of solution finding for an FDE. As mentioned previously, the step method is an effective method for solving some classes of FDEs. But its actual application is sometimes embarrassing, especially for large intervals. O n the other hand, the ordinary differential equations arising in the step method are analytically integrable only in rare cases [57(2), 59,156(3)]. Series development is another method of finding solutions [59,152]. Some asymptotic methods may be used for an FDE with small delay [187( l), 187(2), 234,2351. Applications of all of these methods are rather difficult and these methods are not universal. The most universal method is numerical integration of an initialvalue problem. The solution of problem (3.2), (3.3) is calculated in the following way. First we divide the interval [ t o , T] into N parts, by the points ti = to + iz, z = ( T  t o ) N  ' . The values x i of the solution x(t) at points t i = x ( t , ) are defined by the relation xi+
= xi
+ zf [ t i , x i , x(ti  h)]
(3.17)
If the stepsize z for the numerical integration is chosen such that mz where m is an integer, then relation (3.17) takes the form = x , + T ~ ( ~ ~ , x ~ , x ~  , x, , =) c; p (  k z ) ,
k s O
=
h,
(3.18)
Recurrent relations (3.18) permit us to define the solution of problem (3.4), (3.3) on any time interval. For a survey of numerical methods for FDEs, see Cryer [49].
54. LINEAR EQUATIONS
4.1. General Form of a Linear RFDE Consider the linear RFDE
W =W , xt> + f ( t ) ,
x(t> E R,
(4.1)
Here the operator L: [O, a)x C [  co,01 + R , is continuous and linear in x,; i.e., for any numbers a, p and functions cp, $EC[CO,O] we will have L(t,acp + p$) = ctL(t, cp) + pL(t, $). Using the Riesz theorem, Eq. (4.1) may be written in the following equivalent form [13,186] (4.2)
26
1. Theoretical Foundations of Functional Differential Equations
Here R(t, s) is an ( n x n ) matrix, R(t,0) = 0 and for any t 2 0 there exists o(t) such that R(t, s) = const for o(t) 2 s. For all t 2 0 the function R(t,s) has a bounded variation u(t) in s; the function R[t, o(t)] is continuous and R(t,s) is mean continuous in t . If, in addition, o(t) = h, then we have an RFDE with bounded delay Ph
4.2. A Priori Solution Estimate
Obtain the estimate of the solution of Eq. (4.2) with the initial condition (P(@ E CBC
4, = q(@,
a,01
(4.4)
Show that
If u(t) is a bounded function, then (4.5) coincides with (3.14). If Ix(s)l 5 IIqllB, to 5 s I t , inequality (4.5) is obvious. Otherwise, denote t , € ( t o ,t] such a Integrating Eq. (4.2) we obtain point that m(t) = Ix(t,)l > l/qllB. x(t> = x(t0)
+Jidi
JOrn
IId,R(i, 41x(5

s)
+ 11f(5)d 5
+
61
Hence m(t> = Ix(t,>l 5 IlqllB
+
1:
v(s>m(s)ds
If(s>l ds
(4.6)
From (4.6) and Lemma 3.2, (4.5) follows. Inequality (4.5) means that the solution of (4.2), (4.4) is extendable on the semiaxis [to,a). Let us indicate some properties of RFDEs that distinguish them from ordinary differential equations. Usually the solution space of a linear RFDE is infinite dimensional, while for ordinary equations it is always finite dimensional. However, there are cases in which the solution space of an RFDE is finite dimensional for sufficiently large t ; i.e., a part of the solution space is bound together [59,156(4), 156(5)]. Further, in contrast to equations without delays, the solution of scalar linear homogeneous RFDEs with
27
$4. Linear Equations
bounded coefficients may decrease faster than does any exponent. For example, the equation a(t) = 2tx(t  1) exp(1  2t) has the solution x ( t ) = exp(  t2), t 2 0. The solutions of ordinary differential equations cannot decrease as quickly.
4.3. Linear NFDEs Linear NFDEs have the form i ( t > = L(t, X I ,
i t )
+f(t)
(4.7)
Here the continuous operator L: [0, co) x C [  h, O] x L , [  h, 03 + R , is linear in x , and i t .Like (4.3), Eq. (4.7) may be written as i(t)=
Jr,
[ d , K o ( t , s)]x(t
+ s)
+ j O h [ d S K I ( ts, ) l i ( t + s) + f ( t >
(4.8)
Matrices K O and K , are assumed such that the corresponding operator L(t, x,, i t )is continuous. The initialvalue problem (4.8), (3.8) may be stated for the NFDE as in 53. Theorem 3.2 implies the existence and uniqueness of the solution of problem (4.8), (3.8) under the assumptions ro
(4.9) SUP
J
ttto
~ d , ~ , (S)It ,
=
1 Pi(a, s) i=l
Here the notation to means that the point t o is exc.Jded from the integration interval. Instead of (4.1.2), we introduce an ( n x n ) matrix y ( t , s), y ( t , t ) = I . The matrix y(t,s) satisfies Eq. (4.8) with f ( t ) = 0 as a function of t for any fixed s and as a function of s for any fixed t y ( t , S)
+
s:
y ( t , z)Q(s, Z) dZ = I ,
s
It
s>t
y ( t , s ) = 0,
(4.13) Let Equation (4.8) be of the form rn
i(t)=
C Bix(t

hi) + Boi(t  h ) + f ( t ) ,
t >0
(4.14)
i= 1
The solution of problem (4.14), (4.4) will be 0
x(t) = Y(t0, W
t O )
+ i1 =l
81
1 y ( t , s + hi + jh)Bicp(s)ds
 h .L J'  0
h j=1
+ j=O
J: y(t, s + jh)B',f(s) ds
(4.15)
30
1. Theoretical Foundations of Functional Differential Equations
where = [th'1 + ~ ( h h[t/h]) and [ t ] is the greatest integer in t. The function x ( t ) = 0 for t = 0 and ~ ( t=) 1 for t > 0. The matrix y ( t , s) in (4.15) as a function of t satisfies Eq. (4.14) with f ( t ) = 0, and as a function of s
From (4.15), (4.16) it follows for the RFDE (4.14) with B,
=0
that
where the ( n x n) matrix y(t, s) satisfies the equations
At last for the RFDE (4.Q with K ,
= 0 by virtue of (4.13), we have
where the matrix y ( t , s) is
[d,K,(t, s)]y(t, s)
= at
h
31
$5. Linear Autonomous Equations
55. LINEAR AUTONOMOUS EQUATIONS
5.1. Characteristic Function One of the important and wellstudied classes of FDEs is linear autonomous equations such as r m
i(t) =
J0
[dK(s)]x(t  s)
+ f(t),
x ( ~ ) ER ,
Here all elements of the ( n x n ) matrix K ( s ) have bounded variation on L0, a). Consider a homogeneous linear autonomous RFDE
i ( t )=
[dK(s)]x(t

s)
(5.2)
om !
We shall look for nontrivial solutions x ( t ) = CL exp(zt), where z is a complex number and CY E R , a constant vector. Equation (5.2) has a nontrivial solution of the indicated form if and only if J(z)cc = za,
J(z)
=
jorneZs
dK(s)
Thus the number z must be a solution of the characteristic equation det[zZ  J(z)]
=0
(5.3)
Here Z denotes the identity matrix. The function D(z) = det[zZ  J(z)] is called a characteristic function.
5.2. Zeros of Characteristic Function Indicate some properties of the function D(z). Since for Re z 2 0, max
i: lom 1eZ'
dKij(s)l = y < co
l < i < nj = 1
and then Eq. (5.3) has no zeros in the halfplane Re z > y. Further, if for any z, max
i:
!om/ezs
dKij(s)l < co
lsi m, q > n or p = m, q > n or p > m, q = n. Clearly, not every quasipolynomial has a principal term. Pontryagin [1751 has proved that the quasipolynomial without a principal term always has infinitely many zeros with an arbitrarily large positive real part. It is possible to describe more exactly the distribution of zeros. As a rule, the zeros are divided into those that are asymptotic and those that are nonasymptotic. Nonasymptotic zeros are situated near the origin. They can be found only numerically. For asymptotic zeros some general formulae can be found. First, consider some examples. It is easy to verify that the zeros zk of D(z) = a  exp(z) equal zk
= lnlal
+ i arg a + 2kni,
i2 =  1
(5.5)
Now take the quasipolynomial D(z) = z  a exp(z)
= 0,
a
= ( a (exp(ia)
(5.6)
The graphs of functions z and a exp(  z) cannot have more than two real points of intersection. Consequently, all asymptotic zeros of (5.6) are complex. From (5.6), taking the logarithm we obtain
AS
ReZk
= hlal  Inlz,l,
Im zk
= 12  arg Zk 4
k = 0, & I , f 2 , . . . 2kn
=
2kn
+ o(1)
(Re #?k)/lzkl+ 0, k + & CO, we have IIm zkl = [lZl2

(Re zk)2]1’2 = (zkl
+ o(1)
Thus, for asymptotic zeros of (5.6) we receive zk = lnlal  ln12kn1 + i2kn + O(1)
(5.7)
Formula (5.7) can be made more precise as zk
= lnlal  lnl2knl
+ i[(2k 5 sgn k ) + a] + O(ln k/k) 
(5.8)
Expression (5.8) shows that the distances between asymptotic zeros are greater than some positive number. The asymptotic roots of the general equation (5.3), corresponding to one RFDE, are situated on a finite number of chains of the form (5.8). Therefore,
33
$5. Linear Autonomous Equations
Re zk +  00 for k + & co. But any sector containing the imaginary axis also contains all zeros of (5.3), except a finite number of them. Notice that the distances between zeros of general equation (5.3) can become infinitely small, always being positive [lS, 100,253(4)]
5.3. Representation of a Solution as an Infinite Series We now discuss the representation of the solution of Eq. (5.2) as an infinite series of elementary solutions. Zubov [252( l)] has established the theorem on the asymptotic property of such a series. Any solution x(t) of Eq. (5.2) can be connected with a series
c m
x(t>
(5.9)
Pk(t) exp(zkt)
k=l
Here Pk(t) is a polynomial the degree of which is 1 less than the multiplicity of the root zk of Eq. (5.3). If for some numbers E > 0 and p the strip fi  E < Re z < p + E has no zeros of Eq. (5.3), then the asymptotic formula holds: x(t) =
c
Pk(t) exp(zkt) +
[exp(pt)l
Rezk>B
For a scalar RFDE of order n [18,253(4)] it is proved that series (5.9) converges and can be differentiated n times term by term in this case. For systems of RFDEs, series (5.9) does not necessarily converge. p < 0 From Zubov’s theorem follows an important corollary. If Re zk I (this condition is equivalent to Re zk < 0), then all solutions of the homogeneous equation (5.2) tend to zero for t + 00. In other words, the condition Re zk < 0 is necessary and suficient for asymptotic stability of RFDE (5.2).
5.4. Linear Autonomous NFDEs Linear autonomous NFDEs have the form [dKo(s)]x(t
i(t)=
om !
 S)
+
s:
[dK,(s)]i(t  s),
t 20
(5.10)
All elements of the (n x n) matrices Ko(s) and K,(s) have bounded variation on [0, co) and x(t) E R , . Clearly, the characteristic equation, corresponding to NFDE (5.lo), is exp( zs) dK,(s)
1
=0
(5.1 1)
The lefthand side of Eq. (5.11) is called a characteristic function of N F D E (5.10). State Some properties of Eq. (5.10) solutions with piecewise constant
34
1. Theoretical Foundations of Functional Differential Equations
kernels K o ( s ) and K ,(s), i.e., properties of neutral differential difference equations. In this case the characteristic function (5.1 1) has the form m
D(z) =
r(l)
C= O 1 aljzl exp(zklj)
(5.12)
j =1
1
For commensurable delays kmj in terms corresponding to the highest derivative of Eq. (5.10), asymptotic zeros of (5.12) are situated on some chains of type (5.5) or (5.8). In the case of incommensurable delays kmj, in addition to these chains there exists a finite number of strips that are parallel to the imaginary axis and which contain the zeros of (5.12). The real parts of the zeros fill in the corresponding intervals of the real axis densely everywhere. The imaginary parts of the zeros are almost asymptotically periodic. The multiplicity of each zero of (5.12) is finite and does not surpass (1 + m ) maxOsls m r(1) + m. The distribution of the zeros of the general characteristic function (5.1 1) cannot be studied practically. If the highest derivative of a scalar NFDE enters only with commensurable delays, then any solution of such an equation can be represented as a convergent injnite series of the type (5.9). Incommensurable delays in the highest derivative series (5.9) may diverge. Take, for example, the equations
i ( t )  i ( t  1)  i ( t  k)  i ( t  1  k)
=0
0 It I 1
+k
x(t)
= cp(t),
~ ( t= ) @(t),
0 I t I 1;
q ( t ) = ( i ) ( t  1)2,
~ ( t= ) 0,
(5.13) 1I t 5 1+ k
In (5.13) the number k is defined as follows. Take any number define the continued fraction
k= 1 40
1

1 ~
1

+ 41 + 4 2 + 43 +
c1
> 0 and
...
The numbers qk are given by the expressions qo = 1, q 1 = 1,. . . , qk+
1
=
[Q;
exp(2XaQk) + 13, . . . .
Here Qk is a denominator of the convergent of a continued fraction:
and [p] is the greatest integer in p. In Leontjev [125] it is shown that series (5.9) for Eq. (5.13) is divergent. The representation of the solution of the
35
56. Feedback Systems with Aftereffects
general equation (5.10) as an infinite series is not studied practically. For NFDE
C C
a r s d r ) (t h,)
= 0,
a,,
z0
(5.14)
r = O s=O
under the assumption that Re zk I E < 0, it is proved that all solutions approach zero more rapidly than an exponent; i.e., the trivial solution of Eq. (5.14) is exponentially stable. If only Re zk < 0, then the situation is more complicated, since the NFDE can have a chain of zeros which approach the imaginary axis. There exists a scalar NFDE with one delay, the trivial solution of which is asymptotically, but not exponentially, stable. All solutions in this case have an exact estimation x ( t ) = O(t'), 6 > 0 [78(1)]. In the case of a few delays, if there is a chain of multiple zeros which approach the imaginary axis from the left, the trivial solution may be stable or unstable [72(1), 72(2), 2163. The supercritical case in which Re zk I0 and there are infinitely many zeros on the imaginary axis is studied in Gromova [72(2)]. In this case the trivial solution may be stable or unstable.
56. FEEDBACK SYSTEMS WITH AFTEREFFECTS
6.1. SingleLoop Feedback Systems Feedback systems for the plants with aftereffect can be divided into two types: singleloop and twoloop systems [68,214( l), 242(2)]. A structural scheme of a singleloop system is shown in Fig. 1.17. Assume that transfer function W,(s) of a plant with delay is W,@) = H,(s) exp(  sz),
H,(s)
= A(s)/B(s)
(6.1)
Here A(s) and B(s) are two polynomials in s. Consider the feedback system shown in Fig. 6.1 in the case in which the controller is a standard one of the P,
Fig. 1.17. Singleloop feedback system with delay.
36
1, Theoretical Foundations of Functional Differential Equations
PD, I, PI or PID type. Let the transfer function of the plant be Wp(s>= Kp exp(  s)/(Ts + 4 )
(6.2)
Here K , is a gain, T is a time constant and T = 1 is the delay of the plant. The number q can have two values: q = 0 or q = 1. Here q = 0 corresponds to an astatic plant with delay and q = 1 corresponds to an inertial plant. Air intake is an example of such a plant (see $1). Table 1.2 lists some of the characteristics of a feedback system (6.2) closed by one of the standard controllers. In Table 1.2 we designate by K = K , K , the gain of an open system by I the integral action time and D the derivative action time. For general plant (6.1) similar conclusions hold. If the degree of the polynomial A(s) is one unit less than the degree of B(s), then the type of FDE describing the closed system coincides with the type in Table 1.2. In the case in which the degree of A(s) is two or more units less than the degree of B(s),the closed system is governed by an RFDE. So, a closed feedback system for a plant with delay can be governed either by an RFDE or by an NFDE.
6.2. Effects of Little Delays In the process of mathematical model construction one often neglects the existence of little delays. Sometimes this may lead to paradoxical conclusions [68,75, 111,1331. Consider one example of this type. Figure 1.18 shows an ideal predictor [108(5), 1651 described by the equations u(t) = K , [ u ( t ) ~(t)], ~ ( t=) K 2 u ( t )+ u(t  1). These imply for K , K , = 1 that u(t + 1) =  u ( t ) . Thus the output at moment t depends on the input at a future moment ( t + 1). This contradicts the causality principle. The paradox can be explained by the
+
L 1 j J Fig. 1.18. “Ideal” predictor.
4
I
~~
Proportionalintegralderivative (PID): W,(S)= K,(1 + Ds + l/Is)
+
Proportionalplusintegral(PI): W,(s) = K,(1 l/ls)
Proportionalplusderivative(PD): W,(s) =K,(1 + s D )
W(s)= K ,
Controller type and its transfer function
Table 1.2
+ Kexp(s)
Is( Ts
K , l s exp(  s)
K , l s exp( s)
+ 1) + K (Is + I)exp(s)
+ 1) + K ( I + Is + IDS’)exp( s)
Is(Ts
K J s exp( s) Is(Ts + 1) + K exp(s)
K, exp( 4 Ts+l+K(l+sD)exp(s)
Ts + 1
K , exP( s )
Transfer function of a closed system
1
I
+
+ I i ( t ) + Kx(t  1)

1)
+ K x ( t  1)
+ +rqt) + K x ( t  1) = K P 1 j ( t  1)
TlX(t) + KIDX(t KIl(t  1)
+
TIX(t) l i ( t ) K l i ( t  1) = K,Ij(t  I )
= K,Ij(t  1 )
TlX(t)
T l ( t )+ K D i ( t  1) ~ ( t ) + Kx(t  1 ) = K,y(t  1 )
K,y(t  1)
+ x ( t ) + Kx(t  1)
=
Ti(t)
Equation of a closed system
1
I Neutral
Retarded
Retarded
Neutral
Retarded
Type of equation
4
W
38
1. Theoretical Foundations of Functional Differential Equations
effects of little delays. For example, if one allows a little delay h in unit K 2 , then the predictor is described by the equations
40 = K,Cu(t)
+ 401,
4 t ) = K , v ( t  h)
+~
(t 1)
This means that ~ ( t=) K , [ u ( t )
+ u(t
+ K:[u(t


h)
+ u(t

2h)l
1) + 2u(t  1  h)]
+ .. .
Thus in this case the output is defined only by values of the input at preceding moments. Another example of this type is absolutely invariant systems studied intensively by Schipanov et al. [196]. An absolutely invariant system loses its remarkable properties if one takes little delays into account.
6.3. Analysis of SingleLoop Feedback Systems The knowledge of the transfer function permits us, with the aid of the inverse Laplace transform, to define the output of a feedback system. For example, the response x(t) on the unit step input l(t) for a closed system with W,,(s) = K , exp(  s)/[ Ts K exp(  s)] is equal to
+
Even in this simple case, the analytical expression is sufficiently complicated and is difficult to analyse. Graphical representation of transient processes is more convenient. Figure 1.19 represents transient processes in closed systems for different types of controllers and different time constants T. Notice that the steps of output in Fig. 1.19 may be explained by the presence of a 6 function in corresponding NFDEs. The study of transient processes permits us to make some general conclusions. For all types of controllers the increase of delay z (or the decrease of time constant T ) leads to an increase of overshoot and of oscillating properties. The P controller may be used only for plants with a small delay, the properties of which are similar to those of plants without delay (see also 94). The use of PD controllers leads to a decrease of overshoot and of oscillating properties, but there is a position error. For I or PI controllers there is no position error, but the response time is sufficiently great. The best and most universal controllers are PID controllers. They result in a considerable decrease in overshoot and response time, but their adjustments are the most difficult [68, 1881. Let us now formulate some general rules for choosing a controller. P controllers give response times on the order of 5r and a position error of less
$6. Feedback Systems with Aftereffects
39
20
.
*(
16 '2
08 0 666 04
0
4
8
16
12
20
24
28
32
36
20
24
28
32
36
f
0
4
12
8
16 f
PI PID c
< 0666 0
20
40
60
80
100
t Fig. 1.19. Transient processes in systems with time lag. (Steps ac.)
than 10% only for plants with minor delay 2 < 0.05T. PI controllers d o not give position errors, but their response times are more than 6 2 . The use of PID controllers permits us to decrease the response time to a time on the order of 42.
6.4. TwoLoop Systems with Delay (Smith, Reswick and Other Controllers) Twoloop systems have good performances in comparison with singleloop ones. The cascade circuit shown in Fig. 1.20 can be used when the plant consists of two parts; the first one with transfer function H,(s) has no delay, and the second one has transfer function H 2 ( s )exp( s7). The introduction of local feedback from the input of the second part permits us to compensate for disturbance fl,but not for disturbance f2. Thus the cascade circuit has a restrictive region of applications [12,683.
40
1. Theoretical Foundations of Functional Differential Equations
4
I
Fig. 1.20. Cascade circuit for a system with time lag.
The twoloop system with delay compensation shown in Fig. 1.21 is also considered [134( 1), 134(2), 161). Delay Compensation can be used when the plant has not time lag, but the controller has a certain delay. Let us assume that a closed system containing only plant H ( s ) and controller W,(s)has good performance. In the presence of unit R(s) exp(  sz), system performance grows worse, and some correction is necessary. For delay compensation, one introduces a second feedback containing an amplifier K , and plant model H,(s) (see also Loginov et al. [131]).
Controller
R(s)esT
q 5 )
Plant H(5)
Km
Fig. 1.21. System with delay compensation.
x f
$6. Feedback Systems with Aftereffects
41
Fig. 1.22. Reswick controller for a system with delay.
A few computations show that the transfer function of a closed system equals @(s) = W,(s)H(s)R(s)exp( sz)U'(s)
+ K , W,(s)H(s)  W,(s)R(s)exp( sz)H,(s) + W,(s)R(s)exp(  sz)H(s) = H ( s ) we have U ( s ) = 1 + K,W,(s)H(s). Thus the U(s) = 1
For H,(s) unit R(s) exp( sz) does not exert influence on the performance of the closed system. In the case in which it is impossible to separate the delay from the rest of the plant, Smith or Reswick controllers are usually used. The basic idea of these controllers consists of the construction of a plant model which permits us to obtain plant output without delay. Consider the system containing the Reswick controller, shown in Fig. 1.22 [68,184]:
4s) = As>  4 s X u(s) = u(s)
u 1 (s) = 4S)/Hm(s)
+ Pu(s) exp(  sz)
Eu(s> + f(s)lWs) exp(sr) = x(s)
Excluding from this system variables E(s), u l ( s ) and u(s), we get X(S)
= H ( s ) ~ x P (  ~ T ) { Y+( ~Hm(s)C1 )  P e x ~ (  st) I f ( s) l/u , ( s)
U , ( s ) = H,(s)[l
 fi exp( ST)]
For H,(s) = H ( s ) and p
=
+ H ( s ) exp( st)
1 it follows that
x(s) = y(s) exp(  S T )
+ H(s)[exp( ss)
 exp( 2st)]f(s)
(6.3)
Thus the delay is excluded from the denominator of the transfer function. The action of delay is reduced to the time shift of the input y ( t ) and to the time shift 22 of the transient processes caused by the disturbance f ( t ) . For
42
1. Theoretical Foundations of Functional Differential Equations
Fig. 1.23. Smith controller for a system with delay.
H ( s ) = K , the response time is equal to 22 and is minimal. In fact, the disturbance can reach the controller input only by passing the plant, i.e., with delay z. After a time z the controller output signal reaches the plant output and can liquidate the error. Practical realization of the Reswick controller is difficult because the simulation of unit l/H(s) demands the exact knowledge of plant parameters. Also, it is necessary to take p 2 1, but for p > 1 the closed system becomes unstable. Let us now analyse the Smith controller [214(1), 214(2)]. A structural scheme (Fig. 1.23) yields E ( S ) = y(s) 
x(s),
u(s) = Wp(s)&l(s),
cl(s) = E(S) u 1 ( ~ )=
x(s) = H ( s ) exp(  st)Cu(s)

ul(s)[l

exp(sz)] (6.4)
Hrn(s)u(s)
+ f(4l
Excluding from (6.4) the auxiliary variables ~(s), E ~ ( s ) ,ul(s) and u(s), we get
4 s ) = CW,(W(s)y(s)
exp( sz)I/U,(s>
+ { H ( s )e x ~ (  s z ) + H(s)Wp(s)Hrn(s)Cexp(2sz)If(s)}/U,(s) U,(s) = 1
+ W,(s)H(s) exp(sz) + W,(s)H,(s)[l exp( st)]
Suppose that H,(s) = H(s). Then I W,(jo)H(jw)l 9 1, we obtain x(s) E y(s) exp(  sz)
in
the
frequency
+ H(s)[exp( sz)  exp(

strip where
2sz)]f(s)
(6.5)
Relation (6.5) is just the same as (6.3). Thus, the Smith controller has the same performance as Reswick's. Systems containing Smith controllers have been studied elsewhere [68,161,190,215]. The influence of parameter errors
$6. Feedback Systems with Aftereffects
L
_________
43
_____
J
Fig. 1.24. Twoloop system with auxiliary controller.
between a plant and its model has also been investigated [94,190,215], and the results showed that the Smith controller provides good performances if the time constant error is less than 20 % and the delay error is less than 10 %. If disturbances act after the delay unit, then their action can be decreased by introducing into the controller a differentiator with transfer function W,(s) = Ts/(Ts + 1). The main purpose of the controller proposed by Kabal'nov et al. [94] and shown in Fig. 1.24 is to limit the influence of parameter errors. Introduction of an auxiliary controller R,(s) permits us to conserve good performances for parameter errors on the order of 15 %20 %. If parameter errors are greater, then one can use the adaptive controller described in Chapter 2 $9.
Chapter 2
Stability of Retarded Equations
61. GENERAL DEFINITIONS OF STABILITY 1.1. Definition of Liapunov Stability Denote M a metric space of continuous functions cp(e), cp: (  00, 01 + R, with a metric p and Q H a sphere in M (see Chapter 1, 93). Consider the initialvalue problem i ( t ) = f(t,
x,),
x,(e)= x(t
xtO(e>= ( ~ ( 0do)E M ,
+ el, f:R1
 00
x
QH
+
0 such that all solutions x ( t ) ofproblem (1.1) satisfying the condition p(xto,zto) 5 6 are dejined on [to,co) and also Ix(t)  z(t)J 2 &fort 2 to. Liapunov stability means that an arbitrarily narrow E neighborhood of the solution x(t) contains all those solutions of problem (1.1) which are sufficiently close to zto at initial moment to. It should be noted that one must study the stability of a given solution z ( t ) of problem (1.1) but not the stability of Eq. (1.1). 1.1 [ 3 6 ] . Consider the equation EXAMPLE
+ (x4(t) + 4~~(t))~’~]~(t) (1.2) is x(t) = c sin(xt + d), where c and
i ( t ) = $[x2(t)
( 121
The general solution of d are constants. It is clear that the trivial solution z ( t ) = 0 of Eq. (1.2) is stable, but any other solution will be unstable. Actually, for the solutions xl(t)= c1
44
$1. General Definitions of Stability
45
+
sin(c,t d , ) and x 2 ( t ) = c2 sin(c2t irrational, we have
+ d2), where c1 # 0, c2 # 0 and cJcl

limIx,(t)

X2(t)l = I C l l
is
+ Ic2I
t m
So the solutions x l ( t ) and x 2 ( t )diverge on a finite value as t + cc and this is independent from the proximity of (x,(to), il(to)) to (x2(t0),a2(to)). REMARK 1.1. For linear equations i ( t ) = L(t, x,) + g(t), L(t, 0 ) = 0 (where L(t, cp) is a linear map) from the stability of a solution for some g(t) follows the stability of any solution for arbitrary gl(t). Hence it is possible to call the linear equation stable or unstable. The stability investigation of any solution z ( t ) may be reduced to the study of stability of the trivial solution. For this it is sufficient to introduce the function y ( t ) = x ( t )  z(t). Then i ( t ) = f(4 Y , + z t )  f(t>4 = g(t, Y A
d t , 0) = 0
(1.3) Below we assume that f ( t , 0) = 0 and investigate only the stability of the trivial solution of (1.1). Definition 1.2. The trivial solution x ( t ) = 0 of&. ( 1 . 1 ) is called stable iffor any E > 0 there exists a 6 = a(&, t o ) such that Ix(t, t o , cp)l I E , t 2 to for any initial function cp(0) E Qd. Otherwise the trivial solution is called unstable. Let us make some remarks about the dependence of stability on the choice of initial moment t o . Assume that sufficient conditions of existence, uniqueness and continuous dependence of solutions on initial data are fulfilled on [s, to],s 5 t o . Then, from the trivial solution stability related to disturbances of the initial function at the moment t o , follows stability related to disturbances in any previous moment t , E [s, to].This statement is true by virtue of continuous dependence of the solution on [s, to] from the initial data. At the same time the trivial solution stability related to disturbances at to may be unstable related to disturbances at the moment t , > t o . For ordinary differential equations there is no such effect because all solutions are leftcontinuable.
EXAMPLE 1.2 [253( l ) ] . Consider the equation a(t) = b(t)x(t  3x/2)
(1.4)
Here the continuous function b(t) = 0 for t 5 3x12; b(t) = cos t for 3x12 5 t I 371; and b(t) = 1 for t 2 371. Take to = 0 and let cp be any continuous function: cp E C(3x/2,0]. It is easy to see that the solution x(t, 0, cp) equals x(t, 0, cp) = cp(0) for 0 I t I 3x12 and x(t, 0, cp) =  cp(0) sin t for t 2 3x12. So Ix(t, 0, cp)l I I cp(0)I for t 2 0.
46
2. Stability of Retarded Equations
Hence the trivial solution is stable related to disturbances at the moment to = 0. Now consider another initial moment t, 2 371 and define the initial function cpl(t, + 8) = 6 exp[I(t, 8)] for  3 ~ / 2I 8I 0. Here A is a positive root of the equation y = exp(  37cy/2). We get that x(t, t,, cp,) = 6 exp[l(t  t,)] for t 2 t , 3n/2 and also / x ( t ,t,, ql)l + co as t + co, 6 # 0. Thus the trivial solution is unstable related to the disturbances at any moment't, 2 37c.
+
1.2. Uniform Stability Definition 1.3. The trivial solution of (1.1) is uniformly stable i f for any E E (0,H I there exists a 6 ( ~> ) 0 independentfrom to, cp such that Ix(t, to, rp)l< E for t 2 to and q(8) E Qa, to E R , .
Uniform stability is stronger than the stability in the sense of Definition 1.2. Corresponding examples for nonautonomous equations are available [57(2), 108(5)]. However, for autonomous ( f (t, x r ) = f(x,)) or operiodic ( f ( t + w, x,) = f ( t , x,)) Eq. ( 1 . 1 ) the uniform stability follows from the stability of the trivial solution. For any solution x(t, to, cp) of operiodic Eq. (1.1) and any integer m we have 4 4 tl, X f l ( 4 t o , cp)) x(t
* mo, * mwcp) to
= x(t, t o , cp),
t, > t o
= x(t, to, cp)
Hence for the proof of the uniform stability it is sufficient to investigate only the moments to E [0, a]. Take any E E (0, H ) and define V(E, o)> 0 such that Ix(t,w, cp)l < E, t 2 w , llcpil I~ ( E , w )The . existence of q follows from the stability for the initial moment to = o.Choose a 6 ( ~from ) Definition 1.3 such that max
max Ix(w,
O5foSw
to, cp)l
5 V ( E , 0)
l l ~ l) 0 such that Ix(t, to, cp)J = I x(t, 0,xu(@,to, q>)I I E for to E CO, 01 and I1q II Ih(4.
1.3. Asymptotic Stability Definition 1.4. The trivial solution x(t) = 0 of (1.1) is called asymptotically stable if (1) it is stable (2) for any to there exists a A = A(t,) > 0 such that
47
51. General Definitions of Stability
The set Q(t,) G M ofall initialfunctions q such that lim x(t, to, q) = 0, t + co, is called the attraction domain of the trivial solution at initial moment to. Consider an example showing that in Definition 1.4 the first condition does not follow, in general, from the second one.
1.3. Given the equations EXAMPLE i ( t )= tlx(t) = t'x(t)
+ 2 )  3 ~ 3 ( (+t 2)/3)  27t(t + 2 )  3 ~ 3 (t T(t)) 
27t(t
t 2 to
=
1,
T(t) = (2t  2)/3
(1.6)
For t 2 1 we have z(t) 2 0. So (1.6) is the RFDE for t 2 1 with initial data x(1) = x,. The solution x(t, l , x , ) = x,t exp(xi(t  l ) ) , t 2 1, satisfies condition (1.5). But for Jx,J= b and t = 1 + X 2 we will have Ix(1 + b  2 , 1 , xo)l = 6(1 + b  2 ) exp(b2(1 + X 2 ) )= (6 + 6')/e 2 2/e. Hence, the first condition of Definition 1.4 is not fulfilled. The trivial solution of (1.6) is not stable.
Definition 1.5. The trivial solution of (1.1) is called uniformly asymprotically stable if it is uni$ormly stable and for any y > 0 there exist T(y) > 0 and H , E (0,H ) such that Ix(t, t o , q)l I yfor t 2 to + T(y)and cp(@ E Q H I ,t , E R , . In addition, the sphere Q H , c M lies in the domain of attraction of the trivial solution. Uniform asymptotic stability is stronger than asymptotic stability. But if Eq. ( 1 . 1 ) is autonomous or o periodic and its trivial solution is asymptotically stable, then it will be uniformly asymptotic stable.
1.4, Other Stability Definitions. Stability Under Delay Disturbances Besides the previously mentioned stability definitions one sometimes uses the definitions of global stability, stability under steady acting disturbances, exponential stability, partial stability, stability on$nite time interval, etc. For a detailed discussion of the different stability definitions see Kolmanovskii [108(5)], Malkin [138] and Yoshizawa [249]. However, let us remark that for a linear RFDE, exponential stability follows from uniform asymptotic stability. A specific problem for differentialdifference equations is the stability under delay disturbances [129,212,223(2)]. For linear equations, some involving results are given in El'sgol'tz [58(2)] and Kolmanovskii and Nosov [l08(5)].If the trivial solution of a nonlinear RFDE is uniformly asymptotically stable for a given delay, then it is stable for any delays whose distances
48
2. Stability of Retarded Equations
from the given delay are sufficiently small [I 14(3)114(5)]. In general, this statement is not valid .for NFDEs. The following problem is interesting for equations with variable delay. Assume that a linear system is asymptotically stable for every constant delay from a given interval. Will this system be stable for variable delay from the same interval? Consider an example showing that it is not true in general. EXAMPLE 1.4 [156(23)]. i(t) =
Investigate the scalar equation (1.7)
x(0) = xg, t 2 0
x(t  r(t)),
Take a number a E ($,7c/2). Let the delay ~ ( tin) (1.7) be constant: ~ ( t = ) T = const. Then for 0 5 z 5 a the trivial solution of (1.7) is asymptotically stable (see Fig. 2.1). Define the variable delay S ( t ) E [0, a] ?(t)=
i
O 1. Hence with T = ?(t) that x(k(a Ix(k(a 1))l + co for k + co. So the trivial solution of (1.7) with ?(t) is unstable. Notice that the trivial solution of the scalar equation i ( t ) = a(t)x(t  ~ ( t ) ) where , a(t) > 0 and ~ ( t2) 0, is asymptotically stable if (SUP,,40) (sup,~(t))< 3 [156(23)1.
+ + +
+
$2. LINEAR AUTONOMOUS RETARDED EQUATIONS 2.1. Statement of the Problem Stability of linear autonomous RFDEs has been studied by many authors [16,18,59,252(1)] with the aid of a series expansion of solutions. But usually the arrangement of the characteristic quasipolynomial roots is investigated in detail only for differentialdifference equations. For some results about convergence of expansion in series of basic solutions see 41.5. Krasovskii [114(5)] considered an RFDE with finite delay using semigroup theory. In this paragraph, on the basis of Laplace transform, stability conditions for a linear autonomous RFDE with arbitrary delay are ascertained. Consider the RFDE [dK(s)]x(t
i(t) =

s),
x(~)E R,,
t 2
0
(2.1)
WO!
40)
=
da
 o O < t l I O
(2.2)
49
$2.. Linear Autonomous Retarded Equations
Assume that all elements K , of the matrix K ( s ) are functions with bounded variations on [0, a)and the integral in (2.1) is a RiemannStiltjes one. Moreover
om !
sIdKiJ(s)I < co,
i,j
r
=
1 , . .. , n
(2.3)
The initial function cp is continuous and is such that SUP
IWI < a,
F(t) =
f 2 O
CdK(s)lcp(t  s)
(2.4)
Introduce the norm 11 cp 1) in the space of initial functions satisfying (2.4) and llcpllo =
Icp(0)l +
(jomlroli
J;Ir(r)l'
dt)".,
dt
0 and continuous for Re z 2 0 since function (2.10) does not vanish and R(z) by virtue of (2.3) is bounded for Re z 2 0. Hence, using the Cauchy theorem [123], (2.11)
42. Linear Autonomous Retarded Equations
51
By virtue of Plancherel theorem (1861 it is sufficient for L, stability that
jm
IX(iB)I2 dB
0 r m
and also let the characteristic function (2.10) have no zeros in the halfplane R e z 2 y. Then I x ( t , cp)l IC , exp( yt).
Proof: Consider the function y ( t ) = x ( t ) exp(yt). Substituting y ( t ) in (2.1) we obtain the characteristic function AJz) = det[(z  y ) I  R(z  y ) ] of the e q u 6 n for y(t). This and the assumptions of Thorem 2.3 imply that Ay(z) has no zeros in the halfplane Re z 2 0. So by virtue of Theorem 2.2 we get limt+my ( t ) = 0. Theorem 2.3 is proved.
$3. METHODS OF STABILITY INVESTIGATION OF LINEAR AUTONOMOUS RFDEs By virtue of Theorem 2.2 the stability of RFDE (2.1) depends on the presence of zeros of the characteristic functions (2.1 1) in the halfplane Re z 2 0. We shall call “pzeros” the zeros with positive real parts. For ordinary differential equations, function (2.11) is reduced to a polynomial. Conditions for absence of pzeros for polynomials are given by the wellknown RouthHurwitz theorem. For general characteristic function (2.11) such complete results are not obtained.
3.1. Analytical Methods (Pontriagin’s and Chebotarev’s Theorems) Consider the quasipolynomial
c m
D(z) =
r
aljz exp(bjz)
I=O j = 1
Quasipolynomials (3.1) are characteristic functions for differentialdifference equations. The RouthHurwitz problem for quasipolynomials (3.1) may be studied by different methods. Let us examine some of them. Quasipolynomial with commensurable delays may be presented as
54
2. Stability of Retarded Equations
Consider the quasipolynomial D,(z) for z D,(iw) = g(w) + 80).
= io,
where w is a real number:
Pontriagin’s Theorem [175]. I f quasipolynomial (3.2) has no pzeros then all the zeros of the functions g(w) and f (0) are real, simple and alternating and g(w) f (0)f(w)g(w)> 0,
 cg
< 0 < co
(3.3)
For absence of pzeros of D,(z) one of the following conditions is suflcient: (1) All the zeros of the functions g(w) and f ( o ) are real, simple and alternating and inequality (3.3) isfuljilled for at least one real w ; ( 2 ) all the zeros of thefunction g(w) [or f (w)] are real and simple and for each zero relation (3.3) is satisjied.
The simple proof of Pontriagin’s theorem is given by Postnikov [177]. Consider, as example the equation, i ( t ) = bx(t

h),
b >0
(3.4)
The quasipolynomial D(z) = z exp(hz)  b = 0 corresponds to Eq. (3.4). By setting z = iw we get D(iw) = o sin wh  b
+ iw cos o h = g ( o ) + i f ( w )
g(w) = w sin o h  b,
f ( w ) = o cos o h .
Zeros of the function f ( w ) = o cos wh are equal to w o = 0, o k + , = + n/2)/h;k = 0, 1,. . . . These zeros are real and simple. Relation (3.3) for = wk has the form g(0)f (0)  j’(o)g(o) = b > 0, g(wk+i > f(%+ 1) f(o,+ l)g(o, + 1)= wk+ ,h(wk+  b ) > 0. The last inequality will be fulfilled under the conditions (kz
b < W 1 < ... < wk+l
So the trivial solution of Eq. (3.4) is asymptotically stable for 2bh < n. By analogy we may prove that the equation i ( t ) = ax(t) + bx(t  h ) is asymptotically stable if a < h’, a <  b < (a2 + p2h2)1/2,where p is the root of equation p = ah’tgp such that 0 < p < n. If a = 0 we set p = 4 2 . One of the ways of studying quasipolynomials (3.1) with incommensurable delays is generalization of the RouthHurwitz conditions. Expand the quasipolynomial D(z) in the series D(z) = a, + a,z + a 2 z 2 + . . . and define the functions u(z) and u(z): D(iz) = u(z) + iu(z), u(2) =
u(z) = a.
a,z  a3z3 + a5z5  .

a 2 z 2 + a4z4
+.
$3. Methods of Stability Investigation of Linear Autonomous RFDEs
55
Let us introduce the determinants Qm:
Chebotarev’s Theorem [37]. Assume that thefunctions u(z) and u(z) have no common zeros. Then quasipolynomial (3.1) has no pzeros ifand only i f Qm > 0,
1,2,.
..
(3.5) Applications of this theorem are not effective because an infinite number of inequalities (3.5) must be verified. 3.2. Method of D Subdivision. Vyshnegradskii Diagrams For determination of conditions under which quasipolynomial (3.1) has no pzeros, the method of D subdivision is widely applied [37,59, 160(1), 160(2)]. The method of D subdivision is based on the fact that zeros of quasipolynomial (3.1) are continuous functions of parameters a,j and b j . Construct the subdivision of the coefficient’s space by hypersurfaces, the points of which are quasipolynomials with at least one imaginary root. Such a procedure is called a D subdivision. For continuous variation of quasipolynomial parameters the number of pzeros may change only by passage of some zeros through an imaginary axis. Therefore points of every domain of D subdivision correspond to quasipolynomials with the same number of pzeros counted according to their multiplicities. Define now the number of pzeros at least in one point of every domain of D subdivision of the corresponding quasipolynomial. Then for all points of every domain the number of quasipolynomial pzeros will be just the same. D subdivision has a visual form when the quasipolynomial depends on two parameters. Substitute z = io in characteristic equation (3.1), linearly depending on two parameters a and b. Equating to zero the real and imaginary parts, we have
+ bQ,(o) + R,(o) = 0 q w ) = aP,(w) + b Q 2 ( 0 ) + R 2 ( 0 )= 0
U ( w )= a P , ( w )
(3.6)
where P l ( w ) ,. . . , R2(w)are some continuous functions. From Eqs. (3.6) we obtain
a = AJA, b = AJA, A = P,Q,  P2Q, (3.7) The formulae (3.7) determine the point a , = a(@,), b = b(wl) of D curve for w = wl. If o varies from a to 00 we obtain some curves. Besides these
56
2. Stability of Retarded Equations
Fig.2.1. D subdivision for quasipolynomial (3.9).
Fig. 2.2. Stability domain for the equation i ( t )
+ ax(t  1 ) + bx(t  1) = 0.
$3. Methods of Stability Investigation of Linear Autonomous RFDEs
57
curves, D subdivision contains singular straight lines such that either A = 0, A, = 0 or A = 0, A2 = 0 or A = 00, A, = 00. Consider for example, D subdivision of the equation i(t)
+ U X ( ~ +) bx(t  h) = 0
(3.8)
Characteristic quasipolynomial of Eq. (3.8)
D(z) = z
+ a + b exp(  z h )
(3.9)
+
has the root z = 0 for a b = 0. Quasipolynomial (3.9) has an imaginary b cos wh = 0 and y  b sin o h = 0. So the root z = iw if and only if a i boundaries of D subdivision are singular straight line a b = 0 and parametric curves a = (a cos wh)/(sin oh),b = o/(sin oh), 00 < o < 00. D subdivision for quasipolynomial (3.9) is represented in Fig. 2.1. Quasipolynomial (3.9) has no pzeros for a > 0 and b = 0. Hence Eq. (3.8) is asymptotically stable in the domain shaded in Fig. 2.1. The number of pzeros in other domains of D subdivision is indicated in Fig. 2.1. The stability domain for the equation % ( t )+ a i ( t  1) bx(t  1) = 0 is represented in Fig. 2.2. D subdivision for the equation x ( t ) a i ( t  1) bx(t) = 0 is given
+
+ +
+
b
\
P=2
P=2
P.2 P=l
Fig. 2.3. D subdivision for the equation , t ( t ) + nx(t

1)
+ bx(r) = 0.
58
2. Stability of Retarded Equations
in Fig. 2.3. The stability domain of this equation consists of an infinite number of triangles adjoining the axis a = 0. The stability domain for the equation x ( t ) a i ( t ) bx(t  1) = 0 is represented in Fig. 2.4.
+
+
Fig. 2.4. Stability domain for the equation x(t) + ax(t) + bx(t  1)
= 0.
The formulated results enable us to find the stability domains of the simplest feedback systems from Table 1.2. In the theory of automatic control, these stability domains are often called Vyshnegradskii diagrams [219(2), 241(2)]. Consider for example the system with P controller i(t)
+ T  'x(t) + K T  l x ( t

t) =
K T  l u ( t  t)
0'JI
Fig. 2.5. Vyshnegradskii diagram for system (3.10) in parameters a = T  ' , b = K T  '
(3.10)
59
$3. Methods of Stability Investigation of Linear Autonomous RFDEs
100 
50

20 10

52
Q 0 5 1 2 71
001002 0 0 5 0 1 0 2
Fig. 2.6. Vyshnegradskii diagram for system (3.10) in parameters TI = T', K.
10.5 
0.20.1 0.050.02
0 01
L
5
10
20
50
100
Tl
Fig. 2.7. Vyshnegradskii diagram for a system closed by I controller.
The Vyshnegradskii diagram of system (3.10) for a = T  ' and b = KT' is given in Fig. 2.1. But physically sensible feedback systems are possible only for K > 0 and T > 0. Consequently, the system (3.10) stability domain is less than in Fig. 2.1. The Vyshnegradskii diagram of system (3.10) in parameters a = T  ' and b = KT' is represented in Fig. 2.5. Logarithmic scales are used in Fig. 2.6,2.7 and 2.8. The Vyshnegradskii diagram for system (1.6) with an I controller is represented in Fig. 2.7 and with a PI controller in Fig. 2.8. For other diagrams for singlecircuit feedback systems see Gorecki [ 6 8 ] .
60
2. Stability of Retarded Equations
Fig. 2.8. Vyshnegradskii diagram for a system closed by a PI controller.
3.3. Michailov, Nyquist and Integral Frequency Criteria Frequency methods are used in the theory of automatic control for investigation of the stability of autonomous systems. For RFDEs, frequency methods were developed by Kabakov [93] and Tzypkin [232( 1)232(3)]. These methods are based on the argument principle from the complex analysis [123]. The Michailov and Nyquist criteria are the most frequently used. Consider the characteristic functions (2.11) A(z) = det[lr  ~ o w e  z s d K ( s )=] Z"
+ fi(z)z"' + ... + fn(z)
(3.11)
Here all the functions & 0. Substituting z = io in (3.1 l), we get
+
A ( i o ) = V ( w ) iV(w)
(3.12)
The hodograph of function (3.12) in the complex plane is called a Michailov one.
Michailov Criterion. For asymptotic stability of linear nth order equation with characteristicfunction (3.1l), it is necessary and su.cient that variation of the function A(iw) will be equal to nx/2 when w varies from 0 to co; i.e., arg A(ico)l 0
= nx/2
$3. Methods of Stability Investigation of Linear Autonomous RFDEs
61
The Michailov criterion for RFDEs was proved by Kabakov [93] and Tzypkin [232( l)]. But for RFDEs, Michailov hodographs are more complicated than for ordinary differential equations. Typical Michailov hodographs for stable systems without delays are represented in Fig. 2.9. Michailov hodographs for D(z) = 2z2
+ 0 . 5 ~+ 2 + exp( +zz)
(3.13)
are shown in Fig. 2.10 for z = 1, 5, 10. System (3.13) is unstable for z = 1, 10 and it is stable if T = 5.
I Fig.2.9. Typical Michailov hodographs for nthorder systems without delay.
Let the transfer function of a linear RFDE be equal to W(s).The function W ( i o )is called afrequency response of this system. Nyquist Criterion. Let the openloop system be stable. Then for asymptotic stability of the closedloop system it is necessary and sufJicient that frequency response of the openloop system does not envelope the point ( 1) (see Fig. 2.1 1).
The use of Michailov and Nyquist criteria for RFDEs is difficult because of the complexity of the corresponding curves.
62
2. Stability of Retarded Equations
I
r=10
Fig.2.10. Michailov hodographs for quasipolynomials (3.13).
Fig. 2.11. Nyquist criterion: I, stable system; 11, unstable system.
Integral criteria are more effective. Define the function of logarithmic derivative R(o)for a system with characteristic function A(z) R(w) = Re 
A’(z) A(z)
~
=
d [arg F(iw)] dw ~
U ( w ) V ( w ) V ( w ) U ( w ) U2(w) V 2 ( w )
+
(3.14)
$3. Methods of Stability Investigation of Linear Autonomous RFDEs
63
where A ( i o ) = U(o)+ i V ( o ) . From the argument principle the integral criterion of stability follows. Integral Criterion of Stability [ 1441. For asymptotic stability of RFDE with characteristic function A(z), it is necessary and suficient that
Iom nn 2
R(o)d o = 
(3.15)
Here R(o)is dejned by formula (3.14) The integral stability criterion is more convenient for computer calculations. Choose a number s such that
Ism
R(o) d o < 1
Then instead of condition (3.15) it is sufficient to verify the inequality I s m R ( w )d o >
( n  1)n 2
~
Consider, for example, the system with characteristic function
+
A(z) = 0 . 1 + ~ 0.32 ~ 0.5 + (0.12 + 0.2) exp( zzl) + (0.22 + 0.3) exp(  z t 2 ) We get
+ 0.5 + 0.10 sin zlo + 0.2 cos zlw + 0.20 sin + 0.3 cos z 2 0 V =0 . 3 ~ + 0.10 cos zlw  0.2 sin zlw + 0.20 cos z t o
U
=
0.10’
t20
 0.3 sin z2w
The function R(w) from (3.14) is represented in Fig. 2.12 for z1 = 3.0, z2 = 1.5 and in Fig. 2.13 for t l = 2.5, z2 = 1.7. Computing the integrals we have
[
20
R(w)d o = 3.0455;
z1 = 3.0, z2 = 1.5
At the first case the system is stable and it is unstable at the second one. Remark that studying the function R(o) we obtain some estimations of the transient regimes (see Melkumjan [144] and 44). The wave criterion for RFDEs is developed in Lekus and Rovinskii [124].
64
2. Stability of Retarded Equations
0 5
1 5 _  _ _ _ _ _ _ _ _ _ _ _   _  _ _               
Fig. 2.13. Logarithmic derivative R(w) of unstable system.
3.4. Stability for Arbitrary Delays. Tsypkin Criterion In some problems it is interesting to have the conditions under which an RFDE is stable for arbitrary delays. Consider an openloop system with transfer function W ( s ) = [ R ,  l(s)/Qn(s)] exp( ST), where R, l(s) and Qn(s)
65
$3. Methods of Stability Investigation of Linear Autonomous RFDEs
are polynomials of degrees ( n  1) and n. The closedloop system (Fig. 2.14) has the following transfer function Rn i(s) exp(sz)/CRn i(s) e x ~ (  s z ) + QAs)l
(3.17)
I
I
Fig.2.14. System with separated delay unit.
Tsypkin Criterion [232(2)]. Let Qn(s) be a stable polynomial. Then the closedloop system (3.17) is stable for arbitrary delays 0 T < co ifund only if IQ,,(io)l > l R n  l ( i o ) l ,
co < o < co
(3.18)
For the proof let us remark that the frequency response of system (3.17) may be constructed in two steps. First, one constructs the curve R, l(io)/Q,,(io). By virtue of (3.18) this curve is situated entirely within the unit circle. But the action of the factor exp(  izw) is reduced to the rotation of the vector So the frequency response curve of R,l(iw)/Qn(iw) on the angle (m). system (3.16) is situated entirely within the unit circle also and cannot envelope the point ( 1). By the Nyquist criterion, the closedloop system (3.17) is stable. r ) ~some o,,. Then it is always possible Now let lQn(io)cr)l= ~ R n  l ( i o cfor to choose the rotation angle (zcrocr) such that the frequency response passes through the point (1). In addition, system (3.17) is stable for and unstable for z, < z. 0 0, then by virtue of Remark 3.1 condition (3.21) will be necessary. (c) Let the function kl(s) have a jump a, < 0 in the point s = 0 and
Then the function A2(z) has no pzeros. (d) Assume that Boo < 0, ctl1 + ~ ~ < 1~ and / 2 ( 2/?00)"2 > 0. Then the function A2(z) has no pzeros.
Blo Pol
>
For the proofs of statements (a)(d) see Kolmanovskii [108(5)]. As an example consider the equation X(t) = ax(t)
+ bx(t  h),
t20
(3.22)
Because of (d), the trivial solution of Eq. (3.22) is asymptotically stable if 2b))'I2. This example shows that delay may stabilize the system. The system which is unstable for b = 0 becomes asymptotically stable for some b > 0.
2a > 2b > a > 0 and 0 < h < (b/(a
+
$4. STABILIZATION SYSTEM OF THE
TWOREFLECTOR ANTENNA The tworeflector (or Cassegrainian) antenna is used to form a narrow directional source of radiofrequency radiation. The tworeflector antenna consists of a big immovable reflector and of a little movable one, furnished with a control system. The big reflector is a hemispherical metal surface approximately 50 m in diameter, with the axis forming a certain angle with the horizon. The little reflector has a specially calculated complex surface with a radiator situated in its focus. The radiator is placed on a massive platform having a diameter of 7 m. To reach high guidance accuracy the control system includes a platform stabilization system, the precision of which is about 5" to 6". The antenna has to function outdoors continuously
68
2. Stability of Retarded Equations
by day and night under considerable wind and temperature disturbances. Platform deflections are measured by optical pickups. The actuator of the stabilization system contains three screws which can move three points of the platform in the vertical direction. Hence the stabilization system is multidimensional, more exactly threedimensional, including three separate channels for each screw servomechanism. The position errors in the screw servomechanism are measured by hydrostatic pickups which have considerable lags, until 1 to 2 sec.
Fig.2.15. Separate channel of screw servomechanism.
Fig.2.16. Scheme of multidimensional stabilization system
69
& Stabilization I. System of the TwoReflector Antenna
Fig. 2.17. Matrix structural scheme.
A separate channel of the screw servomechanism is represented in Fig. 2.15. Figure 2.16 shows a general scheme of a multidimensional stabilization system, and Fig. 2.17 gives a matrix structural scheme. The transfermatrix function of the openloop multidimensional system is
G(s) = W(s)CC(s)+ D(4l
(4.1)
From the symmetry of the multidimensional system we have Wll
K, D,,
=
W,,
=
W,,
= K,K,oK,K,, =
D,,
=
+ K2)'
= K,s'(T,s K2 =
1
+ K,oK,K, (4.2)
D , , = K , exp(sz)
Cij = K , exp(sz),
i , j = 1, 2, 3,
i #j
Numerical values for parameters are received experimentally [1441 and are as follows: for the engine T = T, = 0.11 sec; K , = 17.58 rad/V sec; reduction coefficient K , = 9 x mm/rad; pickup gain K , = 0.4 V/mm; interconnection gain K , = 0.05 V/mm; voltage gain K , = 250; power gain K,, = 10; local feedback gain K , = 0.03 V sec/rad. The transfermatrix function of the closedloop system is equal to @(s) = G(s)[I G(s)]  l . The characteristic function of the closedloop multidimensional system is
+
+ G(s)] = P,(s) + Pl(s) exp( sz) + P 2 ( s )exp(  2 sz) + P3(s) exp( 3s.r)
D(s) = det[I

(4.3)
71
§5. Liapunov Direct Method for Equations with Delay
R(wl
w
~v =500 Fig. 2.19. Family of R(w) for different K,.
For Kr ~ 0.02 V/sec rad the system is stable. The number of oscillations N decreases with the increase of Kr. Fig. 2.19 shows a family of graphs R(w) for r = 0.3 sec, Kr = 0.02 V/sec rad and K. = 200, 300, 500. Together with the increase of K. the transient time t, also increases, system stability decreases and finally forK.= 500 the system becomes unstable. For K 1 = 15, K 2 = 6.3, T = 0.11 one can evaluate critical lag ren which is equal to 0.5 sec. If 0 :::;; r :::;; rer• then the system is stable; for r > re., the system is unstable. These results have been confirmed by experiments [144].
§5. LIAPUNOV DIRECT METHOD FOR EQUATIONS WITH DELAY Stability theory of RFDEs is one of the important directions in the investigations of these equations. Stability theory is developed in several directions. First, stability of concrete 'types of equations are studied (for example, linear autonomous equations, linear periodic equations, quasilinear equations and so on). Further, general methods similar to the Liapunov direct method are created. In this section some fundamental ideas and results relating to the Liapunov direct method for RFDEs are presented.
72
2. Stability of Retarded Equations
5.1. Application of Liapunov Functions. RazumikhinType Theorems Consider the initialvalue problem for the RFDE q t ) = f(t,X J , x,,(e)
x,(e) = x(t
t 2 to,
+ e)
h Ie Io
= q(e),
(5.1)
Assume as in $1 that cp(6) E C [  h, 01, the operator f (t, cp), f : R , x QH+R, is continuous and Lipschitzian in cp E Q H and f ( t , 0)
(5.2)
=0
Investigate the stability of the trivial solution x(t) = 0 of Eq. (5.1). This is possible by the method of the Liapunov function [3,65,180(1), 180(2), 1973. Formulate a Razumikhintype theorem.
Theorem 5.1. Let there exist, for Eq. (5.1), the continuous positivedejnite function V(t,x), V : R , x R, +R,, whose derivative computed along the trajectories of Eq. (5.1) is nonpositiue for any solution x ( t ) satisfying the inequality V(s,x(s)) IV(t,x(t)),
s I t, t 2 to
(5.3)
Then the trivial solution of(5.1) is stable. If;in addition thefunction V(t,x) has an infinitesimal upper limit and its derivative V is negativedefinite along any solution x(t) of Eq. (5.1) such that
V(s,4 s ) ) I f ( V ( t , x(t)),
s
I t, t 2 to
(5.4)
where f is a continuous function and also f (u) > u for u > 0. Then the trivial solution of (5.1) is unformly asymptotically stable.
For the proof of Theorem 5.1 see Hale [81(1)], and Razumikhin [180(1), 180(2)].
5.2. Method of LiapunovKrasovskii Functionals for Equations with Bounded Delay Krasovskii has proposed to use for stability investigation the functionals defined on C [  h , O ] . Theorems 5.2 and 5.3 are proved by Krasovskii [114(1)114(5)]. Let K R , x Q H+ R , be some continuous functional such that V(t,0 ) = 0
(5.5)
Denote by o , ( r ) , r 2 0, some scalar, continuous, nondecreasing functions
73
$5. Liapunov Direct Method for Equations with Delay
such that o,(O) = 0 and o i ( r ) > 0 for r > 0. The functional V(t,cp) is positivedefinite if there exists a function o1such that V(t, CP) 2 wi(ldO)l),
~PEQH,
~ E R I
(5.6)
The functional V(t,cp) has an infinitesimal upper limit if there exists a function o2such that U t , cp) 5 f%(lIcp(~)lO
(5.7)
Substitute the trajectory xr(8), h I 8I 0 of (5.1), instead of cp(B), into the functional V(t, cp) and denote V ( t ) by V(t, x J . The derivative $' of the functional V ( t ,x,) computed along the trajectories of (5.1) is equal by definition to
P=
lim sup[V(t
+ At)kft)]/(At)
Ar + O
In other words the derivative $' is the righthand upper derivative of the functional V on the trajectories of (5.1). The derivative Pis called negativedejinite if there exists such a function o3that $'< 03(1x(t)l). Theorem 5.2. Let there exist continuous positivedefinite functional V(t,cp) such that VI 0. Then the trivial solution of Eq. (5.1) is stable. Proof: Assume any that
E E (0, H ) .
By virtue of ( 5 . 5 ) we can take d ( ~ to) , such
From condition $'I 0 it follows that V is nonincreasing along the trajectories of (5.1). Hence using (5.6) and (5.8) we have w,(lx(t)l) 5
v,x,) Iv t o , =
XtJ
t 2 to
V(t0,cp) I W1(E),
The monotonicity of w1 implies that Ix(t)l
E,
t 2 to. Theorem 5.2 is proved.
REMARK. If, in addition to the assumptions of Theorem 5.2, the functional V(t, cp) satisfies (5.7), then the trivial solution is uniformly stable. Theorem 5.3. Let there exist a continuous functional V(t, cp) such that
PI
(5.9) Then the trivial solution of (5.1) is uniformly asymptotically stable. Conversely, ifthe trivial solution of (5.1) is uniformly asymptotically stable then there exists ~l(lcp(0)l) V ( t , cp) I co2(llcp(~)ll),
03(IX(Ql)
74
2. Stability of Retarded Equations
a continuousfunctional V ( t ,cp) satisfying (5.9) and a local Lipschitz condition in cp.
For the proof of Theorem 5.3 see Krasovskii [114(5)] EXAMPLE 5.1. Consider a scalar equation i ( t ) = ax(t)
+ b(t)x(t

a > 0, h > 0, t 2 to
h),
(5.10)
where b(t) is a continuous function. Introduce the functional V ( t ,X J
The derivative
= x2(t)
+ a [hx2(s) ds
i. is such that
f'= 2x(t)[ax(t) I v[x2(t)
+ b(t)x(t
+ x2(t


h)],
h)] + ax2(t)  ax2(t  h )
v
= asup
Ib(t)l
ttto
By Theorem 5.3 the trivial solution of Eq. (5.10) is uniformly asymptotically stable if v > 0.
5.3. Equations with Unbounded Delay Consider an RFDE with unbounded delay (5.11) where the initial function cp(8) is continuous. As in $1 denote M a metric space of continuous functions q(8) cp: (00,01 + R, with a metric p . Denote Q H a ~ , p(q,O) I H ) . Let the map sphere in the space M , i.e., QH= ( c p M f : [0, co) x M + R, satisfy conditions sufficient for existence and uniqueness of the problem (5.1 1) solution. Assume also f ( t 7
0 ) = 0, I f ( t , q)l I F,,
t 2 to,
PEQH
(5.12)
Notice that Theorems 5.2, 5.3 are not directly applied to Eq. (5.11) with unbounded delay. Theorem 5.4. Let there exist a continuous functional V ( t , cp), I/: [0, co) x Q H + R , such that Ol(l
cp(0)I) I v ( t , cp) I 02(P(Cp, 0)) i.II 03(Ix(t)l)
Then the trivial solution of Eq. (5.1 1) is asymptotically stable.
(5.13) (5.14)
75
$5. Liapunov Direct Method for Equations with Delay
Proof: Show first that the solution x ( t ) = 0 of (5.11) is stable. Assume any > 0. Choose 6 > 0 such that w , ( E )= 0,(6). Using (5.13) and (5.14), we obtain for any q E Qa that
E
l . 0 I(
I x ( t )I)
I V ( t ,Xr> I
w o
1
cp)
I o,(P(cp? 0))
I0,(6) = 0 1 ( E )
(5.15)
The monotonicity of o1 implies that Ix(t, t o , cp(0))l s E for cp(0)E Qa. Persuade oneself that ~ ( t o, , q(@) + 0 for t + co and for any q(0) E QHl. Here H I is such that w,(H) = o,(H,). Reasoning as in (5.15) we get that Ix(t, t o ,cp(6))l I H for t 2 0 and cpeQH1.Assume now that the solution x(t, t o , cp(B)) does not tend to zero for some q(8) E QH1.Then for some E > 0 there exists a sequence { t i }such that Ix(ti, t o , cp(6))l 2 E and ti co.By (5.12) F,. Therefore, for all s such the derivative i ( t , t o ,cp(8)) is bounded: ii(t)l I that Is1 I . (5.14) it A = &/(2F0) we have Ix(t, s, t o , cp(6))l 2 ~ / 2 From follows PI  o 3 ( & / 2 )= v < 0 for Is1 2 A. So +
+
V ( t )  V(t0)=
1
Pdt 2
ti+A
1
P(t) dt
= 2 VA N ( t )
fist i i  A
where N ( t ) is a number of t i such that t i It. But N ( t ) * 00 for t + 00. Consequently, V ( t )  V(t,) +  co for t + 00.This contradiction with (5.13) shows that x(t, t o , q(8)) + 0 for t + co and cp E QHl. Theorem 5.4 is proved.
5.4. Stability in the First Approximation and under Steady Acting Perturbations Consider the equation
+
q t ) = ~ ( x , ) ~ ( tx,), ,
x, = x(t
+ e),
h I 8I o
(5.16)
Here L(cp):C[ h, 01 + R, is a continuous linear mapping and R(t, cp): R , x C[  h, 01 + R, is a continuous mapping. Along with (5.16) consider the equation of j r s t approximation q t ) = L(XJ =
L
[dqe)lx(t
+ e),
h Ie I o
(5.17)
where K(8) is a matrix whose elements are functions with bounded variations. The characteristic function of Eq. (5.17) is D(z), D(z) = det[
Iho
[dK(0)l]eze I z ]
(5.18)
Assume that IRfa 9)l
Pllcp(~)ll
(5.19)
76
2. Stability of Retarded Equations
Theorem 5.5. Let zeros zi of function (5.18) satisfy the condition Re ziI  v < 0. Then there exists a constant > 0 such that the trivial solution of (5.16) is uniformly asymptotically stable for any mapping R ( t , cp) complying with (5.19) [114(5)]. Requirement (5.19) in Theorem 5.5 may be weakened until the following: there is a number T > 0 and a continuous function $(s) 2 0 such that
+ jt
t+T
IR(4 cp)l I $(t)llcp(~)II,
*(s)
ds 5 B
For nonautonomous or nonlinear equations of first approximation, complete results are not derived, although there are some theorems in this direction [79(4), 8 1(4)]. In particular, if the trivial solution of the first approximation equation is exponentially stable, then the trivial solution of the full equation is asymptotically stable [108(5), 114(5)]. Let us investigate stability under steadyacting disturbances. Along with Eq. (5.1) consider the disturbed systems 4 t ) = f ( 4 x,)
+ N t , x,)
(5.20)
The trivial solution of (5.1) is called stable under steadyacting disturbances if for any E > 0 there exist 6 > 0 and 'I > 0 such that all solutions x(t, to, cp) satisfy the inequality Ix(t,to,cp)l I E for llcpll I 6 and IR(t,$)I I q, where 11$11 IE. Krasovskii [114(5)] has proved that the trivial solution of (5.1) is stable under steadyacting disturbances if it is uniformly asymptotically stable. It is possible to replace the condition of smallness of the disturbance's norm by the assumption of its smallness in the average [63,128].
5.5. Case of a Nonpositive Derivative Construction of LiapunovKrasovskji functionals satisfying the conditions of Theorems 5.2 and 5.3 involve some difficulties. Sometimes it is easier to construct the functionals with a nonpositive derivative. The corresponding theorems follow. First we formulate a statement which extends the BarbashinKrasovskii theorem. Consider an autonomous RFDE of the form (5.1): q t ) = jyx,),
t 2 0, x t
= x(
+ e),
 h I e Io
(5.21)
By definition, the element $ E C [  h, 01 belongs to an wlimit set corresponding to the initial function cp if x(t, 0, cp) is defined on [0, co) and there exists a sequence (t,,}, t, + 00 for n + co such that IIx,,(cp)  $11 + 0, n ,00, and xt,(q) = x(t, O,O,cp). The set Q c C [  h, OJ is called invariant if x,(cp) E Q, t E [0, 0 0 ) for any cp E Q.
+
77
$5. Liapunov Direct Method for Equations with Delay
It is possible to prove that the wlimit set corresponding to any initial function is situated in the invariant set of Eq. (5.21).
Theorem 5.6 [81(1), 81(4)]. Let there exist the continuous functional V(cp):C [  h, 01 + R , such that w1ClCp(O)l1 IUcp)I
wl(t)+ 00,
~zCllcp(e)lll and
t + 00
ex,) 5 0
Let Z be the set of those elements from C [  h, 07 for which V = 0 and Q is the greatest invariant set situated in 2. Then all solutions of(5.21) tend to Q for t * co. In particular, i f the set Q has the only zero element then the trivial solution of (5.21) is asymptotically stable.
Set forth the theorem extending the Matrosov stability criterion founded on the use of two Liapunov functions [108(5), 162(10)]. Consider problem (5.1). Let V(t,cp) and W(t,cp) be continuous functionals defined on I x QH and denote V ( t ) by V(t,x,) and W ( t ) by W ( t ,x,). The derivative W is called integrally unbounded in a set G E Q H if for any number B > 0 there exist a number T ( B ) > 0 and a continuous function ( ( t ) such that uniformly in X, = x(t 0) E G for t 2 to
+
W I
w,
6"
B ((s)dsI
(5.22)
T(B)
Denote d((p(O), C) the distance between the element cp(Q set G c C [  h, 01.
E C [  h,
01 and the
Theorem 5.7 [l08(5), 162(10)]. It is necessary and suficient for uniform asymptotic stability of the trivial solution of (5.1) that there exist thefunctionals V(t,cp) and W ( t ,cp) such that
ol(icp(o)i) w,
I cp) 4 w2(iicp(e)ii), t 2 t o , N O E Q ~ , ~ , E G (2) V I 0 3 ( x , )I 0, where W3(cp) is a continuousfunctional dejined on Q H ; ( 3 ) I W ( t , ~ ) Il L, t 2 to, d o ) E Q H , to E R i ; (4) for any p E (0, H ) there exists a p > 0 such chat the derivative W is integrally unbounded in the set E(p, p ) c C [  h, 01, where
(1)
Ebu, P ) = { v ( @ E Q H , d ( ( ~ ( e )Q(Gj ,
= 0 ) )5 P,
PI llc~(e)III H)
Here Q(W3 = 0) = (p E Q H , W3(cp) = 0). I n addition, the attraction domain of the trivial solution is the sphere S,, where K < H and w 2 ( K )I w,(H).
A general outline of the proof follows. Conditions (1) and (2) ensure the hit of any trajectory x ( t ) in the neighbourhood E(p, p ) of the set Q. By virtue of ( 3 ) and (4) the trajectory necessarily leaves E(p, p ) after a finite time. Then vwill
78
2. Stability of Retarded Equations
be negative. So the function V decreases and the trajectory x ( t ) approaches the origin. For a full proof of Theorem 5.7 see Kolmanovskii and Nosov [108(5)] and Nosov [162(10)].
EXAMPLE 5.2. Consider the equation
a(t) =  a x ( t )
+ b(t)x(t  k ) + c(t)y(t)
j(t>= yc(t)x(t),
t 2 0,
o
h2
(5.23)
Show that sufficient conditions of uniform asymptotic stability are (a Ib(t)l) I c1 < 0, c3 2 Ic(t)l 2 c2 > 0, y > 0. Introduce the functional
+
V ( t ,x , , y,)
= yx2(t)
+ y2(t) + ya
bih
x 2 ( s ) ds
We have I y(a
+ lb1)(x2(t)+ x2(t

k)) I 0
The set Q(ih3 = 0) s C [  k , 01 is, in our case, the set of functions (q,$) E C [  h, 01 such that q(0) = q(  h) = 0. The set Q(o3= 0) does not coincide with the origin. As a second functional we take
w(4 X t Y J = y(t)xtt>
sgn c(t>
(5.24)
This functional is bounded in any sphere Q Hand also
ri/=
 x(t)(  y c ( t ) x 2 ( t ) ) sgn c(t)  y ( t ) sgn c(t) .[ax(t) b(t)x(t  k ) c(t)y(t)]
+
+
The set E ( p , p ) is defined by the inequalities Ix(t)l I p, Ix(t  k)l I p, p2 I y 2 ( t ) 5 H 2 . In the set E(p, p ) we have
W=
c2p2
+ apH + Jb(t)JpH+ y c 3 p 2 I c 2 p 2 / 2
where p is such that apH
+ Ib(t)IpH + yc,p2
I 2apH
+ y c 3 p 2 I c,p2/2.
Hence the derivative Wis integrally unbounded in the set E(p, p). Asymptotic stability of the trivial solution of (5.23) follows now from Theorem 5.7.
5.6. Global Stability Let the map f(t, q),f : R x C[  h, 01 iR, be continuous and Lipschitzian in q E Q H for arbitrary H . Further, I f ( t , q)l I M , , t E R,, q ( 0 )E Q H .
Definition 5.1. T k e trivial solution of (5.1) is globally uniformly asymptotically stable i f it is uniformly stable and iffor any y > 0 and a sphere Qx there
79
55. Liapunov Direct Method for Equations with Delay
exists an l(y,K) such that l(y, K ) . t 2 6,
+
Ix(t,to,cp)l Iy for any
cp(B)EQ,
and
Theorem 5.8 [108(5)]. Let there exist two functionals V ( t , cp) and Wft, cp) satisfying the conditions of Theorem 5.7 in any Q H and q(s)
~1(1cp(O)I) I V ( t ,cp),
+
m,
s
(5.25)
+
Then the trivial solution of (5.1) is globally uniformly asymptotically stable. Notice that under condition (5.25) and other assumptions sufficient for asymptotic stability (see Theorems 5.1, 5.3 and 5.4) the solution is also globally asymptotically stable.
5.7. Survey of Other Results. Exponential Stability The stability problems for RFDEs are studied in different ways. Stability in the first approximation in various senses is studied by Halanay [79(4)], Hale [8l(l)] and Kolmanovskii and Nosov [108(5)]. Stability under steady acting disturbances is considered by Germaidze and Krasovskii [63]. Instability theorems are proved by Shimanov [205(1)]. Critical cases (i.e., the cases in which some roots of the characteristic function lie on the imaginary axis) are also studied [167,178,205(2)]. The Klimushev papers [ l o l l are devoted to the study of the stability of singular perturbed RFDEs. For applications of the vector Liapunovfunctional to the stability of connected systems see Gromova and Markos [73]. Also available is a survey devoted to the Liapunov direct method for RFDEs [3,108(5)]. Mention some results about exponential stability. The trivial solution of (5.1) is exponentially stable if any solution x(t, to, cp) of (5.1) satisfies the inequality
I x(t, t o , cp)l
I Bllcp(e)II exp(  a(t  to))
B > o, a > o, t 2 t o , iiqo(e)ii I H, IH
(5.26)
A necessary and sufficient condition [114(5)] of exponential stability of the trivial solution of (5.1) is the existence of a functional V ( t ,cp) such that
c,
CI /I (P(e>ll 5 v ( t ?cp) I II cp(@ i.5

c, 11% I/>
I V ( t , CP)

/I
v(t,11/11 I c, lid@ 11/(e>II
(5.27)

where C iare some positive constants. Note that for a linear RFDE the exponential stability is equivalent to the uniform asymptotic stability. Further, for the linear autonomous RFDE (5.17), asymptotic stability implies uniform asymptotic stability and, hence, exponential stability. It is interesting
80
2. Stability of Retarded Equations
to notice that for linear autonomous NFDEs, asymptotic stability does not, in general, imply uniform stability. For example [78(1)], the equation i ( t ) + i ( t  h) x ( t ) = 0, h > 0, has a solution tending to zero for t ,00 as t” where v > 0 (see Chapter 1 45). This means that the trivial solution is not exponentially stable. Hence, asymptotic stability may not be uniform in the initial function.
+
96. CONSTRUCTION OF LIAPUNOVKRASOVSKII FUNCTIONALS
6.1. Linear Autonomous Equations Krasovskii has considered equations i ( t ) = Cx(t)
+ Bx(t  h) +
1;
D(0)x(t
+ 0) do,
XER,
(6.1)
h
A quadratic functional is constructed as [114(6), 183(2)]
+
V(X,,)= x’(O)C~X(O)
x’(s)~(s)x(s) ds j;h
(6.2)
Here a is a constant (n x n) matrix and p, y are (n x n) matrices with piecewise differentiable elements. The derivative P of functional (6.2) along the paths of Eq. (6.1) equals
P = x’(O)ox(O)
+ x’(  h);lx(O)
 h) ds + ~ ~ h ~ ’ ( s ) ~d s( +~ ) x x’(s)cp(s)x( (~) 1;h
(6.3)
Theorem 6.1 [114(6)]. Let the solutions of Eq. (6.1) be asymptotically stable. Then for any functional W of the form (6.3) there exists the uniquefunctional V of the form (6.2) such that V = W. Besides, if the functional W is negative dejinite, then the functional V is positive dejinite. Equations for matrices a, p, y depending on w , A, v, cp and E are derived by Repin [183(2)].
81
$6. Construction of LiapunovKrasovskii Functionals
Quadratic Liapunov functionals giving necessary and sujicient stability conditions for the matrix equation i ( t ) = Ax(t)
+ BX(t  h),
x E R,
(6.4)
are constructed by Castelan [34,35]. These functionals are obtained by passage to the limit in Liapunov functions derived in the usual way for difference equations which approximate RFDE (6.4). Notice that this algorithm is complicated enough and the problem of construction of more simple functionals remains actual. Tzar’kov and Engel’son study the problem of construction of a LiapunovKrasovskii functional for linear autonomous and periodic systems [23 1( l), 232(2)]. The functional is constructed in the form
v,cp) =
11
cp(S)cp(Sl)
dPt
Q
Here pt is a symmetric measure defined on the square Q = [  h, 01 x [ h, 01 for every t. A system of equations is derived for this measure. The equation i ( t ) = + a x ( t  1) is considered as an example by Tzar’kov [231]. LiapunovKrasovskii functional giving the necessary and sufficient condition of asymptotic stability (i.e., the condition n/2 < a < 0, see $3) is
+
cp2(s) ds
V(cp)= ~ ’ ( 0 ) a’
1’
a ( + 1 ‘OS +sina
+ (a2 + 2)
g(s) sin as ds 
1
~
2a
where
6.2. Stability of a Chemical Reactor Closed by a P Controller Investigate the stability of the chemical reactor described in Chapter 1, $1. Assume that the controller is a P controller and the actuator has dead zone, saturatim and hysteresis loops. A structural scheme of this feedback system is represented in Fig. 2.20. The nonlinear element describing the actuator may be defined in different ways. The output u(t) may be considered as a onevalued functional depending on previous history x(s), 0 I s I t ; i.e., u(t) = F(x(t + e)), t 2 8 I0. The distinctive feature of such a description is that
82
2. Stability of Retarded Equations
Fig. 2.20. Structural scheme of feedback system for chemical reactor.
the functional has unbounded delay for t + co.It is obvious that F satisfies the Lipschitz condition:
IF ( d t +

F($(t
+
I
r II cp(t
+ 0)  rL(t + 0) IIC[O,t ]
or IF(cp(t + 6))  F($(t
+ Wl'
[om
jorndsR(r,cp> $)
I&
 s)  $(t  s)l' dsR(s9 cp,
$)
(6.5)
r2
+
The system represented in Fig. 2.20 is described by the equation T i ( t ) x(t) = K,K,F(x(t  z + 0)) + K,K,y(t  7). Let j ( t ) be a fixed input signal, x ( t ) corresponding to j ( t ) be a disturbed solution and ( x ( t ) z(t)) be a disturbed solution. The disturbance z ( t ) satisfies the equation
+
~ i ( t+) z ( t ) = K,K,F,(z(t

+ 8))
(6.6)
where ~,(z(t
+ 8)) = q x ( t  7 + e) + z(t  T + e))  F(x(t  5 + el) z2(t z + e) d , ~ ( s x,, , zS) F:(z(t  + 8)) I 
lom
Consider the functional rr
r m
V
=
T z 2 ( t )+ K , K , Y  ~J
[d,R(s, X,, z,)] 0
J
z2(sl) d s , tr+s
83
96. Construction of LiapunovKrasovskii Functionals
The derivative of this functional is
P = 2 z ( t ) [  z ( t )  K,K,F,(z(t
+ K,K,r'z2(t)
 7
+ 6))]
d,R(s, X,, z,)  K , K p r  ' Jorn
. Iornz2(t T
+ s) d,R(s, X,, z,) I
+ K,K,rlF:(z(t
T
2z2(t)
+ K,K,rz2(t)
+ 6 ) ) + K,K,r1z2(t)r2
. {omz2(t  7 + s) d,R(s, X,,z,)
 K,KpF1
2 z Z ( t ) [  1 + K,K,r]
If K,K,r < 1, then all hypothesis of Theorem 5.4 are fulfilled. Hence, the trivial solution of Eq. (6.6) will be asymptotically stable. In other words, the closed system of a reactor regulation will be stable.
6.3. Scalar Nonlinear Equations Consider the RFDE
Kernel ko(s) has bounded variation on [0, a]. The initial function rp(6) E CB[  co,01; i.e., it is continuous and bounded on (  co,01. The continuous functional a(t, rp): [0, a)x CB[  co,01 + R , is such that a(t, 0) = 0 and
(6.9)

Let the kernel ko(s) have a jump b > 0 in zero and also
Idko(s))+ r,
b> +O
cll0
+
Jams
dR(s) < 00
(6.10)
84
2. Stability of Retarded Equations
Then the trivial solution of Eq. (6.8) is asymptotically stable. For the proof, consider the functional V(t,x,)
= X2(t)
I’,
+ j+mo dko(s)
+I*
ds,
X2(Sl)
jOrndR(.)
6’,
x2(s1) ds,
(6.1 1)
The derivative of the functional (6.1 1) is
v = x2(t)(  2b +
+ 2x(t)
c jm
[
a(t, x,) 
dk,(s)
x ( t  s)
< Ix’(+ 
Idko(s)l 
+O
.]
1 (6.12)
Hence, by virtue of Theorem 5.4,the trivial solution is asymptotically stable under conditions (6.10). Investigate the similarly nonlinear equation i(t) = 
~ ( t S)
d,k,(t,
S)
+ ~ ( tx,),,
t 2
0 (6.13)
lorn
xo(0) =
do)
Here k,(t, s) = 0 for t 2 0, s = 0 and
ko(t, s) =
s:
f ( t , sl) ds,
+
m
ai(t)e(s 7i(s)) i= 1
t 2 0, s > 0, sl(t)= 0, al(t) # 0, 2 4 s ) = 1
+ sgn s
The function ko(t, s) has for all t a bounded variation in s E [0, m). Assume that the functions ai(t) are continuous and that zi(t) are continuously differentiable.Function f ( t , s) is continuous and f ( t , s) = 0 for t I 0, s 2 0. There exists a constant L > 0 and a bounded nondecreasing function g(s) such that for all t 2 0 the series lai(t)l converge uniformly and 1 Iai(t)I < L,
I : ,
I?=
js;lf(t>s)l
ds I g ( 4

g(s,),
Vs,,
s2,
0 < s1 < s 2 a
Let the Eq. (6.12) coefficients meet the formulated conditions and also let
85
$6. Construction of LiapunovKrasovskii Functionals
(a) 0 I ti I q < 1 for all i (hence functions t  t i ( t )have inverse differentiable function ki(t)). (b) r
sup [Iomds
Pm
[+’
m
If(tl, s)l d t ,
+
120
s dR(s) lom
+ f l.l’)/mi(s)Ids]
< co.
i=2
Then the trivial solution of (6.1 3) is asymptotically stable. For the proof, consider the functional r m
+f
prs
kiW
Iai(s)lx2(s zi(s)) ds
i= 2
+
I,
IomdR(s)[;2(tl)
dt,
(6.14) From (6.14) and (a)(c) it follows that x2(t) I
v
I i=2
m
jom
If(t
+ s, s)l
ds  2r 
If(t, s)\ d s om !
1 Iai(ki(t))hi(t)I i=2
1
I P1X2(t)
So, because of Theorem 5.3, the trivial solution of Eq. (6.13) is asymptotically stable. b(t)x(t  h), conditions (a)(c) of For the equation i ( t ) = a(t)x(t) asymptotical stability take the form inf, a(t) > sup,)b(t)/,t 2 0. The cases in which equations of the first approximation have no jump at zero are more complicated. They are studied on the basis of degenerated functionals in Chapter 3, $4.
+

6.4. Nonlinear Equations of Second Order
Some functionals for secondorder equations are constructed by Kolmanovskii [105(1I)] and Kolmanovskii and Nosov [108(5)]. In view of
86
2. Stability of Retarded Equations
their awkwardness we study here only the system i ( t > = Y(tX
P(t) = a (t)y(t)
+ a(t)x(t

(6.15)
h)
Assume that 0 < A I a(t) I C , , 0 < a(t) I C , ,
2a(t) 2 ha(t) + [tt+hu(s)ds,
b(t) I 0
(6.16)
Then the trivial solution of Eq. (6.15) is stable. For the proof, consider the functional
+ y,(t) +
I/ = a(t)xZ(t)
It may be verified that
P I Ci(t)XZ(t)
+ yZ(t)
[
za(t)
+ ha(t)] +
6"hu@) ds]
Now stability follows from Theorem 5.3. If both inequalities (6.16) are strict then by Theorem 5.3, we have asymptotic stability. If only the first inequality (6.16) is strict, then asymptotic stability may be proved by means of the Matrosov criterion 5.7. The set Q (G3 = 0) in this criterion for the functional I/ consists of elements y,(B) E C [  h, 01 such that y ( t ) = 0. Define the second functional
w = y(t)x(t

h)
(6.17)
Functional (6.17) is bounded in any sphere Q H and
w = x(t)y(t)x(t

h)  a(t)xZ(t  h )  y(t)y(t  h )
In the set E ( p , p ) (i.e., for pz < llxt112 I H Z and Iy(t)l I p ) we have W I ApZ
+ p H SUP l ~ ( t )+l p H I Apz/2 120
Here p is such that
p H sup la(t)l
+ pH I Ap2/2
t>O
Now the asymptotic stability of the trivial solution of Eq. (6.15) follows from Theorem 5.7.
87
$7. Stability of Nuclear Reactors
In Sinha [210] stability of some third and fourthorder equations is investigated by using LiapunovKrasovskii functionals. For references devoted to applications of the Liapunov direct method to RFDEs see Alekseevskaja and Gromova [3].
57. STABILITY OF NUCLEAR REACTORS 7.1. SingleTemperature Reactor with Convective Feedback
A mathematical model of the singletemperature reactor with convective feedback described in 41.1.6 is [70] N ( t ) = [ aT(t  z)
T ( t ) = r"(t)


& ( N ( t ) NO)]N(t)
(7.1)
N o  T(t)]
Investigate the stability of the stationary solution T(t) = 0, N ( t ) = N o . The system of the first approximation for (7.1) is ii(t) = bn(t)  aT(t

z),
T ( t ) = n(t)  T ( t )
where a = a N O ,b = EN,,, n(t) = N ( t )  N o , r tion for system (7.2) has the form
=
(7.2)
1. The characteristic equa
z 2 + ( 1 b)z+b+aexp(z(z)=O
(7.3)
D subdivision (see Chapter 2, $3.2) for quasipolynomial (7.3) consists of the straight line a + b = 0 and the curve a(o) = (1
+ 02)o/(o cos oz + sin oz)
b ( o ) = o ( w sin 05  cos O T ) / ( O cos oz + sin oz)
In Fig. 2.21, D, denotes the stability domain for z > 0 and D, u D, is the stability domain for z = 0. The delay z has a certain destabilizing effect on system (7.2), and it is necessary to take it into account. According to Theorem 5.5 about stability in the first approximation, system (7.1) is asymptotically stable if its parameters lie in the stability domain D, of linear system (7.2) ~701.

7.2. Stability of Two Interconnected Reactors
The dynamics of two interconnected reactors exchanging energy by neutrons under the assumption that the internal feedback of each reactor
88
2. Stability of Retarded Equations
tb
C={( t TI’,( Fig. 2.21. Stability domain for singletemperature nuclear reactor.
may be characterized by negative reactivity factors is described as = pl(t) 
” N , ( t ) + cxlz N z ( t  zlZ)+
4
11
ml
1 Ail Cil i=l mz
Nz(t)= pz(t)
 P2
N,(t)
+ clzl N , ( t  z21) + 1 AjzCj2 12
12
Cil(t)=
 A ~ ~ c+~Pi, ~( ,~( t)) , 1
j= 1
i = I , . . . ,m,
11
+
ejz(t) = Pl(t) = P l O  El“l(t)
1,
(7.4) ~ ~ ( t ) j , = 1, . . . , m,
 N101,
P Z ( t ) = P z o  &,“At)  N201 Here N i , pi(i = 1,2) are neutron density and reactivity in the reactor number i, respectively; li is neutron lifetime; C i k ,I , and P i k are concentrations, decay constants and fractions of group number i of retarded neutron emitters in reactor number k, respectively; m iis a number of groups of retarded neutrons in reactor number i; ei is a negative power reactivity factor; and PI, = x ? L l B i k Investigate the stability of the stationary solution = 1,. . . ,m,
Cil = Cilo > 0,
i
C j 2 = C j z O> 0,
j = 1,. . . ,m 2
(7.5)
89
97. Stability of Nuclear Reactors
Let us introduce the new variables
Equations (7.4) can be written in the form
= xl(t)  yil(t),
i = 1, . . . ,m,
A,= (B  YF(t))V(t) C(t) = C* for t 2 0 if co> C*. Really, from the second equation of system (8.2) using the variationofconstants formula and the inequality t(m(s))F(s z)V(s  t) 2 0, we have c ( t >= c *
+ (Co  C*) exp(pct) + a
c2 = (WfP
 9Y2PcC*)/P(aP P J Y )
F 2 = B / y , m2 = SV2/pm.If these values are positive, they can be interpreted as a state of chronic disease.
Theorem 8.2. Stationary solution 1 is asymptotically stable if
B < YF*
(8.5)
Proof: If system (8.2) is linearized in the neighbourhood of the stationary solution y , Ci, F i , mi, i = 1, 2, we have lXt) = (B  YFi)U(t) Y V f ( t )  ~ f ( t ) u ( t )
i ( t ) = CtFiU(t  5 )
+ ffi/,f(t
 t)  pcs(t)
+ au(t  z)f(t
f ( t >= ~ s ( t) (VYY+ ~ f ) f ( t) ?yFiu(t)  ~ f ( t ) o ( t ) b(t>= 04t)

 t)
(8.6)
PmAt)
Here we denote u(t) = V ( t )  6 , s(t) = C(t)  Ci, f ( t ) = F(t)  F i , p ( t ) = m(t)  mi, i = 1, 2. The characteristic equation for the linear part of (8.6) in the neighbourhood of solution 1 can be written in the form
D(4 =
 0, C , = C*, F , = F*, m, = 0.
Theorem 8.3. Let condition (8.5) hold and also let 0 < V, < IB = P V  ' ~  ~ C*P f Y  l V 
(8.7)
Then the component V ( t ) decreases on the interval [O, a).
At t = 0 we have v(0) = (B  yF*) V, < 0. Consequently, V ( t ) decreases on some interval [O, t , ) . If t , # a,then there is a first moment t , such that P(t,) = 0 and v(t,+ A) 2 0, A > 0. Therefore, it follows from the first equation (8.2) that yF(t,) = B, yF(t, + A) IB and P ( t , ) S O . On the other hand, from the third equation (8.2), Theorem 8.1 and Eq. (8.7), we have =m
&Z)
t , >  w F ( t , ) V ( t z )  PfF(t2)> PC*  rB%  P f P Y  > 0
This contradiction proves Theorem 8.3. The value IB is called an immunological barrier. Theorem 8.3 states that if at the initial moment the immunological barrier has not been passed, then the disease is mild and the quantity of viruses penetrating an organism decreases. Consider now the stationary solution 2. In this case the characteristic equation of the linear part of system (8.6) has the form
W )= (Pm + Z ) D , ( Z ) D,(z)=(z3az2bz+d+gzexp(zz)fexp(zz)) a = Pc
d
+ VYV, + Pf > 0,
= YPCVB&
> 0,
b
= Pc(?YV,
9 = WV,,
f
+ P f )  VYBVz
(8.8)
= Bg
Theorem 8.4. Zf pu,t s 1,
0 < (f  d)(a  gz)' < ( b  g  f z )
(8.9)
then stationary solution 2 is asymptotically stable.
Proof. The proof of Theorem 8.4 will follow from Theorem 5.5 about the first approximation stability if we show that Eq. (8.8) has no zeros in the halfplane Re z 2 0. Applying the Michailov criterion (see Chapter 2,§3.4) to the quasipolynomial Dl(z), we obtain

+ d  f cos o~+ g o sin w t V ( o )= Im D,(w) = o3 bw + gw cos o z + f sin wz
U ( w ) = Re D,(w) = ao'
Since U(0) = d  f < 0, V(0)= 0. The inequalities B(a

g z ) > ypC,(l  p c z ) 2 0,
p C a> b 2 br
94
2. Stability of Retarded Equations
imply for w > 0 dU/do >~
+
4  2( f~~ g)s  SZ] > o [ ( u  gz)
+~
( b g  ft)] > 0
Consequently, there is a value o2such that U ( 0 2 )= 0
and
V(0,) < w2[o:  b
+ g + fzl
YF*
1. Mild 2. Sharp
V(t)
+ 0,
t
t
co
t ( t ) is great
/9  yF* > 0.33 PS > rlYf
3. Chronic 4. Mortal
0 < B  F*y < 0.33
V(t)
+
B > YF* SP < rlrf
V(t)
+
V, > 0, t + co co, f + co
$9. DESIGN OF ADAPTIVE CONTROLLER FOR RETARDED SYSTEMS Adaptive controllers, and in particular model references adaptive controllers (MRAC), are often used for quality control of plants with a wide range of variation of dynamical properties or with considerable noise [I108(5), 1221. An MRAC principal scheme is represented in Fig. 2.25. The aim of MRAC design is the construction of an adaptation circuit which provides the ) x ( t )  y(t). Here x ( t ) is system output and y(t) stability in phase error ~ ( t = is model output. 9.1. Scalar Equations
Let the system be described by the scalar RFDE i ( t ) = (a
.~
+ AU + &(t))x(t) + ( b + Ab + 6b(t)) ( t h + Ah + 6h(t)) + f ( t ) , t 2 to
(9.1)
Here Aa, Ab, Ah are a priori unknown quantities from some domains: IAal 5 A , , IAbl I B,, IAhl I H , < h. Parameters 6a(t), 6b(t), 6h(t) are changed by adaptation circuit and f ( t ) is an input, I f ( t ) l I F o . Along with RFDE (9.1) consider the reference model j ( t ) = ay(t)
+ by(t  h) + f ( t ) ,
t 2 to
(9.2)
with zero initial data x ( t ) = y ( t ) = ~ ( t=) 0, t I to. For a > Ibl the trivial solution of Eq. (9.2) is uniformly asymptotically stable (see $3). At first assume that delays in the system and reference model are equal, i.e., Ah 6h(t) = 0. Then equation for ~ ( t=) x ( t )  y ( t ) is
+
i(t) = ae(t)
+ bE(t

h)
+ cr(t)x(t) + B(t)x(t
t 2 t o , a(t) = AU
+ du(t),

h)
P ( t ) = Ab
+ 6b(t)
(9.3)
97
59. Design of Adaptive Controller for Retarded Systems
Y
Fig. 2.25. Principal scheme of adaptive controller.
The design of an adaptation circuit may be done on the basis of the LiapunovKrasovskii functionals method [108(5), 122,163(2)].Consider the functional V t , E,, 4
8) = y(t)E2(t) + U 2 ( t ) + B2(Q
;
+ (4) 
7 ( t ) ) l  h & 2 ( s ) ds
The derivative of functional (9.4) along the paths of Eq. (9.3) is equal:
P = (7  2ay)c2(t) + 2 y b ~ ( t ) ~( th)
+ 2 y x ( t ) ~ ( t ) a ( +t ) 2yx(t  h ) ~ ( t ) B ( t ) Take the adaptation circuits of parameters a and
as
(9.4)
98
2. Stability of Retarded Equations
Model
I
I Multiplier
Plant
L) I
Delay unit
iMultiplier
Multiplier
Fig. 2.26. Scheme of adaptive controller for system with delay.
The principal scheme of an adaptive system using algorithm (9.9, (9.6) is represented in Fig. 2.26. The derivative of functional (9.4) along the paths of Eqs. (9.3), (9.9, (9.6) is
i. = (j  2ay)&Z(t) + 2Yb&(t)&(t h)
99
$9. Design of Adaptive Controller for Retarded Systems
For all t 2 0 let
Theorem 9.1 [108(5), 163(2)].
C, I ay
+ Ibl I  C
a
0 I C, I y(t) I C,,
+ (bly + 37 I C4
)
thjsSt
Relation (9.7) yields
V
I (a
+ 1 b 1 + iv)y(t)s2(t)+ ( ay + I b 1 y + f j ) s 2 ( t

h)
I  C4(&2(t)+ &2(t h)) I 0 Assume now that &(t)does not tend to zero. Then the integral from f' is unbounded but the difference V ( t )  V(0) is bounded. Hence s(t) + 0, t + co. For the proof of the second part of Theorem 9.1 let us use the Matrosov criterion (Theorem 5.7). The equality = 0 is possible only for s(t) = s(t  h) = 0. So functional (9.4) satisfies the first two conditions of Theorem 5.7, As a second functional consider
w,st, a,P)
= (a(t>x(t)
+ P(t)x(t  h))E(t)
(9.10)
Denote E ( p , p ) the domain E(p, p)
B)
CCh, 01 x R,,I4t)l
I P? Is(t  h)l I p, 0 < p 2 < a2(t) + B2(t) < a:},
= ((4 a,
a, < a
The derivative W of the functional W along the paths of Eqs. (9.3), (9.9, (9.6) equals W = (a(t)x(t) p(t)x(t  h))[as(t) bs(t  h) a(t)x(t)

+
+
+
+ P(t)x(t  h)I + s(t)(d/dt)(a(t)x(t)+ B(t)x(t  h)) = (a(t)Y(t)+ bYt)y(t  h))' + c(t)f(a, b, x, i,~ ( t )~, ( t h)) (9.1 1)
100
2. Stability of Retarded Equations
For W in the domain E ( p , p ) the estimate holds
s,
s,
s+ T(B)
s+ T ( B )
W t )dt
2
+ qy(t  h))’ dt  21 B 2 2.5B, p z + q2 2 pz > 0
(py(t)
p = tL(s), q = P(s),
~
Statement of Theorem 9.1 now follows from Theorems 5.7, 5.8. 9.2. Delay Adjustment [162(1 l)]
Let the reference model be described by (9.2) and let the plant be described by the equation i(t)=
ax(t)
+ bx(t  h + Ah + 6h(t))+ f(t) t 2 to
If x(t) is a twice continuously differentiable function and IA all t > t o , then ~ ( t=) ~ ( t) y(t) satisfies the RFDE
+ bs(t  h) + b[x(t  h + z(t))x(t = as@) + bs(t h) + bo(t)i(t  h) + b(zZ(t)/2) X(t  h + Or@)) where z ( t ) = Ah + 6h(t), 0 < O < 1. i(t) = as(t)
(9.12)
+ 6h(t)(< h for  h)]

(9.13)
Consider the following algorithm of delay selfadjustment: t ( t ) = (d/dt)(bh(t)) =
by(t, x)W  h1.m  V(t, x)C4t)/u(t)I
(9.14)
Here u(t) is a sensitivity function of solution x(t) relative to the following bu(t  h) + b i ( t  h), u(t) = 0, t 2 t o . delay variations: u(t) = au(t) Assume that lu(t)l > 0. Then algorithm (9.14) is physically realizable. Algorithm (9.14) is more complicated than (9.5) and (9.16) since, in order for it to be realized, we must know not only x(t) or x(t  h), but also the derivative i ( t  h). Notice that if a system is described by Eq. (9.1), and the variables x(t), x(t  h), i ( t ) may be measured exactly, then Aa and Ab may be determined algebraically from the equation, and we may then set 6a(t) = Aa, 6b(t) = Ab. Otherwise, under these assumptions the problem of parameter adjustments a(t) and B(t) is trivial. But for Eq. (9.12) the knowledge of x(t), x(t  h), i ( t ) and i ( t  h) is not sufficient in general for the determination of z(t) by an elementary method. Therefore, it is meaningful to take into account i ( t  h) in adaptation algorithm (9.14) for z(t).
+
101
$9. Design of Adaptive Controller for Retarded Systems
Theorem 9.2 [l08(5)]. Let thefirst condition of Theorem 9.1 be fuljilled, let the function x ( t ) be twice continuously differentiable and let IX(t)l Ic,
0I c, Iq(t, x) I c,,
lu(t)l > 0
(9.15)
Then the triuial solution of system (9.13), (9.14) will be uniformly asymptotically stable and stable under steadyacting disturbances for all suficiently small Ah such that Iz(t)l < hfor t 2 to.
9.3. Multidimensional Systems Let the reference model have the form j ( t ) = A(qO>y(t) + B(qO)y(t h) + f ( t ) ,
(9.16)
y(t)ERn
The system equation is i ( t ) = A(q)x(t)
+ B(q)x(t  h) + @(t)+ f ( t ) ,
x E Rn
(9.17)
Here matrices A(q) and B(q) are linearly dependent on the parameters q = ( q l , . . . ,qm), and @(t) is the action on the system from the adaptation circuit. We are given 4 = qo + Aq>
Aq
= (Aq,, . . ., Aqm)
(9.18)
Then Eq. (9.17) has the form
rA
+ B(qo)x(t  h)
i ( t ) = A(qO)x(t)
+f i= 1
aqi
Aqi x(t)
+ aB Aqi x(t aqi 

1+
h)
+
@(t) f ( t )
The problem is to choose a function @(t) such that E(t) + 0 for t
(9.19)
co. Put
Equation (9.20) is an adaptation signal for which the functions Gq,(t), i = 1,. . . , m, must be determined. For E = x  y, we get
+
rd
i ( t ) = A(qO)&(t) B(qO)E(t h)
+
f i= 1
aqi
x(t)
+ x(t aB aqi

1
h) ai(t)
(9.21)
(9.22)
102
2. Stability of Retarded Equations
Find functions vi by using the Liapunov direct method. There exists a quadratic functional vl(t,E ( t ) ,
E,) =
r' = r > o
+
&yt)r(t)E(t) T/z(&,),
(9.23)
such that the derivative of functional (9.23) along the paths of the system i(t)
=
+
(9.24)
A(qO)&(t) B(qO)&(t  h)
5 wg( le(t)/), where the subscript indiwill be negativedefinite ( pl)(9.24) cates that the derivative is taken along the path of Eq. (9.24). The existence of such a functional follows from Krasovskii Theorems 5.3, 6.1. Take now the functional rn
V ( t ,4% &,, u(t>,K t ) ) = vI(4
40,E,) +
1
(9.25)
i= 1
The derivative P of functional (9.25) along the trajectories of equations (9.21), (9.22) is equal to
Assume now that
Then the derivative V will be negativedefinite relative to
'
8,:
0
= (p1)(9.24) 5  w 3 ( 1 E ( t ) l )
Finally we get the equations of the adaptation circuit
It may be proved that E ( t )  + O as t  + co (See Kolmanovskii and Nosov [108(5)], $6.1). For the proof of the relation a ( t ) 0 some further assumptions must be made as in Theorem 9.1. +
$10. Stability of Viscoelastic Bodies
103
910. STABILITY OF VISCOELASTIC BODIES
This section is devoted to the study of applications of Liapunov direct methods to the investigation of the stability of viscoelastic bodies. General stability statements formulated in terms of the existence of a Liapunov functional could be easily obtained by adaptation of the theorems from $5. But the greatest interest is in constructing new Liapunov functionals and obtaining the effective stability conditions for concrete problems.
10.1. Constitutive Equations of Viscoelastic Bodies Viscoelastic bodies are characterized by the need to consider their memory effect, i.e., the influence of all past states of bodies from the initial to the current instant. Such phenomena are called hereditary. General features of phenomena under consideration are creep and relaxation. Notice that for a viscoelastic material which has been unstrained at all times prior to t o , the strain c(t) at time t is determined by the stress history o(z), to 5 z It through the constitutive equation. In accord with Reiner’s second axiom of rheology [6], viscoelastic materials are those whose behavior is partly fluidlike and partly solidlike. The principle phenomenon distinguishing viscoelastic substances from purely viscous fluids is the occurrence of strain recovery and viscoelastic substances are distinguished from perfectly elastic solids by the occurrence of creep. It appears that in viscoelastic substances stress produced by strain should relax as in purely viscous fluids, but at a finite rate. In order to obtain the constitutive equation, the Boltzmann superposition principle was used. The evolution of stress and strain are governed by the Constitutive equation
Here E(t) is a strain, o ( t ) a stress, E the Young’s modulus and K ( t , s) a creep kernel. The constitutive equation of nonhomogeneous ageing bodies has the form C6l
Here p(x) is a function characterizing the heterogeneity of an age. The model of nonhomogeneous ageing bodies generalizes the foregoing models of classical viscoelasticity.
104
2. Stability of Retarded Equations
Fig.2.27. Viscoelastic hingedhinged bar on viscoelastic foundation.
10.2. Dynamic Stability of a Viscoelastic Bar Let us derive the stability conditions of a viscoelastic hingedhinged bar on a viscoelastic foundation (Fig. 2.27). The bar is subjected to an axial load P . Denote y ( t , x ) as the lateral deflection of the bar from its straight position. The foundation action on the bar is equal to q =  C y  C l j , where j = ay(t, x)/at and constants C > 0, C , 2 0. The bar stability is investigated with respect to the initial disturbances of its deflection.
Definition 10.1. The bar is called stable iffor any E > 0 there exists a S ( E ) > 0 such that W ( t ) = yo(y(t,x)’ + j ( t , x)’) dx < E as soon as W(to)< 6 ( ~ ) . The bending moment M is given by M ( t , x ) = EJ(Z

R)y” = E J
Here J is the moment of inertia of the cross section and r(t  s) a relaxation kernel. The equation of equilibrium is
M” + Py”  q
=
(10.1)
pj
where p is the density of the bar material. Represent the deflection as a Fourier series m
y(t,x ) =
1 a,(t) sin d 1 x , n= 1
I,
71
=
1
(10.2)
105
$10. Stability of Viscoelastic Bodies
Substituting (10.2) in (lO.l), we obtain an RFDE for a,(t) a,
+ Blnhn+ BOnan+ B2,
I:
r(t  s)a,(s) ds
bOn= [c ~ i f n + 2 ~ ~ i ; ' ~ 4 1 ~  1fiIn , =c l p pzn = pL1(n11)4EJ, an(td = an07
=
0 (10.3)
hn(t0)
= ano.
Hence the investigation of the stability of the trivial solution of Eq. (10.1) is reduced to the investigation of the stability of Eq. (10.3), studied earlier in this chapter. It is convenient to put a,(t) = 0 for t < t o . Then a solution of (10.3) may be considered as a solution of
ii,
+ Blnhn+ POnan+ Bz,,
61
r(s)a,(t  S) ds = 0
(10.4)
Consider the Liapunov functionals V,
2 vn
+ 2a,Z(t)B3n +
= hi(t>
+ l:r(s) B3n
= Bon
ds l  s d t l
+ I,rW
[
hn(t)
+ Blnan(t)  Ilr(s) ds jt:san(tl)
jt:ld(f2) + B3,a;(tdI
dtl]
d t ~
(10.5)
ds
of functionals (10.5) along the trajectories of Eq. (10.4) The derivatives satisfy the inequality
From the positivedefiniteness of functionals (10.5) and the negativedefiniteness of functionals (10.6), we arrive at the following stability conditions
P < max(EJn2lF2,2 m ) C1pL'>
I:
sr(s) ds
(10.7)
Actually, relations (10.5) and (10.7) yield

m
1 (a,2(t)+ h i @ ) )+ 0,
t
+ co
n= 1
Hence, by virtue of the Parseval equality, W ( t )+ 0 for t + 00. But for any finite time interval, W ( t )= O(W(to)).So the bar is stable under conditions (10.7).
106
2. Stability of Retarded Equations
511. ABSOLUTE STABILITY OF SYSTEMS WITH DELAY Absolute (in the LurieAizerman sense) stability has been investigated by Popov and Halanay [176]. Razvan’s work [ l 8 l ] and the appendix to the Russian translation of it by Lihtarnikov and Jakubovich [130] contain the most important results in this domain. We limit our presentation to the statement of the problem and the formulation of two theorems.
a Summer
Fig. 2.28. Feedback system with nonlinear actuator.
Consider a feedback system with delay shown in Fig. 2.28. System phase variables are measured by pickups. The adder generates a control signal (r = c l x l . . . + c,x, = c’x, which governs an actuator. The actuator output 5: is equal to f(a) in the case of direct automatic control without delay, i.e., ( ( t ) = f(a(t)). When there is a delay in the actuator, we assume that t(t>= f ( 4 t  4). Suppose that plant dynamics is described by the RFDE
+
i ( t ) = Ax(t)
+ Bx(t  h) + b((t),
~ ( tE )R,
Here the matrices A, B and vector b are constant. The system closed by an adder and an actuator with delay is described by the equation i(t) =
Ax(t)
+ Bx(t

h) + bf[.(t  h ) ] ,
C’X
(11.1)
107
$11: Absolute Stability of Systems with Delay
If the actuator has no delay, then we get k(t) = Ax(t)
+ Bx(t

h)
+ bf(a(t)),
(11.2)
= C'X
As a rule, the characteristic f(a) is nonlinear and its exact determination is difficult. Assume that a continuous characteristic f ( a ) is defined for all af(0) = 0 and also satisfies the following conditions: (a) There exist constants k,, k, > 0 such that k , 0 2 (b) J:f(a) do = CO.
< of(o)
k,a2.
Such characteristics are called admissible.
Definition 11.1. Control system (11.1) [or (11.2)] is said to be absolutely stable in the angle [k,, k,], ifthe trivial solution ofEquation ( 1 1.1) [or (1 1.2)] is globally asymptotically stable for all admissible characteristics f (a). For determination of absolute stability conditions of concrete systems one uses the ordinary frequency domain criteria of PopovHalanay [1761. We now state these criteria.
Theorem 11.1. Let all the roots of characteristic equation det[A lie in the halfplane Re z that
(l/K)
+ B exp( zh)

<  c i < 0. Also let
zZ]
=0
(11.3)
there exist a number q > 0 such
+ Re(1 + j o q ) exp(jwh)c'(A + B exp(joh) .b>O,
joZ)'
k, h and any t the monodromy operator U ( t ) is completely continuous and its spectrum is at most countable and does not depend on t : a(U(s)) = a(U(t))for each s and t. The eigenvalues of operator U ( t ) also called characteristic (or Floquet) multipliers of (12.1). Theorem 12.1 [223]. For asymptotic stability ofthe solutions ofEq. (12.1) it is necessary and suficient that the spectrum a( U ( t ) )of the monodromy operator U ( t ) lies inside the unit circle. I f all Floquet multipliers p j ofEq. (12.1) are such that lpjl I 1 and the multipliers p k with lpkl = 1 have simple elementary divisors, then the trivial solution of Eq. (12.1) is stable.
12.2. Particular Classes of RFDEs
Theorem 12.1 reduces the problem of the stability of RFDE (12.1) to the investigation of spectrum a( U(t)). Usually the problem of determination or estimation of a(U(t))is very complicated and now there are no significant results in it. But for some particular classes of Eq. (12.1) this problem is
I09
$12. Stability of Linear Periodic Equations
studied better. Consider a scalar equation with delays divisible by period [253(3), 253(5)]: i(t)
+ b,(t)x(t) + b,(t)x(t  mlo)+ b,(t)x(t  rn2w) = 0 bi(t + = bi(t), i = 0, 1, 2, x(t)E R , W)
(12.2)
where rn, and m, are positive integer numbers. Let us look for a solution of the form x(t> = P(t> exp(W,
P(t
+ W ) = P(t>
(12.3)
Such solutions are usually called Floquet solutions. For p ( t ) we get the equation @(t>+ g(4 M t ) = 0 g(t,
A) = b,(t)
+ b , ( t ) exp( m,Aw) + b,(t) exp( m,Aw)
Integrating this equation we conclude that the function
i
p ( t ) = C exp At 
c I g(s,
A) ds
must be w periodic. Hence we have the following equation for A: AW
+ jomg(s,A) ds = 0
(12.4)
Equation (12.4) is a quasipolynomial. We know (see Chapter 1,95) that Eq. (12.4) has at most a countable number of roots and each of them has finite multiplicity. The Floquet solution (12.3) corresponds to every zero of quasipolynomial (12.4). By Theorem 12.1 for asymptotic stability of Eq. (12.2), it is necessary and sufficient that all the roots A j of Eq. (12.4) satisfy the condition Re Aj < 0. For systems of equations with delays divisible by period, it is also possible to write out a characteristic equation, but its coefficients depend on the elements of the fundamental matrix of an ordinary differential equation. In this case, simple assertions are not derived [253(5)]. Nevertheless there are some results for equations with quasiconstant coefficients [115]. Thus for a system
110
2. Stability of Retarded Equations
the particular solution has the form PI = exPC@)tlP(t, 111,
4 0 ) = 1, p(t, p) = P(t
+ w, PI
Consequently, if all the roots I j of Eq. (12.6) satisfy the condition Re < 0, then system (12.5) is asymptotically stable for sufficiently small p. Some stability criteria for periodic equations with small parameters are obtained by a Laplace transform. For example, in Valeev [233] it is proved that the trivial solution of the scalar equation i(t) = 2p(cos t)x(t  h), p > 0, is stable if sin h > 0 and unstable if sin h < 0.
12.3. Floquet Theory Floquet solution of Eq. (12.1) is a solution x(t) such that X,+w =
pxt,
x, # 0, p # 0, a < t < 02
The number p called a Floquet multiplier is also an eigenvalue of the monodromy operator U ( t ) .The Floquet exponent I equals I = o In p. If p is a simple eigenvalue, then the Floquet solution $(t) = p(t) exp(At), p(t o)= p(t) corresponds to p. Let p be an rmultiple eigenvalue of the operator U(t). Then a finitedimensional subspace E,(t) c C [  h, 01 corresponding to p is invariant relative to U(t). The action of the operator U(t) in subspace E,(t) is described by a matrix. This matrix has Jordan form for appropriate basis. The following Y Floquet solutions correspond to an rmultiple eigenvalue
+
ICljl(t) = Pjl(t>ex~(At) = [{l/[m(j> t+hjmcj,(t)

l]!} t m ( j )  lpjl(t) + . . + ~jm(j)(t)Ie~*(12.7) j = 1, . . ., q,
m(1) + . . .
+ m(q) = r
Here q is the number of Jordan cells of the matrix describing the action of operator U(t) in an invariant subspace E,(t) c C [ h, 01 and m ( j ) are dimensions of these cells. All the functions pjl(t), . . . ,pjmcj,(t)in formulae (12.7) are o periodic.
Theorem 12.2. Any solution x(t) of Eq. (12.1) may be expanded in an asymptotic series over Floquet solutions (12.7)for t + 00. I n other words, for every solution x(t) and a number cc > 0 there exists thefinite sum %(t)of Floquet solutions (12.7): (12.8)
111
512. Stability of Linear Periodic Equations
such that
Ix(t)  %(t)l < C exp(@),
fl > a,
t +
00
Coefficients of the expansion (12.8) are determined by the aid of a socalled conjugate system [253(5)]. As an addition to Theorem 12.2 it may be proved that X(t) is a solution of some ordinary differential equation. This equation may be written out if Floquet multipliers are known. Hence, solutions of periodic RFDEs are approximated by solutions of ordinary differential equations with precision to functions vanishing faster than any exponential ones. It is possible to develop the perturbation theory of RFDEs similar to those for ordinary differential equations. Notice that in general Theorem 12.2 may not be made more precise; i.e., asymptotic convergence of a series over Floquet solutions cannot be changed by convergence in the usual sense. Hence, a system of Floquet solutions may be incomplete. EXAMPLE 12.1 [253(5)].
Consider a scalar equation
i(t)=
27c(cos(2nt))x(t  1)
(1 2.9)
Since the mean value of the coefficients equals zero, Eq. (12.4) takes the form A = 0. The Floquet solution $l(t) = exp(sin 27ct) corresponds to the unique root A1 = 0 of Eq. (12.4). Equation (12.9) has no other Floquet solutions, and so the system of Floquet solutions is not complete. Nevertheless, Theorem 12.2 is valid for Eq. (12.9). For any solution x(t) of this equation and arbitrary p > 0, we have x ( t ) = C,$l(t) + O[exp( Pt)]. Consider, for example, the solution x1 of Eq. (12.9) defined by the initial condition xl(t) = I , 0 < t _< 1. By the step method (see Chapter 1, $3) we get
Expanding the solution $l(t) in a Taylor series, we obtain
O12 dP
>'II5 CzlBI',
Il.S(ib)iB Ko(ifi))' 11 5 C , ,
IBI < p < 00 +
 00
Now for the end of the proof it is sufficient to repeat the proofs of Theorems 2.2.1 and 2.2.2.
REMARK1.1. Let condition (1.3) be invalid. Then the trivial solution may be unstable even if all the zeros of function (1.6) lie to the left of the imaginary axis. For example, the trivial solution of the equation L(L(L(x)))= 0,
L(x) = i ( t )  i ( t  1)
+ 7?X(t)
(1.16)
is unstable in the Liapunov sense. But at the same time all the zeros of the quasipolynomial corresponding to Eq. (1.16) lie to the left of the imaginary axis [216].
REMARK1.2. Let requirement (1.3) be fulfilled for some y > 0 and let (1.17) Further, assume that in the halfplane Re z 2 7, function (1.6) has no zeros. Then, similar to Theorem 2.2.3, it may be proved that Ix(t)l I C , exp(  y t ) . 1.2. Methods of Stability Investigation of Linear Autonomous NFDEs All the methods from Chapter 2, $3 may be applied for investigation of characteristic functions corresponding to linear autonomous NFDE. The
117
$ 1 . Stability of Linear Autonomous NFDEs
Fig. 3.Z. Stability domains for Eq. (1.17).
theorems of Pontrjagin und Chebotarev are valid for NFDEs without any changes and the method of D subdivision is also applicable. The stability domains for the equation i ( t ) = ax(t)
+ bx(t  z) + c i ( t  T )
(1.18)
in the space of parameters (a, b ) for c = 0.9, 0.6,0,0.6,0.9 are shown in Fig. 3.1. Equation (1.1) is always unstable for IcI > 1 because the corresponding quasipolynomial has a chain of asymptotic zeros of the form (1.5.5),with the positive real part equal to lnlcl. For all IcI < 1, Eqs. (1.1) have a common part of the stability domains a > Ibl. It follows from the Tsypkin criterion (see Chapter 2,93). Vyshnegradskii diagrams for control system (1.6.l), having the transfer function W ( s )= K ( T s I)' exp( ST), closed by a PD controller and by a PID controller in parameters Tl = T  ' , I , = I / T , D , D/T, are shown in Figs. 3.2 and 3.3 (consequent equations can be found in Table 1.2).
+
a=
118
3. Stability of Neutral Functional Differential Equations
2Stable
0 2
05
2
1
5
T1
Fig 3.2, Vyshnegradskii diagram for a system closed by a PD controller.
5
2
L Stable
0.2
0.5
1
2
r,
Fig.3.3. Vyshnegradskii diagram for a system closed by a PID controller.
119
$1. Stability of Linear Autonomous NFDEs
Fig. 3.4. Michailov hodograph for a stable NFDE.
The frequency stability criteria of Michailov and Nyquist from Chapter 2,93 are not valid for NFDEs in general. It is connected with oscillations in the Michailov hodograph. Consider, for example, the equation i(t)=
ax(t)
= ci(t

a >0
l),
(1.19)
The Michailov hodograph D(iw) = iw  a + ciw exp(  iw), 0 I w < co,for stable Eq. (1.19) when IcI < 1 is represented in Fig. 3.4 and for unstable Eq. (1.19) when IcI > 1 in Fig. 3.5. From Figs. 3.4 and 3.5 it follows that the functions arg D(iw) have no limit for o + co and that their increments on the interval 0 I w < co are not defined. But Michailov criterion is true [7] for a particular characteristic function
[
det I z  H
J(s)eCZsds

lorn

z
Jl(s)e’S ds Jorn
where
/oml
J(s)lep” ds < co,
J,(s)leps ds < &I,
1
p
>0
120
3. Stability of Neutral Functional Differential Equations
Fig. 3.5. Michailov hodograph for an unstable NFDE.
$2. STABILITY OF AEROAUTOELASTICITY EQUATIONS
2.1. Equations of Unsteady Motion of an Elastic Rigid Body The dynamic model of unsteady motion of the elastic flying vehicle must take into account the interaction of aerodynamic, elastic and inertial forces (i.e., aeroelastic effects) as well as the dependence of aerodynamic characteristics on the past states of flow (or on the aerodynamic trace). One such description is the aeroautoelasticity models [19,201. Practically, one can consider only small perturbations of the motion and deformation parameters. The widely used method of known forms is based on the representation of deformations as a series of known forms which are usually the natural oscillation forms of the vehicle in a vacuum. If one is restricted to a finite number n of the oscillation forms, one can construct an approximate finitedimensional model of flying vehicle motion p(ij
+ 2waeq + 0 2 q )= c
(2.1)
121
$2. Stability of Aeroautoelasticity Equations
Here p, w, z are the diagonal (n x n ) matrices of generalized masses, frequencies and damping factors of natural oscillations; q is a vector of generalized coordinates; and C is a vector of generalized forces. The matrices p, w, ae are defined simultaneously with the calculation of natural oscillation forms on the basis of a masselastic model of the vehicle. The knowledge of a vector q permits us to calculate the displacements of any point (5, I], i)of the vehicle and their derivatives according to the expression
Here
are natural oscillation forms of the vehicle in the coordinate system
(t,I], [) fixed in the principal inertial axes of the vehicle. The vector of the generalized forces C for the known aerodynamic pressure p(5, I], [,t), applied to the vehicle surface cr, is
(2.3) Here E ~ j , = 1, . . .,N , are the kinematic parameters, including vectors q(t) and q(t), vector of wind velocity A(t), and also vectors d(t), &t), describing the states and the rates of states of control surfaces. We write the reaction of component C , of the generalized force vector on the step change of E, (i.e., the unit step function) in the form
H;(e)
= C;
+ z;(e)
(2.4)
Hence C: is a quasistationary aerodynamic derivative. The functions fz(8) are called the unsteady unit step functions. They describe the influence of the aerodynamic trace. For the subsonic speed (Mach number M < 1) the function f;(8) are determined for all moments 8, 0 I 8 < co and for sufficiently great 8, I;(O)zAf?*. For the supersonic speed ( M > 1) f38) = 0 for 8 2 8* > 0 [7, 191. The unsteady unit step functions are calculated theoretically on the base of the discrete vortices theory or determined experimentally. Usually they are represented by tables or by diagrams. Figure 3.6 shows the unit step function of aerodynamic moment m, with respect to the diametrical axis, passing by the choice of a rectangular wing with aspect ratio of A = 5. Curve 1 corresponds to M = 0.7 and curve 2 to M = 0.9. The quasistationary values of moment mz = H;=(co) are indicated by dotted lines on Fig. 3.6. Using the Duhamel integral, we can write any aerodynamic characteristic in the form C,(t) =
1 C,,, J
C,,(t) = C & J+ E~
(2.5)
122
3. Stability of Neutral Functional Differential Equations
Fig. 3.6. Unsteady unit step function of aerodynamic moment m,: 1, M
= 0.7; 2,
M
= 0.9.
Substituting (2.5) in (2.1) we obtain the closed linear equations of Jiying vehicle perturbated motion
144 + oaq + 0 2 q )  C4q  Cqq 
sd
@(0)ij(t  0) d0 = C66
+ [:II(0)6(t
 0) d0
s,'
@(0)q(t  0) d0
+ CA' + CAA
+
Equations (2.6) describe the deterministic model of flying vehicle dynamics. The stochastic model is more adequate. It takes into account the atmospheric turbulences and chaotic noises in the pickups. The stochastic differential equations for the turbulent wind velocity A and the model of pickups are taken in the form
A=F,A+ti,,
i=Mr+i+ti2
Here til, Lj2 are vector stochastic processes of the white noise type with known intensity matrices, r is a vector of pickup states and 2 is a vector of output
123
12. Stability of Aeroautoelasticity Equations
pickup signals. The drives of control surfaces are described by the equation 6 = S6 + Gu where u is a vector of control drives signals. Finally, the linear stochastic aeroautoelastic model has the form
ff
Equation (2.7) is an SFDE of the neutral type. In Chapter 4 we shall formulate the theorem of existence and uniqueness of the solutions of such equations and shall give some methods of stability investigation. 2.2. Stability of Aeroautoelastic Equations Study the stability of deterministic equation (2.6) for the control vector
u
= 0:
q(t)
+ A q ( t ) + Bq(t) =
c
Z,(O)G(t
 0) de
+ S:I,(B)q(t
(2.8)
 d) d9
According to Theorem 1.2 for asymptotic stability of the solution of Eq. (2.8), it is necessary and sufficient that the characteristic function D ( z ) has no zeros in the halfplane Re z 2 0. The characteristic function D ( z ) is defined with the aid of the theorem on the Laplace transform of convolution and is equal to
D(z) = det[Zz2
+ Az + B  z J
Io(s)ezSds  z 2
J
0
Il(s)e’” ds] (2.9) 0
The Michailov criterion holds for function (2.9) if for some y > 0
jom
Ii(s)ey”ds < 00,
i = 0, 1
The kernels I,(s) are usually uncertainties Mi(s), i = 0, 1. It is shown by Astapov [7] that if function (2.9) has no zeros in the halfplane Re z 2 I, I > 0 and the value
+
is sufficiently small then the replacement of kernels I , ( @ by Zi(e) Mi(9), i = 0, 1, does not distort the asymptotic stability. Consider in detail the
124
3. Stability of Neutral Functional Differential Equations
Fig. 3.7. Stability domain of triangular wing for M
=
1.2.
onedimensional oscillation of the control surface, the so called control surface buzz [121]. Let a rigid wing move in gas flow and turn with respect to an axis. The angular wing displacements are restricted by an elastic spring of rigidity k. The motion equation is ij
+ aq + bq = kl a=
Io(0)q(t  0) d0 loW
 k mwz,
b = k,
+ k,
 k,m;,
sd
Il(0)ij(t  0) d0
k,
= ~ S 6  ~ / ( 2 5 , (2.10) )
k , = kb2/(J , u ~ ) If one considers the quasistationary model (i.e., J o ( 0 ) = 0, Jl(0) = 0), then Eq. (2.10) is asymptotically stable for a > 0, b > 0. Investigate the stability of a triangular wing with aspect ratio I = 2.5 for M = 1.2 and M = 2. We shall use the results of Astapov [7], and Belotzerkovskii [19], who give the aerodynamic coefficients and unsteady step functions. Figure 3.7 presents the stability regions obtained with the aid of the Michailov criterion for these cases. Figures 3.8 and 3.9 present the stability region of a rectangular wing with I = 5 and I = 2.5 for M = 0.4. It is clear from these figures that aerodynamic aftereffects increase the stability region for subsonic speed and decrease it for supersonic speed. The new qualitative phenomenon is also revealed. The vertical straight line crosses the stability region boundary twice: on Fig. 3.7 for M = 1.2 and 0.73k1< a < 0.79k1,and on Fig. 3.8 for 0.76kl < a < 0.64k1. It means that the wing is stable if the speed is less than a critical speed ukr or more than the other critical speed u : ~ .For the speeds varying from u,l, to the wing is unstable; it oscillates with increasing amplitude and after some time goes to ruin. This effect is observed in practice.
UK
125
42. Stability of Aeroautoelasticity Equations
Fig. 3.8. Stability domain of rectangular wing for 1 = 5 and M = 0.4.
It can be explained in the models which take into account the aerodynamic aftereffects. The twodimensional flexuretorsion flutter of a rigid rectangular wing is considered by Astapov [7], and Belotzerkosvii [19]. The critical flutter speeds are obtained: in the model with aftereffects it is equal to 290 m/sec, and in the quasistationary model it is equal to 190 m/sec. Thus, allowing for the aftereffects of the aerodynamic trace permits us to obtain more adequate models of unsteady motion (aeroautoelasticity model) which gives more precise quantative and qualitative descriptions of the unsteady motion of an elastic flying vehicle. b
Stable
0.11 Fig. 3.9. Stability domain of rectangular wing for 1 = 2.5 and M = 0.4.
126
3. Stability of Neutral Functional Differential Equations
53. LIAPUNOV DIRECT METHOD FOR NEUTRALTYPE EQUATIONS
3.1. Introduction Application of the Liapunov direct method to NFDEs is a nontrivial problem. Several assertions of this type could be obtained by formal extension of the corresponding theorems for equations with delay. It should be noted however that the results thus obtained assuming the existence of positive definite Liapunov functionals of systems’ paths have limited applications. The reason is that it is very difficult to construct Liapunov positivedefinite functionals for NFDEs. Instead of this, another method of stability investigation is stated below based on the use of only positivesemidefinite Liapunov functionals. In this case the investigation consists of two stages. At the first stage a positivesemidefinite (but not, as usual, positivedefinite!) functional is constructed. The second stage is connected with the study of stability of some functional inequalities generated by the positivesemidefinite functional constructed earlier. Some results of this section are established for equations with arbitrary (finite or infinite) aftereffect. The considered method of stability investigation of NFDE with the help of positivesemidefinite functionals was suggested by Kolmanovskii and Nosov [108( l)]. In Kolmanovskii and Nosov [108(2), 108(5)] positivesemidefinite functionals were built for concrete systems and the stability conditions formulated in terms of the systems’ coefficients were obtained. Some problems connected with the application of the Liapunov direct method to NFDEs are considered in other works [48, 81(4), 150, 1791. 3.2. Some Definitions Denote M as the metric space of continuous functions cp(t), cp: (  GO, 01 + R, and Q H is a sphere in M (see Chapter 1, $3). Let F, G: [0, GO] x QH+ RH be two preassigned continuous maps such that IG(t, CP)~ + IF(t, 9)l I F o ,
~ E C O ,GO),
~ ( Q E Q H
(3.1)
Consider the initialvalue problem for the NFDE (d/dt)[x(t)  G(t, x,)] = F ( t , x,),
xo(e)
=
cp(e),
GO
t 2
< e I 0, x, = x(t
0
+ e)EM
(3.2) (3.3)
Assume hereafter that the solutions of problem (3.2), (3.3) exists, is unique and continuously depends on initial data. Let G(t, 0) = F(t, 0) = 0. Then system (3.2) has a trivial solution corresponding to the initial function CP(e)= 0.
127
53. Liapunov Direct Method for NeutralType Equations
Definition 3.1. The trivial solution of problem (3.2), (3.3) is called ( a ) stable if for any E > 0 there exists 6(&) > 0 such that ( x ( t ,cp)( I & f o rt 2 0 as soon as p(cp, 0 ) I b(&);( b ) asymptotically stable if it is stable and lim x(t, cp)
= 0,
V cp(0) E R c M
rm
The region R is referred as the attraction region of the trivial solution.
Later, a great role will be played by functional inequalities Y,)l
=
IY(t> G(t, Y,)l If ( t ) ,
Y d 6 ) = cp(0)
(3.4)
Here f ( t ) is a nonnegative continuous scalar function and cp(6) E M . Designate by y(t, cp) any solution of inequality (3.4).
Definition 3.2. The trivial solution y(t) = 0 of inequality (3.4) is called ( a )f stable i f f o r any E > 0 there exists 6(&)> 0 such that Iy(t, cp)l I E , t 2 0,for all cp and f such that p(cp, 0 ) I &E), sup, f ( t ) I a(&); (b) asymptotically f stable if it isfstable and moreover lim,+my(t, cp) = Ofor all cp(0)~Rc M and a n y f ( t ) such that f ( t ) + 0 at t + 00.
3.3. General Theorems Let V ( t , x,, Z ( t , x,)) be some continuous functional determined for all QH and t 2 0 and such that its derivative P = dV(t, x,, Z ( t , x,))/dt exists along the trajectory of equation (3.2). Since the derivative dZ(t, x,)/dt exists and the derivative i ( t ) may not exist, the requirement of the existence of the derivative t imposes certain restrictions on the dependence of V on x,. As in Chapter 2, $5, we designate by o i ( u ) some continuous nondecreasing functions such that wi(0)= 0 and wi(u) > 0 for u > 0. X,E
Theorem 3.1. Let the functional V ( t ,x,, Z ( t , x,)) exist and satisfy the previously mentioned assumptions. Further, Wl(IZ(t9 x,)l) 5 v t , x,, a t 7 x,)) I W Z ( P ( X , ? 0 ))
P I 0
(3.5) (3.6)
Let the trivial solution of the inequality (3.4) be f stable. Then the trivial solution of equation (3.2) is stable. ProoJ: Take an arbitrary E E ( O , H ) .Since the inequality (3.4) solution g(t) = 0 is fstable, one can find bl(&)> 0 such that any solution of inequality 6,, p(cp, 0 ) I 6,. (3.4) will satisfy the relation ( x ( t ,cp)l I E , t 2 0, if f ( t ) I Now take 6 , ~ ( 0 ,6,) such that ~~(6,) = ~~(6,). Then p(cp, 0) 5 6, by virtue
128
3. Stability of Neutral Functional Differential Equations
of conditions (3.5) and (3.6) we obtain o l ( I Z ( t rx,)l) I V ( t , x,, Z ( t , x,)) I V(0,cp, Z(0, cp)) I w,(6,) = ~ ~ ( 6 ,This ) . and the monotonocity of o1imply f(t)I 6,. Taking into account the f stability of inequality that IZ(t, x,)l I (3.4), we obtain Ix(t, cp)l I E, t 2 0, for all cp(0) such that p(cp, 0) I 6, 5 6,. Theorem 3.1 is proved.
Theorem 3.2 [108(5)]. Let there exist a functional V ( t ,x,, Z ( t , x,)) satisfying requirements of Theorem 3.1 and also V I  4 1 Z ( t r x,>l>
(3.7)
Let the trivial solution of inequality (3.4) be asymptotically f stable. Then the trivial solution of equation (3.2) is asymptotically stable.
As a corollary, consider the ordinary differential equation i ( t ) = F(t, x(t)),
x(t,)
= xg
(3.8)
Here F : [O, co) x R, + R, is a continuous function satisfying the local Lipschitz condition in the second argument, and F(t, 0) = 0. Along with (3.8) consider an NFDE i ( t , x,) = ~ ( tz (, t , x,)), Z ( t , x,)
= x(t) 
X,v)
=~
(0)
G(t, x,)
(3.9)
Here G(t, x,) satisfies the conditions formulated previously.
Theorem 3.3. Let the trivial solution of Eq. (3.8) be asymptotically stable. Further, let the trivial solution of inequality (3.4) be asymptotically f stable. Then the trivial solution of (3.9) is asymptotically stable. For the proofs of Theorems 3.2 and 3.3 see Kolmanovskii and Nosov c108(5)1.
3.4. Global Stability Investigate the problem of global stability of the trivial solution of Eq. (3.2). Assume that the functionals F(t, cp) and G(t, cp) are continuous and determined on the whole space [0, 00) x M. In addition, IF(t, cp)l I F,, IG(t, cp)l I G, for all H > 0, cp eQH and t 2 0. Suppose also that the conditions of existence and uniqueness are fulfilled for an arbitrary sphere S,.
Definition 3.3. The trivial solution of Eq. (3.2) is called globally asymptotically stable if it is stable and lim x(t, cp) = 0, t + co for any initial function 440) E M . Definition 3.4. The trivial solution of inequality (3.4) is called f bounded if bounded solution y(t, cp) corresponds to every bounded function f ( t ) .
I29
$3. Liapunov Direct Method for NeutralType Equations
Theorem 3.4. Let the continuous functional V(t,x,, Z(t, x,)) exist and satisfy all the conditions of Theorems 3.2. In addition, let the trivial solution of inequality (3.4) be f bounded and asymptotically f stable, and also let ol(u)+
u+
00,
co
(3.10)
Then the trivial solution of Eq. (3.2) is globally asymptotically stable.
The proof of this theorem may be obtained by suitable modification of the results from Kolmanovskii and Nosov [108(5)].
3.5. Stability of the Functional Inequalities Establish some conditions off stability for the case in which the space M coincides with CB[  co,O]. Lemma 3.1. Let the continuous functional G(t, cp), dejined on [0, co) x CB[  co,01, satisfy the Lipschitz condition IG(4 CP)  G(t, $)I
VllV  $ 1 1 ~
(3.1 1)
If; in addition, v5 vm(t)+ v l l d e ) l l S
+ sup f ( s )
(3.13)
0,SSl
The estimate (3.13) means that the trivial solution of inequality (3.4) is f stable and f bounded. Formulate some assertions [lOS(5)] about asymptotic f stability of the trivial solution of inequality (3.4) in the case of a finite time lag. Lemma 3.2. Let G(t, y,) = g(t, y(t  h)), where g : R , x R, + R , and h > 0. Then $ Ig(t, y(t  h))l 5 v(y(t  h)(,0 < v < 1, then the trivial solution of the inequality Iy(t)  g(t, y(t  h))l 5 f ( t ) is asymptotically f stable and f bounded. Recall some concepts connected with the theory of almostperiodic functions. The spectrum of the almostperiodic function cp(t) is the set A(cp) such that A(cp) = {A:R,:M[e'Arcp(t)] # 0) t
PT
130
3. Stability of Neutral Functional Differential Equations
The module mod(cp) of the almostperiodic function cp is the set mi&, AieA(q),mi an integer, K a natural number i= 1
Lemma 3.3. Let G(t, y,) = g(t, y(t  h)), g : R , x R , + R,, h > 0, Ig(t, y(t  h))l 5 v(t)ly(t  h)l, v(t) > 0. Let the continuous function v(t) be almost periodic and satisfy the conditions V { k 1Emod(v), A # 0, [hA # O(mod 2z)]),
M[lnv(t)] < 0
Either v(t) is o periodic, h is incommensurable with o and
[ouhv(t) dt < 0 Then the trivial solution of (3.4) is asymptotically f stable and f bounded.
Lemma 3.4. Let the functional G(t, cp) satisfy the Lipschitz condition IG(t, cp)

1
4lso  11/11, 0 < < 2
G(t,
and also G(t, cp) is independent from the values of the function q(8), 8 E [A, 01. Here 0 < A < h. Then the trivial solution of (3.4) is asymptotically f stable and f bounded.
3.6. Equations with Bounded Delay Formulate the stability conditions for the case of Eqs. (3.2) with finite delay and for those with infinite delay such that the initial set coincides with the bounded interval [  h, 01. So F is a continuous functional F : [0, co) x C[  h, t ] + R,.
Theorem 3.5. Assume that the functional G(t, cp), G: [O, R, satisjes the condition IG(t + s, cp)  G(t, $)I
cx))
x C[ h, 01 +
s %(S) + ullq  11/11
O < a < 1, V q , l l / € Q H , O s t
< co, s r 0
(3.14)
Let there exist the functional V ( t ,x,, Z ( t , x,)) satisfying condition (3.5) and
i.
dIx(t)l)
Then the trivial solution of (3.1) is asymptotically stable.
(3.15)
131
$3. Liapunov Direct Method for NeutralType Equations
Proof: By virtue of Theorem 3.1 and Lemma 3.1 the trivial solution of (3.2) is stable. Hence x(t, cp) E Q H for cp(#) E Qa, 6 > 0. Show that x(t, cp) + 0, t + co. Suppose the contrary. Then there exist such a number v > 0 and a sequence of points ti + co such that Ix(ti)I > v. From (3.2) it follows that
~ ( t= ) G(t, xt)
+
1:
F(s, xS)ds
Here it is assumed that x(s) = cp(s) for s I0. Then taking into account (3.1), we have for A 2 0
I x(t + 4  x(t) I = I G(t + A, &+A) G(t, Xt) rf+A
(3.16) Designate
From (3.16) it follows that
+
p(r]) 5 (1  c t  f t l
max max j x ( ~ A)

tp(e)j+ o,(q)
+ F,q
1l2A hSOSO
1
So, due to the uniform continuity of the function x(t) on the closed segment [  h, q ] , the righthand side of this inequality tends to zero as r] + 0. Hence there exists a i j > 0 such that p(ij) 5 v/2. In this case for z E [ti  ij, ti + i j ] we have Ix(z)l > v/2 uniformly for all i. From this follows a contradiction: V + co for t + co. Theorem 3.5 is proved REMARK3.1. The assertion of Theorem 3.5 is valid for the functional G(t, xJ, G: [0, co) x C [  h , t ] + R , satisfying the condition IG(t
+ s, x,+,)

+
G(t, x,)l I o 5 ( s )
CI
max Ix(z hsrst
O 0 for x # 0 and ICI 1. Prove that the trivial solution of (3.17) will be stable. Write (3.17) in the form
+ Cx(t  h),
Z ( t , x,) = x(t)
z
=
w, w =  g ( Z )
Consider the functional
+ Iozg(s)ds
V=
(3.18)
The derivative of functional (3.18) is equal to
P = ww + i g ( Z ) = 0 Hence by virtue of Theorem 3.1 and Lemma 3.1 the trivial solution of Equation (3.17) is asymptotically stable. EXAMPLE 3.2. Derive the stability conditions for the equations X(t)
+ cp(x(t))i(t)+ f ( x ) = 0,
m, + x,)
cp(Z(t,x * > ) z ( tx,) ,
f(0)
=0
+ f(Z(t3 x,))
=0
(3.19) (3.20)
+
where Z(t, x,) = x ( t ) i[exp(sin t ) ] x ( t  1). Under the conditions xf(x) > 0, x # 0, cp(x) > 0, the trivial solution of Eq. (3.19) is uniformly asymptotically stable [lS]. By virtue of Theorem 3.3 and Lemma 3.3 the trivial solution of Eq. (3.20) is also asymptotically stable. If, in addition,
s,’/o)
ds + 00,
x + 00
then the trivial solution of Eq. (3.20) is globally asymptotically stable.
$4. CONSTRUCTION OF DEGENERATED FUNCTIONALS FOR CONCRETE SYSTEMS 4.1. Stability of a Chemical Reactor Closed by a PD Controller
Investigate the stability of the chemical reactor described in Chapter 1, $1 and Chapter 2, $6. Figure 3.10 represents the structural scheme of the investigated system. As in Chapter 2, $6 the output u(t) of a nonlinear actuator is considered as a functional depending on the preceding states of the input x(s), 0 I s I t, i.e., u(t) = F(x(t + e)), t I 8 I 0. Assume also that this functional satisfies Lipschitz condition (2.6.5) from Chapter 2, $6.
W. Construction of Degenerated Functionals for Concrete Systems
133
Fig. 3.10. Structural scheme of a chemical reactor closed by a PD controller.
The functioning of the system illustrated in Fig. 3.10 is described by the equation
Let j ( t ) be some fixed input signal, and let X(t) be the unperturbed solution. Designate the perturbed solution by X(t) + z(t). The perturbation z(t) satisfies the equation
Introduce the following functional
134
3. Stability of Neutral Functional Differential Equations
The derivatives of the functional (4.2) is equal to
V = 2 [ T ~ ( t+) KoK,DF,(z(t
+ yz’(t)
z
+ O))][z(t) + K,K,F,(z(t z2(t  z
jomd,R(s,X,, zs)  y


5
+ e))]
s)d,R(s, X,,z,)
JOrn
I
 Tz2(t)
+ K,K,(D + T ) E1 z2(t) + yr2z2(t) 
+ ~ K ; K , ~ D F : ( ~ (+~ 0)) + K , K , ( D + T)EF:(z(t  y j O r n i i (t 7  s)d,R(s,
Set E
= rl,
y = K,K,(D
VI
[2T

+ 0))
x,,zs)
+ T ) r  l + 2KgK;D. In this case we obtain + 2KoKp(D + T)r + 2KgK,2Dr2]z2(t)
Using Theorem 3.4, we see that the feedback system shown in Fig. 3.10 is asymptotically stable provided that the requirements T > KoKpDr2,
2 T > 2KoKp(D
+ T)r + KgKiDr’
hold. If D = 0, then these requirements coincide with stability conditions of a reactor closed by a P controller, obtained in Chapter 2, $6.2. 4.2. Stability of OneDimensional Nonlinear Systems
Here we investigate the conditions of stability of trivial solutions of scalar equations ~ (t S) dko(s)
+
s:
i(t
 S)
+
t 2 0, X, = ~ ( t 01,
dk,(s)
+ ~ ( tx,),
a < e I o
(4.3)
In some situations system (4.3) is a particular case of equation (3.2). Therefore, in these situations, from the results of Chapter 3,$3 one can extract some conditions of stability of system (4.3). However, owing to the specific features of Eq. (4.3), one can obtain the stability conditions under wider assumptions. These assumptions are connected first of all with the understanding of the solution of Eq. (4.3) and also with the requirement that the derivative of the Liapunov functional be negative only almost everywhere (but not everywhere, as in Chapter 3, 93). At the same time it should be emphasized that the method of stability investigation of Eq. (4.3) is the same as in Chapter 3, $3. Formulate the basic assumptions. The kernel k,(s) has bounded variation
135
&Construction I. of Degenerated Functionals for Concrete Systems
on [O, co), and the corresponding integral in (4.3) is understood in the sense of Stieltjes. The kernel k , ( s ) is determined by kl(s)
=
J
4 t ) dt
+ 1 p,,
0
kl(0)
=0
h,<s
Here summation extends to those values of n for which h, 5 s. The function A(s) is bounded and Riemann integrable on [0, 00). Finally, h, 2 0 and lpl[ Ipzl + ... + Ip,, + ... < co. The functional a(t, cp) is defined and continuous on [0, 00) x CB[  co,01. Also, the functional a(t, cp) satisfies, for any function cp E CB[  co,01, the conditions
+
a(t, 0 ) = 0,
J
r
1 4 4 cp)I2
m
lcp(S)l2
dR,(s, cp)
V cp~CB[co,O]
d R l ( s , cp) I r: < co, 0
Set
It is assumed everywhere below that
The solution x ( t ) of Eq. (4.3) is determined by the initial data
x,(8) = cp(8), i(8) = Q(O), i I0 (4.5) Here cp(8), 8 5 0 is an absolutely continuous bounded function, and the function @ ( O ) is bounded. Under these assumptions there exists the only solution x(t, cp) of problem (4.3),(4.5); i.e., there exists the only function x ( t ) limited and Riemann integrable on each finite interval, which is equal to @(O) at 8 5 0 and such that the function
i(t)
and x ( t ) = cp(0)
will be the solution of problem (4.3), (4.5). Set for each function cp E CB[  co,01 a c p )=
cpm

+
J:
i ( s ) ds, t 2 0,
j0k
dk,(s)
Notice that when requirements (4.4) are met, the trivial solution of the inequality IZ(cp)I 5 Co is f stable (see Chapter 3, 93). By analogy with the proof of Theorem 3.5 we establish Theorem 4.1. Theorem 4.1. The trivial solution of Eq. (4.3) is globally asymptotically stable, i f conditions (4.4) are fuljilled and there exists a fiinctional V(cp), V(cp) =
Wcp) + Z2(cp)
(4.6)
136
3. Stability of Neutral Functional Differential Equations
which satisfies the local Lipschitz conditions, and
0
(4.7)
W(rP) I %(llrPlle)
The derivative of functional (4.6) along the trajectories of system (4.3), exists almost everywhere for all t 2 0 and
vs %(IX(t)l) The proof of Theorem 4.1 can be found in Kolmanovskii and Nosov [108(2), 108(5 ) ] . Obtain on the basis of Theorem 4.1 some concrete stability conditions. Suppose, first, that the kernel ko(s) has a jump in zero, equal to a , > 0, and
Then the trivial solution of Eq. (4.3) is globally asymptotically stable. Consider the functional
W , )= Z’(X1) + (I+ , Moo)
jornIdk1(s)l~  s x 2 ( s lds, )
+O
Estimates (4.7) follow from the inequality j o m ~ d k i ( sl)~, X 2 ( s l ) d s 1 5 ~ 1 i ~ ~ ix= t091 ~ ~ ~ ~
The derivative of functional (4.8) along the trajectories of Eq. (4.3) is
p I2x2(t)Ca,U
 Uo1)
+ (1 + a0A(
jm
Idko(s)l
+O
+ r,)]
From this we arrive at the following assertion.
4.1. Consider the automatic control system represented by the EXAMPLE structural scheme in Fig. 3.11. Let the nonlinear element satisfy the Lipschitz conditions IW(t
+ 0))  F(y(t + 0)) I
jom
Iornd,R(s,x, Y ) I r2,
I(x(t  s)  Y(t  s)ld,R(s, x , Y )
v x, Y E QH
137
$4. Construction of Degenerated Functionals for Concrete Systems
Fig. 3.11. Structural scheme of control system
The work of the system presented in Fig. 3.11 is described by the equation
T i ( t )+ K , K , D i ( t  T) + ~ ( t+) KOK,x(t  T) + K o F ( x ( t  t + 6)) = K,y(t  t) Designate by z(t) the perturbation of the output signal corresponding to some fixed input signal j ( t ) . The function z ( t ) satisfies the equation
Applying the statement established above, we obtain that the system, shown in Fig. 3.1 1 is globally asymptotically stable if
T < K,K,D,
T  K , K , D > ( T + K,K,D)Ko(K,
+ r)
Depending on the concrete form of the systems under consideration, the functional Z ( q ) in Theorem 4.1 can sometimes be reasonably chosen in a different form. Let US formulate some results using this approach. Let
138
3. Stability of Neutral Functional Differential Equations
conditions (4.4) be satisfied, and (4.9)
Then the trivial solution of Eq. (4.3) is globally asymptotically stable. Introduce the functional VX,) = z : ( x t )
+ (P + r d
+ (1 + cll0 + xOl)r;’ + (P + r1)
dk,(s)l
jomdRl(s,cp) ~,‘s*’(tl)
jrnldko(s)l 0
p ( t d dr,
jt
dtl
dtl 11x2(t2) d t 2
(4.10)
tS
Here Z,(cp) is determined by the formula
jo
Zl(cp1 = cp(0)  jorncp(s) dk1W  jorndko(s)
s
cp(tl> dtl
From (4.9) follows the f stability of the solution of the inequality IZ,(Cp)l I C , and the estimates Z t ( 4 5 W t ) 5 clllxtll; Also, it is easy to evaluate that for almost all t > 0
V I 2x2(t)[p(1
 clol 
ale)
+ r,(l + aO1+ cclO)]
From this follows the validity of our statement. Modifying functional (4. lo), one can obtain different conditions of stability of Eq. (4.3). Consider some of them. Use the fact that an arbitrary function ko(s) with bounded variation can be represented in the form of the difference of two limited functions which d o not decrease: ko(s) = k3(s)  k4(s). In this case it is sufficient to add to functional (4.10), in which ko(s) is replaced by k3(s), P by aO3and cll0 by ~(13,the expression (l
+ “01 + “13)
+ “04 + “04
j0rndk4(s)
jt:/2(tl)
dtl
Iomdk,(s) l:2(tl) dtl
JOrndk,(J)
Jtlsdtl
11x2(t2)
139
&Construction I. of Degenerated Functionals for Concrete Systems
With the help of this new functional one can establish the following statement. Let conditions (4.4)be fulfilled and
+ M 1 3 < 1, c103(1
 MO1
+ @'I4 + c(23
0 with the aid of the functional
Under our assumptions, we get
 (1
+ Plo)r’
~ornx2( t s) dR(s)
141
@ Construction I. of Degenerated Functionals for Concrete Systems
Using the inequalities 2x(t)a(t, x,) I rx2(t)
+ rl
x2(t  s) dR(s) lom
2 4 4 x,) Iorndko(s) l  s x ( t l ) dtl I
+ r Jomdko(s) [sxz(tl)
Iornx”(t  s) dR(s)
810rl
dt,
we obtain that 3 I  y x 2 ( t ) . From the assumption Pl0 < 1 and Lemma 3.1 follows that f stability of the trivial solution of the inequality /x(t)  Jorndkds) ]t;sx(tl)
dtli 5 Cl
Thus, from Theorem 3.1 follows the stability of the trivial solution of Eq. (4.13). Some supplementary considerations show that it is asymptotically stable [108(5)]. If the kernel ko(s) is not monotone then it may be represented as a difference of two monotonically nondecreasing bounded functions, k,(s) = kl(s)  k,(s). In this case it is sufficient to add the expression El2
Jorndk1(s)
j‘ ts
+ (1 + B l l )
d t , ],:x’(t,)
dt,
]omdk2(s) jt;sx’(td dtl
to functional (4.14) in which ko(s) is replaced everywhere by kl(s). In this case the following statement holds. Let P11
UOl(1  811) > (1 + 811)(r
< 1, goo
+ clZl + u l 2 + o]*
+ aoz)
s dR(s) < co
Then the trivial solution of (4.13) is globally asymptotically stable.
EXAMPLE 4.2. Consider the equations i ( t ) = ax(t>  bx(t  h)
f ( 0 ) = 0,
If(x1)
+ f(x(t)),
t20
 f(x2)l I C1lx,  ~ 2 1 ,
a
> 0, b > 0.
(4.15)
The trivial solution of (4.15) is asymptotically stable, if bh < 1, b(1  bh) > (1 + bh)(a Cl). Notice that for a > C, the trivial solution of the equation i ( t ) = ax(t) f ( x ( t ) )is unstable. Thus, the time lag may have the stabilizing effect on the system.
+ +
142
3. Stability of Neutral Functional Differential Equations
Modifying functional (4.14) one may establish some other stability conditions. Let El0
< 1% t120
(1  .1o)Poo  (1
+ ao0 +
jams
+ cr1,)r
>0
dR(s) < cc
Then the trivial solution of (4.13) is asymptotically stable. For the proof one uses functional (4.14) in which the second integral contains Idk,(s)l instead of dK0M Consider now one example of a nonautonomous linear equation from which it is seen how general equations may be treated. Consider the equation i(t) =

b(t)x(t  h),
h >0
(4.16)
Here the function b(t) is continuous, bounded and also b(t) > 0. If
PI = sup f t O
l+h
b(s) ds < 1,
infb(t) > 0 f t O
then the trivial solution of (4.16) is asymptotically stable. The proof is based on the functional V ( t,x,)
=
[
x(t) 
jff+h
b(s)x(s  h ) ds
1
+ l  h b ( t l + 2h) d t , l : b ( s + h)x2(s)ds Some applications of degenerated Liapunov functions for study of adaptive systems, governed by ordinary differential equations, are given by Kolmanovskii and Nosov [108(10)].
$5. INSTABILITY OF NEUTRALTY P E EQUATIONS
5.1. Statement of the Problem In this section various problems of instability of NFDEs are studied. The dependence of instability conditions on the set of allowed initial disturbances is marked. We formuiate general instability theorems with the use of degenerate functionals defined on the paths of disturbed motion. It is possible to extend the classical theorem of Chetaev [39] on the systems with aftereffect and to find the selfexcitation conditions for certain distributed selfoscilla
143
$5. Instability of NeutralType Equations
tory systems. For certain equations some instability conditions are obtained. As in Chapter 2, 91 we denote M a metric space of the continuous function rp(t), rp: ( a, 01 + R , with a metric p and Q His a sphere in the space M . Let F and G be two continuous mappings and also IG(t,
CP)I + / F ( t ,CP)I5 F o ,
~ E C O , a), ~ ( Q E Q ,
(5.1)
Consider the initialvalue problem ( d / d t ) [ x ( t ) G(t, x,)]
x,(e)
= F(t, xJ,
= rp(e),
t 20
(5.2)
a < e Io
Below it is assumed that a solution of problem (5.2) exists and that it is unique and continuously depends on the initial data. Let G(t, 0) = F(t, 0) = 0. Then Eq. (5.2) has a trivial solution.
Definition 5.1. The trivial solution of (5.2) is unstable if for any positive E and 6 there exist an initial function rp(0)E M and an instant t , = t,(6) > 0 such that P(CP,0 ) < 6 but Ix(ti, ~ p ) l 2 E5.2. Influence of the Choice of an Admissible Class of Disturbances on Stability
A correct formulation of the stability problem for FDEs must also include the determination of the allowed initial disturbances for the problem under consideration. However, such an analysis depends essentially on the form of the problem studied and, in general, it is hardly realizable. In the theory of RFDE stability, the class of allowed disturbances is usually taken in the form of continuous functions. In this connection we must bear in mind that a solution which is stable under such disturbances will be stable also in cases when actually only a narrower class of initial disturbances is possible. O n the other hand, a solution that is unstable under any continuous disturbances may prove to be stable under actually existing disturbances. This shows that the instability criteria are relatively inferior. We now present an appropriate example. Consider an RFDE which is stable under disturbances actually feasible in such a system and unstable in the case of arbitrary continuous disturbances. In Shimbell [206] the following model was used to describe the behavior of the central nervous system during learning i ( t ) = k [ x ( t )  x(t  I)]"
~(t)],
t>1
(5.3)
i ( t ) = kx(t)[N  ~ ( t ) ] 0I t I 1, ~ ( 0=) x0, 0 < x0 < N , k > 0
(5.4)
144
3. Stability of Neutral Functional Differential Equations
Examine the stability of a solution x(t) = N of Eq. (5.3) [108(8)].As the class of admissible disturbances of the initial function in the interval [O, 11 let us take a set of solutions of Eq. (5.4) which have the form x ( t ) = Nxo[xo
+ ( N  x o ) exp(

kNt)]
(5.5)
From (5.5) it follows that any solution x(t) of Eq. (5.4) is increasing for t E [0, 1 ) . Any solution of Eq. (5.3) which corresponds to the initial function (5.5) in [0, 11 is not decreasing in [0, co),in addition remaining smaller or equal to N . Actually, the solution x(t) cannot cross the line x = N , since in this case i ( t , ) = 0 at the point t,, where x ( t l ) = N ; and for t 2 t , we should have x ( t ) = N . Further, by virtue of (5.3) we have i(1) > 0, since x(1) > x(0) and N > x(1). This means that the solution x ( t ) increases to the right of the point t = 1. Let t , be the first instant such that i ( t , ) = 0. Then either x ( t l )  x ( t ,  1) = 0 or N  x ( t l ) = 0. In the first case, because of Rolle’s theorem there exists a point t , < t , such that i ( t , ) = 0. But this contradicts the assumption that t , is the first moment at which i ( t ) = 0. In the second case, x ( t ) = N for t > t,. Hence the solution x(t) = N of Eq. (5.3) is stable under arbitrary initial disturbances (5.5). We can show, however, that the solution x(t) = N will be unstable if the allowed initial disturbances in [0,1] are taken in the form of space C[O, 13. Take a twice continuously differentiable function cp(8) such that llcp(8)  NII I 6, cp(1) < N , @(8)< 0, (p(8) < 0, 0 I 8 I 1. The derivatives i(1) and x(1) are negative on the basis of (5.3). Hence the solution x(t, 1, cp) decreases to the right of the point t = 1. Show that x(t, 1, cp) +  co for t + co. Assume the contrary. Then there must exist a first point t , > 1 such that x(t,) = 0. But x(t) < 0 for t E [ l , t,). Therefore N  x ( t l ) > 0, i(t, > 0, x(t)  x ( t ,  1 ) < 0, i ( t , )  i ( t l  1) < 0. Hence x(tl) = k [ i ( t , )  i ( t ,  1)][N  x(t,)] k[x(t,)  x(t,  l ) ] [  i ( t , ) ] < 0. The contradiction that results shows that x(t, 1, cp) < 0 for 1 It < 00. In addition, it is obvious that x(t, 1, cp) +  co for t + co.Thus the solution x(t) = N of (5.3) will be unstable if allowed initial disturbances coincide with the space C [  h, 01.
+
5.3. Instability Conditions Consider a continuous functional V ( t ,x,, Z(t, x,)), where Z(t, x,) = x(t) G(t, x,). The region V > 0 is a connected open region in the product [0, co) x QH which is bounded by the surface V = 0 and in which the functional V takes only positive values. Theorem 5.1. Let there exist a continuous functional V ( t )= V( t,x f ,Z(t, x,)) such that for t = 0 the region V > 0 has an open section whose boundary
145
$5. Instability of NeutralType Equations
contains the element cp(0) = 0. Further, in the region V > 0 let the following conditions take place for Eqs. (5.3), (5.4),( 5 . 3 , respectively,
Then the trivial solution of (5.2) is unstable. Proof: Let 6 > 0 be arbitrarily small. There exists an initial function cp(0) such that P(cp(@, 0)
6,
V(0) = V(0,cp, Z(0, cp))
=a
>0
(5.6)
Show that a solution x(t, cp) of (5.2) with such an initial function leaves the sphere QH for t + co. The solution x(t, cp) cannot leave the region V > 0 by passing through the part of the boundary on which V = 0. In fact, if t , > 0 is such that V ( t , ) = 0,
V ( t )> 0
for 0 I t < t ,
(5.7)
then by virtue of (5.4) it follows that P ( t ) 2 0, 0 5 t < t , . Hence V ( t ) 2 V(0)= a > 0, and by continuity V ( t , ) 2 V(0)= a > 0, which contradicts (5.7). Hence the solution x(t, cp) cannot leave the region V > 0 across the part of the boundary on which V = 0. But the solution x(t, cp) cannot always remain in the region V > 0. Assuming this, we would successively obtain V ( t ) 2 0,
V ( t ) 2 V(0)= a > 0,
P ( t ) 2 B(a) > 0
(5.8)
We find by integrating (5.8) that V ( t >xt, Z ( t , x,)) 2 V(0,cp, Z(0, cp))
+ @(a>
(5.9)
If t b ' ( a ) [ q ( H ) a ] , then inequality (5.9) contradicts condition (5.3). So the solution x(t, cp) must leave the region V > 0 through the part of the boundary on which p ( x t ,0) = H . This completes the proof of Theorem 5.1.
REMARK5.1. Suppose that we can select a subregion D which belongs to the region V > 0 and such that all the conditions of Theorem 3.1 are fulfilled in D. Suppose also that x(t, cp) E D for all t > 0. Then the assertion of Theorem 3.1 is valid. A region D which satisfies the preceding conditions is called a sector C1911.
146
3. Stability of Neutral Functional Differential Equations
EXAMPLE 5.1. Consider the equation i(t) ci(t
c 2 0,
a(t) 2 0,

z) = a(t)x(t)
b(t) 2 0,
a(t)
+ b(t) x
( t  z)
+ b(t) 2 A > 0,
z >0
(5.10) (5.1 1)
Introduce the functional V ( t ) = Z ( t , x , ) = x(t)  cx(t  z). Choose the initial function cp(8) such that cp(8) > 0, IIcp(0)ll I 6, Z ( 0 , cp) = cp(8)  ccp(z) = ct > 0. Show that x(t, cp) > 0 for t > 0. Assume the contrary and let t , > 0 be the first instant such that x(t,, cp) = 0. For 0 < t < t,, Z(t, x , = i ( t )  c i ( t  z)
= a(t)x(t)
+ b(t)x(t  z) > 0
Hence Z ( t , x , ) 2 Z ( 0 , cp) = ct > 0. By going to the limit for t + t ,  0, we obtain Z(t,, x , , ) 2 ct > 0. Therefore x ( t l ) = Z ( t , , x,,) c x ( t ,  z) 2 c1 > 0. The latter inequality contradicts the preceding assumption. Thus for any t > 0 we get v ( t )= a(t)x(t) b(t)x(t  z) 2 Act > 0. By virtue of Remark 5.1 the trivial solution of (5.10) will be unstable under conditions (5.11). In the same way it is possible to prove that the trivial solution of the equation i ( t )  c i ( t  z) = g(x,),g(0) = 0, c > 0 is unstable if the functional g(x,(8))> 0 for min, x,(8) > 0, z I 8 I 0.
+
+
5.4. Connection between Instability of NFDEs and Ordinary Differential Equations
As a consequence of Theorem 5.1, investigate the connection between the instability of an ordinary differential equation ~ ( t ,= ) x,,
i ( t ) = F ( t , x,),
to 2 0 , F(t, 0) = 0
(5.12)
and an NFDE i ( t ,X J
=~
( tz(t, , xt)~
xO(0) = ~p(0)
G(t, 0) = 0
Z(t, x , ) = ~ ( t) G(t, x,),
(5.13)
Theorem 5.2. I f the trivial solution of ordinary equation (5.12) is unstable then the trivial solution of N F D E (5.13) is also unstable. Prooj By the inversion theorem for Chetaev’s theorem [114(5)] there exists for (5.12) a function W ( t ,x ) which satisfies the conditions of Chetaev’s theorem. In particular, by virtue of (5.12)
aw
W(t,x ) = __ + at
1 aw axi n
~
i=l
i i
= @(t,x )
(5.14)
Now consider the functional W ( t ,Z ( t , x,)). Using (5.14) and (5.13) we find that W ( t ,Z ( t , x , ) ) = @ (t, xr)). The ‘validity of the other assumptions of
147
$5. Instability of NeutralType Equations
Theorem 5.1 about the functional w ( t , Z(t,x,)) is obvious. Hence the functional W(t,Z(t, x J ) satisfies all the conditions of Theorem 5.1. SO the trivial solution of (5.13) is unstable. EXAMPLE 5.2. Consider the equation (d"/dt")[x(t)+ cx(t  T)]
=f
(x(t)
+ cx(t  7))
(5.15)
Let c be a constant, n 2 3, and let y = 0 be an isolated root of the equation f ( y ) = 0. Then the trivial solution of the equation y"(t) = f ( y ) is unstable ([191], p. 148). According to Theorem 5.2, the trivial solution of (5.15) is also unstable.
5.5. Other Instability Conditions Denote q ( u ) as scalar nondecreasing functions of the argument u 2 0, such that q ( 0 ) = 0 and q ( u ) > 0 for u > 0. Theorem 5.3. Let all the conditions of Theorem 5.1 hold except conditions (5.3) and (5.9, which are replaced in the region V > 0 by v t , x,, Z(t, x,)) 2 %( IZ(4 x,) I
(5.16)
P(t, x,, a t , x,)) 2 %(lZ(t, x,)l
(5.17)
Then the trivial solution of (5.2) is unstable. Proof: Choose any initial function cp satisfying the conditions (5.6). By an argument similar to that made in the proof of Theorem 5.1, it is easy to verify that for this initial function the solution x(t, cp) cannot leave the region across that part of the boundary at which V = 0. However, the solution x(t, cp) cannot remain all the time in the region V > 0. Otherwise V ( t ) 2 0,
~ ( t2) ~ ( 0=) a > O
Take B > 0 such that w2(B) = a. Then 0 2 ( B ) = a = V(0)I V ( t ) I 0 2 ( I Z ( t ,x,)l). Hence /? I IZ(t, x J . Consequently, (5.18)
V ( t ) 2 w3(B) > 0
Integrating (5.18) we obtain a contradiction with the boundedness of the V ( t ,x,, Z(t, x,)) in the sphere Q H :V(r, x,, Z(t, x,)) I functional co2(]Z(t,x,)l) 5 u 2 ( H Fo). Theorem 5.3 is proved.
+
EXAMPLE 5.3. Given the equation (d/dt)[x(t) g(t, ~ (t h))] = U X ( ~ ) , g: R, x R,
+
R,
u > 0, h > 0
(5.19)
I48
3. Stability of Neutral Functional Differential Equations
Let V(t,x,, Z(t, x,)) = Z2(t,x,)  a
L
x2(s)ds (5.20)
Z(t, x,) = x(t)  g(t, x(t  h))
Functional (5.20) satisfies condition (5.16). Further if gz(t, x2(t  h), then
x ( t  h)) I
P = 2ax(t)Z(t,x,)  axZ(t) + uxZ(t  h) =d
+
( t )  2ax(t)g(t,~ (th)) + &(t, ~ (t h)) agz(t,x(t  h)) &(t  h) 2 aZ2(t,x,)
+
Hence by virtue of Theorem 3.3 the trivial solution of (5.19) is unstable. 5.6. Instability of Equations with Bounded Delay
Investigate Eqs. (5.2) with finite or infinite delay but such that the initial set coincides with the interval [  h, 01. So F and G are continuous functionals mapping from [0, 00) x C [  h, t ] into R,. Theorem 5.4. Let there exist a continuous functional V(t,x,, Z(t, x,)) and the region V > 0 satisfy the conditions of Theorem 5.1. Let in the region V > 0 V(4
IG(t
XI,
(5.21)
Z(4 x,)) I w4(11x,ll)
+ s, x,+,)  G(t, x,)l I w5(s) + v
max
IX(T
+ s)  x(z)l
(5.22)
hsrst
Then the trivial solution of (5.2) is unstable. Proofi The solution x(t, 9)cannot leave the region V > 0 through the part of the boundary V = 0. Show that the solution x(t, cp) cannot remain in the region V > 0. Really, let the initial function satisfy (5.6) and p > 0 such that w&) = a. Then w4(p) = CI = V(0)I V ( t ) I w4(llxtll). Consequently, there exists a sequence {ti} such that t i + 00 and Ix(ti)I = p. Show that the function x(t) = x(t, cp) is uniformly continuous in [  h, m). By using (5.1), (5.2) and (5.22) for any A > 0, we get Ix(t
+ A)

x(t)l = IG(t
+ A, x,+J

G(t, x,)
Denote p(q) the function p(q) = sup suplx(t r>O
Asv
+ A)

+
149
4.5. Instability of NeutralType Equations
From (5.24) and (5.22) it follows that (1  v)p(q) 5 max max Ix(0
+ A)  cp(0)l + qFo + w5(q)
ASq h5BSO
The righthand side of this inequality tends to zero for q + 0 due to the uniform continuity of x(t, cp), t E [  h, q]. Hence the uniform continuity of x(t, cp), t E [  h, 00) is proved. Therefore there exists an q such that Ix(s) 1 2 p/2 for ti  q I sI ti q and any i. Integrating (5.23) we obtain for t + co
+ V2u +
s:
o ~ ( ( x ( s ) ( )ds 2
O!
+ 2N(t)~,(p/2)
where N ( t ) is a number of points ti such that ti E [0, t ] . Theorem 5.4 is proved.
EXAMPLE 5.4. Obtain the instability conditions for a scalar NFDE
s : s:
i ( t ) = aOx(t)
k,(s)
=
+
A(t) dt
1:
~ ( t S ) dkO(s) +
+ 1 pn,
i ( t  S)
dk,(s) + ~ ( tx,) ,
(5.25)
h, 2 0
h,st
Here the continuous functional a(t, cp) is defined on [0, co) x CB[  co,01 and also 144 @ I 2
0 ) = 0,
johla(s)12W s )
Kernel R(s) is nondecreasing and bounded. Introduce a functional V,
Here Ph
r2
=
J
Ph
dR(s), 0
Z ( t , x,) = x ( t ) 
aoi =
c
J 0 Idk,(s)I
x ( t  s) dk,(s)
Relations (5.25), yield v(t)2 2x2(t)[ao(l  sol)  (1 + aO1)(r+ aoOl.By virtue of Theorem 5.4 the trivial solution of (5.25) is unstable under condition
150
3. Stability of Neutral Functional Differential Equations
+
+
a,(l  sol) > (1 aol)(r aoo). By using the methods of functionals I/ construction described earlier in this book, it is possible to obtain other conditions of instability of system (5.25).
5.7. Distributed SelfOscillatory Systems Consider the distributed lossless system from Chapter 1, $2 described by the equation
_1 _d [ ~ ( t) K x ( t  T)] C , dt
1 K 21 t =  (5.26) ~ ( t) ~ ( t t)  g(x(t)) + K g ( x ( t  T)), Z Z S Assume that the nonlinear element g( V ) has a negative differential resistance at certain points. Such an element can be the tunnel diode whose currentvoltage curve is plotted in Fig. 3.12 [147]. Find the conditions of selfexcitation of oscillations which are equivalent to the conditions of instability of the solution x ( t ) = uo. The value uo satisfies the equation 0 = CI Z  ' u ,  K Z  ' u ,  g(u,) Kg(u,). Denote y ( t ) = x ( t )  u,. Equation (5.26) in the neighborhood of uo can be written in the form = CI

~
+
= (S  Z  ' ) y ( t ) 
K(S
+ Z  ' ) + O(y(t))+ O(y(t

T)) (5.27)
where S is the steepness of the currentvoltage curve at the point u,; i.e., s = g(uo). Show that the condition of selfexcitation of the distributed
"0
Fig. 3.12. Currentvoltage curve.
V
96. Asymptotic Properties of NeutralType Equations
151
selfoscillatory system represented in Fig. 1.12 has the form S > K 1 .Study for this purpose the directly nonlinear equation (5.27). Take an E > 0 so small that S > (1
+ K + 2~)(1
K)'Z'
[(K(S
+ Z)')

(5.28)
and also in the sphere Q
2 ( S  2'

E)y(t)


~ ] y (t T)
(5.29)
Consider the functionals I/=
y ( t )  Ky(t  t) [ K ( S
+ 2)' + E ]
LZ
Y(S)
ds
(5.30)
The region B = (I/ > 0 ) n ( y > 0) n Q, is a sector. The trajectories of system (5.26) cannot cross the boundary V = 0 because
V
=
[IS  2'
+ K ( S + 2  ' ) + 2~]y(t)2 0.
In ,a manner similar to Example 5.1, we conclude that y ( t ) > 0 for Otherwise, if y ( t J = 0, then V ( t ) 2 V(0)2
ff, y(t1) =
V(t1)
t
> 0.
+ K y ( t ,  t)
+ [ ( K ( S + Z)' +
E]
J;&y(s) ds.
Hence the region B is a sector in which i/ 2 0. By virtue of Remark 5.1, the solution x(t) = uo is unstable.
$6. ASYMPTOTIC PROPERTIES OF NEUTRALTYPE EQUATIONS
6.1. Some Introductory Remarks One of the important problems in the qualitative theory of NFDEs is the study of the asymptotic behaviour of its solutions. Asymptotic behavior for t + a3 of NFDE solutions may be different. They can tend to stationary or periodic solutions or their behavior can be more intricate. Neutral functional differential equations have properties that are similar to the properties of difference equations for some cases and properties that are similar to the
152
3. Stability of Neutral Functional Differential Equations
properties of differential equations for other cases. This duality of NFDEs is essential for a study of the asymptotic behavior of their solutions. In this section some results of both types are described. The method of investigation that we use is based essentially on degenerate Liapunov functionals. It is possible with the aid of degenerate functionals to separate the cases when the asymptotic behaviour of NFDEs is similar to that of difference equations from the cases in which this is not valid.
6.2. Basic Definitions Consider the NFDE [108(7)] Z ( t ) = F(t, Z),
Z ( t ) = x ( t )  G(t, X J ,
t20
(6.1)
Here F and G are given continuous mappings, F(t, x), F : [0, co) x R, + R,; G(t, xt), G : [0, co x C [  h, 01 + R,. The solution of (6.1) is determined by the initial data xo = cp,
cpEcCCh,01
(6.2)
Let lF(t, x)l I M , for 1x1 IH and IG(t, cp)  G(t, $)I I K,llq  $11 for llcpll I H , Il$l/ I H . Assume that there exists a unique solution of problem (6.I), (6.2). Scalar equations (6.1) named completely integrable are considered in Sharkovskii and Romanenko [200]. Let M c C [  h, 01 be a closed set. The distance p between the element cp E C[  h, 01 and the set M is equal to p(cp, M ) = min, 11 cp  Y 11, Y EM . A set M c C [  h, 01 is called an inoariant set for Eq. (6.1) if for arbitrary cp E M , the solution xt(cp)E M .
Definition 6.1. The closed invariant set M of Eq. (6.1) is called stable if for any E > 0 there exists 6 > 0 such that p(x,(cp), M ) E, t > 0,for cp: p(cp, M ) < 6. ZA in addition, lim p(x,(cp), M ) = 0, t + co,V cp ER, then the set M is called asymptotically stable and the set R is called the attraction domain of the set M . Stability of invariant sets is investigated with the aid of the functional inequality Iz(t, Yt)l
=
IY(t> G(t, Yt)l I f ( t > ,
fER1
cpEah,Ol Let M , c C [  h, 01 be a closed invariant set of equation Yo(cp> = cp?
a t ,YJ
= Y(t)  G(t, Y J = 0
(6.3) (6.4)
(6.5)
Definition 6.2. The closed invariant set M of functional Eq. (6.5) is called f stable if for any E > 0 there exists 6 > 0 such that for all solutions of inequality (6.3),(6.4) we have p(y,, MI) < E ifp(cp, M , ) 5 6 a n d f ( t ) 5 6. Thefstable set
153
56. Asymptotic Properties of NeutralType. Equations
M , is called asymptotically f stable if lim p(y,(cp), M , ) = 0, t + 03, for all cp E IR c C [  h, 01 and for all f ( t ) such that f ( t ) + 0, t + 00.
For the case in which the sets M and M , contain the only zero element, Definitions 6.1 and 6.2 coincide with Definitions 3.2.
6.3. Stability of Invariant Sets Theorem 6.1. Let the trivial solution of the ordinary diflerential equation (6.6)
i ( t ) = F(t, x(t))
be uniformly stable (uniformly asymptotically stable). Further assume that the bounded invariant set M of (6.5) is f stable (asymptotically f stable). Then M , will be a stable (asymptotically stable) invariant set of Eq. (6.1).
,
ProoJ By virtue of the inversion theorem [114(5)] there exists for Eq. (6.6) a Liapunov function V(t,x) such that ol(lxl) I v(t,
s 02(IxI),
p(6.6) I O,
cv6.6)
I 03(Ixo]
Here q ( u ) are continuous scalar nondecreasing functions such that oi(0) = 0, q ( u ) > 0 for u > 0 and e 6 . 6 ) is the derivative of the function V along the trajectories of (6.6). Now consider the functional V(t,Z(t, XJ). Clearly, o l ( I Z ( t ,x,)l) IV(t, Z(t, x,)) I m2(1Z(t,x,)l), p(6.1) I 0. Assume any E > 0. From the f stability of the set M , follows the existence E for p(cp, M , ) I 6 and f ( t ) I 6. Choose of a 6 > 0 such that p(y,(cp), M , ) I 6, > 0 such that ~ ~ ( 6K,6,) , I o,(S). Let p(cp, M , ) = 1140  $11 I 6,, llcpll I H and 11$11 I H. Since Z(t, $) = 0 for $ E M , ,
+
IZ(t9 cp)l = IZ(4 $ + (cp  $111 I IZ(t, $)I + IZ(t3 $ + (cp  $))  Z(t?$11 5 (KH
+ 1Pl.
Using the properties of functional V(t, Z(t, xt)), we have
w,a t , 4)IV(O,Z(O,cp))
w,(lZ(t, 41)I
+
I ~ 2 ( 6 , KH6,) I~ , ( 6 )
The monotonicity of the function o l ( u )implies that lZ(t, xt)1 I 6. But the set M , is f stable. Hence p ( X , , M , ) IE ; i.e., the set M , is stable. Similarly, asymptotic stability of the set M , may be established (see Theorem 3.3). Theorem 6.1 is proved.
EXAMPLE 6.1. Consider the scalar equation i ( t ) = f (Z(t)>,
z ( t ) = x ( t )  ax(t  1)[1  x(t  I)] 1 < a < 3, t 2 0
(6.7)
154
3. Stability of Neutral Functional Differential Equations
Assume that the trivial solution of the equation i ( t ) = f ( x ( t ) ) , f ( 0 ) = 0, is asymptotically stable. For Eq. (5.7), difference Eq. (6.5) has the form x(t) = ax(t

1)[1

x(t  l)]
(6.8)
Equation (6.8) was studied in Sharkovskii and Romanenko [200]. For 1 < a < 3 there exists two stationary solutions of (6.8): x , ( t ) = 0 and x 2 ( t ) = 1  a  l . Denote u(t) = x ( t )  x2(t). Then by using (6.8) we get
I u ( t )  (2  a)u(t  1) + auyt lu(t)l I 12  a1 lu(t  s)l

+ ad(t
1)1 I f ( t )

1)
+f ( t )
(6.9)
Let E > 0 be sufficiently small. Choose 6 ~ ( 0E ) , such that 12  a16 + + 6 I E. Remark that such a 6 always exists for 1 < a < 3 and sufficiently small E. For these E and 6 from (6.9) it follows that for cp, llcpll I 6 and J; llfll I 6, lu(t)l I 12  al lcp(t  1>1+ acp2(t  1) + f ( t ) I 12  a16 a62 6 I E, 0 0. The set of such solutions is designated by N ( H ) .
Definition 6.3. The set N ( H ) is f unstable if for any E > 0 there exist cp, II 40 /I 5 H and = q ( ~such ) that I Z(q, x,(cp))  Z(q, Y,) I = I Z(q, x,(cp) I 2 E. If the set N ( H ) is f unstable, then in general it does not define the asymptotic properties of the solutions of Eqs. (6.1). Theorem 6.2. Let the trivial solution Eq. (6.6) be unstable. Then the set N ( H ) is f unstable.
155
96. Asymptotic Properties of NeutralType Equations
Prooj According to the inversion theorem [114(5)], there exists for Eq. (6.6) a function V ( t , x ) satisfying the conditions of the Chetaev theorem. Consider the functional V(t,Z(t, x,)) = V(t).There are elements cp with an arbitrarily small norm in the domain V > 0. In this domain the conditions $ 6 . 1 ) 2 0 and 2 p > 0 if V ( t ) 2 c1 > 0 are fulfilled. In addition, the solution x,(cp) leaves the set N ( H ) . In fact, if V(O,Z(O,cp)) = c1 > 0, then $6.1)(t) 2 0. Hence V ( t )2 V ( 0 )= c1 > 0 and $6.1)( t ) 2 b > 0, V ( t ) 2 P t + 00 for t + 00. But V(t,2)is continuous and bounded in Z for all t . Consequently, the functional 2 cannot be bounded for all t . So Z ( t , x , ) will be greater than any arbitrary E > 0 as soon as t is sufficiently large. Theorem 6.2 is proved.
e6.1)
EXAMPLE 6.2. Consider once more Eq. (6.7). Assume that Eq. (6.7) is unstable and 0 < a < 1. Then there exists the unique asymptotically stable solution y ( t ) = 0 of (6.8). But for NFDE (6.7) this solution will be unstable. So in the considered case, the behaviour of solutions of difference equations (6.8) and NFDE (6.7) are quite different. 6.5. Interconnection between Functional and Differential Equations Sometimes the asymptotic properties of solutions to NFDE (6.1) are determined both by the behaviour of solutions of functional equations (6.5) and by the behavior of differential Eq. (6.6). Consider Eq. (6.1) with Z ( t , x,) = x ( t )
+ q(t, x(t  l), x(t  2), ... , x(t  p ) )
x(t)ER,,
4 : R, x
... x R , + R ,
(6.11)
Here q is a continuous function and Iq@,x(t  11, x(t  21,. . . x(t  P))l I Ix(t  1)1 ... J x ( t p)J 7
+ +
(6.12)
Theorem 6.3. Let the trivial solution of E q . (6.6) be globally exponentially stable and condition (6.12) be fuEfilled. Then all solutions of E q . (6.1) with Z ( t , x,) given by (6.1 1) are bounded for t 2 0 and its trivial solution is stable. Pro05 Global exponential stability of the trivial solution of Eq. (6.6) means that there exist two constants C ( x o ) > 0 and c1 > 0 such that Ix(t, xo)l I C(x0)eFa',
V xo
(6.13)
So any solution of (6.1) by (6.13) satisfies the estimate
Iz(t,x,(cp>>lIC ( Z ( 0 ,cp>>e'' 5 Cl(llcpll)e"',
C,(llrPll) I C(P11cpIl)
(6.14)
156
3. Stability of Neutral Functional Differential Equations
From (6.14) and (6.1 1) it follows that
I4t) + d t , x ( t

I), . . . , x ( t  p))l
s Cl(llcpll)ear
Hence I Ix(t

+
+
+
Ix(t  2)l .... Ix(t  p)I + Cl(llcpll)eat 211 + ... + Ix(t  p)l + Ix(t  p  1)l
Ix(t)l I lx(t  1)l
+ C,(llcpll)[e"' + e"('')] I Iq(t  N ) ( + ... + Iq(t  N  p ) ( + cl(llcpll)[ea'
+ e  Q ( t  1 ) + ... + e"('N+1)
1
N . Designate y Here the number N is such that N  1 < t I Then inequality (6.15) takes the form Ix(t)l I PIlcplI + cl(llcpll)c~Nl +YN2
expCa(t  N
+ 111 5 PIIcpII + C,(llcpll)(l
 y1l
(6.15)
= exp(  a )
+ ... + 11
< 1.
(6.16)
So all solution of (6.1) are bounded. But estimates (6.16) is uniform in cp with llcpll I H.Therefore from (6.16) follows the stability of the trivial solution of Eq. (6.11). The Theorem 6.3 statement is also valid if condition (6.12) is exchanged for the following one
Id t , x ( t
1)), x ( t  2), . . . x(t  P) I I K[Ix(t  1)1 + ... + Ix(t  p)[], 
9
K exp(a) < 1 (6.17)
For K > 1 the trivial solution of the difference equation x ( t ) + q(t, l), . . . , x(t  p ) ) = 0 is in general unstable in the Liapunov sense. Condition (6.17) coordinates the order of this instability with the speed of convergence to zero of all solutions of ordinary differential Eq. (6.6). In addition, all solutions of NFDE (6.1) are bounded and the trivial solution of (6.1) is stable. x(t 
57. LIAPUNOV FUNCTIONALS DEPENDING ON DERIVATIVES
7.1. General Stability Theorem Let us denote by W a normed space of absolutely continuous functions cp(f3) defined on [  h, 01 with square integrable derivatives. Define the norm in W
IIcp11~= cp
157
97. Liapunov Functionals Depending on Derivatives
Denote QH a sphere in W :QH = {cp(B)): IIcp(0)llw 5 H } . Consider the following NFDE i ( t ) = f (t, x,, if), x,,(e)
=
cp(e),
x(t)ER,,
t 2 to
(7.1)
h I e Io
i f ,= @(e),
Let the map f : R , x Q H+ R , be continuous and satisfy the Lipschitz condition in the second and third arguments, and let the Lipschitz constant in the third argument be less than 1. Then, as in Theorem 1.3.2, one can prove that for any cp(0)E W there exists a unique solution of NFDE (7.1) such that x,(e)E Wfor all t 2 to. Suppose that f(t,O,O)=O,
If(t>cp,+,)l2M,
IIcpllwI~,
tER
(7.2)
Definition 7.1. The trivial solution x(t) = 0 of Eq. (7.1) is said to be stable if for any E > 0, t o e R, there is a 6 = b ( ~to) , such that ( ( c p ( l w 2 6(e, to) implies x(t, to, q)l 5 E. The solution x(t) = 0 is said to be asymptotically stable if it is stable and q(t)ER(to)c Wimplies x(t, to, cp) + 0 as t + co.The domain Q(t,) is called an attraction one at moment to. Other definitions of stability can be introduced in a manner analogous to that presented in Chapter 2, $1. Let the functional V(t,cp, @)becontinuous in the sphere Q H and satisfy the , x(t) is a Lipschitz condition in cp and @. Denote by V ( t ) = V(t,x,, i f )where solution of Eq. (7.1). Suppose that the functional V ( t ) is absolutely continuous in t for any solution x(t) of Eq. (7.1). We denote P(t) the upper rightalong the solution of Eq. (7.1): hand derivative of V ( t ,x,, i f ) V(t)=
G
[V(t + At)  V(t)]/(At)
A,+ + O
Denote wi(u) continuous nondecreasing functions such that oi(0)= 0, wi(u) > 0 for u > 0.
Theorem 7.1. Let the continuous functional V ( t ,x,, i, exist ) and satisfy conditions ~ l ( I l X , l l W )2
V t , x,, if)I~ 2 ( I I X , I I w )
V ( t )I 0
(7.3) (7.4)
Then the trivial solution of Eq. (7.1) is stable. Zf P(t) I Wj( Ix(t)I) then the solution x(t) = 0 is asymptotically stable.
(7.5)
158
3. Stability of Neutral Functional Differential Equations
Pro05 Assume E > 0. Take h(e) such that wl(&) = w2(h).From (7.3), (7.4) it follows for llcp(0)llw I 6 ( ~ that )
~l(lIX,IlW)I V ( t )I V ( t 0 )
~2(IlcpIIw) 2 w 2 ( 4 = W,(E)
Ilx,ll$ 2 E ~i.e., ; Ix(t)l I E.This proves stability. To This implies that Ix(t)12 I prove asymptotic stability, choose H , such that w,(H) = w 2 ( H , ) .Then, as H previously demonstrated, \lcpllw I H , and inequality (7.2) implies ( I X , I ~ ~ I and l i ( t ) ( I M . Suppose that a solution x(t, to, cp), Ilcpllw I HI does not tend to zero as t + 03. Then for some E > 0 there exists a sequence ti co,i 03, such that Ix(t, to, cp)l 2 E. Therefore, on the intervals ti  A I t 2 t i A, A = &(2M)’ we have /x(t, to, cp)l 2 4 2 and k(t) I a3(&/2) = y < 0. Therefore, +
c j,LA k(s) d s
+
+
ti+A
V ( t )  V(t0)=
p(4ds 2
2 2y A N ( t )
t,  co as t + co. This contradicts inequality (7.3). The proof of Theorem 7.1 is complete.
7.2. Examples Consider the Eq. [l50] i(t) =
ax(t)
+ c(t)i(t

h),
t20
(7.6)
where c(t) is a continuous function. Define the functional [l50] V ( t , x,, i t )= x 2 ( t )
+
For p ( t ) we get V ( t ) = 2x(t)i(t)
+ a ‘[i’(t)
 i 2 ( t  h)]
Replacing i ( t ) according to (7.6) we obtain for Ic(t)l < 1 that P ( t ) = ux2(t)  a1[1  2(t)]i2(t

h) I ux2(t)
Therefore, the functional V ( t , x,, i t )satisfies all the conditions of Theorem 7.1. This implies the asymptotic stability of the solution x(t) = 0. For other examples see El’sgol’tz and Norkin [59] and Misnik [lSO]. Castelan and Infante [35], for a matrix autonomous NFDE i ( t ) = AX([) f
Bx(t  h) + C i ( t  h),
X(t)ER,
(7.7)
have constructed the functional V ( t ,x,, i , )satiisfying the conditions of Theorem 7.1 under the assumption that the trivial solution of (7.7) is
159
48. Stability and Boundedness of Linear Nonhomogeneous Equations
asymptotically stable. A Liapunov functional was obtained as the limit, in an appropriate sense, of a Liapunov function constructed by wellknown methods for a difference equation approximation of the original NFDE (7.7).
88. STABILITY AND BOUNDEDNESS OF LINEAR NONHOMOGENEOUS EQUATIONS 8.1. Connection between Stability and Boundedness Consider the NFDE
Here the operator L(t,x,, i t ) = A,(t, i t ) + A2(t,x,) is linear in the second and third arguments. Assume that for problem (8.1) conditions of existence, uniqueness and continuous dependence of solutions on initial data are met.
Theorem 8.1 [253(5)]. Zf any solution of problem (8.1) with f ( t ) = 0 is bounded for t 2 0, then its trivial solution is stable. Notice that, in general, the boundedness of all the solutions of a nonlinear equation does not imply their stability (see Example 2.1.1). Let the operators A,(t, i t ) and A,(t, x,) be bounded and o periodic in t. Consider functional equation z(t) = A , ( t , Z t )
+ g(t),
t > 0,
z(e>= 0, e I o
(8.2)
Problem (8.2) is said to satisfy the Perron condition if its solution defined by
Theorem 8.2 [162(13)]. Let operator L(t, x,, iJ be linear in x,, 2, and periodic in t. Assume that problem (8.2) satisjes the Perron condition and that Eq. (8.1) has a bounded solution for the functions f ( t )E L,[O, 00) and q(0) = 0. Then the trivial solution of Eq. (8.1) is asymptotically stable. 8.2. Boundedness of Derivatives Consider an ordinary differential equation with bounded coefficients
+
X(")(t) A,(t)x'"''(t)
+ . . . + A&)
x (t)=f ( t ) X(m) =
It is known the following Esclangon's theorem [36].
d"x/dt"
(8.3)
160
3. Stability of Neutral Functional Differential Equations
Let there exist a bounded solution x(t),  co < t < co of Eq. (8.3) for some bounded functionf(t). Then all derivatives x(t), . . . ,x('")(t), co < t < co,are bounded. This theorem, extended on differentialdifference equations in [1943, is not valid in general for NFDEs. Consider the scalar NFDE (2n It ) i ( t )
i(Jt22)= 2(t  J t 2 _ 2 r r ) x ( J m )
The coefficient of this equation is bounded for t 2 2n and f ( t ) = 0. But this equation has bounded solution x(t) = sin(?), t 2 2n, with unbounded derivative. Prove an Esclangontype theorem for the NFDE x'"t'(t) + A,(t,
X y )
+ A2(t,Xt,...,xi",))
x(e) = q e ) = ... = x ( m ) ( e ) = 0,
t >0
=f(t),
h Ie I o
(8.4)
Here continuous operators A , : R , x L,[h,O] + R,, A2:R, x C [  h, 01 x . . . x C [  h, 01 + R , are linear in all arguments beginning from the second. In addition, A ,(t, cp) E L,[O, T ] for any cp(s) E L,[  h, TI, T > 0, and A2(t, $f", . . . , E L,[O, T ] for any @")(s) E C [  h, T I . Assume also that vrai suplA,(t, $;I), . . . , $$'")I OStST
< a [ ~ ~ $ ( l ) ~ ~ C [  ! r+ , T ~" '
+ ll$(m)llC[h,T])
(8.5)
Condition (8.5) replaces the assumption about the boundedness of the Eq. (8.3) coefficients. Consider the functional equation z(t) + A,(t, Z t ) = q t ) ,
t > 0; R(t)EL,CO,
z(e)= 0,
h Ie Io
col
(8.6)
Theorem 8.3 [l08(5)]. Let operators A , and A , satisfy the formulated requirements and Perron condition be met for Eq. (8.6). If there exists a bounded solution x ( t ) of Eq. (8.4) for a bounded function f ( t )E L,[O, a),then vraisup,,,,,[li(t)] + .  + Ix(")(t>l]< co. $9. LINEAR PERIODIC EQUATIONS 9.1. Some Examples Linear periodic NFDEs are studied less than RFDEs because the monodromy operator for NFDEs is not completely continuous. We now illustrate by examples some features of periodic NFDEs.
161
59. Linear Periodic Equations
EXAMPLE 9.1. The Floquet multiplier of NFDE may have injinite multiplicity. For the equation i ( t )  i ( t  1) = (cos 2nt)[x(t)  x(t  l)], the Floquet multiplier p = 1 has infinite multiplicity since any continuously differentiable 1periodic function x ( t ) is its solution. EXAMPLE 9.2. Consider a scalar equation which has no Floquet solution i(t)  i ( t

2n) = (cos t)x(t  271)
(9.1)
If x ( t ) = exp(It)y(t) is a Floquet solution of (9.1), then y ( t ) must be a 2nperiodic solution of the equation j ( t ) [1  exp(  2nA)] = [(cos t ) exp(  2nA)  I( 1  exp(  25ry))]y(t) (9.2)
But for any complex number I Eq. (9.2) has no nontrivial periodic solutions. Let be I = mi, m = 0, f 1, + 2 , . . . . Then the unique solution of Eq. (9.2) having the form 0 = (cos t)y(t) is y ( t ) = 0. If I # mi, m = 0, f 1,. . . ,then the general solution of Eq. (9.2) is exp(  2nI) cos s  I ] d s } 1  exp(  2nI) This solution will be 2n periodic if and only if
But it is impossible in the domain I # mi, m = 0, f 1,. . . . Hence, there are no nontrivial periodic solutions of Eq. (9.2) and the Floquet solutions of Eq. (9.1). EXAMPLE 9.3. Consider an equation whose Floquet solutions do not give an asymptotic representation for other solutions: 2i(t)  i ( t

271) = (cos t)x(t  27~)
(9.3)
Research of Floquet solutions of Eq. (9.3) is equivalent to the determination of a I such that the equation j(t)[2  exp(  2n;l)l = [(cos t ) exp( 2nI)  4 2  exp(  2nI))]y(t) (9.4) has a nontrivial period solution. If [2  exp(  2nI)l = 0, then Eq. (9.4) has no nontrivial periodic solutions. If [2  exp(  2nI)] # 0, then the periodic solution of Eq. (9.4) exists only for I = 0. Then this periodic solution y ( t ) is simultaneously a Floquet solution x,(t) of Eq. (9.3) x,(t) = y ( t ) = exp(sin t). Now find solution xl(t) of Eq. (9.3) with initial data x l ( t ) = sin t , i l ( t ) = cos t, 0 I t I 271. By the step method we have m+ 1
x,(t) =
C &,(sin tIk,
k= 1
2mn I t I 2n(m + I), m = 0, I , . . . (9.5)
162
3. Stability of Neutral Functional Differential Equations
Here Ak, are defined by the formulae Ak, = Ak, J2
+ Ak,ll1/2,
A f , = 1, A:
= A'",+' = 0
(9.6)
From (9.6) it follows that Ak, > 0, m = 1, 2, ...; 1 I k Im + 1 and m+ 1
max
1 Ak,,
Ix,(t)J= xl(tm)=
k=l
Znrndt
Here the arbitrary processesf,(t), f2(t), t~ [0, T] are measurable relative to (i algebra B  mo(cp) u Bor(dt)and
~oTElfl(t)12 dt < 00 1.2. Existence Theorem
Assume that for any x(8), y(8) E CB[  m,O] and t 2 0
M4
144 x)  44 Y)I2 I
 Y(S)12
dR,(s)
JOm
IW, x)  b(t, Y)12
I Ix(s>

Y(s>12 dR,(s)
( 1.4)
166
4. Stability of Stochastic Functional Differential Equations
where R , and R , are nondecreasing functions with bounded variations r:
=
i = 1, 2
jomdRi(s)< co,
) continuous samples and It is always assumed that the process ( ~ ( 8has
supEIcp(6)I4 < a
eso
Theorem 1.1 [91]. Let Eq. (1.1) and the initial function satisfy the conditions formulated above. Then there exists a unique solution of problem (l.l), (1.2) with bounded fourth moment on any finite time interval. This solution is measurable relative to the processes cp(6), 8 I 0 and ___ = dt
Iom {k(t, t  S ) d,A’(t, S )
Let Eq. (1.7) be of the form d x ( t ) = a(t)x(t  h) dt Then J
= E x ’ ( T ) H x ( T )(where
+ b(t) dt;(t),
t 2 0, h 2 0
T > 0 and a matrix H are given) is equal to
Here P
= B;(O)HB,(O),
Q
= B;(O)HB,(s)
Ns, P ) = B;(S)HB;(P) c1
=
loTTrB;(t)HB,(t)b(t)b’(t)dt
The matrix B , is easily calculated by the step method from the equation B,(t) = B,(t
+ h ) ~ ( +t h),
B,(T) = I, B,(s) = 0, s > T
92. FORMULATION OF STABILITY PROBLEMS FOR SRFDEs. LIAPUNOV DIRECT METHOD
2.1. Basic Definitions of Stability In this paragraph various definitions and theorems relating to stochastic stability are formulated for vector SRFDEs (1.1). It is always assumed that Eq. (1.1) satisfy requirements from $1 and also that 4 t , 0) = 0,
b(t, 0 ) = 0
(2.1)
168
4. Stability of Stochastic Functional Differential Equations
From condition (2.1) it follows that Eq. (1.1) has a trivial solution x ( t ) = 0 defined by a zero initial condition. The stability of the trivial solution of system (1.1) is investigated relative to the disturbances of the initial function cp(0). Denote x(t, cp) the solution of problem (l.l), (1.2).
Definition 2.1. The trivial solution of SRFDE (1.1) is called: ( a )p stable i f for each E > 0 there exists S(E) > 0 such that for any initial process cp(0) with continuous trajectories and independent from Bo,(d(), the inequalities
I
SUPE cp(e)ip 850
0 and E~ > 0 there exists a 6 > 0 such that the solution x(t, cp) of problem (l.l), (1.2) satisfies the inequality
Lo
P supIx(t, q)l I E ,
I
2 1  E2
provided that SuP8coIq(0)l I 6 with probability 1; ( b ) asymptotically stable with respect to probability if it is stable with respect to probability and for any positive number L1 and A2 and arbitrary initial function lp(8)ESH with probability 1, there exists T(H,A,, A 2 ) > 0 such that
Definition 2.3. The trivial solution of SRFDE (1.1) is called exponentially p stable ( p > 0) i f for any positive constants c , and c2 Elx(t, cp)lp Ic,supEIq(0)lP exp(c2t),
t20
850
2.2. Asymptotic p Stability General theorems of the Liapunov direct method for SRFDEs may be obtained by formal generalization of the appropriate theorems for
169
$2. Formulation of Stability Problems for SRFDEs
deterministic equations from Chapter 2, 92. Let us formulate, therefore, the only theorem used in this section.
Theorem 2.1. Let Eq. ( 1 . 1 ) satisfy requirernentsfrorntjl and (2.1). Assume that on [0, 00) x CB[  00,01 there exists the continuous functional V ( t , q) such that V ( t ,q) 2 c*lq(0)lP,
t 2 0,p 2 2
(2.4)
Taking into account (2.4), (2.5) we get
Hence, p stability is proved. Justify (2.3). By (2.7) and It6's formula we conclude that Elx(t, q ) l P as a function of t satisfies the Lipschitz condition and
Therefore relation (2.3) holds true.
REMARK2.1. Theorem 2.1 is valid if, instead of inequality (2.6), the following inequality holds: d W , x,) I c,Ix(t)lP dt
+ U l ( t )dt;(t),
t >0
Here the continuous process a , ( t ) is measurable relative to B  ,o(q) u B0,(d5) and has a bounded second moment on any finite time interval.
170
4. Stability of Stochastic Functional Differential Equations
2.3. Exponential p Stability
Formulate the necessary and sufficient conditions [230(2)] of exponential p stability for a stationary SRFDE with finite delay
+ b(x,) d((t), xo(e)= Cp(e),
t 20
dx(t) = a(x,) dt x(t
+ 0) = xr,
h I
eI o
(2.8)
Theorem 2.2. The necessary and sufJicient condition of exponential p stability of the trivial solution of Eq. (2.8) is the existence of functional V ( q ) with the properties C,IICpIIP
I
UCp) I
 1
lim
r+
+o
;{EV[xr(~p)l

czll’PIIP
~ ( c PI )}C~II(PII~
REMARK2.2. Theorems 2.1 and 2.2 are valid when the stochastic disturbance t ( t ) is a general process with independent increments
2.4. SRFDEs with Random Delays [99, 1291 Consider the system
Here A and B are constant matrices and q(t) E [0, h] is a homogeneous scalar Markovian process with a finite number of states ~ ( 0=) qo. Here q0 is any point from the interval [0, h].
Theorem 2.3. Let the continuous functional V ( 9 ,q ) exist such that, unijormly in ‘l, Wl(llCpII)
lim
r++o
1 
t
5
VCp,‘l)5 ~2(1lCpII),
[EV[x,(cp), d t ) l

‘lE[O,
VCp,‘lo11 I ~AII(PII>,
hl
d o ) = ‘lo
Then the trivial solution of Eq. (2.9) is stable with respect to probability.
Theorem 2.3 permits us to relate the stability of SRFDE (2.9) to the stability of the corresponding deterministic equations. Let, for example, the deterministic system (2.9) with r](t)= r] be exponentially stable uniformly in for arbitrary r] E [0, h] and also the derivative dEl r](t) ‘loI/dt
(2.10)
171
$3. Stability of Scalar Stochastic Equations
be sufficiently small. Then the trivial solution of (2.9) is stable with respect to probability. Notice that in general this statement is not valid if the derivative (2.10) is not small (see Example 2.1.4).
$3. STABILITY OF SCALAR STOCHASTIC EQUATIONS 3.1. Preliminaries Derive some sufficient conditions for meansquare stability of solutions of the scalar SRFDE r rm 1 dx(t) = J x(t  s) dKo(s) a(t, x,) dt
1
+
1
0
+ b(t, x,) dat),
x(e)= cp(e),
t >0
co
c2
+ 2(1 + bh)lal
Note that the trivial solution of Eq. (3.7) with b = 0, a > 0 is meansquare unstable. Consequently it is possible to stabilize a system by introducing time lag.
3.3. Case of Nonautonomous Linear Part Some modifications of functionals (3.2), (3.4), (3.6) permit us to investigate Eq. (3.1) where kernel Ko(s) is exchanged on Ko(t, s). Consider, for example, the equation dx(t)
=
[
~ ( txt) ,
+
s:
~ ( t s)f(t,
+ b(t, X') d t ( t ) ,
S)
1
ds  @l(t)x(t) dt
t20
(3.8)
The functionals a(t, xr)and b(t, x,) in (3.8) are the same as in Eq. (3.1), the the function f ( t , s) is function crl(t) is continuous and bounded for t E [0, a), continuous for all t, s and
pxt,
s)l ds 5 462)  d s d ,
s2
2
s1
where q(s) is a bounded nondecreasing function. The meansquare asymptotic stability of Eq. (3.8) takes place under the conditions 2a1(t)  r:  2r1 
If(t Jom
+ s, s)l ds 
jorn
I f ( t , s)l
1
ds > 0
(3.9)
w. Stability of SecondOrder Equations
175
These stability conditions are proved with the aid of the functional V ( t ,X t ) = X y t )
+
+
s,”
ds p ( t 1
+ s, s)lx2(t1)d t ,
jom
Cr;’ d R , ( s ) + dR,(s)l
lt;sx2(tJ
dt,
For the equation d x ( t ) = a(t)x(t)
t > 0,
+ bx(t  h) @ ( t )
h 2 0,
b a const
where the function a(t) is continuous and bounded, the stability condition is 2a(t) > b2, t 2 0.
54. STABILITY OF SECONDORDER EQUATIONS
4.1. General Assumptions Obtain some conditions for the meansquare stability of the secondorder SRFDE i ( t ) = y(t),
dy(t) =
[

t >0
j o m y ( t  s) dKo(s) 
+ b(t, y,) d&t),
y,
=~
joa
+
1
x ( t  s) d K , ( s )
( t 0),
GO
dt
< 8 I0
(4.1)
The coefficients of Eq. (4.1) will always be assumed to satisfy the conditions from $1. The kernels K,, K O have bounded variations on [0, 00) and the integrals in (4.1) are RiemannStieltjes ones [l05(l l)]. Let, for the arbitrary functions cp,, cp2 E CS[  co,01,
and let
b(t, 0)
=0
(4.3)
where K z ( s ) is a nondecreasing function of bounded variation on [0, 00). Under these assumptions there exists a unique solution [ x ( t ) , y(t)] of system (4.1) satisfying the initial conditions X(O)
= cp(e),
Y(e) =
wo, o I 0
(4.4)
176
4. Stability of Stochastic Functional Differential Equations
Here q ( B ) is a given process with continuously differentiable realization such that sw[v4(e) + ~ ~ ( e ) i (4.5) eso
It is assumed that o algebra B_,,(cp, Q) is independent from the o algebra Bo,(d(). In this case the inclusion B  ,*(x, y ) c [ B  mo((p, 4) u B0,(d5)] is valid. From (4.1)(4.5)it follows that for all t 2 0 the solution of problem (4.1)satisfies the inequality
+
E [ x ~ ( ~ y4(t)] ) I c1supE[v4(8) eso
+ Q4(e)]ec2',
ci 2 0
(4.6)
4.2. Stability Conditions The conditions formulated in this section are expressed in terms of moments of the kernels Kj(s).Denote
r(t) = ECX2(t) + Y2(t)l>
Y(v)= supECv2(@)+ 42(e>l eso
Theorem 4.1. Let the coefficients of system (4.1)satisfy requirements (4.1), (4.2), and let the kernel Ko(s) have a jump at zero of value a > 0, with Do1
> 0, a >
j+mo
IdKOWl
+ E l l + a029
a10
+ a21 + a12 = 00
Then the trioial solution of system (4.1)is asymptotically meansquare stable.
Proof: Introduce the functional w t ,
YJ
= 2B0lX2(t)
+ Y2(t)+ J+rOnIdKo(s)I
[scY2(h)
l;'(tl)
+ BOlX2(tl)ldtl + 2 jorndK2(s)
where
dt,
177
$4.Stability of SecondOrder Equations
On the basis of (4.1) and ItB's formula we have dy2(t) = b2(t,YJ dt
+ 2y(t)b(t, Y,)dt(t)
But by (4.2), (4.3) with probability 1 (4.10) Moreover
(4.11)
[
dy2(t) I lornxz(t s) d K 2 ( s )
+ j+wo y2(t  s)IdKo(s)I
j
m
2x(t)y(t)B01
+ y2(t)(2a + Ull +
IdKo(s)I)
(4.13)
+O
+ [oaldK'(s)l
[Jd,,]mdt + 2Y(t)b(t, Y J dt(t)
It should be noted that inequality (4.13) means that for arbitrary t 2 2 t , 2 0 the integral in the left side of (4.13) from t , to t , is, with probability 1, less than or equal to the integral in the right side of (4.13) between the same limits. From the second equation of Eqs. (4.1) and It8's formula we obtain dK%x,,Y,)
= 2V,(X,,
y,)[ x(t)Bo1 dt
+ b2(t,Y,)dt
+
Y,) d W 1 (4.14)
178
4. Stability of Stochastic Functional Differential Equations
Estimating as in (4.11) the terms on the right side of (4.14) we obtain according to (4.8)
+ j O m x 2 (t
S)
(4.15)
dK,(s) d t
From (4.Q (4.13) and (4.15) follows the formula
(4.16)
Let us integrate inequality (4.16) from zero to t 2 0 and take the expectation on both sides of the result. By (4.9, (4.6), (4.10) and the hypotheses of Theorem 4.1, the expectation of the stochastic integral in the right side of (4.16) is equal to zero. Hence for all t 2 0 W x t , Y,)  W c p , 4) IHere Yl = 2[n 
sd
sm
CYlEY2(s)+ YzEX2(s)lds
IdKo(s)l  all
 a02
+O
Y2
[
= 280, a 
r
o
IdKo(s)l  a l l
I
1
(4.17)
(4.18)
But by the hypotheses of Theorem 4.1 together with (4.8)
w c p ,4)
C,Y(cp)
w x , , Y,) 5 b o , ~ x Z ( t+) EY2(t)
(4.19) (4.20)
From these relations and (4.17) it follows that r(t) I c2y(cp) for arbitrary t 2 0. So the meansquare stability is established. For the proof of asymptotic
v. Stability of SecondOrder Equations
179
meansquare stability it suffices to show that the function r(t) is uniformly continuous on [0, co) and jomr(t)dt < cc
(4.21)
But inequality (4.21) follows immediately from (4.17)(4.20), since they imply that for all t 2 0
j ( ) Y l t Y 2 0 ) + Y z ~ X 2 ( s )ds 1
C,Y(Cp)
Let us prove now that the function r(t) is uniformly continuous on [0, a). For this we apply ItB’s formula to the function x 2 ( t )+ y2(t). From (4.10), (4.11) we obtain that for arbitrary nonnegative t , and t , M t l )  &)I
2C,Y(Cp)(l
+ a00 + a01 + “02)ltl  t z l
Theorem 4.1 is proved For the equation ( t 2 0)
W ) = YW,
dy(t) = [  a l v ( t )
+ a,x(t
 h d l dt
+ a&

h2) d t ( t )
Theorem 4.1 gives that the trivial solution is asymptotically meansquare stable if a, > 0 and a , > a,h, + a:.
Theorem 4.2. Let the conditions (4.2), (4.3) be satisfied and let 41 =
4 = Boo
a10
 P11
+ +a21 < 1,
BOl
>0
> maxC41, (Bo141 + ao,)/(l  4111 f aZl
f a20
+ @3l
0
(4.24)
Sufficient asymptotic meansquare stability conditions for Eq. (4.24) are la/ < [h2(1
 sb’)’
 b2]1’2,
s = c2/2,
c = rnax(Ic,I, Ic21).
Stability domains of (4.24) for different values of parameter sh are represented in Fig. 4.1. The case in which h = 0 is represented in Fig. 4.2. For the equations
+ c,x(t) d0
(4.27)
182
4. Stability of Stochastic Functional Differential Equations
.f
bh
Fig.4.3. Stability domains for system (4.25).
I
10
Fig.4.4. Degenerated case of h = 0.
* a
$4.Stability of SecondOrder Equations
183
1
Fig. 4.5. Stability domains for system (4.26).
Fig.4.6. Degenerated case of h = 0,s > 0.
oh
184
4. Stability of Stochastic Functional Differential Equations
Fig. 4.7. Degenerated case of s = 0, h > 0.
Inequality (4.27) is impossible if s1 = sh3 > 3J3/128 z 0.040595. The stability domain is shown in Fig. 4.5. Degenerated cases in which h = 0, s = 0 or s = h = 0 are shown in Fig. 4.64.8. For the equations % ( t )  [aZ
+ c&t)]x(t) + 2b%(t h) + b2X(t  2h) = 0
(4.28)
the stability domains represented in Fig. 4.9 are given by the inequality b > 0. This inequality is impossible c2/2 = s < a2b[1  h(la1 + lbl)] for sh3 > (219,/%  1871)/128 z 0.001069. Degenerated cases are shown in Fig. 4.104.12.
Fig.4.8. Case of h
= s = 0.
185
$4. Stability of SecondOrder Equations
t’
,sl = 00001 ,s,= 00005 ,sl = 00008
Fig. 4.9. Stability domains for Eq. (4.28)
\ I
I
2
1
\
\ \
/ )
/
/
/
/
I
I
1
2
Fig. 4.10. Degenerated case ( h = 0, s > 0).
L
0
98 I
187
$5. Stationary and Periodic Solutions
g5. STATIONARY AND PERIODIC SOLUTIONS
OF STOCHASTIC RETARDED EQUATIONS This section is devoted to the stationary and periodic solutions of SRFDEs. Some results about existence of stationary solutions are derived on the basis of the Liapunov direct method. 5.1. Existence of Stationary Solutions Let [ ( t )E R, be a continuous stationary (in the strict sense) process, i.e., for every sequence (tl,. .. ,t N )the joint probability distribution of the variables [(tl + T ) ,. . . , [(tN T ) is independent of 7: Consider initialvalue problem
+
i ( t ) = ~ [ x ,[(t)], ,
x(e)= cp(e),
t > 0,
h I eI 0,
X E
(5.1)
R,
cp(e)E
qh,
01
(5.2)
Here continuous functional a(cp, z ) E R, is such that la(cp,,z>
 a(cp2, 41 5 rllcp1  V2IL
cp1,
cPzEC[hh,Ol
Definition 5.1. A stationary process x ( t ) is a stationary solution of Eq. (5.1) if x ( t ) is stationary connected with i ( t ) and satisjes Eq. (5.1). Theorem 5.1. Let there exist such a solution X ( t ) of ( 5 4 , (5.2) such that (1) uniformly in t > 0
lim; s:p(i(s) cmt
( 2 ) for any positive
E ~ E, ~ C ,
> C } ds = 0
and t l , t2 E [ h,
00)
Elx(t1)  x(t2)IE1I C l t ,  t211+E2
(5.3)
Then there exists a stationary solution of Eq. (5.1).
For the proof of Theorem 5.1 see Kolmanovskii [105(1) and Kolmanovskii and Nosov 108(5)]. 5.2. Relation between Stability and Stationarity of Solutions Let a(cp, 5)
= al(cp)
+ [(t), a l ( 0 ) = 0 in Eq. (5.1), k(t) = a,(x,)
+ [(t),
i.e.,
t >0
(5.4)
188
4. Stability of Stochastic Functional Differential Equations
The existence of a stationary solution of Eq. (5.4) depends on the stability “degree” of the deterministic system
Theorem 5.2. Assume that the trivial solution of Eq. (5.5) is unqormly exponentially stable and also that El{(t)(’ < co. Then there exists a stationary solution of Eq. (5.4). Proof. Consider the solution x ( t ) of (5.4) with zero initial condition (5.2). Show that Elx(t)12 I c < co.From the conditions of Theorem 5.2 follow the existence of the continuous functional V ( q ) (Chapter 2, $5) such that
Let the function z ( t ) be determined by
Then for any s 2 0
Hence, applying Lemma 1.4.2 we get
From (5.7) and (5.6) it follows that rs
189
55. Stationary and Periodic Solutions
so
From Chebyshev’s inequality [64] we obtain P { Ix(s)~
> C} ds I c7c’
+
0
for
(5.9)
c +co
By virtue of (5.4), (5.8) Elx(t
+ A)  ~(t)1’ +
I 2A L + A E [ ~ : ( ~ ,[’(s)] ) ds
jt
t+A
5 26
E[r211xsl12
+ (’(s)]
ds 5 c,A2
(5.10)
Theorem 5.2 now follows from Theorem 5.1 and relations (5.9), (5.10).
5.3. Periodic Solutions A stochastic process B(t) E R, is said to be periodic with period T if for every sequence (tl,. . .,t N ) the joint probability distribution of the stochastic variables B(tl + mT),. . . ,B(tN + mT) (where m = f 1k 2 , . . .) is independent of m. Consider the equation i ( t ) = b[t, ~
( t )B(t)], ,
t 2 0, X E R , , P E R ,
(5.11)
Here the continuous functional b(t, cp, z ) for any fixed cp and z is periodic in t with period T and also
INt, cp, 4  b(4 q,z)l I cllv  $11,
v 40,
$EC[hh, 01
Definition 5.2. A periodic process x(t) is called a periodic solution of Eq. (5.11) i f x ( t ) is periodically connected with B(t) and satisjies to Eq. (5.11).
190
4. Stability of Stochastic Functional Differential Equations
Theorem 5.3. Let conditions (5.3) and (5.12) hold and there exist at least one solution x ( t ) of Eq. (5.1 1 ) such that uniformly in k k
= 0,
=
1 , 2,. . .
Then there exists a periodic solution of Eq. (5.1 1).
The proof can be found in Kolmanovskii [l05(1)] and Kolmanovskii and Nosov [108(5)].
5.4. Ergodic Properties of Stationary Solutions We present here some ergodic properties of stationary solutions of the SRFDE. dx(t) = al(x,) dt
+ b[x(t)] d((t), x(t)ER,,
t20 xt = x(t
+ O),
h I 8I 0
(5. 3)
Ergodic properties of Eq. (5.1 3 ) solutions are investigated by means of stationary solutions of a stochastic equation without delay dY(t) = azCY(t)l dt
+ bCY(t)l d m ) ,
t20
(5. 4 )
Here ( ( t ) is a standard Wiener process, the functional al(cp), cp E C [  h, 01 and functions a,(x), b ( x ) are continuous and satisfy the Lipschitz conditions. Any stationary solution x,(cp) of Eq. (5.13) with initial condition (5.2) induces the invariant measure p on (T algebra 3, of Bore1 sets from C [  h, 01.
Definition 5.3. Stationary solution x,(cp) of Eq. (5.13) is called ergodic i f for an arbitrary 3, measurable functional A(cp), cp E C [  h, 01, lim t+m
1 ~
t
' ~ , A [ X s ( c p ) 1 ds =
1
A(cp)P(dcp)
C [  h , 01
almost everywhere with respect to invariant measure p.
Theorem 5.4. Let there exist the unique stationary ergodic solution of Eq. (5.14). I f there exists stationary solution of Eq. (5.13) then it is unique and ergodic. The proof of Theorem 5.4, which is available in Kolmanovskii [105(4)],is based on absolute continuity of measures induced by solutions of Eqs. (5.13) and (5.14). In Kolmanovskii [105(4)], and Kolmanovskii and Nosov [108(5)] the widesense mixing property of stationary solutions of SDFE (5.13) are also obtained.
191
$6. Stability with Respect to the First Approximation
56. STABILITY WITH RESPECT TO THE FIRST
APPROXIMATION 6.1. General Theorem
Conditions of stability theorems with respect to the first approximation depend on the form of the first approximation equations. Establish a theorem for the case in which the equation of the first approximation is a deterministic one
with initial conditions (1.2). Continuous functionals a(t, cp) and b(t, cp) satisfy the conditions
l ~ ( ~ , c p ~   ~ ~ , c p l ~ l I ~ l l I c p  c p , I l , a(t,O)=b(t,O)=O
ucp  cplllt
lb(G cp)  b(4 cpdl I
cp,
cp1 ECLhh,
01
(6.2)
Examine the stability of the trivial solution of the equation
qt>= 44 x,) + w,x,)5(Q
(6.3)
where [ ( t )is a stochastic process with continuous samples. The definitions of stability for Eq. (6.3) are similar to Definitions 2.12.3. Let the trivial solution of Eq. (6.1) be uniformly exponentially stable. Then there exists the functional V(t ,cp) satisfying (5.6). Define the function z(t) by conditions z ( t ) = x(t) for t 5 s and i(t) = a(t, zt) for t > s. Then for any s 2 0 v ( s , x,) =
i 6 (I/A)[V(s
+ A, Z,+A)

v(s,X,)]
A+ + O
+
KG (l/A)[V(s
+ A, x,+J  V ( S+ A,
Z,+A>]
A+ + O
5 (c4/c2)V(s?
xs>
+ c3 I b(s,xs>5(s)l
Hence according to Lemma 1.4.2 from Chapter 1 and (5.5) IIXt
II I (Cz/Cl)IIVII exP[ (C4/C&I
+ (CjL/Cl) j;exPcc4(t Applying Lemma 1.4.1 to the function y ( t )
 ~ ~ / ~ 2 l l l ~d ss l l l ~ ~ ~ ~ l =
(Ix,I(exp[(c4/c2)t] we get
192
4. Stability of Stochastic Functional Differential Equations
Further
Require that  c4t/c,
+ c,L/c,
s:
1
p 2 0 (6.4)
lC(s)lds Iexp(pt),
Then EllxtllZ I (cz/c,)ZEllip112exp(Pt). Thus the following theorem holds.
Theorem 6.1. Let condition (6.4) be fulfilled and the trivial solution of Eq. (6.1) is uniformly exponentially stable. Then the trivial solution of Eq. (6.2) is asymptotically meansquare stable. 6.2. Scalar Equations
Formulate some stability conditions for scalar equations of the form (6.3) f a
i ( t ) =  J x(t  S) dKo(s) 0
+ B[t, x(t)][(t),
t >O
The kernel ko(s) satisfies the inequality [s~idko(s)l2 C , [exp(Cso) exp(Cs)],
s 2 so 2 0
Also assume that one of the following conditions is valid: (1)
Boo
> 0, a10 < 1,
(2) the kernel ko(s) has a jump at zero of value a > 0 and
(3) ko(s) is nondecreasing and ko(s) = ko(h) for s 2 h 2 0,
Denote z(t) the solution of Eq. (6.5) for fl = 0 with initial conditions z(0) = 1, < 0. By virtue of (6.6) and one of conditions (6.7)(6.9) there exist constants y,, y z such that J z ( t ) JI y 1 exp(y,t), where 0 < y z < C , y , > 0. z(s) = 0, s
193
$7, NeutralType Stochastic Functional Differential Equations
Theorem 6.2. Let the above formulated assumptions about Eq. (6.5) be valid. Then: ( a ) the trivial solution of Eq. (6.5) is asymptotically stable with respect to probability if sup,,, El[(t)l < Y ~ ( Y ~ L )  ~
P{:
ccI(.)

E[(s)] ds + 0
I
=
1,
t +
00
(b) the trivial solution of Eq. (6.5) is asymptotically pstable for p I B1(2y,L)', ifthere exists a positive P1, B2 such that
Other stability conditions for equations disturbed by the process [ ( t ) can be found in Kolmanovskii and Nosov [108(5)], and Shaichet [199(2)].
$7. NEUTRALTYPE STOCHASTIC FUNCTIONAL DIFFERENTIAL EQUATIONS 7.1. Definition of SNFDEs An SNFDE may be obtained from deterministic Eq. (3.3.2) by adding random perturbations to their right side. This class of equations was introduced in Kolmanivskii [108(5)]. We now present some results for the following equations d[x(t)  G(t, x,] = a(t, x,) dt
+ b(t, xt) d((t),
t20
(7.1)
Here x E R, and the continuous functional G defined on
[O,
00)
x CB[  00,0]
satisfies the conditions IG(4 CP)  G(t, $)I
1444
 $(s)l
dK,(s)
(7.2)
jom
=
l o m d K l ( s ) 1,
supIG(t, 0)l < 00 rto
where K , ( s ) is a nondecreasing function. The process ( ( t ) is a standard Wiener one. The functionals a(t, x,) and b(t, x,) are just the same as in Eq. (1.1). The initial conditions for Eq. (7.1) has the form (1.2), (1.5). The solution of problem (7.1), (1.2) is understood in the sense of corresponding integral identity.
194
4. Stability of Stochastic Functional Differential Equations
7.2. Existence Theorem
Equation (7.1) differs from Eq. (1.1) only by the term G(t, xt). But it is easy to generalize the theorems about the existence, uniqueness and increase of the solutions of problem (7.1), (1.2) and also the theorems about the existence and properties of stationary solutions known for G = 0. We state here without proof the following theorems.
Theorem 7.1. Let Eq. (7.1) and the initial function satisfy the formulated requirements. Then there exists the unique solution of problem (7.1), (1.2). This solution has a bounded fourth moment on any Jinite time interval and is measurable relative to the processes (p(8), 8 I 0 and &s), 0 I s I t. The process x , is Markovian. The proof of Theorem 7.1 is founded on the method of successive approximations [9l, 108(5)].
Theorem 7.2. Let the requirements of Theorem 7.1 be fuljlled and also the functionals a, b, G be independent o f t . Assume that there exists a solution of Eq. (7.1) with bounded fourth moments. Then there exists a stationary solution of Eq. (7.1). 7.3. Stability Conditions Assume that G(t, 0) = a(t, 0 ) = b(t, 0) = 0. Then there exists a trivial solution of Eq. (7.1). The stability of this solution can be understood in different senses, for example, in the meansquare sense, with probability 1 etc. Here we consider the meansquare stability (see Definition 2.1). Analysis of the stability SNFDE (7.1) as well as the deterministic NFDEs is reasonable to reduce to the considerations to two auxiliary problems. The first one is a construction of a nonnegative functional with a nonpositive derivative. The second one is investigation of the stability of the trivial solution of some auxiliary functional inequality. The form of this inequality depends essentially on the sense in which the stability of Eq. (7.1) is investigated. So for study of meansquare stability it is necessary to consider the inequality
The solution of inequality (7.3) for t 2 0 is determined by initial condition (1.2). The meansquare stability of inequality (7.3) means that for any E > 0 there exists a S(E) > 0 such that Elx(t)I2 < E if C , < d ( ~ )Elq(8)12 , < S(E) S O .
Theorem 7.3. Suppose that Eq. (7.1) satisjies the formulated requirements and the trivial solution of inequality (7.3) is meansquare stable. Further, let there
195
$7. NeutralType Stochastic Functional Differential Equations
exist on [0, 00) x CB[  co,01 a continuous functional V ( t ,cp) of theform V ( t ,cp)
=
W t , cp)
+ Icp(0)  G(t, CP) l2
(7.4)
such that
EV(4 cp) ICsupEIcp(W eso
(7.5)
EV(t,, xtJ  EV(t1, X t , ) 5 0
The functional W(t,cp) in (7.4) is deJined on [0, 00) x C B [  O O , ~ ] , continuous, nonnegative and satisfies estimate (7.5). Then the trivial solution of Eq. (7.1) is meansquare stable.
EXAMPLE 7.1. Consider the neutraltype stochastic equation arising in the theory of aeroelasticity (Chapter 3, $2)
=
[

1 Iomx(t  s) dK,(s)
t 2 0,
1
+ a(t, x r ) dt + b(t, x,) d&t)
X(t)ER,
Assume that K,(s) has a jump at zero of the value b , > 0, the bounded function A(s) is continuously differentiable and
jam 1&)1
A(s) ds < 1,
a,, = !om
ds < co
In addition, conditions (1.5) are fulfilled and
Then trivial solution of Eq. (7.1) is asymptotically meansquare stable. For the proof consider the functional V(t,xt) = b t ) 
lom
x(t  s)A(s) ds]
2
+
lom+ [(a,,
r;
I)
I A(s) I ds
(for details see Kolmanovskii [I 05(5)], Kolmanovskii and Nosov [108(5)]).
196
4. Stability of Stochastic Functional Differential Equations
58. STABILITY OF LINEAR AUTONOMOUS EQUATIONS
8.1. Systems of Linear Equations Derive the asymptotic meansquare stability conditions for systems C199( 1)I. N
d x ( t ) = Ax(t) dt
+ 1 Brx(t  h) d&(t),
h 2 0, t 2 0
(8.1)
r=l
where x E Rnand t r ( t )are scalar independent Wiener processes. The elements aij and bij of matrices A and B, are constant. Notice that necessary and sufficient conditions of asymptotic stability of the deterministic equation 1 = A x are [l55] i = 1, ..., n
Ai>O,
(8.2)
Here Ai are principal diagonal minors of matrix S,
S=
s, s, s, ... 1 s, s, ... 0 s, s, ... 0
0
0 0
0 ... (1)”s”
0
Denote N
O(X)
=
C X’B;B,X = C r= 1
n
dij(x)
=
dij(x)
i,j=l
N
c c
bikbSsxsxk
k,s=l r = l
Lemma 8.1. The trivial solution of Eq. (8.1) is asymptotically meansquare “;:xixj = stable ifthere exists a positivedefinite quadratic form V ( x ) = X’VX such that x’(A V V A ) x =  ~ ( x and ) also the quadratic form ~ ( x=) I j: = d,(x)(l ,  K j ) is positivedejinite.
cZj=
+
Proof. Let L be the generating operator of Eq. (8.1). Consider the functional W(t,x,) = V [ x ( t ) ]
+
N
(8.3)
197
$8. Stability of Linear Autonomous Equations
Then from ItB's formula LW(t,Xt)
(8.4)
= .(X)
Hence the validity of Lemma 8.1 follows from (8.3), (8.4) and Theorem 2.1. REMARK.
Let dij = 0 for i # j . Then a ( x ) is positivedefinite if all
Ki < 1.
Find and a(x). Let A1,,,be the cofactor of the first row and m th column of the determinant A,. Determine the numbers qij by the equation
c oksDik(A)Djs(A)
n1
n
(1)"
=
k,s= 1
qijA2(nr1) r=O
Here oks are coefficients of quadratic form o ( x ) and Dik(A)are cofactors of the determinant D(A) a,,

A ...
a1n
D(A) =
...
a,,
ann A
Then
K j = Aij/2A,, ..
c n
a(x) =
=
aksXkXsr
1
..
bikbSs(l
 I/j)
i, j = l r = 1
k,s= 1
Denote by di principal diagonal minors of the matrix
Theorem 8.1. Let conditions (8.2) be fulfilled and di > 0, i = 1 , . . . , n. Then the trivial solution of system (8.1) is asymptotically meansquare stable. The proof follows from Lemma 8.1 and relation (8.5).
EXAMPLE 8.1. Given two equations 2
aikxk(t)dt
dxi(t) = k= 1
+ bixi(t  h) d(,(t),
i = 1, 2
Sufficient conditions for asymptotic meansquare stability of these equations are
s, = a,, + a,, max[b:(S,
< 0,
+ a:,),
s, = alla2,  a12a21 > 0 b:(S,
+ a;,)]
< 2.S1s2
198
4. Stability of Stochastic Functional Differential Equations
8.2. Corollary for nthOrder Equations Consider a linear SRFDE
i=l
I
N
C fiijx("i)(t h ) t j ( t ) = 0
a,x(""(t) +
j= 1
(8.6)
This equation is equivalent to system of the form (8.1)
dxl(t) = x2(t)d t , . . . , dxn l(t)
c a i x n  i + l ( tdt)
n
n
dx,
= 
= x,(t) d t
N
1 C Pijxni+l(t h) d t j ( t )

(8.7)
i= 1 j = 1
i= 1
For system (8.7) matrices A and S are

...
0
1
Functions o ( w ) and dij(x)are dij(x)= 0 i f i + j < 2n, o ( x ) = dnn(x) n
dnn(x)
=
N
1
(8.9)
fikjPsjXn+lsXn+lk
k,s= 1 j = 1
Lemma 8.2. Trivial solution of Eq. (8.6) is asymptotically meansquare stable if there exists a positive definite quadratic form V ( x ) = x'Vx = j= Kjxixj such that x'(A'V+ V A ) x = dnn(x), v,, < 1
C:,
(0)
qnn
A=
q;'n'
1
a2
0
0
.
. . . 4nn ( n  1' ... 0 ...
an
199
$8. Stability of Linear Autonomous Equations
8.3. Necessary and Sufficient Stability Conditions of Scalar Equations
Consider the scalar equation
(8.10)
+J
~
x(t
+ s>t i ~ ( s ~)
t ) , t 2o
h
with initial condition (1.2). Here K ( s ) and R(s) are functions with bounded variations on the interval 0 < s h. The necessary and sufficient conditions
Fig.4.23. Stability domains for Eq. (8.13) for SE[O,31.
200
4. Stability of Stochastic Functional Differential Equations
for asymptotic meansquare stability of the trivial solution of the system (8.10) are [230(2)]: (1) all roots of the characteristic equation
z
I:,
(8.11)
exp(zs) d K ( s ) = 0
lie in the left half plane; (2)
5 {omI lo
exp (is 0) dR(B)
h
Remark that from condition (1) it follows that deterministic system (8.10) with R = 0 is asymptotically stable. Let, for example, Eq. (8.10) have the form dx(t) = [  u x ( ~ ) bx(t  t)] d t
+ C X ( ~ Z) d t ( t )
rb i
2015
10
t 07
Fig. 4.14. Stability domains for Eq. (8.13) for s
< 10.
(8.13)
20 1
$8. Stability of Linear Autonomous Equations
From (8.11), (8.12) we obtain that system (8.13) is stable in the case Ibl 2 la1 if c2 < 2(1 + bk' sin kz)'(a
+ b cos kj),
k
=
Jbza2
(8.14)
and in the case Ibl 5 la1 if c2 < 2(1 + bk'shkz)'(a
+ bchkz),
k
=
Jazbz
(8.15)
The stability domains of system (8.13) given by inequalities (8.14), (8.15) are represented in Figs. 4.14 and 4.13 for different values of the parameter s = c22/2. The domain of stability is to the right of the boundary.
References
1. Achmerov, R. R., Kamenskii, M. I., Potapov, A. S., Rodkina, A. E. and Sadovskii, B. N., Theory of equations of neutral type. Naucn.Tehn. Znformacija ( V I N I T I ) , 16 (1982). 2. Afanasjev, V. N. and Rodionov, A. M., Stabilization of a class of nonlinear systems with delay. Diflerencial’nye Uraunenija 17 (9,914916 (1981). 3. Alekseevskaja, N. L. and Gromova, P. S., “Liapunov Direct Method for Retarded Differential EquationsDifferential Equations with Deviating Argument,” pp. 1734. Nauka Dumka, 1977. 4. Andronov, A. A. and Mayer, A. G., The simplest linear systems with delay. Automat. Remote Control 7 (23), 95106 (1946). 5. Angelov, V. and Bainov, D. D., Existence and uniqueness of the global solution ofthe initial value problem for neutral type differentialfunctional equations in Banach space. Nonlinear Anal. 3(6) (1979). 6. Arutunian, N. H. and Kolmanovskii, V. B., “Creep Theory of Nonhomogeneous Bodies”, Nauka, Moscow, 1983. 7. Astapov, I. S., Belotzerkovskii, S. M., Kachanov, B. 0. and Kochetkov, Ju. A,, Systems of integrodifferential equations describing unsteady motion of bodies in continuous media. Diflerencial‘nye Uraunenija 18(9), 16281636 (1982). 8. Azbelev, N. B. and Sulavko, T. S., About stability of solutions of differential equations with delay. Diflerencial’nye Uraunenija 10(12), 20912100 (1974). 9. Azjan, Ju. M. and Migulin, V. V., About autooscillation in systems with delayed feedback. Radiotek. Elektron. 1(4), 418430 (1956). 10. Babskii, V. G. and Myshkis, A. D., Mathematic models in biology, connected with regard of delays (Appendix) Murray, J. D., “Lectures on nonlineardifferential equations,” 383394. Models in Biology, MIR, Moscow, 1983. 11. Bailey, H. R. and Reeve, E. B., Mathematical models describing the distribution of I’31albumin in man. J . Lab. Clin. Med. (60), 923943 (1962). 12. Bakli, P. S. Control of the processes with delay. Trudy ZFAK, Moscow, 95111 (1961). 13. Banach, S. “Theorie des Operations Lineaires.” Warszawa, Poland, 1932. 14. Banks, J. T. (1) Representations for solutions of linear functional differential equations. J . Difler. Eq. 5(2), 399409 (1969). (2) “Modeling and Control in the Biomedical Sciences,” Lecture Notes in Biomath. 6 (1975). (3) Parameter identification techniques for physiological control systems. In “Mathematical Aspects of Physiology,” Lectures in Appl. Math. 19, 361385 (1981). 15. Barbashin, E. A,, “Introduction to Stability Theory.” Nauka, Moscow, 1967. 16. Barnea, D. I., A method and new results for stability and instability of autonomous functional differential equations. SIAM J . Appl. Math. (17), 681697 (1969). 17. Bellman, R., On the existence and boundedness of solutions of nonlinear differential difference equations. Ann. of Math. 50(2), 347355 (1949). 18. Bellman, R. and Cooke, K., “DifferentialDifference Equations,” Academic Press, New York, 1963.
202
References
203
19. Belotzerkiovskii, S. M., Kochetkov, Ju. A,, Krasovskii, A. A. and Novitzkii, V. V., “Introduction to aeroautoelasticity.” Nauka. Moscow, 1980. 20. Belotzerkovskii, S . M., Skripach, B. K. and Tabachnikov, V. G., “Wing in unsteady gas flow.” Nauka, Moscow, 1971. 21. Belykh, L. N. and Marchuk, G. I., Chronic forms of a disease and their treatment according to mathematical immune response models. Lecture Notes in Control and Information Sci. 18, 7986 (1979). 22. Bodner, V. A,, “Operator and flying vehicle.” Mashinostroenie, Moscow, 1976. 23. Bodner, V. A,, Rjazanov, Ju. A. and Shaimardanov, F. A., “Systems of control of aircraft engines.” Mashinostroenie, Moscow, 1973. 24. Bodner, V. A., Zakirov, R. A. and Smirnova, I. I., “Aircraft trainers.” Mashinostroenie, Moscow, 1978. 25. Bogomolov, V. L., Control of power of hydrostation. Automat. Remote Control (45), 103129 (1941). 26. Braddock, R. D. and Driessche, P. van den, Population models with time delay. Math. Sci. 5(1), 5566 (1980). 27. Brauer, F., Stability of some population models with delay. Math. Biosci. 33(2), 345358 (1977). 28. Brayton, R., Nonlinear oscillations in a distributed network. Quart. Appl. Math. 24, 289301 (1976). 29. Brewer, D. W., The asymptotic stability of a nonlinear functional differential equation of infinite delay. Houston J . Math. 6(3), 321330 (1980). 30. Burton, T. A,, “Volterra Integral and Difference Equations,” Academic Press, New York, 1983. 31. Butkovskii, A. G., Structural method for system with distributed parameters. Automat. Remote Control, No. 5 (1975). 32. Callender, A. and Stevenson, A. G., Time lag in a control system. Proc. Soc. Chem. rnd. l8(l), 108117 (1939). 33. Cargill, T. F. and Meyer, R. A,, Wages, prices and unemployment: distributed lag estimates. J . Amer. Statist. Assoc. 69(345), 98107 (1974). 34. Castelan, W. B., A Liapunov functional for a matrix retarded differencedifferential equation with several delays. Lecture Notes in Math. (799), 82118 (1980). 35. Castelan, W. B. and Infante, E. F., A Liapunov functional for a matrix neutral differencedifferential equation with one delay. J . Math. Anal. Appl. 71(1), 105130 (1979). 36. Cesari, L., “Asymptotic Behavior and Stability Problems in Ordinary Differential Equations,’’ Springer, New York, 1959. 37. Chebotarev, N. G. and Meiman, N. N., Problem of RouthGurwitz for polynomials and entire functions. Trudy Math. Inst. Steklou 26 (1949). 38. Chernousko, F. L. and Kolmanovskii, V. B. (I) “Optimal Control under Stochastic Disturbances.” Nauka, Moscow, 1978. (2) Computational and approximate methods of optimal control J . Soviet. Math. 12, 310353 (1979). 39. Chetaev, N. G., “Stability of Motion.” Nauka, Moscow, 1965. 40. Chipot, M., On the equations of agedependent population dynamics. Arch. Rational Mechanics Anal. 82(1), 1125 (1983). 41. Chistjakov, P. G., “Accuracy of control system of JRD and TRD.” Mashinostroenie, Moscow, 1977. 42. Coleman, B. D., A general theory of dissipation in materials with memory. Arch. Rational Mech. Anal. 255274 (1968). 43. Coleman, B. D. and Dill, H., On the stability ofcertain motions ofincompressible materials with memory. Arch, Rationat Mech. Anal. 30, 197224 (1968).
204
References
44. Coleman, B. D. and Owen, D. R., On the initialvalue problem for a class of functional differential equations. Arch. Rational Mech. Anal. 55, 275299 (1974). 45. Cooke, K. L. and Yorke, J. A,, Equations modelling population growth, economic growth and gonorrhea epidemiology. In “Ordinary Differential Equations” (L. Weiss, ed.), Academic Press, New York, 1972, pp. 3555. 46. Corduneanu, C. (1) Integral Equations and Stability of Feedback Systems. Academic Press, New York, 1973. (2) Stability problems for some feedback systems with delay. In “Modern Trends in Cybernetics and Systems.” Vol. 2. pp. 321328, SpringerVerlag, Berlin and New York, 1977. 47. Corduneanu, C. and Lakshmikantham, V., Equations with unbounded delay: a survey. Nonlinear Anal. Theory Methods Appl. 4(5), 831878 (1980). 48. Cruz, M. A. and Hale, J. K., Stability of functional differential equations of neutral type. J. Difler. Eq. 7(2), 334355 (1970). 49. Cryer, C. W., Numerical methods for functional differential equations. In “Delay and Functional Differential Equations and Their Applications,” pp. 17101, Academic Press, New York, 1972. 50. Cushing, C. M., “Integrodifferential Equations and Delay Models in Population Dynamics.” Lecture Notes in Biomath. 20 (1977). 51. Danilushkin. A. I. and Rapoport, E. Ja., Algorythm of process functioning of continuouslysuccessive inductive heating. In “Algorythmizing and automatization of technological processes in industrial mountings” Vol. 7. Kuibyschev, 1976. 52. Datko, R., The uniform asymptotic stability of certain neutral differentialdifference equations. J . Math. Anal. Appl. 58(3), 510526 (1977). 53. Differential Equations with Deviating Argument. Nauka Dumka, Kiev, 1977. 54. Delay and Functional Differential Equations and Their Applications. (K. Schmitt, ed.) Academic Press, New York, 1972. 55. Dibrov, B. F., Livshits, M. A. and Volkenstein, M. V. The effect of a time lag in the immune reaction. Lecture Notes in Control and Information Sci. 18, 8794 (1979). 56. Draljuk, B. N. and Sinaiskii, G. V., “Systems of Control for Plants with Transport Delays.” Energia Moscow, (1969). 57. Driver, R. D. (1) A functional differential system of neutral type arising in a twobody problem of classical electrodynamics. In “Nonlinear Differential Equations and Nonlinear Mechanics” pp. 474484. Academic Press, New York, 1963. (2) “Ordinary and Delay Differential Equations” SpringerVerlag, Berlin and New York, 1977. 58. El’sgol’tz, L. E. (1) Stability of solutions of differentialdifferenceequations. Uspehi Mat. Nauk, 9(4), 951 12 (1954). (2) To the influence on stability of small deviation of argument. Trudy Sem. Teor. Differencial. Uravnenii s Otklon. Argumentom Univ. Druzby Narodov Patrisa Lumumby 1, 1141 15, (1962). (3) Qualitative methods in mathematical Analysis. Trans. Math. Monographs 12, (1964). 59. El’sgol’tz, L. E. and Norkin, S. B., “Introduction to the Theory and Application of Differential Equations with Deviating Arguments,” Academic Press, New York, 1973. 60. Faerman, E. Ju., “Problem of LongRange Planning.” Nauka, Moscow, 1971. 61. Fichtengol’tz, G. M., “Course of Differential and Integral Calculus,” Vol. 3. Fizmatgiz, Moscow, 1960. 62. Gerasimov, V. G., “Theoretical Foundation of Control of Heat Processes.” Gosenergoizdat, Moscow, 1949. 63. Germaidze, V. E. and Krasovskii, N. N., Stability under steady acting disturbances. Appl. Marh. Mech. 21(6), 769774 (1957). 64 Gihman, I. I. and Skorohod, A. V., “Stochastic Differential Equations,” Ergeb. Math. Grenzgeb. 72 (1972).
References
205
65. Ginzburg, P. E., Application of Liapunov direct method for investigation of oscillations in linear systems with delay. DiflerenciaPnye Uravnenija 7(10), 19031905 (1971). 66. Gonorovskii, I. S., To the problem of autooscillations in a high frequency generator with delayed feedback. Radiotehnika (Kharkov) 14(1), 2533 (1959). 67. Gorbunov, V. P. and Shichov, S. B., “Nonlinear Dynamics of Nuclear Reactors.” Atomizdat, Moscow, 1975. 68. Gorecki, H., “Analysis and synthesis of controlled systems with delay.” Mashinostroenie, Moscow, 1974. 69. Gorelik, G., To the theory of feedback with delay, J . Tech. Phys. 9(5), 450454 (1939). 70. Gorjachenko, V. D., “Methods of Stability Investigations of Nuclear Reactors.” Atomizdat, Moscow, 1977. 71. Gorjachenko, V. D. and Ivanov, B. N., Dynamics of interaction of two populations as objects with delay. In: “Dynamics of Biological Systems,” pp. 925. Gorjkii, 1978. 72. Gromova, P. S. (1) Investigation of stability of linear differentialdifference equations of neutral type in critical cases. Nonlinear Vibration Problem 4, 149158 (1973). (2) Stability of solutions of linear differentialdifference equations of nth order. Trudy Sem. Teor. Differencial. Uravnenii s Otklon. Argumentom Univ. Druzby. Narodov Patrisa Lumumby 9, 3139 (1975). 73. Gromova, P. S. and Markos, L. P., Method of Liapunov vector functions for retarded systems. Trudy Sem. Teor. Differencial. Uravnenii s Otklon. Argumentom Univ. Druzby. Nardov Patrisa Lumumby, 1422 (1979). 74. Grossman, S. E. and Yorke, J. A,, Asymptotic behavior and exponential stability criteria for differential delay equations. J . Diflerential Equations 12(2), 236255 (1972). 75. Gumowski, I., Sensitivity of certain dynamic systems with respect to a small delay. AutomaticaJ. IFAC 10(6), 659674 (1974). 76. Gurtin, E. M. (1) Thermodynamics and the energy criterion for stability. Arch. Rational Mech. Anal. 52(2), 93103 (1973). (2) Some questions and open problems in continuum mechanics and population dynamics. J . Differential Equations 48(2), 293312 (1983). 77. Hadeler, K. P., “Delay Equations in Biology.” Lecture Notes in Math. 730, 136156 (1979). 78. Hahn, W. (1) Uber DifferentialDifferenzengleichungenmit anomalen Losungen. Math. Ann. 133, (1957). (2) “Stability of Motion,” SpringerVerlag, New York and Berlin, 1967. 79. Halanay, A. (1) Perron’s condition in the theory of general systems with aftereffect. Mathematica (Cluj) 2(2), 257267 (1960). (2) Stability theory of linear periodic systems with delay. Rev. Math. Pures Appl. 6(4), 633653 (1961). (3) Systtmes a retard. Risultats et probltmes. Tagungsber. ICNO111 (Berlin) (1965). (4) “Differential Equations: Stability, Oscillations, Time Lags.” Academic Press, New York, 1966. 80. Halanay, A. and Yorke, J. A., Some new results and problems in the theory of differentialdelay equations. SIAM Rev. 13, 5580 (1971). 81. Hale, J. K. (1) A stability theorem for functionaldifferential equations. Proc. Nat. Acad. Sci. USA, 50(5), 942946 (1963). (2) Asymptotic behavior of the solutions of differentialdifference equations. Trudy ICNOI (Kiev) 2,409426 (1963). (3) Sufficient conditions for stability and instability of autonomous functionaldifferential equations. J . Differential Equations 1, 452482 (1965). (4) “Theory of functional differential equations,” SpringerVerlag, New York and Berlin, 1977. 82. Hale, J. K. and Ize, A. F., On the uniform asymptotic stability of functional differential equations of the neutral type. Proc. Amer. Math. SOC.28(1), 100106 (1971). 83. Hale, J. K. and Meyer, K. R., A class of functional equations of neutral type. Mem. Amer. Math. SOC.(76), 165 (1967). 84. Has’minskii, R. Z., “Stochastic stability of hfferential equations.” Noordhofof, Sijthoff, 1980.
206
References
85. Hausrath, A,, Stability in the critical case of purely imaginery roots for neutral functional differential equations. J . Diferential Equations (13), 329357 (1973). 86. Hino, Y., On stability of some functional differential equations. Funkcial. Ekuac. 14,4760 (1971). 87. Husainov, D. Ja. and Sharkovskii, A. N., Stability of solutions of differential equations with delay. In “Functional and DifferentialDifferenceEquations,” pp. 141147. Inst. of Math. of Ukr. Acad. Sci., Kiev, 1974. 88. Hutchinson,. G. E., Circular causal systems in ecology. Ann. N . Y Acad. Sci. 50, 221246 (1948). 89. Infante, E. F. and Castelan, W. B., A Liapunov functional for a matrix differencedifferential equation. J . Diferential Equations, 29(3), 439451 (1978). 90. Infante, E. F. and Slemrod, M., Asymptotic stability criteria for linear systems of differential difference equations of neutral type and their discrete analogues. J . Math. Anal. Appf. (38), 399415 (1972). 91. It8, K. and Nisio, M., On stationary solutions of a stochastic differential equation. J . Math. Kyoto Uniu. 4, 175 (1964). 92. Ize, A. F. and dos Reis, J. G., Stability of perturbed neutral functional differential equations. Nonlinear Anal. 2(5), 563571 (1978). 93. Kabakov, I. P., Concerning the control process for the steam process for the steam pressure. InPener. Sbornik (2), 2760, 6176 (1946). 94. Kabal’nov, Ju. S., Homjakov, I. M. and Il’jasov, B. G., To the method of control of plants with pure delay. Izu. Vusou. Electromechanica (S), 863866 (1978). 95. Kamenskii, G. A. (1) To the general theory ofequations with deviating argument. Dokl. A N USSR 120(4), 697700 (1958) (2) Existence, uniqueness and continuous dependence from initial data of solutions of systems of differential equations with deviating argument of neutral type. Matem. Sborn. 5544) (1961). 96. Kantorovich, L. V. and Akilov, G. P., “Functional Analysis.” Nauka, Moscow, 1977. 97. Kaplan, J. L. and Yorke, J. A., On the stability of a periodic solution of a differential delay equation. SIAM J . Math. Anal. 6(2), 268282 (1975). 98. Kato, J. (1) On LiapunovRazumikhin type theorems for functional differential equations. Funkcial Ekuac. 16(3), 225239 (1973). (2) Stability problem in functional differential equations with infinite delay. Funkcial. Ekuac. 21, 6380 (1978). 99. Katz, I. Ia., To the stability in the first approximation of systems with stochastic delay. Prikl. M a t . Meh. 31(3), 447452 (1967). 100. Kir’janen, A. I., Stability of solutions of differential equations with deviating argument. Lecture Notes. Leningrad University, 1983. 101. Klimushev, A. I., About asymptotic stability of retarded systems, containing derivatives with small parameters. Prikl. Mat. Meh. 26(1), 5261 (1962). 102. Kobrinskii, N. E. and Kus’min, V. I., “Accuracy of economicmathematical models.” Finansy and Statistica, Moscow, 1981. 103. Kolesov, Ju. S., To the stability of linear differentialdifferenceequations of neutral type. Sibirian Math. J . 20(2), 317321 (1979). 104. Kolesov, Ju. S. and Shvitra, D. I., “Autooscillation in systems with delay.” Mokslas, Vilnuis, 1979. 105. Kolmanovskii, V. B. ( I ) Stationary solutions of equations with delay. Problems Inform. Transmission 3(1), 5057 (1967). (2) Application of the Liapunov method to linear systems with lag. J . Appl. Math. Mech. 31,976980 (1967). (3) Boundedness of solutions of integrodifferential equations with deviating argument. Trudy Sem. Teor. Diferencial. Uraunenii s Otklon Argumentom Uniu. Druzby Narodou Patrisa Lumumby 6,207212 (1968). (4) Ergodic properties of stochastic differential equation solutions. Theory Probab. Appl. 14(1), 142148
References
207
(1969). (5) On the stability of stochastic systems with delay. Problems Inform. Transmission 5(4), 5967 (1969). (6) On the stability of neutraltype linear equations. Differencial’nye Uravnenija 6(7), 12351246 (1970). (7) The stability of nonlinear systems with lag. Math. Notes 7, 446450 (1970). (8) Solutions of constant signs of neutraltype equations. J . Differential Equations 7(6), 11161119 (1971). (9) Representation of solutions of neutral type equations. Ukrain. Mat. Z . 24(2), 171178 (1972). (10) Filtration of stochastic processes with delay. Automat. Remote Control (l), 4249 (1974). (11) On the stability of some stochastic differential equations with retarded argument. Theory Probab. Math. Statist. (2), 110120 (1974). (12) On the equalities for second moments of solutions of stochastic differential equations with delay. Ukrainian Math. J . 27(1), 9497 (1975). (13) Optimal control of stochastic systems with aftereffect. (A survey). Trudy Fourth AllUnion Conf: Theory Appl. Differ. Eq. Deuiat. Argument, pp. 177185. Nauka Dumka, Kiev, 1977. 106. Kolmanovskii, V. B. and Has’minskii, R. Z., Stability of linear systems with retardation. Izu. Vyss. Ucebn. Zaved. Matematika 4(53), 5865 (1966). 107. Kolmanovskii, V. B. and Maizenberg, T. L., Optimal estimation of system states and problems of control of systems with delay. Prikl. Math. Meh. 41(3), 446456 (1977). 108. Kolmanovskii, V. B. and Nosov, V. R. (1) On the stability of nonlinear oscillative systems described by neutraltype equations. Proc. 5th Internat. Con& Nonlinear Osc. 2, Kiev (1969). Issue Inst. Math. Vkr. Acad. Sci. 228232 (1970). (2) On the stability of firstorder neutral type equations. Prikl. Math. Meh. 34, (4) 587594. (3) Stability of stationary systems with aftereffect, Automat. i Telemeh. (l), 918 (1979. (4) Stability of systems with deviating argument of neutral type. Prikl. Math. Meh. 43(2), 209218 (1979. ( 5 ) “Stability and periodic modes of control systems with aftereffect.” Nauka, Moscow, 1981. (6) Stability of neutraltype functional differential equations (a survey) Nonlinear Anal. Theory Methods Appl. 6(9), 873910 (1982). (7) Asymptotic behavior of neutral type equations. In “Oscillation and Solutions Stability of FunctionalDifferential Equations,” pp. 4453. Institute of Math. AN USSR, Kiev, 1982. (8) On instability of systems with aftereffects. Automat. Remote Control 44, (l), part 1,2432 (1983) (9) “Stochastic stability and control.” (Lecture notes). Institute of Electronic Machines, Moscow, 1983. (10) Retarded systems of neutral type. (a survey) Automat. Remote Control 45(1), 536 (1984). 109. Kolmanovskii, V. B. and Shaichet, L. E., Approximate feedback control of quasilinear stochastic retarded systems. Prikl. Math. Meh. 42(6), 978988 (1978). 110. Kostitzin, V. A,, “Symbiose, Parasitisme et Evolution (Etude Mathematique),” Hermann, Paris, 1934. 111. Kozjakin, V. S. and Krasnoselskii, M. A,, To the influence of small delays on dynamic of nonlinear systems. Automat. Remote Control. (I), 58 (1979). 112. Krasnoselskii, M. A., “Shift operator for differential equations.” Nauka, Moscow, 1966. 113. Krasnoselskii, M. A,, Burd, V. Sh. and Kolesov, Ju. S., “Nonlinear AlmostPeriodic Oscillations.” Nauka, Moscow, 1970. 114. Krasovskii, N. N. (1) Inversion of theorems of Liapunov direct method and stability in first approximation. Prikl. Math. Meh. 20(2), 255265 (1956). (2) On the Application of direct Liapunov method to equations with time lag. Prikt. Math. Meh. 20(3), 315327 (1956). (3) On the asymptotic stability of systems with aftereffect. Prikl. Math. Meh. 20(3), 513518 (1956). (4) Stability under large initial disturbances. Prikl. Math. Meh. 21(3), 309319 (1957). ( 5 ) “Stability of motion: Applications of Lyapunov’s Second Method to Differential Systems and Equations with Delay.” Stanford Univ. Press, Stanford, California, 1963. (6) Analytical design of optimal controller in the system with delay. Prikl. Math. Meh. 28(1), 3951 (1964). 115. Krupnova, N. I. and Shimanov, S. N., Test of stability of linear systems with periodic coefficients and timelag. Prikl. Math. Meh. 36(3), 533536 (1972).
2 08
References
116. Kulesko, N. A., About completeness of Floquet solutions of neutral type equations. Mat. Zametki 3(3), 297306 (1968). 117. Kurbatov, V. G., Stability of functionaldifferential equations. Diferen. Uraun. (6), 963972 (1981). 118. Kwong, R. H., A stability theory for the linearquadraticGaussian problem for systems with delays in the state, control and observations. SIAM J . Control Optim. 18(1), 4975 (1980). 119. Lakshmikantham, V. and Leela, S. (1) Differential and Integral Inequalities. Academic Press, New York, 1969. (2) Global Results and Stability of Motion. Univ. Rhode Island, Tech. Rep. No. 8, 1970. (3) A technique in stability theory of delaydifferential equations. Nonlinear Anal. 3(3), 317323 (1979). 120. Lakshmikantham, V. and Rao M. R. H., IntegroDifferential Equations and Extension of Liapunov’s Method. Univ. Rhode Island, Tech. Rep. No. 7, 1969. 121. Lambourne, N. C., Controlsurface buzz Rep. Memorfnda (3364), 117 (1962). 122. Landau, I. D., A survey of model reference adaptive techniques: theory and applications. AutomaticaJ. IFAC 10(4), 353379 (1974). 123. Lavrentjev, M. A. and Shabat, B. V., “Methods of Theory of Functions of Complex Variabele.” Fizmatgiz, Moscow, 1958. 124. Lekus, V. F. and Rovinskii, V. E., “Estimate of Stability of Systems with Delay.” Energoatomizdat, Leningrad, 1982. 125. Leontjev, A. F., Differentialdifferenceequations. Mat. Sb. 24(3), 347374 (1949). 126. Lewis, R. M. and Anderson, B. D. O., Necessary and sufficient conditions for delayindependent stability of linear autonomous system. IEEE Trans. Automat. Control AC25(4), 735739 (1980). 127. Liapunov, A. M., “Stability of Motion.” Gostechizdat, Moscow, 1950. 128. Liberman, L. X., Stability of differentialoperator equations with delay under disturbances bounded in the mean. Siberian Math. J . 4(1), 138144 (1963). 129. Lidskii, E. A,, Stability of systems with stochastic delay. Diferenciarnye Uraunenija l(l), 96101 (1965). 130. Lihtarnikov, A. L. and Jakubovich, V. A,, Absolute stability of nonlinear systems. Appendix to Russian translation of “Absolute Stability of Retarded Control Systems” pp 287357. Rasvan Vb. 131. Loginov, G. V. and Oreshina, S. T., Analysis of retarded systems with predictor by methods of root trajectories. Izu. Vyss. Ucebn. Zaued. Electromehanika (1 I), 11581163 (1978). 132. Lubfer. D. E. and Oglesby, W., Applying deadtime compensation for linear predictor process control. J . Systems Anal. ISA Journal S(11). 5357 (1961). 133. Lunt, S. T. and Tolman, G., Some effects of delays in control systems. Meas. Control 13(5), 184185 (1980). 134. Lutzkiv, N. M. (1) To the synthesis of control systems with compensation of delay. Izu. Vyss. Ucebn. Zaued. Electromehanika (8), 889901 (1972). (2) Systems with compensation of delay and divisor. Izu. Vyss. Ucebn. Zaued. Electromehanika (9), 10031007 (1976). 135. MacDonald, N. “Time Lags in Biological Models.” I n Lecture Notes Biomath. (1978). 136. Mahin, V. A,, Prisnjakov, V. F. and Belik, N. P., “Dynamics of liquid reactive engine.” Mashinostroenie, Moscow, 1969. 137. Maizenberg, T. L., Dirichlet problem for integrodifferential equations. Izu. AN USSR, Ser. Math. 33(3), 570590 (1969). 138. Malkin, I. G., “Stability of Motion.” Nauka, Moscow, 1966. 139. Marchuk, G. I. (1) Some mathematical models in immunology. Proc. 8th IFIP Con5 Opt. Tech. Heidelberg, 41 62 (1978). (2) “Mathematical Models in Immunology and Their Interpretation.” I n Lecture Notes in Control and Information Sci. 18, 114129 (1979). (3) “Mathematical Models in Immunology.” Nauka, Moscow, 1980.
References
209
140. Martynyuk, D. I., “Lectures on the Theory of Stability of Solutions of Systems with Retardations.” Inst. Mat. Acad. Nauk Uk. SSR, Kiev, 1971. 141. “Mathematical Methodes in clinical practice.” Nauka, Novosibirsk, 1978. 142. Matrosov, V. M., About stability of motion. Appl. Math. Mech. 26(6), 9921cO2 (1962). 143. Mazurov, V. M., Malov, D. I. and Salomykov, V. I. Systems of control of PH value in absorbtional column with recycle. Himich. Promysh. (4), 6365 (1974). 144. Melkumjan, D. O., “Analysis of System by the Method of Logarithmic Derivative.” Energoatomizdat, Moscow, 1981. 145. Melvin, W. R., Liapunov’s direct method applied to neutral functional differential equations. J . Math. Anal. Appl. 49(1), 4758 (1975). 146. Michelkevich, V. N. and Chabanov, Ju. A,, Identification and synthesis of system of control of process of infeed grinding, Electromashinostroenie i Electrooborudouanie, (Kiev).31, 4242 (1980). 147. Migulin, V. V., Medvedev, V. I., Mustel’, E. R. and Parygin, V. M., “Foundations of Oscillation Theory, Moscow, Nauka, 1978. 148. Minkin, S. I. and Skljarov, Ju. S., Analysis of transient in long line of direct current by the methods of theory of differential difference equations. Izo. Vyss. Ucebn. Zaoed. Elektromehanika (7), 687694 (1975). 149. Minorsky, N. (1) Control Problems. J . Franklin Inst. 232(6), 519551 (1941). (2) Selfexcited oscillations in dynamical systems possessing retarded actions. J . Appl. Mech. (9), 6571 (1942). 150. Misnik, A. F., Liapunov direct method for neutral equations. Trudy Sem. Teor. Dijierencial. Uraonenii s Otklon. Argumentom Uniu. Druzby Narodou 6, Patrisa Lumumby 78108 (1968). 151. Misnik, A. F. and Nosov, V. R., About stability of differencedifferential equations of neutral type. Sbornik Nauchnyh Rabot Aspirantoo Uniu.Lumumba 1,4355 (1968). 152. Mitropolskii, Ju. A. and Filchakov, P. F., About solutions of nonlinear differential equations with deviating argument with the aid of series. Dokl. AN USSR 212(5), 10591062 (1973). 153. Mitropolskii, Ju. A. and Martynuk, D. I., “Periodic and QuasiPeriodic Oscillations in Systems with Delay.” Vysha Shkola, Kiev, 1979. 154. Mjasnikov, N. N., Theory of direct Vyshnegradskii control and influence of delay. Izu. AN USSR, O7”(9), 12171228 (1953). 155. Mladov, A. G., “Systems of Differential Equations and Liapunov Stability.” Vysshai Shkola, Moscow, 1966. 156. Myshkis, A. D. (1) General theory of differential equations with delay. Uspehi, Mat. Nauk 4(5), 99141 (1949). [Engl. Transl. AMS 55,l62 (1951), and Ser. 1,4,207267 (1962).] (2) “General Theorems of Theory of Ordinary Differential Equations in Nonclassical Cases.” 1 Summer Mathematical School, pp. 45116. Nauka Dumka, Kiev, 1964. (3) “Linear Differential Equations with Delaying Argument.” Nauka, Moscow, 1972. (4) Some problems in the theory of differential equations with deviating arguments. Amer. Math. SOC. Trans. 2(105), 237246 (1976). (5) On some problems of the theory of differential equations with deviating argument. Uspehi. Mat. Nauk. 32(2), 173202 (1977). 157. Myshkis, A. D. and El’sgol’tz, L. E., The status and problems of the theory of differential equations with deviating argument. Uspehi, Mat. Nauk. 22(2), 2157 (1967). 158. Myshkis, A. D., Shimanov, S. N. and El’sgol’tz L. E., Stability and oscillations of systems with timelag. Trudy of ICNOI, Kiev 2, 241267 (1963). 159. Nechaev, Ju. N. and Fedorov, R. M., “Theory of aircraft gasturbine engines.” Mashinostroenie, Moscow, 1977. 160. Neimark, Ju. I. (1) Dsubdivision and spaces of quasipolynomials. Prikl. Mat. Meh. 13(4), 349380 (1949). (2) “Dynamical Systems and Controlled Processes.” Nauka, Moscow, 1978.
210
References
161. Netushil, A. V., Plutes, V. S. and Vlasov, Ju. A,, To the problem of application of SAR with compensation of delays and varying of the object parameters. Izu. Vyss. Ucebn. Zaued. Electromehanika (8), 882891 (1976). 162. Nosov, V. R. (1) On a problem arising in the theory of optimal control with aftereffect. Prikl. Mat. Meh. 2, 399403 (1966). (2) Linear boundaryvalue problems with small lags. Differencial’nye Uraunenija 3, 1025 1028 (1967). (3) Comparison theorems for scalar retarded equations. Trudy Sem. Teor. Differencial. Uraunenii s Otklon. Argumentom Druzby Narodou Patrisa Lumumby 4, 247253 (1967). (4) Sur l’arternative de Fredholm pour les systemes lineaires generaux a coefficients periodiques et a argument deplace. C . R. Acad. Sci. 270(17), 1097 1100 (1970). (5) Solutions periodiques des systtmes differentielsfonctionnels quasilineaires. C.R. Acad. Sci. 270(18), 11701 173 (1970). (6) Periodic solutions of linear equations of general form with deviating arguments. Differencial’nye Uraunenya 7, 639650 (1971). (7) On existence of periodic solutions of linear systems of general form with distributed deviations. Differencial‘nye Uraunenija 7,21682175 (1971). (8) On comparison theorems. Trudy Sem. Teor. Differencial. Uraunenii s Otklon. Argumentum Uniu. Druzby Narodou Patrisa Lumumby 8, 143153 (1972). (9) Periodic solutions of quasilinear functional differential equations. 120. Vyss. Ucebn. Zaved. Matematica 5,5562 (1973). (10) Matrosov criterion of stability for retarded functional differential equations. Diflerencial’nye Uraunenija 7, 12021207 (1977). (1 1) Investigations of stability and quality of transient process in model reference adaptive systems. VII Internationale Konferenz uber Nichtlineare Schwingungen, Band 11, 2, 173182 (1977). (12) Identification of linear and some
163.
164. 165. 166. 167.
168. 169.
nonlinear systems with the aid of adaptive models. I n “Optimization of Dynamic Systems,” 110114. Minsk, 1978. (13) Perron’s theorem for autonomous and periodic systems of functional differential equations. Trudy Sem. Teor. Differenrial. Uraunenii s Otklon. Argumentom Univ. Druzby Narodov Patrisa Lumumby. 4  5 1 (1979). Nosov, V. R. and Prokopov, B. I. (1) Estimation of transient process in model reference adaptive systems. Izu. AN U S S R , Techn. Kybernetika. 4,186191 (1976). (2) On asymptotic stability of adaptive scalar retarded systems. Differencial‘nye Uraunenija 13, 15281531 (1977). (3) Global asymptotic stability of model reference adaptive systems. Prikl. Mat. Meh. 41, 850858 (1977). Nosov, V. R. and Furasov, V. D., Stability of discrete systems relative to given variables and convergence of some optimization algorithms. J . Vycis. Mat. Phys. 19, 316328 (1979). Nyquist, H., Regeneration theory. Bell Syst. Tech. J . 11(1), 126147 (1932). Osipov. IU. S., Stabilization of nonlinear control systems with delay in critical case of one zero root. Diflerencial’nye Uraunenija 1(7), 908922 (1965). Pak, V. E. and Prokopjev, V. P., About stability of systems with aftereffect in critical case of four zero roots. Math. Zapiski Ural. Uniu. 7(4), 7682 (1970). Parks, P. C. Liapunov redesign of model reference adaptive control systems. I E E E Trans. Automat. Control 11(3), 362367. Pazdera, J. S. and Pottinger, H. J., Linear system identification via Liapunov design techniques. (Joint Automat. Control Conference, 10th Univ. Colorado, New York, ASME,
795801, 1969. 170. Pinney, E., “Ordinary differencedifferentialequations,” Univ. of California Press, Berkeley, 1958. 171. Pisarenko, V. G . (1) “Problems of Relativistic Dynamics of Many Bodies and Nonlinear Theory of Field. Nauka Dumka, Kiev, 1974. (2) Equations with deviating argument in the
problem of many gravitated electricity charged bodies with delays of the forces of interaction. I n “Differential Equations with deviating argument,” pp. 255269. Nauka Dumka, Kiev, 1977. 172. Pjatnitzkii, E. S., Structural stability of singlecircuit systems with delay. Automat. Remote Control 23(7), 852862 (1962).
References
21 1
173. Pletnev, G. P., “Automatic control and protection of heatenergetic designs of electric stations.” Energia, Moscow, 1970. 174. Political and related models In “Modeles in applied mathematics,” Vol. 2 (W. F. Lucas, ed.) SpringerVerlag, Berlin and New York, 1983. 175. Pontryagin, L. S., On the zeros of some elementary transcendental functions. Izo. Akad. Nauk SSSR 6(3), 115134 (1942). 176. Popov, V. M. and Halanay, A., About stability of nonlinear controlled systems with delay. Automat. Remote Control, 23(7), 849851 (1962). 177. Postnikov, M. M., “Stable Polynomials.” Nauka, Moscow, 1982. 178. Prokopjev, V. P., About stability of systems with delay in critical case of zero roots. Izv. Vusov, Math. (I), 8894 (1967). 179. Rao, D. and Ramakrishna, Neutral functional differential equations and extensions of Liapunov’s method. Internat. J . Control 10(4), 369375 (1969). 180. Razumikhin, B. S. ( 1 ) On the stability of systems with a delay. Prikl. Mat. Meh. 20(4), 500512 (1956). (2) Application of Liapunov method to problems in the stability of systems with a delay. Automat. i Telemeh. (21), 740749 (1960). 181. Razvan, V., “Absolute Stability of Automatic Systems with Delay.” Nauka, Moscow, 1983. 182. Real, J., Stochastic partial differential equations with delay. Stochastics 8(2), 81102 (1 982). 183. Repin, Yu. M. (1) Stability conditions of systems of linear differential equations for any delays. Proc. Ural Univ. 23(2), 3441 (1960). (2) Quadratic Liapunov functionals for systems with delay. Prikl. M a t . Meh. 29, 564566 (1965). 184. Reswick, J. B., A DelayLine Controller. Trudy First IFAC Congress, Moscow, 1960. 185. Richaud, L., La regulation P.I.D. des processus retardes. 11. Sy st h es du second ordre. Automatisme 15(12), 626638 (1970). 186. Riesz, F., Sz. and Nagy, B., ‘‘LeGons d‘analyse fonctionnelle.” Akademiai Kiado, 1972. 187. Rjabov, Ju. A. ( 1 ) Some asymptotic properties of linear systems with a small delay. Trudy Sem. Teor. Differencial. Uravnenii s Otklon. Argumentom Univ. Druzby Narodov Patrisa Lumumby 3, 153164 (1965). (2) On approximation of solutions of nonlinear retarded differential equations. Trudy Sem. Teor. Differencial. Uravnenii s Otklon. Argumentom Uniu. Druzby Naradov Patrisa Lumumby 3, 165185 (1965). 188. Rohella, R. S. and Chatterjee, B., Effect of time delay on nonlinear systems. J. Inst. Electron. Telecomm. Eng. ( N e w Delhi) 25(9), 386388 (1979). 189. Rotach, V. Ja., “Dynamics of Industrial Automatic Control Systems. Energa, Moscow, 1973. 190. Rotach, V. Ja and Stafeichuk, B. G., Application of linear Smith’s predictor for control of systems with delay. Energetica (2) (1966). 191. Rouche, N. Abetz, P. and Laloy, M., “Liapunov’s Direct Method in Stability Theory.” MIR, Moscow, 1980. 192. Rouche, N. and Mawhin, J., “Equations Differentielles Ordinaires,” Vol. 2. Mason, Paris, 1973. 193. Rubanik, V. P., “Oscillations of Quasilinear Systems with Retardations.” Nauka, Moscow, 1969. 194. Rutman, M. A,, About bounded solutions of linear differential and differentialdifference equations. T r . Odess. Gidromet. Inst. (20), 37 (1959). 195. Sawano, K., Exponential asymptotic stability for functional differential equations with infinite retardations. TohBku Math. J . 31(3), 363382 (1979). 196. Schipanov, G. V., Theory and methods of design of automatic controllers. Automat. Remote Control (l), 4956 (1939). 197. Seifert, G., LiapunovRazumikhin conditions for stability and boundedness of functional differential equations of Volterra type. J . Dtfferential Equations 14(3), 424430 (1973).
212
References
198. Serebrjakova, I. V., Methods of solutions of differential equations with deviating argument in 18th and 19th centuries. Lumumby Trudy Sem. Teor. Diflerencial. Uravnenii s Otklon Argumentom Univ. Druzby Narodov Patrisa 10,4168 (1977). 199. Shaichet, L. E. ( 1 ) Investigation of stability of stochastic systems with delay by the method of Liapunov functionals. Problems of Information Transmission 11(4), 7076 (1975). (2) Stability in the first approximation of stochastic retarded systems. Appl. Math. Mech. 40(6), 11161121 (1976). (3) On the optimal control of one class of stochastic differential equations with partial derivatives. Mat. Zametki 31(6), 933936 (1982). 200. Sharkovskii, A. N. and Romanenko, E. Ju., “Asymptotical Behavior of Solutions of DifferentialDifference Equations,” pp. 171200. Institute of Math. of Acad. of Sci. of Ukr. SSR, Kiev, 1981. 201. Sheridan, T. B. and Ferrell, W. R., “ManMachine Systems: Information, Control and Decision Models of Human Performance.” MIT Press, Cambridge, Massachusetts, 1974. 202. Shevyakov, A. A. (ed.), “Control Theory of Power Plants of Flying Vehicles.” Mashinostroenie, Moscow, 1976. 203. Shigin, E. K. Classification of dynamical models of control objects of chemical technological processes. Automat. Remote Control (6), 145162 (1968). 204. Shil’man, S. V., “Method of Generating Functions in the Theory of Dynamical Systems.” Nauka, Moscow, 1978. 205. Shimanov, S. N. ( 1 ) Instability of motion of a system with time lag. Prikl. Mat. Meh. %(I), 5563 (1960). (2) On stability in the critical case of a zero root for delayed systems. Prikl. Mat. Mech. 24(3), 447457 (1960). (3) To the theory of retarded linear periodic differential equations. Prikl. Mat. Meh. 27(3), 450458 (1963). 206. Shimbell, A., Contributions to the mathematical biophysics of the central nervous system with special reference to learning. Bull. Math. Biophys. (12), 241275 (1950). 207. ShuiNee Chow and Hale, J., “Methods of Bifurcation Theory,” Vol. 251. SpringerVerlag, Berlin and New York, 1982. 208. Sievert, R. M., Zur theoretischmathematischen Behandlung des Problems des biologkchen Strahlenwirkung. Acta Radiologica 22, 237251 (1941). 209. Silkowski, R., A starshaped condition for stability of linear retarded functional differential equations. Proc. Roy. SOC. Edinburgh Sect. A 83, 189198 (1979). 210. Sinha, A. S. C., On stability of solutions of some third and fourth order delaydifferential equations. Information and Control 23(2), 165172 (1973). 211. Slemrod, M. and Infante, E. F., Asymptotic stability criteria for linear systems of differencedifferential equations of neutral type and their discrete analogues. J . Math. Anal. Appl. 38(2), 399415 (1972). 212. Slusarchuk, V. E., Strong absolutely asymptotically stable solutions of linear retarded differential equations in Banach spaces. Diflerenciarnye Uravnenija 15(9), 16141619 (1979). 213. Smirnov, A. I., “Aeroelastic Stability of Flying Vehicles,” Mashinostroenie, Moscow, 1980. 214. Smith, 0. J. M. ( 1 ) “Feedback Control Systems.” MacGrawHill, New York, 1958. (2) A Controller to overcome dead time. J S A Journal, q 2 ) (1959). 215. Smol’nikov, L. P. and Moskvin, V. M., Design of controllers with Smith predictor for plants with delay. Izv. Leningrad Electrotechn. Ins. 111, 7481 (1972). 216. Snow, W., Existence, uniqueness and stability for nonlinear differentialdifferenceequations in the neutral case. Preprint, New York University, New York, 1965. 217. Sokolov, A. A. Stability criterion of linear controlled systems with distributed parameters and its applications. Engin. Sbornik 2(2), 326 (1946). 1980. 218. Solodov, A. V. and Solodova, E. A,, “Systems with Variable Delay.” Nauka, MOSCOW, 219. Solodovnikov, V. A. (1) Use of operator method in studying of the speed control of
References
220. 221.
222. 223.
224. 225. 226.
227.
228. 229. 230.
231.
232.
233. 234. 235.
236.
237. 238.
213
hydroturbines. Automat. i Telemeh. (l), 520 (1941). (2) “Theory of automatic control,” Vol. 1. Mashinostrienie, Moscow, 1967. Solodovnikov, V. V., Filimonov, A. B. Design of controllers for plant with delay. I z u . AN USSR, Tech. Kybernet. (l), 168177 (1979). Starik. L. K., Coupled quasilinear oscillating systems with a source of energy involving retarded connections. Trudy Sern. Teor. Differencial. Uravnenii s Otklon. Argumentom. Univ. Druzby Narodov Patrisa Lumumby 3, 119132 (1965). Stech, H. W., The effect of time lags on the stability of the equilibrium state of a population growth equation. J . Math. Biol. 5(2), 115120 (1978). Stokes, A. (1) A Floquet theory for functional differential equations. Proc. Nut. Acad. Sci. U S A 48(8), 13301334 (1962). (2) Stability of functional differential equations with perturbed lags. J . Math. Anal. Appl. 47(3), 604619 (1974). Svirejev, Ju. P. and Logofet, D. O.,“Stability of Biological Populations.” Nauka, Moscow, 1978. The damping effect of time lag (art. editorial). Engineer 163(5) (1937). Trudy Sern. Teor. Differencial. Uravnenii s Otklon. Argumentom Univ. Druzby Narodov Patrisa Lumumby 1 (1962); 2 (1963); 3 (1965); 4 (1967); 5 (1967); 6 (1968); 7 (1969); 8 (1972); 9 (1975); 10 (1977); 11 (1979). Tychonov, A. N., Sur les equations fonctionelles de Volterra et leurs applications a certains probltmes de la physique mathtmatique. Bull. de rUniu. d’Etat Moscou. Sect. A (l), 125 (1938). Tyshkevich, V. A,, “Some Problems of Stability of Functional Differential Equations.” Nauka Dumka, Kiev, 1981. Tzaljuk, Z. 8. Volterra integral equations. Naucn.Tech. Informacija ( V I N I T I ) 15, 131198 (1977). Tzar’kov, E. F. (1) Asymptotic exponential meansquare stability of the trivial solution of stochastic functionaldifferential equations. Theory Probab. (4), 871875 (1976). (2) Exponential pstability of trivial solution of functionaldifferential equations. Theory Probab. Appl. (2), 445448 (1978). Tzar’kov, E. F. and Engel’son, L. E. (1) Statistical solutions of linear systems of functional differential equations. In “Topological spaces and mappings,” pp. 144 153. Riga University, 1981. (2) Liapunov functional for linear periodic differential equations with aftereffects. In “Topological spaces and mapping,” pp. 117136. Riga University, 1983. Tzypkin, Ja. Z. (1) Stability of system with delayed feedback. Automat. Remote Control 7(2), 107129 (1946). (2) The power of stability of delayed feedback systems Automat. Remote Control 8(3), 145155 (1947). (3) Stability of a class of controlled systems with distributed parameters. Automat. Remote Control 9(3), 176189 (1948). Valeev, K. G., Linear differential equations with sinusoidal coefficients and constant retardations. Proc. ICNOI 2, 1001 19 (1963). Vasiljeva, A. B. and Butusov, V. F., “Asymptotic Expansions of Solutions of Singular Perturbed Equations.” Nauka, Moscow, 1973. Vasiljeva, A. B. and Rodionov, A. M., Applications of the method of perturbations to retarded equations with small delays. Trudy Sern. Teor. Digerencial. Uraonenni s Otklon. Argumentom Univ. Druzby Narodov Patrisa Lumumby 1,2027( 1962), Velarde, M. C., Dissipative structures and oscillation in reactiondiffusion models with and without timedelay. In “Stability of Thermodynamics Systems,” (G. Lebon, ed.), SpringerVerlag, Berlin and New York, 1982. Venkatesh, Y. V., “Energy Methods in TimeVarying System Stability and Instability Analyses.” SpringerVerlag, Berlin and New York, 1977. Volkov, V. Ja. and Kuprijanov, N. S., Stability of linear systems with many delays. TechnicGeskaya Kibernetika (9,170175 (1968).
214
References
239. Volterra, V. ( I ) Sulk equazioni integrodifferenziali della teorie dell’elasticita. Atti Accad. Lincei (18), 295 (1909). (2) Sur la theorie mathematique des phenomenes hereditaires. J . Math. Pures Appl. (7), 249298 (1928). (3) “Theorie mathematique de la Iutte pour la vie.” GauthierVillars, Paris, 1931. 240. Vorobjev, Ju. V., “Investigations of Control of Steam Turbine.” Gosenergoizdat, Moscow, 1950. 241. Voronov, A. A. ( I ) “Stability, Controllability and Observability.” Nauka, Moscow, 1979. (2) “Foundation of the Theory of Automatic Regulation.” Energoizdat, Moscow, 1981. 242. Vostrikov, A. S. and Utkin, V. I., Franzusova, G. A., Systems with derivative of phasevector in control. Automat. Remote Control (3), 2225 (1982). 243. Voznesenskii, I. N., T o choice of control of powerandheat supply turbine. Soviet Energooborudovanie (Moscow) (1934). 244. Walker, J. A., “Dynamical Systems and Evolution Equations: Theory and Applications.” Plenum, New York, 1980. 245. Wangersky, P. J. and Cunningham, W. J., Time lag in preypredation population models. Ecology 38, 136139 (1957). 246. WorzBusekros, A,, Global stability in ecological systems with continuous time delay. s r A M J . ~ p p i Math. . 5s(i),i23134 (1978). 247. Wright, E. M. (1) Linear differential difference equations. Proc. Cambridge Philos. SOC. 44, 179185 (1948). (2) The linear differencedifferentialequation with asymptotically constant coefficients. Amer. J . Math. 70(2), 221238 (1948). (3) The stability of solutions of nonlinear differencedifferential equations. Proc. Roy. SOC. Edinburgh, Sect. A 63(1), 1826 (1950). 248. Yorke, J. ( I ) Some extensions of Liapunov’s second method. SIAM Difl Integ. Equat. Pa., 206207 (1969). (2) Asymptotic stability for one dimensional differentialdelay equations. J . Diflerential Equations 7(1), 189202 (1970). 249. Yoshizawa, T., “Stability Theory and the Existence of Periodic Solutions and Almost Periodic Solutions.” SpringerVerlag, Berlin and New York, 1975. 250. Zaharov, A. A., Kolesov, Ju. S., Spokojnov, A. N. and Fedotov, N. B. Theoretical explanation of ten years oscillation cycles of quantities of animals in Canada and Iakoutia. I n “Studies in Stability and Oscillations,” pp. 82131. Iaroslavl, 1982. 251. Zhadanov, V. I. On the onedimensional symmetric twobody problem of classical electrodynamics.Internat. J . Theoret. Phys. 15, 157167 (1976). 252. Zubov, V. I. (1) “Mathematical Methods of Investigation of Controlled Systems.” Sudpromgiz, 1959. (2) Stochastic behavior of system with finite number of phase states and aftereffect. Diflerencial ’nye Uraunenija 15(3), 387391 (1979). (3) “Analytical Dynamics of Systems of Bodies.” Leningrad Univ., Leningrad, 1983. 253. Zverkin, A. M. (1) Dependence of the stability of solutions of linear differential equations with lag upon the choice of the initial moment. Vestnik Moscou Uniu.Ser. I . Mat. Meh. Astr. Him. 5,1520 (1959). (2) Theorems of existence and uniqueness for equations with deviating argument in critical case. Trudy Sem. Teor. Dlfferencial. Uraunenii s. Otklon. Argumentom Uniu.Druzby Naradou Patrisa Lumumby I, 3746 (1962). (3) Differentialdifference equations with periodic coefficients.Appendix to Russian translation of “Differential Difference Equations” (Bellman and Cooke). MIR, Moscow, 1967.(4) Expansions of solutions of differential difference equations in series. Trudy Sem. Teor. Diflerencial. Urauenii s Otklon. Argumentom Uniu.Druzby Narodou Patrisa Lumumby 4, 350 (1967). (5) Differential equations with deviating argument. Trudy 5 Letney Mat. Shkoly, 307399, Nauka Dumka, Kiev, ( 1968).
Index
A
Aeroautoelasticity, 120 Antenna, 67
random, 170 small, 27 unbounded, 19,21,74 Diagram, Vyshnegradskii, 55, 58, 117, 118 Dynamics, relativistic, 13
B Boundedness, 159 Buss, 124
E
C Condition Perron, 159 selfexcitation, 150 splicing, 22 Controller adaptive, 43, 96 P, 35, 37, 38, 81 PD, 37, 118, 132 PID, 37, 118 Reswick, 4 1 Smith, 42 Criterion frequency, 60 integral, 63, 70 Matrosov, 77 Michailov, 60, 1 19, 123 Nyquist, 6 1, 62, 1 15 Popov  Halanay, 107 Tzypkin, 65
Equation differential difference, 19, 22, 34, 53 for moments, 166 linear, 25 autonomous, 3 1,48, 1 13, 196 periodic, 108, 160 scalar, 83, 134, 159 secondorder, 85, 180 Estimate, Q qriori, 24, 26
F FDE, 36 Flutter, 125 Formula, It6, 166, 169, 172, 177, 180 Function almostperiodic, 129 characteristic, 31, 33, 50, 69, 114 transfer, 1, 35, 37, 69 unsteady unit step, 12 1 Functional degenerated, 132, 140, 152 Liapunov, 126, 156, 155 Krasovskii, 72, 80, 89
D Delay adjustment, 100 bounded, 19,72, 130, 148
H Hodograph, Michailov, 6 1, 62, 1 19, 120
215
Index 1
Ideal predictor, 36 Inequality Chebyshev, 189 functional, 127, 129, 152, 194 Instability, 42, 45, 142, 154
L Lag informational, 4 technological, 1 transportation, 3 Laplace transform, 49, 114
M Method D subdivision, 55, 1 17 Liapunov direct, 7 I , 103, 126, 167, 180, 187 steps, 19,48, 111, 154, 161, 167 Model cutting, 8 immune response, 15 infeed grinding, 8 linear stochastic aeroautoelastic, 123 population dynamics, 13 predatorprey, 15 virus disease, 90 Multiplier, Floquet, 108, 161
N NFDE, 10, 19, 21 linear, 27 autonomous, 33, 1 13 with bounded delay, 130, 148
0 Oscillator. 9
P Problem, initial value, 21, 23, 164 Process Markov, 166, 194
transient, 38 Wiener, 164
Q Quasipolynomial, 32, 53
R Reactor, 1, 8 I , 132 nuclear, 8, 87 Response, frequency, 61 RFDE, 19 linear, 25, 48 with bounded delay, 19, 72 with unbounded delay, 19, 2 1, 74
S
Sector, 145, 151 SFDE, 123, 164 SNFDE, 193 Solution ergodic, 190 fbounded, 128, 130 Floquet, 110, 161 periodic, 189 representation, 29, 33 stationary, 187 Space C[h,O], 19 C[m,O], I8,2 1 CB[m, 01, 19, 21, 165 L[h,OJ, 22 L,[~,O],24 SRFDE, 6, 164 Stability, 44, 113, 164 absolute, 106 domain, 117, 124, 180, 199 exponential, 79, 170, 188, 192 global, 14, 128, 132, 136, 140 in the first approximation, 75, 191 invariant set, 152 Liapunov, 44 linear autonomous NF'DE, 1 13 linear autonomous RFDE, 48 under steady acting disturbances, 75, 79, 99, 101
Index Stabilization, ship, 4 Stable, 44, 45, 73, 127, 157 asymptotically, 46, 5 1, 73, 115, 128 uniformly, 47, 73, 153 f; 127, 129, 130, 152 asymptotically, 127, 129, 130, 152 L,, 49, 50, 114 mean square, 168 asymptotically, 168, 171, 173, 176, 194, 196 P, 168 asymptotically, 168, 169, 193 with respect to probability, 168, 170, 193 System central nervous, 17, 143 distributed selfoscillatory, 150 feedback, 35 singleloop, 35, 38 twoloop, 39 manmachine, 5 with lossless transmission line, 10
217 T Theorem Barbashin Krasovskii, 76 Chebotarev, 55, 117 Chetaev, 142, 146, 155 Esclangon, 159 Krasovskii, 72 Pontriagin, 54, 117 Razumichin, 72 Zubov, 33 Time lag, 2, 3 Turbojet, 6 V
Viscoelasticity, 103
U Uniqueness, 21, 23, 166, 194